From be0ae93c50bfa3e588111585025278daa8cb0694 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Tue, 13 Mar 2018 13:08:12 +0000 Subject: COMPMID-1005: Update Depthwise Convolution form RSH Change-Id: I3033ddb8de183661010d6c71a83f71132037b139 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/124338 Tested-by: Jenkins Reviewed-by: Pablo Tello --- .../kernels/NEDepthwiseConvolutionLayer3x3Kernel.h | 20 +- .../NEON/kernels/convolution/common/profiler.hpp | 31 +- .../kernels/convolution/common/tensor_utils.hpp | 9 +- .../core/NEON/kernels/convolution/common/utils.hpp | 1 - .../kernels/convolution/depthwise/depthwise.hpp | 176 +- .../kernels/convolution/depthwise/impl_base.hpp | 297 +- .../convolution/depthwise/impl_fp32_fp32.hpp | 69 +- .../NEDepthwiseConvolutionLayer3x3Kernel.cpp | 66 +- src/core/NEON/kernels/convolution/common/utils.cpp | 14 +- .../depthwise/depthwise_2x2_3x3_1x1_fp32_fp32.cpp | 937 ++-- .../depthwise/depthwise_2x2_3x3_2x2_fp32_fp32.cpp | 1121 +---- .../depthwise/depthwise_3x3_3x3_1x1_fp32_fp32.cpp | 2058 ++++---- .../depthwise/depthwise_3x3_3x3_2x2_fp32_fp32.cpp | 3994 +++------------ .../depthwise/depthwise_4x4_3x3_1x1_fp32_fp32.cpp | 4115 ++++++--------- .../depthwise/depthwise_4x4_3x3_2x2_fp32_fp32.cpp | 5301 +------------------- tests/datasets/DepthwiseConvolutionLayerDataset.h | 1 + 16 files changed, 4164 insertions(+), 14046 deletions(-) diff --git a/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.h b/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.h index 1367f378f7..5871cc5dcb 100644 --- a/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.h +++ b/arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.h @@ -82,8 +82,24 @@ private: void configure_optimized(); void run_generic(const Window &window, const ThreadInfo &info); void run_optimized(const Window &window, const ThreadInfo &info); - std::unique_ptr create_convolver_object(TensorShape shape, PadStrideInfo conv_info, - const uint8_t *w_ptr, uint8_t *in_ptr, uint8_t *out_ptr); + /** Creates an optimized backend convolver object + * + * @note Convolver of strides 1,2 and convolution size of 3 is currently supported + * + * @param[in] conv_info Padding and stride information to use for the convolution + * @param[in] w Weights tensor + * @param[in] in Input tensor + * @param[in] out Output tensor + * @param[in] setup_strides (Optional) Boolean to enable setting the strides of the tensors + * in the convolver in case of padding. Defaults to false + * + * @return A convolver object or nullptr if the configuration is not supported + */ + std::unique_ptr create_convolver_object(PadStrideInfo conv_info, + const ITensor *w, + const ITensor *in, + ITensor *out, + bool setup_strides = false); private: BorderSize _border_size; diff --git a/arm_compute/core/NEON/kernels/convolution/common/profiler.hpp b/arm_compute/core/NEON/kernels/convolution/common/profiler.hpp index 01fafa9604..c6897e3771 100644 --- a/arm_compute/core/NEON/kernels/convolution/common/profiler.hpp +++ b/arm_compute/core/NEON/kernels/convolution/common/profiler.hpp @@ -24,17 +24,21 @@ #pragma once +#include +#include +#include +#include + +#ifdef CYCLE_PROFILING #include #include -#include -#include #include #include #include #include #include "perf.h" -#include +#endif // CYCLE_PROFILING #ifdef CYCLE_PROFILING class EventIDContainer @@ -295,32 +299,43 @@ public: #endif // CYCLE_PROFILING template - void operator() (const char * event, - T func, - long int bytes_read = 0, - long int ops = 0, - long int bytes_written = 0) { + double operator() (const char * event, + T func, + long int bytes_read = 0, + long int ops = 0, + long int bytes_written = 0) { #ifdef CYCLE_PROFILING if (currentevent==maxevents) { + const std::chrono::high_resolution_clock::time_point start = std::chrono::high_resolution_clock::now(); func(); + const std::chrono::high_resolution_clock::time_point end = std::chrono::high_resolution_clock::now(); + return std::chrono::duration_cast(end - start).count(); } else { const auto countfd = thread_counter_fds.get_counter_fd(); + const std::chrono::high_resolution_clock::time_point start = std::chrono::high_resolution_clock::now(); start_counter(countfd); func(); long long cycs = stop_counter(countfd); + const std::chrono::high_resolution_clock::time_point end = std::chrono::high_resolution_clock::now(); + return std::chrono::duration_cast(end - start).count(); // Store the profiling data std::lock_guard lock_events(event_lock); events[currentevent++] = { get_event_id(event), bytes_read, ops, bytes_written, cycs }; + + return duration_us; } #else (void) event; (void) bytes_read; (void) ops; (void) bytes_written; + const std::chrono::high_resolution_clock::time_point start = std::chrono::high_resolution_clock::now(); func(); + const std::chrono::high_resolution_clock::time_point end = std::chrono::high_resolution_clock::now(); + return std::chrono::duration_cast(end - start).count(); #endif // CYCLE_PROFILING } }; diff --git a/arm_compute/core/NEON/kernels/convolution/common/tensor_utils.hpp b/arm_compute/core/NEON/kernels/convolution/common/tensor_utils.hpp index 68a5c6a178..0c234431b1 100644 --- a/arm_compute/core/NEON/kernels/convolution/common/tensor_utils.hpp +++ b/arm_compute/core/NEON/kernels/convolution/common/tensor_utils.hpp @@ -30,9 +30,12 @@ void PrintTensor(const Tensor4D& tensor); void PrintWeights(const Tensor4D& weights); // Test the equivalence of two tensors -bool CmpTensors(const Tensor4D& a, - const Tensor4D& b, - const float max_delta=0.0f); +// Counts the instances that |a - b|/|a| > max_err +bool CmpTensors( + const Tensor4D& a, + const Tensor4D& b, + const float max_err=0.0f +); // Fill the tensor with a test pattern void TestPattern(Tensor4D& tensor); diff --git a/arm_compute/core/NEON/kernels/convolution/common/utils.hpp b/arm_compute/core/NEON/kernels/convolution/common/utils.hpp index a22809fb58..5f42719119 100644 --- a/arm_compute/core/NEON/kernels/convolution/common/utils.hpp +++ b/arm_compute/core/NEON/kernels/convolution/common/utils.hpp @@ -24,7 +24,6 @@ #pragma once -double TimeInUs(void); void PrintMatrix(const float *const m, const int M, const int N, const int row_stride); inline int iceildiv(const int a, const int b) diff --git a/arm_compute/core/NEON/kernels/convolution/depthwise/depthwise.hpp b/arm_compute/core/NEON/kernels/convolution/depthwise/depthwise.hpp index 80b0614015..4ca68116db 100644 --- a/arm_compute/core/NEON/kernels/convolution/depthwise/depthwise.hpp +++ b/arm_compute/core/NEON/kernels/convolution/depthwise/depthwise.hpp @@ -29,7 +29,7 @@ namespace depthwise class IDepthwiseConvolution { -public: + public: virtual ~IDepthwiseConvolution() = default; virtual int output_size(const int dim_size, const bool padding_same) const = 0; virtual unsigned int get_window(void) const = 0; @@ -59,8 +59,8 @@ class DepthwiseConvolution : public IDepthwiseConvolution static constexpr int kernel_cols = KernelCols; static constexpr int stride_rows = StrideRows; static constexpr int stride_cols = StrideCols; - static constexpr int inner_tile_rows = stride_rows * output_tile_rows + kernel_rows - 1; - static constexpr int inner_tile_cols = stride_cols * output_tile_cols + kernel_cols - 1; + static constexpr int inner_tile_rows = stride_rows * (output_tile_rows - 1) + kernel_rows; + static constexpr int inner_tile_cols = stride_cols * (output_tile_cols - 1) + kernel_cols; /** Create a new depthwise convolution engine. * @@ -79,6 +79,75 @@ class DepthwiseConvolution : public IDepthwiseConvolution const TIn* const weights, const TIn* const input, TOut* const output + ) : DepthwiseConvolution( + n_batches, n_input_rows, n_input_cols, n_channels, padding_same, + weights, input, output, 0 /* column stride = default */ + ) + { + } + + /** Create a new depthwise convolution engine with a specified column stride. + * + * @param[in] n_batches Number of batches tensors. + * @param[in] n_input_rows Number of rows in input tensor. + * @param[in] n_input_cols Number of columns in input tensor. + * @param[in] n_channels Number of channels in input and output tensors. + * @param[in] padding_same True if padding is SAME, else VALID. + * @param[in] weights Pointer to Height x Width x Channel ordered weights. + * @param[in] input Pointer to NHWC ordered input tensor. + * @param[output] output Pointer to NHWC ordered output tensor. + * @param[in] col_stride Stride between columns of the weights, inputs and output tensors. + */ + DepthwiseConvolution( + const int n_batches, const int n_input_rows, const int n_input_cols, + const int n_channels, const bool padding_same, + const TIn* const weights, + const TIn* const input, + TOut* const output, + const int col_stride + ) : DepthwiseConvolution( + n_batches, n_input_rows, n_input_cols, n_channels, padding_same, + weights, input, output, + col_stride, 0, /* Weight row stride = default */ + col_stride, 0, 0, /* Input row stride, batch stride = default */ + col_stride, 0, 0 /* Output row stride, batch stride = default */ + ) + { + } + + /** Create a new depthwise convolution engine. + * + * @param[in] n_batches Number of batches tensors. + * @param[in] n_input_rows Number of rows in input tensor. + * @param[in] n_input_cols Number of columns in input tensor. + * @param[in] n_channels Number of channels in input and output tensors. + * @param[in] padding_same True if padding is SAME, else VALID. + * @param[in] weights Pointer to Height x Width x Channel ordered weights. + * @param[in] input Pointer to NHWC ordered input tensor. + * @param[output] output Pointer to NHWC ordered output tensor. + * @param[in] weight_col_stride Stride between columns of the weights (if 0, defaults appropriately). + * @param[in] weight_row_stride Stride between rows of the weights (if 0, defaults appropriately). + * @param[in] input_col_stride Stride between columns of the input tensor (if 0, defaults appropriately). + * @param[in] input_row_stride Stride between rows of the input tensor (if 0, defaults appropriately). + * @param[in] input_batch_stride Stride between batches of the input tensor (if 0, defaults appropriately). + * @param[in] output_col_stride Stride between columns of the output tensor (if 0, defaults appropriately). + * @param[in] output_row_stride Stride between rows of the output tensor (if 0, defaults appropriately). + * @param[in] output_batch_stride Stride between batches of the output tensor (if 0, defaults appropriately). + */ + DepthwiseConvolution( + const int n_batches, const int n_input_rows, const int n_input_cols, + const int n_channels, const bool padding_same, + const TIn* const weights, + const TIn* const input, + TOut* const output, + const int weight_col_stride, + const int weight_row_stride, + const int input_col_stride, + const int input_row_stride, + const int input_batch_stride, + const int output_col_stride, + const int output_row_stride, + const int output_batch_stride ); // Cannot copy or move a DepthwiseConvolution. @@ -99,14 +168,15 @@ class DepthwiseConvolution : public IDepthwiseConvolution */ int output_size(const int dim_size, const bool padding_same) const override { - return DepthwiseConvolution::get_output_size(dim_size, padding_same); + return DepthwiseConvolution< + OutputTileRows, + OutputTileCols, + KernelRows, + KernelCols, + StrideRows, + StrideCols, + TIn, TOut + >::get_output_size(dim_size, padding_same); } /** Get the window of work to be performed by an instance of the operator. @@ -128,6 +198,8 @@ class DepthwiseConvolution : public IDepthwiseConvolution static void process_tile_row( const int n_channels, const TIn* const weights, + const int weight_row_stride, + const int weight_col_stride, const TIn* const inptr, const int in_row_stride, const int in_col_stride, @@ -143,7 +215,27 @@ class DepthwiseConvolution : public IDepthwiseConvolution const int n_output_cols ); - /** Process a single tile of the tensors. + // Determine the maximum (and minimum) padding values which can be applied + // to tiles of the tensors involved in this class of convolution. + static constexpr int max_in_pad_top = (kernel_rows - 1) / 2; + static constexpr int min_in_pad_top = (kernel_rows - stride_rows) / 2; + + static constexpr int max_in_pad_left = (kernel_cols - 1) / 2; + static constexpr int min_in_pad_left = (kernel_cols - stride_cols) / 2; + + static constexpr int max_in_pad_bottom = inner_tile_rows; + static constexpr int max_in_pad_right = inner_tile_cols; + static constexpr int max_out_pad_bottom = output_tile_rows; + static constexpr int max_out_pad_right = output_tile_cols; + + static constexpr int n_in_pad_top_fns = (max_in_pad_top - min_in_pad_top) + 1; + static constexpr int n_in_pad_left_fns = (max_in_pad_left - min_in_pad_left) + 1; + static constexpr int n_in_pad_bottom_fns = max_in_pad_bottom + 1; + static constexpr int n_in_pad_right_fns = max_in_pad_right + 1; + static constexpr int n_out_pad_bottom_fns = max_out_pad_bottom + 1; + static constexpr int n_out_pad_right_fns = max_out_pad_right + 1; + + /** Pointer to a function which will process a tile. * * @param[in] n_channels Number of channels. * @param[in] weights Pointer to Height x Width x Channels ordered weights. @@ -153,48 +245,47 @@ class DepthwiseConvolution : public IDepthwiseConvolution * @param[out] outptr Pointer to the top-left output value for the tile. * @param[in] out_row_stride Stride between rows of the output tensor. * @param[in] out_col_stride Stride between columns of the output tensor. + * + * The following parameters may be ignored if the function has been + * specialised for specific padding constraints. + * + * @param[in] _in_pad_top Padding to apply to top of input tile. + * @param[in] _in_pad_left Padding to apply to left of input tile. + * @param[in] _in_pad_bottom Padding to apply to bottom of input tile. + * @param[in] _in_pad_right Padding to apply to right of input tile. + * @param[in] _out_pad_bottom Null cells at bottom of output tile. + * @param[in] _out_pad_right Null cells at right of output tile. */ - template < - int in_pad_top, int in_pad_left, int in_pad_bottom, int in_pad_right, - int out_pad_bottom, int out_pad_right - > - static void process_tile( + typedef void (*TileFn)( const int n_channels, const TIn* const weights, + const int weight_row_stride, + const int weight_col_stride, const TIn* const inptr, const int in_row_stride, const int in_col_stride, TOut* const outptr, const int out_row_stride, - const int out_col_stride - ); - - // Type of a pointer to a `process_tile` instance - typedef void (*TileFn)( - const int, - const TIn* const, - const TIn* const, const int, const int, - TOut* const, const int, const int + const int out_col_stride, + const int _in_pad_top, + const int _in_pad_left, + const int _in_pad_bottom, + const int _in_pad_right, + const int _out_pad_bottom, + const int _out_pad_right ); - // Determine the maximum padding values which can be applied to tiles of - // the tensors involved in this class of convolution. - static constexpr int max_in_pad_top = 2; - static constexpr int max_in_pad_left = 2; - static constexpr int max_in_pad_bottom = inner_tile_rows - 1; - static constexpr int max_in_pad_right = inner_tile_cols - 1; - static constexpr int max_out_pad_bottom = output_tile_rows; - static constexpr int max_out_pad_right = output_tile_cols; - - /** Array of methods to process tensor tiles. + /* Arrays of methods to process tensor tiles. * * Allows dynamic dispatch to specialized implementations based on * different padding configurations. */ - static const TileFn tile_fns[ - max_in_pad_top][max_in_pad_left][max_in_pad_bottom][max_in_pad_right][ - max_out_pad_bottom][max_out_pad_right - ]; + static const TileFn tilefn_unpadded; + static const TileFn tilefn_top[n_in_pad_top_fns]; + static const TileFn tilefn_left[n_in_pad_left_fns]; + static const TileFn tilefn_bottom[n_in_pad_bottom_fns][n_out_pad_bottom_fns]; + static const TileFn tilefn_right[n_in_pad_right_fns][n_out_pad_right_fns]; + static const TileFn tilefn_generic; private: // Member variables of instances of a convolution engine. @@ -204,6 +295,11 @@ class DepthwiseConvolution : public IDepthwiseConvolution const int _n_batches, _n_input_rows, _n_input_cols, _n_channels, _n_output_rows, _n_output_cols, _n_tile_rows, _n_tile_cols; const bool _padding_same; + + // Stride information for a convolution instance + const int _weight_col_stride, _weight_row_stride; + const int _input_col_stride, _input_row_stride, _input_batch_stride; + const int _output_col_stride, _output_row_stride, _output_batch_stride; }; } // namespace depthwise diff --git a/arm_compute/core/NEON/kernels/convolution/depthwise/impl_base.hpp b/arm_compute/core/NEON/kernels/convolution/depthwise/impl_base.hpp index f9671fc426..17889849db 100644 --- a/arm_compute/core/NEON/kernels/convolution/depthwise/impl_base.hpp +++ b/arm_compute/core/NEON/kernels/convolution/depthwise/impl_base.hpp @@ -39,6 +39,8 @@ namespace depthwise { +const unsigned int CHANNEL_BLOCK = 16; + template int DepthwiseConvolution::get_output_size( const int dim_size, const bool same_padding @@ -54,7 +56,15 @@ DepthwiseConvolution::DepthwiseConvolution( const int n_channels, const bool padding_same, const TIn* const weights, const TIn* const input, - TOut* const output + TOut* const output, + const int weight_col_stride, + const int weight_row_stride, + const int input_col_stride, + const int input_row_stride, + const int input_batch_stride, + const int output_col_stride, + const int output_row_stride, + const int output_batch_stride ) : _weights(weights), _input(input), _output(output), _n_batches(n_batches), _n_input_rows(n_input_rows), @@ -64,7 +74,15 @@ DepthwiseConvolution::DepthwiseConvolution( _n_output_cols(get_output_size(n_input_cols, padding_same)), _n_tile_rows(iceildiv(_n_output_rows, output_tile_rows)), _n_tile_cols(iceildiv(_n_output_cols, output_tile_cols)), - _padding_same(padding_same) + _padding_same(padding_same), + _weight_col_stride(weight_col_stride ? weight_col_stride : _n_channels), + _weight_row_stride(weight_row_stride ? weight_row_stride : KC * _weight_col_stride), + _input_col_stride(input_col_stride ? input_col_stride : _n_channels), + _input_row_stride(input_row_stride ? input_row_stride : _n_input_cols * _input_col_stride), + _input_batch_stride(input_batch_stride ? input_batch_stride : _n_input_rows * _input_row_stride), + _output_col_stride(output_col_stride ? output_col_stride : _n_channels), + _output_row_stride(output_row_stride ? output_row_stride : _n_output_cols * _output_col_stride), + _output_batch_stride(output_batch_stride ? output_batch_stride : _n_output_rows * _output_row_stride) { } @@ -72,8 +90,8 @@ DepthwiseConvolution::DepthwiseConvolution( template unsigned int DepthwiseConvolution::get_window() const { - // TODO Later support parallelisation over tile rows. - return 1; // _n_tile_rows; + // Parallelise over blocks of channels. + return iceildiv(_n_channels, CHANNEL_BLOCK); } @@ -83,41 +101,31 @@ void DepthwiseConvolution::run( const unsigned int stop ) { - // TODO Later support parallelisation over tile rows. - (void) start; - (void) stop; - - // Compute input striding - const int input_col_stride = _n_channels; - const int input_row_stride = _n_input_cols * input_col_stride; - const int input_batch_stride = _n_input_rows * input_row_stride; - - // Compute output striding - const int output_col_stride = _n_channels; - const int output_row_stride = _n_output_cols * output_col_stride; - const int output_batch_stride = _n_output_rows * output_row_stride; + // Parallelise over blocks of channels + const auto start_channel = CHANNEL_BLOCK * start; + const auto stop_channel = std::min(_n_channels, CHANNEL_BLOCK * stop); // Compute top and bottom padding for input and output const int input_pad_top = _padding_same ? - ((_n_output_rows - 1)*stride_rows + kernel_rows - _n_input_rows) / 2 : 0; + ((_n_output_rows - 1)*stride_rows + kernel_rows - _n_input_rows) / 2 : 0; const int input_pad_left = _padding_same ? - ((_n_output_cols - 1)*stride_cols + kernel_cols - _n_input_cols) / 2 : 0; - constexpr int tile_overlap = kernel_rows - 1; + ((_n_output_cols - 1)*stride_cols + kernel_cols - _n_input_cols) / 2 : 0; + constexpr int tile_overlap = kernel_rows - stride_rows; // Perform the convolution by calling `process_tile_row` for each tile row in // each batch. for (int batch = 0; batch < _n_batches; batch++) { - const TIn* const inptr_batch = _input + batch*input_batch_stride; - TOut* const outptr_batch = _output + batch*output_batch_stride; + const TIn* const inptr_batch = _input + batch*_input_batch_stride; + TOut* const outptr_batch = _output + batch*_output_batch_stride; // Loop over rows of tiles for (int tile_i = 0; tile_i < _n_tile_rows; tile_i++) { // Pointer to the row const int input_row_offset = (tile_i == 0) ? 0 : input_pad_top; - const TIn* const inptr_row = (inptr_batch + ((inner_tile_rows - tile_overlap)*tile_i - input_row_offset)*input_row_stride); - TOut* const outptr_row = outptr_batch + output_tile_rows * tile_i * output_row_stride; + const TIn* const inptr_row = (inptr_batch + ((inner_tile_rows - tile_overlap)*tile_i - input_row_offset)*_input_row_stride); + TOut* const outptr_row = outptr_batch + output_tile_rows * tile_i * _output_row_stride; // Input padding (top + bottom) for the row const int input_row_top = tile_i*(inner_tile_rows - tile_overlap) - input_pad_top; @@ -131,9 +139,10 @@ void DepthwiseConvolution::run( // Process the row process_tile_row( - _n_channels, _weights, - inptr_row, input_row_stride, input_col_stride, - outptr_row, output_row_stride, output_col_stride, + stop_channel - start_channel, + _weights + start_channel, _weight_row_stride, _weight_col_stride, + inptr_row + start_channel, _input_row_stride, _input_col_stride, + outptr_row + start_channel, _output_row_stride, _output_col_stride, input_row_pad_top, input_pad_left, input_row_pad_bottom, output_row_pad_bottom, _n_tile_cols, _n_input_cols, _n_output_cols @@ -147,6 +156,8 @@ template ::process_tile_row( const int n_channels, const TIn* const weights, + const int weight_row_stride, + const int weight_col_stride, const TIn* const inptr, const int in_row_stride, const int in_col_stride, @@ -162,7 +173,7 @@ void DepthwiseConvolution::process_tile_row const int n_output_cols ) { - constexpr int tile_overlap = kernel_cols - 1; + constexpr int tile_overlap = kernel_cols - stride_cols; // Loop over columns of tiles for (int tile_j = 0; tile_j < n_tiles; tile_j++) @@ -183,44 +194,182 @@ void DepthwiseConvolution::process_tile_row TOut* const outptr_col = outptr + tile_j * output_tile_cols * out_col_stride; // Apply the specific tile processing function - tile_fns[row_pad_in_top][t_pad_in_left][row_pad_in_bottom][t_pad_in_right][row_pad_out_bottom][t_pad_out_right]( - n_channels, weights, + const bool pad_top = row_pad_in_top > 0; + const bool pad_left = t_pad_in_left > 0; + const bool pad_bottom = row_pad_in_bottom || row_pad_out_bottom; + const bool pad_right = t_pad_in_right || t_pad_out_right; + + const TileFn tilefn = [&] () { + if (!pad_top && !pad_left && !pad_bottom && !pad_right) + { + // No padding + return tilefn_unpadded; + } + else if (pad_top && !pad_left && !pad_bottom && !pad_right) + { + // Padding on the top only, subtract off the minimum expected padding in + // order to index into the array of specialised methods. + const int index = row_pad_in_top - min_in_pad_top; + return tilefn_top[index]; + } + else if (!pad_top && pad_left && !pad_bottom && !pad_right) + { + // Padding on the left only, subtract off the minimum expected padding in + // order to index into the array of specialised methods. + const int index = t_pad_in_left - min_in_pad_left; + return tilefn_left[index]; + } + else if (!pad_top && !pad_left && pad_bottom && !pad_right) + { + // Padding on the bottom only + return tilefn_bottom[row_pad_in_bottom][row_pad_out_bottom]; + } + else if (!pad_top && !pad_left && !pad_bottom && pad_right) + { + // Padding on the right only + return tilefn_right[t_pad_in_right][t_pad_out_right]; + } + else + { + // Otherwise use generic tile processing method. + return tilefn_generic; + } + }(); + + tilefn( + n_channels, + weights, weight_row_stride, weight_col_stride, inptr_col, in_row_stride, in_col_stride, - outptr_col, out_row_stride, out_col_stride + outptr_col, out_row_stride, out_col_stride, + row_pad_in_top, t_pad_in_left, row_pad_in_bottom, t_pad_in_right, + row_pad_out_bottom, t_pad_out_right ); } } +// New templated struct used solely as a way to provide tile processing +// specialisations. +template +struct DepthwiseConvolutionImpl : public DepthwiseConvolution< + OutputTileRows, OutputTileCols, + KernelRows, KernelCols, + StrideRows, StrideCols, TIn, TOut +> +{ + typedef DepthwiseConvolution< + OutputTileRows, OutputTileCols, + KernelRows, KernelCols, + StrideRows, StrideCols, + TIn, TOut + > DWC; + + /** Perform the depthwise convolution of a tile. + * + * @param[in] n_channels Number of channels. + * @param[in] weights Pointer to Height x Width x Channels ordered weights. + * @param[in] inptr Pointer to the top-left unpadded value of the tile. + * @param[in] in_row_stride Stride between rows of the input tensor. + * @param[in] in_col_stride Stride between columns of the input tensor. + * @param[out] outptr Pointer to the top-left output value for the tile. + * @param[in] out_row_stride Stride between rows of the output tensor. + * @param[in] out_col_stride Stride between columns of the output tensor. + * + * The following parameters may be ignored if the function has been + * specialised for specific padding constraints. + * + * @param[in] _in_pad_top Padding to apply to top of input tile. + * @param[in] _in_pad_left Padding to apply to left of input tile. + * @param[in] _in_pad_bottom Padding to apply to bottom of input tile. + * @param[in] _in_pad_right Padding to apply to right of input tile. + * @param[in] _out_pad_bottom Null cells at bottom of output tile. + * @param[in] _out_pad_right Null cells at right of output tile. + */ + template < + bool Specialize=false, // Specialize (or not) the method + int InPadTop=0, // If specialized, top padding + int InPadLeft=0, // If specialized, left padding + int InPadBottom=0, // If specialized, bottom padding + int InPadRight=0, // If specialized, right padding + int OutPadBottom=0, // If specialized, bottom output padding + int OutPadRight=0 // If specialized, bottom right padding + > + static void process_tile( + const int n_channels, + const TIn* const weights, + const int weight_row_stride, + const int weight_col_stride, + const TIn* const inptr, + const int in_row_stride, + const int in_col_stride, + TOut* const outptr, + const int out_row_stride, + const int out_col_stride, + const int in_pad_top=0, + const int in_pad_left=0, + const int in_pad_bottom=0, + const int in_pad_right=0, + const int out_pad_bottom=0, + const int out_pad_right=0 + ); +}; + + template template < - int in_pad_top, int in_pad_left, int in_pad_bottom, int in_pad_right, - int out_pad_bottom, int out_pad_right + bool Specialize, + int InPadTop, int InPadLeft, int InPadBottom, int InPadRight, + int OutPadBottom, int OutPadRight > -void DepthwiseConvolution::process_tile( +void DepthwiseConvolutionImpl::process_tile( const int n_channels, - const TIn* const weights, - const TIn* const inptr, + const TIn *__restrict__ const weights, + const int weight_row_stride, + const int weight_col_stride, + const TIn *__restrict__ const inptr, const int in_row_stride, const int in_col_stride, - TOut* const outptr, + TOut *__restrict__ const outptr, const int out_row_stride, - const int out_col_stride + const int out_col_stride, + const int _in_pad_top, + const int _in_pad_left, + const int _in_pad_bottom, + const int _in_pad_right, + const int _out_pad_bottom, + const int _out_pad_right ) { + constexpr auto inner_tile_rows = DWC::inner_tile_rows; + constexpr auto inner_tile_cols = DWC::inner_tile_cols; + constexpr auto kernel_rows = DWC::kernel_rows; + constexpr auto kernel_cols = DWC::kernel_cols; + constexpr auto output_tile_rows = DWC::output_tile_rows; + constexpr auto output_tile_cols = DWC::output_tile_cols; + constexpr auto stride_rows = DWC::stride_rows; + constexpr auto stride_cols = DWC::stride_cols; + + // Extract parameters + const int in_pad_top = Specialize ? InPadTop : _in_pad_top; + const int in_pad_left = Specialize ? InPadLeft : _in_pad_left; + const int in_pad_bottom = Specialize ? InPadBottom : _in_pad_bottom; + const int in_pad_right = Specialize ? InPadRight : _in_pad_right; + const int out_pad_bottom = Specialize ? OutPadBottom : _out_pad_bottom; + const int out_pad_right = Specialize ? OutPadRight : _out_pad_right; + // Compute valid ranges of the tile - constexpr int in_cells_i = inner_tile_rows - in_pad_bottom; - constexpr int in_cells_j = inner_tile_cols - in_pad_right; - constexpr int out_cells_i = output_tile_rows - out_pad_bottom; - constexpr int out_cells_j = output_tile_cols - out_pad_right; + const int in_cells_i = inner_tile_rows - in_pad_bottom; + const int in_cells_j = inner_tile_cols - in_pad_right; + const int out_cells_i = output_tile_rows - out_pad_bottom; + const int out_cells_j = output_tile_cols - out_pad_right; // Instantiate pointers - const TIn* inptr_base = inptr; - const TIn* wptr_base = weights; - TOut* outptr_base = outptr; - - const int weight_col_stride = n_channels; - const int weight_row_stride = kernel_cols * n_channels; + const TIn* __restrict__ inptr_base = inptr; + const TIn* __restrict__ wptr_base = weights; + TOut* __restrict__ outptr_base = outptr; // Perform the depthwise convolution int channels_remaining = n_channels; @@ -259,7 +408,7 @@ void DepthwiseConvolution::process_tile( wptr_base++; // Perform the convolution - TOut v[out_cells_i][out_cells_j]; + TOut v[output_tile_rows][output_tile_cols]; for (int out_i = 0; out_i < out_cells_i; out_i++) { for (int out_j = 0; out_j < out_cells_j; out_j++) @@ -287,7 +436,7 @@ void DepthwiseConvolution::process_tile( // Store the output tile for (int i = 0; i < out_cells_i; i++) { - TOut* const outptr_row = outptr_base + i*out_row_stride; + TOut* __restrict__ const outptr_row = outptr_base + i*out_row_stride; for (int j = 0; j < out_cells_j; j++) { *(outptr_row + j*out_col_stride) = v[i][j]; @@ -297,52 +446,4 @@ void DepthwiseConvolution::process_tile( } } - -// New templated struct used solely as a way to provide tile processing -// specialisations. -template -struct DepthwiseConvolutionImpl : public DepthwiseConvolution< - OutputTileRows, OutputTileCols, - KernelRows, KernelCols, - StrideRows, StrideCols, TIn, TOut -> -{ - template < - int in_pad_top, int in_pad_left, int in_pad_bottom, int in_pad_right, - int out_pad_bottom, int out_pad_right - > - static void process_tile( - const int n_channels, - const TIn* const weights, - const TIn* const inptr, - const int in_row_stride, - const int in_col_stride, - TOut* const outptr, - const int out_row_stride, - const int out_col_stride - ) - { - // By default, redirect to parent. Specialised implementations can be added - // by overriding this method. - DepthwiseConvolution:: - template process_tile( - n_channels, - weights, - inptr, - in_row_stride, - in_col_stride, - outptr, - out_row_stride, - out_col_stride - ); - } -}; - } // namespace depthwise diff --git a/arm_compute/core/NEON/kernels/convolution/depthwise/impl_fp32_fp32.hpp b/arm_compute/core/NEON/kernels/convolution/depthwise/impl_fp32_fp32.hpp index e7f0609b0c..7a216ed518 100644 --- a/arm_compute/core/NEON/kernels/convolution/depthwise/impl_fp32_fp32.hpp +++ b/arm_compute/core/NEON/kernels/convolution/depthwise/impl_fp32_fp32.hpp @@ -51,36 +51,58 @@ struct DepthwiseConvolutionImpl DWC; template < - int in_pad_top, int in_pad_left, int in_pad_bottom, int in_pad_right, - int out_pad_bottom, int out_pad_right + bool Specialize=false, // Specialize (or not) the method + int InPadTop=0, // If specialized, top padding + int InPadLeft=0, // If specialized, left padding + int InPadBottom=0, // If specialized, bottom padding + int InPadRight=0, // If specialized, right padding + int OutPadBottom=0, // If specialized, bottom output padding + int OutPadRight=0 // If specialized, bottom right padding > static void process_tile( const int n_channels, const float* const weights, + const int weight_row_stride, + const int weight_col_stride, const float* const inptr, const int in_row_stride, const int in_col_stride, float* const outptr, const int out_row_stride, - const int out_col_stride + const int out_col_stride, + const int in_pad_top=0, + const int in_pad_left=0, + const int in_pad_bottom=0, + const int in_pad_right=0, + const int out_pad_bottom=0, + const int out_pad_right=0 ); }; template template < - int in_pad_top, int in_pad_left, int in_pad_bottom, int in_pad_right, - int out_pad_bottom, int out_pad_right + bool Specialize, + int InPadTop, int InPadLeft, int InPadBottom, int InPadRight, + int OutPadBottom, int OutPadRight > void DepthwiseConvolutionImpl::process_tile( const int n_channels, - const float* const weights, - const float* const inptr, + const float *__restrict__ const weights, + const int weight_row_stride, + const int weight_col_stride, + const float *__restrict__ const inptr, const int in_row_stride, const int in_col_stride, - float* const outptr, + float *__restrict__ const outptr, const int out_row_stride, - const int out_col_stride + const int out_col_stride, + const int _in_pad_top, + const int _in_pad_left, + const int _in_pad_bottom, + const int _in_pad_right, + const int _out_pad_bottom, + const int _out_pad_right ) { constexpr auto inner_tile_rows = DWC::inner_tile_rows; @@ -92,19 +114,24 @@ void DepthwiseConvolutionImpl::process_t constexpr auto stride_rows = DWC::stride_rows; constexpr auto stride_cols = DWC::stride_cols; + // Extract parameters + const int in_pad_top = Specialize ? InPadTop : _in_pad_top; + const int in_pad_left = Specialize ? InPadLeft : _in_pad_left; + const int in_pad_bottom = Specialize ? InPadBottom : _in_pad_bottom; + const int in_pad_right = Specialize ? InPadRight : _in_pad_right; + const int out_pad_bottom = Specialize ? OutPadBottom : _out_pad_bottom; + const int out_pad_right = Specialize ? OutPadRight : _out_pad_right; + // Compute valid ranges of the tile - constexpr int in_cells_i = inner_tile_rows - in_pad_bottom; - constexpr int in_cells_j = inner_tile_cols - in_pad_right; - constexpr int out_cells_i = output_tile_rows - out_pad_bottom; - constexpr int out_cells_j = output_tile_cols - out_pad_right; + const int in_cells_i = inner_tile_rows - in_pad_bottom; + const int in_cells_j = inner_tile_cols - in_pad_right; + const int out_cells_i = output_tile_rows - out_pad_bottom; + const int out_cells_j = output_tile_cols - out_pad_right; // Instantiate pointers - const float* inptr_base = inptr; - const float* wptr_base = weights; - float* outptr_base = outptr; - - const int weight_col_stride = n_channels; - const int weight_row_stride = kernel_cols * n_channels; + const float* __restrict__ inptr_base = inptr; + const float* __restrict__ wptr_base = weights; + float* __restrict__ outptr_base = outptr; // Perform the depthwise convolution int channels_remaining = n_channels; @@ -144,7 +171,7 @@ void DepthwiseConvolutionImpl::process_t wptr_base += 4; // Perform the convolution - float32x4_t v[out_cells_i][out_cells_j]; + float32x4_t v[output_tile_rows][output_tile_cols]; for (int out_i = 0; out_i < out_cells_i; out_i++) { for (int out_j = 0; out_j < out_cells_j; out_j++) @@ -222,7 +249,7 @@ void DepthwiseConvolutionImpl::process_t wptr_base++; // Perform the convolution - float v[out_cells_i][out_cells_j]; + float v[output_tile_rows][output_tile_cols]; for (int out_i = 0; out_i < out_cells_i; out_i++) { for (int out_j = 0; out_j < out_cells_j; out_j++) diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp index f5ee608b60..49c67d19bb 100644 --- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp +++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp @@ -219,8 +219,7 @@ void NEDepthwiseConvolutionLayer3x3Kernel::generate_convolver() ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(_input, _weights); ARM_COMPUTE_ERROR_ON(_weights->info()->dimension(1) != 3 || _weights->info()->dimension(2) != 3); - _convolver = create_convolver_object(_input->info()->tensor_shape(), _conv_info, - _weights->buffer(), _input->buffer(), _output->buffer()); + _convolver = create_convolver_object(_conv_info, _weights, _input, _output, true); } void NEDepthwiseConvolutionLayer3x3Kernel::configure_generic() @@ -282,8 +281,7 @@ void NEDepthwiseConvolutionLayer3x3Kernel::configure_optimized() ARM_COMPUTE_ERROR_ON(_weights->info()->dimension(1) != 3 || _weights->info()->dimension(2) != 3); _border_size = BorderSize(0, 0); - _convolver = create_convolver_object(_input->info()->tensor_shape(), _conv_info, - _weights->buffer(), _input->buffer(), _output->buffer()); + _convolver = create_convolver_object(_conv_info, _weights, _input, _output); // Auto-configure output bool same_padding = _conv_info.has_padding(); @@ -296,6 +294,15 @@ void NEDepthwiseConvolutionLayer3x3Kernel::configure_optimized() auto_init_if_empty(*_output->info(), _input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape)); + // Set padding in channels + const int num_channels = _weights->info()->dimension(0); + if((num_channels >= 128) && (num_channels % 16 == 0)) + { + _input->info()->extend_padding(PaddingSize(0, 4, 0, 0)); + _weights->info()->extend_padding(PaddingSize(0, 4, 0, 0)); + _output->info()->extend_padding(PaddingSize(0, 4, 0, 0)); + } + // Configure window Window win; auto win_last = _convolver->get_window(); @@ -330,41 +337,56 @@ void NEDepthwiseConvolutionLayer3x3Kernel::run_optimized(const Window &window, c _convolver->run(start, end); } -std::unique_ptr NEDepthwiseConvolutionLayer3x3Kernel::create_convolver_object(TensorShape shape, - PadStrideInfo conv_info, - const uint8_t *w_ptr, - uint8_t *in_ptr, - uint8_t *out_ptr) +std::unique_ptr NEDepthwiseConvolutionLayer3x3Kernel::create_convolver_object(PadStrideInfo conv_info, + const ITensor *w, + const ITensor *in, + ITensor *out, + bool setup_strides) { - const int in_rows = shape.z(); - const int in_cols = shape.y(); - const int n_batches = shape[3]; - const int n_channels = shape.x(); - const bool padding_same = conv_info.has_padding(); + const TensorShape shape = in->info()->tensor_shape(); + const int in_rows = shape.z(); + const int in_cols = shape.y(); + const int n_batches = shape[3]; + const int n_channels = shape.x(); + const bool padding_same = conv_info.has_padding(); + const int weight_col_stride = (setup_strides) ? w->info()->strides_in_bytes().y() / w->info()->element_size() : 0; + const int weight_row_stride = (setup_strides) ? w->info()->strides_in_bytes().z() / w->info()->element_size() : 0; + const int input_col_stride = (setup_strides) ? in->info()->strides_in_bytes().y() / in->info()->element_size() : 0; + const int input_row_stride = (setup_strides) ? in->info()->strides_in_bytes().z() / in->info()->element_size() : 0; + const int input_batch_stride = (setup_strides) ? in->info()->strides_in_bytes()[3] / in->info()->element_size() : 0; + const int output_col_stride = (setup_strides) ? out->info()->strides_in_bytes().y() / out->info()->element_size() : 0; + const int output_row_stride = (setup_strides) ? out->info()->strides_in_bytes().z() / out->info()->element_size() : 0; + const int output_batch_stride = (setup_strides) ? out->info()->strides_in_bytes()[3] / out->info()->element_size() : 0; const auto stride_x = conv_info.stride().first; switch(stride_x) { case 1: - return arm_compute::support::cpp14::make_unique>( + return arm_compute::support::cpp14::make_unique>( n_batches, in_rows, in_cols, n_channels, padding_same, - reinterpret_cast(w_ptr), - reinterpret_cast(in_ptr), - reinterpret_cast(out_ptr)); + reinterpret_cast(w->ptr_to_element(Coordinates())), + reinterpret_cast(in->ptr_to_element(Coordinates())), + reinterpret_cast(out->ptr_to_element(Coordinates())), + weight_col_stride, weight_row_stride, + input_col_stride, input_row_stride, input_batch_stride, + output_col_stride, output_row_stride, output_batch_stride); case 2: - return arm_compute::support::cpp14::make_unique>( + return arm_compute::support::cpp14::make_unique>( n_batches, in_rows, in_cols, n_channels, padding_same, - reinterpret_cast(w_ptr), - reinterpret_cast(in_ptr), - reinterpret_cast(out_ptr)); + reinterpret_cast(w->ptr_to_element(Coordinates())), + reinterpret_cast(in->ptr_to_element(Coordinates())), + reinterpret_cast(out->ptr_to_element(Coordinates())), + weight_col_stride, weight_row_stride, + input_col_stride, input_row_stride, input_batch_stride, + output_col_stride, output_row_stride, output_batch_stride); default: return nullptr; } diff --git a/src/core/NEON/kernels/convolution/common/utils.cpp b/src/core/NEON/kernels/convolution/common/utils.cpp index 24d0386c76..45847bbfd9 100644 --- a/src/core/NEON/kernels/convolution/common/utils.cpp +++ b/src/core/NEON/kernels/convolution/common/utils.cpp @@ -23,18 +23,6 @@ */ #include -#include - -double TimeInUs(void) -{ -#ifdef CYCLE_PROFILING - timespec t; - clock_gettime(CLOCK_REALTIME, &t); - return 1e6*t.tv_sec + 1e-3*t.tv_nsec; -#else - return 0; -#endif -} void PrintMatrix(const float* const m, const int M, const int N, const int row_stride) { @@ -47,4 +35,4 @@ void PrintMatrix(const float* const m, const int M, const int N, const int row_s printf("\n"); } printf("\n"); -} +} \ No newline at end of file diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_1x1_fp32_fp32.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_1x1_fp32_fp32.cpp index fa50f79bc5..9b3a60db59 100644 --- a/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_1x1_fp32_fp32.cpp +++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_1x1_fp32_fp32.cpp @@ -28,412 +28,543 @@ namespace depthwise using Conv = DepthwiseConvolution<2, 2, 3, 3, 1, 1, float, float>; using ConvImpl = DepthwiseConvolutionImpl<2, 2, 3, 3, 1, 1, float, float>; +#ifdef __aarch64__ + +template <> template <> -const Conv::TileFn Conv::tile_fns - [max_in_pad_top] - [max_in_pad_left] - [max_in_pad_bottom] - [max_in_pad_right] - [max_out_pad_bottom] - [max_out_pad_right] = { - { // Input pad top = 0 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 0, 0, 0, 0>, - ConvImpl::template process_tile<0, 0, 0, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 0, 0, 1, 0>, - ConvImpl::template process_tile<0, 0, 0, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 0, 1, 0, 0>, - ConvImpl::template process_tile<0, 0, 0, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 0, 1, 1, 0>, - ConvImpl::template process_tile<0, 0, 0, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 0, 2, 0, 0>, - ConvImpl::template process_tile<0, 0, 0, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 0, 2, 1, 0>, - ConvImpl::template process_tile<0, 0, 0, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 1, 0, 0, 0>, - ConvImpl::template process_tile<0, 0, 1, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 1, 0, 1, 0>, - ConvImpl::template process_tile<0, 0, 1, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 1, 1, 0, 0>, - ConvImpl::template process_tile<0, 0, 1, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 1, 1, 1, 0>, - ConvImpl::template process_tile<0, 0, 1, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 1, 2, 0, 0>, - ConvImpl::template process_tile<0, 0, 1, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 1, 2, 1, 0>, - ConvImpl::template process_tile<0, 0, 1, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 2, 0, 0, 0>, - ConvImpl::template process_tile<0, 0, 2, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 2, 0, 1, 0>, - ConvImpl::template process_tile<0, 0, 2, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 2, 1, 0, 0>, - ConvImpl::template process_tile<0, 0, 2, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 2, 1, 1, 0>, - ConvImpl::template process_tile<0, 0, 2, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 2, 2, 0, 0>, - ConvImpl::template process_tile<0, 0, 2, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 2, 2, 1, 0>, - ConvImpl::template process_tile<0, 0, 2, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 2 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 0, 0, 0, 0>, - ConvImpl::template process_tile<0, 1, 0, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 0, 0, 1, 0>, - ConvImpl::template process_tile<0, 1, 0, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 0, 1, 0, 0>, - ConvImpl::template process_tile<0, 1, 0, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 0, 1, 1, 0>, - ConvImpl::template process_tile<0, 1, 0, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 0, 2, 0, 0>, - ConvImpl::template process_tile<0, 1, 0, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 0, 2, 1, 0>, - ConvImpl::template process_tile<0, 1, 0, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 1, 0, 0, 0>, - ConvImpl::template process_tile<0, 1, 1, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 1, 0, 1, 0>, - ConvImpl::template process_tile<0, 1, 1, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 1, 1, 0, 0>, - ConvImpl::template process_tile<0, 1, 1, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 1, 1, 1, 0>, - ConvImpl::template process_tile<0, 1, 1, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 1, 2, 0, 0>, - ConvImpl::template process_tile<0, 1, 1, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 1, 2, 1, 0>, - ConvImpl::template process_tile<0, 1, 1, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 2, 0, 0, 0>, - ConvImpl::template process_tile<0, 1, 2, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 2, 0, 1, 0>, - ConvImpl::template process_tile<0, 1, 2, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 2, 1, 0, 0>, - ConvImpl::template process_tile<0, 1, 2, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 2, 1, 1, 0>, - ConvImpl::template process_tile<0, 1, 2, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 2, 2, 0, 0>, - ConvImpl::template process_tile<0, 1, 2, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 2, 2, 1, 0>, - ConvImpl::template process_tile<0, 1, 2, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 2 - }, // Input pad left = 1 - }, // Input pad top = 0 - { // Input pad top = 1 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 0, 0, 0, 0>, - ConvImpl::template process_tile<1, 0, 0, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 0, 0, 1, 0>, - ConvImpl::template process_tile<1, 0, 0, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 0, 1, 0, 0>, - ConvImpl::template process_tile<1, 0, 0, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 0, 1, 1, 0>, - ConvImpl::template process_tile<1, 0, 0, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 0, 2, 0, 0>, - ConvImpl::template process_tile<1, 0, 0, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 0, 2, 1, 0>, - ConvImpl::template process_tile<1, 0, 0, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 1, 0, 0, 0>, - ConvImpl::template process_tile<1, 0, 1, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 1, 0, 1, 0>, - ConvImpl::template process_tile<1, 0, 1, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 1, 1, 0, 0>, - ConvImpl::template process_tile<1, 0, 1, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 1, 1, 1, 0>, - ConvImpl::template process_tile<1, 0, 1, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 1, 2, 0, 0>, - ConvImpl::template process_tile<1, 0, 1, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 1, 2, 1, 0>, - ConvImpl::template process_tile<1, 0, 1, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 2, 0, 0, 0>, - ConvImpl::template process_tile<1, 0, 2, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 2, 0, 1, 0>, - ConvImpl::template process_tile<1, 0, 2, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 2, 1, 0, 0>, - ConvImpl::template process_tile<1, 0, 2, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 2, 1, 1, 0>, - ConvImpl::template process_tile<1, 0, 2, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 2, 2, 0, 0>, - ConvImpl::template process_tile<1, 0, 2, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 2, 2, 1, 0>, - ConvImpl::template process_tile<1, 0, 2, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 2 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 0, 0, 0, 0>, - ConvImpl::template process_tile<1, 1, 0, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 0, 0, 1, 0>, - ConvImpl::template process_tile<1, 1, 0, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 0, 1, 0, 0>, - ConvImpl::template process_tile<1, 1, 0, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 0, 1, 1, 0>, - ConvImpl::template process_tile<1, 1, 0, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 0, 2, 0, 0>, - ConvImpl::template process_tile<1, 1, 0, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 0, 2, 1, 0>, - ConvImpl::template process_tile<1, 1, 0, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 1, 0, 0, 0>, - ConvImpl::template process_tile<1, 1, 1, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 1, 0, 1, 0>, - ConvImpl::template process_tile<1, 1, 1, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 1, 1, 0, 0>, - ConvImpl::template process_tile<1, 1, 1, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 1, 1, 1, 0>, - ConvImpl::template process_tile<1, 1, 1, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 1, 2, 0, 0>, - ConvImpl::template process_tile<1, 1, 1, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 1, 2, 1, 0>, - ConvImpl::template process_tile<1, 1, 1, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 2, 0, 0, 0>, - ConvImpl::template process_tile<1, 1, 2, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 2, 0, 1, 0>, - ConvImpl::template process_tile<1, 1, 2, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 2, 1, 0, 0>, - ConvImpl::template process_tile<1, 1, 2, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 2, 1, 1, 0>, - ConvImpl::template process_tile<1, 1, 2, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 2, 2, 0, 0>, - ConvImpl::template process_tile<1, 1, 2, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 2, 2, 1, 0>, - ConvImpl::template process_tile<1, 1, 2, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - }, // Input pad bottom = 2 - }, // Input pad left = 1 - }, // Input pad top = 1 +void ConvImpl::process_tile( + const int n_channels, + const float* const weights, + const int weight_row_stride, + const int weight_col_stride, + const float* const inptr, + const int in_row_stride, + const int in_col_stride, + float* const outptr, + const int out_row_stride, + const int out_col_stride, + const int, const int, const int, const int, const int, const int +) +{ + // Copy pointers + const float *uptr0 = inptr; + const float *wptr0 = weights; + float *vptr0 = outptr; + + int channels_remaining = n_channels; + if (channels_remaining >= 4) + { + // Process blocks of 4 channels at a time + int n_iters = ((channels_remaining / 4) + 1)/2 - 1; + const bool odd_tail = (channels_remaining / 4) & 1; + channels_remaining %= 4; + + asm volatile ( + "qW11B .req q0\n" "vW11B .req v0\n" "qW33A .req q1\n" "qU32B .req q1\n" + "vW33A .req v1\n" "vU32B .req v1\n" "qU44B .req q2\n" "qW21A .req q2\n" + "vU44B .req v2\n" "vW21A .req v2\n" "qU21B .req q3\n" "qU32A .req q3\n" + "vU21B .req v3\n" "vU32A .req v3\n" "qU43A .req q4\n" "qV21B .req q4\n" + "vU43A .req v4\n" "vV21B .req v4\n" "qU24A .req q5\n" "qU44A .req q5\n" + "qU33B .req q5\n" "vU24A .req v5\n" "vU44A .req v5\n" "vU33B .req v5\n" + "qU31A .req q6\n" "qV12B .req q6\n" "qU23A .req q6\n" "vU31A .req v6\n" + "vV12B .req v6\n" "vU23A .req v6\n" "qW31B .req q7\n" "qV22A .req q7\n" + "vW31B .req v7\n" "vV22A .req v7\n" "qV12A .req q8\n" "qW21B .req q8\n" + "vV12A .req v8\n" "vW21B .req v8\n" "qU22B .req q9\n" "qU34A .req q9\n" + "vU22B .req v9\n" "vU34A .req v9\n" "qU13B .req q10\n" "qU13A .req q10\n" + "vU13B .req v10\n" "vU13A .req v10\n" "qU34B .req q11\n" "qU22A .req q11\n" + "vU34B .req v11\n" "vU22A .req v11\n" "qU24B .req q12\n" "qU31B .req q12\n" + "vU24B .req v12\n" "vU31B .req v12\n" "qW12B .req q13\n" "qW13A .req q13\n" + "vW12B .req v13\n" "vW13A .req v13\n" "qV21A .req q14\n" "qV11B .req q14\n" + "vV21A .req v14\n" "vV11B .req v14\n" "qW32A .req q15\n" "qW32B .req q15\n" + "vW32A .req v15\n" "vW32B .req v15\n" "qW31A .req q16\n" "qV22B .req q16\n" + "vW31A .req v16\n" "vV22B .req v16\n" + "qW11A .req q17\n" "vW11A .req v17\n" "qW13B .req q18\n" "qU14A .req q18\n" + "vW13B .req v18\n" "vU14A .req v18\n" "qU33A .req q19\n" "qW33B .req q19\n" + "vU33A .req v19\n" "vW33B .req v19\n" "qW22A .req q20\n" "qU23B .req q20\n" + "vW22A .req v20\n" "vU23B .req v20\n" "qU12A .req q21\n" "qU42A .req q21\n" + "vU12A .req v21\n" "vU42A .req v21\n" "qU41A .req q22\n" "qU42B .req q22\n" + "vU41A .req v22\n" "vU42B .req v22\n" "qW23A .req q23\n" "qW23B .req q23\n" + "vW23A .req v23\n" "vW23B .req v23\n" "qU43B .req q24\n" "qU11A .req q24\n" + "vU43B .req v24\n" "vU11A .req v24\n" "qU12B .req q25\n" "qW12A .req q25\n" + "vU12B .req v25\n" "vW12A .req v25\n" "qU41B .req q26\n" "qV11A .req q26\n" + "vU41B .req v26\n" "vV11A .req v26\n" "qW22B .req q27\n" "vW22B .req v27\n" + "qU11B .req q28\n" "qU14B .req q28\n" "vU11B .req v28\n" "vU14B .req v28\n" + "qU21A .req q29\n" "vU21A .req v29\n" + + "u_col_stride1 .req %x[u_col_stride]\n" + "u_col_stride2 .req x0\n" + "u_col_stride3 .req x1\n" + "uptr1 .req x2\n" + "uptr2 .req x3\n" + "uptr3 .req x4\n" + "wptr1 .req x5\n" + "wptr2 .req x6\n" + "vptr1 .req x7\n" + "w_col_stride1 .req %x[w_col_stride]\n" + "w_col_stride2 .req x8\n" + + // Prepare strides and pointers + "add uptr1, %x[uptr0], %x[u_row_stride]\n" + "add uptr2, uptr1 , %x[u_row_stride]\n" + "add uptr3, uptr2 , %x[u_row_stride]\n" + "add wptr1, %x[wptr0], %x[w_row_stride]\n" + "add wptr2, wptr1 , %x[w_row_stride]\n" + "add vptr1, %x[vptr0], %x[v_row_stride]\n" + "add u_col_stride2, %x[u_col_stride], %x[u_col_stride]\n" + "add u_col_stride3, u_col_stride2 , %x[u_col_stride]\n" + "add w_col_stride2, %x[w_col_stride], %x[w_col_stride]\n" + + // Load in preparation for execution + "ldr qU14A, [%x[uptr0], u_col_stride3]\n" + "ldr qW13A, [%x[wptr0], w_col_stride2]\n" + "ldr qU13A, [%x[uptr0], u_col_stride2]\n" + "ldr qW12A, [%x[wptr0], w_col_stride1]\n" + "ldr qU12A, [%x[uptr0], u_col_stride1]\n" + "ldr qW11A, [%x[wptr0]], #0x10\n" + "ldr qU24A, [uptr1, u_col_stride3]\n" + "ldr qW23A, [wptr1, w_col_stride2]\n" + "ldr qU23A, [uptr1, u_col_stride2]\n" + "ldr qW22A, [wptr1, w_col_stride1]\n" + "ldr qU22A, [uptr1, u_col_stride1]\n" + "ldr qW21A, [wptr1], #0x10\n" + "ldr qU34A, [uptr2, u_col_stride3]\n" + "ldr qW33A, [wptr2, w_col_stride2]\n" + "ldr qU33A, [uptr2, u_col_stride2]\n" + "ldr qW32A, [wptr2, w_col_stride1]\n" + "ldr qU32A, [uptr2, u_col_stride1]\n" + "ldr qW31A, [wptr2], #0x10\n" + "fmul vV12A.4s, vU14A.4s, vW13A.4s\n" + "cbz %x[iters], 2f\n" // Jump to tail if doing zero iterations of loop + + "1:" // Main loop body + // A part + "fmul vV11A.4s, vU13A.4s, vW13A.4s\n" + "fmla vV12A.4s, vU13A.4s, vW12A.4s\n" + "fmla vV11A.4s, vU12A.4s, vW12A.4s\n" + "fmla vV12A.4s, vU12A.4s, vW11A.4s\n" + "fmla vV12A.4s, vU24A.4s, vW23A.4s\n" + "fmul vV22A.4s, vU24A.4s, vW13A.4s\n" + "fmla vV11A.4s, vU23A.4s, vW23A.4s\n" + "ldr qU44A, [uptr3, u_col_stride3]\n" + "fmla vV12A.4s, vU23A.4s, vW22A.4s\n" + "ldr qU43A, [uptr3, u_col_stride2]\n" + "fmul vV21A.4s, vU23A.4s, vW13A.4s\n" + "ldr qU42A, [uptr3, u_col_stride1]\n" + "fmla vV22A.4s, vU23A.4s, vW12A.4s\n" + "ldr qU11A, [%x[uptr0]], #0x10\n" + "fmla vV11A.4s, vU22A.4s, vW22A.4s\n" + "ldr qU21A, [uptr1], #0x10\n" + "fmla vV12A.4s, vU22A.4s, vW21A.4s\n" + "ldr qU31A, [uptr2], #0x10\n" + "fmla vV21A.4s, vU22A.4s, vW12A.4s\n" + "ldr qU41A, [uptr3], #0x10\n" + "fmla vV22A.4s, vU22A.4s, vW11A.4s\n" + "ldr qU14B, [%x[uptr0], u_col_stride3]\n" + "fmla vV12A.4s, vU34A.4s, vW33A.4s\n" + "ldr qW13B, [%x[wptr0], w_col_stride2]\n" + "fmla vV22A.4s, vU34A.4s, vW23A.4s\n" + "ldr qU13B, [%x[uptr0], u_col_stride2]\n" + "fmla vV11A.4s, vU33A.4s, vW33A.4s\n" + "ldr qW12B, [%x[wptr0], w_col_stride1]\n" + "fmla vV12A.4s, vU33A.4s, vW32A.4s\n" + "ldr qU12B, [%x[uptr0], u_col_stride1]\n" + "fmla vV21A.4s, vU33A.4s, vW23A.4s\n" + "ldr qW11B, [%x[wptr0]], #0x10\n" + "fmla vV22A.4s, vU33A.4s, vW22A.4s\n" + "ldr qU24B, [uptr1, u_col_stride3]\n" + "fmla vV11A.4s, vU32A.4s, vW32A.4s\n" + "ldr qW23B, [wptr1, w_col_stride2]\n" + "fmla vV12A.4s, vU32A.4s, vW31A.4s\n" + "str qV12A, [%x[vptr0], %x[v_col_stride]]\n" + "fmla vV21A.4s, vU32A.4s, vW22A.4s\n" + "ldr qU23B, [uptr1, u_col_stride2]\n" + "fmla vV22A.4s, vU32A.4s, vW21A.4s\n" + "ldr qW22B, [wptr1, w_col_stride1]\n" + "fmla vV22A.4s, vU44A.4s, vW33A.4s\n" + "ldr qU22B, [uptr1, u_col_stride1]\n" + "fmla vV21A.4s, vU43A.4s, vW33A.4s\n" + "ldr qW21B, [wptr1], #0x10\n" + "fmla vV22A.4s, vU43A.4s, vW32A.4s\n" + "ldr qU34B, [uptr2, u_col_stride3]\n" + "fmla vV21A.4s, vU42A.4s, vW32A.4s\n" + "ldr qW33B, [wptr2, w_col_stride2]\n" + "fmla vV22A.4s, vU42A.4s, vW31A.4s\n" + "str qV22A, [vptr1, %x[v_col_stride]]\n" + "fmla vV11A.4s, vU11A.4s, vW11A.4s\n" + "ldr qU33B, [uptr2, u_col_stride2]\n" + "fmla vV11A.4s, vU21A.4s, vW21A.4s\n" + "ldr qW32B, [wptr2, w_col_stride1]\n" + "fmla vV21A.4s, vU21A.4s, vW11A.4s\n" + "ldr qU32B, [uptr2, u_col_stride1]\n" + "fmla vV11A.4s, vU31A.4s, vW31A.4s\n" + "str qV11A, [%x[vptr0]], #0x10\n" + "fmla vV21A.4s, vU31A.4s, vW21A.4s\n" + "ldr qW31B, [wptr2], #0x10\n" + "fmla vV21A.4s, vU41A.4s, vW31A.4s\n" + "str qV21A, [vptr1], #0x10\n" + + // B part + "fmul vV12B.4s, vU14B.4s, vW13B.4s\n" + "fmul vV11B.4s, vU13B.4s, vW13B.4s\n" + "fmla vV12B.4s, vU13B.4s, vW12B.4s\n" + "fmla vV11B.4s, vU12B.4s, vW12B.4s\n" + "fmla vV12B.4s, vU12B.4s, vW11B.4s\n" + "fmla vV12B.4s, vU24B.4s, vW23B.4s\n" + "fmul vV22B.4s, vU24B.4s, vW13B.4s\n" + "subs %x[iters], %x[iters], #1\n" + "fmla vV11B.4s, vU23B.4s, vW23B.4s\n" + "ldr qU44B, [uptr3, u_col_stride3]\n" + "fmla vV12B.4s, vU23B.4s, vW22B.4s\n" + "ldr qU43B, [uptr3, u_col_stride2]\n" + "fmul vV21B.4s, vU23B.4s, vW13B.4s\n" + "ldr qU42B, [uptr3, u_col_stride1]\n" + "fmla vV22B.4s, vU23B.4s, vW12B.4s\n" + "ldr qU11B, [%x[uptr0]], #0x10\n" + "fmla vV11B.4s, vU22B.4s, vW22B.4s\n" + "ldr qU21B, [uptr1], #0x10\n" + "fmla vV12B.4s, vU22B.4s, vW21B.4s\n" + "ldr qU31B, [uptr2], #0x10\n" + "fmla vV21B.4s, vU22B.4s, vW12B.4s\n" + "ldr qU41B, [uptr3], #0x10\n" + "fmla vV22B.4s, vU22B.4s, vW11B.4s\n" + "ldr qU14A, [%x[uptr0], u_col_stride3]\n" + "fmla vV12B.4s, vU34B.4s, vW33B.4s\n" + "ldr qW13A, [%x[wptr0], w_col_stride2]\n" + "fmla vV22B.4s, vU34B.4s, vW23B.4s\n" + "ldr qU13A, [%x[uptr0], u_col_stride2]\n" + "fmla vV11B.4s, vU33B.4s, vW33B.4s\n" + "ldr qW12A, [%x[wptr0], w_col_stride1]\n" + "fmla vV12B.4s, vU33B.4s, vW32B.4s\n" + "ldr qU12A, [%x[uptr0], u_col_stride1]\n" + "fmla vV21B.4s, vU33B.4s, vW23B.4s\n" + "ldr qW11A, [%x[wptr0]], #0x10\n" + "fmla vV22B.4s, vU33B.4s, vW22B.4s\n" + "ldr qU24A, [uptr1, u_col_stride3]\n" + "fmla vV11B.4s, vU32B.4s, vW32B.4s\n" + "ldr qW23A, [wptr1, w_col_stride2]\n" + "fmla vV12B.4s, vU32B.4s, vW31B.4s\n" + "str qV12B, [%x[vptr0], %x[v_col_stride]]\n" + "fmla vV21B.4s, vU32B.4s, vW22B.4s\n" + "ldr qU23A, [uptr1, u_col_stride2]\n" + "fmla vV22B.4s, vU32B.4s, vW21B.4s\n" + "ldr qW22A, [wptr1, w_col_stride1]\n" + "fmla vV22B.4s, vU44B.4s, vW33B.4s\n" + "ldr qU22A, [uptr1, u_col_stride1]\n" + "fmla vV21B.4s, vU43B.4s, vW33B.4s\n" + "ldr qW21A, [wptr1], #0x10\n" + "fmla vV22B.4s, vU43B.4s, vW32B.4s\n" + "ldr qU34A, [uptr2, u_col_stride3]\n" + "fmla vV21B.4s, vU42B.4s, vW32B.4s\n" + "ldr qW33A, [wptr2, w_col_stride2]\n" + "fmla vV22B.4s, vU42B.4s, vW31B.4s\n" + "str qV22B, [vptr1, %x[v_col_stride]]\n" + "fmla vV11B.4s, vU11B.4s, vW11B.4s\n" + "ldr qU33A, [uptr2, u_col_stride2]\n" + "fmla vV11B.4s, vU21B.4s, vW21B.4s\n" + "ldr qW32A, [wptr2, w_col_stride1]\n" + "fmla vV21B.4s, vU21B.4s, vW11B.4s\n" + "ldr qU32A, [uptr2, u_col_stride1]\n" + "fmla vV11B.4s, vU31B.4s, vW31B.4s\n" + "str qV11B, [%x[vptr0]], #0x10\n" + "fmla vV21B.4s, vU31B.4s, vW21B.4s\n" + "ldr qW31A, [wptr2], #0x10\n" + "fmla vV21B.4s, vU41B.4s, vW31B.4s\n" + "str qV21B, [vptr1], #0x10\n" + "fmul vV12A.4s, vU14A.4s, vW13A.4s\n" + "bne 1b\n" // Loop + + "2:" // Branch destination for zero loops + "cbnz %w[odd_tail], 4f\n" + + "3:" // Even number of iterations + "fmul vV11A.4s, vU13A.4s, vW13A.4s\n" + "fmla vV12A.4s, vU13A.4s, vW12A.4s\n" + "fmla vV11A.4s, vU12A.4s, vW12A.4s\n" + "fmla vV12A.4s, vU12A.4s, vW11A.4s\n" + "fmla vV12A.4s, vU24A.4s, vW23A.4s\n" + "fmul vV22A.4s, vU24A.4s, vW13A.4s\n" + "fmla vV11A.4s, vU23A.4s, vW23A.4s\n" + "ldr qU44A, [uptr3, u_col_stride3]\n" + "fmla vV12A.4s, vU23A.4s, vW22A.4s\n" + "ldr qU43A, [uptr3, u_col_stride2]\n" + "fmul vV21A.4s, vU23A.4s, vW13A.4s\n" + "ldr qU42A, [uptr3, u_col_stride1]\n" + "fmla vV22A.4s, vU23A.4s, vW12A.4s\n" + "ldr qU11A, [%x[uptr0]], #0x10\n" + "fmla vV11A.4s, vU22A.4s, vW22A.4s\n" + "ldr qU21A, [uptr1], #0x10\n" + "fmla vV12A.4s, vU22A.4s, vW21A.4s\n" + "ldr qU31A, [uptr2], #0x10\n" + "fmla vV21A.4s, vU22A.4s, vW12A.4s\n" + "ldr qU41A, [uptr3], #0x10\n" + "fmla vV22A.4s, vU22A.4s, vW11A.4s\n" + "ldr qU14B, [%x[uptr0], u_col_stride3]\n" + "fmla vV12A.4s, vU34A.4s, vW33A.4s\n" + "ldr qW13B, [%x[wptr0], w_col_stride2]\n" + "fmla vV22A.4s, vU34A.4s, vW23A.4s\n" + "ldr qU13B, [%x[uptr0], u_col_stride2]\n" + "fmla vV11A.4s, vU33A.4s, vW33A.4s\n" + "ldr qW12B, [%x[wptr0], w_col_stride1]\n" + "fmla vV12A.4s, vU33A.4s, vW32A.4s\n" + "ldr qU12B, [%x[uptr0], u_col_stride1]\n" + "fmla vV21A.4s, vU33A.4s, vW23A.4s\n" + "ldr qW11B, [%x[wptr0]], #0x10\n" + "fmla vV22A.4s, vU33A.4s, vW22A.4s\n" + "ldr qU24B, [uptr1, u_col_stride3]\n" + "fmla vV11A.4s, vU32A.4s, vW32A.4s\n" + "ldr qW23B, [wptr1, w_col_stride2]\n" + "fmla vV12A.4s, vU32A.4s, vW31A.4s\n" + "str qV12A, [%x[vptr0], %x[v_col_stride]]\n" + "fmla vV21A.4s, vU32A.4s, vW22A.4s\n" + "ldr qU23B, [uptr1, u_col_stride2]\n" + "fmla vV22A.4s, vU32A.4s, vW21A.4s\n" + "ldr qW22B, [wptr1, w_col_stride1]\n" + "fmla vV22A.4s, vU44A.4s, vW33A.4s\n" + "ldr qU22B, [uptr1, u_col_stride1]\n" + "fmla vV21A.4s, vU43A.4s, vW33A.4s\n" + "ldr qW21B, [wptr1], #0x10\n" + "fmla vV22A.4s, vU43A.4s, vW32A.4s\n" + "ldr qU34B, [uptr2, u_col_stride3]\n" + "fmla vV21A.4s, vU42A.4s, vW32A.4s\n" + "ldr qW33B, [wptr2, w_col_stride2]\n" + "fmla vV22A.4s, vU42A.4s, vW31A.4s\n" + "str qV22A, [vptr1, %x[v_col_stride]]\n" + "fmla vV11A.4s, vU11A.4s, vW11A.4s\n" + "ldr qU33B, [uptr2, u_col_stride2]\n" + "fmla vV11A.4s, vU21A.4s, vW21A.4s\n" + "ldr qW32B, [wptr2, w_col_stride1]\n" + "fmla vV21A.4s, vU21A.4s, vW11A.4s\n" + "ldr qU32B, [uptr2, u_col_stride1]\n" + "fmla vV11A.4s, vU31A.4s, vW31A.4s\n" + "str qV11A, [%x[vptr0]], #0x10\n" + "fmla vV21A.4s, vU31A.4s, vW21A.4s\n" + "ldr qW31B, [wptr2], #0x10\n" + "fmla vV21A.4s, vU41A.4s, vW31A.4s\n" + "str qV21A, [vptr1], #0x10\n" + + "fmul vV12B.4s, vU14B.4s, vW13B.4s\n" + "fmul vV11B.4s, vU13B.4s, vW13B.4s\n" + "fmla vV12B.4s, vU13B.4s, vW12B.4s\n" + "fmla vV11B.4s, vU12B.4s, vW12B.4s\n" + "fmla vV12B.4s, vU12B.4s, vW11B.4s\n" + "fmla vV12B.4s, vU24B.4s, vW23B.4s\n" + "fmul vV22B.4s, vU24B.4s, vW13B.4s\n" + "fmla vV11B.4s, vU23B.4s, vW23B.4s\n" + "ldr qU44B, [uptr3, u_col_stride3]\n" + "fmla vV12B.4s, vU23B.4s, vW22B.4s\n" + "ldr qU43B, [uptr3, u_col_stride2]\n" + "fmul vV21B.4s, vU23B.4s, vW13B.4s\n" + "ldr qU42B, [uptr3, u_col_stride1]\n" + "fmla vV22B.4s, vU23B.4s, vW12B.4s\n" + "ldr qU11B, [%x[uptr0]], #0x10\n" + "fmla vV11B.4s, vU22B.4s, vW22B.4s\n" + "ldr qU21B, [uptr1], #0x10\n" + "fmla vV12B.4s, vU22B.4s, vW21B.4s\n" + "ldr qU31B, [uptr2], #0x10\n" + "fmla vV21B.4s, vU22B.4s, vW12B.4s\n" + "ldr qU41B, [uptr3], #0x10\n" + "fmla vV22B.4s, vU22B.4s, vW11B.4s\n" + "fmla vV12B.4s, vU34B.4s, vW33B.4s\n" + "fmla vV22B.4s, vU34B.4s, vW23B.4s\n" + "fmla vV11B.4s, vU33B.4s, vW33B.4s\n" + "fmla vV12B.4s, vU33B.4s, vW32B.4s\n" + "fmla vV21B.4s, vU33B.4s, vW23B.4s\n" + "fmla vV22B.4s, vU33B.4s, vW22B.4s\n" + "fmla vV11B.4s, vU32B.4s, vW32B.4s\n" + "fmla vV12B.4s, vU32B.4s, vW31B.4s\n" + "str qV12B, [%x[vptr0], %x[v_col_stride]]\n" + "fmla vV21B.4s, vU32B.4s, vW22B.4s\n" + "fmla vV22B.4s, vU32B.4s, vW21B.4s\n" + "fmla vV22B.4s, vU44B.4s, vW33B.4s\n" + "fmla vV21B.4s, vU43B.4s, vW33B.4s\n" + "fmla vV22B.4s, vU43B.4s, vW32B.4s\n" + "fmla vV21B.4s, vU42B.4s, vW32B.4s\n" + "fmla vV22B.4s, vU42B.4s, vW31B.4s\n" + "str qV22B, [vptr1, %x[v_col_stride]]\n" + "fmla vV11B.4s, vU11B.4s, vW11B.4s\n" + "fmla vV11B.4s, vU21B.4s, vW21B.4s\n" + "fmla vV21B.4s, vU21B.4s, vW11B.4s\n" + "fmla vV11B.4s, vU31B.4s, vW31B.4s\n" + "str qV11B, [%x[vptr0]], #0x10\n" + "fmla vV21B.4s, vU31B.4s, vW21B.4s\n" + "fmla vV21B.4s, vU41B.4s, vW31B.4s\n" + "str qV21B, [vptr1], #0x10\n" + "b 5f\n" + + "4:" // Odd number of iterations + "fmul vV11A.4s, vU13A.4s, vW13A.4s\n" + "fmla vV12A.4s, vU13A.4s, vW12A.4s\n" + "fmla vV11A.4s, vU12A.4s, vW12A.4s\n" + "fmla vV12A.4s, vU12A.4s, vW11A.4s\n" + "fmla vV12A.4s, vU24A.4s, vW23A.4s\n" + "fmul vV22A.4s, vU24A.4s, vW13A.4s\n" + "fmla vV11A.4s, vU23A.4s, vW23A.4s\n" + "ldr qU44A, [uptr3, u_col_stride3]\n" + "fmla vV12A.4s, vU23A.4s, vW22A.4s\n" + "ldr qU43A, [uptr3, u_col_stride2]\n" + "fmul vV21A.4s, vU23A.4s, vW13A.4s\n" + "ldr qU42A, [uptr3, u_col_stride1]\n" + "fmla vV22A.4s, vU23A.4s, vW12A.4s\n" + "ldr qU11A, [%x[uptr0]], #0x10\n" + "fmla vV11A.4s, vU22A.4s, vW22A.4s\n" + "ldr qU21A, [uptr1], #0x10\n" + "fmla vV12A.4s, vU22A.4s, vW21A.4s\n" + "ldr qU31A, [uptr2], #0x10\n" + "fmla vV21A.4s, vU22A.4s, vW12A.4s\n" + "ldr qU41A, [uptr3], #0x10\n" + "fmla vV22A.4s, vU22A.4s, vW11A.4s\n" + "fmla vV12A.4s, vU34A.4s, vW33A.4s\n" + "fmla vV22A.4s, vU34A.4s, vW23A.4s\n" + "fmla vV11A.4s, vU33A.4s, vW33A.4s\n" + "fmla vV12A.4s, vU33A.4s, vW32A.4s\n" + "fmla vV21A.4s, vU33A.4s, vW23A.4s\n" + "fmla vV22A.4s, vU33A.4s, vW22A.4s\n" + "fmla vV11A.4s, vU32A.4s, vW32A.4s\n" + "fmla vV12A.4s, vU32A.4s, vW31A.4s\n" + "str qV12A, [%x[vptr0], %x[v_col_stride]]\n" + "fmla vV21A.4s, vU32A.4s, vW22A.4s\n" + "fmla vV22A.4s, vU32A.4s, vW21A.4s\n" + "fmla vV22A.4s, vU44A.4s, vW33A.4s\n" + "fmla vV21A.4s, vU43A.4s, vW33A.4s\n" + "fmla vV22A.4s, vU43A.4s, vW32A.4s\n" + "fmla vV21A.4s, vU42A.4s, vW32A.4s\n" + "fmla vV22A.4s, vU42A.4s, vW31A.4s\n" + "str qV22A, [vptr1, %x[v_col_stride]]\n" + "fmla vV11A.4s, vU11A.4s, vW11A.4s\n" + "fmla vV11A.4s, vU21A.4s, vW21A.4s\n" + "fmla vV21A.4s, vU21A.4s, vW11A.4s\n" + "fmla vV11A.4s, vU31A.4s, vW31A.4s\n" + "str qV11A, [%x[vptr0]], #0x10\n" + "fmla vV21A.4s, vU31A.4s, vW21A.4s\n" + "fmla vV21A.4s, vU41A.4s, vW31A.4s\n" + "str qV21A, [vptr1], #0x10\n" + + "5:" // End of method + + ".unreq qW11B\n" ".unreq qW33A\n" ".unreq qU32B\n" + ".unreq qU44B\n" ".unreq qW21A\n" ".unreq qU21B\n" ".unreq qU32A\n" + ".unreq qU43A\n" ".unreq qV21B\n" + ".unreq qU24A\n" ".unreq qU44A\n" ".unreq qU33B\n" + ".unreq qU31A\n" ".unreq qV12B\n" ".unreq qU23A\n" + ".unreq qW31B\n" ".unreq qV22A\n" ".unreq qV12A\n" ".unreq qW21B\n" + ".unreq qU22B\n" ".unreq qU34A\n" ".unreq qU13B\n" ".unreq qU13A\n" + ".unreq qU34B\n" ".unreq qU22A\n" ".unreq qU24B\n" ".unreq qU31B\n" + ".unreq qW12B\n" ".unreq qW13A\n" ".unreq qV21A\n" ".unreq qV11B\n" + ".unreq qW32A\n" ".unreq qW32B\n" ".unreq qW31A\n" ".unreq qV22B\n" + ".unreq qW11A\n" ".unreq qW13B\n" ".unreq qU14A\n" + ".unreq qU33A\n" ".unreq qW33B\n" ".unreq qW22A\n" ".unreq qU23B\n" + ".unreq qU12A\n" ".unreq qU42A\n" ".unreq qU41A\n" ".unreq qU42B\n" + ".unreq qW23A\n" ".unreq qW23B\n" ".unreq qU43B\n" ".unreq qU11A\n" + ".unreq qU12B\n" ".unreq qW12A\n" ".unreq qU41B\n" ".unreq qV11A\n" + ".unreq qW22B\n" ".unreq qU11B\n" ".unreq qU14B\n" ".unreq qU21A\n" + ".unreq vW11B\n" ".unreq vW33A\n" ".unreq vU32B\n" + ".unreq vU44B\n" ".unreq vW21A\n" ".unreq vU21B\n" ".unreq vU32A\n" + ".unreq vU43A\n" ".unreq vV21B\n" + ".unreq vU24A\n" ".unreq vU44A\n" ".unreq vU33B\n" + ".unreq vU31A\n" ".unreq vV12B\n" ".unreq vU23A\n" + ".unreq vW31B\n" ".unreq vV22A\n" ".unreq vV12A\n" ".unreq vW21B\n" + ".unreq vU22B\n" ".unreq vU34A\n" ".unreq vU13B\n" ".unreq vU13A\n" + ".unreq vU34B\n" ".unreq vU22A\n" ".unreq vU24B\n" ".unreq vU31B\n" + ".unreq vW12B\n" ".unreq vW13A\n" ".unreq vV21A\n" ".unreq vV11B\n" + ".unreq vW32A\n" ".unreq vW32B\n" ".unreq vW31A\n" ".unreq vV22B\n" + ".unreq vW11A\n" ".unreq vW13B\n" ".unreq vU14A\n" + ".unreq vU33A\n" ".unreq vW33B\n" ".unreq vW22A\n" ".unreq vU23B\n" + ".unreq vU12A\n" ".unreq vU42A\n" ".unreq vU41A\n" ".unreq vU42B\n" + ".unreq vW23A\n" ".unreq vW23B\n" ".unreq vU43B\n" ".unreq vU11A\n" + ".unreq vU12B\n" ".unreq vW12A\n" ".unreq vU41B\n" ".unreq vV11A\n" + ".unreq vW22B\n" ".unreq vU11B\n" ".unreq vU14B\n" ".unreq vU21A\n" + ".unreq u_col_stride1\n" ".unreq u_col_stride2\n" + ".unreq u_col_stride3\n" + ".unreq uptr1\n" ".unreq uptr2\n" ".unreq uptr3\n" + ".unreq wptr1\n" ".unreq wptr2\n" ".unreq vptr1\n" + ".unreq w_col_stride1\n" ".unreq w_col_stride2\n" + + : [uptr0] "+r" (uptr0), [vptr0] "+r" (vptr0), [wptr0] "+r" (wptr0), + [iters] "+r" (n_iters) + : [u_row_stride] "r" (in_row_stride * sizeof(float)), + [u_col_stride] "r" (in_col_stride * sizeof(float)), + [v_row_stride] "r" (out_row_stride * sizeof(float)), + [v_col_stride] "r" (out_col_stride * sizeof(float)), + [w_row_stride] "r" (weight_row_stride * sizeof(float)), + [w_col_stride] "r" (weight_col_stride * sizeof(float)), + [odd_tail] "r" (odd_tail) + : "x0", "x1", "x2", "x3", "x4", "x5", "x6", "x7", "x8", "cc", + "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", + "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "memory" + ); + } + + if (channels_remaining) + { + // Fall back on the unoptimised version to clean up the tail + ConvImpl::process_tile( + channels_remaining, + wptr0, weight_row_stride, weight_col_stride, + uptr0, in_row_stride, in_col_stride, + vptr0, out_row_stride, out_col_stride, + 0, 0, 0, 0, 0, 0 + ); + } +} + +#endif // __aarch64__ + +template <> +const Conv::TileFn Conv::tilefn_unpadded = ConvImpl::template process_tile; + +template <> +const Conv::TileFn Conv::tilefn_top[n_in_pad_top_fns] = { + ConvImpl::template process_tile, +}; + +template <> +const Conv::TileFn Conv::tilefn_left[n_in_pad_left_fns] = { + ConvImpl::template process_tile, +}; + +template <> +const Conv::TileFn Conv::tilefn_bottom[n_in_pad_bottom_fns][n_out_pad_bottom_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, }; +template <> +const Conv::TileFn Conv::tilefn_right[n_in_pad_right_fns][n_out_pad_right_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, +}; + +template <> +const Conv::TileFn Conv::tilefn_generic = ConvImpl::template process_tile; template class DepthwiseConvolution<2, 2, 3, 3, 1, 1, float, float>; } // namespace depthwise diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_2x2_fp32_fp32.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_2x2_fp32_fp32.cpp index 0ec5a77475..dba2330507 100644 --- a/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_2x2_fp32_fp32.cpp +++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_2x2_3x3_2x2_fp32_fp32.cpp @@ -29,1067 +29,70 @@ using Conv = DepthwiseConvolution<2, 2, 3, 3, 2, 2, float, float>; using ConvImpl = DepthwiseConvolutionImpl<2, 2, 3, 3, 2, 2, float, float>; template <> -const Conv::TileFn Conv::tile_fns - [max_in_pad_top] - [max_in_pad_left] - [max_in_pad_bottom] - [max_in_pad_right] - [max_out_pad_bottom] - [max_out_pad_right] = { - { // Input pad top = 0 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 0, 0, 0>, - Conv::template process_tile<0, 0, 0, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 0, 1, 0>, - Conv::template process_tile<0, 0, 0, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 1, 0, 0>, - Conv::template process_tile<0, 0, 0, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 1, 1, 0>, - Conv::template process_tile<0, 0, 0, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 2, 0, 0>, - Conv::template process_tile<0, 0, 0, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 2, 1, 0>, - Conv::template process_tile<0, 0, 0, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 3, 0, 0>, - Conv::template process_tile<0, 0, 0, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 3, 1, 0>, - Conv::template process_tile<0, 0, 0, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 4, 0, 0>, - Conv::template process_tile<0, 0, 0, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 4, 1, 0>, - Conv::template process_tile<0, 0, 0, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 0, 0, 0>, - Conv::template process_tile<0, 0, 1, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 0, 1, 0>, - Conv::template process_tile<0, 0, 1, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 1, 0, 0>, - Conv::template process_tile<0, 0, 1, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 1, 1, 0>, - Conv::template process_tile<0, 0, 1, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 2, 0, 0>, - Conv::template process_tile<0, 0, 1, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 2, 1, 0>, - Conv::template process_tile<0, 0, 1, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 3, 0, 0>, - Conv::template process_tile<0, 0, 1, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 3, 1, 0>, - Conv::template process_tile<0, 0, 1, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 4, 0, 0>, - Conv::template process_tile<0, 0, 1, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 4, 1, 0>, - Conv::template process_tile<0, 0, 1, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 0, 0, 0>, - Conv::template process_tile<0, 0, 2, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 0, 1, 0>, - Conv::template process_tile<0, 0, 2, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 1, 0, 0>, - Conv::template process_tile<0, 0, 2, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 1, 1, 0>, - Conv::template process_tile<0, 0, 2, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 2, 0, 0>, - Conv::template process_tile<0, 0, 2, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 2, 1, 0>, - Conv::template process_tile<0, 0, 2, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 3, 0, 0>, - Conv::template process_tile<0, 0, 2, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 3, 1, 0>, - Conv::template process_tile<0, 0, 2, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 4, 0, 0>, - Conv::template process_tile<0, 0, 2, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 4, 1, 0>, - Conv::template process_tile<0, 0, 2, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 0, 0, 0>, - Conv::template process_tile<0, 0, 3, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 0, 1, 0>, - Conv::template process_tile<0, 0, 3, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 1, 0, 0>, - Conv::template process_tile<0, 0, 3, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 1, 1, 0>, - Conv::template process_tile<0, 0, 3, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 2, 0, 0>, - Conv::template process_tile<0, 0, 3, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 2, 1, 0>, - Conv::template process_tile<0, 0, 3, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 3, 0, 0>, - Conv::template process_tile<0, 0, 3, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 3, 1, 0>, - Conv::template process_tile<0, 0, 3, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 4, 0, 0>, - Conv::template process_tile<0, 0, 3, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 4, 1, 0>, - Conv::template process_tile<0, 0, 3, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 0, 0, 0>, - Conv::template process_tile<0, 0, 4, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 0, 1, 0>, - Conv::template process_tile<0, 0, 4, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 1, 0, 0>, - Conv::template process_tile<0, 0, 4, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 1, 1, 0>, - Conv::template process_tile<0, 0, 4, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 2, 0, 0>, - Conv::template process_tile<0, 0, 4, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 2, 1, 0>, - Conv::template process_tile<0, 0, 4, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 3, 0, 0>, - Conv::template process_tile<0, 0, 4, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 3, 1, 0>, - Conv::template process_tile<0, 0, 4, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 4, 0, 0>, - Conv::template process_tile<0, 0, 4, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 4, 1, 0>, - Conv::template process_tile<0, 0, 4, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 4 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 0, 0, 0>, - Conv::template process_tile<0, 1, 0, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 0, 1, 0>, - Conv::template process_tile<0, 1, 0, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 1, 0, 0>, - Conv::template process_tile<0, 1, 0, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 1, 1, 0>, - Conv::template process_tile<0, 1, 0, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 2, 0, 0>, - Conv::template process_tile<0, 1, 0, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 2, 1, 0>, - Conv::template process_tile<0, 1, 0, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 3, 0, 0>, - Conv::template process_tile<0, 1, 0, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 3, 1, 0>, - Conv::template process_tile<0, 1, 0, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 4, 0, 0>, - Conv::template process_tile<0, 1, 0, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 4, 1, 0>, - Conv::template process_tile<0, 1, 0, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 0, 0, 0>, - Conv::template process_tile<0, 1, 1, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 0, 1, 0>, - Conv::template process_tile<0, 1, 1, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 1, 0, 0>, - Conv::template process_tile<0, 1, 1, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 1, 1, 0>, - Conv::template process_tile<0, 1, 1, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 2, 0, 0>, - Conv::template process_tile<0, 1, 1, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 2, 1, 0>, - Conv::template process_tile<0, 1, 1, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 3, 0, 0>, - Conv::template process_tile<0, 1, 1, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 3, 1, 0>, - Conv::template process_tile<0, 1, 1, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 4, 0, 0>, - Conv::template process_tile<0, 1, 1, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 4, 1, 0>, - Conv::template process_tile<0, 1, 1, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 0, 0, 0>, - Conv::template process_tile<0, 1, 2, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 0, 1, 0>, - Conv::template process_tile<0, 1, 2, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 1, 0, 0>, - Conv::template process_tile<0, 1, 2, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 1, 1, 0>, - Conv::template process_tile<0, 1, 2, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 2, 0, 0>, - Conv::template process_tile<0, 1, 2, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 2, 1, 0>, - Conv::template process_tile<0, 1, 2, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 3, 0, 0>, - Conv::template process_tile<0, 1, 2, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 3, 1, 0>, - Conv::template process_tile<0, 1, 2, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 4, 0, 0>, - Conv::template process_tile<0, 1, 2, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 4, 1, 0>, - Conv::template process_tile<0, 1, 2, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 0, 0, 0>, - Conv::template process_tile<0, 1, 3, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 0, 1, 0>, - Conv::template process_tile<0, 1, 3, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 1, 0, 0>, - Conv::template process_tile<0, 1, 3, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 1, 1, 0>, - Conv::template process_tile<0, 1, 3, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 2, 0, 0>, - Conv::template process_tile<0, 1, 3, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 2, 1, 0>, - Conv::template process_tile<0, 1, 3, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 3, 0, 0>, - Conv::template process_tile<0, 1, 3, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 3, 1, 0>, - Conv::template process_tile<0, 1, 3, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 4, 0, 0>, - Conv::template process_tile<0, 1, 3, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 4, 1, 0>, - Conv::template process_tile<0, 1, 3, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 0, 0, 0>, - Conv::template process_tile<0, 1, 4, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 0, 1, 0>, - Conv::template process_tile<0, 1, 4, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 1, 0, 0>, - Conv::template process_tile<0, 1, 4, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 1, 1, 0>, - Conv::template process_tile<0, 1, 4, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 2, 0, 0>, - Conv::template process_tile<0, 1, 4, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 2, 1, 0>, - Conv::template process_tile<0, 1, 4, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 3, 0, 0>, - Conv::template process_tile<0, 1, 4, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 3, 1, 0>, - Conv::template process_tile<0, 1, 4, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 4, 0, 0>, - Conv::template process_tile<0, 1, 4, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 4, 1, 0>, - Conv::template process_tile<0, 1, 4, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 4 - }, // Input pad left = 1 - }, // Input pad top = 0 - { // Input pad top = 1 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 0, 0, 0>, - Conv::template process_tile<1, 0, 0, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 0, 1, 0>, - Conv::template process_tile<1, 0, 0, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 1, 0, 0>, - Conv::template process_tile<1, 0, 0, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 1, 1, 0>, - Conv::template process_tile<1, 0, 0, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 2, 0, 0>, - Conv::template process_tile<1, 0, 0, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 2, 1, 0>, - Conv::template process_tile<1, 0, 0, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 3, 0, 0>, - Conv::template process_tile<1, 0, 0, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 3, 1, 0>, - Conv::template process_tile<1, 0, 0, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 4, 0, 0>, - Conv::template process_tile<1, 0, 0, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 4, 1, 0>, - Conv::template process_tile<1, 0, 0, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 0, 0, 0>, - Conv::template process_tile<1, 0, 1, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 0, 1, 0>, - Conv::template process_tile<1, 0, 1, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 1, 0, 0>, - Conv::template process_tile<1, 0, 1, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 1, 1, 0>, - Conv::template process_tile<1, 0, 1, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 2, 0, 0>, - Conv::template process_tile<1, 0, 1, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 2, 1, 0>, - Conv::template process_tile<1, 0, 1, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 3, 0, 0>, - Conv::template process_tile<1, 0, 1, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 3, 1, 0>, - Conv::template process_tile<1, 0, 1, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 4, 0, 0>, - Conv::template process_tile<1, 0, 1, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 4, 1, 0>, - Conv::template process_tile<1, 0, 1, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 0, 0, 0>, - Conv::template process_tile<1, 0, 2, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 0, 1, 0>, - Conv::template process_tile<1, 0, 2, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 1, 0, 0>, - Conv::template process_tile<1, 0, 2, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 1, 1, 0>, - Conv::template process_tile<1, 0, 2, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 2, 0, 0>, - Conv::template process_tile<1, 0, 2, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 2, 1, 0>, - Conv::template process_tile<1, 0, 2, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 3, 0, 0>, - Conv::template process_tile<1, 0, 2, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 3, 1, 0>, - Conv::template process_tile<1, 0, 2, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 4, 0, 0>, - Conv::template process_tile<1, 0, 2, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 4, 1, 0>, - Conv::template process_tile<1, 0, 2, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 0, 0, 0>, - Conv::template process_tile<1, 0, 3, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 0, 1, 0>, - Conv::template process_tile<1, 0, 3, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 1, 0, 0>, - Conv::template process_tile<1, 0, 3, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 1, 1, 0>, - Conv::template process_tile<1, 0, 3, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 2, 0, 0>, - Conv::template process_tile<1, 0, 3, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 2, 1, 0>, - Conv::template process_tile<1, 0, 3, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 3, 0, 0>, - Conv::template process_tile<1, 0, 3, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 3, 1, 0>, - Conv::template process_tile<1, 0, 3, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 4, 0, 0>, - Conv::template process_tile<1, 0, 3, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 4, 1, 0>, - Conv::template process_tile<1, 0, 3, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 0, 0, 0>, - Conv::template process_tile<1, 0, 4, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 0, 1, 0>, - Conv::template process_tile<1, 0, 4, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 1, 0, 0>, - Conv::template process_tile<1, 0, 4, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 1, 1, 0>, - Conv::template process_tile<1, 0, 4, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 2, 0, 0>, - Conv::template process_tile<1, 0, 4, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 2, 1, 0>, - Conv::template process_tile<1, 0, 4, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 3, 0, 0>, - Conv::template process_tile<1, 0, 4, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 3, 1, 0>, - Conv::template process_tile<1, 0, 4, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 4, 0, 0>, - Conv::template process_tile<1, 0, 4, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 4, 1, 0>, - Conv::template process_tile<1, 0, 4, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 4 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 0, 0, 0>, - Conv::template process_tile<1, 1, 0, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 0, 1, 0>, - Conv::template process_tile<1, 1, 0, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 1, 0, 0>, - Conv::template process_tile<1, 1, 0, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 1, 1, 0>, - Conv::template process_tile<1, 1, 0, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 2, 0, 0>, - Conv::template process_tile<1, 1, 0, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 2, 1, 0>, - Conv::template process_tile<1, 1, 0, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 3, 0, 0>, - Conv::template process_tile<1, 1, 0, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 3, 1, 0>, - Conv::template process_tile<1, 1, 0, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 4, 0, 0>, - Conv::template process_tile<1, 1, 0, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 4, 1, 0>, - Conv::template process_tile<1, 1, 0, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 0, 0, 0>, - Conv::template process_tile<1, 1, 1, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 0, 1, 0>, - Conv::template process_tile<1, 1, 1, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 1, 0, 0>, - Conv::template process_tile<1, 1, 1, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 1, 1, 0>, - Conv::template process_tile<1, 1, 1, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 2, 0, 0>, - Conv::template process_tile<1, 1, 1, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 2, 1, 0>, - Conv::template process_tile<1, 1, 1, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 3, 0, 0>, - Conv::template process_tile<1, 1, 1, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 3, 1, 0>, - Conv::template process_tile<1, 1, 1, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 4, 0, 0>, - Conv::template process_tile<1, 1, 1, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 4, 1, 0>, - Conv::template process_tile<1, 1, 1, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 0, 0, 0>, - Conv::template process_tile<1, 1, 2, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 0, 1, 0>, - Conv::template process_tile<1, 1, 2, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 1, 0, 0>, - Conv::template process_tile<1, 1, 2, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 1, 1, 0>, - Conv::template process_tile<1, 1, 2, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 2, 0, 0>, - Conv::template process_tile<1, 1, 2, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 2, 1, 0>, - Conv::template process_tile<1, 1, 2, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 3, 0, 0>, - Conv::template process_tile<1, 1, 2, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 3, 1, 0>, - Conv::template process_tile<1, 1, 2, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 4, 0, 0>, - Conv::template process_tile<1, 1, 2, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 4, 1, 0>, - Conv::template process_tile<1, 1, 2, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 0, 0, 0>, - Conv::template process_tile<1, 1, 3, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 0, 1, 0>, - Conv::template process_tile<1, 1, 3, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 1, 0, 0>, - Conv::template process_tile<1, 1, 3, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 1, 1, 0>, - Conv::template process_tile<1, 1, 3, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 2, 0, 0>, - Conv::template process_tile<1, 1, 3, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 2, 1, 0>, - Conv::template process_tile<1, 1, 3, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 3, 0, 0>, - Conv::template process_tile<1, 1, 3, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 3, 1, 0>, - Conv::template process_tile<1, 1, 3, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 4, 0, 0>, - Conv::template process_tile<1, 1, 3, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 4, 1, 0>, - Conv::template process_tile<1, 1, 3, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 0, 0, 0>, - Conv::template process_tile<1, 1, 4, 0, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 0, 1, 0>, - Conv::template process_tile<1, 1, 4, 0, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 1, 0, 0>, - Conv::template process_tile<1, 1, 4, 1, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 1, 1, 0>, - Conv::template process_tile<1, 1, 4, 1, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 2, 0, 0>, - Conv::template process_tile<1, 1, 4, 2, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 2, 1, 0>, - Conv::template process_tile<1, 1, 4, 2, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 3, 0, 0>, - Conv::template process_tile<1, 1, 4, 3, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 3, 1, 0>, - Conv::template process_tile<1, 1, 4, 3, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 4, 0, 0>, - Conv::template process_tile<1, 1, 4, 4, 0, 1>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 4, 1, 0>, - Conv::template process_tile<1, 1, 4, 4, 1, 1>, - }, // Output pad bottom = 1 - }, // Input pad right = 4 - }, // Input pad bottom = 4 - }, // Input pad left = 1 - }, // Input pad top = 1 +const Conv::TileFn Conv::tilefn_unpadded = ConvImpl::template process_tile; + +template <> +const Conv::TileFn Conv::tilefn_top[n_in_pad_top_fns] = { + ConvImpl::template process_tile, + ConvImpl::template process_tile, +}; + +template <> +const Conv::TileFn Conv::tilefn_left[n_in_pad_left_fns] = { + ConvImpl::template process_tile, + ConvImpl::template process_tile, +}; + +template <> +const Conv::TileFn Conv::tilefn_bottom[n_in_pad_bottom_fns][n_out_pad_bottom_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, +}; + +template <> +const Conv::TileFn Conv::tilefn_right[n_in_pad_right_fns][n_out_pad_right_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, }; +template <> +const Conv::TileFn Conv::tilefn_generic = ConvImpl::template process_tile; template class DepthwiseConvolution<2, 2, 3, 3, 2, 2, float, float>; } // namespace depthwise diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_1x1_fp32_fp32.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_1x1_fp32_fp32.cpp index dc3c383f99..b946e5d19e 100644 --- a/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_1x1_fp32_fp32.cpp +++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_1x1_fp32_fp32.cpp @@ -28,1148 +28,928 @@ namespace depthwise using Conv = DepthwiseConvolution<3, 3, 3, 3, 1, 1, float, float>; using ConvImpl = DepthwiseConvolutionImpl<3, 3, 3, 3, 1, 1, float, float>; +#ifdef __aarch64__ + +template <> template <> -const Conv::TileFn Conv::tile_fns - [max_in_pad_top] - [max_in_pad_left] - [max_in_pad_bottom] - [max_in_pad_right] - [max_out_pad_bottom] - [max_out_pad_right] = { - { // Input pad top = 0 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 0, 0, 0>, - Conv::template process_tile<0, 0, 0, 0, 0, 1>, - Conv::template process_tile<0, 0, 0, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 0, 1, 0>, - Conv::template process_tile<0, 0, 0, 0, 1, 1>, - Conv::template process_tile<0, 0, 0, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 0, 2, 0>, - Conv::template process_tile<0, 0, 0, 0, 2, 1>, - Conv::template process_tile<0, 0, 0, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 1, 0, 0>, - Conv::template process_tile<0, 0, 0, 1, 0, 1>, - Conv::template process_tile<0, 0, 0, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 1, 1, 0>, - Conv::template process_tile<0, 0, 0, 1, 1, 1>, - Conv::template process_tile<0, 0, 0, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 1, 2, 0>, - Conv::template process_tile<0, 0, 0, 1, 2, 1>, - Conv::template process_tile<0, 0, 0, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 2, 0, 0>, - Conv::template process_tile<0, 0, 0, 2, 0, 1>, - Conv::template process_tile<0, 0, 0, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 2, 1, 0>, - Conv::template process_tile<0, 0, 0, 2, 1, 1>, - Conv::template process_tile<0, 0, 0, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 2, 2, 0>, - Conv::template process_tile<0, 0, 0, 2, 2, 1>, - Conv::template process_tile<0, 0, 0, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 3, 0, 0>, - Conv::template process_tile<0, 0, 0, 3, 0, 1>, - Conv::template process_tile<0, 0, 0, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 3, 1, 0>, - Conv::template process_tile<0, 0, 0, 3, 1, 1>, - Conv::template process_tile<0, 0, 0, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 3, 2, 0>, - Conv::template process_tile<0, 0, 0, 3, 2, 1>, - Conv::template process_tile<0, 0, 0, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 0, 0, 0>, - Conv::template process_tile<0, 0, 1, 0, 0, 1>, - Conv::template process_tile<0, 0, 1, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 0, 1, 0>, - Conv::template process_tile<0, 0, 1, 0, 1, 1>, - Conv::template process_tile<0, 0, 1, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 0, 2, 0>, - Conv::template process_tile<0, 0, 1, 0, 2, 1>, - Conv::template process_tile<0, 0, 1, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 1, 0, 0>, - Conv::template process_tile<0, 0, 1, 1, 0, 1>, - Conv::template process_tile<0, 0, 1, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 1, 1, 0>, - Conv::template process_tile<0, 0, 1, 1, 1, 1>, - Conv::template process_tile<0, 0, 1, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 1, 2, 0>, - Conv::template process_tile<0, 0, 1, 1, 2, 1>, - Conv::template process_tile<0, 0, 1, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 2, 0, 0>, - Conv::template process_tile<0, 0, 1, 2, 0, 1>, - Conv::template process_tile<0, 0, 1, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 2, 1, 0>, - Conv::template process_tile<0, 0, 1, 2, 1, 1>, - Conv::template process_tile<0, 0, 1, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 2, 2, 0>, - Conv::template process_tile<0, 0, 1, 2, 2, 1>, - Conv::template process_tile<0, 0, 1, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 3, 0, 0>, - Conv::template process_tile<0, 0, 1, 3, 0, 1>, - Conv::template process_tile<0, 0, 1, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 3, 1, 0>, - Conv::template process_tile<0, 0, 1, 3, 1, 1>, - Conv::template process_tile<0, 0, 1, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 3, 2, 0>, - Conv::template process_tile<0, 0, 1, 3, 2, 1>, - Conv::template process_tile<0, 0, 1, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 0, 0, 0>, - Conv::template process_tile<0, 0, 2, 0, 0, 1>, - Conv::template process_tile<0, 0, 2, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 0, 1, 0>, - Conv::template process_tile<0, 0, 2, 0, 1, 1>, - Conv::template process_tile<0, 0, 2, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 0, 2, 0>, - Conv::template process_tile<0, 0, 2, 0, 2, 1>, - Conv::template process_tile<0, 0, 2, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 1, 0, 0>, - Conv::template process_tile<0, 0, 2, 1, 0, 1>, - Conv::template process_tile<0, 0, 2, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 1, 1, 0>, - Conv::template process_tile<0, 0, 2, 1, 1, 1>, - Conv::template process_tile<0, 0, 2, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 1, 2, 0>, - Conv::template process_tile<0, 0, 2, 1, 2, 1>, - Conv::template process_tile<0, 0, 2, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 2, 0, 0>, - Conv::template process_tile<0, 0, 2, 2, 0, 1>, - Conv::template process_tile<0, 0, 2, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 2, 1, 0>, - Conv::template process_tile<0, 0, 2, 2, 1, 1>, - Conv::template process_tile<0, 0, 2, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 2, 2, 0>, - Conv::template process_tile<0, 0, 2, 2, 2, 1>, - Conv::template process_tile<0, 0, 2, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 3, 0, 0>, - Conv::template process_tile<0, 0, 2, 3, 0, 1>, - Conv::template process_tile<0, 0, 2, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 3, 1, 0>, - Conv::template process_tile<0, 0, 2, 3, 1, 1>, - Conv::template process_tile<0, 0, 2, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 3, 2, 0>, - Conv::template process_tile<0, 0, 2, 3, 2, 1>, - Conv::template process_tile<0, 0, 2, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 0, 0, 0>, - Conv::template process_tile<0, 0, 3, 0, 0, 1>, - Conv::template process_tile<0, 0, 3, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 0, 1, 0>, - Conv::template process_tile<0, 0, 3, 0, 1, 1>, - Conv::template process_tile<0, 0, 3, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 0, 2, 0>, - Conv::template process_tile<0, 0, 3, 0, 2, 1>, - Conv::template process_tile<0, 0, 3, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 1, 0, 0>, - Conv::template process_tile<0, 0, 3, 1, 0, 1>, - Conv::template process_tile<0, 0, 3, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 1, 1, 0>, - Conv::template process_tile<0, 0, 3, 1, 1, 1>, - Conv::template process_tile<0, 0, 3, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 1, 2, 0>, - Conv::template process_tile<0, 0, 3, 1, 2, 1>, - Conv::template process_tile<0, 0, 3, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 2, 0, 0>, - Conv::template process_tile<0, 0, 3, 2, 0, 1>, - Conv::template process_tile<0, 0, 3, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 2, 1, 0>, - Conv::template process_tile<0, 0, 3, 2, 1, 1>, - Conv::template process_tile<0, 0, 3, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 2, 2, 0>, - Conv::template process_tile<0, 0, 3, 2, 2, 1>, - Conv::template process_tile<0, 0, 3, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 3, 0, 0>, - Conv::template process_tile<0, 0, 3, 3, 0, 1>, - Conv::template process_tile<0, 0, 3, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 3, 1, 0>, - Conv::template process_tile<0, 0, 3, 3, 1, 1>, - Conv::template process_tile<0, 0, 3, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 3, 2, 0>, - Conv::template process_tile<0, 0, 3, 3, 2, 1>, - Conv::template process_tile<0, 0, 3, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 3 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 0, 0, 0>, - Conv::template process_tile<0, 1, 0, 0, 0, 1>, - Conv::template process_tile<0, 1, 0, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 0, 1, 0>, - Conv::template process_tile<0, 1, 0, 0, 1, 1>, - Conv::template process_tile<0, 1, 0, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 0, 2, 0>, - Conv::template process_tile<0, 1, 0, 0, 2, 1>, - Conv::template process_tile<0, 1, 0, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 1, 0, 0>, - Conv::template process_tile<0, 1, 0, 1, 0, 1>, - Conv::template process_tile<0, 1, 0, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 1, 1, 0>, - Conv::template process_tile<0, 1, 0, 1, 1, 1>, - Conv::template process_tile<0, 1, 0, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 1, 2, 0>, - Conv::template process_tile<0, 1, 0, 1, 2, 1>, - Conv::template process_tile<0, 1, 0, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 2, 0, 0>, - Conv::template process_tile<0, 1, 0, 2, 0, 1>, - Conv::template process_tile<0, 1, 0, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 2, 1, 0>, - Conv::template process_tile<0, 1, 0, 2, 1, 1>, - Conv::template process_tile<0, 1, 0, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 2, 2, 0>, - Conv::template process_tile<0, 1, 0, 2, 2, 1>, - Conv::template process_tile<0, 1, 0, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 3, 0, 0>, - Conv::template process_tile<0, 1, 0, 3, 0, 1>, - Conv::template process_tile<0, 1, 0, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 3, 1, 0>, - Conv::template process_tile<0, 1, 0, 3, 1, 1>, - Conv::template process_tile<0, 1, 0, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 3, 2, 0>, - Conv::template process_tile<0, 1, 0, 3, 2, 1>, - Conv::template process_tile<0, 1, 0, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 0, 0, 0>, - Conv::template process_tile<0, 1, 1, 0, 0, 1>, - Conv::template process_tile<0, 1, 1, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 0, 1, 0>, - Conv::template process_tile<0, 1, 1, 0, 1, 1>, - Conv::template process_tile<0, 1, 1, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 0, 2, 0>, - Conv::template process_tile<0, 1, 1, 0, 2, 1>, - Conv::template process_tile<0, 1, 1, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 1, 0, 0>, - Conv::template process_tile<0, 1, 1, 1, 0, 1>, - Conv::template process_tile<0, 1, 1, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 1, 1, 0>, - Conv::template process_tile<0, 1, 1, 1, 1, 1>, - Conv::template process_tile<0, 1, 1, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 1, 2, 0>, - Conv::template process_tile<0, 1, 1, 1, 2, 1>, - Conv::template process_tile<0, 1, 1, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 2, 0, 0>, - Conv::template process_tile<0, 1, 1, 2, 0, 1>, - Conv::template process_tile<0, 1, 1, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 2, 1, 0>, - Conv::template process_tile<0, 1, 1, 2, 1, 1>, - Conv::template process_tile<0, 1, 1, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 2, 2, 0>, - Conv::template process_tile<0, 1, 1, 2, 2, 1>, - Conv::template process_tile<0, 1, 1, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 3, 0, 0>, - Conv::template process_tile<0, 1, 1, 3, 0, 1>, - Conv::template process_tile<0, 1, 1, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 3, 1, 0>, - Conv::template process_tile<0, 1, 1, 3, 1, 1>, - Conv::template process_tile<0, 1, 1, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 3, 2, 0>, - Conv::template process_tile<0, 1, 1, 3, 2, 1>, - Conv::template process_tile<0, 1, 1, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 0, 0, 0>, - Conv::template process_tile<0, 1, 2, 0, 0, 1>, - Conv::template process_tile<0, 1, 2, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 0, 1, 0>, - Conv::template process_tile<0, 1, 2, 0, 1, 1>, - Conv::template process_tile<0, 1, 2, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 0, 2, 0>, - Conv::template process_tile<0, 1, 2, 0, 2, 1>, - Conv::template process_tile<0, 1, 2, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 1, 0, 0>, - Conv::template process_tile<0, 1, 2, 1, 0, 1>, - Conv::template process_tile<0, 1, 2, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 1, 1, 0>, - Conv::template process_tile<0, 1, 2, 1, 1, 1>, - Conv::template process_tile<0, 1, 2, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 1, 2, 0>, - Conv::template process_tile<0, 1, 2, 1, 2, 1>, - Conv::template process_tile<0, 1, 2, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 2, 0, 0>, - Conv::template process_tile<0, 1, 2, 2, 0, 1>, - Conv::template process_tile<0, 1, 2, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 2, 1, 0>, - Conv::template process_tile<0, 1, 2, 2, 1, 1>, - Conv::template process_tile<0, 1, 2, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 2, 2, 0>, - Conv::template process_tile<0, 1, 2, 2, 2, 1>, - Conv::template process_tile<0, 1, 2, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 3, 0, 0>, - Conv::template process_tile<0, 1, 2, 3, 0, 1>, - Conv::template process_tile<0, 1, 2, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 3, 1, 0>, - Conv::template process_tile<0, 1, 2, 3, 1, 1>, - Conv::template process_tile<0, 1, 2, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 3, 2, 0>, - Conv::template process_tile<0, 1, 2, 3, 2, 1>, - Conv::template process_tile<0, 1, 2, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 0, 0, 0>, - Conv::template process_tile<0, 1, 3, 0, 0, 1>, - Conv::template process_tile<0, 1, 3, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 0, 1, 0>, - Conv::template process_tile<0, 1, 3, 0, 1, 1>, - Conv::template process_tile<0, 1, 3, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 0, 2, 0>, - Conv::template process_tile<0, 1, 3, 0, 2, 1>, - Conv::template process_tile<0, 1, 3, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 1, 0, 0>, - Conv::template process_tile<0, 1, 3, 1, 0, 1>, - Conv::template process_tile<0, 1, 3, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 1, 1, 0>, - Conv::template process_tile<0, 1, 3, 1, 1, 1>, - Conv::template process_tile<0, 1, 3, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 1, 2, 0>, - Conv::template process_tile<0, 1, 3, 1, 2, 1>, - Conv::template process_tile<0, 1, 3, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 2, 0, 0>, - Conv::template process_tile<0, 1, 3, 2, 0, 1>, - Conv::template process_tile<0, 1, 3, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 2, 1, 0>, - Conv::template process_tile<0, 1, 3, 2, 1, 1>, - Conv::template process_tile<0, 1, 3, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 2, 2, 0>, - Conv::template process_tile<0, 1, 3, 2, 2, 1>, - Conv::template process_tile<0, 1, 3, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 3, 0, 0>, - Conv::template process_tile<0, 1, 3, 3, 0, 1>, - Conv::template process_tile<0, 1, 3, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 3, 1, 0>, - Conv::template process_tile<0, 1, 3, 3, 1, 1>, - Conv::template process_tile<0, 1, 3, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 3, 2, 0>, - Conv::template process_tile<0, 1, 3, 3, 2, 1>, - Conv::template process_tile<0, 1, 3, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 3 - }, // Input pad left = 1 - }, // Input pad top = 0 - { // Input pad top = 1 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 0, 0, 0>, - Conv::template process_tile<1, 0, 0, 0, 0, 1>, - Conv::template process_tile<1, 0, 0, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 0, 1, 0>, - Conv::template process_tile<1, 0, 0, 0, 1, 1>, - Conv::template process_tile<1, 0, 0, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 0, 2, 0>, - Conv::template process_tile<1, 0, 0, 0, 2, 1>, - Conv::template process_tile<1, 0, 0, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 1, 0, 0>, - Conv::template process_tile<1, 0, 0, 1, 0, 1>, - Conv::template process_tile<1, 0, 0, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 1, 1, 0>, - Conv::template process_tile<1, 0, 0, 1, 1, 1>, - Conv::template process_tile<1, 0, 0, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 1, 2, 0>, - Conv::template process_tile<1, 0, 0, 1, 2, 1>, - Conv::template process_tile<1, 0, 0, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 2, 0, 0>, - Conv::template process_tile<1, 0, 0, 2, 0, 1>, - Conv::template process_tile<1, 0, 0, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 2, 1, 0>, - Conv::template process_tile<1, 0, 0, 2, 1, 1>, - Conv::template process_tile<1, 0, 0, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 2, 2, 0>, - Conv::template process_tile<1, 0, 0, 2, 2, 1>, - Conv::template process_tile<1, 0, 0, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 3, 0, 0>, - Conv::template process_tile<1, 0, 0, 3, 0, 1>, - Conv::template process_tile<1, 0, 0, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 3, 1, 0>, - Conv::template process_tile<1, 0, 0, 3, 1, 1>, - Conv::template process_tile<1, 0, 0, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 3, 2, 0>, - Conv::template process_tile<1, 0, 0, 3, 2, 1>, - Conv::template process_tile<1, 0, 0, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 0, 0, 0>, - Conv::template process_tile<1, 0, 1, 0, 0, 1>, - Conv::template process_tile<1, 0, 1, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 0, 1, 0>, - Conv::template process_tile<1, 0, 1, 0, 1, 1>, - Conv::template process_tile<1, 0, 1, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 0, 2, 0>, - Conv::template process_tile<1, 0, 1, 0, 2, 1>, - Conv::template process_tile<1, 0, 1, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 1, 0, 0>, - Conv::template process_tile<1, 0, 1, 1, 0, 1>, - Conv::template process_tile<1, 0, 1, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 1, 1, 0>, - Conv::template process_tile<1, 0, 1, 1, 1, 1>, - Conv::template process_tile<1, 0, 1, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 1, 2, 0>, - Conv::template process_tile<1, 0, 1, 1, 2, 1>, - Conv::template process_tile<1, 0, 1, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 2, 0, 0>, - Conv::template process_tile<1, 0, 1, 2, 0, 1>, - Conv::template process_tile<1, 0, 1, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 2, 1, 0>, - Conv::template process_tile<1, 0, 1, 2, 1, 1>, - Conv::template process_tile<1, 0, 1, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 2, 2, 0>, - Conv::template process_tile<1, 0, 1, 2, 2, 1>, - Conv::template process_tile<1, 0, 1, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 3, 0, 0>, - Conv::template process_tile<1, 0, 1, 3, 0, 1>, - Conv::template process_tile<1, 0, 1, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 3, 1, 0>, - Conv::template process_tile<1, 0, 1, 3, 1, 1>, - Conv::template process_tile<1, 0, 1, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 3, 2, 0>, - Conv::template process_tile<1, 0, 1, 3, 2, 1>, - Conv::template process_tile<1, 0, 1, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 0, 0, 0>, - Conv::template process_tile<1, 0, 2, 0, 0, 1>, - Conv::template process_tile<1, 0, 2, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 0, 1, 0>, - Conv::template process_tile<1, 0, 2, 0, 1, 1>, - Conv::template process_tile<1, 0, 2, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 0, 2, 0>, - Conv::template process_tile<1, 0, 2, 0, 2, 1>, - Conv::template process_tile<1, 0, 2, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 1, 0, 0>, - Conv::template process_tile<1, 0, 2, 1, 0, 1>, - Conv::template process_tile<1, 0, 2, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 1, 1, 0>, - Conv::template process_tile<1, 0, 2, 1, 1, 1>, - Conv::template process_tile<1, 0, 2, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 1, 2, 0>, - Conv::template process_tile<1, 0, 2, 1, 2, 1>, - Conv::template process_tile<1, 0, 2, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 2, 0, 0>, - Conv::template process_tile<1, 0, 2, 2, 0, 1>, - Conv::template process_tile<1, 0, 2, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 2, 1, 0>, - Conv::template process_tile<1, 0, 2, 2, 1, 1>, - Conv::template process_tile<1, 0, 2, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 2, 2, 0>, - Conv::template process_tile<1, 0, 2, 2, 2, 1>, - Conv::template process_tile<1, 0, 2, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 3, 0, 0>, - Conv::template process_tile<1, 0, 2, 3, 0, 1>, - Conv::template process_tile<1, 0, 2, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 3, 1, 0>, - Conv::template process_tile<1, 0, 2, 3, 1, 1>, - Conv::template process_tile<1, 0, 2, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 3, 2, 0>, - Conv::template process_tile<1, 0, 2, 3, 2, 1>, - Conv::template process_tile<1, 0, 2, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 0, 0, 0>, - Conv::template process_tile<1, 0, 3, 0, 0, 1>, - Conv::template process_tile<1, 0, 3, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 0, 1, 0>, - Conv::template process_tile<1, 0, 3, 0, 1, 1>, - Conv::template process_tile<1, 0, 3, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 0, 2, 0>, - Conv::template process_tile<1, 0, 3, 0, 2, 1>, - Conv::template process_tile<1, 0, 3, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 1, 0, 0>, - Conv::template process_tile<1, 0, 3, 1, 0, 1>, - Conv::template process_tile<1, 0, 3, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 1, 1, 0>, - Conv::template process_tile<1, 0, 3, 1, 1, 1>, - Conv::template process_tile<1, 0, 3, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 1, 2, 0>, - Conv::template process_tile<1, 0, 3, 1, 2, 1>, - Conv::template process_tile<1, 0, 3, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 2, 0, 0>, - Conv::template process_tile<1, 0, 3, 2, 0, 1>, - Conv::template process_tile<1, 0, 3, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 2, 1, 0>, - Conv::template process_tile<1, 0, 3, 2, 1, 1>, - Conv::template process_tile<1, 0, 3, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 2, 2, 0>, - Conv::template process_tile<1, 0, 3, 2, 2, 1>, - Conv::template process_tile<1, 0, 3, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 3, 0, 0>, - Conv::template process_tile<1, 0, 3, 3, 0, 1>, - Conv::template process_tile<1, 0, 3, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 3, 1, 0>, - Conv::template process_tile<1, 0, 3, 3, 1, 1>, - Conv::template process_tile<1, 0, 3, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 3, 2, 0>, - Conv::template process_tile<1, 0, 3, 3, 2, 1>, - Conv::template process_tile<1, 0, 3, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 3 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 0, 0, 0>, - Conv::template process_tile<1, 1, 0, 0, 0, 1>, - Conv::template process_tile<1, 1, 0, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 0, 1, 0>, - Conv::template process_tile<1, 1, 0, 0, 1, 1>, - Conv::template process_tile<1, 1, 0, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 0, 2, 0>, - Conv::template process_tile<1, 1, 0, 0, 2, 1>, - Conv::template process_tile<1, 1, 0, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 1, 0, 0>, - Conv::template process_tile<1, 1, 0, 1, 0, 1>, - Conv::template process_tile<1, 1, 0, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 1, 1, 0>, - Conv::template process_tile<1, 1, 0, 1, 1, 1>, - Conv::template process_tile<1, 1, 0, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 1, 2, 0>, - Conv::template process_tile<1, 1, 0, 1, 2, 1>, - Conv::template process_tile<1, 1, 0, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 2, 0, 0>, - Conv::template process_tile<1, 1, 0, 2, 0, 1>, - Conv::template process_tile<1, 1, 0, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 2, 1, 0>, - Conv::template process_tile<1, 1, 0, 2, 1, 1>, - Conv::template process_tile<1, 1, 0, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 2, 2, 0>, - Conv::template process_tile<1, 1, 0, 2, 2, 1>, - Conv::template process_tile<1, 1, 0, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 3, 0, 0>, - Conv::template process_tile<1, 1, 0, 3, 0, 1>, - Conv::template process_tile<1, 1, 0, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 3, 1, 0>, - Conv::template process_tile<1, 1, 0, 3, 1, 1>, - Conv::template process_tile<1, 1, 0, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 3, 2, 0>, - Conv::template process_tile<1, 1, 0, 3, 2, 1>, - Conv::template process_tile<1, 1, 0, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 0, 0, 0>, - Conv::template process_tile<1, 1, 1, 0, 0, 1>, - Conv::template process_tile<1, 1, 1, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 0, 1, 0>, - Conv::template process_tile<1, 1, 1, 0, 1, 1>, - Conv::template process_tile<1, 1, 1, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 0, 2, 0>, - Conv::template process_tile<1, 1, 1, 0, 2, 1>, - Conv::template process_tile<1, 1, 1, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 1, 0, 0>, - Conv::template process_tile<1, 1, 1, 1, 0, 1>, - Conv::template process_tile<1, 1, 1, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 1, 1, 0>, - Conv::template process_tile<1, 1, 1, 1, 1, 1>, - Conv::template process_tile<1, 1, 1, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 1, 2, 0>, - Conv::template process_tile<1, 1, 1, 1, 2, 1>, - Conv::template process_tile<1, 1, 1, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 2, 0, 0>, - Conv::template process_tile<1, 1, 1, 2, 0, 1>, - Conv::template process_tile<1, 1, 1, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 2, 1, 0>, - Conv::template process_tile<1, 1, 1, 2, 1, 1>, - Conv::template process_tile<1, 1, 1, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 2, 2, 0>, - Conv::template process_tile<1, 1, 1, 2, 2, 1>, - Conv::template process_tile<1, 1, 1, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 3, 0, 0>, - Conv::template process_tile<1, 1, 1, 3, 0, 1>, - Conv::template process_tile<1, 1, 1, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 3, 1, 0>, - Conv::template process_tile<1, 1, 1, 3, 1, 1>, - Conv::template process_tile<1, 1, 1, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 3, 2, 0>, - Conv::template process_tile<1, 1, 1, 3, 2, 1>, - Conv::template process_tile<1, 1, 1, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 0, 0, 0>, - Conv::template process_tile<1, 1, 2, 0, 0, 1>, - Conv::template process_tile<1, 1, 2, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 0, 1, 0>, - Conv::template process_tile<1, 1, 2, 0, 1, 1>, - Conv::template process_tile<1, 1, 2, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 0, 2, 0>, - Conv::template process_tile<1, 1, 2, 0, 2, 1>, - Conv::template process_tile<1, 1, 2, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 1, 0, 0>, - Conv::template process_tile<1, 1, 2, 1, 0, 1>, - Conv::template process_tile<1, 1, 2, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 1, 1, 0>, - Conv::template process_tile<1, 1, 2, 1, 1, 1>, - Conv::template process_tile<1, 1, 2, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 1, 2, 0>, - Conv::template process_tile<1, 1, 2, 1, 2, 1>, - Conv::template process_tile<1, 1, 2, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 2, 0, 0>, - Conv::template process_tile<1, 1, 2, 2, 0, 1>, - Conv::template process_tile<1, 1, 2, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 2, 1, 0>, - Conv::template process_tile<1, 1, 2, 2, 1, 1>, - Conv::template process_tile<1, 1, 2, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 2, 2, 0>, - Conv::template process_tile<1, 1, 2, 2, 2, 1>, - Conv::template process_tile<1, 1, 2, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 3, 0, 0>, - Conv::template process_tile<1, 1, 2, 3, 0, 1>, - Conv::template process_tile<1, 1, 2, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 3, 1, 0>, - Conv::template process_tile<1, 1, 2, 3, 1, 1>, - Conv::template process_tile<1, 1, 2, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 3, 2, 0>, - Conv::template process_tile<1, 1, 2, 3, 2, 1>, - Conv::template process_tile<1, 1, 2, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 0, 0, 0>, - Conv::template process_tile<1, 1, 3, 0, 0, 1>, - Conv::template process_tile<1, 1, 3, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 0, 1, 0>, - Conv::template process_tile<1, 1, 3, 0, 1, 1>, - Conv::template process_tile<1, 1, 3, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 0, 2, 0>, - Conv::template process_tile<1, 1, 3, 0, 2, 1>, - Conv::template process_tile<1, 1, 3, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 1, 0, 0>, - Conv::template process_tile<1, 1, 3, 1, 0, 1>, - Conv::template process_tile<1, 1, 3, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 1, 1, 0>, - Conv::template process_tile<1, 1, 3, 1, 1, 1>, - Conv::template process_tile<1, 1, 3, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 1, 2, 0>, - Conv::template process_tile<1, 1, 3, 1, 2, 1>, - Conv::template process_tile<1, 1, 3, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 2, 0, 0>, - Conv::template process_tile<1, 1, 3, 2, 0, 1>, - Conv::template process_tile<1, 1, 3, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 2, 1, 0>, - Conv::template process_tile<1, 1, 3, 2, 1, 1>, - Conv::template process_tile<1, 1, 3, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 2, 2, 0>, - Conv::template process_tile<1, 1, 3, 2, 2, 1>, - Conv::template process_tile<1, 1, 3, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 3, 0, 0>, - Conv::template process_tile<1, 1, 3, 3, 0, 1>, - Conv::template process_tile<1, 1, 3, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 3, 1, 0>, - Conv::template process_tile<1, 1, 3, 3, 1, 1>, - Conv::template process_tile<1, 1, 3, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 3, 2, 0>, - Conv::template process_tile<1, 1, 3, 3, 2, 1>, - Conv::template process_tile<1, 1, 3, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - }, // Input pad bottom = 3 - }, // Input pad left = 1 - }, // Input pad top = 1 +void ConvImpl::process_tile( + const int n_channels, + const float* const weights, + const int weight_row_stride, + const int weight_col_stride, + const float* const inptr, + const int in_row_stride, + const int in_col_stride, + float* const outptr, + const int out_row_stride, + const int out_col_stride, + const int, const int, const int, const int, const int, const int +) +{ + // Copy pointers + const float *uptr0 = inptr; + const float *wptr0 = weights; + float *vptr0 = outptr; + + int channels_remaining = n_channels; + if (channels_remaining >= 4) + { + // Process blocks of 4 channels at a time + int n_iters = ((channels_remaining / 4) + 1)/2 - 1; + const bool odd_tail = (channels_remaining / 4) & 1; + channels_remaining %= 4; + + asm volatile ( + "qU22B .req q0\n" "qU23B .req q0\n" "qW22A .req q0\n" + "vU22B .req v0\n" "vU23B .req v0\n" "vW22A .req v0\n" + "qV12A .req q1\n" "qW11B .req q1\n" + "vV12A .req v1\n" "vW11B .req v1\n" + "qU41A .req q2\n" "qU32B .req q2\n" "qU33A .req q2\n" "qV13B .req q2\n" + "vU41A .req v2\n" "vU32B .req v2\n" "vU33A .req v2\n" "vV13B .req v2\n" + "qU42B .req q3\n" "qU13B .req q3\n" "qU44B .req q3\n" "qU55A .req q3\n" + "vU42B .req v3\n" "vU13B .req v3\n" "vU44B .req v3\n" "vU55A .req v3\n" + "qU34B .req q4\n" "qU15A .req q4\n" "qU42A .req q4\n" "qU44A .req q4\n" "qU12B .req q4\n" + "vU34B .req v4\n" "vU15A .req v4\n" "vU42A .req v4\n" "vU44A .req v4\n" "vU12B .req v4\n" + "qU33B .req q5\n" "qU52A .req q5\n" "qW23A .req q5\n" + "vU33B .req v5\n" "vU52A .req v5\n" "vW23A .req v5\n" + "qV31A .req q6\n" "qU13A .req q6\n" "qV12B .req q6\n" + "vV31A .req v6\n" "vU13A .req v6\n" "vV12B .req v6\n" + "qU35B .req q7\n" "qU51B .req q7\n" "qV11A .req q7\n" "qU53B .req q7\n" + "vU35B .req v7\n" "vU51B .req v7\n" "vV11A .req v7\n" "vU53B .req v7\n" + "qW21A .req q8\n" "qV22B .req q8\n" + "vW21A .req v8\n" "vV22B .req v8\n" + "qV33B .req q9\n" "qU14A .req q9\n" "qV23A .req q9\n" "qU25B .req q9\n" + "vV33B .req v9\n" "vU14A .req v9\n" "vV23A .req v9\n" "vU25B .req v9\n" + "qW21B .req q10\n" "qV32A .req q10\n" "qU35A .req q10\n" + "vW21B .req v10\n" "vV32A .req v10\n" "vU35A .req v10\n" + "qV11B .req q11\n" "qU15B .req q11\n" "qV33A .req q11\n" + "vV11B .req v11\n" "vU15B .req v11\n" "vV33A .req v11\n" + "qU11B .req q12\n" "qW23B .req q12\n" "qU45A .req q12\n" + "vU11B .req v12\n" "vW23B .req v12\n" "vU45A .req v12\n" + "qW11A .req q13\n" "qU45B .req q13\n" "qU52B .req q13\n" + "vW11A .req v13\n" "vU45B .req v13\n" "vU52B .req v13\n" + "qU55B .req q14\n" "qU25A .req q14\n" "qV21A .req q14\n" + "vU55B .req v14\n" "vU25A .req v14\n" "vV21A .req v14\n" + "qU53A .req q15\n" "qV21B .req q15\n" "qU31A .req q15\n" + "vU53A .req v15\n" "vV21B .req v15\n" "vU31A .req v15\n" + "qW13B .req q16\n" "qU23A .req q16\n" + "vW13B .req v16\n" "vU23A .req v16\n" + "qW33B .req q17\n" "qW33A .req q17\n" + "vW33B .req v17\n" "vW33A .req v17\n" + "qU24B .req q18\n" "qU32A .req q18\n" "qV31B .req q18\n" "qV13A .req q18\n" + "vU24B .req v18\n" "vU32A .req v18\n" "vV31B .req v18\n" "vV13A .req v18\n" + "qU31B .req q19\n" "qU11A .req q19\n" "qU54B .req q19\n" "qU43A .req q19\n" + "vU31B .req v19\n" "vU11A .req v19\n" "vU54B .req v19\n" "vU43A .req v19\n" + "qU24A .req q20\n" "qW12B .req q20\n" "qU54A .req q20\n" + "vU24A .req v20\n" "vW12B .req v20\n" "vU54A .req v20\n" + "qV23B .req q21\n" "qW12A .req q21\n" + "vV23B .req v21\n" "vW12A .req v21\n" + "qW32A .req q22\n" "qU43B .req q22\n" + "vW32A .req v22\n" "vU43B .req v22\n" + "qW31A .req q23\n" "qV32B .req q23\n" + "vW31A .req v23\n" "vV32B .req v23\n" + "qU22A .req q24\n" "qW31B .req q24\n" + "vU22A .req v24\n" "vW31B .req v24\n" + "qU21B .req q25\n" "qV22A .req q25\n" + "vU21B .req v25\n" "vV22A .req v25\n" + "qU34A .req q26\n" "qW22B .req q26\n" "qU12A .req q26\n" + "vU34A .req v26\n" "vW22B .req v26\n" "vU12A .req v26\n" + "qW13A .req q27\n" "qU51A .req q27\n" + "vW13A .req v27\n" "vU51A .req v27\n" + "qW32B .req q28\n" + "vW32B .req v28\n" + "qU41B .req q29\n" "qU14B .req q29\n" + "vU41B .req v29\n" "vU14B .req v29\n" + "qU21A .req q30\n" + "vU21A .req v30\n" + + "uptr1 .req x0\n" + "uptr2 .req x1\n" + "uptr3 .req x2\n" + "uptr4 .req x3\n" + + "u_col_stride1 .req %x[u_col_stride]\n" + "u_col_stride2 .req x4\n" + "u_col_stride3 .req x5\n" + "u_col_stride4 .req x6\n" + + "wptr1 .req x7\n" + "wptr2 .req x8\n" + "w_col_stride1 .req %x[w_col_stride]\n" + "w_col_stride2 .req x9\n" + + "vptr1 .req x10\n" + "vptr2 .req x11\n" + "v_col_stride1 .req %x[v_col_stride]\n" + "v_col_stride2 .req x12\n" + + // Prepare strides and pointers + "add uptr1, %x[uptr0], %x[u_row_stride]\n" + "add uptr2, uptr1 , %x[u_row_stride]\n" + "add uptr3, uptr2 , %x[u_row_stride]\n" + "add uptr4, uptr3 , %x[u_row_stride]\n" + "add u_col_stride2, u_col_stride1, u_col_stride1\n" + "add u_col_stride3, u_col_stride2, u_col_stride1\n" + "add u_col_stride4, u_col_stride3, u_col_stride1\n" + + "add wptr1, %x[wptr0], %x[w_row_stride]\n" + "add wptr2, wptr1 , %x[w_row_stride]\n" + "add w_col_stride2, w_col_stride1, w_col_stride1\n" + + "add vptr1, %x[vptr0], %x[v_row_stride]\n" + "add vptr2, vptr1 , %x[v_row_stride]\n" + "add v_col_stride2, v_col_stride1, v_col_stride1\n" + + // Pre-load for A + "ldr qW13A, [%x[wptr0], w_col_stride2]\n" + "ldr qW23A, [wptr1, w_col_stride2]\n" + "ldr qW33A, [wptr2, w_col_stride2]\n" + "ldr qW12A, [%x[wptr0], w_col_stride1]\n" + "ldr qU15A, [%x[uptr0], u_col_stride4]\n" + "ldr qW22A, [wptr1, w_col_stride1]\n" + "ldr qU14A, [%x[uptr0], u_col_stride3]\n" + "ldr qW32A, [wptr2, w_col_stride1]\n" + "ldr qU13A, [%x[uptr0], u_col_stride2]\n" + "ldr qU25A, [uptr1, u_col_stride4]\n" + "ldr qU24A, [uptr1, u_col_stride3]\n" + "ldr qW11A, [%x[wptr0]], #0x10\n" + "ldr qU23A, [uptr1, u_col_stride2]\n" + "ldr qW21A, [wptr1], #0x10\n" + "ldr qW31A, [wptr2], #0x10\n" + "ldr qU34A, [uptr2, u_col_stride3]\n" + "ldr qU35A, [uptr2, u_col_stride4]\n" + + // First part of A + "fmul vV13A.4s, vU15A.4s, vW13A.4s\n" + "ldr qU33A, [uptr2, u_col_stride2]\n" + "fmul vV12A.4s, vU14A.4s, vW13A.4s\n" + "cbz %x[n_iters], 2f\n" // Jump to tail if not looping + + "1:" // Main loop, double unrolled + // A Part + "fmla vV13A.4s, vU14A.4s, vW12A.4s\n" + "ldr qU45A, [uptr3, u_col_stride4]\n" + "fmul vV11A.4s, vU13A.4s, vW13A.4s\n" + "fmla vV12A.4s, vU13A.4s, vW12A.4s\n" + "fmla vV13A.4s, vU13A.4s, vW11A.4s\n" + "ldr qU44A, [uptr3, u_col_stride3]\n" + "fmla vV13A.4s, vU25A.4s, vW23A.4s\n" + "fmul vV23A.4s, vU25A.4s, vW13A.4s\n" + "ldr qU43A, [uptr3, u_col_stride2]\n" + "fmla vV12A.4s, vU24A.4s, vW23A.4s\n" + "fmla vV13A.4s, vU24A.4s, vW22A.4s\n" + "fmul vV22A.4s, vU24A.4s, vW13A.4s\n" + "fmla vV23A.4s, vU24A.4s, vW12A.4s\n" + "ldr qU55A, [uptr4, u_col_stride4]\n" + "fmla vV11A.4s, vU23A.4s, vW23A.4s\n" + "fmla vV12A.4s, vU23A.4s, vW22A.4s\n" + "fmla vV13A.4s, vU23A.4s, vW21A.4s\n" + "fmul vV21A.4s, vU23A.4s, vW13A.4s\n" + "fmla vV22A.4s, vU23A.4s, vW12A.4s\n" + "fmla vV23A.4s, vU23A.4s, vW11A.4s\n" + "ldr qU54A, [uptr4, u_col_stride3]\n" + "fmla vV13A.4s, vU35A.4s, vW33A.4s\n" + "fmla vV23A.4s, vU35A.4s, vW23A.4s\n" + "fmul vV33A.4s, vU35A.4s, vW13A.4s\n" + "ldr qU53A, [uptr4, u_col_stride2]\n" + "fmla vV12A.4s, vU34A.4s, vW33A.4s\n" + "fmla vV13A.4s, vU34A.4s, vW32A.4s\n" + "fmla vV22A.4s, vU34A.4s, vW23A.4s\n" + "fmla vV23A.4s, vU34A.4s, vW22A.4s\n" + "fmul vV32A.4s, vU34A.4s, vW13A.4s\n" + "fmla vV33A.4s, vU34A.4s, vW12A.4s\n" + "ldr qU12A, [%x[uptr0], u_col_stride1]\n" + "fmla vV11A.4s, vU33A.4s, vW33A.4s\n" + "fmla vV12A.4s, vU33A.4s, vW32A.4s\n" + "fmla vV13A.4s, vU33A.4s, vW31A.4s\n" + "str qV13A, [%x[vptr0], v_col_stride2]\n" + "fmla vV21A.4s, vU33A.4s, vW23A.4s\n" + "fmla vV22A.4s, vU33A.4s, vW22A.4s\n" + "fmla vV23A.4s, vU33A.4s, vW21A.4s\n" + "fmul vV31A.4s, vU33A.4s, vW13A.4s\n" + "ldr qW13B, [%x[wptr0], w_col_stride2]\n" + "fmla vV32A.4s, vU33A.4s, vW12A.4s\n" + "fmla vV33A.4s, vU33A.4s, vW11A.4s\n" + "ldr qU22A, [uptr1, u_col_stride1]\n" + "fmla vV23A.4s, vU45A.4s, vW33A.4s\n" + "fmla vV33A.4s, vU45A.4s, vW23A.4s\n" + "ldr qU32A, [uptr2, u_col_stride1]\n" + "fmla vV22A.4s, vU44A.4s, vW33A.4s\n" + "fmla vV23A.4s, vU44A.4s, vW32A.4s\n" + "fmla vV32A.4s, vU44A.4s, vW23A.4s\n" + "fmla vV33A.4s, vU44A.4s, vW22A.4s\n" + "ldr qU42A, [uptr3, u_col_stride1]\n" + "fmla vV21A.4s, vU43A.4s, vW33A.4s\n" + "fmla vV22A.4s, vU43A.4s, vW32A.4s\n" + "fmla vV23A.4s, vU43A.4s, vW31A.4s\n" + "str qV23A, [vptr1, v_col_stride2]\n" + "fmla vV31A.4s, vU43A.4s, vW23A.4s\n" + "ldr qW23B, [wptr1, w_col_stride2]\n" + "fmla vV32A.4s, vU43A.4s, vW22A.4s\n" + "fmla vV33A.4s, vU43A.4s, vW21A.4s\n" + "ldr qU52A, [uptr4, u_col_stride1]\n" + "fmla vV33A.4s, vU55A.4s, vW33A.4s\n" + "ldr qU11A, [%x[uptr0]], #0x10\n" + "fmla vV32A.4s, vU54A.4s, vW33A.4s\n" + "fmla vV33A.4s, vU54A.4s, vW32A.4s\n" + "ldr qU21A, [uptr1], #0x10\n" + "fmla vV31A.4s, vU53A.4s, vW33A.4s\n" + "ldr qW33B, [wptr2, w_col_stride2]\n" + "fmla vV32A.4s, vU53A.4s, vW32A.4s\n" + "fmla vV33A.4s, vU53A.4s, vW31A.4s\n" + "str qV33A, [vptr2, v_col_stride2]\n" + "fmla vV11A.4s, vU12A.4s, vW12A.4s\n" + "ldr qU31A, [uptr2], #0x10\n" + "fmla vV12A.4s, vU12A.4s, vW11A.4s\n" + "ldr qU41A, [uptr3], #0x10\n" + "fmla vV11A.4s, vU22A.4s, vW22A.4s\n" + "ldr qU51A, [uptr4], #0x10\n" + "fmla vV12A.4s, vU22A.4s, vW21A.4s\n" + "ldr qW12B, [%x[wptr0], w_col_stride1]\n" + "fmla vV21A.4s, vU22A.4s, vW12A.4s\n" + "ldr qU15B, [%x[uptr0], u_col_stride4]\n" + "fmla vV22A.4s, vU22A.4s, vW11A.4s\n" + "ldr qW22B, [wptr1, w_col_stride1]\n" + "fmla vV11A.4s, vU32A.4s, vW32A.4s\n" + "ldr qU14B, [%x[uptr0], u_col_stride3]\n" + "fmla vV12A.4s, vU32A.4s, vW31A.4s\n" + "str qV12A, [%x[vptr0], v_col_stride1]\n" + "fmla vV21A.4s, vU32A.4s, vW22A.4s\n" + "ldr qW32B, [wptr2, w_col_stride1]\n" + "fmla vV22A.4s, vU32A.4s, vW21A.4s\n" + "ldr qU13B, [%x[uptr0], u_col_stride2]\n" + "fmla vV31A.4s, vU32A.4s, vW12A.4s\n" + "ldr qU25B, [uptr1, u_col_stride4]\n" + "fmla vV32A.4s, vU32A.4s, vW11A.4s\n" + "ldr qU24B, [uptr1, u_col_stride3]\n" + "fmla vV21A.4s, vU42A.4s, vW32A.4s\n" + "fmla vV22A.4s, vU42A.4s, vW31A.4s\n" + "str qV22A, [vptr1, v_col_stride1]\n" + "fmla vV31A.4s, vU42A.4s, vW22A.4s\n" + "fmla vV32A.4s, vU42A.4s, vW21A.4s\n" + "fmla vV31A.4s, vU52A.4s, vW32A.4s\n" + "fmla vV32A.4s, vU52A.4s, vW31A.4s\n" + "str qV32A, [vptr2, v_col_stride1]\n" + "fmla vV11A.4s, vU11A.4s, vW11A.4s\n" + "ldr qW11B, [%x[wptr0]], #0x10\n" + "fmla vV11A.4s, vU21A.4s, vW21A.4s\n" + "ldr qU23B, [uptr1, u_col_stride2]\n" + "fmla vV21A.4s, vU21A.4s, vW11A.4s\n" + "ldr qW21B, [wptr1], #0x10\n" + "fmla vV11A.4s, vU31A.4s, vW31A.4s\n" + "str qV11A, [%x[vptr0]], #0x10\n" + "fmla vV21A.4s, vU31A.4s, vW21A.4s\n" + "ldr qW31B, [wptr2], #0x10\n" + "fmla vV31A.4s, vU31A.4s, vW11A.4s\n" + "ldr qU34B, [uptr2, u_col_stride3]\n" + "fmla vV21A.4s, vU41A.4s, vW31A.4s\n" + "str qV21A, [vptr1], #0x10\n" + "fmla vV31A.4s, vU41A.4s, vW21A.4s\n" + "ldr qU35B, [uptr2, u_col_stride4]\n" + "fmla vV31A.4s, vU51A.4s, vW31A.4s\n" + "str qV31A, [vptr2], #0x10\n" + + // B Part + "fmul vV13B.4s, vU15B.4s, vW13B.4s\n" + "ldr qU33B, [uptr2, u_col_stride2]\n" + "fmul vV12B.4s, vU14B.4s, vW13B.4s\n" + "fmla vV13B.4s, vU14B.4s, vW12B.4s\n" + "ldr qU45B, [uptr3, u_col_stride4]\n" + "fmul vV11B.4s, vU13B.4s, vW13B.4s\n" + "fmla vV12B.4s, vU13B.4s, vW12B.4s\n" + "fmla vV13B.4s, vU13B.4s, vW11B.4s\n" + "ldr qU44B, [uptr3, u_col_stride3]\n" + "fmla vV13B.4s, vU25B.4s, vW23B.4s\n" + "fmul vV23B.4s, vU25B.4s, vW13B.4s\n" + "ldr qU43B, [uptr3, u_col_stride2]\n" + "fmla vV12B.4s, vU24B.4s, vW23B.4s\n" + "fmla vV13B.4s, vU24B.4s, vW22B.4s\n" + "fmul vV22B.4s, vU24B.4s, vW13B.4s\n" + "fmla vV23B.4s, vU24B.4s, vW12B.4s\n" + "ldr qU55B, [uptr4, u_col_stride4]\n" + "fmla vV11B.4s, vU23B.4s, vW23B.4s\n" + "fmla vV12B.4s, vU23B.4s, vW22B.4s\n" + "fmla vV13B.4s, vU23B.4s, vW21B.4s\n" + "fmul vV21B.4s, vU23B.4s, vW13B.4s\n" + "fmla vV22B.4s, vU23B.4s, vW12B.4s\n" + "fmla vV23B.4s, vU23B.4s, vW11B.4s\n" + "ldr qU54B, [uptr4, u_col_stride3]\n" + "fmla vV13B.4s, vU35B.4s, vW33B.4s\n" + "fmla vV23B.4s, vU35B.4s, vW23B.4s\n" + "fmul vV33B.4s, vU35B.4s, vW13B.4s\n" + "ldr qU53B, [uptr4, u_col_stride2]\n" + "fmla vV12B.4s, vU34B.4s, vW33B.4s\n" + "fmla vV13B.4s, vU34B.4s, vW32B.4s\n" + "fmla vV22B.4s, vU34B.4s, vW23B.4s\n" + "fmla vV23B.4s, vU34B.4s, vW22B.4s\n" + "fmul vV32B.4s, vU34B.4s, vW13B.4s\n" + "fmla vV33B.4s, vU34B.4s, vW12B.4s\n" + "ldr qU12B, [%x[uptr0], u_col_stride1]\n" + "fmla vV11B.4s, vU33B.4s, vW33B.4s\n" + "fmla vV12B.4s, vU33B.4s, vW32B.4s\n" + "fmla vV13B.4s, vU33B.4s, vW31B.4s\n" + "str qV13B, [%x[vptr0], v_col_stride2]\n" + "fmla vV21B.4s, vU33B.4s, vW23B.4s\n" + "fmla vV22B.4s, vU33B.4s, vW22B.4s\n" + "fmla vV23B.4s, vU33B.4s, vW21B.4s\n" + "fmul vV31B.4s, vU33B.4s, vW13B.4s\n" + "ldr qW13A, [%x[wptr0], w_col_stride2]\n" + "fmla vV32B.4s, vU33B.4s, vW12B.4s\n" + "fmla vV33B.4s, vU33B.4s, vW11B.4s\n" + "ldr qU22B, [uptr1, u_col_stride1]\n" + "fmla vV23B.4s, vU45B.4s, vW33B.4s\n" + "fmla vV33B.4s, vU45B.4s, vW23B.4s\n" + "ldr qU32B, [uptr2, u_col_stride1]\n" + "fmla vV22B.4s, vU44B.4s, vW33B.4s\n" + "fmla vV23B.4s, vU44B.4s, vW32B.4s\n" + "fmla vV32B.4s, vU44B.4s, vW23B.4s\n" + "fmla vV33B.4s, vU44B.4s, vW22B.4s\n" + "ldr qU42B, [uptr3, u_col_stride1]\n" + "fmla vV21B.4s, vU43B.4s, vW33B.4s\n" + "fmla vV22B.4s, vU43B.4s, vW32B.4s\n" + "fmla vV23B.4s, vU43B.4s, vW31B.4s\n" + "str qV23B, [vptr1, v_col_stride2]\n" + "fmla vV31B.4s, vU43B.4s, vW23B.4s\n" + "ldr qW23A, [wptr1, w_col_stride2]\n" + "fmla vV32B.4s, vU43B.4s, vW22B.4s\n" + "fmla vV33B.4s, vU43B.4s, vW21B.4s\n" + "ldr qU52B, [uptr4, u_col_stride1]\n" + "fmla vV33B.4s, vU55B.4s, vW33B.4s\n" + "ldr qU11B, [%x[uptr0]], #0x10\n" + "fmla vV32B.4s, vU54B.4s, vW33B.4s\n" + "fmla vV33B.4s, vU54B.4s, vW32B.4s\n" + "ldr qU21B, [uptr1], #0x10\n" + "fmla vV31B.4s, vU53B.4s, vW33B.4s\n" + "ldr qW33A, [wptr2, w_col_stride2]\n" + "fmla vV32B.4s, vU53B.4s, vW32B.4s\n" + "fmla vV33B.4s, vU53B.4s, vW31B.4s\n" + "str qV33B, [vptr2, v_col_stride2]\n" + "fmla vV11B.4s, vU12B.4s, vW12B.4s\n" + "ldr qU31B, [uptr2], #0x10\n" + "fmla vV12B.4s, vU12B.4s, vW11B.4s\n" + "ldr qU41B, [uptr3], #0x10\n" + "fmla vV11B.4s, vU22B.4s, vW22B.4s\n" + "ldr qU51B, [uptr4], #0x10\n" + "fmla vV12B.4s, vU22B.4s, vW21B.4s\n" + "ldr qW12A, [%x[wptr0], w_col_stride1]\n" + "fmla vV21B.4s, vU22B.4s, vW12B.4s\n" + "ldr qU15A, [%x[uptr0], u_col_stride4]\n" + "fmla vV22B.4s, vU22B.4s, vW11B.4s\n" + "ldr qW22A, [wptr1, w_col_stride1]\n" + "fmla vV11B.4s, vU32B.4s, vW32B.4s\n" + "ldr qU14A, [%x[uptr0], u_col_stride3]\n" + "fmla vV12B.4s, vU32B.4s, vW31B.4s\n" + "str qV12B, [%x[vptr0], v_col_stride1]\n" + "fmla vV21B.4s, vU32B.4s, vW22B.4s\n" + "ldr qW32A, [wptr2, w_col_stride1]\n" + "fmla vV22B.4s, vU32B.4s, vW21B.4s\n" + "ldr qU13A, [%x[uptr0], u_col_stride2]\n" + "fmla vV31B.4s, vU32B.4s, vW12B.4s\n" + "ldr qU25A, [uptr1, u_col_stride4]\n" + "fmla vV32B.4s, vU32B.4s, vW11B.4s\n" + "ldr qU24A, [uptr1, u_col_stride3]\n" + "fmla vV21B.4s, vU42B.4s, vW32B.4s\n" + "fmla vV22B.4s, vU42B.4s, vW31B.4s\n" + "str qV22B, [vptr1, v_col_stride1]\n" + "fmla vV31B.4s, vU42B.4s, vW22B.4s\n" + "fmla vV32B.4s, vU42B.4s, vW21B.4s\n" + "fmla vV31B.4s, vU52B.4s, vW32B.4s\n" + "subs %x[n_iters], %x[n_iters], #1\n" + "fmla vV32B.4s, vU52B.4s, vW31B.4s\n" + "str qV32B, [vptr2, v_col_stride1]\n" + "fmla vV11B.4s, vU11B.4s, vW11B.4s\n" + "ldr qW11A, [%x[wptr0]], #0x10\n" + "fmla vV11B.4s, vU21B.4s, vW21B.4s\n" + "ldr qU23A, [uptr1, u_col_stride2]\n" + "fmla vV21B.4s, vU21B.4s, vW11B.4s\n" + "ldr qW21A, [wptr1], #0x10\n" + "fmla vV11B.4s, vU31B.4s, vW31B.4s\n" + "str qV11B, [%x[vptr0]], #0x10\n" + "fmla vV21B.4s, vU31B.4s, vW21B.4s\n" + "ldr qW31A, [wptr2], #0x10\n" + "fmla vV31B.4s, vU31B.4s, vW11B.4s\n" + "ldr qU34A, [uptr2, u_col_stride3]\n" + "fmla vV21B.4s, vU41B.4s, vW31B.4s\n" + "str qV21B, [vptr1], #0x10\n" + "fmla vV31B.4s, vU41B.4s, vW21B.4s\n" + "ldr qU35A, [uptr2, u_col_stride4]\n" + "fmla vV31B.4s, vU51B.4s, vW31B.4s\n" + "str qV31B, [vptr2], #0x10\n" + + // First part of A + "fmul vV13A.4s, vU15A.4s, vW13A.4s\n" + "ldr qU33A, [uptr2, u_col_stride2]\n" + "fmul vV12A.4s, vU14A.4s, vW13A.4s\n" + "bne 1b\n" // Loop + + "2:" // Tail dispatch + "cbnz %w[odd_tail], 3f\n" + + // Even tail + // A Part + "fmla vV13A.4s, vU14A.4s, vW12A.4s\n" + "ldr qU45A, [uptr3, u_col_stride4]\n" + "fmul vV11A.4s, vU13A.4s, vW13A.4s\n" + "fmla vV12A.4s, vU13A.4s, vW12A.4s\n" + "fmla vV13A.4s, vU13A.4s, vW11A.4s\n" + "ldr qU44A, [uptr3, u_col_stride3]\n" + "fmla vV13A.4s, vU25A.4s, vW23A.4s\n" + "fmul vV23A.4s, vU25A.4s, vW13A.4s\n" + "ldr qU43A, [uptr3, u_col_stride2]\n" + "fmla vV12A.4s, vU24A.4s, vW23A.4s\n" + "fmla vV13A.4s, vU24A.4s, vW22A.4s\n" + "fmul vV22A.4s, vU24A.4s, vW13A.4s\n" + "fmla vV23A.4s, vU24A.4s, vW12A.4s\n" + "ldr qU55A, [uptr4, u_col_stride4]\n" + "fmla vV11A.4s, vU23A.4s, vW23A.4s\n" + "fmla vV12A.4s, vU23A.4s, vW22A.4s\n" + "fmla vV13A.4s, vU23A.4s, vW21A.4s\n" + "fmul vV21A.4s, vU23A.4s, vW13A.4s\n" + "fmla vV22A.4s, vU23A.4s, vW12A.4s\n" + "fmla vV23A.4s, vU23A.4s, vW11A.4s\n" + "ldr qU54A, [uptr4, u_col_stride3]\n" + "fmla vV13A.4s, vU35A.4s, vW33A.4s\n" + "fmla vV23A.4s, vU35A.4s, vW23A.4s\n" + "fmul vV33A.4s, vU35A.4s, vW13A.4s\n" + "ldr qU53A, [uptr4, u_col_stride2]\n" + "fmla vV12A.4s, vU34A.4s, vW33A.4s\n" + "fmla vV13A.4s, vU34A.4s, vW32A.4s\n" + "fmla vV22A.4s, vU34A.4s, vW23A.4s\n" + "fmla vV23A.4s, vU34A.4s, vW22A.4s\n" + "fmul vV32A.4s, vU34A.4s, vW13A.4s\n" + "fmla vV33A.4s, vU34A.4s, vW12A.4s\n" + "ldr qU12A, [%x[uptr0], u_col_stride1]\n" + "fmla vV11A.4s, vU33A.4s, vW33A.4s\n" + "fmla vV12A.4s, vU33A.4s, vW32A.4s\n" + "fmla vV13A.4s, vU33A.4s, vW31A.4s\n" + "str qV13A, [%x[vptr0], v_col_stride2]\n" + "fmla vV21A.4s, vU33A.4s, vW23A.4s\n" + "fmla vV22A.4s, vU33A.4s, vW22A.4s\n" + "fmla vV23A.4s, vU33A.4s, vW21A.4s\n" + "fmul vV31A.4s, vU33A.4s, vW13A.4s\n" + "ldr qW13B, [%x[wptr0], w_col_stride2]\n" + "fmla vV32A.4s, vU33A.4s, vW12A.4s\n" + "fmla vV33A.4s, vU33A.4s, vW11A.4s\n" + "ldr qU22A, [uptr1, u_col_stride1]\n" + "fmla vV23A.4s, vU45A.4s, vW33A.4s\n" + "fmla vV33A.4s, vU45A.4s, vW23A.4s\n" + "ldr qU32A, [uptr2, u_col_stride1]\n" + "fmla vV22A.4s, vU44A.4s, vW33A.4s\n" + "fmla vV23A.4s, vU44A.4s, vW32A.4s\n" + "fmla vV32A.4s, vU44A.4s, vW23A.4s\n" + "fmla vV33A.4s, vU44A.4s, vW22A.4s\n" + "ldr qU42A, [uptr3, u_col_stride1]\n" + "fmla vV21A.4s, vU43A.4s, vW33A.4s\n" + "fmla vV22A.4s, vU43A.4s, vW32A.4s\n" + "fmla vV23A.4s, vU43A.4s, vW31A.4s\n" + "str qV23A, [vptr1, v_col_stride2]\n" + "fmla vV31A.4s, vU43A.4s, vW23A.4s\n" + "ldr qW23B, [wptr1, w_col_stride2]\n" + "fmla vV32A.4s, vU43A.4s, vW22A.4s\n" + "fmla vV33A.4s, vU43A.4s, vW21A.4s\n" + "ldr qU52A, [uptr4, u_col_stride1]\n" + "fmla vV33A.4s, vU55A.4s, vW33A.4s\n" + "ldr qU11A, [%x[uptr0]], #0x10\n" + "fmla vV32A.4s, vU54A.4s, vW33A.4s\n" + "fmla vV33A.4s, vU54A.4s, vW32A.4s\n" + "ldr qU21A, [uptr1], #0x10\n" + "fmla vV31A.4s, vU53A.4s, vW33A.4s\n" + "ldr qW33B, [wptr2, w_col_stride2]\n" + "fmla vV32A.4s, vU53A.4s, vW32A.4s\n" + "fmla vV33A.4s, vU53A.4s, vW31A.4s\n" + "str qV33A, [vptr2, v_col_stride2]\n" + "fmla vV11A.4s, vU12A.4s, vW12A.4s\n" + "ldr qU31A, [uptr2], #0x10\n" + "fmla vV12A.4s, vU12A.4s, vW11A.4s\n" + "ldr qU41A, [uptr3], #0x10\n" + "fmla vV11A.4s, vU22A.4s, vW22A.4s\n" + "ldr qU51A, [uptr4], #0x10\n" + "fmla vV12A.4s, vU22A.4s, vW21A.4s\n" + "ldr qW12B, [%x[wptr0], w_col_stride1]\n" + "fmla vV21A.4s, vU22A.4s, vW12A.4s\n" + "ldr qU15B, [%x[uptr0], u_col_stride4]\n" + "fmla vV22A.4s, vU22A.4s, vW11A.4s\n" + "ldr qW22B, [wptr1, w_col_stride1]\n" + "fmla vV11A.4s, vU32A.4s, vW32A.4s\n" + "ldr qU14B, [%x[uptr0], u_col_stride3]\n" + "fmla vV12A.4s, vU32A.4s, vW31A.4s\n" + "str qV12A, [%x[vptr0], v_col_stride1]\n" + "fmla vV21A.4s, vU32A.4s, vW22A.4s\n" + "ldr qW32B, [wptr2, w_col_stride1]\n" + "fmla vV22A.4s, vU32A.4s, vW21A.4s\n" + "ldr qU13B, [%x[uptr0], u_col_stride2]\n" + "fmla vV31A.4s, vU32A.4s, vW12A.4s\n" + "ldr qU25B, [uptr1, u_col_stride4]\n" + "fmla vV32A.4s, vU32A.4s, vW11A.4s\n" + "ldr qU24B, [uptr1, u_col_stride3]\n" + "fmla vV21A.4s, vU42A.4s, vW32A.4s\n" + "fmla vV22A.4s, vU42A.4s, vW31A.4s\n" + "str qV22A, [vptr1, v_col_stride1]\n" + "fmla vV31A.4s, vU42A.4s, vW22A.4s\n" + "fmla vV32A.4s, vU42A.4s, vW21A.4s\n" + "fmla vV31A.4s, vU52A.4s, vW32A.4s\n" + "fmla vV32A.4s, vU52A.4s, vW31A.4s\n" + "str qV32A, [vptr2, v_col_stride1]\n" + "fmla vV11A.4s, vU11A.4s, vW11A.4s\n" + "ldr qW11B, [%x[wptr0]], #0x10\n" + "fmla vV11A.4s, vU21A.4s, vW21A.4s\n" + "ldr qU23B, [uptr1, u_col_stride2]\n" + "fmla vV21A.4s, vU21A.4s, vW11A.4s\n" + "ldr qW21B, [wptr1], #0x10\n" + "fmla vV11A.4s, vU31A.4s, vW31A.4s\n" + "str qV11A, [%x[vptr0]], #0x10\n" + "fmla vV21A.4s, vU31A.4s, vW21A.4s\n" + "ldr qW31B, [wptr2], #0x10\n" + "fmla vV31A.4s, vU31A.4s, vW11A.4s\n" + "ldr qU34B, [uptr2, u_col_stride3]\n" + "fmla vV21A.4s, vU41A.4s, vW31A.4s\n" + "str qV21A, [vptr1], #0x10\n" + "fmla vV31A.4s, vU41A.4s, vW21A.4s\n" + "ldr qU35B, [uptr2, u_col_stride4]\n" + "fmla vV31A.4s, vU51A.4s, vW31A.4s\n" + "str qV31A, [vptr2], #0x10\n" + + // B Part + "fmul vV13B.4s, vU15B.4s, vW13B.4s\n" + "ldr qU33B, [uptr2, u_col_stride2]\n" + "fmul vV12B.4s, vU14B.4s, vW13B.4s\n" + "fmla vV13B.4s, vU14B.4s, vW12B.4s\n" + "ldr qU45B, [uptr3, u_col_stride4]\n" + "fmul vV11B.4s, vU13B.4s, vW13B.4s\n" + "fmla vV12B.4s, vU13B.4s, vW12B.4s\n" + "fmla vV13B.4s, vU13B.4s, vW11B.4s\n" + "ldr qU44B, [uptr3, u_col_stride3]\n" + "fmla vV13B.4s, vU25B.4s, vW23B.4s\n" + "fmul vV23B.4s, vU25B.4s, vW13B.4s\n" + "ldr qU43B, [uptr3, u_col_stride2]\n" + "fmla vV12B.4s, vU24B.4s, vW23B.4s\n" + "fmla vV13B.4s, vU24B.4s, vW22B.4s\n" + "fmul vV22B.4s, vU24B.4s, vW13B.4s\n" + "fmla vV23B.4s, vU24B.4s, vW12B.4s\n" + "ldr qU55B, [uptr4, u_col_stride4]\n" + "fmla vV11B.4s, vU23B.4s, vW23B.4s\n" + "fmla vV12B.4s, vU23B.4s, vW22B.4s\n" + "fmla vV13B.4s, vU23B.4s, vW21B.4s\n" + "fmul vV21B.4s, vU23B.4s, vW13B.4s\n" + "fmla vV22B.4s, vU23B.4s, vW12B.4s\n" + "fmla vV23B.4s, vU23B.4s, vW11B.4s\n" + "ldr qU54B, [uptr4, u_col_stride3]\n" + "fmla vV13B.4s, vU35B.4s, vW33B.4s\n" + "fmla vV23B.4s, vU35B.4s, vW23B.4s\n" + "fmul vV33B.4s, vU35B.4s, vW13B.4s\n" + "ldr qU53B, [uptr4, u_col_stride2]\n" + "fmla vV12B.4s, vU34B.4s, vW33B.4s\n" + "fmla vV13B.4s, vU34B.4s, vW32B.4s\n" + "fmla vV22B.4s, vU34B.4s, vW23B.4s\n" + "fmla vV23B.4s, vU34B.4s, vW22B.4s\n" + "fmul vV32B.4s, vU34B.4s, vW13B.4s\n" + "fmla vV33B.4s, vU34B.4s, vW12B.4s\n" + "ldr qU12B, [%x[uptr0], u_col_stride1]\n" + "fmla vV11B.4s, vU33B.4s, vW33B.4s\n" + "fmla vV12B.4s, vU33B.4s, vW32B.4s\n" + "fmla vV13B.4s, vU33B.4s, vW31B.4s\n" + "str qV13B, [%x[vptr0], v_col_stride2]\n" + "fmla vV21B.4s, vU33B.4s, vW23B.4s\n" + "fmla vV22B.4s, vU33B.4s, vW22B.4s\n" + "fmla vV23B.4s, vU33B.4s, vW21B.4s\n" + "fmul vV31B.4s, vU33B.4s, vW13B.4s\n" + "fmla vV32B.4s, vU33B.4s, vW12B.4s\n" + "fmla vV33B.4s, vU33B.4s, vW11B.4s\n" + "ldr qU22B, [uptr1, u_col_stride1]\n" + "fmla vV23B.4s, vU45B.4s, vW33B.4s\n" + "fmla vV33B.4s, vU45B.4s, vW23B.4s\n" + "ldr qU32B, [uptr2, u_col_stride1]\n" + "fmla vV22B.4s, vU44B.4s, vW33B.4s\n" + "fmla vV23B.4s, vU44B.4s, vW32B.4s\n" + "fmla vV32B.4s, vU44B.4s, vW23B.4s\n" + "fmla vV33B.4s, vU44B.4s, vW22B.4s\n" + "ldr qU42B, [uptr3, u_col_stride1]\n" + "fmla vV21B.4s, vU43B.4s, vW33B.4s\n" + "fmla vV22B.4s, vU43B.4s, vW32B.4s\n" + "fmla vV23B.4s, vU43B.4s, vW31B.4s\n" + "str qV23B, [vptr1, v_col_stride2]\n" + "fmla vV31B.4s, vU43B.4s, vW23B.4s\n" + "fmla vV32B.4s, vU43B.4s, vW22B.4s\n" + "fmla vV33B.4s, vU43B.4s, vW21B.4s\n" + "ldr qU52B, [uptr4, u_col_stride1]\n" + "fmla vV33B.4s, vU55B.4s, vW33B.4s\n" + "ldr qU11B, [%x[uptr0]], #0x10\n" + "fmla vV32B.4s, vU54B.4s, vW33B.4s\n" + "fmla vV33B.4s, vU54B.4s, vW32B.4s\n" + "ldr qU21B, [uptr1], #0x10\n" + "fmla vV31B.4s, vU53B.4s, vW33B.4s\n" + "fmla vV32B.4s, vU53B.4s, vW32B.4s\n" + "fmla vV33B.4s, vU53B.4s, vW31B.4s\n" + "str qV33B, [vptr2, v_col_stride2]\n" + "fmla vV11B.4s, vU12B.4s, vW12B.4s\n" + "ldr qU31B, [uptr2], #0x10\n" + "fmla vV12B.4s, vU12B.4s, vW11B.4s\n" + "ldr qU41B, [uptr3], #0x10\n" + "fmla vV11B.4s, vU22B.4s, vW22B.4s\n" + "ldr qU51B, [uptr4], #0x10\n" + "fmla vV12B.4s, vU22B.4s, vW21B.4s\n" + "fmla vV21B.4s, vU22B.4s, vW12B.4s\n" + "fmla vV22B.4s, vU22B.4s, vW11B.4s\n" + "fmla vV11B.4s, vU32B.4s, vW32B.4s\n" + "fmla vV12B.4s, vU32B.4s, vW31B.4s\n" + "str qV12B, [%x[vptr0], v_col_stride1]\n" + "fmla vV21B.4s, vU32B.4s, vW22B.4s\n" + "fmla vV22B.4s, vU32B.4s, vW21B.4s\n" + "fmla vV31B.4s, vU32B.4s, vW12B.4s\n" + "fmla vV32B.4s, vU32B.4s, vW11B.4s\n" + "fmla vV21B.4s, vU42B.4s, vW32B.4s\n" + "fmla vV22B.4s, vU42B.4s, vW31B.4s\n" + "str qV22B, [vptr1, v_col_stride1]\n" + "fmla vV31B.4s, vU42B.4s, vW22B.4s\n" + "fmla vV32B.4s, vU42B.4s, vW21B.4s\n" + "fmla vV31B.4s, vU52B.4s, vW32B.4s\n" + "subs %x[n_iters], %x[n_iters], #1\n" + "fmla vV32B.4s, vU52B.4s, vW31B.4s\n" + "str qV32B, [vptr2, v_col_stride1]\n" + "fmla vV11B.4s, vU11B.4s, vW11B.4s\n" + "fmla vV11B.4s, vU21B.4s, vW21B.4s\n" + "fmla vV21B.4s, vU21B.4s, vW11B.4s\n" + "fmla vV11B.4s, vU31B.4s, vW31B.4s\n" + "str qV11B, [%x[vptr0]], #0x10\n" + "fmla vV21B.4s, vU31B.4s, vW21B.4s\n" + "fmla vV31B.4s, vU31B.4s, vW11B.4s\n" + "fmla vV21B.4s, vU41B.4s, vW31B.4s\n" + "str qV21B, [vptr1], #0x10\n" + "fmla vV31B.4s, vU41B.4s, vW21B.4s\n" + "fmla vV31B.4s, vU51B.4s, vW31B.4s\n" + "str qV31B, [vptr2], #0x10\n" + + "b 4f\n" // Branch to end of method + + "3:" // Odd tail, finish off A + "fmla vV13A.4s, vU14A.4s, vW12A.4s\n" + "ldr qU45A, [uptr3, u_col_stride4]\n" + "fmul vV11A.4s, vU13A.4s, vW13A.4s\n" + "fmla vV12A.4s, vU13A.4s, vW12A.4s\n" + "fmla vV13A.4s, vU13A.4s, vW11A.4s\n" + "ldr qU44A, [uptr3, u_col_stride3]\n" + "fmla vV13A.4s, vU25A.4s, vW23A.4s\n" + "fmul vV23A.4s, vU25A.4s, vW13A.4s\n" + "ldr qU43A, [uptr3, u_col_stride2]\n" + "fmla vV12A.4s, vU24A.4s, vW23A.4s\n" + "fmla vV13A.4s, vU24A.4s, vW22A.4s\n" + "fmul vV22A.4s, vU24A.4s, vW13A.4s\n" + "fmla vV23A.4s, vU24A.4s, vW12A.4s\n" + "ldr qU55A, [uptr4, u_col_stride4]\n" + "fmla vV11A.4s, vU23A.4s, vW23A.4s\n" + "fmla vV12A.4s, vU23A.4s, vW22A.4s\n" + "fmla vV13A.4s, vU23A.4s, vW21A.4s\n" + "fmul vV21A.4s, vU23A.4s, vW13A.4s\n" + "fmla vV22A.4s, vU23A.4s, vW12A.4s\n" + "fmla vV23A.4s, vU23A.4s, vW11A.4s\n" + "ldr qU54A, [uptr4, u_col_stride3]\n" + "fmla vV13A.4s, vU35A.4s, vW33A.4s\n" + "fmla vV23A.4s, vU35A.4s, vW23A.4s\n" + "fmul vV33A.4s, vU35A.4s, vW13A.4s\n" + "ldr qU53A, [uptr4, u_col_stride2]\n" + "fmla vV12A.4s, vU34A.4s, vW33A.4s\n" + "fmla vV13A.4s, vU34A.4s, vW32A.4s\n" + "fmla vV22A.4s, vU34A.4s, vW23A.4s\n" + "fmla vV23A.4s, vU34A.4s, vW22A.4s\n" + "fmul vV32A.4s, vU34A.4s, vW13A.4s\n" + "fmla vV33A.4s, vU34A.4s, vW12A.4s\n" + "ldr qU12A, [%x[uptr0], u_col_stride1]\n" + "fmla vV11A.4s, vU33A.4s, vW33A.4s\n" + "fmla vV12A.4s, vU33A.4s, vW32A.4s\n" + "fmla vV13A.4s, vU33A.4s, vW31A.4s\n" + "str qV13A, [%x[vptr0], v_col_stride2]\n" + "fmla vV21A.4s, vU33A.4s, vW23A.4s\n" + "fmla vV22A.4s, vU33A.4s, vW22A.4s\n" + "fmla vV23A.4s, vU33A.4s, vW21A.4s\n" + "fmul vV31A.4s, vU33A.4s, vW13A.4s\n" + "fmla vV32A.4s, vU33A.4s, vW12A.4s\n" + "fmla vV33A.4s, vU33A.4s, vW11A.4s\n" + "ldr qU22A, [uptr1, u_col_stride1]\n" + "fmla vV23A.4s, vU45A.4s, vW33A.4s\n" + "fmla vV33A.4s, vU45A.4s, vW23A.4s\n" + "ldr qU32A, [uptr2, u_col_stride1]\n" + "fmla vV22A.4s, vU44A.4s, vW33A.4s\n" + "fmla vV23A.4s, vU44A.4s, vW32A.4s\n" + "fmla vV32A.4s, vU44A.4s, vW23A.4s\n" + "fmla vV33A.4s, vU44A.4s, vW22A.4s\n" + "ldr qU42A, [uptr3, u_col_stride1]\n" + "fmla vV21A.4s, vU43A.4s, vW33A.4s\n" + "fmla vV22A.4s, vU43A.4s, vW32A.4s\n" + "fmla vV23A.4s, vU43A.4s, vW31A.4s\n" + "str qV23A, [vptr1, v_col_stride2]\n" + "fmla vV31A.4s, vU43A.4s, vW23A.4s\n" + "fmla vV32A.4s, vU43A.4s, vW22A.4s\n" + "fmla vV33A.4s, vU43A.4s, vW21A.4s\n" + "ldr qU52A, [uptr4, u_col_stride1]\n" + "fmla vV33A.4s, vU55A.4s, vW33A.4s\n" + "ldr qU11A, [%x[uptr0]], #0x10\n" + "fmla vV32A.4s, vU54A.4s, vW33A.4s\n" + "fmla vV33A.4s, vU54A.4s, vW32A.4s\n" + "ldr qU21A, [uptr1], #0x10\n" + "fmla vV31A.4s, vU53A.4s, vW33A.4s\n" + "fmla vV32A.4s, vU53A.4s, vW32A.4s\n" + "fmla vV33A.4s, vU53A.4s, vW31A.4s\n" + "str qV33A, [vptr2, v_col_stride2]\n" + "fmla vV11A.4s, vU12A.4s, vW12A.4s\n" + "ldr qU31A, [uptr2], #0x10\n" + "fmla vV12A.4s, vU12A.4s, vW11A.4s\n" + "ldr qU41A, [uptr3], #0x10\n" + "fmla vV11A.4s, vU22A.4s, vW22A.4s\n" + "ldr qU51A, [uptr4], #0x10\n" + "fmla vV12A.4s, vU22A.4s, vW21A.4s\n" + "fmla vV21A.4s, vU22A.4s, vW12A.4s\n" + "fmla vV22A.4s, vU22A.4s, vW11A.4s\n" + "fmla vV11A.4s, vU32A.4s, vW32A.4s\n" + "fmla vV12A.4s, vU32A.4s, vW31A.4s\n" + "str qV12A, [%x[vptr0], v_col_stride1]\n" + "fmla vV21A.4s, vU32A.4s, vW22A.4s\n" + "fmla vV22A.4s, vU32A.4s, vW21A.4s\n" + "fmla vV31A.4s, vU32A.4s, vW12A.4s\n" + "fmla vV32A.4s, vU32A.4s, vW11A.4s\n" + "fmla vV21A.4s, vU42A.4s, vW32A.4s\n" + "fmla vV22A.4s, vU42A.4s, vW31A.4s\n" + "str qV22A, [vptr1, v_col_stride1]\n" + "fmla vV31A.4s, vU42A.4s, vW22A.4s\n" + "fmla vV32A.4s, vU42A.4s, vW21A.4s\n" + "fmla vV31A.4s, vU52A.4s, vW32A.4s\n" + "fmla vV32A.4s, vU52A.4s, vW31A.4s\n" + "str qV32A, [vptr2, v_col_stride1]\n" + "fmla vV11A.4s, vU11A.4s, vW11A.4s\n" + "fmla vV11A.4s, vU21A.4s, vW21A.4s\n" + "fmla vV21A.4s, vU21A.4s, vW11A.4s\n" + "fmla vV11A.4s, vU31A.4s, vW31A.4s\n" + "str qV11A, [%x[vptr0]], #0x10\n" + "fmla vV21A.4s, vU31A.4s, vW21A.4s\n" + "fmla vV31A.4s, vU31A.4s, vW11A.4s\n" + "fmla vV21A.4s, vU41A.4s, vW31A.4s\n" + "str qV21A, [vptr1], #0x10\n" + "fmla vV31A.4s, vU41A.4s, vW21A.4s\n" + "fmla vV31A.4s, vU51A.4s, vW31A.4s\n" + "str qV31A, [vptr2], #0x10\n" + + "4:" // End of method + ".unreq uptr1\n" ".unreq uptr2\n" ".unreq uptr3\n" ".unreq uptr4\n" + ".unreq u_col_stride1\n" ".unreq u_col_stride2\n" + ".unreq u_col_stride3\n" ".unreq u_col_stride4\n" + ".unreq wptr1\n" ".unreq wptr2\n" + ".unreq w_col_stride1\n" ".unreq w_col_stride2\n" + ".unreq vptr1\n" ".unreq vptr2\n" + ".unreq v_col_stride1\n" ".unreq v_col_stride2\n" + + ".unreq qU22B\n" ".unreq qW13B\n" ".unreq qW13A\n" ".unreq qU51B\n" + ".unreq qU54B\n" ".unreq qU45A\n" ".unreq qU15A\n" ".unreq qU41B\n" + ".unreq qU24B\n" ".unreq qU21A\n" + ".unreq qV11B\n" ".unreq qU51A\n" ".unreq qU35A\n" ".unreq qU12A\n" + ".unreq qU42B\n" ".unreq qU44B\n" ".unreq qU13B\n" ".unreq qW33A\n" + ".unreq qV31B\n" ".unreq qV23A\n" ".unreq qU31A\n" ".unreq qU35B\n" ".unreq qU13A\n" + ".unreq qV23B\n" ".unreq qU11A\n" ".unreq qU25A\n" ".unreq qU43A\n" ".unreq qU52B\n" + ".unreq qU24A\n" ".unreq qU23B\n" ".unreq qV21A\n" ".unreq qV32B\n" + ".unreq qV33B\n" ".unreq qW11A\n" ".unreq qU31B\n" + ".unreq qW12B\n" ".unreq qU33A\n" ".unreq qU14A\n" ".unreq qU22A\n" + ".unreq qU25B\n" ".unreq qU53B\n" ".unreq qU42A\n" ".unreq qU44A\n" + ".unreq qU43B\n" ".unreq qW31A\n" ".unreq qU11B\n" + ".unreq qW11B\n" ".unreq qW32A\n" + ".unreq qU12B\n" ".unreq qU34B\n" ".unreq qW21A\n" + ".unreq qU14B\n" ".unreq qV21B\n" ".unreq qW22A\n" + ".unreq qW23B\n" ".unreq qW23A\n" ".unreq qU21B\n" + ".unreq qU32B\n" ".unreq qU34A\n" ".unreq qU45B\n" ".unreq qV31A\n" + ".unreq qW12A\n" ".unreq qU33B\n" ".unreq qU15B\n" + ".unreq qW33B\n" ".unreq qU54A\n" ".unreq qU23A\n" + ".unreq qW32B\n" ".unreq qV33A\n" ".unreq qW31B\n" ".unreq qV12A\n" + ".unreq qV12B\n" ".unreq qU41A\n" ".unreq qU53A\n" + ".unreq qV13A\n" ".unreq qU32A\n" ".unreq qW22B\n" + ".unreq qV22B\n" ".unreq qU52A\n" ".unreq qV13B\n" ".unreq qV32A\n" + ".unreq qU55A\n" ".unreq qU55B\n" ".unreq qV22A\n" ".unreq qW21B\n" + ".unreq qV11A\n" + ".unreq vU22B\n" ".unreq vW13B\n" ".unreq vW13A\n" ".unreq vU51B\n" + ".unreq vU54B\n" ".unreq vU45A\n" ".unreq vU15A\n" ".unreq vU41B\n" + ".unreq vU24B\n" ".unreq vU21A\n" + ".unreq vV11B\n" ".unreq vU51A\n" ".unreq vU35A\n" ".unreq vU12A\n" + ".unreq vU42B\n" ".unreq vU44B\n" ".unreq vU13B\n" ".unreq vW33A\n" + ".unreq vV31B\n" ".unreq vV23A\n" ".unreq vU31A\n" ".unreq vU35B\n" ".unreq vU13A\n" + ".unreq vV23B\n" ".unreq vU11A\n" ".unreq vU25A\n" ".unreq vU43A\n" ".unreq vU52B\n" + ".unreq vU24A\n" ".unreq vU23B\n" ".unreq vV21A\n" ".unreq vV32B\n" + ".unreq vV33B\n" ".unreq vW11A\n" ".unreq vU31B\n" + ".unreq vW12B\n" ".unreq vU33A\n" ".unreq vU14A\n" ".unreq vU22A\n" + ".unreq vU25B\n" ".unreq vU53B\n" ".unreq vU42A\n" ".unreq vU44A\n" + ".unreq vU43B\n" ".unreq vW31A\n" ".unreq vU11B\n" + ".unreq vW11B\n" ".unreq vW32A\n" + ".unreq vU12B\n" ".unreq vU34B\n" ".unreq vW21A\n" + ".unreq vU14B\n" ".unreq vV21B\n" ".unreq vW22A\n" + ".unreq vW23B\n" ".unreq vW23A\n" ".unreq vU21B\n" + ".unreq vU32B\n" ".unreq vU34A\n" ".unreq vU45B\n" ".unreq vV31A\n" + ".unreq vW12A\n" ".unreq vU33B\n" ".unreq vU15B\n" + ".unreq vW33B\n" ".unreq vU54A\n" ".unreq vU23A\n" + ".unreq vW32B\n" ".unreq vV33A\n" ".unreq vW31B\n" ".unreq vV12A\n" + ".unreq vV12B\n" ".unreq vU41A\n" ".unreq vU53A\n" + ".unreq vV13A\n" ".unreq vU32A\n" ".unreq vW22B\n" + ".unreq vV22B\n" ".unreq vU52A\n" ".unreq vV13B\n" ".unreq vV32A\n" + ".unreq vU55A\n" ".unreq vU55B\n" ".unreq vV22A\n" ".unreq vW21B\n" + ".unreq vV11A\n" + : [uptr0] "+r" (uptr0), [wptr0] "+r" (wptr0), [vptr0] "+r" (vptr0), + [n_iters] "+r" (n_iters) + : [u_row_stride] "r" (in_row_stride * sizeof(float)), + [u_col_stride] "r" (in_col_stride * sizeof(float)), + [w_row_stride] "r" (weight_row_stride * sizeof(float)), + [w_col_stride] "r" (weight_col_stride * sizeof(float)), + [v_row_stride] "r" (out_row_stride * sizeof(float)), + [v_col_stride] "r" (out_col_stride * sizeof(float)), + [odd_tail] "r" (odd_tail) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", + "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", + "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "x0", + "x1", "x2", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", + "x12", "cc", "memory" + ); + } + if (channels_remaining) + { + // Fall back on the unoptimised version to clean up the tail + ConvImpl::process_tile( + channels_remaining, + wptr0, weight_row_stride, weight_col_stride, + uptr0, in_row_stride, in_col_stride, + vptr0, out_row_stride, out_col_stride, + 0, 0, 0, 0, 0, 0 + ); + } +} + +#endif // __aarch64__ + +template <> +const Conv::TileFn Conv::tilefn_unpadded = ConvImpl::template process_tile; + +template <> +const Conv::TileFn Conv::tilefn_top[n_in_pad_top_fns] = { + ConvImpl::template process_tile, }; +template <> +const Conv::TileFn Conv::tilefn_left[n_in_pad_left_fns] = { + ConvImpl::template process_tile, +}; + +template <> +const Conv::TileFn Conv::tilefn_bottom[n_in_pad_bottom_fns][n_out_pad_bottom_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, +}; + +template <> +const Conv::TileFn Conv::tilefn_right[n_in_pad_right_fns][n_out_pad_right_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, +}; + +template <> +const Conv::TileFn Conv::tilefn_generic = ConvImpl::template process_tile; template class DepthwiseConvolution<3, 3, 3, 3, 1, 1, float, float>; } // namespace depthwise diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_2x2_fp32_fp32.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_2x2_fp32_fp32.cpp index 8d511b1a6c..2510941f35 100644 --- a/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_2x2_fp32_fp32.cpp +++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_2x2_fp32_fp32.cpp @@ -28,3416 +28,596 @@ namespace depthwise using Conv = DepthwiseConvolution<3, 3, 3, 3, 2, 2, float, float>; using ConvImpl = DepthwiseConvolutionImpl<3, 3, 3, 3, 2, 2, float, float>; +#ifdef __aarch64__ + +template <> +template <> +void ConvImpl::process_tile( + const int n_channels, + const float* const weights, + const int weight_row_stride, + const int weight_col_stride, + const float* const inptr, + const int in_row_stride, + const int in_col_stride, + float* const outptr, + const int out_row_stride, + const int out_col_stride, + const int, const int, const int, const int, const int, const int +) +{ + // Copy pointers + const float *uptr0 = inptr; + const float *wptr0 = weights; + float *vptr0 = outptr; + + int channels_remaining = n_channels; + if (channels_remaining >= 4) + { + // Process blocks of 4 channels at a time + int n_iters = channels_remaining / 4 - 1; + channels_remaining %= 4; + + asm volatile( + // Prepare aliases + "qW13 .req q0\n" "vW13 .req v0\n" + "qU15 .req q1\n" "qU73 .req q1\n" "qU45 .req q1\n" "qU14 .req q1\n" + "vU15 .req v1\n" "vU73 .req v1\n" "vU45 .req v1\n" "vU14 .req v1\n" + "qU62 .req q2\n" "qV12 .req q2\n" "vU62 .req v2\n" "vV12 .req v2\n" + "qU51 .req q3\n" "qU43 .req q3\n" "qU55 .req q3\n" + "vU51 .req v3\n" "vU43 .req v3\n" "vU55 .req v3\n" + "qU77 .req q4\n" "qV13 .req q4\n" "qV31 .req q4\n" "qU44 .req q4\n" + "vU77 .req v4\n" "vV13 .req v4\n" "vV31 .req v4\n" "vU44 .req v4\n" + "qV33 .req q5\n" "qU46 .req q5\n" "qU11 .req q5\n" "qU37 .req q5\n" + "vV33 .req v5\n" "vU46 .req v5\n" "vU11 .req v5\n" "vU37 .req v5\n" + "qU56 .req q6\n" "qU25 .req q6\n" "qU32 .req q6\n" + "vU56 .req v6\n" "vU25 .req v6\n" "vU32 .req v6\n" + "qU72 .req q7\n" "qV22 .req q7\n" "vU72 .req v7\n" "vV22 .req v7\n" + "qU67 .req q8\n" "qU61 .req q8\n" "qU13 .req q8\n" + "vU67 .req v8\n" "vU61 .req v8\n" "vU13 .req v8\n" + "qU74 .req q9\n" "qU34 .req q9\n" "qU17 .req q9\n" "qU66 .req q9\n" + "vU74 .req v9\n" "vU34 .req v9\n" "vU17 .req v9\n" "vU66 .req v9\n" + "qU33 .req q10\n" "qU57 .req q10\n" "qU21 .req q10\n" + "vU33 .req v10\n" "vU57 .req v10\n" "vU21 .req v10\n" "qW23 .req q11\n" + "vW23 .req v11\n" "qU42 .req q12\n" "qV23 .req q12\n" "qU23 .req q12\n" + "vU42 .req v12\n" "vV23 .req v12\n" "vU23 .req v12\n" + "qW33 .req q13\n" "vW33 .req v13\n" + "qU76 .req q14\n" "qU47 .req q14\n" "qU64 .req q14\n" "qU41 .req q14\n" + "vU76 .req v14\n" "vU47 .req v14\n" "vU64 .req v14\n" "vU41 .req v14\n" + "qU52 .req q15\n" "qU54 .req q15\n" "qU75 .req q15\n" "qU26 .req q15\n" + "vU52 .req v15\n" "vU54 .req v15\n" "vU75 .req v15\n" "vU26 .req v15\n" + "qU53 .req q16\n" "qU27 .req q16\n" "vU53 .req v16\n" "vU27 .req v16\n" + "qV21 .req q17\n" "qU65 .req q17\n" "vV21 .req v17\n" "vU65 .req v17\n" + "qU31 .req q18\n" "qU24 .req q18\n" "qU36 .req q18\n" + "vU31 .req v18\n" "vU24 .req v18\n" "vU36 .req v18\n" "qU22 .req q19\n" + "vU22 .req v19\n" "qU35 .req q20\n" "qU63 .req q20\n" + "vU35 .req v20\n" "vU63 .req v20\n" "qW12 .req q21\n" + "vW12 .req v21\n" "qV32 .req q22\n" "qU16 .req q22\n" + "vV32 .req v22\n" "vU16 .req v22\n" "qW11 .req q23\n" "vW11 .req v23\n" + "qU12 .req q24\n" "vU12 .req v24\n" "qW31 .req q25\n" "vW31 .req v25\n" + "qW22 .req q26\n" "vW22 .req v26\n" "qU71 .req q27\n" "vU71 .req v27\n" + "qV11 .req q28\n" "vV11 .req v28\n" "qW21 .req q29\n" "vW21 .req v29\n" + "qW32 .req q30\n" "vW32 .req v30\n" + + "uptr1 .req x0\n" + "uptr2 .req x1\n" + "uptr3 .req x2\n" + "uptr4 .req x3\n" + "uptr5 .req x4\n" + "uptr6 .req x5\n" + "u_col_stride1 .req %x[u_col_stride]\n" + "u_col_stride2 .req x6\n" + "u_col_stride3 .req x7\n" + "u_col_stride4 .req x8\n" + "u_col_stride5 .req x9\n" + "u_col_stride6 .req x10\n" + "wptr1 .req x11\n" + "wptr2 .req x12\n" + "w_col_stride1 .req %x[w_col_stride]\n" + "w_col_stride2 .req x13\n" + "vptr1 .req x14\n" + "vptr2 .req x15\n" + "v_col_stride1 .req %x[v_col_stride]\n" + "v_col_stride2 .req x16\n" + + // Prepare strides and pointers + "add uptr1, %x[uptr0], %x[u_row_stride]\n" + "add uptr2, uptr1 , %x[u_row_stride]\n" + "add uptr3, uptr2 , %x[u_row_stride]\n" + "add uptr4, uptr3 , %x[u_row_stride]\n" + "add uptr5, uptr4 , %x[u_row_stride]\n" + "add uptr6, uptr5 , %x[u_row_stride]\n" + "add u_col_stride2, u_col_stride1, u_col_stride1\n" + "add u_col_stride3, u_col_stride2, u_col_stride1\n" + "add u_col_stride4, u_col_stride3, u_col_stride1\n" + "add u_col_stride5, u_col_stride4, u_col_stride1\n" + "add u_col_stride6, u_col_stride5, u_col_stride1\n" + + "add wptr1, %x[wptr0], %x[w_row_stride]\n" + "add wptr2, wptr1 , %x[w_row_stride]\n" + "add w_col_stride2, w_col_stride1, w_col_stride1\n" + + "add vptr1, %x[vptr0], %x[v_row_stride]\n" + "add vptr2, vptr1 , %x[v_row_stride]\n" + "add v_col_stride2, v_col_stride1, v_col_stride1\n" + + // Prepare for first iteration + "ldr qW13, [%x[wptr0], w_col_stride2]\n" + "ldr qW23, [wptr1, w_col_stride2]\n" + "ldr qW33, [wptr2, w_col_stride2]\n" + "ldr qW12, [%x[wptr0], w_col_stride1]\n" + "ldr qW22, [wptr1, w_col_stride1]\n" + "ldr qW32, [wptr2, w_col_stride1]\n" + "ldr qW11, [%x[wptr0]], #0x10\n" + "ldr qW21, [wptr1], #0x10\n" + "ldr qU17, [%x[uptr0], u_col_stride6]\n" + "ldr qU15, [%x[uptr0], u_col_stride4]\n" + "ldr qU16, [%x[uptr0], u_col_stride5]\n" + "ldr qU37, [uptr2, u_col_stride6]\n" + "ldr qU35, [uptr2, u_col_stride4]\n" + "ldr qU36, [uptr2, u_col_stride5]\n" + "ldr qU27, [uptr1, u_col_stride6]\n" + "ldr qU25, [uptr1, u_col_stride4]\n" + "fmul vV13.4s, vU17.4s, vW13.4s\n" + "fmul vV12.4s, vU15.4s, vW13.4s\n" + "fmla vV13.4s, vU15.4s, vW11.4s\n" + "ldr qW31, [wptr2], #0x10\n" + "fmla vV13.4s, vU16.4s, vW12.4s\n" + "ldr qU26, [uptr1, u_col_stride5]\n" + "fmla vV13.4s, vU37.4s, vW33.4s\n" + "ldr qU47, [uptr3, u_col_stride6]\n" + "fmul vV23.4s, vU37.4s, vW13.4s\n" + "ldr qU45, [uptr3, u_col_stride4]\n" + "fmla vV12.4s, vU35.4s, vW33.4s\n" + "ldr qU46, [uptr3, u_col_stride5]\n" + "fmla vV13.4s, vU35.4s, vW31.4s\n" + "ldr qU67, [uptr5, u_col_stride6]\n" + "fmul vV22.4s, vU35.4s, vW13.4s\n" + "cbz %x[n_iters], 2f\n" // Jump to tail if no iterations + + "1:" // Loop body + "fmla vV23.4s, vU35.4s, vW11.4s\n" + "ldr qU65, [uptr5, u_col_stride4]\n" + "fmla vV13.4s, vU36.4s, vW32.4s\n" + "fmla vV23.4s, vU36.4s, vW12.4s\n" + "ldr qU66, [uptr5, u_col_stride5]\n" + "fmla vV13.4s, vU27.4s, vW23.4s\n" + "ldr qU57, [uptr4, u_col_stride6]\n" + "fmla vV12.4s, vU25.4s, vW23.4s\n" + "ldr qU55, [uptr4, u_col_stride4]\n" + "fmla vV13.4s, vU25.4s, vW21.4s\n" + "ldr qU56, [uptr4, u_col_stride5]\n" + "fmla vV13.4s, vU26.4s, vW22.4s\n" + "str qV13, [%x[vptr0], v_col_stride2]\n" + "fmla vV23.4s, vU47.4s, vW23.4s\n" + "ldr qU77, [uptr6, u_col_stride6]\n" + "fmla vV22.4s, vU45.4s, vW23.4s\n" + "fmla vV23.4s, vU45.4s, vW21.4s\n" + "ldr qU75, [uptr6, u_col_stride4]\n" + "fmla vV23.4s, vU46.4s, vW22.4s\n" + "ldr qU76, [uptr6, u_col_stride5]\n" + "fmul vV33.4s, vU67.4s, vW23.4s\n" + "ldr qU14, [%x[uptr0], u_col_stride3]\n" + "fmul vV32.4s, vU65.4s, vW23.4s\n" + "fmla vV33.4s, vU65.4s, vW21.4s\n" + "ldr qU13, [%x[uptr0], u_col_stride2]\n" + "fmla vV33.4s, vU66.4s, vW22.4s\n" + "ldr qU34, [uptr2, u_col_stride3]\n" + "fmla vV23.4s, vU57.4s, vW33.4s\n" + "fmla vV33.4s, vU57.4s, vW13.4s\n" + "ldr qU33, [uptr2, u_col_stride2]\n" + "fmla vV22.4s, vU55.4s, vW33.4s\n" + "fmla vV23.4s, vU55.4s, vW31.4s\n" + "fmla vV32.4s, vU55.4s, vW13.4s\n" + "fmla vV33.4s, vU55.4s, vW11.4s\n" + "ldr qU24, [uptr1, u_col_stride3]\n" + "fmla vV23.4s, vU56.4s, vW32.4s\n" + "str qV23, [vptr1, v_col_stride2]\n" + "fmla vV33.4s, vU56.4s, vW12.4s\n" + "ldr qU23, [uptr1, u_col_stride2]\n" + "fmla vV33.4s, vU77.4s, vW33.4s\n" + "ldr qU44, [uptr3, u_col_stride3]\n" + "fmla vV32.4s, vU75.4s, vW33.4s\n" + "fmla vV33.4s, vU75.4s, vW31.4s\n" + "ldr qU43, [uptr3, u_col_stride2]\n" + "fmla vV33.4s, vU76.4s, vW32.4s\n" + "str qV33, [vptr2, v_col_stride2]\n" + "ldr qU64, [uptr5, u_col_stride3]\n" + "fmla vV12.4s, vU14.4s, vW12.4s\n" + "ldr qU63, [uptr5, u_col_stride2]\n" + "fmul vV11.4s, vU13.4s, vW13.4s\n" + "fmla vV12.4s, vU13.4s, vW11.4s\n" + "ldr qU54, [uptr4, u_col_stride3]\n" + "fmla vV12.4s, vU34.4s, vW32.4s\n" + "fmla vV22.4s, vU34.4s, vW12.4s\n" + "ldr qU53, [uptr4, u_col_stride2]\n" + "fmla vV11.4s, vU33.4s, vW33.4s\n" + "ldr qU74, [uptr6, u_col_stride3]\n" + "fmla vV12.4s, vU33.4s, vW31.4s\n" + "ldr qU73, [uptr6, u_col_stride2]\n" + "fmul vV21.4s, vU33.4s, vW13.4s\n" + "ldr qU12, [%x[uptr0], u_col_stride1]\n" + "fmla vV22.4s, vU33.4s, vW11.4s\n" + "ldr qU11, [%x[uptr0]], #0x10\n" + "fmla vV12.4s, vU24.4s, vW22.4s\n" + "ldr qU32, [uptr2, u_col_stride1]\n" + "fmla vV11.4s, vU23.4s, vW23.4s\n" + "ldr qU31, [uptr2], #0x10\n" + "fmla vV12.4s, vU23.4s, vW21.4s\n" + "str qV12, [%x[vptr0], v_col_stride1]\n" + "fmla vV22.4s, vU44.4s, vW22.4s\n" + "ldr qU22, [uptr1, u_col_stride1]\n" + "fmla vV21.4s, vU43.4s, vW23.4s\n" + "ldr qU21, [uptr1], #0x10\n" + "fmla vV22.4s, vU43.4s, vW21.4s\n" + "ldr qU42, [uptr3, u_col_stride1]\n" + "fmla vV32.4s, vU64.4s, vW22.4s\n" + "ldr qU41, [uptr3], #0x10\n" + "fmul vV31.4s, vU63.4s, vW23.4s\n" + "ldr qW23, [wptr1, w_col_stride2]\n" + "fmla vV32.4s, vU63.4s, vW21.4s\n" + "ldr qU62, [uptr5, u_col_stride1]\n" + "fmla vV22.4s, vU54.4s, vW32.4s\n" + "ldr qU61, [uptr5], #0x10\n" + "fmla vV32.4s, vU54.4s, vW12.4s\n" + "ldr qU52, [uptr4, u_col_stride1]\n" + "fmla vV21.4s, vU53.4s, vW33.4s\n" + "ldr qU51, [uptr4], #0x10\n" + "fmla vV22.4s, vU53.4s, vW31.4s\n" + "str qV22, [vptr1, v_col_stride1]\n" + "fmla vV31.4s, vU53.4s, vW13.4s\n" + "ldr qW13, [%x[wptr0], w_col_stride2]\n" + "fmla vV32.4s, vU53.4s, vW11.4s\n" + "ldr qU72, [uptr6, u_col_stride1]\n" + "fmla vV32.4s, vU74.4s, vW32.4s\n" + "ldr qU71, [uptr6], #0x10\n" + "fmla vV31.4s, vU73.4s, vW33.4s\n" + "ldr qW33, [wptr2, w_col_stride2]\n" + "fmla vV32.4s, vU73.4s, vW31.4s\n" + "str qV32, [vptr2, v_col_stride1]\n" + "fmla vV11.4s, vU12.4s, vW12.4s\n" + "ldr qU17, [%x[uptr0], u_col_stride6]\n" + "fmla vV11.4s, vU11.4s, vW11.4s\n" + "ldr qU15, [%x[uptr0], u_col_stride4]\n" + "fmla vV11.4s, vU32.4s, vW32.4s\n" + "ldr qU16, [%x[uptr0], u_col_stride5]\n" + "fmla vV21.4s, vU32.4s, vW12.4s\n" + "ldr qU37, [uptr2, u_col_stride6]\n" + "fmla vV11.4s, vU31.4s, vW31.4s\n" + "ldr qU35, [uptr2, u_col_stride4]\n" + "fmla vV21.4s, vU31.4s, vW11.4s\n" + "ldr qU36, [uptr2, u_col_stride5]\n" + "fmla vV11.4s, vU22.4s, vW22.4s\n" + "ldr qU27, [uptr1, u_col_stride6]\n" + "fmla vV11.4s, vU21.4s, vW21.4s\n" + "str qV11, [%x[vptr0]], #0x10\n" + "fmla vV21.4s, vU42.4s, vW22.4s\n" + "ldr qU25, [uptr1, u_col_stride4]\n" + "fmla vV21.4s, vU41.4s, vW21.4s\n" + "fmla vV31.4s, vU62.4s, vW22.4s\n" + "ldr qW22, [wptr1, w_col_stride1]\n" + "fmla vV31.4s, vU61.4s, vW21.4s\n" + "ldr qW21, [wptr1], #0x10\n" + "fmla vV21.4s, vU52.4s, vW32.4s\n" + "fmla vV31.4s, vU52.4s, vW12.4s\n" + "ldr qW12, [%x[wptr0], w_col_stride1]\n" + "fmla vV21.4s, vU51.4s, vW31.4s\n" + "str qV21, [vptr1], #0x10\n" + "fmla vV31.4s, vU51.4s, vW11.4s\n" + "ldr qW11, [%x[wptr0]], #0x10\n" + "fmla vV31.4s, vU72.4s, vW32.4s\n" + "ldr qW32, [wptr2, w_col_stride1]\n" + "fmla vV31.4s, vU71.4s, vW31.4s\n" + "str qV31, [vptr2], #0x10\n" + "fmul vV13.4s, vU17.4s, vW13.4s\n" + "fmul vV12.4s, vU15.4s, vW13.4s\n" + "subs %x[n_iters], %x[n_iters], #1\n" + "fmla vV13.4s, vU15.4s, vW11.4s\n" + "ldr qW31, [wptr2], #0x10\n" + "fmla vV13.4s, vU16.4s, vW12.4s\n" + "ldr qU26, [uptr1, u_col_stride5]\n" + "fmla vV13.4s, vU37.4s, vW33.4s\n" + "ldr qU47, [uptr3, u_col_stride6]\n" + "fmul vV23.4s, vU37.4s, vW13.4s\n" + "ldr qU45, [uptr3, u_col_stride4]\n" + "fmla vV12.4s, vU35.4s, vW33.4s\n" + "ldr qU46, [uptr3, u_col_stride5]\n" + "fmla vV13.4s, vU35.4s, vW31.4s\n" + "ldr qU67, [uptr5, u_col_stride6]\n" + "fmul vV22.4s, vU35.4s, vW13.4s\n" + "bne 1b\n" + + "2:" // Tail iteration + "fmla vV23.4s, vU35.4s, vW11.4s\n" + "ldr qU65, [uptr5, u_col_stride4]\n" + "fmla vV13.4s, vU36.4s, vW32.4s\n" + "fmla vV23.4s, vU36.4s, vW12.4s\n" + "ldr qU66, [uptr5, u_col_stride5]\n" + "fmla vV13.4s, vU27.4s, vW23.4s\n" + "ldr qU57, [uptr4, u_col_stride6]\n" + "fmla vV12.4s, vU25.4s, vW23.4s\n" + "ldr qU55, [uptr4, u_col_stride4]\n" + "fmla vV13.4s, vU25.4s, vW21.4s\n" + "ldr qU56, [uptr4, u_col_stride5]\n" + "fmla vV13.4s, vU26.4s, vW22.4s\n" + "str qV13, [%x[vptr0], v_col_stride2]\n" + "fmla vV23.4s, vU47.4s, vW23.4s\n" + "ldr qU77, [uptr6, u_col_stride6]\n" + "fmla vV22.4s, vU45.4s, vW23.4s\n" + "fmla vV23.4s, vU45.4s, vW21.4s\n" + "ldr qU75, [uptr6, u_col_stride4]\n" + "fmla vV23.4s, vU46.4s, vW22.4s\n" + "ldr qU76, [uptr6, u_col_stride5]\n" + "fmul vV33.4s, vU67.4s, vW23.4s\n" + "ldr qU14, [%x[uptr0], u_col_stride3]\n" + "fmul vV32.4s, vU65.4s, vW23.4s\n" + "fmla vV33.4s, vU65.4s, vW21.4s\n" + "ldr qU13, [%x[uptr0], u_col_stride2]\n" + "fmla vV33.4s, vU66.4s, vW22.4s\n" + "ldr qU34, [uptr2, u_col_stride3]\n" + "fmla vV23.4s, vU57.4s, vW33.4s\n" + "fmla vV33.4s, vU57.4s, vW13.4s\n" + "ldr qU33, [uptr2, u_col_stride2]\n" + "fmla vV22.4s, vU55.4s, vW33.4s\n" + "fmla vV23.4s, vU55.4s, vW31.4s\n" + "fmla vV32.4s, vU55.4s, vW13.4s\n" + "fmla vV33.4s, vU55.4s, vW11.4s\n" + "ldr qU24, [uptr1, u_col_stride3]\n" + "fmla vV23.4s, vU56.4s, vW32.4s\n" + "str qV23, [vptr1, v_col_stride2]\n" + "fmla vV33.4s, vU56.4s, vW12.4s\n" + "ldr qU23, [uptr1, u_col_stride2]\n" + "fmla vV33.4s, vU77.4s, vW33.4s\n" + "ldr qU44, [uptr3, u_col_stride3]\n" + "fmla vV32.4s, vU75.4s, vW33.4s\n" + "fmla vV33.4s, vU75.4s, vW31.4s\n" + "ldr qU43, [uptr3, u_col_stride2]\n" + "fmla vV33.4s, vU76.4s, vW32.4s\n" + "str qV33, [vptr2, v_col_stride2]\n" + "ldr qU64, [uptr5, u_col_stride3]\n" + "fmla vV12.4s, vU14.4s, vW12.4s\n" + "ldr qU63, [uptr5, u_col_stride2]\n" + "fmul vV11.4s, vU13.4s, vW13.4s\n" + "fmla vV12.4s, vU13.4s, vW11.4s\n" + "ldr qU54, [uptr4, u_col_stride3]\n" + "fmla vV12.4s, vU34.4s, vW32.4s\n" + "fmla vV22.4s, vU34.4s, vW12.4s\n" + "ldr qU53, [uptr4, u_col_stride2]\n" + "fmla vV11.4s, vU33.4s, vW33.4s\n" + "ldr qU74, [uptr6, u_col_stride3]\n" + "fmla vV12.4s, vU33.4s, vW31.4s\n" + "ldr qU73, [uptr6, u_col_stride2]\n" + "fmul vV21.4s, vU33.4s, vW13.4s\n" + "ldr qU12, [%x[uptr0], u_col_stride1]\n" + "fmla vV22.4s, vU33.4s, vW11.4s\n" + "ldr qU11, [%x[uptr0]], #0x10\n" + "fmla vV12.4s, vU24.4s, vW22.4s\n" + "ldr qU32, [uptr2, u_col_stride1]\n" + "fmla vV11.4s, vU23.4s, vW23.4s\n" + "ldr qU31, [uptr2], #0x10\n" + "fmla vV12.4s, vU23.4s, vW21.4s\n" + "str qV12, [%x[vptr0], v_col_stride1]\n" + "fmla vV22.4s, vU44.4s, vW22.4s\n" + "ldr qU22, [uptr1, u_col_stride1]\n" + "fmla vV21.4s, vU43.4s, vW23.4s\n" + "ldr qU21, [uptr1], #0x10\n" + "fmla vV22.4s, vU43.4s, vW21.4s\n" + "ldr qU42, [uptr3, u_col_stride1]\n" + "fmla vV32.4s, vU64.4s, vW22.4s\n" + "ldr qU41, [uptr3], #0x10\n" + "fmul vV31.4s, vU63.4s, vW23.4s\n" + "fmla vV32.4s, vU63.4s, vW21.4s\n" + "ldr qU62, [uptr5, u_col_stride1]\n" + "fmla vV22.4s, vU54.4s, vW32.4s\n" + "ldr qU61, [uptr5], #0x10\n" + "fmla vV32.4s, vU54.4s, vW12.4s\n" + "ldr qU52, [uptr4, u_col_stride1]\n" + "fmla vV21.4s, vU53.4s, vW33.4s\n" + "ldr qU51, [uptr4], #0x10\n" + "fmla vV22.4s, vU53.4s, vW31.4s\n" + "str qV22, [vptr1, v_col_stride1]\n" + "fmla vV31.4s, vU53.4s, vW13.4s\n" + "fmla vV32.4s, vU53.4s, vW11.4s\n" + "ldr qU72, [uptr6, u_col_stride1]\n" + "fmla vV32.4s, vU74.4s, vW32.4s\n" + "ldr qU71, [uptr6], #0x10\n" + "fmla vV31.4s, vU73.4s, vW33.4s\n" + "fmla vV32.4s, vU73.4s, vW31.4s\n" + "str qV32, [vptr2, v_col_stride1]\n" + "fmla vV11.4s, vU12.4s, vW12.4s\n" + "fmla vV11.4s, vU11.4s, vW11.4s\n" + "fmla vV11.4s, vU32.4s, vW32.4s\n" + "fmla vV21.4s, vU32.4s, vW12.4s\n" + "fmla vV11.4s, vU31.4s, vW31.4s\n" + "fmla vV21.4s, vU31.4s, vW11.4s\n" + "fmla vV11.4s, vU22.4s, vW22.4s\n" + "fmla vV11.4s, vU21.4s, vW21.4s\n" + "str qV11, [%x[vptr0]], #0x10\n" + "fmla vV21.4s, vU42.4s, vW22.4s\n" + "fmla vV21.4s, vU41.4s, vW21.4s\n" + "fmla vV31.4s, vU62.4s, vW22.4s\n" + "fmla vV31.4s, vU61.4s, vW21.4s\n" + "fmla vV21.4s, vU52.4s, vW32.4s\n" + "fmla vV31.4s, vU52.4s, vW12.4s\n" + "fmla vV21.4s, vU51.4s, vW31.4s\n" + "str qV21, [vptr1], #0x10\n" + "fmla vV31.4s, vU51.4s, vW11.4s\n" + "fmla vV31.4s, vU72.4s, vW32.4s\n" + "fmla vV31.4s, vU71.4s, vW31.4s\n" + "str qV31, [vptr2], #0x10\n" + + // Clear aliases + ".unreq uptr1\n" ".unreq uptr2\n" ".unreq uptr3\n" ".unreq uptr4\n" + ".unreq uptr5\n" ".unreq uptr6\n" + ".unreq u_col_stride1\n" ".unreq u_col_stride2\n" ".unreq u_col_stride3\n" + ".unreq u_col_stride4\n" ".unreq u_col_stride5\n" ".unreq u_col_stride6\n" + ".unreq wptr1\n" ".unreq wptr2\n" + ".unreq w_col_stride1\n" ".unreq w_col_stride2\n" + ".unreq vptr1\n" ".unreq vptr2\n" + ".unreq v_col_stride1\n" ".unreq v_col_stride2\n" + ".unreq qU15\n" ".unreq qU73\n" ".unreq qU45\n" ".unreq qU14\n" + ".unreq qW13\n" ".unreq qU62\n" ".unreq qV12\n" + ".unreq qU51\n" ".unreq qU43\n" ".unreq qU55\n" + ".unreq qU77\n" ".unreq qV13\n" ".unreq qV31\n" ".unreq qU44\n" + ".unreq qV33\n" ".unreq qU46\n" ".unreq qU11\n" ".unreq qU37\n" + ".unreq qU56\n" ".unreq qU25\n" ".unreq qU32\n" + ".unreq qU72\n" ".unreq qV22\n" + ".unreq qU67\n" ".unreq qU61\n" ".unreq qU13\n" ".unreq qW33\n" + ".unreq qU74\n" ".unreq qU34\n" ".unreq qU17\n" ".unreq qU66\n" + ".unreq qU33\n" ".unreq qU57\n" ".unreq qU21\n" + ".unreq qW23\n" ".unreq qU42\n" ".unreq qV23\n" ".unreq qU23\n" + ".unreq qU76\n" ".unreq qU47\n" ".unreq qU64\n" ".unreq qU41\n" + ".unreq qU52\n" ".unreq qU54\n" ".unreq qU75\n" ".unreq qU26\n" + ".unreq qU53\n" ".unreq qU27\n" + ".unreq qV21\n" ".unreq qU65\n" + ".unreq qU31\n" ".unreq qU24\n" ".unreq qU36\n" ".unreq qU22\n" + ".unreq qU35\n" ".unreq qU63\n" ".unreq qW12\n" + ".unreq qV32\n" ".unreq qU16\n" ".unreq qW11\n" ".unreq qU12\n" + ".unreq qW31\n" ".unreq qW22\n" ".unreq qU71\n" ".unreq qV11\n" + ".unreq qW21\n" ".unreq qW32\n" ".unreq vW13\n" + ".unreq vU15\n" ".unreq vU73\n" ".unreq vU45\n" ".unreq vU14\n" + ".unreq vU62\n" ".unreq vV12\n" + ".unreq vU51\n" ".unreq vU43\n" ".unreq vU55\n" + ".unreq vU77\n" ".unreq vV13\n" ".unreq vV31\n" ".unreq vU44\n" + ".unreq vV33\n" ".unreq vU46\n" ".unreq vU11\n" ".unreq vU37\n" + ".unreq vU56\n" ".unreq vU25\n" ".unreq vU32\n" + ".unreq vU72\n" ".unreq vV22\n" ".unreq vW21\n" ".unreq vW32\n" + ".unreq vU67\n" ".unreq vU61\n" ".unreq vU13\n" + ".unreq vU74\n" ".unreq vU34\n" ".unreq vU17\n" ".unreq vU66\n" + ".unreq vU33\n" ".unreq vU57\n" ".unreq vU21\n" ".unreq vW23\n" + ".unreq vU42\n" ".unreq vV23\n" ".unreq vU23\n" ".unreq vW33\n" + ".unreq vU76\n" ".unreq vU47\n" ".unreq vU64\n" ".unreq vU41\n" + ".unreq vU52\n" ".unreq vU54\n" ".unreq vU75\n" ".unreq vU26\n" + ".unreq vU53\n" ".unreq vU27\n" ".unreq vV21\n" ".unreq vU65\n" + ".unreq vU31\n" ".unreq vU24\n" ".unreq vU36\n" ".unreq vU22\n" + ".unreq vU35\n" ".unreq vU63\n" ".unreq vW12\n" + ".unreq vV32\n" ".unreq vU16\n" ".unreq vW11\n" ".unreq vU12\n" + ".unreq vW31\n" ".unreq vW22\n" ".unreq vU71\n" ".unreq vV11\n" + : [uptr0] "+r" (uptr0), [wptr0] "+r" (wptr0), [vptr0] "+r" (vptr0), + [n_iters] "+r" (n_iters) + : [u_row_stride] "r" (in_row_stride * sizeof(float)), + [u_col_stride] "r" (in_col_stride * sizeof(float)), + [w_row_stride] "r" (weight_row_stride * sizeof(float)), + [w_col_stride] "r" (weight_col_stride * sizeof(float)), + [v_row_stride] "r" (out_row_stride * sizeof(float)), + [v_col_stride] "r" (out_col_stride * sizeof(float)) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", + "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", + "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "x0", + "x1", "x2", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", + "x12", "x13", "x14", "x15", "x16", "cc", "memory" + ); + } + if (channels_remaining) + { + // Fall back on the unoptimised version to clean up the tail + ConvImpl::process_tile( + channels_remaining, + wptr0, weight_row_stride, weight_col_stride, + uptr0, in_row_stride, in_col_stride, + vptr0, out_row_stride, out_col_stride, + 0, 0, 0, 0, 0, 0 + ); + } +} + +#endif // __aarch64__ + +template <> +const Conv::TileFn Conv::tilefn_unpadded = ConvImpl::template process_tile; + +template <> +const Conv::TileFn Conv::tilefn_top[n_in_pad_top_fns] = { + ConvImpl::template process_tile, + ConvImpl::template process_tile, +}; + template <> -const Conv::TileFn Conv::tile_fns - [max_in_pad_top] - [max_in_pad_left] - [max_in_pad_bottom] - [max_in_pad_right] - [max_out_pad_bottom] - [max_out_pad_right] = { - { // Input pad top = 0 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 0, 0, 0>, - Conv::template process_tile<0, 0, 0, 0, 0, 1>, - Conv::template process_tile<0, 0, 0, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 0, 1, 0>, - Conv::template process_tile<0, 0, 0, 0, 1, 1>, - Conv::template process_tile<0, 0, 0, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 0, 2, 0>, - Conv::template process_tile<0, 0, 0, 0, 2, 1>, - Conv::template process_tile<0, 0, 0, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 1, 0, 0>, - Conv::template process_tile<0, 0, 0, 1, 0, 1>, - Conv::template process_tile<0, 0, 0, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 1, 1, 0>, - Conv::template process_tile<0, 0, 0, 1, 1, 1>, - Conv::template process_tile<0, 0, 0, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 1, 2, 0>, - Conv::template process_tile<0, 0, 0, 1, 2, 1>, - Conv::template process_tile<0, 0, 0, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 2, 0, 0>, - Conv::template process_tile<0, 0, 0, 2, 0, 1>, - Conv::template process_tile<0, 0, 0, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 2, 1, 0>, - Conv::template process_tile<0, 0, 0, 2, 1, 1>, - Conv::template process_tile<0, 0, 0, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 2, 2, 0>, - Conv::template process_tile<0, 0, 0, 2, 2, 1>, - Conv::template process_tile<0, 0, 0, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 3, 0, 0>, - Conv::template process_tile<0, 0, 0, 3, 0, 1>, - Conv::template process_tile<0, 0, 0, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 3, 1, 0>, - Conv::template process_tile<0, 0, 0, 3, 1, 1>, - Conv::template process_tile<0, 0, 0, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 3, 2, 0>, - Conv::template process_tile<0, 0, 0, 3, 2, 1>, - Conv::template process_tile<0, 0, 0, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 4, 0, 0>, - Conv::template process_tile<0, 0, 0, 4, 0, 1>, - Conv::template process_tile<0, 0, 0, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 4, 1, 0>, - Conv::template process_tile<0, 0, 0, 4, 1, 1>, - Conv::template process_tile<0, 0, 0, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 4, 2, 0>, - Conv::template process_tile<0, 0, 0, 4, 2, 1>, - Conv::template process_tile<0, 0, 0, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 5, 0, 0>, - Conv::template process_tile<0, 0, 0, 5, 0, 1>, - Conv::template process_tile<0, 0, 0, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 5, 1, 0>, - Conv::template process_tile<0, 0, 0, 5, 1, 1>, - Conv::template process_tile<0, 0, 0, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 5, 2, 0>, - Conv::template process_tile<0, 0, 0, 5, 2, 1>, - Conv::template process_tile<0, 0, 0, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 6, 0, 0>, - Conv::template process_tile<0, 0, 0, 6, 0, 1>, - Conv::template process_tile<0, 0, 0, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 6, 1, 0>, - Conv::template process_tile<0, 0, 0, 6, 1, 1>, - Conv::template process_tile<0, 0, 0, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 6, 2, 0>, - Conv::template process_tile<0, 0, 0, 6, 2, 1>, - Conv::template process_tile<0, 0, 0, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 0, 0, 0>, - Conv::template process_tile<0, 0, 1, 0, 0, 1>, - Conv::template process_tile<0, 0, 1, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 0, 1, 0>, - Conv::template process_tile<0, 0, 1, 0, 1, 1>, - Conv::template process_tile<0, 0, 1, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 0, 2, 0>, - Conv::template process_tile<0, 0, 1, 0, 2, 1>, - Conv::template process_tile<0, 0, 1, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 1, 0, 0>, - Conv::template process_tile<0, 0, 1, 1, 0, 1>, - Conv::template process_tile<0, 0, 1, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 1, 1, 0>, - Conv::template process_tile<0, 0, 1, 1, 1, 1>, - Conv::template process_tile<0, 0, 1, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 1, 2, 0>, - Conv::template process_tile<0, 0, 1, 1, 2, 1>, - Conv::template process_tile<0, 0, 1, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 2, 0, 0>, - Conv::template process_tile<0, 0, 1, 2, 0, 1>, - Conv::template process_tile<0, 0, 1, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 2, 1, 0>, - Conv::template process_tile<0, 0, 1, 2, 1, 1>, - Conv::template process_tile<0, 0, 1, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 2, 2, 0>, - Conv::template process_tile<0, 0, 1, 2, 2, 1>, - Conv::template process_tile<0, 0, 1, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 3, 0, 0>, - Conv::template process_tile<0, 0, 1, 3, 0, 1>, - Conv::template process_tile<0, 0, 1, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 3, 1, 0>, - Conv::template process_tile<0, 0, 1, 3, 1, 1>, - Conv::template process_tile<0, 0, 1, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 3, 2, 0>, - Conv::template process_tile<0, 0, 1, 3, 2, 1>, - Conv::template process_tile<0, 0, 1, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 4, 0, 0>, - Conv::template process_tile<0, 0, 1, 4, 0, 1>, - Conv::template process_tile<0, 0, 1, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 4, 1, 0>, - Conv::template process_tile<0, 0, 1, 4, 1, 1>, - Conv::template process_tile<0, 0, 1, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 4, 2, 0>, - Conv::template process_tile<0, 0, 1, 4, 2, 1>, - Conv::template process_tile<0, 0, 1, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 5, 0, 0>, - Conv::template process_tile<0, 0, 1, 5, 0, 1>, - Conv::template process_tile<0, 0, 1, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 5, 1, 0>, - Conv::template process_tile<0, 0, 1, 5, 1, 1>, - Conv::template process_tile<0, 0, 1, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 5, 2, 0>, - Conv::template process_tile<0, 0, 1, 5, 2, 1>, - Conv::template process_tile<0, 0, 1, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 6, 0, 0>, - Conv::template process_tile<0, 0, 1, 6, 0, 1>, - Conv::template process_tile<0, 0, 1, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 6, 1, 0>, - Conv::template process_tile<0, 0, 1, 6, 1, 1>, - Conv::template process_tile<0, 0, 1, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 6, 2, 0>, - Conv::template process_tile<0, 0, 1, 6, 2, 1>, - Conv::template process_tile<0, 0, 1, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 0, 0, 0>, - Conv::template process_tile<0, 0, 2, 0, 0, 1>, - Conv::template process_tile<0, 0, 2, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 0, 1, 0>, - Conv::template process_tile<0, 0, 2, 0, 1, 1>, - Conv::template process_tile<0, 0, 2, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 0, 2, 0>, - Conv::template process_tile<0, 0, 2, 0, 2, 1>, - Conv::template process_tile<0, 0, 2, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 1, 0, 0>, - Conv::template process_tile<0, 0, 2, 1, 0, 1>, - Conv::template process_tile<0, 0, 2, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 1, 1, 0>, - Conv::template process_tile<0, 0, 2, 1, 1, 1>, - Conv::template process_tile<0, 0, 2, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 1, 2, 0>, - Conv::template process_tile<0, 0, 2, 1, 2, 1>, - Conv::template process_tile<0, 0, 2, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 2, 0, 0>, - Conv::template process_tile<0, 0, 2, 2, 0, 1>, - Conv::template process_tile<0, 0, 2, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 2, 1, 0>, - Conv::template process_tile<0, 0, 2, 2, 1, 1>, - Conv::template process_tile<0, 0, 2, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 2, 2, 0>, - Conv::template process_tile<0, 0, 2, 2, 2, 1>, - Conv::template process_tile<0, 0, 2, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 3, 0, 0>, - Conv::template process_tile<0, 0, 2, 3, 0, 1>, - Conv::template process_tile<0, 0, 2, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 3, 1, 0>, - Conv::template process_tile<0, 0, 2, 3, 1, 1>, - Conv::template process_tile<0, 0, 2, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 3, 2, 0>, - Conv::template process_tile<0, 0, 2, 3, 2, 1>, - Conv::template process_tile<0, 0, 2, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 4, 0, 0>, - Conv::template process_tile<0, 0, 2, 4, 0, 1>, - Conv::template process_tile<0, 0, 2, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 4, 1, 0>, - Conv::template process_tile<0, 0, 2, 4, 1, 1>, - Conv::template process_tile<0, 0, 2, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 4, 2, 0>, - Conv::template process_tile<0, 0, 2, 4, 2, 1>, - Conv::template process_tile<0, 0, 2, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 5, 0, 0>, - Conv::template process_tile<0, 0, 2, 5, 0, 1>, - Conv::template process_tile<0, 0, 2, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 5, 1, 0>, - Conv::template process_tile<0, 0, 2, 5, 1, 1>, - Conv::template process_tile<0, 0, 2, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 5, 2, 0>, - Conv::template process_tile<0, 0, 2, 5, 2, 1>, - Conv::template process_tile<0, 0, 2, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 6, 0, 0>, - Conv::template process_tile<0, 0, 2, 6, 0, 1>, - Conv::template process_tile<0, 0, 2, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 6, 1, 0>, - Conv::template process_tile<0, 0, 2, 6, 1, 1>, - Conv::template process_tile<0, 0, 2, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 6, 2, 0>, - Conv::template process_tile<0, 0, 2, 6, 2, 1>, - Conv::template process_tile<0, 0, 2, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 0, 0, 0>, - Conv::template process_tile<0, 0, 3, 0, 0, 1>, - Conv::template process_tile<0, 0, 3, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 0, 1, 0>, - Conv::template process_tile<0, 0, 3, 0, 1, 1>, - Conv::template process_tile<0, 0, 3, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 0, 2, 0>, - Conv::template process_tile<0, 0, 3, 0, 2, 1>, - Conv::template process_tile<0, 0, 3, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 1, 0, 0>, - Conv::template process_tile<0, 0, 3, 1, 0, 1>, - Conv::template process_tile<0, 0, 3, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 1, 1, 0>, - Conv::template process_tile<0, 0, 3, 1, 1, 1>, - Conv::template process_tile<0, 0, 3, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 1, 2, 0>, - Conv::template process_tile<0, 0, 3, 1, 2, 1>, - Conv::template process_tile<0, 0, 3, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 2, 0, 0>, - Conv::template process_tile<0, 0, 3, 2, 0, 1>, - Conv::template process_tile<0, 0, 3, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 2, 1, 0>, - Conv::template process_tile<0, 0, 3, 2, 1, 1>, - Conv::template process_tile<0, 0, 3, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 2, 2, 0>, - Conv::template process_tile<0, 0, 3, 2, 2, 1>, - Conv::template process_tile<0, 0, 3, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 3, 0, 0>, - Conv::template process_tile<0, 0, 3, 3, 0, 1>, - Conv::template process_tile<0, 0, 3, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 3, 1, 0>, - Conv::template process_tile<0, 0, 3, 3, 1, 1>, - Conv::template process_tile<0, 0, 3, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 3, 2, 0>, - Conv::template process_tile<0, 0, 3, 3, 2, 1>, - Conv::template process_tile<0, 0, 3, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 4, 0, 0>, - Conv::template process_tile<0, 0, 3, 4, 0, 1>, - Conv::template process_tile<0, 0, 3, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 4, 1, 0>, - Conv::template process_tile<0, 0, 3, 4, 1, 1>, - Conv::template process_tile<0, 0, 3, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 4, 2, 0>, - Conv::template process_tile<0, 0, 3, 4, 2, 1>, - Conv::template process_tile<0, 0, 3, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 5, 0, 0>, - Conv::template process_tile<0, 0, 3, 5, 0, 1>, - Conv::template process_tile<0, 0, 3, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 5, 1, 0>, - Conv::template process_tile<0, 0, 3, 5, 1, 1>, - Conv::template process_tile<0, 0, 3, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 5, 2, 0>, - Conv::template process_tile<0, 0, 3, 5, 2, 1>, - Conv::template process_tile<0, 0, 3, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 6, 0, 0>, - Conv::template process_tile<0, 0, 3, 6, 0, 1>, - Conv::template process_tile<0, 0, 3, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 6, 1, 0>, - Conv::template process_tile<0, 0, 3, 6, 1, 1>, - Conv::template process_tile<0, 0, 3, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 6, 2, 0>, - Conv::template process_tile<0, 0, 3, 6, 2, 1>, - Conv::template process_tile<0, 0, 3, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 0, 0, 0>, - Conv::template process_tile<0, 0, 4, 0, 0, 1>, - Conv::template process_tile<0, 0, 4, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 0, 1, 0>, - Conv::template process_tile<0, 0, 4, 0, 1, 1>, - Conv::template process_tile<0, 0, 4, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 0, 2, 0>, - Conv::template process_tile<0, 0, 4, 0, 2, 1>, - Conv::template process_tile<0, 0, 4, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 1, 0, 0>, - Conv::template process_tile<0, 0, 4, 1, 0, 1>, - Conv::template process_tile<0, 0, 4, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 1, 1, 0>, - Conv::template process_tile<0, 0, 4, 1, 1, 1>, - Conv::template process_tile<0, 0, 4, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 1, 2, 0>, - Conv::template process_tile<0, 0, 4, 1, 2, 1>, - Conv::template process_tile<0, 0, 4, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 2, 0, 0>, - Conv::template process_tile<0, 0, 4, 2, 0, 1>, - Conv::template process_tile<0, 0, 4, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 2, 1, 0>, - Conv::template process_tile<0, 0, 4, 2, 1, 1>, - Conv::template process_tile<0, 0, 4, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 2, 2, 0>, - Conv::template process_tile<0, 0, 4, 2, 2, 1>, - Conv::template process_tile<0, 0, 4, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 3, 0, 0>, - Conv::template process_tile<0, 0, 4, 3, 0, 1>, - Conv::template process_tile<0, 0, 4, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 3, 1, 0>, - Conv::template process_tile<0, 0, 4, 3, 1, 1>, - Conv::template process_tile<0, 0, 4, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 3, 2, 0>, - Conv::template process_tile<0, 0, 4, 3, 2, 1>, - Conv::template process_tile<0, 0, 4, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 4, 0, 0>, - Conv::template process_tile<0, 0, 4, 4, 0, 1>, - Conv::template process_tile<0, 0, 4, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 4, 1, 0>, - Conv::template process_tile<0, 0, 4, 4, 1, 1>, - Conv::template process_tile<0, 0, 4, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 4, 2, 0>, - Conv::template process_tile<0, 0, 4, 4, 2, 1>, - Conv::template process_tile<0, 0, 4, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 5, 0, 0>, - Conv::template process_tile<0, 0, 4, 5, 0, 1>, - Conv::template process_tile<0, 0, 4, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 5, 1, 0>, - Conv::template process_tile<0, 0, 4, 5, 1, 1>, - Conv::template process_tile<0, 0, 4, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 5, 2, 0>, - Conv::template process_tile<0, 0, 4, 5, 2, 1>, - Conv::template process_tile<0, 0, 4, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 6, 0, 0>, - Conv::template process_tile<0, 0, 4, 6, 0, 1>, - Conv::template process_tile<0, 0, 4, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 6, 1, 0>, - Conv::template process_tile<0, 0, 4, 6, 1, 1>, - Conv::template process_tile<0, 0, 4, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 6, 2, 0>, - Conv::template process_tile<0, 0, 4, 6, 2, 1>, - Conv::template process_tile<0, 0, 4, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 4 - { // Input pad bottom = 5 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 0, 0, 0>, - Conv::template process_tile<0, 0, 5, 0, 0, 1>, - Conv::template process_tile<0, 0, 5, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 0, 1, 0>, - Conv::template process_tile<0, 0, 5, 0, 1, 1>, - Conv::template process_tile<0, 0, 5, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 0, 2, 0>, - Conv::template process_tile<0, 0, 5, 0, 2, 1>, - Conv::template process_tile<0, 0, 5, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 1, 0, 0>, - Conv::template process_tile<0, 0, 5, 1, 0, 1>, - Conv::template process_tile<0, 0, 5, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 1, 1, 0>, - Conv::template process_tile<0, 0, 5, 1, 1, 1>, - Conv::template process_tile<0, 0, 5, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 1, 2, 0>, - Conv::template process_tile<0, 0, 5, 1, 2, 1>, - Conv::template process_tile<0, 0, 5, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 2, 0, 0>, - Conv::template process_tile<0, 0, 5, 2, 0, 1>, - Conv::template process_tile<0, 0, 5, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 2, 1, 0>, - Conv::template process_tile<0, 0, 5, 2, 1, 1>, - Conv::template process_tile<0, 0, 5, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 2, 2, 0>, - Conv::template process_tile<0, 0, 5, 2, 2, 1>, - Conv::template process_tile<0, 0, 5, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 3, 0, 0>, - Conv::template process_tile<0, 0, 5, 3, 0, 1>, - Conv::template process_tile<0, 0, 5, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 3, 1, 0>, - Conv::template process_tile<0, 0, 5, 3, 1, 1>, - Conv::template process_tile<0, 0, 5, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 3, 2, 0>, - Conv::template process_tile<0, 0, 5, 3, 2, 1>, - Conv::template process_tile<0, 0, 5, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 4, 0, 0>, - Conv::template process_tile<0, 0, 5, 4, 0, 1>, - Conv::template process_tile<0, 0, 5, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 4, 1, 0>, - Conv::template process_tile<0, 0, 5, 4, 1, 1>, - Conv::template process_tile<0, 0, 5, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 4, 2, 0>, - Conv::template process_tile<0, 0, 5, 4, 2, 1>, - Conv::template process_tile<0, 0, 5, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 5, 0, 0>, - Conv::template process_tile<0, 0, 5, 5, 0, 1>, - Conv::template process_tile<0, 0, 5, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 5, 1, 0>, - Conv::template process_tile<0, 0, 5, 5, 1, 1>, - Conv::template process_tile<0, 0, 5, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 5, 2, 0>, - Conv::template process_tile<0, 0, 5, 5, 2, 1>, - Conv::template process_tile<0, 0, 5, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 6, 0, 0>, - Conv::template process_tile<0, 0, 5, 6, 0, 1>, - Conv::template process_tile<0, 0, 5, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 6, 1, 0>, - Conv::template process_tile<0, 0, 5, 6, 1, 1>, - Conv::template process_tile<0, 0, 5, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 6, 2, 0>, - Conv::template process_tile<0, 0, 5, 6, 2, 1>, - Conv::template process_tile<0, 0, 5, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 5 - { // Input pad bottom = 6 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 0, 0, 0>, - Conv::template process_tile<0, 0, 6, 0, 0, 1>, - Conv::template process_tile<0, 0, 6, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 0, 1, 0>, - Conv::template process_tile<0, 0, 6, 0, 1, 1>, - Conv::template process_tile<0, 0, 6, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 0, 2, 0>, - Conv::template process_tile<0, 0, 6, 0, 2, 1>, - Conv::template process_tile<0, 0, 6, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 1, 0, 0>, - Conv::template process_tile<0, 0, 6, 1, 0, 1>, - Conv::template process_tile<0, 0, 6, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 1, 1, 0>, - Conv::template process_tile<0, 0, 6, 1, 1, 1>, - Conv::template process_tile<0, 0, 6, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 1, 2, 0>, - Conv::template process_tile<0, 0, 6, 1, 2, 1>, - Conv::template process_tile<0, 0, 6, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 2, 0, 0>, - Conv::template process_tile<0, 0, 6, 2, 0, 1>, - Conv::template process_tile<0, 0, 6, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 2, 1, 0>, - Conv::template process_tile<0, 0, 6, 2, 1, 1>, - Conv::template process_tile<0, 0, 6, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 2, 2, 0>, - Conv::template process_tile<0, 0, 6, 2, 2, 1>, - Conv::template process_tile<0, 0, 6, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 3, 0, 0>, - Conv::template process_tile<0, 0, 6, 3, 0, 1>, - Conv::template process_tile<0, 0, 6, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 3, 1, 0>, - Conv::template process_tile<0, 0, 6, 3, 1, 1>, - Conv::template process_tile<0, 0, 6, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 3, 2, 0>, - Conv::template process_tile<0, 0, 6, 3, 2, 1>, - Conv::template process_tile<0, 0, 6, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 4, 0, 0>, - Conv::template process_tile<0, 0, 6, 4, 0, 1>, - Conv::template process_tile<0, 0, 6, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 4, 1, 0>, - Conv::template process_tile<0, 0, 6, 4, 1, 1>, - Conv::template process_tile<0, 0, 6, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 4, 2, 0>, - Conv::template process_tile<0, 0, 6, 4, 2, 1>, - Conv::template process_tile<0, 0, 6, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 5, 0, 0>, - Conv::template process_tile<0, 0, 6, 5, 0, 1>, - Conv::template process_tile<0, 0, 6, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 5, 1, 0>, - Conv::template process_tile<0, 0, 6, 5, 1, 1>, - Conv::template process_tile<0, 0, 6, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 5, 2, 0>, - Conv::template process_tile<0, 0, 6, 5, 2, 1>, - Conv::template process_tile<0, 0, 6, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 6, 0, 0>, - Conv::template process_tile<0, 0, 6, 6, 0, 1>, - Conv::template process_tile<0, 0, 6, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 6, 1, 0>, - Conv::template process_tile<0, 0, 6, 6, 1, 1>, - Conv::template process_tile<0, 0, 6, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 6, 2, 0>, - Conv::template process_tile<0, 0, 6, 6, 2, 1>, - Conv::template process_tile<0, 0, 6, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 6 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 0, 0, 0>, - Conv::template process_tile<0, 1, 0, 0, 0, 1>, - Conv::template process_tile<0, 1, 0, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 0, 1, 0>, - Conv::template process_tile<0, 1, 0, 0, 1, 1>, - Conv::template process_tile<0, 1, 0, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 0, 2, 0>, - Conv::template process_tile<0, 1, 0, 0, 2, 1>, - Conv::template process_tile<0, 1, 0, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 1, 0, 0>, - Conv::template process_tile<0, 1, 0, 1, 0, 1>, - Conv::template process_tile<0, 1, 0, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 1, 1, 0>, - Conv::template process_tile<0, 1, 0, 1, 1, 1>, - Conv::template process_tile<0, 1, 0, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 1, 2, 0>, - Conv::template process_tile<0, 1, 0, 1, 2, 1>, - Conv::template process_tile<0, 1, 0, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 2, 0, 0>, - Conv::template process_tile<0, 1, 0, 2, 0, 1>, - Conv::template process_tile<0, 1, 0, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 2, 1, 0>, - Conv::template process_tile<0, 1, 0, 2, 1, 1>, - Conv::template process_tile<0, 1, 0, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 2, 2, 0>, - Conv::template process_tile<0, 1, 0, 2, 2, 1>, - Conv::template process_tile<0, 1, 0, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 3, 0, 0>, - Conv::template process_tile<0, 1, 0, 3, 0, 1>, - Conv::template process_tile<0, 1, 0, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 3, 1, 0>, - Conv::template process_tile<0, 1, 0, 3, 1, 1>, - Conv::template process_tile<0, 1, 0, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 3, 2, 0>, - Conv::template process_tile<0, 1, 0, 3, 2, 1>, - Conv::template process_tile<0, 1, 0, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 4, 0, 0>, - Conv::template process_tile<0, 1, 0, 4, 0, 1>, - Conv::template process_tile<0, 1, 0, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 4, 1, 0>, - Conv::template process_tile<0, 1, 0, 4, 1, 1>, - Conv::template process_tile<0, 1, 0, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 4, 2, 0>, - Conv::template process_tile<0, 1, 0, 4, 2, 1>, - Conv::template process_tile<0, 1, 0, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 5, 0, 0>, - Conv::template process_tile<0, 1, 0, 5, 0, 1>, - Conv::template process_tile<0, 1, 0, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 5, 1, 0>, - Conv::template process_tile<0, 1, 0, 5, 1, 1>, - Conv::template process_tile<0, 1, 0, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 5, 2, 0>, - Conv::template process_tile<0, 1, 0, 5, 2, 1>, - Conv::template process_tile<0, 1, 0, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 6, 0, 0>, - Conv::template process_tile<0, 1, 0, 6, 0, 1>, - Conv::template process_tile<0, 1, 0, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 6, 1, 0>, - Conv::template process_tile<0, 1, 0, 6, 1, 1>, - Conv::template process_tile<0, 1, 0, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 6, 2, 0>, - Conv::template process_tile<0, 1, 0, 6, 2, 1>, - Conv::template process_tile<0, 1, 0, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 0, 0, 0>, - Conv::template process_tile<0, 1, 1, 0, 0, 1>, - Conv::template process_tile<0, 1, 1, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 0, 1, 0>, - Conv::template process_tile<0, 1, 1, 0, 1, 1>, - Conv::template process_tile<0, 1, 1, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 0, 2, 0>, - Conv::template process_tile<0, 1, 1, 0, 2, 1>, - Conv::template process_tile<0, 1, 1, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 1, 0, 0>, - Conv::template process_tile<0, 1, 1, 1, 0, 1>, - Conv::template process_tile<0, 1, 1, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 1, 1, 0>, - Conv::template process_tile<0, 1, 1, 1, 1, 1>, - Conv::template process_tile<0, 1, 1, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 1, 2, 0>, - Conv::template process_tile<0, 1, 1, 1, 2, 1>, - Conv::template process_tile<0, 1, 1, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 2, 0, 0>, - Conv::template process_tile<0, 1, 1, 2, 0, 1>, - Conv::template process_tile<0, 1, 1, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 2, 1, 0>, - Conv::template process_tile<0, 1, 1, 2, 1, 1>, - Conv::template process_tile<0, 1, 1, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 2, 2, 0>, - Conv::template process_tile<0, 1, 1, 2, 2, 1>, - Conv::template process_tile<0, 1, 1, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 3, 0, 0>, - Conv::template process_tile<0, 1, 1, 3, 0, 1>, - Conv::template process_tile<0, 1, 1, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 3, 1, 0>, - Conv::template process_tile<0, 1, 1, 3, 1, 1>, - Conv::template process_tile<0, 1, 1, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 3, 2, 0>, - Conv::template process_tile<0, 1, 1, 3, 2, 1>, - Conv::template process_tile<0, 1, 1, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 4, 0, 0>, - Conv::template process_tile<0, 1, 1, 4, 0, 1>, - Conv::template process_tile<0, 1, 1, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 4, 1, 0>, - Conv::template process_tile<0, 1, 1, 4, 1, 1>, - Conv::template process_tile<0, 1, 1, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 4, 2, 0>, - Conv::template process_tile<0, 1, 1, 4, 2, 1>, - Conv::template process_tile<0, 1, 1, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 5, 0, 0>, - Conv::template process_tile<0, 1, 1, 5, 0, 1>, - Conv::template process_tile<0, 1, 1, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 5, 1, 0>, - Conv::template process_tile<0, 1, 1, 5, 1, 1>, - Conv::template process_tile<0, 1, 1, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 5, 2, 0>, - Conv::template process_tile<0, 1, 1, 5, 2, 1>, - Conv::template process_tile<0, 1, 1, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 6, 0, 0>, - Conv::template process_tile<0, 1, 1, 6, 0, 1>, - Conv::template process_tile<0, 1, 1, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 6, 1, 0>, - Conv::template process_tile<0, 1, 1, 6, 1, 1>, - Conv::template process_tile<0, 1, 1, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 6, 2, 0>, - Conv::template process_tile<0, 1, 1, 6, 2, 1>, - Conv::template process_tile<0, 1, 1, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 0, 0, 0>, - Conv::template process_tile<0, 1, 2, 0, 0, 1>, - Conv::template process_tile<0, 1, 2, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 0, 1, 0>, - Conv::template process_tile<0, 1, 2, 0, 1, 1>, - Conv::template process_tile<0, 1, 2, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 0, 2, 0>, - Conv::template process_tile<0, 1, 2, 0, 2, 1>, - Conv::template process_tile<0, 1, 2, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 1, 0, 0>, - Conv::template process_tile<0, 1, 2, 1, 0, 1>, - Conv::template process_tile<0, 1, 2, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 1, 1, 0>, - Conv::template process_tile<0, 1, 2, 1, 1, 1>, - Conv::template process_tile<0, 1, 2, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 1, 2, 0>, - Conv::template process_tile<0, 1, 2, 1, 2, 1>, - Conv::template process_tile<0, 1, 2, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 2, 0, 0>, - Conv::template process_tile<0, 1, 2, 2, 0, 1>, - Conv::template process_tile<0, 1, 2, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 2, 1, 0>, - Conv::template process_tile<0, 1, 2, 2, 1, 1>, - Conv::template process_tile<0, 1, 2, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 2, 2, 0>, - Conv::template process_tile<0, 1, 2, 2, 2, 1>, - Conv::template process_tile<0, 1, 2, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 3, 0, 0>, - Conv::template process_tile<0, 1, 2, 3, 0, 1>, - Conv::template process_tile<0, 1, 2, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 3, 1, 0>, - Conv::template process_tile<0, 1, 2, 3, 1, 1>, - Conv::template process_tile<0, 1, 2, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 3, 2, 0>, - Conv::template process_tile<0, 1, 2, 3, 2, 1>, - Conv::template process_tile<0, 1, 2, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 4, 0, 0>, - Conv::template process_tile<0, 1, 2, 4, 0, 1>, - Conv::template process_tile<0, 1, 2, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 4, 1, 0>, - Conv::template process_tile<0, 1, 2, 4, 1, 1>, - Conv::template process_tile<0, 1, 2, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 4, 2, 0>, - Conv::template process_tile<0, 1, 2, 4, 2, 1>, - Conv::template process_tile<0, 1, 2, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 5, 0, 0>, - Conv::template process_tile<0, 1, 2, 5, 0, 1>, - Conv::template process_tile<0, 1, 2, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 5, 1, 0>, - Conv::template process_tile<0, 1, 2, 5, 1, 1>, - Conv::template process_tile<0, 1, 2, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 5, 2, 0>, - Conv::template process_tile<0, 1, 2, 5, 2, 1>, - Conv::template process_tile<0, 1, 2, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 6, 0, 0>, - Conv::template process_tile<0, 1, 2, 6, 0, 1>, - Conv::template process_tile<0, 1, 2, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 6, 1, 0>, - Conv::template process_tile<0, 1, 2, 6, 1, 1>, - Conv::template process_tile<0, 1, 2, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 6, 2, 0>, - Conv::template process_tile<0, 1, 2, 6, 2, 1>, - Conv::template process_tile<0, 1, 2, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 0, 0, 0>, - Conv::template process_tile<0, 1, 3, 0, 0, 1>, - Conv::template process_tile<0, 1, 3, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 0, 1, 0>, - Conv::template process_tile<0, 1, 3, 0, 1, 1>, - Conv::template process_tile<0, 1, 3, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 0, 2, 0>, - Conv::template process_tile<0, 1, 3, 0, 2, 1>, - Conv::template process_tile<0, 1, 3, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 1, 0, 0>, - Conv::template process_tile<0, 1, 3, 1, 0, 1>, - Conv::template process_tile<0, 1, 3, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 1, 1, 0>, - Conv::template process_tile<0, 1, 3, 1, 1, 1>, - Conv::template process_tile<0, 1, 3, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 1, 2, 0>, - Conv::template process_tile<0, 1, 3, 1, 2, 1>, - Conv::template process_tile<0, 1, 3, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 2, 0, 0>, - Conv::template process_tile<0, 1, 3, 2, 0, 1>, - Conv::template process_tile<0, 1, 3, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 2, 1, 0>, - Conv::template process_tile<0, 1, 3, 2, 1, 1>, - Conv::template process_tile<0, 1, 3, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 2, 2, 0>, - Conv::template process_tile<0, 1, 3, 2, 2, 1>, - Conv::template process_tile<0, 1, 3, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 3, 0, 0>, - Conv::template process_tile<0, 1, 3, 3, 0, 1>, - Conv::template process_tile<0, 1, 3, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 3, 1, 0>, - Conv::template process_tile<0, 1, 3, 3, 1, 1>, - Conv::template process_tile<0, 1, 3, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 3, 2, 0>, - Conv::template process_tile<0, 1, 3, 3, 2, 1>, - Conv::template process_tile<0, 1, 3, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 4, 0, 0>, - Conv::template process_tile<0, 1, 3, 4, 0, 1>, - Conv::template process_tile<0, 1, 3, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 4, 1, 0>, - Conv::template process_tile<0, 1, 3, 4, 1, 1>, - Conv::template process_tile<0, 1, 3, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 4, 2, 0>, - Conv::template process_tile<0, 1, 3, 4, 2, 1>, - Conv::template process_tile<0, 1, 3, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 5, 0, 0>, - Conv::template process_tile<0, 1, 3, 5, 0, 1>, - Conv::template process_tile<0, 1, 3, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 5, 1, 0>, - Conv::template process_tile<0, 1, 3, 5, 1, 1>, - Conv::template process_tile<0, 1, 3, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 5, 2, 0>, - Conv::template process_tile<0, 1, 3, 5, 2, 1>, - Conv::template process_tile<0, 1, 3, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 6, 0, 0>, - Conv::template process_tile<0, 1, 3, 6, 0, 1>, - Conv::template process_tile<0, 1, 3, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 6, 1, 0>, - Conv::template process_tile<0, 1, 3, 6, 1, 1>, - Conv::template process_tile<0, 1, 3, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 6, 2, 0>, - Conv::template process_tile<0, 1, 3, 6, 2, 1>, - Conv::template process_tile<0, 1, 3, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 0, 0, 0>, - Conv::template process_tile<0, 1, 4, 0, 0, 1>, - Conv::template process_tile<0, 1, 4, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 0, 1, 0>, - Conv::template process_tile<0, 1, 4, 0, 1, 1>, - Conv::template process_tile<0, 1, 4, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 0, 2, 0>, - Conv::template process_tile<0, 1, 4, 0, 2, 1>, - Conv::template process_tile<0, 1, 4, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 1, 0, 0>, - Conv::template process_tile<0, 1, 4, 1, 0, 1>, - Conv::template process_tile<0, 1, 4, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 1, 1, 0>, - Conv::template process_tile<0, 1, 4, 1, 1, 1>, - Conv::template process_tile<0, 1, 4, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 1, 2, 0>, - Conv::template process_tile<0, 1, 4, 1, 2, 1>, - Conv::template process_tile<0, 1, 4, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 2, 0, 0>, - Conv::template process_tile<0, 1, 4, 2, 0, 1>, - Conv::template process_tile<0, 1, 4, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 2, 1, 0>, - Conv::template process_tile<0, 1, 4, 2, 1, 1>, - Conv::template process_tile<0, 1, 4, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 2, 2, 0>, - Conv::template process_tile<0, 1, 4, 2, 2, 1>, - Conv::template process_tile<0, 1, 4, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 3, 0, 0>, - Conv::template process_tile<0, 1, 4, 3, 0, 1>, - Conv::template process_tile<0, 1, 4, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 3, 1, 0>, - Conv::template process_tile<0, 1, 4, 3, 1, 1>, - Conv::template process_tile<0, 1, 4, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 3, 2, 0>, - Conv::template process_tile<0, 1, 4, 3, 2, 1>, - Conv::template process_tile<0, 1, 4, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 4, 0, 0>, - Conv::template process_tile<0, 1, 4, 4, 0, 1>, - Conv::template process_tile<0, 1, 4, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 4, 1, 0>, - Conv::template process_tile<0, 1, 4, 4, 1, 1>, - Conv::template process_tile<0, 1, 4, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 4, 2, 0>, - Conv::template process_tile<0, 1, 4, 4, 2, 1>, - Conv::template process_tile<0, 1, 4, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 5, 0, 0>, - Conv::template process_tile<0, 1, 4, 5, 0, 1>, - Conv::template process_tile<0, 1, 4, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 5, 1, 0>, - Conv::template process_tile<0, 1, 4, 5, 1, 1>, - Conv::template process_tile<0, 1, 4, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 5, 2, 0>, - Conv::template process_tile<0, 1, 4, 5, 2, 1>, - Conv::template process_tile<0, 1, 4, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 6, 0, 0>, - Conv::template process_tile<0, 1, 4, 6, 0, 1>, - Conv::template process_tile<0, 1, 4, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 6, 1, 0>, - Conv::template process_tile<0, 1, 4, 6, 1, 1>, - Conv::template process_tile<0, 1, 4, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 6, 2, 0>, - Conv::template process_tile<0, 1, 4, 6, 2, 1>, - Conv::template process_tile<0, 1, 4, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 4 - { // Input pad bottom = 5 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 0, 0, 0>, - Conv::template process_tile<0, 1, 5, 0, 0, 1>, - Conv::template process_tile<0, 1, 5, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 0, 1, 0>, - Conv::template process_tile<0, 1, 5, 0, 1, 1>, - Conv::template process_tile<0, 1, 5, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 0, 2, 0>, - Conv::template process_tile<0, 1, 5, 0, 2, 1>, - Conv::template process_tile<0, 1, 5, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 1, 0, 0>, - Conv::template process_tile<0, 1, 5, 1, 0, 1>, - Conv::template process_tile<0, 1, 5, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 1, 1, 0>, - Conv::template process_tile<0, 1, 5, 1, 1, 1>, - Conv::template process_tile<0, 1, 5, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 1, 2, 0>, - Conv::template process_tile<0, 1, 5, 1, 2, 1>, - Conv::template process_tile<0, 1, 5, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 2, 0, 0>, - Conv::template process_tile<0, 1, 5, 2, 0, 1>, - Conv::template process_tile<0, 1, 5, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 2, 1, 0>, - Conv::template process_tile<0, 1, 5, 2, 1, 1>, - Conv::template process_tile<0, 1, 5, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 2, 2, 0>, - Conv::template process_tile<0, 1, 5, 2, 2, 1>, - Conv::template process_tile<0, 1, 5, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 3, 0, 0>, - Conv::template process_tile<0, 1, 5, 3, 0, 1>, - Conv::template process_tile<0, 1, 5, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 3, 1, 0>, - Conv::template process_tile<0, 1, 5, 3, 1, 1>, - Conv::template process_tile<0, 1, 5, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 3, 2, 0>, - Conv::template process_tile<0, 1, 5, 3, 2, 1>, - Conv::template process_tile<0, 1, 5, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 4, 0, 0>, - Conv::template process_tile<0, 1, 5, 4, 0, 1>, - Conv::template process_tile<0, 1, 5, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 4, 1, 0>, - Conv::template process_tile<0, 1, 5, 4, 1, 1>, - Conv::template process_tile<0, 1, 5, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 4, 2, 0>, - Conv::template process_tile<0, 1, 5, 4, 2, 1>, - Conv::template process_tile<0, 1, 5, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 5, 0, 0>, - Conv::template process_tile<0, 1, 5, 5, 0, 1>, - Conv::template process_tile<0, 1, 5, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 5, 1, 0>, - Conv::template process_tile<0, 1, 5, 5, 1, 1>, - Conv::template process_tile<0, 1, 5, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 5, 2, 0>, - Conv::template process_tile<0, 1, 5, 5, 2, 1>, - Conv::template process_tile<0, 1, 5, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 6, 0, 0>, - Conv::template process_tile<0, 1, 5, 6, 0, 1>, - Conv::template process_tile<0, 1, 5, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 6, 1, 0>, - Conv::template process_tile<0, 1, 5, 6, 1, 1>, - Conv::template process_tile<0, 1, 5, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 6, 2, 0>, - Conv::template process_tile<0, 1, 5, 6, 2, 1>, - Conv::template process_tile<0, 1, 5, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 5 - { // Input pad bottom = 6 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 0, 0, 0>, - Conv::template process_tile<0, 1, 6, 0, 0, 1>, - Conv::template process_tile<0, 1, 6, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 0, 1, 0>, - Conv::template process_tile<0, 1, 6, 0, 1, 1>, - Conv::template process_tile<0, 1, 6, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 0, 2, 0>, - Conv::template process_tile<0, 1, 6, 0, 2, 1>, - Conv::template process_tile<0, 1, 6, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 1, 0, 0>, - Conv::template process_tile<0, 1, 6, 1, 0, 1>, - Conv::template process_tile<0, 1, 6, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 1, 1, 0>, - Conv::template process_tile<0, 1, 6, 1, 1, 1>, - Conv::template process_tile<0, 1, 6, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 1, 2, 0>, - Conv::template process_tile<0, 1, 6, 1, 2, 1>, - Conv::template process_tile<0, 1, 6, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 2, 0, 0>, - Conv::template process_tile<0, 1, 6, 2, 0, 1>, - Conv::template process_tile<0, 1, 6, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 2, 1, 0>, - Conv::template process_tile<0, 1, 6, 2, 1, 1>, - Conv::template process_tile<0, 1, 6, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 2, 2, 0>, - Conv::template process_tile<0, 1, 6, 2, 2, 1>, - Conv::template process_tile<0, 1, 6, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 3, 0, 0>, - Conv::template process_tile<0, 1, 6, 3, 0, 1>, - Conv::template process_tile<0, 1, 6, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 3, 1, 0>, - Conv::template process_tile<0, 1, 6, 3, 1, 1>, - Conv::template process_tile<0, 1, 6, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 3, 2, 0>, - Conv::template process_tile<0, 1, 6, 3, 2, 1>, - Conv::template process_tile<0, 1, 6, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 4, 0, 0>, - Conv::template process_tile<0, 1, 6, 4, 0, 1>, - Conv::template process_tile<0, 1, 6, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 4, 1, 0>, - Conv::template process_tile<0, 1, 6, 4, 1, 1>, - Conv::template process_tile<0, 1, 6, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 4, 2, 0>, - Conv::template process_tile<0, 1, 6, 4, 2, 1>, - Conv::template process_tile<0, 1, 6, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 5, 0, 0>, - Conv::template process_tile<0, 1, 6, 5, 0, 1>, - Conv::template process_tile<0, 1, 6, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 5, 1, 0>, - Conv::template process_tile<0, 1, 6, 5, 1, 1>, - Conv::template process_tile<0, 1, 6, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 5, 2, 0>, - Conv::template process_tile<0, 1, 6, 5, 2, 1>, - Conv::template process_tile<0, 1, 6, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 6, 0, 0>, - Conv::template process_tile<0, 1, 6, 6, 0, 1>, - Conv::template process_tile<0, 1, 6, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 6, 1, 0>, - Conv::template process_tile<0, 1, 6, 6, 1, 1>, - Conv::template process_tile<0, 1, 6, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 6, 2, 0>, - Conv::template process_tile<0, 1, 6, 6, 2, 1>, - Conv::template process_tile<0, 1, 6, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 6 - }, // Input pad left = 1 - }, // Input pad top = 0 - { // Input pad top = 1 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 0, 0, 0>, - Conv::template process_tile<1, 0, 0, 0, 0, 1>, - Conv::template process_tile<1, 0, 0, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 0, 1, 0>, - Conv::template process_tile<1, 0, 0, 0, 1, 1>, - Conv::template process_tile<1, 0, 0, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 0, 2, 0>, - Conv::template process_tile<1, 0, 0, 0, 2, 1>, - Conv::template process_tile<1, 0, 0, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 1, 0, 0>, - Conv::template process_tile<1, 0, 0, 1, 0, 1>, - Conv::template process_tile<1, 0, 0, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 1, 1, 0>, - Conv::template process_tile<1, 0, 0, 1, 1, 1>, - Conv::template process_tile<1, 0, 0, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 1, 2, 0>, - Conv::template process_tile<1, 0, 0, 1, 2, 1>, - Conv::template process_tile<1, 0, 0, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 2, 0, 0>, - Conv::template process_tile<1, 0, 0, 2, 0, 1>, - Conv::template process_tile<1, 0, 0, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 2, 1, 0>, - Conv::template process_tile<1, 0, 0, 2, 1, 1>, - Conv::template process_tile<1, 0, 0, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 2, 2, 0>, - Conv::template process_tile<1, 0, 0, 2, 2, 1>, - Conv::template process_tile<1, 0, 0, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 3, 0, 0>, - Conv::template process_tile<1, 0, 0, 3, 0, 1>, - Conv::template process_tile<1, 0, 0, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 3, 1, 0>, - Conv::template process_tile<1, 0, 0, 3, 1, 1>, - Conv::template process_tile<1, 0, 0, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 3, 2, 0>, - Conv::template process_tile<1, 0, 0, 3, 2, 1>, - Conv::template process_tile<1, 0, 0, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 4, 0, 0>, - Conv::template process_tile<1, 0, 0, 4, 0, 1>, - Conv::template process_tile<1, 0, 0, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 4, 1, 0>, - Conv::template process_tile<1, 0, 0, 4, 1, 1>, - Conv::template process_tile<1, 0, 0, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 4, 2, 0>, - Conv::template process_tile<1, 0, 0, 4, 2, 1>, - Conv::template process_tile<1, 0, 0, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 5, 0, 0>, - Conv::template process_tile<1, 0, 0, 5, 0, 1>, - Conv::template process_tile<1, 0, 0, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 5, 1, 0>, - Conv::template process_tile<1, 0, 0, 5, 1, 1>, - Conv::template process_tile<1, 0, 0, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 5, 2, 0>, - Conv::template process_tile<1, 0, 0, 5, 2, 1>, - Conv::template process_tile<1, 0, 0, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 6, 0, 0>, - Conv::template process_tile<1, 0, 0, 6, 0, 1>, - Conv::template process_tile<1, 0, 0, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 6, 1, 0>, - Conv::template process_tile<1, 0, 0, 6, 1, 1>, - Conv::template process_tile<1, 0, 0, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 6, 2, 0>, - Conv::template process_tile<1, 0, 0, 6, 2, 1>, - Conv::template process_tile<1, 0, 0, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 0, 0, 0>, - Conv::template process_tile<1, 0, 1, 0, 0, 1>, - Conv::template process_tile<1, 0, 1, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 0, 1, 0>, - Conv::template process_tile<1, 0, 1, 0, 1, 1>, - Conv::template process_tile<1, 0, 1, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 0, 2, 0>, - Conv::template process_tile<1, 0, 1, 0, 2, 1>, - Conv::template process_tile<1, 0, 1, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 1, 0, 0>, - Conv::template process_tile<1, 0, 1, 1, 0, 1>, - Conv::template process_tile<1, 0, 1, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 1, 1, 0>, - Conv::template process_tile<1, 0, 1, 1, 1, 1>, - Conv::template process_tile<1, 0, 1, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 1, 2, 0>, - Conv::template process_tile<1, 0, 1, 1, 2, 1>, - Conv::template process_tile<1, 0, 1, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 2, 0, 0>, - Conv::template process_tile<1, 0, 1, 2, 0, 1>, - Conv::template process_tile<1, 0, 1, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 2, 1, 0>, - Conv::template process_tile<1, 0, 1, 2, 1, 1>, - Conv::template process_tile<1, 0, 1, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 2, 2, 0>, - Conv::template process_tile<1, 0, 1, 2, 2, 1>, - Conv::template process_tile<1, 0, 1, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 3, 0, 0>, - Conv::template process_tile<1, 0, 1, 3, 0, 1>, - Conv::template process_tile<1, 0, 1, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 3, 1, 0>, - Conv::template process_tile<1, 0, 1, 3, 1, 1>, - Conv::template process_tile<1, 0, 1, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 3, 2, 0>, - Conv::template process_tile<1, 0, 1, 3, 2, 1>, - Conv::template process_tile<1, 0, 1, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 4, 0, 0>, - Conv::template process_tile<1, 0, 1, 4, 0, 1>, - Conv::template process_tile<1, 0, 1, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 4, 1, 0>, - Conv::template process_tile<1, 0, 1, 4, 1, 1>, - Conv::template process_tile<1, 0, 1, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 4, 2, 0>, - Conv::template process_tile<1, 0, 1, 4, 2, 1>, - Conv::template process_tile<1, 0, 1, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 5, 0, 0>, - Conv::template process_tile<1, 0, 1, 5, 0, 1>, - Conv::template process_tile<1, 0, 1, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 5, 1, 0>, - Conv::template process_tile<1, 0, 1, 5, 1, 1>, - Conv::template process_tile<1, 0, 1, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 5, 2, 0>, - Conv::template process_tile<1, 0, 1, 5, 2, 1>, - Conv::template process_tile<1, 0, 1, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 6, 0, 0>, - Conv::template process_tile<1, 0, 1, 6, 0, 1>, - Conv::template process_tile<1, 0, 1, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 6, 1, 0>, - Conv::template process_tile<1, 0, 1, 6, 1, 1>, - Conv::template process_tile<1, 0, 1, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 6, 2, 0>, - Conv::template process_tile<1, 0, 1, 6, 2, 1>, - Conv::template process_tile<1, 0, 1, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 0, 0, 0>, - Conv::template process_tile<1, 0, 2, 0, 0, 1>, - Conv::template process_tile<1, 0, 2, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 0, 1, 0>, - Conv::template process_tile<1, 0, 2, 0, 1, 1>, - Conv::template process_tile<1, 0, 2, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 0, 2, 0>, - Conv::template process_tile<1, 0, 2, 0, 2, 1>, - Conv::template process_tile<1, 0, 2, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 1, 0, 0>, - Conv::template process_tile<1, 0, 2, 1, 0, 1>, - Conv::template process_tile<1, 0, 2, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 1, 1, 0>, - Conv::template process_tile<1, 0, 2, 1, 1, 1>, - Conv::template process_tile<1, 0, 2, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 1, 2, 0>, - Conv::template process_tile<1, 0, 2, 1, 2, 1>, - Conv::template process_tile<1, 0, 2, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 2, 0, 0>, - Conv::template process_tile<1, 0, 2, 2, 0, 1>, - Conv::template process_tile<1, 0, 2, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 2, 1, 0>, - Conv::template process_tile<1, 0, 2, 2, 1, 1>, - Conv::template process_tile<1, 0, 2, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 2, 2, 0>, - Conv::template process_tile<1, 0, 2, 2, 2, 1>, - Conv::template process_tile<1, 0, 2, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 3, 0, 0>, - Conv::template process_tile<1, 0, 2, 3, 0, 1>, - Conv::template process_tile<1, 0, 2, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 3, 1, 0>, - Conv::template process_tile<1, 0, 2, 3, 1, 1>, - Conv::template process_tile<1, 0, 2, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 3, 2, 0>, - Conv::template process_tile<1, 0, 2, 3, 2, 1>, - Conv::template process_tile<1, 0, 2, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 4, 0, 0>, - Conv::template process_tile<1, 0, 2, 4, 0, 1>, - Conv::template process_tile<1, 0, 2, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 4, 1, 0>, - Conv::template process_tile<1, 0, 2, 4, 1, 1>, - Conv::template process_tile<1, 0, 2, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 4, 2, 0>, - Conv::template process_tile<1, 0, 2, 4, 2, 1>, - Conv::template process_tile<1, 0, 2, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 5, 0, 0>, - Conv::template process_tile<1, 0, 2, 5, 0, 1>, - Conv::template process_tile<1, 0, 2, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 5, 1, 0>, - Conv::template process_tile<1, 0, 2, 5, 1, 1>, - Conv::template process_tile<1, 0, 2, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 5, 2, 0>, - Conv::template process_tile<1, 0, 2, 5, 2, 1>, - Conv::template process_tile<1, 0, 2, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 6, 0, 0>, - Conv::template process_tile<1, 0, 2, 6, 0, 1>, - Conv::template process_tile<1, 0, 2, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 6, 1, 0>, - Conv::template process_tile<1, 0, 2, 6, 1, 1>, - Conv::template process_tile<1, 0, 2, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 6, 2, 0>, - Conv::template process_tile<1, 0, 2, 6, 2, 1>, - Conv::template process_tile<1, 0, 2, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 0, 0, 0>, - Conv::template process_tile<1, 0, 3, 0, 0, 1>, - Conv::template process_tile<1, 0, 3, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 0, 1, 0>, - Conv::template process_tile<1, 0, 3, 0, 1, 1>, - Conv::template process_tile<1, 0, 3, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 0, 2, 0>, - Conv::template process_tile<1, 0, 3, 0, 2, 1>, - Conv::template process_tile<1, 0, 3, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 1, 0, 0>, - Conv::template process_tile<1, 0, 3, 1, 0, 1>, - Conv::template process_tile<1, 0, 3, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 1, 1, 0>, - Conv::template process_tile<1, 0, 3, 1, 1, 1>, - Conv::template process_tile<1, 0, 3, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 1, 2, 0>, - Conv::template process_tile<1, 0, 3, 1, 2, 1>, - Conv::template process_tile<1, 0, 3, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 2, 0, 0>, - Conv::template process_tile<1, 0, 3, 2, 0, 1>, - Conv::template process_tile<1, 0, 3, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 2, 1, 0>, - Conv::template process_tile<1, 0, 3, 2, 1, 1>, - Conv::template process_tile<1, 0, 3, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 2, 2, 0>, - Conv::template process_tile<1, 0, 3, 2, 2, 1>, - Conv::template process_tile<1, 0, 3, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 3, 0, 0>, - Conv::template process_tile<1, 0, 3, 3, 0, 1>, - Conv::template process_tile<1, 0, 3, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 3, 1, 0>, - Conv::template process_tile<1, 0, 3, 3, 1, 1>, - Conv::template process_tile<1, 0, 3, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 3, 2, 0>, - Conv::template process_tile<1, 0, 3, 3, 2, 1>, - Conv::template process_tile<1, 0, 3, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 4, 0, 0>, - Conv::template process_tile<1, 0, 3, 4, 0, 1>, - Conv::template process_tile<1, 0, 3, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 4, 1, 0>, - Conv::template process_tile<1, 0, 3, 4, 1, 1>, - Conv::template process_tile<1, 0, 3, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 4, 2, 0>, - Conv::template process_tile<1, 0, 3, 4, 2, 1>, - Conv::template process_tile<1, 0, 3, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 5, 0, 0>, - Conv::template process_tile<1, 0, 3, 5, 0, 1>, - Conv::template process_tile<1, 0, 3, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 5, 1, 0>, - Conv::template process_tile<1, 0, 3, 5, 1, 1>, - Conv::template process_tile<1, 0, 3, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 5, 2, 0>, - Conv::template process_tile<1, 0, 3, 5, 2, 1>, - Conv::template process_tile<1, 0, 3, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 6, 0, 0>, - Conv::template process_tile<1, 0, 3, 6, 0, 1>, - Conv::template process_tile<1, 0, 3, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 6, 1, 0>, - Conv::template process_tile<1, 0, 3, 6, 1, 1>, - Conv::template process_tile<1, 0, 3, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 6, 2, 0>, - Conv::template process_tile<1, 0, 3, 6, 2, 1>, - Conv::template process_tile<1, 0, 3, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 0, 0, 0>, - Conv::template process_tile<1, 0, 4, 0, 0, 1>, - Conv::template process_tile<1, 0, 4, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 0, 1, 0>, - Conv::template process_tile<1, 0, 4, 0, 1, 1>, - Conv::template process_tile<1, 0, 4, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 0, 2, 0>, - Conv::template process_tile<1, 0, 4, 0, 2, 1>, - Conv::template process_tile<1, 0, 4, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 1, 0, 0>, - Conv::template process_tile<1, 0, 4, 1, 0, 1>, - Conv::template process_tile<1, 0, 4, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 1, 1, 0>, - Conv::template process_tile<1, 0, 4, 1, 1, 1>, - Conv::template process_tile<1, 0, 4, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 1, 2, 0>, - Conv::template process_tile<1, 0, 4, 1, 2, 1>, - Conv::template process_tile<1, 0, 4, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 2, 0, 0>, - Conv::template process_tile<1, 0, 4, 2, 0, 1>, - Conv::template process_tile<1, 0, 4, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 2, 1, 0>, - Conv::template process_tile<1, 0, 4, 2, 1, 1>, - Conv::template process_tile<1, 0, 4, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 2, 2, 0>, - Conv::template process_tile<1, 0, 4, 2, 2, 1>, - Conv::template process_tile<1, 0, 4, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 3, 0, 0>, - Conv::template process_tile<1, 0, 4, 3, 0, 1>, - Conv::template process_tile<1, 0, 4, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 3, 1, 0>, - Conv::template process_tile<1, 0, 4, 3, 1, 1>, - Conv::template process_tile<1, 0, 4, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 3, 2, 0>, - Conv::template process_tile<1, 0, 4, 3, 2, 1>, - Conv::template process_tile<1, 0, 4, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 4, 0, 0>, - Conv::template process_tile<1, 0, 4, 4, 0, 1>, - Conv::template process_tile<1, 0, 4, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 4, 1, 0>, - Conv::template process_tile<1, 0, 4, 4, 1, 1>, - Conv::template process_tile<1, 0, 4, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 4, 2, 0>, - Conv::template process_tile<1, 0, 4, 4, 2, 1>, - Conv::template process_tile<1, 0, 4, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 5, 0, 0>, - Conv::template process_tile<1, 0, 4, 5, 0, 1>, - Conv::template process_tile<1, 0, 4, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 5, 1, 0>, - Conv::template process_tile<1, 0, 4, 5, 1, 1>, - Conv::template process_tile<1, 0, 4, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 5, 2, 0>, - Conv::template process_tile<1, 0, 4, 5, 2, 1>, - Conv::template process_tile<1, 0, 4, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 6, 0, 0>, - Conv::template process_tile<1, 0, 4, 6, 0, 1>, - Conv::template process_tile<1, 0, 4, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 6, 1, 0>, - Conv::template process_tile<1, 0, 4, 6, 1, 1>, - Conv::template process_tile<1, 0, 4, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 6, 2, 0>, - Conv::template process_tile<1, 0, 4, 6, 2, 1>, - Conv::template process_tile<1, 0, 4, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 4 - { // Input pad bottom = 5 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 0, 0, 0>, - Conv::template process_tile<1, 0, 5, 0, 0, 1>, - Conv::template process_tile<1, 0, 5, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 0, 1, 0>, - Conv::template process_tile<1, 0, 5, 0, 1, 1>, - Conv::template process_tile<1, 0, 5, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 0, 2, 0>, - Conv::template process_tile<1, 0, 5, 0, 2, 1>, - Conv::template process_tile<1, 0, 5, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 1, 0, 0>, - Conv::template process_tile<1, 0, 5, 1, 0, 1>, - Conv::template process_tile<1, 0, 5, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 1, 1, 0>, - Conv::template process_tile<1, 0, 5, 1, 1, 1>, - Conv::template process_tile<1, 0, 5, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 1, 2, 0>, - Conv::template process_tile<1, 0, 5, 1, 2, 1>, - Conv::template process_tile<1, 0, 5, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 2, 0, 0>, - Conv::template process_tile<1, 0, 5, 2, 0, 1>, - Conv::template process_tile<1, 0, 5, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 2, 1, 0>, - Conv::template process_tile<1, 0, 5, 2, 1, 1>, - Conv::template process_tile<1, 0, 5, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 2, 2, 0>, - Conv::template process_tile<1, 0, 5, 2, 2, 1>, - Conv::template process_tile<1, 0, 5, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 3, 0, 0>, - Conv::template process_tile<1, 0, 5, 3, 0, 1>, - Conv::template process_tile<1, 0, 5, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 3, 1, 0>, - Conv::template process_tile<1, 0, 5, 3, 1, 1>, - Conv::template process_tile<1, 0, 5, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 3, 2, 0>, - Conv::template process_tile<1, 0, 5, 3, 2, 1>, - Conv::template process_tile<1, 0, 5, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 4, 0, 0>, - Conv::template process_tile<1, 0, 5, 4, 0, 1>, - Conv::template process_tile<1, 0, 5, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 4, 1, 0>, - Conv::template process_tile<1, 0, 5, 4, 1, 1>, - Conv::template process_tile<1, 0, 5, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 4, 2, 0>, - Conv::template process_tile<1, 0, 5, 4, 2, 1>, - Conv::template process_tile<1, 0, 5, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 5, 0, 0>, - Conv::template process_tile<1, 0, 5, 5, 0, 1>, - Conv::template process_tile<1, 0, 5, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 5, 1, 0>, - Conv::template process_tile<1, 0, 5, 5, 1, 1>, - Conv::template process_tile<1, 0, 5, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 5, 2, 0>, - Conv::template process_tile<1, 0, 5, 5, 2, 1>, - Conv::template process_tile<1, 0, 5, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 6, 0, 0>, - Conv::template process_tile<1, 0, 5, 6, 0, 1>, - Conv::template process_tile<1, 0, 5, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 6, 1, 0>, - Conv::template process_tile<1, 0, 5, 6, 1, 1>, - Conv::template process_tile<1, 0, 5, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 6, 2, 0>, - Conv::template process_tile<1, 0, 5, 6, 2, 1>, - Conv::template process_tile<1, 0, 5, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 5 - { // Input pad bottom = 6 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 0, 0, 0>, - Conv::template process_tile<1, 0, 6, 0, 0, 1>, - Conv::template process_tile<1, 0, 6, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 0, 1, 0>, - Conv::template process_tile<1, 0, 6, 0, 1, 1>, - Conv::template process_tile<1, 0, 6, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 0, 2, 0>, - Conv::template process_tile<1, 0, 6, 0, 2, 1>, - Conv::template process_tile<1, 0, 6, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 1, 0, 0>, - Conv::template process_tile<1, 0, 6, 1, 0, 1>, - Conv::template process_tile<1, 0, 6, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 1, 1, 0>, - Conv::template process_tile<1, 0, 6, 1, 1, 1>, - Conv::template process_tile<1, 0, 6, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 1, 2, 0>, - Conv::template process_tile<1, 0, 6, 1, 2, 1>, - Conv::template process_tile<1, 0, 6, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 2, 0, 0>, - Conv::template process_tile<1, 0, 6, 2, 0, 1>, - Conv::template process_tile<1, 0, 6, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 2, 1, 0>, - Conv::template process_tile<1, 0, 6, 2, 1, 1>, - Conv::template process_tile<1, 0, 6, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 2, 2, 0>, - Conv::template process_tile<1, 0, 6, 2, 2, 1>, - Conv::template process_tile<1, 0, 6, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 3, 0, 0>, - Conv::template process_tile<1, 0, 6, 3, 0, 1>, - Conv::template process_tile<1, 0, 6, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 3, 1, 0>, - Conv::template process_tile<1, 0, 6, 3, 1, 1>, - Conv::template process_tile<1, 0, 6, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 3, 2, 0>, - Conv::template process_tile<1, 0, 6, 3, 2, 1>, - Conv::template process_tile<1, 0, 6, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 4, 0, 0>, - Conv::template process_tile<1, 0, 6, 4, 0, 1>, - Conv::template process_tile<1, 0, 6, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 4, 1, 0>, - Conv::template process_tile<1, 0, 6, 4, 1, 1>, - Conv::template process_tile<1, 0, 6, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 4, 2, 0>, - Conv::template process_tile<1, 0, 6, 4, 2, 1>, - Conv::template process_tile<1, 0, 6, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 5, 0, 0>, - Conv::template process_tile<1, 0, 6, 5, 0, 1>, - Conv::template process_tile<1, 0, 6, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 5, 1, 0>, - Conv::template process_tile<1, 0, 6, 5, 1, 1>, - Conv::template process_tile<1, 0, 6, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 5, 2, 0>, - Conv::template process_tile<1, 0, 6, 5, 2, 1>, - Conv::template process_tile<1, 0, 6, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 6, 0, 0>, - Conv::template process_tile<1, 0, 6, 6, 0, 1>, - Conv::template process_tile<1, 0, 6, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 6, 1, 0>, - Conv::template process_tile<1, 0, 6, 6, 1, 1>, - Conv::template process_tile<1, 0, 6, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 6, 2, 0>, - Conv::template process_tile<1, 0, 6, 6, 2, 1>, - Conv::template process_tile<1, 0, 6, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 6 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 0, 0, 0>, - Conv::template process_tile<1, 1, 0, 0, 0, 1>, - Conv::template process_tile<1, 1, 0, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 0, 1, 0>, - Conv::template process_tile<1, 1, 0, 0, 1, 1>, - Conv::template process_tile<1, 1, 0, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 0, 2, 0>, - Conv::template process_tile<1, 1, 0, 0, 2, 1>, - Conv::template process_tile<1, 1, 0, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 1, 0, 0>, - Conv::template process_tile<1, 1, 0, 1, 0, 1>, - Conv::template process_tile<1, 1, 0, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 1, 1, 0>, - Conv::template process_tile<1, 1, 0, 1, 1, 1>, - Conv::template process_tile<1, 1, 0, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 1, 2, 0>, - Conv::template process_tile<1, 1, 0, 1, 2, 1>, - Conv::template process_tile<1, 1, 0, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 2, 0, 0>, - Conv::template process_tile<1, 1, 0, 2, 0, 1>, - Conv::template process_tile<1, 1, 0, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 2, 1, 0>, - Conv::template process_tile<1, 1, 0, 2, 1, 1>, - Conv::template process_tile<1, 1, 0, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 2, 2, 0>, - Conv::template process_tile<1, 1, 0, 2, 2, 1>, - Conv::template process_tile<1, 1, 0, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 3, 0, 0>, - Conv::template process_tile<1, 1, 0, 3, 0, 1>, - Conv::template process_tile<1, 1, 0, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 3, 1, 0>, - Conv::template process_tile<1, 1, 0, 3, 1, 1>, - Conv::template process_tile<1, 1, 0, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 3, 2, 0>, - Conv::template process_tile<1, 1, 0, 3, 2, 1>, - Conv::template process_tile<1, 1, 0, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 4, 0, 0>, - Conv::template process_tile<1, 1, 0, 4, 0, 1>, - Conv::template process_tile<1, 1, 0, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 4, 1, 0>, - Conv::template process_tile<1, 1, 0, 4, 1, 1>, - Conv::template process_tile<1, 1, 0, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 4, 2, 0>, - Conv::template process_tile<1, 1, 0, 4, 2, 1>, - Conv::template process_tile<1, 1, 0, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 5, 0, 0>, - Conv::template process_tile<1, 1, 0, 5, 0, 1>, - Conv::template process_tile<1, 1, 0, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 5, 1, 0>, - Conv::template process_tile<1, 1, 0, 5, 1, 1>, - Conv::template process_tile<1, 1, 0, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 5, 2, 0>, - Conv::template process_tile<1, 1, 0, 5, 2, 1>, - Conv::template process_tile<1, 1, 0, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 6, 0, 0>, - Conv::template process_tile<1, 1, 0, 6, 0, 1>, - Conv::template process_tile<1, 1, 0, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 6, 1, 0>, - Conv::template process_tile<1, 1, 0, 6, 1, 1>, - Conv::template process_tile<1, 1, 0, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 6, 2, 0>, - Conv::template process_tile<1, 1, 0, 6, 2, 1>, - Conv::template process_tile<1, 1, 0, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 0, 0, 0>, - Conv::template process_tile<1, 1, 1, 0, 0, 1>, - Conv::template process_tile<1, 1, 1, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 0, 1, 0>, - Conv::template process_tile<1, 1, 1, 0, 1, 1>, - Conv::template process_tile<1, 1, 1, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 0, 2, 0>, - Conv::template process_tile<1, 1, 1, 0, 2, 1>, - Conv::template process_tile<1, 1, 1, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 1, 0, 0>, - Conv::template process_tile<1, 1, 1, 1, 0, 1>, - Conv::template process_tile<1, 1, 1, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 1, 1, 0>, - Conv::template process_tile<1, 1, 1, 1, 1, 1>, - Conv::template process_tile<1, 1, 1, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 1, 2, 0>, - Conv::template process_tile<1, 1, 1, 1, 2, 1>, - Conv::template process_tile<1, 1, 1, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 2, 0, 0>, - Conv::template process_tile<1, 1, 1, 2, 0, 1>, - Conv::template process_tile<1, 1, 1, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 2, 1, 0>, - Conv::template process_tile<1, 1, 1, 2, 1, 1>, - Conv::template process_tile<1, 1, 1, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 2, 2, 0>, - Conv::template process_tile<1, 1, 1, 2, 2, 1>, - Conv::template process_tile<1, 1, 1, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 3, 0, 0>, - Conv::template process_tile<1, 1, 1, 3, 0, 1>, - Conv::template process_tile<1, 1, 1, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 3, 1, 0>, - Conv::template process_tile<1, 1, 1, 3, 1, 1>, - Conv::template process_tile<1, 1, 1, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 3, 2, 0>, - Conv::template process_tile<1, 1, 1, 3, 2, 1>, - Conv::template process_tile<1, 1, 1, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 4, 0, 0>, - Conv::template process_tile<1, 1, 1, 4, 0, 1>, - Conv::template process_tile<1, 1, 1, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 4, 1, 0>, - Conv::template process_tile<1, 1, 1, 4, 1, 1>, - Conv::template process_tile<1, 1, 1, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 4, 2, 0>, - Conv::template process_tile<1, 1, 1, 4, 2, 1>, - Conv::template process_tile<1, 1, 1, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 5, 0, 0>, - Conv::template process_tile<1, 1, 1, 5, 0, 1>, - Conv::template process_tile<1, 1, 1, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 5, 1, 0>, - Conv::template process_tile<1, 1, 1, 5, 1, 1>, - Conv::template process_tile<1, 1, 1, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 5, 2, 0>, - Conv::template process_tile<1, 1, 1, 5, 2, 1>, - Conv::template process_tile<1, 1, 1, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 6, 0, 0>, - Conv::template process_tile<1, 1, 1, 6, 0, 1>, - Conv::template process_tile<1, 1, 1, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 6, 1, 0>, - Conv::template process_tile<1, 1, 1, 6, 1, 1>, - Conv::template process_tile<1, 1, 1, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 6, 2, 0>, - Conv::template process_tile<1, 1, 1, 6, 2, 1>, - Conv::template process_tile<1, 1, 1, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 0, 0, 0>, - Conv::template process_tile<1, 1, 2, 0, 0, 1>, - Conv::template process_tile<1, 1, 2, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 0, 1, 0>, - Conv::template process_tile<1, 1, 2, 0, 1, 1>, - Conv::template process_tile<1, 1, 2, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 0, 2, 0>, - Conv::template process_tile<1, 1, 2, 0, 2, 1>, - Conv::template process_tile<1, 1, 2, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 1, 0, 0>, - Conv::template process_tile<1, 1, 2, 1, 0, 1>, - Conv::template process_tile<1, 1, 2, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 1, 1, 0>, - Conv::template process_tile<1, 1, 2, 1, 1, 1>, - Conv::template process_tile<1, 1, 2, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 1, 2, 0>, - Conv::template process_tile<1, 1, 2, 1, 2, 1>, - Conv::template process_tile<1, 1, 2, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 2, 0, 0>, - Conv::template process_tile<1, 1, 2, 2, 0, 1>, - Conv::template process_tile<1, 1, 2, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 2, 1, 0>, - Conv::template process_tile<1, 1, 2, 2, 1, 1>, - Conv::template process_tile<1, 1, 2, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 2, 2, 0>, - Conv::template process_tile<1, 1, 2, 2, 2, 1>, - Conv::template process_tile<1, 1, 2, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 3, 0, 0>, - Conv::template process_tile<1, 1, 2, 3, 0, 1>, - Conv::template process_tile<1, 1, 2, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 3, 1, 0>, - Conv::template process_tile<1, 1, 2, 3, 1, 1>, - Conv::template process_tile<1, 1, 2, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 3, 2, 0>, - Conv::template process_tile<1, 1, 2, 3, 2, 1>, - Conv::template process_tile<1, 1, 2, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 4, 0, 0>, - Conv::template process_tile<1, 1, 2, 4, 0, 1>, - Conv::template process_tile<1, 1, 2, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 4, 1, 0>, - Conv::template process_tile<1, 1, 2, 4, 1, 1>, - Conv::template process_tile<1, 1, 2, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 4, 2, 0>, - Conv::template process_tile<1, 1, 2, 4, 2, 1>, - Conv::template process_tile<1, 1, 2, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 5, 0, 0>, - Conv::template process_tile<1, 1, 2, 5, 0, 1>, - Conv::template process_tile<1, 1, 2, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 5, 1, 0>, - Conv::template process_tile<1, 1, 2, 5, 1, 1>, - Conv::template process_tile<1, 1, 2, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 5, 2, 0>, - Conv::template process_tile<1, 1, 2, 5, 2, 1>, - Conv::template process_tile<1, 1, 2, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 6, 0, 0>, - Conv::template process_tile<1, 1, 2, 6, 0, 1>, - Conv::template process_tile<1, 1, 2, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 6, 1, 0>, - Conv::template process_tile<1, 1, 2, 6, 1, 1>, - Conv::template process_tile<1, 1, 2, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 6, 2, 0>, - Conv::template process_tile<1, 1, 2, 6, 2, 1>, - Conv::template process_tile<1, 1, 2, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 0, 0, 0>, - Conv::template process_tile<1, 1, 3, 0, 0, 1>, - Conv::template process_tile<1, 1, 3, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 0, 1, 0>, - Conv::template process_tile<1, 1, 3, 0, 1, 1>, - Conv::template process_tile<1, 1, 3, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 0, 2, 0>, - Conv::template process_tile<1, 1, 3, 0, 2, 1>, - Conv::template process_tile<1, 1, 3, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 1, 0, 0>, - Conv::template process_tile<1, 1, 3, 1, 0, 1>, - Conv::template process_tile<1, 1, 3, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 1, 1, 0>, - Conv::template process_tile<1, 1, 3, 1, 1, 1>, - Conv::template process_tile<1, 1, 3, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 1, 2, 0>, - Conv::template process_tile<1, 1, 3, 1, 2, 1>, - Conv::template process_tile<1, 1, 3, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 2, 0, 0>, - Conv::template process_tile<1, 1, 3, 2, 0, 1>, - Conv::template process_tile<1, 1, 3, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 2, 1, 0>, - Conv::template process_tile<1, 1, 3, 2, 1, 1>, - Conv::template process_tile<1, 1, 3, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 2, 2, 0>, - Conv::template process_tile<1, 1, 3, 2, 2, 1>, - Conv::template process_tile<1, 1, 3, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 3, 0, 0>, - Conv::template process_tile<1, 1, 3, 3, 0, 1>, - Conv::template process_tile<1, 1, 3, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 3, 1, 0>, - Conv::template process_tile<1, 1, 3, 3, 1, 1>, - Conv::template process_tile<1, 1, 3, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 3, 2, 0>, - Conv::template process_tile<1, 1, 3, 3, 2, 1>, - Conv::template process_tile<1, 1, 3, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 4, 0, 0>, - Conv::template process_tile<1, 1, 3, 4, 0, 1>, - Conv::template process_tile<1, 1, 3, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 4, 1, 0>, - Conv::template process_tile<1, 1, 3, 4, 1, 1>, - Conv::template process_tile<1, 1, 3, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 4, 2, 0>, - Conv::template process_tile<1, 1, 3, 4, 2, 1>, - Conv::template process_tile<1, 1, 3, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 5, 0, 0>, - Conv::template process_tile<1, 1, 3, 5, 0, 1>, - Conv::template process_tile<1, 1, 3, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 5, 1, 0>, - Conv::template process_tile<1, 1, 3, 5, 1, 1>, - Conv::template process_tile<1, 1, 3, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 5, 2, 0>, - Conv::template process_tile<1, 1, 3, 5, 2, 1>, - Conv::template process_tile<1, 1, 3, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 6, 0, 0>, - Conv::template process_tile<1, 1, 3, 6, 0, 1>, - Conv::template process_tile<1, 1, 3, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 6, 1, 0>, - Conv::template process_tile<1, 1, 3, 6, 1, 1>, - Conv::template process_tile<1, 1, 3, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 6, 2, 0>, - Conv::template process_tile<1, 1, 3, 6, 2, 1>, - Conv::template process_tile<1, 1, 3, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 0, 0, 0>, - Conv::template process_tile<1, 1, 4, 0, 0, 1>, - Conv::template process_tile<1, 1, 4, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 0, 1, 0>, - Conv::template process_tile<1, 1, 4, 0, 1, 1>, - Conv::template process_tile<1, 1, 4, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 0, 2, 0>, - Conv::template process_tile<1, 1, 4, 0, 2, 1>, - Conv::template process_tile<1, 1, 4, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 1, 0, 0>, - Conv::template process_tile<1, 1, 4, 1, 0, 1>, - Conv::template process_tile<1, 1, 4, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 1, 1, 0>, - Conv::template process_tile<1, 1, 4, 1, 1, 1>, - Conv::template process_tile<1, 1, 4, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 1, 2, 0>, - Conv::template process_tile<1, 1, 4, 1, 2, 1>, - Conv::template process_tile<1, 1, 4, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 2, 0, 0>, - Conv::template process_tile<1, 1, 4, 2, 0, 1>, - Conv::template process_tile<1, 1, 4, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 2, 1, 0>, - Conv::template process_tile<1, 1, 4, 2, 1, 1>, - Conv::template process_tile<1, 1, 4, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 2, 2, 0>, - Conv::template process_tile<1, 1, 4, 2, 2, 1>, - Conv::template process_tile<1, 1, 4, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 3, 0, 0>, - Conv::template process_tile<1, 1, 4, 3, 0, 1>, - Conv::template process_tile<1, 1, 4, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 3, 1, 0>, - Conv::template process_tile<1, 1, 4, 3, 1, 1>, - Conv::template process_tile<1, 1, 4, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 3, 2, 0>, - Conv::template process_tile<1, 1, 4, 3, 2, 1>, - Conv::template process_tile<1, 1, 4, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 4, 0, 0>, - Conv::template process_tile<1, 1, 4, 4, 0, 1>, - Conv::template process_tile<1, 1, 4, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 4, 1, 0>, - Conv::template process_tile<1, 1, 4, 4, 1, 1>, - Conv::template process_tile<1, 1, 4, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 4, 2, 0>, - Conv::template process_tile<1, 1, 4, 4, 2, 1>, - Conv::template process_tile<1, 1, 4, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 5, 0, 0>, - Conv::template process_tile<1, 1, 4, 5, 0, 1>, - Conv::template process_tile<1, 1, 4, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 5, 1, 0>, - Conv::template process_tile<1, 1, 4, 5, 1, 1>, - Conv::template process_tile<1, 1, 4, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 5, 2, 0>, - Conv::template process_tile<1, 1, 4, 5, 2, 1>, - Conv::template process_tile<1, 1, 4, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 6, 0, 0>, - Conv::template process_tile<1, 1, 4, 6, 0, 1>, - Conv::template process_tile<1, 1, 4, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 6, 1, 0>, - Conv::template process_tile<1, 1, 4, 6, 1, 1>, - Conv::template process_tile<1, 1, 4, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 6, 2, 0>, - Conv::template process_tile<1, 1, 4, 6, 2, 1>, - Conv::template process_tile<1, 1, 4, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 4 - { // Input pad bottom = 5 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 0, 0, 0>, - Conv::template process_tile<1, 1, 5, 0, 0, 1>, - Conv::template process_tile<1, 1, 5, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 0, 1, 0>, - Conv::template process_tile<1, 1, 5, 0, 1, 1>, - Conv::template process_tile<1, 1, 5, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 0, 2, 0>, - Conv::template process_tile<1, 1, 5, 0, 2, 1>, - Conv::template process_tile<1, 1, 5, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 1, 0, 0>, - Conv::template process_tile<1, 1, 5, 1, 0, 1>, - Conv::template process_tile<1, 1, 5, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 1, 1, 0>, - Conv::template process_tile<1, 1, 5, 1, 1, 1>, - Conv::template process_tile<1, 1, 5, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 1, 2, 0>, - Conv::template process_tile<1, 1, 5, 1, 2, 1>, - Conv::template process_tile<1, 1, 5, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 2, 0, 0>, - Conv::template process_tile<1, 1, 5, 2, 0, 1>, - Conv::template process_tile<1, 1, 5, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 2, 1, 0>, - Conv::template process_tile<1, 1, 5, 2, 1, 1>, - Conv::template process_tile<1, 1, 5, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 2, 2, 0>, - Conv::template process_tile<1, 1, 5, 2, 2, 1>, - Conv::template process_tile<1, 1, 5, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 3, 0, 0>, - Conv::template process_tile<1, 1, 5, 3, 0, 1>, - Conv::template process_tile<1, 1, 5, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 3, 1, 0>, - Conv::template process_tile<1, 1, 5, 3, 1, 1>, - Conv::template process_tile<1, 1, 5, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 3, 2, 0>, - Conv::template process_tile<1, 1, 5, 3, 2, 1>, - Conv::template process_tile<1, 1, 5, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 4, 0, 0>, - Conv::template process_tile<1, 1, 5, 4, 0, 1>, - Conv::template process_tile<1, 1, 5, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 4, 1, 0>, - Conv::template process_tile<1, 1, 5, 4, 1, 1>, - Conv::template process_tile<1, 1, 5, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 4, 2, 0>, - Conv::template process_tile<1, 1, 5, 4, 2, 1>, - Conv::template process_tile<1, 1, 5, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 5, 0, 0>, - Conv::template process_tile<1, 1, 5, 5, 0, 1>, - Conv::template process_tile<1, 1, 5, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 5, 1, 0>, - Conv::template process_tile<1, 1, 5, 5, 1, 1>, - Conv::template process_tile<1, 1, 5, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 5, 2, 0>, - Conv::template process_tile<1, 1, 5, 5, 2, 1>, - Conv::template process_tile<1, 1, 5, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 6, 0, 0>, - Conv::template process_tile<1, 1, 5, 6, 0, 1>, - Conv::template process_tile<1, 1, 5, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 6, 1, 0>, - Conv::template process_tile<1, 1, 5, 6, 1, 1>, - Conv::template process_tile<1, 1, 5, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 6, 2, 0>, - Conv::template process_tile<1, 1, 5, 6, 2, 1>, - Conv::template process_tile<1, 1, 5, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 5 - { // Input pad bottom = 6 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 0, 0, 0>, - Conv::template process_tile<1, 1, 6, 0, 0, 1>, - Conv::template process_tile<1, 1, 6, 0, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 0, 1, 0>, - Conv::template process_tile<1, 1, 6, 0, 1, 1>, - Conv::template process_tile<1, 1, 6, 0, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 0, 2, 0>, - Conv::template process_tile<1, 1, 6, 0, 2, 1>, - Conv::template process_tile<1, 1, 6, 0, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 1, 0, 0>, - Conv::template process_tile<1, 1, 6, 1, 0, 1>, - Conv::template process_tile<1, 1, 6, 1, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 1, 1, 0>, - Conv::template process_tile<1, 1, 6, 1, 1, 1>, - Conv::template process_tile<1, 1, 6, 1, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 1, 2, 0>, - Conv::template process_tile<1, 1, 6, 1, 2, 1>, - Conv::template process_tile<1, 1, 6, 1, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 2, 0, 0>, - Conv::template process_tile<1, 1, 6, 2, 0, 1>, - Conv::template process_tile<1, 1, 6, 2, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 2, 1, 0>, - Conv::template process_tile<1, 1, 6, 2, 1, 1>, - Conv::template process_tile<1, 1, 6, 2, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 2, 2, 0>, - Conv::template process_tile<1, 1, 6, 2, 2, 1>, - Conv::template process_tile<1, 1, 6, 2, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 3, 0, 0>, - Conv::template process_tile<1, 1, 6, 3, 0, 1>, - Conv::template process_tile<1, 1, 6, 3, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 3, 1, 0>, - Conv::template process_tile<1, 1, 6, 3, 1, 1>, - Conv::template process_tile<1, 1, 6, 3, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 3, 2, 0>, - Conv::template process_tile<1, 1, 6, 3, 2, 1>, - Conv::template process_tile<1, 1, 6, 3, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 4, 0, 0>, - Conv::template process_tile<1, 1, 6, 4, 0, 1>, - Conv::template process_tile<1, 1, 6, 4, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 4, 1, 0>, - Conv::template process_tile<1, 1, 6, 4, 1, 1>, - Conv::template process_tile<1, 1, 6, 4, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 4, 2, 0>, - Conv::template process_tile<1, 1, 6, 4, 2, 1>, - Conv::template process_tile<1, 1, 6, 4, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 5, 0, 0>, - Conv::template process_tile<1, 1, 6, 5, 0, 1>, - Conv::template process_tile<1, 1, 6, 5, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 5, 1, 0>, - Conv::template process_tile<1, 1, 6, 5, 1, 1>, - Conv::template process_tile<1, 1, 6, 5, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 5, 2, 0>, - Conv::template process_tile<1, 1, 6, 5, 2, 1>, - Conv::template process_tile<1, 1, 6, 5, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 6, 0, 0>, - Conv::template process_tile<1, 1, 6, 6, 0, 1>, - Conv::template process_tile<1, 1, 6, 6, 0, 2>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 6, 1, 0>, - Conv::template process_tile<1, 1, 6, 6, 1, 1>, - Conv::template process_tile<1, 1, 6, 6, 1, 2>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 6, 2, 0>, - Conv::template process_tile<1, 1, 6, 6, 2, 1>, - Conv::template process_tile<1, 1, 6, 6, 2, 2>, - }, // Output pad bottom = 2 - }, // Input pad right = 6 - }, // Input pad bottom = 6 - }, // Input pad left = 1 - }, // Input pad top = 1 +const Conv::TileFn Conv::tilefn_left[n_in_pad_left_fns] = { + ConvImpl::template process_tile, + ConvImpl::template process_tile, }; +template <> +const Conv::TileFn Conv::tilefn_bottom[n_in_pad_bottom_fns][n_out_pad_bottom_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, +}; + +template <> +const Conv::TileFn Conv::tilefn_right[n_in_pad_right_fns][n_out_pad_right_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, +}; + +template <> +const Conv::TileFn Conv::tilefn_generic = ConvImpl::template process_tile; template class DepthwiseConvolution<3, 3, 3, 3, 2, 2, float, float>; } // namespace depthwise diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_4x4_3x3_1x1_fp32_fp32.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_4x4_3x3_1x1_fp32_fp32.cpp index a1aaaa078c..44b93a12d6 100644 --- a/src/core/NEON/kernels/convolution/depthwise/depthwise_4x4_3x3_1x1_fp32_fp32.cpp +++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_4x4_3x3_1x1_fp32_fp32.cpp @@ -28,2668 +28,1465 @@ namespace depthwise using Conv = DepthwiseConvolution<4, 4, 3, 3, 1, 1, float, float>; using ConvImpl = DepthwiseConvolutionImpl<4, 4, 3, 3, 1, 1, float, float>; +#ifdef __aarch64__ + +template <> template <> -const Conv::TileFn Conv::tile_fns - [max_in_pad_top] - [max_in_pad_left] - [max_in_pad_bottom] - [max_in_pad_right] - [max_out_pad_bottom] - [max_out_pad_right] = { - { // Input pad top = 0 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 0, 0, 0, 0>, - ConvImpl::template process_tile<0, 0, 0, 0, 0, 1>, - ConvImpl::template process_tile<0, 0, 0, 0, 0, 2>, - ConvImpl::template process_tile<0, 0, 0, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 0, 0, 1, 0>, - ConvImpl::template process_tile<0, 0, 0, 0, 1, 1>, - ConvImpl::template process_tile<0, 0, 0, 0, 1, 2>, - ConvImpl::template process_tile<0, 0, 0, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 0, 0, 2, 0>, - ConvImpl::template process_tile<0, 0, 0, 0, 2, 1>, - ConvImpl::template process_tile<0, 0, 0, 0, 2, 2>, - ConvImpl::template process_tile<0, 0, 0, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 0, 0, 3, 0>, - ConvImpl::template process_tile<0, 0, 0, 0, 3, 1>, - ConvImpl::template process_tile<0, 0, 0, 0, 3, 2>, - ConvImpl::template process_tile<0, 0, 0, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 0, 1, 0, 0>, - ConvImpl::template process_tile<0, 0, 0, 1, 0, 1>, - ConvImpl::template process_tile<0, 0, 0, 1, 0, 2>, - ConvImpl::template process_tile<0, 0, 0, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 0, 1, 1, 0>, - ConvImpl::template process_tile<0, 0, 0, 1, 1, 1>, - ConvImpl::template process_tile<0, 0, 0, 1, 1, 2>, - ConvImpl::template process_tile<0, 0, 0, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 0, 1, 2, 0>, - ConvImpl::template process_tile<0, 0, 0, 1, 2, 1>, - ConvImpl::template process_tile<0, 0, 0, 1, 2, 2>, - ConvImpl::template process_tile<0, 0, 0, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 0, 1, 3, 0>, - ConvImpl::template process_tile<0, 0, 0, 1, 3, 1>, - ConvImpl::template process_tile<0, 0, 0, 1, 3, 2>, - ConvImpl::template process_tile<0, 0, 0, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 0, 2, 0, 0>, - ConvImpl::template process_tile<0, 0, 0, 2, 0, 1>, - ConvImpl::template process_tile<0, 0, 0, 2, 0, 2>, - ConvImpl::template process_tile<0, 0, 0, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 0, 2, 1, 0>, - ConvImpl::template process_tile<0, 0, 0, 2, 1, 1>, - ConvImpl::template process_tile<0, 0, 0, 2, 1, 2>, - ConvImpl::template process_tile<0, 0, 0, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 0, 2, 2, 0>, - ConvImpl::template process_tile<0, 0, 0, 2, 2, 1>, - ConvImpl::template process_tile<0, 0, 0, 2, 2, 2>, - ConvImpl::template process_tile<0, 0, 0, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 0, 2, 3, 0>, - ConvImpl::template process_tile<0, 0, 0, 2, 3, 1>, - ConvImpl::template process_tile<0, 0, 0, 2, 3, 2>, - ConvImpl::template process_tile<0, 0, 0, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 0, 3, 0, 0>, - ConvImpl::template process_tile<0, 0, 0, 3, 0, 1>, - ConvImpl::template process_tile<0, 0, 0, 3, 0, 2>, - ConvImpl::template process_tile<0, 0, 0, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 0, 3, 1, 0>, - ConvImpl::template process_tile<0, 0, 0, 3, 1, 1>, - ConvImpl::template process_tile<0, 0, 0, 3, 1, 2>, - ConvImpl::template process_tile<0, 0, 0, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 0, 3, 2, 0>, - ConvImpl::template process_tile<0, 0, 0, 3, 2, 1>, - ConvImpl::template process_tile<0, 0, 0, 3, 2, 2>, - ConvImpl::template process_tile<0, 0, 0, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 0, 3, 3, 0>, - ConvImpl::template process_tile<0, 0, 0, 3, 3, 1>, - ConvImpl::template process_tile<0, 0, 0, 3, 3, 2>, - ConvImpl::template process_tile<0, 0, 0, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 0, 4, 0, 0>, - ConvImpl::template process_tile<0, 0, 0, 4, 0, 1>, - ConvImpl::template process_tile<0, 0, 0, 4, 0, 2>, - ConvImpl::template process_tile<0, 0, 0, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 0, 4, 1, 0>, - ConvImpl::template process_tile<0, 0, 0, 4, 1, 1>, - ConvImpl::template process_tile<0, 0, 0, 4, 1, 2>, - ConvImpl::template process_tile<0, 0, 0, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 0, 4, 2, 0>, - ConvImpl::template process_tile<0, 0, 0, 4, 2, 1>, - ConvImpl::template process_tile<0, 0, 0, 4, 2, 2>, - ConvImpl::template process_tile<0, 0, 0, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 0, 4, 3, 0>, - ConvImpl::template process_tile<0, 0, 0, 4, 3, 1>, - ConvImpl::template process_tile<0, 0, 0, 4, 3, 2>, - ConvImpl::template process_tile<0, 0, 0, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 1, 0, 0, 0>, - ConvImpl::template process_tile<0, 0, 1, 0, 0, 1>, - ConvImpl::template process_tile<0, 0, 1, 0, 0, 2>, - ConvImpl::template process_tile<0, 0, 1, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 1, 0, 1, 0>, - ConvImpl::template process_tile<0, 0, 1, 0, 1, 1>, - ConvImpl::template process_tile<0, 0, 1, 0, 1, 2>, - ConvImpl::template process_tile<0, 0, 1, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 1, 0, 2, 0>, - ConvImpl::template process_tile<0, 0, 1, 0, 2, 1>, - ConvImpl::template process_tile<0, 0, 1, 0, 2, 2>, - ConvImpl::template process_tile<0, 0, 1, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 1, 0, 3, 0>, - ConvImpl::template process_tile<0, 0, 1, 0, 3, 1>, - ConvImpl::template process_tile<0, 0, 1, 0, 3, 2>, - ConvImpl::template process_tile<0, 0, 1, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 1, 1, 0, 0>, - ConvImpl::template process_tile<0, 0, 1, 1, 0, 1>, - ConvImpl::template process_tile<0, 0, 1, 1, 0, 2>, - ConvImpl::template process_tile<0, 0, 1, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 1, 1, 1, 0>, - ConvImpl::template process_tile<0, 0, 1, 1, 1, 1>, - ConvImpl::template process_tile<0, 0, 1, 1, 1, 2>, - ConvImpl::template process_tile<0, 0, 1, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 1, 1, 2, 0>, - ConvImpl::template process_tile<0, 0, 1, 1, 2, 1>, - ConvImpl::template process_tile<0, 0, 1, 1, 2, 2>, - ConvImpl::template process_tile<0, 0, 1, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 1, 1, 3, 0>, - ConvImpl::template process_tile<0, 0, 1, 1, 3, 1>, - ConvImpl::template process_tile<0, 0, 1, 1, 3, 2>, - ConvImpl::template process_tile<0, 0, 1, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 1, 2, 0, 0>, - ConvImpl::template process_tile<0, 0, 1, 2, 0, 1>, - ConvImpl::template process_tile<0, 0, 1, 2, 0, 2>, - ConvImpl::template process_tile<0, 0, 1, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 1, 2, 1, 0>, - ConvImpl::template process_tile<0, 0, 1, 2, 1, 1>, - ConvImpl::template process_tile<0, 0, 1, 2, 1, 2>, - ConvImpl::template process_tile<0, 0, 1, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 1, 2, 2, 0>, - ConvImpl::template process_tile<0, 0, 1, 2, 2, 1>, - ConvImpl::template process_tile<0, 0, 1, 2, 2, 2>, - ConvImpl::template process_tile<0, 0, 1, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 1, 2, 3, 0>, - ConvImpl::template process_tile<0, 0, 1, 2, 3, 1>, - ConvImpl::template process_tile<0, 0, 1, 2, 3, 2>, - ConvImpl::template process_tile<0, 0, 1, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 1, 3, 0, 0>, - ConvImpl::template process_tile<0, 0, 1, 3, 0, 1>, - ConvImpl::template process_tile<0, 0, 1, 3, 0, 2>, - ConvImpl::template process_tile<0, 0, 1, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 1, 3, 1, 0>, - ConvImpl::template process_tile<0, 0, 1, 3, 1, 1>, - ConvImpl::template process_tile<0, 0, 1, 3, 1, 2>, - ConvImpl::template process_tile<0, 0, 1, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 1, 3, 2, 0>, - ConvImpl::template process_tile<0, 0, 1, 3, 2, 1>, - ConvImpl::template process_tile<0, 0, 1, 3, 2, 2>, - ConvImpl::template process_tile<0, 0, 1, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 1, 3, 3, 0>, - ConvImpl::template process_tile<0, 0, 1, 3, 3, 1>, - ConvImpl::template process_tile<0, 0, 1, 3, 3, 2>, - ConvImpl::template process_tile<0, 0, 1, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 1, 4, 0, 0>, - ConvImpl::template process_tile<0, 0, 1, 4, 0, 1>, - ConvImpl::template process_tile<0, 0, 1, 4, 0, 2>, - ConvImpl::template process_tile<0, 0, 1, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 1, 4, 1, 0>, - ConvImpl::template process_tile<0, 0, 1, 4, 1, 1>, - ConvImpl::template process_tile<0, 0, 1, 4, 1, 2>, - ConvImpl::template process_tile<0, 0, 1, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 1, 4, 2, 0>, - ConvImpl::template process_tile<0, 0, 1, 4, 2, 1>, - ConvImpl::template process_tile<0, 0, 1, 4, 2, 2>, - ConvImpl::template process_tile<0, 0, 1, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 1, 4, 3, 0>, - ConvImpl::template process_tile<0, 0, 1, 4, 3, 1>, - ConvImpl::template process_tile<0, 0, 1, 4, 3, 2>, - ConvImpl::template process_tile<0, 0, 1, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 2, 0, 0, 0>, - ConvImpl::template process_tile<0, 0, 2, 0, 0, 1>, - ConvImpl::template process_tile<0, 0, 2, 0, 0, 2>, - ConvImpl::template process_tile<0, 0, 2, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 2, 0, 1, 0>, - ConvImpl::template process_tile<0, 0, 2, 0, 1, 1>, - ConvImpl::template process_tile<0, 0, 2, 0, 1, 2>, - ConvImpl::template process_tile<0, 0, 2, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 2, 0, 2, 0>, - ConvImpl::template process_tile<0, 0, 2, 0, 2, 1>, - ConvImpl::template process_tile<0, 0, 2, 0, 2, 2>, - ConvImpl::template process_tile<0, 0, 2, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 2, 0, 3, 0>, - ConvImpl::template process_tile<0, 0, 2, 0, 3, 1>, - ConvImpl::template process_tile<0, 0, 2, 0, 3, 2>, - ConvImpl::template process_tile<0, 0, 2, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 2, 1, 0, 0>, - ConvImpl::template process_tile<0, 0, 2, 1, 0, 1>, - ConvImpl::template process_tile<0, 0, 2, 1, 0, 2>, - ConvImpl::template process_tile<0, 0, 2, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 2, 1, 1, 0>, - ConvImpl::template process_tile<0, 0, 2, 1, 1, 1>, - ConvImpl::template process_tile<0, 0, 2, 1, 1, 2>, - ConvImpl::template process_tile<0, 0, 2, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 2, 1, 2, 0>, - ConvImpl::template process_tile<0, 0, 2, 1, 2, 1>, - ConvImpl::template process_tile<0, 0, 2, 1, 2, 2>, - ConvImpl::template process_tile<0, 0, 2, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 2, 1, 3, 0>, - ConvImpl::template process_tile<0, 0, 2, 1, 3, 1>, - ConvImpl::template process_tile<0, 0, 2, 1, 3, 2>, - ConvImpl::template process_tile<0, 0, 2, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 2, 2, 0, 0>, - ConvImpl::template process_tile<0, 0, 2, 2, 0, 1>, - ConvImpl::template process_tile<0, 0, 2, 2, 0, 2>, - ConvImpl::template process_tile<0, 0, 2, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 2, 2, 1, 0>, - ConvImpl::template process_tile<0, 0, 2, 2, 1, 1>, - ConvImpl::template process_tile<0, 0, 2, 2, 1, 2>, - ConvImpl::template process_tile<0, 0, 2, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 2, 2, 2, 0>, - ConvImpl::template process_tile<0, 0, 2, 2, 2, 1>, - ConvImpl::template process_tile<0, 0, 2, 2, 2, 2>, - ConvImpl::template process_tile<0, 0, 2, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 2, 2, 3, 0>, - ConvImpl::template process_tile<0, 0, 2, 2, 3, 1>, - ConvImpl::template process_tile<0, 0, 2, 2, 3, 2>, - ConvImpl::template process_tile<0, 0, 2, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 2, 3, 0, 0>, - ConvImpl::template process_tile<0, 0, 2, 3, 0, 1>, - ConvImpl::template process_tile<0, 0, 2, 3, 0, 2>, - ConvImpl::template process_tile<0, 0, 2, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 2, 3, 1, 0>, - ConvImpl::template process_tile<0, 0, 2, 3, 1, 1>, - ConvImpl::template process_tile<0, 0, 2, 3, 1, 2>, - ConvImpl::template process_tile<0, 0, 2, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 2, 3, 2, 0>, - ConvImpl::template process_tile<0, 0, 2, 3, 2, 1>, - ConvImpl::template process_tile<0, 0, 2, 3, 2, 2>, - ConvImpl::template process_tile<0, 0, 2, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 2, 3, 3, 0>, - ConvImpl::template process_tile<0, 0, 2, 3, 3, 1>, - ConvImpl::template process_tile<0, 0, 2, 3, 3, 2>, - ConvImpl::template process_tile<0, 0, 2, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 2, 4, 0, 0>, - ConvImpl::template process_tile<0, 0, 2, 4, 0, 1>, - ConvImpl::template process_tile<0, 0, 2, 4, 0, 2>, - ConvImpl::template process_tile<0, 0, 2, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 2, 4, 1, 0>, - ConvImpl::template process_tile<0, 0, 2, 4, 1, 1>, - ConvImpl::template process_tile<0, 0, 2, 4, 1, 2>, - ConvImpl::template process_tile<0, 0, 2, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 2, 4, 2, 0>, - ConvImpl::template process_tile<0, 0, 2, 4, 2, 1>, - ConvImpl::template process_tile<0, 0, 2, 4, 2, 2>, - ConvImpl::template process_tile<0, 0, 2, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 2, 4, 3, 0>, - ConvImpl::template process_tile<0, 0, 2, 4, 3, 1>, - ConvImpl::template process_tile<0, 0, 2, 4, 3, 2>, - ConvImpl::template process_tile<0, 0, 2, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 3, 0, 0, 0>, - ConvImpl::template process_tile<0, 0, 3, 0, 0, 1>, - ConvImpl::template process_tile<0, 0, 3, 0, 0, 2>, - ConvImpl::template process_tile<0, 0, 3, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 3, 0, 1, 0>, - ConvImpl::template process_tile<0, 0, 3, 0, 1, 1>, - ConvImpl::template process_tile<0, 0, 3, 0, 1, 2>, - ConvImpl::template process_tile<0, 0, 3, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 3, 0, 2, 0>, - ConvImpl::template process_tile<0, 0, 3, 0, 2, 1>, - ConvImpl::template process_tile<0, 0, 3, 0, 2, 2>, - ConvImpl::template process_tile<0, 0, 3, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 3, 0, 3, 0>, - ConvImpl::template process_tile<0, 0, 3, 0, 3, 1>, - ConvImpl::template process_tile<0, 0, 3, 0, 3, 2>, - ConvImpl::template process_tile<0, 0, 3, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 3, 1, 0, 0>, - ConvImpl::template process_tile<0, 0, 3, 1, 0, 1>, - ConvImpl::template process_tile<0, 0, 3, 1, 0, 2>, - ConvImpl::template process_tile<0, 0, 3, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 3, 1, 1, 0>, - ConvImpl::template process_tile<0, 0, 3, 1, 1, 1>, - ConvImpl::template process_tile<0, 0, 3, 1, 1, 2>, - ConvImpl::template process_tile<0, 0, 3, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 3, 1, 2, 0>, - ConvImpl::template process_tile<0, 0, 3, 1, 2, 1>, - ConvImpl::template process_tile<0, 0, 3, 1, 2, 2>, - ConvImpl::template process_tile<0, 0, 3, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 3, 1, 3, 0>, - ConvImpl::template process_tile<0, 0, 3, 1, 3, 1>, - ConvImpl::template process_tile<0, 0, 3, 1, 3, 2>, - ConvImpl::template process_tile<0, 0, 3, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 3, 2, 0, 0>, - ConvImpl::template process_tile<0, 0, 3, 2, 0, 1>, - ConvImpl::template process_tile<0, 0, 3, 2, 0, 2>, - ConvImpl::template process_tile<0, 0, 3, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 3, 2, 1, 0>, - ConvImpl::template process_tile<0, 0, 3, 2, 1, 1>, - ConvImpl::template process_tile<0, 0, 3, 2, 1, 2>, - ConvImpl::template process_tile<0, 0, 3, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 3, 2, 2, 0>, - ConvImpl::template process_tile<0, 0, 3, 2, 2, 1>, - ConvImpl::template process_tile<0, 0, 3, 2, 2, 2>, - ConvImpl::template process_tile<0, 0, 3, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 3, 2, 3, 0>, - ConvImpl::template process_tile<0, 0, 3, 2, 3, 1>, - ConvImpl::template process_tile<0, 0, 3, 2, 3, 2>, - ConvImpl::template process_tile<0, 0, 3, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 3, 3, 0, 0>, - ConvImpl::template process_tile<0, 0, 3, 3, 0, 1>, - ConvImpl::template process_tile<0, 0, 3, 3, 0, 2>, - ConvImpl::template process_tile<0, 0, 3, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 3, 3, 1, 0>, - ConvImpl::template process_tile<0, 0, 3, 3, 1, 1>, - ConvImpl::template process_tile<0, 0, 3, 3, 1, 2>, - ConvImpl::template process_tile<0, 0, 3, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 3, 3, 2, 0>, - ConvImpl::template process_tile<0, 0, 3, 3, 2, 1>, - ConvImpl::template process_tile<0, 0, 3, 3, 2, 2>, - ConvImpl::template process_tile<0, 0, 3, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 3, 3, 3, 0>, - ConvImpl::template process_tile<0, 0, 3, 3, 3, 1>, - ConvImpl::template process_tile<0, 0, 3, 3, 3, 2>, - ConvImpl::template process_tile<0, 0, 3, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 3, 4, 0, 0>, - ConvImpl::template process_tile<0, 0, 3, 4, 0, 1>, - ConvImpl::template process_tile<0, 0, 3, 4, 0, 2>, - ConvImpl::template process_tile<0, 0, 3, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 3, 4, 1, 0>, - ConvImpl::template process_tile<0, 0, 3, 4, 1, 1>, - ConvImpl::template process_tile<0, 0, 3, 4, 1, 2>, - ConvImpl::template process_tile<0, 0, 3, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 3, 4, 2, 0>, - ConvImpl::template process_tile<0, 0, 3, 4, 2, 1>, - ConvImpl::template process_tile<0, 0, 3, 4, 2, 2>, - ConvImpl::template process_tile<0, 0, 3, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 3, 4, 3, 0>, - ConvImpl::template process_tile<0, 0, 3, 4, 3, 1>, - ConvImpl::template process_tile<0, 0, 3, 4, 3, 2>, - ConvImpl::template process_tile<0, 0, 3, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 4, 0, 0, 0>, - ConvImpl::template process_tile<0, 0, 4, 0, 0, 1>, - ConvImpl::template process_tile<0, 0, 4, 0, 0, 2>, - ConvImpl::template process_tile<0, 0, 4, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 4, 0, 1, 0>, - ConvImpl::template process_tile<0, 0, 4, 0, 1, 1>, - ConvImpl::template process_tile<0, 0, 4, 0, 1, 2>, - ConvImpl::template process_tile<0, 0, 4, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 4, 0, 2, 0>, - ConvImpl::template process_tile<0, 0, 4, 0, 2, 1>, - ConvImpl::template process_tile<0, 0, 4, 0, 2, 2>, - ConvImpl::template process_tile<0, 0, 4, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 4, 0, 3, 0>, - ConvImpl::template process_tile<0, 0, 4, 0, 3, 1>, - ConvImpl::template process_tile<0, 0, 4, 0, 3, 2>, - ConvImpl::template process_tile<0, 0, 4, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 4, 1, 0, 0>, - ConvImpl::template process_tile<0, 0, 4, 1, 0, 1>, - ConvImpl::template process_tile<0, 0, 4, 1, 0, 2>, - ConvImpl::template process_tile<0, 0, 4, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 4, 1, 1, 0>, - ConvImpl::template process_tile<0, 0, 4, 1, 1, 1>, - ConvImpl::template process_tile<0, 0, 4, 1, 1, 2>, - ConvImpl::template process_tile<0, 0, 4, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 4, 1, 2, 0>, - ConvImpl::template process_tile<0, 0, 4, 1, 2, 1>, - ConvImpl::template process_tile<0, 0, 4, 1, 2, 2>, - ConvImpl::template process_tile<0, 0, 4, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 4, 1, 3, 0>, - ConvImpl::template process_tile<0, 0, 4, 1, 3, 1>, - ConvImpl::template process_tile<0, 0, 4, 1, 3, 2>, - ConvImpl::template process_tile<0, 0, 4, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 4, 2, 0, 0>, - ConvImpl::template process_tile<0, 0, 4, 2, 0, 1>, - ConvImpl::template process_tile<0, 0, 4, 2, 0, 2>, - ConvImpl::template process_tile<0, 0, 4, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 4, 2, 1, 0>, - ConvImpl::template process_tile<0, 0, 4, 2, 1, 1>, - ConvImpl::template process_tile<0, 0, 4, 2, 1, 2>, - ConvImpl::template process_tile<0, 0, 4, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 4, 2, 2, 0>, - ConvImpl::template process_tile<0, 0, 4, 2, 2, 1>, - ConvImpl::template process_tile<0, 0, 4, 2, 2, 2>, - ConvImpl::template process_tile<0, 0, 4, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 4, 2, 3, 0>, - ConvImpl::template process_tile<0, 0, 4, 2, 3, 1>, - ConvImpl::template process_tile<0, 0, 4, 2, 3, 2>, - ConvImpl::template process_tile<0, 0, 4, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 4, 3, 0, 0>, - ConvImpl::template process_tile<0, 0, 4, 3, 0, 1>, - ConvImpl::template process_tile<0, 0, 4, 3, 0, 2>, - ConvImpl::template process_tile<0, 0, 4, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 4, 3, 1, 0>, - ConvImpl::template process_tile<0, 0, 4, 3, 1, 1>, - ConvImpl::template process_tile<0, 0, 4, 3, 1, 2>, - ConvImpl::template process_tile<0, 0, 4, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 4, 3, 2, 0>, - ConvImpl::template process_tile<0, 0, 4, 3, 2, 1>, - ConvImpl::template process_tile<0, 0, 4, 3, 2, 2>, - ConvImpl::template process_tile<0, 0, 4, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 4, 3, 3, 0>, - ConvImpl::template process_tile<0, 0, 4, 3, 3, 1>, - ConvImpl::template process_tile<0, 0, 4, 3, 3, 2>, - ConvImpl::template process_tile<0, 0, 4, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 0, 4, 4, 0, 0>, - ConvImpl::template process_tile<0, 0, 4, 4, 0, 1>, - ConvImpl::template process_tile<0, 0, 4, 4, 0, 2>, - ConvImpl::template process_tile<0, 0, 4, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 0, 4, 4, 1, 0>, - ConvImpl::template process_tile<0, 0, 4, 4, 1, 1>, - ConvImpl::template process_tile<0, 0, 4, 4, 1, 2>, - ConvImpl::template process_tile<0, 0, 4, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 0, 4, 4, 2, 0>, - ConvImpl::template process_tile<0, 0, 4, 4, 2, 1>, - ConvImpl::template process_tile<0, 0, 4, 4, 2, 2>, - ConvImpl::template process_tile<0, 0, 4, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 0, 4, 4, 3, 0>, - ConvImpl::template process_tile<0, 0, 4, 4, 3, 1>, - ConvImpl::template process_tile<0, 0, 4, 4, 3, 2>, - ConvImpl::template process_tile<0, 0, 4, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 4 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 0, 0, 0, 0>, - ConvImpl::template process_tile<0, 1, 0, 0, 0, 1>, - ConvImpl::template process_tile<0, 1, 0, 0, 0, 2>, - ConvImpl::template process_tile<0, 1, 0, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 0, 0, 1, 0>, - ConvImpl::template process_tile<0, 1, 0, 0, 1, 1>, - ConvImpl::template process_tile<0, 1, 0, 0, 1, 2>, - ConvImpl::template process_tile<0, 1, 0, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 0, 0, 2, 0>, - ConvImpl::template process_tile<0, 1, 0, 0, 2, 1>, - ConvImpl::template process_tile<0, 1, 0, 0, 2, 2>, - ConvImpl::template process_tile<0, 1, 0, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 0, 0, 3, 0>, - ConvImpl::template process_tile<0, 1, 0, 0, 3, 1>, - ConvImpl::template process_tile<0, 1, 0, 0, 3, 2>, - ConvImpl::template process_tile<0, 1, 0, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 0, 1, 0, 0>, - ConvImpl::template process_tile<0, 1, 0, 1, 0, 1>, - ConvImpl::template process_tile<0, 1, 0, 1, 0, 2>, - ConvImpl::template process_tile<0, 1, 0, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 0, 1, 1, 0>, - ConvImpl::template process_tile<0, 1, 0, 1, 1, 1>, - ConvImpl::template process_tile<0, 1, 0, 1, 1, 2>, - ConvImpl::template process_tile<0, 1, 0, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 0, 1, 2, 0>, - ConvImpl::template process_tile<0, 1, 0, 1, 2, 1>, - ConvImpl::template process_tile<0, 1, 0, 1, 2, 2>, - ConvImpl::template process_tile<0, 1, 0, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 0, 1, 3, 0>, - ConvImpl::template process_tile<0, 1, 0, 1, 3, 1>, - ConvImpl::template process_tile<0, 1, 0, 1, 3, 2>, - ConvImpl::template process_tile<0, 1, 0, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 0, 2, 0, 0>, - ConvImpl::template process_tile<0, 1, 0, 2, 0, 1>, - ConvImpl::template process_tile<0, 1, 0, 2, 0, 2>, - ConvImpl::template process_tile<0, 1, 0, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 0, 2, 1, 0>, - ConvImpl::template process_tile<0, 1, 0, 2, 1, 1>, - ConvImpl::template process_tile<0, 1, 0, 2, 1, 2>, - ConvImpl::template process_tile<0, 1, 0, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 0, 2, 2, 0>, - ConvImpl::template process_tile<0, 1, 0, 2, 2, 1>, - ConvImpl::template process_tile<0, 1, 0, 2, 2, 2>, - ConvImpl::template process_tile<0, 1, 0, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 0, 2, 3, 0>, - ConvImpl::template process_tile<0, 1, 0, 2, 3, 1>, - ConvImpl::template process_tile<0, 1, 0, 2, 3, 2>, - ConvImpl::template process_tile<0, 1, 0, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 0, 3, 0, 0>, - ConvImpl::template process_tile<0, 1, 0, 3, 0, 1>, - ConvImpl::template process_tile<0, 1, 0, 3, 0, 2>, - ConvImpl::template process_tile<0, 1, 0, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 0, 3, 1, 0>, - ConvImpl::template process_tile<0, 1, 0, 3, 1, 1>, - ConvImpl::template process_tile<0, 1, 0, 3, 1, 2>, - ConvImpl::template process_tile<0, 1, 0, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 0, 3, 2, 0>, - ConvImpl::template process_tile<0, 1, 0, 3, 2, 1>, - ConvImpl::template process_tile<0, 1, 0, 3, 2, 2>, - ConvImpl::template process_tile<0, 1, 0, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 0, 3, 3, 0>, - ConvImpl::template process_tile<0, 1, 0, 3, 3, 1>, - ConvImpl::template process_tile<0, 1, 0, 3, 3, 2>, - ConvImpl::template process_tile<0, 1, 0, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 0, 4, 0, 0>, - ConvImpl::template process_tile<0, 1, 0, 4, 0, 1>, - ConvImpl::template process_tile<0, 1, 0, 4, 0, 2>, - ConvImpl::template process_tile<0, 1, 0, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 0, 4, 1, 0>, - ConvImpl::template process_tile<0, 1, 0, 4, 1, 1>, - ConvImpl::template process_tile<0, 1, 0, 4, 1, 2>, - ConvImpl::template process_tile<0, 1, 0, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 0, 4, 2, 0>, - ConvImpl::template process_tile<0, 1, 0, 4, 2, 1>, - ConvImpl::template process_tile<0, 1, 0, 4, 2, 2>, - ConvImpl::template process_tile<0, 1, 0, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 0, 4, 3, 0>, - ConvImpl::template process_tile<0, 1, 0, 4, 3, 1>, - ConvImpl::template process_tile<0, 1, 0, 4, 3, 2>, - ConvImpl::template process_tile<0, 1, 0, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 1, 0, 0, 0>, - ConvImpl::template process_tile<0, 1, 1, 0, 0, 1>, - ConvImpl::template process_tile<0, 1, 1, 0, 0, 2>, - ConvImpl::template process_tile<0, 1, 1, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 1, 0, 1, 0>, - ConvImpl::template process_tile<0, 1, 1, 0, 1, 1>, - ConvImpl::template process_tile<0, 1, 1, 0, 1, 2>, - ConvImpl::template process_tile<0, 1, 1, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 1, 0, 2, 0>, - ConvImpl::template process_tile<0, 1, 1, 0, 2, 1>, - ConvImpl::template process_tile<0, 1, 1, 0, 2, 2>, - ConvImpl::template process_tile<0, 1, 1, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 1, 0, 3, 0>, - ConvImpl::template process_tile<0, 1, 1, 0, 3, 1>, - ConvImpl::template process_tile<0, 1, 1, 0, 3, 2>, - ConvImpl::template process_tile<0, 1, 1, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 1, 1, 0, 0>, - ConvImpl::template process_tile<0, 1, 1, 1, 0, 1>, - ConvImpl::template process_tile<0, 1, 1, 1, 0, 2>, - ConvImpl::template process_tile<0, 1, 1, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 1, 1, 1, 0>, - ConvImpl::template process_tile<0, 1, 1, 1, 1, 1>, - ConvImpl::template process_tile<0, 1, 1, 1, 1, 2>, - ConvImpl::template process_tile<0, 1, 1, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 1, 1, 2, 0>, - ConvImpl::template process_tile<0, 1, 1, 1, 2, 1>, - ConvImpl::template process_tile<0, 1, 1, 1, 2, 2>, - ConvImpl::template process_tile<0, 1, 1, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 1, 1, 3, 0>, - ConvImpl::template process_tile<0, 1, 1, 1, 3, 1>, - ConvImpl::template process_tile<0, 1, 1, 1, 3, 2>, - ConvImpl::template process_tile<0, 1, 1, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 1, 2, 0, 0>, - ConvImpl::template process_tile<0, 1, 1, 2, 0, 1>, - ConvImpl::template process_tile<0, 1, 1, 2, 0, 2>, - ConvImpl::template process_tile<0, 1, 1, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 1, 2, 1, 0>, - ConvImpl::template process_tile<0, 1, 1, 2, 1, 1>, - ConvImpl::template process_tile<0, 1, 1, 2, 1, 2>, - ConvImpl::template process_tile<0, 1, 1, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 1, 2, 2, 0>, - ConvImpl::template process_tile<0, 1, 1, 2, 2, 1>, - ConvImpl::template process_tile<0, 1, 1, 2, 2, 2>, - ConvImpl::template process_tile<0, 1, 1, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 1, 2, 3, 0>, - ConvImpl::template process_tile<0, 1, 1, 2, 3, 1>, - ConvImpl::template process_tile<0, 1, 1, 2, 3, 2>, - ConvImpl::template process_tile<0, 1, 1, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 1, 3, 0, 0>, - ConvImpl::template process_tile<0, 1, 1, 3, 0, 1>, - ConvImpl::template process_tile<0, 1, 1, 3, 0, 2>, - ConvImpl::template process_tile<0, 1, 1, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 1, 3, 1, 0>, - ConvImpl::template process_tile<0, 1, 1, 3, 1, 1>, - ConvImpl::template process_tile<0, 1, 1, 3, 1, 2>, - ConvImpl::template process_tile<0, 1, 1, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 1, 3, 2, 0>, - ConvImpl::template process_tile<0, 1, 1, 3, 2, 1>, - ConvImpl::template process_tile<0, 1, 1, 3, 2, 2>, - ConvImpl::template process_tile<0, 1, 1, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 1, 3, 3, 0>, - ConvImpl::template process_tile<0, 1, 1, 3, 3, 1>, - ConvImpl::template process_tile<0, 1, 1, 3, 3, 2>, - ConvImpl::template process_tile<0, 1, 1, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 1, 4, 0, 0>, - ConvImpl::template process_tile<0, 1, 1, 4, 0, 1>, - ConvImpl::template process_tile<0, 1, 1, 4, 0, 2>, - ConvImpl::template process_tile<0, 1, 1, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 1, 4, 1, 0>, - ConvImpl::template process_tile<0, 1, 1, 4, 1, 1>, - ConvImpl::template process_tile<0, 1, 1, 4, 1, 2>, - ConvImpl::template process_tile<0, 1, 1, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 1, 4, 2, 0>, - ConvImpl::template process_tile<0, 1, 1, 4, 2, 1>, - ConvImpl::template process_tile<0, 1, 1, 4, 2, 2>, - ConvImpl::template process_tile<0, 1, 1, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 1, 4, 3, 0>, - ConvImpl::template process_tile<0, 1, 1, 4, 3, 1>, - ConvImpl::template process_tile<0, 1, 1, 4, 3, 2>, - ConvImpl::template process_tile<0, 1, 1, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 2, 0, 0, 0>, - ConvImpl::template process_tile<0, 1, 2, 0, 0, 1>, - ConvImpl::template process_tile<0, 1, 2, 0, 0, 2>, - ConvImpl::template process_tile<0, 1, 2, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 2, 0, 1, 0>, - ConvImpl::template process_tile<0, 1, 2, 0, 1, 1>, - ConvImpl::template process_tile<0, 1, 2, 0, 1, 2>, - ConvImpl::template process_tile<0, 1, 2, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 2, 0, 2, 0>, - ConvImpl::template process_tile<0, 1, 2, 0, 2, 1>, - ConvImpl::template process_tile<0, 1, 2, 0, 2, 2>, - ConvImpl::template process_tile<0, 1, 2, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 2, 0, 3, 0>, - ConvImpl::template process_tile<0, 1, 2, 0, 3, 1>, - ConvImpl::template process_tile<0, 1, 2, 0, 3, 2>, - ConvImpl::template process_tile<0, 1, 2, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 2, 1, 0, 0>, - ConvImpl::template process_tile<0, 1, 2, 1, 0, 1>, - ConvImpl::template process_tile<0, 1, 2, 1, 0, 2>, - ConvImpl::template process_tile<0, 1, 2, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 2, 1, 1, 0>, - ConvImpl::template process_tile<0, 1, 2, 1, 1, 1>, - ConvImpl::template process_tile<0, 1, 2, 1, 1, 2>, - ConvImpl::template process_tile<0, 1, 2, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 2, 1, 2, 0>, - ConvImpl::template process_tile<0, 1, 2, 1, 2, 1>, - ConvImpl::template process_tile<0, 1, 2, 1, 2, 2>, - ConvImpl::template process_tile<0, 1, 2, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 2, 1, 3, 0>, - ConvImpl::template process_tile<0, 1, 2, 1, 3, 1>, - ConvImpl::template process_tile<0, 1, 2, 1, 3, 2>, - ConvImpl::template process_tile<0, 1, 2, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 2, 2, 0, 0>, - ConvImpl::template process_tile<0, 1, 2, 2, 0, 1>, - ConvImpl::template process_tile<0, 1, 2, 2, 0, 2>, - ConvImpl::template process_tile<0, 1, 2, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 2, 2, 1, 0>, - ConvImpl::template process_tile<0, 1, 2, 2, 1, 1>, - ConvImpl::template process_tile<0, 1, 2, 2, 1, 2>, - ConvImpl::template process_tile<0, 1, 2, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 2, 2, 2, 0>, - ConvImpl::template process_tile<0, 1, 2, 2, 2, 1>, - ConvImpl::template process_tile<0, 1, 2, 2, 2, 2>, - ConvImpl::template process_tile<0, 1, 2, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 2, 2, 3, 0>, - ConvImpl::template process_tile<0, 1, 2, 2, 3, 1>, - ConvImpl::template process_tile<0, 1, 2, 2, 3, 2>, - ConvImpl::template process_tile<0, 1, 2, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 2, 3, 0, 0>, - ConvImpl::template process_tile<0, 1, 2, 3, 0, 1>, - ConvImpl::template process_tile<0, 1, 2, 3, 0, 2>, - ConvImpl::template process_tile<0, 1, 2, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 2, 3, 1, 0>, - ConvImpl::template process_tile<0, 1, 2, 3, 1, 1>, - ConvImpl::template process_tile<0, 1, 2, 3, 1, 2>, - ConvImpl::template process_tile<0, 1, 2, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 2, 3, 2, 0>, - ConvImpl::template process_tile<0, 1, 2, 3, 2, 1>, - ConvImpl::template process_tile<0, 1, 2, 3, 2, 2>, - ConvImpl::template process_tile<0, 1, 2, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 2, 3, 3, 0>, - ConvImpl::template process_tile<0, 1, 2, 3, 3, 1>, - ConvImpl::template process_tile<0, 1, 2, 3, 3, 2>, - ConvImpl::template process_tile<0, 1, 2, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 2, 4, 0, 0>, - ConvImpl::template process_tile<0, 1, 2, 4, 0, 1>, - ConvImpl::template process_tile<0, 1, 2, 4, 0, 2>, - ConvImpl::template process_tile<0, 1, 2, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 2, 4, 1, 0>, - ConvImpl::template process_tile<0, 1, 2, 4, 1, 1>, - ConvImpl::template process_tile<0, 1, 2, 4, 1, 2>, - ConvImpl::template process_tile<0, 1, 2, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 2, 4, 2, 0>, - ConvImpl::template process_tile<0, 1, 2, 4, 2, 1>, - ConvImpl::template process_tile<0, 1, 2, 4, 2, 2>, - ConvImpl::template process_tile<0, 1, 2, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 2, 4, 3, 0>, - ConvImpl::template process_tile<0, 1, 2, 4, 3, 1>, - ConvImpl::template process_tile<0, 1, 2, 4, 3, 2>, - ConvImpl::template process_tile<0, 1, 2, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 3, 0, 0, 0>, - ConvImpl::template process_tile<0, 1, 3, 0, 0, 1>, - ConvImpl::template process_tile<0, 1, 3, 0, 0, 2>, - ConvImpl::template process_tile<0, 1, 3, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 3, 0, 1, 0>, - ConvImpl::template process_tile<0, 1, 3, 0, 1, 1>, - ConvImpl::template process_tile<0, 1, 3, 0, 1, 2>, - ConvImpl::template process_tile<0, 1, 3, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 3, 0, 2, 0>, - ConvImpl::template process_tile<0, 1, 3, 0, 2, 1>, - ConvImpl::template process_tile<0, 1, 3, 0, 2, 2>, - ConvImpl::template process_tile<0, 1, 3, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 3, 0, 3, 0>, - ConvImpl::template process_tile<0, 1, 3, 0, 3, 1>, - ConvImpl::template process_tile<0, 1, 3, 0, 3, 2>, - ConvImpl::template process_tile<0, 1, 3, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 3, 1, 0, 0>, - ConvImpl::template process_tile<0, 1, 3, 1, 0, 1>, - ConvImpl::template process_tile<0, 1, 3, 1, 0, 2>, - ConvImpl::template process_tile<0, 1, 3, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 3, 1, 1, 0>, - ConvImpl::template process_tile<0, 1, 3, 1, 1, 1>, - ConvImpl::template process_tile<0, 1, 3, 1, 1, 2>, - ConvImpl::template process_tile<0, 1, 3, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 3, 1, 2, 0>, - ConvImpl::template process_tile<0, 1, 3, 1, 2, 1>, - ConvImpl::template process_tile<0, 1, 3, 1, 2, 2>, - ConvImpl::template process_tile<0, 1, 3, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 3, 1, 3, 0>, - ConvImpl::template process_tile<0, 1, 3, 1, 3, 1>, - ConvImpl::template process_tile<0, 1, 3, 1, 3, 2>, - ConvImpl::template process_tile<0, 1, 3, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 3, 2, 0, 0>, - ConvImpl::template process_tile<0, 1, 3, 2, 0, 1>, - ConvImpl::template process_tile<0, 1, 3, 2, 0, 2>, - ConvImpl::template process_tile<0, 1, 3, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 3, 2, 1, 0>, - ConvImpl::template process_tile<0, 1, 3, 2, 1, 1>, - ConvImpl::template process_tile<0, 1, 3, 2, 1, 2>, - ConvImpl::template process_tile<0, 1, 3, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 3, 2, 2, 0>, - ConvImpl::template process_tile<0, 1, 3, 2, 2, 1>, - ConvImpl::template process_tile<0, 1, 3, 2, 2, 2>, - ConvImpl::template process_tile<0, 1, 3, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 3, 2, 3, 0>, - ConvImpl::template process_tile<0, 1, 3, 2, 3, 1>, - ConvImpl::template process_tile<0, 1, 3, 2, 3, 2>, - ConvImpl::template process_tile<0, 1, 3, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 3, 3, 0, 0>, - ConvImpl::template process_tile<0, 1, 3, 3, 0, 1>, - ConvImpl::template process_tile<0, 1, 3, 3, 0, 2>, - ConvImpl::template process_tile<0, 1, 3, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 3, 3, 1, 0>, - ConvImpl::template process_tile<0, 1, 3, 3, 1, 1>, - ConvImpl::template process_tile<0, 1, 3, 3, 1, 2>, - ConvImpl::template process_tile<0, 1, 3, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 3, 3, 2, 0>, - ConvImpl::template process_tile<0, 1, 3, 3, 2, 1>, - ConvImpl::template process_tile<0, 1, 3, 3, 2, 2>, - ConvImpl::template process_tile<0, 1, 3, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 3, 3, 3, 0>, - ConvImpl::template process_tile<0, 1, 3, 3, 3, 1>, - ConvImpl::template process_tile<0, 1, 3, 3, 3, 2>, - ConvImpl::template process_tile<0, 1, 3, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 3, 4, 0, 0>, - ConvImpl::template process_tile<0, 1, 3, 4, 0, 1>, - ConvImpl::template process_tile<0, 1, 3, 4, 0, 2>, - ConvImpl::template process_tile<0, 1, 3, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 3, 4, 1, 0>, - ConvImpl::template process_tile<0, 1, 3, 4, 1, 1>, - ConvImpl::template process_tile<0, 1, 3, 4, 1, 2>, - ConvImpl::template process_tile<0, 1, 3, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 3, 4, 2, 0>, - ConvImpl::template process_tile<0, 1, 3, 4, 2, 1>, - ConvImpl::template process_tile<0, 1, 3, 4, 2, 2>, - ConvImpl::template process_tile<0, 1, 3, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 3, 4, 3, 0>, - ConvImpl::template process_tile<0, 1, 3, 4, 3, 1>, - ConvImpl::template process_tile<0, 1, 3, 4, 3, 2>, - ConvImpl::template process_tile<0, 1, 3, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 4, 0, 0, 0>, - ConvImpl::template process_tile<0, 1, 4, 0, 0, 1>, - ConvImpl::template process_tile<0, 1, 4, 0, 0, 2>, - ConvImpl::template process_tile<0, 1, 4, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 4, 0, 1, 0>, - ConvImpl::template process_tile<0, 1, 4, 0, 1, 1>, - ConvImpl::template process_tile<0, 1, 4, 0, 1, 2>, - ConvImpl::template process_tile<0, 1, 4, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 4, 0, 2, 0>, - ConvImpl::template process_tile<0, 1, 4, 0, 2, 1>, - ConvImpl::template process_tile<0, 1, 4, 0, 2, 2>, - ConvImpl::template process_tile<0, 1, 4, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 4, 0, 3, 0>, - ConvImpl::template process_tile<0, 1, 4, 0, 3, 1>, - ConvImpl::template process_tile<0, 1, 4, 0, 3, 2>, - ConvImpl::template process_tile<0, 1, 4, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 4, 1, 0, 0>, - ConvImpl::template process_tile<0, 1, 4, 1, 0, 1>, - ConvImpl::template process_tile<0, 1, 4, 1, 0, 2>, - ConvImpl::template process_tile<0, 1, 4, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 4, 1, 1, 0>, - ConvImpl::template process_tile<0, 1, 4, 1, 1, 1>, - ConvImpl::template process_tile<0, 1, 4, 1, 1, 2>, - ConvImpl::template process_tile<0, 1, 4, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 4, 1, 2, 0>, - ConvImpl::template process_tile<0, 1, 4, 1, 2, 1>, - ConvImpl::template process_tile<0, 1, 4, 1, 2, 2>, - ConvImpl::template process_tile<0, 1, 4, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 4, 1, 3, 0>, - ConvImpl::template process_tile<0, 1, 4, 1, 3, 1>, - ConvImpl::template process_tile<0, 1, 4, 1, 3, 2>, - ConvImpl::template process_tile<0, 1, 4, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 4, 2, 0, 0>, - ConvImpl::template process_tile<0, 1, 4, 2, 0, 1>, - ConvImpl::template process_tile<0, 1, 4, 2, 0, 2>, - ConvImpl::template process_tile<0, 1, 4, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 4, 2, 1, 0>, - ConvImpl::template process_tile<0, 1, 4, 2, 1, 1>, - ConvImpl::template process_tile<0, 1, 4, 2, 1, 2>, - ConvImpl::template process_tile<0, 1, 4, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 4, 2, 2, 0>, - ConvImpl::template process_tile<0, 1, 4, 2, 2, 1>, - ConvImpl::template process_tile<0, 1, 4, 2, 2, 2>, - ConvImpl::template process_tile<0, 1, 4, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 4, 2, 3, 0>, - ConvImpl::template process_tile<0, 1, 4, 2, 3, 1>, - ConvImpl::template process_tile<0, 1, 4, 2, 3, 2>, - ConvImpl::template process_tile<0, 1, 4, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 4, 3, 0, 0>, - ConvImpl::template process_tile<0, 1, 4, 3, 0, 1>, - ConvImpl::template process_tile<0, 1, 4, 3, 0, 2>, - ConvImpl::template process_tile<0, 1, 4, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 4, 3, 1, 0>, - ConvImpl::template process_tile<0, 1, 4, 3, 1, 1>, - ConvImpl::template process_tile<0, 1, 4, 3, 1, 2>, - ConvImpl::template process_tile<0, 1, 4, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 4, 3, 2, 0>, - ConvImpl::template process_tile<0, 1, 4, 3, 2, 1>, - ConvImpl::template process_tile<0, 1, 4, 3, 2, 2>, - ConvImpl::template process_tile<0, 1, 4, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 4, 3, 3, 0>, - ConvImpl::template process_tile<0, 1, 4, 3, 3, 1>, - ConvImpl::template process_tile<0, 1, 4, 3, 3, 2>, - ConvImpl::template process_tile<0, 1, 4, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<0, 1, 4, 4, 0, 0>, - ConvImpl::template process_tile<0, 1, 4, 4, 0, 1>, - ConvImpl::template process_tile<0, 1, 4, 4, 0, 2>, - ConvImpl::template process_tile<0, 1, 4, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<0, 1, 4, 4, 1, 0>, - ConvImpl::template process_tile<0, 1, 4, 4, 1, 1>, - ConvImpl::template process_tile<0, 1, 4, 4, 1, 2>, - ConvImpl::template process_tile<0, 1, 4, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<0, 1, 4, 4, 2, 0>, - ConvImpl::template process_tile<0, 1, 4, 4, 2, 1>, - ConvImpl::template process_tile<0, 1, 4, 4, 2, 2>, - ConvImpl::template process_tile<0, 1, 4, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<0, 1, 4, 4, 3, 0>, - ConvImpl::template process_tile<0, 1, 4, 4, 3, 1>, - ConvImpl::template process_tile<0, 1, 4, 4, 3, 2>, - ConvImpl::template process_tile<0, 1, 4, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 4 - }, // Input pad left = 1 - }, // Input pad top = 0 - { // Input pad top = 1 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 0, 0, 0, 0>, - ConvImpl::template process_tile<1, 0, 0, 0, 0, 1>, - ConvImpl::template process_tile<1, 0, 0, 0, 0, 2>, - ConvImpl::template process_tile<1, 0, 0, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 0, 0, 1, 0>, - ConvImpl::template process_tile<1, 0, 0, 0, 1, 1>, - ConvImpl::template process_tile<1, 0, 0, 0, 1, 2>, - ConvImpl::template process_tile<1, 0, 0, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 0, 0, 2, 0>, - ConvImpl::template process_tile<1, 0, 0, 0, 2, 1>, - ConvImpl::template process_tile<1, 0, 0, 0, 2, 2>, - ConvImpl::template process_tile<1, 0, 0, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 0, 0, 3, 0>, - ConvImpl::template process_tile<1, 0, 0, 0, 3, 1>, - ConvImpl::template process_tile<1, 0, 0, 0, 3, 2>, - ConvImpl::template process_tile<1, 0, 0, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 0, 1, 0, 0>, - ConvImpl::template process_tile<1, 0, 0, 1, 0, 1>, - ConvImpl::template process_tile<1, 0, 0, 1, 0, 2>, - ConvImpl::template process_tile<1, 0, 0, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 0, 1, 1, 0>, - ConvImpl::template process_tile<1, 0, 0, 1, 1, 1>, - ConvImpl::template process_tile<1, 0, 0, 1, 1, 2>, - ConvImpl::template process_tile<1, 0, 0, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 0, 1, 2, 0>, - ConvImpl::template process_tile<1, 0, 0, 1, 2, 1>, - ConvImpl::template process_tile<1, 0, 0, 1, 2, 2>, - ConvImpl::template process_tile<1, 0, 0, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 0, 1, 3, 0>, - ConvImpl::template process_tile<1, 0, 0, 1, 3, 1>, - ConvImpl::template process_tile<1, 0, 0, 1, 3, 2>, - ConvImpl::template process_tile<1, 0, 0, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 0, 2, 0, 0>, - ConvImpl::template process_tile<1, 0, 0, 2, 0, 1>, - ConvImpl::template process_tile<1, 0, 0, 2, 0, 2>, - ConvImpl::template process_tile<1, 0, 0, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 0, 2, 1, 0>, - ConvImpl::template process_tile<1, 0, 0, 2, 1, 1>, - ConvImpl::template process_tile<1, 0, 0, 2, 1, 2>, - ConvImpl::template process_tile<1, 0, 0, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 0, 2, 2, 0>, - ConvImpl::template process_tile<1, 0, 0, 2, 2, 1>, - ConvImpl::template process_tile<1, 0, 0, 2, 2, 2>, - ConvImpl::template process_tile<1, 0, 0, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 0, 2, 3, 0>, - ConvImpl::template process_tile<1, 0, 0, 2, 3, 1>, - ConvImpl::template process_tile<1, 0, 0, 2, 3, 2>, - ConvImpl::template process_tile<1, 0, 0, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 0, 3, 0, 0>, - ConvImpl::template process_tile<1, 0, 0, 3, 0, 1>, - ConvImpl::template process_tile<1, 0, 0, 3, 0, 2>, - ConvImpl::template process_tile<1, 0, 0, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 0, 3, 1, 0>, - ConvImpl::template process_tile<1, 0, 0, 3, 1, 1>, - ConvImpl::template process_tile<1, 0, 0, 3, 1, 2>, - ConvImpl::template process_tile<1, 0, 0, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 0, 3, 2, 0>, - ConvImpl::template process_tile<1, 0, 0, 3, 2, 1>, - ConvImpl::template process_tile<1, 0, 0, 3, 2, 2>, - ConvImpl::template process_tile<1, 0, 0, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 0, 3, 3, 0>, - ConvImpl::template process_tile<1, 0, 0, 3, 3, 1>, - ConvImpl::template process_tile<1, 0, 0, 3, 3, 2>, - ConvImpl::template process_tile<1, 0, 0, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 0, 4, 0, 0>, - ConvImpl::template process_tile<1, 0, 0, 4, 0, 1>, - ConvImpl::template process_tile<1, 0, 0, 4, 0, 2>, - ConvImpl::template process_tile<1, 0, 0, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 0, 4, 1, 0>, - ConvImpl::template process_tile<1, 0, 0, 4, 1, 1>, - ConvImpl::template process_tile<1, 0, 0, 4, 1, 2>, - ConvImpl::template process_tile<1, 0, 0, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 0, 4, 2, 0>, - ConvImpl::template process_tile<1, 0, 0, 4, 2, 1>, - ConvImpl::template process_tile<1, 0, 0, 4, 2, 2>, - ConvImpl::template process_tile<1, 0, 0, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 0, 4, 3, 0>, - ConvImpl::template process_tile<1, 0, 0, 4, 3, 1>, - ConvImpl::template process_tile<1, 0, 0, 4, 3, 2>, - ConvImpl::template process_tile<1, 0, 0, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 1, 0, 0, 0>, - ConvImpl::template process_tile<1, 0, 1, 0, 0, 1>, - ConvImpl::template process_tile<1, 0, 1, 0, 0, 2>, - ConvImpl::template process_tile<1, 0, 1, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 1, 0, 1, 0>, - ConvImpl::template process_tile<1, 0, 1, 0, 1, 1>, - ConvImpl::template process_tile<1, 0, 1, 0, 1, 2>, - ConvImpl::template process_tile<1, 0, 1, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 1, 0, 2, 0>, - ConvImpl::template process_tile<1, 0, 1, 0, 2, 1>, - ConvImpl::template process_tile<1, 0, 1, 0, 2, 2>, - ConvImpl::template process_tile<1, 0, 1, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 1, 0, 3, 0>, - ConvImpl::template process_tile<1, 0, 1, 0, 3, 1>, - ConvImpl::template process_tile<1, 0, 1, 0, 3, 2>, - ConvImpl::template process_tile<1, 0, 1, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 1, 1, 0, 0>, - ConvImpl::template process_tile<1, 0, 1, 1, 0, 1>, - ConvImpl::template process_tile<1, 0, 1, 1, 0, 2>, - ConvImpl::template process_tile<1, 0, 1, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 1, 1, 1, 0>, - ConvImpl::template process_tile<1, 0, 1, 1, 1, 1>, - ConvImpl::template process_tile<1, 0, 1, 1, 1, 2>, - ConvImpl::template process_tile<1, 0, 1, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 1, 1, 2, 0>, - ConvImpl::template process_tile<1, 0, 1, 1, 2, 1>, - ConvImpl::template process_tile<1, 0, 1, 1, 2, 2>, - ConvImpl::template process_tile<1, 0, 1, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 1, 1, 3, 0>, - ConvImpl::template process_tile<1, 0, 1, 1, 3, 1>, - ConvImpl::template process_tile<1, 0, 1, 1, 3, 2>, - ConvImpl::template process_tile<1, 0, 1, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 1, 2, 0, 0>, - ConvImpl::template process_tile<1, 0, 1, 2, 0, 1>, - ConvImpl::template process_tile<1, 0, 1, 2, 0, 2>, - ConvImpl::template process_tile<1, 0, 1, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 1, 2, 1, 0>, - ConvImpl::template process_tile<1, 0, 1, 2, 1, 1>, - ConvImpl::template process_tile<1, 0, 1, 2, 1, 2>, - ConvImpl::template process_tile<1, 0, 1, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 1, 2, 2, 0>, - ConvImpl::template process_tile<1, 0, 1, 2, 2, 1>, - ConvImpl::template process_tile<1, 0, 1, 2, 2, 2>, - ConvImpl::template process_tile<1, 0, 1, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 1, 2, 3, 0>, - ConvImpl::template process_tile<1, 0, 1, 2, 3, 1>, - ConvImpl::template process_tile<1, 0, 1, 2, 3, 2>, - ConvImpl::template process_tile<1, 0, 1, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 1, 3, 0, 0>, - ConvImpl::template process_tile<1, 0, 1, 3, 0, 1>, - ConvImpl::template process_tile<1, 0, 1, 3, 0, 2>, - ConvImpl::template process_tile<1, 0, 1, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 1, 3, 1, 0>, - ConvImpl::template process_tile<1, 0, 1, 3, 1, 1>, - ConvImpl::template process_tile<1, 0, 1, 3, 1, 2>, - ConvImpl::template process_tile<1, 0, 1, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 1, 3, 2, 0>, - ConvImpl::template process_tile<1, 0, 1, 3, 2, 1>, - ConvImpl::template process_tile<1, 0, 1, 3, 2, 2>, - ConvImpl::template process_tile<1, 0, 1, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 1, 3, 3, 0>, - ConvImpl::template process_tile<1, 0, 1, 3, 3, 1>, - ConvImpl::template process_tile<1, 0, 1, 3, 3, 2>, - ConvImpl::template process_tile<1, 0, 1, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 1, 4, 0, 0>, - ConvImpl::template process_tile<1, 0, 1, 4, 0, 1>, - ConvImpl::template process_tile<1, 0, 1, 4, 0, 2>, - ConvImpl::template process_tile<1, 0, 1, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 1, 4, 1, 0>, - ConvImpl::template process_tile<1, 0, 1, 4, 1, 1>, - ConvImpl::template process_tile<1, 0, 1, 4, 1, 2>, - ConvImpl::template process_tile<1, 0, 1, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 1, 4, 2, 0>, - ConvImpl::template process_tile<1, 0, 1, 4, 2, 1>, - ConvImpl::template process_tile<1, 0, 1, 4, 2, 2>, - ConvImpl::template process_tile<1, 0, 1, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 1, 4, 3, 0>, - ConvImpl::template process_tile<1, 0, 1, 4, 3, 1>, - ConvImpl::template process_tile<1, 0, 1, 4, 3, 2>, - ConvImpl::template process_tile<1, 0, 1, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 2, 0, 0, 0>, - ConvImpl::template process_tile<1, 0, 2, 0, 0, 1>, - ConvImpl::template process_tile<1, 0, 2, 0, 0, 2>, - ConvImpl::template process_tile<1, 0, 2, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 2, 0, 1, 0>, - ConvImpl::template process_tile<1, 0, 2, 0, 1, 1>, - ConvImpl::template process_tile<1, 0, 2, 0, 1, 2>, - ConvImpl::template process_tile<1, 0, 2, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 2, 0, 2, 0>, - ConvImpl::template process_tile<1, 0, 2, 0, 2, 1>, - ConvImpl::template process_tile<1, 0, 2, 0, 2, 2>, - ConvImpl::template process_tile<1, 0, 2, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 2, 0, 3, 0>, - ConvImpl::template process_tile<1, 0, 2, 0, 3, 1>, - ConvImpl::template process_tile<1, 0, 2, 0, 3, 2>, - ConvImpl::template process_tile<1, 0, 2, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 2, 1, 0, 0>, - ConvImpl::template process_tile<1, 0, 2, 1, 0, 1>, - ConvImpl::template process_tile<1, 0, 2, 1, 0, 2>, - ConvImpl::template process_tile<1, 0, 2, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 2, 1, 1, 0>, - ConvImpl::template process_tile<1, 0, 2, 1, 1, 1>, - ConvImpl::template process_tile<1, 0, 2, 1, 1, 2>, - ConvImpl::template process_tile<1, 0, 2, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 2, 1, 2, 0>, - ConvImpl::template process_tile<1, 0, 2, 1, 2, 1>, - ConvImpl::template process_tile<1, 0, 2, 1, 2, 2>, - ConvImpl::template process_tile<1, 0, 2, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 2, 1, 3, 0>, - ConvImpl::template process_tile<1, 0, 2, 1, 3, 1>, - ConvImpl::template process_tile<1, 0, 2, 1, 3, 2>, - ConvImpl::template process_tile<1, 0, 2, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 2, 2, 0, 0>, - ConvImpl::template process_tile<1, 0, 2, 2, 0, 1>, - ConvImpl::template process_tile<1, 0, 2, 2, 0, 2>, - ConvImpl::template process_tile<1, 0, 2, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 2, 2, 1, 0>, - ConvImpl::template process_tile<1, 0, 2, 2, 1, 1>, - ConvImpl::template process_tile<1, 0, 2, 2, 1, 2>, - ConvImpl::template process_tile<1, 0, 2, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 2, 2, 2, 0>, - ConvImpl::template process_tile<1, 0, 2, 2, 2, 1>, - ConvImpl::template process_tile<1, 0, 2, 2, 2, 2>, - ConvImpl::template process_tile<1, 0, 2, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 2, 2, 3, 0>, - ConvImpl::template process_tile<1, 0, 2, 2, 3, 1>, - ConvImpl::template process_tile<1, 0, 2, 2, 3, 2>, - ConvImpl::template process_tile<1, 0, 2, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 2, 3, 0, 0>, - ConvImpl::template process_tile<1, 0, 2, 3, 0, 1>, - ConvImpl::template process_tile<1, 0, 2, 3, 0, 2>, - ConvImpl::template process_tile<1, 0, 2, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 2, 3, 1, 0>, - ConvImpl::template process_tile<1, 0, 2, 3, 1, 1>, - ConvImpl::template process_tile<1, 0, 2, 3, 1, 2>, - ConvImpl::template process_tile<1, 0, 2, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 2, 3, 2, 0>, - ConvImpl::template process_tile<1, 0, 2, 3, 2, 1>, - ConvImpl::template process_tile<1, 0, 2, 3, 2, 2>, - ConvImpl::template process_tile<1, 0, 2, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 2, 3, 3, 0>, - ConvImpl::template process_tile<1, 0, 2, 3, 3, 1>, - ConvImpl::template process_tile<1, 0, 2, 3, 3, 2>, - ConvImpl::template process_tile<1, 0, 2, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 2, 4, 0, 0>, - ConvImpl::template process_tile<1, 0, 2, 4, 0, 1>, - ConvImpl::template process_tile<1, 0, 2, 4, 0, 2>, - ConvImpl::template process_tile<1, 0, 2, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 2, 4, 1, 0>, - ConvImpl::template process_tile<1, 0, 2, 4, 1, 1>, - ConvImpl::template process_tile<1, 0, 2, 4, 1, 2>, - ConvImpl::template process_tile<1, 0, 2, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 2, 4, 2, 0>, - ConvImpl::template process_tile<1, 0, 2, 4, 2, 1>, - ConvImpl::template process_tile<1, 0, 2, 4, 2, 2>, - ConvImpl::template process_tile<1, 0, 2, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 2, 4, 3, 0>, - ConvImpl::template process_tile<1, 0, 2, 4, 3, 1>, - ConvImpl::template process_tile<1, 0, 2, 4, 3, 2>, - ConvImpl::template process_tile<1, 0, 2, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 3, 0, 0, 0>, - ConvImpl::template process_tile<1, 0, 3, 0, 0, 1>, - ConvImpl::template process_tile<1, 0, 3, 0, 0, 2>, - ConvImpl::template process_tile<1, 0, 3, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 3, 0, 1, 0>, - ConvImpl::template process_tile<1, 0, 3, 0, 1, 1>, - ConvImpl::template process_tile<1, 0, 3, 0, 1, 2>, - ConvImpl::template process_tile<1, 0, 3, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 3, 0, 2, 0>, - ConvImpl::template process_tile<1, 0, 3, 0, 2, 1>, - ConvImpl::template process_tile<1, 0, 3, 0, 2, 2>, - ConvImpl::template process_tile<1, 0, 3, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 3, 0, 3, 0>, - ConvImpl::template process_tile<1, 0, 3, 0, 3, 1>, - ConvImpl::template process_tile<1, 0, 3, 0, 3, 2>, - ConvImpl::template process_tile<1, 0, 3, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 3, 1, 0, 0>, - ConvImpl::template process_tile<1, 0, 3, 1, 0, 1>, - ConvImpl::template process_tile<1, 0, 3, 1, 0, 2>, - ConvImpl::template process_tile<1, 0, 3, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 3, 1, 1, 0>, - ConvImpl::template process_tile<1, 0, 3, 1, 1, 1>, - ConvImpl::template process_tile<1, 0, 3, 1, 1, 2>, - ConvImpl::template process_tile<1, 0, 3, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 3, 1, 2, 0>, - ConvImpl::template process_tile<1, 0, 3, 1, 2, 1>, - ConvImpl::template process_tile<1, 0, 3, 1, 2, 2>, - ConvImpl::template process_tile<1, 0, 3, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 3, 1, 3, 0>, - ConvImpl::template process_tile<1, 0, 3, 1, 3, 1>, - ConvImpl::template process_tile<1, 0, 3, 1, 3, 2>, - ConvImpl::template process_tile<1, 0, 3, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 3, 2, 0, 0>, - ConvImpl::template process_tile<1, 0, 3, 2, 0, 1>, - ConvImpl::template process_tile<1, 0, 3, 2, 0, 2>, - ConvImpl::template process_tile<1, 0, 3, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 3, 2, 1, 0>, - ConvImpl::template process_tile<1, 0, 3, 2, 1, 1>, - ConvImpl::template process_tile<1, 0, 3, 2, 1, 2>, - ConvImpl::template process_tile<1, 0, 3, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 3, 2, 2, 0>, - ConvImpl::template process_tile<1, 0, 3, 2, 2, 1>, - ConvImpl::template process_tile<1, 0, 3, 2, 2, 2>, - ConvImpl::template process_tile<1, 0, 3, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 3, 2, 3, 0>, - ConvImpl::template process_tile<1, 0, 3, 2, 3, 1>, - ConvImpl::template process_tile<1, 0, 3, 2, 3, 2>, - ConvImpl::template process_tile<1, 0, 3, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 3, 3, 0, 0>, - ConvImpl::template process_tile<1, 0, 3, 3, 0, 1>, - ConvImpl::template process_tile<1, 0, 3, 3, 0, 2>, - ConvImpl::template process_tile<1, 0, 3, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 3, 3, 1, 0>, - ConvImpl::template process_tile<1, 0, 3, 3, 1, 1>, - ConvImpl::template process_tile<1, 0, 3, 3, 1, 2>, - ConvImpl::template process_tile<1, 0, 3, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 3, 3, 2, 0>, - ConvImpl::template process_tile<1, 0, 3, 3, 2, 1>, - ConvImpl::template process_tile<1, 0, 3, 3, 2, 2>, - ConvImpl::template process_tile<1, 0, 3, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 3, 3, 3, 0>, - ConvImpl::template process_tile<1, 0, 3, 3, 3, 1>, - ConvImpl::template process_tile<1, 0, 3, 3, 3, 2>, - ConvImpl::template process_tile<1, 0, 3, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 3, 4, 0, 0>, - ConvImpl::template process_tile<1, 0, 3, 4, 0, 1>, - ConvImpl::template process_tile<1, 0, 3, 4, 0, 2>, - ConvImpl::template process_tile<1, 0, 3, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 3, 4, 1, 0>, - ConvImpl::template process_tile<1, 0, 3, 4, 1, 1>, - ConvImpl::template process_tile<1, 0, 3, 4, 1, 2>, - ConvImpl::template process_tile<1, 0, 3, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 3, 4, 2, 0>, - ConvImpl::template process_tile<1, 0, 3, 4, 2, 1>, - ConvImpl::template process_tile<1, 0, 3, 4, 2, 2>, - ConvImpl::template process_tile<1, 0, 3, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 3, 4, 3, 0>, - ConvImpl::template process_tile<1, 0, 3, 4, 3, 1>, - ConvImpl::template process_tile<1, 0, 3, 4, 3, 2>, - ConvImpl::template process_tile<1, 0, 3, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 4, 0, 0, 0>, - ConvImpl::template process_tile<1, 0, 4, 0, 0, 1>, - ConvImpl::template process_tile<1, 0, 4, 0, 0, 2>, - ConvImpl::template process_tile<1, 0, 4, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 4, 0, 1, 0>, - ConvImpl::template process_tile<1, 0, 4, 0, 1, 1>, - ConvImpl::template process_tile<1, 0, 4, 0, 1, 2>, - ConvImpl::template process_tile<1, 0, 4, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 4, 0, 2, 0>, - ConvImpl::template process_tile<1, 0, 4, 0, 2, 1>, - ConvImpl::template process_tile<1, 0, 4, 0, 2, 2>, - ConvImpl::template process_tile<1, 0, 4, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 4, 0, 3, 0>, - ConvImpl::template process_tile<1, 0, 4, 0, 3, 1>, - ConvImpl::template process_tile<1, 0, 4, 0, 3, 2>, - ConvImpl::template process_tile<1, 0, 4, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 4, 1, 0, 0>, - ConvImpl::template process_tile<1, 0, 4, 1, 0, 1>, - ConvImpl::template process_tile<1, 0, 4, 1, 0, 2>, - ConvImpl::template process_tile<1, 0, 4, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 4, 1, 1, 0>, - ConvImpl::template process_tile<1, 0, 4, 1, 1, 1>, - ConvImpl::template process_tile<1, 0, 4, 1, 1, 2>, - ConvImpl::template process_tile<1, 0, 4, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 4, 1, 2, 0>, - ConvImpl::template process_tile<1, 0, 4, 1, 2, 1>, - ConvImpl::template process_tile<1, 0, 4, 1, 2, 2>, - ConvImpl::template process_tile<1, 0, 4, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 4, 1, 3, 0>, - ConvImpl::template process_tile<1, 0, 4, 1, 3, 1>, - ConvImpl::template process_tile<1, 0, 4, 1, 3, 2>, - ConvImpl::template process_tile<1, 0, 4, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 4, 2, 0, 0>, - ConvImpl::template process_tile<1, 0, 4, 2, 0, 1>, - ConvImpl::template process_tile<1, 0, 4, 2, 0, 2>, - ConvImpl::template process_tile<1, 0, 4, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 4, 2, 1, 0>, - ConvImpl::template process_tile<1, 0, 4, 2, 1, 1>, - ConvImpl::template process_tile<1, 0, 4, 2, 1, 2>, - ConvImpl::template process_tile<1, 0, 4, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 4, 2, 2, 0>, - ConvImpl::template process_tile<1, 0, 4, 2, 2, 1>, - ConvImpl::template process_tile<1, 0, 4, 2, 2, 2>, - ConvImpl::template process_tile<1, 0, 4, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 4, 2, 3, 0>, - ConvImpl::template process_tile<1, 0, 4, 2, 3, 1>, - ConvImpl::template process_tile<1, 0, 4, 2, 3, 2>, - ConvImpl::template process_tile<1, 0, 4, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 4, 3, 0, 0>, - ConvImpl::template process_tile<1, 0, 4, 3, 0, 1>, - ConvImpl::template process_tile<1, 0, 4, 3, 0, 2>, - ConvImpl::template process_tile<1, 0, 4, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 4, 3, 1, 0>, - ConvImpl::template process_tile<1, 0, 4, 3, 1, 1>, - ConvImpl::template process_tile<1, 0, 4, 3, 1, 2>, - ConvImpl::template process_tile<1, 0, 4, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 4, 3, 2, 0>, - ConvImpl::template process_tile<1, 0, 4, 3, 2, 1>, - ConvImpl::template process_tile<1, 0, 4, 3, 2, 2>, - ConvImpl::template process_tile<1, 0, 4, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 4, 3, 3, 0>, - ConvImpl::template process_tile<1, 0, 4, 3, 3, 1>, - ConvImpl::template process_tile<1, 0, 4, 3, 3, 2>, - ConvImpl::template process_tile<1, 0, 4, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 0, 4, 4, 0, 0>, - ConvImpl::template process_tile<1, 0, 4, 4, 0, 1>, - ConvImpl::template process_tile<1, 0, 4, 4, 0, 2>, - ConvImpl::template process_tile<1, 0, 4, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 0, 4, 4, 1, 0>, - ConvImpl::template process_tile<1, 0, 4, 4, 1, 1>, - ConvImpl::template process_tile<1, 0, 4, 4, 1, 2>, - ConvImpl::template process_tile<1, 0, 4, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 0, 4, 4, 2, 0>, - ConvImpl::template process_tile<1, 0, 4, 4, 2, 1>, - ConvImpl::template process_tile<1, 0, 4, 4, 2, 2>, - ConvImpl::template process_tile<1, 0, 4, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 0, 4, 4, 3, 0>, - ConvImpl::template process_tile<1, 0, 4, 4, 3, 1>, - ConvImpl::template process_tile<1, 0, 4, 4, 3, 2>, - ConvImpl::template process_tile<1, 0, 4, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 4 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 0, 0, 0, 0>, - ConvImpl::template process_tile<1, 1, 0, 0, 0, 1>, - ConvImpl::template process_tile<1, 1, 0, 0, 0, 2>, - ConvImpl::template process_tile<1, 1, 0, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 0, 0, 1, 0>, - ConvImpl::template process_tile<1, 1, 0, 0, 1, 1>, - ConvImpl::template process_tile<1, 1, 0, 0, 1, 2>, - ConvImpl::template process_tile<1, 1, 0, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 0, 0, 2, 0>, - ConvImpl::template process_tile<1, 1, 0, 0, 2, 1>, - ConvImpl::template process_tile<1, 1, 0, 0, 2, 2>, - ConvImpl::template process_tile<1, 1, 0, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 0, 0, 3, 0>, - ConvImpl::template process_tile<1, 1, 0, 0, 3, 1>, - ConvImpl::template process_tile<1, 1, 0, 0, 3, 2>, - ConvImpl::template process_tile<1, 1, 0, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 0, 1, 0, 0>, - ConvImpl::template process_tile<1, 1, 0, 1, 0, 1>, - ConvImpl::template process_tile<1, 1, 0, 1, 0, 2>, - ConvImpl::template process_tile<1, 1, 0, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 0, 1, 1, 0>, - ConvImpl::template process_tile<1, 1, 0, 1, 1, 1>, - ConvImpl::template process_tile<1, 1, 0, 1, 1, 2>, - ConvImpl::template process_tile<1, 1, 0, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 0, 1, 2, 0>, - ConvImpl::template process_tile<1, 1, 0, 1, 2, 1>, - ConvImpl::template process_tile<1, 1, 0, 1, 2, 2>, - ConvImpl::template process_tile<1, 1, 0, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 0, 1, 3, 0>, - ConvImpl::template process_tile<1, 1, 0, 1, 3, 1>, - ConvImpl::template process_tile<1, 1, 0, 1, 3, 2>, - ConvImpl::template process_tile<1, 1, 0, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 0, 2, 0, 0>, - ConvImpl::template process_tile<1, 1, 0, 2, 0, 1>, - ConvImpl::template process_tile<1, 1, 0, 2, 0, 2>, - ConvImpl::template process_tile<1, 1, 0, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 0, 2, 1, 0>, - ConvImpl::template process_tile<1, 1, 0, 2, 1, 1>, - ConvImpl::template process_tile<1, 1, 0, 2, 1, 2>, - ConvImpl::template process_tile<1, 1, 0, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 0, 2, 2, 0>, - ConvImpl::template process_tile<1, 1, 0, 2, 2, 1>, - ConvImpl::template process_tile<1, 1, 0, 2, 2, 2>, - ConvImpl::template process_tile<1, 1, 0, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 0, 2, 3, 0>, - ConvImpl::template process_tile<1, 1, 0, 2, 3, 1>, - ConvImpl::template process_tile<1, 1, 0, 2, 3, 2>, - ConvImpl::template process_tile<1, 1, 0, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 0, 3, 0, 0>, - ConvImpl::template process_tile<1, 1, 0, 3, 0, 1>, - ConvImpl::template process_tile<1, 1, 0, 3, 0, 2>, - ConvImpl::template process_tile<1, 1, 0, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 0, 3, 1, 0>, - ConvImpl::template process_tile<1, 1, 0, 3, 1, 1>, - ConvImpl::template process_tile<1, 1, 0, 3, 1, 2>, - ConvImpl::template process_tile<1, 1, 0, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 0, 3, 2, 0>, - ConvImpl::template process_tile<1, 1, 0, 3, 2, 1>, - ConvImpl::template process_tile<1, 1, 0, 3, 2, 2>, - ConvImpl::template process_tile<1, 1, 0, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 0, 3, 3, 0>, - ConvImpl::template process_tile<1, 1, 0, 3, 3, 1>, - ConvImpl::template process_tile<1, 1, 0, 3, 3, 2>, - ConvImpl::template process_tile<1, 1, 0, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 0, 4, 0, 0>, - ConvImpl::template process_tile<1, 1, 0, 4, 0, 1>, - ConvImpl::template process_tile<1, 1, 0, 4, 0, 2>, - ConvImpl::template process_tile<1, 1, 0, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 0, 4, 1, 0>, - ConvImpl::template process_tile<1, 1, 0, 4, 1, 1>, - ConvImpl::template process_tile<1, 1, 0, 4, 1, 2>, - ConvImpl::template process_tile<1, 1, 0, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 0, 4, 2, 0>, - ConvImpl::template process_tile<1, 1, 0, 4, 2, 1>, - ConvImpl::template process_tile<1, 1, 0, 4, 2, 2>, - ConvImpl::template process_tile<1, 1, 0, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 0, 4, 3, 0>, - ConvImpl::template process_tile<1, 1, 0, 4, 3, 1>, - ConvImpl::template process_tile<1, 1, 0, 4, 3, 2>, - ConvImpl::template process_tile<1, 1, 0, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 1, 0, 0, 0>, - ConvImpl::template process_tile<1, 1, 1, 0, 0, 1>, - ConvImpl::template process_tile<1, 1, 1, 0, 0, 2>, - ConvImpl::template process_tile<1, 1, 1, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 1, 0, 1, 0>, - ConvImpl::template process_tile<1, 1, 1, 0, 1, 1>, - ConvImpl::template process_tile<1, 1, 1, 0, 1, 2>, - ConvImpl::template process_tile<1, 1, 1, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 1, 0, 2, 0>, - ConvImpl::template process_tile<1, 1, 1, 0, 2, 1>, - ConvImpl::template process_tile<1, 1, 1, 0, 2, 2>, - ConvImpl::template process_tile<1, 1, 1, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 1, 0, 3, 0>, - ConvImpl::template process_tile<1, 1, 1, 0, 3, 1>, - ConvImpl::template process_tile<1, 1, 1, 0, 3, 2>, - ConvImpl::template process_tile<1, 1, 1, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 1, 1, 0, 0>, - ConvImpl::template process_tile<1, 1, 1, 1, 0, 1>, - ConvImpl::template process_tile<1, 1, 1, 1, 0, 2>, - ConvImpl::template process_tile<1, 1, 1, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 1, 1, 1, 0>, - ConvImpl::template process_tile<1, 1, 1, 1, 1, 1>, - ConvImpl::template process_tile<1, 1, 1, 1, 1, 2>, - ConvImpl::template process_tile<1, 1, 1, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 1, 1, 2, 0>, - ConvImpl::template process_tile<1, 1, 1, 1, 2, 1>, - ConvImpl::template process_tile<1, 1, 1, 1, 2, 2>, - ConvImpl::template process_tile<1, 1, 1, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 1, 1, 3, 0>, - ConvImpl::template process_tile<1, 1, 1, 1, 3, 1>, - ConvImpl::template process_tile<1, 1, 1, 1, 3, 2>, - ConvImpl::template process_tile<1, 1, 1, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 1, 2, 0, 0>, - ConvImpl::template process_tile<1, 1, 1, 2, 0, 1>, - ConvImpl::template process_tile<1, 1, 1, 2, 0, 2>, - ConvImpl::template process_tile<1, 1, 1, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 1, 2, 1, 0>, - ConvImpl::template process_tile<1, 1, 1, 2, 1, 1>, - ConvImpl::template process_tile<1, 1, 1, 2, 1, 2>, - ConvImpl::template process_tile<1, 1, 1, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 1, 2, 2, 0>, - ConvImpl::template process_tile<1, 1, 1, 2, 2, 1>, - ConvImpl::template process_tile<1, 1, 1, 2, 2, 2>, - ConvImpl::template process_tile<1, 1, 1, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 1, 2, 3, 0>, - ConvImpl::template process_tile<1, 1, 1, 2, 3, 1>, - ConvImpl::template process_tile<1, 1, 1, 2, 3, 2>, - ConvImpl::template process_tile<1, 1, 1, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 1, 3, 0, 0>, - ConvImpl::template process_tile<1, 1, 1, 3, 0, 1>, - ConvImpl::template process_tile<1, 1, 1, 3, 0, 2>, - ConvImpl::template process_tile<1, 1, 1, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 1, 3, 1, 0>, - ConvImpl::template process_tile<1, 1, 1, 3, 1, 1>, - ConvImpl::template process_tile<1, 1, 1, 3, 1, 2>, - ConvImpl::template process_tile<1, 1, 1, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 1, 3, 2, 0>, - ConvImpl::template process_tile<1, 1, 1, 3, 2, 1>, - ConvImpl::template process_tile<1, 1, 1, 3, 2, 2>, - ConvImpl::template process_tile<1, 1, 1, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 1, 3, 3, 0>, - ConvImpl::template process_tile<1, 1, 1, 3, 3, 1>, - ConvImpl::template process_tile<1, 1, 1, 3, 3, 2>, - ConvImpl::template process_tile<1, 1, 1, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 1, 4, 0, 0>, - ConvImpl::template process_tile<1, 1, 1, 4, 0, 1>, - ConvImpl::template process_tile<1, 1, 1, 4, 0, 2>, - ConvImpl::template process_tile<1, 1, 1, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 1, 4, 1, 0>, - ConvImpl::template process_tile<1, 1, 1, 4, 1, 1>, - ConvImpl::template process_tile<1, 1, 1, 4, 1, 2>, - ConvImpl::template process_tile<1, 1, 1, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 1, 4, 2, 0>, - ConvImpl::template process_tile<1, 1, 1, 4, 2, 1>, - ConvImpl::template process_tile<1, 1, 1, 4, 2, 2>, - ConvImpl::template process_tile<1, 1, 1, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 1, 4, 3, 0>, - ConvImpl::template process_tile<1, 1, 1, 4, 3, 1>, - ConvImpl::template process_tile<1, 1, 1, 4, 3, 2>, - ConvImpl::template process_tile<1, 1, 1, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 2, 0, 0, 0>, - ConvImpl::template process_tile<1, 1, 2, 0, 0, 1>, - ConvImpl::template process_tile<1, 1, 2, 0, 0, 2>, - ConvImpl::template process_tile<1, 1, 2, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 2, 0, 1, 0>, - ConvImpl::template process_tile<1, 1, 2, 0, 1, 1>, - ConvImpl::template process_tile<1, 1, 2, 0, 1, 2>, - ConvImpl::template process_tile<1, 1, 2, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 2, 0, 2, 0>, - ConvImpl::template process_tile<1, 1, 2, 0, 2, 1>, - ConvImpl::template process_tile<1, 1, 2, 0, 2, 2>, - ConvImpl::template process_tile<1, 1, 2, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 2, 0, 3, 0>, - ConvImpl::template process_tile<1, 1, 2, 0, 3, 1>, - ConvImpl::template process_tile<1, 1, 2, 0, 3, 2>, - ConvImpl::template process_tile<1, 1, 2, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 2, 1, 0, 0>, - ConvImpl::template process_tile<1, 1, 2, 1, 0, 1>, - ConvImpl::template process_tile<1, 1, 2, 1, 0, 2>, - ConvImpl::template process_tile<1, 1, 2, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 2, 1, 1, 0>, - ConvImpl::template process_tile<1, 1, 2, 1, 1, 1>, - ConvImpl::template process_tile<1, 1, 2, 1, 1, 2>, - ConvImpl::template process_tile<1, 1, 2, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 2, 1, 2, 0>, - ConvImpl::template process_tile<1, 1, 2, 1, 2, 1>, - ConvImpl::template process_tile<1, 1, 2, 1, 2, 2>, - ConvImpl::template process_tile<1, 1, 2, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 2, 1, 3, 0>, - ConvImpl::template process_tile<1, 1, 2, 1, 3, 1>, - ConvImpl::template process_tile<1, 1, 2, 1, 3, 2>, - ConvImpl::template process_tile<1, 1, 2, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 2, 2, 0, 0>, - ConvImpl::template process_tile<1, 1, 2, 2, 0, 1>, - ConvImpl::template process_tile<1, 1, 2, 2, 0, 2>, - ConvImpl::template process_tile<1, 1, 2, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 2, 2, 1, 0>, - ConvImpl::template process_tile<1, 1, 2, 2, 1, 1>, - ConvImpl::template process_tile<1, 1, 2, 2, 1, 2>, - ConvImpl::template process_tile<1, 1, 2, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 2, 2, 2, 0>, - ConvImpl::template process_tile<1, 1, 2, 2, 2, 1>, - ConvImpl::template process_tile<1, 1, 2, 2, 2, 2>, - ConvImpl::template process_tile<1, 1, 2, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 2, 2, 3, 0>, - ConvImpl::template process_tile<1, 1, 2, 2, 3, 1>, - ConvImpl::template process_tile<1, 1, 2, 2, 3, 2>, - ConvImpl::template process_tile<1, 1, 2, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 2, 3, 0, 0>, - ConvImpl::template process_tile<1, 1, 2, 3, 0, 1>, - ConvImpl::template process_tile<1, 1, 2, 3, 0, 2>, - ConvImpl::template process_tile<1, 1, 2, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 2, 3, 1, 0>, - ConvImpl::template process_tile<1, 1, 2, 3, 1, 1>, - ConvImpl::template process_tile<1, 1, 2, 3, 1, 2>, - ConvImpl::template process_tile<1, 1, 2, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 2, 3, 2, 0>, - ConvImpl::template process_tile<1, 1, 2, 3, 2, 1>, - ConvImpl::template process_tile<1, 1, 2, 3, 2, 2>, - ConvImpl::template process_tile<1, 1, 2, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 2, 3, 3, 0>, - ConvImpl::template process_tile<1, 1, 2, 3, 3, 1>, - ConvImpl::template process_tile<1, 1, 2, 3, 3, 2>, - ConvImpl::template process_tile<1, 1, 2, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 2, 4, 0, 0>, - ConvImpl::template process_tile<1, 1, 2, 4, 0, 1>, - ConvImpl::template process_tile<1, 1, 2, 4, 0, 2>, - ConvImpl::template process_tile<1, 1, 2, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 2, 4, 1, 0>, - ConvImpl::template process_tile<1, 1, 2, 4, 1, 1>, - ConvImpl::template process_tile<1, 1, 2, 4, 1, 2>, - ConvImpl::template process_tile<1, 1, 2, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 2, 4, 2, 0>, - ConvImpl::template process_tile<1, 1, 2, 4, 2, 1>, - ConvImpl::template process_tile<1, 1, 2, 4, 2, 2>, - ConvImpl::template process_tile<1, 1, 2, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 2, 4, 3, 0>, - ConvImpl::template process_tile<1, 1, 2, 4, 3, 1>, - ConvImpl::template process_tile<1, 1, 2, 4, 3, 2>, - ConvImpl::template process_tile<1, 1, 2, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 3, 0, 0, 0>, - ConvImpl::template process_tile<1, 1, 3, 0, 0, 1>, - ConvImpl::template process_tile<1, 1, 3, 0, 0, 2>, - ConvImpl::template process_tile<1, 1, 3, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 3, 0, 1, 0>, - ConvImpl::template process_tile<1, 1, 3, 0, 1, 1>, - ConvImpl::template process_tile<1, 1, 3, 0, 1, 2>, - ConvImpl::template process_tile<1, 1, 3, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 3, 0, 2, 0>, - ConvImpl::template process_tile<1, 1, 3, 0, 2, 1>, - ConvImpl::template process_tile<1, 1, 3, 0, 2, 2>, - ConvImpl::template process_tile<1, 1, 3, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 3, 0, 3, 0>, - ConvImpl::template process_tile<1, 1, 3, 0, 3, 1>, - ConvImpl::template process_tile<1, 1, 3, 0, 3, 2>, - ConvImpl::template process_tile<1, 1, 3, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 3, 1, 0, 0>, - ConvImpl::template process_tile<1, 1, 3, 1, 0, 1>, - ConvImpl::template process_tile<1, 1, 3, 1, 0, 2>, - ConvImpl::template process_tile<1, 1, 3, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 3, 1, 1, 0>, - ConvImpl::template process_tile<1, 1, 3, 1, 1, 1>, - ConvImpl::template process_tile<1, 1, 3, 1, 1, 2>, - ConvImpl::template process_tile<1, 1, 3, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 3, 1, 2, 0>, - ConvImpl::template process_tile<1, 1, 3, 1, 2, 1>, - ConvImpl::template process_tile<1, 1, 3, 1, 2, 2>, - ConvImpl::template process_tile<1, 1, 3, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 3, 1, 3, 0>, - ConvImpl::template process_tile<1, 1, 3, 1, 3, 1>, - ConvImpl::template process_tile<1, 1, 3, 1, 3, 2>, - ConvImpl::template process_tile<1, 1, 3, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 3, 2, 0, 0>, - ConvImpl::template process_tile<1, 1, 3, 2, 0, 1>, - ConvImpl::template process_tile<1, 1, 3, 2, 0, 2>, - ConvImpl::template process_tile<1, 1, 3, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 3, 2, 1, 0>, - ConvImpl::template process_tile<1, 1, 3, 2, 1, 1>, - ConvImpl::template process_tile<1, 1, 3, 2, 1, 2>, - ConvImpl::template process_tile<1, 1, 3, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 3, 2, 2, 0>, - ConvImpl::template process_tile<1, 1, 3, 2, 2, 1>, - ConvImpl::template process_tile<1, 1, 3, 2, 2, 2>, - ConvImpl::template process_tile<1, 1, 3, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 3, 2, 3, 0>, - ConvImpl::template process_tile<1, 1, 3, 2, 3, 1>, - ConvImpl::template process_tile<1, 1, 3, 2, 3, 2>, - ConvImpl::template process_tile<1, 1, 3, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 3, 3, 0, 0>, - ConvImpl::template process_tile<1, 1, 3, 3, 0, 1>, - ConvImpl::template process_tile<1, 1, 3, 3, 0, 2>, - ConvImpl::template process_tile<1, 1, 3, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 3, 3, 1, 0>, - ConvImpl::template process_tile<1, 1, 3, 3, 1, 1>, - ConvImpl::template process_tile<1, 1, 3, 3, 1, 2>, - ConvImpl::template process_tile<1, 1, 3, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 3, 3, 2, 0>, - ConvImpl::template process_tile<1, 1, 3, 3, 2, 1>, - ConvImpl::template process_tile<1, 1, 3, 3, 2, 2>, - ConvImpl::template process_tile<1, 1, 3, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 3, 3, 3, 0>, - ConvImpl::template process_tile<1, 1, 3, 3, 3, 1>, - ConvImpl::template process_tile<1, 1, 3, 3, 3, 2>, - ConvImpl::template process_tile<1, 1, 3, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 3, 4, 0, 0>, - ConvImpl::template process_tile<1, 1, 3, 4, 0, 1>, - ConvImpl::template process_tile<1, 1, 3, 4, 0, 2>, - ConvImpl::template process_tile<1, 1, 3, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 3, 4, 1, 0>, - ConvImpl::template process_tile<1, 1, 3, 4, 1, 1>, - ConvImpl::template process_tile<1, 1, 3, 4, 1, 2>, - ConvImpl::template process_tile<1, 1, 3, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 3, 4, 2, 0>, - ConvImpl::template process_tile<1, 1, 3, 4, 2, 1>, - ConvImpl::template process_tile<1, 1, 3, 4, 2, 2>, - ConvImpl::template process_tile<1, 1, 3, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 3, 4, 3, 0>, - ConvImpl::template process_tile<1, 1, 3, 4, 3, 1>, - ConvImpl::template process_tile<1, 1, 3, 4, 3, 2>, - ConvImpl::template process_tile<1, 1, 3, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 4, 0, 0, 0>, - ConvImpl::template process_tile<1, 1, 4, 0, 0, 1>, - ConvImpl::template process_tile<1, 1, 4, 0, 0, 2>, - ConvImpl::template process_tile<1, 1, 4, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 4, 0, 1, 0>, - ConvImpl::template process_tile<1, 1, 4, 0, 1, 1>, - ConvImpl::template process_tile<1, 1, 4, 0, 1, 2>, - ConvImpl::template process_tile<1, 1, 4, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 4, 0, 2, 0>, - ConvImpl::template process_tile<1, 1, 4, 0, 2, 1>, - ConvImpl::template process_tile<1, 1, 4, 0, 2, 2>, - ConvImpl::template process_tile<1, 1, 4, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 4, 0, 3, 0>, - ConvImpl::template process_tile<1, 1, 4, 0, 3, 1>, - ConvImpl::template process_tile<1, 1, 4, 0, 3, 2>, - ConvImpl::template process_tile<1, 1, 4, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 4, 1, 0, 0>, - ConvImpl::template process_tile<1, 1, 4, 1, 0, 1>, - ConvImpl::template process_tile<1, 1, 4, 1, 0, 2>, - ConvImpl::template process_tile<1, 1, 4, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 4, 1, 1, 0>, - ConvImpl::template process_tile<1, 1, 4, 1, 1, 1>, - ConvImpl::template process_tile<1, 1, 4, 1, 1, 2>, - ConvImpl::template process_tile<1, 1, 4, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 4, 1, 2, 0>, - ConvImpl::template process_tile<1, 1, 4, 1, 2, 1>, - ConvImpl::template process_tile<1, 1, 4, 1, 2, 2>, - ConvImpl::template process_tile<1, 1, 4, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 4, 1, 3, 0>, - ConvImpl::template process_tile<1, 1, 4, 1, 3, 1>, - ConvImpl::template process_tile<1, 1, 4, 1, 3, 2>, - ConvImpl::template process_tile<1, 1, 4, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 4, 2, 0, 0>, - ConvImpl::template process_tile<1, 1, 4, 2, 0, 1>, - ConvImpl::template process_tile<1, 1, 4, 2, 0, 2>, - ConvImpl::template process_tile<1, 1, 4, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 4, 2, 1, 0>, - ConvImpl::template process_tile<1, 1, 4, 2, 1, 1>, - ConvImpl::template process_tile<1, 1, 4, 2, 1, 2>, - ConvImpl::template process_tile<1, 1, 4, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 4, 2, 2, 0>, - ConvImpl::template process_tile<1, 1, 4, 2, 2, 1>, - ConvImpl::template process_tile<1, 1, 4, 2, 2, 2>, - ConvImpl::template process_tile<1, 1, 4, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 4, 2, 3, 0>, - ConvImpl::template process_tile<1, 1, 4, 2, 3, 1>, - ConvImpl::template process_tile<1, 1, 4, 2, 3, 2>, - ConvImpl::template process_tile<1, 1, 4, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 4, 3, 0, 0>, - ConvImpl::template process_tile<1, 1, 4, 3, 0, 1>, - ConvImpl::template process_tile<1, 1, 4, 3, 0, 2>, - ConvImpl::template process_tile<1, 1, 4, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 4, 3, 1, 0>, - ConvImpl::template process_tile<1, 1, 4, 3, 1, 1>, - ConvImpl::template process_tile<1, 1, 4, 3, 1, 2>, - ConvImpl::template process_tile<1, 1, 4, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 4, 3, 2, 0>, - ConvImpl::template process_tile<1, 1, 4, 3, 2, 1>, - ConvImpl::template process_tile<1, 1, 4, 3, 2, 2>, - ConvImpl::template process_tile<1, 1, 4, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 4, 3, 3, 0>, - ConvImpl::template process_tile<1, 1, 4, 3, 3, 1>, - ConvImpl::template process_tile<1, 1, 4, 3, 3, 2>, - ConvImpl::template process_tile<1, 1, 4, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - ConvImpl::template process_tile<1, 1, 4, 4, 0, 0>, - ConvImpl::template process_tile<1, 1, 4, 4, 0, 1>, - ConvImpl::template process_tile<1, 1, 4, 4, 0, 2>, - ConvImpl::template process_tile<1, 1, 4, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - ConvImpl::template process_tile<1, 1, 4, 4, 1, 0>, - ConvImpl::template process_tile<1, 1, 4, 4, 1, 1>, - ConvImpl::template process_tile<1, 1, 4, 4, 1, 2>, - ConvImpl::template process_tile<1, 1, 4, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - ConvImpl::template process_tile<1, 1, 4, 4, 2, 0>, - ConvImpl::template process_tile<1, 1, 4, 4, 2, 1>, - ConvImpl::template process_tile<1, 1, 4, 4, 2, 2>, - ConvImpl::template process_tile<1, 1, 4, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - ConvImpl::template process_tile<1, 1, 4, 4, 3, 0>, - ConvImpl::template process_tile<1, 1, 4, 4, 3, 1>, - ConvImpl::template process_tile<1, 1, 4, 4, 3, 2>, - ConvImpl::template process_tile<1, 1, 4, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - }, // Input pad bottom = 4 - }, // Input pad left = 1 - }, // Input pad top = 1 +void ConvImpl::process_tile( + const int n_channels, + const float* const weights, + const int weight_row_stride, + const int weight_col_stride, + const float* const inptr, + const int in_row_stride, + const int in_col_stride, + float* const outptr, + const int out_row_stride, + const int out_col_stride, + const int, const int, const int, const int, const int, const int +) +{ + constexpr auto inner_tile_rows = DWC::inner_tile_rows; + constexpr auto inner_tile_cols = DWC::inner_tile_cols; + constexpr auto kernel_rows = DWC::kernel_rows; + constexpr auto kernel_cols = DWC::kernel_cols; + constexpr auto output_tile_rows = DWC::output_tile_rows; + constexpr auto output_tile_cols = DWC::output_tile_cols; + constexpr auto stride_rows = DWC::stride_rows; + constexpr auto stride_cols = DWC::stride_cols; + + // Extract parameters + const int in_pad_top = 0; + const int in_pad_left = 0; + const int in_pad_bottom = 0; + const int in_pad_right = 0; + const int out_pad_bottom = 0; + const int out_pad_right = 0; + + // Compute valid ranges of the tile + const int in_cells_i = inner_tile_rows - in_pad_bottom; + const int in_cells_j = inner_tile_cols - in_pad_right; + const int out_cells_i = output_tile_rows - out_pad_bottom; + const int out_cells_j = output_tile_cols - out_pad_right; + + // Copy pointers + const float *uptr0 = inptr; + const float *wptr0 = weights; + float *vptr0 = outptr; + const bool same_strides = ( + weight_col_stride == in_col_stride && + weight_col_stride == out_col_stride + ); + + int channels_remaining = n_channels; + if (channels_remaining >= 4 && same_strides) + { + int c4_rem = channels_remaining / 4; + channels_remaining %= 4; + const int prefetch_depth = 8; + + asm volatile ( + "qW22 .req q0\n" "vW22 .req v0\n" + "qU64 .req q1\n" "qU35 .req q1\n" "qV41 .req q1\n" + "vU64 .req v1\n" "vU35 .req v1\n" "vV41 .req v1\n" + "qU34 .req q2\n" "qU21 .req q2\n" "qV43 .req q2\n" + "vU34 .req v2\n" "vU21 .req v2\n" "vV43 .req v2\n" + "qW21 .req q3\n" "vW21 .req v3\n" + "qU24 .req q4\n" "qU54 .req q4\n" "qV31 .req q4\n" + "vU24 .req v4\n" "vU54 .req v4\n" "vV31 .req v4\n" + "qV12 .req q5\n" "qU61 .req q5\n" "vV12 .req v5\n" "vU61 .req v5\n" + "qU26 .req q6\n" "qV32 .req q6\n" "vU26 .req v6\n" "vV32 .req v6\n" + "qU36 .req q7\n" "qU51 .req q7\n" "qU66 .req q7\n" "qU12 .req q7\n" + "vU36 .req v7\n" "vU51 .req v7\n" "vU66 .req v7\n" "vU12 .req v7\n" + "qV14 .req q8\n" "qV11 .req q8\n" "qU65 .req q8\n" + "vV14 .req v8\n" "vV11 .req v8\n" "vU65 .req v8\n" + "qU15 .req q9\n" "qU22 .req q9\n" "qU45 .req q9\n" + "vU15 .req v9\n" "vU22 .req v9\n" "vU45 .req v9\n" + "qV22 .req q10\n" "qU14 .req q10\n" "vV22 .req v10\n" "vU14 .req v10\n" + "qU44 .req q11\n" "qU43 .req q11\n" "qU11 .req q11\n" + "vU44 .req v11\n" "vU43 .req v11\n" "vU11 .req v11\n" + "qV24 .req q12\n" "qV42 .req q12\n" "vV24 .req v12\n" "vV42 .req v12\n" + "qW31 .req q13\n" "vW31 .req v13\n" "qW13 .req q14\n" "vW13 .req v14\n" + "qU33 .req q15\n" "qU62 .req q15\n" "qU25 .req q15\n" "qU56 .req q15\n" + "vU33 .req v15\n" "vU62 .req v15\n" "vU25 .req v15\n" "vU56 .req v15\n" + "qW33 .req q16\n" "vW33 .req v16\n" + "qU42 .req q17\n" "qU16 .req q17\n" "qV44 .req q17\n" + "vU42 .req v17\n" "vU16 .req v17\n" "vV44 .req v17\n" + "qU63 .req q18\n" "qU31 .req q18\n" "qV34 .req q18\n" + "vU63 .req v18\n" "vU31 .req v18\n" "vV34 .req v18\n" + "qW11 .req q19\n" "vW11 .req v19\n" "qU41 .req q20\n" "qV13 .req q20\n" + "vU41 .req v20\n" "vV13 .req v20\n" "qV33 .req q21\n" "vV33 .req v21\n" + "qU46 .req q22\n" "qU32 .req q22\n" "qU13 .req q22\n" + "vU46 .req v22\n" "vU32 .req v22\n" "vU13 .req v22\n" "qW23 .req q23\n" + "vW23 .req v23\n" "qV23 .req q24\n" "vV23 .req v24\n" + "qV21 .req q25\n" "qU55 .req q25\n" "vV21 .req v25\n" "vU55 .req v25\n" + "qW12 .req q26\n" "vW12 .req v26\n" "qW32 .req q27\n" "vW32 .req v27\n" + "qU23 .req q28\n" "qU52 .req q28\n" + "vU23 .req v28\n" "vU52 .req v28\n" "qU53 .req q29\n" "vU53 .req v29\n" + + "uptr1 .req x0\n" + "uptr2 .req x1\n" + "uptr3 .req x2\n" + "uptr4 .req x3\n" + "uptr5 .req x4\n" + + "vptr1 .req x5\n" + "vptr2 .req x6\n" + "vptr3 .req x7\n" + + "wptr1 .req x8\n" + "wptr2 .req x9\n" + + // Prepare pointers and strides + "add uptr1, %x[uptr0], %x[u_row_stride]\n" + "add uptr2, uptr1 , %x[u_row_stride]\n" + "add uptr3, uptr2 , %x[u_row_stride]\n" + "add uptr4, uptr3 , %x[u_row_stride]\n" + "add uptr5, uptr4 , %x[u_row_stride]\n" + + "add vptr1, %x[vptr0], %x[v_row_stride]\n" + "add vptr2, vptr1 , %x[v_row_stride]\n" + "add vptr3, vptr2 , %x[v_row_stride]\n" + + "add wptr1, %x[wptr0], %x[w_row_stride]\n" + "add wptr2, wptr1 , %x[w_row_stride]\n" + + // Load initial operands + "ldr qU16, [%x[uptr0], %x[uvw_col_stride5]]\n" + "ldr qW13, [%x[wptr0], %x[uvw_col_stride2]]\n" + "subs %x[c4_rem], %x[c4_rem], #1\n" + "ldr qU15, [%x[uptr0], %x[uvw_col_stride4]]\n" + "ldr qW23, [wptr1, %x[uvw_col_stride2]]\n" + "ldr qU14, [%x[uptr0], %x[uvw_col_stride3]]\n" + "ldr qW33, [wptr2, %x[uvw_col_stride2]]\n" + "ldr qU26, [uptr1, %x[uvw_col_stride5]]\n" + "ldr qW12, [%x[wptr0], %x[uvw_col_stride1]]\n" + "ldr qU25, [uptr1, %x[uvw_col_stride4]]\n" + "ldr qW22, [wptr1, %x[uvw_col_stride1]]\n" + "ldr qU36, [uptr2, %x[uvw_col_stride5]]\n" + "ldr qW32, [wptr2, %x[uvw_col_stride1]]\n" + "ldr qW11, [%x[wptr0]], #0x10\n" + "fmul vV14.4s, vU16.4s, vW13.4s\n" + "ldr qU24, [uptr1, %x[uvw_col_stride3]]\n" + "fmul vV13.4s, vU15.4s, vW13.4s\n" + "ldr qW31, [wptr2], #0x10\n" + "fmla vV14.4s, vU15.4s, vW12.4s\n" + "ldr qW21, [wptr1], #0x10\n" + "fmul vV12.4s, vU14.4s, vW13.4s\n" + "ldr qU34, [uptr2, %x[uvw_col_stride3]]\n" + "fmla vV13.4s, vU14.4s, vW12.4s\n" + "ldr qU46, [uptr3, %x[uvw_col_stride5]]\n" + "fmla vV14.4s, vU14.4s, vW11.4s\n" + "ldr qU45, [uptr3, %x[uvw_col_stride4]]\n" + "fmla vV14.4s, vU26.4s, vW23.4s\n" + "ldr qU35, [uptr2, %x[uvw_col_stride4]]\n" + "fmul vV24.4s, vU26.4s, vW13.4s\n" + "ldr qU44, [uptr3, %x[uvw_col_stride3]]\n" + "fmla vV13.4s, vU25.4s, vW23.4s\n" + "beq 2f\n" // Single iteration only + + "1:" // Loop body + "fmla vV14.4s, vU25.4s, vW22.4s\n" + "prfm pldl1keep, [%x[wptr0], %[prftch]]\n" + "fmul vV23.4s, vU25.4s, vW13.4s\n" + "prfm pldl1keep, [%x[wptr0], %x[prftch_uvw_col_stride1]]\n" + "fmla vV24.4s, vU25.4s, vW12.4s\n" + "ldr qU56, [uptr4, %x[uvw_col_stride5]]\n" + "fmla vV12.4s, vU24.4s, vW23.4s\n" + "prfm pldl1keep, [%x[wptr0], %x[prftch_uvw_col_stride2] ]\n" + "fmla vV13.4s, vU24.4s, vW22.4s\n" + "prfm pldl1keep, [ wptr1 , %[prftch]]\n" + "fmla vV14.4s, vU24.4s, vW21.4s\n" + "prfm pldl1keep, [ wptr1 , %x[prftch_uvw_col_stride1]]\n" + "fmul vV22.4s, vU24.4s, vW13.4s\n" + "prfm pldl1keep, [ wptr1 , %x[prftch_uvw_col_stride2] ]\n" + "fmla vV23.4s, vU24.4s, vW12.4s\n" + "prfm pldl1keep, [ wptr2 , %x[prftch]]\n" + "fmla vV24.4s, vU24.4s, vW11.4s\n" + "ldr qU55, [uptr4, %x[uvw_col_stride4]]\n" + "fmla vV14.4s, vU36.4s, vW33.4s\n" + "prfm pldl1keep, [ wptr2 , %x[prftch_uvw_col_stride1]]\n" + "fmla vV24.4s, vU36.4s, vW23.4s\n" + "prfm pldl1keep, [ wptr2 , %x[prftch_uvw_col_stride2] ]\n" + "fmul vV34.4s, vU36.4s, vW13.4s\n" + "ldr qU54, [uptr4, %x[uvw_col_stride3]]\n" + "fmla vV13.4s, vU35.4s, vW33.4s\n" + "prfm pldl1keep, [ uptr2 , %x[prftch_uvw_col_stride1]]\n" + "fmla vV14.4s, vU35.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr2 , %x[prftch_uvw_col_stride2] ]\n" + "fmla vV23.4s, vU35.4s, vW23.4s\n" + "prfm pldl1keep, [ uptr2 , %x[prftch_uvw_col_stride3] ]\n" + "fmla vV24.4s, vU35.4s, vW22.4s\n" + "prfm pldl1keep, [ uptr2 , %x[prftch_uvw_col_stride4] ]\n" + "fmul vV33.4s, vU35.4s, vW13.4s\n" + "prfm pldl1keep, [ uptr2 , %x[prftch_uvw_col_stride5] ]\n" + "fmla vV34.4s, vU35.4s, vW12.4s\n" + "ldr qU66, [uptr5, %x[uvw_col_stride5]]\n" + "fmla vV12.4s, vU34.4s, vW33.4s\n" + "prfm pldl1keep, [ uptr3 , %[prftch]]\n" + "fmla vV13.4s, vU34.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr3 , %x[prftch_uvw_col_stride1]]\n" + "fmla vV14.4s, vU34.4s, vW31.4s\n" + "str qV14, [%x[vptr0], %x[uvw_col_stride3]]\n" + "fmla vV22.4s, vU34.4s, vW23.4s\n" + "prfm pldl1keep, [ uptr3 , %x[prftch_uvw_col_stride2] ]\n" + "fmla vV23.4s, vU34.4s, vW22.4s\n" + "prfm pldl1keep, [ uptr3 , %x[prftch_uvw_col_stride3] ]\n" + "fmla vV24.4s, vU34.4s, vW21.4s\n" + "prfm pldl1keep, [ uptr3 , %x[prftch_uvw_col_stride4] ]\n" + "fmul vV32.4s, vU34.4s, vW13.4s\n" + "prfm pldl1keep, [ uptr3 , %x[prftch_uvw_col_stride5] ]\n" + "fmla vV33.4s, vU34.4s, vW12.4s\n" + "prfm pldl1keep, [ uptr4 , %[prftch]]\n" + "fmla vV34.4s, vU34.4s, vW11.4s\n" + "ldr qU65, [uptr5, %x[uvw_col_stride4]]\n" + "fmla vV24.4s, vU46.4s, vW33.4s\n" + "prfm pldl1keep, [ uptr4 , %x[prftch_uvw_col_stride1]]\n" + "fmla vV34.4s, vU46.4s, vW23.4s\n" + "prfm pldl1keep, [ uptr4 , %x[prftch_uvw_col_stride2] ]\n" + "fmul vV44.4s, vU46.4s, vW13.4s\n" + "ldr qU64, [uptr5, %x[uvw_col_stride3]]\n" + "fmla vV23.4s, vU45.4s, vW33.4s\n" + "prfm pldl1keep, [ uptr4 , %x[prftch_uvw_col_stride3] ]\n" + "fmla vV24.4s, vU45.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr4 , %x[prftch_uvw_col_stride4] ]\n" + "fmla vV33.4s, vU45.4s, vW23.4s\n" + "prfm pldl1keep, [ uptr4 , %x[prftch_uvw_col_stride5] ]\n" + "fmla vV34.4s, vU45.4s, vW22.4s\n" + "prfm pldl1keep, [ uptr5 , %[prftch]]\n" + "fmul vV43.4s, vU45.4s, vW13.4s\n" + "prfm pldl1keep, [ uptr5 , %x[prftch_uvw_col_stride1]]\n" + "fmla vV44.4s, vU45.4s, vW12.4s\n" + "ldr qU13, [%x[uptr0], %x[uvw_col_stride2]]\n" + "fmla vV22.4s, vU44.4s, vW33.4s\n" + "prfm pldl1keep, [ uptr5 , %x[prftch_uvw_col_stride2] ]\n" + "fmla vV23.4s, vU44.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr5 , %x[prftch_uvw_col_stride3] ]\n" + "fmla vV24.4s, vU44.4s, vW31.4s\n" + "str qV24, [vptr1, %x[uvw_col_stride3]]\n" + "fmla vV32.4s, vU44.4s, vW23.4s\n" + "prfm pldl1keep, [ uptr5 , %x[prftch_uvw_col_stride4] ]\n" + "fmla vV33.4s, vU44.4s, vW22.4s\n" + "prfm pldl1keep, [ uptr5 , %x[prftch_uvw_col_stride5] ]\n" + "fmla vV34.4s, vU44.4s, vW21.4s\n" + "prfm pstl1keep, [%x[vptr0], %[prftch]]\n" + "fmul vV42.4s, vU44.4s, vW13.4s\n" + "prfm pstl1keep, [%x[vptr0], %x[prftch_uvw_col_stride1]]\n" + "fmla vV43.4s, vU44.4s, vW12.4s\n" + "prfm pstl1keep, [%x[vptr0], %x[prftch_uvw_col_stride2] ]\n" + "fmla vV44.4s, vU44.4s, vW11.4s\n" + "ldr qU23, [uptr1, %x[uvw_col_stride2]]\n" + "fmla vV34.4s, vU56.4s, vW33.4s\n" + "prfm pstl1keep, [%x[vptr0], %x[prftch_uvw_col_stride3] ]\n" + "fmla vV44.4s, vU56.4s, vW23.4s\n" + "ldr qU33, [uptr2, %x[uvw_col_stride2]]\n" + "fmla vV33.4s, vU55.4s, vW33.4s\n" + "prfm pstl1keep, [ vptr1 , %[prftch]]\n" + "fmla vV34.4s, vU55.4s, vW32.4s\n" + "prfm pstl1keep, [ vptr1 , %x[prftch_uvw_col_stride1]]\n" + "fmla vV43.4s, vU55.4s, vW23.4s\n" + "prfm pstl1keep, [ vptr1 , %x[prftch_uvw_col_stride2] ]\n" + "fmla vV44.4s, vU55.4s, vW22.4s\n" + "ldr qU43, [uptr3, %x[uvw_col_stride2]]\n" + "fmla vV32.4s, vU54.4s, vW33.4s\n" + "prfm pstl1keep, [ vptr1 , %x[prftch_uvw_col_stride3] ]\n" + "fmla vV33.4s, vU54.4s, vW32.4s\n" + "prfm pstl1keep, [ vptr2 , %[prftch]]\n" + "fmla vV34.4s, vU54.4s, vW31.4s\n" + "str qV34, [vptr2, %x[uvw_col_stride3]]\n" + "fmla vV42.4s, vU54.4s, vW23.4s\n" + "prfm pstl1keep, [ vptr2 , %x[prftch_uvw_col_stride1]]\n" + "fmla vV43.4s, vU54.4s, vW22.4s\n" + "prfm pstl1keep, [ vptr2 , %x[prftch_uvw_col_stride2] ]\n" + "fmla vV44.4s, vU54.4s, vW21.4s\n" + "ldr qU53, [uptr4, %x[uvw_col_stride2]]\n" + "fmla vV44.4s, vU66.4s, vW33.4s\n" + "ldr qU63, [uptr5, %x[uvw_col_stride2]]\n" + "fmla vV43.4s, vU65.4s, vW33.4s\n" + "prfm pstl1keep, [ vptr2 , %x[prftch_uvw_col_stride3] ]\n" + "fmla vV44.4s, vU65.4s, vW32.4s\n" + "ldr qU12, [%x[uptr0], %x[uvw_col_stride1]]\n" + "fmla vV42.4s, vU64.4s, vW33.4s\n" + "prfm pstl1keep, [ vptr3 , %[prftch]]\n" + "fmla vV43.4s, vU64.4s, vW32.4s\n" + "prfm pstl1keep, [ vptr3 , %x[prftch_uvw_col_stride1]]\n" + "fmla vV44.4s, vU64.4s, vW31.4s\n" + "str qV44, [vptr3, %x[uvw_col_stride3]]\n" + "fmul vV11.4s, vU13.4s, vW13.4s\n" + "ldr qU22, [uptr1, %x[uvw_col_stride1]]\n" + "fmla vV12.4s, vU13.4s, vW12.4s\n" + "prfm pstl1keep, [ vptr3 , %x[prftch_uvw_col_stride2] ]\n" + "fmla vV13.4s, vU13.4s, vW11.4s\n" + "ldr qU32, [uptr2, %x[uvw_col_stride1]]\n" + "fmla vV11.4s, vU23.4s, vW23.4s\n" + "prfm pstl1keep, [ vptr3 , %x[prftch_uvw_col_stride3] ]\n" + "fmla vV12.4s, vU23.4s, vW22.4s\n" + "fmla vV13.4s, vU23.4s, vW21.4s\n" + "fmul vV21.4s, vU23.4s, vW13.4s\n" + "fmla vV22.4s, vU23.4s, vW12.4s\n" + "fmla vV23.4s, vU23.4s, vW11.4s\n" + "ldr qU42, [uptr3, %x[uvw_col_stride1]]\n" + "fmla vV11.4s, vU33.4s, vW33.4s\n" + "fmla vV12.4s, vU33.4s, vW32.4s\n" + "fmla vV13.4s, vU33.4s, vW31.4s\n" + "str qV13, [%x[vptr0], %x[uvw_col_stride2]]\n" + "fmla vV21.4s, vU33.4s, vW23.4s\n" + "fmla vV22.4s, vU33.4s, vW22.4s\n" + "fmla vV23.4s, vU33.4s, vW21.4s\n" + "fmul vV31.4s, vU33.4s, vW13.4s\n" + "fmla vV32.4s, vU33.4s, vW12.4s\n" + "fmla vV33.4s, vU33.4s, vW11.4s\n" + "ldr qU52, [uptr4, %x[uvw_col_stride1]]\n" + "fmla vV21.4s, vU43.4s, vW33.4s\n" + "fmla vV22.4s, vU43.4s, vW32.4s\n" + "fmla vV23.4s, vU43.4s, vW31.4s\n" + "str qV23, [vptr1, %x[uvw_col_stride2]]\n" + "fmla vV31.4s, vU43.4s, vW23.4s\n" + "fmla vV32.4s, vU43.4s, vW22.4s\n" + "fmla vV33.4s, vU43.4s, vW21.4s\n" + "fmul vV41.4s, vU43.4s, vW13.4s\n" + "ldr qW13, [%x[wptr0], %x[uvw_col_stride2]]\n" + "fmla vV42.4s, vU43.4s, vW12.4s\n" + "fmla vV43.4s, vU43.4s, vW11.4s\n" + "ldr qU62, [uptr5, %x[uvw_col_stride1]]\n" + "fmla vV31.4s, vU53.4s, vW33.4s\n" + "fmla vV32.4s, vU53.4s, vW32.4s\n" + "fmla vV33.4s, vU53.4s, vW31.4s\n" + "str qV33, [vptr2, %x[uvw_col_stride2]]\n" + "fmla vV41.4s, vU53.4s, vW23.4s\n" + "ldr qW23, [wptr1, %x[uvw_col_stride2]]\n" + "fmla vV42.4s, vU53.4s, vW22.4s\n" + "fmla vV43.4s, vU53.4s, vW21.4s\n" + "ldr qU11, [%x[uptr0]], #0x10\n" + "fmla vV41.4s, vU63.4s, vW33.4s\n" + "ldr qW33, [wptr2, %x[uvw_col_stride2]]\n" + "fmla vV42.4s, vU63.4s, vW32.4s\n" + "prfm pldl1keep, [%x[uptr0], %[prftch]]\n" + "fmla vV43.4s, vU63.4s, vW31.4s\n" + "str qV43, [vptr3, %x[uvw_col_stride2]]\n" + "fmla vV11.4s, vU12.4s, vW12.4s\n" + "ldr qU21, [uptr1], #0x10\n" + "fmla vV12.4s, vU12.4s, vW11.4s\n" + "ldr qU31, [uptr2], #0x10\n" + "fmla vV11.4s, vU22.4s, vW22.4s\n" + "prfm pldl1keep, [%x[uptr0], %x[prftch_uvw_col_stride1]]\n" + "fmla vV12.4s, vU22.4s, vW21.4s\n" + "prfm pldl1keep, [%x[uptr0], %x[prftch_uvw_col_stride2] ]\n" + "fmla vV21.4s, vU22.4s, vW12.4s\n" + "prfm pldl1keep, [%x[uptr0], %x[prftch_uvw_col_stride3] ]\n" + "fmla vV22.4s, vU22.4s, vW11.4s\n" + "ldr qU41, [uptr3], #0x10\n" + "fmla vV11.4s, vU32.4s, vW32.4s\n" + "prfm pldl1keep, [%x[uptr0], %x[prftch_uvw_col_stride4] ]\n" + "fmla vV12.4s, vU32.4s, vW31.4s\n" + "str qV12, [%x[vptr0], %x[uvw_col_stride1]]\n" + "fmla vV21.4s, vU32.4s, vW22.4s\n" + "prfm pldl1keep, [%x[uptr0], %x[prftch_uvw_col_stride5] ]\n" + "fmla vV22.4s, vU32.4s, vW21.4s\n" + "prfm pldl1keep, [ uptr1 , %[prftch]]\n" + "fmla vV31.4s, vU32.4s, vW12.4s\n" + "prfm pldl1keep, [ uptr1 , %x[prftch_uvw_col_stride1]]\n" + "fmla vV32.4s, vU32.4s, vW11.4s\n" + "ldr qU51, [uptr4], #0x10\n" + "fmla vV21.4s, vU42.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr1 , %x[prftch_uvw_col_stride2] ]\n" + "fmla vV22.4s, vU42.4s, vW31.4s\n" + "str qV22, [vptr1, %x[uvw_col_stride1]]\n" + "fmla vV31.4s, vU42.4s, vW22.4s\n" + "prfm pldl1keep, [ uptr1 , %x[prftch_uvw_col_stride3] ]\n" + "fmla vV32.4s, vU42.4s, vW21.4s\n" + "subs %x[c4_rem], %x[c4_rem], #1\n" + "fmla vV41.4s, vU42.4s, vW12.4s\n" + "ldr qW12, [%x[wptr0], %x[uvw_col_stride1]]\n" + "fmla vV42.4s, vU42.4s, vW11.4s\n" + "ldr qU61, [uptr5], #0x10\n" + "fmla vV31.4s, vU52.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr1 , %x[prftch_uvw_col_stride4] ]\n" + "fmla vV32.4s, vU52.4s, vW31.4s\n" + "str qV32, [vptr2, %x[uvw_col_stride1]]\n" + "fmla vV41.4s, vU52.4s, vW22.4s\n" + "ldr qW22, [wptr1, %x[uvw_col_stride1]]\n" + "fmla vV42.4s, vU52.4s, vW21.4s\n" + "ldr qU16, [%x[uptr0], %x[uvw_col_stride5]]\n" + "fmla vV41.4s, vU62.4s, vW32.4s\n" + "ldr qW32, [wptr2, %x[uvw_col_stride1]]\n" + "fmla vV42.4s, vU62.4s, vW31.4s\n" + "str qV42, [vptr3, %x[uvw_col_stride1]]\n" + "fmla vV11.4s, vU11.4s, vW11.4s\n" + "ldr qU15, [%x[uptr0], %x[uvw_col_stride4]]\n" + "fmla vV11.4s, vU21.4s, vW21.4s\n" + "ldr qU14, [%x[uptr0], %x[uvw_col_stride3]]\n" + "fmla vV21.4s, vU21.4s, vW11.4s\n" + "ldr qU26, [uptr1, %x[uvw_col_stride5]]\n" + "fmla vV11.4s, vU31.4s, vW31.4s\n" + "str qV11, [%x[vptr0]], #0x10\n" + "fmla vV21.4s, vU31.4s, vW21.4s\n" + "prfm pldl1keep, [ uptr1 , %x[prftch_uvw_col_stride5] ]\n" + "fmla vV31.4s, vU31.4s, vW11.4s\n" + "ldr qU25, [uptr1, %x[uvw_col_stride4]]\n" + "fmla vV21.4s, vU41.4s, vW31.4s\n" + "str qV21, [vptr1], #0x10\n" + "fmla vV31.4s, vU41.4s, vW21.4s\n" + "prfm pldl1keep, [ uptr2 , %[prftch]]\n" + "fmla vV41.4s, vU41.4s, vW11.4s\n" + "ldr qW11, [%x[wptr0]], #0x10\n" + "fmla vV31.4s, vU51.4s, vW31.4s\n" + "str qV31, [vptr2], #0x10\n" + "fmla vV41.4s, vU51.4s, vW21.4s\n" + "ldr qU36, [uptr2, %x[uvw_col_stride5]]\n" + "fmla vV41.4s, vU61.4s, vW31.4s\n" + "str qV41, [vptr3], #0x10\n" + "fmul vV14.4s, vU16.4s, vW13.4s\n" + "ldr qU24, [uptr1, %x[uvw_col_stride3]]\n" + "fmul vV13.4s, vU15.4s, vW13.4s\n" + "ldr qW31, [wptr2], #0x10\n" + "fmla vV14.4s, vU15.4s, vW12.4s\n" + "ldr qW21, [wptr1], #0x10\n" + "fmul vV12.4s, vU14.4s, vW13.4s\n" + "ldr qU34, [uptr2, %x[uvw_col_stride3]]\n" + "fmla vV13.4s, vU14.4s, vW12.4s\n" + "ldr qU46, [uptr3, %x[uvw_col_stride5]]\n" + "fmla vV14.4s, vU14.4s, vW11.4s\n" + "ldr qU45, [uptr3, %x[uvw_col_stride4]]\n" + "fmla vV14.4s, vU26.4s, vW23.4s\n" + "ldr qU35, [uptr2, %x[uvw_col_stride4]]\n" + "fmul vV24.4s, vU26.4s, vW13.4s\n" + "ldr qU44, [uptr3, %x[uvw_col_stride3]]\n" + "fmla vV13.4s, vU25.4s, vW23.4s\n" + "bne 1b\n" + + "2:" // Final iteration + "fmla vV14.4s, vU25.4s, vW22.4s\n" + "fmul vV23.4s, vU25.4s, vW13.4s\n" + "fmla vV24.4s, vU25.4s, vW12.4s\n" + "ldr qU56, [uptr4, %x[uvw_col_stride5]]\n" + "fmla vV12.4s, vU24.4s, vW23.4s\n" + "fmla vV13.4s, vU24.4s, vW22.4s\n" + "fmla vV14.4s, vU24.4s, vW21.4s\n" + "fmul vV22.4s, vU24.4s, vW13.4s\n" + "fmla vV23.4s, vU24.4s, vW12.4s\n" + "fmla vV24.4s, vU24.4s, vW11.4s\n" + "ldr qU55, [uptr4, %x[uvw_col_stride4]]\n" + "fmla vV14.4s, vU36.4s, vW33.4s\n" + "fmla vV24.4s, vU36.4s, vW23.4s\n" + "fmul vV34.4s, vU36.4s, vW13.4s\n" + "ldr qU54, [uptr4, %x[uvw_col_stride3]]\n" + "fmla vV13.4s, vU35.4s, vW33.4s\n" + "fmla vV14.4s, vU35.4s, vW32.4s\n" + "fmla vV23.4s, vU35.4s, vW23.4s\n" + "fmla vV24.4s, vU35.4s, vW22.4s\n" + "fmul vV33.4s, vU35.4s, vW13.4s\n" + "fmla vV34.4s, vU35.4s, vW12.4s\n" + "ldr qU66, [uptr5, %x[uvw_col_stride5]]\n" + "fmla vV12.4s, vU34.4s, vW33.4s\n" + "fmla vV13.4s, vU34.4s, vW32.4s\n" + "fmla vV14.4s, vU34.4s, vW31.4s\n" + "str qV14, [%x[vptr0], %x[uvw_col_stride3]]\n" + "fmla vV22.4s, vU34.4s, vW23.4s\n" + "fmla vV23.4s, vU34.4s, vW22.4s\n" + "fmla vV24.4s, vU34.4s, vW21.4s\n" + "fmul vV32.4s, vU34.4s, vW13.4s\n" + "fmla vV33.4s, vU34.4s, vW12.4s\n" + "fmla vV34.4s, vU34.4s, vW11.4s\n" + "ldr qU65, [uptr5, %x[uvw_col_stride4]]\n" + "fmla vV24.4s, vU46.4s, vW33.4s\n" + "fmla vV34.4s, vU46.4s, vW23.4s\n" + "fmul vV44.4s, vU46.4s, vW13.4s\n" + "ldr qU64, [uptr5, %x[uvw_col_stride3]]\n" + "fmla vV23.4s, vU45.4s, vW33.4s\n" + "fmla vV24.4s, vU45.4s, vW32.4s\n" + "fmla vV33.4s, vU45.4s, vW23.4s\n" + "fmla vV34.4s, vU45.4s, vW22.4s\n" + "fmul vV43.4s, vU45.4s, vW13.4s\n" + "fmla vV44.4s, vU45.4s, vW12.4s\n" + "ldr qU13, [%x[uptr0], %x[uvw_col_stride2]]\n" + "fmla vV22.4s, vU44.4s, vW33.4s\n" + "fmla vV23.4s, vU44.4s, vW32.4s\n" + "fmla vV24.4s, vU44.4s, vW31.4s\n" + "str qV24, [vptr1, %x[uvw_col_stride3]]\n" + "fmla vV32.4s, vU44.4s, vW23.4s\n" + "fmla vV33.4s, vU44.4s, vW22.4s\n" + "fmla vV34.4s, vU44.4s, vW21.4s\n" + "fmul vV42.4s, vU44.4s, vW13.4s\n" + "fmla vV43.4s, vU44.4s, vW12.4s\n" + "fmla vV44.4s, vU44.4s, vW11.4s\n" + "ldr qU23, [uptr1, %x[uvw_col_stride2]]\n" + "fmla vV34.4s, vU56.4s, vW33.4s\n" + "fmla vV44.4s, vU56.4s, vW23.4s\n" + "ldr qU33, [uptr2, %x[uvw_col_stride2]]\n" + "fmla vV33.4s, vU55.4s, vW33.4s\n" + "fmla vV34.4s, vU55.4s, vW32.4s\n" + "fmla vV43.4s, vU55.4s, vW23.4s\n" + "fmla vV44.4s, vU55.4s, vW22.4s\n" + "ldr qU43, [uptr3, %x[uvw_col_stride2]]\n" + "fmla vV32.4s, vU54.4s, vW33.4s\n" + "fmla vV33.4s, vU54.4s, vW32.4s\n" + "fmla vV34.4s, vU54.4s, vW31.4s\n" + "str qV34, [vptr2, %x[uvw_col_stride3]]\n" + "fmla vV42.4s, vU54.4s, vW23.4s\n" + "fmla vV43.4s, vU54.4s, vW22.4s\n" + "fmla vV44.4s, vU54.4s, vW21.4s\n" + "ldr qU53, [uptr4, %x[uvw_col_stride2]]\n" + "fmla vV44.4s, vU66.4s, vW33.4s\n" + "ldr qU63, [uptr5, %x[uvw_col_stride2]]\n" + "fmla vV43.4s, vU65.4s, vW33.4s\n" + "fmla vV44.4s, vU65.4s, vW32.4s\n" + "ldr qU12, [%x[uptr0], %x[uvw_col_stride1]]\n" + "fmla vV42.4s, vU64.4s, vW33.4s\n" + "fmla vV43.4s, vU64.4s, vW32.4s\n" + "fmla vV44.4s, vU64.4s, vW31.4s\n" + "str qV44, [vptr3, %x[uvw_col_stride3]]\n" + "fmul vV11.4s, vU13.4s, vW13.4s\n" + "ldr qU22, [uptr1, %x[uvw_col_stride1]]\n" + "fmla vV12.4s, vU13.4s, vW12.4s\n" + "fmla vV13.4s, vU13.4s, vW11.4s\n" + "ldr qU32, [uptr2, %x[uvw_col_stride1]]\n" + "fmla vV11.4s, vU23.4s, vW23.4s\n" + "fmla vV12.4s, vU23.4s, vW22.4s\n" + "fmla vV13.4s, vU23.4s, vW21.4s\n" + "fmul vV21.4s, vU23.4s, vW13.4s\n" + "fmla vV22.4s, vU23.4s, vW12.4s\n" + "fmla vV23.4s, vU23.4s, vW11.4s\n" + "ldr qU42, [uptr3, %x[uvw_col_stride1]]\n" + "fmla vV11.4s, vU33.4s, vW33.4s\n" + "fmla vV12.4s, vU33.4s, vW32.4s\n" + "fmla vV13.4s, vU33.4s, vW31.4s\n" + "str qV13, [%x[vptr0], %x[uvw_col_stride2]]\n" + "fmla vV21.4s, vU33.4s, vW23.4s\n" + "fmla vV22.4s, vU33.4s, vW22.4s\n" + "fmla vV23.4s, vU33.4s, vW21.4s\n" + "fmul vV31.4s, vU33.4s, vW13.4s\n" + "fmla vV32.4s, vU33.4s, vW12.4s\n" + "fmla vV33.4s, vU33.4s, vW11.4s\n" + "ldr qU52, [uptr4, %x[uvw_col_stride1]]\n" + "fmla vV21.4s, vU43.4s, vW33.4s\n" + "fmla vV22.4s, vU43.4s, vW32.4s\n" + "fmla vV23.4s, vU43.4s, vW31.4s\n" + "str qV23, [vptr1, %x[uvw_col_stride2]]\n" + "fmla vV31.4s, vU43.4s, vW23.4s\n" + "fmla vV32.4s, vU43.4s, vW22.4s\n" + "fmla vV33.4s, vU43.4s, vW21.4s\n" + "fmul vV41.4s, vU43.4s, vW13.4s\n" + "fmla vV42.4s, vU43.4s, vW12.4s\n" + "fmla vV43.4s, vU43.4s, vW11.4s\n" + "ldr qU62, [uptr5, %x[uvw_col_stride1]]\n" + "fmla vV31.4s, vU53.4s, vW33.4s\n" + "fmla vV32.4s, vU53.4s, vW32.4s\n" + "fmla vV33.4s, vU53.4s, vW31.4s\n" + "str qV33, [vptr2, %x[uvw_col_stride2]]\n" + "fmla vV41.4s, vU53.4s, vW23.4s\n" + "fmla vV42.4s, vU53.4s, vW22.4s\n" + "fmla vV43.4s, vU53.4s, vW21.4s\n" + "ldr qU11, [%x[uptr0]], #0x10\n" + "fmla vV41.4s, vU63.4s, vW33.4s\n" + "fmla vV42.4s, vU63.4s, vW32.4s\n" + "fmla vV43.4s, vU63.4s, vW31.4s\n" + "str qV43, [vptr3, %x[uvw_col_stride2]]\n" + "fmla vV11.4s, vU12.4s, vW12.4s\n" + "ldr qU21, [uptr1], #0x10\n" + "fmla vV12.4s, vU12.4s, vW11.4s\n" + "ldr qU31, [uptr2], #0x10\n" + "fmla vV11.4s, vU22.4s, vW22.4s\n" + "fmla vV12.4s, vU22.4s, vW21.4s\n" + "fmla vV21.4s, vU22.4s, vW12.4s\n" + "fmla vV22.4s, vU22.4s, vW11.4s\n" + "ldr qU41, [uptr3], #0x10\n" + "fmla vV11.4s, vU32.4s, vW32.4s\n" + "fmla vV12.4s, vU32.4s, vW31.4s\n" + "str qV12, [%x[vptr0], %x[uvw_col_stride1]]\n" + "fmla vV21.4s, vU32.4s, vW22.4s\n" + "fmla vV22.4s, vU32.4s, vW21.4s\n" + "fmla vV31.4s, vU32.4s, vW12.4s\n" + "fmla vV32.4s, vU32.4s, vW11.4s\n" + "ldr qU51, [uptr4], #0x10\n" + "fmla vV21.4s, vU42.4s, vW32.4s\n" + "fmla vV22.4s, vU42.4s, vW31.4s\n" + "str qV22, [vptr1, %x[uvw_col_stride1]]\n" + "fmla vV31.4s, vU42.4s, vW22.4s\n" + "fmla vV32.4s, vU42.4s, vW21.4s\n" + "subs %x[c4_rem], %x[c4_rem], #1\n" + "fmla vV41.4s, vU42.4s, vW12.4s\n" + "fmla vV42.4s, vU42.4s, vW11.4s\n" + "ldr qU61, [uptr5], #0x10\n" + "fmla vV31.4s, vU52.4s, vW32.4s\n" + "fmla vV32.4s, vU52.4s, vW31.4s\n" + "str qV32, [vptr2, %x[uvw_col_stride1]]\n" + "fmla vV41.4s, vU52.4s, vW22.4s\n" + "fmla vV42.4s, vU52.4s, vW21.4s\n" + "fmla vV41.4s, vU62.4s, vW32.4s\n" + "fmla vV42.4s, vU62.4s, vW31.4s\n" + "str qV42, [vptr3, %x[uvw_col_stride1]]\n" + "fmla vV11.4s, vU11.4s, vW11.4s\n" + "fmla vV11.4s, vU21.4s, vW21.4s\n" + "fmla vV21.4s, vU21.4s, vW11.4s\n" + "fmla vV11.4s, vU31.4s, vW31.4s\n" + "str qV11, [%x[vptr0]], #0x10\n" + "fmla vV21.4s, vU31.4s, vW21.4s\n" + "fmla vV31.4s, vU31.4s, vW11.4s\n" + "fmla vV21.4s, vU41.4s, vW31.4s\n" + "str qV21, [vptr1], #0x10\n" + "fmla vV31.4s, vU41.4s, vW21.4s\n" + "fmla vV41.4s, vU41.4s, vW11.4s\n" + "fmla vV31.4s, vU51.4s, vW31.4s\n" + "str qV31, [vptr2], #0x10\n" + "fmla vV41.4s, vU51.4s, vW21.4s\n" + "fmla vV41.4s, vU61.4s, vW31.4s\n" + "str qV41, [vptr3], #0x10\n" + + ".unreq qW22\n" ".unreq qU64\n" ".unreq qU35\n" ".unreq qV41\n" + ".unreq qU34\n" ".unreq qU21\n" ".unreq qV43\n" ".unreq qW21\n" + ".unreq qU24\n" ".unreq qU54\n" ".unreq qV31\n" ".unreq qV12\n" + ".unreq qU61\n" ".unreq qU26\n" ".unreq qV32\n" + ".unreq qU36\n" ".unreq qU51\n" ".unreq qU66\n" ".unreq qU12\n" + ".unreq qV14\n" ".unreq qV11\n" ".unreq qU65\n" + ".unreq qU15\n" ".unreq qU22\n" ".unreq qU45\n" + ".unreq qV22\n" ".unreq qU14\n" + ".unreq qU44\n" ".unreq qU43\n" ".unreq qU11\n" + ".unreq qV24\n" ".unreq qV42\n" ".unreq qW31\n" ".unreq qW13\n" + ".unreq qU33\n" ".unreq qU62\n" ".unreq qU25\n" ".unreq qU56\n" + ".unreq qW33\n" + ".unreq qU42\n" ".unreq qU16\n" ".unreq qV44\n" + ".unreq qU63\n" ".unreq qU31\n" ".unreq qV34\n" + ".unreq qW11\n" ".unreq qU41\n" ".unreq qV13\n" ".unreq qV33\n" + ".unreq qU46\n" ".unreq qU32\n" ".unreq qU13\n" + ".unreq qW23\n" ".unreq qV23\n" ".unreq qV21\n" ".unreq qU55\n" + ".unreq qW12\n" ".unreq qW32\n" ".unreq qU23\n" ".unreq qU52\n" + ".unreq qU53\n" ".unreq vW22\n" + ".unreq vU64\n" ".unreq vU35\n" ".unreq vV41\n" + ".unreq vU34\n" ".unreq vU21\n" ".unreq vV43\n" ".unreq vW21\n" + ".unreq vU24\n" ".unreq vU54\n" ".unreq vV31\n" + ".unreq vV12\n" ".unreq vU61\n" + ".unreq vU26\n" ".unreq vV32\n" + ".unreq vU36\n" ".unreq vU51\n" ".unreq vU66\n" ".unreq vU12\n" + ".unreq vV14\n" ".unreq vV11\n" ".unreq vU65\n" + ".unreq vU15\n" ".unreq vU22\n" ".unreq vU45\n" + ".unreq vV22\n" ".unreq vU14\n" + ".unreq vU44\n" ".unreq vU43\n" ".unreq vU11\n" + ".unreq vV24\n" ".unreq vV42\n" ".unreq vW31\n" ".unreq vW13\n" + ".unreq vU33\n" ".unreq vU62\n" ".unreq vU25\n" ".unreq vU56\n" + ".unreq vW33\n" ".unreq vU42\n" ".unreq vU16\n" ".unreq vV44\n" + ".unreq vU63\n" ".unreq vU31\n" ".unreq vV34\n" ".unreq vW11\n" + ".unreq vU41\n" ".unreq vV13\n" ".unreq vV33\n" + ".unreq vU46\n" ".unreq vU32\n" ".unreq vU13\n" ".unreq vW23\n" + ".unreq vV23\n" ".unreq vV21\n" ".unreq vU55\n" ".unreq vW12\n" + ".unreq vW32\n" ".unreq vU23\n" ".unreq vU52\n" ".unreq vU53\n" + : [uptr0] "+r" (uptr0), [vptr0] "+r" (vptr0), [wptr0] "+r" (wptr0), + [c4_rem] "+r" (c4_rem) + : [u_row_stride] "r" (in_row_stride * sizeof(float)), + [v_row_stride] "r" (out_row_stride * sizeof(float)), + [w_row_stride] "r" (weight_row_stride * sizeof(float)), + [uvw_col_stride1] "r" (1 * in_col_stride * sizeof(float)), + [uvw_col_stride2] "r" (2 * in_col_stride * sizeof(float)), + [uvw_col_stride3] "r" (3 * in_col_stride * sizeof(float)), + [uvw_col_stride4] "r" (4 * in_col_stride * sizeof(float)), + [uvw_col_stride5] "r" (5 * in_col_stride * sizeof(float)), + [prftch] "i" (prefetch_depth * sizeof(float)), + [prftch_uvw_col_stride1] "r" ((prefetch_depth + 1 * in_col_stride) * sizeof(float)), + [prftch_uvw_col_stride2] "r" ((prefetch_depth + 2 * in_col_stride) * sizeof(float)), + [prftch_uvw_col_stride3] "r" ((prefetch_depth + 3 * in_col_stride) * sizeof(float)), + [prftch_uvw_col_stride4] "r" ((prefetch_depth + 4 * in_col_stride) * sizeof(float)), + [prftch_uvw_col_stride5] "r" ((prefetch_depth + 5 * in_col_stride) * sizeof(float)) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", + "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "x0", + "x1", "x2", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "cc", "memory" + ); + } + else if (channels_remaining >= 4) + { + int c4_rem = channels_remaining / 4; + channels_remaining %= 4; + + asm volatile ( + "qW22 .req q0\n" "vW22 .req v0\n" + "qU64 .req q1\n" "qU35 .req q1\n" "qV41 .req q1\n" + "vU64 .req v1\n" "vU35 .req v1\n" "vV41 .req v1\n" + "qU34 .req q2\n" "qU21 .req q2\n" "qV43 .req q2\n" + "vU34 .req v2\n" "vU21 .req v2\n" "vV43 .req v2\n" + "qW21 .req q3\n" "vW21 .req v3\n" + "qU24 .req q4\n" "qU54 .req q4\n" "qV31 .req q4\n" + "vU24 .req v4\n" "vU54 .req v4\n" "vV31 .req v4\n" + "qV12 .req q5\n" "qU61 .req q5\n" "vV12 .req v5\n" "vU61 .req v5\n" + "qU26 .req q6\n" "qV32 .req q6\n" "vU26 .req v6\n" "vV32 .req v6\n" + "qU36 .req q7\n" "qU51 .req q7\n" "qU66 .req q7\n" "qU12 .req q7\n" + "vU36 .req v7\n" "vU51 .req v7\n" "vU66 .req v7\n" "vU12 .req v7\n" + "qV14 .req q8\n" "qV11 .req q8\n" "qU65 .req q8\n" + "vV14 .req v8\n" "vV11 .req v8\n" "vU65 .req v8\n" + "qU15 .req q9\n" "qU22 .req q9\n" "qU45 .req q9\n" + "vU15 .req v9\n" "vU22 .req v9\n" "vU45 .req v9\n" + "qV22 .req q10\n" "qU14 .req q10\n" "vV22 .req v10\n" "vU14 .req v10\n" + "qU44 .req q11\n" "qU43 .req q11\n" "qU11 .req q11\n" + "vU44 .req v11\n" "vU43 .req v11\n" "vU11 .req v11\n" + "qV24 .req q12\n" "qV42 .req q12\n" "vV24 .req v12\n" "vV42 .req v12\n" + "qW31 .req q13\n" "vW31 .req v13\n" "qW13 .req q14\n" "vW13 .req v14\n" + "qU33 .req q15\n" "qU62 .req q15\n" "qU25 .req q15\n" "qU56 .req q15\n" + "vU33 .req v15\n" "vU62 .req v15\n" "vU25 .req v15\n" "vU56 .req v15\n" + "qW33 .req q16\n" "vW33 .req v16\n" + "qU42 .req q17\n" "qU16 .req q17\n" "qV44 .req q17\n" + "vU42 .req v17\n" "vU16 .req v17\n" "vV44 .req v17\n" + "qU63 .req q18\n" "qU31 .req q18\n" "qV34 .req q18\n" + "vU63 .req v18\n" "vU31 .req v18\n" "vV34 .req v18\n" + "qW11 .req q19\n" "vW11 .req v19\n" "qU41 .req q20\n" "qV13 .req q20\n" + "vU41 .req v20\n" "vV13 .req v20\n" "qV33 .req q21\n" "vV33 .req v21\n" + "qU46 .req q22\n" "qU32 .req q22\n" "qU13 .req q22\n" + "vU46 .req v22\n" "vU32 .req v22\n" "vU13 .req v22\n" "qW23 .req q23\n" + "vW23 .req v23\n" "qV23 .req q24\n" "vV23 .req v24\n" + "qV21 .req q25\n" "qU55 .req q25\n" "vV21 .req v25\n" "vU55 .req v25\n" + "qW12 .req q26\n" "vW12 .req v26\n" "qW32 .req q27\n" "vW32 .req v27\n" + "qU23 .req q28\n" "qU52 .req q28\n" + "vU23 .req v28\n" "vU52 .req v28\n" "qU53 .req q29\n" "vU53 .req v29\n" + + "uptr1 .req x0\n" + "uptr2 .req x1\n" + "uptr3 .req x2\n" + "uptr4 .req x3\n" + "uptr5 .req x4\n" + + "vptr1 .req x5\n" + "vptr2 .req x6\n" + "vptr3 .req x7\n" + + "wptr1 .req x8\n" + "wptr2 .req x9\n" + + "u_col_stride2 .req x10\n" + "u_col_stride3 .req x11\n" + "u_col_stride4 .req x12\n" + "u_col_stride5 .req x13\n" + + "v_col_stride2 .req x14\n" + "v_col_stride3 .req x15\n" + + "w_col_stride2 .req x16\n" + + // Prepare pointers and strides + "add uptr1, %x[uptr0], %x[u_row_stride]\n" + "add uptr2, uptr1 , %x[u_row_stride]\n" + "add uptr3, uptr2 , %x[u_row_stride]\n" + "add uptr4, uptr3 , %x[u_row_stride]\n" + "add uptr5, uptr4 , %x[u_row_stride]\n" + + "add vptr1, %x[vptr0], %x[v_row_stride]\n" + "add vptr2, vptr1 , %x[v_row_stride]\n" + "add vptr3, vptr2 , %x[v_row_stride]\n" + + "add wptr1, %x[wptr0], %x[w_row_stride]\n" + "add wptr2, wptr1 , %x[w_row_stride]\n" + + "add u_col_stride2, %x[u_col_stride1], %x[u_col_stride1]\n" + "add u_col_stride3, u_col_stride2 , %x[u_col_stride1]\n" + "add u_col_stride4, u_col_stride3 , %x[u_col_stride1]\n" + "add u_col_stride5, u_col_stride4 , %x[u_col_stride1]\n" + + "add v_col_stride2, %x[v_col_stride1], %x[v_col_stride1]\n" + "add v_col_stride3, v_col_stride2 , %x[v_col_stride1]\n" + + "add w_col_stride2, %x[w_col_stride1], %x[w_col_stride1]\n" + + // Load initial operands + "ldr qU16, [%x[uptr0], u_col_stride5]\n" + "ldr qW13, [%x[wptr0], w_col_stride2]\n" + "subs %x[c4_rem], %x[c4_rem], #1\n" + "ldr qU15, [%x[uptr0], u_col_stride4]\n" + "ldr qW23, [wptr1, w_col_stride2]\n" + "ldr qU14, [%x[uptr0], u_col_stride3]\n" + "ldr qW33, [wptr2, w_col_stride2]\n" + "ldr qU26, [uptr1, u_col_stride5]\n" + "ldr qW12, [%x[wptr0], %x[w_col_stride1]]\n" + "ldr qU25, [uptr1, u_col_stride4]\n" + "ldr qW22, [wptr1, %x[w_col_stride1]]\n" + "ldr qU36, [uptr2, u_col_stride5]\n" + "ldr qW32, [wptr2, %x[w_col_stride1]]\n" + "ldr qW11, [%x[wptr0]], #0x10\n" + "fmul vV14.4s, vU16.4s, vW13.4s\n" + "ldr qU24, [uptr1, u_col_stride3]\n" + "fmul vV13.4s, vU15.4s, vW13.4s\n" + "ldr qW31, [wptr2], #0x10\n" + "fmla vV14.4s, vU15.4s, vW12.4s\n" + "ldr qW21, [wptr1], #0x10\n" + "fmul vV12.4s, vU14.4s, vW13.4s\n" + "ldr qU34, [uptr2, u_col_stride3]\n" + "fmla vV13.4s, vU14.4s, vW12.4s\n" + "ldr qU46, [uptr3, u_col_stride5]\n" + "fmla vV14.4s, vU14.4s, vW11.4s\n" + "ldr qU45, [uptr3, u_col_stride4]\n" + "fmla vV14.4s, vU26.4s, vW23.4s\n" + "ldr qU35, [uptr2, u_col_stride4]\n" + "fmul vV24.4s, vU26.4s, vW13.4s\n" + "ldr qU44, [uptr3, u_col_stride3]\n" + "fmla vV13.4s, vU25.4s, vW23.4s\n" + "beq 2f\n" // Single iteration only + + "1:" // Loop body + "fmla vV14.4s, vU25.4s, vW22.4s\n" + "prfm pldl1keep, [%x[wptr0]]\n" + "fmul vV23.4s, vU25.4s, vW13.4s\n" + "prfm pldl1keep, [%x[wptr0], %x[w_col_stride1]]\n" + "fmla vV24.4s, vU25.4s, vW12.4s\n" + "ldr qU56, [uptr4, u_col_stride5]\n" + "fmla vV12.4s, vU24.4s, vW23.4s\n" + "prfm pldl1keep, [%x[wptr0], w_col_stride2 ]\n" + "fmla vV13.4s, vU24.4s, vW22.4s\n" + "prfm pldl1keep, [ wptr1 ]\n" + "fmla vV14.4s, vU24.4s, vW21.4s\n" + "prfm pldl1keep, [ wptr1 , %x[w_col_stride1]]\n" + "fmul vV22.4s, vU24.4s, vW13.4s\n" + "prfm pldl1keep, [ wptr1 , w_col_stride2 ]\n" + "fmla vV23.4s, vU24.4s, vW12.4s\n" + "prfm pldl1keep, [ wptr2 ]\n" + "fmla vV24.4s, vU24.4s, vW11.4s\n" + "ldr qU55, [uptr4, u_col_stride4]\n" + "fmla vV14.4s, vU36.4s, vW33.4s\n" + "prfm pldl1keep, [ wptr2 , %x[w_col_stride1]]\n" + "fmla vV24.4s, vU36.4s, vW23.4s\n" + "prfm pldl1keep, [ wptr2 , w_col_stride2 ]\n" + "fmul vV34.4s, vU36.4s, vW13.4s\n" + "ldr qU54, [uptr4, u_col_stride3]\n" + "fmla vV13.4s, vU35.4s, vW33.4s\n" + "prfm pldl1keep, [ uptr2 , %x[u_col_stride1]]\n" + "fmla vV14.4s, vU35.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr2 , u_col_stride2 ]\n" + "fmla vV23.4s, vU35.4s, vW23.4s\n" + "prfm pldl1keep, [ uptr2 , u_col_stride3 ]\n" + "fmla vV24.4s, vU35.4s, vW22.4s\n" + "prfm pldl1keep, [ uptr2 , u_col_stride4 ]\n" + "fmul vV33.4s, vU35.4s, vW13.4s\n" + "prfm pldl1keep, [ uptr2 , u_col_stride5 ]\n" + "fmla vV34.4s, vU35.4s, vW12.4s\n" + "ldr qU66, [uptr5, u_col_stride5]\n" + "fmla vV12.4s, vU34.4s, vW33.4s\n" + "prfm pldl1keep, [ uptr3 ]\n" + "fmla vV13.4s, vU34.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr3 , %x[u_col_stride1]]\n" + "fmla vV14.4s, vU34.4s, vW31.4s\n" + "str qV14, [%x[vptr0], v_col_stride3]\n" + "fmla vV22.4s, vU34.4s, vW23.4s\n" + "prfm pldl1keep, [ uptr3 , u_col_stride2 ]\n" + "fmla vV23.4s, vU34.4s, vW22.4s\n" + "prfm pldl1keep, [ uptr3 , u_col_stride3 ]\n" + "fmla vV24.4s, vU34.4s, vW21.4s\n" + "prfm pldl1keep, [ uptr3 , u_col_stride4 ]\n" + "fmul vV32.4s, vU34.4s, vW13.4s\n" + "prfm pldl1keep, [ uptr3 , u_col_stride5 ]\n" + "fmla vV33.4s, vU34.4s, vW12.4s\n" + "prfm pldl1keep, [ uptr4 ]\n" + "fmla vV34.4s, vU34.4s, vW11.4s\n" + "ldr qU65, [uptr5, u_col_stride4]\n" + "fmla vV24.4s, vU46.4s, vW33.4s\n" + "prfm pldl1keep, [ uptr4 , %x[u_col_stride1]]\n" + "fmla vV34.4s, vU46.4s, vW23.4s\n" + "prfm pldl1keep, [ uptr4 , u_col_stride2 ]\n" + "fmul vV44.4s, vU46.4s, vW13.4s\n" + "ldr qU64, [uptr5, u_col_stride3]\n" + "fmla vV23.4s, vU45.4s, vW33.4s\n" + "prfm pldl1keep, [ uptr4 , u_col_stride3 ]\n" + "fmla vV24.4s, vU45.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr4 , u_col_stride4 ]\n" + "fmla vV33.4s, vU45.4s, vW23.4s\n" + "prfm pldl1keep, [ uptr4 , u_col_stride5 ]\n" + "fmla vV34.4s, vU45.4s, vW22.4s\n" + "prfm pldl1keep, [ uptr5 ]\n" + "fmul vV43.4s, vU45.4s, vW13.4s\n" + "prfm pldl1keep, [ uptr5 , %x[u_col_stride1]]\n" + "fmla vV44.4s, vU45.4s, vW12.4s\n" + "ldr qU13, [%x[uptr0], u_col_stride2]\n" + "fmla vV22.4s, vU44.4s, vW33.4s\n" + "prfm pldl1keep, [ uptr5 , u_col_stride2 ]\n" + "fmla vV23.4s, vU44.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr5 , u_col_stride3 ]\n" + "fmla vV24.4s, vU44.4s, vW31.4s\n" + "str qV24, [vptr1, v_col_stride3]\n" + "fmla vV32.4s, vU44.4s, vW23.4s\n" + "prfm pldl1keep, [ uptr5 , u_col_stride4 ]\n" + "fmla vV33.4s, vU44.4s, vW22.4s\n" + "prfm pldl1keep, [ uptr5 , u_col_stride5 ]\n" + "fmla vV34.4s, vU44.4s, vW21.4s\n" + "prfm pstl1keep, [%x[vptr0]]\n" + "fmul vV42.4s, vU44.4s, vW13.4s\n" + "prfm pstl1keep, [%x[vptr0], %x[v_col_stride1]]\n" + "fmla vV43.4s, vU44.4s, vW12.4s\n" + "prfm pstl1keep, [%x[vptr0], v_col_stride2 ]\n" + "fmla vV44.4s, vU44.4s, vW11.4s\n" + "ldr qU23, [uptr1, u_col_stride2]\n" + "fmla vV34.4s, vU56.4s, vW33.4s\n" + "prfm pstl1keep, [%x[vptr0], v_col_stride3 ]\n" + "fmla vV44.4s, vU56.4s, vW23.4s\n" + "ldr qU33, [uptr2, u_col_stride2]\n" + "fmla vV33.4s, vU55.4s, vW33.4s\n" + "prfm pstl1keep, [ vptr1 ]\n" + "fmla vV34.4s, vU55.4s, vW32.4s\n" + "prfm pstl1keep, [ vptr1 , %x[v_col_stride1]]\n" + "fmla vV43.4s, vU55.4s, vW23.4s\n" + "prfm pstl1keep, [ vptr1 , v_col_stride2 ]\n" + "fmla vV44.4s, vU55.4s, vW22.4s\n" + "ldr qU43, [uptr3, u_col_stride2]\n" + "fmla vV32.4s, vU54.4s, vW33.4s\n" + "prfm pstl1keep, [ vptr1 , v_col_stride3 ]\n" + "fmla vV33.4s, vU54.4s, vW32.4s\n" + "prfm pstl1keep, [ vptr2 ]\n" + "fmla vV34.4s, vU54.4s, vW31.4s\n" + "str qV34, [vptr2, v_col_stride3]\n" + "fmla vV42.4s, vU54.4s, vW23.4s\n" + "prfm pstl1keep, [ vptr2 , %x[v_col_stride1]]\n" + "fmla vV43.4s, vU54.4s, vW22.4s\n" + "prfm pstl1keep, [ vptr2 , v_col_stride2 ]\n" + "fmla vV44.4s, vU54.4s, vW21.4s\n" + "ldr qU53, [uptr4, u_col_stride2]\n" + "fmla vV44.4s, vU66.4s, vW33.4s\n" + "ldr qU63, [uptr5, u_col_stride2]\n" + "fmla vV43.4s, vU65.4s, vW33.4s\n" + "prfm pstl1keep, [ vptr2 , v_col_stride3 ]\n" + "fmla vV44.4s, vU65.4s, vW32.4s\n" + "ldr qU12, [%x[uptr0], %x[u_col_stride1]]\n" + "fmla vV42.4s, vU64.4s, vW33.4s\n" + "prfm pstl1keep, [ vptr3 ]\n" + "fmla vV43.4s, vU64.4s, vW32.4s\n" + "prfm pstl1keep, [ vptr3 , %x[v_col_stride1]]\n" + "fmla vV44.4s, vU64.4s, vW31.4s\n" + "str qV44, [vptr3, v_col_stride3]\n" + "fmul vV11.4s, vU13.4s, vW13.4s\n" + "ldr qU22, [uptr1, %x[u_col_stride1]]\n" + "fmla vV12.4s, vU13.4s, vW12.4s\n" + "prfm pstl1keep, [ vptr3 , v_col_stride2 ]\n" + "fmla vV13.4s, vU13.4s, vW11.4s\n" + "ldr qU32, [uptr2, %x[u_col_stride1]]\n" + "fmla vV11.4s, vU23.4s, vW23.4s\n" + "prfm pstl1keep, [ vptr3 , v_col_stride3 ]\n" + "fmla vV12.4s, vU23.4s, vW22.4s\n" + "fmla vV13.4s, vU23.4s, vW21.4s\n" + "fmul vV21.4s, vU23.4s, vW13.4s\n" + "fmla vV22.4s, vU23.4s, vW12.4s\n" + "fmla vV23.4s, vU23.4s, vW11.4s\n" + "ldr qU42, [uptr3, %x[u_col_stride1]]\n" + "fmla vV11.4s, vU33.4s, vW33.4s\n" + "fmla vV12.4s, vU33.4s, vW32.4s\n" + "fmla vV13.4s, vU33.4s, vW31.4s\n" + "str qV13, [%x[vptr0], v_col_stride2]\n" + "fmla vV21.4s, vU33.4s, vW23.4s\n" + "fmla vV22.4s, vU33.4s, vW22.4s\n" + "fmla vV23.4s, vU33.4s, vW21.4s\n" + "fmul vV31.4s, vU33.4s, vW13.4s\n" + "fmla vV32.4s, vU33.4s, vW12.4s\n" + "fmla vV33.4s, vU33.4s, vW11.4s\n" + "ldr qU52, [uptr4, %x[u_col_stride1]]\n" + "fmla vV21.4s, vU43.4s, vW33.4s\n" + "fmla vV22.4s, vU43.4s, vW32.4s\n" + "fmla vV23.4s, vU43.4s, vW31.4s\n" + "str qV23, [vptr1, v_col_stride2]\n" + "fmla vV31.4s, vU43.4s, vW23.4s\n" + "fmla vV32.4s, vU43.4s, vW22.4s\n" + "fmla vV33.4s, vU43.4s, vW21.4s\n" + "fmul vV41.4s, vU43.4s, vW13.4s\n" + "ldr qW13, [%x[wptr0], w_col_stride2]\n" + "fmla vV42.4s, vU43.4s, vW12.4s\n" + "fmla vV43.4s, vU43.4s, vW11.4s\n" + "ldr qU62, [uptr5, %x[u_col_stride1]]\n" + "fmla vV31.4s, vU53.4s, vW33.4s\n" + "fmla vV32.4s, vU53.4s, vW32.4s\n" + "fmla vV33.4s, vU53.4s, vW31.4s\n" + "str qV33, [vptr2, v_col_stride2]\n" + "fmla vV41.4s, vU53.4s, vW23.4s\n" + "ldr qW23, [wptr1, w_col_stride2]\n" + "fmla vV42.4s, vU53.4s, vW22.4s\n" + "fmla vV43.4s, vU53.4s, vW21.4s\n" + "ldr qU11, [%x[uptr0]], #0x10\n" + "fmla vV41.4s, vU63.4s, vW33.4s\n" + "ldr qW33, [wptr2, w_col_stride2]\n" + "fmla vV42.4s, vU63.4s, vW32.4s\n" + "prfm pldl1keep, [%x[uptr0]]\n" + "fmla vV43.4s, vU63.4s, vW31.4s\n" + "str qV43, [vptr3, v_col_stride2]\n" + "fmla vV11.4s, vU12.4s, vW12.4s\n" + "ldr qU21, [uptr1], #0x10\n" + "fmla vV12.4s, vU12.4s, vW11.4s\n" + "ldr qU31, [uptr2], #0x10\n" + "fmla vV11.4s, vU22.4s, vW22.4s\n" + "prfm pldl1keep, [%x[uptr0], %x[u_col_stride1]]\n" + "fmla vV12.4s, vU22.4s, vW21.4s\n" + "prfm pldl1keep, [%x[uptr0], u_col_stride2 ]\n" + "fmla vV21.4s, vU22.4s, vW12.4s\n" + "prfm pldl1keep, [%x[uptr0], u_col_stride3 ]\n" + "fmla vV22.4s, vU22.4s, vW11.4s\n" + "ldr qU41, [uptr3], #0x10\n" + "fmla vV11.4s, vU32.4s, vW32.4s\n" + "prfm pldl1keep, [%x[uptr0], u_col_stride4 ]\n" + "fmla vV12.4s, vU32.4s, vW31.4s\n" + "str qV12, [%x[vptr0], %x[v_col_stride1]]\n" + "fmla vV21.4s, vU32.4s, vW22.4s\n" + "prfm pldl1keep, [%x[uptr0], u_col_stride5 ]\n" + "fmla vV22.4s, vU32.4s, vW21.4s\n" + "prfm pldl1keep, [ uptr1 ]\n" + "fmla vV31.4s, vU32.4s, vW12.4s\n" + "prfm pldl1keep, [ uptr1 , %x[u_col_stride1]]\n" + "fmla vV32.4s, vU32.4s, vW11.4s\n" + "ldr qU51, [uptr4], #0x10\n" + "fmla vV21.4s, vU42.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr1 , u_col_stride2 ]\n" + "fmla vV22.4s, vU42.4s, vW31.4s\n" + "str qV22, [vptr1, %x[v_col_stride1]]\n" + "fmla vV31.4s, vU42.4s, vW22.4s\n" + "prfm pldl1keep, [ uptr1 , u_col_stride3 ]\n" + "fmla vV32.4s, vU42.4s, vW21.4s\n" + "subs %x[c4_rem], %x[c4_rem], #1\n" + "fmla vV41.4s, vU42.4s, vW12.4s\n" + "ldr qW12, [%x[wptr0], %x[w_col_stride1]]\n" + "fmla vV42.4s, vU42.4s, vW11.4s\n" + "ldr qU61, [uptr5], #0x10\n" + "fmla vV31.4s, vU52.4s, vW32.4s\n" + "prfm pldl1keep, [ uptr1 , u_col_stride4 ]\n" + "fmla vV32.4s, vU52.4s, vW31.4s\n" + "str qV32, [vptr2, %x[v_col_stride1]]\n" + "fmla vV41.4s, vU52.4s, vW22.4s\n" + "ldr qW22, [wptr1, %x[w_col_stride1]]\n" + "fmla vV42.4s, vU52.4s, vW21.4s\n" + "ldr qU16, [%x[uptr0], u_col_stride5]\n" + "fmla vV41.4s, vU62.4s, vW32.4s\n" + "ldr qW32, [wptr2, %x[w_col_stride1]]\n" + "fmla vV42.4s, vU62.4s, vW31.4s\n" + "str qV42, [vptr3, %x[v_col_stride1]]\n" + "fmla vV11.4s, vU11.4s, vW11.4s\n" + "ldr qU15, [%x[uptr0], u_col_stride4]\n" + "fmla vV11.4s, vU21.4s, vW21.4s\n" + "ldr qU14, [%x[uptr0], u_col_stride3]\n" + "fmla vV21.4s, vU21.4s, vW11.4s\n" + "ldr qU26, [uptr1, u_col_stride5]\n" + "fmla vV11.4s, vU31.4s, vW31.4s\n" + "str qV11, [%x[vptr0]], #0x10\n" + "fmla vV21.4s, vU31.4s, vW21.4s\n" + "prfm pldl1keep, [ uptr1 , u_col_stride5 ]\n" + "fmla vV31.4s, vU31.4s, vW11.4s\n" + "ldr qU25, [uptr1, u_col_stride4]\n" + "fmla vV21.4s, vU41.4s, vW31.4s\n" + "str qV21, [vptr1], #0x10\n" + "fmla vV31.4s, vU41.4s, vW21.4s\n" + "prfm pldl1keep, [ uptr2 ]\n" + "fmla vV41.4s, vU41.4s, vW11.4s\n" + "ldr qW11, [%x[wptr0]], #0x10\n" + "fmla vV31.4s, vU51.4s, vW31.4s\n" + "str qV31, [vptr2], #0x10\n" + "fmla vV41.4s, vU51.4s, vW21.4s\n" + "ldr qU36, [uptr2, u_col_stride5]\n" + "fmla vV41.4s, vU61.4s, vW31.4s\n" + "str qV41, [vptr3], #0x10\n" + "fmul vV14.4s, vU16.4s, vW13.4s\n" + "ldr qU24, [uptr1, u_col_stride3]\n" + "fmul vV13.4s, vU15.4s, vW13.4s\n" + "ldr qW31, [wptr2], #0x10\n" + "fmla vV14.4s, vU15.4s, vW12.4s\n" + "ldr qW21, [wptr1], #0x10\n" + "fmul vV12.4s, vU14.4s, vW13.4s\n" + "ldr qU34, [uptr2, u_col_stride3]\n" + "fmla vV13.4s, vU14.4s, vW12.4s\n" + "ldr qU46, [uptr3, u_col_stride5]\n" + "fmla vV14.4s, vU14.4s, vW11.4s\n" + "ldr qU45, [uptr3, u_col_stride4]\n" + "fmla vV14.4s, vU26.4s, vW23.4s\n" + "ldr qU35, [uptr2, u_col_stride4]\n" + "fmul vV24.4s, vU26.4s, vW13.4s\n" + "ldr qU44, [uptr3, u_col_stride3]\n" + "fmla vV13.4s, vU25.4s, vW23.4s\n" + "bne 1b\n" + + "2:" // Final iteration + "fmla vV14.4s, vU25.4s, vW22.4s\n" + "fmul vV23.4s, vU25.4s, vW13.4s\n" + "fmla vV24.4s, vU25.4s, vW12.4s\n" + "ldr qU56, [uptr4, u_col_stride5]\n" + "fmla vV12.4s, vU24.4s, vW23.4s\n" + "fmla vV13.4s, vU24.4s, vW22.4s\n" + "fmla vV14.4s, vU24.4s, vW21.4s\n" + "fmul vV22.4s, vU24.4s, vW13.4s\n" + "fmla vV23.4s, vU24.4s, vW12.4s\n" + "fmla vV24.4s, vU24.4s, vW11.4s\n" + "ldr qU55, [uptr4, u_col_stride4]\n" + "fmla vV14.4s, vU36.4s, vW33.4s\n" + "fmla vV24.4s, vU36.4s, vW23.4s\n" + "fmul vV34.4s, vU36.4s, vW13.4s\n" + "ldr qU54, [uptr4, u_col_stride3]\n" + "fmla vV13.4s, vU35.4s, vW33.4s\n" + "fmla vV14.4s, vU35.4s, vW32.4s\n" + "fmla vV23.4s, vU35.4s, vW23.4s\n" + "fmla vV24.4s, vU35.4s, vW22.4s\n" + "fmul vV33.4s, vU35.4s, vW13.4s\n" + "fmla vV34.4s, vU35.4s, vW12.4s\n" + "ldr qU66, [uptr5, u_col_stride5]\n" + "fmla vV12.4s, vU34.4s, vW33.4s\n" + "fmla vV13.4s, vU34.4s, vW32.4s\n" + "fmla vV14.4s, vU34.4s, vW31.4s\n" + "str qV14, [%x[vptr0], v_col_stride3]\n" + "fmla vV22.4s, vU34.4s, vW23.4s\n" + "fmla vV23.4s, vU34.4s, vW22.4s\n" + "fmla vV24.4s, vU34.4s, vW21.4s\n" + "fmul vV32.4s, vU34.4s, vW13.4s\n" + "fmla vV33.4s, vU34.4s, vW12.4s\n" + "fmla vV34.4s, vU34.4s, vW11.4s\n" + "ldr qU65, [uptr5, u_col_stride4]\n" + "fmla vV24.4s, vU46.4s, vW33.4s\n" + "fmla vV34.4s, vU46.4s, vW23.4s\n" + "fmul vV44.4s, vU46.4s, vW13.4s\n" + "ldr qU64, [uptr5, u_col_stride3]\n" + "fmla vV23.4s, vU45.4s, vW33.4s\n" + "fmla vV24.4s, vU45.4s, vW32.4s\n" + "fmla vV33.4s, vU45.4s, vW23.4s\n" + "fmla vV34.4s, vU45.4s, vW22.4s\n" + "fmul vV43.4s, vU45.4s, vW13.4s\n" + "fmla vV44.4s, vU45.4s, vW12.4s\n" + "ldr qU13, [%x[uptr0], u_col_stride2]\n" + "fmla vV22.4s, vU44.4s, vW33.4s\n" + "fmla vV23.4s, vU44.4s, vW32.4s\n" + "fmla vV24.4s, vU44.4s, vW31.4s\n" + "str qV24, [vptr1, v_col_stride3]\n" + "fmla vV32.4s, vU44.4s, vW23.4s\n" + "fmla vV33.4s, vU44.4s, vW22.4s\n" + "fmla vV34.4s, vU44.4s, vW21.4s\n" + "fmul vV42.4s, vU44.4s, vW13.4s\n" + "fmla vV43.4s, vU44.4s, vW12.4s\n" + "fmla vV44.4s, vU44.4s, vW11.4s\n" + "ldr qU23, [uptr1, u_col_stride2]\n" + "fmla vV34.4s, vU56.4s, vW33.4s\n" + "fmla vV44.4s, vU56.4s, vW23.4s\n" + "ldr qU33, [uptr2, u_col_stride2]\n" + "fmla vV33.4s, vU55.4s, vW33.4s\n" + "fmla vV34.4s, vU55.4s, vW32.4s\n" + "fmla vV43.4s, vU55.4s, vW23.4s\n" + "fmla vV44.4s, vU55.4s, vW22.4s\n" + "ldr qU43, [uptr3, u_col_stride2]\n" + "fmla vV32.4s, vU54.4s, vW33.4s\n" + "fmla vV33.4s, vU54.4s, vW32.4s\n" + "fmla vV34.4s, vU54.4s, vW31.4s\n" + "str qV34, [vptr2, v_col_stride3]\n" + "fmla vV42.4s, vU54.4s, vW23.4s\n" + "fmla vV43.4s, vU54.4s, vW22.4s\n" + "fmla vV44.4s, vU54.4s, vW21.4s\n" + "ldr qU53, [uptr4, u_col_stride2]\n" + "fmla vV44.4s, vU66.4s, vW33.4s\n" + "ldr qU63, [uptr5, u_col_stride2]\n" + "fmla vV43.4s, vU65.4s, vW33.4s\n" + "fmla vV44.4s, vU65.4s, vW32.4s\n" + "ldr qU12, [%x[uptr0], %x[u_col_stride1]]\n" + "fmla vV42.4s, vU64.4s, vW33.4s\n" + "fmla vV43.4s, vU64.4s, vW32.4s\n" + "fmla vV44.4s, vU64.4s, vW31.4s\n" + "str qV44, [vptr3, v_col_stride3]\n" + "fmul vV11.4s, vU13.4s, vW13.4s\n" + "ldr qU22, [uptr1, %x[u_col_stride1]]\n" + "fmla vV12.4s, vU13.4s, vW12.4s\n" + "fmla vV13.4s, vU13.4s, vW11.4s\n" + "ldr qU32, [uptr2, %x[u_col_stride1]]\n" + "fmla vV11.4s, vU23.4s, vW23.4s\n" + "fmla vV12.4s, vU23.4s, vW22.4s\n" + "fmla vV13.4s, vU23.4s, vW21.4s\n" + "fmul vV21.4s, vU23.4s, vW13.4s\n" + "fmla vV22.4s, vU23.4s, vW12.4s\n" + "fmla vV23.4s, vU23.4s, vW11.4s\n" + "ldr qU42, [uptr3, %x[u_col_stride1]]\n" + "fmla vV11.4s, vU33.4s, vW33.4s\n" + "fmla vV12.4s, vU33.4s, vW32.4s\n" + "fmla vV13.4s, vU33.4s, vW31.4s\n" + "str qV13, [%x[vptr0], v_col_stride2]\n" + "fmla vV21.4s, vU33.4s, vW23.4s\n" + "fmla vV22.4s, vU33.4s, vW22.4s\n" + "fmla vV23.4s, vU33.4s, vW21.4s\n" + "fmul vV31.4s, vU33.4s, vW13.4s\n" + "fmla vV32.4s, vU33.4s, vW12.4s\n" + "fmla vV33.4s, vU33.4s, vW11.4s\n" + "ldr qU52, [uptr4, %x[u_col_stride1]]\n" + "fmla vV21.4s, vU43.4s, vW33.4s\n" + "fmla vV22.4s, vU43.4s, vW32.4s\n" + "fmla vV23.4s, vU43.4s, vW31.4s\n" + "str qV23, [vptr1, v_col_stride2]\n" + "fmla vV31.4s, vU43.4s, vW23.4s\n" + "fmla vV32.4s, vU43.4s, vW22.4s\n" + "fmla vV33.4s, vU43.4s, vW21.4s\n" + "fmul vV41.4s, vU43.4s, vW13.4s\n" + "fmla vV42.4s, vU43.4s, vW12.4s\n" + "fmla vV43.4s, vU43.4s, vW11.4s\n" + "ldr qU62, [uptr5, %x[u_col_stride1]]\n" + "fmla vV31.4s, vU53.4s, vW33.4s\n" + "fmla vV32.4s, vU53.4s, vW32.4s\n" + "fmla vV33.4s, vU53.4s, vW31.4s\n" + "str qV33, [vptr2, v_col_stride2]\n" + "fmla vV41.4s, vU53.4s, vW23.4s\n" + "fmla vV42.4s, vU53.4s, vW22.4s\n" + "fmla vV43.4s, vU53.4s, vW21.4s\n" + "ldr qU11, [%x[uptr0]], #0x10\n" + "fmla vV41.4s, vU63.4s, vW33.4s\n" + "fmla vV42.4s, vU63.4s, vW32.4s\n" + "fmla vV43.4s, vU63.4s, vW31.4s\n" + "str qV43, [vptr3, v_col_stride2]\n" + "fmla vV11.4s, vU12.4s, vW12.4s\n" + "ldr qU21, [uptr1], #0x10\n" + "fmla vV12.4s, vU12.4s, vW11.4s\n" + "ldr qU31, [uptr2], #0x10\n" + "fmla vV11.4s, vU22.4s, vW22.4s\n" + "fmla vV12.4s, vU22.4s, vW21.4s\n" + "fmla vV21.4s, vU22.4s, vW12.4s\n" + "fmla vV22.4s, vU22.4s, vW11.4s\n" + "ldr qU41, [uptr3], #0x10\n" + "fmla vV11.4s, vU32.4s, vW32.4s\n" + "fmla vV12.4s, vU32.4s, vW31.4s\n" + "str qV12, [%x[vptr0], %x[v_col_stride1]]\n" + "fmla vV21.4s, vU32.4s, vW22.4s\n" + "fmla vV22.4s, vU32.4s, vW21.4s\n" + "fmla vV31.4s, vU32.4s, vW12.4s\n" + "fmla vV32.4s, vU32.4s, vW11.4s\n" + "ldr qU51, [uptr4], #0x10\n" + "fmla vV21.4s, vU42.4s, vW32.4s\n" + "fmla vV22.4s, vU42.4s, vW31.4s\n" + "str qV22, [vptr1, %x[v_col_stride1]]\n" + "fmla vV31.4s, vU42.4s, vW22.4s\n" + "fmla vV32.4s, vU42.4s, vW21.4s\n" + "subs %x[c4_rem], %x[c4_rem], #1\n" + "fmla vV41.4s, vU42.4s, vW12.4s\n" + "fmla vV42.4s, vU42.4s, vW11.4s\n" + "ldr qU61, [uptr5], #0x10\n" + "fmla vV31.4s, vU52.4s, vW32.4s\n" + "fmla vV32.4s, vU52.4s, vW31.4s\n" + "str qV32, [vptr2, %x[v_col_stride1]]\n" + "fmla vV41.4s, vU52.4s, vW22.4s\n" + "fmla vV42.4s, vU52.4s, vW21.4s\n" + "fmla vV41.4s, vU62.4s, vW32.4s\n" + "fmla vV42.4s, vU62.4s, vW31.4s\n" + "str qV42, [vptr3, %x[v_col_stride1]]\n" + "fmla vV11.4s, vU11.4s, vW11.4s\n" + "fmla vV11.4s, vU21.4s, vW21.4s\n" + "fmla vV21.4s, vU21.4s, vW11.4s\n" + "fmla vV11.4s, vU31.4s, vW31.4s\n" + "str qV11, [%x[vptr0]], #0x10\n" + "fmla vV21.4s, vU31.4s, vW21.4s\n" + "fmla vV31.4s, vU31.4s, vW11.4s\n" + "fmla vV21.4s, vU41.4s, vW31.4s\n" + "str qV21, [vptr1], #0x10\n" + "fmla vV31.4s, vU41.4s, vW21.4s\n" + "fmla vV41.4s, vU41.4s, vW11.4s\n" + "fmla vV31.4s, vU51.4s, vW31.4s\n" + "str qV31, [vptr2], #0x10\n" + "fmla vV41.4s, vU51.4s, vW21.4s\n" + "fmla vV41.4s, vU61.4s, vW31.4s\n" + "str qV41, [vptr3], #0x10\n" + + ".unreq qW22\n" ".unreq qU64\n" ".unreq qU35\n" ".unreq qV41\n" + ".unreq qU34\n" ".unreq qU21\n" ".unreq qV43\n" ".unreq qW21\n" + ".unreq qU24\n" ".unreq qU54\n" ".unreq qV31\n" ".unreq qV12\n" + ".unreq qU61\n" ".unreq qU26\n" ".unreq qV32\n" + ".unreq qU36\n" ".unreq qU51\n" ".unreq qU66\n" ".unreq qU12\n" + ".unreq qV14\n" ".unreq qV11\n" ".unreq qU65\n" + ".unreq qU15\n" ".unreq qU22\n" ".unreq qU45\n" + ".unreq qV22\n" ".unreq qU14\n" + ".unreq qU44\n" ".unreq qU43\n" ".unreq qU11\n" + ".unreq qV24\n" ".unreq qV42\n" ".unreq qW31\n" ".unreq qW13\n" + ".unreq qU33\n" ".unreq qU62\n" ".unreq qU25\n" ".unreq qU56\n" + ".unreq qW33\n" + ".unreq qU42\n" ".unreq qU16\n" ".unreq qV44\n" + ".unreq qU63\n" ".unreq qU31\n" ".unreq qV34\n" + ".unreq qW11\n" ".unreq qU41\n" ".unreq qV13\n" ".unreq qV33\n" + ".unreq qU46\n" ".unreq qU32\n" ".unreq qU13\n" + ".unreq qW23\n" ".unreq qV23\n" ".unreq qV21\n" ".unreq qU55\n" + ".unreq qW12\n" ".unreq qW32\n" ".unreq qU23\n" ".unreq qU52\n" + ".unreq qU53\n" ".unreq vW22\n" + ".unreq vU64\n" ".unreq vU35\n" ".unreq vV41\n" + ".unreq vU34\n" ".unreq vU21\n" ".unreq vV43\n" ".unreq vW21\n" + ".unreq vU24\n" ".unreq vU54\n" ".unreq vV31\n" + ".unreq vV12\n" ".unreq vU61\n" + ".unreq vU26\n" ".unreq vV32\n" + ".unreq vU36\n" ".unreq vU51\n" ".unreq vU66\n" ".unreq vU12\n" + ".unreq vV14\n" ".unreq vV11\n" ".unreq vU65\n" + ".unreq vU15\n" ".unreq vU22\n" ".unreq vU45\n" + ".unreq vV22\n" ".unreq vU14\n" + ".unreq vU44\n" ".unreq vU43\n" ".unreq vU11\n" + ".unreq vV24\n" ".unreq vV42\n" ".unreq vW31\n" ".unreq vW13\n" + ".unreq vU33\n" ".unreq vU62\n" ".unreq vU25\n" ".unreq vU56\n" + ".unreq vW33\n" ".unreq vU42\n" ".unreq vU16\n" ".unreq vV44\n" + ".unreq vU63\n" ".unreq vU31\n" ".unreq vV34\n" ".unreq vW11\n" + ".unreq vU41\n" ".unreq vV13\n" ".unreq vV33\n" + ".unreq vU46\n" ".unreq vU32\n" ".unreq vU13\n" ".unreq vW23\n" + ".unreq vV23\n" ".unreq vV21\n" ".unreq vU55\n" ".unreq vW12\n" + ".unreq vW32\n" ".unreq vU23\n" ".unreq vU52\n" ".unreq vU53\n" + : [uptr0] "+r" (uptr0), [vptr0] "+r" (vptr0), [wptr0] "+r" (wptr0), + [c4_rem] "+r" (c4_rem) + : [u_row_stride] "r" (in_row_stride * sizeof(float)), + [u_col_stride1] "r" (in_col_stride * sizeof(float)), + [v_row_stride] "r" (out_row_stride * sizeof(float)), + [v_col_stride1] "r" (out_col_stride * sizeof(float)), + [w_row_stride] "r" (weight_row_stride * sizeof(float)), + [w_col_stride1] "r" (weight_col_stride * sizeof(float)) + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", + "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "x0", + "x1", "x2", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", + "x12", "x13", "x14", "x15", "x16", "cc", "memory" + ); + } + for (; channels_remaining; channels_remaining--) + { + // Load input tile + float u[inner_tile_rows][inner_tile_cols]; + for (int i = 0; i < inner_tile_rows; i++) + { + const float* const inptr_row = uptr0 + (i - in_pad_top)*in_row_stride; + for (int j = 0; j < inner_tile_cols; j++) + { + if (i < in_pad_top || in_cells_i <= i || + j < in_pad_left || in_cells_j <= j) + { + u[i][j] = static_cast(0); + } + else + { + u[i][j] = *(inptr_row + (j - in_pad_left)*in_col_stride); + } + } + } + uptr0++; + + // Load weights tile + float w[kernel_rows][kernel_cols]; + for (int i = 0; i < kernel_rows; i++) + { + const float* const wptr_row = wptr0 + i*weight_row_stride; + for (int j = 0; j < kernel_cols; j++) + { + w[i][j] = *(wptr_row + j*weight_col_stride); + } + } + wptr0++; + + // Perform the convolution + float v[output_tile_rows][output_tile_cols]; + for (int out_i = 0; out_i < out_cells_i; out_i++) + { + for (int out_j = 0; out_j < out_cells_j; out_j++) + { + // Clear the accumulator + v[out_i][out_j] = static_cast(0); + + // Base co-ordinate + const int base_i = out_i * stride_rows; + const int base_j = out_j * stride_cols; + + // Fill the accumulator + for (int in_i = 0; in_i < kernel_rows; in_i++) + { + const int i = base_i + in_i; + for (int in_j = 0; in_j < kernel_cols; in_j++) + { + const int j = base_j + in_j; + v[out_i][out_j] += w[in_i][in_j] * u[i][j]; + } + } + } + } + + // Store the output tile + for (int i = 0; i < out_cells_i; i++) + { + float* const outptr_row = vptr0 + i*out_row_stride; + for (int j = 0; j < out_cells_j; j++) + { + *(outptr_row + j*out_col_stride) = v[i][j]; + } + } + vptr0++; + } +} + +#endif // __aarch64__ + +template <> +const Conv::TileFn Conv::tilefn_unpadded = ConvImpl::template process_tile; + +template <> +const Conv::TileFn Conv::tilefn_top[n_in_pad_top_fns] = { + ConvImpl::template process_tile, }; +template <> +const Conv::TileFn Conv::tilefn_left[n_in_pad_left_fns] = { + ConvImpl::template process_tile, +}; + +template <> +const Conv::TileFn Conv::tilefn_bottom[n_in_pad_bottom_fns][n_out_pad_bottom_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, +}; + +template <> +const Conv::TileFn Conv::tilefn_right[n_in_pad_right_fns][n_out_pad_right_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, +}; + +template <> +const Conv::TileFn Conv::tilefn_generic = ConvImpl::template process_tile; template class DepthwiseConvolution<4, 4, 3, 3, 1, 1, float, float>; } // namespace depthwise diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_4x4_3x3_2x2_fp32_fp32.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_4x4_3x3_2x2_fp32_fp32.cpp index 2104c0bbf7..8eb53a66b4 100644 --- a/src/core/NEON/kernels/convolution/depthwise/depthwise_4x4_3x3_2x2_fp32_fp32.cpp +++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_4x4_3x3_2x2_fp32_fp32.cpp @@ -29,5179 +29,138 @@ using Conv = DepthwiseConvolution<4, 4, 3, 3, 2, 2, float, float>; using ConvImpl = DepthwiseConvolutionImpl<4, 4, 3, 3, 2, 2, float, float>; template <> -const Conv::TileFn Conv::tile_fns - [max_in_pad_top] - [max_in_pad_left] - [max_in_pad_bottom] - [max_in_pad_right] - [max_out_pad_bottom] - [max_out_pad_right] = { - { // Input pad top = 0 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 0, 0, 0>, - Conv::template process_tile<0, 0, 0, 0, 0, 1>, - Conv::template process_tile<0, 0, 0, 0, 0, 2>, - Conv::template process_tile<0, 0, 0, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 0, 1, 0>, - Conv::template process_tile<0, 0, 0, 0, 1, 1>, - Conv::template process_tile<0, 0, 0, 0, 1, 2>, - Conv::template process_tile<0, 0, 0, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 0, 2, 0>, - Conv::template process_tile<0, 0, 0, 0, 2, 1>, - Conv::template process_tile<0, 0, 0, 0, 2, 2>, - Conv::template process_tile<0, 0, 0, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 0, 0, 3, 0>, - Conv::template process_tile<0, 0, 0, 0, 3, 1>, - Conv::template process_tile<0, 0, 0, 0, 3, 2>, - Conv::template process_tile<0, 0, 0, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 1, 0, 0>, - Conv::template process_tile<0, 0, 0, 1, 0, 1>, - Conv::template process_tile<0, 0, 0, 1, 0, 2>, - Conv::template process_tile<0, 0, 0, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 1, 1, 0>, - Conv::template process_tile<0, 0, 0, 1, 1, 1>, - Conv::template process_tile<0, 0, 0, 1, 1, 2>, - Conv::template process_tile<0, 0, 0, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 1, 2, 0>, - Conv::template process_tile<0, 0, 0, 1, 2, 1>, - Conv::template process_tile<0, 0, 0, 1, 2, 2>, - Conv::template process_tile<0, 0, 0, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 0, 1, 3, 0>, - Conv::template process_tile<0, 0, 0, 1, 3, 1>, - Conv::template process_tile<0, 0, 0, 1, 3, 2>, - Conv::template process_tile<0, 0, 0, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 2, 0, 0>, - Conv::template process_tile<0, 0, 0, 2, 0, 1>, - Conv::template process_tile<0, 0, 0, 2, 0, 2>, - Conv::template process_tile<0, 0, 0, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 2, 1, 0>, - Conv::template process_tile<0, 0, 0, 2, 1, 1>, - Conv::template process_tile<0, 0, 0, 2, 1, 2>, - Conv::template process_tile<0, 0, 0, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 2, 2, 0>, - Conv::template process_tile<0, 0, 0, 2, 2, 1>, - Conv::template process_tile<0, 0, 0, 2, 2, 2>, - Conv::template process_tile<0, 0, 0, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 0, 2, 3, 0>, - Conv::template process_tile<0, 0, 0, 2, 3, 1>, - Conv::template process_tile<0, 0, 0, 2, 3, 2>, - Conv::template process_tile<0, 0, 0, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 3, 0, 0>, - Conv::template process_tile<0, 0, 0, 3, 0, 1>, - Conv::template process_tile<0, 0, 0, 3, 0, 2>, - Conv::template process_tile<0, 0, 0, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 3, 1, 0>, - Conv::template process_tile<0, 0, 0, 3, 1, 1>, - Conv::template process_tile<0, 0, 0, 3, 1, 2>, - Conv::template process_tile<0, 0, 0, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 3, 2, 0>, - Conv::template process_tile<0, 0, 0, 3, 2, 1>, - Conv::template process_tile<0, 0, 0, 3, 2, 2>, - Conv::template process_tile<0, 0, 0, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 0, 3, 3, 0>, - Conv::template process_tile<0, 0, 0, 3, 3, 1>, - Conv::template process_tile<0, 0, 0, 3, 3, 2>, - Conv::template process_tile<0, 0, 0, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 4, 0, 0>, - Conv::template process_tile<0, 0, 0, 4, 0, 1>, - Conv::template process_tile<0, 0, 0, 4, 0, 2>, - Conv::template process_tile<0, 0, 0, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 4, 1, 0>, - Conv::template process_tile<0, 0, 0, 4, 1, 1>, - Conv::template process_tile<0, 0, 0, 4, 1, 2>, - Conv::template process_tile<0, 0, 0, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 4, 2, 0>, - Conv::template process_tile<0, 0, 0, 4, 2, 1>, - Conv::template process_tile<0, 0, 0, 4, 2, 2>, - Conv::template process_tile<0, 0, 0, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 0, 4, 3, 0>, - Conv::template process_tile<0, 0, 0, 4, 3, 1>, - Conv::template process_tile<0, 0, 0, 4, 3, 2>, - Conv::template process_tile<0, 0, 0, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 5, 0, 0>, - Conv::template process_tile<0, 0, 0, 5, 0, 1>, - Conv::template process_tile<0, 0, 0, 5, 0, 2>, - Conv::template process_tile<0, 0, 0, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 5, 1, 0>, - Conv::template process_tile<0, 0, 0, 5, 1, 1>, - Conv::template process_tile<0, 0, 0, 5, 1, 2>, - Conv::template process_tile<0, 0, 0, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 5, 2, 0>, - Conv::template process_tile<0, 0, 0, 5, 2, 1>, - Conv::template process_tile<0, 0, 0, 5, 2, 2>, - Conv::template process_tile<0, 0, 0, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 0, 5, 3, 0>, - Conv::template process_tile<0, 0, 0, 5, 3, 1>, - Conv::template process_tile<0, 0, 0, 5, 3, 2>, - Conv::template process_tile<0, 0, 0, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 0, 6, 0, 0>, - Conv::template process_tile<0, 0, 0, 6, 0, 1>, - Conv::template process_tile<0, 0, 0, 6, 0, 2>, - Conv::template process_tile<0, 0, 0, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 0, 6, 1, 0>, - Conv::template process_tile<0, 0, 0, 6, 1, 1>, - Conv::template process_tile<0, 0, 0, 6, 1, 2>, - Conv::template process_tile<0, 0, 0, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 0, 6, 2, 0>, - Conv::template process_tile<0, 0, 0, 6, 2, 1>, - Conv::template process_tile<0, 0, 0, 6, 2, 2>, - Conv::template process_tile<0, 0, 0, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 0, 6, 3, 0>, - Conv::template process_tile<0, 0, 0, 6, 3, 1>, - Conv::template process_tile<0, 0, 0, 6, 3, 2>, - Conv::template process_tile<0, 0, 0, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 0, 0, 0>, - Conv::template process_tile<0, 0, 1, 0, 0, 1>, - Conv::template process_tile<0, 0, 1, 0, 0, 2>, - Conv::template process_tile<0, 0, 1, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 0, 1, 0>, - Conv::template process_tile<0, 0, 1, 0, 1, 1>, - Conv::template process_tile<0, 0, 1, 0, 1, 2>, - Conv::template process_tile<0, 0, 1, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 0, 2, 0>, - Conv::template process_tile<0, 0, 1, 0, 2, 1>, - Conv::template process_tile<0, 0, 1, 0, 2, 2>, - Conv::template process_tile<0, 0, 1, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 1, 0, 3, 0>, - Conv::template process_tile<0, 0, 1, 0, 3, 1>, - Conv::template process_tile<0, 0, 1, 0, 3, 2>, - Conv::template process_tile<0, 0, 1, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 1, 0, 0>, - Conv::template process_tile<0, 0, 1, 1, 0, 1>, - Conv::template process_tile<0, 0, 1, 1, 0, 2>, - Conv::template process_tile<0, 0, 1, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 1, 1, 0>, - Conv::template process_tile<0, 0, 1, 1, 1, 1>, - Conv::template process_tile<0, 0, 1, 1, 1, 2>, - Conv::template process_tile<0, 0, 1, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 1, 2, 0>, - Conv::template process_tile<0, 0, 1, 1, 2, 1>, - Conv::template process_tile<0, 0, 1, 1, 2, 2>, - Conv::template process_tile<0, 0, 1, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 1, 1, 3, 0>, - Conv::template process_tile<0, 0, 1, 1, 3, 1>, - Conv::template process_tile<0, 0, 1, 1, 3, 2>, - Conv::template process_tile<0, 0, 1, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 2, 0, 0>, - Conv::template process_tile<0, 0, 1, 2, 0, 1>, - Conv::template process_tile<0, 0, 1, 2, 0, 2>, - Conv::template process_tile<0, 0, 1, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 2, 1, 0>, - Conv::template process_tile<0, 0, 1, 2, 1, 1>, - Conv::template process_tile<0, 0, 1, 2, 1, 2>, - Conv::template process_tile<0, 0, 1, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 2, 2, 0>, - Conv::template process_tile<0, 0, 1, 2, 2, 1>, - Conv::template process_tile<0, 0, 1, 2, 2, 2>, - Conv::template process_tile<0, 0, 1, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 1, 2, 3, 0>, - Conv::template process_tile<0, 0, 1, 2, 3, 1>, - Conv::template process_tile<0, 0, 1, 2, 3, 2>, - Conv::template process_tile<0, 0, 1, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 3, 0, 0>, - Conv::template process_tile<0, 0, 1, 3, 0, 1>, - Conv::template process_tile<0, 0, 1, 3, 0, 2>, - Conv::template process_tile<0, 0, 1, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 3, 1, 0>, - Conv::template process_tile<0, 0, 1, 3, 1, 1>, - Conv::template process_tile<0, 0, 1, 3, 1, 2>, - Conv::template process_tile<0, 0, 1, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 3, 2, 0>, - Conv::template process_tile<0, 0, 1, 3, 2, 1>, - Conv::template process_tile<0, 0, 1, 3, 2, 2>, - Conv::template process_tile<0, 0, 1, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 1, 3, 3, 0>, - Conv::template process_tile<0, 0, 1, 3, 3, 1>, - Conv::template process_tile<0, 0, 1, 3, 3, 2>, - Conv::template process_tile<0, 0, 1, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 4, 0, 0>, - Conv::template process_tile<0, 0, 1, 4, 0, 1>, - Conv::template process_tile<0, 0, 1, 4, 0, 2>, - Conv::template process_tile<0, 0, 1, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 4, 1, 0>, - Conv::template process_tile<0, 0, 1, 4, 1, 1>, - Conv::template process_tile<0, 0, 1, 4, 1, 2>, - Conv::template process_tile<0, 0, 1, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 4, 2, 0>, - Conv::template process_tile<0, 0, 1, 4, 2, 1>, - Conv::template process_tile<0, 0, 1, 4, 2, 2>, - Conv::template process_tile<0, 0, 1, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 1, 4, 3, 0>, - Conv::template process_tile<0, 0, 1, 4, 3, 1>, - Conv::template process_tile<0, 0, 1, 4, 3, 2>, - Conv::template process_tile<0, 0, 1, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 5, 0, 0>, - Conv::template process_tile<0, 0, 1, 5, 0, 1>, - Conv::template process_tile<0, 0, 1, 5, 0, 2>, - Conv::template process_tile<0, 0, 1, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 5, 1, 0>, - Conv::template process_tile<0, 0, 1, 5, 1, 1>, - Conv::template process_tile<0, 0, 1, 5, 1, 2>, - Conv::template process_tile<0, 0, 1, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 5, 2, 0>, - Conv::template process_tile<0, 0, 1, 5, 2, 1>, - Conv::template process_tile<0, 0, 1, 5, 2, 2>, - Conv::template process_tile<0, 0, 1, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 1, 5, 3, 0>, - Conv::template process_tile<0, 0, 1, 5, 3, 1>, - Conv::template process_tile<0, 0, 1, 5, 3, 2>, - Conv::template process_tile<0, 0, 1, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 1, 6, 0, 0>, - Conv::template process_tile<0, 0, 1, 6, 0, 1>, - Conv::template process_tile<0, 0, 1, 6, 0, 2>, - Conv::template process_tile<0, 0, 1, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 1, 6, 1, 0>, - Conv::template process_tile<0, 0, 1, 6, 1, 1>, - Conv::template process_tile<0, 0, 1, 6, 1, 2>, - Conv::template process_tile<0, 0, 1, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 1, 6, 2, 0>, - Conv::template process_tile<0, 0, 1, 6, 2, 1>, - Conv::template process_tile<0, 0, 1, 6, 2, 2>, - Conv::template process_tile<0, 0, 1, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 1, 6, 3, 0>, - Conv::template process_tile<0, 0, 1, 6, 3, 1>, - Conv::template process_tile<0, 0, 1, 6, 3, 2>, - Conv::template process_tile<0, 0, 1, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 0, 0, 0>, - Conv::template process_tile<0, 0, 2, 0, 0, 1>, - Conv::template process_tile<0, 0, 2, 0, 0, 2>, - Conv::template process_tile<0, 0, 2, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 0, 1, 0>, - Conv::template process_tile<0, 0, 2, 0, 1, 1>, - Conv::template process_tile<0, 0, 2, 0, 1, 2>, - Conv::template process_tile<0, 0, 2, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 0, 2, 0>, - Conv::template process_tile<0, 0, 2, 0, 2, 1>, - Conv::template process_tile<0, 0, 2, 0, 2, 2>, - Conv::template process_tile<0, 0, 2, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 2, 0, 3, 0>, - Conv::template process_tile<0, 0, 2, 0, 3, 1>, - Conv::template process_tile<0, 0, 2, 0, 3, 2>, - Conv::template process_tile<0, 0, 2, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 1, 0, 0>, - Conv::template process_tile<0, 0, 2, 1, 0, 1>, - Conv::template process_tile<0, 0, 2, 1, 0, 2>, - Conv::template process_tile<0, 0, 2, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 1, 1, 0>, - Conv::template process_tile<0, 0, 2, 1, 1, 1>, - Conv::template process_tile<0, 0, 2, 1, 1, 2>, - Conv::template process_tile<0, 0, 2, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 1, 2, 0>, - Conv::template process_tile<0, 0, 2, 1, 2, 1>, - Conv::template process_tile<0, 0, 2, 1, 2, 2>, - Conv::template process_tile<0, 0, 2, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 2, 1, 3, 0>, - Conv::template process_tile<0, 0, 2, 1, 3, 1>, - Conv::template process_tile<0, 0, 2, 1, 3, 2>, - Conv::template process_tile<0, 0, 2, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 2, 0, 0>, - Conv::template process_tile<0, 0, 2, 2, 0, 1>, - Conv::template process_tile<0, 0, 2, 2, 0, 2>, - Conv::template process_tile<0, 0, 2, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 2, 1, 0>, - Conv::template process_tile<0, 0, 2, 2, 1, 1>, - Conv::template process_tile<0, 0, 2, 2, 1, 2>, - Conv::template process_tile<0, 0, 2, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 2, 2, 0>, - Conv::template process_tile<0, 0, 2, 2, 2, 1>, - Conv::template process_tile<0, 0, 2, 2, 2, 2>, - Conv::template process_tile<0, 0, 2, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 2, 2, 3, 0>, - Conv::template process_tile<0, 0, 2, 2, 3, 1>, - Conv::template process_tile<0, 0, 2, 2, 3, 2>, - Conv::template process_tile<0, 0, 2, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 3, 0, 0>, - Conv::template process_tile<0, 0, 2, 3, 0, 1>, - Conv::template process_tile<0, 0, 2, 3, 0, 2>, - Conv::template process_tile<0, 0, 2, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 3, 1, 0>, - Conv::template process_tile<0, 0, 2, 3, 1, 1>, - Conv::template process_tile<0, 0, 2, 3, 1, 2>, - Conv::template process_tile<0, 0, 2, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 3, 2, 0>, - Conv::template process_tile<0, 0, 2, 3, 2, 1>, - Conv::template process_tile<0, 0, 2, 3, 2, 2>, - Conv::template process_tile<0, 0, 2, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 2, 3, 3, 0>, - Conv::template process_tile<0, 0, 2, 3, 3, 1>, - Conv::template process_tile<0, 0, 2, 3, 3, 2>, - Conv::template process_tile<0, 0, 2, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 4, 0, 0>, - Conv::template process_tile<0, 0, 2, 4, 0, 1>, - Conv::template process_tile<0, 0, 2, 4, 0, 2>, - Conv::template process_tile<0, 0, 2, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 4, 1, 0>, - Conv::template process_tile<0, 0, 2, 4, 1, 1>, - Conv::template process_tile<0, 0, 2, 4, 1, 2>, - Conv::template process_tile<0, 0, 2, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 4, 2, 0>, - Conv::template process_tile<0, 0, 2, 4, 2, 1>, - Conv::template process_tile<0, 0, 2, 4, 2, 2>, - Conv::template process_tile<0, 0, 2, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 2, 4, 3, 0>, - Conv::template process_tile<0, 0, 2, 4, 3, 1>, - Conv::template process_tile<0, 0, 2, 4, 3, 2>, - Conv::template process_tile<0, 0, 2, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 5, 0, 0>, - Conv::template process_tile<0, 0, 2, 5, 0, 1>, - Conv::template process_tile<0, 0, 2, 5, 0, 2>, - Conv::template process_tile<0, 0, 2, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 5, 1, 0>, - Conv::template process_tile<0, 0, 2, 5, 1, 1>, - Conv::template process_tile<0, 0, 2, 5, 1, 2>, - Conv::template process_tile<0, 0, 2, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 5, 2, 0>, - Conv::template process_tile<0, 0, 2, 5, 2, 1>, - Conv::template process_tile<0, 0, 2, 5, 2, 2>, - Conv::template process_tile<0, 0, 2, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 2, 5, 3, 0>, - Conv::template process_tile<0, 0, 2, 5, 3, 1>, - Conv::template process_tile<0, 0, 2, 5, 3, 2>, - Conv::template process_tile<0, 0, 2, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 2, 6, 0, 0>, - Conv::template process_tile<0, 0, 2, 6, 0, 1>, - Conv::template process_tile<0, 0, 2, 6, 0, 2>, - Conv::template process_tile<0, 0, 2, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 2, 6, 1, 0>, - Conv::template process_tile<0, 0, 2, 6, 1, 1>, - Conv::template process_tile<0, 0, 2, 6, 1, 2>, - Conv::template process_tile<0, 0, 2, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 2, 6, 2, 0>, - Conv::template process_tile<0, 0, 2, 6, 2, 1>, - Conv::template process_tile<0, 0, 2, 6, 2, 2>, - Conv::template process_tile<0, 0, 2, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 2, 6, 3, 0>, - Conv::template process_tile<0, 0, 2, 6, 3, 1>, - Conv::template process_tile<0, 0, 2, 6, 3, 2>, - Conv::template process_tile<0, 0, 2, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 0, 0, 0>, - Conv::template process_tile<0, 0, 3, 0, 0, 1>, - Conv::template process_tile<0, 0, 3, 0, 0, 2>, - Conv::template process_tile<0, 0, 3, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 0, 1, 0>, - Conv::template process_tile<0, 0, 3, 0, 1, 1>, - Conv::template process_tile<0, 0, 3, 0, 1, 2>, - Conv::template process_tile<0, 0, 3, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 0, 2, 0>, - Conv::template process_tile<0, 0, 3, 0, 2, 1>, - Conv::template process_tile<0, 0, 3, 0, 2, 2>, - Conv::template process_tile<0, 0, 3, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 3, 0, 3, 0>, - Conv::template process_tile<0, 0, 3, 0, 3, 1>, - Conv::template process_tile<0, 0, 3, 0, 3, 2>, - Conv::template process_tile<0, 0, 3, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 1, 0, 0>, - Conv::template process_tile<0, 0, 3, 1, 0, 1>, - Conv::template process_tile<0, 0, 3, 1, 0, 2>, - Conv::template process_tile<0, 0, 3, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 1, 1, 0>, - Conv::template process_tile<0, 0, 3, 1, 1, 1>, - Conv::template process_tile<0, 0, 3, 1, 1, 2>, - Conv::template process_tile<0, 0, 3, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 1, 2, 0>, - Conv::template process_tile<0, 0, 3, 1, 2, 1>, - Conv::template process_tile<0, 0, 3, 1, 2, 2>, - Conv::template process_tile<0, 0, 3, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 3, 1, 3, 0>, - Conv::template process_tile<0, 0, 3, 1, 3, 1>, - Conv::template process_tile<0, 0, 3, 1, 3, 2>, - Conv::template process_tile<0, 0, 3, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 2, 0, 0>, - Conv::template process_tile<0, 0, 3, 2, 0, 1>, - Conv::template process_tile<0, 0, 3, 2, 0, 2>, - Conv::template process_tile<0, 0, 3, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 2, 1, 0>, - Conv::template process_tile<0, 0, 3, 2, 1, 1>, - Conv::template process_tile<0, 0, 3, 2, 1, 2>, - Conv::template process_tile<0, 0, 3, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 2, 2, 0>, - Conv::template process_tile<0, 0, 3, 2, 2, 1>, - Conv::template process_tile<0, 0, 3, 2, 2, 2>, - Conv::template process_tile<0, 0, 3, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 3, 2, 3, 0>, - Conv::template process_tile<0, 0, 3, 2, 3, 1>, - Conv::template process_tile<0, 0, 3, 2, 3, 2>, - Conv::template process_tile<0, 0, 3, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 3, 0, 0>, - Conv::template process_tile<0, 0, 3, 3, 0, 1>, - Conv::template process_tile<0, 0, 3, 3, 0, 2>, - Conv::template process_tile<0, 0, 3, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 3, 1, 0>, - Conv::template process_tile<0, 0, 3, 3, 1, 1>, - Conv::template process_tile<0, 0, 3, 3, 1, 2>, - Conv::template process_tile<0, 0, 3, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 3, 2, 0>, - Conv::template process_tile<0, 0, 3, 3, 2, 1>, - Conv::template process_tile<0, 0, 3, 3, 2, 2>, - Conv::template process_tile<0, 0, 3, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 3, 3, 3, 0>, - Conv::template process_tile<0, 0, 3, 3, 3, 1>, - Conv::template process_tile<0, 0, 3, 3, 3, 2>, - Conv::template process_tile<0, 0, 3, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 4, 0, 0>, - Conv::template process_tile<0, 0, 3, 4, 0, 1>, - Conv::template process_tile<0, 0, 3, 4, 0, 2>, - Conv::template process_tile<0, 0, 3, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 4, 1, 0>, - Conv::template process_tile<0, 0, 3, 4, 1, 1>, - Conv::template process_tile<0, 0, 3, 4, 1, 2>, - Conv::template process_tile<0, 0, 3, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 4, 2, 0>, - Conv::template process_tile<0, 0, 3, 4, 2, 1>, - Conv::template process_tile<0, 0, 3, 4, 2, 2>, - Conv::template process_tile<0, 0, 3, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 3, 4, 3, 0>, - Conv::template process_tile<0, 0, 3, 4, 3, 1>, - Conv::template process_tile<0, 0, 3, 4, 3, 2>, - Conv::template process_tile<0, 0, 3, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 5, 0, 0>, - Conv::template process_tile<0, 0, 3, 5, 0, 1>, - Conv::template process_tile<0, 0, 3, 5, 0, 2>, - Conv::template process_tile<0, 0, 3, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 5, 1, 0>, - Conv::template process_tile<0, 0, 3, 5, 1, 1>, - Conv::template process_tile<0, 0, 3, 5, 1, 2>, - Conv::template process_tile<0, 0, 3, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 5, 2, 0>, - Conv::template process_tile<0, 0, 3, 5, 2, 1>, - Conv::template process_tile<0, 0, 3, 5, 2, 2>, - Conv::template process_tile<0, 0, 3, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 3, 5, 3, 0>, - Conv::template process_tile<0, 0, 3, 5, 3, 1>, - Conv::template process_tile<0, 0, 3, 5, 3, 2>, - Conv::template process_tile<0, 0, 3, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 3, 6, 0, 0>, - Conv::template process_tile<0, 0, 3, 6, 0, 1>, - Conv::template process_tile<0, 0, 3, 6, 0, 2>, - Conv::template process_tile<0, 0, 3, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 3, 6, 1, 0>, - Conv::template process_tile<0, 0, 3, 6, 1, 1>, - Conv::template process_tile<0, 0, 3, 6, 1, 2>, - Conv::template process_tile<0, 0, 3, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 3, 6, 2, 0>, - Conv::template process_tile<0, 0, 3, 6, 2, 1>, - Conv::template process_tile<0, 0, 3, 6, 2, 2>, - Conv::template process_tile<0, 0, 3, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 3, 6, 3, 0>, - Conv::template process_tile<0, 0, 3, 6, 3, 1>, - Conv::template process_tile<0, 0, 3, 6, 3, 2>, - Conv::template process_tile<0, 0, 3, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 0, 0, 0>, - Conv::template process_tile<0, 0, 4, 0, 0, 1>, - Conv::template process_tile<0, 0, 4, 0, 0, 2>, - Conv::template process_tile<0, 0, 4, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 0, 1, 0>, - Conv::template process_tile<0, 0, 4, 0, 1, 1>, - Conv::template process_tile<0, 0, 4, 0, 1, 2>, - Conv::template process_tile<0, 0, 4, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 0, 2, 0>, - Conv::template process_tile<0, 0, 4, 0, 2, 1>, - Conv::template process_tile<0, 0, 4, 0, 2, 2>, - Conv::template process_tile<0, 0, 4, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 4, 0, 3, 0>, - Conv::template process_tile<0, 0, 4, 0, 3, 1>, - Conv::template process_tile<0, 0, 4, 0, 3, 2>, - Conv::template process_tile<0, 0, 4, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 1, 0, 0>, - Conv::template process_tile<0, 0, 4, 1, 0, 1>, - Conv::template process_tile<0, 0, 4, 1, 0, 2>, - Conv::template process_tile<0, 0, 4, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 1, 1, 0>, - Conv::template process_tile<0, 0, 4, 1, 1, 1>, - Conv::template process_tile<0, 0, 4, 1, 1, 2>, - Conv::template process_tile<0, 0, 4, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 1, 2, 0>, - Conv::template process_tile<0, 0, 4, 1, 2, 1>, - Conv::template process_tile<0, 0, 4, 1, 2, 2>, - Conv::template process_tile<0, 0, 4, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 4, 1, 3, 0>, - Conv::template process_tile<0, 0, 4, 1, 3, 1>, - Conv::template process_tile<0, 0, 4, 1, 3, 2>, - Conv::template process_tile<0, 0, 4, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 2, 0, 0>, - Conv::template process_tile<0, 0, 4, 2, 0, 1>, - Conv::template process_tile<0, 0, 4, 2, 0, 2>, - Conv::template process_tile<0, 0, 4, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 2, 1, 0>, - Conv::template process_tile<0, 0, 4, 2, 1, 1>, - Conv::template process_tile<0, 0, 4, 2, 1, 2>, - Conv::template process_tile<0, 0, 4, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 2, 2, 0>, - Conv::template process_tile<0, 0, 4, 2, 2, 1>, - Conv::template process_tile<0, 0, 4, 2, 2, 2>, - Conv::template process_tile<0, 0, 4, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 4, 2, 3, 0>, - Conv::template process_tile<0, 0, 4, 2, 3, 1>, - Conv::template process_tile<0, 0, 4, 2, 3, 2>, - Conv::template process_tile<0, 0, 4, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 3, 0, 0>, - Conv::template process_tile<0, 0, 4, 3, 0, 1>, - Conv::template process_tile<0, 0, 4, 3, 0, 2>, - Conv::template process_tile<0, 0, 4, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 3, 1, 0>, - Conv::template process_tile<0, 0, 4, 3, 1, 1>, - Conv::template process_tile<0, 0, 4, 3, 1, 2>, - Conv::template process_tile<0, 0, 4, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 3, 2, 0>, - Conv::template process_tile<0, 0, 4, 3, 2, 1>, - Conv::template process_tile<0, 0, 4, 3, 2, 2>, - Conv::template process_tile<0, 0, 4, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 4, 3, 3, 0>, - Conv::template process_tile<0, 0, 4, 3, 3, 1>, - Conv::template process_tile<0, 0, 4, 3, 3, 2>, - Conv::template process_tile<0, 0, 4, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 4, 0, 0>, - Conv::template process_tile<0, 0, 4, 4, 0, 1>, - Conv::template process_tile<0, 0, 4, 4, 0, 2>, - Conv::template process_tile<0, 0, 4, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 4, 1, 0>, - Conv::template process_tile<0, 0, 4, 4, 1, 1>, - Conv::template process_tile<0, 0, 4, 4, 1, 2>, - Conv::template process_tile<0, 0, 4, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 4, 2, 0>, - Conv::template process_tile<0, 0, 4, 4, 2, 1>, - Conv::template process_tile<0, 0, 4, 4, 2, 2>, - Conv::template process_tile<0, 0, 4, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 4, 4, 3, 0>, - Conv::template process_tile<0, 0, 4, 4, 3, 1>, - Conv::template process_tile<0, 0, 4, 4, 3, 2>, - Conv::template process_tile<0, 0, 4, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 5, 0, 0>, - Conv::template process_tile<0, 0, 4, 5, 0, 1>, - Conv::template process_tile<0, 0, 4, 5, 0, 2>, - Conv::template process_tile<0, 0, 4, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 5, 1, 0>, - Conv::template process_tile<0, 0, 4, 5, 1, 1>, - Conv::template process_tile<0, 0, 4, 5, 1, 2>, - Conv::template process_tile<0, 0, 4, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 5, 2, 0>, - Conv::template process_tile<0, 0, 4, 5, 2, 1>, - Conv::template process_tile<0, 0, 4, 5, 2, 2>, - Conv::template process_tile<0, 0, 4, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 4, 5, 3, 0>, - Conv::template process_tile<0, 0, 4, 5, 3, 1>, - Conv::template process_tile<0, 0, 4, 5, 3, 2>, - Conv::template process_tile<0, 0, 4, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 4, 6, 0, 0>, - Conv::template process_tile<0, 0, 4, 6, 0, 1>, - Conv::template process_tile<0, 0, 4, 6, 0, 2>, - Conv::template process_tile<0, 0, 4, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 4, 6, 1, 0>, - Conv::template process_tile<0, 0, 4, 6, 1, 1>, - Conv::template process_tile<0, 0, 4, 6, 1, 2>, - Conv::template process_tile<0, 0, 4, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 4, 6, 2, 0>, - Conv::template process_tile<0, 0, 4, 6, 2, 1>, - Conv::template process_tile<0, 0, 4, 6, 2, 2>, - Conv::template process_tile<0, 0, 4, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 4, 6, 3, 0>, - Conv::template process_tile<0, 0, 4, 6, 3, 1>, - Conv::template process_tile<0, 0, 4, 6, 3, 2>, - Conv::template process_tile<0, 0, 4, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 4 - { // Input pad bottom = 5 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 0, 0, 0>, - Conv::template process_tile<0, 0, 5, 0, 0, 1>, - Conv::template process_tile<0, 0, 5, 0, 0, 2>, - Conv::template process_tile<0, 0, 5, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 0, 1, 0>, - Conv::template process_tile<0, 0, 5, 0, 1, 1>, - Conv::template process_tile<0, 0, 5, 0, 1, 2>, - Conv::template process_tile<0, 0, 5, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 0, 2, 0>, - Conv::template process_tile<0, 0, 5, 0, 2, 1>, - Conv::template process_tile<0, 0, 5, 0, 2, 2>, - Conv::template process_tile<0, 0, 5, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 5, 0, 3, 0>, - Conv::template process_tile<0, 0, 5, 0, 3, 1>, - Conv::template process_tile<0, 0, 5, 0, 3, 2>, - Conv::template process_tile<0, 0, 5, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 1, 0, 0>, - Conv::template process_tile<0, 0, 5, 1, 0, 1>, - Conv::template process_tile<0, 0, 5, 1, 0, 2>, - Conv::template process_tile<0, 0, 5, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 1, 1, 0>, - Conv::template process_tile<0, 0, 5, 1, 1, 1>, - Conv::template process_tile<0, 0, 5, 1, 1, 2>, - Conv::template process_tile<0, 0, 5, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 1, 2, 0>, - Conv::template process_tile<0, 0, 5, 1, 2, 1>, - Conv::template process_tile<0, 0, 5, 1, 2, 2>, - Conv::template process_tile<0, 0, 5, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 5, 1, 3, 0>, - Conv::template process_tile<0, 0, 5, 1, 3, 1>, - Conv::template process_tile<0, 0, 5, 1, 3, 2>, - Conv::template process_tile<0, 0, 5, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 2, 0, 0>, - Conv::template process_tile<0, 0, 5, 2, 0, 1>, - Conv::template process_tile<0, 0, 5, 2, 0, 2>, - Conv::template process_tile<0, 0, 5, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 2, 1, 0>, - Conv::template process_tile<0, 0, 5, 2, 1, 1>, - Conv::template process_tile<0, 0, 5, 2, 1, 2>, - Conv::template process_tile<0, 0, 5, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 2, 2, 0>, - Conv::template process_tile<0, 0, 5, 2, 2, 1>, - Conv::template process_tile<0, 0, 5, 2, 2, 2>, - Conv::template process_tile<0, 0, 5, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 5, 2, 3, 0>, - Conv::template process_tile<0, 0, 5, 2, 3, 1>, - Conv::template process_tile<0, 0, 5, 2, 3, 2>, - Conv::template process_tile<0, 0, 5, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 3, 0, 0>, - Conv::template process_tile<0, 0, 5, 3, 0, 1>, - Conv::template process_tile<0, 0, 5, 3, 0, 2>, - Conv::template process_tile<0, 0, 5, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 3, 1, 0>, - Conv::template process_tile<0, 0, 5, 3, 1, 1>, - Conv::template process_tile<0, 0, 5, 3, 1, 2>, - Conv::template process_tile<0, 0, 5, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 3, 2, 0>, - Conv::template process_tile<0, 0, 5, 3, 2, 1>, - Conv::template process_tile<0, 0, 5, 3, 2, 2>, - Conv::template process_tile<0, 0, 5, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 5, 3, 3, 0>, - Conv::template process_tile<0, 0, 5, 3, 3, 1>, - Conv::template process_tile<0, 0, 5, 3, 3, 2>, - Conv::template process_tile<0, 0, 5, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 4, 0, 0>, - Conv::template process_tile<0, 0, 5, 4, 0, 1>, - Conv::template process_tile<0, 0, 5, 4, 0, 2>, - Conv::template process_tile<0, 0, 5, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 4, 1, 0>, - Conv::template process_tile<0, 0, 5, 4, 1, 1>, - Conv::template process_tile<0, 0, 5, 4, 1, 2>, - Conv::template process_tile<0, 0, 5, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 4, 2, 0>, - Conv::template process_tile<0, 0, 5, 4, 2, 1>, - Conv::template process_tile<0, 0, 5, 4, 2, 2>, - Conv::template process_tile<0, 0, 5, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 5, 4, 3, 0>, - Conv::template process_tile<0, 0, 5, 4, 3, 1>, - Conv::template process_tile<0, 0, 5, 4, 3, 2>, - Conv::template process_tile<0, 0, 5, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 5, 0, 0>, - Conv::template process_tile<0, 0, 5, 5, 0, 1>, - Conv::template process_tile<0, 0, 5, 5, 0, 2>, - Conv::template process_tile<0, 0, 5, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 5, 1, 0>, - Conv::template process_tile<0, 0, 5, 5, 1, 1>, - Conv::template process_tile<0, 0, 5, 5, 1, 2>, - Conv::template process_tile<0, 0, 5, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 5, 2, 0>, - Conv::template process_tile<0, 0, 5, 5, 2, 1>, - Conv::template process_tile<0, 0, 5, 5, 2, 2>, - Conv::template process_tile<0, 0, 5, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 5, 5, 3, 0>, - Conv::template process_tile<0, 0, 5, 5, 3, 1>, - Conv::template process_tile<0, 0, 5, 5, 3, 2>, - Conv::template process_tile<0, 0, 5, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 5, 6, 0, 0>, - Conv::template process_tile<0, 0, 5, 6, 0, 1>, - Conv::template process_tile<0, 0, 5, 6, 0, 2>, - Conv::template process_tile<0, 0, 5, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 5, 6, 1, 0>, - Conv::template process_tile<0, 0, 5, 6, 1, 1>, - Conv::template process_tile<0, 0, 5, 6, 1, 2>, - Conv::template process_tile<0, 0, 5, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 5, 6, 2, 0>, - Conv::template process_tile<0, 0, 5, 6, 2, 1>, - Conv::template process_tile<0, 0, 5, 6, 2, 2>, - Conv::template process_tile<0, 0, 5, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 5, 6, 3, 0>, - Conv::template process_tile<0, 0, 5, 6, 3, 1>, - Conv::template process_tile<0, 0, 5, 6, 3, 2>, - Conv::template process_tile<0, 0, 5, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 5 - { // Input pad bottom = 6 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 0, 0, 0>, - Conv::template process_tile<0, 0, 6, 0, 0, 1>, - Conv::template process_tile<0, 0, 6, 0, 0, 2>, - Conv::template process_tile<0, 0, 6, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 0, 1, 0>, - Conv::template process_tile<0, 0, 6, 0, 1, 1>, - Conv::template process_tile<0, 0, 6, 0, 1, 2>, - Conv::template process_tile<0, 0, 6, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 0, 2, 0>, - Conv::template process_tile<0, 0, 6, 0, 2, 1>, - Conv::template process_tile<0, 0, 6, 0, 2, 2>, - Conv::template process_tile<0, 0, 6, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 6, 0, 3, 0>, - Conv::template process_tile<0, 0, 6, 0, 3, 1>, - Conv::template process_tile<0, 0, 6, 0, 3, 2>, - Conv::template process_tile<0, 0, 6, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 1, 0, 0>, - Conv::template process_tile<0, 0, 6, 1, 0, 1>, - Conv::template process_tile<0, 0, 6, 1, 0, 2>, - Conv::template process_tile<0, 0, 6, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 1, 1, 0>, - Conv::template process_tile<0, 0, 6, 1, 1, 1>, - Conv::template process_tile<0, 0, 6, 1, 1, 2>, - Conv::template process_tile<0, 0, 6, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 1, 2, 0>, - Conv::template process_tile<0, 0, 6, 1, 2, 1>, - Conv::template process_tile<0, 0, 6, 1, 2, 2>, - Conv::template process_tile<0, 0, 6, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 6, 1, 3, 0>, - Conv::template process_tile<0, 0, 6, 1, 3, 1>, - Conv::template process_tile<0, 0, 6, 1, 3, 2>, - Conv::template process_tile<0, 0, 6, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 2, 0, 0>, - Conv::template process_tile<0, 0, 6, 2, 0, 1>, - Conv::template process_tile<0, 0, 6, 2, 0, 2>, - Conv::template process_tile<0, 0, 6, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 2, 1, 0>, - Conv::template process_tile<0, 0, 6, 2, 1, 1>, - Conv::template process_tile<0, 0, 6, 2, 1, 2>, - Conv::template process_tile<0, 0, 6, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 2, 2, 0>, - Conv::template process_tile<0, 0, 6, 2, 2, 1>, - Conv::template process_tile<0, 0, 6, 2, 2, 2>, - Conv::template process_tile<0, 0, 6, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 6, 2, 3, 0>, - Conv::template process_tile<0, 0, 6, 2, 3, 1>, - Conv::template process_tile<0, 0, 6, 2, 3, 2>, - Conv::template process_tile<0, 0, 6, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 3, 0, 0>, - Conv::template process_tile<0, 0, 6, 3, 0, 1>, - Conv::template process_tile<0, 0, 6, 3, 0, 2>, - Conv::template process_tile<0, 0, 6, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 3, 1, 0>, - Conv::template process_tile<0, 0, 6, 3, 1, 1>, - Conv::template process_tile<0, 0, 6, 3, 1, 2>, - Conv::template process_tile<0, 0, 6, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 3, 2, 0>, - Conv::template process_tile<0, 0, 6, 3, 2, 1>, - Conv::template process_tile<0, 0, 6, 3, 2, 2>, - Conv::template process_tile<0, 0, 6, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 6, 3, 3, 0>, - Conv::template process_tile<0, 0, 6, 3, 3, 1>, - Conv::template process_tile<0, 0, 6, 3, 3, 2>, - Conv::template process_tile<0, 0, 6, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 4, 0, 0>, - Conv::template process_tile<0, 0, 6, 4, 0, 1>, - Conv::template process_tile<0, 0, 6, 4, 0, 2>, - Conv::template process_tile<0, 0, 6, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 4, 1, 0>, - Conv::template process_tile<0, 0, 6, 4, 1, 1>, - Conv::template process_tile<0, 0, 6, 4, 1, 2>, - Conv::template process_tile<0, 0, 6, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 4, 2, 0>, - Conv::template process_tile<0, 0, 6, 4, 2, 1>, - Conv::template process_tile<0, 0, 6, 4, 2, 2>, - Conv::template process_tile<0, 0, 6, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 6, 4, 3, 0>, - Conv::template process_tile<0, 0, 6, 4, 3, 1>, - Conv::template process_tile<0, 0, 6, 4, 3, 2>, - Conv::template process_tile<0, 0, 6, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 5, 0, 0>, - Conv::template process_tile<0, 0, 6, 5, 0, 1>, - Conv::template process_tile<0, 0, 6, 5, 0, 2>, - Conv::template process_tile<0, 0, 6, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 5, 1, 0>, - Conv::template process_tile<0, 0, 6, 5, 1, 1>, - Conv::template process_tile<0, 0, 6, 5, 1, 2>, - Conv::template process_tile<0, 0, 6, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 5, 2, 0>, - Conv::template process_tile<0, 0, 6, 5, 2, 1>, - Conv::template process_tile<0, 0, 6, 5, 2, 2>, - Conv::template process_tile<0, 0, 6, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 6, 5, 3, 0>, - Conv::template process_tile<0, 0, 6, 5, 3, 1>, - Conv::template process_tile<0, 0, 6, 5, 3, 2>, - Conv::template process_tile<0, 0, 6, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 0, 6, 6, 0, 0>, - Conv::template process_tile<0, 0, 6, 6, 0, 1>, - Conv::template process_tile<0, 0, 6, 6, 0, 2>, - Conv::template process_tile<0, 0, 6, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 0, 6, 6, 1, 0>, - Conv::template process_tile<0, 0, 6, 6, 1, 1>, - Conv::template process_tile<0, 0, 6, 6, 1, 2>, - Conv::template process_tile<0, 0, 6, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 0, 6, 6, 2, 0>, - Conv::template process_tile<0, 0, 6, 6, 2, 1>, - Conv::template process_tile<0, 0, 6, 6, 2, 2>, - Conv::template process_tile<0, 0, 6, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 0, 6, 6, 3, 0>, - Conv::template process_tile<0, 0, 6, 6, 3, 1>, - Conv::template process_tile<0, 0, 6, 6, 3, 2>, - Conv::template process_tile<0, 0, 6, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 6 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 0, 0, 0>, - Conv::template process_tile<0, 1, 0, 0, 0, 1>, - Conv::template process_tile<0, 1, 0, 0, 0, 2>, - Conv::template process_tile<0, 1, 0, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 0, 1, 0>, - Conv::template process_tile<0, 1, 0, 0, 1, 1>, - Conv::template process_tile<0, 1, 0, 0, 1, 2>, - Conv::template process_tile<0, 1, 0, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 0, 2, 0>, - Conv::template process_tile<0, 1, 0, 0, 2, 1>, - Conv::template process_tile<0, 1, 0, 0, 2, 2>, - Conv::template process_tile<0, 1, 0, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 0, 0, 3, 0>, - Conv::template process_tile<0, 1, 0, 0, 3, 1>, - Conv::template process_tile<0, 1, 0, 0, 3, 2>, - Conv::template process_tile<0, 1, 0, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 1, 0, 0>, - Conv::template process_tile<0, 1, 0, 1, 0, 1>, - Conv::template process_tile<0, 1, 0, 1, 0, 2>, - Conv::template process_tile<0, 1, 0, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 1, 1, 0>, - Conv::template process_tile<0, 1, 0, 1, 1, 1>, - Conv::template process_tile<0, 1, 0, 1, 1, 2>, - Conv::template process_tile<0, 1, 0, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 1, 2, 0>, - Conv::template process_tile<0, 1, 0, 1, 2, 1>, - Conv::template process_tile<0, 1, 0, 1, 2, 2>, - Conv::template process_tile<0, 1, 0, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 0, 1, 3, 0>, - Conv::template process_tile<0, 1, 0, 1, 3, 1>, - Conv::template process_tile<0, 1, 0, 1, 3, 2>, - Conv::template process_tile<0, 1, 0, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 2, 0, 0>, - Conv::template process_tile<0, 1, 0, 2, 0, 1>, - Conv::template process_tile<0, 1, 0, 2, 0, 2>, - Conv::template process_tile<0, 1, 0, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 2, 1, 0>, - Conv::template process_tile<0, 1, 0, 2, 1, 1>, - Conv::template process_tile<0, 1, 0, 2, 1, 2>, - Conv::template process_tile<0, 1, 0, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 2, 2, 0>, - Conv::template process_tile<0, 1, 0, 2, 2, 1>, - Conv::template process_tile<0, 1, 0, 2, 2, 2>, - Conv::template process_tile<0, 1, 0, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 0, 2, 3, 0>, - Conv::template process_tile<0, 1, 0, 2, 3, 1>, - Conv::template process_tile<0, 1, 0, 2, 3, 2>, - Conv::template process_tile<0, 1, 0, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 3, 0, 0>, - Conv::template process_tile<0, 1, 0, 3, 0, 1>, - Conv::template process_tile<0, 1, 0, 3, 0, 2>, - Conv::template process_tile<0, 1, 0, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 3, 1, 0>, - Conv::template process_tile<0, 1, 0, 3, 1, 1>, - Conv::template process_tile<0, 1, 0, 3, 1, 2>, - Conv::template process_tile<0, 1, 0, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 3, 2, 0>, - Conv::template process_tile<0, 1, 0, 3, 2, 1>, - Conv::template process_tile<0, 1, 0, 3, 2, 2>, - Conv::template process_tile<0, 1, 0, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 0, 3, 3, 0>, - Conv::template process_tile<0, 1, 0, 3, 3, 1>, - Conv::template process_tile<0, 1, 0, 3, 3, 2>, - Conv::template process_tile<0, 1, 0, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 4, 0, 0>, - Conv::template process_tile<0, 1, 0, 4, 0, 1>, - Conv::template process_tile<0, 1, 0, 4, 0, 2>, - Conv::template process_tile<0, 1, 0, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 4, 1, 0>, - Conv::template process_tile<0, 1, 0, 4, 1, 1>, - Conv::template process_tile<0, 1, 0, 4, 1, 2>, - Conv::template process_tile<0, 1, 0, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 4, 2, 0>, - Conv::template process_tile<0, 1, 0, 4, 2, 1>, - Conv::template process_tile<0, 1, 0, 4, 2, 2>, - Conv::template process_tile<0, 1, 0, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 0, 4, 3, 0>, - Conv::template process_tile<0, 1, 0, 4, 3, 1>, - Conv::template process_tile<0, 1, 0, 4, 3, 2>, - Conv::template process_tile<0, 1, 0, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 5, 0, 0>, - Conv::template process_tile<0, 1, 0, 5, 0, 1>, - Conv::template process_tile<0, 1, 0, 5, 0, 2>, - Conv::template process_tile<0, 1, 0, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 5, 1, 0>, - Conv::template process_tile<0, 1, 0, 5, 1, 1>, - Conv::template process_tile<0, 1, 0, 5, 1, 2>, - Conv::template process_tile<0, 1, 0, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 5, 2, 0>, - Conv::template process_tile<0, 1, 0, 5, 2, 1>, - Conv::template process_tile<0, 1, 0, 5, 2, 2>, - Conv::template process_tile<0, 1, 0, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 0, 5, 3, 0>, - Conv::template process_tile<0, 1, 0, 5, 3, 1>, - Conv::template process_tile<0, 1, 0, 5, 3, 2>, - Conv::template process_tile<0, 1, 0, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 0, 6, 0, 0>, - Conv::template process_tile<0, 1, 0, 6, 0, 1>, - Conv::template process_tile<0, 1, 0, 6, 0, 2>, - Conv::template process_tile<0, 1, 0, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 0, 6, 1, 0>, - Conv::template process_tile<0, 1, 0, 6, 1, 1>, - Conv::template process_tile<0, 1, 0, 6, 1, 2>, - Conv::template process_tile<0, 1, 0, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 0, 6, 2, 0>, - Conv::template process_tile<0, 1, 0, 6, 2, 1>, - Conv::template process_tile<0, 1, 0, 6, 2, 2>, - Conv::template process_tile<0, 1, 0, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 0, 6, 3, 0>, - Conv::template process_tile<0, 1, 0, 6, 3, 1>, - Conv::template process_tile<0, 1, 0, 6, 3, 2>, - Conv::template process_tile<0, 1, 0, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 0, 0, 0>, - Conv::template process_tile<0, 1, 1, 0, 0, 1>, - Conv::template process_tile<0, 1, 1, 0, 0, 2>, - Conv::template process_tile<0, 1, 1, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 0, 1, 0>, - Conv::template process_tile<0, 1, 1, 0, 1, 1>, - Conv::template process_tile<0, 1, 1, 0, 1, 2>, - Conv::template process_tile<0, 1, 1, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 0, 2, 0>, - Conv::template process_tile<0, 1, 1, 0, 2, 1>, - Conv::template process_tile<0, 1, 1, 0, 2, 2>, - Conv::template process_tile<0, 1, 1, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 1, 0, 3, 0>, - Conv::template process_tile<0, 1, 1, 0, 3, 1>, - Conv::template process_tile<0, 1, 1, 0, 3, 2>, - Conv::template process_tile<0, 1, 1, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 1, 0, 0>, - Conv::template process_tile<0, 1, 1, 1, 0, 1>, - Conv::template process_tile<0, 1, 1, 1, 0, 2>, - Conv::template process_tile<0, 1, 1, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 1, 1, 0>, - Conv::template process_tile<0, 1, 1, 1, 1, 1>, - Conv::template process_tile<0, 1, 1, 1, 1, 2>, - Conv::template process_tile<0, 1, 1, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 1, 2, 0>, - Conv::template process_tile<0, 1, 1, 1, 2, 1>, - Conv::template process_tile<0, 1, 1, 1, 2, 2>, - Conv::template process_tile<0, 1, 1, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 1, 1, 3, 0>, - Conv::template process_tile<0, 1, 1, 1, 3, 1>, - Conv::template process_tile<0, 1, 1, 1, 3, 2>, - Conv::template process_tile<0, 1, 1, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 2, 0, 0>, - Conv::template process_tile<0, 1, 1, 2, 0, 1>, - Conv::template process_tile<0, 1, 1, 2, 0, 2>, - Conv::template process_tile<0, 1, 1, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 2, 1, 0>, - Conv::template process_tile<0, 1, 1, 2, 1, 1>, - Conv::template process_tile<0, 1, 1, 2, 1, 2>, - Conv::template process_tile<0, 1, 1, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 2, 2, 0>, - Conv::template process_tile<0, 1, 1, 2, 2, 1>, - Conv::template process_tile<0, 1, 1, 2, 2, 2>, - Conv::template process_tile<0, 1, 1, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 1, 2, 3, 0>, - Conv::template process_tile<0, 1, 1, 2, 3, 1>, - Conv::template process_tile<0, 1, 1, 2, 3, 2>, - Conv::template process_tile<0, 1, 1, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 3, 0, 0>, - Conv::template process_tile<0, 1, 1, 3, 0, 1>, - Conv::template process_tile<0, 1, 1, 3, 0, 2>, - Conv::template process_tile<0, 1, 1, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 3, 1, 0>, - Conv::template process_tile<0, 1, 1, 3, 1, 1>, - Conv::template process_tile<0, 1, 1, 3, 1, 2>, - Conv::template process_tile<0, 1, 1, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 3, 2, 0>, - Conv::template process_tile<0, 1, 1, 3, 2, 1>, - Conv::template process_tile<0, 1, 1, 3, 2, 2>, - Conv::template process_tile<0, 1, 1, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 1, 3, 3, 0>, - Conv::template process_tile<0, 1, 1, 3, 3, 1>, - Conv::template process_tile<0, 1, 1, 3, 3, 2>, - Conv::template process_tile<0, 1, 1, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 4, 0, 0>, - Conv::template process_tile<0, 1, 1, 4, 0, 1>, - Conv::template process_tile<0, 1, 1, 4, 0, 2>, - Conv::template process_tile<0, 1, 1, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 4, 1, 0>, - Conv::template process_tile<0, 1, 1, 4, 1, 1>, - Conv::template process_tile<0, 1, 1, 4, 1, 2>, - Conv::template process_tile<0, 1, 1, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 4, 2, 0>, - Conv::template process_tile<0, 1, 1, 4, 2, 1>, - Conv::template process_tile<0, 1, 1, 4, 2, 2>, - Conv::template process_tile<0, 1, 1, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 1, 4, 3, 0>, - Conv::template process_tile<0, 1, 1, 4, 3, 1>, - Conv::template process_tile<0, 1, 1, 4, 3, 2>, - Conv::template process_tile<0, 1, 1, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 5, 0, 0>, - Conv::template process_tile<0, 1, 1, 5, 0, 1>, - Conv::template process_tile<0, 1, 1, 5, 0, 2>, - Conv::template process_tile<0, 1, 1, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 5, 1, 0>, - Conv::template process_tile<0, 1, 1, 5, 1, 1>, - Conv::template process_tile<0, 1, 1, 5, 1, 2>, - Conv::template process_tile<0, 1, 1, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 5, 2, 0>, - Conv::template process_tile<0, 1, 1, 5, 2, 1>, - Conv::template process_tile<0, 1, 1, 5, 2, 2>, - Conv::template process_tile<0, 1, 1, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 1, 5, 3, 0>, - Conv::template process_tile<0, 1, 1, 5, 3, 1>, - Conv::template process_tile<0, 1, 1, 5, 3, 2>, - Conv::template process_tile<0, 1, 1, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 1, 6, 0, 0>, - Conv::template process_tile<0, 1, 1, 6, 0, 1>, - Conv::template process_tile<0, 1, 1, 6, 0, 2>, - Conv::template process_tile<0, 1, 1, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 1, 6, 1, 0>, - Conv::template process_tile<0, 1, 1, 6, 1, 1>, - Conv::template process_tile<0, 1, 1, 6, 1, 2>, - Conv::template process_tile<0, 1, 1, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 1, 6, 2, 0>, - Conv::template process_tile<0, 1, 1, 6, 2, 1>, - Conv::template process_tile<0, 1, 1, 6, 2, 2>, - Conv::template process_tile<0, 1, 1, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 1, 6, 3, 0>, - Conv::template process_tile<0, 1, 1, 6, 3, 1>, - Conv::template process_tile<0, 1, 1, 6, 3, 2>, - Conv::template process_tile<0, 1, 1, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 0, 0, 0>, - Conv::template process_tile<0, 1, 2, 0, 0, 1>, - Conv::template process_tile<0, 1, 2, 0, 0, 2>, - Conv::template process_tile<0, 1, 2, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 0, 1, 0>, - Conv::template process_tile<0, 1, 2, 0, 1, 1>, - Conv::template process_tile<0, 1, 2, 0, 1, 2>, - Conv::template process_tile<0, 1, 2, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 0, 2, 0>, - Conv::template process_tile<0, 1, 2, 0, 2, 1>, - Conv::template process_tile<0, 1, 2, 0, 2, 2>, - Conv::template process_tile<0, 1, 2, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 2, 0, 3, 0>, - Conv::template process_tile<0, 1, 2, 0, 3, 1>, - Conv::template process_tile<0, 1, 2, 0, 3, 2>, - Conv::template process_tile<0, 1, 2, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 1, 0, 0>, - Conv::template process_tile<0, 1, 2, 1, 0, 1>, - Conv::template process_tile<0, 1, 2, 1, 0, 2>, - Conv::template process_tile<0, 1, 2, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 1, 1, 0>, - Conv::template process_tile<0, 1, 2, 1, 1, 1>, - Conv::template process_tile<0, 1, 2, 1, 1, 2>, - Conv::template process_tile<0, 1, 2, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 1, 2, 0>, - Conv::template process_tile<0, 1, 2, 1, 2, 1>, - Conv::template process_tile<0, 1, 2, 1, 2, 2>, - Conv::template process_tile<0, 1, 2, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 2, 1, 3, 0>, - Conv::template process_tile<0, 1, 2, 1, 3, 1>, - Conv::template process_tile<0, 1, 2, 1, 3, 2>, - Conv::template process_tile<0, 1, 2, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 2, 0, 0>, - Conv::template process_tile<0, 1, 2, 2, 0, 1>, - Conv::template process_tile<0, 1, 2, 2, 0, 2>, - Conv::template process_tile<0, 1, 2, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 2, 1, 0>, - Conv::template process_tile<0, 1, 2, 2, 1, 1>, - Conv::template process_tile<0, 1, 2, 2, 1, 2>, - Conv::template process_tile<0, 1, 2, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 2, 2, 0>, - Conv::template process_tile<0, 1, 2, 2, 2, 1>, - Conv::template process_tile<0, 1, 2, 2, 2, 2>, - Conv::template process_tile<0, 1, 2, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 2, 2, 3, 0>, - Conv::template process_tile<0, 1, 2, 2, 3, 1>, - Conv::template process_tile<0, 1, 2, 2, 3, 2>, - Conv::template process_tile<0, 1, 2, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 3, 0, 0>, - Conv::template process_tile<0, 1, 2, 3, 0, 1>, - Conv::template process_tile<0, 1, 2, 3, 0, 2>, - Conv::template process_tile<0, 1, 2, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 3, 1, 0>, - Conv::template process_tile<0, 1, 2, 3, 1, 1>, - Conv::template process_tile<0, 1, 2, 3, 1, 2>, - Conv::template process_tile<0, 1, 2, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 3, 2, 0>, - Conv::template process_tile<0, 1, 2, 3, 2, 1>, - Conv::template process_tile<0, 1, 2, 3, 2, 2>, - Conv::template process_tile<0, 1, 2, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 2, 3, 3, 0>, - Conv::template process_tile<0, 1, 2, 3, 3, 1>, - Conv::template process_tile<0, 1, 2, 3, 3, 2>, - Conv::template process_tile<0, 1, 2, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 4, 0, 0>, - Conv::template process_tile<0, 1, 2, 4, 0, 1>, - Conv::template process_tile<0, 1, 2, 4, 0, 2>, - Conv::template process_tile<0, 1, 2, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 4, 1, 0>, - Conv::template process_tile<0, 1, 2, 4, 1, 1>, - Conv::template process_tile<0, 1, 2, 4, 1, 2>, - Conv::template process_tile<0, 1, 2, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 4, 2, 0>, - Conv::template process_tile<0, 1, 2, 4, 2, 1>, - Conv::template process_tile<0, 1, 2, 4, 2, 2>, - Conv::template process_tile<0, 1, 2, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 2, 4, 3, 0>, - Conv::template process_tile<0, 1, 2, 4, 3, 1>, - Conv::template process_tile<0, 1, 2, 4, 3, 2>, - Conv::template process_tile<0, 1, 2, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 5, 0, 0>, - Conv::template process_tile<0, 1, 2, 5, 0, 1>, - Conv::template process_tile<0, 1, 2, 5, 0, 2>, - Conv::template process_tile<0, 1, 2, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 5, 1, 0>, - Conv::template process_tile<0, 1, 2, 5, 1, 1>, - Conv::template process_tile<0, 1, 2, 5, 1, 2>, - Conv::template process_tile<0, 1, 2, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 5, 2, 0>, - Conv::template process_tile<0, 1, 2, 5, 2, 1>, - Conv::template process_tile<0, 1, 2, 5, 2, 2>, - Conv::template process_tile<0, 1, 2, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 2, 5, 3, 0>, - Conv::template process_tile<0, 1, 2, 5, 3, 1>, - Conv::template process_tile<0, 1, 2, 5, 3, 2>, - Conv::template process_tile<0, 1, 2, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 2, 6, 0, 0>, - Conv::template process_tile<0, 1, 2, 6, 0, 1>, - Conv::template process_tile<0, 1, 2, 6, 0, 2>, - Conv::template process_tile<0, 1, 2, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 2, 6, 1, 0>, - Conv::template process_tile<0, 1, 2, 6, 1, 1>, - Conv::template process_tile<0, 1, 2, 6, 1, 2>, - Conv::template process_tile<0, 1, 2, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 2, 6, 2, 0>, - Conv::template process_tile<0, 1, 2, 6, 2, 1>, - Conv::template process_tile<0, 1, 2, 6, 2, 2>, - Conv::template process_tile<0, 1, 2, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 2, 6, 3, 0>, - Conv::template process_tile<0, 1, 2, 6, 3, 1>, - Conv::template process_tile<0, 1, 2, 6, 3, 2>, - Conv::template process_tile<0, 1, 2, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 0, 0, 0>, - Conv::template process_tile<0, 1, 3, 0, 0, 1>, - Conv::template process_tile<0, 1, 3, 0, 0, 2>, - Conv::template process_tile<0, 1, 3, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 0, 1, 0>, - Conv::template process_tile<0, 1, 3, 0, 1, 1>, - Conv::template process_tile<0, 1, 3, 0, 1, 2>, - Conv::template process_tile<0, 1, 3, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 0, 2, 0>, - Conv::template process_tile<0, 1, 3, 0, 2, 1>, - Conv::template process_tile<0, 1, 3, 0, 2, 2>, - Conv::template process_tile<0, 1, 3, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 3, 0, 3, 0>, - Conv::template process_tile<0, 1, 3, 0, 3, 1>, - Conv::template process_tile<0, 1, 3, 0, 3, 2>, - Conv::template process_tile<0, 1, 3, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 1, 0, 0>, - Conv::template process_tile<0, 1, 3, 1, 0, 1>, - Conv::template process_tile<0, 1, 3, 1, 0, 2>, - Conv::template process_tile<0, 1, 3, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 1, 1, 0>, - Conv::template process_tile<0, 1, 3, 1, 1, 1>, - Conv::template process_tile<0, 1, 3, 1, 1, 2>, - Conv::template process_tile<0, 1, 3, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 1, 2, 0>, - Conv::template process_tile<0, 1, 3, 1, 2, 1>, - Conv::template process_tile<0, 1, 3, 1, 2, 2>, - Conv::template process_tile<0, 1, 3, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 3, 1, 3, 0>, - Conv::template process_tile<0, 1, 3, 1, 3, 1>, - Conv::template process_tile<0, 1, 3, 1, 3, 2>, - Conv::template process_tile<0, 1, 3, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 2, 0, 0>, - Conv::template process_tile<0, 1, 3, 2, 0, 1>, - Conv::template process_tile<0, 1, 3, 2, 0, 2>, - Conv::template process_tile<0, 1, 3, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 2, 1, 0>, - Conv::template process_tile<0, 1, 3, 2, 1, 1>, - Conv::template process_tile<0, 1, 3, 2, 1, 2>, - Conv::template process_tile<0, 1, 3, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 2, 2, 0>, - Conv::template process_tile<0, 1, 3, 2, 2, 1>, - Conv::template process_tile<0, 1, 3, 2, 2, 2>, - Conv::template process_tile<0, 1, 3, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 3, 2, 3, 0>, - Conv::template process_tile<0, 1, 3, 2, 3, 1>, - Conv::template process_tile<0, 1, 3, 2, 3, 2>, - Conv::template process_tile<0, 1, 3, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 3, 0, 0>, - Conv::template process_tile<0, 1, 3, 3, 0, 1>, - Conv::template process_tile<0, 1, 3, 3, 0, 2>, - Conv::template process_tile<0, 1, 3, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 3, 1, 0>, - Conv::template process_tile<0, 1, 3, 3, 1, 1>, - Conv::template process_tile<0, 1, 3, 3, 1, 2>, - Conv::template process_tile<0, 1, 3, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 3, 2, 0>, - Conv::template process_tile<0, 1, 3, 3, 2, 1>, - Conv::template process_tile<0, 1, 3, 3, 2, 2>, - Conv::template process_tile<0, 1, 3, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 3, 3, 3, 0>, - Conv::template process_tile<0, 1, 3, 3, 3, 1>, - Conv::template process_tile<0, 1, 3, 3, 3, 2>, - Conv::template process_tile<0, 1, 3, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 4, 0, 0>, - Conv::template process_tile<0, 1, 3, 4, 0, 1>, - Conv::template process_tile<0, 1, 3, 4, 0, 2>, - Conv::template process_tile<0, 1, 3, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 4, 1, 0>, - Conv::template process_tile<0, 1, 3, 4, 1, 1>, - Conv::template process_tile<0, 1, 3, 4, 1, 2>, - Conv::template process_tile<0, 1, 3, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 4, 2, 0>, - Conv::template process_tile<0, 1, 3, 4, 2, 1>, - Conv::template process_tile<0, 1, 3, 4, 2, 2>, - Conv::template process_tile<0, 1, 3, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 3, 4, 3, 0>, - Conv::template process_tile<0, 1, 3, 4, 3, 1>, - Conv::template process_tile<0, 1, 3, 4, 3, 2>, - Conv::template process_tile<0, 1, 3, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 5, 0, 0>, - Conv::template process_tile<0, 1, 3, 5, 0, 1>, - Conv::template process_tile<0, 1, 3, 5, 0, 2>, - Conv::template process_tile<0, 1, 3, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 5, 1, 0>, - Conv::template process_tile<0, 1, 3, 5, 1, 1>, - Conv::template process_tile<0, 1, 3, 5, 1, 2>, - Conv::template process_tile<0, 1, 3, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 5, 2, 0>, - Conv::template process_tile<0, 1, 3, 5, 2, 1>, - Conv::template process_tile<0, 1, 3, 5, 2, 2>, - Conv::template process_tile<0, 1, 3, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 3, 5, 3, 0>, - Conv::template process_tile<0, 1, 3, 5, 3, 1>, - Conv::template process_tile<0, 1, 3, 5, 3, 2>, - Conv::template process_tile<0, 1, 3, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 3, 6, 0, 0>, - Conv::template process_tile<0, 1, 3, 6, 0, 1>, - Conv::template process_tile<0, 1, 3, 6, 0, 2>, - Conv::template process_tile<0, 1, 3, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 3, 6, 1, 0>, - Conv::template process_tile<0, 1, 3, 6, 1, 1>, - Conv::template process_tile<0, 1, 3, 6, 1, 2>, - Conv::template process_tile<0, 1, 3, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 3, 6, 2, 0>, - Conv::template process_tile<0, 1, 3, 6, 2, 1>, - Conv::template process_tile<0, 1, 3, 6, 2, 2>, - Conv::template process_tile<0, 1, 3, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 3, 6, 3, 0>, - Conv::template process_tile<0, 1, 3, 6, 3, 1>, - Conv::template process_tile<0, 1, 3, 6, 3, 2>, - Conv::template process_tile<0, 1, 3, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 0, 0, 0>, - Conv::template process_tile<0, 1, 4, 0, 0, 1>, - Conv::template process_tile<0, 1, 4, 0, 0, 2>, - Conv::template process_tile<0, 1, 4, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 0, 1, 0>, - Conv::template process_tile<0, 1, 4, 0, 1, 1>, - Conv::template process_tile<0, 1, 4, 0, 1, 2>, - Conv::template process_tile<0, 1, 4, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 0, 2, 0>, - Conv::template process_tile<0, 1, 4, 0, 2, 1>, - Conv::template process_tile<0, 1, 4, 0, 2, 2>, - Conv::template process_tile<0, 1, 4, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 4, 0, 3, 0>, - Conv::template process_tile<0, 1, 4, 0, 3, 1>, - Conv::template process_tile<0, 1, 4, 0, 3, 2>, - Conv::template process_tile<0, 1, 4, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 1, 0, 0>, - Conv::template process_tile<0, 1, 4, 1, 0, 1>, - Conv::template process_tile<0, 1, 4, 1, 0, 2>, - Conv::template process_tile<0, 1, 4, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 1, 1, 0>, - Conv::template process_tile<0, 1, 4, 1, 1, 1>, - Conv::template process_tile<0, 1, 4, 1, 1, 2>, - Conv::template process_tile<0, 1, 4, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 1, 2, 0>, - Conv::template process_tile<0, 1, 4, 1, 2, 1>, - Conv::template process_tile<0, 1, 4, 1, 2, 2>, - Conv::template process_tile<0, 1, 4, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 4, 1, 3, 0>, - Conv::template process_tile<0, 1, 4, 1, 3, 1>, - Conv::template process_tile<0, 1, 4, 1, 3, 2>, - Conv::template process_tile<0, 1, 4, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 2, 0, 0>, - Conv::template process_tile<0, 1, 4, 2, 0, 1>, - Conv::template process_tile<0, 1, 4, 2, 0, 2>, - Conv::template process_tile<0, 1, 4, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 2, 1, 0>, - Conv::template process_tile<0, 1, 4, 2, 1, 1>, - Conv::template process_tile<0, 1, 4, 2, 1, 2>, - Conv::template process_tile<0, 1, 4, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 2, 2, 0>, - Conv::template process_tile<0, 1, 4, 2, 2, 1>, - Conv::template process_tile<0, 1, 4, 2, 2, 2>, - Conv::template process_tile<0, 1, 4, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 4, 2, 3, 0>, - Conv::template process_tile<0, 1, 4, 2, 3, 1>, - Conv::template process_tile<0, 1, 4, 2, 3, 2>, - Conv::template process_tile<0, 1, 4, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 3, 0, 0>, - Conv::template process_tile<0, 1, 4, 3, 0, 1>, - Conv::template process_tile<0, 1, 4, 3, 0, 2>, - Conv::template process_tile<0, 1, 4, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 3, 1, 0>, - Conv::template process_tile<0, 1, 4, 3, 1, 1>, - Conv::template process_tile<0, 1, 4, 3, 1, 2>, - Conv::template process_tile<0, 1, 4, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 3, 2, 0>, - Conv::template process_tile<0, 1, 4, 3, 2, 1>, - Conv::template process_tile<0, 1, 4, 3, 2, 2>, - Conv::template process_tile<0, 1, 4, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 4, 3, 3, 0>, - Conv::template process_tile<0, 1, 4, 3, 3, 1>, - Conv::template process_tile<0, 1, 4, 3, 3, 2>, - Conv::template process_tile<0, 1, 4, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 4, 0, 0>, - Conv::template process_tile<0, 1, 4, 4, 0, 1>, - Conv::template process_tile<0, 1, 4, 4, 0, 2>, - Conv::template process_tile<0, 1, 4, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 4, 1, 0>, - Conv::template process_tile<0, 1, 4, 4, 1, 1>, - Conv::template process_tile<0, 1, 4, 4, 1, 2>, - Conv::template process_tile<0, 1, 4, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 4, 2, 0>, - Conv::template process_tile<0, 1, 4, 4, 2, 1>, - Conv::template process_tile<0, 1, 4, 4, 2, 2>, - Conv::template process_tile<0, 1, 4, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 4, 4, 3, 0>, - Conv::template process_tile<0, 1, 4, 4, 3, 1>, - Conv::template process_tile<0, 1, 4, 4, 3, 2>, - Conv::template process_tile<0, 1, 4, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 5, 0, 0>, - Conv::template process_tile<0, 1, 4, 5, 0, 1>, - Conv::template process_tile<0, 1, 4, 5, 0, 2>, - Conv::template process_tile<0, 1, 4, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 5, 1, 0>, - Conv::template process_tile<0, 1, 4, 5, 1, 1>, - Conv::template process_tile<0, 1, 4, 5, 1, 2>, - Conv::template process_tile<0, 1, 4, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 5, 2, 0>, - Conv::template process_tile<0, 1, 4, 5, 2, 1>, - Conv::template process_tile<0, 1, 4, 5, 2, 2>, - Conv::template process_tile<0, 1, 4, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 4, 5, 3, 0>, - Conv::template process_tile<0, 1, 4, 5, 3, 1>, - Conv::template process_tile<0, 1, 4, 5, 3, 2>, - Conv::template process_tile<0, 1, 4, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 4, 6, 0, 0>, - Conv::template process_tile<0, 1, 4, 6, 0, 1>, - Conv::template process_tile<0, 1, 4, 6, 0, 2>, - Conv::template process_tile<0, 1, 4, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 4, 6, 1, 0>, - Conv::template process_tile<0, 1, 4, 6, 1, 1>, - Conv::template process_tile<0, 1, 4, 6, 1, 2>, - Conv::template process_tile<0, 1, 4, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 4, 6, 2, 0>, - Conv::template process_tile<0, 1, 4, 6, 2, 1>, - Conv::template process_tile<0, 1, 4, 6, 2, 2>, - Conv::template process_tile<0, 1, 4, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 4, 6, 3, 0>, - Conv::template process_tile<0, 1, 4, 6, 3, 1>, - Conv::template process_tile<0, 1, 4, 6, 3, 2>, - Conv::template process_tile<0, 1, 4, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 4 - { // Input pad bottom = 5 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 0, 0, 0>, - Conv::template process_tile<0, 1, 5, 0, 0, 1>, - Conv::template process_tile<0, 1, 5, 0, 0, 2>, - Conv::template process_tile<0, 1, 5, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 0, 1, 0>, - Conv::template process_tile<0, 1, 5, 0, 1, 1>, - Conv::template process_tile<0, 1, 5, 0, 1, 2>, - Conv::template process_tile<0, 1, 5, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 0, 2, 0>, - Conv::template process_tile<0, 1, 5, 0, 2, 1>, - Conv::template process_tile<0, 1, 5, 0, 2, 2>, - Conv::template process_tile<0, 1, 5, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 5, 0, 3, 0>, - Conv::template process_tile<0, 1, 5, 0, 3, 1>, - Conv::template process_tile<0, 1, 5, 0, 3, 2>, - Conv::template process_tile<0, 1, 5, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 1, 0, 0>, - Conv::template process_tile<0, 1, 5, 1, 0, 1>, - Conv::template process_tile<0, 1, 5, 1, 0, 2>, - Conv::template process_tile<0, 1, 5, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 1, 1, 0>, - Conv::template process_tile<0, 1, 5, 1, 1, 1>, - Conv::template process_tile<0, 1, 5, 1, 1, 2>, - Conv::template process_tile<0, 1, 5, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 1, 2, 0>, - Conv::template process_tile<0, 1, 5, 1, 2, 1>, - Conv::template process_tile<0, 1, 5, 1, 2, 2>, - Conv::template process_tile<0, 1, 5, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 5, 1, 3, 0>, - Conv::template process_tile<0, 1, 5, 1, 3, 1>, - Conv::template process_tile<0, 1, 5, 1, 3, 2>, - Conv::template process_tile<0, 1, 5, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 2, 0, 0>, - Conv::template process_tile<0, 1, 5, 2, 0, 1>, - Conv::template process_tile<0, 1, 5, 2, 0, 2>, - Conv::template process_tile<0, 1, 5, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 2, 1, 0>, - Conv::template process_tile<0, 1, 5, 2, 1, 1>, - Conv::template process_tile<0, 1, 5, 2, 1, 2>, - Conv::template process_tile<0, 1, 5, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 2, 2, 0>, - Conv::template process_tile<0, 1, 5, 2, 2, 1>, - Conv::template process_tile<0, 1, 5, 2, 2, 2>, - Conv::template process_tile<0, 1, 5, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 5, 2, 3, 0>, - Conv::template process_tile<0, 1, 5, 2, 3, 1>, - Conv::template process_tile<0, 1, 5, 2, 3, 2>, - Conv::template process_tile<0, 1, 5, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 3, 0, 0>, - Conv::template process_tile<0, 1, 5, 3, 0, 1>, - Conv::template process_tile<0, 1, 5, 3, 0, 2>, - Conv::template process_tile<0, 1, 5, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 3, 1, 0>, - Conv::template process_tile<0, 1, 5, 3, 1, 1>, - Conv::template process_tile<0, 1, 5, 3, 1, 2>, - Conv::template process_tile<0, 1, 5, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 3, 2, 0>, - Conv::template process_tile<0, 1, 5, 3, 2, 1>, - Conv::template process_tile<0, 1, 5, 3, 2, 2>, - Conv::template process_tile<0, 1, 5, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 5, 3, 3, 0>, - Conv::template process_tile<0, 1, 5, 3, 3, 1>, - Conv::template process_tile<0, 1, 5, 3, 3, 2>, - Conv::template process_tile<0, 1, 5, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 4, 0, 0>, - Conv::template process_tile<0, 1, 5, 4, 0, 1>, - Conv::template process_tile<0, 1, 5, 4, 0, 2>, - Conv::template process_tile<0, 1, 5, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 4, 1, 0>, - Conv::template process_tile<0, 1, 5, 4, 1, 1>, - Conv::template process_tile<0, 1, 5, 4, 1, 2>, - Conv::template process_tile<0, 1, 5, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 4, 2, 0>, - Conv::template process_tile<0, 1, 5, 4, 2, 1>, - Conv::template process_tile<0, 1, 5, 4, 2, 2>, - Conv::template process_tile<0, 1, 5, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 5, 4, 3, 0>, - Conv::template process_tile<0, 1, 5, 4, 3, 1>, - Conv::template process_tile<0, 1, 5, 4, 3, 2>, - Conv::template process_tile<0, 1, 5, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 5, 0, 0>, - Conv::template process_tile<0, 1, 5, 5, 0, 1>, - Conv::template process_tile<0, 1, 5, 5, 0, 2>, - Conv::template process_tile<0, 1, 5, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 5, 1, 0>, - Conv::template process_tile<0, 1, 5, 5, 1, 1>, - Conv::template process_tile<0, 1, 5, 5, 1, 2>, - Conv::template process_tile<0, 1, 5, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 5, 2, 0>, - Conv::template process_tile<0, 1, 5, 5, 2, 1>, - Conv::template process_tile<0, 1, 5, 5, 2, 2>, - Conv::template process_tile<0, 1, 5, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 5, 5, 3, 0>, - Conv::template process_tile<0, 1, 5, 5, 3, 1>, - Conv::template process_tile<0, 1, 5, 5, 3, 2>, - Conv::template process_tile<0, 1, 5, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 5, 6, 0, 0>, - Conv::template process_tile<0, 1, 5, 6, 0, 1>, - Conv::template process_tile<0, 1, 5, 6, 0, 2>, - Conv::template process_tile<0, 1, 5, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 5, 6, 1, 0>, - Conv::template process_tile<0, 1, 5, 6, 1, 1>, - Conv::template process_tile<0, 1, 5, 6, 1, 2>, - Conv::template process_tile<0, 1, 5, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 5, 6, 2, 0>, - Conv::template process_tile<0, 1, 5, 6, 2, 1>, - Conv::template process_tile<0, 1, 5, 6, 2, 2>, - Conv::template process_tile<0, 1, 5, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 5, 6, 3, 0>, - Conv::template process_tile<0, 1, 5, 6, 3, 1>, - Conv::template process_tile<0, 1, 5, 6, 3, 2>, - Conv::template process_tile<0, 1, 5, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 5 - { // Input pad bottom = 6 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 0, 0, 0>, - Conv::template process_tile<0, 1, 6, 0, 0, 1>, - Conv::template process_tile<0, 1, 6, 0, 0, 2>, - Conv::template process_tile<0, 1, 6, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 0, 1, 0>, - Conv::template process_tile<0, 1, 6, 0, 1, 1>, - Conv::template process_tile<0, 1, 6, 0, 1, 2>, - Conv::template process_tile<0, 1, 6, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 0, 2, 0>, - Conv::template process_tile<0, 1, 6, 0, 2, 1>, - Conv::template process_tile<0, 1, 6, 0, 2, 2>, - Conv::template process_tile<0, 1, 6, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 6, 0, 3, 0>, - Conv::template process_tile<0, 1, 6, 0, 3, 1>, - Conv::template process_tile<0, 1, 6, 0, 3, 2>, - Conv::template process_tile<0, 1, 6, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 1, 0, 0>, - Conv::template process_tile<0, 1, 6, 1, 0, 1>, - Conv::template process_tile<0, 1, 6, 1, 0, 2>, - Conv::template process_tile<0, 1, 6, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 1, 1, 0>, - Conv::template process_tile<0, 1, 6, 1, 1, 1>, - Conv::template process_tile<0, 1, 6, 1, 1, 2>, - Conv::template process_tile<0, 1, 6, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 1, 2, 0>, - Conv::template process_tile<0, 1, 6, 1, 2, 1>, - Conv::template process_tile<0, 1, 6, 1, 2, 2>, - Conv::template process_tile<0, 1, 6, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 6, 1, 3, 0>, - Conv::template process_tile<0, 1, 6, 1, 3, 1>, - Conv::template process_tile<0, 1, 6, 1, 3, 2>, - Conv::template process_tile<0, 1, 6, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 2, 0, 0>, - Conv::template process_tile<0, 1, 6, 2, 0, 1>, - Conv::template process_tile<0, 1, 6, 2, 0, 2>, - Conv::template process_tile<0, 1, 6, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 2, 1, 0>, - Conv::template process_tile<0, 1, 6, 2, 1, 1>, - Conv::template process_tile<0, 1, 6, 2, 1, 2>, - Conv::template process_tile<0, 1, 6, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 2, 2, 0>, - Conv::template process_tile<0, 1, 6, 2, 2, 1>, - Conv::template process_tile<0, 1, 6, 2, 2, 2>, - Conv::template process_tile<0, 1, 6, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 6, 2, 3, 0>, - Conv::template process_tile<0, 1, 6, 2, 3, 1>, - Conv::template process_tile<0, 1, 6, 2, 3, 2>, - Conv::template process_tile<0, 1, 6, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 3, 0, 0>, - Conv::template process_tile<0, 1, 6, 3, 0, 1>, - Conv::template process_tile<0, 1, 6, 3, 0, 2>, - Conv::template process_tile<0, 1, 6, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 3, 1, 0>, - Conv::template process_tile<0, 1, 6, 3, 1, 1>, - Conv::template process_tile<0, 1, 6, 3, 1, 2>, - Conv::template process_tile<0, 1, 6, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 3, 2, 0>, - Conv::template process_tile<0, 1, 6, 3, 2, 1>, - Conv::template process_tile<0, 1, 6, 3, 2, 2>, - Conv::template process_tile<0, 1, 6, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 6, 3, 3, 0>, - Conv::template process_tile<0, 1, 6, 3, 3, 1>, - Conv::template process_tile<0, 1, 6, 3, 3, 2>, - Conv::template process_tile<0, 1, 6, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 4, 0, 0>, - Conv::template process_tile<0, 1, 6, 4, 0, 1>, - Conv::template process_tile<0, 1, 6, 4, 0, 2>, - Conv::template process_tile<0, 1, 6, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 4, 1, 0>, - Conv::template process_tile<0, 1, 6, 4, 1, 1>, - Conv::template process_tile<0, 1, 6, 4, 1, 2>, - Conv::template process_tile<0, 1, 6, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 4, 2, 0>, - Conv::template process_tile<0, 1, 6, 4, 2, 1>, - Conv::template process_tile<0, 1, 6, 4, 2, 2>, - Conv::template process_tile<0, 1, 6, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 6, 4, 3, 0>, - Conv::template process_tile<0, 1, 6, 4, 3, 1>, - Conv::template process_tile<0, 1, 6, 4, 3, 2>, - Conv::template process_tile<0, 1, 6, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 5, 0, 0>, - Conv::template process_tile<0, 1, 6, 5, 0, 1>, - Conv::template process_tile<0, 1, 6, 5, 0, 2>, - Conv::template process_tile<0, 1, 6, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 5, 1, 0>, - Conv::template process_tile<0, 1, 6, 5, 1, 1>, - Conv::template process_tile<0, 1, 6, 5, 1, 2>, - Conv::template process_tile<0, 1, 6, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 5, 2, 0>, - Conv::template process_tile<0, 1, 6, 5, 2, 1>, - Conv::template process_tile<0, 1, 6, 5, 2, 2>, - Conv::template process_tile<0, 1, 6, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 6, 5, 3, 0>, - Conv::template process_tile<0, 1, 6, 5, 3, 1>, - Conv::template process_tile<0, 1, 6, 5, 3, 2>, - Conv::template process_tile<0, 1, 6, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<0, 1, 6, 6, 0, 0>, - Conv::template process_tile<0, 1, 6, 6, 0, 1>, - Conv::template process_tile<0, 1, 6, 6, 0, 2>, - Conv::template process_tile<0, 1, 6, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<0, 1, 6, 6, 1, 0>, - Conv::template process_tile<0, 1, 6, 6, 1, 1>, - Conv::template process_tile<0, 1, 6, 6, 1, 2>, - Conv::template process_tile<0, 1, 6, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<0, 1, 6, 6, 2, 0>, - Conv::template process_tile<0, 1, 6, 6, 2, 1>, - Conv::template process_tile<0, 1, 6, 6, 2, 2>, - Conv::template process_tile<0, 1, 6, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<0, 1, 6, 6, 3, 0>, - Conv::template process_tile<0, 1, 6, 6, 3, 1>, - Conv::template process_tile<0, 1, 6, 6, 3, 2>, - Conv::template process_tile<0, 1, 6, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 6 - }, // Input pad left = 1 - }, // Input pad top = 0 - { // Input pad top = 1 - { // Input pad left = 0 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 0, 0, 0>, - Conv::template process_tile<1, 0, 0, 0, 0, 1>, - Conv::template process_tile<1, 0, 0, 0, 0, 2>, - Conv::template process_tile<1, 0, 0, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 0, 1, 0>, - Conv::template process_tile<1, 0, 0, 0, 1, 1>, - Conv::template process_tile<1, 0, 0, 0, 1, 2>, - Conv::template process_tile<1, 0, 0, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 0, 2, 0>, - Conv::template process_tile<1, 0, 0, 0, 2, 1>, - Conv::template process_tile<1, 0, 0, 0, 2, 2>, - Conv::template process_tile<1, 0, 0, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 0, 0, 3, 0>, - Conv::template process_tile<1, 0, 0, 0, 3, 1>, - Conv::template process_tile<1, 0, 0, 0, 3, 2>, - Conv::template process_tile<1, 0, 0, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 1, 0, 0>, - Conv::template process_tile<1, 0, 0, 1, 0, 1>, - Conv::template process_tile<1, 0, 0, 1, 0, 2>, - Conv::template process_tile<1, 0, 0, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 1, 1, 0>, - Conv::template process_tile<1, 0, 0, 1, 1, 1>, - Conv::template process_tile<1, 0, 0, 1, 1, 2>, - Conv::template process_tile<1, 0, 0, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 1, 2, 0>, - Conv::template process_tile<1, 0, 0, 1, 2, 1>, - Conv::template process_tile<1, 0, 0, 1, 2, 2>, - Conv::template process_tile<1, 0, 0, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 0, 1, 3, 0>, - Conv::template process_tile<1, 0, 0, 1, 3, 1>, - Conv::template process_tile<1, 0, 0, 1, 3, 2>, - Conv::template process_tile<1, 0, 0, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 2, 0, 0>, - Conv::template process_tile<1, 0, 0, 2, 0, 1>, - Conv::template process_tile<1, 0, 0, 2, 0, 2>, - Conv::template process_tile<1, 0, 0, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 2, 1, 0>, - Conv::template process_tile<1, 0, 0, 2, 1, 1>, - Conv::template process_tile<1, 0, 0, 2, 1, 2>, - Conv::template process_tile<1, 0, 0, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 2, 2, 0>, - Conv::template process_tile<1, 0, 0, 2, 2, 1>, - Conv::template process_tile<1, 0, 0, 2, 2, 2>, - Conv::template process_tile<1, 0, 0, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 0, 2, 3, 0>, - Conv::template process_tile<1, 0, 0, 2, 3, 1>, - Conv::template process_tile<1, 0, 0, 2, 3, 2>, - Conv::template process_tile<1, 0, 0, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 3, 0, 0>, - Conv::template process_tile<1, 0, 0, 3, 0, 1>, - Conv::template process_tile<1, 0, 0, 3, 0, 2>, - Conv::template process_tile<1, 0, 0, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 3, 1, 0>, - Conv::template process_tile<1, 0, 0, 3, 1, 1>, - Conv::template process_tile<1, 0, 0, 3, 1, 2>, - Conv::template process_tile<1, 0, 0, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 3, 2, 0>, - Conv::template process_tile<1, 0, 0, 3, 2, 1>, - Conv::template process_tile<1, 0, 0, 3, 2, 2>, - Conv::template process_tile<1, 0, 0, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 0, 3, 3, 0>, - Conv::template process_tile<1, 0, 0, 3, 3, 1>, - Conv::template process_tile<1, 0, 0, 3, 3, 2>, - Conv::template process_tile<1, 0, 0, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 4, 0, 0>, - Conv::template process_tile<1, 0, 0, 4, 0, 1>, - Conv::template process_tile<1, 0, 0, 4, 0, 2>, - Conv::template process_tile<1, 0, 0, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 4, 1, 0>, - Conv::template process_tile<1, 0, 0, 4, 1, 1>, - Conv::template process_tile<1, 0, 0, 4, 1, 2>, - Conv::template process_tile<1, 0, 0, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 4, 2, 0>, - Conv::template process_tile<1, 0, 0, 4, 2, 1>, - Conv::template process_tile<1, 0, 0, 4, 2, 2>, - Conv::template process_tile<1, 0, 0, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 0, 4, 3, 0>, - Conv::template process_tile<1, 0, 0, 4, 3, 1>, - Conv::template process_tile<1, 0, 0, 4, 3, 2>, - Conv::template process_tile<1, 0, 0, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 5, 0, 0>, - Conv::template process_tile<1, 0, 0, 5, 0, 1>, - Conv::template process_tile<1, 0, 0, 5, 0, 2>, - Conv::template process_tile<1, 0, 0, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 5, 1, 0>, - Conv::template process_tile<1, 0, 0, 5, 1, 1>, - Conv::template process_tile<1, 0, 0, 5, 1, 2>, - Conv::template process_tile<1, 0, 0, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 5, 2, 0>, - Conv::template process_tile<1, 0, 0, 5, 2, 1>, - Conv::template process_tile<1, 0, 0, 5, 2, 2>, - Conv::template process_tile<1, 0, 0, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 0, 5, 3, 0>, - Conv::template process_tile<1, 0, 0, 5, 3, 1>, - Conv::template process_tile<1, 0, 0, 5, 3, 2>, - Conv::template process_tile<1, 0, 0, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 0, 6, 0, 0>, - Conv::template process_tile<1, 0, 0, 6, 0, 1>, - Conv::template process_tile<1, 0, 0, 6, 0, 2>, - Conv::template process_tile<1, 0, 0, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 0, 6, 1, 0>, - Conv::template process_tile<1, 0, 0, 6, 1, 1>, - Conv::template process_tile<1, 0, 0, 6, 1, 2>, - Conv::template process_tile<1, 0, 0, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 0, 6, 2, 0>, - Conv::template process_tile<1, 0, 0, 6, 2, 1>, - Conv::template process_tile<1, 0, 0, 6, 2, 2>, - Conv::template process_tile<1, 0, 0, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 0, 6, 3, 0>, - Conv::template process_tile<1, 0, 0, 6, 3, 1>, - Conv::template process_tile<1, 0, 0, 6, 3, 2>, - Conv::template process_tile<1, 0, 0, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 0, 0, 0>, - Conv::template process_tile<1, 0, 1, 0, 0, 1>, - Conv::template process_tile<1, 0, 1, 0, 0, 2>, - Conv::template process_tile<1, 0, 1, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 0, 1, 0>, - Conv::template process_tile<1, 0, 1, 0, 1, 1>, - Conv::template process_tile<1, 0, 1, 0, 1, 2>, - Conv::template process_tile<1, 0, 1, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 0, 2, 0>, - Conv::template process_tile<1, 0, 1, 0, 2, 1>, - Conv::template process_tile<1, 0, 1, 0, 2, 2>, - Conv::template process_tile<1, 0, 1, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 1, 0, 3, 0>, - Conv::template process_tile<1, 0, 1, 0, 3, 1>, - Conv::template process_tile<1, 0, 1, 0, 3, 2>, - Conv::template process_tile<1, 0, 1, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 1, 0, 0>, - Conv::template process_tile<1, 0, 1, 1, 0, 1>, - Conv::template process_tile<1, 0, 1, 1, 0, 2>, - Conv::template process_tile<1, 0, 1, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 1, 1, 0>, - Conv::template process_tile<1, 0, 1, 1, 1, 1>, - Conv::template process_tile<1, 0, 1, 1, 1, 2>, - Conv::template process_tile<1, 0, 1, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 1, 2, 0>, - Conv::template process_tile<1, 0, 1, 1, 2, 1>, - Conv::template process_tile<1, 0, 1, 1, 2, 2>, - Conv::template process_tile<1, 0, 1, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 1, 1, 3, 0>, - Conv::template process_tile<1, 0, 1, 1, 3, 1>, - Conv::template process_tile<1, 0, 1, 1, 3, 2>, - Conv::template process_tile<1, 0, 1, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 2, 0, 0>, - Conv::template process_tile<1, 0, 1, 2, 0, 1>, - Conv::template process_tile<1, 0, 1, 2, 0, 2>, - Conv::template process_tile<1, 0, 1, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 2, 1, 0>, - Conv::template process_tile<1, 0, 1, 2, 1, 1>, - Conv::template process_tile<1, 0, 1, 2, 1, 2>, - Conv::template process_tile<1, 0, 1, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 2, 2, 0>, - Conv::template process_tile<1, 0, 1, 2, 2, 1>, - Conv::template process_tile<1, 0, 1, 2, 2, 2>, - Conv::template process_tile<1, 0, 1, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 1, 2, 3, 0>, - Conv::template process_tile<1, 0, 1, 2, 3, 1>, - Conv::template process_tile<1, 0, 1, 2, 3, 2>, - Conv::template process_tile<1, 0, 1, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 3, 0, 0>, - Conv::template process_tile<1, 0, 1, 3, 0, 1>, - Conv::template process_tile<1, 0, 1, 3, 0, 2>, - Conv::template process_tile<1, 0, 1, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 3, 1, 0>, - Conv::template process_tile<1, 0, 1, 3, 1, 1>, - Conv::template process_tile<1, 0, 1, 3, 1, 2>, - Conv::template process_tile<1, 0, 1, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 3, 2, 0>, - Conv::template process_tile<1, 0, 1, 3, 2, 1>, - Conv::template process_tile<1, 0, 1, 3, 2, 2>, - Conv::template process_tile<1, 0, 1, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 1, 3, 3, 0>, - Conv::template process_tile<1, 0, 1, 3, 3, 1>, - Conv::template process_tile<1, 0, 1, 3, 3, 2>, - Conv::template process_tile<1, 0, 1, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 4, 0, 0>, - Conv::template process_tile<1, 0, 1, 4, 0, 1>, - Conv::template process_tile<1, 0, 1, 4, 0, 2>, - Conv::template process_tile<1, 0, 1, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 4, 1, 0>, - Conv::template process_tile<1, 0, 1, 4, 1, 1>, - Conv::template process_tile<1, 0, 1, 4, 1, 2>, - Conv::template process_tile<1, 0, 1, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 4, 2, 0>, - Conv::template process_tile<1, 0, 1, 4, 2, 1>, - Conv::template process_tile<1, 0, 1, 4, 2, 2>, - Conv::template process_tile<1, 0, 1, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 1, 4, 3, 0>, - Conv::template process_tile<1, 0, 1, 4, 3, 1>, - Conv::template process_tile<1, 0, 1, 4, 3, 2>, - Conv::template process_tile<1, 0, 1, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 5, 0, 0>, - Conv::template process_tile<1, 0, 1, 5, 0, 1>, - Conv::template process_tile<1, 0, 1, 5, 0, 2>, - Conv::template process_tile<1, 0, 1, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 5, 1, 0>, - Conv::template process_tile<1, 0, 1, 5, 1, 1>, - Conv::template process_tile<1, 0, 1, 5, 1, 2>, - Conv::template process_tile<1, 0, 1, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 5, 2, 0>, - Conv::template process_tile<1, 0, 1, 5, 2, 1>, - Conv::template process_tile<1, 0, 1, 5, 2, 2>, - Conv::template process_tile<1, 0, 1, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 1, 5, 3, 0>, - Conv::template process_tile<1, 0, 1, 5, 3, 1>, - Conv::template process_tile<1, 0, 1, 5, 3, 2>, - Conv::template process_tile<1, 0, 1, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 1, 6, 0, 0>, - Conv::template process_tile<1, 0, 1, 6, 0, 1>, - Conv::template process_tile<1, 0, 1, 6, 0, 2>, - Conv::template process_tile<1, 0, 1, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 1, 6, 1, 0>, - Conv::template process_tile<1, 0, 1, 6, 1, 1>, - Conv::template process_tile<1, 0, 1, 6, 1, 2>, - Conv::template process_tile<1, 0, 1, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 1, 6, 2, 0>, - Conv::template process_tile<1, 0, 1, 6, 2, 1>, - Conv::template process_tile<1, 0, 1, 6, 2, 2>, - Conv::template process_tile<1, 0, 1, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 1, 6, 3, 0>, - Conv::template process_tile<1, 0, 1, 6, 3, 1>, - Conv::template process_tile<1, 0, 1, 6, 3, 2>, - Conv::template process_tile<1, 0, 1, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 0, 0, 0>, - Conv::template process_tile<1, 0, 2, 0, 0, 1>, - Conv::template process_tile<1, 0, 2, 0, 0, 2>, - Conv::template process_tile<1, 0, 2, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 0, 1, 0>, - Conv::template process_tile<1, 0, 2, 0, 1, 1>, - Conv::template process_tile<1, 0, 2, 0, 1, 2>, - Conv::template process_tile<1, 0, 2, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 0, 2, 0>, - Conv::template process_tile<1, 0, 2, 0, 2, 1>, - Conv::template process_tile<1, 0, 2, 0, 2, 2>, - Conv::template process_tile<1, 0, 2, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 2, 0, 3, 0>, - Conv::template process_tile<1, 0, 2, 0, 3, 1>, - Conv::template process_tile<1, 0, 2, 0, 3, 2>, - Conv::template process_tile<1, 0, 2, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 1, 0, 0>, - Conv::template process_tile<1, 0, 2, 1, 0, 1>, - Conv::template process_tile<1, 0, 2, 1, 0, 2>, - Conv::template process_tile<1, 0, 2, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 1, 1, 0>, - Conv::template process_tile<1, 0, 2, 1, 1, 1>, - Conv::template process_tile<1, 0, 2, 1, 1, 2>, - Conv::template process_tile<1, 0, 2, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 1, 2, 0>, - Conv::template process_tile<1, 0, 2, 1, 2, 1>, - Conv::template process_tile<1, 0, 2, 1, 2, 2>, - Conv::template process_tile<1, 0, 2, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 2, 1, 3, 0>, - Conv::template process_tile<1, 0, 2, 1, 3, 1>, - Conv::template process_tile<1, 0, 2, 1, 3, 2>, - Conv::template process_tile<1, 0, 2, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 2, 0, 0>, - Conv::template process_tile<1, 0, 2, 2, 0, 1>, - Conv::template process_tile<1, 0, 2, 2, 0, 2>, - Conv::template process_tile<1, 0, 2, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 2, 1, 0>, - Conv::template process_tile<1, 0, 2, 2, 1, 1>, - Conv::template process_tile<1, 0, 2, 2, 1, 2>, - Conv::template process_tile<1, 0, 2, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 2, 2, 0>, - Conv::template process_tile<1, 0, 2, 2, 2, 1>, - Conv::template process_tile<1, 0, 2, 2, 2, 2>, - Conv::template process_tile<1, 0, 2, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 2, 2, 3, 0>, - Conv::template process_tile<1, 0, 2, 2, 3, 1>, - Conv::template process_tile<1, 0, 2, 2, 3, 2>, - Conv::template process_tile<1, 0, 2, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 3, 0, 0>, - Conv::template process_tile<1, 0, 2, 3, 0, 1>, - Conv::template process_tile<1, 0, 2, 3, 0, 2>, - Conv::template process_tile<1, 0, 2, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 3, 1, 0>, - Conv::template process_tile<1, 0, 2, 3, 1, 1>, - Conv::template process_tile<1, 0, 2, 3, 1, 2>, - Conv::template process_tile<1, 0, 2, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 3, 2, 0>, - Conv::template process_tile<1, 0, 2, 3, 2, 1>, - Conv::template process_tile<1, 0, 2, 3, 2, 2>, - Conv::template process_tile<1, 0, 2, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 2, 3, 3, 0>, - Conv::template process_tile<1, 0, 2, 3, 3, 1>, - Conv::template process_tile<1, 0, 2, 3, 3, 2>, - Conv::template process_tile<1, 0, 2, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 4, 0, 0>, - Conv::template process_tile<1, 0, 2, 4, 0, 1>, - Conv::template process_tile<1, 0, 2, 4, 0, 2>, - Conv::template process_tile<1, 0, 2, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 4, 1, 0>, - Conv::template process_tile<1, 0, 2, 4, 1, 1>, - Conv::template process_tile<1, 0, 2, 4, 1, 2>, - Conv::template process_tile<1, 0, 2, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 4, 2, 0>, - Conv::template process_tile<1, 0, 2, 4, 2, 1>, - Conv::template process_tile<1, 0, 2, 4, 2, 2>, - Conv::template process_tile<1, 0, 2, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 2, 4, 3, 0>, - Conv::template process_tile<1, 0, 2, 4, 3, 1>, - Conv::template process_tile<1, 0, 2, 4, 3, 2>, - Conv::template process_tile<1, 0, 2, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 5, 0, 0>, - Conv::template process_tile<1, 0, 2, 5, 0, 1>, - Conv::template process_tile<1, 0, 2, 5, 0, 2>, - Conv::template process_tile<1, 0, 2, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 5, 1, 0>, - Conv::template process_tile<1, 0, 2, 5, 1, 1>, - Conv::template process_tile<1, 0, 2, 5, 1, 2>, - Conv::template process_tile<1, 0, 2, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 5, 2, 0>, - Conv::template process_tile<1, 0, 2, 5, 2, 1>, - Conv::template process_tile<1, 0, 2, 5, 2, 2>, - Conv::template process_tile<1, 0, 2, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 2, 5, 3, 0>, - Conv::template process_tile<1, 0, 2, 5, 3, 1>, - Conv::template process_tile<1, 0, 2, 5, 3, 2>, - Conv::template process_tile<1, 0, 2, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 2, 6, 0, 0>, - Conv::template process_tile<1, 0, 2, 6, 0, 1>, - Conv::template process_tile<1, 0, 2, 6, 0, 2>, - Conv::template process_tile<1, 0, 2, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 2, 6, 1, 0>, - Conv::template process_tile<1, 0, 2, 6, 1, 1>, - Conv::template process_tile<1, 0, 2, 6, 1, 2>, - Conv::template process_tile<1, 0, 2, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 2, 6, 2, 0>, - Conv::template process_tile<1, 0, 2, 6, 2, 1>, - Conv::template process_tile<1, 0, 2, 6, 2, 2>, - Conv::template process_tile<1, 0, 2, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 2, 6, 3, 0>, - Conv::template process_tile<1, 0, 2, 6, 3, 1>, - Conv::template process_tile<1, 0, 2, 6, 3, 2>, - Conv::template process_tile<1, 0, 2, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 0, 0, 0>, - Conv::template process_tile<1, 0, 3, 0, 0, 1>, - Conv::template process_tile<1, 0, 3, 0, 0, 2>, - Conv::template process_tile<1, 0, 3, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 0, 1, 0>, - Conv::template process_tile<1, 0, 3, 0, 1, 1>, - Conv::template process_tile<1, 0, 3, 0, 1, 2>, - Conv::template process_tile<1, 0, 3, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 0, 2, 0>, - Conv::template process_tile<1, 0, 3, 0, 2, 1>, - Conv::template process_tile<1, 0, 3, 0, 2, 2>, - Conv::template process_tile<1, 0, 3, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 3, 0, 3, 0>, - Conv::template process_tile<1, 0, 3, 0, 3, 1>, - Conv::template process_tile<1, 0, 3, 0, 3, 2>, - Conv::template process_tile<1, 0, 3, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 1, 0, 0>, - Conv::template process_tile<1, 0, 3, 1, 0, 1>, - Conv::template process_tile<1, 0, 3, 1, 0, 2>, - Conv::template process_tile<1, 0, 3, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 1, 1, 0>, - Conv::template process_tile<1, 0, 3, 1, 1, 1>, - Conv::template process_tile<1, 0, 3, 1, 1, 2>, - Conv::template process_tile<1, 0, 3, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 1, 2, 0>, - Conv::template process_tile<1, 0, 3, 1, 2, 1>, - Conv::template process_tile<1, 0, 3, 1, 2, 2>, - Conv::template process_tile<1, 0, 3, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 3, 1, 3, 0>, - Conv::template process_tile<1, 0, 3, 1, 3, 1>, - Conv::template process_tile<1, 0, 3, 1, 3, 2>, - Conv::template process_tile<1, 0, 3, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 2, 0, 0>, - Conv::template process_tile<1, 0, 3, 2, 0, 1>, - Conv::template process_tile<1, 0, 3, 2, 0, 2>, - Conv::template process_tile<1, 0, 3, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 2, 1, 0>, - Conv::template process_tile<1, 0, 3, 2, 1, 1>, - Conv::template process_tile<1, 0, 3, 2, 1, 2>, - Conv::template process_tile<1, 0, 3, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 2, 2, 0>, - Conv::template process_tile<1, 0, 3, 2, 2, 1>, - Conv::template process_tile<1, 0, 3, 2, 2, 2>, - Conv::template process_tile<1, 0, 3, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 3, 2, 3, 0>, - Conv::template process_tile<1, 0, 3, 2, 3, 1>, - Conv::template process_tile<1, 0, 3, 2, 3, 2>, - Conv::template process_tile<1, 0, 3, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 3, 0, 0>, - Conv::template process_tile<1, 0, 3, 3, 0, 1>, - Conv::template process_tile<1, 0, 3, 3, 0, 2>, - Conv::template process_tile<1, 0, 3, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 3, 1, 0>, - Conv::template process_tile<1, 0, 3, 3, 1, 1>, - Conv::template process_tile<1, 0, 3, 3, 1, 2>, - Conv::template process_tile<1, 0, 3, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 3, 2, 0>, - Conv::template process_tile<1, 0, 3, 3, 2, 1>, - Conv::template process_tile<1, 0, 3, 3, 2, 2>, - Conv::template process_tile<1, 0, 3, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 3, 3, 3, 0>, - Conv::template process_tile<1, 0, 3, 3, 3, 1>, - Conv::template process_tile<1, 0, 3, 3, 3, 2>, - Conv::template process_tile<1, 0, 3, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 4, 0, 0>, - Conv::template process_tile<1, 0, 3, 4, 0, 1>, - Conv::template process_tile<1, 0, 3, 4, 0, 2>, - Conv::template process_tile<1, 0, 3, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 4, 1, 0>, - Conv::template process_tile<1, 0, 3, 4, 1, 1>, - Conv::template process_tile<1, 0, 3, 4, 1, 2>, - Conv::template process_tile<1, 0, 3, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 4, 2, 0>, - Conv::template process_tile<1, 0, 3, 4, 2, 1>, - Conv::template process_tile<1, 0, 3, 4, 2, 2>, - Conv::template process_tile<1, 0, 3, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 3, 4, 3, 0>, - Conv::template process_tile<1, 0, 3, 4, 3, 1>, - Conv::template process_tile<1, 0, 3, 4, 3, 2>, - Conv::template process_tile<1, 0, 3, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 5, 0, 0>, - Conv::template process_tile<1, 0, 3, 5, 0, 1>, - Conv::template process_tile<1, 0, 3, 5, 0, 2>, - Conv::template process_tile<1, 0, 3, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 5, 1, 0>, - Conv::template process_tile<1, 0, 3, 5, 1, 1>, - Conv::template process_tile<1, 0, 3, 5, 1, 2>, - Conv::template process_tile<1, 0, 3, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 5, 2, 0>, - Conv::template process_tile<1, 0, 3, 5, 2, 1>, - Conv::template process_tile<1, 0, 3, 5, 2, 2>, - Conv::template process_tile<1, 0, 3, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 3, 5, 3, 0>, - Conv::template process_tile<1, 0, 3, 5, 3, 1>, - Conv::template process_tile<1, 0, 3, 5, 3, 2>, - Conv::template process_tile<1, 0, 3, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 3, 6, 0, 0>, - Conv::template process_tile<1, 0, 3, 6, 0, 1>, - Conv::template process_tile<1, 0, 3, 6, 0, 2>, - Conv::template process_tile<1, 0, 3, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 3, 6, 1, 0>, - Conv::template process_tile<1, 0, 3, 6, 1, 1>, - Conv::template process_tile<1, 0, 3, 6, 1, 2>, - Conv::template process_tile<1, 0, 3, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 3, 6, 2, 0>, - Conv::template process_tile<1, 0, 3, 6, 2, 1>, - Conv::template process_tile<1, 0, 3, 6, 2, 2>, - Conv::template process_tile<1, 0, 3, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 3, 6, 3, 0>, - Conv::template process_tile<1, 0, 3, 6, 3, 1>, - Conv::template process_tile<1, 0, 3, 6, 3, 2>, - Conv::template process_tile<1, 0, 3, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 0, 0, 0>, - Conv::template process_tile<1, 0, 4, 0, 0, 1>, - Conv::template process_tile<1, 0, 4, 0, 0, 2>, - Conv::template process_tile<1, 0, 4, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 0, 1, 0>, - Conv::template process_tile<1, 0, 4, 0, 1, 1>, - Conv::template process_tile<1, 0, 4, 0, 1, 2>, - Conv::template process_tile<1, 0, 4, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 0, 2, 0>, - Conv::template process_tile<1, 0, 4, 0, 2, 1>, - Conv::template process_tile<1, 0, 4, 0, 2, 2>, - Conv::template process_tile<1, 0, 4, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 4, 0, 3, 0>, - Conv::template process_tile<1, 0, 4, 0, 3, 1>, - Conv::template process_tile<1, 0, 4, 0, 3, 2>, - Conv::template process_tile<1, 0, 4, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 1, 0, 0>, - Conv::template process_tile<1, 0, 4, 1, 0, 1>, - Conv::template process_tile<1, 0, 4, 1, 0, 2>, - Conv::template process_tile<1, 0, 4, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 1, 1, 0>, - Conv::template process_tile<1, 0, 4, 1, 1, 1>, - Conv::template process_tile<1, 0, 4, 1, 1, 2>, - Conv::template process_tile<1, 0, 4, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 1, 2, 0>, - Conv::template process_tile<1, 0, 4, 1, 2, 1>, - Conv::template process_tile<1, 0, 4, 1, 2, 2>, - Conv::template process_tile<1, 0, 4, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 4, 1, 3, 0>, - Conv::template process_tile<1, 0, 4, 1, 3, 1>, - Conv::template process_tile<1, 0, 4, 1, 3, 2>, - Conv::template process_tile<1, 0, 4, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 2, 0, 0>, - Conv::template process_tile<1, 0, 4, 2, 0, 1>, - Conv::template process_tile<1, 0, 4, 2, 0, 2>, - Conv::template process_tile<1, 0, 4, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 2, 1, 0>, - Conv::template process_tile<1, 0, 4, 2, 1, 1>, - Conv::template process_tile<1, 0, 4, 2, 1, 2>, - Conv::template process_tile<1, 0, 4, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 2, 2, 0>, - Conv::template process_tile<1, 0, 4, 2, 2, 1>, - Conv::template process_tile<1, 0, 4, 2, 2, 2>, - Conv::template process_tile<1, 0, 4, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 4, 2, 3, 0>, - Conv::template process_tile<1, 0, 4, 2, 3, 1>, - Conv::template process_tile<1, 0, 4, 2, 3, 2>, - Conv::template process_tile<1, 0, 4, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 3, 0, 0>, - Conv::template process_tile<1, 0, 4, 3, 0, 1>, - Conv::template process_tile<1, 0, 4, 3, 0, 2>, - Conv::template process_tile<1, 0, 4, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 3, 1, 0>, - Conv::template process_tile<1, 0, 4, 3, 1, 1>, - Conv::template process_tile<1, 0, 4, 3, 1, 2>, - Conv::template process_tile<1, 0, 4, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 3, 2, 0>, - Conv::template process_tile<1, 0, 4, 3, 2, 1>, - Conv::template process_tile<1, 0, 4, 3, 2, 2>, - Conv::template process_tile<1, 0, 4, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 4, 3, 3, 0>, - Conv::template process_tile<1, 0, 4, 3, 3, 1>, - Conv::template process_tile<1, 0, 4, 3, 3, 2>, - Conv::template process_tile<1, 0, 4, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 4, 0, 0>, - Conv::template process_tile<1, 0, 4, 4, 0, 1>, - Conv::template process_tile<1, 0, 4, 4, 0, 2>, - Conv::template process_tile<1, 0, 4, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 4, 1, 0>, - Conv::template process_tile<1, 0, 4, 4, 1, 1>, - Conv::template process_tile<1, 0, 4, 4, 1, 2>, - Conv::template process_tile<1, 0, 4, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 4, 2, 0>, - Conv::template process_tile<1, 0, 4, 4, 2, 1>, - Conv::template process_tile<1, 0, 4, 4, 2, 2>, - Conv::template process_tile<1, 0, 4, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 4, 4, 3, 0>, - Conv::template process_tile<1, 0, 4, 4, 3, 1>, - Conv::template process_tile<1, 0, 4, 4, 3, 2>, - Conv::template process_tile<1, 0, 4, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 5, 0, 0>, - Conv::template process_tile<1, 0, 4, 5, 0, 1>, - Conv::template process_tile<1, 0, 4, 5, 0, 2>, - Conv::template process_tile<1, 0, 4, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 5, 1, 0>, - Conv::template process_tile<1, 0, 4, 5, 1, 1>, - Conv::template process_tile<1, 0, 4, 5, 1, 2>, - Conv::template process_tile<1, 0, 4, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 5, 2, 0>, - Conv::template process_tile<1, 0, 4, 5, 2, 1>, - Conv::template process_tile<1, 0, 4, 5, 2, 2>, - Conv::template process_tile<1, 0, 4, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 4, 5, 3, 0>, - Conv::template process_tile<1, 0, 4, 5, 3, 1>, - Conv::template process_tile<1, 0, 4, 5, 3, 2>, - Conv::template process_tile<1, 0, 4, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 4, 6, 0, 0>, - Conv::template process_tile<1, 0, 4, 6, 0, 1>, - Conv::template process_tile<1, 0, 4, 6, 0, 2>, - Conv::template process_tile<1, 0, 4, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 4, 6, 1, 0>, - Conv::template process_tile<1, 0, 4, 6, 1, 1>, - Conv::template process_tile<1, 0, 4, 6, 1, 2>, - Conv::template process_tile<1, 0, 4, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 4, 6, 2, 0>, - Conv::template process_tile<1, 0, 4, 6, 2, 1>, - Conv::template process_tile<1, 0, 4, 6, 2, 2>, - Conv::template process_tile<1, 0, 4, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 4, 6, 3, 0>, - Conv::template process_tile<1, 0, 4, 6, 3, 1>, - Conv::template process_tile<1, 0, 4, 6, 3, 2>, - Conv::template process_tile<1, 0, 4, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 4 - { // Input pad bottom = 5 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 0, 0, 0>, - Conv::template process_tile<1, 0, 5, 0, 0, 1>, - Conv::template process_tile<1, 0, 5, 0, 0, 2>, - Conv::template process_tile<1, 0, 5, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 0, 1, 0>, - Conv::template process_tile<1, 0, 5, 0, 1, 1>, - Conv::template process_tile<1, 0, 5, 0, 1, 2>, - Conv::template process_tile<1, 0, 5, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 0, 2, 0>, - Conv::template process_tile<1, 0, 5, 0, 2, 1>, - Conv::template process_tile<1, 0, 5, 0, 2, 2>, - Conv::template process_tile<1, 0, 5, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 5, 0, 3, 0>, - Conv::template process_tile<1, 0, 5, 0, 3, 1>, - Conv::template process_tile<1, 0, 5, 0, 3, 2>, - Conv::template process_tile<1, 0, 5, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 1, 0, 0>, - Conv::template process_tile<1, 0, 5, 1, 0, 1>, - Conv::template process_tile<1, 0, 5, 1, 0, 2>, - Conv::template process_tile<1, 0, 5, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 1, 1, 0>, - Conv::template process_tile<1, 0, 5, 1, 1, 1>, - Conv::template process_tile<1, 0, 5, 1, 1, 2>, - Conv::template process_tile<1, 0, 5, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 1, 2, 0>, - Conv::template process_tile<1, 0, 5, 1, 2, 1>, - Conv::template process_tile<1, 0, 5, 1, 2, 2>, - Conv::template process_tile<1, 0, 5, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 5, 1, 3, 0>, - Conv::template process_tile<1, 0, 5, 1, 3, 1>, - Conv::template process_tile<1, 0, 5, 1, 3, 2>, - Conv::template process_tile<1, 0, 5, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 2, 0, 0>, - Conv::template process_tile<1, 0, 5, 2, 0, 1>, - Conv::template process_tile<1, 0, 5, 2, 0, 2>, - Conv::template process_tile<1, 0, 5, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 2, 1, 0>, - Conv::template process_tile<1, 0, 5, 2, 1, 1>, - Conv::template process_tile<1, 0, 5, 2, 1, 2>, - Conv::template process_tile<1, 0, 5, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 2, 2, 0>, - Conv::template process_tile<1, 0, 5, 2, 2, 1>, - Conv::template process_tile<1, 0, 5, 2, 2, 2>, - Conv::template process_tile<1, 0, 5, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 5, 2, 3, 0>, - Conv::template process_tile<1, 0, 5, 2, 3, 1>, - Conv::template process_tile<1, 0, 5, 2, 3, 2>, - Conv::template process_tile<1, 0, 5, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 3, 0, 0>, - Conv::template process_tile<1, 0, 5, 3, 0, 1>, - Conv::template process_tile<1, 0, 5, 3, 0, 2>, - Conv::template process_tile<1, 0, 5, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 3, 1, 0>, - Conv::template process_tile<1, 0, 5, 3, 1, 1>, - Conv::template process_tile<1, 0, 5, 3, 1, 2>, - Conv::template process_tile<1, 0, 5, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 3, 2, 0>, - Conv::template process_tile<1, 0, 5, 3, 2, 1>, - Conv::template process_tile<1, 0, 5, 3, 2, 2>, - Conv::template process_tile<1, 0, 5, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 5, 3, 3, 0>, - Conv::template process_tile<1, 0, 5, 3, 3, 1>, - Conv::template process_tile<1, 0, 5, 3, 3, 2>, - Conv::template process_tile<1, 0, 5, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 4, 0, 0>, - Conv::template process_tile<1, 0, 5, 4, 0, 1>, - Conv::template process_tile<1, 0, 5, 4, 0, 2>, - Conv::template process_tile<1, 0, 5, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 4, 1, 0>, - Conv::template process_tile<1, 0, 5, 4, 1, 1>, - Conv::template process_tile<1, 0, 5, 4, 1, 2>, - Conv::template process_tile<1, 0, 5, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 4, 2, 0>, - Conv::template process_tile<1, 0, 5, 4, 2, 1>, - Conv::template process_tile<1, 0, 5, 4, 2, 2>, - Conv::template process_tile<1, 0, 5, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 5, 4, 3, 0>, - Conv::template process_tile<1, 0, 5, 4, 3, 1>, - Conv::template process_tile<1, 0, 5, 4, 3, 2>, - Conv::template process_tile<1, 0, 5, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 5, 0, 0>, - Conv::template process_tile<1, 0, 5, 5, 0, 1>, - Conv::template process_tile<1, 0, 5, 5, 0, 2>, - Conv::template process_tile<1, 0, 5, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 5, 1, 0>, - Conv::template process_tile<1, 0, 5, 5, 1, 1>, - Conv::template process_tile<1, 0, 5, 5, 1, 2>, - Conv::template process_tile<1, 0, 5, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 5, 2, 0>, - Conv::template process_tile<1, 0, 5, 5, 2, 1>, - Conv::template process_tile<1, 0, 5, 5, 2, 2>, - Conv::template process_tile<1, 0, 5, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 5, 5, 3, 0>, - Conv::template process_tile<1, 0, 5, 5, 3, 1>, - Conv::template process_tile<1, 0, 5, 5, 3, 2>, - Conv::template process_tile<1, 0, 5, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 5, 6, 0, 0>, - Conv::template process_tile<1, 0, 5, 6, 0, 1>, - Conv::template process_tile<1, 0, 5, 6, 0, 2>, - Conv::template process_tile<1, 0, 5, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 5, 6, 1, 0>, - Conv::template process_tile<1, 0, 5, 6, 1, 1>, - Conv::template process_tile<1, 0, 5, 6, 1, 2>, - Conv::template process_tile<1, 0, 5, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 5, 6, 2, 0>, - Conv::template process_tile<1, 0, 5, 6, 2, 1>, - Conv::template process_tile<1, 0, 5, 6, 2, 2>, - Conv::template process_tile<1, 0, 5, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 5, 6, 3, 0>, - Conv::template process_tile<1, 0, 5, 6, 3, 1>, - Conv::template process_tile<1, 0, 5, 6, 3, 2>, - Conv::template process_tile<1, 0, 5, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 5 - { // Input pad bottom = 6 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 0, 0, 0>, - Conv::template process_tile<1, 0, 6, 0, 0, 1>, - Conv::template process_tile<1, 0, 6, 0, 0, 2>, - Conv::template process_tile<1, 0, 6, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 0, 1, 0>, - Conv::template process_tile<1, 0, 6, 0, 1, 1>, - Conv::template process_tile<1, 0, 6, 0, 1, 2>, - Conv::template process_tile<1, 0, 6, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 0, 2, 0>, - Conv::template process_tile<1, 0, 6, 0, 2, 1>, - Conv::template process_tile<1, 0, 6, 0, 2, 2>, - Conv::template process_tile<1, 0, 6, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 6, 0, 3, 0>, - Conv::template process_tile<1, 0, 6, 0, 3, 1>, - Conv::template process_tile<1, 0, 6, 0, 3, 2>, - Conv::template process_tile<1, 0, 6, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 1, 0, 0>, - Conv::template process_tile<1, 0, 6, 1, 0, 1>, - Conv::template process_tile<1, 0, 6, 1, 0, 2>, - Conv::template process_tile<1, 0, 6, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 1, 1, 0>, - Conv::template process_tile<1, 0, 6, 1, 1, 1>, - Conv::template process_tile<1, 0, 6, 1, 1, 2>, - Conv::template process_tile<1, 0, 6, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 1, 2, 0>, - Conv::template process_tile<1, 0, 6, 1, 2, 1>, - Conv::template process_tile<1, 0, 6, 1, 2, 2>, - Conv::template process_tile<1, 0, 6, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 6, 1, 3, 0>, - Conv::template process_tile<1, 0, 6, 1, 3, 1>, - Conv::template process_tile<1, 0, 6, 1, 3, 2>, - Conv::template process_tile<1, 0, 6, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 2, 0, 0>, - Conv::template process_tile<1, 0, 6, 2, 0, 1>, - Conv::template process_tile<1, 0, 6, 2, 0, 2>, - Conv::template process_tile<1, 0, 6, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 2, 1, 0>, - Conv::template process_tile<1, 0, 6, 2, 1, 1>, - Conv::template process_tile<1, 0, 6, 2, 1, 2>, - Conv::template process_tile<1, 0, 6, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 2, 2, 0>, - Conv::template process_tile<1, 0, 6, 2, 2, 1>, - Conv::template process_tile<1, 0, 6, 2, 2, 2>, - Conv::template process_tile<1, 0, 6, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 6, 2, 3, 0>, - Conv::template process_tile<1, 0, 6, 2, 3, 1>, - Conv::template process_tile<1, 0, 6, 2, 3, 2>, - Conv::template process_tile<1, 0, 6, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 3, 0, 0>, - Conv::template process_tile<1, 0, 6, 3, 0, 1>, - Conv::template process_tile<1, 0, 6, 3, 0, 2>, - Conv::template process_tile<1, 0, 6, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 3, 1, 0>, - Conv::template process_tile<1, 0, 6, 3, 1, 1>, - Conv::template process_tile<1, 0, 6, 3, 1, 2>, - Conv::template process_tile<1, 0, 6, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 3, 2, 0>, - Conv::template process_tile<1, 0, 6, 3, 2, 1>, - Conv::template process_tile<1, 0, 6, 3, 2, 2>, - Conv::template process_tile<1, 0, 6, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 6, 3, 3, 0>, - Conv::template process_tile<1, 0, 6, 3, 3, 1>, - Conv::template process_tile<1, 0, 6, 3, 3, 2>, - Conv::template process_tile<1, 0, 6, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 4, 0, 0>, - Conv::template process_tile<1, 0, 6, 4, 0, 1>, - Conv::template process_tile<1, 0, 6, 4, 0, 2>, - Conv::template process_tile<1, 0, 6, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 4, 1, 0>, - Conv::template process_tile<1, 0, 6, 4, 1, 1>, - Conv::template process_tile<1, 0, 6, 4, 1, 2>, - Conv::template process_tile<1, 0, 6, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 4, 2, 0>, - Conv::template process_tile<1, 0, 6, 4, 2, 1>, - Conv::template process_tile<1, 0, 6, 4, 2, 2>, - Conv::template process_tile<1, 0, 6, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 6, 4, 3, 0>, - Conv::template process_tile<1, 0, 6, 4, 3, 1>, - Conv::template process_tile<1, 0, 6, 4, 3, 2>, - Conv::template process_tile<1, 0, 6, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 5, 0, 0>, - Conv::template process_tile<1, 0, 6, 5, 0, 1>, - Conv::template process_tile<1, 0, 6, 5, 0, 2>, - Conv::template process_tile<1, 0, 6, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 5, 1, 0>, - Conv::template process_tile<1, 0, 6, 5, 1, 1>, - Conv::template process_tile<1, 0, 6, 5, 1, 2>, - Conv::template process_tile<1, 0, 6, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 5, 2, 0>, - Conv::template process_tile<1, 0, 6, 5, 2, 1>, - Conv::template process_tile<1, 0, 6, 5, 2, 2>, - Conv::template process_tile<1, 0, 6, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 6, 5, 3, 0>, - Conv::template process_tile<1, 0, 6, 5, 3, 1>, - Conv::template process_tile<1, 0, 6, 5, 3, 2>, - Conv::template process_tile<1, 0, 6, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 0, 6, 6, 0, 0>, - Conv::template process_tile<1, 0, 6, 6, 0, 1>, - Conv::template process_tile<1, 0, 6, 6, 0, 2>, - Conv::template process_tile<1, 0, 6, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 0, 6, 6, 1, 0>, - Conv::template process_tile<1, 0, 6, 6, 1, 1>, - Conv::template process_tile<1, 0, 6, 6, 1, 2>, - Conv::template process_tile<1, 0, 6, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 0, 6, 6, 2, 0>, - Conv::template process_tile<1, 0, 6, 6, 2, 1>, - Conv::template process_tile<1, 0, 6, 6, 2, 2>, - Conv::template process_tile<1, 0, 6, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 0, 6, 6, 3, 0>, - Conv::template process_tile<1, 0, 6, 6, 3, 1>, - Conv::template process_tile<1, 0, 6, 6, 3, 2>, - Conv::template process_tile<1, 0, 6, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 6 - }, // Input pad left = 0 - { // Input pad left = 1 - { // Input pad bottom = 0 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 0, 0, 0>, - Conv::template process_tile<1, 1, 0, 0, 0, 1>, - Conv::template process_tile<1, 1, 0, 0, 0, 2>, - Conv::template process_tile<1, 1, 0, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 0, 1, 0>, - Conv::template process_tile<1, 1, 0, 0, 1, 1>, - Conv::template process_tile<1, 1, 0, 0, 1, 2>, - Conv::template process_tile<1, 1, 0, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 0, 2, 0>, - Conv::template process_tile<1, 1, 0, 0, 2, 1>, - Conv::template process_tile<1, 1, 0, 0, 2, 2>, - Conv::template process_tile<1, 1, 0, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 0, 0, 3, 0>, - Conv::template process_tile<1, 1, 0, 0, 3, 1>, - Conv::template process_tile<1, 1, 0, 0, 3, 2>, - Conv::template process_tile<1, 1, 0, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 1, 0, 0>, - Conv::template process_tile<1, 1, 0, 1, 0, 1>, - Conv::template process_tile<1, 1, 0, 1, 0, 2>, - Conv::template process_tile<1, 1, 0, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 1, 1, 0>, - Conv::template process_tile<1, 1, 0, 1, 1, 1>, - Conv::template process_tile<1, 1, 0, 1, 1, 2>, - Conv::template process_tile<1, 1, 0, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 1, 2, 0>, - Conv::template process_tile<1, 1, 0, 1, 2, 1>, - Conv::template process_tile<1, 1, 0, 1, 2, 2>, - Conv::template process_tile<1, 1, 0, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 0, 1, 3, 0>, - Conv::template process_tile<1, 1, 0, 1, 3, 1>, - Conv::template process_tile<1, 1, 0, 1, 3, 2>, - Conv::template process_tile<1, 1, 0, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 2, 0, 0>, - Conv::template process_tile<1, 1, 0, 2, 0, 1>, - Conv::template process_tile<1, 1, 0, 2, 0, 2>, - Conv::template process_tile<1, 1, 0, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 2, 1, 0>, - Conv::template process_tile<1, 1, 0, 2, 1, 1>, - Conv::template process_tile<1, 1, 0, 2, 1, 2>, - Conv::template process_tile<1, 1, 0, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 2, 2, 0>, - Conv::template process_tile<1, 1, 0, 2, 2, 1>, - Conv::template process_tile<1, 1, 0, 2, 2, 2>, - Conv::template process_tile<1, 1, 0, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 0, 2, 3, 0>, - Conv::template process_tile<1, 1, 0, 2, 3, 1>, - Conv::template process_tile<1, 1, 0, 2, 3, 2>, - Conv::template process_tile<1, 1, 0, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 3, 0, 0>, - Conv::template process_tile<1, 1, 0, 3, 0, 1>, - Conv::template process_tile<1, 1, 0, 3, 0, 2>, - Conv::template process_tile<1, 1, 0, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 3, 1, 0>, - Conv::template process_tile<1, 1, 0, 3, 1, 1>, - Conv::template process_tile<1, 1, 0, 3, 1, 2>, - Conv::template process_tile<1, 1, 0, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 3, 2, 0>, - Conv::template process_tile<1, 1, 0, 3, 2, 1>, - Conv::template process_tile<1, 1, 0, 3, 2, 2>, - Conv::template process_tile<1, 1, 0, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 0, 3, 3, 0>, - Conv::template process_tile<1, 1, 0, 3, 3, 1>, - Conv::template process_tile<1, 1, 0, 3, 3, 2>, - Conv::template process_tile<1, 1, 0, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 4, 0, 0>, - Conv::template process_tile<1, 1, 0, 4, 0, 1>, - Conv::template process_tile<1, 1, 0, 4, 0, 2>, - Conv::template process_tile<1, 1, 0, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 4, 1, 0>, - Conv::template process_tile<1, 1, 0, 4, 1, 1>, - Conv::template process_tile<1, 1, 0, 4, 1, 2>, - Conv::template process_tile<1, 1, 0, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 4, 2, 0>, - Conv::template process_tile<1, 1, 0, 4, 2, 1>, - Conv::template process_tile<1, 1, 0, 4, 2, 2>, - Conv::template process_tile<1, 1, 0, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 0, 4, 3, 0>, - Conv::template process_tile<1, 1, 0, 4, 3, 1>, - Conv::template process_tile<1, 1, 0, 4, 3, 2>, - Conv::template process_tile<1, 1, 0, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 5, 0, 0>, - Conv::template process_tile<1, 1, 0, 5, 0, 1>, - Conv::template process_tile<1, 1, 0, 5, 0, 2>, - Conv::template process_tile<1, 1, 0, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 5, 1, 0>, - Conv::template process_tile<1, 1, 0, 5, 1, 1>, - Conv::template process_tile<1, 1, 0, 5, 1, 2>, - Conv::template process_tile<1, 1, 0, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 5, 2, 0>, - Conv::template process_tile<1, 1, 0, 5, 2, 1>, - Conv::template process_tile<1, 1, 0, 5, 2, 2>, - Conv::template process_tile<1, 1, 0, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 0, 5, 3, 0>, - Conv::template process_tile<1, 1, 0, 5, 3, 1>, - Conv::template process_tile<1, 1, 0, 5, 3, 2>, - Conv::template process_tile<1, 1, 0, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 0, 6, 0, 0>, - Conv::template process_tile<1, 1, 0, 6, 0, 1>, - Conv::template process_tile<1, 1, 0, 6, 0, 2>, - Conv::template process_tile<1, 1, 0, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 0, 6, 1, 0>, - Conv::template process_tile<1, 1, 0, 6, 1, 1>, - Conv::template process_tile<1, 1, 0, 6, 1, 2>, - Conv::template process_tile<1, 1, 0, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 0, 6, 2, 0>, - Conv::template process_tile<1, 1, 0, 6, 2, 1>, - Conv::template process_tile<1, 1, 0, 6, 2, 2>, - Conv::template process_tile<1, 1, 0, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 0, 6, 3, 0>, - Conv::template process_tile<1, 1, 0, 6, 3, 1>, - Conv::template process_tile<1, 1, 0, 6, 3, 2>, - Conv::template process_tile<1, 1, 0, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 0 - { // Input pad bottom = 1 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 0, 0, 0>, - Conv::template process_tile<1, 1, 1, 0, 0, 1>, - Conv::template process_tile<1, 1, 1, 0, 0, 2>, - Conv::template process_tile<1, 1, 1, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 0, 1, 0>, - Conv::template process_tile<1, 1, 1, 0, 1, 1>, - Conv::template process_tile<1, 1, 1, 0, 1, 2>, - Conv::template process_tile<1, 1, 1, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 0, 2, 0>, - Conv::template process_tile<1, 1, 1, 0, 2, 1>, - Conv::template process_tile<1, 1, 1, 0, 2, 2>, - Conv::template process_tile<1, 1, 1, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 1, 0, 3, 0>, - Conv::template process_tile<1, 1, 1, 0, 3, 1>, - Conv::template process_tile<1, 1, 1, 0, 3, 2>, - Conv::template process_tile<1, 1, 1, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 1, 0, 0>, - Conv::template process_tile<1, 1, 1, 1, 0, 1>, - Conv::template process_tile<1, 1, 1, 1, 0, 2>, - Conv::template process_tile<1, 1, 1, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 1, 1, 0>, - Conv::template process_tile<1, 1, 1, 1, 1, 1>, - Conv::template process_tile<1, 1, 1, 1, 1, 2>, - Conv::template process_tile<1, 1, 1, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 1, 2, 0>, - Conv::template process_tile<1, 1, 1, 1, 2, 1>, - Conv::template process_tile<1, 1, 1, 1, 2, 2>, - Conv::template process_tile<1, 1, 1, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 1, 1, 3, 0>, - Conv::template process_tile<1, 1, 1, 1, 3, 1>, - Conv::template process_tile<1, 1, 1, 1, 3, 2>, - Conv::template process_tile<1, 1, 1, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 2, 0, 0>, - Conv::template process_tile<1, 1, 1, 2, 0, 1>, - Conv::template process_tile<1, 1, 1, 2, 0, 2>, - Conv::template process_tile<1, 1, 1, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 2, 1, 0>, - Conv::template process_tile<1, 1, 1, 2, 1, 1>, - Conv::template process_tile<1, 1, 1, 2, 1, 2>, - Conv::template process_tile<1, 1, 1, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 2, 2, 0>, - Conv::template process_tile<1, 1, 1, 2, 2, 1>, - Conv::template process_tile<1, 1, 1, 2, 2, 2>, - Conv::template process_tile<1, 1, 1, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 1, 2, 3, 0>, - Conv::template process_tile<1, 1, 1, 2, 3, 1>, - Conv::template process_tile<1, 1, 1, 2, 3, 2>, - Conv::template process_tile<1, 1, 1, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 3, 0, 0>, - Conv::template process_tile<1, 1, 1, 3, 0, 1>, - Conv::template process_tile<1, 1, 1, 3, 0, 2>, - Conv::template process_tile<1, 1, 1, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 3, 1, 0>, - Conv::template process_tile<1, 1, 1, 3, 1, 1>, - Conv::template process_tile<1, 1, 1, 3, 1, 2>, - Conv::template process_tile<1, 1, 1, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 3, 2, 0>, - Conv::template process_tile<1, 1, 1, 3, 2, 1>, - Conv::template process_tile<1, 1, 1, 3, 2, 2>, - Conv::template process_tile<1, 1, 1, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 1, 3, 3, 0>, - Conv::template process_tile<1, 1, 1, 3, 3, 1>, - Conv::template process_tile<1, 1, 1, 3, 3, 2>, - Conv::template process_tile<1, 1, 1, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 4, 0, 0>, - Conv::template process_tile<1, 1, 1, 4, 0, 1>, - Conv::template process_tile<1, 1, 1, 4, 0, 2>, - Conv::template process_tile<1, 1, 1, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 4, 1, 0>, - Conv::template process_tile<1, 1, 1, 4, 1, 1>, - Conv::template process_tile<1, 1, 1, 4, 1, 2>, - Conv::template process_tile<1, 1, 1, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 4, 2, 0>, - Conv::template process_tile<1, 1, 1, 4, 2, 1>, - Conv::template process_tile<1, 1, 1, 4, 2, 2>, - Conv::template process_tile<1, 1, 1, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 1, 4, 3, 0>, - Conv::template process_tile<1, 1, 1, 4, 3, 1>, - Conv::template process_tile<1, 1, 1, 4, 3, 2>, - Conv::template process_tile<1, 1, 1, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 5, 0, 0>, - Conv::template process_tile<1, 1, 1, 5, 0, 1>, - Conv::template process_tile<1, 1, 1, 5, 0, 2>, - Conv::template process_tile<1, 1, 1, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 5, 1, 0>, - Conv::template process_tile<1, 1, 1, 5, 1, 1>, - Conv::template process_tile<1, 1, 1, 5, 1, 2>, - Conv::template process_tile<1, 1, 1, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 5, 2, 0>, - Conv::template process_tile<1, 1, 1, 5, 2, 1>, - Conv::template process_tile<1, 1, 1, 5, 2, 2>, - Conv::template process_tile<1, 1, 1, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 1, 5, 3, 0>, - Conv::template process_tile<1, 1, 1, 5, 3, 1>, - Conv::template process_tile<1, 1, 1, 5, 3, 2>, - Conv::template process_tile<1, 1, 1, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 1, 6, 0, 0>, - Conv::template process_tile<1, 1, 1, 6, 0, 1>, - Conv::template process_tile<1, 1, 1, 6, 0, 2>, - Conv::template process_tile<1, 1, 1, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 1, 6, 1, 0>, - Conv::template process_tile<1, 1, 1, 6, 1, 1>, - Conv::template process_tile<1, 1, 1, 6, 1, 2>, - Conv::template process_tile<1, 1, 1, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 1, 6, 2, 0>, - Conv::template process_tile<1, 1, 1, 6, 2, 1>, - Conv::template process_tile<1, 1, 1, 6, 2, 2>, - Conv::template process_tile<1, 1, 1, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 1, 6, 3, 0>, - Conv::template process_tile<1, 1, 1, 6, 3, 1>, - Conv::template process_tile<1, 1, 1, 6, 3, 2>, - Conv::template process_tile<1, 1, 1, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 1 - { // Input pad bottom = 2 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 0, 0, 0>, - Conv::template process_tile<1, 1, 2, 0, 0, 1>, - Conv::template process_tile<1, 1, 2, 0, 0, 2>, - Conv::template process_tile<1, 1, 2, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 0, 1, 0>, - Conv::template process_tile<1, 1, 2, 0, 1, 1>, - Conv::template process_tile<1, 1, 2, 0, 1, 2>, - Conv::template process_tile<1, 1, 2, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 0, 2, 0>, - Conv::template process_tile<1, 1, 2, 0, 2, 1>, - Conv::template process_tile<1, 1, 2, 0, 2, 2>, - Conv::template process_tile<1, 1, 2, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 2, 0, 3, 0>, - Conv::template process_tile<1, 1, 2, 0, 3, 1>, - Conv::template process_tile<1, 1, 2, 0, 3, 2>, - Conv::template process_tile<1, 1, 2, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 1, 0, 0>, - Conv::template process_tile<1, 1, 2, 1, 0, 1>, - Conv::template process_tile<1, 1, 2, 1, 0, 2>, - Conv::template process_tile<1, 1, 2, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 1, 1, 0>, - Conv::template process_tile<1, 1, 2, 1, 1, 1>, - Conv::template process_tile<1, 1, 2, 1, 1, 2>, - Conv::template process_tile<1, 1, 2, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 1, 2, 0>, - Conv::template process_tile<1, 1, 2, 1, 2, 1>, - Conv::template process_tile<1, 1, 2, 1, 2, 2>, - Conv::template process_tile<1, 1, 2, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 2, 1, 3, 0>, - Conv::template process_tile<1, 1, 2, 1, 3, 1>, - Conv::template process_tile<1, 1, 2, 1, 3, 2>, - Conv::template process_tile<1, 1, 2, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 2, 0, 0>, - Conv::template process_tile<1, 1, 2, 2, 0, 1>, - Conv::template process_tile<1, 1, 2, 2, 0, 2>, - Conv::template process_tile<1, 1, 2, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 2, 1, 0>, - Conv::template process_tile<1, 1, 2, 2, 1, 1>, - Conv::template process_tile<1, 1, 2, 2, 1, 2>, - Conv::template process_tile<1, 1, 2, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 2, 2, 0>, - Conv::template process_tile<1, 1, 2, 2, 2, 1>, - Conv::template process_tile<1, 1, 2, 2, 2, 2>, - Conv::template process_tile<1, 1, 2, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 2, 2, 3, 0>, - Conv::template process_tile<1, 1, 2, 2, 3, 1>, - Conv::template process_tile<1, 1, 2, 2, 3, 2>, - Conv::template process_tile<1, 1, 2, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 3, 0, 0>, - Conv::template process_tile<1, 1, 2, 3, 0, 1>, - Conv::template process_tile<1, 1, 2, 3, 0, 2>, - Conv::template process_tile<1, 1, 2, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 3, 1, 0>, - Conv::template process_tile<1, 1, 2, 3, 1, 1>, - Conv::template process_tile<1, 1, 2, 3, 1, 2>, - Conv::template process_tile<1, 1, 2, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 3, 2, 0>, - Conv::template process_tile<1, 1, 2, 3, 2, 1>, - Conv::template process_tile<1, 1, 2, 3, 2, 2>, - Conv::template process_tile<1, 1, 2, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 2, 3, 3, 0>, - Conv::template process_tile<1, 1, 2, 3, 3, 1>, - Conv::template process_tile<1, 1, 2, 3, 3, 2>, - Conv::template process_tile<1, 1, 2, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 4, 0, 0>, - Conv::template process_tile<1, 1, 2, 4, 0, 1>, - Conv::template process_tile<1, 1, 2, 4, 0, 2>, - Conv::template process_tile<1, 1, 2, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 4, 1, 0>, - Conv::template process_tile<1, 1, 2, 4, 1, 1>, - Conv::template process_tile<1, 1, 2, 4, 1, 2>, - Conv::template process_tile<1, 1, 2, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 4, 2, 0>, - Conv::template process_tile<1, 1, 2, 4, 2, 1>, - Conv::template process_tile<1, 1, 2, 4, 2, 2>, - Conv::template process_tile<1, 1, 2, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 2, 4, 3, 0>, - Conv::template process_tile<1, 1, 2, 4, 3, 1>, - Conv::template process_tile<1, 1, 2, 4, 3, 2>, - Conv::template process_tile<1, 1, 2, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 5, 0, 0>, - Conv::template process_tile<1, 1, 2, 5, 0, 1>, - Conv::template process_tile<1, 1, 2, 5, 0, 2>, - Conv::template process_tile<1, 1, 2, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 5, 1, 0>, - Conv::template process_tile<1, 1, 2, 5, 1, 1>, - Conv::template process_tile<1, 1, 2, 5, 1, 2>, - Conv::template process_tile<1, 1, 2, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 5, 2, 0>, - Conv::template process_tile<1, 1, 2, 5, 2, 1>, - Conv::template process_tile<1, 1, 2, 5, 2, 2>, - Conv::template process_tile<1, 1, 2, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 2, 5, 3, 0>, - Conv::template process_tile<1, 1, 2, 5, 3, 1>, - Conv::template process_tile<1, 1, 2, 5, 3, 2>, - Conv::template process_tile<1, 1, 2, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 2, 6, 0, 0>, - Conv::template process_tile<1, 1, 2, 6, 0, 1>, - Conv::template process_tile<1, 1, 2, 6, 0, 2>, - Conv::template process_tile<1, 1, 2, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 2, 6, 1, 0>, - Conv::template process_tile<1, 1, 2, 6, 1, 1>, - Conv::template process_tile<1, 1, 2, 6, 1, 2>, - Conv::template process_tile<1, 1, 2, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 2, 6, 2, 0>, - Conv::template process_tile<1, 1, 2, 6, 2, 1>, - Conv::template process_tile<1, 1, 2, 6, 2, 2>, - Conv::template process_tile<1, 1, 2, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 2, 6, 3, 0>, - Conv::template process_tile<1, 1, 2, 6, 3, 1>, - Conv::template process_tile<1, 1, 2, 6, 3, 2>, - Conv::template process_tile<1, 1, 2, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 2 - { // Input pad bottom = 3 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 0, 0, 0>, - Conv::template process_tile<1, 1, 3, 0, 0, 1>, - Conv::template process_tile<1, 1, 3, 0, 0, 2>, - Conv::template process_tile<1, 1, 3, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 0, 1, 0>, - Conv::template process_tile<1, 1, 3, 0, 1, 1>, - Conv::template process_tile<1, 1, 3, 0, 1, 2>, - Conv::template process_tile<1, 1, 3, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 0, 2, 0>, - Conv::template process_tile<1, 1, 3, 0, 2, 1>, - Conv::template process_tile<1, 1, 3, 0, 2, 2>, - Conv::template process_tile<1, 1, 3, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 3, 0, 3, 0>, - Conv::template process_tile<1, 1, 3, 0, 3, 1>, - Conv::template process_tile<1, 1, 3, 0, 3, 2>, - Conv::template process_tile<1, 1, 3, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 1, 0, 0>, - Conv::template process_tile<1, 1, 3, 1, 0, 1>, - Conv::template process_tile<1, 1, 3, 1, 0, 2>, - Conv::template process_tile<1, 1, 3, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 1, 1, 0>, - Conv::template process_tile<1, 1, 3, 1, 1, 1>, - Conv::template process_tile<1, 1, 3, 1, 1, 2>, - Conv::template process_tile<1, 1, 3, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 1, 2, 0>, - Conv::template process_tile<1, 1, 3, 1, 2, 1>, - Conv::template process_tile<1, 1, 3, 1, 2, 2>, - Conv::template process_tile<1, 1, 3, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 3, 1, 3, 0>, - Conv::template process_tile<1, 1, 3, 1, 3, 1>, - Conv::template process_tile<1, 1, 3, 1, 3, 2>, - Conv::template process_tile<1, 1, 3, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 2, 0, 0>, - Conv::template process_tile<1, 1, 3, 2, 0, 1>, - Conv::template process_tile<1, 1, 3, 2, 0, 2>, - Conv::template process_tile<1, 1, 3, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 2, 1, 0>, - Conv::template process_tile<1, 1, 3, 2, 1, 1>, - Conv::template process_tile<1, 1, 3, 2, 1, 2>, - Conv::template process_tile<1, 1, 3, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 2, 2, 0>, - Conv::template process_tile<1, 1, 3, 2, 2, 1>, - Conv::template process_tile<1, 1, 3, 2, 2, 2>, - Conv::template process_tile<1, 1, 3, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 3, 2, 3, 0>, - Conv::template process_tile<1, 1, 3, 2, 3, 1>, - Conv::template process_tile<1, 1, 3, 2, 3, 2>, - Conv::template process_tile<1, 1, 3, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 3, 0, 0>, - Conv::template process_tile<1, 1, 3, 3, 0, 1>, - Conv::template process_tile<1, 1, 3, 3, 0, 2>, - Conv::template process_tile<1, 1, 3, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 3, 1, 0>, - Conv::template process_tile<1, 1, 3, 3, 1, 1>, - Conv::template process_tile<1, 1, 3, 3, 1, 2>, - Conv::template process_tile<1, 1, 3, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 3, 2, 0>, - Conv::template process_tile<1, 1, 3, 3, 2, 1>, - Conv::template process_tile<1, 1, 3, 3, 2, 2>, - Conv::template process_tile<1, 1, 3, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 3, 3, 3, 0>, - Conv::template process_tile<1, 1, 3, 3, 3, 1>, - Conv::template process_tile<1, 1, 3, 3, 3, 2>, - Conv::template process_tile<1, 1, 3, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 4, 0, 0>, - Conv::template process_tile<1, 1, 3, 4, 0, 1>, - Conv::template process_tile<1, 1, 3, 4, 0, 2>, - Conv::template process_tile<1, 1, 3, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 4, 1, 0>, - Conv::template process_tile<1, 1, 3, 4, 1, 1>, - Conv::template process_tile<1, 1, 3, 4, 1, 2>, - Conv::template process_tile<1, 1, 3, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 4, 2, 0>, - Conv::template process_tile<1, 1, 3, 4, 2, 1>, - Conv::template process_tile<1, 1, 3, 4, 2, 2>, - Conv::template process_tile<1, 1, 3, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 3, 4, 3, 0>, - Conv::template process_tile<1, 1, 3, 4, 3, 1>, - Conv::template process_tile<1, 1, 3, 4, 3, 2>, - Conv::template process_tile<1, 1, 3, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 5, 0, 0>, - Conv::template process_tile<1, 1, 3, 5, 0, 1>, - Conv::template process_tile<1, 1, 3, 5, 0, 2>, - Conv::template process_tile<1, 1, 3, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 5, 1, 0>, - Conv::template process_tile<1, 1, 3, 5, 1, 1>, - Conv::template process_tile<1, 1, 3, 5, 1, 2>, - Conv::template process_tile<1, 1, 3, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 5, 2, 0>, - Conv::template process_tile<1, 1, 3, 5, 2, 1>, - Conv::template process_tile<1, 1, 3, 5, 2, 2>, - Conv::template process_tile<1, 1, 3, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 3, 5, 3, 0>, - Conv::template process_tile<1, 1, 3, 5, 3, 1>, - Conv::template process_tile<1, 1, 3, 5, 3, 2>, - Conv::template process_tile<1, 1, 3, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 3, 6, 0, 0>, - Conv::template process_tile<1, 1, 3, 6, 0, 1>, - Conv::template process_tile<1, 1, 3, 6, 0, 2>, - Conv::template process_tile<1, 1, 3, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 3, 6, 1, 0>, - Conv::template process_tile<1, 1, 3, 6, 1, 1>, - Conv::template process_tile<1, 1, 3, 6, 1, 2>, - Conv::template process_tile<1, 1, 3, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 3, 6, 2, 0>, - Conv::template process_tile<1, 1, 3, 6, 2, 1>, - Conv::template process_tile<1, 1, 3, 6, 2, 2>, - Conv::template process_tile<1, 1, 3, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 3, 6, 3, 0>, - Conv::template process_tile<1, 1, 3, 6, 3, 1>, - Conv::template process_tile<1, 1, 3, 6, 3, 2>, - Conv::template process_tile<1, 1, 3, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 3 - { // Input pad bottom = 4 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 0, 0, 0>, - Conv::template process_tile<1, 1, 4, 0, 0, 1>, - Conv::template process_tile<1, 1, 4, 0, 0, 2>, - Conv::template process_tile<1, 1, 4, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 0, 1, 0>, - Conv::template process_tile<1, 1, 4, 0, 1, 1>, - Conv::template process_tile<1, 1, 4, 0, 1, 2>, - Conv::template process_tile<1, 1, 4, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 0, 2, 0>, - Conv::template process_tile<1, 1, 4, 0, 2, 1>, - Conv::template process_tile<1, 1, 4, 0, 2, 2>, - Conv::template process_tile<1, 1, 4, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 4, 0, 3, 0>, - Conv::template process_tile<1, 1, 4, 0, 3, 1>, - Conv::template process_tile<1, 1, 4, 0, 3, 2>, - Conv::template process_tile<1, 1, 4, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 1, 0, 0>, - Conv::template process_tile<1, 1, 4, 1, 0, 1>, - Conv::template process_tile<1, 1, 4, 1, 0, 2>, - Conv::template process_tile<1, 1, 4, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 1, 1, 0>, - Conv::template process_tile<1, 1, 4, 1, 1, 1>, - Conv::template process_tile<1, 1, 4, 1, 1, 2>, - Conv::template process_tile<1, 1, 4, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 1, 2, 0>, - Conv::template process_tile<1, 1, 4, 1, 2, 1>, - Conv::template process_tile<1, 1, 4, 1, 2, 2>, - Conv::template process_tile<1, 1, 4, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 4, 1, 3, 0>, - Conv::template process_tile<1, 1, 4, 1, 3, 1>, - Conv::template process_tile<1, 1, 4, 1, 3, 2>, - Conv::template process_tile<1, 1, 4, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 2, 0, 0>, - Conv::template process_tile<1, 1, 4, 2, 0, 1>, - Conv::template process_tile<1, 1, 4, 2, 0, 2>, - Conv::template process_tile<1, 1, 4, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 2, 1, 0>, - Conv::template process_tile<1, 1, 4, 2, 1, 1>, - Conv::template process_tile<1, 1, 4, 2, 1, 2>, - Conv::template process_tile<1, 1, 4, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 2, 2, 0>, - Conv::template process_tile<1, 1, 4, 2, 2, 1>, - Conv::template process_tile<1, 1, 4, 2, 2, 2>, - Conv::template process_tile<1, 1, 4, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 4, 2, 3, 0>, - Conv::template process_tile<1, 1, 4, 2, 3, 1>, - Conv::template process_tile<1, 1, 4, 2, 3, 2>, - Conv::template process_tile<1, 1, 4, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 3, 0, 0>, - Conv::template process_tile<1, 1, 4, 3, 0, 1>, - Conv::template process_tile<1, 1, 4, 3, 0, 2>, - Conv::template process_tile<1, 1, 4, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 3, 1, 0>, - Conv::template process_tile<1, 1, 4, 3, 1, 1>, - Conv::template process_tile<1, 1, 4, 3, 1, 2>, - Conv::template process_tile<1, 1, 4, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 3, 2, 0>, - Conv::template process_tile<1, 1, 4, 3, 2, 1>, - Conv::template process_tile<1, 1, 4, 3, 2, 2>, - Conv::template process_tile<1, 1, 4, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 4, 3, 3, 0>, - Conv::template process_tile<1, 1, 4, 3, 3, 1>, - Conv::template process_tile<1, 1, 4, 3, 3, 2>, - Conv::template process_tile<1, 1, 4, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 4, 0, 0>, - Conv::template process_tile<1, 1, 4, 4, 0, 1>, - Conv::template process_tile<1, 1, 4, 4, 0, 2>, - Conv::template process_tile<1, 1, 4, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 4, 1, 0>, - Conv::template process_tile<1, 1, 4, 4, 1, 1>, - Conv::template process_tile<1, 1, 4, 4, 1, 2>, - Conv::template process_tile<1, 1, 4, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 4, 2, 0>, - Conv::template process_tile<1, 1, 4, 4, 2, 1>, - Conv::template process_tile<1, 1, 4, 4, 2, 2>, - Conv::template process_tile<1, 1, 4, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 4, 4, 3, 0>, - Conv::template process_tile<1, 1, 4, 4, 3, 1>, - Conv::template process_tile<1, 1, 4, 4, 3, 2>, - Conv::template process_tile<1, 1, 4, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 5, 0, 0>, - Conv::template process_tile<1, 1, 4, 5, 0, 1>, - Conv::template process_tile<1, 1, 4, 5, 0, 2>, - Conv::template process_tile<1, 1, 4, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 5, 1, 0>, - Conv::template process_tile<1, 1, 4, 5, 1, 1>, - Conv::template process_tile<1, 1, 4, 5, 1, 2>, - Conv::template process_tile<1, 1, 4, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 5, 2, 0>, - Conv::template process_tile<1, 1, 4, 5, 2, 1>, - Conv::template process_tile<1, 1, 4, 5, 2, 2>, - Conv::template process_tile<1, 1, 4, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 4, 5, 3, 0>, - Conv::template process_tile<1, 1, 4, 5, 3, 1>, - Conv::template process_tile<1, 1, 4, 5, 3, 2>, - Conv::template process_tile<1, 1, 4, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 4, 6, 0, 0>, - Conv::template process_tile<1, 1, 4, 6, 0, 1>, - Conv::template process_tile<1, 1, 4, 6, 0, 2>, - Conv::template process_tile<1, 1, 4, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 4, 6, 1, 0>, - Conv::template process_tile<1, 1, 4, 6, 1, 1>, - Conv::template process_tile<1, 1, 4, 6, 1, 2>, - Conv::template process_tile<1, 1, 4, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 4, 6, 2, 0>, - Conv::template process_tile<1, 1, 4, 6, 2, 1>, - Conv::template process_tile<1, 1, 4, 6, 2, 2>, - Conv::template process_tile<1, 1, 4, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 4, 6, 3, 0>, - Conv::template process_tile<1, 1, 4, 6, 3, 1>, - Conv::template process_tile<1, 1, 4, 6, 3, 2>, - Conv::template process_tile<1, 1, 4, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 4 - { // Input pad bottom = 5 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 0, 0, 0>, - Conv::template process_tile<1, 1, 5, 0, 0, 1>, - Conv::template process_tile<1, 1, 5, 0, 0, 2>, - Conv::template process_tile<1, 1, 5, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 0, 1, 0>, - Conv::template process_tile<1, 1, 5, 0, 1, 1>, - Conv::template process_tile<1, 1, 5, 0, 1, 2>, - Conv::template process_tile<1, 1, 5, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 0, 2, 0>, - Conv::template process_tile<1, 1, 5, 0, 2, 1>, - Conv::template process_tile<1, 1, 5, 0, 2, 2>, - Conv::template process_tile<1, 1, 5, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 5, 0, 3, 0>, - Conv::template process_tile<1, 1, 5, 0, 3, 1>, - Conv::template process_tile<1, 1, 5, 0, 3, 2>, - Conv::template process_tile<1, 1, 5, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 1, 0, 0>, - Conv::template process_tile<1, 1, 5, 1, 0, 1>, - Conv::template process_tile<1, 1, 5, 1, 0, 2>, - Conv::template process_tile<1, 1, 5, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 1, 1, 0>, - Conv::template process_tile<1, 1, 5, 1, 1, 1>, - Conv::template process_tile<1, 1, 5, 1, 1, 2>, - Conv::template process_tile<1, 1, 5, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 1, 2, 0>, - Conv::template process_tile<1, 1, 5, 1, 2, 1>, - Conv::template process_tile<1, 1, 5, 1, 2, 2>, - Conv::template process_tile<1, 1, 5, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 5, 1, 3, 0>, - Conv::template process_tile<1, 1, 5, 1, 3, 1>, - Conv::template process_tile<1, 1, 5, 1, 3, 2>, - Conv::template process_tile<1, 1, 5, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 2, 0, 0>, - Conv::template process_tile<1, 1, 5, 2, 0, 1>, - Conv::template process_tile<1, 1, 5, 2, 0, 2>, - Conv::template process_tile<1, 1, 5, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 2, 1, 0>, - Conv::template process_tile<1, 1, 5, 2, 1, 1>, - Conv::template process_tile<1, 1, 5, 2, 1, 2>, - Conv::template process_tile<1, 1, 5, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 2, 2, 0>, - Conv::template process_tile<1, 1, 5, 2, 2, 1>, - Conv::template process_tile<1, 1, 5, 2, 2, 2>, - Conv::template process_tile<1, 1, 5, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 5, 2, 3, 0>, - Conv::template process_tile<1, 1, 5, 2, 3, 1>, - Conv::template process_tile<1, 1, 5, 2, 3, 2>, - Conv::template process_tile<1, 1, 5, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 3, 0, 0>, - Conv::template process_tile<1, 1, 5, 3, 0, 1>, - Conv::template process_tile<1, 1, 5, 3, 0, 2>, - Conv::template process_tile<1, 1, 5, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 3, 1, 0>, - Conv::template process_tile<1, 1, 5, 3, 1, 1>, - Conv::template process_tile<1, 1, 5, 3, 1, 2>, - Conv::template process_tile<1, 1, 5, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 3, 2, 0>, - Conv::template process_tile<1, 1, 5, 3, 2, 1>, - Conv::template process_tile<1, 1, 5, 3, 2, 2>, - Conv::template process_tile<1, 1, 5, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 5, 3, 3, 0>, - Conv::template process_tile<1, 1, 5, 3, 3, 1>, - Conv::template process_tile<1, 1, 5, 3, 3, 2>, - Conv::template process_tile<1, 1, 5, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 4, 0, 0>, - Conv::template process_tile<1, 1, 5, 4, 0, 1>, - Conv::template process_tile<1, 1, 5, 4, 0, 2>, - Conv::template process_tile<1, 1, 5, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 4, 1, 0>, - Conv::template process_tile<1, 1, 5, 4, 1, 1>, - Conv::template process_tile<1, 1, 5, 4, 1, 2>, - Conv::template process_tile<1, 1, 5, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 4, 2, 0>, - Conv::template process_tile<1, 1, 5, 4, 2, 1>, - Conv::template process_tile<1, 1, 5, 4, 2, 2>, - Conv::template process_tile<1, 1, 5, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 5, 4, 3, 0>, - Conv::template process_tile<1, 1, 5, 4, 3, 1>, - Conv::template process_tile<1, 1, 5, 4, 3, 2>, - Conv::template process_tile<1, 1, 5, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 5, 0, 0>, - Conv::template process_tile<1, 1, 5, 5, 0, 1>, - Conv::template process_tile<1, 1, 5, 5, 0, 2>, - Conv::template process_tile<1, 1, 5, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 5, 1, 0>, - Conv::template process_tile<1, 1, 5, 5, 1, 1>, - Conv::template process_tile<1, 1, 5, 5, 1, 2>, - Conv::template process_tile<1, 1, 5, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 5, 2, 0>, - Conv::template process_tile<1, 1, 5, 5, 2, 1>, - Conv::template process_tile<1, 1, 5, 5, 2, 2>, - Conv::template process_tile<1, 1, 5, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 5, 5, 3, 0>, - Conv::template process_tile<1, 1, 5, 5, 3, 1>, - Conv::template process_tile<1, 1, 5, 5, 3, 2>, - Conv::template process_tile<1, 1, 5, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 5, 6, 0, 0>, - Conv::template process_tile<1, 1, 5, 6, 0, 1>, - Conv::template process_tile<1, 1, 5, 6, 0, 2>, - Conv::template process_tile<1, 1, 5, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 5, 6, 1, 0>, - Conv::template process_tile<1, 1, 5, 6, 1, 1>, - Conv::template process_tile<1, 1, 5, 6, 1, 2>, - Conv::template process_tile<1, 1, 5, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 5, 6, 2, 0>, - Conv::template process_tile<1, 1, 5, 6, 2, 1>, - Conv::template process_tile<1, 1, 5, 6, 2, 2>, - Conv::template process_tile<1, 1, 5, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 5, 6, 3, 0>, - Conv::template process_tile<1, 1, 5, 6, 3, 1>, - Conv::template process_tile<1, 1, 5, 6, 3, 2>, - Conv::template process_tile<1, 1, 5, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 5 - { // Input pad bottom = 6 - { // Input pad right = 0 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 0, 0, 0>, - Conv::template process_tile<1, 1, 6, 0, 0, 1>, - Conv::template process_tile<1, 1, 6, 0, 0, 2>, - Conv::template process_tile<1, 1, 6, 0, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 0, 1, 0>, - Conv::template process_tile<1, 1, 6, 0, 1, 1>, - Conv::template process_tile<1, 1, 6, 0, 1, 2>, - Conv::template process_tile<1, 1, 6, 0, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 0, 2, 0>, - Conv::template process_tile<1, 1, 6, 0, 2, 1>, - Conv::template process_tile<1, 1, 6, 0, 2, 2>, - Conv::template process_tile<1, 1, 6, 0, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 6, 0, 3, 0>, - Conv::template process_tile<1, 1, 6, 0, 3, 1>, - Conv::template process_tile<1, 1, 6, 0, 3, 2>, - Conv::template process_tile<1, 1, 6, 0, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 0 - { // Input pad right = 1 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 1, 0, 0>, - Conv::template process_tile<1, 1, 6, 1, 0, 1>, - Conv::template process_tile<1, 1, 6, 1, 0, 2>, - Conv::template process_tile<1, 1, 6, 1, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 1, 1, 0>, - Conv::template process_tile<1, 1, 6, 1, 1, 1>, - Conv::template process_tile<1, 1, 6, 1, 1, 2>, - Conv::template process_tile<1, 1, 6, 1, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 1, 2, 0>, - Conv::template process_tile<1, 1, 6, 1, 2, 1>, - Conv::template process_tile<1, 1, 6, 1, 2, 2>, - Conv::template process_tile<1, 1, 6, 1, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 6, 1, 3, 0>, - Conv::template process_tile<1, 1, 6, 1, 3, 1>, - Conv::template process_tile<1, 1, 6, 1, 3, 2>, - Conv::template process_tile<1, 1, 6, 1, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 1 - { // Input pad right = 2 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 2, 0, 0>, - Conv::template process_tile<1, 1, 6, 2, 0, 1>, - Conv::template process_tile<1, 1, 6, 2, 0, 2>, - Conv::template process_tile<1, 1, 6, 2, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 2, 1, 0>, - Conv::template process_tile<1, 1, 6, 2, 1, 1>, - Conv::template process_tile<1, 1, 6, 2, 1, 2>, - Conv::template process_tile<1, 1, 6, 2, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 2, 2, 0>, - Conv::template process_tile<1, 1, 6, 2, 2, 1>, - Conv::template process_tile<1, 1, 6, 2, 2, 2>, - Conv::template process_tile<1, 1, 6, 2, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 6, 2, 3, 0>, - Conv::template process_tile<1, 1, 6, 2, 3, 1>, - Conv::template process_tile<1, 1, 6, 2, 3, 2>, - Conv::template process_tile<1, 1, 6, 2, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 2 - { // Input pad right = 3 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 3, 0, 0>, - Conv::template process_tile<1, 1, 6, 3, 0, 1>, - Conv::template process_tile<1, 1, 6, 3, 0, 2>, - Conv::template process_tile<1, 1, 6, 3, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 3, 1, 0>, - Conv::template process_tile<1, 1, 6, 3, 1, 1>, - Conv::template process_tile<1, 1, 6, 3, 1, 2>, - Conv::template process_tile<1, 1, 6, 3, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 3, 2, 0>, - Conv::template process_tile<1, 1, 6, 3, 2, 1>, - Conv::template process_tile<1, 1, 6, 3, 2, 2>, - Conv::template process_tile<1, 1, 6, 3, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 6, 3, 3, 0>, - Conv::template process_tile<1, 1, 6, 3, 3, 1>, - Conv::template process_tile<1, 1, 6, 3, 3, 2>, - Conv::template process_tile<1, 1, 6, 3, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 3 - { // Input pad right = 4 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 4, 0, 0>, - Conv::template process_tile<1, 1, 6, 4, 0, 1>, - Conv::template process_tile<1, 1, 6, 4, 0, 2>, - Conv::template process_tile<1, 1, 6, 4, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 4, 1, 0>, - Conv::template process_tile<1, 1, 6, 4, 1, 1>, - Conv::template process_tile<1, 1, 6, 4, 1, 2>, - Conv::template process_tile<1, 1, 6, 4, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 4, 2, 0>, - Conv::template process_tile<1, 1, 6, 4, 2, 1>, - Conv::template process_tile<1, 1, 6, 4, 2, 2>, - Conv::template process_tile<1, 1, 6, 4, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 6, 4, 3, 0>, - Conv::template process_tile<1, 1, 6, 4, 3, 1>, - Conv::template process_tile<1, 1, 6, 4, 3, 2>, - Conv::template process_tile<1, 1, 6, 4, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 4 - { // Input pad right = 5 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 5, 0, 0>, - Conv::template process_tile<1, 1, 6, 5, 0, 1>, - Conv::template process_tile<1, 1, 6, 5, 0, 2>, - Conv::template process_tile<1, 1, 6, 5, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 5, 1, 0>, - Conv::template process_tile<1, 1, 6, 5, 1, 1>, - Conv::template process_tile<1, 1, 6, 5, 1, 2>, - Conv::template process_tile<1, 1, 6, 5, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 5, 2, 0>, - Conv::template process_tile<1, 1, 6, 5, 2, 1>, - Conv::template process_tile<1, 1, 6, 5, 2, 2>, - Conv::template process_tile<1, 1, 6, 5, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 6, 5, 3, 0>, - Conv::template process_tile<1, 1, 6, 5, 3, 1>, - Conv::template process_tile<1, 1, 6, 5, 3, 2>, - Conv::template process_tile<1, 1, 6, 5, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 5 - { // Input pad right = 6 - { // Output pad bottom = 0 - Conv::template process_tile<1, 1, 6, 6, 0, 0>, - Conv::template process_tile<1, 1, 6, 6, 0, 1>, - Conv::template process_tile<1, 1, 6, 6, 0, 2>, - Conv::template process_tile<1, 1, 6, 6, 0, 3>, - }, // Output pad bottom = 0 - { // Output pad bottom = 1 - Conv::template process_tile<1, 1, 6, 6, 1, 0>, - Conv::template process_tile<1, 1, 6, 6, 1, 1>, - Conv::template process_tile<1, 1, 6, 6, 1, 2>, - Conv::template process_tile<1, 1, 6, 6, 1, 3>, - }, // Output pad bottom = 1 - { // Output pad bottom = 2 - Conv::template process_tile<1, 1, 6, 6, 2, 0>, - Conv::template process_tile<1, 1, 6, 6, 2, 1>, - Conv::template process_tile<1, 1, 6, 6, 2, 2>, - Conv::template process_tile<1, 1, 6, 6, 2, 3>, - }, // Output pad bottom = 2 - { // Output pad bottom = 3 - Conv::template process_tile<1, 1, 6, 6, 3, 0>, - Conv::template process_tile<1, 1, 6, 6, 3, 1>, - Conv::template process_tile<1, 1, 6, 6, 3, 2>, - Conv::template process_tile<1, 1, 6, 6, 3, 3>, - }, // Output pad bottom = 3 - }, // Input pad right = 6 - }, // Input pad bottom = 6 - }, // Input pad left = 1 - }, // Input pad top = 1 +const Conv::TileFn Conv::tilefn_unpadded = ConvImpl::template process_tile; + +template <> +const Conv::TileFn Conv::tilefn_top[n_in_pad_top_fns] = { + ConvImpl::template process_tile, + ConvImpl::template process_tile, +}; + +template <> +const Conv::TileFn Conv::tilefn_left[n_in_pad_left_fns] = { + ConvImpl::template process_tile, + ConvImpl::template process_tile, +}; + +template <> +const Conv::TileFn Conv::tilefn_bottom[n_in_pad_bottom_fns][n_out_pad_bottom_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, +}; + +template <> +const Conv::TileFn Conv::tilefn_right[n_in_pad_right_fns][n_out_pad_right_fns] = { + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, + { + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + ConvImpl::template process_tile, + }, }; +template <> +const Conv::TileFn Conv::tilefn_generic = ConvImpl::template process_tile; template class DepthwiseConvolution<4, 4, 3, 3, 2, 2, float, float>; } // namespace depthwise diff --git a/tests/datasets/DepthwiseConvolutionLayerDataset.h b/tests/datasets/DepthwiseConvolutionLayerDataset.h index b8a16a7c3d..4a0637d77f 100644 --- a/tests/datasets/DepthwiseConvolutionLayerDataset.h +++ b/tests/datasets/DepthwiseConvolutionLayerDataset.h @@ -193,6 +193,7 @@ public: add_config(TensorShape(7U, 7U, 32U), TensorShape(3U, 3U, 32U), TensorShape(4U, 4U, 32U), PadStrideInfo(2, 2, 1, 1, 1, 1, DimensionRoundingType::CEIL)); add_config(TensorShape(8U, 8U, 32U), TensorShape(3U, 3U, 32U), TensorShape(3U, 3U, 32U), PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)); add_config(TensorShape(8U, 8U, 32U), TensorShape(3U, 3U, 32U), TensorShape(4U, 4U, 32U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL)); + add_config(TensorShape(64U, 64U, 128U), TensorShape(3U, 3U, 128U), TensorShape(32U, 32U, 128U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL)); } }; } // namespace datasets -- cgit v1.2.1