From bda6e4b51bc4045c97100bb9d562164ba7c6c28f Mon Sep 17 00:00:00 2001 From: Pablo Tello Date: Wed, 22 Aug 2018 11:40:33 +0100 Subject: COMPMID-1247:Integrate kernel size 1x3 & 3x1 support in NEWinogradLayer. Change-Id: I6fe198881230e49864c841a3b2366ccf2a9247f9 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/145210 Tested-by: Jenkins Reviewed-by: Georgios Pinitas --- .../kernels/NEWinogradConvolutionLayerKernel.cpp | 74 +++++-- .../winograd/transforms/input_2x2_3x3_fp32.cpp | 2 +- .../winograd/transforms/input_2x2_5x5_fp32.cpp | 2 +- .../winograd/transforms/input_4x4_3x3_fp32.cpp | 2 +- .../winograd/transforms/input_6_3_fp32.cpp | 226 +++++++++++++++++++++ .../winograd/transforms/output_6_3_fp32.cpp | 186 +++++++++++++++++ .../winograd/transforms/weights_6_3_fp32.cpp | 125 ++++++++++++ .../kernels/convolution/winograd/winograd_gemm.cpp | 3 + 8 files changed, 597 insertions(+), 23 deletions(-) create mode 100644 src/core/NEON/kernels/convolution/winograd/transforms/input_6_3_fp32.cpp create mode 100644 src/core/NEON/kernels/convolution/winograd/transforms/output_6_3_fp32.cpp create mode 100644 src/core/NEON/kernels/convolution/winograd/transforms/weights_6_3_fp32.cpp (limited to 'src/core/NEON') diff --git a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp index 3d7a16dd45..8f990712e8 100644 --- a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp +++ b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp @@ -40,19 +40,27 @@ namespace arm_compute namespace { +inline bool is_kernel_size_supported(Size2D size) +{ + const std::array supported_input_sizes = { { Size2D(1, 3), Size2D(3, 1), Size2D(5, 5), Size2D(3, 3) } }; + return std::end(supported_input_sizes) != std::find(std::begin(supported_input_sizes), std::end(supported_input_sizes), size); +} + Status validate_arguments_winograd_weight_trans(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); - const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != 3 && input->dimension(idx_width) != 5); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != input->dimension(idx_height)); + const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); + const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); + const auto input_width = input->dimension(idx_width); + const auto input_height = input->dimension(idx_height); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_kernel_size_supported(Size2D(input_width, input_height)), "Only 1x3, 3x1, 3x3 and 5x5 kernels are supported"); ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4); const Size2D &output_tile = winograd_info.output_tile_size; - ARM_COMPUTE_RETURN_ERROR_ON(output_tile != Size2D(2U, 2U) && output_tile != Size2D(4U, 4U)); + const std::array supported_tile_sizes = { { Size2D(2U, 2U), Size2D(4U, 4U), Size2D(1U, 6U), Size2D(6U, 1U) } }; + ARM_COMPUTE_RETURN_ERROR_ON(std::end(supported_tile_sizes) == std::find(std::begin(supported_tile_sizes), std::end(supported_tile_sizes), output_tile)); // Checks performed when output is configured if(output->total_size() != 0) @@ -98,8 +106,8 @@ Status validate_arguments_winograd_input_trans(const ITensorInfo *input, const I ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd input transform only supports unit strides"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != 3U && kernel_dims.width != 5U), "Winograd input transform only supports 3x3 and 5x5 kernels"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != kernel_dims.height), "Winograd input transform only supports 3x3 and 5x5 kernels"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_kernel_size_supported(Size2D(kernel_dims.width, kernel_dims.height)), + "Only 1x3, 3x1, 3x3 and 5x5 kernels are supported"); // Validate configured output if(output->total_size() != 0) @@ -151,9 +159,11 @@ Status validate_arguments_winograd_output_trans(const ITensorInfo *input, const ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != num_tiles.area()); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != 3U && kernel_dims.width != 5U), "Winograd output transform only supports 3x3 and 5x5 kernels"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != kernel_dims.height), "Winograd output transform only supports 3x3 and 5x5 kernels"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((input->dimension(2) != size_t(16U)) && (input->dimension(2) != size_t(36U))), "Only 2x2 and 4x4 output tile is supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_kernel_size_supported(Size2D(kernel_dims.width, kernel_dims.height)), + "Only 1x3, 3x1, 3x3 and 5x5 kernels are supported"); + + const std::array supported_gemm_sizes = { { 8U, 16U, 36U } }; + ARM_COMPUTE_RETURN_ERROR_ON(std::end(supported_gemm_sizes) == std::find(std::begin(supported_gemm_sizes), std::end(supported_gemm_sizes), input->dimension(2))); ARM_COMPUTE_UNUSED(kernel_dims); if(bias != nullptr) { @@ -201,7 +211,21 @@ std::pair validate_and_configure_window_winograd_output_trans(IT } } // namespace -// Weights transform +template +Status INEWinogradLayerTransformWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *weights) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); + const DataLayout data_layout = input->data_layout(); + const unsigned int width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const unsigned int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_kernel_size_supported(Size2D(weights->dimension(width_idx), weights->dimension(height_idx))), + "Only 1x3, 3x1, 3x3 and 5x5 kernels are supported"); + ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); + return Status{}; +} + +template class INEWinogradLayerTransformWeightsKernel; template unsigned int NEWinogradLayerTransformWeightsKernel::get_weight_storage_size(int num_output_channels, int num_input_channels) const @@ -278,6 +302,8 @@ Status NEWinogradLayerTransformWeightsKernel; template class NEWinogradLayerTransformWeightsKernel; template class NEWinogradLayerTransformWeightsKernel; +template class NEWinogradLayerTransformWeightsKernel; +template class NEWinogradLayerTransformWeightsKernel; // Input transform @@ -343,14 +369,15 @@ void NEWinogradLayerTransformInputKernelinfo()->element_size(); - const int input_col_stride = _input_nhwc->info()->strides_in_bytes().y() / element_size_in_bytes; - const int input_row_stride = _input_nhwc->info()->strides_in_bytes().z() / element_size_in_bytes; - const int input_batch_stride = _input_nhwc->info()->strides_in_bytes()[3] / element_size_in_bytes; - - InputTransform input_transform(reinterpret_cast(_input_nhwc->buffer() + _input_nhwc->info()->offset_first_element_in_bytes()), + const int element_size_in_bytes = _input_nhwc->info()->element_size(); + const int input_col_stride = _input_nhwc->info()->strides_in_bytes().y() / element_size_in_bytes; + const int input_row_stride = _input_nhwc->info()->strides_in_bytes().z() / element_size_in_bytes; + const int input_batch_stride = _input_nhwc->info()->strides_in_bytes()[3] / element_size_in_bytes; + const auto input_nhwc_ptr = reinterpret_cast(_input_nhwc->buffer() + _input_nhwc->info()->offset_first_element_in_bytes()); + auto output_ptr = reinterpret_cast(_output->buffer() + _output->info()->offset_first_element_in_bytes()); + InputTransform input_transform(input_nhwc_ptr, _num_batches, _num_rows, _num_cols, _num_channels, _padding, - reinterpret_cast(_output->buffer() + _output->info()->offset_first_element_in_bytes()), + output_ptr, _matrix_stride, _num_channels, input_batch_stride, input_row_stride, input_col_stride); // The code below cannot be moved to configure because biases hasn't been allocated at that point @@ -371,6 +398,8 @@ Status NEWinogradLayerTransformInputKernel; template class NEWinogradLayerTransformInputKernel; template class NEWinogradLayerTransformInputKernel; +template class NEWinogradLayerTransformInputKernel; +template class NEWinogradLayerTransformInputKernel; // Output transform @@ -438,7 +467,6 @@ void NEWinogradLayerTransformOutputKernelinfo()->set_valid_region(ValidRegion(Coordinates(), _output_nhwc->info()->tensor_shape())); INEKernel::configure(win); @@ -452,10 +480,14 @@ void NEWinogradLayerTransformOutputKernelinfo()->strides_in_bytes()[2] / sizeof(T); + const int out_col_stride = _output_nhwc->info()->strides_in_bytes()[1] / sizeof(T); + OutputTransform output_transform(reinterpret_cast(_output_workspace->buffer()), _matrix_stride, _matrix_row_stride, (_biases ? reinterpret_cast(_biases->buffer() + _biases->info()->offset_first_element_in_bytes()) : nullptr), reinterpret_cast(_output_nhwc->buffer() + _output_nhwc->info()->offset_first_element_in_bytes()), - _num_batches, _num_rows, _num_cols, _num_channels, 0, _output_nhwc->info()->strides_in_bytes()[2] / sizeof(T), _output_nhwc->info()->strides_in_bytes()[1] / sizeof(T)); + _num_batches, _num_rows, _num_cols, _num_channels, out_batch_stride, out_row_stride, out_col_stride); // The code below cannot be moved to configure because biases hasn't been allocated at that point const size_t fst = window.x().start(); @@ -478,5 +510,7 @@ Status NEWinogradLayerTransformOutputKernel; template class NEWinogradLayerTransformOutputKernel; template class NEWinogradLayerTransformOutputKernel; +template class NEWinogradLayerTransformOutputKernel; +template class NEWinogradLayerTransformOutputKernel; } // namespace arm_compute diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp index 6d8afc0def..97b2695d69 100644 --- a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp +++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp @@ -329,7 +329,7 @@ void Transform::process_tile( template <> template <> -const Transform::TileFn Transform::tile_fns[2][2][max_pad_bottom][max_pad_right] = +const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = { { { diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_5x5_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_5x5_fp32.cpp index fd30b6118e..30c9463bb8 100644 --- a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_5x5_fp32.cpp +++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_5x5_fp32.cpp @@ -298,7 +298,7 @@ void Transform::process_tile( template <> template <> -const Transform::TileFn Transform::tile_fns[2][2][max_pad_bottom][max_pad_right] = +const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = { { { diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_4x4_3x3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_4x4_3x3_fp32.cpp index 04d1573e4c..7f93187132 100644 --- a/src/core/NEON/kernels/convolution/winograd/transforms/input_4x4_3x3_fp32.cpp +++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_4x4_3x3_fp32.cpp @@ -326,7 +326,7 @@ void Transform::process_tile( */ template <> template <> -const Transform::TileFn Transform::tile_fns[2][2][max_pad_bottom][max_pad_right] = +const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = { { { diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_6_3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_6_3_fp32.cpp new file mode 100644 index 0000000000..67e46499cd --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_6_3_fp32.cpp @@ -0,0 +1,226 @@ +/* + * Copyright (c) 2017 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp" +#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" +#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp" + + +namespace winograd +{ + +using Transform = WinogradGEMM<1, 6, 1, 3>::InputTransform; + +template <> +template <> +int Transform::ops_performed(const Tensor4DShape &input_shape) +{ + (void) input_shape; + return 0; // TODO +} + +template <> +template <> +template +void Transform::process_tile( + int n_channels, + const float* const input_base, + const int input_row_stride, + const int input_col_stride, + float* const matrix_base, + const int matrix_stride +) +{ + (void) input_row_stride; // No rows over which to stride + constexpr int inner_tile_j = 8; + constexpr int cells_j = inner_tile_j - pad_right; + + float *outptr = matrix_base; + + // Get pointers into the input tile + const float *x_ptrs[inner_tile_j]; + for (int j = pad_left, xj = 0; j < cells_j; j++, xj++) + { + x_ptrs[j] = input_base + xj*input_col_stride; + } + + // Vectors used/computed in this kernel. + float x[inner_tile_j]; + float U[inner_tile_j]; + + for (int j = 0; j < inner_tile_j; j++) + { + x[j] = 0.0f; + } + + // Perform the Winograd input transformation for each channel in the input + // tensor. + int channels_remaining = n_channels; +#ifdef __arm_any__ + for (; channels_remaining >= 4; channels_remaining -= 4) + { + float32x4_t x[inner_tile_j], U[inner_tile_j]; + for (int j = 0; j < inner_tile_cols; j++) + { + x[j] = vdupq_n_f32(0.0f); + } + + // Load x + for (int j = pad_left; j < cells_j; j++) + { + x[j] = vld1q_f32(x_ptrs[j]); + x_ptrs[j] += 4; + } + + // Compute U = x . X + U[0] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[2], 49), x[4], -14), x[0], -36); + U[1] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[2], 36), x[3], 13), x[4], -13), x[1], -36), x[5], -1); + U[2] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[5], 1), x[2], 36), x[1], 36), x[4], -13), x[3], -13); + U[3] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[3], 20), x[2], 9), x[5], -2), x[4], -10), x[1], -18); + U[4] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[1], 18), x[2], 9), x[5], 2), x[4], -10), x[3], -20); + U[5] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[3], 15), x[2], 4), x[5], -3), x[4], -5), x[1], -12); + U[6] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[1], 12), x[2], 4), x[5], 3), x[4], -5), x[3], -15); + U[7] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[7], 1), x[3], 49), x[5], -14), x[1], -36); + + // Store the transformed vector + for (int j = 0; j < inner_tile_j; j++) + { + vst1q_f32(outptr + j*matrix_stride, U[j]); + } + outptr += 4; + } + + for (; channels_remaining >= 2; channels_remaining -= 2) + { + float32x2_t x[inner_tile_j], U[inner_tile_j]; + for (int j = 0; j < inner_tile_cols; j++) + { + x[j] = vdup_n_f32(0.0f); + } + + // Load x + for (int j = pad_left; j < cells_j; j++) + { + x[j] = vld1_f32(x_ptrs[j]); + x_ptrs[j] += 2; + } + + // Compute U = x . X + U[0] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[2], 49), x[4], -14), x[0], -36); + U[1] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[2], 36), x[3], 13), x[4], -13), x[1], -36), x[5], -1); + U[2] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[5], 1), x[2], 36), x[1], 36), x[4], -13), x[3], -13); + U[3] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[3], 20), x[2], 9), x[5], -2), x[4], -10), x[1], -18); + U[4] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[1], 18), x[2], 9), x[5], 2), x[4], -10), x[3], -20); + U[5] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[3], 15), x[2], 4), x[5], -3), x[4], -5), x[1], -12); + U[6] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[1], 12), x[2], 4), x[5], 3), x[4], -5), x[3], -15); + U[7] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[7], 1), x[3], 49), x[5], -14), x[1], -36); + + // Store the transformed vector + for (int j = 0; j < inner_tile_j; j++) + { + vst1_f32(outptr + j*matrix_stride, U[j]); + } + outptr += 2; + } +#endif // __arm_any__ + for (; channels_remaining; channels_remaining--) + { + // Load x + for (int j = pad_left; j < cells_j; j++) + { + x[j] = *(x_ptrs[j]++); + } + + // Compute U = x . X + U[0] = x[0]*-36 + x[4]*-14 + x[2]*49 + x[6]*1; + U[1] = x[5]*-1 + x[1]*-36 + x[4]*-13 + x[3]*13 + x[2]*36 + x[6]*1; + U[2] = x[3]*-13 + x[4]*-13 + x[1]*36 + x[2]*36 + x[5]*1 + x[6]*1; + U[3] = x[1]*-18 + x[4]*-10 + x[5]*-2 + x[2]*9 + x[3]*20 + x[6]*1; + U[4] = x[3]*-20 + x[4]*-10 + x[5]*2 + x[2]*9 + x[1]*18 + x[6]*1; + U[5] = x[1]*-12 + x[4]*-5 + x[5]*-3 + x[2]*4 + x[3]*15 + x[6]*1; + U[6] = x[3]*-15 + x[4]*-5 + x[5]*3 + x[2]*4 + x[1]*12 + x[6]*1; + U[7] = x[1]*-36 + x[5]*-14 + x[3]*49 + x[7]*1; + + // Store the transformed vector + for (int j = 0; j < inner_tile_j; j++) + { + *(outptr + j*matrix_stride) = U[j]; + } + outptr++; + } +} + +template <> +template <> +const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = +{ + { + { + { + Transform::template process_tile<0, 0, 0, 0>, + Transform::template process_tile<0, 0, 0, 1>, + Transform::template process_tile<0, 0, 0, 2>, + Transform::template process_tile<0, 0, 0, 3>, + Transform::template process_tile<0, 0, 0, 4>, + Transform::template process_tile<0, 0, 0, 5>, + Transform::template process_tile<0, 0, 0, 6>, + } + }, + { + { + Transform::template process_tile<0, 1, 0, 0>, + Transform::template process_tile<0, 1, 0, 1>, + Transform::template process_tile<0, 1, 0, 2>, + Transform::template process_tile<0, 1, 0, 3>, + Transform::template process_tile<0, 1, 0, 4>, + Transform::template process_tile<0, 1, 0, 5>, + Transform::template process_tile<0, 1, 0, 6>, + } + } + } +}; + +template +using TransformTransposed = typename WinogradGEMM::template InputTransform; + +template <> +template <> +const TransformTransposed<6, 3>::TileFn + TransformTransposed<6, 3>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = {}; + +template <> +template <> +const TransformTransposed<4, 5>::TileFn + TransformTransposed<4, 5>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = {}; + +template <> +template <> +const TransformTransposed<2, 7>::TileFn + TransformTransposed<2, 7>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = {}; + + + +template struct WinogradGEMM<1, 6, 1, 3>::InputTransform; +template struct WinogradGEMM<6, 1, 3, 1>::InputTransform; +} // namespace winograd diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/output_6_3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/output_6_3_fp32.cpp new file mode 100644 index 0000000000..16667ccdb6 --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/transforms/output_6_3_fp32.cpp @@ -0,0 +1,186 @@ +/* + * Copyright (c) 2017 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/NEON/kernels/convolution/winograd/transforms/output.hpp" +#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" +#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp" + +namespace winograd +{ + +using Transform = WinogradGEMM<1, 6, 1, 3>::OutputTransform; +using TransformTransposed = WinogradGEMM<6, 1, 3, 1>::OutputTransform; + +template <> +template <> +int Transform::ops_performed(const Tensor4DShape &shape) +{ + (void) shape; + return 0; // TODO +} + +template <> +template <> +template +void Transform::process_tile( + const int n_channels, + const float* const matrix_base, + const int matrix_stride, + const float* const biases, + float* const output, + const int output_row_stride, + const int output_col_stride +) +{ + (void) output_row_stride; + constexpr int cells_j = output_tile_cols - pad_right; + + // Construct a map to the output cells + float *outptrs[cells_j]; + for (int j = 0; j < cells_j; j++) + { + outptrs[j] = output + j*output_col_stride; + } + const float *inptr = matrix_base; + const float *bptr = biases; + + // For each channel of the output + int channels_remaining = n_channels; +#ifdef __arm_any__ + for (; channels_remaining >= 4; channels_remaining -= 4) + { + // Matrices used and computed during this transform + float32x4_t F[inner_tile_cols], f[output_tile_cols], b = vdupq_n_f32(0.0f); + + // Read a 1x8 tile in the Winograd domain + for (int j = 0; j < inner_tile_cols; j++) + { + F[j] = vld1q_f32(inptr + j*matrix_stride); + } + inptr += 4; + + f[0] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(F[6], 1), F[5], 1), F[4], 1), F[3], 1), F[2], 1), F[1], 1), F[0], 1); + f[1] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(F[2], 1), F[6], 3), F[4], 2), F[3], -2), F[5], -3), F[1], -1); + f[2] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(F[2], 1), F[1], 1), F[6], 9), F[5], 9), F[4], 4), F[3], 4); + f[3] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(F[2], 1), F[6], 27), F[4], 8), F[3], -8), F[5], -27), F[1], -1); + f[4] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(F[2], 1), F[1], 1), F[6], 81), F[5], 81), F[4], 16), F[3], 16); + f[5] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(F[7], 1), F[2], 1), F[6], 243), F[4], 32), F[3], -32), F[5], -243), F[1], -1); + + // Write out the output tile + if (bptr != 0) + { + b = vld1q_f32(bptr); + bptr += 4; + } + for (int j = 0; j < cells_j; j++) + { + vst1q_f32(outptrs[j], f[j] + b); + outptrs[j] += 4; + } + } + for (; channels_remaining >= 2; channels_remaining -= 2) + { + // Matrices used and computed during this transform + float32x2_t F[inner_tile_cols], f[output_tile_cols], b = vdup_n_f32(0.0f); + + // Read a 1x8 tile in the Winograd domain + for (int j = 0; j < inner_tile_cols; j++) + { + F[j] = vld1_f32(inptr + j*matrix_stride); + } + inptr += 2; + + f[0] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(F[6], 1), F[5], 1), F[4], 1), F[3], 1), F[2], 1), F[1], 1), F[0], 1); + f[1] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(F[2], 1), F[6], 3), F[4], 2), F[3], -2), F[5], -3), F[1], -1); + f[2] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(F[2], 1), F[1], 1), F[6], 9), F[5], 9), F[4], 4), F[3], 4); + f[3] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(F[2], 1), F[6], 27), F[4], 8), F[3], -8), F[5], -27), F[1], -1); + f[4] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(F[2], 1), F[1], 1), F[6], 81), F[5], 81), F[4], 16), F[3], 16); + f[5] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(F[7], 1), F[2], 1), F[6], 243), F[4], 32), F[3], -32), F[5], -243), F[1], -1); + + // Write out the output tile + if (bptr != 0) + { + b = vld1_f32(bptr); + bptr += 2; + } + for (int j = 0; j < cells_j; j++) + { + vst1_f32(outptrs[j], f[j] + b); + outptrs[j] += 2; + } + } +#endif // __arm_any__ + for (; channels_remaining; channels_remaining--) + { + // Matrices used and computed during this transform + float F[inner_tile_cols], f[output_tile_cols], b = 0.0f; + + // Read a 1x8 tile in the Winograd domain + for (int j = 0; j < inner_tile_cols; j++) + { + F[j] = *(inptr + j*matrix_stride); + } + inptr++; + + f[0] = F[0]*1 + F[1]*1 + F[2]*1 + F[3]*1 + F[4]*1 + F[5]*1 + F[6]*1; + f[1] = F[1]*-1 + F[5]*-3 + F[3]*-2 + F[4]*2 + F[6]*3 + F[2]*1; + f[2] = F[3]*4 + F[4]*4 + F[5]*9 + F[6]*9 + F[1]*1 + F[2]*1; + f[3] = F[1]*-1 + F[5]*-27 + F[3]*-8 + F[4]*8 + F[6]*27 + F[2]*1; + f[4] = F[3]*16 + F[4]*16 + F[5]*81 + F[6]*81 + F[1]*1 + F[2]*1; + f[5] = F[1]*-1 + F[5]*-243 + F[3]*-32 + F[4]*32 + F[6]*243 + F[2]*1 + F[7]*1; + + // Write out the output tile + if (bptr != 0) + { + b = *(bptr++); + } + for (int j = 0; j < cells_j; j++) + { + *(outptrs[j]++) = f[j] + b; + } + } +} + +template <> +template <> +const Transform::TileFn Transform::tile_fns[max_pad_bottom][max_pad_right] = +{ + { + Transform::template process_tile<0, 0>, + Transform::template process_tile<0, 1>, + Transform::template process_tile<0, 2>, + Transform::template process_tile<0, 3>, + Transform::template process_tile<0, 4>, + Transform::template process_tile<0, 5>, + }, +}; + +template <> +template <> +const TransformTransposed::TileFn TransformTransposed::tile_fns[max_pad_bottom][max_pad_right] = {}; + + +template struct WinogradGEMM<1, 6, 1, 3>::OutputTransform; +template struct WinogradGEMM<6, 1, 3, 1>::OutputTransform; +} // namespace winograd diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/weights_6_3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/weights_6_3_fp32.cpp new file mode 100644 index 0000000000..c560aa8c8f --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/transforms/weights_6_3_fp32.cpp @@ -0,0 +1,125 @@ +/* + * Copyright (c) 2017 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp" +#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" +#include "arm_compute/core/NEON/kernels/convolution/winograd/transforms/kernel.hpp" + + +namespace winograd +{ + template <> + template <> + void WinogradGEMM<1, 6, 1, 3>::WeightsTransform::execute( + const int n_output_channels, + const int n_input_channels, + const float* const input, // NOTE: Data in HWIO order + float* const output, + const int matrix_stride, + const int matrix_row_stride + ) + { + // Get pointers to each cell of the weight tensor + const auto weight_col_stride = n_input_channels * n_output_channels; + const float *inptrs[3]; + for (int j = 0; j < 3; j++) + { + inptrs[j] = input + j*weight_col_stride; + } + + // For each input channel + for (int ic = 0; ic < n_input_channels; ic++) + { + float *outptr = output + ic * matrix_row_stride; + + // For each output channel + int channels_remaining = n_output_channels; + for (; channels_remaining; channels_remaining--) + { + // Matrices used and computed in this kernel + float w[3], V[inner_tile_cols]; + + // Read weights + for (int j = 0; j < 3; j++) + { + w[j] = *(inptrs[j]++); + } + + // Compute V = w WT + V[0] = (w[0]*-1) / 36.0f; + V[1] = (w[1]*-1 + w[0]*1 + w[2]*1) / 48.0f; + V[2] = (w[0]*1 + w[1]*1 + w[2]*1) / 48.0f; + V[3] = (w[0]*-1 + w[2]*-4 + w[1]*2) / 120.0f; + V[4] = (w[0]*-1 + w[2]*-4 + w[1]*-2) / 120.0f; + V[5] = (w[1]*-3 + w[2]*9 + w[0]*1) / 720.0f; + V[6] = (w[1]*3 + w[2]*9 + w[0]*1) / 720.0f; + V[7] = (w[2]*1) / 1; + + // Store the transformed weights + for (int j = 0; j < inner_tile_cols; j++) + { + *(outptr + j*matrix_stride) = V[j]; + } + outptr++; + } + } + } + + template <> + template <> + int WinogradGEMM<1, 6, 1, 3>::WeightsTransform::ops_performed(const KernelShape &shape) + { + (void) shape; + return 0; // TODO + } + + template <> + template <> + void WinogradGEMM<6, 1, 3, 1>::WeightsTransform::execute( + const int n_output_channels, + const int n_input_channels, + const float* const input, // NOTE: Data in HWIO order + float* const output, + const int matrix_stride, + const int matrix_row_stride + ) + { + // Redirect to the 1xN implementation + WinogradGEMM<1, 6, 1, 3>::template WeightsTransform::execute( + n_output_channels, n_input_channels, input, output, matrix_stride, + matrix_row_stride + ); + } + + template <> + template <> + int WinogradGEMM<6, 1, 3, 1>::WeightsTransform::ops_performed(const KernelShape &shape) + { + (void) shape; + return 0; // TODO + } + + template struct WinogradGEMM<1, 6, 1, 3>::WeightsTransform; + template struct WinogradGEMM<6, 1, 3, 1>::WeightsTransform; +} diff --git a/src/core/NEON/kernels/convolution/winograd/winograd_gemm.cpp b/src/core/NEON/kernels/convolution/winograd/winograd_gemm.cpp index a5d43024a4..d544fd5710 100644 --- a/src/core/NEON/kernels/convolution/winograd/winograd_gemm.cpp +++ b/src/core/NEON/kernels/convolution/winograd/winograd_gemm.cpp @@ -225,4 +225,7 @@ int WinogradGEMM:: template class WinogradGEMM<2, 2, 3, 3>::Convolution; template class WinogradGEMM<4, 4, 3, 3>::Convolution; +template class WinogradGEMM<1, 6, 1, 3>::Convolution; +template class WinogradGEMM<6, 1, 3, 1>::Convolution; + template class WinogradGEMM<2, 2, 5, 5>::Convolution; -- cgit v1.2.1