diff options
Diffstat (limited to 'src')
-rw-r--r-- | src/core/NEON/NEKernels.h | 1 | ||||
-rw-r--r-- | src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp (renamed from src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp) | 282 | ||||
-rw-r--r-- | src/core/cpu/kernels/CpuWinogradConv2dKernel.h (renamed from src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h) | 180 | ||||
-rw-r--r-- | src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp | 754 | ||||
-rw-r--r-- | src/runtime/cpu/operators/CpuWinogradConv2d.cpp | 848 | ||||
-rw-r--r-- | src/runtime/cpu/operators/CpuWinogradConv2d.h | 137 |
6 files changed, 1254 insertions, 948 deletions
diff --git a/src/core/NEON/NEKernels.h b/src/core/NEON/NEKernels.h index cd09544d31..6c6c51dd87 100644 --- a/src/core/NEON/NEKernels.h +++ b/src/core/NEON/NEKernels.h @@ -66,6 +66,5 @@ #include "src/core/NEON/kernels/NEStridedSliceKernel.h" #include "src/core/NEON/kernels/NETileKernel.h" #include "src/core/NEON/kernels/NEWeightsReshapeKernel.h" -#include "src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h" #endif /* ARM_COMPUTE_NEKERNELS_H */ diff --git a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp b/src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp index be34980663..74b031b226 100644 --- a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp +++ b/src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp @@ -21,7 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h" +#include "src/core/cpu/kernels/CpuWinogradConv2dKernel.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" @@ -39,6 +39,8 @@ namespace arm_compute { +namespace cpu +{ //Batched Gemms namespace @@ -175,7 +177,7 @@ std::pair<Status, Window> validate_and_configure_window_winograd_output_trans(IT } } // namespace -Status INEWinogradLayerTransformWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *weights) +Status ICpuWinogradConv2dTransformWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *weights) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); @@ -189,7 +191,7 @@ Status INEWinogradLayerTransformWeightsKernel::validate(const ITensorInfo *input } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -unsigned int NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_weight_storage_size(int num_output_channels, int num_input_channels) const +unsigned int CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_weight_storage_size(int num_output_channels, int num_input_channels) const { const KernelShape shape(num_output_channels, KernelRows, KernelCols, num_input_channels); return static_cast<unsigned int>( @@ -198,89 +200,94 @@ unsigned int NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTile } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::NEWinogradLayerTransformWeightsKernel() - : _transform(nullptr), _weights_hwio(nullptr), _output(nullptr), _matrix_stride(0), _num_output_channels(0), _num_input_channels(0) +CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::CpuWinogradConv2dTransformWeightsKernel() + : _transform(nullptr), _num_output_channels(0), _matrix_stride(0) { } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -int NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_matrix_stride(int num_output_channels, int num_input_channels) const +int CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_matrix_stride(int num_output_channels, int num_input_channels) const { return WinogradConv::get_kernel_matrix_stride(num_input_channels, num_output_channels); } #ifndef DOXYGEN_SKIP_THIS template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure( - const ITensor *weights_hwio, - ITensor *output, - const int matrix_stride, /** Stride across matrices in the output. */ - const int num_output_channels, /** Number of filters. */ - const int num_input_channels) /** Number of channels in each filter. */ -{ - _weights_hwio = weights_hwio; - _output = output; - _matrix_stride = matrix_stride; - _num_output_channels = num_output_channels; - _num_input_channels = num_input_channels; +void CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure( + const ITensorInfo *weights_hwio, + ITensorInfo *output, + const int matrix_stride, /** Stride across matrices in the output. */ + const int num_output_channels, /** Number of filters. */ + const int num_input_channels) /** Number of channels in each filter. */ +{ + ARM_COMPUTE_UNUSED(weights_hwio, output); + _transform = std::make_unique<WeightsTransform>(num_output_channels, num_input_channels); + _num_output_channels = num_output_channels; + _matrix_stride = matrix_stride; Window win; auto win_last = _transform->get_window(); win.set(Window::DimX, Window::Dimension(0, win_last, 1)); - INEKernel::configure(win); + ICpuKernel::configure(win); } #endif /* DOXYGEN_SKIP_THIS */ template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::run(const Window &window, const ThreadInfo &info) +void CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) { ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON(tensors.empty()); + const size_t fst = window.x().start(); const size_t lst = window.x().end(); - _transform->set_weight_tensor(_weights_hwio->buffer()); + + const ITensor *weights_hwio = tensors.get_const_tensor(TensorType::ACL_SRC); + ITensor *output = tensors.get_tensor(TensorType::ACL_DST); + + _transform->set_weight_tensor(weights_hwio->buffer()); const int matrix_row_stride = roundup(_num_output_channels, WinogradConv::N_BLOCK); - _transform->set_output_matrices(_output->buffer(), _matrix_stride, matrix_row_stride); - _transform->set_working_space(_output->buffer()); + _transform->set_output_matrices(output->buffer(), _matrix_stride, matrix_row_stride); + _transform->set_working_space(output->buffer()); _transform->run(fst, lst); } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -bool NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::is_parallelisable() const +bool CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::is_parallelisable() const { return false; } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -Status NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::validate(const ITensorInfo *input, const ITensorInfo *output, - const WinogradInfo &winograd_info) +Status CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::validate(const ITensorInfo *input, const ITensorInfo *output, + const WinogradInfo &winograd_info) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_weight_trans(input, output, winograd_info)); ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_weight_trans(input->clone().get(), output->clone().get(), winograd_info).first); return Status{}; } -template class NEWinogradLayerTransformWeightsKernel<float, 2, 2, 3, 3>; -template class NEWinogradLayerTransformWeightsKernel<float, 4, 4, 3, 3>; -template class NEWinogradLayerTransformWeightsKernel<float, 2, 2, 5, 5>; -template class NEWinogradLayerTransformWeightsKernel<float, 1, 6, 1, 3>; -template class NEWinogradLayerTransformWeightsKernel<float, 6, 1, 3, 1>; +template class CpuWinogradConv2dTransformWeightsKernel<float, 2, 2, 3, 3>; +template class CpuWinogradConv2dTransformWeightsKernel<float, 4, 4, 3, 3>; +template class CpuWinogradConv2dTransformWeightsKernel<float, 2, 2, 5, 5>; +template class CpuWinogradConv2dTransformWeightsKernel<float, 1, 6, 1, 3>; +template class CpuWinogradConv2dTransformWeightsKernel<float, 6, 1, 3, 1>; -template class NEWinogradLayerTransformWeightsKernel<float, 1, 4, 1, 5>; -template class NEWinogradLayerTransformWeightsKernel<float, 4, 1, 5, 1>; -template class NEWinogradLayerTransformWeightsKernel<float, 1, 2, 1, 7>; -template class NEWinogradLayerTransformWeightsKernel<float, 2, 1, 7, 1>; +template class CpuWinogradConv2dTransformWeightsKernel<float, 1, 4, 1, 5>; +template class CpuWinogradConv2dTransformWeightsKernel<float, 4, 1, 5, 1>; +template class CpuWinogradConv2dTransformWeightsKernel<float, 1, 2, 1, 7>; +template class CpuWinogradConv2dTransformWeightsKernel<float, 2, 1, 7, 1>; #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template class NEWinogradLayerTransformWeightsKernel<__fp16, 4, 4, 3, 3>; +template class CpuWinogradConv2dTransformWeightsKernel<__fp16, 4, 4, 3, 3>; #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC // Input transform template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -unsigned int NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_input_storage_size( +unsigned int CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_input_storage_size( int num_batches, /* Number of batches in the input tensor. */ int num_channels, /* Number of feature maps in the input tensor. */ int num_rows, /* Number of rows in each feature map. */ @@ -296,13 +303,13 @@ unsigned int NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCo } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -unsigned int NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_working_space_size(unsigned int num_threads) const +unsigned int CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_working_space_size(unsigned int num_threads) const { return _transform->get_working_space_size(num_threads) / sizeof(T); } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -int NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_matrix_stride( +int CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_matrix_stride( int num_batches, /* Number of batches in the input tensor. */ int num_channels, /* Number of feature maps in the input tensor. */ int num_rows, /* Number of rows in each feature map. */ @@ -313,38 +320,32 @@ int NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, Kerne } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::NEWinogradLayerTransformInputKernel() - : _transform(nullptr), _input_nhwc(nullptr), _num_batches(0), _num_rows(0), _num_cols(0), _num_channels(0), _padding(), _output(nullptr), _matrix_stride(0), _padding_top(), _padding_left(), - _padding_right(), _padding_bottom(), _workspace(nullptr) +CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::CpuWinogradConv2dTransformInputKernel() + : _transform(nullptr), _num_channels(0), _matrix_stride(0) { } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure( - const ITensor *input_nhwc, - const int num_batches, /* Number of batches in input tensor. */ - const int num_rows, /* Number of rows in input tensor. */ - const int num_cols, /* Number of columns in input tensor. */ - const int num_channels, /* Number of channels in input tensor. */ - const PaddingType padding, /* Padding type. */ - ITensor *output, /* Base of output matrices. */ - const int matrix_stride, /* Stride between output matrices. */ - ITensor *workspace) -{ - _input_nhwc = input_nhwc; - _num_batches = num_batches; - _num_rows = num_rows; - _num_cols = num_cols; +void CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure( + const ITensorInfo *input_nhwc, + const int num_batches, /* Number of batches in input tensor. */ + const int num_rows, /* Number of rows in input tensor. */ + const int num_cols, /* Number of columns in input tensor. */ + const int num_channels, /* Number of channels in input tensor. */ + const PaddingType padding, /* Padding type. */ + ITensorInfo *output, /* Base of output matrices. */ + const int matrix_stride, /* Stride between output matrices. */ + ITensorInfo *workspace) +{ + ARM_COMPUTE_UNUSED(input_nhwc, output, matrix_stride, workspace); + _num_channels = num_channels; - _padding = padding; - _output = output; _matrix_stride = matrix_stride; - _workspace = workspace; - _padding_top = (padding == PADDING_SAME) ? (KernelRows - 1) / 2 : 0; - _padding_left = (padding == PADDING_SAME) ? (KernelCols - 1) / 2 : 0; - _padding_bottom = (padding == PADDING_SAME) ? iceildiv(KernelRows - 1, 2) : 0; - _padding_right = (padding == PADDING_SAME) ? iceildiv(KernelCols - 1, 2) : 0; + const int padding_top = (padding == PADDING_SAME) ? (KernelRows - 1) / 2 : 0; + const int padding_left = (padding == PADDING_SAME) ? (KernelCols - 1) / 2 : 0; + const int padding_bottom = (padding == PADDING_SAME) ? iceildiv(KernelRows - 1, 2) : 0; + const int padding_right = (padding == PADDING_SAME) ? iceildiv(KernelCols - 1, 2) : 0; _transform = std::make_unique<InputTransform>( KernelRows, @@ -353,37 +354,41 @@ void NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, Kern num_rows, num_cols, num_channels, - _padding_top, /**< Padding to apply to the top of the image. */ - _padding_left, /**< Padding to apply to the left of the image. */ - _padding_bottom, /**< Padding to apply to the bottom of the image. */ - _padding_right /**< Padding to apply to the right of the image. */ + padding_top, /**< Padding to apply to the top of the image. */ + padding_left, /**< Padding to apply to the left of the image. */ + padding_bottom, /**< Padding to apply to the bottom of the image. */ + padding_right /**< Padding to apply to the right of the image. */ ); Window win; auto win_last = _transform->get_window(); win.set(Window::DimX, Window::Dimension(0, win_last, 1)); - INEKernel::configure(win); + ICpuKernel::configure(win); } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::run(const Window &window, const ThreadInfo &info) +void CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) { ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_NULLPTR(_workspace); - - 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<const T *>(_input_nhwc->buffer() + _input_nhwc->info()->offset_first_element_in_bytes()); - auto output_ptr = reinterpret_cast<T *>(_output->buffer() + _output->info()->offset_first_element_in_bytes()); + ARM_COMPUTE_ERROR_ON(tensors.empty()); + + const ITensor *input_nhwc = tensors.get_const_tensor(TensorType::ACL_SRC); + const ITensor *workspace = tensors.get_const_tensor(TensorType::ACL_INT); + ITensor *output = tensors.get_tensor(TensorType::ACL_DST); + + 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<const T *>(input_nhwc->buffer() + input_nhwc->info()->offset_first_element_in_bytes()); + auto output_ptr = reinterpret_cast<T *>(output->buffer() + output->info()->offset_first_element_in_bytes()); ARM_COMPUTE_ERROR_ON_NULLPTR(output_ptr); _transform->set_input_tensor(input_nhwc_ptr, input_batch_stride, input_row_stride, input_col_stride); _transform->set_output_matrices(output_ptr, _matrix_stride, _num_channels); - _transform->set_working_space(_workspace->buffer()); + _transform->set_working_space(workspace->buffer()); // The code below cannot be moved to configure because biases hasn't been allocated at that point const size_t fst = window.x().start(); @@ -392,7 +397,8 @@ void NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, Kern } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -Status NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) +Status CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::validate(const ITensorInfo *input, const ITensorInfo *output, + const WinogradInfo &winograd_info) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_input_trans(input, output, winograd_info)); ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_input_trans(input->clone().get(), output->clone().get(), winograd_info).first); @@ -400,25 +406,25 @@ Status NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, Ke return Status{}; } -template class NEWinogradLayerTransformInputKernel<float, 2, 2, 3, 3>; -template class NEWinogradLayerTransformInputKernel<float, 4, 4, 3, 3>; -template class NEWinogradLayerTransformInputKernel<float, 2, 2, 5, 5>; -template class NEWinogradLayerTransformInputKernel<float, 1, 6, 1, 3>; -template class NEWinogradLayerTransformInputKernel<float, 6, 1, 3, 1>; +template class CpuWinogradConv2dTransformInputKernel<float, 2, 2, 3, 3>; +template class CpuWinogradConv2dTransformInputKernel<float, 4, 4, 3, 3>; +template class CpuWinogradConv2dTransformInputKernel<float, 2, 2, 5, 5>; +template class CpuWinogradConv2dTransformInputKernel<float, 1, 6, 1, 3>; +template class CpuWinogradConv2dTransformInputKernel<float, 6, 1, 3, 1>; -template class NEWinogradLayerTransformInputKernel<float, 1, 4, 1, 5>; -template class NEWinogradLayerTransformInputKernel<float, 4, 1, 5, 1>; -template class NEWinogradLayerTransformInputKernel<float, 1, 2, 1, 7>; -template class NEWinogradLayerTransformInputKernel<float, 2, 1, 7, 1>; +template class CpuWinogradConv2dTransformInputKernel<float, 1, 4, 1, 5>; +template class CpuWinogradConv2dTransformInputKernel<float, 4, 1, 5, 1>; +template class CpuWinogradConv2dTransformInputKernel<float, 1, 2, 1, 7>; +template class CpuWinogradConv2dTransformInputKernel<float, 2, 1, 7, 1>; #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template class NEWinogradLayerTransformInputKernel<__fp16, 4, 4, 3, 3>; +template class CpuWinogradConv2dTransformInputKernel<__fp16, 4, 4, 3, 3>; #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC // Output transform template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -unsigned int NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_output_storage_size( +unsigned int CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_output_storage_size( int num_batches, /* Number of batches in the output tensor. */ int num_rows, /* Number of rows in each feature map of the input tensor. */ int num_cols, /* Number of columns in each feature map of the input tensor. */ @@ -434,20 +440,19 @@ unsigned int NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileC } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::NEWinogradLayerTransformOutputKernel() - : _transform(nullptr), _biases(nullptr), _transformed_output(nullptr), _workspace(nullptr), _matrix_stride(0), _matrix_row_stride(0), _output_nhwc(nullptr), _num_batches(0), _num_rows(0), - _num_cols(0), _num_channels(0) +CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::CpuWinogradConv2dTransformOutputKernel() + : _transform(nullptr), _matrix_stride(0), _matrix_row_stride(0) { } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -unsigned int NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_working_space_size(unsigned int num_threads) const +unsigned int CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_working_space_size(unsigned int num_threads) const { return _transform->get_working_space_size(num_threads) / sizeof(T); } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -int NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_matrix_stride( +int CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_matrix_stride( int num_batches, /* Number of batches in the output tensor. */ int num_rows, /* Number of rows in each feature map of the input tensor. */ int num_cols, /* Number of columns in each feature map of the input tensor. */ @@ -458,7 +463,7 @@ int NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, Kern } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -std::pair<unsigned int, unsigned int> NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_output_shape( +std::pair<unsigned int, unsigned int> CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_output_shape( int num_rows, /* Number of rows in each feature map of the input tensor. */ int num_cols, /* Number of columns in each feature map of the input tensor. */ bool padding_same) const @@ -467,54 +472,52 @@ std::pair<unsigned int, unsigned int> NEWinogradLayerTransformOutputKernel<T, Ou } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure( - const ITensor *biases, - const ITensor *transformed_output, +void CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure( + const ITensorInfo *biases, + const ITensorInfo *transformed_output, const int matrix_stride, - ITensor *output_nhwc, + ITensorInfo *output_nhwc, const int num_batches, const int num_rows, const int num_cols, const int num_channels, - ITensor *workspace, + ITensorInfo *workspace, const arm_gemm::Activation &activation) { - _biases = biases; - _workspace = workspace; - _transformed_output = transformed_output; - _matrix_stride = matrix_stride; - _matrix_row_stride = roundup(num_channels, WinogradConv::N_BLOCK); - _output_nhwc = output_nhwc; - _num_batches = num_batches; - _num_rows = num_rows; - _num_cols = num_cols; - _num_channels = num_channels; + ARM_COMPUTE_UNUSED(biases, transformed_output, output_nhwc, num_batches, num_rows, num_cols, workspace, activation); + + _matrix_stride = matrix_stride; + _matrix_row_stride = roundup(num_channels, WinogradConv::N_BLOCK); + // We don't have the biases buffer at this stage as it hasn't been allocated, we pass in nullptr OutputTransform is only used here to compute the window _transform = std::make_unique<OutputTransform>(num_batches, num_rows, num_cols, num_channels, activation); Window win; auto win_last = _transform->get_window(); win.set(Window::DimX, Window::Dimension(0, win_last, 1)); - INEKernel::configure(win); + ICpuKernel::configure(win); } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::run(const Window &window, const ThreadInfo &info) +void CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) { - ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_NULLPTR(_workspace); - ARM_COMPUTE_ERROR_ON_NULLPTR(_transformed_output); - ARM_COMPUTE_ERROR_ON_NULLPTR(_output_nhwc); - - const int out_batch_stride = _output_nhwc->info()->strides_in_bytes()[3] / sizeof(T); - const int out_row_stride = _output_nhwc->info()->strides_in_bytes()[2] / sizeof(T); - const int out_col_stride = _output_nhwc->info()->strides_in_bytes()[1] / sizeof(T); - - _transform->set_input_matrices(_transformed_output->buffer(), _matrix_stride, _matrix_row_stride); - _transform->set_bias((_biases ? reinterpret_cast<T *>(_biases->buffer() + _biases->info()->offset_first_element_in_bytes()) : nullptr)); - _transform->set_output_tensor(_output_nhwc->buffer() + _output_nhwc->info()->offset_first_element_in_bytes(), out_batch_stride, out_row_stride, out_col_stride); - _transform->set_working_space(_workspace->buffer()); + ARM_COMPUTE_ERROR_ON(tensors.empty()); + + const ITensor *biases = tensors.get_const_tensor(TensorType::ACL_SRC_0); + const ITensor *transformed_output = tensors.get_const_tensor(TensorType::ACL_SRC_1); + ITensor *workspace = tensors.get_tensor(TensorType::ACL_INT); + ITensor *dst_nhwc = tensors.get_tensor(TensorType::ACL_DST); + + const int out_batch_stride = dst_nhwc->info()->strides_in_bytes()[3] / sizeof(T); + const int out_row_stride = dst_nhwc->info()->strides_in_bytes()[2] / sizeof(T); + const int out_col_stride = dst_nhwc->info()->strides_in_bytes()[1] / sizeof(T); + + _transform->set_input_matrices(transformed_output->buffer(), _matrix_stride, _matrix_row_stride); + _transform->set_bias((biases ? reinterpret_cast<T *>(biases->buffer() + biases->info()->offset_first_element_in_bytes()) : nullptr)); + _transform->set_output_tensor(dst_nhwc->buffer() + dst_nhwc->info()->offset_first_element_in_bytes(), out_batch_stride, out_row_stride, out_col_stride); + _transform->set_working_space(workspace->buffer()); + // The code below cannot be moved to configure because biases hasn't been allocated at that point const size_t fst = window.x().start(); const size_t lst = window.x().end(); @@ -522,8 +525,8 @@ void NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, Ker } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -Status NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, - const WinogradInfo &winograd_info) +Status CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, + const WinogradInfo &winograd_info) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_output_trans(input, (bias != nullptr ? bias->clone().get() : nullptr), output, winograd_info)); ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_output_trans(input->clone().get(), output->clone().get(), winograd_info).first); @@ -531,18 +534,19 @@ Status NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, K return Status{}; } -template class NEWinogradLayerTransformOutputKernel<float, 2, 2, 3, 3>; -template class NEWinogradLayerTransformOutputKernel<float, 4, 4, 3, 3>; -template class NEWinogradLayerTransformOutputKernel<float, 2, 2, 5, 5>; -template class NEWinogradLayerTransformOutputKernel<float, 1, 6, 1, 3>; -template class NEWinogradLayerTransformOutputKernel<float, 6, 1, 3, 1>; +template class CpuWinogradConv2dTransformOutputKernel<float, 2, 2, 3, 3>; +template class CpuWinogradConv2dTransformOutputKernel<float, 4, 4, 3, 3>; +template class CpuWinogradConv2dTransformOutputKernel<float, 2, 2, 5, 5>; +template class CpuWinogradConv2dTransformOutputKernel<float, 1, 6, 1, 3>; +template class CpuWinogradConv2dTransformOutputKernel<float, 6, 1, 3, 1>; -template class NEWinogradLayerTransformOutputKernel<float, 1, 4, 1, 5>; -template class NEWinogradLayerTransformOutputKernel<float, 4, 1, 5, 1>; -template class NEWinogradLayerTransformOutputKernel<float, 1, 2, 1, 7>; -template class NEWinogradLayerTransformOutputKernel<float, 2, 1, 7, 1>; +template class CpuWinogradConv2dTransformOutputKernel<float, 1, 4, 1, 5>; +template class CpuWinogradConv2dTransformOutputKernel<float, 4, 1, 5, 1>; +template class CpuWinogradConv2dTransformOutputKernel<float, 1, 2, 1, 7>; +template class CpuWinogradConv2dTransformOutputKernel<float, 2, 1, 7, 1>; #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template class NEWinogradLayerTransformOutputKernel<__fp16, 4, 4, 3, 3>; +template class CpuWinogradConv2dTransformOutputKernel<__fp16, 4, 4, 3, 3>; #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +} // namespace cpu } // namespace arm_compute diff --git a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h b/src/core/cpu/kernels/CpuWinogradConv2dKernel.h index 75d257de4b..b5a29ffd02 100644 --- a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h +++ b/src/core/cpu/kernels/CpuWinogradConv2dKernel.h @@ -21,22 +21,21 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_NEGEMMWINOGRADCONVOLUTIONLAYERKERNEL_H -#define ARM_COMPUTE_NEGEMMWINOGRADCONVOLUTIONLAYERKERNEL_H +#ifndef ARM_COMPUTE_CPUWINOGRADCONV2DKERNEL_H +#define ARM_COMPUTE_CPUWINOGRADCONV2DKERNEL_H -#include "src/core/NEON/INEKernel.h" #include "src/core/NEON/kernels/convolution/common/convolution.hpp" #include "src/core/NEON/kernels/convolution/common/tensor.hpp" +#include "src/core/cpu/ICpuKernel.h" #include "src/core/NEON/kernels/convolution/winograd/winograd_layer.hpp" namespace arm_compute { -// Forward declarations -class ITensor; - +namespace cpu +{ /** Interface for the kernel to perform Winograd input transform. */ -class INEWinogradLayerTransformInputKernel : public INEKernel +class ICpuWinogradConv2dTransformInputKernel : public ICpuKernel { public: /** Get the working space required to perform the transformation. @@ -87,30 +86,30 @@ public: * @param[in] matrix_stride Stride between output matrices. * @param[in] workspace Tensor to be used as the working space during the computation. */ - virtual void configure(const ITensor *input_nhwc, const int num_batches, const int num_rows, const int num_cols, const int num_channels, - const PaddingType padding, ITensor *output, const int matrix_stride, ITensor *workspace) = 0; + virtual void configure(const ITensorInfo *input_nhwc, const int num_batches, const int num_rows, const int num_cols, const int num_channels, + const PaddingType padding, ITensorInfo *output, const int matrix_stride, ITensorInfo *workspace) = 0; /** Destructor */ - virtual ~INEWinogradLayerTransformInputKernel() + virtual ~ICpuWinogradConv2dTransformInputKernel() { } }; /** Kernel to perform Winograd input transform. */ template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -class NEWinogradLayerTransformInputKernel : public INEWinogradLayerTransformInputKernel +class CpuWinogradConv2dTransformInputKernel : public ICpuWinogradConv2dTransformInputKernel { public: /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformInputKernel(const NEWinogradLayerTransformInputKernel &) = delete; + CpuWinogradConv2dTransformInputKernel(const CpuWinogradConv2dTransformInputKernel &) = delete; /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformInputKernel &operator=(const NEWinogradLayerTransformInputKernel &) = delete; + CpuWinogradConv2dTransformInputKernel &operator=(const CpuWinogradConv2dTransformInputKernel &) = delete; /** Allow instances of this class to be moved */ - NEWinogradLayerTransformInputKernel(NEWinogradLayerTransformInputKernel &&) = default; + CpuWinogradConv2dTransformInputKernel(CpuWinogradConv2dTransformInputKernel &&) = default; /** Allow instances of this class to be moved */ - NEWinogradLayerTransformInputKernel &operator=(NEWinogradLayerTransformInputKernel &&) = default; + CpuWinogradConv2dTransformInputKernel &operator=(CpuWinogradConv2dTransformInputKernel &&) = default; /** Default destructor */ - ~NEWinogradLayerTransformInputKernel() = default; + ~CpuWinogradConv2dTransformInputKernel() = default; /** Determine how much memory (in units of TIn) to allocate for the * transformed input. @@ -160,11 +159,11 @@ public: bool same_padding) const override; /** Default constructor */ - NEWinogradLayerTransformInputKernel(); + CpuWinogradConv2dTransformInputKernel(); const char *name() const override { - return "NEWinogradLayerTransformInputKernel"; + return "CpuWinogradConv2dTransformInputKernel"; } /** Configure the output transform kernel. @@ -180,25 +179,25 @@ public: * @param[in] workspace Tensor to be used as the working space during the computation. */ void configure( - const ITensor *input_nhwc, - const int num_batches, - const int num_rows, - const int num_cols, - const int num_channels, - const PaddingType padding, - ITensor *output, - const int matrix_stride, - ITensor *workspace) override; + const ITensorInfo *input_nhwc, + const int num_batches, + const int num_rows, + const int num_cols, + const int num_channels, + const PaddingType padding, + ITensorInfo *output, + const int matrix_stride, + ITensorInfo *workspace) override; // Inherited methods overridden: - void run(const Window &window, const ThreadInfo &info) override; + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; /** Winograd base kernel */ using WinogradBase = winograd::WinogradGEMM<OutputTileRows, OutputTileCols, KernelRows, KernelCols, winograd::WinogradRoots::Integers>; /** Winograd convolution kernel */ using WinogradConv = typename WinogradBase::template Convolution<T, T>; - /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformInputKernel + /** Static function to check if given info will lead to a valid configuration of @ref CpuWinogradConv2dTransformInputKernel * * @param[in] input First tensor input info. Data types supported: F16/F32. * @param[in] output Output tensor info. Data types supported: same as @p input. @@ -212,23 +211,12 @@ private: using InputTransform = typename WinogradBase::template InputTransform<T, T>; std::unique_ptr<InputTransform> _transform{ nullptr }; - const ITensor *_input_nhwc; - int _num_batches; /**< Number of batches in input tensor. */ - int _num_rows; /**< Number of rows in input tensor. */ - int _num_cols; /**< Number of columns in input tensor. */ - int _num_channels; /**< Number of channels in input tensor. */ - PaddingType _padding; /**< Padding type. */ - ITensor *_output; /**< Base of output matrices. */ - int _matrix_stride; /**< Stride between output matrices. */ - int _padding_top; /**< Padding to apply to the top of the image. */ - int _padding_left; /**< Padding to apply to the left of the image. */ - int _padding_right; /**< Padding to apply to the right of the image. */ - int _padding_bottom; /**< Padding to apply to the bottom of the image. */ - ITensor *_workspace; + int _num_channels; /**< Number of channels in input tensor. */ + int _matrix_stride; /**< Stride between output matrices. */ }; /** Interface for the kernel to perform Winograd output transform. */ -class INEWinogradLayerTransformOutputKernel : public INEKernel +class ICpuWinogradConv2dTransformOutputKernel : public ICpuKernel { public: /** Get the working space required to perform the transformation. @@ -294,44 +282,44 @@ public: * @param[in] activation Activation to be used */ virtual void configure( - const ITensor *biases, - const ITensor *transformed_output, + const ITensorInfo *biases, + const ITensorInfo *transformed_output, const int matrix_stride, - ITensor *output_nhwc, + ITensorInfo *output_nhwc, const int num_batches, const int num_rows, const int num_cols, const int num_channels, - ITensor *workspace, + ITensorInfo *workspace, const arm_gemm::Activation &activation) = 0; - virtual ~INEWinogradLayerTransformOutputKernel() + virtual ~ICpuWinogradConv2dTransformOutputKernel() { } }; /** Kernel to perform Winograd output transform. */ template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -class NEWinogradLayerTransformOutputKernel : public INEWinogradLayerTransformOutputKernel +class CpuWinogradConv2dTransformOutputKernel : public ICpuWinogradConv2dTransformOutputKernel { public: const char *name() const override { - return "NEWinogradLayerTransformOutputKernel"; + return "CpuWinogradConv2dTransformOutputKernel"; } /** Constructor */ - NEWinogradLayerTransformOutputKernel(); + CpuWinogradConv2dTransformOutputKernel(); /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformOutputKernel(const NEWinogradLayerTransformOutputKernel &) = delete; + CpuWinogradConv2dTransformOutputKernel(const CpuWinogradConv2dTransformOutputKernel &) = delete; /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformOutputKernel &operator=(const NEWinogradLayerTransformOutputKernel &) = delete; + CpuWinogradConv2dTransformOutputKernel &operator=(const CpuWinogradConv2dTransformOutputKernel &) = delete; /** Allow instances of this class to be moved */ - NEWinogradLayerTransformOutputKernel(NEWinogradLayerTransformOutputKernel &&) = default; + CpuWinogradConv2dTransformOutputKernel(CpuWinogradConv2dTransformOutputKernel &&) = default; /** Allow instances of this class to be moved */ - NEWinogradLayerTransformOutputKernel &operator=(NEWinogradLayerTransformOutputKernel &&) = default; + CpuWinogradConv2dTransformOutputKernel &operator=(CpuWinogradConv2dTransformOutputKernel &&) = default; /** Default destructor */ - ~NEWinogradLayerTransformOutputKernel() = default; + ~CpuWinogradConv2dTransformOutputKernel() = default; // Inherited methods overridden: /** Determine how much memory (in units of TOut) to allocate for the @@ -395,20 +383,20 @@ public: * @param[in] activation Activation to be used */ void configure( - const ITensor *biases, - const ITensor *transformed_output, + const ITensorInfo *biases, + const ITensorInfo *transformed_output, const int matrix_stride, - ITensor *output_nhwc, + ITensorInfo *output_nhwc, const int num_batches, const int num_rows, const int num_cols, const int num_channels, - ITensor *workspace, + ITensorInfo *workspace, const arm_gemm::Activation &activation) override; - void run(const Window &window, const ThreadInfo &info) override; + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; - /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformOutputKernel + /** Static function to check if given info will lead to a valid configuration of @ref CpuWinogradConv2dTransformOutputKernel * * @param[in] input Source tensor info with shape [C, N, 16, batches] or [C, N, 36, batches]. Data types supported: F16/F32. * @param[in] bias Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input @@ -425,35 +413,27 @@ private: using OutputTransform = typename WinogradBase::template OutputTransform<T, T>; std::unique_ptr<OutputTransform> _transform{ nullptr }; - const ITensor *_biases; - const ITensor *_transformed_output; - ITensor *_workspace; int _matrix_stride; int _matrix_row_stride; - ITensor *_output_nhwc; - int _num_batches; - int _num_rows; - int _num_cols; - int _num_channels; }; /** Interface for the kernel to perform Winograd weights transform. */ -class INEWinogradLayerTransformWeightsKernel : public INEKernel +class ICpuWinogradConv2dTransformWeightsKernel : public ICpuKernel { public: /** Prevent instances of this class from being copied (As this class contains pointers) */ - INEWinogradLayerTransformWeightsKernel(const INEWinogradLayerTransformWeightsKernel &) = default; + ICpuWinogradConv2dTransformWeightsKernel(const ICpuWinogradConv2dTransformWeightsKernel &) = default; /** Prevent instances of this class from being copied (As this class contains pointers) */ - INEWinogradLayerTransformWeightsKernel &operator=(const INEWinogradLayerTransformWeightsKernel &) = default; + ICpuWinogradConv2dTransformWeightsKernel &operator=(const ICpuWinogradConv2dTransformWeightsKernel &) = default; /** Allow instances of this class to be moved */ - INEWinogradLayerTransformWeightsKernel(INEWinogradLayerTransformWeightsKernel &&) = default; + ICpuWinogradConv2dTransformWeightsKernel(ICpuWinogradConv2dTransformWeightsKernel &&) = default; /** Allow instances of this class to be moved */ - INEWinogradLayerTransformWeightsKernel &operator=(INEWinogradLayerTransformWeightsKernel &&) = default; + ICpuWinogradConv2dTransformWeightsKernel &operator=(ICpuWinogradConv2dTransformWeightsKernel &&) = default; - INEWinogradLayerTransformWeightsKernel() + ICpuWinogradConv2dTransformWeightsKernel() { } - virtual ~INEWinogradLayerTransformWeightsKernel() + virtual ~ICpuWinogradConv2dTransformWeightsKernel() { } /** Determine how much memory (in units of T) to allocate for the @@ -476,16 +456,16 @@ public: /** Configure the weights transform kernel. * - * @param[in] weights_hwio Pointer to the weights tensor + * @param[in] weights_hwio Pointer to the weights tensor info * @param[out] output Pointer to working space for the output tensor in the Winograd domain. * @param[in] matrix_stride Stride across matrices in the output workspace. * @param[in] num_output_channels Number of filters. * @param[in] num_input_channels Number of channels in each filter. */ - virtual void configure(const ITensor *weights_hwio, ITensor *output, const int matrix_stride, const int num_output_channels, const int num_input_channels) = 0; + virtual void configure(const ITensorInfo *weights_hwio, ITensorInfo *output, const int matrix_stride, const int num_output_channels, const int num_input_channels) = 0; - /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformWeightsKernel + /** Static function to check if given info will lead to a valid configuration of @ref CpuWinogradConv2dTransformWeightsKernel * * @param[in] input First tensor input info. Data types supported: F16/F32. * @param[in] weights Weights tensor info. Data types supported: same as @p input. @@ -497,28 +477,28 @@ public: /** Kernel to perform Winograd weights transform. */ template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -class NEWinogradLayerTransformWeightsKernel final : public INEWinogradLayerTransformWeightsKernel +class CpuWinogradConv2dTransformWeightsKernel final : public ICpuWinogradConv2dTransformWeightsKernel { public: /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformWeightsKernel(const NEWinogradLayerTransformWeightsKernel &) = delete; + CpuWinogradConv2dTransformWeightsKernel(const CpuWinogradConv2dTransformWeightsKernel &) = delete; /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformWeightsKernel &operator=(const NEWinogradLayerTransformWeightsKernel &) = delete; + CpuWinogradConv2dTransformWeightsKernel &operator=(const CpuWinogradConv2dTransformWeightsKernel &) = delete; /** Allow instances of this class to be moved */ - NEWinogradLayerTransformWeightsKernel(NEWinogradLayerTransformWeightsKernel &&) = default; + CpuWinogradConv2dTransformWeightsKernel(CpuWinogradConv2dTransformWeightsKernel &&) = default; /** Allow instances of this class to be moved */ - NEWinogradLayerTransformWeightsKernel &operator=(NEWinogradLayerTransformWeightsKernel &&) = default; + CpuWinogradConv2dTransformWeightsKernel &operator=(CpuWinogradConv2dTransformWeightsKernel &&) = default; /** Default destructor */ - ~NEWinogradLayerTransformWeightsKernel() = default; + ~CpuWinogradConv2dTransformWeightsKernel() = default; /** Default constructor. */ - NEWinogradLayerTransformWeightsKernel(); + CpuWinogradConv2dTransformWeightsKernel(); const char *name() const override { - return "NEWinogradLayerTransformWeightsKernel"; + return "CpuWinogradConv2dTransformWeightsKernel"; } - /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformWeightsKernel + /** Static function to check if given info will lead to a valid configuration of @ref CpuWinogradConv2dTransformWeightsKernel * * @param[in] input Source tensor info. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout). * kernel_x must be 3 and equal to kernel_y. Data types supported: F16/F32. @@ -534,13 +514,13 @@ public: #ifndef DOXYGEN_SKIP_THIS /** Configure the weights transform kernel. * - * @param[in] weights_hwio Pointer to the weights tensor + * @param[in] weights_hwio Pointer to the weights tensor info * @param[out] output Pointer to working space for the output tensor in the Winograd domain. * @param[in] matrix_stride Stride across matrices in the output workspace. * @param[in] num_output_channels Number of filters. * @param[in] num_input_channels Number of channels in each filter. */ - void configure(const ITensor *weights_hwio, ITensor *output, const int matrix_stride, const int num_output_channels, const int num_input_channels) override; + void configure(const ITensorInfo *weights_hwio, ITensorInfo *output, const int matrix_stride, const int num_output_channels, const int num_input_channels) override; #endif /* DOXYGEN_SKIP_THIS */ /** Determine how much memory (in units of T) to allocate for the @@ -561,7 +541,7 @@ public: * @return Stride expressed in bytes. */ int get_matrix_stride(int num_output_channels, int num_input_channels) const override; - void run(const Window &window, const ThreadInfo &info) override; + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; bool is_parallelisable() const override; private: @@ -570,16 +550,13 @@ private: using WeightsTransform = typename WinogradBase::template WeightsTransform<T, T>; std::unique_ptr<WeightsTransform> _transform{ nullptr }; - const ITensor *_weights_hwio; - ITensor *_output; - int _matrix_stride; int _num_output_channels; - int _num_input_channels; + int _matrix_stride; }; /** Kernel to perform Winograd. */ template <typename TIn, typename TOut, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -class NEWinogradLayerConfiguration +class CpuWinogradConv2dConfiguration { public: /** Winograd base kernel */ @@ -588,10 +565,11 @@ public: using WinogradConv = typename WinogradBase::template Convolution<TIn, TOut>; - using TransformInputKernel = NEWinogradLayerTransformInputKernel<TIn, OutputTileRows, OutputTileCols, KernelRows, KernelCols>; - using TransformWeightsKernel = NEWinogradLayerTransformWeightsKernel<TIn, OutputTileRows, OutputTileCols, KernelRows, KernelCols>; - using TransformOutputKernel = NEWinogradLayerTransformOutputKernel<TOut, OutputTileRows, OutputTileCols, KernelRows, KernelCols>; + using TransformInputKernel = CpuWinogradConv2dTransformInputKernel<TIn, OutputTileRows, OutputTileCols, KernelRows, KernelCols>; + using TransformWeightsKernel = CpuWinogradConv2dTransformWeightsKernel<TIn, OutputTileRows, OutputTileCols, KernelRows, KernelCols>; + using TransformOutputKernel = CpuWinogradConv2dTransformOutputKernel<TOut, OutputTileRows, OutputTileCols, KernelRows, KernelCols>; }; +} // namespace cpu } // namespace arm_compute -#endif /*ARM_COMPUTE_NEGEMMWINOGRADCONVOLUTIONLAYERKERNEL_H*/ +#endif /*ARM_COMPUTE_CPUWINOGRADCONV2DKERNEL_H*/ diff --git a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp index 57950d5126..745179c050 100644 --- a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp @@ -24,754 +24,94 @@ #include "arm_compute/runtime/NEON/functions/NEWinogradConvolutionLayer.h" #include "arm_compute/core/Error.h" +#include "arm_compute/core/ITensorPack.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "arm_compute/runtime/NEON/NEScheduler.h" #include "src/core/CPP/Validate.h" -#include "src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h" +#include "src/core/cpu/kernels/CpuWinogradConv2dKernel.h" +#include "src/core/helpers/MemoryHelpers.h" +#include "src/runtime/cpu/operators/CpuWinogradConv2d.h" #include "src/core/NEON/kernels/convolution/common/utils.hpp" #include "src/core/NEON/kernels/convolution/winograd/winograd.hpp" namespace arm_compute { -namespace -{ -inline Status validate_kernel_3x3(const Size2D input_dims, const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); - - if(input->data_type() == DataType::F32) - { - if(input_dims.width > 4 && input_dims.height > 4) - { - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<float, 4, 4, 3, 3>::validate(input, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<float, 4, 4, 3, 3>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<float, 4, 4, 3, 3>::validate(batched_mm_output, biases, output, winograd_info))); - } - else - { - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<float, 2, 2, 3, 3>::validate(input, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<float, 2, 2, 3, 3>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<float, 2, 2, 3, 3>::validate(batched_mm_output, biases, output, winograd_info))); - } - } -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - else if(input->data_type() == DataType::F16) - { - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<__fp16, 4, 4, 3, 3>::validate(input, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<__fp16, 4, 4, 3, 3>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<__fp16, 4, 4, 3, 3>::validate(batched_mm_output, biases, output, winograd_info))); - } -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - - if(act_info.enabled()) - { - NEActivationLayer::validate(output, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_5x5(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<float, 2, 2, 5, 5>::validate(input, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<float, 2, 2, 5, 5>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<float, 2, 2, 5, 5>::validate(batched_mm_output, biases, output, winograd_info))); - if(act_info.enabled()) - { - NEActivationLayer::validate(output, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_3x1(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<float, 1, 6, 1, 3>::validate(input, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<float, 1, 6, 1, 3>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<float, 1, 6, 1, 3>::validate(batched_mm_output, biases, output, winograd_info))); - if(act_info.enabled()) - { - NEActivationLayer::validate(output, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_1x3(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<float, 6, 1, 3, 1>::validate(input, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<float, 6, 1, 3, 1>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<float, 6, 1, 3, 1>::validate(batched_mm_output, biases, output, winograd_info))); - - if(act_info.enabled()) - { - NEActivationLayer::validate(output, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_5x1(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<float, 1, 4, 1, 5>::validate(input, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<float, 1, 4, 1, 5>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<float, 1, 4, 1, 5>::validate(batched_mm_output, biases, output, winograd_info))); - if(act_info.enabled()) - { - NEActivationLayer::validate(output, nullptr, act_info); - } - return Status{}; -} -inline Status validate_kernel_1x5(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<float, 4, 1, 5, 1>::validate(input, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<float, 4, 1, 5, 1>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<float, 4, 1, 5, 1>::validate(batched_mm_output, biases, output, winograd_info))); - if(act_info.enabled()) - { - NEActivationLayer::validate(output, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_7x1(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<float, 1, 2, 1, 7>::validate(input, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<float, 1, 2, 1, 7>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<float, 1, 2, 1, 7>::validate(batched_mm_output, biases, output, winograd_info))); - if(act_info.enabled()) - { - NEActivationLayer::validate(output, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_1x7(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel<float, 2, 1, 7, 1>::validate(input, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel<float, 2, 1, 7, 1>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel<float, 2, 1, 7, 1>::validate(batched_mm_output, biases, output, winograd_info))); - - if(act_info.enabled()) - { - NEActivationLayer::validate(output, nullptr, act_info); - } - return Status{}; -} +using namespace arm_compute::experimental; -inline Tensor4DShape internal_get_input_shape(const arm_compute::ITensor *input) +struct NEWinogradConvolutionLayer::Impl { - const DataLayout data_layout = input->info()->data_layout(); - const int in_width = input->info()->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH)); - const int in_height = input->info()->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT)); - const int in_channels = input->info()->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL)); - const int in_batches = input->info()->dimension(3); - - return Tensor4DShape{ in_batches, in_height, in_width, in_channels }; -} - -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info) -{ - ARM_COMPUTE_UNUSED(output); - ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); - - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd layer only supports unit strides."); - if(biases != nullptr) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); - } - return INEWinogradLayerTransformWeightsKernel::validate(input, weights); -} - -Size2D winograd_output_tile(const Size2D &input_dims, const Size2D &kernel_dims, DataType data_type) -{ - Size2D output_tile = Size2D{}; - if(kernel_dims == Size2D(3U, 3U)) - { - output_tile = (input_dims.width <= 4 || input_dims.height <= 4) ? Size2D(2U, 2U) : Size2D(4U, 4U); - if(data_type == DataType::F16) - { - output_tile = Size2D(4U, 4U); - } - } - else if(kernel_dims == Size2D(5U, 5U)) - { - output_tile = Size2D(2U, 2U); - } - else if(kernel_dims == Size2D(1U, 3U)) - { - output_tile = Size2D(1U, 6U); - } - else if(kernel_dims == Size2D(3U, 1U)) - { - output_tile = Size2D(6U, 1U); - } - else if(kernel_dims == Size2D(1U, 5U)) - { - output_tile = Size2D(1U, 4U); - } - else if(kernel_dims == Size2D(5U, 1U)) - { - output_tile = Size2D(4U, 1U); - } - else if(kernel_dims == Size2D(7U, 1U)) - { - output_tile = Size2D(2U, 1U); - } - else if(kernel_dims == Size2D(1U, 7U)) - { - output_tile = Size2D(1U, 2U); - } - return output_tile; -} - -bool check_support_fast_math(const Size2D &output_tile, const Size2D &kernel_size, DataType data_type) -{ - // Check if we want to configure a Winograd configuration which requires fast math - using WinogradConfiguration = std::pair<std::pair<int, int>, std::pair<int, int>>; - - const std::vector<WinogradConfiguration> fast_math_winograd_f16 = - { - WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3)) - }; - - const std::vector<WinogradConfiguration> fast_math_winograd_f32 = - { - WinogradConfiguration(std::pair<int, int>(2, 2), std::pair<int, int>(5, 5)), - WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5)) - }; - - auto p = std::make_pair(std::pair<int, int>(output_tile.width, output_tile.height), - std::pair<int, int>(kernel_size.width, kernel_size.height)); - - switch(data_type) - { - case DataType::F16: - return std::find(fast_math_winograd_f16.begin(), fast_math_winograd_f16.end(), p) != fast_math_winograd_f16.end(); - case DataType::F32: - return std::find(fast_math_winograd_f32.begin(), fast_math_winograd_f32.end(), p) != fast_math_winograd_f32.end(); - default: - return false; - } -} - -inline bool fuse_function_supported(const ActivationLayerInfo &act_info) -{ - return act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU || act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU; -} - -arm_gemm::Activation arm_gemm_activation_from_acl_activation(const ActivationLayerInfo &act_info) -{ - switch(act_info.activation()) - { - case ActivationLayerInfo::ActivationFunction::RELU: - { - return arm_gemm::Activation(arm_gemm::Activation::Type::ReLU, act_info.a(), act_info.b()); - } - case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: - { - return arm_gemm::Activation(arm_gemm::Activation::Type::BoundedReLU, act_info.a(), act_info.b()); - } - default: - { - return arm_gemm::Activation(arm_gemm::Activation::Type::None); - } - } -} -} //namespace + MemoryGroup memory_group{}; + std::unique_ptr<cpu::CpuWinogradConv2d> op{ nullptr }; + ITensorPack run_pack{}; + ITensorPack prep_pack{}; + WorkspaceData<Tensor> workspace{}; + experimental::MemoryRequirements aux_mem_req{}; + const ITensor *original_weights{ nullptr }; + bool is_prepared{ false }; + bool is_activationlayer_enabled{ false }; + DataLayout data_layout{}; +}; NEWinogradConvolutionLayer::NEWinogradConvolutionLayer(const std::shared_ptr<IMemoryManager> &memory_manager) - : _memory_group(memory_manager), _gemm_function(memory_manager), _transform_input_kernel(nullptr), _transform_output_kernel(nullptr), _transform_weights_kernel(nullptr), _activationlayer_function(), - _permute_input(), _permute_weights(), _permute_output(), _input_transformed(), _output_transformed(), _input_workspace(), _output_workspace(), _kernel_storage(), _input_nhwc(), _output_nhwc(), - _weights_hwio(), _input(), _weights(), _output(), _is_prepared(false), _is_activationlayer_enabled(false), _data_layout() + : _impl(std::make_unique<Impl>()) { + _impl->memory_group = MemoryGroup(std::move(memory_manager)); } +NEWinogradConvolutionLayer::~NEWinogradConvolutionLayer() = default; + void NEWinogradConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math) { - ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info)); - - // Get indices for the width and height - _data_layout = input->info()->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); - const unsigned int channel_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL); - - const Size2D input_dims = Size2D(input->info()->dimension(width_idx), input->info()->dimension(height_idx)); - const Size2D kernel_size = Size2D(weights->info()->dimension(width_idx), weights->info()->dimension(height_idx)); - const DataType data_type = input->info()->data_type(); - const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, data_type); - - // Check if the Winograd configuration requires fast math - if(!enable_fast_math) - { - ARM_COMPUTE_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size, data_type), - "This Winograd configuration requires enable_fast_math=true"); - } + _impl->original_weights = weights; + _impl->op = std::make_unique<cpu::CpuWinogradConv2d>(); + _impl->op->configure(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info, act_info, enable_fast_math); - _weights = weights; - _input = input; - _output = output; - _is_prepared = false; - - int n_gemms = 1; - int N_BLOCK = 1; // Size of block used by GEMM. - - std::unique_ptr<INEWinogradLayerTransformInputKernel> transform_input_kernel; - std::unique_ptr<INEWinogradLayerTransformWeightsKernel> transform_weights_kernel; - std::unique_ptr<INEWinogradLayerTransformOutputKernel> transform_output_kernel; - - if(data_type == DataType::F32) - { - if(kernel_size == Size2D(3, 3)) - { - if(input->info()->dimension(width_idx) > 4 && input->info()->dimension(height_idx) > 4) - { - using config = NEWinogradLayerConfiguration<float, float, 4, 4, 3, 3>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else - { - using config = NEWinogradLayerConfiguration<float, float, 2, 2, 3, 3>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - } - else if(kernel_size == Size2D(5, 5)) - { - using config = NEWinogradLayerConfiguration<float, float, 2, 2, 5, 5>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(1, 3)) - { - using config = NEWinogradLayerConfiguration<float, float, 6, 1, 3, 1>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(3, 1)) - { - using config = NEWinogradLayerConfiguration<float, float, 1, 6, 1, 3>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(1, 5)) - { - using config = NEWinogradLayerConfiguration<float, float, 4, 1, 5, 1>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(5, 1)) - { - using config = NEWinogradLayerConfiguration<float, float, 1, 4, 1, 5>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(1, 7)) - { - using config = NEWinogradLayerConfiguration<float, float, 2, 1, 7, 1>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(7, 1)) - { - using config = NEWinogradLayerConfiguration<float, float, 1, 2, 1, 7>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else - { - ARM_COMPUTE_ERROR("Not supported."); - } - } -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - else if(data_type == DataType::F16) - { - if(kernel_size == Size2D(3, 3)) - { - using config = NEWinogradLayerConfiguration<__fp16, __fp16, 4, 4, 3, 3>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else - { - ARM_COMPUTE_ERROR("Not supported."); - } - } -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - else - { - ARM_COMPUTE_ERROR("Not supported."); - } - - const PaddingType use_padding_type = (conv_info.pad_top() != 0u || conv_info.pad_left() != 0) ? PADDING_SAME : PADDING_VALID; - const bool use_same_padding = use_padding_type == PADDING_SAME; - - // Get convolved dimensions - const int in_channels = input->info()->dimension(channel_idx); - const int out_channels = output->info()->dimension(channel_idx); - - const Tensor4DShape in_shape(internal_get_input_shape(input)); - const size_t data_type_size = input->info()->element_size(); - // Get the memory required to instantiate a new Winograd operator. - constexpr size_t storage_alignment = 64; - - // Kernel Storage - const size_t kernel_storage_size = transform_weights_kernel->get_weight_storage_size(out_channels, - in_channels) - * data_type_size; - - // Input storage - const size_t input_storage_size = transform_input_kernel->get_input_storage_size(in_shape.n_batches, in_shape.n_channels, in_shape.n_rows, in_shape.n_cols, - use_same_padding) - * data_type_size; - - // Output storage - const size_t output_storage_size = transform_output_kernel->get_output_storage_size(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels) * data_type_size; - const int kernel_matrix_stride = transform_weights_kernel->get_matrix_stride(out_channels, in_channels); - const int output_matrix_stride = transform_output_kernel->get_matrix_stride(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels); - const auto output_shape = transform_output_kernel->get_output_shape(in_shape.n_rows, in_shape.n_cols, use_padding_type == PADDING_SAME); - const int input_matrix_stride = transform_input_kernel->get_matrix_stride(in_shape.n_batches, in_channels, in_shape.n_rows, in_shape.n_cols, use_padding_type == PADDING_SAME); - - // Configure GEMM - const int tile_rows = iceildiv(output_shape.first, output_tile.height); - const int tile_cols = iceildiv(output_shape.second, output_tile.width); - const int m = in_shape.n_batches * tile_rows * tile_cols; - const int k = in_shape.n_channels; - const int n = out_channels; - const int kernel_matrix_row_stride = roundup(out_channels, N_BLOCK); - const int output_matrix_row_stride = kernel_matrix_row_stride; - - TensorShape a_shape(k, m, 1, n_gemms); - Strides a_strides(data_type_size); - a_strides.set(1, a_strides[0] * k); - //a_strides.set(2, data_type_size * input_matrix_stride / n_gemms); FIXME: This is the real batch size, but RSH's code crashes if it's not 0. - a_strides.set(2, 0); - a_strides.set(3, data_type_size * input_matrix_stride); - - TensorShape b_shape(n, k, n_gemms); - Strides b_strides(data_type_size); - b_strides.set(1, data_type_size * kernel_matrix_row_stride); - b_strides.set(2, data_type_size * kernel_matrix_stride); - - TensorShape d_shape(n, m, 1, n_gemms); - Strides d_strides(data_type_size); - d_strides.set(1, data_type_size * output_matrix_row_stride); - //d_strides.set(2, data_type_size * output_matrix_stride / n_gemms); FIXME: This is the real batch size, but RSH's code crashes if it's not 0. - d_strides.set(2, 0); - d_strides.set(3, data_type_size * output_matrix_stride); - - TensorInfo a_info{}; - TensorInfo b_info{}; - TensorInfo d_info{}; - a_info.init(a_shape, 1, data_type, a_strides, 0, input_storage_size); - b_info.init(b_shape, 1, data_type, b_strides, 0, kernel_storage_size); - d_info.init(d_shape, 1, data_type, d_strides, 0, output_storage_size); - - _input_transformed.allocator()->init(a_info, storage_alignment); - _kernel_storage.allocator()->init(b_info, storage_alignment); - _output_transformed.allocator()->init(d_info, storage_alignment); - - // configure and allocate dst tensor to be used to convert from winograd domain to spatial domain when calling to reshape_output() - TensorInfo info(TensorShape(_output->info()->dimension(2), _output->info()->dimension(0), - _output->info()->dimension(1), _output->info()->dimension(3)), - 1, _output->info()->data_type()); - _output_nhwc.allocator()->init(info); - - const ITensor *input_to_use = _input; - ITensor *output_to_use = _output; - PermutationVector weights_permutation_vector(3U, 0U, 1U, 2U); - const unsigned int max_num_threads = NEScheduler::get().num_threads(); - - // Configure the kernel to transform the input tensor from NCHW -> NHWC - if(_data_layout == DataLayout::NCHW) - { - _memory_group.manage(&_input_nhwc); - _permute_input.configure(input, &_input_nhwc, PermutationVector(2U, 0U, 1U)); - input_to_use = &_input_nhwc; - weights_permutation_vector = PermutationVector(3U, 2U, 0U, 1U); - } - - // Configure input transform kernel - _memory_group.manage(&_input_transformed); - _memory_group.manage(&_input_workspace); - transform_input_kernel->configure(input_to_use, in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, in_shape.n_channels, use_padding_type, - &_input_transformed, input_matrix_stride, &_input_workspace); - const size_t input_workspace_size = transform_input_kernel->get_working_space_size(max_num_threads); - TensorInfo input_workspace_info(TensorShape(input_workspace_size), 1, _input->info()->data_type()); - _input_workspace.allocator()->init(input_workspace_info); - _input_workspace.allocator()->allocate(); - if(_data_layout == DataLayout::NCHW) - { - _input_nhwc.allocator()->allocate(); - } - - // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map] - _permute_weights.configure(weights, &_weights_hwio, weights_permutation_vector); - transform_weights_kernel->configure(&_weights_hwio, &_kernel_storage, kernel_matrix_stride, out_channels, in_channels); - - // Configure GEMM function - _memory_group.manage(&_output_transformed); - _gemm_function.configure(&_input_transformed, &_kernel_storage, nullptr, &_output_transformed, 1.0f, 0.f); - _input_transformed.allocator()->allocate(); - - // Configure output transform function - // The biases tensor has not been allocated at this point in time, the output transform will add the biases to the final result in the run() method - if(_data_layout == DataLayout::NCHW) - { - _memory_group.manage(&_output_nhwc); - output_to_use = &_output_nhwc; - } - const arm_gemm::Activation activation = arm_gemm_activation_from_acl_activation(act_info); - - transform_output_kernel->configure(biases, - &_output_transformed, - output_matrix_stride, - output_to_use, - in_shape.n_batches, - output_shape.first, - output_shape.second, - out_channels, - &_output_workspace, - activation); - - const size_t output_workspace_size = transform_output_kernel->get_working_space_size(max_num_threads); - TensorInfo output_workspace_info(TensorShape(output_workspace_size), 1, _output->info()->data_type()); - _output_workspace.allocator()->init(output_workspace_info); - _output_workspace.allocator()->allocate(); - _output_transformed.allocator()->allocate(); - - // Reorder the convoluted output to ACL's ordering NCHW - if(_data_layout == DataLayout::NCHW) - { - _permute_output.configure(&_output_nhwc, _output, PermutationVector(1U, 2U, 0U)); - _output_nhwc.allocator()->allocate(); - } - - _transform_input_kernel = std::move(transform_input_kernel); - _transform_weights_kernel = std::move(transform_weights_kernel); - _transform_output_kernel = std::move(transform_output_kernel); - - //Configure Activation Layer - _is_activationlayer_enabled = act_info.enabled() && !fuse_function_supported(act_info); - if(_is_activationlayer_enabled) - { - _activationlayer_function.configure(_output, nullptr, act_info); - } + _impl->aux_mem_req = _impl->op->workspace(); + _impl->run_pack = { { ACL_SRC_0, input }, { ACL_SRC_1, weights }, { ACL_SRC_2, biases }, { ACL_DST, output } }; + _impl->prep_pack = { { ACL_SRC_1, weights }, { ACL_SRC_2, biases } }; + _impl->workspace = manage_workspace<Tensor>(_impl->aux_mem_req, _impl->memory_group, _impl->run_pack, _impl->prep_pack); } void NEWinogradConvolutionLayer::run() { prepare(); - MemoryGroupResourceScope scope_mg(_memory_group); - - if(_data_layout == DataLayout::NCHW) - { - //Bring channels to the front as Winograd code expects the tensor to be in the format NHWC - _permute_input.run(); - } - - // Transform input tensor to the winograd domain - NEScheduler::get().schedule(_transform_input_kernel.get(), Window::DimX); - - //Run 16 GEMMs in multiple threads, each kernel runs one or more GEMMs - _gemm_function.run(); - - // Transform output tensor to the spatial domain - NEScheduler::get().schedule(_transform_output_kernel.get(), Window::DimX); - - if(_data_layout == DataLayout::NCHW) - { - // Reorder the convoluted output to ACL's ordering NCHW - _permute_output.run(); - } - - if(_is_activationlayer_enabled) - { - _activationlayer_function.run(); - } + MemoryGroupResourceScope scope_mg(_impl->memory_group); + _impl->op->run(_impl->run_pack); } Status NEWinogradConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math) { - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info)); - - // Get indices for the width and 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); - - // Input shape, kernel size and output tile - const Size2D input_dims = Size2D(input->dimension(idx_width), input->dimension(idx_height)); - const Size2D kernel_size = Size2D(weights->dimension(idx_width), weights->dimension(idx_height)); - const DataType data_type = input->data_type(); - const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, data_type); - - // Check if the Winograd configuration requires fast math - if(!enable_fast_math) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size, data_type), - "This Winograd configuration requires enable_fast_math=true"); - } - - const WinogradInfo winograd_info = WinogradInfo(output_tile, - kernel_size, - input_dims, - conv_info, - input->data_layout()); - - // Validate input transform - const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info); - const TensorInfo input0 = input->clone()->set_tensor_shape(input0_shape); - // Validate filter transform - const TensorShape input1_shape = misc::shape_calculator::compute_winograd_filter_transform_shape(*weights, winograd_info); - const TensorInfo input1 = weights->clone()->set_tensor_shape(input1_shape); - // Validate batched matrix multiply - TensorShape batched_mm_output_shape = input0.tensor_shape(); - batched_mm_output_shape[0] = input1.tensor_shape()[0]; - const TensorInfo batched_mm_output = input0.clone()->set_tensor_shape(batched_mm_output_shape); - - if(kernel_size == Size2D(3, 3)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != conv_info.pad_left(), "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_bottom(), "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_left(), "Only SAME or VALID padding supported"); - return validate_kernel_3x3(input_dims, input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); - } - else if(kernel_size == Size2D(5, 5)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != conv_info.pad_left(), "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_bottom(), "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_left(), "Only SAME or VALID padding supported"); - return validate_kernel_5x5(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); - } - if(kernel_size == Size2D(3, 1)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_bottom() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_3x1(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); - } - else if(kernel_size == Size2D(1, 3)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_right() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_1x3(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); - } - else if(kernel_size == Size2D(5, 1)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_bottom() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_5x1(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); - } - else if(kernel_size == Size2D(1, 5)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_right() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_1x5(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); - } - else if(kernel_size == Size2D(7, 1)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 3, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 3, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_bottom() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_7x1(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); - } - else if(kernel_size == Size2D(1, 7)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 3, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 3, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_right() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_1x7(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); - } - else - { - ARM_COMPUTE_RETURN_ERROR_MSG("Kernel shape not supported"); - } + return cpu::CpuWinogradConv2d::validate(input, weights, biases, output, conv_info, act_info, enable_fast_math); } void NEWinogradConvolutionLayer::prepare() { - if(!_is_prepared) + if(!_impl->is_prepared) { - // Permute weights - _weights_hwio.allocator()->allocate(); - _permute_weights.run(); - _weights->mark_as_unused(); + _impl->op->prepare(_impl->prep_pack); + _impl->original_weights->mark_as_unused(); - // Transform weights - _kernel_storage.allocator()->allocate(); - NEScheduler::get().schedule(_transform_weights_kernel.get(), Window::DimX); - _weights_hwio.allocator()->free(); - - _gemm_function.prepare(); - if(!_kernel_storage.is_used()) + // Release temporary tensors that are only used in prepare stage + for(auto &ws : _impl->workspace) { - _kernel_storage.allocator()->free(); + const int slot = ws.first; + for(auto &m : _impl->aux_mem_req) + { + if(m.slot == slot && m.lifetime == MemoryLifetime::Prepare) + { + auto tensor = ws.second.get(); + tensor->allocator()->free(); + break; + } + } } - _is_prepared = true; + _impl->is_prepared = true; } } } // namespace arm_compute diff --git a/src/runtime/cpu/operators/CpuWinogradConv2d.cpp b/src/runtime/cpu/operators/CpuWinogradConv2d.cpp new file mode 100644 index 0000000000..bf105d5880 --- /dev/null +++ b/src/runtime/cpu/operators/CpuWinogradConv2d.cpp @@ -0,0 +1,848 @@ +/* + * Copyright (c) 2021 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 "src/runtime/cpu/operators/CpuWinogradConv2d.h" +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/core/utils/quantization/AsymmHelpers.h" +#include "arm_compute/runtime/FunctionDescriptors.h" +#include "arm_compute/runtime/NEON/NEScheduler.h" +#include "src/core/CPP/Validate.h" +#include "src/core/NEON/kernels/convolution/common/utils.hpp" +#include "src/core/NEON/kernels/convolution/winograd/winograd.hpp" +#include "src/core/cpu/kernels/CpuWinogradConv2dKernel.h" +#include "src/core/helpers/MemoryHelpers.h" +#include "src/runtime/cpu/operators/CpuActivation.h" +#include "src/runtime/cpu/operators/CpuPermute.h" +#include "src/runtime/cpu/operators/CpuWinogradConv2d.h" +#include "src/runtime/cpu/utils/CpuAuxTensorHandler.h" + +#include "support/Cast.h" + +#include <set> + +namespace arm_compute +{ +namespace cpu +{ +using namespace arm_compute::experimental; +using namespace arm_compute::utils::cast; + +namespace +{ +arm_gemm::Activation arm_gemm_activation_from_acl_activation(const ActivationLayerInfo &act_info) +{ + switch(act_info.activation()) + { + case ActivationLayerInfo::ActivationFunction::RELU: + { + return arm_gemm::Activation(arm_gemm::Activation::Type::ReLU, act_info.a(), act_info.b()); + } + case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: + { + return arm_gemm::Activation(arm_gemm::Activation::Type::BoundedReLU, act_info.a(), act_info.b()); + } + default: + { + return arm_gemm::Activation(arm_gemm::Activation::Type::None); + } + } +} + +inline Status validate_kernel_3x3(const Size2D input_dims, const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, + const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); + + if(input->data_type() == DataType::F32) + { + if(input_dims.width > 4 && input_dims.height > 4) + { + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 4, 4, 3, 3>::validate(input, input0, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 4, 4, 3, 3>::validate(weights, input1, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 4, 4, 3, 3>::validate(batched_mm_output, biases, output, winograd_info))); + } + else + { + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 2, 2, 3, 3>::validate(input, input0, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 2, 2, 3, 3>::validate(weights, input1, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 2, 2, 3, 3>::validate(batched_mm_output, biases, output, winograd_info))); + } + } +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + else if(input->data_type() == DataType::F16) + { + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<__fp16, 4, 4, 3, 3>::validate(input, input0, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<__fp16, 4, 4, 3, 3>::validate(weights, input1, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<__fp16, 4, 4, 3, 3>::validate(batched_mm_output, biases, output, winograd_info))); + } +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + + if(act_info.enabled()) + { + CpuActivation::validate(output, nullptr, act_info); + } + return Status{}; +} + +inline Status validate_kernel_5x5(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, + const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) +{ + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 2, 2, 5, 5>::validate(input, input0, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 2, 2, 5, 5>::validate(weights, input1, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 2, 2, 5, 5>::validate(batched_mm_output, biases, output, winograd_info))); + if(act_info.enabled()) + { + CpuActivation::validate(output, nullptr, act_info); + } + return Status{}; +} + +inline Status validate_kernel_3x1(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, + const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 1, 6, 1, 3>::validate(input, input0, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 1, 6, 1, 3>::validate(weights, input1, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 1, 6, 1, 3>::validate(batched_mm_output, biases, output, winograd_info))); + if(act_info.enabled()) + { + CpuActivation::validate(output, nullptr, act_info); + } + return Status{}; +} + +inline Status validate_kernel_1x3(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, + const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 6, 1, 3, 1>::validate(input, input0, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 6, 1, 3, 1>::validate(weights, input1, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 6, 1, 3, 1>::validate(batched_mm_output, biases, output, winograd_info))); + + if(act_info.enabled()) + { + CpuActivation::validate(output, nullptr, act_info); + } + return Status{}; +} + +inline Status validate_kernel_5x1(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, + const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 1, 4, 1, 5>::validate(input, input0, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 1, 4, 1, 5>::validate(weights, input1, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 1, 4, 1, 5>::validate(batched_mm_output, biases, output, winograd_info))); + if(act_info.enabled()) + { + CpuActivation::validate(output, nullptr, act_info); + } + return Status{}; +} +inline Status validate_kernel_1x5(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, + const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 4, 1, 5, 1>::validate(input, input0, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 4, 1, 5, 1>::validate(weights, input1, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 4, 1, 5, 1>::validate(batched_mm_output, biases, output, winograd_info))); + if(act_info.enabled()) + { + CpuActivation::validate(output, nullptr, act_info); + } + return Status{}; +} + +inline Status validate_kernel_7x1(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, + const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 1, 2, 1, 7>::validate(input, input0, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 1, 2, 1, 7>::validate(weights, input1, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 1, 2, 1, 7>::validate(batched_mm_output, biases, output, winograd_info))); + if(act_info.enabled()) + { + CpuActivation::validate(output, nullptr, act_info); + } + return Status{}; +} + +inline Status validate_kernel_1x7(const ITensorInfo *input, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, + const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 2, 1, 7, 1>::validate(input, input0, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 2, 1, 7, 1>::validate(weights, input1, winograd_info))); + ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 2, 1, 7, 1>::validate(batched_mm_output, biases, output, winograd_info))); + + if(act_info.enabled()) + { + CpuActivation::validate(output, nullptr, act_info); + } + return Status{}; +} + +inline Tensor4DShape internal_get_input_shape(const ITensorInfo *input) +{ + const DataLayout data_layout = input->data_layout(); + const int in_width = input->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH)); + const int in_height = input->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT)); + const int in_channels = input->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL)); + const int in_batches = input->dimension(3); + + return Tensor4DShape{ in_batches, in_height, in_width, in_channels }; +} + +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info) +{ + ARM_COMPUTE_UNUSED(output); + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd layer only supports unit strides."); + if(biases != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + return ICpuWinogradConv2dTransformWeightsKernel::validate(input, weights); +} +Size2D winograd_output_tile(const Size2D &input_dims, const Size2D &kernel_dims, DataType data_type) +{ + Size2D output_tile = Size2D{}; + if(kernel_dims == Size2D(3U, 3U)) + { + output_tile = (input_dims.width <= 4 || input_dims.height <= 4) ? Size2D(2U, 2U) : Size2D(4U, 4U); + if(data_type == DataType::F16) + { + output_tile = Size2D(4U, 4U); + } + } + else if(kernel_dims == Size2D(5U, 5U)) + { + output_tile = Size2D(2U, 2U); + } + else if(kernel_dims == Size2D(1U, 3U)) + { + output_tile = Size2D(1U, 6U); + } + else if(kernel_dims == Size2D(3U, 1U)) + { + output_tile = Size2D(6U, 1U); + } + else if(kernel_dims == Size2D(1U, 5U)) + { + output_tile = Size2D(1U, 4U); + } + else if(kernel_dims == Size2D(5U, 1U)) + { + output_tile = Size2D(4U, 1U); + } + else if(kernel_dims == Size2D(7U, 1U)) + { + output_tile = Size2D(2U, 1U); + } + else if(kernel_dims == Size2D(1U, 7U)) + { + output_tile = Size2D(1U, 2U); + } + return output_tile; +} + +bool check_support_fast_math(const Size2D &output_tile, const Size2D &kernel_size, DataType data_type) +{ + // Check if we want to configure a Winograd configuration which requires fast math + using WinogradConfiguration = std::pair<std::pair<int, int>, std::pair<int, int>>; + + const std::vector<WinogradConfiguration> fast_math_winograd_f16 = + { + WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3)) + }; + + const std::vector<WinogradConfiguration> fast_math_winograd_f32 = + { + WinogradConfiguration(std::pair<int, int>(2, 2), std::pair<int, int>(5, 5)), + WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5)) + }; + + auto p = std::make_pair(std::pair<int, int>(output_tile.width, output_tile.height), + std::pair<int, int>(kernel_size.width, kernel_size.height)); + + switch(data_type) + { + case DataType::F16: + return std::find(fast_math_winograd_f16.begin(), fast_math_winograd_f16.end(), p) != fast_math_winograd_f16.end(); + case DataType::F32: + return std::find(fast_math_winograd_f32.begin(), fast_math_winograd_f32.end(), p) != fast_math_winograd_f32.end(); + default: + return false; + } +} + +inline bool fuse_function_supported(const ActivationLayerInfo &act_info) +{ + return act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU || act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU; +} + +} // namespace + +CpuWinogradConv2d::CpuWinogradConv2d() + : _gemm_function(std::make_unique<CpuGemm>()), + _activation_func(std::make_unique<CpuActivation>()), + _permute_input(std::make_unique<CpuPermute>()), + _permute_output(std::make_unique<CpuPermute>()), + _permute_weights(std::make_unique<CpuPermute>()), + _transform_input_kernel(nullptr), + _transform_weights_kernel(nullptr), + _transform_output_kernel(nullptr), + _data_layout(), + _aux_mem(AuxTensorIdx::Count), + _input_nhwc(), + _output_nhwc(), + _input_workspace(), + _kernel_storage(), + _output_workspace(), + _input_transformed(), + _output_transformed(), + _weights_hwio(), + _run_activation(false), + _is_prepared(false) +{ +} + +CpuWinogradConv2d::~CpuWinogradConv2d() = default; + +void CpuWinogradConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, + const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, weights, biases, dst, conv_info)); + + // Get indices for the width and height + _data_layout = src->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); + const unsigned int channel_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL); + + const Size2D input_dims = Size2D(src->dimension(width_idx), src->dimension(height_idx)); + const Size2D kernel_size = Size2D(weights->dimension(width_idx), weights->dimension(height_idx)); + const DataType data_type = src->data_type(); + const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, data_type); + + // Check if the Winograd configuration requires fast math + if(!enable_fast_math) + { + ARM_COMPUTE_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size, data_type), + "This Winograd configuration requires enable_fast_math=true"); + } + + _is_prepared = false; + + std::unique_ptr<ICpuWinogradConv2dTransformInputKernel> transform_input_kernel; + std::unique_ptr<ICpuWinogradConv2dTransformWeightsKernel> transform_weights_kernel; + std::unique_ptr<ICpuWinogradConv2dTransformOutputKernel> transform_output_kernel; + + int n_gemms = 1; + int N_BLOCK = 1; // Size of block used by GEMM. + if(data_type == DataType::F32) + { + if(kernel_size == Size2D(3, 3)) + { + if(src->dimension(width_idx) > 4 && src->dimension(height_idx) > 4) + { + using config = CpuWinogradConv2dConfiguration<float, float, 4, 4, 3, 3>; + transform_input_kernel = std::make_unique<config::TransformInputKernel>(); + transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); + transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); + n_gemms = config::WinogradBase::N_GEMMS; + N_BLOCK = config::WinogradConv::N_BLOCK; + } + else + { + using config = CpuWinogradConv2dConfiguration<float, float, 2, 2, 3, 3>; + transform_input_kernel = std::make_unique<config::TransformInputKernel>(); + transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); + transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); + n_gemms = config::WinogradBase::N_GEMMS; + N_BLOCK = config::WinogradConv::N_BLOCK; + } + } + else if(kernel_size == Size2D(5, 5)) + { + using config = CpuWinogradConv2dConfiguration<float, float, 2, 2, 5, 5>; + transform_input_kernel = std::make_unique<config::TransformInputKernel>(); + transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); + transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); + n_gemms = config::WinogradBase::N_GEMMS; + N_BLOCK = config::WinogradConv::N_BLOCK; + } + else if(kernel_size == Size2D(1, 3)) + { + using config = CpuWinogradConv2dConfiguration<float, float, 6, 1, 3, 1>; + transform_input_kernel = std::make_unique<config::TransformInputKernel>(); + transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); + transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); + n_gemms = config::WinogradBase::N_GEMMS; + N_BLOCK = config::WinogradConv::N_BLOCK; + } + else if(kernel_size == Size2D(3, 1)) + { + using config = CpuWinogradConv2dConfiguration<float, float, 1, 6, 1, 3>; + transform_input_kernel = std::make_unique<config::TransformInputKernel>(); + transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); + transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); + n_gemms = config::WinogradBase::N_GEMMS; + N_BLOCK = config::WinogradConv::N_BLOCK; + } + else if(kernel_size == Size2D(1, 5)) + { + using config = CpuWinogradConv2dConfiguration<float, float, 4, 1, 5, 1>; + transform_input_kernel = std::make_unique<config::TransformInputKernel>(); + transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); + transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); + n_gemms = config::WinogradBase::N_GEMMS; + N_BLOCK = config::WinogradConv::N_BLOCK; + } + else if(kernel_size == Size2D(5, 1)) + { + using config = CpuWinogradConv2dConfiguration<float, float, 1, 4, 1, 5>; + transform_input_kernel = std::make_unique<config::TransformInputKernel>(); + transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); + transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); + n_gemms = config::WinogradBase::N_GEMMS; + N_BLOCK = config::WinogradConv::N_BLOCK; + } + else if(kernel_size == Size2D(1, 7)) + { + using config = CpuWinogradConv2dConfiguration<float, float, 2, 1, 7, 1>; + transform_input_kernel = std::make_unique<config::TransformInputKernel>(); + transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); + transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); + n_gemms = config::WinogradBase::N_GEMMS; + N_BLOCK = config::WinogradConv::N_BLOCK; + } + else if(kernel_size == Size2D(7, 1)) + { + using config = CpuWinogradConv2dConfiguration<float, float, 1, 2, 1, 7>; + transform_input_kernel = std::make_unique<config::TransformInputKernel>(); + transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); + transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); + n_gemms = config::WinogradBase::N_GEMMS; + N_BLOCK = config::WinogradConv::N_BLOCK; + } + else + { + ARM_COMPUTE_ERROR("Not supported."); + } + } +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + else if(data_type == DataType::F16) + { + if(kernel_size == Size2D(3, 3)) + { + using config = CpuWinogradConv2dConfiguration<__fp16, __fp16, 4, 4, 3, 3>; + transform_input_kernel = std::make_unique<config::TransformInputKernel>(); + transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); + transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); + n_gemms = config::WinogradBase::N_GEMMS; + N_BLOCK = config::WinogradConv::N_BLOCK; + } + else + { + ARM_COMPUTE_ERROR("Not supported."); + } + } +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + else + { + ARM_COMPUTE_ERROR("Not supported."); + } + + const PaddingType use_padding_type = (conv_info.pad_top() != 0u || conv_info.pad_left() != 0) ? PADDING_SAME : PADDING_VALID; + const bool use_same_padding = use_padding_type == PADDING_SAME; + + // Get convolved dimensions + const int in_channels = src->dimension(channel_idx); + const int out_channels = dst->dimension(channel_idx); + + const Tensor4DShape in_shape(internal_get_input_shape(src)); + const size_t data_type_size = src->element_size(); + // Get the memory required to instantiate a new Winograd operator. + constexpr size_t storage_alignment = 64; + + // Kernel Storage + const size_t kernel_storage_size = transform_weights_kernel->get_weight_storage_size(out_channels, + in_channels) + * data_type_size; + + // Input storage + const size_t input_storage_size = transform_input_kernel->get_input_storage_size(in_shape.n_batches, in_shape.n_channels, in_shape.n_rows, in_shape.n_cols, + use_same_padding) + * data_type_size; + + // Output storage + const size_t output_storage_size = transform_output_kernel->get_output_storage_size(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels) * data_type_size; + const int kernel_matrix_stride = transform_weights_kernel->get_matrix_stride(out_channels, in_channels); + const int output_matrix_stride = transform_output_kernel->get_matrix_stride(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels); + const auto output_shape = transform_output_kernel->get_output_shape(in_shape.n_rows, in_shape.n_cols, use_padding_type == PADDING_SAME); + const int input_matrix_stride = transform_input_kernel->get_matrix_stride(in_shape.n_batches, in_channels, in_shape.n_rows, in_shape.n_cols, use_padding_type == PADDING_SAME); + + // Configure GEMM + const int tile_rows = iceildiv(output_shape.first, output_tile.height); + const int tile_cols = iceildiv(output_shape.second, output_tile.width); + const int m = in_shape.n_batches * tile_rows * tile_cols; + const int k = in_shape.n_channels; + const int n = out_channels; + const int kernel_matrix_row_stride = roundup(out_channels, N_BLOCK); + const int output_matrix_row_stride = kernel_matrix_row_stride; + + TensorShape a_shape(k, m, 1, n_gemms); + Strides a_strides(data_type_size); + a_strides.set(1, a_strides[0] * k); + //a_strides.set(2, data_type_size * input_matrix_stride / n_gemms); FIXME: This is the real batch size, but RSH's code crashes if it's not 0. + a_strides.set(2, 0); + a_strides.set(3, data_type_size * input_matrix_stride); + + TensorShape b_shape(n, k, n_gemms); + Strides b_strides(data_type_size); + b_strides.set(1, data_type_size * kernel_matrix_row_stride); + b_strides.set(2, data_type_size * kernel_matrix_stride); + + TensorShape d_shape(n, m, 1, n_gemms); + Strides d_strides(data_type_size); + d_strides.set(1, data_type_size * output_matrix_row_stride); + //d_strides.set(2, data_type_size * output_matrix_stride / n_gemms); FIXME: This is the real batch size, but RSH's code crashes if it's not 0. + d_strides.set(2, 0); + d_strides.set(3, data_type_size * output_matrix_stride); + + TensorInfo a_info{}; + TensorInfo b_info{}; + TensorInfo d_info{}; + a_info.init(a_shape, 1, data_type, a_strides, 0, input_storage_size); + b_info.init(b_shape, 1, data_type, b_strides, 0, kernel_storage_size); + d_info.init(d_shape, 1, data_type, d_strides, 0, output_storage_size); + + _input_transformed = a_info; + _kernel_storage = b_info; + _output_transformed = d_info; + + // configure and allocate dst tensor to be used to convert from winograd domain to spatial domain when calling to reshape_output() + TensorInfo info(TensorShape(dst->dimension(2), dst->dimension(0), + dst->dimension(1), dst->dimension(3)), + 1, dst->data_type()); + _output_nhwc = info; + + const ITensorInfo *input_to_use = src; + ITensorInfo *output_to_use = dst; + PermutationVector weights_permutation_vector(3U, 0U, 1U, 2U); + const unsigned int max_num_threads = NEScheduler::get().num_threads(); + + // Configure the kernel to transform the input tensor from NCHW -> NHWC + if(_data_layout == DataLayout::NCHW) + { + _permute_input->configure(src, &_input_nhwc, PermutationVector(2U, 0U, 1U)); + _aux_mem[PermutedInput] = MemoryInfo(offset_int_vec(PermutedInput), MemoryLifetime::Temporary, src->total_size()); + input_to_use = &_input_nhwc; + weights_permutation_vector = PermutationVector(3U, 2U, 0U, 1U); + } + + // Configure input transform kernel + transform_input_kernel->configure(input_to_use, in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, in_shape.n_channels, use_padding_type, + &_input_transformed, input_matrix_stride, &_input_workspace); + const size_t input_workspace_size = transform_input_kernel->get_working_space_size(max_num_threads); + TensorInfo input_workspace_info(TensorShape(input_workspace_size), 1, src->data_type()); + _input_workspace = input_workspace_info; + + // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map] + _permute_weights->configure(weights, &_weights_hwio, weights_permutation_vector); + transform_weights_kernel->configure(&_weights_hwio, &_kernel_storage, kernel_matrix_stride, out_channels, in_channels); + + // Configure GEMM function + _gemm_function->configure(&_input_transformed, &_kernel_storage, nullptr, &_output_transformed, 1.0f, 0.f); + + // Configure output transform function + // The biases tensor has not been allocated at this point in time, the output transform will add the biases to the final result in the run() method + if(_data_layout == DataLayout::NCHW) + { + output_to_use = &_output_nhwc; + } + const arm_gemm::Activation activation = arm_gemm_activation_from_acl_activation(act_info); + + transform_output_kernel->configure(biases, + &_output_transformed, + output_matrix_stride, + output_to_use, + in_shape.n_batches, + output_shape.first, + output_shape.second, + out_channels, + &_output_workspace, + activation); + + const size_t output_workspace_size = transform_output_kernel->get_working_space_size(max_num_threads); + TensorInfo output_workspace_info(TensorShape(output_workspace_size), 1, dst->data_type()); + _output_workspace = output_workspace_info; + + // Reorder the convoluted output to ACL's ordering NCHW + if(_data_layout == DataLayout::NCHW) + { + _permute_output->configure(&_output_nhwc, dst, PermutationVector(1U, 2U, 0U)); + _aux_mem[PermutedOutput] = MemoryInfo(offset_int_vec(PermutedOutput), MemoryLifetime::Temporary, dst->total_size()); + } + + _transform_input_kernel = std::move(transform_input_kernel); + _transform_weights_kernel = std::move(transform_weights_kernel); + _transform_output_kernel = std::move(transform_output_kernel); + + //Configure Activation Layer + _run_activation = act_info.enabled() && !fuse_function_supported(act_info); + if(_run_activation) + { + _activation_func->configure(dst, nullptr, act_info); + } + + auto asm_mem_req = _gemm_function->workspace(); + _aux_mem[GemmWorkspace] = asm_mem_req[GemmWorkspace]; + _aux_mem[Pretranspose] = asm_mem_req[Pretranspose]; + _aux_mem[InterleavedLHS] = asm_mem_req[InterleavedLHS]; + _aux_mem[TransposedRHS] = asm_mem_req[TransposedRHS]; + _aux_mem[TempResult] = asm_mem_req[TempResult]; + + _aux_mem[InputTransformed] = MemoryInfo(offset_int_vec(InputTransformed), MemoryLifetime::Persistent, input_storage_size, storage_alignment); + _aux_mem[InputWorkspace] = MemoryInfo(offset_int_vec(InputWorkspace), MemoryLifetime::Persistent, input_workspace_size); + if(_aux_mem[Pretranspose].size > 0) + { + // Release permuted weights at the of prepare as they are further transposed by the assembly dispatch + _aux_mem[PermutedWeights] = MemoryInfo(offset_int_vec(PermutedWeights), MemoryLifetime::Prepare, _weights_hwio.total_size()); + } + else + { + _aux_mem[PermutedWeights] = MemoryInfo(offset_int_vec(PermutedWeights), MemoryLifetime::Persistent, _weights_hwio.total_size()); + } + _aux_mem[WeightsTransformed] = MemoryInfo(offset_int_vec(WeightsTransformed), MemoryLifetime::Persistent, kernel_storage_size, storage_alignment); + _aux_mem[OutputTransformed] = MemoryInfo(offset_int_vec(OutputTransformed), MemoryLifetime::Persistent, output_storage_size, storage_alignment); + _aux_mem[OutputWorkspace] = MemoryInfo(offset_int_vec(OutputWorkspace), MemoryLifetime::Persistent, output_workspace_size); +} + +Status CpuWinogradConv2d::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, + const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info)); + + // Get indices for the width and 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); + + // Input shape, kernel size and output tile + const Size2D input_dims = Size2D(input->dimension(idx_width), input->dimension(idx_height)); + const Size2D kernel_size = Size2D(weights->dimension(idx_width), weights->dimension(idx_height)); + const DataType data_type = input->data_type(); + const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, data_type); + + // Check if the Winograd configuration requires fast math + if(!enable_fast_math) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size, data_type), + "This Winograd configuration requires enable_fast_math=true"); + } + + const WinogradInfo winograd_info = WinogradInfo(output_tile, + kernel_size, + input_dims, + conv_info, + input->data_layout()); + + // Validate input transform + const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info); + const TensorInfo input0 = input->clone()->set_tensor_shape(input0_shape); + // Validate filter transform + const TensorShape input1_shape = misc::shape_calculator::compute_winograd_filter_transform_shape(*weights, winograd_info); + const TensorInfo input1 = weights->clone()->set_tensor_shape(input1_shape); + // Validate batched matrix multiply + TensorShape batched_mm_output_shape = input0.tensor_shape(); + batched_mm_output_shape[0] = input1.tensor_shape()[0]; + const TensorInfo batched_mm_output = input0.clone()->set_tensor_shape(batched_mm_output_shape); + + if(kernel_size == Size2D(3, 3)) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 1, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 1, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 1, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 1, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != conv_info.pad_left(), "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_bottom(), "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_left(), "Only SAME or VALID padding supported"); + return validate_kernel_3x3(input_dims, input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); + } + else if(kernel_size == Size2D(5, 5)) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 2, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 2, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 2, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 2, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != conv_info.pad_left(), "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_bottom(), "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_left(), "Only SAME or VALID padding supported"); + return validate_kernel_5x5(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); + } + if(kernel_size == Size2D(3, 1)) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 1, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 1, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_bottom() != 0, "Only SAME or VALID padding supported"); + return validate_kernel_3x1(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); + } + else if(kernel_size == Size2D(1, 3)) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 1, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 1, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_right() != 0, "Only SAME or VALID padding supported"); + return validate_kernel_1x3(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); + } + else if(kernel_size == Size2D(5, 1)) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 2, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 2, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_bottom() != 0, "Only SAME or VALID padding supported"); + return validate_kernel_5x1(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); + } + else if(kernel_size == Size2D(1, 5)) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 2, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 2, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_right() != 0, "Only SAME or VALID padding supported"); + return validate_kernel_1x5(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); + } + else if(kernel_size == Size2D(7, 1)) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 3, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 3, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_bottom() != 0, "Only SAME or VALID padding supported"); + return validate_kernel_7x1(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); + } + else if(kernel_size == Size2D(1, 7)) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 3, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 3, "Only SAME or VALID padding supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_right() != 0, "Only SAME or VALID padding supported"); + return validate_kernel_1x7(input, &input0, &input1, &batched_mm_output, weights, biases, output, winograd_info, act_info); + } + else + { + ARM_COMPUTE_RETURN_ERROR_MSG("Kernel shape not supported"); + } +} + +void CpuWinogradConv2d::run(ITensorPack &tensors) +{ + prepare(tensors); + + auto a = tensors.get_const_tensor(ACL_SRC_0); + auto c = tensors.get_const_tensor(ACL_SRC_2); + auto d = tensors.get_tensor(ACL_DST); + + CpuAuxTensorHandler input_nhwc(offset_int_vec(PermutedInput), _input_nhwc, tensors, true); + CpuAuxTensorHandler output_nhwc(offset_int_vec(PermutedOutput), _output_nhwc, tensors, true); + CpuAuxTensorHandler input_transformed(offset_int_vec(InputTransformed), _input_transformed, tensors, true); + CpuAuxTensorHandler input_workspace(offset_int_vec(InputWorkspace), _input_workspace, tensors, true); + + const bool is_nchw = _data_layout == DataLayout::NCHW; + if(is_nchw) + { + //Bring channels to the front as Winograd code expects the tensor to be in the format NHWC + ITensorPack pack{ { ACL_SRC, a }, { ACL_DST, input_nhwc.get() } }; + _permute_input->run(pack); + } + + // Transform input tensor to the winograd domain + ITensorPack transform_input_pack{ { ACL_SRC, is_nchw ? input_nhwc.get() : a }, { ACL_DST, input_transformed.get() }, { ACL_INT, input_workspace.get() } }; + NEScheduler::get().schedule_op(_transform_input_kernel.get(), Window::DimX, _transform_input_kernel->window(), transform_input_pack); + + CpuAuxTensorHandler output_transformed(offset_int_vec(OutputTransformed), _output_transformed, tensors, true); + CpuAuxTensorHandler weights_transformed(offset_int_vec(WeightsTransformed), _kernel_storage, tensors, true); + + // Run 16 GEMMs in multiple threads, each kernel runs one or more GEMMs + ITensorPack gemm_pack{ { ACL_SRC, input_transformed.get() }, { ACL_SRC_1, weights_transformed.get() }, { ACL_DST, output_transformed.get() } }; + _gemm_function->run(gemm_pack); + + // Transform output tensor to the spatial domain + CpuAuxTensorHandler output_workspace(offset_int_vec(OutputWorkspace), _output_workspace, tensors, true); + ITensorPack transform_output_pack{ { ACL_SRC_0, c }, { ACL_SRC_1, output_transformed.get() }, { ACL_DST, is_nchw ? output_nhwc.get() : d }, { ACL_INT, output_workspace.get() } }; + NEScheduler::get().schedule_op(_transform_output_kernel.get(), Window::DimX, _transform_output_kernel->window(), transform_output_pack); + + if(is_nchw) + { + // Reorder the convoluted output to ACL's ordering NCHW + ITensorPack pack{ { ACL_SRC, output_nhwc.get() }, { ACL_DST, d } }; + _permute_output->run(pack); + } + + if(_run_activation) + { + ITensorPack pack{ { ACL_SRC, d }, { ACL_DST, d } }; + _activation_func->run(pack); + } +} + +void CpuWinogradConv2d::prepare(ITensorPack &tensors) +{ + if(!_is_prepared) + { + // Permute weights + const ITensor *weights = tensors.get_const_tensor(ACL_SRC_1); + ITensor *weights_aux = utils::cast::polymorphic_cast<ITensor *>(tensors.get_tensor(offset_int_vec(PermutedWeights))); + ARM_COMPUTE_ERROR_ON_NULLPTR(weights, weights_aux); + + CpuAuxTensorHandler permuted_weights(_weights_hwio, *weights_aux); + ITensorPack permute_tensors{ { ACL_SRC, weights }, { ACL_DST, permuted_weights.get() } }; + _permute_weights->run(permute_tensors); + + // Transform weights + ITensor *weights_transf = utils::cast::polymorphic_cast<ITensor *>(tensors.get_tensor(offset_int_vec(WeightsTransformed))); + ARM_COMPUTE_ERROR_ON_NULLPTR(weights_transf); + + CpuAuxTensorHandler transformed_weights(_kernel_storage, *weights_transf); + ITensorPack transform_tensors{ { ACL_SRC, permuted_weights.get() }, { ACL_DST, transformed_weights.get() } }; + NEScheduler::get().schedule_op(_transform_weights_kernel.get(), Window::DimX, _transform_weights_kernel->window(), transform_tensors); + + CpuAuxTensorHandler input_transformed(offset_int_vec(InputTransformed), _input_transformed, tensors, true); + CpuAuxTensorHandler output_transformed(offset_int_vec(OutputTransformed), _output_transformed, tensors, true); + ITensorPack gemm_pack = tensors; + gemm_pack.add_const_tensor(ACL_SRC_0, input_transformed.get()); + gemm_pack.add_const_tensor(ACL_SRC_1, transformed_weights.get()); + _gemm_function->prepare(gemm_pack); + + _is_prepared = true; + } +} + +experimental::MemoryRequirements CpuWinogradConv2d::workspace() const +{ + return _aux_mem; +} +} // namespace cpu +} // namespace arm_compute
\ No newline at end of file diff --git a/src/runtime/cpu/operators/CpuWinogradConv2d.h b/src/runtime/cpu/operators/CpuWinogradConv2d.h new file mode 100644 index 0000000000..14c61f7355 --- /dev/null +++ b/src/runtime/cpu/operators/CpuWinogradConv2d.h @@ -0,0 +1,137 @@ +/* + * Copyright (c) 2021 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. + */ +#ifndef ARM_COMPUTE_CPU_WINOGRAD_CONV2D_KERNEL_H +#define ARM_COMPUTE_CPU_WINOGRAD_CONV2D_KERNEL_H + +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/runtime/FunctionDescriptors.h" +#include "src/core/common/Macros.h" +#include "src/core/cpu/kernels/CpuWinogradConv2dKernel.h" +#include "src/runtime/cpu/ICpuOperator.h" +#include "src/runtime/cpu/operators/CpuActivation.h" +#include "src/runtime/cpu/operators/CpuGemm.h" +#include "src/runtime/cpu/operators/CpuPermute.h" +#include "src/runtime/cpu/operators/internal/CpuGemmAssemblyDispatch.h" + +namespace arm_compute +{ +namespace cpu +{ +class CpuWinogradConv2d : public ICpuOperator +{ +public: + /** Constructor */ + CpuWinogradConv2d(); + ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuWinogradConv2d); + /** Destructor */ + ~CpuWinogradConv2d(); + + /** Set the input and output tensors. + * + * Valid data layouts: + * - NHWC + * - NCHW + * + * Valid data type configurations: + * |src0 |src1 |src2 |dst | + * |:--------------|:--------------|:------|:--------------| + * |F16 |F16 |F16 |F16 | + * |F32 |F32 |F32 |F32 | + * + * @param[in] src Source tensor info. 3 lower dimensions represent a single input [width, height, IFM], + * while every optional dimension from 4 and above represent a batch of inputs. + * Data types supported: F16/F32. + * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input. + * Currently only 3x3 and 5x5 kernels are supported. + * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights. + * @param[out] dst Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. + * Data types supported: Same as @p input. + * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. + * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation + * available which may introduce a drop of accuracy as well. Default is false + */ + void configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info, + const ActivationLayerInfo &act_info = ActivationLayerInfo(), + bool enable_fast_math = false); + /** Static function to check if given info will lead to a valid configuration of @ref CpuWinogradConv2d + * + * Similar to CpuWinogradConv2d::configure() + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + const ActivationLayerInfo &act_info = ActivationLayerInfo(), + bool enable_fast_math = false); + + // Inherited methods overridden: + void run(ITensorPack &tensors) override; + void prepare(ITensorPack &constants) override; + experimental::MemoryRequirements workspace() const override; + +private: + enum AuxTensorIdx + { + GemmWorkspace = 0, + Pretranspose, + InterleavedLHS, + TransposedRHS, + TempResult, + PermutedInput, + InputTransformed, + InputWorkspace, + PermutedOutput, + PermutedWeights, + WeightsTransformed, + OutputTransformed, + OutputWorkspace, + Count + }; + + std::unique_ptr<CpuGemm> _gemm_function; + std::unique_ptr<CpuActivation> _activation_func; + std::unique_ptr<CpuPermute> _permute_input; + std::unique_ptr<CpuPermute> _permute_output; + std::unique_ptr<CpuPermute> _permute_weights; + std::unique_ptr<ICPPKernel> _transform_input_kernel; + std::unique_ptr<ICPPKernel> _transform_weights_kernel; + std::unique_ptr<ICPPKernel> _transform_output_kernel; + + DataLayout _data_layout; + experimental::MemoryRequirements _aux_mem{ Count }; + TensorInfo _input_nhwc; + TensorInfo _output_nhwc; + TensorInfo _input_workspace; + TensorInfo _kernel_storage; + TensorInfo _output_workspace; + TensorInfo _input_transformed; + TensorInfo _output_transformed; + TensorInfo _weights_hwio; + bool _run_activation; + bool _is_prepared; +}; +} // namespace cpu +} // namespace arm_compute + +#endif /* ARM_COMPUTE_CPU_WINOGRAD_CONV2D_KERNEL_H */ |