diff options
Diffstat (limited to 'src/runtime/NEON/functions/NEWinogradLayer.cpp')
| -rw-r--r-- | src/runtime/NEON/functions/NEWinogradLayer.cpp | 110 |
1 files changed, 75 insertions, 35 deletions
diff --git a/src/runtime/NEON/functions/NEWinogradLayer.cpp b/src/runtime/NEON/functions/NEWinogradLayer.cpp index 215f1bfddf..e343583b36 100644 --- a/src/runtime/NEON/functions/NEWinogradLayer.cpp +++ b/src/runtime/NEON/functions/NEWinogradLayer.cpp @@ -28,6 +28,8 @@ #include "arm_compute/runtime/NEON/NEScheduler.h" #include "support/ToolchainSupport.h" +#include "arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h" + #include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" namespace @@ -45,8 +47,9 @@ inline Tensor4DShape internal_get_input_shape(const arm_compute::ITensor *input) namespace arm_compute { NEWinogradLayer::NEWinogradLayer(std::shared_ptr<IMemoryManager> memory_manager) - : _memory_group(std::move(memory_manager)), _winograd_kernel(), _transform_input_kernel(), _transform_output_kernel(), _transform_weights_kernel(), _permute_input(), _permute_weights(), - _permute_output(), _input_workspace(), _output_workspace(), _kernel_storage(), _input_nhwc(), _output_nhwc(), _weights_hwio(), _input(), _weights(), _output(), _reshaped_kernel(false) + : _memory_group(std::move(memory_manager)), _batched_gemm_kernel(nullptr), _transform_input_kernel(nullptr), _transform_output_kernel(nullptr), _transform_weights_kernel(nullptr), _permute_input(), + _permute_weights(), _permute_output(), _input_workspace(), _output_workspace(), _kernel_storage(), _input_nhwc(), _output_nhwc(), _weights_hwio(), _input(), _weights(), _output(), + _reshaped_kernel(false) { } /* arm_compute */ @@ -54,7 +57,7 @@ void NEWinogradLayer::configure(const ITensor *input, const ITensor *weights, co { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, biases); - ARM_COMPUTE_ERROR_ON_MSG(weights->info()->dimension(1) != 3 || weights->info()->dimension(0) != 3, "Only 3x3 kernels are supported"); + ARM_COMPUTE_ERROR_ON_MSG(weights->info()->dimension(0) != 3 && weights->info()->dimension(0) != 5, "Only 3 and 5 kernels are supported"); ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); if(biases != nullptr) @@ -67,6 +70,36 @@ void NEWinogradLayer::configure(const ITensor *input, const ITensor *weights, co _input = input; _output = output; + std::unique_ptr<INEWinogradLayerBatchedGEMMKernel<float, float>> batched_gemm_kernel; + std::unique_ptr<INEWinogradLayerTransformInputKernel<float>> transform_input_kernel; + std::unique_ptr<INEWinogradLayerTransformWeightsKernel<float>> transform_weights_kernel; + std::unique_ptr<INEWinogradLayerTransformOutputKernel<float>> transform_output_kernel; + + switch(weights->info()->dimension(0)) + { + case 3: + { + batched_gemm_kernel = support::cpp14::make_unique<NEWinogradLayerBatchedGEMMKernel<float, float, 2, 2, 3, 3>>(); + transform_input_kernel = support::cpp14::make_unique<NEWinogradLayerTransformInputKernel<float, 2, 2, 3, 3>>(); + transform_weights_kernel = support::cpp14::make_unique<NEWinogradLayerTransformWeightsKernel<float, 2, 2, 3, 3>>(); + transform_output_kernel = support::cpp14::make_unique<NEWinogradLayerTransformOutputKernel<float, 2, 2, 3, 3>>(); + break; + } + case 5: + { + batched_gemm_kernel = support::cpp14::make_unique<NEWinogradLayerBatchedGEMMKernel<float, float, 2, 2, 5, 5>>(); + transform_input_kernel = support::cpp14::make_unique<NEWinogradLayerTransformInputKernel<float, 2, 2, 5, 5>>(); + transform_weights_kernel = support::cpp14::make_unique<NEWinogradLayerTransformWeightsKernel<float, 2, 2, 5, 5>>(); + transform_output_kernel = support::cpp14::make_unique<NEWinogradLayerTransformOutputKernel<float, 2, 2, 5, 5>>(); + break; + } + default: + { + ARM_COMPUTE_ERROR("Not supported."); + break; + } + } + const PaddingType use_padding_type = (conv_info.pad_left() != 0u) ? PADDING_SAME : PADDING_VALID; const bool use_same_padding = use_padding_type == PADDING_SAME; @@ -84,22 +117,19 @@ void NEWinogradLayer::configure(const ITensor *input, const ITensor *weights, co 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; - const size_t kernel_storage_size = NEWinogradLayerTransformWeightsKernel<2, 2, 3, 3>::get_weight_storage_size(out_channels, in_channels) * data_type_size; + const size_t kernel_storage_size = transform_weights_kernel->get_weight_storage_size(out_channels, in_channels) * data_type_size; _kernel_storage.allocator()->init(TensorInfo(TensorShape{ (kernel_storage_size + storage_alignment - 1) }, 1, DataType::U8)); _memory_group.manage(&_kernel_storage); _memory_group.manage(&_input_nhwc); _kernel_storage.allocator()->allocate(); // Input storage - - using IT = NEWinogradLayerTransformInputKernel<2, 2, 3, 3>; - const size_t input_storage_size = IT::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; + 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; _input_workspace.allocator()->init(TensorInfo(TensorShape{ (input_storage_size + storage_alignment - 1) }, 1, DataType::U8)); _memory_group.manage(&_input_workspace); _input_workspace.allocator()->allocate(); // Output storage - using OT = NEWinogradLayerTransformOutputKernel<2, 2, 3, 3>; - const size_t output_storage_size = OT::get_output_storage_size(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels, use_same_padding) * data_type_size; + 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, use_same_padding) * data_type_size; _output_workspace.allocator()->init(TensorInfo(TensorShape{ (output_storage_size + storage_alignment - 1) }, 1, DataType::U8)); _memory_group.manage(&_output_workspace); _output_workspace.allocator()->allocate(); @@ -119,47 +149,57 @@ void NEWinogradLayer::configure(const ITensor *input, const ITensor *weights, co _permute_input.configure(input, &_input_nhwc, PermutationVector(2U, 0U, 1U)); _input_nhwc.allocator()->allocate(); - using T = winograd::WinogradGEMM<2, 2, 3, 3>::Convolution<float, float>; const int weights_width = weights->info()->dimension(0); const int weights_height = weights->info()->dimension(1); const KernelShape kernel_shape({ out_channels, weights_height, weights_width, in_channels }); // Configure the InputTransform - const int input_matrix_stride = T::get_input_matrix_stride(kernel_shape, in_shape, use_padding_type); - _transform_input_kernel.configure(reinterpret_cast<float *>(_input_nhwc.buffer()), in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, in_shape.n_channels, use_padding_type, + const int input_matrix_stride = transform_input_kernel->get_matrix_stride(kernel_shape, in_shape, use_padding_type); + transform_input_kernel->configure(reinterpret_cast<float *>(_input_nhwc.buffer()), in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, in_shape.n_channels, use_padding_type, reinterpret_cast<float *>(_input_workspace.buffer()), input_matrix_stride); // Configure WeightsTransform - const int kernel_matrix_stride = T::get_kernel_matrix_stride(kernel_shape); - _transform_weights_kernel.configure(&_weights_hwio, reinterpret_cast<float *>(_kernel_storage.buffer()), kernel_matrix_stride, out_channels, in_channels); + const int kernel_matrix_stride = transform_weights_kernel->get_matrix_stride(kernel_shape); + transform_weights_kernel->configure(&_weights_hwio, reinterpret_cast<float *>(_kernel_storage.buffer()), kernel_matrix_stride, out_channels, in_channels); // Configure OutputTransform //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 - const int output_matrix_stride = T::get_output_matrix_stride(kernel_shape, in_shape, use_padding_type); - const auto output_shape(T::get_output_shape(kernel_shape, in_shape, use_padding_type)); + const int output_matrix_stride = transform_output_kernel->get_matrix_stride(kernel_shape, in_shape, use_padding_type); + const auto output_shape(transform_output_kernel->get_output_shape(kernel_shape, in_shape, use_padding_type)); - _transform_output_kernel.configure(biases, reinterpret_cast<float *>(_output_workspace.buffer()), + transform_output_kernel->configure(biases, reinterpret_cast<float *>(_output_workspace.buffer()), output_matrix_stride, reinterpret_cast<float *>(_output_nhwc.buffer()), in_shape.n_batches, output_shape.n_rows, output_shape.n_cols, out_channels); // Configure Batched GEMMs - const int tile_rows = iceildiv(output_shape.n_rows, NEWinogradLayerKernel<2, 2, 3, 3>::_output_tile_rows); - const int tile_cols = iceildiv(output_shape.n_cols, NEWinogradLayerKernel<2, 2, 3, 3>::_output_tile_cols); - 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 input_matrix_row_stride = in_shape.n_channels; - const int kernel_matrix_row_stride = roundup(out_channels, NEWinogradLayerKernel<2, 2, 3, 3>::WinogradConv::N_BLOCK); - const int output_matrix_row_stride = kernel_matrix_row_stride; - - _winograd_kernel.configure(NEWinogradLayerKernel<2, 2, 3, 3>::WinogradBase::N_GEMMS, m, k, n, - input_matrix_stride, input_matrix_row_stride, - kernel_matrix_stride, kernel_matrix_row_stride, - output_matrix_stride, output_matrix_row_stride, - reinterpret_cast<float *>(_input_workspace.buffer()), reinterpret_cast<float *>(_kernel_storage.buffer()), reinterpret_cast<float *>(_output_workspace.buffer())); + const int output_tile_rows = batched_gemm_kernel->get_output_tile_rows(); + const int output_tile_cols = batched_gemm_kernel->get_output_tile_cols(); + const int n_block = batched_gemm_kernel->get_number_blocks(); + const int tile_rows = iceildiv(output_shape.n_rows, output_tile_rows); + const int tile_cols = iceildiv(output_shape.n_cols, output_tile_cols); + 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 input_matrix_row_stride = in_shape.n_channels; + const int kernel_matrix_row_stride = roundup(out_channels, n_block); + const int output_matrix_row_stride = kernel_matrix_row_stride; + const unsigned n_gemms = batched_gemm_kernel->get_number_gemms(); + + batched_gemm_kernel->configure(n_gemms, m, k, n, + input_matrix_stride, input_matrix_row_stride, + kernel_matrix_stride, kernel_matrix_row_stride, + output_matrix_stride, output_matrix_row_stride, + reinterpret_cast<float *>(_input_workspace.buffer()), + reinterpret_cast<float *>(_kernel_storage.buffer()), + reinterpret_cast<float *>(_output_workspace.buffer())); // Reorder the convoluted output to ACL's ordering NCHW _permute_output.configure(&_output_nhwc, _output, PermutationVector(1U, 2U, 0U)); + + _transform_input_kernel = std::move(transform_input_kernel); + _transform_weights_kernel = std::move(transform_weights_kernel); + _transform_output_kernel = std::move(transform_output_kernel); + _batched_gemm_kernel = std::move(batched_gemm_kernel); } void NEWinogradLayer::run() @@ -169,19 +209,19 @@ void NEWinogradLayer::run() { _reshaped_kernel = true; _permute_weights.run(); - NEScheduler::get().schedule(&_transform_weights_kernel, Window::DimX); + NEScheduler::get().schedule(_transform_weights_kernel.get(), Window::DimX); } //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, Window::DimX); + NEScheduler::get().schedule(_transform_input_kernel.get(), Window::DimX); //Run 16 GEMMs in multiple threads, each kernel runs one or more GEMMs - NEScheduler::get().schedule(&_winograd_kernel, Window::DimX); + NEScheduler::get().schedule(_batched_gemm_kernel.get(), Window::DimX); // Transform output tensor to the spatial domain - NEScheduler::get().schedule(&_transform_output_kernel, Window::DimX); + NEScheduler::get().schedule(_transform_output_kernel.get(), Window::DimX); // Reorder the convoluted output to ACL's ordering NCHW _permute_output.run(); |