From beabe3bdf47306d0940ddf2ddf52ada6903a0875 Mon Sep 17 00:00:00 2001 From: Moritz Pflanzer Date: Thu, 31 Aug 2017 14:56:32 +0100 Subject: COMPMID-481: Add AArch64 GEMM Change-Id: I34f94f99cb05f0eabafee13c5e623ee779b72360 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/83741 Tested-by: Kaizen Reviewed-by: Anthony Barbier Reviewed-by: Pablo Tello --- src/runtime/NEON/functions/NEConvolutionLayer.cpp | 150 ++++++++++++++++----- src/runtime/NEON/functions/NEGEMM.cpp | 152 +++++++++++++++------- 2 files changed, 218 insertions(+), 84 deletions(-) (limited to 'src/runtime/NEON/functions') diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp index 0466a4a501..44bf2de70c 100644 --- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp @@ -23,17 +23,25 @@ */ #include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h" +#include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.h" #include "arm_compute/core/PixelValue.h" #include "arm_compute/core/Size2D.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" #include "arm_compute/runtime/NEON/NEScheduler.h" +#include "support/ToolchainSupport.h" + +namespace arm_compute +{ +#include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp" +#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_sgemm_12x8.hpp" +} // namespace arm_compute #include #include -using namespace arm_compute; - +namespace arm_compute +{ NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights(std::shared_ptr memory_manager) : _memory_group(std::move(memory_manager)), _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false) { @@ -69,8 +77,10 @@ void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const I _weights_reshaped.allocator()->init(info_wr); _memory_group.manage(&_weights_reshaped); + _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped); _weights_transposed_kernel.configure(&_weights_reshaped, output); + _weights_reshaped.allocator()->allocate(); } else @@ -84,6 +94,7 @@ void NEConvolutionLayerReshapeWeights::run() _memory_group.acquire(); NEScheduler::get().schedule(&_weights_reshape_kernel, 3); + if(_transpose1xW) { NEScheduler::get().schedule(&_weights_transposed_kernel, Window::DimY); @@ -93,8 +104,8 @@ void NEConvolutionLayerReshapeWeights::run() } NEConvolutionLayer::NEConvolutionLayer(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), - _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _has_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false) + : _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_optimised_kernel(nullptr), _output_col2im_kernel(), + _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _workspace(), _has_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false) { } @@ -137,45 +148,72 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height, conv_info); - // Check if its a "fully connected" convolution + // Check if its a "fully connected" convolution, i.e. the output size is 1x1xnum_kernels _is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1)); +#if defined(__aarch64__) + if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32) + { + _mm_optimised_kernel = support::cpp14::make_unique(); + } +#endif /* defined(__aarch64__) */ + unsigned int mat_weights_cols = weights->info()->dimension(3); unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + (_has_bias ? 1 : 0); // Reshape weights if needed - if(_are_weights_reshaped) + if(_mm_optimised_kernel != nullptr) { - mat_weights_cols = weights_info.num_kernels(); - const unsigned int quarter_reshaped_cols = weights->info()->dimension(0) / 4; - mat_weights_rows = (_has_bias ? 1 + quarter_reshaped_cols : quarter_reshaped_cols); + if(_are_weights_reshaped) + { + mat_weights_cols = weights_info.num_kernels(); + mat_weights_rows = weights->info()->dimension(1); + } + else + { + TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows }; + + // Create tensor to store the reshaped weights + _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); + _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */); + weights = &_weights_reshaped; + } } else { - if(_is_fully_connected_convolution) + if(_are_weights_reshaped) { - // Create tensor to store the reshaped weights - TensorShape shape_wr(mat_weights_cols, mat_weights_rows); - TensorInfo info_wr(shape_wr, 1, dt, fixed_point_position); - _weights_reshaped.allocator()->init(info_wr); - _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */); + mat_weights_cols = weights_info.num_kernels(); + mat_weights_rows = weights->info()->dimension(0) / 4 + (_has_bias ? 1 : 0); } else { - // Create tensor to store transposed weights - const float transpose_width = 16.0f / input->info()->element_size(); - TensorShape shape_wt(mat_weights_rows * static_cast(transpose_width), static_cast(std::ceil(mat_weights_cols / transpose_width))); - TensorInfo info_wt(shape_wt, 1, dt, fixed_point_position); - _weights_reshaped.allocator()->init(info_wt); - _reshape_weights.configure(weights, biases, &_weights_reshaped, true /* 1xW transpose */); + TensorShape reshaped_weights_shape; + + if(_is_fully_connected_convolution) + { + reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows }; + } + else + { + // Create tensor to store transposed weights + const float transpose_width = 16.0f / input->info()->element_size(); + reshaped_weights_shape = TensorShape{ mat_weights_rows *static_cast(transpose_width), + static_cast(std::ceil(mat_weights_cols / transpose_width)) }; + } + + // Create tensor to store the reshaped weights + _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); + _reshape_weights.configure(weights, biases, &_weights_reshaped, !_is_fully_connected_convolution /* 1xW transpose */); + weights = &_weights_reshaped; } - weights = &_weights_reshaped; } // Create tensor to store im2col reshaped inputs const unsigned int mat_input_cols = mat_weights_rows; const unsigned int mat_input_rows = conv_w * conv_h; - TensorShape shape_im2col = input->info()->tensor_shape(); + + TensorShape shape_im2col(input->info()->tensor_shape()); shape_im2col.set(0, mat_input_cols); shape_im2col.set(1, mat_input_rows); shape_im2col.set(2, 1); @@ -185,7 +223,7 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, // Create tensor (interleave) to prepare input tensor for GEMM if(!_is_fully_connected_convolution) { - TensorShape shape_interleaved = shape_im2col; + TensorShape shape_interleaved(shape_im2col); shape_interleaved.set(0, shape_interleaved.x() * 4); shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); _input_interleaved_reshaped.allocator()->init(TensorInfo(shape_interleaved, 1, dt, fixed_point_position)); @@ -193,7 +231,7 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, } // Create GEMM output tensor - TensorShape shape_gemm = _input_im2col_reshaped.info()->tensor_shape(); + TensorShape shape_gemm(_input_im2col_reshaped.info()->tensor_shape()); shape_gemm.set(0, mat_weights_cols); shape_gemm.set(1, mat_input_rows); _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, dt, fixed_point_position)); @@ -201,16 +239,49 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, // Configure kernels _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias); - if(_is_fully_connected_convolution) + +#if defined(__aarch64__) + if(_mm_optimised_kernel != nullptr) { - _mm_kernel.configure(&_input_im2col_reshaped, weights, &_gemm_output, 1.0f); + struct CPUInfo ci = NEScheduler::get().cpu_info(); + + const int M = _gemm_output.info()->tensor_shape().y(); + const int N = _gemm_output.info()->tensor_shape().x(); + const int K = _input_im2col_reshaped.info()->tensor_shape().x(); + + GemmInterleaved gemm(&ci, M, N, K, false, false); + + constexpr size_t alignment = 4096; + _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8)); + _memory_group.manage(&_workspace); + + // Configure matrix multiplication kernel + if(_is_fully_connected_convolution) + { + _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace, 1.f, 0.f, false, false); + } + else + { + _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace); + } + + _workspace.allocator()->allocate(); } else +#endif /* defined(__aarch64__) */ { - _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped); - _mm_kernel.configure(&_input_interleaved_reshaped, weights, &_gemm_output, 1.0f); - _input_interleaved_reshaped.allocator()->allocate(); + if(_is_fully_connected_convolution) + { + _mm_kernel.configure(&_input_im2col_reshaped, weights, &_gemm_output, 1.0f); + } + else + { + _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped); + _mm_kernel.configure(&_input_interleaved_reshaped, weights, &_gemm_output, 1.0f); + _input_interleaved_reshaped.allocator()->allocate(); + } } + _input_im2col_reshaped.allocator()->allocate(); _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h)); _gemm_output.allocator()->allocate(); @@ -237,17 +308,26 @@ void NEConvolutionLayer::run() // Run input reshaping NEScheduler::get().schedule(&_input_im2col_kernel, Window::DimY); - if(!_is_fully_connected_convolution) + + // Runs matrix multiply on reshaped matrices + if(_mm_optimised_kernel != nullptr) { - // Run interleave - NEScheduler::get().schedule(&_input_interleave_kernel, Window::DimY); + NEScheduler::get().schedule(_mm_optimised_kernel.get(), Window::DimY); } + else + { + if(!_is_fully_connected_convolution) + { + // Run interleave + NEScheduler::get().schedule(&_input_interleave_kernel, Window::DimY); + } - // Runs matrix multiply on reshaped matrices - NEScheduler::get().schedule(&_mm_kernel, Window::DimY); + NEScheduler::get().schedule(&_mm_kernel, Window::DimY); + } // Reshape output matrix NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY); _memory_group.release(); } +} // namespace arm_compute diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp index 85b283cd41..1d6aa65e37 100644 --- a/src/runtime/NEON/functions/NEGEMM.cpp +++ b/src/runtime/NEON/functions/NEGEMM.cpp @@ -26,18 +26,27 @@ #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" +#include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/Validate.h" #include "arm_compute/runtime/NEON/NEScheduler.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "support/ToolchainSupport.h" -#include +namespace arm_compute +{ +#include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp" +#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_sgemm_12x8.hpp" +} // namespace arm_compute -using namespace arm_compute; +#include +namespace arm_compute +{ NEGEMM::NEGEMM(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _run_vector_matrix_multiplication(false), _run_addition(false) + : _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _mm_optimised_kernel(nullptr), _ma_kernel(), _tmp_a(), _tmp_b(), _workspace(), + _run_vector_matrix_multiplication(false), _run_addition(false) { } @@ -57,57 +66,94 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(1) != d->info()->dimension(1), "The C matrix must have the same number of columns as the output matrix"); } - // Check if the first input tensor is a vector. If so, all the kernels for reshaping the tensors can be skipped - if((a->info()->dimension(1) == 1)) + _run_vector_matrix_multiplication = a->info()->dimension(1) < 2; + +#if defined(__aarch64__) + if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && a->info()->data_type() == DataType::F32 && (c == nullptr || beta == 0.f)) { - _run_vector_matrix_multiplication = true; + _mm_optimised_kernel = support::cpp14::make_unique(); + } +#endif /* defined(__aarch64__) */ + // Check if the first input tensor is a vector. + // If so, all the kernels for reshaping the tensors can be skipped + if(_run_vector_matrix_multiplication) + { // Configure the matrix multiply kernel _mm_kernel.configure(a, b, d, alpha); + + // Configure matrix addition kernel + if(beta != 0 && c != nullptr) + { + _ma_kernel.configure(c, d, beta); + _run_addition = true; + } } else { - _run_vector_matrix_multiplication = false; +#if defined(__aarch64__) + if(_mm_optimised_kernel != nullptr) + { + struct CPUInfo ci = NEScheduler::get().cpu_info(); - TensorShape shape_tmp_a = a->info()->tensor_shape(); - TensorShape shape_tmp_b = b->info()->tensor_shape(); + const int M = d->info()->tensor_shape().y(); + const int N = d->info()->tensor_shape().x(); + const int K = a->info()->tensor_shape().x(); - shape_tmp_a.set(0, a->info()->dimension(0) * 4); - shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.0f)); + GemmInterleaved gemm(&ci, M, N, K, false, false); - const unsigned int transpose_w = 16 / data_size_from_type(b->info()->data_type()); - shape_tmp_b.set(0, b->info()->dimension(1) * transpose_w); - shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / static_cast(transpose_w))); + constexpr size_t alignment = 4096; + _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8)); + _memory_group.manage(&_workspace); - TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position()); - TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), a->info()->fixed_point_position()); + // Configure matrix multiplication kernel + _mm_optimised_kernel->configure(a, b, d, &_workspace, alpha, 0.f); - _tmp_a.allocator()->init(info_a); - _tmp_b.allocator()->init(info_b); + _workspace.allocator()->allocate(); + } + else +#endif /* defined(__aarch64__) */ + { + TensorShape shape_tmp_a = a->info()->tensor_shape(); + TensorShape shape_tmp_b = b->info()->tensor_shape(); - // Manage intermediate buffers - _memory_group.manage(&_tmp_a); - _memory_group.manage(&_tmp_b); + shape_tmp_a.set(0, a->info()->dimension(0) * 4); + shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.0f)); - // Configure interleave kernel - _interleave_kernel.configure(a, &_tmp_a); + const unsigned int transpose_w = 16 / data_size_from_type(b->info()->data_type()); + shape_tmp_b.set(0, b->info()->dimension(1) * transpose_w); + shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / static_cast(transpose_w))); - // Configure transpose kernel - _transpose_kernel.configure(b, &_tmp_b); + TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position()); + TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), a->info()->fixed_point_position()); - // Configure matrix multiplication kernel - _mm_kernel.configure(&_tmp_a, &_tmp_b, d, alpha); + _tmp_a.allocator()->init(info_a); + _tmp_b.allocator()->init(info_b); - // Allocate once the all configure methods have been called - _tmp_a.allocator()->allocate(); - _tmp_b.allocator()->allocate(); - } + // Manage intermediate buffers + _memory_group.manage(&_tmp_a); + _memory_group.manage(&_tmp_b); - // Configure matrix addition kernel - if(beta != 0 && c != nullptr) - { - _ma_kernel.configure(c, d, beta); - _run_addition = true; + // Configure interleave kernel + _interleave_kernel.configure(a, &_tmp_a); + + // Configure transpose kernel + _transpose_kernel.configure(b, &_tmp_b); + + // Configure matrix multiplication kernel + _mm_kernel.configure(&_tmp_a, &_tmp_b, d, alpha); + + // Allocate once the all configure methods have been called + _tmp_a.allocator()->allocate(); + _tmp_b.allocator()->allocate(); + + // Configure matrix addition kernel + if(beta != 0 && c != nullptr) + { + _ma_kernel.configure(c, d, beta); + _run_addition = true; + } + } } } @@ -115,23 +161,31 @@ void NEGEMM::run() { _memory_group.acquire(); - if(!_run_vector_matrix_multiplication) + if(_mm_optimised_kernel != nullptr) { - // Run interleave kernel - NEScheduler::get().schedule(&_interleave_kernel, Window::DimY); - - // Run transpose kernel - NEScheduler::get().schedule(&_transpose_kernel, Window::DimY); + NEScheduler::get().schedule(_mm_optimised_kernel.get(), Window::DimY); + _memory_group.release(); } + else + { + if(!_run_vector_matrix_multiplication) + { + // Run interleave kernel + NEScheduler::get().schedule(&_interleave_kernel, Window::DimY); - // Run matrix multiply kernel - NEScheduler::get().schedule(&_mm_kernel, _run_vector_matrix_multiplication ? Window::DimX : Window::DimY); + // Run transpose kernel + NEScheduler::get().schedule(&_transpose_kernel, Window::DimY); + } - _memory_group.release(); + NEScheduler::get().schedule(&_mm_kernel, _run_vector_matrix_multiplication ? Window::DimX : Window::DimY); - // Run matrix addition kernel - if(_run_addition) - { - NEScheduler::get().schedule(&_ma_kernel, Window::DimY); + _memory_group.release(); + + // Run matrix addition kernel + if(_run_addition) + { + NEScheduler::get().schedule(&_ma_kernel, Window::DimY); + } } } +} // namespace arm_compute -- cgit v1.2.1