From 7cd26d4a1b14bc4bf7c61496803416ab3d84791f Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Wed, 9 Jan 2019 18:35:17 +0000 Subject: COMPMID-1867: Add NEON/SVE GEMM Hybrid kernels. Change-Id: Ib40a9921e7f9a6a8be6c38872d6b3a0f24ed0cd3 Reviewed-on: https://review.mlplatform.org/515 Reviewed-by: Anthony Barbier Tested-by: Arm Jenkins --- .../NEGEMMInterleavedMatrixMultiplyWrapper.h | 130 ++++++++++++++++++--- 1 file changed, 111 insertions(+), 19 deletions(-) (limited to 'arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedMatrixMultiplyWrapper.h') diff --git a/arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedMatrixMultiplyWrapper.h b/arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedMatrixMultiplyWrapper.h index 46a05abcdb..e2b849aa3d 100644 --- a/arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedMatrixMultiplyWrapper.h +++ b/arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedMatrixMultiplyWrapper.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -26,8 +26,13 @@ #include "arm_compute/core/NEON/kernels/assembly/Helpers.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" #include "arm_compute/core/NEON/kernels/assembly/INEGEMMWrapperKernel.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" +#include "arm_compute/core/WindowIterator.h" namespace arm_compute { @@ -84,7 +89,7 @@ public: }; /** Equivalent to arm_gemm::GemmInterleaved's strategy::kernel() but using Compute Library types. */ -template +template class NEGEMMInterleavedMatrixMultiplyWrapperTemplate : public NEGEMMInterleavedMatrixMultiplyWrapper { public: @@ -94,7 +99,7 @@ public: * @param[in] transformed_b Already reshaped matrix B. * @param[out] tmp_c Temporary buffer to be used to store intermediate results. * @param[in,out] c Result matrix C. - * @param[in] batch_window Window containing iteration information for the M and batch dimensions. + * @param[in] block_walker Window containing iteration information for the M and batch dimensions. * @param[in] block_sizes Block sizes to use for the matrix multiplication (A & B must have been reshaped using these same block sizes). * @param[in] params M, N, K sizes. * @param[in] is_pretransposed Is B also pretransposed ? @@ -102,30 +107,117 @@ public: * @param[in] beta Beta value * @param[in] max_num_threads Maximum number of threads that might be used for the calculations. */ - void configure(const ITensor *prepared_a, const ITensor *transformed_b, ITensor *tmp_c, ITensor *c, const Window &batch_window, const BlockSizes &block_sizes, - const INEGEMMWrapperKernel::Params ¶ms, bool b_is_pretransposed, float alpha, float beta, unsigned int max_num_threads); + void configure(const ITensor *prepared_a, const ITensor *transformed_b, ITensor *tmp_c, ITensor *c, const Window &block_walker, const BlockSizes &block_sizes, + const INEGEMMWrapperKernel::Params ¶ms, bool b_is_pretransposed, float alpha, float beta, unsigned int max_num_threads) + { + _prepared_a = prepared_a; + _transformed_b = transformed_b; + _tmp_c = tmp_c; + _c = c; + _block_walker = block_walker; + _block_sizes = block_sizes; + _params = params; + _b_is_pretransposed = b_is_pretransposed; + _alpha = alpha; + _beta = beta; + + auto_init_if_empty(*_tmp_c->info(), c->info()->clone()->set_tensor_shape(TensorShape{ _block_sizes.x_block * strategy::out_height(), max_num_threads })); + } // Inherited methods overridden: - void transform(const MatrixMultiplyWorkload &wl, const ThreadInfo &info, const Window &batch_window, const Coordinates &start_offset, const Coordinates &end_offset) override; - void create_workloads(std::vector &workloads) override; + void transform(const MatrixMultiplyWorkload &wl, const ThreadInfo &info, const Window &batch_window, const Coordinates &start_offset, const Coordinates &end_offset) override + { + strategy strat(info.cpu_info); + TensorAccessor prepared_a(*_prepared_a); + TensorAccessor transformed_b(*_transformed_b); + TensorAccessor c(*_c); + TensorAccessor tmp_c(*_tmp_c); + + int prev_batch = -1; + typename strategy::operand_type *a_ptr = nullptr; + auto window_iterator = arm_compute::create_window_iterator(batch_window, start_offset, end_offset, [&](const Coordinates & id) + { + const unsigned int y = id.x(); + const unsigned int batch = id.y(); + const unsigned int ymax = std::min(_params.M, y + strategy::out_height()); + + // If it's the first block of a new batch then reset the pointer to A. + if(prev_batch != static_cast(batch)) + { + const unsigned int first_m = id.x(); + a_ptr = prepared_a(0, first_m, batch); + prev_batch = batch; + } + + // Call matrix multiply assembly routine to process the block: + strat.kernel(a_ptr, transformed_b(wl._offset_transformed_b), tmp_c(0, info.thread_id), 1, wl._bblocks, wl._kern_k); + a_ptr += strategy::out_height() * wl._kern_k; + + // Merge the result with the other blocks' results: + strat.transforms.Merge(c(0, 0, batch, wl._multi), tmp_c(0, info.thread_id), c.stride(1), y, ymax, wl._x0, wl._xmax, _alpha, (wl._k0 == 0 ? _beta : static_cast(1))); + }); + auto on_new_row_size = [&](unsigned int start, unsigned int end) + { + //Nothing to do + }; + window_iterator.iterate_2D(on_new_row_size); + } + void create_workloads(std::vector &workloads) override + { + unsigned int offset_transformed_b = 0; + unsigned int wl_index = 0; + unsigned int num_buffers = 0, reshaped_block_size = 0; + + if(!_b_is_pretransposed) + { + num_buffers = _transformed_b->info()->tensor_shape()[1]; + reshaped_block_size = _transformed_b->info()->tensor_shape()[0]; + } + execute_window_loop(_block_walker, [&](const Coordinates & id) + { + const unsigned int x0 = id.x(); + const unsigned int k0 = id.y(); + const unsigned int multi = id.z(); + + const unsigned int xmax = std::min(x0 + _block_walker.x().step(), _params.N); + const unsigned int kmax = std::min(k0 + _block_walker.y().step(), _params.K); + + // Figure out how many "K" the kernel will actually process. + const int kern_k = ceil_to_multiple(kmax - k0, strategy::k_unroll()); + const int bblocks = DIV_CEIL(xmax - x0, strategy::out_width()); + + workloads.push_back(MatrixMultiplyWorkload(offset_transformed_b, x0, xmax, k0, kmax, multi, kern_k, bblocks)); + + if(_b_is_pretransposed) + { + offset_transformed_b += bblocks * strategy::out_width() * kern_k; + } + else + { + // Rotate through the BufferManager's buffers: + wl_index++; + offset_transformed_b = (wl_index % num_buffers) * reshaped_block_size; + } + }); + } private: const ITensor *_prepared_a { nullptr }; - const ITensor *_transformed_b{ nullptr }; - ITensor *_tmp_c{ nullptr }; - ITensor *_c{ nullptr }; - unsigned int _Nsize{ 0 }; - unsigned int _Ksize{ 0 }; - bool _transpose_b{ false }; - BlockSizes _block_sizes{}; - INEGEMMWrapperKernel::Params _params{}; - Window _block_walker{}; - bool _b_is_pretransposed{ false }; - Tr _alpha{}; - Tr _beta{}; + const ITensor *_transformed_b{ nullptr }; + ITensor *_tmp_c{ nullptr }; + ITensor *_c{ nullptr }; + unsigned int _Nsize{ 0 }; + unsigned int _Ksize{ 0 }; + bool _transpose_b{ false }; + BlockSizes _block_sizes{}; + INEGEMMWrapperKernel::Params _params{}; + Window _block_walker{}; + bool _b_is_pretransposed{ false }; + typename strategy::result_type _alpha{}; + typename strategy::result_type _beta{}; }; } // namespace arm_compute -- cgit v1.2.1