diff options
author | Georgios Pinitas <georgios.pinitas@arm.com> | 2020-07-06 19:10:38 +0100 |
---|---|---|
committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2020-07-07 15:07:37 +0000 |
commit | 0cc50ed757f06f4f076e261cb7253dd67264dec6 (patch) | |
tree | 866306f780f7529ea172d78210355aed761ad7a4 /src/core/NEON/kernels/assembly | |
parent | ab23dd0fbc632063235a6ad408241dc79a35d3e4 (diff) | |
download | ComputeLibrary-0cc50ed757f06f4f076e261cb7253dd67264dec6.tar.gz |
COMPMID-3324: Remove pretransposed support from NEON backend
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I394c6c539969940e0119cbc14174909d47e65de6
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3519
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/assembly')
-rw-r--r-- | src/core/NEON/kernels/assembly/Helpers.cpp | 67 | ||||
-rw-r--r-- | src/core/NEON/kernels/assembly/Helpers.h | 122 | ||||
-rw-r--r-- | src/core/NEON/kernels/assembly/arm_gemm.hpp | 10 |
3 files changed, 3 insertions, 196 deletions
diff --git a/src/core/NEON/kernels/assembly/Helpers.cpp b/src/core/NEON/kernels/assembly/Helpers.cpp deleted file mode 100644 index 5990505a59..0000000000 --- a/src/core/NEON/kernels/assembly/Helpers.cpp +++ /dev/null @@ -1,67 +0,0 @@ -/* - * Copyright (c) 2018-2020 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/core/NEON/kernels/assembly/Helpers.h" - -namespace arm_compute -{ -arm_gemm::KernelDescription get_gemm_info(DataType input_type, - const CPUInfo &ci, - const unsigned int num_threads, - const INEGEMMWrapperKernel::Params &p, - arm_gemm::Activation activation, - bool pretranspose_hint) -{ - switch(input_type) - { -#ifdef __aarch64__ - case DataType::QASYMM8: - case DataType::U8: - { - arm_gemm::GemmArgs args(&ci, p.M, p.N, p.K, p.batches, p.multis, false, false, activation, num_threads, pretranspose_hint); - return arm_gemm::get_gemm_method<uint8_t, uint32_t>(args); - } - case DataType::S8: - { - arm_gemm::GemmArgs args(&ci, p.M, p.N, p.K, p.batches, p.multis, false, false, activation, num_threads, pretranspose_hint); - return arm_gemm::get_gemm_method<int8_t, int32_t>(args); - } -#endif // __aarch64__ -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - { - arm_gemm::GemmArgs args(&ci, p.M, p.N, p.K, p.batches, p.multis, false, false, activation, num_threads, pretranspose_hint); - return arm_gemm::get_gemm_method<__fp16, __fp16>(args); - } -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - case DataType::F32: - { - arm_gemm::GemmArgs args(&ci, p.M, p.N, p.K, p.batches, p.multis, false, false, activation, num_threads, pretranspose_hint); - return arm_gemm::get_gemm_method<float, float>(args); - } - default: - return arm_gemm::KernelDescription(); - } -} -} // namespace arm_compute diff --git a/src/core/NEON/kernels/assembly/Helpers.h b/src/core/NEON/kernels/assembly/Helpers.h deleted file mode 100644 index 09c0446ada..0000000000 --- a/src/core/NEON/kernels/assembly/Helpers.h +++ /dev/null @@ -1,122 +0,0 @@ -/* - * Copyright (c) 2018-2020 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_ASSEMBLY_HELPERS_H -#define ARM_COMPUTE_ASSEMBLY_HELPERS_H - -#include "arm_compute/core/CPP/CPPTypes.h" -#include "arm_compute/core/Utils.h" - -#include "arm_compute/core/NEON/kernels/assembly/INEGEMMWrapperKernel.h" -#include "arm_gemm.hpp" - -namespace arm_compute -{ -/** Block sizes to use to break the M, N, K dimension */ -struct BlockSizes -{ - unsigned int k_block{ 0 }; /**< Block size alon the K dimension */ - unsigned int x_block{ 0 }; /**< Block size along the N (x) dimension */ - unsigned int m_round{ 0 }; /**< Block size along the M dimension (Must be a multiple of strategy_out_height) */ - unsigned int strategy_out_height{ 0 }; /**< Number of rows (M) processed by the selected strategy */ -}; - -/** Extracts the kernel description of the selected kernel by the GEMM backend heuristics - * - * @param[in] input_type Data type of the input tensor. - * @param[in] ci CPU information. - * @param[in] num_threads Maximum number of threads that might be used for the calculations. - * @param[in] p M, N, K sizes. - * @param[in] activation Activation struct - * @param[in] pretranspose_hint Is B also pretransposed ? - * - * @return Kernel description that the assembly heuristics picked for the given configuration - */ -arm_gemm::KernelDescription get_gemm_info(DataType input_type, - const CPUInfo &ci, - const unsigned int num_threads, - const INEGEMMWrapperKernel::Params &p, - arm_gemm::Activation activation, - bool pretranspose_hint); - -/** Calculate the recommended block sizes to use based on the CPU cache sizes and the strategy which will be used - * - * @param[in] ci CPU information. - * @param[in] M M dimension. - * @param[in] N N dimension. - * @param[in] K K dimension. - * - * @return Recommeded block sizes to use for the given M, N, K dimensions. - */ -template <typename strategy> -BlockSizes calculate_block_sizes(const CPUInfo &ci, unsigned int M, unsigned int N, unsigned int K) -{ - BlockSizes bs; - - using Toi = typename strategy::operand_type; - - const unsigned int L1_size = ci.get_L1_cache_size(); - const unsigned int L2_size = ci.get_L2_cache_size(); - - // Work out blocking parameters - - // k_block: Find out how much of the larger array can be loaded into half the cache. - // This should account for associative caches. - bs.k_block = (L1_size / 2) / (sizeof(Toi) * (std::max(strategy::out_width(), strategy::out_height()))); - - // Needs to be (at least a single) multiple of the K unroll level. - bs.k_block /= strategy::k_unroll(); - bs.k_block = std::max(bs.k_block, 1U) * strategy::k_unroll(); - - // Now tune to presented problem size; this is how many blocks we need. - int num_k_blocks = DIV_CEIL(K, bs.k_block); - - // So divide the space equally into that many blocks. - bs.k_block = DIV_CEIL(K, num_k_blocks); - - // And round UP to the K unroll level required. - bs.k_block = ceil_to_multiple(bs.k_block, strategy::k_unroll()); - - // x_block: Work out how many rows (of length k_block) will fit in the L2 - // Don't allocate more than 90% of the L2 to allow for overheads, and subtract off the L1 contents. - bs.x_block = (((L2_size * 9) / 10) - (bs.k_block * sizeof(Toi) * (strategy::out_width() + strategy::out_height()))) / (sizeof(Toi) * bs.k_block); - - // Needs to be (at least a single) multiple of the kernel output width. - bs.x_block /= strategy::out_width(); - bs.x_block = std::max(bs.x_block, 1U) * strategy::out_width(); - - // And tune to the presented problem size. - int num_x_blocks = DIV_CEIL(N, bs.x_block); - bs.x_block = DIV_CEIL(N, num_x_blocks); - - bs.x_block = ceil_to_multiple(bs.x_block, strategy::out_width()); - - // Work out the rounded size of M - needed for some buffers. - bs.m_round = ceil_to_multiple(M, strategy::out_height()); - bs.strategy_out_height = strategy::out_height(); - - return bs; -} - -} // namespace arm_compute -#endif /* ARM_COMPUTE_ASSEMBLY_HELPERS_H */ diff --git a/src/core/NEON/kernels/assembly/arm_gemm.hpp b/src/core/NEON/kernels/assembly/arm_gemm.hpp index 2df7132500..58bc9a51d1 100644 --- a/src/core/NEON/kernels/assembly/arm_gemm.hpp +++ b/src/core/NEON/kernels/assembly/arm_gemm.hpp @@ -104,19 +104,15 @@ public: unsigned int _Ksize; unsigned int _nbatches; unsigned int _nmulti; - bool _trA; - bool _trB; Activation _act; int _maxthreads; - bool _pretransposed_hint; const GemmConfig *_cfg; GemmArgs(const CPUInfo *ci, const unsigned int M, const unsigned int N, const unsigned int K, const unsigned int nbatches, - const unsigned int nmulti, const bool trA, const bool trB, - Activation act, const int maxthreads, - const bool pretransposed_hint, const GemmConfig *cfg = nullptr) - : _ci(ci), _Msize(M), _Nsize(N), _Ksize(K), _nbatches(nbatches), _nmulti(nmulti), _trA(trA), _trB(trB), _act(act), _maxthreads(maxthreads), _pretransposed_hint(pretransposed_hint), _cfg(cfg) + const unsigned int nmulti, Activation act, const int maxthreads, + const GemmConfig *cfg = nullptr) + : _ci(ci), _Msize(M), _Nsize(N), _Ksize(K), _nbatches(nbatches), _nmulti(nmulti), _act(act), _maxthreads(maxthreads), _cfg(cfg) { } }; |