diff options
author | David Mansell <David.Mansell@arm.com> | 2018-05-17 18:51:26 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:52:54 +0000 |
commit | ce8f60510210efc0cf1c921fac75efc49bc70edc (patch) | |
tree | c9f6fb303593198d783639cce25e09ed160e2d0b /src/core/NEON/kernels/arm_gemm/gemv_batched.hpp | |
parent | 2d008a476f5f09f63574990a93e8bf606ae5629e (diff) | |
download | ComputeLibrary-ce8f60510210efc0cf1c921fac75efc49bc70edc.tar.gz |
COMPMID-1177: Improved native GEMM.
Improve the native GEMM so it can cope with any value for M. Also
change the selection code so that the native GEMM is selected if M is
small and nmulti is large - Winograd needs GEMMs like this and they
don't thread properly with the blocked GEMM.
(also rename gemm_batched.hpp back to gemv_batched.hpp)
Change-Id: I736c33373ada562cbc0c00540520a58103faa9d5
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/131739
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemv_batched.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemv_batched.hpp | 110 |
1 files changed, 110 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemv_batched.hpp b/src/core/NEON/kernels/arm_gemm/gemv_batched.hpp new file mode 100644 index 0000000000..bb09770efc --- /dev/null +++ b/src/core/NEON/kernels/arm_gemm/gemv_batched.hpp @@ -0,0 +1,110 @@ +/* + * Copyright (c) 2017-2018 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. + */ +#pragma once + +#include "arm_gemm.hpp" + +namespace arm_gemm +{ + +/* "Batched GEMV" (where M=1 and nbatches>1) can be executed much more + * efficiently as a GEMM (with M'=nbatches and nbatches'=1). This wrapper + * implements this. */ +template <typename To, typename Tr> +class GemvBatched : public GemmCommon<To, Tr> +{ +private: + UniqueGemmCommon<To, Tr> _subgemm = nullptr; + +public: + GemvBatched(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, + const To alpha, const To beta, const int maxthreads, const bool pretransposed_hint) + { + /* Just create a subgemm with batches->M */ + _subgemm = gemm<To, Tr>(ci, nbatches, N, K, 1, nmulti, trA, trB, alpha, beta, maxthreads, pretransposed_hint); + } + + void set_arrays(const To *A, const int lda, const int A_batch_stride, const int A_multi_stride, + const To *B, const int ldb, const int B_multi_stride, + Tr *C, const int ldc, const int C_batch_stride, const int C_multi_stride) override + { + /* A and C's batch stride becomes their new row stride. New batch stride is 0 as nbatches for subgemm is always 1. */ + _subgemm->set_arrays(A, A_batch_stride, 0, A_multi_stride, + B, ldb, B_multi_stride, + C, C_batch_stride, 0, C_multi_stride); + } + + unsigned int get_window_size() const override + { + return _subgemm->get_window_size(); + } + + void set_nthreads(int nthreads) override + { + _subgemm->set_nthreads(nthreads); + } + + void execute(unsigned int start, unsigned int end, int threadid) override + { + _subgemm->execute(start, end, threadid); + } + + size_t get_working_size() const override + { + return _subgemm->get_working_size(); + } + + void set_working_space(void *space) override + { + _subgemm->set_working_space(space); + } + + bool B_is_pretransposed() const override + { + return _subgemm->B_is_pretransposed(); + } + + bool B_pretranspose_required() const override + { + return _subgemm->B_pretranspose_required(); + } + + size_t get_B_pretransposed_array_size() const override + { + return _subgemm->get_B_pretransposed_array_size(); + } + + void pretranspose_B_array(void *buffer, const To *B, const int ldb, const int B_multi_stride) override + { + _subgemm->pretranspose_B_array(buffer, B, ldb, B_multi_stride); + } + + void set_pretransposed_B_data(void *buffer) override + { + _subgemm->set_pretransposed_B_data(buffer); + } +}; + +} // namespace arm_gemm |