diff options
| author | Georgios Pinitas <georgios.pinitas@arm.com> | 2019-01-23 11:24:50 +0000 |
|---|---|---|
| committer | Michalis Spyrou <michalis.spyrou@arm.com> | 2019-01-24 10:19:46 +0000 |
| commit | 1d480652b820317fc97ccbc3cb517e3b9e8be197 (patch) | |
| tree | b3c845ec02cccf89430b95186ed3e3f2ae65e2bd /src/core/NEON/kernels/arm_gemm/gemm_int8.cpp | |
| parent | 20b527a7029d02d0edda78fd92002cbc430dbe05 (diff) | |
| download | ComputeLibrary-1d480652b820317fc97ccbc3cb517e3b9e8be197.tar.gz | |
COMPMID-1867: Add u8 and s8 hybrid assembly kernels.
Change-Id: Ifeb005f9d18d19feff11949474cce84d9e03749c
Reviewed-on: https://review.mlplatform.org/565
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_int8.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_int8.cpp | 11 |
1 files changed, 10 insertions, 1 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp index 34dc8bc341..5811c2a1ce 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 ARM Limited. + * Copyright (c) 2017-2019 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,6 +25,7 @@ #include "arm_gemm.hpp" #include "gemm_common.hpp" +#include "gemm_hybrid.hpp" #include "gemm_implementation.hpp" #include "gemm_interleaved.hpp" #include "gemm_native.hpp" @@ -32,6 +33,7 @@ #include "kernels/a64_gemm_s16_12x8.hpp" #include "kernels/a64_gemm_s8_12x8.hpp" #include "kernels/a64_gemm_s8_4x4.hpp" +#include "kernels/a64_hybrid_s8s32_dot_16x4.hpp" #include "kernels/sve_interleaved_s8s32_dot_3VLx8.hpp" #include "kernels/sve_native_s8s32_dot_4VLx4.hpp" @@ -55,6 +57,13 @@ static const GemmImplementation<int8_t, int32_t> gemm_s8_methods[] = { }, #endif { + GemmMethod::GEMM_HYBRID, + "hybrid_s8s32_dot_16x4", + [](const GemmArgs<int32_t> &args) { return args._ci->has_dotprod() && args._Ksize>=16 && (args._Ksize % 16 == 0) && (args._Nsize % 16 == 0) && !args._trA && !args._trB && args._pretransposed_hint; }, + [](const GemmArgs<int32_t> &args) { return args._Nsize<=256 && args._Ksize>128; }, + [](const GemmArgs<int32_t> &args) { return new GemmHybrid<hybrid_s8s32_dot_16x4, int8_t, int32_t>(args); } +}, +{ GemmMethod::GEMM_INTERLEAVED, "gemm_s8_12x8", [](const GemmArgs<int32_t> &args) { return args._ci->has_dotprod(); }, |