From 48b3ef89de5f21a0169d8416e3d54081f82c7bf8 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 14 Oct 2019 19:03:09 +0100 Subject: COMPMID-2577: Fuse bias addition and activation in gemm assembly kernels Change-Id: I7f52112d2d05b1ea3d3f3d4b19b8eafab05d6c44 Signed-off-by: Georgios Pinitas Reviewed-on: https://review.mlplatform.org/c/2141 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Pablo Marquez --- src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp | 32 +++++++++++++------------- 1 file changed, 16 insertions(+), 16 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp index a402332552..079c04ae06 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp @@ -42,45 +42,45 @@ static const GemmImplementation gemm_quint { GemmMethod::GEMM_HYBRID_QUANTIZED, "smallK_hybrid_u8u32_dot_1VLx8", - [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._Ksize<=64 && args._alpha==1 && args._beta==0 && !args._trA && args._pretransposed_hint; }, + [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._Ksize<=64 && !args._trA && args._pretransposed_hint; }, nullptr, - [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized(args, qp); } + [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized(args, qp); } }, { GemmMethod::GEMM_HYBRID_QUANTIZED, "hybrid_u8u32_dot_4VLx4", - [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._Ksize>=16 && args._alpha==1 && !args._trA && !args._trB && args._pretransposed_hint; }, - [](const GemmArgs &args, const ARequantizeLayer32 &) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); }, - [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized(args, qp); } + [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; }, + [](const GemmArgs &args, const ARequantizeLayer32 &) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); }, + [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized(args, qp); } }, #endif { GemmMethod::GEMM_HYBRID_QUANTIZED, "smallK_hybrid_u8u32_dot_4x8", - [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && args._alpha==1 && args._beta==0 && !args._trA && args._pretransposed_hint; }, + [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._trA && args._pretransposed_hint; }, nullptr, - [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized(args, qp); } + [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized(args, qp); } }, { GemmMethod::GEMM_HYBRID_QUANTIZED, "smallK_hybrid_u8u32_dot_4x6", - [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && args._alpha==1 && args._beta==0 && !args._trA && args._pretransposed_hint; }, + [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._trA && args._pretransposed_hint; }, nullptr, - [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized(args, qp); } + [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized(args, qp); } }, { GemmMethod::GEMM_HYBRID_QUANTIZED, "hybrid_u8u32_dot_16x4", - [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._ci->has_dotprod() && args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; }, - [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._Nsize<=256 && args._Ksize>128; }, - [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized(args, qp); } + [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._ci->has_dotprod() && args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; }, + [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._Nsize<=256 && args._Ksize>128; }, + [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized(args, qp); } }, { GemmMethod::QUANTIZE_WRAPPER, "quantized_wrapper", nullptr, nullptr, - [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new QuantizeWrapper(args, qp); } + [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new QuantizeWrapper(args, qp); } }, { GemmMethod::DEFAULT, @@ -96,9 +96,9 @@ const GemmImplementation *gemm_implementat return gemm_quint8_methods; } -template UniqueGemmCommon gemm(const GemmArgs &args, const ARequantizeLayer32 &os); -template KernelDescription get_gemm_method(const GemmArgs &args, const ARequantizeLayer32 &os); -template std::vector get_compatible_kernels(const GemmArgs &args, const ARequantizeLayer32 &os); +template UniqueGemmCommon gemm(const GemmArgs &args, const ARequantizeLayer32 &os); +template KernelDescription get_gemm_method(const GemmArgs &args, const ARequantizeLayer32 &os); +template std::vector get_compatible_kernels(const GemmArgs &args, const ARequantizeLayer32 &os); } // namespace arm_gemm -- cgit v1.2.1