diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp | 46 |
1 files changed, 23 insertions, 23 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp index f5a30720d2..aead814d7e 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp @@ -48,66 +48,66 @@ static const GemmImplementation<uint8_t, uint32_t> gemm_u8_methods[] = { { GemmMethod::GEMM_HYBRID, "smallK_hybrid_u8u32_dot_1VLx8", - [](const GemmArgs<uint32_t> &args) { return args._Ksize<=64 && args._alpha==1 && args._beta==0 && !args._trA && args._pretransposed_hint; }, + [](const GemmArgs &args) { return args._Ksize<=64 && !args._trA && args._pretransposed_hint; }, nullptr, - [](const GemmArgs<uint32_t> &args) { return new GemmHybrid<smallK_hybrid_u8u32_dot_1VLx8, uint8_t, uint32_t>(args); } + [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_u8u32_dot_1VLx8, uint8_t, uint32_t>(args); } }, { GemmMethod::GEMM_HYBRID, "hybrid_u8u32_dot_4VLx4", - [](const GemmArgs<uint32_t> &args) { return args._Ksize>=16 && args._alpha==1 && !args._trA && args._pretransposed_hint; }, - [](const GemmArgs<uint32_t> &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); }, - [](const GemmArgs<uint32_t> &args) { return new GemmHybrid<hybrid_u8u32_dot_4VLx4, uint8_t, uint32_t>(args); } + [](const GemmArgs &args) { return args._Ksize>=16 && !args._trA && args._pretransposed_hint; }, + [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); }, + [](const GemmArgs &args) { return new GemmHybrid<hybrid_u8u32_dot_4VLx4, uint8_t, uint32_t>(args); } }, { GemmMethod::GEMM_NATIVE, "native_u8u32_dot_4VLx4", - [](const GemmArgs<uint32_t> &args) { return (args._Ksize>=16 && args._alpha==1 && !args._trA && !args._trB); }, - [](const GemmArgs<uint32_t> &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); }, - [](const GemmArgs<uint32_t> &args) { return new GemmNative<native_u8u32_dot_4VLx4, uint8_t, uint32_t>(args); } + [](const GemmArgs &args) { return (args._Ksize>=16 && !args._trA && !args._trB); }, + [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); }, + [](const GemmArgs &args) { return new GemmNative<native_u8u32_dot_4VLx4, uint8_t, uint32_t>(args); } }, { GemmMethod::GEMM_INTERLEAVED, "interleaved_u8u32_dot_3VLx8", - [](const GemmArgs<uint32_t> &args) { return (args._Ksize>4) && args._alpha==1 && (args._beta==0 || args._beta==1); }, + [](const GemmArgs &args) { return (args._Ksize>4); }, nullptr, - [](const GemmArgs<uint32_t> &args) { return new GemmInterleaved<interleaved_u8u32_dot_3VLx8, uint8_t, uint32_t>(args); } + [](const GemmArgs &args) { return new GemmInterleaved<interleaved_u8u32_dot_3VLx8, uint8_t, uint32_t>(args); } }, #endif { GemmMethod::GEMM_HYBRID, "smallK_hybrid_u8u32_dot_4x8", - [](const GemmArgs<uint32_t> &args) { 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) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._trA && args._pretransposed_hint; }, nullptr, - [](const GemmArgs<uint32_t> &args) { return new GemmHybrid<smallK_hybrid_u8u32_dot_4x8, uint8_t, uint32_t>(args); } + [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_u8u32_dot_4x8, uint8_t, uint32_t>(args); } }, { GemmMethod::GEMM_HYBRID, "smallK_hybrid_u8u32_dot_4x6", - [](const GemmArgs<uint32_t> &args) { 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) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._trA && args._pretransposed_hint; }, nullptr, - [](const GemmArgs<uint32_t> &args) { return new GemmHybrid<smallK_hybrid_u8u32_dot_4x6, uint8_t, uint32_t>(args); } + [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_u8u32_dot_4x6, uint8_t, uint32_t>(args); } }, { GemmMethod::GEMM_HYBRID, "hybrid_u8u32_dot_16x4", - [](const GemmArgs<uint32_t> &args) { return args._ci->has_dotprod() && args._Ksize>=16 && args._alpha==1 && !args._trA && !args._trB && args._pretransposed_hint; }, - [](const GemmArgs<uint32_t> &args) { return args._Nsize<=256 && args._Ksize>128; }, - [](const GemmArgs<uint32_t> &args) { return new GemmHybrid<hybrid_u8u32_dot_16x4, uint8_t, uint32_t>(args); } + [](const GemmArgs &args) { return args._ci->has_dotprod() && args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; }, + [](const GemmArgs &args) { return args._Nsize<=256 && args._Ksize>128; }, + [](const GemmArgs &args) { return new GemmHybrid<hybrid_u8u32_dot_16x4, uint8_t, uint32_t>(args); } }, { GemmMethod::GEMM_INTERLEAVED, "gemm_u8_12x8", - [](const GemmArgs<uint32_t> &args) { return args._ci->has_dotprod() && args._alpha==1 && (args._beta==0 || args._beta==1); }, + [](const GemmArgs &args) { return args._ci->has_dotprod(); }, nullptr, - [](const GemmArgs<uint32_t> &args) { return new GemmInterleaved<gemm_u8_12x8, uint8_t, uint32_t>(args); } + [](const GemmArgs &args) { return new GemmInterleaved<gemm_u8_12x8, uint8_t, uint32_t>(args); } }, { GemmMethod::GEMM_INTERLEAVED, "gemm_u8_4x4", nullptr, nullptr, - [](const GemmArgs<uint32_t> &args) { return new GemmInterleaved<gemm_u8_4x4, uint8_t, uint32_t>(args); } + [](const GemmArgs &args) { return new GemmInterleaved<gemm_u8_4x4, uint8_t, uint32_t>(args); } }, { GemmMethod::DEFAULT, @@ -124,9 +124,9 @@ const GemmImplementation<uint8_t, uint32_t> *gemm_implementation_list<uint8_t, u } /* Explicitly instantiate the external functions for these types. */ -template UniqueGemmCommon<uint8_t, uint32_t> gemm<uint8_t, uint32_t, Nothing>(const GemmArgs<uint32_t> &args, const Nothing &); -template KernelDescription get_gemm_method<uint8_t, uint32_t, Nothing>(const GemmArgs<uint32_t> &args, const Nothing &); -template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint32_t, Nothing> (const GemmArgs<uint32_t> &args, const Nothing &); +template UniqueGemmCommon<uint8_t, uint32_t> gemm<uint8_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &); +template KernelDescription get_gemm_method<uint8_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &); +template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint32_t, Nothing> (const GemmArgs &args, const Nothing &); } // namespace arm_gemm |