diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_int8.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_int8.cpp | 8 |
1 files changed, 4 insertions, 4 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp index fd20e53f60..fedda3a47a 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) 2017-2020, 2022-2023 Arm Limited. + * Copyright (c) 2017-2020, 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -63,7 +63,7 @@ static const GemmImplementation<int8_t, int32_t> gemm_s8_methods[] = { "sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL", [](const GemmArgs &args) { return args._ci->has_sme2(); }, [](const GemmArgs &args) { const auto VL = sme::get_vector_length<int32_t>(); - return args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); }, + return args._Nsize >= 8*VL || args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); }, [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL, int8_t, int32_t>(args); } }, { @@ -128,14 +128,14 @@ GemmImplementation<int8_t, int32_t>::with_estimate( { GemmMethod::GEMM_HYBRID, "a64_smallK_hybrid_s8s32_dot_8x4", - [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input; }, + [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input && !args._accumulate; }, [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); }, [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_s8s32_dot_8x4, int8_t, int32_t>(args); } }, { GemmMethod::GEMM_HYBRID, "a64_smallK_hybrid_s8s32_dot_6x4", - [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input; }, + [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input && !args._accumulate; }, [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); }, [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_s8s32_dot_6x4, int8_t, int32_t>(args); } }, |