diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp | 18 |
1 files changed, 4 insertions, 14 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp index fad0e84bbb..f3b66528a4 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp @@ -27,9 +27,7 @@ #include "gemm_hybrid.hpp" #include "gemm_implementation.hpp" #include "gemm_interleaved.hpp" -#include "gemm_native.hpp" #include "gemv_batched.hpp" -#include "gemv_native_transposed.hpp" #include "gemv_pretransposed.hpp" #include "kernels/a64_interleaved_bf16fp32_dot_12x8.hpp" @@ -38,7 +36,6 @@ #include "kernels/a32_sgemm_8x6.hpp" #include "kernels/sve_interleaved_bf16fp32_dot_3VLx8.hpp" #include "kernels/sve_interleaved_bf16fp32_mmla_3VLx8.hpp" -#include "kernels/sve_native_bf16fp32_dot_4VLx4.hpp" #include "kernels/sve_hybrid_bf16fp32_dot_4VLx4.hpp" #include "kernels/sve_hybrid_bf16fp32_mmla_4VLx4.hpp" #include "kernels/sve_hybrid_bf16fp32_mmla_6VLx2.hpp" @@ -53,38 +50,31 @@ static const GemmImplementation<bfloat16, float> gemm_bf16_methods[] = { GemmMethod::GEMM_HYBRID, "hybrid_bf16fp32_mmla_6VLx2", - [](const GemmArgs &args) { return (args._Ksize>=8 && !args._trA && args._pretransposed_hint); }, + [](const GemmArgs &args) { return (args._Ksize>=8); }, [](const GemmArgs &args) { return ((args._Msize <= 4) && (args._Nsize <= hybrid_bf16fp32_mmla_6VLx2::out_width())); }, [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_mmla_6VLx2, bfloat16, float>(args); } }, { GemmMethod::GEMM_HYBRID, "hybrid_bf16fp32_mmla_8VLx2", - [](const GemmArgs &args) { return (args._Ksize>=8 && !args._trA && args._pretransposed_hint); }, + [](const GemmArgs &args) { return (args._Ksize>=8); }, [](const GemmArgs &args) { return (args._Msize <= 4); }, [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_mmla_8VLx2, bfloat16, float>(args); } }, { GemmMethod::GEMM_HYBRID, "hybrid_bf16fp32_mmla_4VLx4", - [](const GemmArgs &args) { return (args._Ksize>=8 && !args._trA && args._pretransposed_hint); }, + [](const GemmArgs &args) { return (args._Ksize>=8); }, [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); }, [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_mmla_4VLx4, bfloat16, float>(args); } }, { GemmMethod::GEMM_HYBRID, "hybrid_bf16fp32_dot_4VLx4", - [](const GemmArgs &args) { return (args._Ksize>=8 && !args._trA && args._pretransposed_hint); }, + [](const GemmArgs &args) { return (args._Ksize>=8); }, [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); }, [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_dot_4VLx4, bfloat16, float>(args); } }, -{ // gemm_bf16_native - GemmMethod::GEMM_NATIVE, - "native_bf16fp32_dot_4VLx4", - [](const GemmArgs &args) { return (args._Ksize>=8 && !args._trA && !args._trB); }, - [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); }, - [](const GemmArgs &args) { return new GemmNative<native_bf16fp32_dot_4VLx4, bfloat16, float>(args); } -}, { // gemm_bf16_interleaved GemmMethod::GEMM_INTERLEAVED, "interleaved_bf16fp32_mmla_3VLx8", |