diff options
author | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-07-16 16:16:43 +0100 |
---|---|---|
committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-07-22 02:25:50 +0000 |
commit | 4ee8b1599dbaf7634d25607fa5ac96ba3dc6b0f2 (patch) | |
tree | 2f8362d33cdad4212f4b96995681c68184c759e1 /src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL.hpp | |
parent | 59fd7a722e5bc7e85309d6200bc37a772721a719 (diff) | |
download | ComputeLibrary-4ee8b1599dbaf7634d25607fa5ac96ba3dc6b0f2.tar.gz |
Update GEMM assembly kernels
- Introduce Fp32 kernels with internal calculations in Bfloat16 when
fast_mode is enabled
- Improve kernel selection heuristics
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I68a9e7e862b6fd2721b46e0d7cc791091c4ab279
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5965
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL.hpp | 18 |
1 files changed, 14 insertions, 4 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL.hpp index 87f063d224..c0718b1e75 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL.hpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL.hpp @@ -22,8 +22,8 @@ * IN THE SOFTWARE. */ #pragma once -#ifdef ARM_COMPUTE_ENABLE_SVE +#ifdef ARM_COMPUTE_ENABLE_SVE #include "../std_transforms_sve.hpp" #define ARGLIST \ @@ -38,11 +38,13 @@ namespace arm_gemm { // Actual kernel implementations void sve_hybrid_fp32_mla_8x1VL( ARGLIST ); +void sve_hybrid_fp32_mla_8x1VL_a64fx( ARGLIST ); class cls_sve_hybrid_fp32_mla_8x1VL { public: - typedef float operand_type; + typedef float lhs_operand_type; + typedef float rhs_operand_type; typedef float result_type; typedef void (*kern_type)( ARGLIST ); @@ -68,16 +70,24 @@ public: return true; } - StdTransformsSVE<operand_type, result_type, 8, 1, 1> transforms = {}; + StdTransformsSVE<rhs_operand_type, result_type, 8, 1, 1> transforms = {}; // Default to the generic kernel kern_type kernel=sve_hybrid_fp32_mla_8x1VL; - cls_sve_hybrid_fp32_mla_8x1VL(const CPUInfo *) + cls_sve_hybrid_fp32_mla_8x1VL(const CPUInfo *ci) { + switch(ci->get_cpu_model()) { + default: + break; + case CPUModel::A64FX: + kernel=sve_hybrid_fp32_mla_8x1VL_a64fx; + break; + } } }; } // namespace arm_gemm #undef ARGLIST + #endif // ARM_COMPUTE_ENABLE_SVE |