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_u8qa_dot_4x4VL.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_u8qa_dot_4x4VL.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL.hpp | 24 |
1 files changed, 21 insertions, 3 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL.hpp index 3de8d178cd..c66ebedc4d 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL.hpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL.hpp @@ -22,9 +22,10 @@ * IN THE SOFTWARE. */ #pragma once -#ifdef ARM_COMPUTE_ENABLE_SVE +#ifdef ARM_COMPUTE_ENABLE_SVE #include "../std_transforms_sve.hpp" +#include "../performance_parameters.hpp" #define ARGLIST \ unsigned int, const unsigned int *, \ @@ -42,7 +43,8 @@ void sve_hybrid_u8qa_dot_4x4VL( ARGLIST ); class cls_sve_hybrid_u8qa_dot_4x4VL { public: - typedef uint8_t operand_type; + typedef uint8_t lhs_operand_type; + typedef uint8_t rhs_operand_type; typedef uint8_t result_type; typedef void (*kern_type)( ARGLIST ); @@ -68,7 +70,22 @@ public: return false; } - StdTransformsSVE<operand_type, result_type, 4, 4, 4> transforms = {}; + StdTransformsSVE<rhs_operand_type, result_type, 4, 4, 4> transforms = {}; + template<typename T> + static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci) + { + + if (std::is_same<T, uint8_t>::value) { + switch (ci->get_cpu_model()) { + default: + return { 29.89 }; + case CPUModel::A510: + return { 17.12 }; + } + } + + return { 1.0 }; + } // Default to the generic kernel kern_type kernel=sve_hybrid_u8qa_dot_4x4VL; @@ -80,4 +97,5 @@ public: } // namespace arm_gemm #undef ARGLIST + #endif // ARM_COMPUTE_ENABLE_SVE |