From 387f80ed5b4c8ef32cfdba4ea27734dd257f87f0 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Tue, 27 Jul 2021 10:46:13 +0100 Subject: Update GEMM assembly performance parameters Resolves: COMPMID-4709 Signed-off-by: Georgios Pinitas Change-Id: I20913c6153b8b86edf6e83d5f64175dcdf76dab4 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6002 Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio Comments-Addressed: Arm Jenkins --- .../NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12.hpp | 15 ++++++++++++--- .../NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp | 4 ++-- .../NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12.hpp | 7 +++++-- .../kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp | 2 +- .../kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp | 2 +- 5 files changed, 21 insertions(+), 9 deletions(-) diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12.hpp index 83ccb4681b..758f2b1f8c 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12.hpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12.hpp @@ -66,11 +66,14 @@ public: if (std::is_same::value) { switch (ci->get_cpu_model()) { case CPUModel::A510: - return { 19.73, 2.81, 0.27 }; + return { 19.73, 3.38, 0.27 }; case CPUModel::A55r1: return { 15.361, 0.9341, 0.1636 }; + case CPUModel::V1: + return { 62.40, 4.71, 0.67 }; + default: return { 29.0698, 3.9793, 0.4003 }; } @@ -79,10 +82,16 @@ public: if (std::is_same::value) { switch (ci->get_cpu_model()) { case CPUModel::A510: - return { 19.73, 3.41, 3.70 }; + return { 19.73, 3.38, 3.70 }; + + case CPUModel::A55r1: + return { 14.286, 1.171, 1.209 }; + + case CPUModel::V1: + return { 61.58, 4.78, 10.83 }; default: - return { 31.81, 3.68, 8.01 }; + return { 31.82, 3.51, 8.03 }; } } } diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp index 07c4769479..45ba8dc423 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp @@ -69,7 +69,7 @@ public: template static PerformanceParameters get_performance_parameters(const CPUInfo *ci) { - if (std::is_same::value) { + if (std::is_same::value) { switch (ci->get_cpu_model()) { case CPUModel::A510: return { 2.64, 2.72, 2.64 }; @@ -79,7 +79,7 @@ public: } } - if (std::is_same::value) { + if (std::is_same::value) { switch(ci->get_cpu_model()) { case CPUModel::A510: return { 2.64, 1.79, 0.10 }; diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12.hpp index 0329f57615..21c9f59661 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12.hpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12.hpp @@ -71,7 +71,7 @@ public: template static PerformanceParameters get_performance_parameters(const CPUInfo *ci) { - if (std::is_same::value) { + if (std::is_same::value) { switch (ci->get_cpu_model()) { case CPUModel::A510: return { 19.73, 3.38, 0.27 }; @@ -87,11 +87,14 @@ public: } } - if (std::is_same::value) { + if (std::is_same::value) { switch (ci->get_cpu_model()) { case CPUModel::A510: return { 19.73, 3.38, 3.70 }; + case CPUModel::A55r1: + return { 14.286, 1.171, 1.209 }; + case CPUModel::V1: return { 61.58, 4.78, 10.83 }; diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp index 2142f1067d..30e265fbc0 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp @@ -83,7 +83,7 @@ public: case CPUModel::A510: return { 5.42 }; case CPUModel::V1: - return { 28.40 }; + return { 34.56 }; } } diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp index d941ccc0e9..61c7ad17e7 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp @@ -83,7 +83,7 @@ public: case CPUModel::A510: return { 5.31 }; case CPUModel::V1: - return { 26.64 }; + return { 28.93 }; } } -- cgit v1.2.1