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/a64_hybrid_s8qs_dot_6x16/a55.cpp | |
| 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/a64_hybrid_s8qs_dot_6x16/a55.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp | 12 |
1 files changed, 6 insertions, 6 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp index 6e3a00ed72..ba8a2ccb1d 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp @@ -309,8 +309,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v0.4s }, [x25]\n" "ld1r { v4.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "mov v2.16b, v0.16b\n" "mov v5.16b, v4.16b\n" + "mov v2.16b, v0.16b\n" "mov v6.16b, v4.16b\n" "mov v3.16b, v0.16b\n" "mov v7.16b, v4.16b\n" @@ -693,8 +693,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v0.4s }, [x25]\n" "ld1r { v4.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "mov v2.16b, v0.16b\n" "mov v5.16b, v4.16b\n" + "mov v2.16b, v0.16b\n" "mov v6.16b, v4.16b\n" "mov v3.16b, v0.16b\n" "mov v7.16b, v4.16b\n" @@ -1193,8 +1193,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v0.4s }, [x25]\n" "ld1r { v4.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "mov v2.16b, v0.16b\n" "mov v5.16b, v4.16b\n" + "mov v2.16b, v0.16b\n" "mov v6.16b, v4.16b\n" "mov v3.16b, v0.16b\n" "mov v7.16b, v4.16b\n" @@ -1809,8 +1809,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v0.4s }, [x25]\n" "ld1r { v4.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "mov v2.16b, v0.16b\n" "mov v5.16b, v4.16b\n" + "mov v2.16b, v0.16b\n" "mov v6.16b, v4.16b\n" "mov v3.16b, v0.16b\n" "mov v7.16b, v4.16b\n" @@ -2541,8 +2541,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v0.4s }, [x25]\n" "ld1r { v4.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "mov v2.16b, v0.16b\n" "mov v5.16b, v4.16b\n" + "mov v2.16b, v0.16b\n" "mov v6.16b, v4.16b\n" "mov v3.16b, v0.16b\n" "mov v7.16b, v4.16b\n" @@ -3392,8 +3392,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v0.4s }, [x25]\n" "ld1r { v4.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "mov v2.16b, v0.16b\n" "mov v5.16b, v4.16b\n" + "mov v2.16b, v0.16b\n" "mov v6.16b, v4.16b\n" "mov v3.16b, v0.16b\n" "mov v7.16b, v4.16b\n" |