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_fp16_mla_6x32/generic.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_fp16_mla_6x32/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp | 5 |
1 files changed, 1 insertions, 4 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp index 6e51773166..8877306f40 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp @@ -92,9 +92,6 @@ void a64_hybrid_fp16_mla_6x32 ( break; } __asm__ __volatile__( -#ifndef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - ".arch armv8.2-a+fp16\n" -#endif "1:" // Row loop "cmp %x[M], #0x6\n" "bge 246f\n" @@ -4068,12 +4065,12 @@ void a64_hybrid_fp16_mla_6x32 ( "ld1 { v16.8h }, [x23], #0x10\n" "ld1 { v20.8h }, [x22], #0x10\n" "ld1 { v24.8h }, [x21], #0x10\n" - "ld1 { v28.8h }, [x20], #0x10\n" "ld1 { v9.8h }, [x28], #0x10\n" "ld1 { v13.8h }, [x24], #0x10\n" "ld1 { v17.8h }, [x23], #0x10\n" "ld1 { v21.8h }, [x22], #0x10\n" "ld1 { v25.8h }, [x21], #0x10\n" + "ld1 { v28.8h }, [x20], #0x10\n" "ld1 { v29.8h }, [x20], #0x10\n" "tbz x11, #3, 252f\n" "ld1 { v10.8h }, [x28], #0x10\n" |