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_fp32_mla_6x16/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_fp32_mla_6x16/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp | 10 |
1 files changed, 5 insertions, 5 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp index f5504b44d4..28e9be4cb7 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp @@ -1893,8 +1893,8 @@ void a64_hybrid_fp32_mla_6x16 ( "ld1 { v22.4s }, [x22], #0x10\n" "ld1 { v26.4s }, [x21], #0x10\n" "tbz x11, #1, 136f\n" - "mov x19, #0x38\n" "ldr d11, [x28], #0x8\n" + "mov x19, #0x38\n" "ldr d15, [x24], #0x8\n" "ldr d19, [x23], #0x8\n" "ldr d23, [x22], #0x8\n" @@ -1947,8 +1947,8 @@ void a64_hybrid_fp32_mla_6x16 ( "ld1 { v20.4s }, [x22], #0x10\n" "ld1 { v24.4s }, [x21], #0x10\n" "tbz x11, #1, 140f\n" - "mov x19, #0x18\n" "ldr d9, [x28], #0x8\n" + "mov x19, #0x18\n" "ldr d13, [x24], #0x8\n" "ldr d17, [x23], #0x8\n" "ldr d21, [x22], #0x8\n" @@ -2586,12 +2586,12 @@ void a64_hybrid_fp32_mla_6x16 ( "ld1 { v16.4s }, [x23], #0x10\n" "ld1 { v20.4s }, [x22], #0x10\n" "ld1 { v24.4s }, [x21], #0x10\n" - "ld1 { v28.4s }, [x20], #0x10\n" "ld1 { v9.4s }, [x28], #0x10\n" "ld1 { v13.4s }, [x24], #0x10\n" "ld1 { v17.4s }, [x23], #0x10\n" "ld1 { v21.4s }, [x22], #0x10\n" "ld1 { v25.4s }, [x21], #0x10\n" + "ld1 { v28.4s }, [x20], #0x10\n" "ld1 { v29.4s }, [x20], #0x10\n" "tbz x11, #2, 170f\n" "ld1 { v10.4s }, [x28], #0x10\n" @@ -2601,8 +2601,8 @@ void a64_hybrid_fp32_mla_6x16 ( "ld1 { v26.4s }, [x21], #0x10\n" "ld1 { v30.4s }, [x20], #0x10\n" "tbz x11, #1, 169f\n" - "mov x19, #0x38\n" "ldr d11, [x28], #0x8\n" + "mov x19, #0x38\n" "ldr d15, [x24], #0x8\n" "ldr d19, [x23], #0x8\n" "ldr d23, [x22], #0x8\n" @@ -2662,8 +2662,8 @@ void a64_hybrid_fp32_mla_6x16 ( "ld1 { v24.4s }, [x21], #0x10\n" "ld1 { v28.4s }, [x20], #0x10\n" "tbz x11, #1, 173f\n" - "mov x19, #0x18\n" "ldr d9, [x28], #0x8\n" + "mov x19, #0x18\n" "ldr d13, [x24], #0x8\n" "ldr d17, [x23], #0x8\n" "ldr d21, [x22], #0x8\n" |