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_u8qa_dot_4x16/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_u8qa_dot_4x16/a55.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp | 18 |
1 files changed, 9 insertions, 9 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp index 954e2891fb..c410374357 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp @@ -406,10 +406,10 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "b 122f\n" "31:" // Height 2 "movi v11.4s, #0x0\n" - "movi v12.4s, #0x0\n" - "movi v15.16b, #0x1\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "movi v12.4s, #0x0\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "movi v15.16b, #0x1\n" "mov x9, %x[col_bias]\n" "bic %x[flags], %x[flags], #0x80000000\n" "mov x28, %x[output_ptr]\n" @@ -853,12 +853,12 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "b 122f\n" "61:" // Height 3 "movi v11.4s, #0x0\n" + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "movi v12.4s, #0x0\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" "movi v13.4s, #0x0\n" - "movi v15.16b, #0x1\n" - "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "mov x9, %x[col_bias]\n" - "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "movi v15.16b, #0x1\n" "bic %x[flags], %x[flags], #0x80000000\n" "mov x28, %x[output_ptr]\n" "62:" // Height 3: Column loop @@ -1426,14 +1426,14 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "b 122f\n" "91:" // Height 4 "movi v11.4s, #0x0\n" + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "movi v12.4s, #0x0\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" "movi v13.4s, #0x0\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "movi v14.4s, #0x0\n" - "movi v15.16b, #0x1\n" - "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" "mov x9, %x[col_bias]\n" - "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "movi v15.16b, #0x1\n" "bic %x[flags], %x[flags], #0x80000000\n" "mov x28, %x[output_ptr]\n" "mov x19, #0x4\n" |