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_hgemm_8x24 | |
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_hgemm_8x24')
3 files changed, 0 insertions, 12 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp index 29cdd33893..e5728beba8 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp @@ -71,10 +71,6 @@ void a64_hgemm_asimd_8x24_a55r1(const __fp16 *Apanel, const __fp16 *Bpanel, __fp register float16x8_t b2 asm("v6"); __asm __volatile ( - // Enable FP16 instruction support (but only if it's not already on). -#ifndef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - ".arch armv8.2-a+fp16\n" -#endif // Initialize result registers, load initial operands, prime prefetches. "movi v8.8h, #0x0\n" "ldr %d[a0], [%[a_ptr]]\n" diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp index c9c48dd1c0..23b87fa192 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp @@ -66,10 +66,6 @@ void a64_hgemm_asimd_8x24(const __fp16 *Apanel, const __fp16 *Bpanel, __fp16 *Cp register float16x8_t b2a asm("v7"); __asm __volatile ( - // Enable FP16 instruction support (but only if it's not already on). -#ifndef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - ".arch armv8.2-a+fp16\n" -#endif // Initialize result registers, load initial operands, prime prefetches. "movi v8.8h, #0x0\n" "ldr %q[a0], [%[a_ptr]]\n" diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp index a6d2405e7e..b47fa6a2d7 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp @@ -63,10 +63,6 @@ void a64_hgemm_asimd_8x24_x1(const __fp16 *Apanel, const __fp16 *Bpanel, __fp16 register float16x8_t b2 asm("v4"); __asm __volatile ( - // Enable FP16 instruction support (but only if it's not already on). -#ifndef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - ".arch armv8.2-a+fp16\n" -#endif // Initialize result registers, load initial operands, prime prefetches. "movi v8.8h, #0x0\n" "ldr %q[a0], [%[a_ptr]]\n" |