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/transforms/a32_transpose_interleave_8way_32bit.hpp | |
| 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/transforms/a32_transpose_interleave_8way_32bit.hpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/transforms/a32_transpose_interleave_8way_32bit.hpp | 14 |
1 files changed, 7 insertions, 7 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a32_transpose_interleave_8way_32bit.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a32_transpose_interleave_8way_32bit.hpp index 3ce1d328a7..b50c240a3a 100644 --- a/src/core/NEON/kernels/arm_gemm/transforms/a32_transpose_interleave_8way_32bit.hpp +++ b/src/core/NEON/kernels/arm_gemm/transforms/a32_transpose_interleave_8way_32bit.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 Arm Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -30,7 +30,7 @@ // Generic unblocked transposed 8x32-bit sized specialisation template <> template <typename T> -inline void TransformImpl<8, 1, true, 4, 4, VLType::None>::Transform( +void TransformImpl<8, 1, true, 4, 4, VLType::None>::Transform( T* out, const T* const in, const int stride, const int x0, const int xmax, const int k0, const int kmax ) { @@ -45,7 +45,7 @@ inline void TransformImpl<8, 1, true, 4, 4, VLType::None>::Transform( // Generic 16x16-bit sized specialisation template <> template <typename T> -inline void TransformImpl<16, 1, true, 2, 2, VLType::None>::Transform( +void TransformImpl<16, 1, true, 2, 2, VLType::None>::Transform( T* out, const T* const in, const int stride, const int x0, const int xmax, const int k0, const int kmax ) { @@ -59,7 +59,7 @@ inline void TransformImpl<16, 1, true, 2, 2, VLType::None>::Transform( // Specialised 16 x uint16_t version template <> -inline void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x1(const uint16_t *&in0, uint16_t *out) { +void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x1(const uint16_t *&in0, uint16_t *out) { __asm volatile ( "VLD1.32 {d0-d3}, [%[in0]]!\n" "VST1.32 {d0-d3}, [%[out]]\n" @@ -72,7 +72,7 @@ inline void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x1(con } template <> -inline void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x2(const uint16_t *&in0, const uint16_t *&in1, uint16_t *out) { +void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x2(const uint16_t *&in0, const uint16_t *&in1, uint16_t *out) { __asm volatile ( "VLD1.32 {d0-d3}, [%[in0]]!\n" "VST1.32 {d0-d3}, [%[out]]!\n" @@ -90,7 +90,7 @@ inline void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x2(con } template <> -inline void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x4(const uint16_t *&in0, const uint16_t *&in1, const uint16_t *&in2, const uint16_t *&in3, uint16_t *out) { +void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x4(const uint16_t *&in0, const uint16_t *&in1, const uint16_t *&in2, const uint16_t *&in3, uint16_t *out) { __asm __volatile ( "VLD1.32 {d0-d3}, [%[in0]]!\n" "VST1.32 {d0-d3}, [%[out]]!\n" @@ -117,7 +117,7 @@ inline void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x4(con template <> template <> -inline void TransformImpl<16, 1, true, 2, 2, VLType::None>::Transform( +void TransformImpl<16, 1, true, 2, 2, VLType::None>::Transform( uint16_t* out, const uint16_t* const in, const int stride, const int x0, const int xmax, const int k0, const int kmax ) { |