diff options
author | Michael Tyler <michael.tyler@arm.com> | 2024-06-04 15:47:37 +0100 |
---|---|---|
committer | Michael Tyler <michael.tyler@arm.com> | 2024-06-25 09:10:13 +0000 |
commit | fc94f4d23abd4bc427b701f54ad85282e9ec7872 (patch) | |
tree | 5e2980599256e2b2f4374e5beb61596fc95c9d5a /src/core/NEON/kernels/arm_gemm/std_transforms_sve.hpp | |
parent | c2237ec4094c7824f8f7e61bc89504d01c5b59ff (diff) | |
download | ComputeLibrary-fc94f4d23abd4bc427b701f54ad85282e9ec7872.tar.gz |
Update CPU kernels and add mixed sign GEMM support
- Add support for mixed sign quantized convolution.
- Add support for mixed sign dequantized GEMM.
- Add SME FP16 GEMV kernel.
- Change SME vector length function to use RDSVL instead of static variable.
- Add GEMM dilation support internally (not exposed yet).
- Remove unused "get_default_activation_values" functions.
- Add SVE fixed format interleaved BF16 DOT kernel.
- Updates and optimizations to assembly kernels.
Resolves COMPMID-6926
Change-Id: I227f502502611d4cc4111c89e30c53ce94079544
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11570
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/std_transforms_sve.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/std_transforms_sve.hpp | 10 |
1 files changed, 5 insertions, 5 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/std_transforms_sve.hpp b/src/core/NEON/kernels/arm_gemm/std_transforms_sve.hpp index c516bfc456..32d597f4af 100644 --- a/src/core/NEON/kernels/arm_gemm/std_transforms_sve.hpp +++ b/src/core/NEON/kernels/arm_gemm/std_transforms_sve.hpp @@ -39,24 +39,24 @@ namespace arm_gemm { * The optional 'block' parameter is for kernels using dot-product type * instructions like UDOT and SDOT. */ -template<typename TOperand, typename TResult, unsigned int height, unsigned int width_vectors, unsigned int block=1, unsigned int mmla=1, bool integrate_sums=false> +template<typename TInput, typename TWeight, typename TResult, unsigned int height, unsigned int width_vectors, unsigned int block=1, unsigned int mmla=1, bool integrate_sums=false> class StdTransformsSVE { public: template<typename TIn> - void PrepareA(TOperand *out, const TIn *in, const int stride, const int y0, + void PrepareA(TInput *out, const TIn *in, const int stride, const int y0, const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) { Interleave<height, block, VLType::None>(out, in, stride, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier); } template<typename TIn> - void PrepareA_indirect(TOperand *out, const TIn * const * const *ptr, size_t stringlen, size_t rounded_stringlen, const int y0, + void PrepareA_indirect(TInput *out, const TIn * const * const *ptr, size_t stringlen, size_t rounded_stringlen, const int y0, const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) { IndirectInterleave<height, block, VLType::None>(out, ptr, stringlen, rounded_stringlen, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier); } template<typename TIn> - void PrepareA_convolution(TOperand *out, const TIn *ptr, size_t stride, const convolver<TIn> &conv, size_t rounded_stringlen, + void PrepareA_convolution(TInput *out, const TIn *ptr, size_t stride, const convolver<TIn> &conv, size_t rounded_stringlen, const int y0, const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) { ConvolutionInterleave<height, block, VLType::None>(out, ptr, stride, conv, rounded_stringlen, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier); } @@ -66,7 +66,7 @@ public: } template<typename TIn> - void PrepareB(TOperand *out, const TIn *in, const int stride, const int x0, + void PrepareB(TWeight *out, const TIn *in, const int stride, const int x0, const int xmax, const int k0, const int kmax, bool transposed) { assert (!transposed); Transform<width_vectors, block, true, VLType::SVE>(out, in, stride, x0, xmax, k0, kmax); |