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/convolution/winograd/output_transforms | |
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/convolution/winograd/output_transforms')
8 files changed, 38 insertions, 38 deletions
diff --git a/src/core/NEON/kernels/convolution/winograd/output_transforms/a64_fp16_4x4_3x3.cpp b/src/core/NEON/kernels/convolution/winograd/output_transforms/a64_fp16_4x4_3x3.cpp index 295005a2ee..4218b754b4 100644 --- a/src/core/NEON/kernels/convolution/winograd/output_transforms/a64_fp16_4x4_3x3.cpp +++ b/src/core/NEON/kernels/convolution/winograd/output_transforms/a64_fp16_4x4_3x3.cpp @@ -34,13 +34,13 @@ namespace output_transform { void a64_fp16_4x4_3x3( unsigned int n_channels, const __fp16* inptr, - size_t matrix_stride, + const size_t matrix_stride, const __fp16* bptr, __fp16* const output, - size_t output_row_stride, - size_t output_col_stride, - __fp16 output_min, - __fp16 output_max + const size_t output_row_stride, + const size_t output_col_stride, + const __fp16 output_min, + const __fp16 output_max ) { constexpr int output_tile_rows = 4, output_tile_cols = 4; diff --git a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x2_1x7.cpp b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x2_1x7.cpp index 8c6cf9725e..c52df266a5 100644 --- a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x2_1x7.cpp +++ b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x2_1x7.cpp @@ -33,13 +33,13 @@ namespace output_transform { void arm_fp32_1x2_1x7( unsigned int n_channels, const float* inptr, - size_t matrix_stride, + const size_t matrix_stride, const float* bptr, float *outptr, size_t, // No need to stride across rows - size_t output_col_stride, - float output_min, - float output_max + const size_t output_col_stride, + const float output_min, + const float output_max ) { constexpr auto inner_tile_cols = 8u, output_tile_cols = 2u; diff --git a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x4_1x5.cpp b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x4_1x5.cpp index ac05f23221..7d771abeee 100644 --- a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x4_1x5.cpp +++ b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x4_1x5.cpp @@ -33,13 +33,13 @@ namespace output_transform { void arm_fp32_1x4_1x5( unsigned int n_channels, const float* inptr, - size_t matrix_stride, + const size_t matrix_stride, const float* bptr, float *outptr, size_t, // No need to stride across rows - size_t output_col_stride, - float output_min, - float output_max + const size_t output_col_stride, + const float output_min, + const float output_max ) { constexpr auto inner_tile_cols = 8u, output_tile_cols = 4u; diff --git a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x6_1x3.cpp b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x6_1x3.cpp index 154dc6fe1a..513908190a 100644 --- a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x6_1x3.cpp +++ b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_1x6_1x3.cpp @@ -34,13 +34,13 @@ namespace output_transform { void arm_fp32_1x6_1x3( unsigned int n_channels, const float* inptr, - size_t matrix_stride, + const size_t matrix_stride, const float* bptr, float *outptr, size_t, // No need to stride across rows - size_t output_col_stride, - float output_min, - float output_max + const size_t output_col_stride, + const float output_min, + const float output_max ) { constexpr unsigned int inner_tile_cols = 8, output_tile_cols = 6; diff --git a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_3x3.cpp b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_3x3.cpp index 28f042bcbf..4c7376bef8 100644 --- a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_3x3.cpp +++ b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_3x3.cpp @@ -33,13 +33,13 @@ namespace output_transform { void arm_fp32_2x2_3x3( unsigned int n_channels, const float* inptr, - size_t matrix_stride, + const size_t matrix_stride, const float* bptr, float *outptr, - size_t output_row_stride, - size_t output_col_stride, - float output_min, - float output_max + const size_t output_row_stride, + const size_t output_col_stride, + const float output_min, + const float output_max ) { constexpr auto output_tile_rows = 2u, output_tile_cols = 2u; diff --git a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_5x5.cpp b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_5x5.cpp index 8e5ba74ac3..d5649b8a18 100644 --- a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_5x5.cpp +++ b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_2x2_5x5.cpp @@ -33,13 +33,13 @@ namespace output_transform { void arm_fp32_2x2_5x5( unsigned int n_channels, const float* inptr, - size_t matrix_stride, + const size_t matrix_stride, const float* bptr, float *outptr, - size_t output_row_stride, - size_t output_col_stride, - float output_min, - float output_max + const size_t output_row_stride, + const size_t output_col_stride, + const float output_min, + const float output_max ) { constexpr auto output_tile_rows = 2u, output_tile_cols = 2u; diff --git a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_4x4_3x3.cpp b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_4x4_3x3.cpp index 72c43019fa..6a32f67b5d 100644 --- a/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_4x4_3x3.cpp +++ b/src/core/NEON/kernels/convolution/winograd/output_transforms/arm_fp32_4x4_3x3.cpp @@ -33,13 +33,13 @@ namespace output_transform { void arm_fp32_4x4_3x3( unsigned int n_channels, const float* inptr, - size_t matrix_stride, + const size_t matrix_stride, const float* bptr, float *outptr, - size_t output_row_stride, - size_t output_col_stride, - float output_min, - float output_max + const size_t output_row_stride, + const size_t output_col_stride, + const float output_min, + const float output_max ) { constexpr auto output_tile_rows = 4u, output_tile_cols = 4u; diff --git a/src/core/NEON/kernels/convolution/winograd/output_transforms/sme_fp32_mopa_4x4_3x3.cpp b/src/core/NEON/kernels/convolution/winograd/output_transforms/sme_fp32_mopa_4x4_3x3.cpp index 043914d590..8d2b00c1fb 100644 --- a/src/core/NEON/kernels/convolution/winograd/output_transforms/sme_fp32_mopa_4x4_3x3.cpp +++ b/src/core/NEON/kernels/convolution/winograd/output_transforms/sme_fp32_mopa_4x4_3x3.cpp @@ -31,15 +31,15 @@ namespace winograd { namespace output_transform { void sme_fp32_mopa_4x4_3x3( - unsigned int n_channels, + const unsigned int n_channels, const float* inptr, - size_t matrix_stride, + const size_t matrix_stride, const float* bptr, float* const output, - size_t output_row_stride, - size_t output_col_stride, - float output_min, - float output_max + const size_t output_row_stride, + const size_t output_col_stride, + const float output_min, + const float output_max ) { // The below assembler uses the Kronecker product and the "vec trick" to |