diff options
author | ramelg01 <ramy.elgammal@arm.com> | 2022-06-29 16:28:10 +0100 |
---|---|---|
committer | Ramy Elgammal <ramy.elgammal@arm.com> | 2022-07-14 17:56:01 +0000 |
commit | a1f7851e2f776610019db8725c2963c36b0c85eb (patch) | |
tree | eddf90d87594bec2a88d9ad76bf4d03907ff5958 /src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x4_1x5.cpp | |
parent | 4bfc70e31766587c951204c93a127a486e007d0c (diff) | |
download | ComputeLibrary-a1f7851e2f776610019db8725c2963c36b0c85eb.tar.gz |
Integrate new winograd APIs from MLTech
Resolves: COMPMID-5400
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: Ib4428436dd7a6e40d8b2d8a2f8dac1b079154551
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7894
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x4_1x5.cpp')
-rw-r--r-- | src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x4_1x5.cpp | 77 |
1 files changed, 77 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x4_1x5.cpp b/src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x4_1x5.cpp new file mode 100644 index 0000000000..585fb2516b --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/weight_transforms/cpp_fp32_1x4_1x5.cpp @@ -0,0 +1,77 @@ +/* + * Copyright (c) 2022 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include <cstddef> + +namespace arm_conv { +namespace winograd { +namespace weight_transform { + +void cpp_fp32_1x4_1x5( + unsigned int n_channels, + const float *inptr, + size_t, // ld_weight_row + size_t ld_weight_col, + float *outptr, + size_t matrix_stride +) +{ + constexpr auto kernel_cols = 5u, inner_tile_cols = 8u; + + // For each output channel + for (; n_channels; n_channels--) + { + // Matrices used and computed in this kernel + float w[kernel_cols], V[inner_tile_cols]; + + // Read weights + for (auto j = 0u; j < kernel_cols; j++) + { + w[j] = *(inptr + j * ld_weight_col); + } + + // Compute V = w WT + V[0] = (w[0]*-1) / 36; + V[1] = (w[1]*-1 + w[3]*-1 + w[0]*1 + w[2]*1 + w[4]*1) / 48; + V[2] = (w[0]*1 + w[1]*1 + w[2]*1 + w[3]*1 + w[4]*1) / 48; + V[3] = (w[0]*-1 + w[4]*-16 + w[2]*-4 + w[1]*2 + w[3]*8) / 120; + V[4] = (w[0]*-1 + w[4]*-16 + w[3]*-8 + w[2]*-4 + w[1]*-2) / 120; + V[5] = (w[3]*-27 + w[1]*-3 + w[2]*9 + w[4]*81 + w[0]*1) / 720; + V[6] = (w[1]*3 + w[2]*9 + w[3]*27 + w[4]*81 + w[0]*1) / 720; + V[7] = (w[4]*1) / 1; + + // Store the transformed weights + for (auto j = 0u; j < inner_tile_cols; j++) + { + *(outptr + j*matrix_stride) = V[j]; + } + + inptr++; + outptr++; + } +} + +} // namespace weight_transform +} // namespace winograd +} // namespace arm_conv |