From 72686fa6ee0f04d458ed2274b4d34917628ef14d Mon Sep 17 00:00:00 2001 From: Pablo Tello Date: Mon, 3 Sep 2018 11:40:33 +0100 Subject: COMPMID-1550: Winograd integrate RSH changes. Refactors the transforms to make use of partial specialization. Change-Id: Idff68d22817a00a7ee9eef5351a5a9fd33147540 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/146635 Tested-by: Jenkins Reviewed-by: Georgios Pinitas --- .../winograd/transforms/input_1x8_fp32.cpp | 274 ++++++++++ .../winograd/transforms/input_2x2_3x3_fp32.cpp | 21 +- .../winograd/transforms/input_2x2_5x5_fp32.cpp | 458 ---------------- .../winograd/transforms/input_4x4_3x3_fp32.cpp | 486 ---------------- .../winograd/transforms/input_6_3_fp32.cpp | 226 -------- .../winograd/transforms/input_6x6_fp32.cpp | 608 +++++++++++++++++++++ 6 files changed, 884 insertions(+), 1189 deletions(-) create mode 100644 src/core/NEON/kernels/convolution/winograd/transforms/input_1x8_fp32.cpp delete mode 100644 src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_5x5_fp32.cpp delete mode 100644 src/core/NEON/kernels/convolution/winograd/transforms/input_4x4_3x3_fp32.cpp delete mode 100644 src/core/NEON/kernels/convolution/winograd/transforms/input_6_3_fp32.cpp create mode 100644 src/core/NEON/kernels/convolution/winograd/transforms/input_6x6_fp32.cpp (limited to 'src/core/NEON/kernels/convolution/winograd') diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_1x8_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_1x8_fp32.cpp new file mode 100644 index 0000000000..042d4debbc --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_1x8_fp32.cpp @@ -0,0 +1,274 @@ +/* + * Copyright (c) 2017 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 "arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp" +#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" +#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp" + +namespace +{ + +template +void winograd_input_transform_1x8_fp32_process_tile( + int n_channels, + const float* const input_base, + const int input_row_stride, + const int input_col_stride, + float* const matrix_base, + const int matrix_stride +) +{ + (void) input_row_stride; // No rows over which to stride + constexpr int inner_tile_cols = 8; + constexpr int cells_j = inner_tile_cols - pad_right; + + float *outptr = matrix_base; + + // Get pointers into the input tile + const float *x_ptrs[inner_tile_cols]; + for (int j = pad_left, xj = 0; j < cells_j; j++, xj++) + { + x_ptrs[j] = input_base + xj*input_col_stride; + } + + // Vectors used/computed in this kernel. + float x[inner_tile_cols]; + float U[inner_tile_cols]; + + for (int j = 0; j < inner_tile_cols; j++) + { + x[j] = 0.0f; + } + + // Perform the Winograd input transformation for each channel in the input + // tensor. + int channels_remaining = n_channels; +#ifdef __arm_any__ + for (; channels_remaining >= 4; channels_remaining -= 4) + { + float32x4_t x[inner_tile_cols], U[inner_tile_cols]; + for (int j = 0; j < inner_tile_cols; j++) + { + x[j] = vdupq_n_f32(0.0f); + } + + // Load x + for (int j = pad_left; j < cells_j; j++) + { + x[j] = vld1q_f32(x_ptrs[j]); + x_ptrs[j] += 4; + } + + // Compute U = x . X + U[0] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[2], 49), x[4], -14), x[0], -36); + U[1] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[2], 36), x[3], 13), x[4], -13), x[1], -36), x[5], -1); + U[2] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[5], 1), x[2], 36), x[1], 36), x[4], -13), x[3], -13); + U[3] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[3], 20), x[2], 9), x[5], -2), x[4], -10), x[1], -18); + U[4] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[1], 18), x[2], 9), x[5], 2), x[4], -10), x[3], -20); + U[5] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[3], 15), x[2], 4), x[5], -3), x[4], -5), x[1], -12); + U[6] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[1], 12), x[2], 4), x[5], 3), x[4], -5), x[3], -15); + U[7] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[7], 1), x[3], 49), x[5], -14), x[1], -36); + + // Store the transformed vector + for (int j = 0; j < inner_tile_cols; j++) + { + vst1q_f32(outptr + j*matrix_stride, U[j]); + } + outptr += 4; + } + for (; channels_remaining >= 2; channels_remaining -= 2) + { + float32x2_t x[inner_tile_cols], U[inner_tile_cols]; + for (int j = 0; j < inner_tile_cols; j++) + { + x[j] = vdup_n_f32(0.0f); + } + + // Load x + for (int j = pad_left; j < cells_j; j++) + { + x[j] = vld1_f32(x_ptrs[j]); + x_ptrs[j] += 2; + } + + // Compute U = x . X + U[0] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[2], 49), x[4], -14), x[0], -36); + U[1] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[2], 36), x[3], 13), x[4], -13), x[1], -36), x[5], -1); + U[2] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[5], 1), x[2], 36), x[1], 36), x[4], -13), x[3], -13); + U[3] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[3], 20), x[2], 9), x[5], -2), x[4], -10), x[1], -18); + U[4] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[1], 18), x[2], 9), x[5], 2), x[4], -10), x[3], -20); + U[5] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[3], 15), x[2], 4), x[5], -3), x[4], -5), x[1], -12); + U[6] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[1], 12), x[2], 4), x[5], 3), x[4], -5), x[3], -15); + U[7] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[7], 1), x[3], 49), x[5], -14), x[1], -36); + + // Store the transformed vector + for (int j = 0; j < inner_tile_cols; j++) + { + vst1_f32(outptr + j*matrix_stride, U[j]); + } + outptr += 2; + } +#endif // __arm_any__ + for (; channels_remaining; channels_remaining--) + { + // Load x + for (int j = pad_left; j < cells_j; j++) + { + x[j] = *(x_ptrs[j]++); + } + + // Compute U = x . X + U[0] = x[0]*-36 + x[4]*-14 + x[2]*49 + x[6]*1; + U[1] = x[5]*-1 + x[1]*-36 + x[4]*-13 + x[3]*13 + x[2]*36 + x[6]*1; + U[2] = x[3]*-13 + x[4]*-13 + x[1]*36 + x[2]*36 + x[5]*1 + x[6]*1; + U[3] = x[1]*-18 + x[4]*-10 + x[5]*-2 + x[2]*9 + x[3]*20 + x[6]*1; + U[4] = x[3]*-20 + x[4]*-10 + x[5]*2 + x[2]*9 + x[1]*18 + x[6]*1; + U[5] = x[1]*-12 + x[4]*-5 + x[5]*-3 + x[2]*4 + x[3]*15 + x[6]*1; + U[6] = x[3]*-15 + x[4]*-5 + x[5]*3 + x[2]*4 + x[1]*12 + x[6]*1; + U[7] = x[1]*-36 + x[5]*-14 + x[3]*49 + x[7]*1; + + // Store the transformed vector + for (int j = 0; j < inner_tile_cols; j++) + { + *(outptr + j*matrix_stride) = U[j]; + } + outptr++; + } +} + +} + +namespace winograd +{ +template +using Transform = InputTransformImpl<1, x, 1, 8, float>; + +template <> +const Transform<3>::TileFn + Transform<3>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = +{ + { + { + { + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 0>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 1>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 2>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 3>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 4>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 5>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 6>, + } + }, + { + { + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 0>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 1>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 2>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 3>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 4>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 5>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 6>, + } + } + } +}; + +template <> +const Transform<5>::TileFn + Transform<5>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = +{ + { + { + { + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 0>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 1>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 2>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 3>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 4>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 5>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 6>, + } + }, + { + { + winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 0>, + winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 1>, + winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 2>, + winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 3>, + winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 4>, + winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 5>, + winograd_input_transform_1x8_fp32_process_tile<0, 2, 0, 6>, + } + } + } +}; + +template <> +const Transform<7>::TileFn + Transform<7>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = +{ + { + { + { + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 0>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 1>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 2>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 3>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 4>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 5>, + winograd_input_transform_1x8_fp32_process_tile<0, 0, 0, 6>, + } + }, + { + { + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 0>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 1>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 2>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 3>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 4>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 5>, + winograd_input_transform_1x8_fp32_process_tile<0, 1, 0, 6>, + } + }, + { + { + winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 0>, + winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 1>, + winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 2>, + winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 3>, + winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 4>, + winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 5>, + winograd_input_transform_1x8_fp32_process_tile<0, 3, 0, 6>, + } + } + } +}; + +template class InputTransform<1, 3, 1, 8, float>; +template class InputTransform<3, 1, 8, 1, float>; +template class InputTransform<1, 5, 1, 8, float>; +template class InputTransform<5, 1, 8, 1, float>; +template class InputTransform<1, 7, 1, 8, float>; +template class InputTransform<7, 1, 8, 1, float>; +} // namespace winograd diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp index 97b2695d69..a9d5d52d15 100644 --- a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp +++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp @@ -29,22 +29,7 @@ namespace winograd { -using Transform = WinogradGEMM<2, 2, 3, 3>::InputTransform; - -/****************************************************************************** - * Cost methods for the input transform. - * ===================================== - */ -template <> -template <> -int Transform::ops_performed(const Tensor4DShape &input_shape) -{ - // NOTE: Cost in FLOPs rather than instructions or uops. - const int tile_M = iceildiv(input_shape.n_rows, inner_tile_rows); - const int tile_N = iceildiv(input_shape.n_cols, inner_tile_cols); - return 16 * 16 * tile_M * tile_N * input_shape.n_channels; -} -/*****************************************************************************/ +using Transform = InputTransformImpl<3, 3, 4, 4, float>; /***************************************************************************** * F(2x2, 3x3) implies the use of a 4x4 input tile. Such tiles can require a @@ -100,7 +85,6 @@ int Transform::ops_performed(const Tensor4DShape &input_shape) * Padding right in {0, 1, 2} */ template <> -template <> template void Transform::process_tile( int n_channels, @@ -327,7 +311,6 @@ void Transform::process_tile( } } -template <> template <> const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = { @@ -405,5 +388,5 @@ const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom] } }; -template struct WinogradGEMM<2, 2, 3, 3>::InputTransform; +template class InputTransform<3, 3, 4, 4, float>; } // namespace winograd diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_5x5_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_5x5_fp32.cpp deleted file mode 100644 index 30c9463bb8..0000000000 --- a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_5x5_fp32.cpp +++ /dev/null @@ -1,458 +0,0 @@ -/* - * Copyright (c) 2017 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 "arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp" -#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" -#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp" - -namespace winograd -{ - -using Transform = WinogradGEMM<2, 2, 5, 5>::InputTransform; - -template <> -template <> -int Transform::ops_performed(const Tensor4DShape &input_shape) -{ - (void) input_shape; - return 0; // TODO -} - -/***************************************************************************** -* F(2x2, 5x5) implies the use of a 6x6 input tile. -* -* Build an array of the specialised methods that deal with each of the -* different padding combinations which may be required. These padding -* constraints are the space: -* -* Padding top in {0, 2} -* Padding left in {0, 2} -* Padding bottom in {0, 1, 2, 3, 4} -* Padding right in {0, 1, 2, 3, 4} -*/ -template <> -template <> -template -void Transform::process_tile( - int n_channels, - const float* const input_base, - const int input_row_stride, - const int input_col_stride, - float* const matrix_base, - const int matrix_stride -) -{ - constexpr int cells_i = 6 - pad_bottom; - constexpr int cells_j = 6 - pad_right; - - float *outptr = matrix_base; - - // Get pointers into the input tile - const float *x_ptrs[6][6]; - for (int i = pad_top, xi = 0; i < cells_i; i++, xi++) - { - // Get a pointer into the row - const float* const row_ptr = input_base + xi*input_row_stride; - - for (int j = pad_left, xj = 0; j < cells_j; j++, xj++) - { - x_ptrs[i][j] = row_ptr + xj*input_col_stride; - } - } - - // Matrices used/computed in this kernel. - float x[6][6], XTx[6][6], U[6][6]; - for (int i = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++) - { - x[i][j] = XTx[i][j] = 0.0f; - } - } - - // Perform the Winograd input transformation for each channel in the input - // tensor. - int channels_remaining = n_channels; -#ifdef __aarch64__ - for (; channels_remaining >= 4; channels_remaining -= 4) - { - // Matrices used/computed in this kernel - float32x4_t x[6][6], XTx[6][6], U[6][6]; - for (int i = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++) - { - x[i][j] = vdupq_n_f32(0.0f); - XTx[i][j] = vdupq_n_f32(0.0f); - } - } - - // Read a 6x6 tile in the Winograd domain - for (int i = pad_top; i < cells_i; i++) - { - for (int j = pad_left; j < cells_j; j++) - { - x[i][j] = vld1q_f32(x_ptrs[i][j]); - x_ptrs[i][j] += 4; - } - } - - // Compute XT . x - for (int j = pad_left; j < cells_j; j++) - { - // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j]; - XTx[0][j] = vmlsq_n_f32(vmlaq_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f); - - // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j]; - XTx[1][j] = vmlsq_n_f32(vaddq_f32(x[3][j], x[4][j]), vaddq_f32(x[1][j], x[2][j]), 4.0f); - - // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j]; - XTx[2][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[3][j]), vsubq_f32(x[1][j], x[2][j]), 4.0f); - - // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j]; - XTx[3][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[3][j], x[1][j]), 2.0f); - - // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j]; - XTx[4][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[1][j], x[3][j]), 2.0f); - - // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j]; - XTx[5][j] = vmlsq_n_f32(vmlaq_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f); - } - - // Compute U = XT . x . X - for (int i = 0; i < 6; i++) - { - // U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4]; - U[i][0] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f); - - // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4]; - U[i][1] = vmlsq_n_f32(vaddq_f32(XTx[i][3], XTx[i][4]), vaddq_f32(XTx[i][1], XTx[i][2]), 4.0f); - - // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4]; - U[i][2] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][3]), vsubq_f32(XTx[i][1], XTx[i][2]), 4.0f); - - // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4]; - U[i][3] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][3], XTx[i][1]), 2.0f); - - // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4]; - U[i][4] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][1], XTx[i][3]), 2.0f); - - // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5]; - U[i][5] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f); - } - - // Store the transformed matrix - for (int i = 0, m = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++, m++) - { - vst1q_f32(outptr + m*matrix_stride, U[i][j]); - } - } - outptr += 4; - } -#endif // __aarch64__ -#ifdef __arm_any__ - for (; channels_remaining >= 2; channels_remaining -= 2) - { - // Matrices used/computed in this kernel - float32x2_t x[6][6], XTx[6][6], U[6][6]; - for (int i = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++) - { - x[i][j] = vdup_n_f32(0.0f); - XTx[i][j] = vdup_n_f32(0.0f); - } - } - - // Read a 6x6 tile in the Winograd domain - for (int i = pad_top; i < cells_i; i++) - { - for (int j = pad_left; j < cells_j; j++) - { - x[i][j] = vld1_f32(x_ptrs[i][j]); - x_ptrs[i][j] += 2; - } - } - - // Compute XT . x - for (int j = pad_left; j < cells_j; j++) - { - // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j]; - XTx[0][j] = vmls_n_f32(vmla_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f); - - // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j]; - XTx[1][j] = vmls_n_f32(vadd_f32(x[3][j], x[4][j]), vadd_f32(x[1][j], x[2][j]), 4.0f); - - // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j]; - XTx[2][j] = vmla_n_f32(vsub_f32(x[4][j], x[3][j]), vsub_f32(x[1][j], x[2][j]), 4.0f); - - // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j]; - XTx[3][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[3][j], x[1][j]), 2.0f); - - // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j]; - XTx[4][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[1][j], x[3][j]), 2.0f); - - // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j]; - XTx[5][j] = vmls_n_f32(vmla_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f); - } - - // Compute U = XT . x . X - for (int i = 0; i < 6; i++) - { - // U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4]; - U[i][0] = vmls_n_f32(vmla_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f); - - // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4]; - U[i][1] = vmls_n_f32(vadd_f32(XTx[i][3], XTx[i][4]), vadd_f32(XTx[i][1], XTx[i][2]), 4.0f); - - // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4]; - U[i][2] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][3]), vsub_f32(XTx[i][1], XTx[i][2]), 4.0f); - - // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4]; - U[i][3] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][3], XTx[i][1]), 2.0f); - - // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4]; - U[i][4] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][1], XTx[i][3]), 2.0f); - - // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5]; - U[i][5] = vmls_n_f32(vmla_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f); - } - - // Store the transformed matrix - for (int i = 0, m = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++, m++) - { - vst1_f32(outptr + m*matrix_stride, U[i][j]); - } - } - outptr += 2; - } -#endif // __arm_any__ - for (; channels_remaining; channels_remaining--) - { - // Load x - for (int i = pad_top; i < cells_i; i++) - { - for (int j = pad_left; j < cells_j; j++) - { - x[i][j] = *(x_ptrs[i][j]++); - } - } - - // Compute XT . x - for (int j = pad_left; j < cells_j; j++) - { - XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j]; - XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j]; - XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j]; - XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j]; - XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j]; - XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j]; - } - - // Compute U = XT . x . X - for (int i = 0; i < 6; i++) - { - U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4]; - U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4]; - U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4]; - U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4]; - U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4]; - U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5]; - } - - // Store the transformed matrix - for (int i = 0, m = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++, m++) - { - *(outptr + m*matrix_stride) = U[i][j]; - } - } - outptr++; - } -} - -template <> -template <> -const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = -{ - { - { - { - Transform::template process_tile<0, 0, 0, 0>, // No padding - Transform::template process_tile<0, 0, 0, 1>, // Right - Transform::template process_tile<0, 0, 0, 2>, // " " - Transform::template process_tile<0, 0, 0, 3>, // " " - Transform::template process_tile<0, 0, 0, 4>, // " " - }, - { - Transform::template process_tile<0, 0, 1, 0>, // Bottom - Transform::template process_tile<0, 0, 1, 1>, // Bottom right - Transform::template process_tile<0, 0, 1, 2>, // " " - Transform::template process_tile<0, 0, 1, 3>, // " " - Transform::template process_tile<0, 0, 1, 4>, // " " - }, - { - Transform::template process_tile<0, 0, 2, 0>, // Bottom - Transform::template process_tile<0, 0, 2, 1>, // Bottom right - Transform::template process_tile<0, 0, 2, 2>, // " " - Transform::template process_tile<0, 0, 2, 3>, // " " - Transform::template process_tile<0, 0, 2, 4>, // " " - }, - { - Transform::template process_tile<0, 0, 3, 0>, // Bottom - Transform::template process_tile<0, 0, 3, 1>, // Bottom right - Transform::template process_tile<0, 0, 3, 2>, // " " - Transform::template process_tile<0, 0, 3, 3>, // " " - Transform::template process_tile<0, 0, 3, 4>, // " " - }, - { - Transform::template process_tile<0, 0, 4, 0>, // Bottom - Transform::template process_tile<0, 0, 4, 1>, // Bottom right - Transform::template process_tile<0, 0, 4, 2>, // " " - Transform::template process_tile<0, 0, 4, 3>, // " " - Transform::template process_tile<0, 0, 4, 4>, // " " - } - }, - { - { - Transform::template process_tile<0, 2, 0, 0>, // Left - Transform::template process_tile<0, 2, 0, 1>, - Transform::template process_tile<0, 2, 0, 2>, - Transform::template process_tile<0, 2, 0, 3>, - Transform::template process_tile<0, 2, 0, 4>, - }, - { - Transform::template process_tile<0, 2, 1, 0>, // Bottom left - Transform::template process_tile<0, 2, 1, 1>, - Transform::template process_tile<0, 2, 1, 2>, - Transform::template process_tile<0, 2, 1, 3>, - Transform::template process_tile<0, 2, 1, 4>, - }, - { - Transform::template process_tile<0, 2, 2, 0>, // " " - Transform::template process_tile<0, 2, 2, 1>, - Transform::template process_tile<0, 2, 2, 2>, - Transform::template process_tile<0, 2, 2, 3>, - Transform::template process_tile<0, 2, 2, 4>, - }, - { - Transform::template process_tile<0, 2, 3, 0>, // " " - Transform::template process_tile<0, 2, 3, 1>, - Transform::template process_tile<0, 2, 3, 2>, - Transform::template process_tile<0, 2, 3, 3>, - Transform::template process_tile<0, 2, 3, 4>, - }, - { - Transform::template process_tile<0, 2, 4, 0>, // " " - Transform::template process_tile<0, 2, 4, 1>, - Transform::template process_tile<0, 2, 4, 2>, - Transform::template process_tile<0, 2, 4, 3>, - Transform::template process_tile<0, 2, 4, 4>, - } - } - }, - { - { - { - Transform::template process_tile<2, 0, 0, 0>, // Top - Transform::template process_tile<2, 0, 0, 1>, // Top right - Transform::template process_tile<2, 0, 0, 2>, // " " - Transform::template process_tile<2, 0, 0, 3>, // " " - Transform::template process_tile<2, 0, 0, 4>, // " " - }, - { - Transform::template process_tile<2, 0, 1, 0>, - Transform::template process_tile<2, 0, 1, 1>, - Transform::template process_tile<2, 0, 1, 2>, - Transform::template process_tile<2, 0, 1, 3>, - Transform::template process_tile<2, 0, 1, 4>, - }, - { - Transform::template process_tile<2, 0, 2, 0>, - Transform::template process_tile<2, 0, 2, 1>, - Transform::template process_tile<2, 0, 2, 2>, - Transform::template process_tile<2, 0, 2, 3>, - Transform::template process_tile<2, 0, 2, 4>, - }, - { - Transform::template process_tile<2, 0, 3, 0>, - Transform::template process_tile<2, 0, 3, 1>, - Transform::template process_tile<2, 0, 3, 2>, - Transform::template process_tile<2, 0, 3, 3>, - Transform::template process_tile<2, 0, 3, 4>, - }, - { - Transform::template process_tile<2, 0, 4, 0>, - Transform::template process_tile<2, 0, 4, 1>, - Transform::template process_tile<2, 0, 4, 2>, - Transform::template process_tile<2, 0, 4, 3>, - Transform::template process_tile<2, 0, 4, 4>, - }, - }, - { - { - Transform::template process_tile<2, 2, 0, 0>, // Top left - Transform::template process_tile<2, 2, 0, 1>, - Transform::template process_tile<2, 2, 0, 2>, - Transform::template process_tile<2, 2, 0, 3>, - Transform::template process_tile<2, 2, 0, 4>, - }, - { - Transform::template process_tile<2, 2, 1, 0>, - Transform::template process_tile<2, 2, 1, 1>, - Transform::template process_tile<2, 2, 1, 2>, - Transform::template process_tile<2, 2, 1, 3>, - Transform::template process_tile<2, 2, 1, 4>, - }, - { - Transform::template process_tile<2, 2, 2, 0>, - Transform::template process_tile<2, 2, 2, 1>, - Transform::template process_tile<2, 2, 2, 2>, - Transform::template process_tile<2, 2, 2, 3>, - Transform::template process_tile<2, 2, 2, 4>, - }, - { - Transform::template process_tile<2, 2, 3, 0>, - Transform::template process_tile<2, 2, 3, 1>, - Transform::template process_tile<2, 2, 3, 2>, - Transform::template process_tile<2, 2, 3, 3>, - Transform::template process_tile<2, 2, 3, 4>, - }, - { - Transform::template process_tile<2, 2, 4, 0>, - Transform::template process_tile<2, 2, 4, 1>, - Transform::template process_tile<2, 2, 4, 2>, - Transform::template process_tile<2, 2, 4, 3>, - Transform::template process_tile<2, 2, 4, 4>, - } - } - } -}; - -template struct WinogradGEMM<2, 2, 5, 5>::InputTransform; -} // namespace winograd diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_4x4_3x3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_4x4_3x3_fp32.cpp deleted file mode 100644 index 7f93187132..0000000000 --- a/src/core/NEON/kernels/convolution/winograd/transforms/input_4x4_3x3_fp32.cpp +++ /dev/null @@ -1,486 +0,0 @@ -/* - * Copyright (c) 2017 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 "arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp" -#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" -#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp" - -namespace winograd -{ - -using Transform = WinogradGEMM<4, 4, 3, 3>::InputTransform; - -template <> -template <> -int Transform::ops_performed(const Tensor4DShape &input_shape) -{ - // NOTE: Cost in FLOPs rather than instructions or uops. - const int tile_M = iceildiv(input_shape.n_rows, inner_tile_rows); - const int tile_N = iceildiv(input_shape.n_cols, inner_tile_cols); - return 12 * 24 * tile_M * tile_N * input_shape.n_channels; -} - -/* F(4x4, 3x3) implies the use of a 6x6 input tile. Such tiles can require a -* variety of padding types. For example, tiles at the top and left of an -* image can require one row or column of padding on their top and left sides -* if the padding type is SAME (where X represents a padded value): -* -* ___________ ___________ -* |X X X X X X| |X X X X X X| -* |X | | | -* |X | | | -* |X | | | -* |X | | | -* |X__________| |___________| -* ___________ -* |X | -* |X | -* |X | -* |X | -* |X | -* |X__________| -* -* For tiles near the right or bottom of the image it is more complicated. -* Such tiles might require padding by 0, 1, 2 or 3 rows or columns if the -* padding type is VALID or 1, 2, 3 or 4 rows or columns if the padding -* type is SAME. -* -* Build an array of the specialised methods that deal with each of the -* different padding combinations which may be required. These padding -* constraints are the space: -* -* Padding top in {0, 1} -* Padding left in {0, 1} -* Padding bottom in {0, 1, 2, 3, 4} -* Padding right in {0, 1, 2, 3, 4} -*/ -template <> -template <> -template -void Transform::process_tile( - int n_channels, - const float* const input_base, - const int input_row_stride, - const int input_col_stride, - float* const matrix_base, - const int matrix_stride -) -{ - constexpr int cells_i = 6 - pad_bottom; - constexpr int cells_j = 6 - pad_right; - - float *outptr = matrix_base; - - // Get pointers into the input tile - const float *x_ptrs[6][6]; - for (int i = pad_top, xi = 0; i < cells_i; i++, xi++) - { - // Get a pointer into the row - const float* const row_ptr = input_base + xi*input_row_stride; - - for (int j = pad_left, xj = 0; j < cells_j; j++, xj++) - { - x_ptrs[i][j] = row_ptr + xj*input_col_stride; - } - } - - // Matrices used/computed in this kernel. - float x[6][6], XTx[6][6], U[6][6]; - for (int i = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++) - { - x[i][j] = XTx[i][j] = 0.0f; - } - } - - // Perform the Winograd input transformation for each channel in the input - // tensor. - int channels_remaining = n_channels; -#ifdef __aarch64__ - for (; channels_remaining >= 4; channels_remaining -= 4) - { - // Matrices used/computed in this kernel - float32x4_t x[6][6], XTx[6][6], U[6][6]; - for (int i = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++) - { - x[i][j] = vdupq_n_f32(0.0f); - XTx[i][j] = vdupq_n_f32(0.0f); - } - } - - // Read a 6x6 tile in the Winograd domain - for (int i = pad_top; i < cells_i; i++) - { - for (int j = pad_left; j < cells_j; j++) - { - x[i][j] = vld1q_f32(x_ptrs[i][j]); - x_ptrs[i][j] += 4; - } - } - - // Compute XT . x - for (int j = pad_left; j < cells_j; j++) - { - // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j]; - XTx[0][j] = vmlsq_n_f32(vmlaq_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f); - - // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j]; - XTx[1][j] = vmlsq_n_f32(vaddq_f32(x[3][j], x[4][j]), vaddq_f32(x[1][j], x[2][j]), 4.0f); - - // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j]; - XTx[2][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[3][j]), vsubq_f32(x[1][j], x[2][j]), 4.0f); - - // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j]; - XTx[3][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[3][j], x[1][j]), 2.0f); - - // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j]; - XTx[4][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[1][j], x[3][j]), 2.0f); - - // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j]; - XTx[5][j] = vmlsq_n_f32(vmlaq_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f); - } - - // Compute U = XT . x . X - for (int i = 0; i < 6; i++) - { - // U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4]; - U[i][0] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f); - - // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4]; - U[i][1] = vmlsq_n_f32(vaddq_f32(XTx[i][3], XTx[i][4]), vaddq_f32(XTx[i][1], XTx[i][2]), 4.0f); - - // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4]; - U[i][2] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][3]), vsubq_f32(XTx[i][1], XTx[i][2]), 4.0f); - - // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4]; - U[i][3] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][3], XTx[i][1]), 2.0f); - - // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4]; - U[i][4] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][1], XTx[i][3]), 2.0f); - - // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5]; - U[i][5] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f); - } - - // Store the transformed matrix - for (int i = 0, m = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++, m++) - { - vst1q_f32(outptr + m*matrix_stride, U[i][j]); - } - } - outptr += 4; - } -#endif // __aarch64__ -#ifdef __arm_any__ - for (; channels_remaining >= 2; channels_remaining -= 2) - { - // Matrices used/computed in this kernel - float32x2_t x[6][6], XTx[6][6], U[6][6]; - for (int i = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++) - { - x[i][j] = vdup_n_f32(0.0f); - XTx[i][j] = vdup_n_f32(0.0f); - } - } - - // Read a 6x6 tile in the Winograd domain - for (int i = pad_top; i < cells_i; i++) - { - for (int j = pad_left; j < cells_j; j++) - { - x[i][j] = vld1_f32(x_ptrs[i][j]); - x_ptrs[i][j] += 2; - } - } - - // Compute XT . x - for (int j = pad_left; j < cells_j; j++) - { - // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j]; - XTx[0][j] = vmls_n_f32(vmla_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f); - - // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j]; - XTx[1][j] = vmls_n_f32(vadd_f32(x[3][j], x[4][j]), vadd_f32(x[1][j], x[2][j]), 4.0f); - - // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j]; - XTx[2][j] = vmla_n_f32(vsub_f32(x[4][j], x[3][j]), vsub_f32(x[1][j], x[2][j]), 4.0f); - - // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j]; - XTx[3][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[3][j], x[1][j]), 2.0f); - - // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j]; - XTx[4][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[1][j], x[3][j]), 2.0f); - - // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j]; - XTx[5][j] = vmls_n_f32(vmla_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f); - } - - // Compute U = XT . x . X - for (int i = 0; i < 6; i++) - { - // U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4]; - U[i][0] = vmls_n_f32(vmla_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f); - - // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4]; - U[i][1] = vmls_n_f32(vadd_f32(XTx[i][3], XTx[i][4]), vadd_f32(XTx[i][1], XTx[i][2]), 4.0f); - - // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4]; - U[i][2] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][3]), vsub_f32(XTx[i][1], XTx[i][2]), 4.0f); - - // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4]; - U[i][3] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][3], XTx[i][1]), 2.0f); - - // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4]; - U[i][4] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][1], XTx[i][3]), 2.0f); - - // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5]; - U[i][5] = vmls_n_f32(vmla_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f); - } - - // Store the transformed matrix - for (int i = 0, m = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++, m++) - { - vst1_f32(outptr + m*matrix_stride, U[i][j]); - } - } - outptr += 2; - } -#endif // __arm_any__ - for (; channels_remaining; channels_remaining--) - { - // Load x - for (int i = pad_top; i < cells_i; i++) - { - for (int j = pad_left; j < cells_j; j++) - { - x[i][j] = *(x_ptrs[i][j]++); - } - } - - // Compute XT . x - for (int j = pad_left; j < cells_j; j++) - { - XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j]; - XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j]; - XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j]; - XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j]; - XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j]; - XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j]; - } - - // Compute U = XT . x . X - for (int i = 0; i < 6; i++) - { - U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4]; - U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4]; - U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4]; - U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4]; - U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4]; - U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5]; - } - - // Store the transformed matrix - for (int i = 0, m = 0; i < 6; i++) - { - for (int j = 0; j < 6; j++, m++) - { - *(outptr + m*matrix_stride) = U[i][j]; - } - } - outptr++; - } -} - -/* In the below, unusual or especially small tiles are routed via the slow - * path whereas common or large tiles are routed through a faster path. - */ -template <> -template <> -const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = -{ - { - { - { - Transform::template process_tile<0, 0, 0, 0>, // No padding - Transform::template process_tile<0, 0, 0, 1>, // Right - Transform::template process_tile<0, 0, 0, 2>, // " " - Transform::template process_tile<0, 0, 0, 3>, // " " - Transform::template process_tile<0, 0, 0, 4>, // " " - }, - { - Transform::template process_tile<0, 0, 1, 0>, // Bottom - Transform::template process_tile<0, 0, 1, 1>, // Bottom right - Transform::template process_tile<0, 0, 1, 2>, // " " - Transform::template process_tile<0, 0, 1, 3>, // " " - Transform::template process_tile<0, 0, 1, 4>, // " " - }, - { - Transform::template process_tile<0, 0, 2, 0>, // Bottom - Transform::template process_tile<0, 0, 2, 1>, // Bottom right - Transform::template process_tile<0, 0, 2, 2>, // " " - Transform::template process_tile<0, 0, 2, 3>, // " " - Transform::template process_tile<0, 0, 2, 4>, // " " - }, - { - Transform::template process_tile<0, 0, 3, 0>, // Bottom - Transform::template process_tile<0, 0, 3, 1>, // Bottom right - Transform::template process_tile<0, 0, 3, 2>, // " " - Transform::template process_tile<0, 0, 3, 3>, // " " - Transform::template process_tile<0, 0, 3, 4>, // " " - }, - { - Transform::template process_tile<0, 0, 4, 0>, // Bottom - Transform::template process_tile<0, 0, 4, 1>, // Bottom right - Transform::template process_tile<0, 0, 4, 2>, // " " - Transform::template process_tile<0, 0, 4, 3>, // " " - Transform::template process_tile<0, 0, 4, 4>, // " " - } - }, - { - { - Transform::template process_tile<0, 1, 0, 0>, // Left - Transform::template process_tile<0, 1, 0, 1>, - Transform::template process_tile<0, 1, 0, 2>, - Transform::template process_tile<0, 1, 0, 3>, - Transform::template process_tile<0, 1, 0, 4>, - }, - { - Transform::template process_tile<0, 1, 1, 0>, // Bottom left - Transform::template process_tile<0, 1, 1, 1>, - Transform::template process_tile<0, 1, 1, 2>, - Transform::template process_tile<0, 1, 1, 3>, - Transform::template process_tile<0, 1, 1, 4>, - }, - { - Transform::template process_tile<0, 1, 2, 0>, // " " - Transform::template process_tile<0, 1, 2, 1>, - Transform::template process_tile<0, 1, 2, 2>, - Transform::template process_tile<0, 1, 2, 3>, - Transform::template process_tile<0, 1, 2, 4>, - }, - { - Transform::template process_tile<0, 1, 3, 0>, // " " - Transform::template process_tile<0, 1, 3, 1>, - Transform::template process_tile<0, 1, 3, 2>, - Transform::template process_tile<0, 1, 3, 3>, - Transform::template process_tile<0, 1, 3, 4>, - }, - { - Transform::template process_tile<0, 1, 4, 0>, // " " - Transform::template process_tile<0, 1, 4, 1>, - Transform::template process_tile<0, 1, 4, 2>, - Transform::template process_tile<0, 1, 4, 3>, - Transform::template process_tile<0, 1, 4, 4>, - } - } - }, - { - { - { - Transform::template process_tile<1, 0, 0, 0>, // Top - Transform::template process_tile<1, 0, 0, 1>, // Top right - Transform::template process_tile<1, 0, 0, 2>, // " " - Transform::template process_tile<1, 0, 0, 3>, // " " - Transform::template process_tile<1, 0, 0, 4>, // " " - }, - { - Transform::template process_tile<1, 0, 1, 0>, - Transform::template process_tile<1, 0, 1, 1>, - Transform::template process_tile<1, 0, 1, 2>, - Transform::template process_tile<1, 0, 1, 3>, - Transform::template process_tile<1, 0, 1, 4>, - }, - { - Transform::template process_tile<1, 0, 2, 0>, - Transform::template process_tile<1, 0, 2, 1>, - Transform::template process_tile<1, 0, 2, 2>, - Transform::template process_tile<1, 0, 2, 3>, - Transform::template process_tile<1, 0, 2, 4>, - }, - { - Transform::template process_tile<1, 0, 3, 0>, - Transform::template process_tile<1, 0, 3, 1>, - Transform::template process_tile<1, 0, 3, 2>, - Transform::template process_tile<1, 0, 3, 3>, - Transform::template process_tile<1, 0, 3, 4>, - }, - { - Transform::template process_tile<1, 0, 4, 0>, - Transform::template process_tile<1, 0, 4, 1>, - Transform::template process_tile<1, 0, 4, 2>, - Transform::template process_tile<1, 0, 4, 3>, - Transform::template process_tile<1, 0, 4, 4>, - }, - }, - { - { - Transform::template process_tile<1, 1, 0, 0>, // Top left - Transform::template process_tile<1, 1, 0, 1>, - Transform::template process_tile<1, 1, 0, 2>, - Transform::template process_tile<1, 1, 0, 3>, - Transform::template process_tile<1, 1, 0, 4>, - }, - { - Transform::template process_tile<1, 1, 1, 0>, - Transform::template process_tile<1, 1, 1, 1>, - Transform::template process_tile<1, 1, 1, 2>, - Transform::template process_tile<1, 1, 1, 3>, - Transform::template process_tile<1, 1, 1, 4>, - }, - { - Transform::template process_tile<1, 1, 2, 0>, - Transform::template process_tile<1, 1, 2, 1>, - Transform::template process_tile<1, 1, 2, 2>, - Transform::template process_tile<1, 1, 2, 3>, - Transform::template process_tile<1, 1, 2, 4>, - }, - { - Transform::template process_tile<1, 1, 3, 0>, - Transform::template process_tile<1, 1, 3, 1>, - Transform::template process_tile<1, 1, 3, 2>, - Transform::template process_tile<1, 1, 3, 3>, - Transform::template process_tile<1, 1, 3, 4>, - }, - { - Transform::template process_tile<1, 1, 4, 0>, - Transform::template process_tile<1, 1, 4, 1>, - Transform::template process_tile<1, 1, 4, 2>, - Transform::template process_tile<1, 1, 4, 3>, - Transform::template process_tile<1, 1, 4, 4>, - } - } - } -}; - -template struct WinogradGEMM<4, 4, 3, 3>::InputTransform; -} // namespace winograd diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_6_3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_6_3_fp32.cpp deleted file mode 100644 index 67e46499cd..0000000000 --- a/src/core/NEON/kernels/convolution/winograd/transforms/input_6_3_fp32.cpp +++ /dev/null @@ -1,226 +0,0 @@ -/* - * Copyright (c) 2017 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 "arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp" -#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" -#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp" - - -namespace winograd -{ - -using Transform = WinogradGEMM<1, 6, 1, 3>::InputTransform; - -template <> -template <> -int Transform::ops_performed(const Tensor4DShape &input_shape) -{ - (void) input_shape; - return 0; // TODO -} - -template <> -template <> -template -void Transform::process_tile( - int n_channels, - const float* const input_base, - const int input_row_stride, - const int input_col_stride, - float* const matrix_base, - const int matrix_stride -) -{ - (void) input_row_stride; // No rows over which to stride - constexpr int inner_tile_j = 8; - constexpr int cells_j = inner_tile_j - pad_right; - - float *outptr = matrix_base; - - // Get pointers into the input tile - const float *x_ptrs[inner_tile_j]; - for (int j = pad_left, xj = 0; j < cells_j; j++, xj++) - { - x_ptrs[j] = input_base + xj*input_col_stride; - } - - // Vectors used/computed in this kernel. - float x[inner_tile_j]; - float U[inner_tile_j]; - - for (int j = 0; j < inner_tile_j; j++) - { - x[j] = 0.0f; - } - - // Perform the Winograd input transformation for each channel in the input - // tensor. - int channels_remaining = n_channels; -#ifdef __arm_any__ - for (; channels_remaining >= 4; channels_remaining -= 4) - { - float32x4_t x[inner_tile_j], U[inner_tile_j]; - for (int j = 0; j < inner_tile_cols; j++) - { - x[j] = vdupq_n_f32(0.0f); - } - - // Load x - for (int j = pad_left; j < cells_j; j++) - { - x[j] = vld1q_f32(x_ptrs[j]); - x_ptrs[j] += 4; - } - - // Compute U = x . X - U[0] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[2], 49), x[4], -14), x[0], -36); - U[1] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[2], 36), x[3], 13), x[4], -13), x[1], -36), x[5], -1); - U[2] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[5], 1), x[2], 36), x[1], 36), x[4], -13), x[3], -13); - U[3] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[3], 20), x[2], 9), x[5], -2), x[4], -10), x[1], -18); - U[4] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[1], 18), x[2], 9), x[5], 2), x[4], -10), x[3], -20); - U[5] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[3], 15), x[2], 4), x[5], -3), x[4], -5), x[1], -12); - U[6] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[6], 1), x[1], 12), x[2], 4), x[5], 3), x[4], -5), x[3], -15); - U[7] = vmlaq_n_f32(vmlaq_n_f32(vmlaq_n_f32(vmulq_n_f32(x[7], 1), x[3], 49), x[5], -14), x[1], -36); - - // Store the transformed vector - for (int j = 0; j < inner_tile_j; j++) - { - vst1q_f32(outptr + j*matrix_stride, U[j]); - } - outptr += 4; - } - - for (; channels_remaining >= 2; channels_remaining -= 2) - { - float32x2_t x[inner_tile_j], U[inner_tile_j]; - for (int j = 0; j < inner_tile_cols; j++) - { - x[j] = vdup_n_f32(0.0f); - } - - // Load x - for (int j = pad_left; j < cells_j; j++) - { - x[j] = vld1_f32(x_ptrs[j]); - x_ptrs[j] += 2; - } - - // Compute U = x . X - U[0] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[2], 49), x[4], -14), x[0], -36); - U[1] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[2], 36), x[3], 13), x[4], -13), x[1], -36), x[5], -1); - U[2] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[5], 1), x[2], 36), x[1], 36), x[4], -13), x[3], -13); - U[3] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[3], 20), x[2], 9), x[5], -2), x[4], -10), x[1], -18); - U[4] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[1], 18), x[2], 9), x[5], 2), x[4], -10), x[3], -20); - U[5] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[3], 15), x[2], 4), x[5], -3), x[4], -5), x[1], -12); - U[6] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[6], 1), x[1], 12), x[2], 4), x[5], 3), x[4], -5), x[3], -15); - U[7] = vmla_n_f32(vmla_n_f32(vmla_n_f32(vmul_n_f32(x[7], 1), x[3], 49), x[5], -14), x[1], -36); - - // Store the transformed vector - for (int j = 0; j < inner_tile_j; j++) - { - vst1_f32(outptr + j*matrix_stride, U[j]); - } - outptr += 2; - } -#endif // __arm_any__ - for (; channels_remaining; channels_remaining--) - { - // Load x - for (int j = pad_left; j < cells_j; j++) - { - x[j] = *(x_ptrs[j]++); - } - - // Compute U = x . X - U[0] = x[0]*-36 + x[4]*-14 + x[2]*49 + x[6]*1; - U[1] = x[5]*-1 + x[1]*-36 + x[4]*-13 + x[3]*13 + x[2]*36 + x[6]*1; - U[2] = x[3]*-13 + x[4]*-13 + x[1]*36 + x[2]*36 + x[5]*1 + x[6]*1; - U[3] = x[1]*-18 + x[4]*-10 + x[5]*-2 + x[2]*9 + x[3]*20 + x[6]*1; - U[4] = x[3]*-20 + x[4]*-10 + x[5]*2 + x[2]*9 + x[1]*18 + x[6]*1; - U[5] = x[1]*-12 + x[4]*-5 + x[5]*-3 + x[2]*4 + x[3]*15 + x[6]*1; - U[6] = x[3]*-15 + x[4]*-5 + x[5]*3 + x[2]*4 + x[1]*12 + x[6]*1; - U[7] = x[1]*-36 + x[5]*-14 + x[3]*49 + x[7]*1; - - // Store the transformed vector - for (int j = 0; j < inner_tile_j; j++) - { - *(outptr + j*matrix_stride) = U[j]; - } - outptr++; - } -} - -template <> -template <> -const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = -{ - { - { - { - Transform::template process_tile<0, 0, 0, 0>, - Transform::template process_tile<0, 0, 0, 1>, - Transform::template process_tile<0, 0, 0, 2>, - Transform::template process_tile<0, 0, 0, 3>, - Transform::template process_tile<0, 0, 0, 4>, - Transform::template process_tile<0, 0, 0, 5>, - Transform::template process_tile<0, 0, 0, 6>, - } - }, - { - { - Transform::template process_tile<0, 1, 0, 0>, - Transform::template process_tile<0, 1, 0, 1>, - Transform::template process_tile<0, 1, 0, 2>, - Transform::template process_tile<0, 1, 0, 3>, - Transform::template process_tile<0, 1, 0, 4>, - Transform::template process_tile<0, 1, 0, 5>, - Transform::template process_tile<0, 1, 0, 6>, - } - } - } -}; - -template -using TransformTransposed = typename WinogradGEMM::template InputTransform; - -template <> -template <> -const TransformTransposed<6, 3>::TileFn - TransformTransposed<6, 3>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = {}; - -template <> -template <> -const TransformTransposed<4, 5>::TileFn - TransformTransposed<4, 5>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = {}; - -template <> -template <> -const TransformTransposed<2, 7>::TileFn - TransformTransposed<2, 7>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = {}; - - - -template struct WinogradGEMM<1, 6, 1, 3>::InputTransform; -template struct WinogradGEMM<6, 1, 3, 1>::InputTransform; -} // namespace winograd diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_6x6_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_6x6_fp32.cpp new file mode 100644 index 0000000000..908613068a --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_6x6_fp32.cpp @@ -0,0 +1,608 @@ +/* + * Copyright (c) 2017 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 "arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp" +#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" +#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp" + +namespace +{ + +template +void winograd_input_transform_6x6_fp32_process_tile( + int n_channels, + const float* const input_base, + const int input_row_stride, + const int input_col_stride, + float* const matrix_base, + const int matrix_stride +) +{ + constexpr int inner_tile_rows = 6; + constexpr int inner_tile_cols = 6; + constexpr int cells_i = inner_tile_rows - pad_bottom; + constexpr int cells_j = inner_tile_cols - pad_right; + + float *outptr = matrix_base; + + // Get pointers into the input tile + const float *x_ptrs[inner_tile_rows][inner_tile_cols]; + for (int i = pad_top, xi = 0; i < cells_i; i++, xi++) + { + // Get a pointer into the row + const float* const row_ptr = input_base + xi*input_row_stride; + + for (int j = pad_left, xj = 0; j < cells_j; j++, xj++) + { + x_ptrs[i][j] = row_ptr + xj*input_col_stride; + } + } + + // Matrices used/computed in this kernel. + float x[inner_tile_rows][inner_tile_cols]; + float XTx[inner_tile_rows][inner_tile_cols]; + float U[inner_tile_rows][inner_tile_cols]; + for (int i = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++) + { + x[i][j] = XTx[i][j] = 0.0f; + } + } + + // Perform the Winograd input transformation for each channel in the input + // tensor. + int channels_remaining = n_channels; +#ifdef __aarch64__ + for (; channels_remaining >= 4; channels_remaining -= 4) + { + // Matrices used/computed in this kernel + float32x4_t x[inner_tile_rows][inner_tile_cols]; + float32x4_t XTx[inner_tile_rows][inner_tile_cols]; + float32x4_t U[inner_tile_rows][inner_tile_cols]; + for (int i = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++) + { + x[i][j] = vdupq_n_f32(0.0f); + XTx[i][j] = vdupq_n_f32(0.0f); + } + } + + // Read a 6x6 tile in the Winograd domain + for (int i = pad_top; i < cells_i; i++) + { + for (int j = pad_left; j < cells_j; j++) + { + x[i][j] = vld1q_f32(x_ptrs[i][j]); + x_ptrs[i][j] += 4; + } + } + + // Compute XT . x + for (int j = pad_left; j < cells_j; j++) + { + // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j]; + XTx[0][j] = vmlsq_n_f32(vmlaq_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f); + + // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j]; + XTx[1][j] = vmlsq_n_f32(vaddq_f32(x[3][j], x[4][j]), vaddq_f32(x[1][j], x[2][j]), 4.0f); + + // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j]; + XTx[2][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[3][j]), vsubq_f32(x[1][j], x[2][j]), 4.0f); + + // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j]; + XTx[3][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[3][j], x[1][j]), 2.0f); + + // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j]; + XTx[4][j] = vmlaq_n_f32(vsubq_f32(x[4][j], x[2][j]), vsubq_f32(x[1][j], x[3][j]), 2.0f); + + // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j]; + XTx[5][j] = vmlsq_n_f32(vmlaq_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f); + } + + // Compute U = XT . x . X + for (int i = 0; i < inner_tile_rows; i++) + { + // U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4]; + U[i][0] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f); + + // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4]; + U[i][1] = vmlsq_n_f32(vaddq_f32(XTx[i][3], XTx[i][4]), vaddq_f32(XTx[i][1], XTx[i][2]), 4.0f); + + // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4]; + U[i][2] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][3]), vsubq_f32(XTx[i][1], XTx[i][2]), 4.0f); + + // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4]; + U[i][3] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][3], XTx[i][1]), 2.0f); + + // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4]; + U[i][4] = vmlaq_n_f32(vsubq_f32(XTx[i][4], XTx[i][2]), vsubq_f32(XTx[i][1], XTx[i][3]), 2.0f); + + // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5]; + U[i][5] = vmlsq_n_f32(vmlaq_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f); + } + + // Store the transformed matrix + for (int i = 0, m = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++, m++) + { + vst1q_f32(outptr + m*matrix_stride, U[i][j]); + } + } + outptr += 4; + } +#endif // __aarch64__ +#ifdef __arm_any__ + for (; channels_remaining >= 2; channels_remaining -= 2) + { + // Matrices used/computed in this kernel + float32x2_t x[inner_tile_rows][inner_tile_cols]; + float32x2_t XTx[inner_tile_rows][inner_tile_cols]; + float32x2_t U[inner_tile_rows][inner_tile_cols]; + for (int i = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++) + { + x[i][j] = vdup_n_f32(0.0f); + XTx[i][j] = vdup_n_f32(0.0f); + } + } + + // Read a 6x6 tile in the Winograd domain + for (int i = pad_top; i < cells_i; i++) + { + for (int j = pad_left; j < cells_j; j++) + { + x[i][j] = vld1_f32(x_ptrs[i][j]); + x_ptrs[i][j] += 2; + } + } + + // Compute XT . x + for (int j = pad_left; j < cells_j; j++) + { + // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j]; + XTx[0][j] = vmls_n_f32(vmla_n_f32(x[4][j], x[0][j], 4.0f), x[2][j], 5.0f); + + // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j]; + XTx[1][j] = vmls_n_f32(vadd_f32(x[3][j], x[4][j]), vadd_f32(x[1][j], x[2][j]), 4.0f); + + // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j]; + XTx[2][j] = vmla_n_f32(vsub_f32(x[4][j], x[3][j]), vsub_f32(x[1][j], x[2][j]), 4.0f); + + // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j]; + XTx[3][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[3][j], x[1][j]), 2.0f); + + // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j]; + XTx[4][j] = vmla_n_f32(vsub_f32(x[4][j], x[2][j]), vsub_f32(x[1][j], x[3][j]), 2.0f); + + // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j]; + XTx[5][j] = vmls_n_f32(vmla_n_f32(x[5][j], x[1][j], 4.0f), x[3][j], 5.0f); + } + + // Compute U = XT . x . X + for (int i = 0; i < inner_tile_rows; i++) + { + // U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4]; + U[i][0] = vmls_n_f32(vmla_n_f32(XTx[i][4], XTx[i][0], 4.0f), XTx[i][2], 5.0f); + + // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4]; + U[i][1] = vmls_n_f32(vadd_f32(XTx[i][3], XTx[i][4]), vadd_f32(XTx[i][1], XTx[i][2]), 4.0f); + + // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4]; + U[i][2] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][3]), vsub_f32(XTx[i][1], XTx[i][2]), 4.0f); + + // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4]; + U[i][3] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][3], XTx[i][1]), 2.0f); + + // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4]; + U[i][4] = vmla_n_f32(vsub_f32(XTx[i][4], XTx[i][2]), vsub_f32(XTx[i][1], XTx[i][3]), 2.0f); + + // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5]; + U[i][5] = vmls_n_f32(vmla_n_f32(XTx[i][5], XTx[i][1], 4.0f), XTx[i][3], 5.0f); + } + + // Store the transformed matrix + for (int i = 0, m = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++, m++) + { + vst1_f32(outptr + m*matrix_stride, U[i][j]); + } + } + outptr += 2; + } +#endif // __arm_any__ + for (; channels_remaining; channels_remaining--) + { + // Load x + for (int i = pad_top; i < cells_i; i++) + { + for (int j = pad_left; j < cells_j; j++) + { + x[i][j] = *(x_ptrs[i][j]++); + } + } + + // Compute XT . x + for (int j = pad_left; j < cells_j; j++) + { + XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j]; + XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j]; + XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j]; + XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j]; + XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j]; + XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j]; + } + + // Compute U = XT . x . X + for (int i = 0; i < inner_tile_rows; i++) + { + U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4]; + U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4]; + U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4]; + U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4]; + U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4]; + U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5]; + } + + // Store the transformed matrix + for (int i = 0, m = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++, m++) + { + *(outptr + m*matrix_stride) = U[i][j]; + } + } + outptr++; + } +} +} + +namespace winograd +{ +template +using Transform = InputTransformImpl; + +template <> +const Transform<3>::TileFn + Transform<3>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = +{ + { + { + { + winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 0>, // No padding + winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 1>, // Right + winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 4>, // " " + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 0>, // Bottom + winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 1>, // Bottom right + winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 4>, // " " + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 0>, // Bottom + winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 1>, // Bottom right + winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 4>, // " " + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 0>, // Bottom + winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 1>, // Bottom right + winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 4>, // " " + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 0>, // Bottom + winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 1>, // Bottom right + winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 4>, // " " + } + }, + { + { + winograd_input_transform_6x6_fp32_process_tile<0, 1, 0, 0>, // Left + winograd_input_transform_6x6_fp32_process_tile<0, 1, 0, 1>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 0, 2>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 0, 3>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 0, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 1, 1, 0>, // Bottom left + winograd_input_transform_6x6_fp32_process_tile<0, 1, 1, 1>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 1, 2>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 1, 3>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 1, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 1, 2, 0>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 1, 2, 1>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 2, 2>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 2, 3>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 2, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 1, 3, 0>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 1, 3, 1>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 3, 2>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 3, 3>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 3, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 1, 4, 0>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 1, 4, 1>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 4, 2>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 4, 3>, + winograd_input_transform_6x6_fp32_process_tile<0, 1, 4, 4>, + } + } + }, + { + { + { + winograd_input_transform_6x6_fp32_process_tile<1, 0, 0, 0>, // Top + winograd_input_transform_6x6_fp32_process_tile<1, 0, 0, 1>, // Top right + winograd_input_transform_6x6_fp32_process_tile<1, 0, 0, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<1, 0, 0, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<1, 0, 0, 4>, // " " + }, + { + winograd_input_transform_6x6_fp32_process_tile<1, 0, 1, 0>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 1, 1>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 1, 2>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 1, 3>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 1, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<1, 0, 2, 0>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 2, 1>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 2, 2>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 2, 3>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 2, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<1, 0, 3, 0>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 3, 1>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 3, 2>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 3, 3>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 3, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<1, 0, 4, 0>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 4, 1>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 4, 2>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 4, 3>, + winograd_input_transform_6x6_fp32_process_tile<1, 0, 4, 4>, + }, + }, + { + { + winograd_input_transform_6x6_fp32_process_tile<1, 1, 0, 0>, // Top left + winograd_input_transform_6x6_fp32_process_tile<1, 1, 0, 1>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 0, 2>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 0, 3>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 0, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<1, 1, 1, 0>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 1, 1>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 1, 2>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 1, 3>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 1, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<1, 1, 2, 0>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 2, 1>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 2, 2>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 2, 3>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 2, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<1, 1, 3, 0>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 3, 1>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 3, 2>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 3, 3>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 3, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<1, 1, 4, 0>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 4, 1>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 4, 2>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 4, 3>, + winograd_input_transform_6x6_fp32_process_tile<1, 1, 4, 4>, + } + } + } +}; + +template <> +const Transform<5>::TileFn + Transform<5>::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = +{ + { + { + { + winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 0>, // No padding + winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 1>, // Right + winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 0, 4>, // " " + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 0>, // Bottom + winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 1>, // Bottom right + winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 1, 4>, // " " + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 0>, // Bottom + winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 1>, // Bottom right + winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 2, 4>, // " " + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 0>, // Bottom + winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 1>, // Bottom right + winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 3, 4>, // " " + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 0>, // Bottom + winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 1>, // Bottom right + winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 0, 4, 4>, // " " + } + }, + { + { + winograd_input_transform_6x6_fp32_process_tile<0, 2, 0, 0>, // Left + winograd_input_transform_6x6_fp32_process_tile<0, 2, 0, 1>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 0, 2>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 0, 3>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 0, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 2, 1, 0>, // Bottom left + winograd_input_transform_6x6_fp32_process_tile<0, 2, 1, 1>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 1, 2>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 1, 3>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 1, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 2, 2, 0>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 2, 2, 1>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 2, 2>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 2, 3>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 2, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 2, 3, 0>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 2, 3, 1>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 3, 2>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 3, 3>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 3, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<0, 2, 4, 0>, // " " + winograd_input_transform_6x6_fp32_process_tile<0, 2, 4, 1>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 4, 2>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 4, 3>, + winograd_input_transform_6x6_fp32_process_tile<0, 2, 4, 4>, + } + } + }, + { + { + { + winograd_input_transform_6x6_fp32_process_tile<2, 0, 0, 0>, // Top + winograd_input_transform_6x6_fp32_process_tile<2, 0, 0, 1>, // Top right + winograd_input_transform_6x6_fp32_process_tile<2, 0, 0, 2>, // " " + winograd_input_transform_6x6_fp32_process_tile<2, 0, 0, 3>, // " " + winograd_input_transform_6x6_fp32_process_tile<2, 0, 0, 4>, // " " + }, + { + winograd_input_transform_6x6_fp32_process_tile<2, 0, 1, 0>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 1, 1>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 1, 2>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 1, 3>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 1, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<2, 0, 2, 0>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 2, 1>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 2, 2>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 2, 3>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 2, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<2, 0, 3, 0>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 3, 1>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 3, 2>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 3, 3>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 3, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<2, 0, 4, 0>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 4, 1>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 4, 2>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 4, 3>, + winograd_input_transform_6x6_fp32_process_tile<2, 0, 4, 4>, + }, + }, + { + { + winograd_input_transform_6x6_fp32_process_tile<2, 2, 0, 0>, // Top left + winograd_input_transform_6x6_fp32_process_tile<2, 2, 0, 1>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 0, 2>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 0, 3>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 0, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<2, 2, 1, 0>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 1, 1>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 1, 2>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 1, 3>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 1, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<2, 2, 2, 0>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 2, 1>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 2, 2>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 2, 3>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 2, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<2, 2, 3, 0>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 3, 1>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 3, 2>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 3, 3>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 3, 4>, + }, + { + winograd_input_transform_6x6_fp32_process_tile<2, 2, 4, 0>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 4, 1>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 4, 2>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 4, 3>, + winograd_input_transform_6x6_fp32_process_tile<2, 2, 4, 4>, + } + } + } +}; + +template class InputTransform<3, 3, 6, 6, float>; +template class InputTransform<5, 5, 6, 6, float>; +} -- cgit v1.2.1