diff options
Diffstat (limited to 'src/core/NEON/kernels/convolution/winograd/input_transforms')
6 files changed, 2389 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp16_6x6.cpp b/src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp16_6x6.cpp new file mode 100644 index 0000000000..ad759b225e --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp16_6x6.cpp @@ -0,0 +1,280 @@ +/* + * 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. + */ +#if defined(__aarch64__) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) + +#include <arm_neon.h> +#include <cstddef> + +namespace arm_conv { +namespace winograd { +namespace input_transform { + +void a64_fp16_6x6( + const unsigned int n_channels, + const __fp16* const input_base, + const size_t input_row_stride, + const size_t input_col_stride, + __fp16* outptr, + const size_t matrix_stride +) +{ + constexpr int inner_tile_rows = 6; + constexpr int inner_tile_cols = 6; + + // Get pointers into the input tile + const __fp16 *x_ptrs[inner_tile_rows][inner_tile_cols]; + for (int i = 0, xi = 0; i < inner_tile_rows; i++, xi++) + { + // Get a pointer into the row + const __fp16* const row_ptr = input_base + xi*input_row_stride; + + for (int j = 0, xj = 0; j < inner_tile_cols; j++, xj++) + { + x_ptrs[i][j] = row_ptr + xj*input_col_stride; + } + } + + // Matrices used/computed in this kernel. + __fp16 x[inner_tile_rows][inner_tile_cols]; + __fp16 XTx[inner_tile_rows][inner_tile_cols]; + __fp16 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; + for (; channels_remaining >= 8; channels_remaining -= 8) + { + // Matrices used/computed in this kernel + float16x8_t x[inner_tile_rows][inner_tile_cols]; + float16x8_t XTx[inner_tile_rows][inner_tile_cols]; + float16x8_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_f16(0.0f); + XTx[i][j] = vdupq_n_f16(0.0f); + } + } + + // Read a 6x6 tile in the Winograd domain + for (int i = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++) + { + x[i][j] = vld1q_f16(x_ptrs[i][j]); + x_ptrs[i][j] += 8; + } + } + + // Compute XT . x + for (int j = 0; j < inner_tile_cols; j++) + { + // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j]; + XTx[0][j] = vsubq_f16(vaddq_f16(x[4][j], vmulq_f16(x[0][j], vdupq_n_f16(4.0f))), vmulq_f16(x[2][j], vdupq_n_f16(5.0f))); + + // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j]; + XTx[1][j] = vsubq_f16(vaddq_f16(x[3][j], x[4][j]), vmulq_f16(vaddq_f16(x[1][j], x[2][j]), vdupq_n_f16(4.0f))); + + // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j]; + XTx[2][j] = vaddq_f16(vsubq_f16(x[4][j], x[3][j]), vmulq_f16(vsubq_f16(x[1][j], x[2][j]), vdupq_n_f16(4.0f))); + + // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j]; + XTx[3][j] = vaddq_f16(vsubq_f16(x[4][j], x[2][j]), vmulq_f16(vsubq_f16(x[3][j], x[1][j]), vdupq_n_f16(2.0f))); + + // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j]; + XTx[4][j] = vaddq_f16(vsubq_f16(x[4][j], x[2][j]), vmulq_f16(vsubq_f16(x[1][j], x[3][j]), vdupq_n_f16(2.0f))); + + // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j]; + XTx[5][j] = vsubq_f16(vaddq_f16(x[5][j], vmulq_f16(x[1][j], vdupq_n_f16(4.0f))), vmulq_f16(x[3][j], vdupq_n_f16(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] = vsubq_f16(vaddq_f16(XTx[i][4], vmulq_f16(XTx[i][0], vdupq_n_f16(4.0f))), vmulq_f16(XTx[i][2], vdupq_n_f16(5.0f))); + + // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4]; + U[i][1] = vsubq_f16(vaddq_f16(XTx[i][3], XTx[i][4]), vmulq_f16(vaddq_f16(XTx[i][1], XTx[i][2]), vdupq_n_f16(4.0f))); + + // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4]; + U[i][2] = vaddq_f16(vsubq_f16(XTx[i][4], XTx[i][3]), vmulq_f16(vsubq_f16(XTx[i][1], XTx[i][2]), vdupq_n_f16(4.0f))); + + // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4]; + U[i][3] = vaddq_f16(vsubq_f16(XTx[i][4], XTx[i][2]), vmulq_f16(vsubq_f16(XTx[i][3], XTx[i][1]), vdupq_n_f16(2.0f))); + + // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4]; + U[i][4] = vaddq_f16(vsubq_f16(XTx[i][4], XTx[i][2]), vmulq_f16(vsubq_f16(XTx[i][1], XTx[i][3]), vdupq_n_f16(2.0f))); + + // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5]; + U[i][5] = vsubq_f16(vaddq_f16(XTx[i][5], vmulq_f16(XTx[i][1], vdupq_n_f16(4.0f))), vmulq_f16(XTx[i][3], vdupq_n_f16(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_f16(outptr + m*matrix_stride, U[i][j]); + } + } + outptr += 8; + } + for (; channels_remaining >= 4; channels_remaining -= 4) + { + // Matrices used/computed in this kernel + float16x4_t x[inner_tile_rows][inner_tile_cols]; + float16x4_t XTx[inner_tile_rows][inner_tile_cols]; + float16x4_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_f16(0.0f); + XTx[i][j] = vdup_n_f16(0.0f); + } + } + + // Read a 6x6 tile in the Winograd domain + for (int i = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++) + { + x[i][j] = vld1_f16(x_ptrs[i][j]); + x_ptrs[i][j] += 4; + } + } + + // Compute XT . x + for (int j = 0; j < inner_tile_cols; j++) + { + // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j]; + XTx[0][j] = vsub_f16(vadd_f16(x[4][j], vmul_f16(x[0][j], vdup_n_f16(4.0f))), vmul_f16(x[2][j], vdup_n_f16(5.0f))); + + // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j]; + XTx[1][j] = vsub_f16(vadd_f16(x[3][j], x[4][j]), vmul_f16(vadd_f16(x[1][j], x[2][j]), vdup_n_f16(4.0f))); + + // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j]; + XTx[2][j] = vadd_f16(vsub_f16(x[4][j], x[3][j]), vmul_f16(vsub_f16(x[1][j], x[2][j]), vdup_n_f16(4.0f))); + + // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j]; + XTx[3][j] = vadd_f16(vsub_f16(x[4][j], x[2][j]), vmul_f16(vsub_f16(x[3][j], x[1][j]), vdup_n_f16(2.0f))); + + // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j]; + XTx[4][j] = vadd_f16(vsub_f16(x[4][j], x[2][j]), vmul_f16(vsub_f16(x[1][j], x[3][j]), vdup_n_f16(2.0f))); + + // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j]; + XTx[5][j] = vsub_f16(vadd_f16(x[5][j], vmul_f16(x[1][j], vdup_n_f16(4.0f))), vmul_f16(x[3][j], vdup_n_f16(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] = vsub_f16(vadd_f16(XTx[i][4], vmul_f16(XTx[i][0], vdup_n_f16(4.0f))), vmul_f16(XTx[i][2], vdup_n_f16(5.0f))); + + // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4]; + U[i][1] = vsub_f16(vadd_f16(XTx[i][3], XTx[i][4]), vmul_f16(vadd_f16(XTx[i][1], XTx[i][2]), vdup_n_f16(4.0f))); + + // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4]; + U[i][2] = vadd_f16(vsub_f16(XTx[i][4], XTx[i][3]), vmul_f16(vsub_f16(XTx[i][1], XTx[i][2]), vdup_n_f16(4.0f))); + + // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4]; + U[i][3] = vadd_f16(vsub_f16(XTx[i][4], XTx[i][2]), vmul_f16(vsub_f16(XTx[i][3], XTx[i][1]), vdup_n_f16(2.0f))); + + // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4]; + U[i][4] = vadd_f16(vsub_f16(XTx[i][4], XTx[i][2]), vmul_f16(vsub_f16(XTx[i][1], XTx[i][3]), vdup_n_f16(2.0f))); + + // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5]; + U[i][5] = vsub_f16(vadd_f16(XTx[i][5], vmul_f16(XTx[i][1], vdup_n_f16(4.0f))), vmul_f16(XTx[i][3], vdup_n_f16(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_f16(outptr + m*matrix_stride, U[i][j]); + } + } + outptr += 4; + } + for (; channels_remaining; channels_remaining--) + { + // Load x + for (int i = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++) + { + x[i][j] = *(x_ptrs[i][j]++); + } + } + + // Compute XT . x + for (int j = 0; j < inner_tile_cols; 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 input_transform +} // namespace winograd +} // namespace arm_conv + +#endif // defined(__aarch64__) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) diff --git a/src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp32_6x6.cpp b/src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp32_6x6.cpp new file mode 100644 index 0000000000..6f818c69ff --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/input_transforms/a64_fp32_6x6.cpp @@ -0,0 +1,1140 @@ +/* + * 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. + */ + +#ifdef __aarch64__ + +#include <cstddef> + +namespace arm_conv { +namespace winograd { +namespace input_transform { + +void a64_fp32_6x6( + unsigned int n_channels, + const float *input_base, + const size_t input_row_stride, + const size_t input_col_stride, + float *matrix_base, + const size_t matrix_stride +) +{ + const float pcoeffs[4] = {1.0f, 2.0f, 4.0f, 5.0f}; + __asm__ __volatile__( + "ldr q0, [%[pcoeffs]]\n" + "add x25, %[inptr0], %[input_row_stride]\n" + "add x10, %[input_col_stride1], %[input_col_stride1]\n" + "add x16, x25, %[input_row_stride]\n" + "add x19, x10, %[input_col_stride1]\n" + "add x26, x16, %[input_row_stride]\n" + "add x20, x19, %[input_col_stride1]\n" + "add x17, x26, %[input_row_stride]\n" + "add x21, x20, %[input_col_stride1]\n" + "add x27, x17, %[input_row_stride]\n" + "add x28, %[outptr0], %[output_row_stride]\n" + "add x11, %[output_col_stride1], %[output_col_stride1]\n" + "add x22, x28, %[output_row_stride]\n" + "add x13, x11, %[output_col_stride1]\n" + "add x12, x22, %[output_row_stride]\n" + "add x23, x13, %[output_col_stride1]\n" + "add x14, x12, %[output_row_stride]\n" + "add x15, x23, %[output_col_stride1]\n" + "add x24, x14, %[output_row_stride]\n" + "cmp %w[n_channels], #4\n" + "blt 2f\n" + "1:\n" + "ldr q8, [%[inptr0], x20]\n" + "ldr q2, [%[inptr0], x10]\n" + "mov v14.16b, v8.16b\n" + "ldr q9, [%[inptr0]]\n" + "mov v10.16b, v8.16b\n" + "ldr q1, [%[inptr0], x21]\n" + "fmla v14.4s, v9.4s, v0.s[2]\n" + "ldr q4, [%[inptr0], x19]\n" + "mov v9.16b, v8.16b\n" + "ldr q12, [%[inptr0], %[input_col_stride1]]\n" + "fmls v10.4s, v12.4s, v0.s[2]\n" + "ldr q5, [x16, x20]\n" + "fmls v14.4s, v2.4s, v0.s[3]\n" + "ldr q20, [x16, x10]\n" + "fmla v9.4s, v12.4s, v0.s[2]\n" + "ldr q3, [x16]\n" + "fmls v10.4s, v2.4s, v0.s[2]\n" + "ldr q6, [x16, x21]\n" + "mov v7.16b, v8.16b\n" + "ldr q16, [x16, x19]\n" + "fmls v9.4s, v2.4s, v0.s[2]\n" + "ldr q22, [x16, %[input_col_stride1]]\n" + "fadd v10.4s, v10.4s, v4.4s\n" + "ldr q17, [x17, x20]\n" + "fmls v7.4s, v12.4s, v0.s[1]\n" + "ldr q15, [x17, x10]\n" + "fsub v9.4s, v9.4s, v4.4s\n" + "ldr q19, [x17]\n" + "mov v8.16b, v8.16b\n" + "ldr q18, [x17, x21]\n" + "fsub v7.4s, v7.4s, v2.4s\n" + "ldr q13, [x17, x19]\n" + "fmla v7.4s, v4.4s, v0.s[1]\n" + "ldr q21, [x17, %[input_col_stride1]]\n" + "fmla v8.4s, v12.4s, v0.s[1]\n" + "add %[inptr0], %[inptr0], #16\n" + "mov v11.16b, v1.16b\n" + "add x16, x16, #16\n" + "mov v1.16b, v5.16b\n" + "add x17, x17, #16\n" + "fsub v8.4s, v8.4s, v2.4s\n" + "fmla v11.4s, v12.4s, v0.s[2]\n" + "fmls v8.4s, v4.4s, v0.s[1]\n" + "fmla v1.4s, v3.4s, v0.s[2]\n" + "mov v2.16b, v5.16b\n" + "mov v3.16b, v5.16b\n" + "fmls v11.4s, v4.4s, v0.s[3]\n" + "mov v4.16b, v5.16b\n" + "fmls v1.4s, v20.4s, v0.s[3]\n" + "fmls v2.4s, v22.4s, v0.s[2]\n" + "fmla v3.4s, v22.4s, v0.s[2]\n" + "fmls v4.4s, v22.4s, v0.s[1]\n" + "mov v5.16b, v5.16b\n" + "mov v6.16b, v6.16b\n" + "fmls v2.4s, v20.4s, v0.s[2]\n" + "mov v12.16b, v17.16b\n" + "fmls v3.4s, v20.4s, v0.s[2]\n" + "fsub v4.4s, v4.4s, v20.4s\n" + "fmla v4.4s, v16.4s, v0.s[1]\n" + "fmla v5.4s, v22.4s, v0.s[1]\n" + "fadd v2.4s, v2.4s, v16.4s\n" + "fmla v6.4s, v22.4s, v0.s[2]\n" + "fsub v3.4s, v3.4s, v16.4s\n" + "fmla v12.4s, v19.4s, v0.s[2]\n" + "fsub v5.4s, v5.4s, v20.4s\n" + "mov v19.16b, v17.16b\n" + "fmls v5.4s, v16.4s, v0.s[1]\n" + "fmls v6.4s, v16.4s, v0.s[3]\n" + "fmls v12.4s, v15.4s, v0.s[3]\n" + "fmls v19.4s, v21.4s, v0.s[2]\n" + "mov v20.16b, v17.16b\n" + "mov v16.16b, v17.16b\n" + "mov v17.16b, v17.16b\n" + "mov v18.16b, v18.16b\n" + "fmls v19.4s, v15.4s, v0.s[2]\n" + "fmla v20.4s, v21.4s, v0.s[2]\n" + "fmls v16.4s, v21.4s, v0.s[1]\n" + "fmla v17.4s, v21.4s, v0.s[1]\n" + "fmla v18.4s, v21.4s, v0.s[2]\n" + "mov v23.16b, v12.16b\n" + "fadd v19.4s, v19.4s, v13.4s\n" + "fmls v20.4s, v15.4s, v0.s[2]\n" + "fsub v16.4s, v16.4s, v15.4s\n" + "fsub v17.4s, v17.4s, v15.4s\n" + "fmla v16.4s, v13.4s, v0.s[1]\n" + "fmls v17.4s, v13.4s, v0.s[1]\n" + "fsub v20.4s, v20.4s, v13.4s\n" + "fmls v18.4s, v13.4s, v0.s[3]\n" + "fmla v23.4s, v14.4s, v0.s[2]\n" + "mov v15.16b, v19.16b\n" + "mov v14.16b, v20.16b\n" + "mov v24.16b, v16.16b\n" + "fmla v15.4s, v10.4s, v0.s[2]\n" + "mov v10.16b, v17.16b\n" + "fmls v23.4s, v1.4s, v0.s[3]\n" + "fmla v14.4s, v9.4s, v0.s[2]\n" + "fmla v24.4s, v7.4s, v0.s[2]\n" + "fmla v10.4s, v8.4s, v0.s[2]\n" + "fmls v15.4s, v2.4s, v0.s[3]\n" + "mov v7.16b, v18.16b\n" + "str q23, [%[outptr0]]\n" + "fmls v14.4s, v3.4s, v0.s[3]\n" + "fmls v24.4s, v4.4s, v0.s[3]\n" + "fmls v10.4s, v5.4s, v0.s[3]\n" + "str q15, [%[outptr0], %[output_col_stride1]]\n" + "fmla v7.4s, v11.4s, v0.s[2]\n" + "str q14, [%[outptr0], x11]\n" + "str q24, [%[outptr0], x13]\n" + "str q10, [%[outptr0], x23]\n" + "fmls v7.4s, v6.4s, v0.s[3]\n" + "str q7, [%[outptr0], x15]\n" + "add %[outptr0], %[outptr0], #16\n" + "mov v26.16b, v12.16b\n" + "mov v25.16b, v19.16b\n" + "ldr q11, [x25, x20]\n" + "mov v10.16b, v11.16b\n" + "ldr q23, [x25, x10]\n" + "mov v9.16b, v11.16b\n" + "ldr q7, [x25]\n" + "fmla v10.4s, v7.4s, v0.s[2]\n" + "ldr q13, [x25, x21]\n" + "mov v7.16b, v11.16b\n" + "ldr q31, [x25, x19]\n" + "mov v8.16b, v11.16b\n" + "ldr q21, [x25, %[input_col_stride1]]\n" + "fmls v10.4s, v23.4s, v0.s[3]\n" + "ldr q30, [x26, x20]\n" + "fmls v9.4s, v21.4s, v0.s[2]\n" + "ldr q29, [x26, x10]\n" + "fmla v7.4s, v21.4s, v0.s[2]\n" + "ldr q22, [x26]\n" + "fmls v8.4s, v21.4s, v0.s[1]\n" + "ldr q24, [x26, x21]\n" + "fmls v9.4s, v23.4s, v0.s[2]\n" + "ldr q27, [x26, x19]\n" + "fmls v7.4s, v23.4s, v0.s[2]\n" + "ldr q28, [x26, %[input_col_stride1]]\n" + "fsub v8.4s, v8.4s, v23.4s\n" + "add x25, x25, #16\n" + "fadd v9.4s, v9.4s, v31.4s\n" + "add x26, x26, #16\n" + "fsub v7.4s, v7.4s, v31.4s\n" + "fmla v8.4s, v31.4s, v0.s[1]\n" + "mov v11.16b, v11.16b\n" + "mov v15.16b, v13.16b\n" + "mov v14.16b, v30.16b\n" + "mov v13.16b, v30.16b\n" + "fmla v11.4s, v21.4s, v0.s[1]\n" + "fmla v15.4s, v21.4s, v0.s[2]\n" + "fmla v14.4s, v22.4s, v0.s[2]\n" + "fmls v13.4s, v28.4s, v0.s[2]\n" + "mov v21.16b, v30.16b\n" + "mov v22.16b, v30.16b\n" + "fsub v11.4s, v11.4s, v23.4s\n" + "fmls v15.4s, v31.4s, v0.s[3]\n" + "fmls v11.4s, v31.4s, v0.s[1]\n" + "fmls v14.4s, v29.4s, v0.s[3]\n" + "fmls v13.4s, v29.4s, v0.s[2]\n" + "fmla v21.4s, v28.4s, v0.s[2]\n" + "fmls v22.4s, v28.4s, v0.s[1]\n" + "mov v23.16b, v30.16b\n" + "mov v24.16b, v24.16b\n" + "fmls v26.4s, v10.4s, v0.s[2]\n" + "fadd v13.4s, v13.4s, v27.4s\n" + "fmls v21.4s, v29.4s, v0.s[2]\n" + "fsub v22.4s, v22.4s, v29.4s\n" + "fmla v23.4s, v28.4s, v0.s[1]\n" + "fmla v22.4s, v27.4s, v0.s[1]\n" + "fmla v24.4s, v28.4s, v0.s[2]\n" + "fsub v21.4s, v21.4s, v27.4s\n" + "fmls v26.4s, v1.4s, v0.s[2]\n" + "fsub v23.4s, v23.4s, v29.4s\n" + "fmls v25.4s, v9.4s, v0.s[2]\n" + "fmls v23.4s, v27.4s, v0.s[1]\n" + "fmls v24.4s, v27.4s, v0.s[3]\n" + "fadd v26.4s, v26.4s, v14.4s\n" + "mov v27.16b, v20.16b\n" + "str q26, [x28]\n" + "fmls v25.4s, v2.4s, v0.s[2]\n" + "fmls v27.4s, v7.4s, v0.s[2]\n" + "mov v31.16b, v16.16b\n" + "mov v30.16b, v17.16b\n" + "mov v29.16b, v18.16b\n" + "fadd v25.4s, v25.4s, v13.4s\n" + "fmls v31.4s, v8.4s, v0.s[2]\n" + "str q25, [x28, %[output_col_stride1]]\n" + "fmls v27.4s, v3.4s, v0.s[2]\n" + "fmls v30.4s, v11.4s, v0.s[2]\n" + "fmls v29.4s, v15.4s, v0.s[2]\n" + "fmls v31.4s, v4.4s, v0.s[2]\n" + "mov v26.16b, v12.16b\n" + "fadd v27.4s, v27.4s, v21.4s\n" + "mov v25.16b, v19.16b\n" + "str q27, [x28, x11]\n" + "fmls v30.4s, v5.4s, v0.s[2]\n" + "fadd v31.4s, v31.4s, v22.4s\n" + "fmls v29.4s, v6.4s, v0.s[2]\n" + "str q31, [x28, x13]\n" + "fmla v26.4s, v10.4s, v0.s[2]\n" + "fadd v30.4s, v30.4s, v23.4s\n" + "fmla v25.4s, v9.4s, v0.s[2]\n" + "str q30, [x28, x23]\n" + "fadd v29.4s, v29.4s, v24.4s\n" + "str q29, [x28, x15]\n" + "fmls v26.4s, v1.4s, v0.s[2]\n" + "fmls v25.4s, v2.4s, v0.s[2]\n" + "add x28, x28, #16\n" + "mov v30.16b, v20.16b\n" + "mov v29.16b, v16.16b\n" + "fsub v26.4s, v26.4s, v14.4s\n" + "mov v28.16b, v17.16b\n" + "str q26, [x22]\n" + "fsub v25.4s, v25.4s, v13.4s\n" + "str q25, [x22, %[output_col_stride1]]\n" + "fmla v30.4s, v7.4s, v0.s[2]\n" + "fmla v29.4s, v8.4s, v0.s[2]\n" + "fmla v28.4s, v11.4s, v0.s[2]\n" + "mov v26.16b, v18.16b\n" + "mov v25.16b, v12.16b\n" + "fmls v30.4s, v3.4s, v0.s[2]\n" + "mov v31.16b, v19.16b\n" + "fmls v29.4s, v4.4s, v0.s[2]\n" + "fmls v28.4s, v5.4s, v0.s[2]\n" + "fmla v26.4s, v15.4s, v0.s[2]\n" + "fmls v25.4s, v10.4s, v0.s[1]\n" + "fsub v30.4s, v30.4s, v21.4s\n" + "fmls v31.4s, v9.4s, v0.s[1]\n" + "str q30, [x22, x11]\n" + "fsub v29.4s, v29.4s, v22.4s\n" + "str q29, [x22, x13]\n" + "fsub v28.4s, v28.4s, v23.4s\n" + "str q28, [x22, x23]\n" + "fmls v26.4s, v6.4s, v0.s[2]\n" + "fsub v25.4s, v25.4s, v1.4s\n" + "fsub v31.4s, v31.4s, v2.4s\n" + "fmla v25.4s, v14.4s, v0.s[1]\n" + "fmla v31.4s, v13.4s, v0.s[1]\n" + "fsub v26.4s, v26.4s, v24.4s\n" + "mov v27.16b, v20.16b\n" + "str q26, [x22, x15]\n" + "mov v26.16b, v16.16b\n" + "str q25, [x12]\n" + "fmls v27.4s, v7.4s, v0.s[1]\n" + "str q31, [x12, %[output_col_stride1]]\n" + "fmls v26.4s, v8.4s, v0.s[1]\n" + "mov v25.16b, v17.16b\n" + "add x22, x22, #16\n" + "fsub v27.4s, v27.4s, v3.4s\n" + "mov v28.16b, v18.16b\n" + "fmla v27.4s, v21.4s, v0.s[1]\n" + "fsub v26.4s, v26.4s, v4.4s\n" + "fmla v26.4s, v22.4s, v0.s[1]\n" + "fmls v25.4s, v11.4s, v0.s[1]\n" + "fmls v28.4s, v15.4s, v0.s[1]\n" + "mov v12.16b, v12.16b\n" + "str q27, [x12, x11]\n" + "mov v19.16b, v19.16b\n" + "str q26, [x12, x13]\n" + "fsub v25.4s, v25.4s, v5.4s\n" + "fmla v25.4s, v23.4s, v0.s[1]\n" + "fsub v28.4s, v28.4s, v6.4s\n" + "fmla v28.4s, v24.4s, v0.s[1]\n" + "fmla v12.4s, v10.4s, v0.s[1]\n" + "fmla v19.4s, v9.4s, v0.s[1]\n" + "mov v20.16b, v20.16b\n" + "str q25, [x12, x23]\n" + "mov v16.16b, v16.16b\n" + "str q28, [x12, x15]\n" + "fsub v12.4s, v12.4s, v1.4s\n" + "fmls v12.4s, v14.4s, v0.s[1]\n" + "add x12, x12, #16\n" + "fsub v19.4s, v19.4s, v2.4s\n" + "fmla v20.4s, v7.4s, v0.s[1]\n" + "fmls v19.4s, v13.4s, v0.s[1]\n" + "fmla v16.4s, v8.4s, v0.s[1]\n" + "str q12, [x14]\n" + "mov v1.16b, v17.16b\n" + "fsub v20.4s, v20.4s, v3.4s\n" + "mov v17.16b, v18.16b\n" + "str q19, [x14, %[output_col_stride1]]\n" + "fmls v20.4s, v21.4s, v0.s[1]\n" + "fsub v16.4s, v16.4s, v4.4s\n" + "fmla v1.4s, v11.4s, v0.s[1]\n" + "fmls v16.4s, v22.4s, v0.s[1]\n" + "fmla v17.4s, v15.4s, v0.s[1]\n" + "str q20, [x14, x11]\n" + "fsub v1.4s, v1.4s, v5.4s\n" + "str q16, [x14, x13]\n" + "fmls v1.4s, v23.4s, v0.s[1]\n" + "fsub v17.4s, v17.4s, v6.4s\n" + "fmls v17.4s, v24.4s, v0.s[1]\n" + "str q1, [x14, x23]\n" + "str q17, [x14, x15]\n" + "add x14, x14, #16\n" + "ldr q2, [x27, x20]\n" + "mov v4.16b, v2.16b\n" + "ldr q17, [x27, x10]\n" + "mov v12.16b, v2.16b\n" + "ldr q18, [x27]\n" + "fmla v4.4s, v18.4s, v0.s[2]\n" + "ldr q3, [x27, x21]\n" + "mov v6.16b, v2.16b\n" + "ldr q5, [x27, x19]\n" + "mov v1.16b, v2.16b\n" + "ldr q18, [x27, %[input_col_stride1]]\n" + "fmls v4.4s, v17.4s, v0.s[3]\n" + "add x27, x27, #16\n" + "fmls v12.4s, v18.4s, v0.s[2]\n" + "sub %w[n_channels], %w[n_channels], #4\n" + "fmla v6.4s, v18.4s, v0.s[2]\n" + "cmp %w[n_channels], #4\n" + "fmls v1.4s, v18.4s, v0.s[1]\n" + "mov v2.16b, v2.16b\n" + "fmls v12.4s, v17.4s, v0.s[2]\n" + "mov v3.16b, v3.16b\n" + "fmls v6.4s, v17.4s, v0.s[2]\n" + "fmla v2.4s, v18.4s, v0.s[1]\n" + "fsub v1.4s, v1.4s, v17.4s\n" + "fmla v3.4s, v18.4s, v0.s[2]\n" + "fadd v12.4s, v12.4s, v5.4s\n" + "fmla v1.4s, v5.4s, v0.s[1]\n" + "fsub v6.4s, v6.4s, v5.4s\n" + "fsub v2.4s, v2.4s, v17.4s\n" + "fmls v2.4s, v5.4s, v0.s[1]\n" + "fmls v3.4s, v5.4s, v0.s[3]\n" + "mov v4.16b, v4.16b\n" + "mov v16.16b, v12.16b\n" + "mov v5.16b, v6.16b\n" + "mov v6.16b, v1.16b\n" + "fmla v4.4s, v10.4s, v0.s[2]\n" + "fmla v16.4s, v9.4s, v0.s[2]\n" + "fmla v5.4s, v7.4s, v0.s[2]\n" + "fmla v6.4s, v8.4s, v0.s[2]\n" + "mov v9.16b, v2.16b\n" + "mov v10.16b, v3.16b\n" + "fmls v4.4s, v14.4s, v0.s[3]\n" + "fmls v16.4s, v13.4s, v0.s[3]\n" + "fmls v5.4s, v21.4s, v0.s[3]\n" + "fmls v6.4s, v22.4s, v0.s[3]\n" + "fmla v9.4s, v11.4s, v0.s[2]\n" + "fmla v10.4s, v15.4s, v0.s[2]\n" + "str q4, [x24]\n" + "str q16, [x24, %[output_col_stride1]]\n" + "str q5, [x24, x11]\n" + "str q6, [x24, x13]\n" + "fmls v9.4s, v23.4s, v0.s[3]\n" + "fmls v10.4s, v24.4s, v0.s[3]\n" + "str q9, [x24, x23]\n" + "str q10, [x24, x15]\n" + "add x24, x24, #16\n" + "bge 1b\n" + "2:\n" + "cmp %w[n_channels], #2\n" + "blt 3f\n" + "ldr d8, [%[inptr0], x20]\n" + "mov v14.16b, v8.16b\n" + "ldr d2, [%[inptr0], x10]\n" + "mov v10.16b, v8.16b\n" + "ldr d9, [%[inptr0]]\n" + "fmla v14.4s, v9.4s, v0.s[2]\n" + "ldr d1, [%[inptr0], x21]\n" + "mov v9.16b, v8.16b\n" + "ldr d4, [%[inptr0], x19]\n" + "mov v7.16b, v8.16b\n" + "ldr d12, [%[inptr0], %[input_col_stride1]]\n" + "fmls v14.4s, v2.4s, v0.s[3]\n" + "ldr d5, [x16, x20]\n" + "fmls v10.4s, v12.4s, v0.s[2]\n" + "ldr d20, [x16, x10]\n" + "fmla v9.4s, v12.4s, v0.s[2]\n" + "ldr d3, [x16]\n" + "fmls v7.4s, v12.4s, v0.s[1]\n" + "ldr d6, [x16, x21]\n" + "fmls v10.4s, v2.4s, v0.s[2]\n" + "ldr d16, [x16, x19]\n" + "fmls v9.4s, v2.4s, v0.s[2]\n" + "ldr d22, [x16, %[input_col_stride1]]\n" + "fsub v7.4s, v7.4s, v2.4s\n" + "ldr d17, [x17, x20]\n" + "fadd v10.4s, v10.4s, v4.4s\n" + "ldr d15, [x17, x10]\n" + "fsub v9.4s, v9.4s, v4.4s\n" + "ldr d19, [x17]\n" + "fmla v7.4s, v4.4s, v0.s[1]\n" + "ldr d18, [x17, x21]\n" + "mov v8.16b, v8.16b\n" + "ldr d13, [x17, x19]\n" + "mov v11.16b, v1.16b\n" + "ldr d21, [x17, %[input_col_stride1]]\n" + "fmla v8.4s, v12.4s, v0.s[1]\n" + "add %[inptr0], %[inptr0], #8\n" + "fmla v11.4s, v12.4s, v0.s[2]\n" + "add x16, x16, #8\n" + "mov v1.16b, v5.16b\n" + "add x17, x17, #8\n" + "fsub v8.4s, v8.4s, v2.4s\n" + "mov v2.16b, v5.16b\n" + "fmls v8.4s, v4.4s, v0.s[1]\n" + "fmls v11.4s, v4.4s, v0.s[3]\n" + "fmla v1.4s, v3.4s, v0.s[2]\n" + "fmls v2.4s, v22.4s, v0.s[2]\n" + "mov v3.16b, v5.16b\n" + "mov v4.16b, v5.16b\n" + "mov v5.16b, v5.16b\n" + "mov v6.16b, v6.16b\n" + "fmls v1.4s, v20.4s, v0.s[3]\n" + "fmls v2.4s, v20.4s, v0.s[2]\n" + "fmla v3.4s, v22.4s, v0.s[2]\n" + "fmls v4.4s, v22.4s, v0.s[1]\n" + "fmla v5.4s, v22.4s, v0.s[1]\n" + "fmla v6.4s, v22.4s, v0.s[2]\n" + "fadd v2.4s, v2.4s, v16.4s\n" + "mov v12.16b, v17.16b\n" + "fmls v3.4s, v20.4s, v0.s[2]\n" + "fsub v4.4s, v4.4s, v20.4s\n" + "fmla v4.4s, v16.4s, v0.s[1]\n" + "fsub v5.4s, v5.4s, v20.4s\n" + "fmls v5.4s, v16.4s, v0.s[1]\n" + "fmls v6.4s, v16.4s, v0.s[3]\n" + "fsub v3.4s, v3.4s, v16.4s\n" + "fmla v12.4s, v19.4s, v0.s[2]\n" + "mov v19.16b, v17.16b\n" + "mov v20.16b, v17.16b\n" + "mov v16.16b, v17.16b\n" + "mov v17.16b, v17.16b\n" + "fmls v12.4s, v15.4s, v0.s[3]\n" + "fmls v19.4s, v21.4s, v0.s[2]\n" + "fmla v20.4s, v21.4s, v0.s[2]\n" + "fmls v16.4s, v21.4s, v0.s[1]\n" + "fmla v17.4s, v21.4s, v0.s[1]\n" + "mov v18.16b, v18.16b\n" + "fmls v19.4s, v15.4s, v0.s[2]\n" + "mov v23.16b, v12.16b\n" + "fmls v20.4s, v15.4s, v0.s[2]\n" + "fsub v16.4s, v16.4s, v15.4s\n" + "fmla v16.4s, v13.4s, v0.s[1]\n" + "fsub v17.4s, v17.4s, v15.4s\n" + "fadd v19.4s, v19.4s, v13.4s\n" + "fmls v17.4s, v13.4s, v0.s[1]\n" + "fsub v20.4s, v20.4s, v13.4s\n" + "fmla v18.4s, v21.4s, v0.s[2]\n" + "fmla v23.4s, v14.4s, v0.s[2]\n" + "mov v15.16b, v19.16b\n" + "mov v14.16b, v20.16b\n" + "mov v24.16b, v16.16b\n" + "fmls v18.4s, v13.4s, v0.s[3]\n" + "fmla v15.4s, v10.4s, v0.s[2]\n" + "fmls v23.4s, v1.4s, v0.s[3]\n" + "fmla v14.4s, v9.4s, v0.s[2]\n" + "fmla v24.4s, v7.4s, v0.s[2]\n" + "mov v10.16b, v17.16b\n" + "fmls v15.4s, v2.4s, v0.s[3]\n" + "mov v7.16b, v18.16b\n" + "str d23, [%[outptr0]]\n" + "fmls v14.4s, v3.4s, v0.s[3]\n" + "fmls v24.4s, v4.4s, v0.s[3]\n" + "fmla v10.4s, v8.4s, v0.s[2]\n" + "str d15, [%[outptr0], %[output_col_stride1]]\n" + "fmla v7.4s, v11.4s, v0.s[2]\n" + "str d14, [%[outptr0], x11]\n" + "fmls v10.4s, v5.4s, v0.s[3]\n" + "str d24, [%[outptr0], x13]\n" + "fmls v7.4s, v6.4s, v0.s[3]\n" + "str d10, [%[outptr0], x23]\n" + "str d7, [%[outptr0], x15]\n" + "add %[outptr0], %[outptr0], #8\n" + "mov v26.16b, v12.16b\n" + "mov v25.16b, v19.16b\n" + "ldr d11, [x25, x20]\n" + "mov v10.16b, v11.16b\n" + "ldr d23, [x25, x10]\n" + "mov v9.16b, v11.16b\n" + "ldr d7, [x25]\n" + "fmla v10.4s, v7.4s, v0.s[2]\n" + "ldr d13, [x25, x21]\n" + "mov v7.16b, v11.16b\n" + "ldr d31, [x25, x19]\n" + "mov v8.16b, v11.16b\n" + "ldr d21, [x25, %[input_col_stride1]]\n" + "fmls v10.4s, v23.4s, v0.s[3]\n" + "ldr d30, [x26, x20]\n" + "fmls v9.4s, v21.4s, v0.s[2]\n" + "ldr d29, [x26, x10]\n" + "fmla v7.4s, v21.4s, v0.s[2]\n" + "ldr d22, [x26]\n" + "fmls v8.4s, v21.4s, v0.s[1]\n" + "ldr d24, [x26, x21]\n" + "fmls v9.4s, v23.4s, v0.s[2]\n" + "ldr d27, [x26, x19]\n" + "fmls v7.4s, v23.4s, v0.s[2]\n" + "ldr d28, [x26, %[input_col_stride1]]\n" + "fsub v8.4s, v8.4s, v23.4s\n" + "add x25, x25, #8\n" + "fadd v9.4s, v9.4s, v31.4s\n" + "add x26, x26, #8\n" + "fsub v7.4s, v7.4s, v31.4s\n" + "fmla v8.4s, v31.4s, v0.s[1]\n" + "mov v11.16b, v11.16b\n" + "mov v15.16b, v13.16b\n" + "mov v14.16b, v30.16b\n" + "mov v13.16b, v30.16b\n" + "fmla v11.4s, v21.4s, v0.s[1]\n" + "fmla v15.4s, v21.4s, v0.s[2]\n" + "fmla v14.4s, v22.4s, v0.s[2]\n" + "fmls v13.4s, v28.4s, v0.s[2]\n" + "mov v21.16b, v30.16b\n" + "mov v22.16b, v30.16b\n" + "fsub v11.4s, v11.4s, v23.4s\n" + "fmls v15.4s, v31.4s, v0.s[3]\n" + "fmls v11.4s, v31.4s, v0.s[1]\n" + "fmls v14.4s, v29.4s, v0.s[3]\n" + "fmls v13.4s, v29.4s, v0.s[2]\n" + "fmla v21.4s, v28.4s, v0.s[2]\n" + "fmls v22.4s, v28.4s, v0.s[1]\n" + "mov v23.16b, v30.16b\n" + "mov v24.16b, v24.16b\n" + "fmls v26.4s, v10.4s, v0.s[2]\n" + "fadd v13.4s, v13.4s, v27.4s\n" + "fmls v21.4s, v29.4s, v0.s[2]\n" + "fsub v22.4s, v22.4s, v29.4s\n" + "fmla v23.4s, v28.4s, v0.s[1]\n" + "fmla v22.4s, v27.4s, v0.s[1]\n" + "fmla v24.4s, v28.4s, v0.s[2]\n" + "fsub v21.4s, v21.4s, v27.4s\n" + "fmls v26.4s, v1.4s, v0.s[2]\n" + "fsub v23.4s, v23.4s, v29.4s\n" + "fmls v25.4s, v9.4s, v0.s[2]\n" + "fmls v23.4s, v27.4s, v0.s[1]\n" + "fmls v24.4s, v27.4s, v0.s[3]\n" + "fadd v26.4s, v26.4s, v14.4s\n" + "mov v27.16b, v20.16b\n" + "str d26, [x28]\n" + "fmls v25.4s, v2.4s, v0.s[2]\n" + "fmls v27.4s, v7.4s, v0.s[2]\n" + "mov v31.16b, v16.16b\n" + "mov v30.16b, v17.16b\n" + "mov v29.16b, v18.16b\n" + "fadd v25.4s, v25.4s, v13.4s\n" + "fmls v31.4s, v8.4s, v0.s[2]\n" + "str d25, [x28, %[output_col_stride1]]\n" + "fmls v27.4s, v3.4s, v0.s[2]\n" + "fmls v30.4s, v11.4s, v0.s[2]\n" + "fmls v29.4s, v15.4s, v0.s[2]\n" + "fmls v31.4s, v4.4s, v0.s[2]\n" + "mov v26.16b, v12.16b\n" + "fadd v27.4s, v27.4s, v21.4s\n" + "mov v25.16b, v19.16b\n" + "str d27, [x28, x11]\n" + "fmls v30.4s, v5.4s, v0.s[2]\n" + "fadd v31.4s, v31.4s, v22.4s\n" + "fmls v29.4s, v6.4s, v0.s[2]\n" + "str d31, [x28, x13]\n" + "fmla v26.4s, v10.4s, v0.s[2]\n" + "fadd v30.4s, v30.4s, v23.4s\n" + "fmla v25.4s, v9.4s, v0.s[2]\n" + "str d30, [x28, x23]\n" + "fadd v29.4s, v29.4s, v24.4s\n" + "str d29, [x28, x15]\n" + "fmls v26.4s, v1.4s, v0.s[2]\n" + "fmls v25.4s, v2.4s, v0.s[2]\n" + "add x28, x28, #8\n" + "mov v30.16b, v20.16b\n" + "mov v29.16b, v16.16b\n" + "fsub v26.4s, v26.4s, v14.4s\n" + "mov v28.16b, v17.16b\n" + "str d26, [x22]\n" + "fsub v25.4s, v25.4s, v13.4s\n" + "str d25, [x22, %[output_col_stride1]]\n" + "fmla v30.4s, v7.4s, v0.s[2]\n" + "fmla v29.4s, v8.4s, v0.s[2]\n" + "fmla v28.4s, v11.4s, v0.s[2]\n" + "mov v26.16b, v18.16b\n" + "mov v25.16b, v12.16b\n" + "fmls v30.4s, v3.4s, v0.s[2]\n" + "mov v31.16b, v19.16b\n" + "fmls v29.4s, v4.4s, v0.s[2]\n" + "fmls v28.4s, v5.4s, v0.s[2]\n" + "fmla v26.4s, v15.4s, v0.s[2]\n" + "fmls v25.4s, v10.4s, v0.s[1]\n" + "fsub v30.4s, v30.4s, v21.4s\n" + "fmls v31.4s, v9.4s, v0.s[1]\n" + "str d30, [x22, x11]\n" + "fsub v29.4s, v29.4s, v22.4s\n" + "str d29, [x22, x13]\n" + "fsub v28.4s, v28.4s, v23.4s\n" + "str d28, [x22, x23]\n" + "fmls v26.4s, v6.4s, v0.s[2]\n" + "fsub v25.4s, v25.4s, v1.4s\n" + "fsub v31.4s, v31.4s, v2.4s\n" + "fmla v25.4s, v14.4s, v0.s[1]\n" + "fmla v31.4s, v13.4s, v0.s[1]\n" + "fsub v26.4s, v26.4s, v24.4s\n" + "mov v27.16b, v20.16b\n" + "str d26, [x22, x15]\n" + "mov v26.16b, v16.16b\n" + "str d25, [x12]\n" + "fmls v27.4s, v7.4s, v0.s[1]\n" + "str d31, [x12, %[output_col_stride1]]\n" + "fmls v26.4s, v8.4s, v0.s[1]\n" + "mov v25.16b, v17.16b\n" + "add x22, x22, #8\n" + "fsub v27.4s, v27.4s, v3.4s\n" + "mov v28.16b, v18.16b\n" + "fmla v27.4s, v21.4s, v0.s[1]\n" + "fsub v26.4s, v26.4s, v4.4s\n" + "fmla v26.4s, v22.4s, v0.s[1]\n" + "fmls v25.4s, v11.4s, v0.s[1]\n" + "fmls v28.4s, v15.4s, v0.s[1]\n" + "mov v12.16b, v12.16b\n" + "str d27, [x12, x11]\n" + "mov v19.16b, v19.16b\n" + "str d26, [x12, x13]\n" + "fsub v25.4s, v25.4s, v5.4s\n" + "fmla v25.4s, v23.4s, v0.s[1]\n" + "fsub v28.4s, v28.4s, v6.4s\n" + "fmla v28.4s, v24.4s, v0.s[1]\n" + "fmla v12.4s, v10.4s, v0.s[1]\n" + "fmla v19.4s, v9.4s, v0.s[1]\n" + "mov v20.16b, v20.16b\n" + "str d25, [x12, x23]\n" + "mov v16.16b, v16.16b\n" + "str d28, [x12, x15]\n" + "fsub v12.4s, v12.4s, v1.4s\n" + "fmls v12.4s, v14.4s, v0.s[1]\n" + "add x12, x12, #8\n" + "fsub v19.4s, v19.4s, v2.4s\n" + "fmla v20.4s, v7.4s, v0.s[1]\n" + "fmls v19.4s, v13.4s, v0.s[1]\n" + "fmla v16.4s, v8.4s, v0.s[1]\n" + "str d12, [x14]\n" + "mov v1.16b, v17.16b\n" + "fsub v20.4s, v20.4s, v3.4s\n" + "mov v17.16b, v18.16b\n" + "str d19, [x14, %[output_col_stride1]]\n" + "fmls v20.4s, v21.4s, v0.s[1]\n" + "fsub v16.4s, v16.4s, v4.4s\n" + "fmla v1.4s, v11.4s, v0.s[1]\n" + "fmls v16.4s, v22.4s, v0.s[1]\n" + "fmla v17.4s, v15.4s, v0.s[1]\n" + "str d20, [x14, x11]\n" + "fsub v1.4s, v1.4s, v5.4s\n" + "str d16, [x14, x13]\n" + "fmls v1.4s, v23.4s, v0.s[1]\n" + "fsub v17.4s, v17.4s, v6.4s\n" + "fmls v17.4s, v24.4s, v0.s[1]\n" + "str d1, [x14, x23]\n" + "str d17, [x14, x15]\n" + "add x14, x14, #8\n" + "ldr d2, [x27, x20]\n" + "mov v4.16b, v2.16b\n" + "ldr d17, [x27, x10]\n" + "mov v12.16b, v2.16b\n" + "ldr d18, [x27]\n" + "fmla v4.4s, v18.4s, v0.s[2]\n" + "ldr d3, [x27, x21]\n" + "mov v6.16b, v2.16b\n" + "ldr d5, [x27, x19]\n" + "mov v1.16b, v2.16b\n" + "ldr d18, [x27, %[input_col_stride1]]\n" + "fmls v4.4s, v17.4s, v0.s[3]\n" + "add x27, x27, #8\n" + "fmls v12.4s, v18.4s, v0.s[2]\n" + "sub %w[n_channels], %w[n_channels], #2\n" + "fmla v6.4s, v18.4s, v0.s[2]\n" + "fmls v1.4s, v18.4s, v0.s[1]\n" + "mov v2.16b, v2.16b\n" + "mov v3.16b, v3.16b\n" + "fmls v12.4s, v17.4s, v0.s[2]\n" + "mov v4.16b, v4.16b\n" + "fmls v6.4s, v17.4s, v0.s[2]\n" + "fsub v1.4s, v1.4s, v17.4s\n" + "fmla v1.4s, v5.4s, v0.s[1]\n" + "fmla v2.4s, v18.4s, v0.s[1]\n" + "fadd v12.4s, v12.4s, v5.4s\n" + "fmla v3.4s, v18.4s, v0.s[2]\n" + "fsub v6.4s, v6.4s, v5.4s\n" + "fmla v4.4s, v10.4s, v0.s[2]\n" + "fsub v2.4s, v2.4s, v17.4s\n" + "mov v16.16b, v12.16b\n" + "fmls v2.4s, v5.4s, v0.s[1]\n" + "fmls v3.4s, v5.4s, v0.s[3]\n" + "fmls v4.4s, v14.4s, v0.s[3]\n" + "fmla v16.4s, v9.4s, v0.s[2]\n" + "mov v5.16b, v6.16b\n" + "mov v6.16b, v1.16b\n" + "mov v9.16b, v2.16b\n" + "mov v10.16b, v3.16b\n" + "str d4, [x24]\n" + "fmls v16.4s, v13.4s, v0.s[3]\n" + "fmla v5.4s, v7.4s, v0.s[2]\n" + "fmla v6.4s, v8.4s, v0.s[2]\n" + "fmla v9.4s, v11.4s, v0.s[2]\n" + "fmla v10.4s, v15.4s, v0.s[2]\n" + "str d16, [x24, %[output_col_stride1]]\n" + "fmls v5.4s, v21.4s, v0.s[3]\n" + "fmls v6.4s, v22.4s, v0.s[3]\n" + "fmls v9.4s, v23.4s, v0.s[3]\n" + "fmls v10.4s, v24.4s, v0.s[3]\n" + "str d5, [x24, x11]\n" + "str d6, [x24, x13]\n" + "str d9, [x24, x23]\n" + "str d10, [x24, x15]\n" + "add x24, x24, #8\n" + "3:\n" + "cbz %w[n_channels], 4f\n" + "ldr s8, [%[inptr0], x20]\n" + "mov v14.16b, v8.16b\n" + "ldr s2, [%[inptr0], x10]\n" + "mov v10.16b, v8.16b\n" + "ldr s9, [%[inptr0]]\n" + "fmla v14.4s, v9.4s, v0.s[2]\n" + "ldr s1, [%[inptr0], x21]\n" + "mov v9.16b, v8.16b\n" + "ldr s4, [%[inptr0], x19]\n" + "mov v7.16b, v8.16b\n" + "ldr s12, [%[inptr0], %[input_col_stride1]]\n" + "fmls v14.4s, v2.4s, v0.s[3]\n" + "ldr s5, [x16, x20]\n" + "fmls v10.4s, v12.4s, v0.s[2]\n" + "ldr s20, [x16, x10]\n" + "fmla v9.4s, v12.4s, v0.s[2]\n" + "ldr s3, [x16]\n" + "fmls v7.4s, v12.4s, v0.s[1]\n" + "ldr s6, [x16, x21]\n" + "fmls v10.4s, v2.4s, v0.s[2]\n" + "ldr s16, [x16, x19]\n" + "fmls v9.4s, v2.4s, v0.s[2]\n" + "ldr s22, [x16, %[input_col_stride1]]\n" + "fsub v7.4s, v7.4s, v2.4s\n" + "ldr s17, [x17, x20]\n" + "fadd v10.4s, v10.4s, v4.4s\n" + "ldr s15, [x17, x10]\n" + "fsub v9.4s, v9.4s, v4.4s\n" + "ldr s19, [x17]\n" + "fmla v7.4s, v4.4s, v0.s[1]\n" + "ldr s18, [x17, x21]\n" + "mov v8.16b, v8.16b\n" + "ldr s13, [x17, x19]\n" + "mov v11.16b, v1.16b\n" + "ldr s21, [x17, %[input_col_stride1]]\n" + "fmla v8.4s, v12.4s, v0.s[1]\n" + "add %[inptr0], %[inptr0], #4\n" + "fmla v11.4s, v12.4s, v0.s[2]\n" + "add x16, x16, #4\n" + "mov v1.16b, v5.16b\n" + "add x17, x17, #4\n" + "fsub v8.4s, v8.4s, v2.4s\n" + "mov v2.16b, v5.16b\n" + "fmls v8.4s, v4.4s, v0.s[1]\n" + "fmls v11.4s, v4.4s, v0.s[3]\n" + "fmla v1.4s, v3.4s, v0.s[2]\n" + "fmls v2.4s, v22.4s, v0.s[2]\n" + "mov v3.16b, v5.16b\n" + "mov v4.16b, v5.16b\n" + "mov v5.16b, v5.16b\n" + "mov v6.16b, v6.16b\n" + "fmls v1.4s, v20.4s, v0.s[3]\n" + "fmls v2.4s, v20.4s, v0.s[2]\n" + "fmla v3.4s, v22.4s, v0.s[2]\n" + "fmls v4.4s, v22.4s, v0.s[1]\n" + "fmla v5.4s, v22.4s, v0.s[1]\n" + "fmla v6.4s, v22.4s, v0.s[2]\n" + "fadd v2.4s, v2.4s, v16.4s\n" + "mov v12.16b, v17.16b\n" + "fmls v3.4s, v20.4s, v0.s[2]\n" + "fsub v4.4s, v4.4s, v20.4s\n" + "fmla v4.4s, v16.4s, v0.s[1]\n" + "fsub v5.4s, v5.4s, v20.4s\n" + "fmls v5.4s, v16.4s, v0.s[1]\n" + "fmls v6.4s, v16.4s, v0.s[3]\n" + "fsub v3.4s, v3.4s, v16.4s\n" + "fmla v12.4s, v19.4s, v0.s[2]\n" + "mov v19.16b, v17.16b\n" + "mov v20.16b, v17.16b\n" + "mov v16.16b, v17.16b\n" + "mov v17.16b, v17.16b\n" + "fmls v12.4s, v15.4s, v0.s[3]\n" + "fmls v19.4s, v21.4s, v0.s[2]\n" + "fmla v20.4s, v21.4s, v0.s[2]\n" + "fmls v16.4s, v21.4s, v0.s[1]\n" + "fmla v17.4s, v21.4s, v0.s[1]\n" + "mov v18.16b, v18.16b\n" + "fmls v19.4s, v15.4s, v0.s[2]\n" + "mov v23.16b, v12.16b\n" + "fmls v20.4s, v15.4s, v0.s[2]\n" + "fsub v16.4s, v16.4s, v15.4s\n" + "fmla v16.4s, v13.4s, v0.s[1]\n" + "fsub v17.4s, v17.4s, v15.4s\n" + "fadd v19.4s, v19.4s, v13.4s\n" + "fmls v17.4s, v13.4s, v0.s[1]\n" + "fsub v20.4s, v20.4s, v13.4s\n" + "fmla v18.4s, v21.4s, v0.s[2]\n" + "fmla v23.4s, v14.4s, v0.s[2]\n" + "mov v15.16b, v19.16b\n" + "mov v14.16b, v20.16b\n" + "mov v24.16b, v16.16b\n" + "fmls v18.4s, v13.4s, v0.s[3]\n" + "fmla v15.4s, v10.4s, v0.s[2]\n" + "fmls v23.4s, v1.4s, v0.s[3]\n" + "fmla v14.4s, v9.4s, v0.s[2]\n" + "fmla v24.4s, v7.4s, v0.s[2]\n" + "mov v10.16b, v17.16b\n" + "fmls v15.4s, v2.4s, v0.s[3]\n" + "mov v7.16b, v18.16b\n" + "str s23, [%[outptr0]]\n" + "fmls v14.4s, v3.4s, v0.s[3]\n" + "fmls v24.4s, v4.4s, v0.s[3]\n" + "fmla v10.4s, v8.4s, v0.s[2]\n" + "str s15, [%[outptr0], %[output_col_stride1]]\n" + "fmla v7.4s, v11.4s, v0.s[2]\n" + "str s14, [%[outptr0], x11]\n" + "fmls v10.4s, v5.4s, v0.s[3]\n" + "str s24, [%[outptr0], x13]\n" + "fmls v7.4s, v6.4s, v0.s[3]\n" + "str s10, [%[outptr0], x23]\n" + "str s7, [%[outptr0], x15]\n" + "add %[outptr0], %[outptr0], #4\n" + "mov v26.16b, v12.16b\n" + "mov v25.16b, v19.16b\n" + "ldr s11, [x25, x20]\n" + "mov v10.16b, v11.16b\n" + "ldr s23, [x25, x10]\n" + "mov v9.16b, v11.16b\n" + "ldr s7, [x25]\n" + "fmla v10.4s, v7.4s, v0.s[2]\n" + "ldr s13, [x25, x21]\n" + "mov v7.16b, v11.16b\n" + "ldr s31, [x25, x19]\n" + "mov v8.16b, v11.16b\n" + "ldr s21, [x25, %[input_col_stride1]]\n" + "fmls v10.4s, v23.4s, v0.s[3]\n" + "ldr s30, [x26, x20]\n" + "fmls v9.4s, v21.4s, v0.s[2]\n" + "ldr s29, [x26, x10]\n" + "fmla v7.4s, v21.4s, v0.s[2]\n" + "ldr s22, [x26]\n" + "fmls v8.4s, v21.4s, v0.s[1]\n" + "ldr s24, [x26, x21]\n" + "fmls v9.4s, v23.4s, v0.s[2]\n" + "ldr s27, [x26, x19]\n" + "fmls v7.4s, v23.4s, v0.s[2]\n" + "ldr s28, [x26, %[input_col_stride1]]\n" + "fsub v8.4s, v8.4s, v23.4s\n" + "add x25, x25, #4\n" + "fadd v9.4s, v9.4s, v31.4s\n" + "add x26, x26, #4\n" + "fsub v7.4s, v7.4s, v31.4s\n" + "fmla v8.4s, v31.4s, v0.s[1]\n" + "mov v11.16b, v11.16b\n" + "mov v15.16b, v13.16b\n" + "mov v14.16b, v30.16b\n" + "mov v13.16b, v30.16b\n" + "fmla v11.4s, v21.4s, v0.s[1]\n" + "fmla v15.4s, v21.4s, v0.s[2]\n" + "fmla v14.4s, v22.4s, v0.s[2]\n" + "fmls v13.4s, v28.4s, v0.s[2]\n" + "mov v21.16b, v30.16b\n" + "mov v22.16b, v30.16b\n" + "fsub v11.4s, v11.4s, v23.4s\n" + "fmls v15.4s, v31.4s, v0.s[3]\n" + "fmls v11.4s, v31.4s, v0.s[1]\n" + "fmls v14.4s, v29.4s, v0.s[3]\n" + "fmls v13.4s, v29.4s, v0.s[2]\n" + "fmla v21.4s, v28.4s, v0.s[2]\n" + "fmls v22.4s, v28.4s, v0.s[1]\n" + "mov v23.16b, v30.16b\n" + "mov v24.16b, v24.16b\n" + "fmls v26.4s, v10.4s, v0.s[2]\n" + "fadd v13.4s, v13.4s, v27.4s\n" + "fmls v21.4s, v29.4s, v0.s[2]\n" + "fsub v22.4s, v22.4s, v29.4s\n" + "fmla v23.4s, v28.4s, v0.s[1]\n" + "fmla v22.4s, v27.4s, v0.s[1]\n" + "fmla v24.4s, v28.4s, v0.s[2]\n" + "fsub v21.4s, v21.4s, v27.4s\n" + "fmls v26.4s, v1.4s, v0.s[2]\n" + "fsub v23.4s, v23.4s, v29.4s\n" + "fmls v25.4s, v9.4s, v0.s[2]\n" + "fmls v23.4s, v27.4s, v0.s[1]\n" + "fmls v24.4s, v27.4s, v0.s[3]\n" + "fadd v26.4s, v26.4s, v14.4s\n" + "mov v27.16b, v20.16b\n" + "str s26, [x28]\n" + "fmls v25.4s, v2.4s, v0.s[2]\n" + "fmls v27.4s, v7.4s, v0.s[2]\n" + "mov v31.16b, v16.16b\n" + "mov v30.16b, v17.16b\n" + "mov v29.16b, v18.16b\n" + "fadd v25.4s, v25.4s, v13.4s\n" + "fmls v31.4s, v8.4s, v0.s[2]\n" + "str s25, [x28, %[output_col_stride1]]\n" + "fmls v27.4s, v3.4s, v0.s[2]\n" + "fmls v30.4s, v11.4s, v0.s[2]\n" + "fmls v29.4s, v15.4s, v0.s[2]\n" + "fmls v31.4s, v4.4s, v0.s[2]\n" + "mov v26.16b, v12.16b\n" + "fadd v27.4s, v27.4s, v21.4s\n" + "mov v25.16b, v19.16b\n" + "str s27, [x28, x11]\n" + "fmls v30.4s, v5.4s, v0.s[2]\n" + "fadd v31.4s, v31.4s, v22.4s\n" + "fmls v29.4s, v6.4s, v0.s[2]\n" + "str s31, [x28, x13]\n" + "fmla v26.4s, v10.4s, v0.s[2]\n" + "fadd v30.4s, v30.4s, v23.4s\n" + "fmla v25.4s, v9.4s, v0.s[2]\n" + "str s30, [x28, x23]\n" + "fadd v29.4s, v29.4s, v24.4s\n" + "str s29, [x28, x15]\n" + "fmls v26.4s, v1.4s, v0.s[2]\n" + "fmls v25.4s, v2.4s, v0.s[2]\n" + "add x28, x28, #4\n" + "mov v30.16b, v20.16b\n" + "mov v29.16b, v16.16b\n" + "fsub v26.4s, v26.4s, v14.4s\n" + "mov v28.16b, v17.16b\n" + "str s26, [x22]\n" + "fsub v25.4s, v25.4s, v13.4s\n" + "str s25, [x22, %[output_col_stride1]]\n" + "fmla v30.4s, v7.4s, v0.s[2]\n" + "fmla v29.4s, v8.4s, v0.s[2]\n" + "fmla v28.4s, v11.4s, v0.s[2]\n" + "mov v26.16b, v18.16b\n" + "mov v25.16b, v12.16b\n" + "fmls v30.4s, v3.4s, v0.s[2]\n" + "mov v31.16b, v19.16b\n" + "fmls v29.4s, v4.4s, v0.s[2]\n" + "fmls v28.4s, v5.4s, v0.s[2]\n" + "fmla v26.4s, v15.4s, v0.s[2]\n" + "fmls v25.4s, v10.4s, v0.s[1]\n" + "fsub v30.4s, v30.4s, v21.4s\n" + "fmls v31.4s, v9.4s, v0.s[1]\n" + "str s30, [x22, x11]\n" + "fsub v29.4s, v29.4s, v22.4s\n" + "str s29, [x22, x13]\n" + "fsub v28.4s, v28.4s, v23.4s\n" + "str s28, [x22, x23]\n" + "fmls v26.4s, v6.4s, v0.s[2]\n" + "fsub v25.4s, v25.4s, v1.4s\n" + "fsub v31.4s, v31.4s, v2.4s\n" + "fmla v25.4s, v14.4s, v0.s[1]\n" + "fmla v31.4s, v13.4s, v0.s[1]\n" + "fsub v26.4s, v26.4s, v24.4s\n" + "mov v27.16b, v20.16b\n" + "str s26, [x22, x15]\n" + "mov v26.16b, v16.16b\n" + "str s25, [x12]\n" + "fmls v27.4s, v7.4s, v0.s[1]\n" + "str s31, [x12, %[output_col_stride1]]\n" + "fmls v26.4s, v8.4s, v0.s[1]\n" + "mov v25.16b, v17.16b\n" + "add x22, x22, #4\n" + "fsub v27.4s, v27.4s, v3.4s\n" + "mov v28.16b, v18.16b\n" + "fmla v27.4s, v21.4s, v0.s[1]\n" + "fsub v26.4s, v26.4s, v4.4s\n" + "fmla v26.4s, v22.4s, v0.s[1]\n" + "fmls v25.4s, v11.4s, v0.s[1]\n" + "fmls v28.4s, v15.4s, v0.s[1]\n" + "mov v12.16b, v12.16b\n" + "str s27, [x12, x11]\n" + "mov v19.16b, v19.16b\n" + "str s26, [x12, x13]\n" + "fsub v25.4s, v25.4s, v5.4s\n" + "fmla v25.4s, v23.4s, v0.s[1]\n" + "fsub v28.4s, v28.4s, v6.4s\n" + "fmla v28.4s, v24.4s, v0.s[1]\n" + "fmla v12.4s, v10.4s, v0.s[1]\n" + "fmla v19.4s, v9.4s, v0.s[1]\n" + "mov v20.16b, v20.16b\n" + "str s25, [x12, x23]\n" + "mov v16.16b, v16.16b\n" + "str s28, [x12, x15]\n" + "fsub v12.4s, v12.4s, v1.4s\n" + "fmls v12.4s, v14.4s, v0.s[1]\n" + "add x12, x12, #4\n" + "fsub v19.4s, v19.4s, v2.4s\n" + "fmla v20.4s, v7.4s, v0.s[1]\n" + "fmls v19.4s, v13.4s, v0.s[1]\n" + "fmla v16.4s, v8.4s, v0.s[1]\n" + "str s12, [x14]\n" + "mov v1.16b, v17.16b\n" + "fsub v20.4s, v20.4s, v3.4s\n" + "mov v17.16b, v18.16b\n" + "str s19, [x14, %[output_col_stride1]]\n" + "fmls v20.4s, v21.4s, v0.s[1]\n" + "fsub v16.4s, v16.4s, v4.4s\n" + "fmla v1.4s, v11.4s, v0.s[1]\n" + "fmls v16.4s, v22.4s, v0.s[1]\n" + "fmla v17.4s, v15.4s, v0.s[1]\n" + "str s20, [x14, x11]\n" + "fsub v1.4s, v1.4s, v5.4s\n" + "str s16, [x14, x13]\n" + "fmls v1.4s, v23.4s, v0.s[1]\n" + "fsub v17.4s, v17.4s, v6.4s\n" + "fmls v17.4s, v24.4s, v0.s[1]\n" + "str s1, [x14, x23]\n" + "str s17, [x14, x15]\n" + "add x14, x14, #4\n" + "ldr s2, [x27, x20]\n" + "mov v4.16b, v2.16b\n" + "ldr s17, [x27, x10]\n" + "mov v12.16b, v2.16b\n" + "ldr s18, [x27]\n" + "fmla v4.4s, v18.4s, v0.s[2]\n" + "ldr s3, [x27, x21]\n" + "mov v6.16b, v2.16b\n" + "ldr s5, [x27, x19]\n" + "mov v1.16b, v2.16b\n" + "ldr s18, [x27, %[input_col_stride1]]\n" + "fmls v4.4s, v17.4s, v0.s[3]\n" + "add x27, x27, #4\n" + "fmls v12.4s, v18.4s, v0.s[2]\n" + "fmla v6.4s, v18.4s, v0.s[2]\n" + "fmls v1.4s, v18.4s, v0.s[1]\n" + "mov v2.16b, v2.16b\n" + "mov v3.16b, v3.16b\n" + "mov v4.16b, v4.16b\n" + "fmls v12.4s, v17.4s, v0.s[2]\n" + "fmls v6.4s, v17.4s, v0.s[2]\n" + "fsub v1.4s, v1.4s, v17.4s\n" + "fmla v2.4s, v18.4s, v0.s[1]\n" + "fmla v1.4s, v5.4s, v0.s[1]\n" + "fmla v3.4s, v18.4s, v0.s[2]\n" + "fadd v12.4s, v12.4s, v5.4s\n" + "fsub v6.4s, v6.4s, v5.4s\n" + "fsub v2.4s, v2.4s, v17.4s\n" + "fmla v4.4s, v10.4s, v0.s[2]\n" + "fmls v2.4s, v5.4s, v0.s[1]\n" + "fmls v3.4s, v5.4s, v0.s[3]\n" + "mov v16.16b, v12.16b\n" + "mov v5.16b, v6.16b\n" + "fmls v4.4s, v14.4s, v0.s[3]\n" + "mov v6.16b, v1.16b\n" + "fmla v16.4s, v9.4s, v0.s[2]\n" + "fmla v5.4s, v7.4s, v0.s[2]\n" + "fmla v6.4s, v8.4s, v0.s[2]\n" + "mov v9.16b, v2.16b\n" + "str s4, [x24]\n" + "mov v10.16b, v3.16b\n" + "fmls v16.4s, v13.4s, v0.s[3]\n" + "fmls v5.4s, v21.4s, v0.s[3]\n" + "fmls v6.4s, v22.4s, v0.s[3]\n" + "fmla v9.4s, v11.4s, v0.s[2]\n" + "fmla v10.4s, v15.4s, v0.s[2]\n" + "str s16, [x24, %[output_col_stride1]]\n" + "str s5, [x24, x11]\n" + "fmls v9.4s, v23.4s, v0.s[3]\n" + "str s6, [x24, x13]\n" + "fmls v10.4s, v24.4s, v0.s[3]\n" + "str s9, [x24, x23]\n" + "str s10, [x24, x15]\n" + "add x24, x24, #4\n" + "4:\n" + : [outptr0] "+r" (matrix_base), + [n_channels] "+r" (n_channels), + [inptr0] "+r" (input_base) + : [pcoeffs] "r" (pcoeffs), + [output_row_stride] "r" (6 * matrix_stride * sizeof(float)), + [output_col_stride1] "r" (matrix_stride * sizeof(float)), + [input_row_stride] "r" (input_row_stride * sizeof(float)), + [input_col_stride1] "r" (input_col_stride * sizeof(float)) + : "cc", "v0", "v1", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", + "v18", "v19", "v2", "v20", "v21", "v22", "v23", "v24", "v25", "v26", + "v27", "v28", "v29", "v3", "v30", "v31", "v4", "v5", "v6", "v7", "v8", + "v9", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x10", "x19", + "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "memory" + ); +} + +} // namespace input_transform +} // namespace winograd +} // namespace arm_conv + +#endif // __aarch64__ diff --git a/src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_1x8.cpp b/src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_1x8.cpp new file mode 100644 index 0000000000..2d6b333a59 --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_1x8.cpp @@ -0,0 +1,155 @@ +/* + * 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> +#include <arm_neon.h> + +namespace arm_conv { +namespace winograd { +namespace input_transform { + +void arm_fp32_1x8( + const unsigned int n_channels, + const float *const input_base, + size_t, // We don't need to stride over rows + const size_t input_col_stride, + float *outptr, + const size_t matrix_stride +) +{ + constexpr int inner_tile_cols = 8; + + // Get pointers into the input tile + const float *x_ptrs[inner_tile_cols]; + for (int j = 0, xj = 0; j < inner_tile_cols; 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; + 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 = 0; j < inner_tile_cols; 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 = 0; j < inner_tile_cols; 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; + } + for (; channels_remaining; channels_remaining--) + { + // Load x + for (int j = 0; j < inner_tile_cols; 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 input_transform +} // namespace winograd +} // namespace arm_conv diff --git a/src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_4x4.cpp b/src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_4x4.cpp new file mode 100644 index 0000000000..fae0173374 --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_4x4.cpp @@ -0,0 +1,251 @@ +/* + * 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> +#include <arm_neon.h> + +namespace arm_conv { +namespace winograd { +namespace input_transform { + +void arm_fp32_4x4( + const unsigned int n_channels, + const float *input_base, + const size_t input_row_stride, + const size_t input_col_stride, + float *outptr, + const size_t matrix_stride +) +{ + constexpr int inner_tile_rows = 4, inner_tile_cols = 4; + + // Get pointers into the input tile + const float *x_ptrs[inner_tile_rows][inner_tile_cols]; + for (int i = 0, xi = 0; i < inner_tile_rows; i++, xi++) + { + // Get a pointer into the row + const float* const row_ptr = input_base + xi*input_row_stride; + + for (int j = 0, xj = 0; j < inner_tile_cols; 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; + 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); + } + } + + // Load x + for (int i = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++) + { + x[i][j] = vld1q_f32(x_ptrs[i][j]); + x_ptrs[i][j] += 4; + } + } + + // Compute XT . x + for (int j = 0; j < inner_tile_cols; j++) + { + // XTx[0][j] = x[0][j] - x[2][j]; + XTx[0][j] = vsubq_f32(x[0][j], x[2][j]); + + // XTx[1][j] = x[1][j] + x[2][j]; + XTx[1][j] = vaddq_f32(x[1][j], x[2][j]); + + // XTx[2][j] = x[2][j] - x[1][j]; + XTx[2][j] = vsubq_f32(x[2][j], x[1][j]); + + // XTx[3][j] = x[1][j] - x[3][j]; + XTx[3][j] = vsubq_f32(x[1][j], x[3][j]); + } + + // Compute U = XT . x . X + for (int i = 0; i < inner_tile_rows; i++) + { + // U[i][0] = XTx[i][0] - XTx[i][2]; + U[i][0] = vsubq_f32(XTx[i][0], XTx[i][2]); + + // U[i][1] = XTx[i][1] + XTx[i][2]; + U[i][1] = vaddq_f32(XTx[i][1], XTx[i][2]); + + // U[i][2] = XTx[i][2] - XTx[i][1]; + U[i][2] = vsubq_f32(XTx[i][2], XTx[i][1]); + + // U[i][3] = XTx[i][1] - XTx[i][3]; + U[i][3] = vsubq_f32(XTx[i][1], XTx[i][3]); + } + + // 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; + } + 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); + } + } + + // Load x + for (int i = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++) + { + x[i][j] = vld1_f32(x_ptrs[i][j]); + x_ptrs[i][j] += 2; + } + } + + // Compute XT . x + for (int j = 0; j < inner_tile_cols; j++) + { + // XTx[0][j] = x[0][j] - x[2][j]; + XTx[0][j] = vsub_f32(x[0][j], x[2][j]); + + // XTx[1][j] = x[1][j] + x[2][j]; + XTx[1][j] = vadd_f32(x[1][j], x[2][j]); + + // XTx[2][j] = x[2][j] - x[1][j]; + XTx[2][j] = vsub_f32(x[2][j], x[1][j]); + + // XTx[3][j] = x[1][j] - x[3][j]; + XTx[3][j] = vsub_f32(x[1][j], x[3][j]); + } + + // Compute U = XT . x . X + for (int i = 0; i < inner_tile_rows; i++) + { + // U[i][0] = XTx[i][0] - XTx[i][2]; + U[i][0] = vsub_f32(XTx[i][0], XTx[i][2]); + + // U[i][1] = XTx[i][1] + XTx[i][2]; + U[i][1] = vadd_f32(XTx[i][1], XTx[i][2]); + + // U[i][2] = XTx[i][2] - XTx[i][1]; + U[i][2] = vsub_f32(XTx[i][2], XTx[i][1]); + + // U[i][3] = XTx[i][1] - XTx[i][3]; + U[i][3] = vsub_f32(XTx[i][1], XTx[i][3]); + } + + // 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; + } + for (; channels_remaining; channels_remaining--) + { + // Load x + for (int i = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++) + { + x[i][j] = *(x_ptrs[i][j]++); + } + } + + // Compute XT . x + for (int j = 0; j < inner_tile_cols; j++) + { + XTx[0][j] = x[0][j] - x[2][j]; + XTx[1][j] = x[1][j] + x[2][j]; + XTx[2][j] = x[2][j] - x[1][j]; + XTx[3][j] = x[1][j] - x[3][j]; + } + + // Compute U = XT . x . X + for (int i = 0; i < inner_tile_rows; i++) + { + U[i][0] = XTx[i][0] - XTx[i][2]; + U[i][1] = XTx[i][1] + XTx[i][2]; + U[i][2] = XTx[i][2] - XTx[i][1]; + U[i][3] = XTx[i][1] - XTx[i][3]; + } + + // 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 input_transform +} // namespace winograd +} // namespace arm_conv diff --git a/src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_6x6.cpp b/src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_6x6.cpp new file mode 100644 index 0000000000..4adc45768e --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/input_transforms/arm_fp32_6x6.cpp @@ -0,0 +1,202 @@ +/* + * 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. + */ + +#ifndef __aarch64__ + +#include <arm_neon.h> +#include <cstddef> + +namespace arm_conv { +namespace winograd { +namespace input_transform { + +void arm_fp32_6x6( + unsigned int n_channels, + const float* const input_base, + const size_t input_row_stride, + const size_t input_col_stride, + float* outptr, + const size_t matrix_stride +) +{ + constexpr int inner_tile_rows = 6; + constexpr int inner_tile_cols = 6; + + // Get pointers into the input tile + const float *x_ptrs[inner_tile_rows][inner_tile_cols]; + for (int i = 0, xi = 0; i < inner_tile_rows; i++, xi++) + { + // Get a pointer into the row + const float* const row_ptr = input_base + xi*input_row_stride; + + for (int j = 0, xj = 0; j < inner_tile_cols; 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; + 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 = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++) + { + x[i][j] = vld1_f32(x_ptrs[i][j]); + x_ptrs[i][j] += 2; + } + } + + // Compute XT . x + for (int j = 0; j < inner_tile_cols; 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; + } + for (; channels_remaining; channels_remaining--) + { + // Load x + for (int i = 0; i < inner_tile_rows; i++) + { + for (int j = 0; j < inner_tile_cols; j++) + { + x[i][j] = *(x_ptrs[i][j]++); + } + } + + // Compute XT . x + for (int j = 0; j < inner_tile_cols; 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 input_transform +} // namespace winograd +} // namespace arm_conv + +#endif // ! __aarch64__ diff --git a/src/core/NEON/kernels/convolution/winograd/input_transforms/sve_fp32_6x6.cpp b/src/core/NEON/kernels/convolution/winograd/input_transforms/sve_fp32_6x6.cpp new file mode 100644 index 0000000000..a2f096f489 --- /dev/null +++ b/src/core/NEON/kernels/convolution/winograd/input_transforms/sve_fp32_6x6.cpp @@ -0,0 +1,361 @@ +/* + * 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. + */ + +#if __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE) +#include <cstddef> + +namespace arm_conv { +namespace winograd { +namespace input_transform { + +void sve_fp32_6x6( + const unsigned int num_channels, + const float *input, + const size_t input_row_stride, + const size_t input_col_stride, + float *output, + const size_t output_col_stride +) +{ + const float B_values[4] = { 1.0f, 2.0f, 4.0f, 5.0f }; + long long_channels = num_channels; + + // Generated by armasmgen (February 04th, 2021) + __asm__ __volatile__( + "fmov z16.s, #4.0\n" + "ptrue p1.b\n" + "ld1rqw { z2.s }, p1/Z, [%x[B_values]]\n" + "add x16, %x[input_row_0], %x[input_row_stride], LSL #2\n" + "add x15, %x[output_row_0], %x[output_row_stride], LSL #2\n" + "add x14, %x[input_row_0], %x[input_row_stride], LSL #3\n" + "add x13, %x[output_row_0], %x[output_row_stride], LSL #3\n" + "add x12, x14, %x[input_row_stride], LSL #2\n" + "add x11, x13, %x[output_row_stride], LSL #2\n" + "add x10, %x[input_row_0], %x[input_row_stride], LSL #4\n" + "add x9, %x[output_row_0], %x[output_row_stride], LSL #4\n" + "add x28, x10, %x[input_row_stride], LSL #2\n" + "add x27, x9, %x[output_row_stride], LSL #2\n" + "lsl x26, %x[input_col_1_stride], #0x1\n" + "lsl x25, %x[output_col_1_stride], #0x1\n" + "add x24, x26, %x[input_col_1_stride]\n" + "add x23, x25, %x[output_col_1_stride]\n" + "lsl x22, %x[input_col_1_stride], #0x2\n" + "lsl x21, %x[output_col_1_stride], #0x2\n" + "add x20, x22, %x[input_col_1_stride]\n" + "add x19, x21, %x[output_col_1_stride]\n" + "whilelt p0.s, XZR, %x[num_channels]\n" + "beq 2f\n" + "1:" // channel_loop + "ld1w { z31.s }, p0/Z, [%x[input_row_0]]\n" + "decw %x[num_channels]\n" + "ld1w { z28.s }, p0/Z, [%x[input_row_0], %x[input_col_1_stride], LSL #2]\n" + "fmul z13.s, z28.s, z2.s[1]\n" + "ld1w { z27.s }, p0/Z, [%x[input_row_0], x26, LSL #2]\n" + "ld1w { z11.s }, p0/Z, [%x[input_row_0], x24, LSL #2]\n" + "fneg z13.s, p1/M, z13.s\n" + "ld1w { z7.s }, p0/Z, [%x[input_row_0], x22, LSL #2]\n" + "fsub z15.s, z7.s, z27.s\n" + "fmad z31.s, p1/M, z16.s, z7.s\n" + "ld1w { z3.s }, p0/Z, [%x[input_row_0], x20, LSL #2]\n" + "fmla z13.s, z11.s, z2.s[1]\n" + "ld1w { z12.s }, p0/Z, [x14]\n" + "incb %x[input_row_0]\n" + "fmls z31.s, z27.s, z2.s[3]\n" + "ld1w { z14.s }, p0/Z, [x14, %x[input_col_1_stride], LSL #2]\n" + "fsub z25.s, z15.s, z13.s\n" + "fadd z8.s, z13.s, z15.s\n" + "ld1w { z24.s }, p0/Z, [x14, x26, LSL #2]\n" + "fmsb z27.s, p1/M, z16.s, z7.s\n" + "ld1w { z22.s }, p0/Z, [x14, x24, LSL #2]\n" + "fmul z7.s, z28.s, z2.s[2]\n" + "ld1w { z1.s }, p0/Z, [x14, x22, LSL #2]\n" + "fsub z15.s, z1.s, z24.s\n" + "fneg z7.s, p1/M, z7.s\n" + "ld1w { z20.s }, p0/Z, [x14, x20, LSL #2]\n" + "fadd z7.s, z7.s, z11.s\n" + "ld1w { z29.s }, p0/Z, [x10]\n" + "incb x14\n" + "fmad z28.s, p1/M, z16.s, z3.s\n" + "ld1w { z10.s }, p0/Z, [x10, %x[input_col_1_stride], LSL #2]\n" + "fmad z12.s, p1/M, z16.s, z1.s\n" + "ld1w { z18.s }, p0/Z, [x10, x26, LSL #2]\n" + "fmul z13.s, z14.s, z2.s[1]\n" + "ld1w { z19.s }, p0/Z, [x10, x24, LSL #2]\n" + "fadd z17.s, z7.s, z27.s\n" + "ld1w { z9.s }, p0/Z, [x10, x22, LSL #2]\n" + "fsub z27.s, z27.s, z7.s\n" + "fmls z28.s, z11.s, z2.s[3]\n" + "ld1w { z21.s }, p0/Z, [x10, x20, LSL #2]\n" + "incb x10\n" + "fmls z12.s, z24.s, z2.s[3]\n" + "fneg z13.s, p1/M, z13.s\n" + "fmla z13.s, z22.s, z2.s[1]\n" + "fsub z30.s, z15.s, z13.s\n" + "fadd z4.s, z13.s, z15.s\n" + "fmsb z24.s, p1/M, z16.s, z1.s\n" + "fsub z15.s, z9.s, z18.s\n" + "fmul z1.s, z14.s, z2.s[2]\n" + "fmad z14.s, p1/M, z16.s, z20.s\n" + "fmad z29.s, p1/M, z16.s, z9.s\n" + "fmul z13.s, z10.s, z2.s[1]\n" + "fneg z1.s, p1/M, z1.s\n" + "fadd z1.s, z1.s, z22.s\n" + "fmls z14.s, z22.s, z2.s[3]\n" + "fmls z29.s, z18.s, z2.s[3]\n" + "fadd z5.s, z1.s, z24.s\n" + "fsub z24.s, z24.s, z1.s\n" + "fneg z13.s, p1/M, z13.s\n" + "fmla z13.s, z19.s, z2.s[1]\n" + "fsub z23.s, z15.s, z13.s\n" + "fadd z11.s, z13.s, z15.s\n" + "fmsb z18.s, p1/M, z16.s, z9.s\n" + "fmul z9.s, z10.s, z2.s[2]\n" + "fmad z10.s, p1/M, z16.s, z21.s\n" + "fmad z31.s, p1/M, z16.s, z29.s\n" + "fmad z8.s, p1/M, z16.s, z11.s\n" + "fneg z9.s, p1/M, z9.s\n" + "fadd z9.s, z9.s, z19.s\n" + "fmls z10.s, z19.s, z2.s[3]\n" + "fmls z31.s, z12.s, z2.s[3]\n" + "st1w { z31.s }, p0, [%x[output_row_0]]\n" + "fadd z26.s, z9.s, z18.s\n" + "fsub z18.s, z18.s, z9.s\n" + "fmls z8.s, z4.s, z2.s[3]\n" + "fmad z25.s, p1/M, z16.s, z23.s\n" + "fmad z28.s, p1/M, z16.s, z10.s\n" + "fmad z17.s, p1/M, z16.s, z26.s\n" + "fmad z27.s, p1/M, z16.s, z18.s\n" + "fmls z25.s, z30.s, z2.s[3]\n" + "fmls z28.s, z14.s, z2.s[3]\n" + "fmls z17.s, z5.s, z2.s[3]\n" + "st1w { z17.s }, p0, [%x[output_row_0], %x[output_col_1_stride], LSL #2]\n" + "fmls z27.s, z24.s, z2.s[3]\n" + "st1w { z27.s }, p0, [%x[output_row_0], x25, LSL #2]\n" + "st1w { z8.s }, p0, [%x[output_row_0], x23, LSL #2]\n" + "st1w { z25.s }, p0, [%x[output_row_0], x21, LSL #2]\n" + "st1w { z28.s }, p0, [%x[output_row_0], x19, LSL #2]\n" + "incb %x[output_row_0]\n" + "ld1w { z19.s }, p0/Z, [x16]\n" + "ld1w { z7.s }, p0/Z, [x16, %x[input_col_1_stride], LSL #2]\n" + "fmul z13.s, z7.s, z2.s[1]\n" + "ld1w { z6.s }, p0/Z, [x16, x26, LSL #2]\n" + "ld1w { z27.s }, p0/Z, [x16, x24, LSL #2]\n" + "fneg z13.s, p1/M, z13.s\n" + "ld1w { z25.s }, p0/Z, [x16, x22, LSL #2]\n" + "fsub z15.s, z25.s, z6.s\n" + "fmad z19.s, p1/M, z16.s, z25.s\n" + "ld1w { z20.s }, p0/Z, [x16, x20, LSL #2]\n" + "fmla z13.s, z27.s, z2.s[1]\n" + "ld1w { z0.s }, p0/Z, [x12]\n" + "incb x16\n" + "fmls z19.s, z6.s, z2.s[3]\n" + "ld1w { z31.s }, p0/Z, [x12, %x[input_col_1_stride], LSL #2]\n" + "fsub z8.s, z15.s, z13.s\n" + "fadd z28.s, z13.s, z15.s\n" + "ld1w { z1.s }, p0/Z, [x12, x26, LSL #2]\n" + "fmsb z6.s, p1/M, z16.s, z25.s\n" + "ld1w { z21.s }, p0/Z, [x12, x24, LSL #2]\n" + "fmul z25.s, z7.s, z2.s[2]\n" + "ld1w { z22.s }, p0/Z, [x12, x22, LSL #2]\n" + "fsub z15.s, z22.s, z1.s\n" + "fneg z25.s, p1/M, z25.s\n" + "ld1w { z17.s }, p0/Z, [x12, x20, LSL #2]\n" + "fadd z25.s, z25.s, z27.s\n" + "incb x12\n" + "fmad z7.s, p1/M, z16.s, z20.s\n" + "fmad z0.s, p1/M, z16.s, z22.s\n" + "fmul z13.s, z31.s, z2.s[1]\n" + "fadd z3.s, z25.s, z6.s\n" + "fsub z6.s, z6.s, z25.s\n" + "fmls z7.s, z27.s, z2.s[3]\n" + "fmls z0.s, z1.s, z2.s[3]\n" + "fneg z13.s, p1/M, z13.s\n" + "fmla z13.s, z21.s, z2.s[1]\n" + "fsub z9.s, z15.s, z13.s\n" + "fadd z27.s, z13.s, z15.s\n" + "fmsb z1.s, p1/M, z16.s, z22.s\n" + "fsub z15.s, z29.s, z12.s\n" + "fmul z22.s, z31.s, z2.s[2]\n" + "fmad z31.s, p1/M, z16.s, z17.s\n" + "fmul z13.s, z19.s, z2.s[1]\n" + "fmsb z12.s, p1/M, z16.s, z29.s\n" + "fneg z22.s, p1/M, z22.s\n" + "fadd z22.s, z22.s, z21.s\n" + "fmls z31.s, z21.s, z2.s[3]\n" + "fneg z13.s, p1/M, z13.s\n" + "fadd z25.s, z22.s, z1.s\n" + "fsub z1.s, z1.s, z22.s\n" + "fmla z13.s, z0.s, z2.s[1]\n" + "fmul z29.s, z19.s, z2.s[2]\n" + "fadd z22.s, z13.s, z15.s\n" + "st1w { z22.s }, p0, [x11]\n" + "fneg z29.s, p1/M, z29.s\n" + "fsub z22.s, z15.s, z13.s\n" + "fadd z29.s, z29.s, z0.s\n" + "st1w { z22.s }, p0, [x9]\n" + "fadd z22.s, z29.s, z12.s\n" + "fsub z15.s, z26.s, z5.s\n" + "fmul z13.s, z3.s, z2.s[1]\n" + "fsub z12.s, z12.s, z29.s\n" + "fmsb z5.s, p1/M, z16.s, z26.s\n" + "fmul z26.s, z3.s, z2.s[2]\n" + "fneg z13.s, p1/M, z13.s\n" + "fmla z13.s, z25.s, z2.s[1]\n" + "fneg z26.s, p1/M, z26.s\n" + "fadd z26.s, z26.s, z25.s\n" + "fadd z21.s, z13.s, z15.s\n" + "st1w { z21.s }, p0, [x11, %x[output_col_1_stride], LSL #2]\n" + "fsub z21.s, z15.s, z13.s\n" + "fmul z13.s, z6.s, z2.s[1]\n" + "fneg z13.s, p1/M, z13.s\n" + "st1w { z21.s }, p0, [x9, %x[output_col_1_stride], LSL #2]\n" + "fadd z21.s, z26.s, z5.s\n" + "fsub z15.s, z18.s, z24.s\n" + "fmla z13.s, z1.s, z2.s[1]\n" + "fsub z5.s, z5.s, z26.s\n" + "fmsb z24.s, p1/M, z16.s, z18.s\n" + "fmul z18.s, z6.s, z2.s[2]\n" + "fadd z20.s, z13.s, z15.s\n" + "st1w { z20.s }, p0, [x11, x25, LSL #2]\n" + "fneg z18.s, p1/M, z18.s\n" + "fsub z20.s, z15.s, z13.s\n" + "fadd z18.s, z18.s, z1.s\n" + "st1w { z20.s }, p0, [x9, x25, LSL #2]\n" + "fadd z20.s, z18.s, z24.s\n" + "fsub z15.s, z11.s, z4.s\n" + "fmul z13.s, z28.s, z2.s[1]\n" + "fsub z24.s, z24.s, z18.s\n" + "fmsb z4.s, p1/M, z16.s, z11.s\n" + "fmul z11.s, z28.s, z2.s[2]\n" + "fneg z13.s, p1/M, z13.s\n" + "fmla z13.s, z27.s, z2.s[1]\n" + "fneg z11.s, p1/M, z11.s\n" + "fadd z11.s, z11.s, z27.s\n" + "fadd z26.s, z13.s, z15.s\n" + "st1w { z26.s }, p0, [x11, x23, LSL #2]\n" + "fsub z26.s, z15.s, z13.s\n" + "fmul z13.s, z8.s, z2.s[1]\n" + "fneg z13.s, p1/M, z13.s\n" + "st1w { z26.s }, p0, [x9, x23, LSL #2]\n" + "fadd z26.s, z11.s, z4.s\n" + "fsub z15.s, z23.s, z30.s\n" + "fmla z13.s, z9.s, z2.s[1]\n" + "fsub z4.s, z4.s, z11.s\n" + "fmsb z30.s, p1/M, z16.s, z23.s\n" + "fmul z23.s, z8.s, z2.s[2]\n" + "fadd z18.s, z13.s, z15.s\n" + "st1w { z18.s }, p0, [x11, x21, LSL #2]\n" + "fneg z23.s, p1/M, z23.s\n" + "fsub z18.s, z15.s, z13.s\n" + "fadd z23.s, z23.s, z9.s\n" + "st1w { z18.s }, p0, [x9, x21, LSL #2]\n" + "fadd z18.s, z23.s, z30.s\n" + "fsub z15.s, z10.s, z14.s\n" + "fmul z13.s, z7.s, z2.s[1]\n" + "fsub z30.s, z30.s, z23.s\n" + "fmsb z14.s, p1/M, z16.s, z10.s\n" + "fmul z10.s, z7.s, z2.s[2]\n" + "fneg z13.s, p1/M, z13.s\n" + "fmla z13.s, z31.s, z2.s[1]\n" + "fneg z10.s, p1/M, z10.s\n" + "fadd z10.s, z10.s, z31.s\n" + "fadd z17.s, z13.s, z15.s\n" + "st1w { z17.s }, p0, [x11, x19, LSL #2]\n" + "fsub z17.s, z15.s, z13.s\n" + "incb x11\n" + "st1w { z17.s }, p0, [x9, x19, LSL #2]\n" + "fadd z17.s, z10.s, z14.s\n" + "fsub z14.s, z14.s, z10.s\n" + "st1w { z22.s }, p0, [x15]\n" + "incb x9\n" + "st1w { z12.s }, p0, [x13]\n" + "st1w { z21.s }, p0, [x15, %x[output_col_1_stride], LSL #2]\n" + "st1w { z5.s }, p0, [x13, %x[output_col_1_stride], LSL #2]\n" + "st1w { z20.s }, p0, [x15, x25, LSL #2]\n" + "st1w { z24.s }, p0, [x13, x25, LSL #2]\n" + "st1w { z26.s }, p0, [x15, x23, LSL #2]\n" + "st1w { z4.s }, p0, [x13, x23, LSL #2]\n" + "st1w { z18.s }, p0, [x15, x21, LSL #2]\n" + "st1w { z30.s }, p0, [x13, x21, LSL #2]\n" + "st1w { z17.s }, p0, [x15, x19, LSL #2]\n" + "incb x15\n" + "st1w { z14.s }, p0, [x13, x19, LSL #2]\n" + "incb x13\n" + "ld1w { z23.s }, p0/Z, [x28]\n" + "ld1w { z22.s }, p0/Z, [x28, %x[input_col_1_stride], LSL #2]\n" + "fmul z13.s, z22.s, z2.s[1]\n" + "ld1w { z21.s }, p0/Z, [x28, x26, LSL #2]\n" + "ld1w { z20.s }, p0/Z, [x28, x24, LSL #2]\n" + "fneg z13.s, p1/M, z13.s\n" + "ld1w { z26.s }, p0/Z, [x28, x22, LSL #2]\n" + "fsub z15.s, z26.s, z21.s\n" + "fmad z23.s, p1/M, z16.s, z26.s\n" + "ld1w { z18.s }, p0/Z, [x28, x20, LSL #2]\n" + "fmla z13.s, z20.s, z2.s[1]\n" + "incb x28\n" + "fmls z23.s, z21.s, z2.s[3]\n" + "fsub z17.s, z15.s, z13.s\n" + "fadd z30.s, z13.s, z15.s\n" + "fmsb z21.s, p1/M, z16.s, z26.s\n" + "fmul z26.s, z22.s, z2.s[2]\n" + "fmad z22.s, p1/M, z16.s, z18.s\n" + "fmad z19.s, p1/M, z16.s, z23.s\n" + "fmad z28.s, p1/M, z16.s, z30.s\n" + "fneg z26.s, p1/M, z26.s\n" + "fadd z26.s, z26.s, z20.s\n" + "fmls z22.s, z20.s, z2.s[3]\n" + "fmls z19.s, z0.s, z2.s[3]\n" + "st1w { z19.s }, p0, [x27]\n" + "fadd z23.s, z26.s, z21.s\n" + "fsub z21.s, z21.s, z26.s\n" + "fmls z28.s, z27.s, z2.s[3]\n" + "fmad z8.s, p1/M, z16.s, z17.s\n" + "fmad z7.s, p1/M, z16.s, z22.s\n" + "fmad z3.s, p1/M, z16.s, z23.s\n" + "fmad z6.s, p1/M, z16.s, z21.s\n" + "fmls z8.s, z9.s, z2.s[3]\n" + "fmls z7.s, z31.s, z2.s[3]\n" + "fmls z3.s, z25.s, z2.s[3]\n" + "st1w { z3.s }, p0, [x27, %x[output_col_1_stride], LSL #2]\n" + "fmls z6.s, z1.s, z2.s[3]\n" + "st1w { z6.s }, p0, [x27, x25, LSL #2]\n" + "st1w { z28.s }, p0, [x27, x23, LSL #2]\n" + "st1w { z8.s }, p0, [x27, x21, LSL #2]\n" + "st1w { z7.s }, p0, [x27, x19, LSL #2]\n" + "incb x27\n" + "whilelt p0.s, XZR, %x[num_channels]\n" + "bne 1b\n" + "2:" // channel_loop_end + + : [input_row_0] "+&r" (input), [num_channels] "+&r" (long_channels), [output_row_0] "+&r" (output) + : [B_values] "r" (B_values), [input_col_1_stride] "r" ((long) input_col_stride), [input_row_stride] "r" ((long) input_row_stride), [output_col_1_stride] "r" ((long) output_col_stride), [output_row_stride] "r" (6 * (long) output_col_stride) + : "cc", "memory", "p0", "p1", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + ); +} + +} // namespace input_transform +} // namespace winograd +} // namespace arm_conv + +#endif // __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE) |