COMPMID-3108: Add Winograd 3x3,4x4 FP16 support for NEON
Change-Id: I20680dc74a3d709297539e2132417308a7aecc9d
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3159
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_4x4_3x3_fp16_fp16_integers.cpp b/src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_4x4_3x3_fp16_fp16_integers.cpp
new file mode 100644
index 0000000..3c4f8b4
--- /dev/null
+++ b/src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_4x4_3x3_fp16_fp16_integers.cpp
@@ -0,0 +1,259 @@
+/*
+ * Copyright (c) 2020 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 __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+#include "arm.hpp"
+#include "kernel.hpp"
+
+namespace winograd
+{
+
+template <>
+void WeightTransform<3, 3, 6, 6, __fp16, __fp16, WinogradRoots::Integers>::execute(
+ const int n_output_channels,
+ const int n_input_channels,
+ const __fp16* const input, // NOTE: Data in HWIO order
+ __fp16* const output,
+ const int matrix_stride,
+ const int matrix_row_stride
+)
+{
+ // Get pointers to each cell of the weight tensor
+ const auto weight_col_stride = n_input_channels * n_output_channels;
+ const auto weight_row_stride = 3 * weight_col_stride;
+ const __fp16 *inptrs[3][3];
+ for (int i = 0; i < 3; i++)
+ {
+ for (int j = 0; j < 3; j++)
+ {
+ inptrs[i][j] = input + i*weight_row_stride + j*weight_col_stride;
+ }
+ }
+
+ // For each input channel
+ for (int ic = 0; ic < n_input_channels; ic++)
+ {
+ __fp16 *outptr = output + ic * matrix_row_stride;
+
+ // For each output channel
+ int channels_remaining = n_output_channels;
+#ifdef __aarch64__
+ for (; channels_remaining >= 8; channels_remaining -= 8)
+ {
+ // Matrices used and computed in this kernel
+ float16x8_t w[3][3], Ww[6][3], V[6][6];
+
+ // Read weights
+ for (int i = 0; i < 3; i++)
+ {
+ for (int j = 0; j < 3; j++)
+ {
+ w[i][j] = vld1q_f16(inptrs[i][j]);
+ inptrs[i][j] += 8;
+ }
+ }
+
+ // Compute the matrix W w
+ for (int j = 0; j < 3; j++)
+ {
+ // Ww[0][j] = 6*w[0][j];
+ Ww[0][j] = vmulq_n_f16(w[0][j], 6.0);
+
+ // Ww[1][j] = -4*w[0][j] + -4*w[1][j] + -4*w[2][j];
+ Ww[1][j] = vmulq_n_f16(vaddq_f16(vaddq_f16(w[0][j], w[1][j]), w[2][j]), -4.0);
+
+ // Ww[2][j] = -4*w[0][j] + 4*w[1][j] + -4*w[2][j];
+ Ww[2][j] = vmulq_n_f16(vsubq_f16(vsubq_f16(w[1][j], w[0][j]), w[2][j]), 4.0);
+
+ // Ww[3][j] = 1*w[0][j] + 2*w[1][j] + 4*w[2][j];
+ Ww[3][j] = vaddq_f16(vaddq_f16(w[0][j], vmulq_f16(w[1][j], vdupq_n_f16(2.0f))), vmulq_f16(w[2][j], vdupq_n_f16(4.0f)));
+
+ // Ww[4][j] = 1*w[0][j] + -2*w[1][j] + 4*w[2][j];
+ Ww[4][j] = vaddq_f16(vsubq_f16(w[0][j], vmulq_f16(w[1][j], vdupq_n_f16(2.0f))), vmulq_f16(w[2][j], vdupq_n_f16(4.0f)));
+
+ // Ww[5][j] = 24*w[2][j];
+ Ww[5][j] = vmulq_n_f16(w[2][j], 24.0f);
+ }
+
+ // Compute V = W w WT
+ for (int i = 0; i < 6; i++)
+ {
+ const float recip576 = 1.0f / 576.0f;
+
+ // V[i][0] = 6*Ww[i][0];
+ V[i][0] = vmulq_n_f16(vmulq_n_f16(Ww[i][0], 6.0), recip576);
+
+ // V[i][1] = -4*Ww[i][0] + -4*Ww[i][1] + -4*Ww[i][2];
+ V[i][1] = vmulq_n_f16(vmulq_n_f16(vaddq_f16(vaddq_f16(Ww[i][0], Ww[i][1]), Ww[i][2]), -4.0), recip576);
+
+ // V[i][2] = -4*Ww[i][0] + 4*Ww[i][1] + -4*Ww[i][2];
+ V[i][2] = vmulq_n_f16(vmulq_n_f16(vsubq_f16(vsubq_f16(Ww[i][1], Ww[i][0]), Ww[i][2]), 4.0), recip576);
+
+ // V[i][3] = 1*Ww[i][0] + 2*Ww[i][1] + 4*Ww[i][2];
+ V[i][3] = vmulq_n_f16(vaddq_f16(vaddq_f16(Ww[i][0], vmulq_f16(Ww[i][1], vdupq_n_f16(2.0f))), vmulq_f16(Ww[i][2], vdupq_n_f16(4.0f))), recip576);
+
+ // V[i][4] = 1*Ww[i][0] + -2*Ww[i][1] + 4*Ww[i][2];
+ V[i][4] = vmulq_n_f16(vaddq_f16(vsubq_f16(Ww[i][0], vmulq_f16(Ww[i][1], vdupq_n_f16(2.0f))), vmulq_f16(Ww[i][2], vdupq_n_f16(4.0f))), recip576);
+
+ // V[i][5] = 24*Ww[i][2];
+ V[i][5] = vmulq_n_f16(vmulq_n_f16(Ww[i][2], 24.0f), recip576);
+ }
+
+ // Store the transformed weights
+ for (int i = 0, m = 0; i < 6; i++)
+ {
+ for (int j = 0; j < 6; j++, m++)
+ {
+ vst1q_f16(outptr + m*matrix_stride, V[i][j]);
+ }
+ }
+ outptr += 8;
+ }
+#endif // __aarch64__
+#ifdef __arm_any__
+ for (; channels_remaining >= 4; channels_remaining -= 4)
+ {
+ // Matrices used and computed in this kernel
+ float16x4_t w[3][3], Ww[6][3], V[6][6];
+
+ // Read weights
+ for (int i = 0; i < 3; i++)
+ {
+ for (int j = 0; j < 3; j++)
+ {
+ w[i][j] = vld1_f16(inptrs[i][j]);
+ inptrs[i][j] += 4;
+ }
+ }
+
+ // Compute the matrix W w
+ for (int j = 0; j < 3; j++)
+ {
+ // Ww[0][j] = 6*w[0][j];
+ Ww[0][j] = vmul_n_f16(w[0][j], 6.0);
+
+ // Ww[1][j] = -4*w[0][j] + -4*w[1][j] + -4*w[2][j];
+ Ww[1][j] = vmul_n_f16(vadd_f16(vadd_f16(w[0][j], w[1][j]), w[2][j]), -4.0);
+
+ // Ww[2][j] = -4*w[0][j] + 4*w[1][j] + -4*w[2][j];
+ Ww[2][j] = vmul_n_f16(vsub_f16(vsub_f16(w[1][j], w[0][j]), w[2][j]), 4.0);
+
+ // Ww[3][j] = 1*w[0][j] + 2*w[1][j] + 4*w[2][j];
+ Ww[3][j] = vadd_f16(vadd_f16(w[0][j], vmul_f16(w[1][j], vdup_n_f16(2.0f))), vmul_f16(w[2][j], vdup_n_f16(4.0f)));
+
+ // Ww[4][j] = 1*w[0][j] + -2*w[1][j] + 4*w[2][j];
+ Ww[4][j] = vadd_f16(vsub_f16(w[0][j], vmul_f16(w[1][j], vdup_n_f16(2.0f))), vmul_f16(w[2][j], vdup_n_f16(4.0f)));
+
+ // Ww[5][j] = 24*w[2][j];
+ Ww[5][j] = vmul_n_f16(w[2][j], 24.0f);
+ }
+
+ // Compute V = W w WT
+ for (int i = 0; i < 6; i++)
+ {
+ const float recip576 = 1.0f / 576.0f;
+
+ // V[i][0] = 6*Ww[i][0];
+ V[i][0] = vmul_n_f16(vmul_n_f16(Ww[i][0], 6.0), recip576);
+
+ // V[i][1] = -4*Ww[i][0] + -4*Ww[i][1] + -4*Ww[i][2];
+ V[i][1] = vmul_n_f16(vmul_n_f16(vadd_f16(vadd_f16(Ww[i][0], Ww[i][1]), Ww[i][2]), -4.0), recip576);
+
+ // V[i][2] = -4*Ww[i][0] + 4*Ww[i][1] + -4*Ww[i][2];
+ V[i][2] = vmul_n_f16(vmul_n_f16(vsub_f16(vsub_f16(Ww[i][1], Ww[i][0]), Ww[i][2]), 4.0), recip576);
+
+ // V[i][3] = 1*Ww[i][0] + 2*Ww[i][1] + 4*Ww[i][2];
+ V[i][3] = vmul_n_f16(vadd_f16(vadd_f16(Ww[i][0], vmul_f16(Ww[i][1], vdup_n_f16(2.0f))), vmul_f16(Ww[i][2], vdup_n_f16(4.0f))), recip576);
+
+ // V[i][4] = 1*Ww[i][0] + -2*Ww[i][1] + 4*Ww[i][2];
+ V[i][4] = vmul_n_f16(vadd_f16(vsub_f16(Ww[i][0], vmul_f16(Ww[i][1], vdup_n_f16(2.0f))), vmul_f16(Ww[i][2], vdup_n_f16(4.0f))), recip576);
+
+ // V[i][5] = 24*Ww[i][2];
+ V[i][5] = vmul_n_f16(vmul_n_f16(Ww[i][2], 24.0f), recip576);
+ }
+
+ // Store the transformed weights
+ for (int i = 0, m = 0; i < 6; i++)
+ {
+ for (int j = 0; j < 6; j++, m++)
+ {
+ vst1_f16(outptr + m*matrix_stride, V[i][j]);
+ }
+ }
+ outptr += 4;
+ }
+#endif // __arm_any__
+ for (; channels_remaining; channels_remaining--)
+ {
+ // Matrices used and computed in this kernel
+ __fp16 w[3][3], Ww[6][3], V[6][6];
+
+ // Read weights
+ for (int i = 0; i < 3; i++)
+ {
+ for (int j = 0; j < 3; j++)
+ {
+ w[i][j] = *(inptrs[i][j]++);
+ }
+ }
+
+ // Compute the matrix W w
+ for (int j = 0; j < 3; j++)
+ {
+ Ww[0][j] = 6*w[0][j];
+ Ww[1][j] = -4*w[0][j] + -4*w[1][j] + -4*w[2][j];
+ Ww[2][j] = -4*w[0][j] + 4*w[1][j] + -4*w[2][j];
+ Ww[3][j] = 1*w[0][j] + 2*w[1][j] + 4*w[2][j];
+ Ww[4][j] = 1*w[0][j] + -2*w[1][j] + 4*w[2][j];
+ Ww[5][j] = 24*w[2][j];
+ }
+
+ // Compute V = W w WT
+ for (int i = 0; i < 6; i++)
+ {
+ V[i][0] = ( 6*Ww[i][0]) / 576.0;
+ V[i][1] = (-4*Ww[i][0] + -4*Ww[i][1] + -4*Ww[i][2]) / 576.0;
+ V[i][2] = (-4*Ww[i][0] + 4*Ww[i][1] + -4*Ww[i][2]) / 576.0;
+ V[i][3] = ( 1*Ww[i][0] + 2*Ww[i][1] + 4*Ww[i][2]) / 576.0;
+ V[i][4] = ( 1*Ww[i][0] + -2*Ww[i][1] + 4*Ww[i][2]) / 576.0;
+ V[i][5] = (24*Ww[i][2]) / 576.0;
+ }
+
+ // Store the transformed weights
+ for (int i = 0, m = 0; i < 6; i++)
+ {
+ for (int j = 0; j < 6; j++, m++)
+ {
+ *(outptr + m*matrix_stride) = V[i][j];
+ }
+ }
+ outptr++;
+ }
+ }
+}
+
+template class WeightTransform<3, 3, 6, 6, __fp16, __fp16, WinogradRoots::Integers>;
+
+} // namespace
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC