From 5264b7d5555ec980f9c52c719122479d0d676af8 Mon Sep 17 00:00:00 2001 From: Pablo Tello Date: Mon, 21 Oct 2019 14:25:41 +0100 Subject: COMPMID-2576: Fuse activation in Winograd output transform. Change-Id: I26dd1307847adeaaefae0a7374b9858c07d71372 Signed-off-by: Pablo Tello Reviewed-on: https://review.mlplatform.org/c/2172 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice --- .../output_2x2_3x3_fp32_fp32_integers.cpp | 19 ++++++++++++++----- 1 file changed, 14 insertions(+), 5 deletions(-) (limited to 'src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_2x2_3x3_fp32_fp32_integers.cpp') diff --git a/src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_2x2_3x3_fp32_fp32_integers.cpp b/src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_2x2_3x3_fp32_fp32_integers.cpp index d6ebf44f41..5136bc15c4 100644 --- a/src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_2x2_3x3_fp32_fp32_integers.cpp +++ b/src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_2x2_3x3_fp32_fp32_integers.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -36,7 +36,9 @@ void OutputTransform<3, 3, 4, 4, float, float, WinogradRoots::Integers>::transfo const float* bptr, float* const output, const int output_row_stride, - const int output_col_stride + const int output_col_stride, + const float output_min, + const float output_max ) { // Construct a map to the output cells @@ -103,7 +105,10 @@ void OutputTransform<3, 3, 4, 4, float, float, WinogradRoots::Integers>::transfo { for (int j = 0; j < output_tile_cols; j++) { - vst1q_f32(outptrs[i][j], vaddq_f32(f[i][j], b)); + const auto y = + vmaxq_f32(vminq_f32(vaddq_f32(f[i][j], b), vdupq_n_f32(output_max)), + vdupq_n_f32(output_min)); + vst1q_f32(outptrs[i][j], y); outptrs[i][j] += 4; } } @@ -161,7 +166,10 @@ void OutputTransform<3, 3, 4, 4, float, float, WinogradRoots::Integers>::transfo { for (int j = 0; j < output_tile_cols; j++) { - vst1_f32(outptrs[i][j], vadd_f32(f[i][j], b)); + const auto y = + vmax_f32(vmin_f32(vadd_f32(f[i][j], b), vdup_n_f32(output_max)), + vdup_n_f32(output_min)); + vst1_f32(outptrs[i][j], y); outptrs[i][j] += 2; } } @@ -211,7 +219,8 @@ void OutputTransform<3, 3, 4, 4, float, float, WinogradRoots::Integers>::transfo { for (int j = 0; j < output_tile_cols; j++) { - *(outptrs[i][j]++) = f[i][j] + b; + const auto y = std::max(std::min(f[i][j] + b, output_max), output_min); + *(outptrs[i][j]++) = y; } } } -- cgit v1.2.1