From patchwork Tue Jul 16 05:55:53 2019 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: "Guo, Yejun" X-Patchwork-Id: 13961 Return-Path: X-Original-To: patchwork@ffaux-bg.ffmpeg.org Delivered-To: patchwork@ffaux-bg.ffmpeg.org Received: from ffbox0-bg.mplayerhq.hu (ffbox0-bg.ffmpeg.org [79.124.17.100]) by ffaux.localdomain (Postfix) with ESMTP id D5716448391 for ; Tue, 16 Jul 2019 08:58:38 +0300 (EEST) Received: from [127.0.1.1] (localhost [127.0.0.1]) by ffbox0-bg.mplayerhq.hu (Postfix) with ESMTP id B0B1C68AD27; Tue, 16 Jul 2019 08:58:38 +0300 (EEST) X-Original-To: ffmpeg-devel@ffmpeg.org Delivered-To: ffmpeg-devel@ffmpeg.org Received: from mga05.intel.com (mga05.intel.com [192.55.52.43]) by ffbox0-bg.mplayerhq.hu (Postfix) with ESMTPS id 7C02368A830 for ; Tue, 16 Jul 2019 08:58:32 +0300 (EEST) X-Amp-Result: SKIPPED(no attachment in message) X-Amp-File-Uploaded: False Received: from fmsmga008.fm.intel.com ([10.253.24.58]) by fmsmga105.fm.intel.com with ESMTP/TLS/DHE-RSA-AES256-GCM-SHA384; 15 Jul 2019 22:58:30 -0700 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.63,496,1557212400"; d="scan'208";a="167552299" Received: from yguo18-skl-u1604.sh.intel.com ([10.239.13.25]) by fmsmga008.fm.intel.com with ESMTP; 15 Jul 2019 22:58:30 -0700 From: "Guo, Yejun" To: ffmpeg-devel@ffmpeg.org Date: Tue, 16 Jul 2019 13:55:53 +0800 Message-Id: <1563256553-28073-1-git-send-email-yejun.guo@intel.com> X-Mailer: git-send-email 2.7.4 Subject: [FFmpeg-devel] [PATCH 2/4] dnn: add layer pad which is equivalent to tf.pad X-BeenThere: ffmpeg-devel@ffmpeg.org X-Mailman-Version: 2.1.20 Precedence: list List-Id: FFmpeg development discussions and patches List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Reply-To: FFmpeg development discussions and patches Cc: yejun.guo@intel.com MIME-Version: 1.0 Errors-To: ffmpeg-devel-bounces@ffmpeg.org Sender: "ffmpeg-devel" the reason to add this layer first is that vf_sr uses it in its tensorflow model, and the next plan is to update the python script to convert tf.pad into native model. Signed-off-by: Guo, Yejun --- libavfilter/dnn/Makefile | 1 + libavfilter/dnn/dnn_backend_native_layer_pad.c | 211 +++++++++++++++++++++++++ libavfilter/dnn/dnn_backend_native_layer_pad.h | 40 +++++ 3 files changed, 252 insertions(+) create mode 100644 libavfilter/dnn/dnn_backend_native_layer_pad.c create mode 100644 libavfilter/dnn/dnn_backend_native_layer_pad.h diff --git a/libavfilter/dnn/Makefile b/libavfilter/dnn/Makefile index 1d12ade..83938e5 100644 --- a/libavfilter/dnn/Makefile +++ b/libavfilter/dnn/Makefile @@ -1,5 +1,6 @@ OBJS-$(CONFIG_DNN) += dnn/dnn_interface.o OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native.o +OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_pad.o DNN-OBJS-$(CONFIG_LIBTENSORFLOW) += dnn/dnn_backend_tf.o diff --git a/libavfilter/dnn/dnn_backend_native_layer_pad.c b/libavfilter/dnn/dnn_backend_native_layer_pad.c new file mode 100644 index 0000000..aa12f7f --- /dev/null +++ b/libavfilter/dnn/dnn_backend_native_layer_pad.c @@ -0,0 +1,211 @@ +/* + * Copyright (c) 2019 Guo Yejun + * + * This file is part of FFmpeg. + * + * FFmpeg is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * FFmpeg is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#include +#include "libavutil/avassert.h" +#include "dnn_backend_native_layer_pad.h" + +static int before_get_buddy(int given, int paddings, LayerPadModeParam mode) +{ + if (mode == LPMP_SYMMETRIC) { + return (2 * paddings - 1 - given); + } else if (mode == LPMP_REFLECT) { + return (2 * paddings - given); + } else { + av_assert0(!"should not reach here"); + return 0; + } +} + +static int after_get_buddy(int given, int border, LayerPadModeParam mode) +{ + if (mode == LPMP_SYMMETRIC) { + int offset = given - border; + return (border - 1 - offset); + } else if (mode == LPMP_REFLECT) { + int offset = given - border; + return (border - 2 - offset); + } else { + av_assert0(!"should not reach here"); + return 0; + } +} + +void dnn_execute_layer_pad(const float *input, float *output, const LayerPadParams *params, int number, int height, int width, int channel) +{ + int32_t before_paddings; + int32_t after_paddings; + + // suppose format is + int new_number = number + params->paddings[0][0] + params->paddings[0][1]; + int new_height = height + params->paddings[1][0] + params->paddings[1][1]; + int new_width = width + params->paddings[2][0] + params->paddings[2][1]; + int new_channel = channel + params->paddings[3][0] + params->paddings[3][1]; + + int c_stride = channel; + int wc_stride = c_stride * width; + int hwc_stride = wc_stride * height; + + int new_c_stride = new_channel; + int new_wc_stride = new_c_stride * new_width; + int new_hwc_stride = new_wc_stride * new_height; + + // copy the original data + for (int n = 0; n < number; n++) { + for (int h = 0; h < height; h++) { + for (int w = 0; w < width; w++) { + const float *src = input + n * hwc_stride + h * wc_stride + w * c_stride; + float *dst = output + (n + params->paddings[0][0]) * new_hwc_stride + + (h + params->paddings[1][0]) * new_wc_stride + + (w + params->paddings[2][0]) * new_c_stride + + params->paddings[3][0]; + memcpy(dst, src, channel * sizeof(float)); + } + } + } + + // handle the first dimension + before_paddings = params->paddings[0][0]; + after_paddings = params->paddings[0][1]; + for (int n = 0; n < before_paddings; n++) { + float *dst = output + n * new_hwc_stride; + if (params->mode == LPMP_CONSTANT) { + for (int i = 0; i < new_hwc_stride; i++) { + *dst = params->constant_values; + } + } + else { + int buddy = before_get_buddy(n, before_paddings, params->mode); + float *src = output + buddy * new_hwc_stride; + memcpy(dst, src, new_hwc_stride * sizeof(float)); + } + } + for (int n = 0; n < after_paddings; n++) { + int given = number + before_paddings + n; + float *dst = output + given * new_hwc_stride; + if (params->mode == LPMP_CONSTANT) { + for (int i = 0; i < new_hwc_stride; i++) { + *dst = params->constant_values; + } + } else { + int buddy = after_get_buddy(given, number + before_paddings, params->mode); + float *src = output + buddy * new_hwc_stride; + memcpy(dst, src, new_hwc_stride * sizeof(float)); + } + } + + // handle the second dimension + before_paddings = params->paddings[1][0]; + after_paddings = params->paddings[1][1]; + for (int n = 0; n < new_number; n++) { + float *start = output + n * new_hwc_stride; + for (int h = 0; h < before_paddings; h++) { + float *dst = start + h * new_wc_stride; + if (params->mode == LPMP_CONSTANT) { + for (int i = 0; i < new_wc_stride; i++) { + *dst = params->constant_values; + } + } else { + int buddy = before_get_buddy(h, before_paddings, params->mode); + float *src = start + buddy * new_wc_stride; + memcpy(dst, src, new_wc_stride * sizeof(float)); + } + } + for (int h = 0; h < after_paddings; h++) { + int given = height + before_paddings + h; + float *dst = start + given * new_wc_stride; + if (params->mode == LPMP_CONSTANT) { + for (int i = 0; i < new_wc_stride; i++) { + *dst = params->constant_values; + } + } else { + int buddy = after_get_buddy(given, height + before_paddings, params->mode); + float *src = start + buddy * new_wc_stride; + memcpy(dst, src, new_wc_stride * sizeof(float)); + } + } + } + + // handle the third dimension + before_paddings = params->paddings[2][0]; + after_paddings = params->paddings[2][1]; + for (int n = 0; n < new_number; n++) { + for (int h = 0; h < new_height; h++) { + float *start = output + n * new_hwc_stride + h * new_wc_stride; + for (int w = 0; w < before_paddings; w++) { + float *dst = start + w * new_c_stride; + if (params->mode == LPMP_CONSTANT) { + for (int i = 0; i < new_c_stride; i++) { + *dst = params->constant_values; + } + } else { + int buddy = before_get_buddy(w, before_paddings, params->mode); + float *src = start + buddy * new_c_stride; + memcpy(dst, src, new_c_stride * sizeof(float)); + } + } + for (int w = 0; w < after_paddings; w++) { + int given = width + before_paddings + w; + float *dst = start + given * new_c_stride; + if (params->mode == LPMP_CONSTANT) { + for (int i = 0; i < new_c_stride; i++) { + *dst = params->constant_values; + } + } else { + int buddy = after_get_buddy(given, width + before_paddings, params->mode); + float *src = start + buddy * new_c_stride; + memcpy(dst, src, new_c_stride * sizeof(float)); + } + } + } + } + + // handle the fourth dimension + before_paddings = params->paddings[3][0]; + after_paddings = params->paddings[3][1]; + for (int n = 0; n < new_number; n++) { + for (int h = 0; h < new_height; h++) { + for (int w = 0; w < new_width; w++) { + float *start = output + n * new_hwc_stride + h * new_wc_stride + w * new_c_stride; + for (int c = 0; c < before_paddings; c++) { + float *dst = start + c; + if (params->mode == LPMP_CONSTANT) { + *dst = params->constant_values; + } else { + int buddy = before_get_buddy(c, before_paddings, params->mode); + float *src = start + buddy; + *dst = *src; + } + } + for (int c = 0; c < after_paddings; c++) { + int given = channel + before_paddings + c; + float *dst = start + given; + if (params->mode == LPMP_CONSTANT) { + *dst = params->constant_values; + } else { + int buddy = after_get_buddy(given, channel + before_paddings, params->mode); + float *src = start + buddy; + *dst = *src; + } + } + } + } + } +} diff --git a/libavfilter/dnn/dnn_backend_native_layer_pad.h b/libavfilter/dnn/dnn_backend_native_layer_pad.h new file mode 100644 index 0000000..465e96a --- /dev/null +++ b/libavfilter/dnn/dnn_backend_native_layer_pad.h @@ -0,0 +1,40 @@ +/* + * Copyright (c) 2019 Guo Yejun + * + * This file is part of FFmpeg. + * + * FFmpeg is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * FFmpeg is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +/** + * @file + * layer pad (equivalent to tf.pad) for native backend. + */ +#include + +#ifndef AVFILTER_DNN_BACKEND_NATIVE_LAYER_PAD_H +#define AVFILTER_DNN_BACKEND_NATIVE_LAYER_PAD_H + +typedef enum {LPMP_CONSTANT, LPMP_REFLECT, LPMP_SYMMETRIC} LayerPadModeParam; + +typedef struct LayerPadParams{ + int32_t paddings[4][2]; + LayerPadModeParam mode; + float constant_values; +} LayerPadParams; + +void dnn_execute_layer_pad(const float *input, float *output, const LayerPadParams *params, int number, int height, int width, int channel); + +#endif