2 * Copyright (c) 2019 Guo Yejun
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * implementing a generic image processing filter using deep learning networks.
26 #include "libavformat/avio.h"
27 #include "libavutil/opt.h"
28 #include "libavutil/pixdesc.h"
29 #include "libavutil/avassert.h"
30 #include "libavutil/imgutils.h"
32 #include "dnn_interface.h"
35 #include "libswscale/swscale.h"
37 typedef struct DnnProcessingContext {
41 DNNBackendType backend_type;
42 char *model_inputname;
43 char *model_outputname;
45 DNNModule *dnn_module;
48 // input & output of the model at execution time
52 struct SwsContext *sws_gray8_to_grayf32;
53 struct SwsContext *sws_grayf32_to_gray8;
54 struct SwsContext *sws_uv_scale;
56 } DnnProcessingContext;
58 #define OFFSET(x) offsetof(DnnProcessingContext, x)
59 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
60 static const AVOption dnn_processing_options[] = {
61 { "dnn_backend", "DNN backend", OFFSET(backend_type), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, FLAGS, "backend" },
62 { "native", "native backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, 0, 0, FLAGS, "backend" },
63 #if (CONFIG_LIBTENSORFLOW == 1)
64 { "tensorflow", "tensorflow backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, 0, 0, FLAGS, "backend" },
66 #if (CONFIG_LIBOPENVINO == 1)
67 { "openvino", "openvino backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 2 }, 0, 0, FLAGS, "backend" },
69 { "model", "path to model file", OFFSET(model_filename), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
70 { "input", "input name of the model", OFFSET(model_inputname), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
71 { "output", "output name of the model", OFFSET(model_outputname), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
75 AVFILTER_DEFINE_CLASS(dnn_processing);
77 static av_cold int init(AVFilterContext *context)
79 DnnProcessingContext *ctx = context->priv;
81 if (!ctx->model_filename) {
82 av_log(ctx, AV_LOG_ERROR, "model file for network is not specified\n");
83 return AVERROR(EINVAL);
85 if (!ctx->model_inputname) {
86 av_log(ctx, AV_LOG_ERROR, "input name of the model network is not specified\n");
87 return AVERROR(EINVAL);
89 if (!ctx->model_outputname) {
90 av_log(ctx, AV_LOG_ERROR, "output name of the model network is not specified\n");
91 return AVERROR(EINVAL);
94 ctx->dnn_module = ff_get_dnn_module(ctx->backend_type);
95 if (!ctx->dnn_module) {
96 av_log(ctx, AV_LOG_ERROR, "could not create DNN module for requested backend\n");
97 return AVERROR(ENOMEM);
99 if (!ctx->dnn_module->load_model) {
100 av_log(ctx, AV_LOG_ERROR, "load_model for network is not specified\n");
101 return AVERROR(EINVAL);
104 ctx->model = (ctx->dnn_module->load_model)(ctx->model_filename);
106 av_log(ctx, AV_LOG_ERROR, "could not load DNN model\n");
107 return AVERROR(EINVAL);
113 static int query_formats(AVFilterContext *context)
115 static const enum AVPixelFormat pix_fmts[] = {
116 AV_PIX_FMT_RGB24, AV_PIX_FMT_BGR24,
117 AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAYF32,
118 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
119 AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
122 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
123 return ff_set_common_formats(context, fmts_list);
126 #define LOG_FORMAT_CHANNEL_MISMATCH() \
127 av_log(ctx, AV_LOG_ERROR, \
128 "the frame's format %s does not match " \
129 "the model input channel %d\n", \
130 av_get_pix_fmt_name(fmt), \
131 model_input->channels);
133 static int check_modelinput_inlink(const DNNData *model_input, const AVFilterLink *inlink)
135 AVFilterContext *ctx = inlink->dst;
136 enum AVPixelFormat fmt = inlink->format;
138 // the design is to add explicit scale filter before this filter
139 if (model_input->height != -1 && model_input->height != inlink->h) {
140 av_log(ctx, AV_LOG_ERROR, "the model requires frame height %d but got %d\n",
141 model_input->height, inlink->h);
144 if (model_input->width != -1 && model_input->width != inlink->w) {
145 av_log(ctx, AV_LOG_ERROR, "the model requires frame width %d but got %d\n",
146 model_input->width, inlink->w);
151 case AV_PIX_FMT_RGB24:
152 case AV_PIX_FMT_BGR24:
153 if (model_input->channels != 3) {
154 LOG_FORMAT_CHANNEL_MISMATCH();
157 if (model_input->dt != DNN_FLOAT && model_input->dt != DNN_UINT8) {
158 av_log(ctx, AV_LOG_ERROR, "only support dnn models with input data type as float32 and uint8.\n");
162 case AV_PIX_FMT_GRAY8:
163 if (model_input->channels != 1) {
164 LOG_FORMAT_CHANNEL_MISMATCH();
167 if (model_input->dt != DNN_UINT8) {
168 av_log(ctx, AV_LOG_ERROR, "only support dnn models with input data type uint8.\n");
172 case AV_PIX_FMT_GRAYF32:
173 case AV_PIX_FMT_YUV420P:
174 case AV_PIX_FMT_YUV422P:
175 case AV_PIX_FMT_YUV444P:
176 case AV_PIX_FMT_YUV410P:
177 case AV_PIX_FMT_YUV411P:
178 if (model_input->channels != 1) {
179 LOG_FORMAT_CHANNEL_MISMATCH();
182 if (model_input->dt != DNN_FLOAT) {
183 av_log(ctx, AV_LOG_ERROR, "only support dnn models with input data type float32.\n");
188 av_log(ctx, AV_LOG_ERROR, "%s not supported.\n", av_get_pix_fmt_name(fmt));
195 static int config_input(AVFilterLink *inlink)
197 AVFilterContext *context = inlink->dst;
198 DnnProcessingContext *ctx = context->priv;
199 DNNReturnType result;
203 result = ctx->model->get_input(ctx->model->model, &model_input, ctx->model_inputname);
204 if (result != DNN_SUCCESS) {
205 av_log(ctx, AV_LOG_ERROR, "could not get input from the model\n");
209 check = check_modelinput_inlink(&model_input, inlink);
214 ctx->input.width = inlink->w;
215 ctx->input.height = inlink->h;
216 ctx->input.channels = model_input.channels;
217 ctx->input.dt = model_input.dt;
219 result = (ctx->model->set_input_output)(ctx->model->model,
220 &ctx->input, ctx->model_inputname,
221 (const char **)&ctx->model_outputname, 1);
222 if (result != DNN_SUCCESS) {
223 av_log(ctx, AV_LOG_ERROR, "could not set input and output for the model\n");
230 static int prepare_sws_context(AVFilterLink *outlink)
232 AVFilterContext *context = outlink->src;
233 DnnProcessingContext *ctx = context->priv;
234 AVFilterLink *inlink = context->inputs[0];
235 enum AVPixelFormat fmt = inlink->format;
236 DNNDataType input_dt = ctx->input.dt;
237 DNNDataType output_dt = ctx->output.dt;
240 case AV_PIX_FMT_RGB24:
241 case AV_PIX_FMT_BGR24:
242 if (input_dt == DNN_FLOAT) {
243 ctx->sws_gray8_to_grayf32 = sws_getContext(inlink->w * 3,
249 0, NULL, NULL, NULL);
251 if (output_dt == DNN_FLOAT) {
252 ctx->sws_grayf32_to_gray8 = sws_getContext(outlink->w * 3,
258 0, NULL, NULL, NULL);
261 case AV_PIX_FMT_YUV420P:
262 case AV_PIX_FMT_YUV422P:
263 case AV_PIX_FMT_YUV444P:
264 case AV_PIX_FMT_YUV410P:
265 case AV_PIX_FMT_YUV411P:
266 av_assert0(input_dt == DNN_FLOAT);
267 av_assert0(output_dt == DNN_FLOAT);
268 ctx->sws_gray8_to_grayf32 = sws_getContext(inlink->w,
274 0, NULL, NULL, NULL);
275 ctx->sws_grayf32_to_gray8 = sws_getContext(outlink->w,
281 0, NULL, NULL, NULL);
283 if (inlink->w != outlink->w || inlink->h != outlink->h) {
284 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(fmt);
285 int sws_src_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
286 int sws_src_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
287 int sws_dst_h = AV_CEIL_RSHIFT(outlink->h, desc->log2_chroma_h);
288 int sws_dst_w = AV_CEIL_RSHIFT(outlink->w, desc->log2_chroma_w);
289 ctx->sws_uv_scale = sws_getContext(sws_src_w, sws_src_h, AV_PIX_FMT_GRAY8,
290 sws_dst_w, sws_dst_h, AV_PIX_FMT_GRAY8,
291 SWS_BICUBIC, NULL, NULL, NULL);
292 ctx->sws_uv_height = sws_src_h;
303 static int config_output(AVFilterLink *outlink)
305 AVFilterContext *context = outlink->src;
306 DnnProcessingContext *ctx = context->priv;
307 DNNReturnType result;
309 // have a try run in case that the dnn model resize the frame
310 result = (ctx->dnn_module->execute_model)(ctx->model, &ctx->output, 1);
311 if (result != DNN_SUCCESS){
312 av_log(ctx, AV_LOG_ERROR, "failed to execute model\n");
316 outlink->w = ctx->output.width;
317 outlink->h = ctx->output.height;
319 prepare_sws_context(outlink);
324 static int copy_from_frame_to_dnn(DnnProcessingContext *ctx, const AVFrame *frame)
326 int bytewidth = av_image_get_linesize(frame->format, frame->width, 0);
327 DNNData *dnn_input = &ctx->input;
329 switch (frame->format) {
330 case AV_PIX_FMT_RGB24:
331 case AV_PIX_FMT_BGR24:
332 if (dnn_input->dt == DNN_FLOAT) {
333 sws_scale(ctx->sws_gray8_to_grayf32, (const uint8_t **)frame->data, frame->linesize,
334 0, frame->height, (uint8_t * const*)(&dnn_input->data),
335 (const int [4]){frame->width * 3 * sizeof(float), 0, 0, 0});
337 av_assert0(dnn_input->dt == DNN_UINT8);
338 av_image_copy_plane(dnn_input->data, bytewidth,
339 frame->data[0], frame->linesize[0],
340 bytewidth, frame->height);
343 case AV_PIX_FMT_GRAY8:
344 case AV_PIX_FMT_GRAYF32:
345 av_image_copy_plane(dnn_input->data, bytewidth,
346 frame->data[0], frame->linesize[0],
347 bytewidth, frame->height);
349 case AV_PIX_FMT_YUV420P:
350 case AV_PIX_FMT_YUV422P:
351 case AV_PIX_FMT_YUV444P:
352 case AV_PIX_FMT_YUV410P:
353 case AV_PIX_FMT_YUV411P:
354 sws_scale(ctx->sws_gray8_to_grayf32, (const uint8_t **)frame->data, frame->linesize,
355 0, frame->height, (uint8_t * const*)(&dnn_input->data),
356 (const int [4]){frame->width * sizeof(float), 0, 0, 0});
365 static int copy_from_dnn_to_frame(DnnProcessingContext *ctx, AVFrame *frame)
367 int bytewidth = av_image_get_linesize(frame->format, frame->width, 0);
368 DNNData *dnn_output = &ctx->output;
370 switch (frame->format) {
371 case AV_PIX_FMT_RGB24:
372 case AV_PIX_FMT_BGR24:
373 if (dnn_output->dt == DNN_FLOAT) {
374 sws_scale(ctx->sws_grayf32_to_gray8, (const uint8_t *[4]){(const uint8_t *)dnn_output->data, 0, 0, 0},
375 (const int[4]){frame->width * 3 * sizeof(float), 0, 0, 0},
376 0, frame->height, (uint8_t * const*)frame->data, frame->linesize);
379 av_assert0(dnn_output->dt == DNN_UINT8);
380 av_image_copy_plane(frame->data[0], frame->linesize[0],
381 dnn_output->data, bytewidth,
382 bytewidth, frame->height);
385 case AV_PIX_FMT_GRAY8:
386 // it is possible that data type of dnn output is float32,
387 // need to add support for such case when needed.
388 av_assert0(dnn_output->dt == DNN_UINT8);
389 av_image_copy_plane(frame->data[0], frame->linesize[0],
390 dnn_output->data, bytewidth,
391 bytewidth, frame->height);
393 case AV_PIX_FMT_GRAYF32:
394 av_assert0(dnn_output->dt == DNN_FLOAT);
395 av_image_copy_plane(frame->data[0], frame->linesize[0],
396 dnn_output->data, bytewidth,
397 bytewidth, frame->height);
399 case AV_PIX_FMT_YUV420P:
400 case AV_PIX_FMT_YUV422P:
401 case AV_PIX_FMT_YUV444P:
402 case AV_PIX_FMT_YUV410P:
403 case AV_PIX_FMT_YUV411P:
404 sws_scale(ctx->sws_grayf32_to_gray8, (const uint8_t *[4]){(const uint8_t *)dnn_output->data, 0, 0, 0},
405 (const int[4]){frame->width * sizeof(float), 0, 0, 0},
406 0, frame->height, (uint8_t * const*)frame->data, frame->linesize);
415 static av_always_inline int isPlanarYUV(enum AVPixelFormat pix_fmt)
417 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
419 return !(desc->flags & AV_PIX_FMT_FLAG_RGB) && desc->nb_components == 3;
422 static int copy_uv_planes(DnnProcessingContext *ctx, AVFrame *out, const AVFrame *in)
424 const AVPixFmtDescriptor *desc;
427 if (!ctx->sws_uv_scale) {
428 av_assert0(in->height == out->height && in->width == out->width);
429 desc = av_pix_fmt_desc_get(in->format);
430 uv_height = AV_CEIL_RSHIFT(in->height, desc->log2_chroma_h);
431 for (int i = 1; i < 3; ++i) {
432 int bytewidth = av_image_get_linesize(in->format, in->width, i);
433 av_image_copy_plane(out->data[i], out->linesize[i],
434 in->data[i], in->linesize[i],
435 bytewidth, uv_height);
438 sws_scale(ctx->sws_uv_scale, (const uint8_t **)(in->data + 1), in->linesize + 1,
439 0, ctx->sws_uv_height, out->data + 1, out->linesize + 1);
440 sws_scale(ctx->sws_uv_scale, (const uint8_t **)(in->data + 2), in->linesize + 2,
441 0, ctx->sws_uv_height, out->data + 2, out->linesize + 2);
447 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
449 AVFilterContext *context = inlink->dst;
450 AVFilterLink *outlink = context->outputs[0];
451 DnnProcessingContext *ctx = context->priv;
452 DNNReturnType dnn_result;
455 copy_from_frame_to_dnn(ctx, in);
457 dnn_result = (ctx->dnn_module->execute_model)(ctx->model, &ctx->output, 1);
458 if (dnn_result != DNN_SUCCESS){
459 av_log(ctx, AV_LOG_ERROR, "failed to execute model\n");
464 out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
467 return AVERROR(ENOMEM);
470 av_frame_copy_props(out, in);
471 copy_from_dnn_to_frame(ctx, out);
473 if (isPlanarYUV(in->format))
474 copy_uv_planes(ctx, out, in);
477 return ff_filter_frame(outlink, out);
480 static av_cold void uninit(AVFilterContext *ctx)
482 DnnProcessingContext *context = ctx->priv;
484 sws_freeContext(context->sws_gray8_to_grayf32);
485 sws_freeContext(context->sws_grayf32_to_gray8);
486 sws_freeContext(context->sws_uv_scale);
488 if (context->dnn_module)
489 (context->dnn_module->free_model)(&context->model);
491 av_freep(&context->dnn_module);
494 static const AVFilterPad dnn_processing_inputs[] = {
497 .type = AVMEDIA_TYPE_VIDEO,
498 .config_props = config_input,
499 .filter_frame = filter_frame,
504 static const AVFilterPad dnn_processing_outputs[] = {
507 .type = AVMEDIA_TYPE_VIDEO,
508 .config_props = config_output,
513 AVFilter ff_vf_dnn_processing = {
514 .name = "dnn_processing",
515 .description = NULL_IF_CONFIG_SMALL("Apply DNN processing filter to the input."),
516 .priv_size = sizeof(DnnProcessingContext),
519 .query_formats = query_formats,
520 .inputs = dnn_processing_inputs,
521 .outputs = dnn_processing_outputs,
522 .priv_class = &dnn_processing_class,