1 /*
2 * Copyright (c) 2019 Guo Yejun
3 *
4 * This file is part of FFmpeg.
5 *
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.
10 *
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.
15 *
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
19 */
20
21 /**
22 * @file
23 * implementing a generic image processing filter using deep learning networks.
24 */
25
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"
31 #include "filters.h"
32 #include "dnn_filter_common.h"
33 #include "formats.h"
34 #include "internal.h"
35 #include "libswscale/swscale.h"
36 #include "libavutil/time.h"
37
38 typedef struct DnnProcessingContext {
39 const AVClass *class;
40 DnnContext dnnctx;
41 struct SwsContext *sws_uv_scale;
42 int sws_uv_height;
43 } DnnProcessingContext;
44
45 #define OFFSET(x) offsetof(DnnProcessingContext, dnnctx.x)
46 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
47 static const AVOption dnn_processing_options[] = {
48 { "dnn_backend", "DNN backend", OFFSET(backend_type), AV_OPT_TYPE_INT, { .i64 = 0 }, INT_MIN, INT_MAX, FLAGS, "backend" },
49 { "native", "native backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, 0, 0, FLAGS, "backend" },
50 #if (CONFIG_LIBTENSORFLOW == 1)
51 { "tensorflow", "tensorflow backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, 0, 0, FLAGS, "backend" },
52 #endif
53 #if (CONFIG_LIBOPENVINO == 1)
54 { "openvino", "openvino backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 2 }, 0, 0, FLAGS, "backend" },
55 #endif
56 DNN_COMMON_OPTIONS
57 { NULL }
58 };
59
60 AVFILTER_DEFINE_CLASS(dnn_processing);
61
init(AVFilterContext * context)62 static av_cold int init(AVFilterContext *context)
63 {
64 DnnProcessingContext *ctx = context->priv;
65 return ff_dnn_init(&ctx->dnnctx, DFT_PROCESS_FRAME, context);
66 }
67
query_formats(AVFilterContext * context)68 static int query_formats(AVFilterContext *context)
69 {
70 static const enum AVPixelFormat pix_fmts[] = {
71 AV_PIX_FMT_RGB24, AV_PIX_FMT_BGR24,
72 AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAYF32,
73 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
74 AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
75 AV_PIX_FMT_NV12,
76 AV_PIX_FMT_NONE
77 };
78 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
79 return ff_set_common_formats(context, fmts_list);
80 }
81
82 #define LOG_FORMAT_CHANNEL_MISMATCH() \
83 av_log(ctx, AV_LOG_ERROR, \
84 "the frame's format %s does not match " \
85 "the model input channel %d\n", \
86 av_get_pix_fmt_name(fmt), \
87 model_input->channels);
88
check_modelinput_inlink(const DNNData * model_input,const AVFilterLink * inlink)89 static int check_modelinput_inlink(const DNNData *model_input, const AVFilterLink *inlink)
90 {
91 AVFilterContext *ctx = inlink->dst;
92 enum AVPixelFormat fmt = inlink->format;
93
94 // the design is to add explicit scale filter before this filter
95 if (model_input->height != -1 && model_input->height != inlink->h) {
96 av_log(ctx, AV_LOG_ERROR, "the model requires frame height %d but got %d\n",
97 model_input->height, inlink->h);
98 return AVERROR(EIO);
99 }
100 if (model_input->width != -1 && model_input->width != inlink->w) {
101 av_log(ctx, AV_LOG_ERROR, "the model requires frame width %d but got %d\n",
102 model_input->width, inlink->w);
103 return AVERROR(EIO);
104 }
105 if (model_input->dt != DNN_FLOAT) {
106 avpriv_report_missing_feature(ctx, "data type rather than DNN_FLOAT");
107 return AVERROR(EIO);
108 }
109
110 switch (fmt) {
111 case AV_PIX_FMT_RGB24:
112 case AV_PIX_FMT_BGR24:
113 if (model_input->channels != 3) {
114 LOG_FORMAT_CHANNEL_MISMATCH();
115 return AVERROR(EIO);
116 }
117 return 0;
118 case AV_PIX_FMT_GRAYF32:
119 case AV_PIX_FMT_YUV420P:
120 case AV_PIX_FMT_YUV422P:
121 case AV_PIX_FMT_YUV444P:
122 case AV_PIX_FMT_YUV410P:
123 case AV_PIX_FMT_YUV411P:
124 case AV_PIX_FMT_NV12:
125 if (model_input->channels != 1) {
126 LOG_FORMAT_CHANNEL_MISMATCH();
127 return AVERROR(EIO);
128 }
129 return 0;
130 default:
131 avpriv_report_missing_feature(ctx, "%s", av_get_pix_fmt_name(fmt));
132 return AVERROR(EIO);
133 }
134
135 return 0;
136 }
137
config_input(AVFilterLink * inlink)138 static int config_input(AVFilterLink *inlink)
139 {
140 AVFilterContext *context = inlink->dst;
141 DnnProcessingContext *ctx = context->priv;
142 DNNReturnType result;
143 DNNData model_input;
144 int check;
145
146 result = ff_dnn_get_input(&ctx->dnnctx, &model_input);
147 if (result != DNN_SUCCESS) {
148 av_log(ctx, AV_LOG_ERROR, "could not get input from the model\n");
149 return AVERROR(EIO);
150 }
151
152 check = check_modelinput_inlink(&model_input, inlink);
153 if (check != 0) {
154 return check;
155 }
156
157 return 0;
158 }
159
isPlanarYUV(enum AVPixelFormat pix_fmt)160 static av_always_inline int isPlanarYUV(enum AVPixelFormat pix_fmt)
161 {
162 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
163 av_assert0(desc);
164 return !(desc->flags & AV_PIX_FMT_FLAG_RGB) && desc->nb_components == 3;
165 }
166
prepare_uv_scale(AVFilterLink * outlink)167 static int prepare_uv_scale(AVFilterLink *outlink)
168 {
169 AVFilterContext *context = outlink->src;
170 DnnProcessingContext *ctx = context->priv;
171 AVFilterLink *inlink = context->inputs[0];
172 enum AVPixelFormat fmt = inlink->format;
173
174 if (isPlanarYUV(fmt)) {
175 if (inlink->w != outlink->w || inlink->h != outlink->h) {
176 if (fmt == AV_PIX_FMT_NV12) {
177 ctx->sws_uv_scale = sws_getContext(inlink->w >> 1, inlink->h >> 1, AV_PIX_FMT_YA8,
178 outlink->w >> 1, outlink->h >> 1, AV_PIX_FMT_YA8,
179 SWS_BICUBIC, NULL, NULL, NULL);
180 ctx->sws_uv_height = inlink->h >> 1;
181 } else {
182 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(fmt);
183 int sws_src_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
184 int sws_src_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
185 int sws_dst_h = AV_CEIL_RSHIFT(outlink->h, desc->log2_chroma_h);
186 int sws_dst_w = AV_CEIL_RSHIFT(outlink->w, desc->log2_chroma_w);
187 ctx->sws_uv_scale = sws_getContext(sws_src_w, sws_src_h, AV_PIX_FMT_GRAY8,
188 sws_dst_w, sws_dst_h, AV_PIX_FMT_GRAY8,
189 SWS_BICUBIC, NULL, NULL, NULL);
190 ctx->sws_uv_height = sws_src_h;
191 }
192 }
193 }
194
195 return 0;
196 }
197
config_output(AVFilterLink * outlink)198 static int config_output(AVFilterLink *outlink)
199 {
200 AVFilterContext *context = outlink->src;
201 DnnProcessingContext *ctx = context->priv;
202 DNNReturnType result;
203 AVFilterLink *inlink = context->inputs[0];
204
205 // have a try run in case that the dnn model resize the frame
206 result = ff_dnn_get_output(&ctx->dnnctx, inlink->w, inlink->h, &outlink->w, &outlink->h);
207 if (result != DNN_SUCCESS) {
208 av_log(ctx, AV_LOG_ERROR, "could not get output from the model\n");
209 return AVERROR(EIO);
210 }
211
212 prepare_uv_scale(outlink);
213
214 return 0;
215 }
216
copy_uv_planes(DnnProcessingContext * ctx,AVFrame * out,const AVFrame * in)217 static int copy_uv_planes(DnnProcessingContext *ctx, AVFrame *out, const AVFrame *in)
218 {
219 const AVPixFmtDescriptor *desc;
220 int uv_height;
221
222 if (!ctx->sws_uv_scale) {
223 av_assert0(in->height == out->height && in->width == out->width);
224 desc = av_pix_fmt_desc_get(in->format);
225 uv_height = AV_CEIL_RSHIFT(in->height, desc->log2_chroma_h);
226 for (int i = 1; i < 3; ++i) {
227 int bytewidth = av_image_get_linesize(in->format, in->width, i);
228 av_image_copy_plane(out->data[i], out->linesize[i],
229 in->data[i], in->linesize[i],
230 bytewidth, uv_height);
231 }
232 } else if (in->format == AV_PIX_FMT_NV12) {
233 sws_scale(ctx->sws_uv_scale, (const uint8_t **)(in->data + 1), in->linesize + 1,
234 0, ctx->sws_uv_height, out->data + 1, out->linesize + 1);
235 } else {
236 sws_scale(ctx->sws_uv_scale, (const uint8_t **)(in->data + 1), in->linesize + 1,
237 0, ctx->sws_uv_height, out->data + 1, out->linesize + 1);
238 sws_scale(ctx->sws_uv_scale, (const uint8_t **)(in->data + 2), in->linesize + 2,
239 0, ctx->sws_uv_height, out->data + 2, out->linesize + 2);
240 }
241
242 return 0;
243 }
244
filter_frame(AVFilterLink * inlink,AVFrame * in)245 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
246 {
247 AVFilterContext *context = inlink->dst;
248 AVFilterLink *outlink = context->outputs[0];
249 DnnProcessingContext *ctx = context->priv;
250 DNNReturnType dnn_result;
251 AVFrame *out;
252
253 out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
254 if (!out) {
255 av_frame_free(&in);
256 return AVERROR(ENOMEM);
257 }
258 av_frame_copy_props(out, in);
259
260 dnn_result = ff_dnn_execute_model(&ctx->dnnctx, in, out);
261 if (dnn_result != DNN_SUCCESS){
262 av_log(ctx, AV_LOG_ERROR, "failed to execute model\n");
263 av_frame_free(&in);
264 av_frame_free(&out);
265 return AVERROR(EIO);
266 }
267
268 if (isPlanarYUV(in->format))
269 copy_uv_planes(ctx, out, in);
270
271 av_frame_free(&in);
272 return ff_filter_frame(outlink, out);
273 }
274
activate_sync(AVFilterContext * filter_ctx)275 static int activate_sync(AVFilterContext *filter_ctx)
276 {
277 AVFilterLink *inlink = filter_ctx->inputs[0];
278 AVFilterLink *outlink = filter_ctx->outputs[0];
279 AVFrame *in = NULL;
280 int64_t pts;
281 int ret, status;
282 int got_frame = 0;
283
284 FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);
285
286 do {
287 // drain all input frames
288 ret = ff_inlink_consume_frame(inlink, &in);
289 if (ret < 0)
290 return ret;
291 if (ret > 0) {
292 ret = filter_frame(inlink, in);
293 if (ret < 0)
294 return ret;
295 got_frame = 1;
296 }
297 } while (ret > 0);
298
299 // if frame got, schedule to next filter
300 if (got_frame)
301 return 0;
302
303 if (ff_inlink_acknowledge_status(inlink, &status, &pts)) {
304 if (status == AVERROR_EOF) {
305 ff_outlink_set_status(outlink, status, pts);
306 return ret;
307 }
308 }
309
310 FF_FILTER_FORWARD_WANTED(outlink, inlink);
311
312 return FFERROR_NOT_READY;
313 }
314
flush_frame(AVFilterLink * outlink,int64_t pts,int64_t * out_pts)315 static int flush_frame(AVFilterLink *outlink, int64_t pts, int64_t *out_pts)
316 {
317 DnnProcessingContext *ctx = outlink->src->priv;
318 int ret;
319 DNNAsyncStatusType async_state;
320
321 ret = ff_dnn_flush(&ctx->dnnctx);
322 if (ret != DNN_SUCCESS) {
323 return -1;
324 }
325
326 do {
327 AVFrame *in_frame = NULL;
328 AVFrame *out_frame = NULL;
329 async_state = ff_dnn_get_async_result(&ctx->dnnctx, &in_frame, &out_frame);
330 if (out_frame) {
331 if (isPlanarYUV(in_frame->format))
332 copy_uv_planes(ctx, out_frame, in_frame);
333 av_frame_free(&in_frame);
334 ret = ff_filter_frame(outlink, out_frame);
335 if (ret < 0)
336 return ret;
337 if (out_pts)
338 *out_pts = out_frame->pts + pts;
339 }
340 av_usleep(5000);
341 } while (async_state >= DAST_NOT_READY);
342
343 return 0;
344 }
345
activate_async(AVFilterContext * filter_ctx)346 static int activate_async(AVFilterContext *filter_ctx)
347 {
348 AVFilterLink *inlink = filter_ctx->inputs[0];
349 AVFilterLink *outlink = filter_ctx->outputs[0];
350 DnnProcessingContext *ctx = filter_ctx->priv;
351 AVFrame *in = NULL, *out = NULL;
352 int64_t pts;
353 int ret, status;
354 int got_frame = 0;
355 int async_state;
356
357 FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);
358
359 do {
360 // drain all input frames
361 ret = ff_inlink_consume_frame(inlink, &in);
362 if (ret < 0)
363 return ret;
364 if (ret > 0) {
365 out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
366 if (!out) {
367 av_frame_free(&in);
368 return AVERROR(ENOMEM);
369 }
370 av_frame_copy_props(out, in);
371 if (ff_dnn_execute_model_async(&ctx->dnnctx, in, out) != DNN_SUCCESS) {
372 return AVERROR(EIO);
373 }
374 }
375 } while (ret > 0);
376
377 // drain all processed frames
378 do {
379 AVFrame *in_frame = NULL;
380 AVFrame *out_frame = NULL;
381 async_state = ff_dnn_get_async_result(&ctx->dnnctx, &in_frame, &out_frame);
382 if (out_frame) {
383 if (isPlanarYUV(in_frame->format))
384 copy_uv_planes(ctx, out_frame, in_frame);
385 av_frame_free(&in_frame);
386 ret = ff_filter_frame(outlink, out_frame);
387 if (ret < 0)
388 return ret;
389 got_frame = 1;
390 }
391 } while (async_state == DAST_SUCCESS);
392
393 // if frame got, schedule to next filter
394 if (got_frame)
395 return 0;
396
397 if (ff_inlink_acknowledge_status(inlink, &status, &pts)) {
398 if (status == AVERROR_EOF) {
399 int64_t out_pts = pts;
400 ret = flush_frame(outlink, pts, &out_pts);
401 ff_outlink_set_status(outlink, status, out_pts);
402 return ret;
403 }
404 }
405
406 FF_FILTER_FORWARD_WANTED(outlink, inlink);
407
408 return 0;
409 }
410
activate(AVFilterContext * filter_ctx)411 static int activate(AVFilterContext *filter_ctx)
412 {
413 DnnProcessingContext *ctx = filter_ctx->priv;
414
415 if (ctx->dnnctx.async)
416 return activate_async(filter_ctx);
417 else
418 return activate_sync(filter_ctx);
419 }
420
uninit(AVFilterContext * ctx)421 static av_cold void uninit(AVFilterContext *ctx)
422 {
423 DnnProcessingContext *context = ctx->priv;
424
425 sws_freeContext(context->sws_uv_scale);
426 ff_dnn_uninit(&context->dnnctx);
427 }
428
429 static const AVFilterPad dnn_processing_inputs[] = {
430 {
431 .name = "default",
432 .type = AVMEDIA_TYPE_VIDEO,
433 .config_props = config_input,
434 },
435 { NULL }
436 };
437
438 static const AVFilterPad dnn_processing_outputs[] = {
439 {
440 .name = "default",
441 .type = AVMEDIA_TYPE_VIDEO,
442 .config_props = config_output,
443 },
444 { NULL }
445 };
446
447 AVFilter ff_vf_dnn_processing = {
448 .name = "dnn_processing",
449 .description = NULL_IF_CONFIG_SMALL("Apply DNN processing filter to the input."),
450 .priv_size = sizeof(DnnProcessingContext),
451 .init = init,
452 .uninit = uninit,
453 .query_formats = query_formats,
454 .inputs = dnn_processing_inputs,
455 .outputs = dnn_processing_outputs,
456 .priv_class = &dnn_processing_class,
457 .activate = activate,
458 };
459