1 /*
2 * Copyright (c) 2015 Paul B Mahol
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 * Adaptive Temporal Averaging Denoiser,
24 * based on paper "Video Denoising Based on Adaptive Temporal Averaging" by
25 * David Bartovčak and Miroslav Vrankić
26 */
27
28 #include "libavutil/imgutils.h"
29 #include "libavutil/opt.h"
30 #include "libavutil/pixdesc.h"
31 #include "avfilter.h"
32
33 #define FF_BUFQUEUE_SIZE 129
34 #include "bufferqueue.h"
35
36 #include "atadenoise.h"
37 #include "formats.h"
38 #include "internal.h"
39 #include "video.h"
40
41 #define SIZE FF_BUFQUEUE_SIZE
42
43 typedef struct ATADenoiseContext {
44 const AVClass *class;
45
46 float fthra[4], fthrb[4];
47 int thra[4], thrb[4];
48 int algorithm;
49
50 int planes;
51 int nb_planes;
52 int planewidth[4];
53 int planeheight[4];
54
55 struct FFBufQueue q;
56 void *data[4][SIZE];
57 int linesize[4][SIZE];
58 int size, mid;
59 int available;
60
61 int (*filter_slice)(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs);
62
63 ATADenoiseDSPContext dsp;
64 } ATADenoiseContext;
65
66 #define OFFSET(x) offsetof(ATADenoiseContext, x)
67 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_RUNTIME_PARAM
68 #define VF AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
69
70 static const AVOption atadenoise_options[] = {
71 { "0a", "set threshold A for 1st plane", OFFSET(fthra[0]), AV_OPT_TYPE_FLOAT, {.dbl=0.02}, 0, 0.3, FLAGS },
72 { "0b", "set threshold B for 1st plane", OFFSET(fthrb[0]), AV_OPT_TYPE_FLOAT, {.dbl=0.04}, 0, 5.0, FLAGS },
73 { "1a", "set threshold A for 2nd plane", OFFSET(fthra[1]), AV_OPT_TYPE_FLOAT, {.dbl=0.02}, 0, 0.3, FLAGS },
74 { "1b", "set threshold B for 2nd plane", OFFSET(fthrb[1]), AV_OPT_TYPE_FLOAT, {.dbl=0.04}, 0, 5.0, FLAGS },
75 { "2a", "set threshold A for 3rd plane", OFFSET(fthra[2]), AV_OPT_TYPE_FLOAT, {.dbl=0.02}, 0, 0.3, FLAGS },
76 { "2b", "set threshold B for 3rd plane", OFFSET(fthrb[2]), AV_OPT_TYPE_FLOAT, {.dbl=0.04}, 0, 5.0, FLAGS },
77 { "s", "set how many frames to use", OFFSET(size), AV_OPT_TYPE_INT, {.i64=9}, 5, SIZE, VF },
78 { "p", "set what planes to filter", OFFSET(planes), AV_OPT_TYPE_FLAGS, {.i64=7}, 0, 15, FLAGS },
79 { "a", "set variant of algorithm", OFFSET(algorithm),AV_OPT_TYPE_INT, {.i64=PARALLEL}, 0, NB_ATAA-1, FLAGS, "a" },
80 { "p", "parallel", 0, AV_OPT_TYPE_CONST, {.i64=PARALLEL}, 0, 0, FLAGS, "a" },
81 { "s", "serial", 0, AV_OPT_TYPE_CONST, {.i64=SERIAL}, 0, 0, FLAGS, "a" },
82 { NULL }
83 };
84
85 AVFILTER_DEFINE_CLASS(atadenoise);
86
query_formats(AVFilterContext * ctx)87 static int query_formats(AVFilterContext *ctx)
88 {
89 static const enum AVPixelFormat pixel_fmts[] = {
90 AV_PIX_FMT_GRAY8,
91 AV_PIX_FMT_GRAY9,
92 AV_PIX_FMT_GRAY10,
93 AV_PIX_FMT_GRAY12,
94 AV_PIX_FMT_GRAY14,
95 AV_PIX_FMT_GRAY16,
96 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
97 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
98 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
99 AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
100 AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_YUVJ444P,
101 AV_PIX_FMT_YUVJ411P,
102 AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
103 AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
104 AV_PIX_FMT_YUV440P10,
105 AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV420P12,
106 AV_PIX_FMT_YUV440P12,
107 AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV420P14,
108 AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
109 AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRP9, AV_PIX_FMT_GBRP10,
110 AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRP14, AV_PIX_FMT_GBRP16,
111 AV_PIX_FMT_YUVA420P, AV_PIX_FMT_YUVA422P, AV_PIX_FMT_YUVA444P,
112 AV_PIX_FMT_YUVA444P9, AV_PIX_FMT_YUVA444P10, AV_PIX_FMT_YUVA444P12, AV_PIX_FMT_YUVA444P16,
113 AV_PIX_FMT_YUVA422P9, AV_PIX_FMT_YUVA422P10, AV_PIX_FMT_YUVA422P12, AV_PIX_FMT_YUVA422P16,
114 AV_PIX_FMT_YUVA420P9, AV_PIX_FMT_YUVA420P10, AV_PIX_FMT_YUVA420P16,
115 AV_PIX_FMT_GBRAP, AV_PIX_FMT_GBRAP10, AV_PIX_FMT_GBRAP12, AV_PIX_FMT_GBRAP16,
116 AV_PIX_FMT_NONE
117 };
118 AVFilterFormats *formats = ff_make_format_list(pixel_fmts);
119 if (!formats)
120 return AVERROR(ENOMEM);
121 return ff_set_common_formats(ctx, formats);
122 }
123
init(AVFilterContext * ctx)124 static av_cold int init(AVFilterContext *ctx)
125 {
126 ATADenoiseContext *s = ctx->priv;
127
128 if (!(s->size & 1)) {
129 av_log(ctx, AV_LOG_WARNING, "size %d is invalid. Must be an odd value, setting it to %d.\n", s->size, s->size|1);
130 s->size |= 1;
131 }
132 s->mid = s->size / 2 + 1;
133
134 return 0;
135 }
136
137 typedef struct ThreadData {
138 AVFrame *in, *out;
139 } ThreadData;
140
141 #define FILTER_ROW(type, name) \
142 static void filter_row##name(const uint8_t *ssrc, uint8_t *ddst, \
143 const uint8_t *ssrcf[SIZE], \
144 int w, int mid, int size, \
145 int thra, int thrb) \
146 { \
147 const type *src = (const type *)ssrc; \
148 const type **srcf = (const type **)ssrcf; \
149 type *dst = (type *)ddst; \
150 \
151 for (int x = 0; x < w; x++) { \
152 const int srcx = src[x]; \
153 unsigned lsumdiff = 0, rsumdiff = 0; \
154 unsigned ldiff, rdiff; \
155 unsigned sum = srcx; \
156 int l = 0, r = 0; \
157 int srcjx, srcix; \
158 \
159 for (int j = mid - 1, i = mid + 1; j >= 0 && i < size; j--, i++) { \
160 srcjx = srcf[j][x]; \
161 \
162 ldiff = FFABS(srcx - srcjx); \
163 lsumdiff += ldiff; \
164 if (ldiff > thra || \
165 lsumdiff > thrb) \
166 break; \
167 l++; \
168 sum += srcjx; \
169 \
170 srcix = srcf[i][x]; \
171 \
172 rdiff = FFABS(srcx - srcix); \
173 rsumdiff += rdiff; \
174 if (rdiff > thra || \
175 rsumdiff > thrb) \
176 break; \
177 r++; \
178 sum += srcix; \
179 } \
180 \
181 dst[x] = (sum + ((r + l + 1) >> 1)) / (r + l + 1); \
182 } \
183 }
184
185 FILTER_ROW(uint8_t, 8)
186 FILTER_ROW(uint16_t, 16)
187
188 #define FILTER_ROW_SERIAL(type, name) \
189 static void filter_row##name##_serial(const uint8_t *ssrc, uint8_t *ddst, \
190 const uint8_t *ssrcf[SIZE], \
191 int w, int mid, int size, \
192 int thra, int thrb) \
193 { \
194 const type *src = (const type *)ssrc; \
195 const type **srcf = (const type **)ssrcf; \
196 type *dst = (type *)ddst; \
197 \
198 for (int x = 0; x < w; x++) { \
199 const int srcx = src[x]; \
200 unsigned lsumdiff = 0, rsumdiff = 0; \
201 unsigned ldiff, rdiff; \
202 unsigned sum = srcx; \
203 int l = 0, r = 0; \
204 int srcjx, srcix; \
205 \
206 for (int j = mid - 1; j >= 0; j--) { \
207 srcjx = srcf[j][x]; \
208 \
209 ldiff = FFABS(srcx - srcjx); \
210 lsumdiff += ldiff; \
211 if (ldiff > thra || \
212 lsumdiff > thrb) \
213 break; \
214 l++; \
215 sum += srcjx; \
216 } \
217 \
218 for (int i = mid + 1; i < size; i++) { \
219 srcix = srcf[i][x]; \
220 \
221 rdiff = FFABS(srcx - srcix); \
222 rsumdiff += rdiff; \
223 if (rdiff > thra || \
224 rsumdiff > thrb) \
225 break; \
226 r++; \
227 sum += srcix; \
228 } \
229 \
230 dst[x] = (sum + ((r + l + 1) >> 1)) / (r + l + 1); \
231 } \
232 }
233
234 FILTER_ROW_SERIAL(uint8_t, 8)
235 FILTER_ROW_SERIAL(uint16_t, 16)
236
filter_slice(AVFilterContext * ctx,void * arg,int jobnr,int nb_jobs)237 static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
238 {
239 ATADenoiseContext *s = ctx->priv;
240 ThreadData *td = arg;
241 AVFrame *in = td->in;
242 AVFrame *out = td->out;
243 const int size = s->size;
244 const int mid = s->mid;
245 int p, y, i;
246
247 for (p = 0; p < s->nb_planes; p++) {
248 const int h = s->planeheight[p];
249 const int w = s->planewidth[p];
250 const int slice_start = (h * jobnr) / nb_jobs;
251 const int slice_end = (h * (jobnr+1)) / nb_jobs;
252 const uint8_t *src = in->data[p] + slice_start * in->linesize[p];
253 uint8_t *dst = out->data[p] + slice_start * out->linesize[p];
254 const int thra = s->thra[p];
255 const int thrb = s->thrb[p];
256 const uint8_t **data = (const uint8_t **)s->data[p];
257 const int *linesize = (const int *)s->linesize[p];
258 const uint8_t *srcf[SIZE];
259
260 if (!((1 << p) & s->planes)) {
261 av_image_copy_plane(dst, out->linesize[p], src, in->linesize[p],
262 w, slice_end - slice_start);
263 continue;
264 }
265
266 for (i = 0; i < size; i++)
267 srcf[i] = data[i] + slice_start * linesize[i];
268
269 for (y = slice_start; y < slice_end; y++) {
270 s->dsp.filter_row(src, dst, srcf, w, mid, size, thra, thrb);
271
272 dst += out->linesize[p];
273 src += in->linesize[p];
274
275 for (i = 0; i < size; i++)
276 srcf[i] += linesize[i];
277 }
278 }
279
280 return 0;
281 }
282
config_input(AVFilterLink * inlink)283 static int config_input(AVFilterLink *inlink)
284 {
285 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
286 AVFilterContext *ctx = inlink->dst;
287 ATADenoiseContext *s = ctx->priv;
288 int depth;
289
290 s->nb_planes = desc->nb_components;
291
292 s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
293 s->planeheight[0] = s->planeheight[3] = inlink->h;
294 s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
295 s->planewidth[0] = s->planewidth[3] = inlink->w;
296
297 depth = desc->comp[0].depth;
298 s->filter_slice = filter_slice;
299 if (depth == 8)
300 s->dsp.filter_row = s->algorithm == PARALLEL ? filter_row8 : filter_row8_serial;
301 else
302 s->dsp.filter_row = s->algorithm == PARALLEL ? filter_row16 : filter_row16_serial;
303
304 s->thra[0] = s->fthra[0] * (1 << depth) - 1;
305 s->thra[1] = s->fthra[1] * (1 << depth) - 1;
306 s->thra[2] = s->fthra[2] * (1 << depth) - 1;
307 s->thrb[0] = s->fthrb[0] * (1 << depth) - 1;
308 s->thrb[1] = s->fthrb[1] * (1 << depth) - 1;
309 s->thrb[2] = s->fthrb[2] * (1 << depth) - 1;
310
311 if (ARCH_X86)
312 ff_atadenoise_init_x86(&s->dsp, depth, s->algorithm);
313
314 return 0;
315 }
316
filter_frame(AVFilterLink * inlink,AVFrame * buf)317 static int filter_frame(AVFilterLink *inlink, AVFrame *buf)
318 {
319 AVFilterContext *ctx = inlink->dst;
320 AVFilterLink *outlink = ctx->outputs[0];
321 ATADenoiseContext *s = ctx->priv;
322 AVFrame *out, *in;
323 int i;
324
325 if (s->q.available != s->size) {
326 if (s->q.available < s->mid) {
327 for (i = 0; i < s->mid; i++) {
328 out = av_frame_clone(buf);
329 if (!out) {
330 av_frame_free(&buf);
331 return AVERROR(ENOMEM);
332 }
333 ff_bufqueue_add(ctx, &s->q, out);
334 }
335 }
336 if (s->q.available < s->size) {
337 ff_bufqueue_add(ctx, &s->q, buf);
338 s->available++;
339 }
340 return 0;
341 }
342
343 in = ff_bufqueue_peek(&s->q, s->mid);
344
345 if (!ctx->is_disabled) {
346 ThreadData td;
347
348 out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
349 if (!out) {
350 av_frame_free(&buf);
351 return AVERROR(ENOMEM);
352 }
353
354 for (i = 0; i < s->size; i++) {
355 AVFrame *frame = ff_bufqueue_peek(&s->q, i);
356
357 s->data[0][i] = frame->data[0];
358 s->data[1][i] = frame->data[1];
359 s->data[2][i] = frame->data[2];
360 s->linesize[0][i] = frame->linesize[0];
361 s->linesize[1][i] = frame->linesize[1];
362 s->linesize[2][i] = frame->linesize[2];
363 }
364
365 td.in = in; td.out = out;
366 ctx->internal->execute(ctx, s->filter_slice, &td, NULL,
367 FFMIN3(s->planeheight[1],
368 s->planeheight[2],
369 ff_filter_get_nb_threads(ctx)));
370 av_frame_copy_props(out, in);
371 } else {
372 out = av_frame_clone(in);
373 if (!out) {
374 av_frame_free(&buf);
375 return AVERROR(ENOMEM);
376 }
377 }
378
379 in = ff_bufqueue_get(&s->q);
380 av_frame_free(&in);
381 ff_bufqueue_add(ctx, &s->q, buf);
382
383 return ff_filter_frame(outlink, out);
384 }
385
request_frame(AVFilterLink * outlink)386 static int request_frame(AVFilterLink *outlink)
387 {
388 AVFilterContext *ctx = outlink->src;
389 ATADenoiseContext *s = ctx->priv;
390 int ret = 0;
391
392 ret = ff_request_frame(ctx->inputs[0]);
393
394 if (ret == AVERROR_EOF && !ctx->is_disabled && s->available) {
395 AVFrame *buf = av_frame_clone(ff_bufqueue_peek(&s->q, s->available));
396 if (!buf)
397 return AVERROR(ENOMEM);
398
399 ret = filter_frame(ctx->inputs[0], buf);
400 s->available--;
401 }
402
403 return ret;
404 }
405
uninit(AVFilterContext * ctx)406 static av_cold void uninit(AVFilterContext *ctx)
407 {
408 ATADenoiseContext *s = ctx->priv;
409
410 ff_bufqueue_discard_all(&s->q);
411 }
412
process_command(AVFilterContext * ctx,const char * cmd,const char * arg,char * res,int res_len,int flags)413 static int process_command(AVFilterContext *ctx,
414 const char *cmd,
415 const char *arg,
416 char *res,
417 int res_len,
418 int flags)
419 {
420 int ret = ff_filter_process_command(ctx, cmd, arg, res, res_len, flags);
421
422 if (ret < 0)
423 return ret;
424
425 return config_input(ctx->inputs[0]);
426 }
427
428 static const AVFilterPad inputs[] = {
429 {
430 .name = "default",
431 .type = AVMEDIA_TYPE_VIDEO,
432 .filter_frame = filter_frame,
433 .config_props = config_input,
434 },
435 { NULL }
436 };
437
438 static const AVFilterPad outputs[] = {
439 {
440 .name = "default",
441 .type = AVMEDIA_TYPE_VIDEO,
442 .request_frame = request_frame,
443 },
444 { NULL }
445 };
446
447 AVFilter ff_vf_atadenoise = {
448 .name = "atadenoise",
449 .description = NULL_IF_CONFIG_SMALL("Apply an Adaptive Temporal Averaging Denoiser."),
450 .priv_size = sizeof(ATADenoiseContext),
451 .priv_class = &atadenoise_class,
452 .init = init,
453 .uninit = uninit,
454 .query_formats = query_formats,
455 .inputs = inputs,
456 .outputs = outputs,
457 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL | AVFILTER_FLAG_SLICE_THREADS,
458 .process_command = process_command,
459 };
460