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