]> git.sesse.net Git - ffmpeg/blob - libavfilter/vf_atadenoise.c
avfilter/vf_atadenoise: add support for commands
[ffmpeg] / 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|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
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_NONE
112     };
113     AVFilterFormats *formats = ff_make_format_list(pixel_fmts);
114     if (!formats)
115         return AVERROR(ENOMEM);
116     return ff_set_common_formats(ctx, formats);
117 }
118
119 static av_cold int init(AVFilterContext *ctx)
120 {
121     ATADenoiseContext *s = ctx->priv;
122
123     if (!(s->size & 1)) {
124         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);
125         s->size |= 1;
126     }
127     s->mid = s->size / 2 + 1;
128
129     return 0;
130 }
131
132 typedef struct ThreadData {
133     AVFrame *in, *out;
134 } ThreadData;
135
136 #define FILTER_ROW(type, name)                                              \
137 static void filter_row##name(const uint8_t *ssrc, uint8_t *ddst,            \
138                              const uint8_t *ssrcf[SIZE],                    \
139                              int w, int mid, int size,                      \
140                              int thra, int thrb)                            \
141 {                                                                           \
142     const type *src = (const type *)ssrc;                                   \
143     const type **srcf = (const type **)ssrcf;                               \
144     type *dst = (type *)ddst;                                               \
145                                                                             \
146     for (int x = 0; x < w; x++) {                                           \
147        const int srcx = src[x];                                             \
148        unsigned lsumdiff = 0, rsumdiff = 0;                                 \
149        unsigned ldiff, rdiff;                                               \
150        unsigned sum = srcx;                                                 \
151        int l = 0, r = 0;                                                    \
152        int srcjx, srcix;                                                    \
153                                                                             \
154        for (int j = mid - 1, i = mid + 1; j >= 0 && i < size; j--, i++) {   \
155            srcjx = srcf[j][x];                                              \
156                                                                             \
157            ldiff = FFABS(srcx - srcjx);                                     \
158            lsumdiff += ldiff;                                               \
159            if (ldiff > thra ||                                              \
160                lsumdiff > thrb)                                             \
161                break;                                                       \
162            l++;                                                             \
163            sum += srcjx;                                                    \
164                                                                             \
165            srcix = srcf[i][x];                                              \
166                                                                             \
167            rdiff = FFABS(srcx - srcix);                                     \
168            rsumdiff += rdiff;                                               \
169            if (rdiff > thra ||                                              \
170                rsumdiff > thrb)                                             \
171                break;                                                       \
172            r++;                                                             \
173            sum += srcix;                                                    \
174        }                                                                    \
175                                                                             \
176        dst[x] = (sum + ((r + l + 1) >> 1)) / (r + l + 1);                   \
177    }                                                                        \
178 }
179
180 FILTER_ROW(uint8_t, 8)
181 FILTER_ROW(uint16_t, 16)
182
183 #define FILTER_ROW_SERIAL(type, name)                                       \
184 static void filter_row##name##_serial(const uint8_t *ssrc, uint8_t *ddst,   \
185                                       const uint8_t *ssrcf[SIZE],           \
186                                       int w, int mid, int size,             \
187                                       int thra, int thrb)                   \
188 {                                                                           \
189     const type *src = (const type *)ssrc;                                   \
190     const type **srcf = (const type **)ssrcf;                               \
191     type *dst = (type *)ddst;                                               \
192                                                                             \
193     for (int x = 0; x < w; x++) {                                           \
194        const int srcx = src[x];                                             \
195        unsigned lsumdiff = 0, rsumdiff = 0;                                 \
196        unsigned ldiff, rdiff;                                               \
197        unsigned sum = srcx;                                                 \
198        int l = 0, r = 0;                                                    \
199        int srcjx, srcix;                                                    \
200                                                                             \
201        for (int j = mid - 1; j >= 0; j--) {                                 \
202            srcjx = srcf[j][x];                                              \
203                                                                             \
204            ldiff = FFABS(srcx - srcjx);                                     \
205            lsumdiff += ldiff;                                               \
206            if (ldiff > thra ||                                              \
207                lsumdiff > thrb)                                             \
208                break;                                                       \
209            l++;                                                             \
210            sum += srcjx;                                                    \
211        }                                                                    \
212                                                                             \
213        for (int i = mid + 1; i < size; i++) {                               \
214            srcix = srcf[i][x];                                              \
215                                                                             \
216            rdiff = FFABS(srcx - srcix);                                     \
217            rsumdiff += rdiff;                                               \
218            if (rdiff > thra ||                                              \
219                rsumdiff > thrb)                                             \
220                break;                                                       \
221            r++;                                                             \
222            sum += srcix;                                                    \
223        }                                                                    \
224                                                                             \
225        dst[x] = (sum + ((r + l + 1) >> 1)) / (r + l + 1);                   \
226    }                                                                        \
227 }
228
229 FILTER_ROW_SERIAL(uint8_t, 8)
230 FILTER_ROW_SERIAL(uint16_t, 16)
231
232 static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
233 {
234     ATADenoiseContext *s = ctx->priv;
235     ThreadData *td = arg;
236     AVFrame *in = td->in;
237     AVFrame *out = td->out;
238     const int size = s->size;
239     const int mid = s->mid;
240     int p, y, i;
241
242     for (p = 0; p < s->nb_planes; p++) {
243         const int h = s->planeheight[p];
244         const int w = s->planewidth[p];
245         const int slice_start = (h * jobnr) / nb_jobs;
246         const int slice_end = (h * (jobnr+1)) / nb_jobs;
247         const uint8_t *src = in->data[p] + slice_start * in->linesize[p];
248         uint8_t *dst = out->data[p] + slice_start * out->linesize[p];
249         const int thra = s->thra[p];
250         const int thrb = s->thrb[p];
251         const uint8_t **data = (const uint8_t **)s->data[p];
252         const int *linesize = (const int *)s->linesize[p];
253         const uint8_t *srcf[SIZE];
254
255         if (!((1 << p) & s->planes)) {
256             av_image_copy_plane(dst, out->linesize[p], src, in->linesize[p],
257                                 w, slice_end - slice_start);
258             continue;
259         }
260
261         for (i = 0; i < size; i++)
262             srcf[i] = data[i] + slice_start * linesize[i];
263
264         for (y = slice_start; y < slice_end; y++) {
265             s->dsp.filter_row(src, dst, srcf, w, mid, size, thra, thrb);
266
267             dst += out->linesize[p];
268             src += in->linesize[p];
269
270             for (i = 0; i < size; i++)
271                 srcf[i] += linesize[i];
272         }
273     }
274
275     return 0;
276 }
277
278 static int config_input(AVFilterLink *inlink)
279 {
280     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
281     AVFilterContext *ctx = inlink->dst;
282     ATADenoiseContext *s = ctx->priv;
283     int depth;
284
285     s->nb_planes = desc->nb_components;
286
287     s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
288     s->planeheight[0] = s->planeheight[3] = inlink->h;
289     s->planewidth[1]  = s->planewidth[2]  = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
290     s->planewidth[0]  = s->planewidth[3]  = inlink->w;
291
292     depth = desc->comp[0].depth;
293     s->filter_slice = filter_slice;
294     if (depth == 8)
295         s->dsp.filter_row = s->algorithm == PARALLEL ? filter_row8 : filter_row8_serial;
296     else
297         s->dsp.filter_row = s->algorithm == PARALLEL ? filter_row16 : filter_row16_serial;
298
299     s->thra[0] = s->fthra[0] * (1 << depth) - 1;
300     s->thra[1] = s->fthra[1] * (1 << depth) - 1;
301     s->thra[2] = s->fthra[2] * (1 << depth) - 1;
302     s->thrb[0] = s->fthrb[0] * (1 << depth) - 1;
303     s->thrb[1] = s->fthrb[1] * (1 << depth) - 1;
304     s->thrb[2] = s->fthrb[2] * (1 << depth) - 1;
305
306     if (ARCH_X86)
307         ff_atadenoise_init_x86(&s->dsp, depth, s->algorithm);
308
309     return 0;
310 }
311
312 static int filter_frame(AVFilterLink *inlink, AVFrame *buf)
313 {
314     AVFilterContext *ctx = inlink->dst;
315     AVFilterLink *outlink = ctx->outputs[0];
316     ATADenoiseContext *s = ctx->priv;
317     AVFrame *out, *in;
318     int i;
319
320     if (s->q.available != s->size) {
321         if (s->q.available < s->mid) {
322             for (i = 0; i < s->mid; i++) {
323                 out = av_frame_clone(buf);
324                 if (!out) {
325                     av_frame_free(&buf);
326                     return AVERROR(ENOMEM);
327                 }
328                 ff_bufqueue_add(ctx, &s->q, out);
329             }
330         }
331         if (s->q.available < s->size) {
332             ff_bufqueue_add(ctx, &s->q, buf);
333             s->available++;
334         }
335         return 0;
336     }
337
338     in = ff_bufqueue_peek(&s->q, s->mid);
339
340     if (!ctx->is_disabled) {
341         ThreadData td;
342
343         out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
344         if (!out) {
345             av_frame_free(&buf);
346             return AVERROR(ENOMEM);
347         }
348
349         for (i = 0; i < s->size; i++) {
350             AVFrame *frame = ff_bufqueue_peek(&s->q, i);
351
352             s->data[0][i] = frame->data[0];
353             s->data[1][i] = frame->data[1];
354             s->data[2][i] = frame->data[2];
355             s->linesize[0][i] = frame->linesize[0];
356             s->linesize[1][i] = frame->linesize[1];
357             s->linesize[2][i] = frame->linesize[2];
358         }
359
360         td.in = in; td.out = out;
361         ctx->internal->execute(ctx, s->filter_slice, &td, NULL,
362                                FFMIN3(s->planeheight[1],
363                                       s->planeheight[2],
364                                       ff_filter_get_nb_threads(ctx)));
365         av_frame_copy_props(out, in);
366     } else {
367         out = av_frame_clone(in);
368         if (!out) {
369             av_frame_free(&buf);
370             return AVERROR(ENOMEM);
371         }
372     }
373
374     in = ff_bufqueue_get(&s->q);
375     av_frame_free(&in);
376     ff_bufqueue_add(ctx, &s->q, buf);
377
378     return ff_filter_frame(outlink, out);
379 }
380
381 static int request_frame(AVFilterLink *outlink)
382 {
383     AVFilterContext *ctx = outlink->src;
384     ATADenoiseContext *s = ctx->priv;
385     int ret = 0;
386
387     ret = ff_request_frame(ctx->inputs[0]);
388
389     if (ret == AVERROR_EOF && !ctx->is_disabled && s->available) {
390         AVFrame *buf = av_frame_clone(ff_bufqueue_peek(&s->q, s->available));
391         if (!buf)
392             return AVERROR(ENOMEM);
393
394         ret = filter_frame(ctx->inputs[0], buf);
395         s->available--;
396     }
397
398     return ret;
399 }
400
401 static av_cold void uninit(AVFilterContext *ctx)
402 {
403     ATADenoiseContext *s = ctx->priv;
404
405     ff_bufqueue_discard_all(&s->q);
406 }
407
408 static int process_command(AVFilterContext *ctx,
409                            const char *cmd,
410                            const char *arg,
411                            char *res,
412                            int res_len,
413                            int flags)
414 {
415     int ret = ff_filter_process_command(ctx, cmd, arg, res, res_len, flags);
416
417     if (ret < 0)
418         return ret;
419
420     return config_input(ctx->inputs[0]);
421 }
422
423 static const AVFilterPad inputs[] = {
424     {
425         .name         = "default",
426         .type         = AVMEDIA_TYPE_VIDEO,
427         .filter_frame = filter_frame,
428         .config_props = config_input,
429     },
430     { NULL }
431 };
432
433 static const AVFilterPad outputs[] = {
434     {
435         .name          = "default",
436         .type          = AVMEDIA_TYPE_VIDEO,
437         .request_frame = request_frame,
438     },
439     { NULL }
440 };
441
442 AVFilter ff_vf_atadenoise = {
443     .name          = "atadenoise",
444     .description   = NULL_IF_CONFIG_SMALL("Apply an Adaptive Temporal Averaging Denoiser."),
445     .priv_size     = sizeof(ATADenoiseContext),
446     .priv_class    = &atadenoise_class,
447     .init          = init,
448     .uninit        = uninit,
449     .query_formats = query_formats,
450     .inputs        = inputs,
451     .outputs       = outputs,
452     .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL | AVFILTER_FLAG_SLICE_THREADS,
453     .process_command = process_command,
454 };