avfilter/af_sofalizer: speed up fast convolution
[ffmpeg.git] / libavfilter / vf_remap.c
1 /*
2  * Copyright (c) 2016 Floris Sluiter
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  * Pixel remap filter
24  * This filter copies pixel by pixel a source frame to a target frame.
25  * It remaps the pixels to a new x,y destination based on two files ymap/xmap.
26  * Map files are passed as a parameter and are in PGM format (P2 or P5),
27  * where the values are y(rows)/x(cols) coordinates of the source_frame.
28  * The *target* frame dimension is based on mapfile dimensions: specified in the
29  * header of the mapfile and reflected in the number of datavalues.
30  * Dimensions of ymap and xmap must be equal. Datavalues must be positive or zero.
31  * Any datavalue in the ymap or xmap which value is higher
32  * then the *source* frame height or width is silently ignored, leaving a
33  * blank/chromakey pixel. This can safely be used as a feature to create overlays.
34  *
35  * Algorithm digest:
36  * Target_frame[y][x] = Source_frame[ ymap[y][x] ][ [xmap[y][x] ];
37  */
38
39 #include "libavutil/imgutils.h"
40 #include "libavutil/pixdesc.h"
41 #include "libavutil/opt.h"
42 #include "avfilter.h"
43 #include "formats.h"
44 #include "framesync.h"
45 #include "internal.h"
46 #include "video.h"
47
48 typedef struct RemapContext {
49     const AVClass *class;
50     int nb_planes;
51     int nb_components;
52     int step;
53     FFFrameSync fs;
54
55     int (*remap_slice)(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs);
56 } RemapContext;
57
58 #define OFFSET(x) offsetof(RemapContext, x)
59 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
60
61 static const AVOption remap_options[] = {
62     { NULL }
63 };
64
65 AVFILTER_DEFINE_CLASS(remap);
66
67 typedef struct ThreadData {
68     AVFrame *in, *xin, *yin, *out;
69     int nb_planes;
70     int nb_components;
71     int step;
72 } ThreadData;
73
74 static int query_formats(AVFilterContext *ctx)
75 {
76     static const enum AVPixelFormat pix_fmts[] = {
77         AV_PIX_FMT_YUVA444P,
78         AV_PIX_FMT_YUV444P,
79         AV_PIX_FMT_YUVJ444P,
80         AV_PIX_FMT_RGB24, AV_PIX_FMT_BGR24,
81         AV_PIX_FMT_ARGB, AV_PIX_FMT_ABGR, AV_PIX_FMT_RGBA, AV_PIX_FMT_BGRA,
82         AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRAP,
83         AV_PIX_FMT_YUV444P9, AV_PIX_FMT_YUV444P10, AV_PIX_FMT_YUV444P12,
84         AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV444P16,
85         AV_PIX_FMT_YUVA444P9, AV_PIX_FMT_YUVA444P10, AV_PIX_FMT_YUVA444P16,
86         AV_PIX_FMT_GBRP9, AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRP12,
87         AV_PIX_FMT_GBRP14, AV_PIX_FMT_GBRP16,
88         AV_PIX_FMT_GBRAP10, AV_PIX_FMT_GBRAP12, AV_PIX_FMT_GBRAP16,
89         AV_PIX_FMT_RGB48, AV_PIX_FMT_BGR48,
90         AV_PIX_FMT_RGBA64, AV_PIX_FMT_BGRA64,
91         AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAY9,
92         AV_PIX_FMT_GRAY10, AV_PIX_FMT_GRAY12,
93         AV_PIX_FMT_GRAY14, AV_PIX_FMT_GRAY16,
94         AV_PIX_FMT_NONE
95     };
96     static const enum AVPixelFormat map_fmts[] = {
97         AV_PIX_FMT_GRAY16,
98         AV_PIX_FMT_NONE
99     };
100     AVFilterFormats *pix_formats = NULL, *map_formats = NULL;
101     int ret;
102
103     if (!(pix_formats = ff_make_format_list(pix_fmts)) ||
104         !(map_formats = ff_make_format_list(map_fmts))) {
105         ret = AVERROR(ENOMEM);
106         goto fail;
107     }
108     if ((ret = ff_formats_ref(pix_formats, &ctx->inputs[0]->out_formats)) < 0 ||
109         (ret = ff_formats_ref(map_formats, &ctx->inputs[1]->out_formats)) < 0 ||
110         (ret = ff_formats_ref(map_formats, &ctx->inputs[2]->out_formats)) < 0 ||
111         (ret = ff_formats_ref(pix_formats, &ctx->outputs[0]->in_formats)) < 0)
112         goto fail;
113     return 0;
114 fail:
115     if (pix_formats)
116         av_freep(&pix_formats->formats);
117     av_freep(&pix_formats);
118     if (map_formats)
119         av_freep(&map_formats->formats);
120     av_freep(&map_formats);
121     return ret;
122 }
123
124 /**
125  * remap_planar algorithm expects planes of same size
126  * pixels are copied from source to target using :
127  * Target_frame[y][x] = Source_frame[ ymap[y][x] ][ [xmap[y][x] ];
128  */
129 #define DEFINE_REMAP_PLANAR_FUNC(name, bits, div)                                           \
130 static int remap_planar##bits##_##name##_slice(AVFilterContext *ctx, void *arg,             \
131                                                int jobnr, int nb_jobs)                      \
132 {                                                                                           \
133     const ThreadData *td = (ThreadData*)arg;                                                \
134     const AVFrame *in  = td->in;                                                            \
135     const AVFrame *xin = td->xin;                                                           \
136     const AVFrame *yin = td->yin;                                                           \
137     const AVFrame *out = td->out;                                                           \
138     const int slice_start = (out->height *  jobnr   ) / nb_jobs;                            \
139     const int slice_end   = (out->height * (jobnr+1)) / nb_jobs;                            \
140     const int xlinesize = xin->linesize[0] / 2;                                             \
141     const int ylinesize = yin->linesize[0] / 2;                                             \
142     int x , y, plane;                                                                       \
143                                                                                             \
144     for (plane = 0; plane < td->nb_planes ; plane++) {                                      \
145         const int dlinesize  = out->linesize[plane] / div;                                  \
146         const uint##bits##_t *src = (const uint##bits##_t *)in->data[plane];                \
147         uint##bits##_t *dst = (uint##bits##_t *)out->data[plane] + slice_start * dlinesize; \
148         const int slinesize  = in->linesize[plane] / div;                                   \
149         const uint16_t *xmap = (const uint16_t *)xin->data[0] + slice_start * xlinesize;    \
150         const uint16_t *ymap = (const uint16_t *)yin->data[0] + slice_start * ylinesize;    \
151                                                                                             \
152         for (y = slice_start; y < slice_end; y++) {                                         \
153             for (x = 0; x < out->width; x++) {                                              \
154                 if (ymap[x] < in->height && xmap[x] < in->width) {                          \
155                     dst[x] = src[ymap[x] * slinesize + xmap[x]];                            \
156                 } else {                                                                    \
157                     dst[x] = 0;                                                             \
158                 }                                                                           \
159             }                                                                               \
160             dst  += dlinesize;                                                              \
161             xmap += xlinesize;                                                              \
162             ymap += ylinesize;                                                              \
163         }                                                                                   \
164     }                                                                                       \
165                                                                                             \
166     return 0;                                                                               \
167 }
168
169 DEFINE_REMAP_PLANAR_FUNC(nearest, 8, 1)
170 DEFINE_REMAP_PLANAR_FUNC(nearest, 16, 2)
171
172 /**
173  * remap_packed algorithm expects pixels with both padded bits (step) and
174  * number of components correctly set.
175  * pixels are copied from source to target using :
176  * Target_frame[y][x] = Source_frame[ ymap[y][x] ][ [xmap[y][x] ];
177  */
178 #define DEFINE_REMAP_PACKED_FUNC(name, bits, div)                                           \
179 static int remap_packed##bits##_##name##_slice(AVFilterContext *ctx, void *arg,             \
180                                                int jobnr, int nb_jobs)                      \
181 {                                                                                           \
182     const ThreadData *td = (ThreadData*)arg;                                                \
183     const AVFrame *in  = td->in;                                                            \
184     const AVFrame *xin = td->xin;                                                           \
185     const AVFrame *yin = td->yin;                                                           \
186     const AVFrame *out = td->out;                                                           \
187     const int slice_start = (out->height *  jobnr   ) / nb_jobs;                            \
188     const int slice_end   = (out->height * (jobnr+1)) / nb_jobs;                            \
189     const int dlinesize  = out->linesize[0] / div;                                          \
190     const int slinesize  = in->linesize[0] / div;                                           \
191     const int xlinesize  = xin->linesize[0] / 2;                                            \
192     const int ylinesize  = yin->linesize[0] / 2;                                            \
193     const uint##bits##_t *src = (const uint##bits##_t *)in->data[0];                        \
194     uint##bits##_t *dst = (uint##bits##_t *)out->data[0] + slice_start * dlinesize;         \
195     const uint16_t *xmap = (const uint16_t *)xin->data[0] + slice_start * xlinesize;        \
196     const uint16_t *ymap = (const uint16_t *)yin->data[0] + slice_start * ylinesize;        \
197     const int step       = td->step / div;                                                  \
198     int c, x, y;                                                                            \
199                                                                                             \
200     for (y = slice_start; y < slice_end; y++) {                                             \
201         for (x = 0; x < out->width; x++) {                                                  \
202             for (c = 0; c < td->nb_components; c++) {                                       \
203                 if (ymap[x] < in->height && xmap[x] < in->width) {                          \
204                     dst[x * step + c] = src[ymap[x] * slinesize + xmap[x] * step + c];      \
205                 } else {                                                                    \
206                     dst[x * step + c] = 0;                                                  \
207                 }                                                                           \
208             }                                                                               \
209         }                                                                                   \
210         dst  += dlinesize;                                                                  \
211         xmap += xlinesize;                                                                  \
212         ymap += ylinesize;                                                                  \
213     }                                                                                       \
214                                                                                             \
215     return 0;                                                                               \
216 }
217
218 DEFINE_REMAP_PACKED_FUNC(nearest, 8, 1)
219 DEFINE_REMAP_PACKED_FUNC(nearest, 16, 2)
220
221 static int config_input(AVFilterLink *inlink)
222 {
223     AVFilterContext *ctx = inlink->dst;
224     RemapContext *s = ctx->priv;
225     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
226
227     s->nb_planes = av_pix_fmt_count_planes(inlink->format);
228     s->nb_components = desc->nb_components;
229
230     if (desc->comp[0].depth == 8) {
231         if (s->nb_planes > 1 || s->nb_components == 1) {
232             s->remap_slice = remap_planar8_nearest_slice;
233         } else {
234             s->remap_slice = remap_packed8_nearest_slice;
235         }
236     } else {
237         if (s->nb_planes > 1 || s->nb_components == 1) {
238             s->remap_slice = remap_planar16_nearest_slice;
239         } else {
240             s->remap_slice = remap_packed16_nearest_slice;
241         }
242     }
243
244     s->step = av_get_padded_bits_per_pixel(desc) >> 3;
245     return 0;
246 }
247
248 static int process_frame(FFFrameSync *fs)
249 {
250     AVFilterContext *ctx = fs->parent;
251     RemapContext *s = fs->opaque;
252     AVFilterLink *outlink = ctx->outputs[0];
253     AVFrame *out, *in, *xpic, *ypic;
254     int ret;
255
256     if ((ret = ff_framesync_get_frame(&s->fs, 0, &in,   0)) < 0 ||
257         (ret = ff_framesync_get_frame(&s->fs, 1, &xpic, 0)) < 0 ||
258         (ret = ff_framesync_get_frame(&s->fs, 2, &ypic, 0)) < 0)
259         return ret;
260
261     if (ctx->is_disabled) {
262         out = av_frame_clone(in);
263         if (!out)
264             return AVERROR(ENOMEM);
265     } else {
266         ThreadData td;
267
268         out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
269         if (!out)
270             return AVERROR(ENOMEM);
271         av_frame_copy_props(out, in);
272
273         td.in  = in;
274         td.xin = xpic;
275         td.yin = ypic;
276         td.out = out;
277         td.nb_planes = s->nb_planes;
278         td.nb_components = s->nb_components;
279         td.step = s->step;
280         ctx->internal->execute(ctx, s->remap_slice, &td, NULL, FFMIN(outlink->h, ff_filter_get_nb_threads(ctx)));
281     }
282     out->pts = av_rescale_q(in->pts, s->fs.time_base, outlink->time_base);
283
284     return ff_filter_frame(outlink, out);
285 }
286
287 static int config_output(AVFilterLink *outlink)
288 {
289     AVFilterContext *ctx = outlink->src;
290     RemapContext *s = ctx->priv;
291     AVFilterLink *srclink = ctx->inputs[0];
292     AVFilterLink *xlink = ctx->inputs[1];
293     AVFilterLink *ylink = ctx->inputs[2];
294     FFFrameSyncIn *in;
295     int ret;
296
297     if (xlink->w != ylink->w || xlink->h != ylink->h) {
298         av_log(ctx, AV_LOG_ERROR, "Second input link %s parameters "
299                "(size %dx%d) do not match the corresponding "
300                "third input link %s parameters (%dx%d)\n",
301                ctx->input_pads[1].name, xlink->w, xlink->h,
302                ctx->input_pads[2].name, ylink->w, ylink->h);
303         return AVERROR(EINVAL);
304     }
305
306     outlink->w = xlink->w;
307     outlink->h = xlink->h;
308     outlink->time_base = srclink->time_base;
309     outlink->sample_aspect_ratio = srclink->sample_aspect_ratio;
310     outlink->frame_rate = srclink->frame_rate;
311
312     ret = ff_framesync_init(&s->fs, ctx, 3);
313     if (ret < 0)
314         return ret;
315
316     in = s->fs.in;
317     in[0].time_base = srclink->time_base;
318     in[1].time_base = xlink->time_base;
319     in[2].time_base = ylink->time_base;
320     in[0].sync   = 2;
321     in[0].before = EXT_STOP;
322     in[0].after  = EXT_STOP;
323     in[1].sync   = 1;
324     in[1].before = EXT_NULL;
325     in[1].after  = EXT_INFINITY;
326     in[2].sync   = 1;
327     in[2].before = EXT_NULL;
328     in[2].after  = EXT_INFINITY;
329     s->fs.opaque   = s;
330     s->fs.on_event = process_frame;
331
332     return ff_framesync_configure(&s->fs);
333 }
334
335 static int activate(AVFilterContext *ctx)
336 {
337     RemapContext *s = ctx->priv;
338     return ff_framesync_activate(&s->fs);
339 }
340
341 static av_cold void uninit(AVFilterContext *ctx)
342 {
343     RemapContext *s = ctx->priv;
344
345     ff_framesync_uninit(&s->fs);
346 }
347
348 static const AVFilterPad remap_inputs[] = {
349     {
350         .name         = "source",
351         .type         = AVMEDIA_TYPE_VIDEO,
352         .config_props = config_input,
353     },
354     {
355         .name         = "xmap",
356         .type         = AVMEDIA_TYPE_VIDEO,
357     },
358     {
359         .name         = "ymap",
360         .type         = AVMEDIA_TYPE_VIDEO,
361     },
362     { NULL }
363 };
364
365 static const AVFilterPad remap_outputs[] = {
366     {
367         .name          = "default",
368         .type          = AVMEDIA_TYPE_VIDEO,
369         .config_props  = config_output,
370     },
371     { NULL }
372 };
373
374 AVFilter ff_vf_remap = {
375     .name          = "remap",
376     .description   = NULL_IF_CONFIG_SMALL("Remap pixels."),
377     .priv_size     = sizeof(RemapContext),
378     .uninit        = uninit,
379     .query_formats = query_formats,
380     .activate      = activate,
381     .inputs        = remap_inputs,
382     .outputs       = remap_outputs,
383     .priv_class    = &remap_class,
384     .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS,
385 };