1 /*
2 * Copyright (c) 2011 Roger Pau Monné <roger.pau@entel.upc.edu>
3 * Copyright (c) 2011 Stefano Sabatini
4 * Copyright (c) 2013 Paul B Mahol
5 *
6 * This file is part of FFmpeg.
7 *
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23 /**
24 * @file
25 * Caculate the PSNR between two input videos.
26 */
27
28 #include "libavutil/avstring.h"
29 #include "libavutil/opt.h"
30 #include "libavutil/pixdesc.h"
31 #include "avfilter.h"
32 #include "drawutils.h"
33 #include "formats.h"
34 #include "framesync.h"
35 #include "internal.h"
36 #include "psnr.h"
37 #include "video.h"
38
39 typedef struct PSNRContext {
40 const AVClass *class;
41 FFFrameSync fs;
42 double mse, min_mse, max_mse, mse_comp[4];
43 uint64_t nb_frames;
44 FILE *stats_file;
45 char *stats_file_str;
46 int stats_version;
47 int stats_header_written;
48 int stats_add_max;
49 int max[4], average_max;
50 int is_rgb;
51 uint8_t rgba_map[4];
52 char comps[4];
53 int nb_components;
54 int planewidth[4];
55 int planeheight[4];
56 double planeweight[4];
57 PSNRDSPContext dsp;
58 } PSNRContext;
59
60 #define OFFSET(x) offsetof(PSNRContext, x)
61 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
62
63 static const AVOption psnr_options[] = {
64 {"stats_file", "Set file where to store per-frame difference information", OFFSET(stats_file_str), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
65 {"f", "Set file where to store per-frame difference information", OFFSET(stats_file_str), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
66 {"stats_version", "Set the format version for the stats file.", OFFSET(stats_version), AV_OPT_TYPE_INT, {.i64=1}, 1, 2, FLAGS },
67 {"output_max", "Add raw stats (max values) to the output log.", OFFSET(stats_add_max), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS},
68 { NULL }
69 };
70
71 FRAMESYNC_DEFINE_CLASS(psnr, PSNRContext, fs);
72
pow_2(unsigned base)73 static inline unsigned pow_2(unsigned base)
74 {
75 return base*base;
76 }
77
get_psnr(double mse,uint64_t nb_frames,int max)78 static inline double get_psnr(double mse, uint64_t nb_frames, int max)
79 {
80 return 10.0 * log10(pow_2(max) / (mse / nb_frames));
81 }
82
sse_line_8bit(const uint8_t * main_line,const uint8_t * ref_line,int outw)83 static uint64_t sse_line_8bit(const uint8_t *main_line, const uint8_t *ref_line, int outw)
84 {
85 int j;
86 unsigned m2 = 0;
87
88 for (j = 0; j < outw; j++)
89 m2 += pow_2(main_line[j] - ref_line[j]);
90
91 return m2;
92 }
93
sse_line_16bit(const uint8_t * _main_line,const uint8_t * _ref_line,int outw)94 static uint64_t sse_line_16bit(const uint8_t *_main_line, const uint8_t *_ref_line, int outw)
95 {
96 int j;
97 uint64_t m2 = 0;
98 const uint16_t *main_line = (const uint16_t *) _main_line;
99 const uint16_t *ref_line = (const uint16_t *) _ref_line;
100
101 for (j = 0; j < outw; j++)
102 m2 += pow_2(main_line[j] - ref_line[j]);
103
104 return m2;
105 }
106
107 static inline
compute_images_mse(PSNRContext * s,const uint8_t * main_data[4],const int main_linesizes[4],const uint8_t * ref_data[4],const int ref_linesizes[4],int w,int h,double mse[4])108 void compute_images_mse(PSNRContext *s,
109 const uint8_t *main_data[4], const int main_linesizes[4],
110 const uint8_t *ref_data[4], const int ref_linesizes[4],
111 int w, int h, double mse[4])
112 {
113 int i, c;
114
115 for (c = 0; c < s->nb_components; c++) {
116 const int outw = s->planewidth[c];
117 const int outh = s->planeheight[c];
118 const uint8_t *main_line = main_data[c];
119 const uint8_t *ref_line = ref_data[c];
120 const int ref_linesize = ref_linesizes[c];
121 const int main_linesize = main_linesizes[c];
122 uint64_t m = 0;
123 for (i = 0; i < outh; i++) {
124 m += s->dsp.sse_line(main_line, ref_line, outw);
125 ref_line += ref_linesize;
126 main_line += main_linesize;
127 }
128 mse[c] = m / (double)(outw * outh);
129 }
130 }
131
set_meta(AVDictionary ** metadata,const char * key,char comp,float d)132 static void set_meta(AVDictionary **metadata, const char *key, char comp, float d)
133 {
134 char value[128];
135 snprintf(value, sizeof(value), "%0.2f", d);
136 if (comp) {
137 char key2[128];
138 snprintf(key2, sizeof(key2), "%s%c", key, comp);
139 av_dict_set(metadata, key2, value, 0);
140 } else {
141 av_dict_set(metadata, key, value, 0);
142 }
143 }
144
do_psnr(FFFrameSync * fs)145 static int do_psnr(FFFrameSync *fs)
146 {
147 AVFilterContext *ctx = fs->parent;
148 PSNRContext *s = ctx->priv;
149 AVFrame *master, *ref;
150 double comp_mse[4], mse = 0;
151 int ret, j, c;
152 AVDictionary **metadata;
153
154 ret = ff_framesync_dualinput_get(fs, &master, &ref);
155 if (ret < 0)
156 return ret;
157 if (!ref)
158 return ff_filter_frame(ctx->outputs[0], master);
159 metadata = &master->metadata;
160
161 compute_images_mse(s, (const uint8_t **)master->data, master->linesize,
162 (const uint8_t **)ref->data, ref->linesize,
163 master->width, master->height, comp_mse);
164
165 for (j = 0; j < s->nb_components; j++)
166 mse += comp_mse[j] * s->planeweight[j];
167
168 s->min_mse = FFMIN(s->min_mse, mse);
169 s->max_mse = FFMAX(s->max_mse, mse);
170
171 s->mse += mse;
172 for (j = 0; j < s->nb_components; j++)
173 s->mse_comp[j] += comp_mse[j];
174 s->nb_frames++;
175
176 for (j = 0; j < s->nb_components; j++) {
177 c = s->is_rgb ? s->rgba_map[j] : j;
178 set_meta(metadata, "lavfi.psnr.mse.", s->comps[j], comp_mse[c]);
179 set_meta(metadata, "lavfi.psnr.psnr.", s->comps[j], get_psnr(comp_mse[c], 1, s->max[c]));
180 }
181 set_meta(metadata, "lavfi.psnr.mse_avg", 0, mse);
182 set_meta(metadata, "lavfi.psnr.psnr_avg", 0, get_psnr(mse, 1, s->average_max));
183
184 if (s->stats_file) {
185 if (s->stats_version == 2 && !s->stats_header_written) {
186 fprintf(s->stats_file, "psnr_log_version:2 fields:n");
187 fprintf(s->stats_file, ",mse_avg");
188 for (j = 0; j < s->nb_components; j++) {
189 fprintf(s->stats_file, ",mse_%c", s->comps[j]);
190 }
191 fprintf(s->stats_file, ",psnr_avg");
192 for (j = 0; j < s->nb_components; j++) {
193 fprintf(s->stats_file, ",psnr_%c", s->comps[j]);
194 }
195 if (s->stats_add_max) {
196 fprintf(s->stats_file, ",max_avg");
197 for (j = 0; j < s->nb_components; j++) {
198 fprintf(s->stats_file, ",max_%c", s->comps[j]);
199 }
200 }
201 fprintf(s->stats_file, "\n");
202 s->stats_header_written = 1;
203 }
204 fprintf(s->stats_file, "n:%"PRId64" mse_avg:%0.2f ", s->nb_frames, mse);
205 for (j = 0; j < s->nb_components; j++) {
206 c = s->is_rgb ? s->rgba_map[j] : j;
207 fprintf(s->stats_file, "mse_%c:%0.2f ", s->comps[j], comp_mse[c]);
208 }
209 fprintf(s->stats_file, "psnr_avg:%0.2f ", get_psnr(mse, 1, s->average_max));
210 for (j = 0; j < s->nb_components; j++) {
211 c = s->is_rgb ? s->rgba_map[j] : j;
212 fprintf(s->stats_file, "psnr_%c:%0.2f ", s->comps[j],
213 get_psnr(comp_mse[c], 1, s->max[c]));
214 }
215 if (s->stats_version == 2 && s->stats_add_max) {
216 fprintf(s->stats_file, "max_avg:%d ", s->average_max);
217 for (j = 0; j < s->nb_components; j++) {
218 c = s->is_rgb ? s->rgba_map[j] : j;
219 fprintf(s->stats_file, "max_%c:%d ", s->comps[j], s->max[c]);
220 }
221 }
222 fprintf(s->stats_file, "\n");
223 }
224
225 return ff_filter_frame(ctx->outputs[0], master);
226 }
227
init(AVFilterContext * ctx)228 static av_cold int init(AVFilterContext *ctx)
229 {
230 PSNRContext *s = ctx->priv;
231
232 s->min_mse = +INFINITY;
233 s->max_mse = -INFINITY;
234
235 if (s->stats_file_str) {
236 if (s->stats_version < 2 && s->stats_add_max) {
237 av_log(ctx, AV_LOG_ERROR,
238 "stats_add_max was specified but stats_version < 2.\n" );
239 return AVERROR(EINVAL);
240 }
241 if (!strcmp(s->stats_file_str, "-")) {
242 s->stats_file = stdout;
243 } else {
244 s->stats_file = fopen(s->stats_file_str, "w");
245 if (!s->stats_file) {
246 int err = AVERROR(errno);
247 char buf[128];
248 av_strerror(err, buf, sizeof(buf));
249 av_log(ctx, AV_LOG_ERROR, "Could not open stats file %s: %s\n",
250 s->stats_file_str, buf);
251 return err;
252 }
253 }
254 }
255
256 s->fs.on_event = do_psnr;
257 return 0;
258 }
259
query_formats(AVFilterContext * ctx)260 static int query_formats(AVFilterContext *ctx)
261 {
262 static const enum AVPixelFormat pix_fmts[] = {
263 AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAY9, AV_PIX_FMT_GRAY10, AV_PIX_FMT_GRAY12, AV_PIX_FMT_GRAY14, AV_PIX_FMT_GRAY16,
264 #define PF_NOALPHA(suf) AV_PIX_FMT_YUV420##suf, AV_PIX_FMT_YUV422##suf, AV_PIX_FMT_YUV444##suf
265 #define PF_ALPHA(suf) AV_PIX_FMT_YUVA420##suf, AV_PIX_FMT_YUVA422##suf, AV_PIX_FMT_YUVA444##suf
266 #define PF(suf) PF_NOALPHA(suf), PF_ALPHA(suf)
267 PF(P), PF(P9), PF(P10), PF_NOALPHA(P12), PF_NOALPHA(P14), PF(P16),
268 AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_YUV410P,
269 AV_PIX_FMT_YUVJ411P, AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
270 AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_YUVJ444P,
271 AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRP9, AV_PIX_FMT_GBRP10,
272 AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRP14, AV_PIX_FMT_GBRP16,
273 AV_PIX_FMT_GBRAP, AV_PIX_FMT_GBRAP10, AV_PIX_FMT_GBRAP12, AV_PIX_FMT_GBRAP16,
274 AV_PIX_FMT_NONE
275 };
276
277 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
278 if (!fmts_list)
279 return AVERROR(ENOMEM);
280 return ff_set_common_formats(ctx, fmts_list);
281 }
282
config_input_ref(AVFilterLink * inlink)283 static int config_input_ref(AVFilterLink *inlink)
284 {
285 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
286 AVFilterContext *ctx = inlink->dst;
287 PSNRContext *s = ctx->priv;
288 double average_max;
289 unsigned sum;
290 int j;
291
292 s->nb_components = desc->nb_components;
293 if (ctx->inputs[0]->w != ctx->inputs[1]->w ||
294 ctx->inputs[0]->h != ctx->inputs[1]->h) {
295 av_log(ctx, AV_LOG_ERROR, "Width and height of input videos must be same.\n");
296 return AVERROR(EINVAL);
297 }
298 if (ctx->inputs[0]->format != ctx->inputs[1]->format) {
299 av_log(ctx, AV_LOG_ERROR, "Inputs must be of same pixel format.\n");
300 return AVERROR(EINVAL);
301 }
302
303 s->max[0] = (1 << desc->comp[0].depth) - 1;
304 s->max[1] = (1 << desc->comp[1].depth) - 1;
305 s->max[2] = (1 << desc->comp[2].depth) - 1;
306 s->max[3] = (1 << desc->comp[3].depth) - 1;
307
308 s->is_rgb = ff_fill_rgba_map(s->rgba_map, inlink->format) >= 0;
309 s->comps[0] = s->is_rgb ? 'r' : 'y' ;
310 s->comps[1] = s->is_rgb ? 'g' : 'u' ;
311 s->comps[2] = s->is_rgb ? 'b' : 'v' ;
312 s->comps[3] = 'a';
313
314 s->planeheight[1] = s->planeheight[2] = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
315 s->planeheight[0] = s->planeheight[3] = inlink->h;
316 s->planewidth[1] = s->planewidth[2] = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
317 s->planewidth[0] = s->planewidth[3] = inlink->w;
318 sum = 0;
319 for (j = 0; j < s->nb_components; j++)
320 sum += s->planeheight[j] * s->planewidth[j];
321 average_max = 0;
322 for (j = 0; j < s->nb_components; j++) {
323 s->planeweight[j] = (double) s->planeheight[j] * s->planewidth[j] / sum;
324 average_max += s->max[j] * s->planeweight[j];
325 }
326 s->average_max = lrint(average_max);
327
328 s->dsp.sse_line = desc->comp[0].depth > 8 ? sse_line_16bit : sse_line_8bit;
329 if (ARCH_X86)
330 ff_psnr_init_x86(&s->dsp, desc->comp[0].depth);
331
332 return 0;
333 }
334
config_output(AVFilterLink * outlink)335 static int config_output(AVFilterLink *outlink)
336 {
337 AVFilterContext *ctx = outlink->src;
338 PSNRContext *s = ctx->priv;
339 AVFilterLink *mainlink = ctx->inputs[0];
340 int ret;
341
342 ret = ff_framesync_init_dualinput(&s->fs, ctx);
343 if (ret < 0)
344 return ret;
345 outlink->w = mainlink->w;
346 outlink->h = mainlink->h;
347 outlink->time_base = mainlink->time_base;
348 outlink->sample_aspect_ratio = mainlink->sample_aspect_ratio;
349 outlink->frame_rate = mainlink->frame_rate;
350 if ((ret = ff_framesync_configure(&s->fs)) < 0)
351 return ret;
352
353 return 0;
354 }
355
activate(AVFilterContext * ctx)356 static int activate(AVFilterContext *ctx)
357 {
358 PSNRContext *s = ctx->priv;
359 return ff_framesync_activate(&s->fs);
360 }
361
uninit(AVFilterContext * ctx)362 static av_cold void uninit(AVFilterContext *ctx)
363 {
364 PSNRContext *s = ctx->priv;
365
366 if (s->nb_frames > 0) {
367 int j;
368 char buf[256];
369
370 buf[0] = 0;
371 for (j = 0; j < s->nb_components; j++) {
372 int c = s->is_rgb ? s->rgba_map[j] : j;
373 av_strlcatf(buf, sizeof(buf), " %c:%f", s->comps[j],
374 get_psnr(s->mse_comp[c], s->nb_frames, s->max[c]));
375 }
376 av_log(ctx, AV_LOG_INFO, "PSNR%s average:%f min:%f max:%f\n",
377 buf,
378 get_psnr(s->mse, s->nb_frames, s->average_max),
379 get_psnr(s->max_mse, 1, s->average_max),
380 get_psnr(s->min_mse, 1, s->average_max));
381 }
382
383 ff_framesync_uninit(&s->fs);
384
385 if (s->stats_file && s->stats_file != stdout)
386 fclose(s->stats_file);
387 }
388
389 static const AVFilterPad psnr_inputs[] = {
390 {
391 .name = "main",
392 .type = AVMEDIA_TYPE_VIDEO,
393 },{
394 .name = "reference",
395 .type = AVMEDIA_TYPE_VIDEO,
396 .config_props = config_input_ref,
397 },
398 { NULL }
399 };
400
401 static const AVFilterPad psnr_outputs[] = {
402 {
403 .name = "default",
404 .type = AVMEDIA_TYPE_VIDEO,
405 .config_props = config_output,
406 },
407 { NULL }
408 };
409
410 AVFilter ff_vf_psnr = {
411 .name = "psnr",
412 .description = NULL_IF_CONFIG_SMALL("Calculate the PSNR between two video streams."),
413 .preinit = psnr_framesync_preinit,
414 .init = init,
415 .uninit = uninit,
416 .query_formats = query_formats,
417 .activate = activate,
418 .priv_size = sizeof(PSNRContext),
419 .priv_class = &psnr_class,
420 .inputs = psnr_inputs,
421 .outputs = psnr_outputs,
422 };
423