History

lavfi/lut: Add slice threading support

Used the command for 1080p h264 clip as follow: a). ffmpeg -i input -vf lutyuv="u=128:v=128" -f null /dev/null b). ffmpeg -i input -vf lutrgb="g=0:b=0" -f null /dev/null after enabled the slice threading, the fps change from: a). 144fps to 258fps (lutyuv) b). 94fps to 153fps (lutrgb) in Intel(R) Core(TM) i5-8265U CPU @ 1.60GHz

Jun Zhao committed 7 years ago

Commit bbad0bc1 parent 360bee81 file · +197 −113Browse files

Unified · Split

libavfilter/vf_lut.c+197−113
@@ -337,13 +337,194 @@ static int config_props(AVFilterLink *inlink) · 336 unmodified lines
337337 return 0;
338338 }
339339
340+struct thread_data {
341+ AVFrame *in;
342+ AVFrame *out;
343+
344+ int w;
345+ int h;
346+};
347+
348+#define LOAD_PACKED_COMMON\
349+ LutContext *s = ctx->priv;\
350+ const struct thread_data *td = arg;\
351+\
352+ int i, j;\
353+ const int w = td->w;\
354+ const int h = td->h;\
355+ AVFrame *in = td->in;\
356+ AVFrame *out = td->out;\
357+ const uint16_t (*tab)[256*256] = (const uint16_t (*)[256*256])s->lut;\
358+ const int step = s->step;\
359+\
360+ const int slice_start = (h * jobnr ) / nb_jobs;\
361+ const int slice_end = (h * (jobnr+1)) / nb_jobs;\
362+
363+/* packed, 16-bit */
364+static int lut_packed_16bits(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
365+{
366+ LOAD_PACKED_COMMON
367+
368+ uint16_t *inrow, *outrow, *inrow0, *outrow0;
369+ const int in_linesize = in->linesize[0] / 2;
370+ const int out_linesize = out->linesize[0] / 2;
371+ inrow0 = (uint16_t *)in ->data[0];
372+ outrow0 = (uint16_t *)out->data[0];
373+
374+ for (i = slice_start; i < slice_end; i++) {
375+ inrow = inrow0 + i * in_linesize;
376+ outrow = outrow0 + i * out_linesize;
377+ for (j = 0; j < w; j++) {
378+
379+ switch (step) {
380+#if HAVE_BIGENDIAN
381+ case 4: outrow[3] = av_bswap16(tab[3][av_bswap16(inrow[3])]); // Fall-through
382+ case 3: outrow[2] = av_bswap16(tab[2][av_bswap16(inrow[2])]); // Fall-through
383+ case 2: outrow[1] = av_bswap16(tab[1][av_bswap16(inrow[1])]); // Fall-through
384+ default: outrow[0] = av_bswap16(tab[0][av_bswap16(inrow[0])]);
385+#else
386+ case 4: outrow[3] = tab[3][inrow[3]]; // Fall-through
387+ case 3: outrow[2] = tab[2][inrow[2]]; // Fall-through
388+ case 2: outrow[1] = tab[1][inrow[1]]; // Fall-through
389+ default: outrow[0] = tab[0][inrow[0]];
390+#endif
391+ }
392+ outrow += step;
393+ inrow += step;
394+ }
395+ }
396+
397+ return 0;
398+}
399+
400+/* packed, 8-bit */
401+static int lut_packed_8bits(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
402+{
403+ LOAD_PACKED_COMMON
404+
405+ uint8_t *inrow, *outrow, *inrow0, *outrow0;
406+ const int in_linesize = in->linesize[0];
407+ const int out_linesize = out->linesize[0];
408+ inrow0 = in ->data[0];
409+ outrow0 = out->data[0];
410+
411+ for (i = slice_start; i < slice_end; i++) {
412+ inrow = inrow0 + i * in_linesize;
413+ outrow = outrow0 + i * out_linesize;
414+ for (j = 0; j < w; j++) {
415+ switch (step) {
416+ case 4: outrow[3] = tab[3][inrow[3]]; // Fall-through
417+ case 3: outrow[2] = tab[2][inrow[2]]; // Fall-through
418+ case 2: outrow[1] = tab[1][inrow[1]]; // Fall-through
419+ default: outrow[0] = tab[0][inrow[0]];
420+ }
421+ outrow += step;
422+ inrow += step;
423+ }
424+ }
425+
426+ return 0;
427+}
428+
429+#define LOAD_PLANAR_COMMON\
430+ LutContext *s = ctx->priv;\
431+ const struct thread_data *td = arg;\
432+ int i, j, plane;\
433+ AVFrame *in = td->in;\
434+ AVFrame *out = td->out;\
435+
436+#define PLANAR_COMMON\
437+ int vsub = plane == 1 || plane == 2 ? s->vsub : 0;\
438+ int hsub = plane == 1 || plane == 2 ? s->hsub : 0;\
439+ int h = AV_CEIL_RSHIFT(td->h, vsub);\
440+ int w = AV_CEIL_RSHIFT(td->w, hsub);\
441+ const uint16_t *tab = s->lut[plane];\
442+\
443+ const int slice_start = (h * jobnr ) / nb_jobs;\
444+ const int slice_end = (h * (jobnr+1)) / nb_jobs;\
445+
446+/* planar >8 bit depth */
447+static int lut_planar_16bits(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
448+{
449+ LOAD_PLANAR_COMMON
450+
451+ uint16_t *inrow, *outrow;
452+
453+ for (plane = 0; plane < 4 && in->data[plane] && in->linesize[plane]; plane++) {
454+ PLANAR_COMMON
455+
456+ const int in_linesize = in->linesize[plane] / 2;
457+ const int out_linesize = out->linesize[plane] / 2;
458+
459+ inrow = (uint16_t *)(in ->data[plane] + slice_start * in_linesize);
460+ outrow = (uint16_t *)(out->data[plane] + slice_start * out_linesize);
461+
462+ for (i = slice_start; i < slice_end; i++) {
463+ for (j = 0; j < w; j++) {
464+#if HAVE_BIGENDIAN
465+ outrow[j] = av_bswap16(tab[av_bswap16(inrow[j])]);
466+#else
467+ outrow[j] = tab[inrow[j]];
468+#endif
469+ }
470+ inrow += in_linesize;
471+ outrow += out_linesize;
472+ }
473+ }
474+
475+ return 0;
476+}
477+
478+/* planar 8bit depth */
479+static int lut_planar_8bits(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
480+{
481+ LOAD_PLANAR_COMMON
482+
483+ uint8_t *inrow, *outrow;
484+
485+ for (plane = 0; plane < 4 && in->data[plane] && in->linesize[plane]; plane++) {
486+ PLANAR_COMMON
487+
488+ const int in_linesize = in->linesize[plane];
489+ const int out_linesize = out->linesize[plane];
490+
491+ inrow = in ->data[plane] + slice_start * in_linesize;
492+ outrow = out->data[plane] + slice_start * out_linesize;
493+
494+ for (i = slice_start; i < slice_end; i++) {
495+ for (j = 0; j < w; j++)
496+ outrow[j] = tab[inrow[j]];
497+ inrow += in_linesize;
498+ outrow += out_linesize;
499+ }
500+ }
501+
502+ return 0;
503+}
504+
505+#define PACKED_THREAD_DATA\
506+ struct thread_data td = {\
507+ .in = in,\
508+ .out = out,\
509+ .w = inlink->w,\
510+ .h = in->height,\
511+ };\
512+
513+#define PLANAR_THREAD_DATA\
514+ struct thread_data td = {\
515+ .in = in,\
516+ .out = out,\
517+ .w = inlink->w,\
518+ .h = inlink->h,\
519+ };\
520+
340521 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
341522 {
342523 AVFilterContext *ctx = inlink->dst;
343524 LutContext *s = ctx->priv;
344525 AVFilterLink *outlink = ctx->outputs[0];
345526 AVFrame *out;
346− int i, j, plane, direct = 0;
527+ int direct = 0;
347528
348529 if (av_frame_is_writable(in)) {
349530 direct = 1;
@@ -359,121 +540,24 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in) · 9 unmodified lines
359540
360541 if (s->is_rgb && s->is_16bit && !s->is_planar) {
361542 /* packed, 16-bit */
362− uint16_t *inrow, *outrow, *inrow0, *outrow0;
363− const int w = inlink->w;
364− const int h = in->height;
365− const uint16_t (*tab)[256*256] = (const uint16_t (*)[256*256])s->lut;
366− const int in_linesize = in->linesize[0] / 2;
367− const int out_linesize = out->linesize[0] / 2;
368− const int step = s->step;
369−
370− inrow0 = (uint16_t*) in ->data[0];
371− outrow0 = (uint16_t*) out->data[0];
372−
373− for (i = 0; i < h; i ++) {
374− inrow = inrow0;
375− outrow = outrow0;
376− for (j = 0; j < w; j++) {
377−
378− switch (step) {
379−#if HAVE_BIGENDIAN
380− case 4: outrow[3] = av_bswap16(tab[3][av_bswap16(inrow[3])]); // Fall-through
381− case 3: outrow[2] = av_bswap16(tab[2][av_bswap16(inrow[2])]); // Fall-through
382− case 2: outrow[1] = av_bswap16(tab[1][av_bswap16(inrow[1])]); // Fall-through
383− default: outrow[0] = av_bswap16(tab[0][av_bswap16(inrow[0])]);
384−#else
385− case 4: outrow[3] = tab[3][inrow[3]]; // Fall-through
386− case 3: outrow[2] = tab[2][inrow[2]]; // Fall-through
387− case 2: outrow[1] = tab[1][inrow[1]]; // Fall-through
388− default: outrow[0] = tab[0][inrow[0]];
389−#endif
390− }
391− outrow += step;
392− inrow += step;
393− }
394− inrow0 += in_linesize;
395− outrow0 += out_linesize;
396− }
543+ PACKED_THREAD_DATA
544+ ctx->internal->execute(ctx, lut_packed_16bits, &td, NULL,
545+ FFMIN(in->height, ff_filter_get_nb_threads(ctx)));
397546 } else if (s->is_rgb && !s->is_planar) {
398− /* packed */
399− uint8_t *inrow, *outrow, *inrow0, *outrow0;
400− const int w = inlink->w;
401− const int h = in->height;
402− const uint16_t (*tab)[256*256] = (const uint16_t (*)[256*256])s->lut;
403− const int in_linesize = in->linesize[0];
404− const int out_linesize = out->linesize[0];
405− const int step = s->step;
406−
407− inrow0 = in ->data[0];
408− outrow0 = out->data[0];
409−
410− for (i = 0; i < h; i ++) {
411− inrow = inrow0;
412− outrow = outrow0;
413− for (j = 0; j < w; j++) {
414− switch (step) {
415− case 4: outrow[3] = tab[3][inrow[3]]; // Fall-through
416− case 3: outrow[2] = tab[2][inrow[2]]; // Fall-through
417− case 2: outrow[1] = tab[1][inrow[1]]; // Fall-through
418− default: outrow[0] = tab[0][inrow[0]];
419− }
420− outrow += step;
421− inrow += step;
422− }
423− inrow0 += in_linesize;
424− outrow0 += out_linesize;
425− }
547+ /* packed 8 bits */
548+ PACKED_THREAD_DATA
549+ ctx->internal->execute(ctx, lut_packed_8bits, &td, NULL,
550+ FFMIN(in->height, ff_filter_get_nb_threads(ctx)));
426551 } else if (s->is_16bit) {
427− // planar >8 bit depth
428− uint16_t *inrow, *outrow;
429−
430− for (plane = 0; plane < 4 && in->data[plane] && in->linesize[plane]; plane++) {
431− int vsub = plane == 1 || plane == 2 ? s->vsub : 0;
432− int hsub = plane == 1 || plane == 2 ? s->hsub : 0;
433− int h = AV_CEIL_RSHIFT(inlink->h, vsub);
434− int w = AV_CEIL_RSHIFT(inlink->w, hsub);
435− const uint16_t *tab = s->lut[plane];
436− const int in_linesize = in->linesize[plane] / 2;
437− const int out_linesize = out->linesize[plane] / 2;
438−
439− inrow = (uint16_t *)in ->data[plane];
440− outrow = (uint16_t *)out->data[plane];
441−
442− for (i = 0; i < h; i++) {
443− for (j = 0; j < w; j++) {
444−#if HAVE_BIGENDIAN
445− outrow[j] = av_bswap16(tab[av_bswap16(inrow[j])]);
446−#else
447− outrow[j] = tab[inrow[j]];
448−#endif
449− }
450− inrow += in_linesize;
451− outrow += out_linesize;
452− }
453− }
552+ /* planar >8 bit depth */
553+ PLANAR_THREAD_DATA
554+ ctx->internal->execute(ctx, lut_planar_16bits, &td, NULL,
555+ FFMIN(in->height, ff_filter_get_nb_threads(ctx)));
454556 } else {
455557 /* planar 8bit depth */
456− uint8_t *inrow, *outrow;
457−
458− for (plane = 0; plane < 4 && in->data[plane] && in->linesize[plane]; plane++) {
459− int vsub = plane == 1 || plane == 2 ? s->vsub : 0;
460− int hsub = plane == 1 || plane == 2 ? s->hsub : 0;
461− int h = AV_CEIL_RSHIFT(inlink->h, vsub);
462− int w = AV_CEIL_RSHIFT(inlink->w, hsub);
463− const uint16_t *tab = s->lut[plane];
464− const int in_linesize = in->linesize[plane];
465− const int out_linesize = out->linesize[plane];
466−
467− inrow = in ->data[plane];
468− outrow = out->data[plane];
469−
470− for (i = 0; i < h; i++) {
471− for (j = 0; j < w; j++)
472− outrow[j] = tab[inrow[j]];
473− inrow += in_linesize;
474− outrow += out_linesize;
475− }
476− }
558+ PLANAR_THREAD_DATA
559+ ctx->internal->execute(ctx, lut_planar_8bits, &td, NULL,
560+ FFMIN(in->height, ff_filter_get_nb_threads(ctx)));
477561 }
478562
479563 if (!direct)
@@ -508,7 +592,7 @@ static const AVFilterPad outputs[] = { · 28 unmodified lines
508592 .query_formats = query_formats, \
509593 .inputs = inputs, \
510594 .outputs = outputs, \
511− .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC, \
595+ .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS, \
512596 }
513597
514598 #if CONFIG_LUT_FILTER

Keyboard shortcuts

General

Show keyboard shortcuts
?
Focus the current search
/

Issues and pull requests

Open filters
f

Issues, pull requests and Inbox

Select next item
j↓
Select previous item
k↑
Open selected item
Enter
Dismiss filters or clear selection
Esc

Repository code

Go to file
t