hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-24 02:24:07 +00:00
Code Issues Actions Projects Releases Wiki Activity

Improve default for number of parallel frames

Browse source
This commit is contained in:
Ari Koivula 2016-09-27 23:14:57 +03:00
parent 19d423ab29
commit d7391a9593
1 changed files with 54 additions and 30 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

84
src/encoder.c
View file

@ -36,41 +36,65 @@ static int size_of_wpp_ends(int threads)
static int select_owf_auto(const kvz_config *const cfg) static int select_owf_auto(const kvz_config *const cfg)
{ {
if (cfg->wpp) { if (cfg->intra_period == 1) {
// If wpp is on, select owf such that less than 15% of the if (cfg->wpp) {
// frame is covered by the are threads can not work at the same time. // If wpp is on, select owf such that less than 15% of the
// frame is covered by the are threads can not work at the same time.
const int lcu_width = CEILDIV(cfg->width, LCU_WIDTH);
const int lcu_height = CEILDIV(cfg->height, LCU_WIDTH);
// Find the largest number of threads per frame that satifies the
// the condition: wpp start/stop inefficiency takes up less than 15%
// of frame area.
int threads_per_frame = 1;
const int wpp_treshold = lcu_width * lcu_height * 15 / 100;
while ((threads_per_frame + 1) * 2 < lcu_width &&
threads_per_frame + 1 < lcu_height &&
size_of_wpp_ends(threads_per_frame + 1) < wpp_treshold) {
++threads_per_frame;
}
const int threads = MAX(cfg->threads, 1);
const int frames = CEILDIV(threads, threads_per_frame);
// Convert from number of parallel frames to number of additional frames.
return CLIP(0, threads - 1, frames - 1);
} else {
// If wpp is not on, select owf such that there is enough
// tiles for twice the number of threads.
int tiles_per_frame = cfg->tiles_width_count * cfg->tiles_height_count;
int threads = (cfg->threads > 1 ? cfg->threads : 1);
int frames = CEILDIV(threads * 4, tiles_per_frame);
// Limit number of frames to 1.25x the number of threads for the case
// where there is only 1 tile per frame.
frames = CLIP(1, threads * 4 / 3, frames);
return frames - 1;
}
} else {
// Try and estimate a good number of parallel frames for inter.
const int lcu_width = CEILDIV(cfg->width, LCU_WIDTH); const int lcu_width = CEILDIV(cfg->width, LCU_WIDTH);
const int lcu_height = CEILDIV(cfg->height, LCU_WIDTH); const int lcu_height = CEILDIV(cfg->height, LCU_WIDTH);
int threads_per_frame = MIN(lcu_width / 2, lcu_height);
int threads = cfg->threads;
// Find the largest number of threads per frame that satifies the // If all threads fit into one frame, at least two parallel frames should
// the condition: wpp start/stop inefficiency takes up less than 15% // be used to reduce the effect of WPP spin-up and wind-down.
// of frame area. int frames = 1;
int threads_per_frame = 1;
const int wpp_treshold = lcu_width * lcu_height * 15 / 100; while (threads > 0 && threads_per_frame > 0) {
while ((threads_per_frame + 1) * 2 < lcu_width && frames += 1;
threads_per_frame + 1 < lcu_height && threads -= threads_per_frame;
size_of_wpp_ends(threads_per_frame + 1) < wpp_treshold) threads_per_frame -= 2;
{
++threads_per_frame;
} }
const int threads = MAX(cfg->threads, 1); if (cfg->gop_lowdelay && cfg->gop_lp_definition.t > 1) {
const int frames = CEILDIV(threads, threads_per_frame); // Temporal skipping makes every other frame very fast to encode so
// more parallel frames should be used.
// Convert from number of parallel frames to number of additional frames. frames *= 2;
return CLIP(0, threads - 1, frames - 1); }
} else { return CLIP(0, cfg->threads * 2 - 1, frames - 1);
// If wpp is not on, select owf such that there is enough
// tiles for twice the number of threads.
int tiles_per_frame= cfg->tiles_width_count * cfg->tiles_height_count;
int threads = (cfg->threads > 1 ? cfg->threads : 1);
int frames = CEILDIV(threads * 4, tiles_per_frame);
// Limit number of frames to 1.25x the number of threads for the case
// where there is only 1 tile per frame.
frames = CLIP(1, threads * 4 / 3, frames);
return frames - 1;
} }
} }