| Message ID | 20200306000939.1823-1-jamrial@gmail.com |
|---|---|
| State | Accepted |
| Headers | show |
| Series | [FFmpeg-devel] avcodec/vp9: use a buffer pool to allocate VP9Frame extradata | expand |
| Context | Check | Description |
|---|---|---|
| andriy/ffmpeg-patchwork | success | Make fate finished |
Quoting James Almer (2020-03-06 01:09:39) > Signed-off-by: James Almer <jamrial@gmail.com> > --- > libavcodec/vp9.c | 13 ++++++++++++- > libavcodec/vp9dec.h | 4 ++++ > 2 files changed, 16 insertions(+), 1 deletion(-) > Looks reasonable. Any measurable performance difference?
On 3/26/2020 11:30 AM, Anton Khirnov wrote: > Quoting James Almer (2020-03-06 01:09:39) >> Signed-off-by: James Almer <jamrial@gmail.com> >> --- >> libavcodec/vp9.c | 13 ++++++++++++- >> libavcodec/vp9dec.h | 4 ++++ >> 2 files changed, 16 insertions(+), 1 deletion(-) >> > > Looks reasonable. Any measurable performance difference? Very noticeable when i use the START_TIMER/STOP_TIMER macros within the function, of course, but despite only allocating about three buffers for the entire decoding process instead of one per frame, i didn't notice any reduction in overall decoding time. Someone better at benchmarking may want to give it a try, though. Will apply, thanks.
diff --git a/libavcodec/vp9.c b/libavcodec/vp9.c index 7aaae9b792..7ee375d4d0 100644 --- a/libavcodec/vp9.c +++ b/libavcodec/vp9.c @@ -112,10 +112,20 @@ static int vp9_frame_alloc(AVCodecContext *avctx, VP9Frame *f) return ret; sz = 64 * s->sb_cols * s->sb_rows; - f->extradata = av_buffer_allocz(sz * (1 + sizeof(VP9mvrefPair))); + if (sz != s->frame_extradata_pool_size) { + av_buffer_pool_uninit(&s->frame_extradata_pool); + s->frame_extradata_pool = av_buffer_pool_init(sz * (1 + sizeof(VP9mvrefPair)), NULL); + if (!s->frame_extradata_pool) { + s->frame_extradata_pool_size = 0; + goto fail; + } + s->frame_extradata_pool_size = sz; + } + f->extradata = av_buffer_pool_get(s->frame_extradata_pool); if (!f->extradata) { goto fail; } + memset(f->extradata->data, 0, f->extradata->size); f->segmentation_map = f->extradata->data; f->mv = (VP9mvrefPair *) (f->extradata->data + sz); @@ -1210,6 +1220,7 @@ static av_cold int vp9_decode_free(AVCodecContext *avctx) vp9_frame_unref(avctx, &s->s.frames[i]); av_frame_free(&s->s.frames[i].tf.f); } + av_buffer_pool_uninit(&s->frame_extradata_pool); for (i = 0; i < 8; i++) { if (s->s.refs[i].f->buf[0]) ff_thread_release_buffer(avctx, &s->s.refs[i]); diff --git a/libavcodec/vp9dec.h b/libavcodec/vp9dec.h index 66573edc79..de02b146f0 100644 --- a/libavcodec/vp9dec.h +++ b/libavcodec/vp9dec.h @@ -152,6 +152,10 @@ typedef struct VP9Context { int block_alloc_using_2pass; uint16_t mvscale[3][2]; uint8_t mvstep[3][2]; + + // frame specific buffer pools + AVBufferPool *frame_extradata_pool; + int frame_extradata_pool_size; } VP9Context; struct VP9TileData {
Signed-off-by: James Almer <jamrial@gmail.com> --- libavcodec/vp9.c | 13 ++++++++++++- libavcodec/vp9dec.h | 4 ++++ 2 files changed, 16 insertions(+), 1 deletion(-)