vp9: split superframes in the filtering stage before actual decoding

Significantly increases the efficiency of frame threading, since
individual frames in a superframe can now be decoded in parallel.
This commit is contained in:
Anton Khirnov 2016-11-13 11:24:16 +01:00
parent 03a80925ef
commit fa1749dd34
2 changed files with 14 additions and 62 deletions

2
configure vendored
View File

@ -2105,7 +2105,7 @@ vp6a_decoder_select="vp6_decoder"
vp6f_decoder_select="vp6_decoder" vp6f_decoder_select="vp6_decoder"
vp7_decoder_select="h264pred videodsp vp8dsp" vp7_decoder_select="h264pred videodsp vp8dsp"
vp8_decoder_select="h264pred videodsp vp8dsp" vp8_decoder_select="h264pred videodsp vp8dsp"
vp9_decoder_select="videodsp" vp9_decoder_select="videodsp vp9_superframe_split_bsf"
webp_decoder_select="vp8_decoder" webp_decoder_select="vp8_decoder"
wmapro_decoder_select="mdct sinewin wma_freqs" wmapro_decoder_select="mdct sinewin wma_freqs"
wmav1_decoder_select="mdct sinewin wma_freqs" wmav1_decoder_select="mdct sinewin wma_freqs"

View File

@ -1188,14 +1188,18 @@ static int update_refs(AVCodecContext *avctx)
return 0; return 0;
} }
static int vp9_decode_frame(AVCodecContext *avctx, AVFrame *frame, static int vp9_decode_frame(AVCodecContext *avctx, void *output,
int *got_frame, const uint8_t *data, int size, int *got_frame, AVPacket *pkt)
int can_finish_setup)
{ {
VP9Context *s = avctx->priv_data; VP9Context *s = avctx->priv_data;
AVFrame *frame = output;
const uint8_t *data = pkt->data;
int size = pkt->size;
AVFrame *f; AVFrame *f;
int ret, tile_row, tile_col, i, ref = -1, row, col; int ret, tile_row, tile_col, i, ref = -1, row, col;
s->setup_finished = 0;
ret = decode_frame_header(avctx, data, size, &ref); ret = decode_frame_header(avctx, data, size, &ref);
if (ret < 0) { if (ret < 0) {
return ret; return ret;
@ -1210,7 +1214,7 @@ static int vp9_decode_frame(AVCodecContext *avctx, AVFrame *frame,
if (ret < 0) if (ret < 0)
return ret; return ret;
*got_frame = 1; *got_frame = 1;
return 0; return pkt->size;
} }
data += ret; data += ret;
size -= ret; size -= ret;
@ -1261,7 +1265,7 @@ static int vp9_decode_frame(AVCodecContext *avctx, AVFrame *frame,
s->prob_ctx[s->framectxid].p = s->prob.p; s->prob_ctx[s->framectxid].p = s->prob.p;
} }
if ((s->parallelmode || !s->refreshctx) && if ((s->parallelmode || !s->refreshctx) &&
can_finish_setup && avctx->active_thread_type & FF_THREAD_FRAME) { avctx->active_thread_type & FF_THREAD_FRAME) {
ff_thread_finish_setup(avctx); ff_thread_finish_setup(avctx);
s->setup_finished = 1; s->setup_finished = 1;
} }
@ -1402,7 +1406,7 @@ static int vp9_decode_frame(AVCodecContext *avctx, AVFrame *frame,
if (s->pass < 2 && s->refreshctx && !s->parallelmode) { if (s->pass < 2 && s->refreshctx && !s->parallelmode) {
ff_vp9_adapt_probs(s); ff_vp9_adapt_probs(s);
if (can_finish_setup && avctx->active_thread_type & FF_THREAD_FRAME) { if (avctx->active_thread_type & FF_THREAD_FRAME) {
ff_thread_finish_setup(avctx); ff_thread_finish_setup(avctx);
s->setup_finished = 1; s->setup_finished = 1;
} }
@ -1428,60 +1432,7 @@ fail:
*got_frame = 1; *got_frame = 1;
} }
return 0; return pkt->size;
}
static int vp9_decode_packet(AVCodecContext *avctx, void *frame,
int *got_frame, AVPacket *avpkt)
{
VP9Context *s = avctx->priv_data;
const uint8_t *data = avpkt->data;
int size = avpkt->size;
int marker, ret;
s->setup_finished = 0;
/* Read superframe index - this is a collection of individual frames
* that together lead to one visible frame */
marker = data[size - 1];
if ((marker & 0xe0) == 0xc0) {
int nbytes = 1 + ((marker >> 3) & 0x3);
int n_frames = 1 + (marker & 0x7);
int idx_sz = 2 + n_frames * nbytes;
if (size >= idx_sz && data[size - idx_sz] == marker) {
const uint8_t *idx = data + size + 1 - idx_sz;
while (n_frames--) {
unsigned sz = AV_RL32(idx);
if (nbytes < 4)
sz &= (1 << (8 * nbytes)) - 1;
idx += nbytes;
if (sz > size) {
av_log(avctx, AV_LOG_ERROR,
"Superframe packet size too big: %u > %d\n",
sz, size);
return AVERROR_INVALIDDATA;
}
ret = vp9_decode_frame(avctx, frame, got_frame, data, sz,
!n_frames);
if (ret < 0)
return ret;
data += sz;
size -= sz;
}
return avpkt->size;
}
}
/* If we get here, there was no valid superframe index, i.e. this is just
* one whole single frame. Decode it as such from the complete input buf. */
if ((ret = vp9_decode_frame(avctx, frame, got_frame, data, size, 1)) < 0)
return ret;
return size;
} }
static av_cold int vp9_decode_free(AVCodecContext *avctx) static av_cold int vp9_decode_free(AVCodecContext *avctx)
@ -1590,10 +1541,11 @@ AVCodec ff_vp9_decoder = {
.id = AV_CODEC_ID_VP9, .id = AV_CODEC_ID_VP9,
.priv_data_size = sizeof(VP9Context), .priv_data_size = sizeof(VP9Context),
.init = vp9_decode_init, .init = vp9_decode_init,
.decode = vp9_decode_packet, .decode = vp9_decode_frame,
.flush = vp9_decode_flush, .flush = vp9_decode_flush,
.close = vp9_decode_free, .close = vp9_decode_free,
.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS, .capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS,
.init_thread_copy = vp9_decode_init, .init_thread_copy = vp9_decode_init,
.update_thread_context = vp9_decode_update_thread_context, .update_thread_context = vp9_decode_update_thread_context,
.bsfs = "vp9_superframe_split",
}; };