| Message ID | 20230319141514.28134-2-michael@niedermayer.cc |
|---|---|
| State | New |
| Headers | show |
| Series | [FFmpeg-devel,1/2] avcodec/snowenc: AV_CODEC_CAP_ENCODER_RECON_FRAME support | expand |
On 3/19/2023 11:15 AM, Michael Niedermayer wrote: > about 50% faster (based on command line fps) You should not allocate the 256 decoders if the encoder is a AV_CODEC_FLAG_RECON_FRAME one, as they will not be used. Doing so would save memory and speed up initialization. > > Signed-off-by: Michael Niedermayer <michael@niedermayer.cc> > --- > libavfilter/vf_uspp.c | 41 ++++++++++++++++++++++++++++------------- > 1 file changed, 28 insertions(+), 13 deletions(-) > > diff --git a/libavfilter/vf_uspp.c b/libavfilter/vf_uspp.c > index f60eb230a2..9318410089 100644 > --- a/libavfilter/vf_uspp.c > +++ b/libavfilter/vf_uspp.c > @@ -231,16 +231,25 @@ static int filter_1phase(AVFilterContext *ctx, void *arg, int i, int nb_jobs) > return ret; > } > > - ret = avcodec_send_packet(p->avctx_dec[i], pkt); > - av_packet_unref(pkt); > - if (ret < 0) { > - av_log(p->avctx_dec[i], AV_LOG_ERROR, "Error sending a packet for decoding\n"); > - return ret; > - } > - ret = avcodec_receive_frame(p->avctx_dec[i], p->frame_dec[i]); > - if (ret < 0) { > - av_log(p->avctx_dec[i], AV_LOG_ERROR, "Error receiving a frame from decoding\n"); > - return ret; > + if (p->avctx_enc[i]->flags & AV_CODEC_FLAG_RECON_FRAME) { > + av_packet_unref(pkt); > + ret = avcodec_receive_frame(p->avctx_enc[i], p->frame_dec[i]); > + if (ret < 0) { > + av_log(p->avctx_dec[i], AV_LOG_ERROR, "Error receiving a frame from encoding\n"); > + return ret; > + } > + } else { > + ret = avcodec_send_packet(p->avctx_dec[i], pkt); > + av_packet_unref(pkt); > + if (ret < 0) { > + av_log(p->avctx_dec[i], AV_LOG_ERROR, "Error sending a packet for decoding\n"); > + return ret; > + } > + ret = avcodec_receive_frame(p->avctx_dec[i], p->frame_dec[i]); > + if (ret < 0) { > + av_log(p->avctx_dec[i], AV_LOG_ERROR, "Error receiving a frame from decoding\n"); > + return ret; > + } > } > > offset = (BLOCK-x1) + (BLOCK-y1) * p->frame_dec[i]->linesize[0]; > @@ -397,6 +406,10 @@ static int config_input(AVFilterLink *inlink) > avctx_enc->max_b_frames = 0; > avctx_enc->pix_fmt = inlink->format; > avctx_enc->flags = AV_CODEC_FLAG_QSCALE | AV_CODEC_FLAG_LOW_DELAY; > + if (enc->capabilities & AV_CODEC_CAP_ENCODER_RECON_FRAME) { > + avctx_enc->flags |= AV_CODEC_FLAG_RECON_FRAME; > + av_dict_set(&opts, "no_bitstream", "1", 0); > + } > avctx_enc->strict_std_compliance = FF_COMPLIANCE_EXPERIMENTAL; > avctx_enc->global_quality = 123; > avctx_dec->thread_count = > @@ -408,9 +421,11 @@ static int config_input(AVFilterLink *inlink) > av_assert0(avctx_enc->codec); > > > - ret = avcodec_open2(avctx_dec, dec, NULL); > - if (ret < 0) > - return ret; > + if (!(enc->capabilities & AV_CODEC_CAP_ENCODER_RECON_FRAME)) { > + ret = avcodec_open2(avctx_dec, dec, NULL); > + if (ret < 0) > + return ret; > + } > > if (!(uspp->frame[i] = av_frame_alloc())) > return AVERROR(ENOMEM);
On Sun, Mar 19, 2023 at 11:24:54AM -0300, James Almer wrote: > On 3/19/2023 11:15 AM, Michael Niedermayer wrote: > > about 50% faster (based on command line fps) > > You should not allocate the 256 decoders if the encoder is a > AV_CODEC_FLAG_RECON_FRAME one, as they will not be used. > Doing so would save memory and speed up initialization. ok will post a new patchset thx [...]
diff --git a/libavfilter/vf_uspp.c b/libavfilter/vf_uspp.c index f60eb230a2..9318410089 100644 --- a/libavfilter/vf_uspp.c +++ b/libavfilter/vf_uspp.c @@ -231,16 +231,25 @@ static int filter_1phase(AVFilterContext *ctx, void *arg, int i, int nb_jobs) return ret; } - ret = avcodec_send_packet(p->avctx_dec[i], pkt); - av_packet_unref(pkt); - if (ret < 0) { - av_log(p->avctx_dec[i], AV_LOG_ERROR, "Error sending a packet for decoding\n"); - return ret; - } - ret = avcodec_receive_frame(p->avctx_dec[i], p->frame_dec[i]); - if (ret < 0) { - av_log(p->avctx_dec[i], AV_LOG_ERROR, "Error receiving a frame from decoding\n"); - return ret; + if (p->avctx_enc[i]->flags & AV_CODEC_FLAG_RECON_FRAME) { + av_packet_unref(pkt); + ret = avcodec_receive_frame(p->avctx_enc[i], p->frame_dec[i]); + if (ret < 0) { + av_log(p->avctx_dec[i], AV_LOG_ERROR, "Error receiving a frame from encoding\n"); + return ret; + } + } else { + ret = avcodec_send_packet(p->avctx_dec[i], pkt); + av_packet_unref(pkt); + if (ret < 0) { + av_log(p->avctx_dec[i], AV_LOG_ERROR, "Error sending a packet for decoding\n"); + return ret; + } + ret = avcodec_receive_frame(p->avctx_dec[i], p->frame_dec[i]); + if (ret < 0) { + av_log(p->avctx_dec[i], AV_LOG_ERROR, "Error receiving a frame from decoding\n"); + return ret; + } } offset = (BLOCK-x1) + (BLOCK-y1) * p->frame_dec[i]->linesize[0]; @@ -397,6 +406,10 @@ static int config_input(AVFilterLink *inlink) avctx_enc->max_b_frames = 0; avctx_enc->pix_fmt = inlink->format; avctx_enc->flags = AV_CODEC_FLAG_QSCALE | AV_CODEC_FLAG_LOW_DELAY; + if (enc->capabilities & AV_CODEC_CAP_ENCODER_RECON_FRAME) { + avctx_enc->flags |= AV_CODEC_FLAG_RECON_FRAME; + av_dict_set(&opts, "no_bitstream", "1", 0); + } avctx_enc->strict_std_compliance = FF_COMPLIANCE_EXPERIMENTAL; avctx_enc->global_quality = 123; avctx_dec->thread_count = @@ -408,9 +421,11 @@ static int config_input(AVFilterLink *inlink) av_assert0(avctx_enc->codec); - ret = avcodec_open2(avctx_dec, dec, NULL); - if (ret < 0) - return ret; + if (!(enc->capabilities & AV_CODEC_CAP_ENCODER_RECON_FRAME)) { + ret = avcodec_open2(avctx_dec, dec, NULL); + if (ret < 0) + return ret; + } if (!(uspp->frame[i] = av_frame_alloc())) return AVERROR(ENOMEM);
about 50% faster (based on command line fps) Signed-off-by: Michael Niedermayer <michael@niedermayer.cc> --- libavfilter/vf_uspp.c | 41 ++++++++++++++++++++++++++++------------- 1 file changed, 28 insertions(+), 13 deletions(-)