[FFmpeg-devel] [PATCH v5] avcodec/v210dec: add the frame and slice threading support

[Limin Wang]

On Mon, Oct 14, 2019 at 10:12:45PM +0200, Michael Niedermayer wrote:
> On Sun, Oct 13, 2019 at 06:45:16AM +0800, Limin Wang wrote:
> > On Sat, Oct 12, 2019 at 11:46:58PM +0200, Michael Niedermayer wrote:
> > > On Sat, Oct 12, 2019 at 06:58:21PM +0800, lance.lmwang at gmail.com wrote:
> > > > From: Limin Wang <lance.lmwang at gmail.com>
> > > > 
> > > > The multithread is avoid one core cpu is full with other filter like scale etc.
> > > > About the performance, the gain is very small, below is my testing for
> > > > performance.
> > > > In order to avoid the disk bottleneck, I'll use stream_loop mode for 10 frame
> > > > only.
> > > > 
> > > > ./ffmpeg -y -i ~/Movies/4k_Rec709_ProResHQ.mov -c:v v210 -f rawvideo -frames 10
> > > > ~/Movies/1.v210
> > > > 
> > > > master:
> > > > ./ffmpeg -threads 1 -s 4096x3072 -stream_loop 100 -i ~/Movies/1.v210 -benchmark
> > > > -f null -
> > > > frame= 1010 fps= 42 q=-0.0 Lsize=N/A time=00:00:40.40 bitrate=N/A speed=1.69x
> > > > video:529kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB muxing
> > > > overhead: unknown
> > > > bench: utime=10.082s stime=13.784s rtime=23.889s
> > > > bench: maxrss=147836928kB
> > > > 
> > > > patch applied:
> > > > ./ffmpeg -threads 4 -thread_type frame+slice  -s 4096x3072 -stream_loop 100 -i
> > > > ~/Movies/1.v210 -benchmark -f null -
> > > > 
> > > > frame= 1010 fps= 55 q=-0.0 Lsize=N/A time=00:00:40.40 bitrate=N/A speed=2.22x
> > > > video:529kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB muxing
> > > > overhead: unknown
> > > > bench: utime=11.407s stime=17.258s rtime=18.279s
> > > > bench: maxrss=442884096kB
> > > > 
> > > > Signed-off-by: Limin Wang <lance.lmwang at gmail.com>
> > > > ---
> > > >  libavcodec/v210dec.c | 132 +++++++++++++++++++++++++++----------------
> > > >  libavcodec/v210dec.h |   1 +
> > > >  2 files changed, 85 insertions(+), 48 deletions(-)
> > > > 
> > > > diff --git a/libavcodec/v210dec.c b/libavcodec/v210dec.c
> > > > index 5a33d8c089..c3ef8051e6 100644
> > > > --- a/libavcodec/v210dec.c
> > > > +++ b/libavcodec/v210dec.c
> > > > @@ -28,6 +28,7 @@
> > > >  #include "libavutil/internal.h"
> > > >  #include "libavutil/mem.h"
> > > >  #include "libavutil/intreadwrite.h"
> > > > +#include "thread.h"
> > > >  
> > > >  #define READ_PIXELS(a, b, c)         \
> > > >      do {                             \
> > > > @@ -37,6 +38,12 @@
> > > >          *c++ = (val >> 20) & 0x3FF;  \
> > > >      } while (0)
> > > >  
> > > > +typedef struct ThreadData {
> > > > +    AVFrame *frame;
> > > > +    uint8_t *buf;
> > > > +    int stride;
> > > > +} ThreadData;
> > > > +
> > > >  static void v210_planar_unpack_c(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width)
> > > >  {
> > > >      uint32_t val;
> > > > @@ -64,21 +71,87 @@ static av_cold int decode_init(AVCodecContext *avctx)
> > > >      avctx->pix_fmt             = AV_PIX_FMT_YUV422P10;
> > > >      avctx->bits_per_raw_sample = 10;
> > > >  
> > > > +    s->thread_count  = av_clip(avctx->thread_count, 1, avctx->height/4);
> > > >      s->aligned_input = 0;
> > > >      ff_v210dec_init(s);
> > > >  
> > > >      return 0;
> > > >  }
> > > >  
> > > > +static int v210_decode_slice(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
> > > > +{
> > > > +    V210DecContext *s = avctx->priv_data;
> > > > +    int h, w;
> > > > +    ThreadData *td = arg;
> > > > +    AVFrame *frame = td->frame;
> > > > +    int stride = td->stride;
> > > > +    int slice_h = avctx->height / s->thread_count;
> > > > +    int slice_m = avctx->height % s->thread_count;
> > > > +    int slice_start = jobnr * slice_h;
> > > 
> > > this is still not correct
> > > the height of a slice is not the same for all slices if the frame height
> > > is not divisible by the number of slices
> > Yes, so the last slice is processed different by the following code. I have tested with
> > different thread number to verify the fate-v210 result. 
> > 
> > make fate-v210 SAMPLES=../fate-suite thread_type=frame+slice threads=[1-7]
> 
> What the code should do is split the frame evenly not
> have many small slices and a really large last one
> 
> a really large last slice would need much longer to be processed making
> the work distribution uneven

That's true, most of my testing are >=4K resolution, for other
resolution, the process is very faster and no need to enable 
thread. I'll do the split the frame evenly by your suggestion.


> 
> thx
> 
> [...]
> -- 
> Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
> 
> If you fake or manipulate statistics in a paper in physics you will never
> get a job again.
> If you fake or manipulate statistics in a paper in medicin you will get
> a job for life at the pharma industry.



> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".