[FFmpeg-devel] [PATCH] lavfi/scale: add scale opencl filter

[Wei Gao]

2013/6/27 Michael Niedermayer <michaelni at gmx.at>

> On Wed, Jun 26, 2013 at 07:18:15PM +0800, Wei Gao wrote:
> > 2013/6/26 Michael Niedermayer <michaelni at gmx.at>
> >
> > > On Tue, Jun 25, 2013 at 09:55:45AM +0800, Wei Gao wrote:
> > > > 2013/6/20 Michael Niedermayer <michaelni at gmx.at>
> > > >
> > > > > On Thu, Jun 20, 2013 at 09:58:55AM +0800, Wei Gao wrote:
> > > > > > 2013/6/8 Michael Niedermayer <michaelni at gmx.at>
> > > > > >
> > > > > > > On Wed, May 29, 2013 at 04:55:47PM +0800, Wei Gao wrote:
> > > > >
> > >
> > > >
> > > > which flags should we test is better?
> > >
> > > bicubic and lanczos are popular scaler variants IIRC
> > >
> > Hi
> >
> > Thanks for your reply, the attachment is the benchmark and md5 code
>
> "
> ffmpeg-area_asm.log:frame=  912 fps=212 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-area_opencl.log:frame=  912 fps=177 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-bicubic_asm.log:frame=  912 fps=166 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-bicubic_opencl.log:frame=  912 fps=163 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-bicublin_asm.log:frame=  912 fps=185 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-bicublin_opencl.log:frame=  912 fps=166 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-bilinear_asm.log:frame=  912 fps=205 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-bilinear_opencl.log:frame=  912 fps=172 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-default_asm.log:frame=  912 fps=166 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-default_opencl.log:frame=  912 fps=163 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-fast_bilinear_asm.log:frame=  912 fps=216 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-fast_bilinear_opencl.log:frame=  912 fps=181 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-neighbor_asm.log:frame=  912 fps=224 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-neighbor_opencl.log:frame=  912 fps=179 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-spline_asm.log:frame=  912 fps=104 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> ffmpeg-spline_opencl.log:frame=  912 fps=137 q=0.0 Lsize=       0kB
> time=00:00:38.03 bitrate=   0.0kbits/s
> "
>
> so basically, its not faster
> and that is against single threaded swscale, which could be made to
> use multiple threads
>
Hi,

The performance of opencl scale is releated with GPU device, some low
performance is not only the kernel run time, but also the DMA(copy data
between CPU and GPU), and this will not cost time on HSA computer.

Thanks
Best regards

>
> [...]
>
> --
> Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
>
> He who knows, does not speak. He who speaks, does not know. -- Lao Tsu
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
>