[Beignet] darktable support

[Zhigang Gong]

On Sat, May 03, 2014 at 08:31:09PM +0300, Vasily Khoruzhick wrote:
> Hi Zhigang,
> 
> I've checked OpenCL performance with darktable and unfortunatelly it's
> slower than CPU.
> 
> Exporting same photo takes 0.522 secs on CPU and 2.133 secs on GPU.
> 
> Here's "darktable -d opencl -d perf" output for CPU:
> 
> [export] creating pixelpipe took 0.172 secs (0.170 CPU)
> [pixelpipe_process] [export] using device -1
> [dev_pixelpipe] took 0.000 secs (-0.000 CPU) initing base buffer [export]
> [dev_pixelpipe] took 0.007 secs (0.010 CPU) processing `white balance' [export]
> [dev_pixelpipe] took 0.009 secs (0.017 CPU) processing `highlight
> reconstruction' [export]
> [dev_pixelpipe] took 0.264 secs (0.637 CPU) processing `demosaic' [export]
> [dev_pixelpipe] took 0.035 secs (0.103 CPU) processing `base curve' [export]
> [dev_pixelpipe] took 0.045 secs (0.160 CPU) processing `input color
> profile' [export]
> [dev_pixelpipe] took 0.070 secs (0.213 CPU) processing `sharpen' [export]
> [dev_pixelpipe] took 0.068 secs (0.227 CPU) processing `output color
> profile' [export]
> [dev_pixelpipe] took 0.021 secs (0.070 CPU) processing `gamma' [export]
> [dev_process_export] pixel pipeline processing took 0.522 secs (1.437 CPU)
> [export_job] exported to `/home/anarsoul/Photos/test/DSC_2563_07.jpg'
> 
> And here's for GPU:
> 
> [export] creating pixelpipe took 0.227 secs (0.240 CPU)
> [pixelpipe_process] [export] using device 0
> [dev_pixelpipe] took 0.000 secs (0.000 CPU) initing base buffer [export]
> [dev_pixelpipe] took 0.077 secs (0.020 CPU) processing `white balance' [export]
> [dev_pixelpipe] took 0.017 secs (-0.000 CPU) processing `highlight
> reconstruction' [export]
> [dev_pixelpipe] took 0.175 secs (-0.000 CPU) processing `demosaic' [export]
> [dev_pixelpipe] took 0.079 secs (-0.000 CPU) processing `base curve' [export]
> [dev_pixelpipe] took 0.236 secs (-0.000 CPU) processing `input color
> profile' [export]
> [dev_pixelpipe] took 0.510 secs (-0.000 CPU) processing `sharpen' [export]
> [dev_pixelpipe] took 0.230 secs (-0.000 CPU) processing `output color
> profile' [export]
> [dev_pixelpipe] took 0.808 secs (0.810 CPU) processing `gamma' [export]
> [opencl_profiling] spent  0.0001 seconds in [Write Image (from host to device)]
> [opencl_profiling] spent  0.8581 seconds in whitebalance_1ui
> [opencl_profiling] spent  0.1243 seconds in highlights_1f
> [opencl_profiling] spent  0.5701 seconds in ppg_demosaic_green
> [opencl_profiling] spent  1.1415 seconds in ppg_demosaic_redblue
> [opencl_profiling] spent  0.1549 seconds in border_interpolate
> [opencl_profiling] spent  0.3583 seconds in clip_and_zoom
> [opencl_profiling] spent  1.2348 seconds in basecurve
> [opencl_profiling] spent  3.7282 seconds in colorin
> [opencl_profiling] spent  3.9993 seconds in sharpen_hblur
> [opencl_profiling] spent  3.8118 seconds in sharpen_vblur
> [opencl_profiling] spent  0.3095 seconds in sharpen_mix
> [opencl_profiling] spent  3.6290 seconds in colorout
> [opencl_profiling] spent  0.0000 seconds in [Read Image (from device to host)]
> [opencl_profiling] spent 19.9198 seconds totally in command queue
> (with 0 events missing)
> [dev_process_export] pixel pipeline processing took 2.133 secs (0.830 CPU)
> [export_job] exported to `/home/anarsoul/Photos/test/DSC_2563_13.jpg'
> 
> Is it possible to improve performance with OpenCL enabled somehow?
Yes, it is possible. But currently, the highest priority case is the luxmark,
once we finish this round performance tuning for luxmark. We will turn to see
some other cases and the darktable seems a good candidate.

> 
> Thanks!
> 
> Regards,
> Vasily
> 
> On Sat, May 3, 2014 at 8:04 PM, Vasily Khoruzhick <anarsoul at gmail.com> wrote:
> > Hi Zhigang,
> >
> > It works for me when beignet is built without ICD loader support, so I
> > guess something's wrong with it.
We found some bugs in some early versions of ocl-icd. You may try to upgrade your
ocl-icd to version 2.1.3-2 or after. And you can easily check whether it is crashed
in the ocl-icd internal or crashed in beignet part.

> >
> > Thanks!
> >
> > Regards
> > Vasily
> >
> > On Sat, May 3, 2014 at 7:26 PM, Vasily Khoruzhick <anarsoul at gmail.com> wrote:
> >> Hi Zhigang,
> >>
> >> On Sat, May 3, 2014 at 6:00 PM, Zhigang Gong <zhigang.gong at gmail.com> wrote:
> >>>
> >>>   From the log, I found the opencl is loaded successfully and it works well.
> >>>   The only big difference from your environment is that you are using a GT1
> >>> device,
> >>>   but I am using a GT1. I will try to find a GT1 device to check whether it
> >>> has
> >>>   the same issue you met.
> >>>   And if you could provide a back trace of the crash,  it may helpful to
> >>> identify what's
> >>>   going wrong there.
> >>
> >> I have GT2 as well, so I can try it.
> >>
> >> Did you try to import some photo and switch to "darkroom" tab?
> >>
> >> I'll try to produce backtrace.
> >>
> >> Regards
> >> Vasily
> >>
> >>>
> >>>   Thanks for reporting this issue.
> >>>>
> >>>>
> >>>> Thanks!
> >>>>
> >>>> Regards
> >>>> Vasily
> >>>> _______________________________________________
> >>>> Beignet mailing list
> >>>> Beignet at lists.freedesktop.org
> >>>> http://lists.freedesktop.org/mailman/listinfo/beignet
> >>>
> >>>
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet