Re: [Gegl-developer] Motion blur strange behavior

[Victor Oliveira <victormatheus gmail com>]

On Thu, Dec 27, 2012 at 1:38 AM, Pavel Roschin <roshin scriptumplus ru> wrote:
> Hello, I noticed two strange things in using gegl's motion blur:
>
> 1. Opencl and normal versions are different - try to use ImageMagick's
> compare

This is because the OpenCL path doesn't support the new abyss
policies, so it just uses a (0,0,0,0) pixel outside the geglbuffer
boundaries, while the "normal" version is probably using clamp.

> 2. I have to insert node "crop" because filter produces too big buffer
> (bigger than initial image)
>
> I attached test images and source code.
>
> I see too much bugs in filters, that must work with neighborhood
> pixels, it seems that this is due to complicated tile nature of gegl...

Yeah... it seems some of the filters are broken, I have to dig up what
is the problem, probably some of the recent changes in GEGL.

> And so, some tests:
>
> 1. OpenCL version of motion-blur:  1.0 s  (GeForce GT 430,   year 2012)
> 2. CPU version of motion-blur:     >10 s  (Pentium 4 3.0GHz, year 2003)
> 3. GIMP's old version motion blur: ~2 s (faster, much faster gegl's)
>
> 10x times faster. But my CPU significantly slower and older than GPU.
>
> In this case I hope GIMP will keep his fast (even if 8bit-per-color)
> filters, sometimes speed is the main thing in image processing. Or
> may be GEGL will have faster operations for 8-bit data?

But that's the point of using the GPU! We can have good and awesome
floating-point precision and speed :)

> --
> Pavel