Am 22.06.2017 um 15:58 schrieb Dan Torop:
Ulrich Pegelow <ulrich.pege...@tongareva.de> writes:
Am 21.06.2017 um 18:03 schrieb Dan Torop:
I'd say that is well withing the time frame – provided it's not too invasive.
That is great. Should be a tweak to the Bayer downscale code (making
start/endpoints of sample right), bringing that work to X-Trans, and then SSE
variants.
Please consider that changes in that place also will require changes to
the equivalent OpenCL code which might be anything but trivial.
Oh no! I was thinking of the dt_iop_clip_and_zoom_demosaic_{half,third}_*()
functions in imageop_math.c. Those don't have OpenCL implementations, though?
Certainly getting into revising OpenCL code would be a way more invasive effort.
Well there are:
demosaic_vng.cl:clip_and_zoom_demosaic_third_size_xtrans()
demosaic_ppg.cl:clip_and_zoom_demosaic_half_size()
I also remember now that I wanted to look into whether taking advantage of box
filters being separable would speed up the code. But that would involve
allocating a bit of memory so might also be something about which to be
conservative.
I have not looked into your code, so I don't know how large the base of
the box filters is. But probably you implement them as gliding window
filters. That's exactly something you can't implement efficiently in OpenCL.
We have that situation in some modules like highpass coming from times
much before OpenCL. In the end I implemented a gaussian filter in OpenCL
that tries to mimic the box filter as close as possible. Still that is
not an ideal way ...
Ulrich
Ulrich
[...]
___________________________________________________________________________
darktable developer mailing list
to unsubscribe send a mail to darktable-dev+unsubscr...@lists.darktable.org
