[gdal-dev] Resampling on the GPU using Apple Metal

[Even Rouault]

On jeudi 5 mars 2020 22:05:59 CET John Daniel wrote:
> Hello everyone,
> I am investigating adding support for Apple Metal to GPU resampling in GDAL.
> This is what I've learned so far:
> 
> 1) Currently, OpenCL is only used for resampling operations via gdalwarp.
> RasterIO resampling via gdal_translate does not use OpenCL. This makes
> gdal_translate, and RasterIO, much slower for resampling operations. 2) Not
> even taking OpenCL into account, there are two independent resampling
> paths. Resampling via RasterIO uses routines in gcore/overview.cpp.
> Resampling via warp uses a different set of resampling implementations in
> alg/gdalwarpkernel.cpp, for the same resampling algorithms.
> 
> So, if I wanted to implement a new set of resampling routines using Apple
> Metal, is there a reason why I should not implement it in the overview.cpp
> path and then have warp use that same path for resampling? Am I missing
> something?

> Is one set of resampling implementations better than the other?

Not really. The RasterIO / overview resampling can make extra assumptions in 
comparison to the warping resampling that make its CPU implementation normally 
faster. In warping mode, pixels of a same target line do not correspond to a 
single line in the source image, whereas for overviews, they do.

The fact that warping resampling has a OpenCL acceleration path is just 
because only it was targetted during the work where it was implemented. 
There's no fundamental reason why RasterIO / overview resampling couldn't have 
one.

Even

-- 
Spatialys - Geospatial professional services
http://www.spatialys.com