[gdal-dev] OpenCL Performance

[Even Rouault]

Le mardi 22 juillet 2014 22:10:56, Anderson, Gregory a écrit :
> Hello,
> 
> 
> I recently installed GDAL with OpenCL and my tests have shown that running
> gdalwarp with '-wo "USE_OPENCL=FALSE"' gives better performance then
> running it without. I'm using Ubuntu 14.04 64-bit on an Intel i7-2720QM
> with an AMD Radeon HD 6400M series graphics card. My OpenCL support is the
> AMD APP SDK and the AMD Catalyst driver, and I've confirmed that both the
> CPU and the GPU are seen as compute devices using clinfo. I've tried using
> both the CPU and the GPU as the OpenCL device and gotten the same results
> in both cases. I've also tried on both a hard disk and an SSD with no
> luck. My test data is a large (3.4GB) GeoTIFF file, and the command is
> 'time gdalwarp -t_srs EPSG:3395 -r cubic gnc-world.tif gnc-world3395.tif'.
> Does anyone have any ideas why the performance might be so bad?

Gregory,

I had tried a long time ago with a ATI Radeon HD 5450 and I had issues with 
the AMD OpenCL driver (the GUI became almost totally unresponsible when OpenCL 
was used). I somehow remember that it might be beneficial when using lanczos 
which was really slow on the CPU (but I've pushed improvements in recent 
versions to speed it up). Perhaps the benefit of faster computation on the GPU 
is lost by the time to move the data between the main RAM and the GPU RAM. Or 
perhaps that in your case most time is spent on I/O.
What is strange however is that you get the same timing when using CPU or GPU 
as OpenCL device.

Not related with OpenCL, but in GDAL 1.11, you can also try the NUM_THREADS 
warping option to speed warping :
http://lists.osgeo.org/pipermail/gdal-dev/2012-June/033084.html

Even

-- 
Geospatial professional services
http://even.rouault.free.fr/services.html