[PyOpenCL] Barrier caused limited groupsize

[Andreas Kloeckner]

Hi Kyle,

On Mon, 5 Jul 2010 13:00:18 -0700, Kyle Hanson <hanooter at gmail.com> wrote:
> To fix a CAL version mismatch, I installed the latest Catalyst (10.6) and
> reinstalled pyopencl. Now it gives me a choice to run on processor or GPU,
> however the GPU version doesnt work, ending with an error that I have
> included.
> 
> 
> ][][][][]][[][][][]][][][][][][][[][]
> C:\Users\phun\Downloads\pyopencl-0.91.5.tar\pyopencl-0.91.5\pyopencl-0.91.5\exam
> ples>python matrix-multiply.py
> Choose device(s) from these choices:
> [0] <pyopencl.Device 'Dual Core AMD Opteron(tm) Processor 280
>      ' at 0x351b450>
> [1] <pyopencl.Device 'ATI RV770
>      ' at 0x351b640>
> Choice, comma-separated [0]:1
> Traceback (most recent call last):
>   File "matrix-multiply.py", line 164, in <module>
>     ).build(options="-cl-mad-enable -cl-fast-relaxed-math")
>   File
> "C:\Python26\lib\site-packages\pyopencl-0.91.5-py2.6-win32.egg\pyopencl\_
> _init__.py", line 130, in program_build
>     "Build on %s:\n\n%s" % (dev, log) for dev, log in build_logs))
> pyopencl.RuntimeError: clBuildProgram failed: build program failure
> 
> Build on <pyopencl.Device 'ATI RV770
>           ' at 0x351b640>:
> 
> Warning: not supported option: -cl-mad-enable
> 
> Warning: not supported option: -cl-fast-relaxed-math
> 
> Warning:W000:Barrier caused limited groupsize
> Warning:W000:Barrier caused limited groupsize
> Warning:W000:Barrier caused limited groupsize
> Error: Requested compile size is bigger than the required workgroup size of

It seems that 4xxx-generation AMD hardware only supports barriers for
work groups of 64 items or fewer. It might be sufficient to change
block_size to 8 in the matrix multiply example (and also change both of
the hard-coded 16s--fixed in git), making for 8x8 tile sizes, which
should work. Can you try this and report back?

(Unrelatedly, I've fixed the git version of the example to only use 
-cl-mad-enable and -cl-fast-relaxed-math if the code is running on
an Nvidia device, which should fix the first two warnings.)

HTH,
Andreas

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
URL: <http://lists.tiker.net/pipermail/pyopencl/attachments/20100709/46381dac/attachment.pgp>