Dnia 2010-09-27, pon o godzinie 04:42 -0700, jmcarval pisze: > Hi. > Installed PyCUDA 0.94.1 in several Linux boxes. > All have Ubuntu 10.4 with CUDA 3.1 (drv 256.40) and python 2.6.5 > > Boxes with 1.1 capability GPUs like 8600GT, 9400 GT or FX850 are ok and some > user's are already trying them. > Boxes with 1.3 (GTX280) and 2.0 (GTX480) have dificulties just running the > supplied tests. On these: > > test_cumath.py passes all tests but is 5 times slower in the GTX280 and > 40(!) times slower in the GTX480 > As far as I can see, this test never uses float64 > > test_driver.py passes all tests but is 4 times slower in the GTX280 and 9(!) > times slower in the GTX480 > As far as I can see, this test also doesn't use native float64. Testing > fp_textures seem to be done with some sort of emulation, but I may be wrong > here. > > test_gpuarray.py is 2 times slower in the GTX280 and fails the dot, sum, > minmax and subset_minmax tests on the GTX480.
The same problem was mentioned by Peter Schmidtke on 2010-07-21 on this mailing list. IIRC this was because of increased accuracy of floating point operations on Fermi, and relaxed (by Andreas) checking of results of operations: Julien Cornebise (2010-07-21): > To be honest, I don't have enough experience in pyCUDA to know how > much the precision issue is a problem -- for what I know, it could > even be that Fermi are too precise, as I haven't checked how the test > is done :) OTOH Andreas wrote on 2010-08-03: > I must be getting old. The loosening of the test bounds was in PyOpenCL, > not PyCUDA. Sorry for the confusion. In any case, test_gpuarray fully > passes for me on Fermi with current git. Peter, Julien, anyone: are you > still seeing problems/failures? I do not see significant changes in git between versions from 2010-08-02 (ccdc6b) and current one (79c10e). > > Can anybody please point me to some mail thread, FAQ or instalation guide > that helps me correct whatever I have done wrong? > I would propose to add something to FAQ/wiki, after resolving this issue. I have tested PyOpenCL and PyCUDA on ION (9400) with Debian 64 drivers 256.53. PyOpenCL. test_clmath and test_array - no problems test_wrapper - segmentation fault (?) PyCUDA: test_cumath, test_driver, test_gpuarray - OK -- Tomasz Rybak <[email protected]> GPG/PGP key ID: 2AD5 9860 Fingerprint A481 824E 7DD3 9C0E C40A 488E C654 FB33 2AD5 9860 http://member.acm.org/~tomaszrybak
signature.asc
Description: This is a digitally signed message part
_______________________________________________ PyCUDA mailing list [email protected] http://lists.tiker.net/listinfo/pycuda
