Dnia 2014-08-08, pią o godzinie 02:16 +0200, Michael Clerx pisze:
Just managed to install pycuda under fedora 20. Once I got going,
everything was quite simple: ./configure without any option switches
worked fine and the alternative option (python setup.py build and
install) worked too. Only a few small issues:
- The website says cuda 2.0 or higher. Fedora switched to gcc 4 a few
update cycles ago. Cuda 5 for some reason won't accept gcc 4, so to make
this work you had to hack around with gcc a little so that any cuda code
uses gcc 3. However... pycuda uses some gcc switches that only exist in
gcc 4, so using cuda 5 on fedora is not an option. Cuda 6 doesn't mind
gcc 4 (although you still have to add an --override switch when
installing) and then works fine. So for Fedora at least, I'd say (the
current version of) pycuda only works with cuda 6 and higher.
You would need to be more specific when talking about versions.
I know that you were probably talking about names of Fedora packages
but for CUDA exact gcc version matters. For example:
CUDA 5.5 works with gcc 4.9, CUDA 5.0 required gcc 4.6, the same for
CUDA 4.2. The 5.0 situation caused problems for Debian as CUDA
required other gcc version that the one used to build NVIDIA
- PyCuda builds and installs under python3 (using "python3 setup.py
build"), but then fails some tests. Mainly, the error message "ndarray
is not C-contiguous" crops up.
Hope this helps some people looking to install under fedora. Sorry if
this is all well known and I was the only one unable to find it on
Thanks for all the information. Can you put it on the wiki?
Tomasz Rybak GPG/PGP key ID: 2AD5 9860
Fingerprint A481 824E 7DD3 9C0E C40A 488E C654 FB33 2AD5 9860