Closed vincefn closed 3 months ago
I added a new commit as I found an issue when accessing a sub-array using a[0,Ø]
, which overflowed because one of the shape dimensions was an np.int32
and with numpy 2 int+np.int32
returns a np.int32
, which can easily overflow...
Thanks! Here's a corresponding PR on Gitlab (with CI): https://gitlab.tiker.net/inducer/pycuda/-/merge_requests/92
Ok, looking at the CI results
If I read correctly the K40 raw log , there is the failure of TestDriver.test_register_host_memory
(expected, see above, not sure if that can be corrected), and the TestGPUArray.test_curand_wrappers_8gb
failure with an ECC error which leads to the following device context errors.
The Titan V and X have failures on tests involving texture - cuda 12 issue ?
Finally had a chance to look this over, the changes look good to me. I've skip/xfailed (as appropriate) the failing tests. The ECC failure may well be hardware problem on our end.
(Just waiting to see if anything else pops up on the Gitlab CI.)
This looks good to go. Thanks for working on it. I'll roll a release with these changes.
Thanks for the quick release !
See https://github.com/inducer/pycuda/issues/450
All pycuda tests are passing (with python 12 and numpy 2.0, and cuda 11.8) , except
TestDriver.test_register_host_memory
which fails withValueError: Cannot set the NumPy array 'base' dependency more than once