Andreas Klöckner
Andreas Klöckner
But `matrixMulDrv` must be calling `cuInit`, and that must be succeeding, given how far it gets. Could check (maybe with "Dependency Walker") that pycuda's `_driver` DLL and `matrixMulDrv` find the...
That's very strange. There aren't any binaries up on the package index ([see](https://pypi.org/project/pycuda/#files)?), so I wonder where your original binary was compiled, and what went wrong with it...
I don't see what else could have happened.
I'm generally supportive, and I'd be happy to consider a patch/PR. Writing one shouldn't be super hard.
Is there a canonical Python wrapper for dlpack? If so, you'd just need to import that and fill it with data from a `GPUArray` (or fill a `GPUArray` with data...
Check `dmesg`. Your kernel may have crashed due to invalid memory access and killed the context.
Numpy does not document the `np.float` and `np.complex` types as something that can be instantiated: https://docs.scipy.org/doc/numpy/user/basics.types.html. As such, it seems more like undefined behavior that you're able to do so....
I'd be happy to take a (tested) patch of either issue, but I don't have the time to work on this myself.
For PyCUDA, `async` was never documented, to it's an easy change: https://gitlab.tiker.net/inducer/pycuda/merge_requests/9 One issue that I ran into with this change is that numpy 1.13.3 doesn't even seem to build...
Oh, OK. Thanks. Trying: https://gitlab.tiker.net/inducer/pycuda/-/jobs/39546