Andreas Klöckner
Andreas Klöckner
I suspect the most fruitful approach would be to modify the kernel caching layer to support this. Maybe allow setting a mode where all used kernels can be "recorded". (This...
* OpenCL is reference counted, so even at the OpenCL API level, the only way to free resources is to lose all references. * Since each Python object acting as...
> sorry I have no time for doc PR. No worries. > I am trying to create pytorch-like ML framework based on pyopencl. Cool. Looking forward to seeing what you...
Some tips: - Add a bunch of CL build options ``` "-cl-denorms-are-zero", "-cl-fast-relaxed-math", "-cl-finite-math-only", "-cl-mad-enable", "-cl-no-signed-zeros", ``` - `get_group_id()` and `get_local_id()` return 64-bit integer, which can lead to 64-bit index...
TBH, I'm a bit torn about this situation. (Btw hope you're OK with my retitling of this issue.) The additional conditional you are asking for is the only reasonable thing...
> For a kernel with a kernel-wide outer loop (say an FEM action kernel), all the inames would get condensed into a single basic set which would lead to pretty...
The rule I'm proposing would prohibit separate loops in the tree from sharing a domain, in a way a sharpening of the "no branching" rule that exists. What it would...
I would welcome a contributed CI job! Maybe the existing Linux wheel build can help as an example.
Thanks for the offer! As of #421, it's just Github Actions all around, which AFAIK supports Windows just fine.