Andreas Klöckner

That's because they're generated at runtime: https://github.com/inducer/pyopencl/blob/291649892f4972e46c5823bfa359b37ad94d6cd5/pyopencl/cltypes.py#L72-L83 I suspect it'd be a simple exercise to rewrite this so that it generates source code that can then be read by an...

> It's impossible not to declare the variable name before runtime. I agree. That's why I proposed that we generate the source at install time.

I'm not sure how that could be done. The features relies on being able to evaluate the offsets of various struct members, and the only way I know to do...

7f20c759f47278ef14b69cc3d5546993cbb52529 is a WIP commit that doesn't even compile. It's intended to prevent further duplication of work with @matthiasdiener (cf. #587). Sorry about not getting this pushed sooner.

0cf016ff73bc3f7985b7ca227c4a8194e847ba0e has broken compatibility with `examples/svm.py`. I'll look into it.

@nchristensen > Segmentation fault Are you able to get a backtrace (e.g. with `gdb`?)

It appears that a good part of the slower execution of SVM comes from the ordering of the argument type checks in `Kernel.set_arg`: https://github.com/inducer/pyopencl/blob/8bce4f6cce70bf22baa8fd238784f3557e942b55/src/wrap_cl.hpp#L4504-L4518 We check for memory objects (images,...

With the changes in https://github.com/inducer/pyopencl/compare/7ef5ce5ec280038559295d45986c96302c193d6d..87420b1334806b7fa0fb0104e80a09366869b9f0, this performs about equivalently to buffers (with pocl) in my unscientific benchmarking.

Correct, it should be `int_ptr`. I have pushed a fix there.

The latest changes fix a double-free situation that led to crashes during lazy exec, specifically these here: https://github.com/inducer/pyopencl/compare/87420b1334806b7fa0fb0104e80a09366869b9f0..ae4e76fd09103f9d99f13065eb5599c98190b861