Graham Markall comments

Results 455 comments of


                                            Graham Markall

Fix issue 8370

Also, apologies for the delay in this round of review.

Fix issue 8370

Many thanks - just waiting on CI before I approve.

On target cuda, return type of function may not match type given in signature string

Starter patch: ```diff diff --git a/numba/cuda/decorators.py b/numba/cuda/decorators.py index a755ea0c8..9e514f5d4 100644 --- a/numba/cuda/decorators.py +++ b/numba/cuda/decorators.py @@ -121,7 +121,7 @@ def jit(func_or_sig=None, device=False, inline=False, link=[], debug=None, if device: from numba.core import typeinfer...

On target cuda, return type of function may not match type given in signature string

> RE: [#8400 (comment)](https://github.com/numba/numba/issues/8400#issuecomment-1229502785) @cstyl > > The behavior of `i32 + i32` promoting to `i64 + i64` is due to Numba's lack of bigint support and we opted for...

On target cuda, return type of function may not match type given in signature string

Now that #8788 is open to fix this particular issue, I've opened #8789 to track the `compile_ptx` use case separately (cc @cstyl).

Testhound/fp16 math functions

Hopefully #8572 will help with testing on the buildfarm - perhaps this could have another try once that is merged?

Setting cache=True with a recursive function results in Segmentation fault

This is an issue that may be resolved with an eventual move to OrcJIT, as we'd not need to have special handling (Numba's "linker") for unresolved symbols with OrcJIT.