Use pragma to disable execution checks in cuda::proclaim_return_type. #448
Conversation
|
Thanks a lot for the contribution 🎉 I am adding some cleanup because we want to centralize a macro in our |
|
There was a GitHub outage this morning. Some commits from @miscco are not showing in this PR. I am going to open this PR from draft state, since that fixed the problem for another PR where I saw this earlier today. |
bdice
force-pushed
the
cuda-proclaim-return-type-nv-exec-check-disable
branch
from
3480817 to
b1deea2
Compare
| @@ -394,6 +400,7 @@ __invoke(_Fp&& __f, _A0&& __a0) | |||
|
|
|||
| // bullet 7 | |||
|
|
|||
| _LIBCUDACXX_DISABLE_EXEC_CHECK | |||
There was a problem hiding this comment.
We should be aware, that this will disable execution checks for most of the library.
I am all in favor for that, but it will lead to bugs
There was a problem hiding this comment.
I was wondering about the potential for this problem. But I'm not sure what can be done about it... 🤔
There was a problem hiding this comment.
Unavoidable I'm afraid. This has been true in Thrust for time immemorial. It's nearly impossible to have generic C++ code without heavy use of this pragma.
…to fix problem with nested device lambdas.
Also cleanup the tests a bit
miscco
force-pushed
the
cuda-proclaim-return-type-nv-exec-check-disable
branch
from
b1deea2 to
6912090
Compare
|
@bdice Thanks a lot for the contribution! I have rebased the PR on latest master to resolve some minor merge conflicts. |
|
@miscco Thank you very much! |
Use pragma suggested by @jrhemstad to disable execution checks in
cuda::proclaim_return_type. This fixes a problem with nested device lambdas. Resolves #447.I'm not familiar with the layout of libcudacxx so I'm not sure what kind of test is appropriate to add here.
There is a minimal reproducer at https://godbolt.org/z/oaMhrcoPv, also noted in issue #447. I tested this locally with a more complex source file from libcudf (see here) and it compiled successfully after applying this patch.