Fix <complex> on NVRTC
#109
Conversation
wmaxey
force-pushed
the
bugfix/nvrtc_complex
branch
2 times, most recently
from
f2ea8f4
to
64b2727
Compare
wmaxey
force-pushed
the
bugfix/nvrtc_complex
branch
from
64b2727
to
7256707
Compare
| @@ -16,166 +16,200 @@ | |||
| #include <cuda/std/complex> | |||
| #include <cuda/std/cassert> | |||
|
|
|||
| #ifdef __CUDA_ARCH__ | |||
There was a problem hiding this comment.
This needs to be brought to attention, __CUDA_ARCH__ globals are not going to be supportable with NVC++. So this change serves a dual purpose in the future.
Immediately, however, NVRTC cannot initialize the testcases constant using the dynamic floating point values from the likes of CUDART_INF_F.
libcxx/include/cmath
Outdated
| @@ -312,33 +312,19 @@ long double truncl(long double x); | |||
| #pragma GCC system_header | |||
| #endif | |||
|
|
|||
| #ifdef _LIBCUDACXX_COMPILER_NVRTC | |||
| #include <math_constants.h> | |||
There was a problem hiding this comment.
Just FYI: I am not sure what's the release strategy of libcu++, so maybe this can be ignored, but including this header has caused trouble in CuPy when a user runs CUDA + NVRTC + CuPy in a "runtime" environment (such as the CUDA Runtime Docker, or RAPIDS's favorite Conda env), because none of the standard CUDA headers are present there due to EULA.
There was a problem hiding this comment.
Thanks Leo!
That's insightful, I'm recalling now similar issues with cooperative_groups being unavailable in NVRTC.
I have a more preferable solution now that I think about it.
There was a problem hiding this comment.
Wesley, the change in 6696f20 is cool 😄
.upstream-tests/test/std/numerics/complex.number/cmplx.over/UDT_is_rejected.fail.cpp
Show resolved
Hide resolved
.upstream-tests/test/std/numerics/complex.number/cmplx.over/UDT_is_rejected.fail.cpp
Outdated
Show resolved
Hide resolved
Changes made in this patchset to fix #101:
<cmath>exposes a subset ofcuda::std::math functions under NVRTC.<cmath>creates defines forINFINITYandNANwhen NVRTC is detected