implement new CUDA compiler checks #2983
Conversation
|
Please properly rebase this PR onto |
| @@ -45,12 +45,12 @@ _CCCL_DIAG_SUPPRESS_CLANG("-Wunused-function") | |||
| # include <cuda_bf16.h> | |||
| _CCCL_DIAG_POP | |||
|
|
|||
| # if _CCCL_CUDACC_AT_LEAST(11, 8) | |||
| # if _CCCL_CUDA_COMPILER_AT_LEAST(11, 8) | |||
There was a problem hiding this comment.
We cannot change that macro, This relates to the __CUDACC__ macro, aka the version of the cuda headers used
This is different from the version of e.g clang-cuda and also nvc++ can consume cuda headers from different versions.
So we should leave that macro alone.
So for the currently supported compilers we have
- nvcc: nvcc version + cudacc version more or less equivalent
- nvrtc: same as nvcc
- clang-cuda: clang version + cudacc version. slightly coupled, I believe the cudacc version is bound to the clang version
- nvc++: nvc++ version and cudacc version are completely uncoupled, nvc++ can consume different cuda header versions
There was a problem hiding this comment.
@bernhardmgruber @gevtushenko Do you have a opinions here?
There was a problem hiding this comment.
@miscco it seems you already have a far better understanding what __CUDACC__ means, so I leave this decision to you.
I agree that if __CUDACC__ does not designate a CUDA compiler version, then renaming _CCCL_CUDACC_AT_LEAST to _CCCL_CUDA_COMPILER_AT_LEAST is misleading.
|
The PR is replaced by #3057. |
This PR introduces a new CUDA compiler handling, following the C++ compiler checks.
Changes:
_CCCL_CUDA_COMPILER_*now expands to the CUDA compiler version (does what_CCCL_CUDACC_VERdid)_CCCL_CUDA_COMPILER(_name [, _op, _maj [, min]])macro_CCCL_CUDA_COMPILERis replaced with_CCCL_HAS_CUDA_COMPILERflag_CCCL_CUDACC_*are replaced by_CCCL_CUDA_COMPILER_*to match the naming_CCCL_CUDA_COMPILATIONflag can be used to check if CUDA source file is being compiled_CCCL_DEVICE_COMPILATIONflag can be used to check if the device code is being compiled (should replacedefined(__CUDA_ARCH__)_CCCL_PTX_ARCHshould be valid for compilation with NVHPC compiler nowThere are still many places where the CUDA compiler version is meant to be NVCC only and should be fixed.