Verifying the architectures during compiler identification is redundant,
and requires a lot more up-front information than we should need.
It also causes unsupported architectures to break the compiler id and
version detection, so the resulting output from CMake does not report
the compiler version, which is useful information to know why the
specified architectures are not supported.
The "detecting compiler ABI info" and "check for working compiler" steps
already pass `CMAKE_CUDA_ARCHITECTURES` into their test projects.
Therefore we can just drop the earlier architecture testing. Bad
architectures will be reported as a not-working compiler, and the
output will include the compiler's error message.
This reverts the approach from:
* commit 19cc5bc296 (CUDA: Throw error if user-specified architectures
don't work, 2020-05-26, v3.18.0-rc1~79^2)
* commit 650c1029a0 (CUDA: Detect non-working user-specified architectures
on NVCC, 2020-05-28, v3.18.0-rc1~51^2)
* commit 01428c5560 (CUDA: Fail fast if CMAKE_CUDA_ARCHITECTURES
doesn't work during detection,
2020-08-29, v3.19.0-rc1~241^2).
Their goal was in part to avoid waiting until the test for working
compiler to detect unsupported architectures. However, experience has
shown that failing earlier is more trouble than it's worth.
Fixes: #23161
Issue: #20756
CUDA 11.6 added the `nvcc -arch=native` flag to automatically compile
for the host GPUs' architectures. Add support for specifying this
special `native` value in `CMAKE_CUDA_ARCHITECTURES` and
`CUDA_ARCHITECTURES`. During the compiler ABI detection step,
detect the native architectures so we can pass them explicitly
when using Clang or older versions of nvcc.
Fixes: #22375
Clang does not automatically choose an architecture supported
by the SDK. The `OFF` value is meant for users to provide their
own architecture flags, so provide one as part of the test.
The `CudaOnly.All` test only sets these in project code after CUDA is
enabled. Add another case to test the values during compiler detection.
Issue: #23161
If empty we otherwise treat it the same as unset in most places, but still end
up failing eventually with a confusing "Failed to find a working CUDA
architecture".
This also detects some other basic invalid ones (e.g. "al").
Robert Maynard