CMake CUDA features

[chuckatkins]

This adds a few useful features to the CMake code wrt how it interacts with CUDA:

1. Remove old policies and conditional branches for CMAKE_VERSION < 3.18
2. Allow the builtin CMAKE_CUDA_ARCHITECTURES variable to be used instead of GPU_COMPUTE_VER, falling back to GPU_COMPUTE_VER if CMake is too old or CMAKE_CUDA_ARCHITECTURES is not specified.
3. Add a USE_CUDA_LTO option to enable device-code link-time-optimization.
4. Allow CUDA_HOST_COMPILER and CUDA_RUNTIME_LIBRARY to be user-overridden.

The above features are implemented in such a way to preserve existing behavior when they are not specified or enabled.

[trivialfis]

Thank you for working on modernizing the CMake code! The CMAKE_CUDA_ARCHITECTURES change looks good to me, and I agree the LTO might be useful. But I have a couple questions about related changes in comments.

[chuckatkins]

@trivialfis I think this is good to go now

[trivialfis]

Let me take another look tomorrow, thank you for the patience!

[robertmaynard]

Looks great!

[trivialfis]

Hi, could you please take a look into the errors in the CI?

[chuckatkins]

Hi, could you please take a look into the errors in the CI?

Looks like tests/ci_build/prune_libnccl.sh is using some of the modified bits that needed to be changed. Should be fixed now.

[chuckatkins]

Hi, could you please take a look into the errors in the CI?

Looks like tests/ci_build/prune_libnccl.sh is using some of the modified bits that needed to be changed. Should be fixed now.

@trivialfis @robertmaynard coincidentally, in fixing this I think this might have come a cross a bug in nvprune and it's argumet parsing. Initially the changes to compute CMAKE_CUDA_ARCHITECTURES would generate something like 60-real;70-real;80 to generate cubin for 60, 70, and 80 and ptx for 80. CMake translates this to nvcc flags --generate-code=arch=compute_60,code=[sm_60] --generate-code=arch=compute_70,code=[sm_70] --generate-code=arch=compute_80,code=[compute_80,sm_80]. According to nvprune --help it's supposed to accept the same --generate-code code flags as nvcc but it chokes on the last --generate-code=arch=compute_80,code=[compute_80,sm_80] with multiple elements in the code= portion. Adjusting the CMake code to instead compute 60-real;70-real;80-real;80-virtual changes the nvcc flags for 80 to --generate-code=arch=compute_80,code=[sm_80] --generate-code=arch=compute_80,code=[compute_80] which is effectively the same thing but parsed by nvprune without issue.

[robertmaynard]

Hi, could you please take a look into the errors in the CI?

Looks like tests/ci_build/prune_libnccl.sh is using some of the modified bits that needed to be changed. Should be fixed now.

@trivialfis @robertmaynard coincidentally, in fixing this I think this might have come a cross a bug in nvprune and it's argumet parsing. Initially the changes to compute CMAKE_CUDA_ARCHITECTURES would generate something like 60-real;70-real;80 to generate cubin for 60, 70, and 80 and ptx for 80. CMake translates this to nvcc flags --generate-code=arch=compute_60,code=[sm_60] --generate-code=arch=compute_70,code=[sm_70] --generate-code=arch=compute_80,code=[compute_80,sm_80]. According to nvprune --help it's supposed to accept the same --generate-code code flags as nvcc but it chokes on the last --generate-code=arch=compute_80,code=[compute_80,sm_80] with multiple elements in the code= portion. Adjusting the CMake code to instead compute 60-real;70-real;80-real;80-virtual changes the nvcc flags for 80 to --generate-code=arch=compute_80,code=[sm_80] --generate-code=arch=compute_80,code=[compute_80] which is effectively the same thing but parsed by nvprune without issue.

I will file a bug on nvprune about this, it does look like a regression of support

[trivialfis]

Thank you for your excellent work on enabling LTO and removing the old GPU arch!