Enable architecture selection for DPCTL_TARGET_CUDA

[vlad-perevezentsev]

This PR proposes to change DPCTL_TARGET_CUDA CMake option from a boolean to a string allowing users to specify a CUDA architecture (e.g. sm_80). If not specified, it defaults to sm_50.

$ python scripts/build_locally.py --verbose --cmake-opts="-DDPCTL_TARGET_CUDA=<cuda_arch>"
# or
$ python scripts/build_locally.py --verbose --cmake-opts="-DDPCTL_TARGET_CUDA=ON"

The specified architecture is used to construct a SYCL alias target (e.g. nvidia_gpu_sm_80) and passed via -fsycl-targets option, following OneAPI for NVIDIA GPUs

Additionally removing DPCTL_TARGET_CUDA env handling logic

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to an issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?
• Have you checked performance impact of proposed changes?
• Have you added documentation for your changes, if necessary?
• Have you added your changes to the changelog?
• If this PR is a work in progress, are you opening the PR as a draft?

[github-actions]

View rendered docs @ https://intelpython.github.io/dpctl/pulls/2096/index.html

[github-actions]

Array API standard conformance tests for dpctl=0.21.0dev0=py310h93fe807_8 ran successfully.
Passed: 1115
Failed: 6
Skipped: 119

[coveralls]

coverage: 84.972% (-0.01%) from 84.984%
when pulling feee948 on update_cuda_build
into 556a5c6 on master.

[github-actions]

Array API standard conformance tests for dpctl=0.21.0dev0=py310h93fe807_9 ran successfully.
Passed: 1114
Failed: 7
Skipped: 119

[github-actions]

Array API standard conformance tests for dpctl=0.21.0dev0=py310h93fe807_10 ran successfully.
Passed: 1114
Failed: 7
Skipped: 119

[antonwolfy]

Is it fair to validate DPCTL_TARGET_CUDA only in case when empty DPCTL_SYCL_TARGETS?

[ndgrigorian]

it was how we were doing it before—but it looks like current logical flow will add HIP targets even when DPCTL_SYCL_TARGETS is not none, but not CUDA

so that should probably be changed, either make DPCTL_SYCL_TARGETS exclusive from both or check DPCTL_TARGET_CUDA as well

[vlad-perevezentsev]

In my understanding, when the user passes DPCTL_SYCL_TARGETS he is responsible for the correctness of the flags.

The logic of checking if (NOT “x${DPCTL_TARGET_HIP}” STREQUAL “x”) when DPCTL_SYCL_TARGETS is not none was added to pass the correct compile and link options.

if(_dpctl_amd_targets)
      list(APPEND _dpctl_sycl_target_compile_options -Xsycl-target-backend=amdgcn-amd-amdhsa --offload-arch=${_dpctl_amd_targets})
      list(APPEND _dpctl_sycl_target_link_options -Xsycl-target-backend=amdgcn-amd-amdhsa --offload-arch=${_dpctl_amd_targets})
   endif()

I am already working on PR that will refresh the logic for AMD build using aliases to remove if(_dpctl_amd_targets) branch.

[antonwolfy]

The mapping is not clear from the reference doc.

[ndgrigorian]

Yes, seems they aren't necessarily related either (see here and below it)

[vlad-perevezentsev]

A CUDA developer notes that sm_XX refers to machine code for a specific GPU hardware architecture. Since each Compute Capability version corresponds to a particular architecture (CC 8.0 -> Ampere A100) it is reasonable to say that sm_80 corresponds to CC 8.0

I changed the text a bit

[github-actions]

Array API standard conformance tests for dpctl=0.21.0dev0=py310h93fe807_17 ran successfully.
Passed: 1113
Failed: 8
Skipped: 119