Check for support of CUDA Memory Pools at runtime (#4679)

[ao2]

Type

• Bug fix (non-breaking change which fixes an issue): Fixes CUDA runtime error - open3D v0.14.1 #4679
• New feature (non-breaking change which adds functionality). Resolves #
• Breaking change (fix or feature that would cause existing functionality to not work as expected) Resolves #

Motivation and Context

Some CUDA GPUs, like the Quadro M3000M don't support Memory Pools operations like cudaMallocAsync/cudaFreeAsync even on driver versions newer than 11020, and this can result in errors like:

CUDA runtime error: operation not supported

So check for support at runtime instead of compile time.

Checklist:

• I have run python util/check_style.py --apply to apply Open3D code style
  to my code.
• This PR changes Open3D behavior or adds new functionality.
  • Both C++ (Doxygen) and Python (Sphinx / Google style) documentation is
    updated accordingly.
  • I have added or updated C++ and / or Python unit tests OR included test
    results (e.g. screenshots or numbers) here.
• I will follow up and update the code if CI fails.
• For fork PRs, I have selected Allow edits from maintainers.

Description

A new core::cuda::SupportsMemoryPools() helper function is added, and it is used to decide whether to use cudaMallocAsync/cudaMallocFree or not.

Initially I though about adding this to the Device class (e.g. Device::IsCUDAMemoryPoolsSupported()), but this is quite CUDA specific so I decided not to.

Please note that currently I cannot test the code on non-Quadro GPUs so I would like someone else to verify that it does not break already working cases.

Thanks,
Antonio

---

This change is 

[update-docs]

Thanks for submitting this pull request! The maintainers of this repository would appreciate if you could update the CHANGELOG.md based on your changes.

[benjaminum]

Thank you for this contribution! Please have a look at the checks. There are some problems with unknown identifiers error C3861: 'cudaFreeAsync': identifier not found.

[ao2]

I pushed an update with a compilation fix and I was able to build and run locally.

[ao2]

OK, in the latest version I kept the compile time check as well to be able to build with CUDA versions older than 11.2, apparently some CI jobs are using 11.0.3

Open3D/.github/workflows/windows.yml

Line 48 in 881ae76

CUDA_VERSION: 11.0.3

[ao2]

Pushed another update to also guard cudaDevAttrMemoryPoolsSupported behind cuda version >= 11.2

Hopefully this fixes the CI windows build issues with older cuda versions.

I have no clue about the failure in the headless-docs job tho, any suggestion?

Thanks, Antonio

[benjaminum]

Pushed another update to also guard cudaDevAttrMemoryPoolsSupported behind cuda version >= 11.2

Hopefully this fixes the CI windows build issues with older cuda versions.

I have no clue about the failure in the headless-docs job tho, any suggestion?

Thanks, Antonio

Thanks! The headless-docs problem does not seem related to this PR

[benjaminum]

Reviewable status: 0 of 4 files reviewed, all discussions resolved