Support CUDA 11.5.1

Tasks

• Read CUDA Release Notes.
• Update CUDA driver in FlexCI (Linux/Windows) test infrastructure (ask @kmaehashi). -> not needed
• ~Add FlexCI configuration for the CUDA version.~
• Update compute capability support.
  • Check NVRTC Reference (–gpu-architecture option) and change the maximum value of supported compute capability defined in cupy/cuda/compiler.py (_nvrtc_max_compute_capability) and _cc_major_map in cupy/fft/_callback.pyx, if needed. #6166
  • Check NVCC Reference (–generate-code) and change the list of architectures in cupy_setup_build.py (_arch_list) if needed. #6166
  • Backport. #6331
• ~Update bundled FP16 header files.~
  • ~Copy FP16 header files to cupy/_core/include/cupy/_cuda/, then update cupy/_core/core.pyx to use it.~
  • ~Backport.~
• ~Update library installer.~
  • ~Fix cupyx/tools/install_library.py.~
  • ~Backport.~
• ~Wheel distribution~
  • ~Update cupy-release-tools. https://github.com/cupy/cupy-release-tools/pull/221~
  • ~Add the package to duplicate detection.~
  • ~Backport.~
• ~Documentation~
  • ~Add new wheel package to Installation Guide and README.md. Update requirements in the installation guide.~
  • ~Backport.~
  • ~Add new wheel package to the website.~
• ~Implement or create an issue to support new features, if applicable.~

[leofang]

Yes, please check with nvrtcGetSupportedArchs(). NVRTC docs are not fully up-to-date and I’ve brought it to internal attention a while ago. For example, sm_86/compute_86 are also missing (but they’ve been supported since early 11.x).

[takagi]

>>> nvrtc.getSupportedArchs()
(35, 37, 50, 52, 53, 60, 61, 62, 70, 72, 75, 80, 86, 87)