Document minimum CUDA version of 11.4

[harrism]

Description

With the addition of libcudacxx 2.1.0, minimum CUDA version required to build RMM is now 11.4. This PR updates the readme to reflect this.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[jrhemstad]

Not that it matters, but libcu++ should support 11.1.

[bdice]

Approving with one question for @raydouglass about whether RAPIDS should document driver support or only CTK support.

[jakirkham]

Think we will want to update these constraints to match whatever minimum we use:

rmm/conda/recipes/librmm/meta.yaml

Line 78 in 682e2f3

- cuda-version {{ cuda_spec }}

rmm/conda/recipes/rmm/meta.yaml

Line 71 in 682e2f3

- {{ pin_compatible('cuda-version', max_pin='x', min_pin='x') }}

[bdice]

@jakirkham This PR is changing the docs for build requirements, not run requirements. Those conda recipe specs are relevant for runtime requirements but not for setting the minimum version to build.

[bdice]

Not that it matters, but libcu++ should support 11.1.

@benfred reported that libcudacxx 2.1.0 includes code that requires CUDA 11.4 to compile. @benfred Can you confirm what piece requires CUDA 11.4?

[bdice]

Looks like the failure was here: https://github.com/benfred/implicit/actions/runs/6947998720/job/18903049451?pr=703#step:4:2193

        FAILED: implicit/gpu/CMakeFiles/_cuda.dir/als.cu.o
        /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler -DFMT_HEADER_ONLY=1 -DLIBCUDACXX_ENABLE_EXPERIMENTAL_MEMORY_RESOURCE -DSPDLOG_FMT_EXTERNAL -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_CUDA -DTHRUST_HOST_SYSTEM=THRUST_HOST_SYSTEM_CPP -D_cuda_EXPORTS -I/project/. -I/opt/_internal/cpython-3.8.18/include/python3.8 -I/project/_skbuild/linux-x86_64-3.8/cmake-build/_deps/raft-src/cpp/include -I/project/_skbuild/linux-x86_64-3.8/cmake-build/_deps/rmm-src/include -I/project/_skbuild/linux-x86_64-3.8/cmake-build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include -I/project/_skbuild/linux-x86_64-3.8/cmake-build/_deps/thrust-src -I/project/_skbuild/linux-x86_64-3.8/cmake-build/_deps/thrust-src/dependencies/cub -I/project/_skbuild/linux-x86_64-3.8/cmake-build/_deps/fmt-src/include -I/project/_skbuild/linux-x86_64-3.8/cmake-build/_deps/spdlog-src/include -isystem /usr/local/cuda/include --extended-lambda -Wno-deprecated-gpu-targets -Xfatbin=-compress-all --expt-relaxed-constexpr -O3 -DNDEBUG -std=c++17 "--generate-code=arch=compute_60,code=[compute_60,sm_60]" "--generate-code=arch=compute_70,code=[compute_70,sm_70]" "--generate-code=arch=compute_80,code=[compute_80,sm_80]" "--generate-code=arch=compute_86,code=[compute_86,sm_86]" -Xcompiler=-fPIC -MD -MT implicit/gpu/CMakeFiles/_cuda.dir/als.cu.o -MF implicit/gpu/CMakeFiles/_cuda.dir/als.cu.o.d -x cu -c /project/implicit/gpu/als.cu -o implicit/gpu/CMakeFiles/_cuda.dir/als.cu.o
        /project/_skbuild/linux-x86_64-3.8/cmake-build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/detail/libcxx/include/__concepts/../__concepts/../__concepts/swappable.h(173): error: A __device__ variable cannot be marked constexpr
        
        /project/_skbuild/linux-x86_64-3.8/cmake-build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/memory_resource(203): error: A __device__ variable cannot be marked constexpr
        
        2 errors detected in the compilation of "/project/implicit/gpu/als.cu".

Coming from memory_resource:203, which uses _LIBCUDACXX_CPO_ACCESSIBILITY to make it __device__ constexpr (or possibly const on CUDA <=11.2?).

[bdice]

Support for early CUDA 11 versions has been fixed in libcudacxx: NVIDIA/libcudacxx@015cd67

We will have this fix once we migrate to CCCL 2.2.0, but we only test RAPIDS with CUDA 11.4 so it's safer to require that. The changes in this PR should be fine.

[harrism]

The constexpr error in libcu++ is fixed in 2.2.0, which RMM does not use yet. In any case we test only on 11.4+, so let's go with documenting that as the minimum.