[jira] [Updated] (ARROW-12300) ArrowCUDA erroneously links to CUDA Runtime while only using CUDA Driver API

["Keith Kraus (Jira)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/ARROW-12300?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Keith Kraus updated ARROW-12300:
--------------------------------
    Summary: ArrowCUDA erroneously links to CUDA Runtime while only using CUDA Driver API  (was: ArrowCUDA erroneously links to CUDA Runtime while only using Driver API)

> ArrowCUDA erroneously links to CUDA Runtime while only using CUDA Driver API
> ----------------------------------------------------------------------------
>
>                 Key: ARROW-12300
>                 URL: https://issues.apache.org/jira/browse/ARROW-12300
>             Project: Apache Arrow
>          Issue Type: Bug
>          Components: GPU, Packaging
>    Affects Versions: 3.0.0
>            Reporter: Keith Kraus
>            Priority: Major
>
> As far as I can tell the Arrow GPU code only uses the CUDA driver API, but the CMake links it to the CUDA Runtime: https://github.com/apache/arrow/blob/apache-arrow-3.0.0/cpp/src/arrow/gpu/CMakeLists.txt#L39 ({{${CUDA_LIBRARIES}}} is defined via the CMake {{FindCUDA}} module here: https://cmake.org/cmake/help/latest/module/FindCUDA.html#result-variables)
> By default the static library for the CUDA Runtime will be used and things will get statically linked inside of the Arrow GPU library unnecessarily. Additionally, this can cause issues related to CUDA initialization as shown here: https://github.com/rapidsai/cudf/issues/7600



--
This message was sent by Atlassian Jira
(v8.3.4#803005)