Version: v0.5
Standalone version of nvcuda based from Wine-Staging
This builds ONLY the 64bit version of nvcuda. For 32bit version
version see:
https://github.com/SveSop/nvcuda/tree/32bit_only
The master branch is meant for keeping up-to-date with
current driver implementation of nvcuda, but are a work in
progress and not all functions are implemented.
The DirectX implementation in the CUDA driver for linux is not
available for this relay.
Required driver is 525 series or newer.
The recommended driver is always most recent. (580+).
- [WINE] (version >= 10.0) https://www.winehq.org/
- [Meson] http://mesonbuild.com/
- [NINJA] https://ninja-build.org/
- [MINGW-W64] https://www.mingw-w64.org/
Build by running the included script:
./package-release.sh packagename destdir
If you put --enable-tests after the buildscript like this:
./package-release.sh packagename destdir --enable-tests
An executable will be placed in destdir/bin named cudatest.exe
Run this using wine: wine ./cudatest.exe and it will perform some minor
function tests to verify that your adapter is working with cuda.
Building this test executable requires the build system have mingw-w64 installed.
If you put --fakedll after the buildscript like this:
./package-release.sh packagename destdir --fakedll
The library will be built as a winelib dll.so and a fakedll .dll placed in the
output folder in typical wine folderstructure eg.
nvcuda-packagename/lib/wine/x86_64-windows and nvcuda-packagename/lib/wine/x86_64-unix
These can preferrably be used with the wine env variable WINEDLLPATH like this:
export WINEDLLPATH=nvcuda-packagename/lib/wine
This should make wineboot -u copy the fakedll to your WINEPREFIX automatically if you
use wine > 10.0. Be aware that you must use this ENV variable whenever you use that
WINEPREFIX for CUDA.
The files can also be copied directly into the wine binary folders in the same folderstructure,
and it should work the same way.
In certain cases where there is virtual or emulated hardware, nvcuda is not able to
obtain the proper PCI ID of the adapter through wine and the call to cuDeviceGetLuid
will fail. This can be overriden by using this env variable:
CUDA_FAKE_LUID=1
This will generate a LUID for the cuda adapter and return CUDA_SUCCESS from the call.
PS. This should ONLY be used in those cases where this call fails due to types of
virtual/emulated hardware as this can potentially cause other issues if enabled.