CUDAlink does not support nVIDIA CUDA Toolkit

[bertramaerts (Bert Aerts)]

Hi,
Currently both Mathematica 13.1 and 13.2 only support CUDA Toolkit 11.8 from nVIDIA.
When will CUDAlink support CUDA Toolkit 12.0 from nVIDIA?

Kind regards,
Bert

[valzapodvz (ZAQU zaqu)]

Are you sure it is not a fake limitation, cause major version changed?

[bertramaerts (Bert Aerts)]

I contacted Wolfram support and got as answer:

Our developers tell me that they are currently working on this and support could be added in 13.2.1.

[valzapodvz (ZAQU zaqu)]

Meanwhile installed 11.8. CUDA works now, but OpenCL does not I was also surprised it apparently needs CUDA now…

Looks like 13.2 broke OpenCL.

[bertramaerts (Bert Aerts)]

On my laptop running Ubuntu 22.10 both CUDA 11.8 and OpenCL 3.0 are working fine. I start Mathematica 13.2 with this shell script:

cat mathematica.sh 
export NVIDIA_DRIVER_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu/libnvidia-tls.so.525.60.11
export CUDA_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu/libcuda.so
export PATH=/home/bert/bin_gcc_11:/usr/local/cuda-11.8/bin:$PATH$
g++ --version
mathematica &

Then OpenCL looks like:

[valzapodvz (ZAQU zaqu)]

Yeah, on windows 11 OpenCLQ[] fails already, so does info command.

There is also an issue that even though Visual Studio 2022 is used by Mathematica to compile CUDA, and it works great, CUDACCompilers[]

still prints (but that is old issue)

{{“Name” → “Visual Studio”,
“Compiler” →
CCompilerDriverVisualStudioCompilerVisualStudioCompiler,
“CompilerInstallation” →
“C:\Program Files (x86)\Microsoft Visual
Studio\2019\Enterprise”, “CompilerName” → Automatic}}

[bertramaerts (Bert Aerts)]

Unfortunately Mathematica 13.2.1 still doesn’t support nVIDIA CUDA Toolkit 12.0.1

Wolfram support says they are still working on it…

[valzapodvz (ZAQU zaqu)]

So do you know how to make OpenCL work on windows? It still does not work with 13.2.1 and latest Nvidia driver and Toolkit 11.8.

[bertramaerts (Bert Aerts)]

CUDA Toolkit 12.2 released today. Mathematica 13.3.0 released today. On Ubuntu 23.04 with
nVIDIA driver 535.54.03 on GeForce RTX 2060 laptop.

Example from Wolframs Documentation FAILS: CUDAMap.

And CUDADot seems to work without error message, but gives wrong results.

Only specifying a CUDA kernel function gives correct results

What am I doing wrong?
Can anyone confirm the same behaviour?
[1]: https://community.wolfram.com//c/portal/getImageAttachment?filename=Screenshotfrom2023-06-2822-07-27.png&userId=340012
[2]: https://community.wolfram.com//c/portal/getImageAttachment?filename=Screenshotfrom2023-06-2822-07-11.png&userId=340012
[3]: https://community.wolfram.com//c/portal/getImageAttachment?filename=Screenshotfrom2023-06-2822-10-10.png&userId=340012
[4]: https://community.wolfram.com//c/portal/getImageAttachment?filename=Screenshotfrom2023-06-2822-02-53.png&userId=340012
[5]: https://community.wolfram.com//c/portal/getImageAttachment?filename=Screenshotfrom2023-06-2822-14-12.png&userId=340012

[valzapodvz (ZAQU zaqu)]

Well, on 13.3 and Toolkit 11.8 {0.540302, -0.416147, -0.989992, -0.653644, 0.283662, 0.96017,
0.753902, -0.1455, -0.91113, -0.839072} is printed.

Did not test 12.2 toolkit yet.

[fahimc (Fahim Chandurwala)]

Very sorry about this. There should be a paclet update soon to resolve this issue on Linux (WL v13.3).
In the meantime, you could try renaming the underlying Examples library:

e.g.:

mv ~/Wolfram/Mathematica/13.3/SystemFiles/Links/GPUTools/LibraryResources/Linux-x86-64/Examples_12.so ~/Wolfram/Mathematica/13.3/SystemFiles/Links/GPUTools/LibraryResources/Linux
-x86-64/libExamples_12.so

restart WL and retry:

Needs["CUDALink`"]
CUDAQ[]
CUDAMap[Cos, 1.*Range[10]]

Also, you mention CUDADot not giving the correct result, could you please provide an example.

[devb (Dev Budhathoki)]

This behavior is reproducible and a new paclet release with the fix will release soon. For now, the workaround would be renaming following two files “Examples_12.so” and “Examples_11.so” present in

FileNameJoin[{$InstallationDirectory, "SystemFiles", "Links", "GPUTools", "LibraryResources", "Linux-x86-64"}]

to “libExample_12.so” and “libExample_11.so” respectively.

[bertramaerts (Bert Aerts)]

OK thanks for the fix of renaming the library files.
CUDAMap now gives correct results.

But please look at below image : CUDADot gives wrong results:

[bertramaerts (Bert Aerts)]

Mathematica 13.3.1 finally fully supports nVIDIA CUDA Toolkit 12.2.1
The CUDADot issue is solved.
Tested on Ubuntu 23.04 with nVIDIA 535.98 driver.

[valzapodvz (ZAQU zaqu)]

Meawhile

Get["OpenCLLink`OpenCLLink`"]

still prints and OpenCL does not work.

SetDelayed::write: Tag NVCCCompiler in NVCCCompiler[Available] is Protected.

SetDelayed::write: Tag NVCCCompiler in NVCCCompiler[Name][] is Protected.

SetDelayed::write: Tag NVCCCompiler in NVCCCompiler[Name][] is Protected.
General::stop: Further output of SetDelayed::write will be suppressed during this calculation.
Set::write: Tag NVCCCompiler in Options[NVCCCompiler] is Protected.
Set::write: Tag NVCCCompiler in Options[NVCCCompiler] is Protected.
Table::iterb: Iterator {CCompilerDriver`CCompilerDriverRegistry`Private`installation,NVCCCompiler[Installations][]} does not have appropriate bounds.
Table::iterb: Iterator {CCompilerDriver`CCompilerDriverRegistry`Private`installation,NVCCCompiler[Installations][]} does not have appropriate bounds.