Matlab gpuArray doesn't work with sparse arrays

The Matlab gpuArray works perfectly for non-sparse matrices, e.g., x=gpuArray(eye(1)).

But when the input changed to sparse arrays, e.g., x=gpuArray(speye(1)), I got the following error message

Error using gpuArray

An unexpected error occurred on the device. The error code was: UNKNOWN_ERROR.

How can I fix it?

My gpu is RTX 4090 with nvidia game ready driver 528.29, cuda version 12.0,

My gpuDevice outputs the following results

CUDADevice with properties:

Name: 'NVIDIA GeForce RTX 4090'

Index: 1

ComputeCapability: '8.9'

SupportsDouble: 1

DriverVersion: 12

ToolkitVersion: 11.2000

MaxThreadsPerBlock: 1024

MaxShmemPerBlock: 49152 (49.15 KB)

MaxThreadBlockSize: [1024 1024 64]

MaxGridSize: [2.1475e+09 65535 65535]

SIMDWidth: 32

TotalMemory: 25756696576 (25.76 GB)

AvailableMemory: 23916261376 (23.92 GB)

MultiprocessorCount: 128

ClockRateKHz: 2535000

ComputeMode: 'Default'

GPUOverlapsTransfers: 1

KernelExecutionTimeout: 1

CanMapHostMemory: 1

DeviceSupported: 1

DeviceAvailable: 1

DeviceSelected: 1

[Walter Roberson]

Have you installed Update 4 of R2022b?

[Bruno Luong]

AFAIK the current update for R2022b is still Update 3.

It looks like the premature Update 4 had issue and they withdraw it.

[Joss Knight]

This looks like a problem specific to Ada cards (compute capability 8.9). Can you please contact MathWorks support?

[Walter Roberson]

unfortunately I cannot verify from my phone, but the recent Question about that had someone from Mathworks say that Update 4 had been restored except for offline installation.

[雨文]

I installed Update 3 of R2022b just now but still got the same error when the version was R2022b update 1

I updated my Matlab using the upper right notification icon and it was shown that Udpate 3 was avaiable (no Update 4).

Do I really need to install Update 4?

[Bruno Luong]

@Walter Roberson "someone from Mathworks say that Update 4 had been restored"

Not seeing anything showed up on my side.

[雨文]

Well, I manually downloaded R2022b Update 4 from Download MATLAB, Simulink, Stateflow and Other MathWorks Products and reinstalled it.

Unfortunately, it doesn't work neither. As Joss said, this might be a problem specific to gpu with Ada Lovelace architechture.

Answer 1

[Animesh]

Hi,

I understand that when you try to create sparse array/matrix on a GPU, it pops an error whereas when try for normal arrays/matrices on GPU it works fine.

This is a known issue with Ada Lovelace graphic cards and CUDA 11.2.

As a workaround the following steps can be helpful:

On Windows machine:

1. Install the CUDA Toolkit version 11.3.
2. Open MATLAB.
3. Before executing any other command, execute the below two commands and ignore the warning message:

>> loadlibrary('E:\3rdparty\R2023a\9378850\win64\CUDA\bin\cusparse64_11.dll', 'E:\3rdparty\R2023a\9378850\win64\CUDA\include\cuComplex.h', 'addheader', 'E:\3rdparty\R2023a\9378850\win64\CUDA\include\cusparse.h',  'addheader','E:\3rdparty\R2023a\9378850\win64\CUDA\include\cusparse.h'); 
>> loadlibrary('E:\3rdparty\R2023a\9378850\win64\CUDA\bin\cusolver64_11.dll', 'E:\3rdparty\R2023a\9378850\win64\CUDA\include\cusolver_common.h', 'addheader', 'E:\3rdparty\R2023a\9378850\win64\CUDA\include\cusolverDn.h', 'addheader','E:\3rdparty\R2023a\9378850\win64\CUDA\include\cusolverMg.h', 'addheader', 'E:\3rdparty\R2023a\9378850\win64\CUDA\include\cusolverRf.h', 'addheader','E:\3rdparty\R2023a\9378850\win64\CUDA\include\cusolverSp.h', 'addheader', 'E:\3rdparty\R2023a\9378850\win64\CUDA\include\cusolverSp_LOWLEVEL_PREVIEW.h');

4. Execute your code.

On Linux machine:

1. Install the CUDA Toolkit version 11.3.
2. Before opening MATLAB, open a terminal and set the environment variable ‘LD_PRELOAD’ to point to paths of cuSPARSE and cuSOLVER libraries of the CUDA version 11.3.

setenv LD_PRELOAD /usr/local/cuda-11.3/lib64/libcusolver.so.11.1.1.58:/usr/local/cuda-11.3/lib64/libcusolver.so:/usr/local/cuda-11.3/lib64/libcusolver.so.11:/usr/local/cuda-11.3/lib64/libcusparse.so:/usr/local/cuda-11.3/lib64/libcusparse.so.11:/usr/local/cuda-11.3/lib64/libcusparse.so.11.5.0.58

3. Open MATLAB and execute your code.

I hope this helps.