Loop with GPU arrayfun

I'm looking at the Matlab example on rendering Fractals with GPUs ( https://www.mathworks.com/help/distcomp/examples/illustrating-three-approaches-to-gpu-computing-the-mandelbrot-set.html ) and it works fine. But I want to loop this, and I get this error:

Error using gpuArray.linspace
An unexpected error occurred trying to launch a kernel. The CUDA
error was:
invalid configuration argument

Error in matlabArrFunLoop (line 11)
x = gpuArray.linspace( xlim(1)/k, xlim(2)/k, gridSize );

What's wrong with looping this code? Thanks

Code:

clear
close all
gridSize = 3000;
xlim = [-0.748766713922161, -0.748766707771757];
ylim = [ 0.123640844894862,  0.123640851045266];

for k=1:1:3
% Setup
tic
k
x = gpuArray.linspace( xlim(1)/k, xlim(2)/k, gridSize );
y = gpuArray.linspace( ylim(1)/k, ylim(2)/k, gridSize );
[xGrid,yGrid] = meshgrid( x, y );

% Calculate
count = arrayfun( @iterFunc,xGrid, yGrid );

% Show
count = gather( count ); % Fetch the data back from the GPU
figure(3)
imagesc( x, y, count )
axis off
toc
end

function count = iterFunc(x0,y0)
maxIterations=5000;
z0 = complex(x0,y0);
z = z0;
count = 1;
while (count <= maxIterations) && (abs(z) <= 2)
    count = count + 1;
    z = z*z + z0;
end
count = log(count);
end

[Walter Roberson]

I get that message, but first I get

Error using gpuArray/arrayfun
The GPU has timed out. See www.mathworks.com/gputimeout for details about GPU timeouts and how to avoid them.

I have some other things going on with my computer at the moment which might be trying to use some of the GPU cycles. But just in general, gpu can time out if the work it is asked to do takes longer then the kernel hold time that is imposed for GPUs that are not in exclusive mode.

Answer 1

[Matt J]

The code runs fine for me. Try restarting MATLAB and/or the computer.

[Maxime Rivard]

Did you run it multiple times?

[Matt J]

About 5 or 6 times. How many times was I supposed to run it?

[Maxime Rivard]

I get it the first time. What gpu do you have? I have a 1080 ti.

                    Name: 'GeForce GTX 1080 Ti'
                       Index: 1
           ComputeCapability: '6.1'
              SupportsDouble: 1
               DriverVersion: 8
              ToolkitVersion: 8
          MaxThreadsPerBlock: 1024
            MaxShmemPerBlock: 49152
          MaxThreadBlockSize: [1024 1024 64]
                 MaxGridSize: [2.1475e+09 65535 65535]
                   SIMDWidth: 32
                 TotalMemory: 1.1811e+10
             AvailableMemory: 9.7107e+09
         MultiprocessorCount: 28
                ClockRateKHz: 1582000
                 ComputeMode: 'Default'
        GPUOverlapsTransfers: 1
      KernelExecutionTimeout: 1
            CanMapHostMemory: 1
             DeviceSupported: 1
              DeviceSelected: 1

[Walter Roberson]

I do not get this the first time on my 650M: I get a response that the GPU has timed out. When you have a GPU device that is shared between graphics and compute usage, then kernels typically have a fairly limited timeout.

http://developer.download.nvidia.com/NsightVisualStudio/2.2/Documentation/UserGuide/HTML/Content/Timeout_Detection_Recovery.htm

https://www.mathworks.com/matlabcentral/answers/98712-why-am-i-not-able-to-access-a-gpu-card-on-a-windows-vista-windows-7-machine-managed-under-job-manage

The information I find suggests that the GeForce GTX 1080 Ti does not support TCC.

https://devtalk.nvidia.com/default/topic/1023458/announcements/can-anyone-confirm-that-the-1080-ti-does-not-support-tcc-/

[Maxime Rivard]

As I have a very powerful card, I don't seem to get the timeout. However, when I loop large matrices, I get

Error using gpuArray.linspace
An unexpected error occurred trying to launch a kernel. The CUDA
error was:
invalid configuration argument

For example, I can run my algorithm (not the one posted) with 1080p frames to the end, but if I run the same algorithm with 4k frames, I invariably get this error at the 191th frame...

[Matt J]

I have a Titan X, but I've disabled the timeout.

[Joss Knight]

Timeouts can have some very odd side-effects. When I have timeouts enabled I get a launch failure issue with your code. But when I disable timeouts, I get no problems. So I recommend you try that first. The simplest way is to create a new DWORD registry key called TdrLevel in the appropriate place and set to 0, following the instructions on this page:

https://docs.microsoft.com/en-us/windows-hardware/drivers/display/tdr-registry-keys

Then reboot and try again.

[Maxime Rivard]

It solved the issue. Thanks!

[Hoang Nguyen]

Hello Joss, is there any problem if we disable timeouts by your suggestion. In fact, I could run any examples utilizing GPU but not so sure about consequences since that registry key is not built-in.

[Joss Knight]

There is potential for system lock-up if you are using perhaps less rigorous GPU code that may contain bugs. If the graphics card can't draw graphics because it's locked up running some CUDA kernel then to all intents and purposes your system has crashed.

But this is very rarely an issue for most people.