Neural networks - CUDAKernel​/setConsta​ntMemory - the data supplied is too big for constant 'hintsD'

On R2015a with Parallel Computing Toolbox and Neural Network Toolbox.

Using the following code with GPU Nvidia GeForce GTX980 Ti:

net1 = feedforwardnet(20);

net1.trainFcn = 'trainscg';

x = inputs(1:4284,2:2000)'; % if I reduce this to 2:1900, it will work

t = double(targets'); % casting to double for GPU

t = t(:,1:4284);

% preparing for GPU xg = nndata2gpu(x); tg = nndata2gpu(t);

net1.input.processFcns = {'mapminmax'}; net1.output.processFcns = {'mapminmax'};

net2 = configure(net1,x,t); % Configure with MATLAB arrays

net2 = train(net2,xg,tg);

As you can see, this is not a big dataset. When I run this, it generates this error:

Error using parallel.gpu.CUDAKernel/setConstantMemory The data supplied is too big for constant 'hintsD'.

Error in nnGPU.codeHints (line 33) setConstantMemory(hints.yKernel,'hintsD',hints.double);

Error in nncalc.setup2 (line 13) calcHints = calcMode.codeHints(calcHints);

Error in nncalc.setup (line 17) [calcLib,calcNet] = nncalc.setup2(calcMode,calcNet,calcData,calcHints);

Error in network/train (line 357) [calcLib,calcNet,net,resourceText] = nncalc.setup(calcMode,net,data);

gpuDevice is showing this:

                Name: 'GeForce GTX 980 Ti'

                     Index: 1

         ComputeCapability: '5.2'

            SupportsDouble: 1

             DriverVersion: 8

            ToolkitVersion: 6.5000

        MaxThreadsPerBlock: 1024

          MaxShmemPerBlock: 49152

        MaxThreadBlockSize: [1024 1024 64]

               MaxGridSize: [2.1475e+09 65535 65535]

                 SIMDWidth: 32

               TotalMemory: 6.4425e+09

           AvailableMemory: 5.1520e+09

       MultiprocessorCount: 22

              ClockRateKHz: 1139500

               ComputeMode: 'Default'

      GPUOverlapsTransfers: 1

    KernelExecutionTimeout: 1

          CanMapHostMemory: 1

           DeviceSupported: 1

            DeviceSelected: 1

As noted in the code above, if I reduce x marginally, it will run.

I don't understand why data of this size would generate a memory error?

Am I missing a step in preparing this for GPU?

Answer 1

[Amanjit Dulai]

I was able to reproduce your issue. The best solution is to do the GPU training a different way by using the 'useGPU' flag. This does not use the shared memory in this way, and side-steps this issue. Your example code would look like this:

net1 = feedforwardnet(20);
net1.trainFcn = 'trainscg';

x = inputs(1:4284,2:2000)';

t = double(targets'); % casting to double for GPU

t = t(:,1:4284);

net1.input.processFcns = {'mapminmax'};
net1.output.processFcns = {'mapminmax'};

net1 = train(net1,x,t,'useGPU','yes');

Answer 2

[Joss Knight]

Constant memory is a special fast read-only cache with 64KB of space. That's enough to store about 8000 elements of double-precision data. Perhaps you want to use shared memory, which will give you 16 or 48MB depending on your device configuration.

[Jacob Townsend]

Thanks a lot for the suggestion Joss. It seems the default methods are trying to allocate this to Constant Memory (in the error chain, it is running from nncalc all the way through to nnGPU.CodeHints). Are you aware of a way to change these options to default to using shared memory only?

[Joss Knight]

Sorry, I don't understand this code to know what hints.double is and why it needs to be so large. I'll see if I can get someone who knows to help you.

Answer 3

[Jacob Townsend]

Thanks for that - solved that step!