CUDA_ERROR_LAUNCH_FAILED when using "gather" function

Dear all,

I created two cuda functions, each one in a separated .cuh file, that are called from my main .cu program. I am trying to call then inside my Matlab code and everything seems to work fine, until I call the "gather" function, in order to get the variable back to the cpu. It gives me the following error message:

Error using gpuArray/gather An unexpected error occurred during CUDA execution. The CUDA error was: CUDA_ERROR_LAUNCH_FAILED

And this is the only message that Matlab returns to me. The strange thing is that if I call just one function, it works fine. The problem arises if I call the two functions sequentially and try to gather the resulting vector from the graphics card. Here is my example file:

clear all; clc;

kernelK = parallel.gpu.CUDAKernel('femCuda.ptx', 'fem.cuh', 'setK');

kernelKGlobal = parallel.gpu.CUDAKernel('femCuda.ptx', 'fem2.cuh', 'assemblyGlobalStiffness');

nEl = 2;

ndofs = 2;

nNodes = 4;

x = [0. 1. 1. 0.];

y = [0. 0. 1. 1.];

vecE = [1.];

vecNu = [0.];

conect = int32([0 0 1 3 0 1 2 3]);

x = gpuArray(x);

y = gpuArray(y);

conect = gpuArray(conect);

vecE = gpuArray(vecE);

vecNu = gpuArray(vecNu);

K = gpuArray.zeros(36*nEl, 1);

Kglobal = gpuArray.zeros((nNodes*ndofs)*(nNodes*ndofs), 1);

K = feval(kernelK, x, y, conect, vecE, vecNu, nNodes, nEl, nEl, K);

Kglobal = feval(kernelKGlobal, K, nEl, nNodes, conect, nNodes, Kglobal);

If I just use K = feval(kernelK, x, y, conect, vecE, vecNu, nNodes, nEl, nEl, K);, everything works fine. The problem is when I call the two kernels. What can possibly be wrong?

Thank you in advance,

[Joss Knight]

Can you confirm that the second kernel also works in isolation?

If you put wait(gpuDevice) between your two calls, does it start working?

Can you provide your CUDA code, or a minimal version which reproduces the error?