Accessing complex workspace data from a CUDA kernel

In my MATLAB workspace, I have some complex data that I wish to transfer to the GPU for further processing with a CUDA kernel. Transferring the data to the GPU is easy with the gpuArray command.

However, what data type should I specify in the .cu kernel prototype to access a complex gpuArray? In the Parallel Computing Toolbox documentation, cuComplex is not listed as a supported data type for the C prototype, and when I use float2 I receive an error message about the complexity of the input not matching the prototype:

"parallel.gpu.GPUArray must match the complexity specified on the kernel prototype"

Regards, Jeremy Dillon

[King Fisher]

Have you written CUDA Kernel? and you are activating it by parallel.gpu.kernel('.ptx','.cu') command? if yes Would you please share your code?

Answer 1

[Edric Ellis]

You should use 'double2' to match complex double MATLAB data, and 'float2' for complex single MATLAB data.

[Jeremy Dillon]

Thank you Edric, that worked. When I tried float2 I had forgotten to initialize my output array as complex, e.g.

o = complex(parallel.gpu.GPUArray.zeros([1 N],'single'));