CudaFFT 2D Array transform

[SmokhCR]

Dear Kunzmi,

I can see and test CudaDeviceVariable'<'cuFloatComplex'>' d_signal 1D array transformation.
However, in my project I need to use 2D array of cuFloatComplex.

Could you please suggest ways of implementing this?

Sincerely,

[kunzmi]

Hi,

I'm not sure how to understand your question. Do you have general problems in extending a 1D FFT to 2D or is it the 2D representation of your data in memory? Or do you have troubles setting up the CUFFT-plan for 2D transforms? Or something completely different?

Michael

[SmokhCR]

Dear Michael,

I have found the only class compatible with 2D C# arrays cuFloatComplex[,].
public CudaPitchedDeviceVariable<cuFloatComplex> d_signal;

In my example I have 2D C# matrix 4x4

.....
       cuFloatComplex[,] h_signal = new cuFloatComplex[4,4];
.....
       plan = new CudaFFTPlan2D(4, 4, cufftType.C2C);
       d_signal.CopyToDevice(h_signal);
       plan.Exec(d_signal.DevicePointer,  TransformDirection.Forward); 
       d_out_signal.CopyToHost(h_signal);

unfortunately, after Forward transform, I don't see real transformed data. Resulting array is just mixture of random or zero
values. I'm pretty sure I'm missing some important instructions of just using wrong classes.
Please help me to start.

Sincerely,

[kunzmi]

Hi
The problem is that you mix pitched and continuous memory. CudaPitchedDeviceVariable adds a few bytes at each line to ensure that all lines start at a “good” memory address – that's for example necessary for binding memory to texture units etc.

CudaFFTPlan2D on the other hand cannot deal with pitched memory allocations, that is why your results are wrong.

Either change CudaPitchedDeviceVariable to CudaDeviceVariable to get continuous memory allocations that fit to CudaFFTPlan2D or use CudaFFTPlanMany which allows to set the pitch parameters:

//note that cufft has "inverted" width/height order:
int[] size = new int[2] { data.Height, data.Width };
int[] inEmbed = new int[2] { data.Height, data.Pitch / (int)cuFloatComplex.SizeOf };
int[] onEmbed = new int[2] { data.Height, data.Pitch / (int)cuFloatComplex.SizeOf };

//create a cufft plan:
CudaFFTPlanMany fftPlan = new CudaFFTPlanMany(2, size, 1, cufftType.C2C, inEmbed, 1, 1, onEmbed, 1, 1);

[SmokhCR]

Hi Michael,
thank you very much! It makes great sense. Simple CudaDeviceVariable works fine with 2D C# arrays and CudaFFTPlan2D.
Could you please, tell a bit more about Pitched size. What it can be?
E.g. we have 256x256 matrix, what Pitch value would be or recommended?

Thanks in advance,