4

I am trying to figure out how to use the batch mode offered in the CUFFT library.

I basically have an image that is 5300 pixels wide and 3500 tall. Currently this means I am running 3500 1D FFT's on those 5300 elements using FFTW.

Is this a good candidate problem to run the CUFFT library in batch mode? How does the data have to be set up to do this problem?

Thanks

Derek
  • 11,715
  • 32
  • 127
  • 228

2 Answers2

5

yes, you can use the batch mode.

To use the batch mode,the 5300 elements should be stored continuously.

That means the distance between adjacent batches is 5300. You can go this way:

..........
cufftComplex *host;
cufftComplex *device;
CudaMallocHost((void **)&host,sizeof(cufftComplex)*5300*3500);
CudaMalloc((void **)&devcie,sizeof(cufftComplex)*5300*3500);
//here add the elements,like this:
//host[0-5299] the first batch, host[5300-10599] the second batch ,and up to the 3500th batch.
CudaMemcpy(device,host,sizeof(cufftComplex)*5300*3500,......);
CufftPlan1d(&device,5300,type,3500);
CufftExecC2C(......);
......

For more details see the CUFFT Manual.

user1676582
  • 61
  • 1
  • 1
2

yes this is a good problem.

You should go the following way:

  1. create a array with size: sizeof(cufftComplex)*5300*3500 at the gpu(here I assume that you have complex input data)
  2. copy your data to the gpu
  3. create a plan with cufftPlan1d()
  4. execute the plan for example with cufftExecC2C()

For more Information you must have a look at the CUFFT Manual

moggi
  • 1,466
  • 4
  • 18
  • 29