topic in Intel® oneAPI Math Kernel Library

about 2D FFT

Dawn_L_ — Thu, 20 Nov 2014 19:28:05 GMT

Hi,

Below is a block of codes for using 2D FFT from MKL:

MKL_LONG flen1[2]={nffty,nfftx};
MKL_LONG status;

DFTI_DESCRIPTOR_HANDLE hand1;
status=DftiCreateDescriptor(&hand1,DFTI_SINGLE,DFTI_COMPLEX,2,flen1);
status=DftiCommitDescriptor(hand1);
status=DftiComputeForward(hand1,tmp);

tmp is a 2D complex array.

problems:

tmp in DftiComputeForward: if tmp is given by tmp[ ][ ] the compiling can go through and the block codes works well. If tmp is given by **tmp and alloc memory to it, the compiling can go through but executing is crushed at DftiComputeForward.

So, the problem is in/out parameter in DftiComputeForward has to be hard-coding 2D array and cannot be pointer type?

Any answer is welcomed. Thank you very much.

You are correct, the i/o

Dmitry_B_Intel — Fri, 21 Nov 2014 03:44:14 GMT

You are correct, the i/o pointers should point directly to the data, not to the pointers to the rows/columns. See also what FFTW3 documentation says on this.

Thanks
Dima

Dawn_L_ — Fri, 21 Nov 2014 16:41:03 GMT

Thank you, Dima, for your reply.

Same block code with different problem as below:

complex tmp[nffty][nfftx],swap[nffty][nfftx];

MKL_LONG flen1[2];
flen1[0]=nffty;
flen1[1]=nfftx;

MKL_LONG status;

DFTI_DESCRIPTOR_HANDLE hand1;
status=DftiCreateDescriptor(&hand1,DFTI_SINGLE,DFTI_COMPLEX,2,flen1);
status=DftiCommitDescriptor(hand1);
status=DftiComputeForward(hand1,tmp);

Problem: If I define nffty=1024 and nfftx=1024, the executing of program is crushed at "complex tmp[nffty][nfftx],swap[nffty][nfftx];"

Do you think there is size limitation for 2D complex array of tmp and swap?

Thank you very much in advanced.

Yes, 16 megabytes exceed the

Evgueni_P_Intel — Tue, 25 Nov 2014 04:46:40 GMT

Yes, 16 megabytes exceed the typical size of thread stack.

Try allocating the arrays at run time or extend the size of stack using an appropriate compiler switch.