New to MKL, Memory Crashes Occuring

[Thomas_M_3]

I am new to the MKL.  My platform is Win7 (64bit), Visual Studio 2010, and I am using the MSVC compiler / linker, not Intel.  I am building a 32 bit project.  I have version 11.1 of the Intel Compiler and whatever the associated MKL is.  I just made my first project in MSVC using MKL functions, and I'm getting fatal memory errors.

Running this project causes a fatal error.  I've tried it in the National Instruments C IDE as well, and it causes "dynamic memory corruption" and general protection faults.  What's going on?? I configured MSVC 2010 to use MKL by the following steps: including the 32 bit static libraries: mkl_core.lib, mkl_intel_c.lib, mkl_intel_thread.lib, and libiomp5md.lib.  In addition, I placed libiomp5md.dll in my executable directory and added the MKL ia32 include directory to my include paths.

#include "mkl.h"
#include <stdio.h>
#include <string.h>
#include <math.h>#define ARRAYSIZE 8096static double* fftArray = NULL;
static double* zeros = NULL;
static MKL_LONG status;
static DFTI_DESCRIPTOR_HANDLE descriptorHandle;void initializeFFT()
{
int i;for (i = 0; i < ARRAYSIZE; i++)
fftArray = sin((double)i/3) + sin((double)i/30) + sin((double)i/100);memset(zeros, 0, ARRAYSIZE * sizeof(double));
}int main()
{
fftArray = (double*)mkl_malloc(ARRAYSIZE * sizeof(double), 16);
zeros = (double*)mkl_malloc(ARRAYSIZE * sizeof(double), 16);
intializeFFT();
DftiCreateDescriptor(&descriptorHandle, DFTI_DOUBLE, DFTI_COMPLEX, 1, ARRAYSIZE);
DftiCommitDescriptor(descriptorHandle);
status = DftiComputeForward(descriptorHandle, fftArray, zeros);
DftiFreeDescriptor(&descriptorHandle);mkl_free_buffers();
}

[Evgueni_P_Intel]

Hi Thomas, If you expect MKL to do complex-to-complex FFT of length ARRAYSIZE, please allocate 2*sizeof(double)*ARRAYSIZE bytes for the input array and 2*sizeof(double)*ARRAYSIZE bytes for the output array. Thanks, Evgueni.

[Thomas_M_3]

Thanks, that might be the problem. Let me ask a quick clarification though. If I'm performing the complex FFT in place, two buffers are required: real in/out, and imaginary in/out. If my real and imaginary data that I need transformed are both, say, 4096 pts long, I have two buffers that are 4096. You are saying I cannot simply pass these two 4096 element buffers to the FFT function, that I need to double the length of both the real and imaginary input buffers to 8096? I was under the impression that the output from an FFT was always the length of the input. Therefore if you had a real and an imaginary buffer both of ARRAYSIZE, the output would also be two buffers of ARRAYSIZE. What part of this is wrong?

[Evgueni_P_Intel]

If you want MKL to do an FFT in-place on real and imaginary parts stored in different arrays (split-complex numbers), please add [cpp] DftiSetValue(&descriptorHandle, DFTI_COMPLEX_STORAGE, DFTI_REAL_REAL); [/cpp] between the following lines [cpp] DftiCreateDescriptor(&descriptorHandle, DFTI_DOUBLE, DFTI_COMPLEX, 1, ARRAYSIZE); DftiCommitDescriptor(descriptorHandle); [/cpp] Without DFTI_COMPLEX_STORAGE set to DFTI_REAL_REAL, MKL interprets DftiComputeForward(handle, x, y) as out-of-place FFT on interleaved complex data. You may also want to check the examples examples/dftc/source/config_complex_storage.c in the MKL intallation directory. Thanks, Evgueni.

[Thomas_M_3]

Thanks - this is what I think I was missing! I was just reading the documentation for DftiComputeForward, which lists: status = DftiComputeForward(desc_handle, xre_inout, xim_inout); as one of the four parameter options. However, it was missing the key configuration information that you provided! I looked over the examples again as you suggested, and indeed the 1D in place complex transform uses an array of type "_Complex", which I did not notice before.

[Evgueni_P_Intel]

We will pass your input to the MKL documentation team. Thank you for trying Intel MKL!

[Thomas_M_3]

I just wanted to update the thread that the suggested fix works. As an aside, I tested the speed of both interleaved (COMPLEX_COMPLEX) and separate-array (REAL_REAL) storage for the complex forward transform. As expected, the interleaved data storage leads to faster execution time.