Bad Results from MKL FFT with large inverse FFT

[happyIntelCamper]

Go forward with a power of two MKL FFT single precision real to complex.
But do a reverse fft with also a power of two but eight times larger in size, but using a separate descriptor.
Program finishes but results incorrect.

Valgrind gives this:

==27556== Invalid write of size 4
==27556== at 0x9ED7FBF: ipps_initTabBitRevNorm (in /cm/apps-3.3/intel/mkl/10.0.011/lib/em64t/libmkl_def.so)
==27556== by 0x9F4BF28: ipps_initTabTwd_Large_32f (in /cm/apps-3.3/intel/mkl/10.0.011/lib/em64t/libmkl_def.so)
==27556== Address 0x4 is not stack'd, malloc'd or (recently) free'd
==27556==

How does this table get overwritten?

[Vladimir_Petrov__Int]

Hello,

The following information could shed some light on this issue:
- what is the dimension of the transform?
- what are the sizes in question?
- how the descriptors were created and what parameters were modified via DftiSetValue?

Best regards,
-Vladimir

[happyIntelCamper]

Quoting - Vladimir Petrov (Intel)

Hi,

Transform is 1D. Real to complex in place.
Forward size is 4096, and inverse fft is 16384.
DFTI forward
Status = DftiCreateDescriptor(&handle1, DFTI_SINGLE, DFTI_REAL, 1, fftLen1);
DFTI inverse
Status = DftiCreateDescriptor(&handle2, DFTI_SINGLE, DFTI_REAL, 1, fftLen2);
Only set value:
Status = DftiSetValue(&handle2, DFTI_BACKWARD_SCALE, Scale);

The above problem size fails, but 2048 forward and 4 * 2048 inverse works, as well
as 8192 forward and 4 * 8192 inverse also works?

Hello,

The following information could shed some light on this issue:
- what is the dimension of the transform?
- what are the sizes in question?
- how the descriptors were created and what parameters were modified via DftiSetValue?

Best regards,
-Vladimir

[Dmitry_B_Intel]

Hi,

It would be nice if you posted a self-contained example that fails as you described. I have not managed to reproduce the problem with the information you have given.

Thanks
Dima

[Gennady_F_Intel]

Hi Jim,

If i am not mistaken, i believe the similar problem was fixed in one of the latest versions.

btw - what's your linking line?

--Gennady