topic status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FO in Intel® oneAPI Math Kernel Library

status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FORMAT);

Erich_Zwyssig — Fri, 18 Jun 2010 11:18:06 GMT

Hello all

Am I right in assuming that the good ault dzfft1d used the CCS format (by default) and the new Dfti procedures use Pack (by default)?

Therefore
- status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FORMAT);
should get me the same DFT output in Dfti as in dzfft1d?

Do I set the CCS format before or after DftiCommitDescriptor?

I tried before but cannot see a change in the DFT output.

Any advice?

Thanks for your help

Erich

status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FO

Dmitry_B_Intel — Fri, 18 Jun 2010 11:44:39 GMT

Erich,

For 1D real-to-complex transform, DFTI assumes CCS storage as the default, the same as dzfft1d did. That is probably why you don't see a change in the output when you correctly set it before DftiCommitDescriptor.

Thanks
Dima

status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FO

Erich_Zwyssig — Fri, 18 Jun 2010 12:43:19 GMT

Hi Dima

Thank you for your reply.

> DFTI assumes CCS storage as the default

If that is the case, then I'd expect to see 0's in the data, i.e. in position '(0).im', but I don't.

Any idea what's wrong?

Thanks for your help

Erich

status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FO

Dmitry_B_Intel — Fri, 18 Jun 2010 13:23:54 GMT

Erich,
This looks like a weird bug. Could you share how do the transform,what floating point precision, what language, what MKL version do you use, what platform, what is compile/link line?
Dima

status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FO

Erich_Zwyssig — Fri, 18 Jun 2010 14:28:26 GMT

Hi Dima

MKL 10.2.2.025
Linux 2.6.18-164.15.1.el5.inf.1PAE #1 SMP Tue Apr 27 10:03:58 BST 2010 i686 i686 i386 GNU/Linux
C++

then

MKL_LONG status;
status = DftiCreateDescriptor( &pHandle, DFTI_DOUBLE, DFTI_COMPLEX, 1, L);
status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FORMAT);
status = DftiCommitDescriptor( pHandle);

later

status = DftiComputeForward( pHandle, x_fft );

for compile and link I got a makefile. Are you interested in anything special? I'd rather not share the complete file, sorry.

Thanks for your help

Erich

status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FO

barragan_villanueva_ — Mon, 21 Jun 2010 05:41:39 GMT

Hi Erich,

For real-to-complex transform you need to use

status = DftiCreateDescriptor( &pHandle, DFTI_DOUBLE, DFTI_REAL, 1, L);

Please take a look at $MKLROOT/examples/dftc/source/real_1d_ccs_double_ex*.c for more details.

Also, CCS format is by default for such transforms.

status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FO

Erich_Zwyssig — Tue, 06 Jul 2010 12:13:28 GMT

Hi Victor

Thanks for your help.

Question, what does the order of FFT, iFFT, setting the DFTI_BACKWARD_SCALE factor and
committing need to be?

status = DftiSetValue( pHandle, DFTI_BACKWARD_SCALE, Scale);
status = DftiCommitDescriptor( pHandle);

Can I set all values, then commit, i.e.
- create
- set #1 (format)
- set #2 (backward scale)
- commit

- run loads of FFTs
- run loads of iFFTs

or do
- set (backward scale)
- commit
need to be done after the FFT, right before the iFFT?

Thanks for your help

Kind regards

Erich

status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FO

barragan_villanueva_ — Tue, 06 Jul 2010 13:29:39 GMT

Hi Erich,

The most correct way is

- create
- set all values you want (no commits are needed during settings)
- commit at last

and use FFTs or iFFTs.

But if you want to change some value again than commit will be required.

status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FO

Erich_Zwyssig — Tue, 06 Jul 2010 14:10:49 GMT

Hi Victor

Thanks for your reply.

It all seems to work fine now.

One final question. Would you expect to get the exact same values from mkl_crfft1d and DftiComputeBackward if Scale = 1.0/(double)n; is used?

Kind regards

Erich

status = DftiSetValue( pHandle, DFTI_PACKED_FORMAT, DFTI_CCS_FO

barragan_villanueva_ — Wed, 07 Jul 2010 10:09:50 GMT

Erich

All DFTI transforms are not normalized and therefore you can use/(or not) scaling if needed.
For example,for 1D with using scaling=1/n you will get approximately the same (not exacltly the same) results after forwardplus backward transforms.