Speeding up IPP FFT

[fbasile]

I'm benchmarking different implementations of 1D-FFT, and I've written a simple code for testing IPP:

#include
#include
#include

int main() {
for (unsigned int myOrder = 10; myOrder < 21; myOrder++) {
const unsigned int myLength = (1 << myOrder);
std::cout << "Length = " << myLength << std::endl;

IppsFFTSpec_C_32fc* mySpec;
ippsFFTInitAlloc_C_32fc(&mySpec, myOrder, IPP_FFT_NODIV_BY_ANY, ippAlgHintFast);

int myBufferSize = 0;
ippsFFTGetBufSize_C_32fc(mySpec, &myBufferSize);
Ipp8u *myBuffer = ippsMalloc_8u(myBufferSize);

Ipp32fc *myA = ippsMalloc_32fc(myLength);
Ipp32fc *myB = ippsMalloc_32fc(myLength);
for(unsigned int n = 0; n < myLength; ++n) {
myA.re = rand()/RAND_MAX;
myA.im = rand()/RAND_MAX;
}

struct timeval myStartingTime;
struct timeval myEndingTime;
gettimeofday(&myStartingTime, 0);

for (unsigned int myRepetition = 0; myRepetition < 1000; ++myRepetition) {
ippsFFTFwd_CToC_32fc(myA, myB, mySpec, myBuffer);
}

gettimeofday(&myEndingTime, 0);
double myElapsedSeconds = myEndingTime.tv_sec - myStartingTime.tv_sec +
(myEndingTime.tv_usec - myStartingTime.tv_usec) / 1000000.0;
std::cout << "\t\testimated Mflops: "
<< 2.5 * myOrder * myLength / myElapsedSeconds / 1000
<< std::endl;

ippsFree(myA);
ippsFree(myB);
ippsFree(myBuffer);
ippsFFTFree_C_32fc(mySpec);
}
}

And I compile it with icc.

It seems to be drammatically slower then FFTW3 implementation, BUT I'm not happy with this result: I've found benchmarks (http://www.fftw.org/speed/CoreDuo-3.0GHz-icc64/) showing performances to be similar.

I'm afraid (pretty sure) I'm not fully exploiting IPP. How should I speed up this code?

Thanks
francesco

[fbasile]

...my mistake, I was using different normalizations, the performances are comparable now.

Apart from this mistake, any advice or trick speeding up IPP-FFT is very wellcome!

Francesco