My code works without SIMD, but gives the wrong answer when enabled about pffft HOT 3 CLOSED

OK, I checked the sources and you are right, the ordering in the results of pffft[d]_transform of course depends on SIMD packet size. If performance is paramount for your use case, you could reverse engineer the code of pffft[d]_zreorder and write a function with the signature pffft[d]_extract(SETUP_STRUCT *setup, const float *in, int index, float *out, pffft_direction_t direction); which extracts the i'th value of pffft[d]_transform's results.

from pffft.

Yes, I can confirm your results on Apple M1. @hayguen can you see anything obviously wrong here?

Without SIMD:

% clang -Ilib pffft/pf*.c test.cpp -o test && ./test

result=3443.696594
Success.

With SIMD:

% clang -Ilib pffft/pf*.c test.cpp -o test -DPFFFT_ENABLE_NEON=1 && ./test

result=3793.724228
Success.

from pffft.

Got it!

I changed all three pffftd_transform to pffftd_transform_ordered. Now the results are exactly the same.

If I measure yin_pitch with gcc using the provided code without compiler optimizations (-Ox) I get:

SIMD disabled: ~3600 μs
SIMD enabled:  ~2400 μs

But with -O3 the results are roughly the same at around ~1000 μs. Anyways, it's all really fast and your library works great.

Thanks for taking the time to verify the issue.

from pffft.