Hi All, I prototyped a piece of code using FFTW and it works great and I'm now trying to port it to the iPhone. The iPhone has it's own DSP library written by Apple which they say is 5x faster than the best competitor. However, when I pass the same data (exactly) into their routine I get a different result set (though interestingly if I graph the data they look similar). My question is... should all FFT implementations return the same results not withstanding a little rounding? Or can the results vary so dramatically? The values coming back from the Apple library are about 2x those of FFTW. Thanks Ray
FFT Different APIs Returning Different Results
Started by ●November 15, 2010
Reply by ●November 15, 20102010-11-15
On Nov 15, 8:22�pm, Raeldor <rael...@gmail.com> wrote:> Hi All, > > I prototyped a piece of code using FFTW and it works great and I'm now > trying to port it to the iPhone. �The iPhone has it's own DSP library > written by Apple which they say is 5x faster than the best > competitor. �However, when I pass the same data (exactly) into their > routine I get a different result set (though interestingly if I graph > the data they look similar). > > My question is... should all FFT implementations return the same > results not withstanding a little rounding? �Or can the results vary > so dramatically? �The values coming back from the Apple library are > about 2x those of FFTW. > > Thanks > RayCheck the documentation. Different FFTs use different scalings. John
Reply by ●November 16, 20102010-11-16
On Nov 15, 5:40�pm, John <sampson...@gmail.com> wrote:> On Nov 15, 8:22�pm, Raeldor <rael...@gmail.com> wrote: > > > > > > > > > > > Hi All, > > > I prototyped a piece of code using FFTW and it works great and I'm now > > trying to port it to the iPhone. �The iPhone has it's own DSP library > > written by Apple which they say is 5x faster than the best > > competitor. �However, when I pass the same data (exactly) into their > > routine I get a different result set (though interestingly if I graph > > the data they look similar). > > > My question is... should all FFT implementations return the same > > results not withstanding a little rounding? �Or can the results vary > > so dramatically? �The values coming back from the Apple library are > > about 2x those of FFTW. > > > Thanks > > Ray > > Check the documentation. Different FFTs use different scalings. > > JohnYou were absolutely right. It was scaling. The results from analyzing the frequencies come out to be the same. Odd I expected this to be based on a standard formula so the result would be the same.
Reply by ●November 17, 20102010-11-17
On Nov 16, 3:57�pm, Raeldor <rael...@gmail.com> wrote:> On Nov 15, 5:40�pm, John <sampson...@gmail.com> wrote: > > > > > On Nov 15, 8:22�pm, Raeldor <rael...@gmail.com> wrote: > > > > Hi All, > > > > I prototyped a piece of code using FFTW and it works great and I'm now > > > trying to port it to the iPhone. �The iPhone has it's own DSP library > > > written by Apple which they say is 5x faster than the best > > > competitor. �However, when I pass the same data (exactly) into their > > > routine I get a different result set (though interestingly if I graph > > > the data they look similar). > > > > My question is... should all FFT implementations return the same > > > results not withstanding a little rounding? �Or can the results vary > > > so dramatically? �The values coming back from the Apple library are > > > about 2x those of FFTW. > > > > Thanks > > > Ray > > > Check the documentation. Different FFTs use different scalings. > > > John > > You were absolutely right. �It was scaling. �The results from > analyzing the frequencies come out to be the same. �Odd I expected > this to be based on a standard formula so the result would be the same.There is no standard formula for an FFT. It is all smoke and mirrors. Chris ======================= Chris Bore BORES Signal Processing www.bores.com
Reply by ●November 17, 20102010-11-17
On Nov 17, 4:00=A0am, Chris Bore <chris.b...@gmail.com> wrote:> On Nov 16, 3:57=A0pm, Raeldor <rael...@gmail.com> wrote: > > > > > > > > > > > On Nov 15, 5:40=A0pm, John <sampson...@gmail.com> wrote: > > > > On Nov 15, 8:22=A0pm, Raeldor <rael...@gmail.com> wrote: > > > > > Hi All, > > > > > I prototyped a piece of code using FFTW and it works great and I'm =now> > > > trying to port it to the iPhone. =A0The iPhone has it's own DSP lib=rary> > > > written by Apple which they say is 5x faster than the best > > > > competitor. =A0However, when I pass the same data (exactly) into th=eir> > > > routine I get a different result set (though interestingly if I gra=ph> > > > the data they look similar). > > > > > My question is... should all FFT implementations return the same > > > > results not withstanding a little rounding? =A0Or can the results v=ary> > > > so dramatically? =A0The values coming back from the Apple library a=re> > > > about 2x those of FFTW. > > > > > Thanks > > > > Ray > > > > Check the documentation. Different FFTs use different scalings. > > > > John > > > You were absolutely right. =A0It was scaling. =A0The results from > > analyzing the frequencies come out to be the same. =A0Odd I expected > > this to be based on a standard formula so the result would be the same. > > There is no standard formula for an FFT. It is all smoke and mirrors. > > Chris > =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D > Chris Bore > BORES Signal Processingwww.bores.comOut of curiosity, did you benchmark FFTW against Apple's DSP library? I'm curious as to whether the claim of "5x faster" is actually legit.
