Subscribe: scitech Mailing List
http://rss.lists.apple.com/scitech.rss
Preview: scitech Mailing List

scitech Mailing List



scitech Mailing List



Published: 2016-02-29T23:35:02-08:00

 



Re: [apple scitech] Large-scale FFT's and convolutions on Apple hardware
From: Adriaan van Os
Reply to list

[...]

I have to correct myself with regard to the stride after some more tests. A vDSP_fft2d_zrop or
vDSP_fft2d_zrip with a stride is up to ten times slower than calling it without the stride (where
it is 0 or 1). In other words, the stride parameter is useless. [...]



Re: [apple scitech] Large-scale FFT's and convolutions on Apple hardware
From: Adriaan van Os
Reply to list

[...]

This brings me to the following questions. What makes a processor run hot ? Standard instructions
compared to vector code ? Memory stalls versus no memory stalls ? And what is the best way to cool
it down with software (probably user configurable) ? [...]



Re: [apple scitech] Large-scale FFT's and convolutions on Apple hardware
From: Adriaan van Os
Reply to list

[...]

Incidentally, I compared MatrixFFT with vDSP this weekend. An FFT of a 4096 x 2048 pixel planar
with mfftExecute was ten times (!) faster than vDSP_fft2d_zrop or vDSP_fft2d_zrip. An FFT of a 512
x 512 pixel planar with mfftExecute was two to threes times faster than with vDSP. [...]



Re: [apple scitech] Large-scale FFT's and convolutions on Apple hardware
From: Kiran Pradeep
Reply to list
[...]

For some one who might take the same route, I have a starter sample
which tries to adhere to steps/optimizations mentioned in above email.
Sample at https://github.com/kiranpradeep/vDSPxcorr2D/blob/master/2DCorrelation/xcorr2D.c#L13