At 18:22 +0200 5/26/04, Andrea Riciputi wrote:
>Thanks for all your answers. I've read the Apple's web pages and documentation regarding the topic. Now I have more questions for you:
>
>1) My data are double precision, does this prevent from Altivec accelleration?
Short answer. Yes. Altivec only works on single precision numbers.
Long answer. No. The Accelerate framework has been tuned to optimize double precision calculations as well. So the Accelerate framework is more than Altivec. It is tuned also to perform well and take advantage of G3/G4/G5 specific features.
For very specific tuning Atlas might be an improvement since it auto tunes itself to different cache sizes which I don't believe that the Accellerate framework does.
>2) My 2-d data are real and even both respect to x and y axes. vDSP library doesn't have an implementation of discrete cosine/sine trasform (that should enhance performances by a factor 4). Any good reason for this?
Only do an FFT on one quadrant (ie 1/4 of your data). Then phase shift the output to compensate for the offset of the origin.
>3) Any hope to see convolution functions for 2d data appearing in vDSP soon?
Above my pay grade... :-)
>4) Is there any efficient way to multiply two matrix element wise (I mean not rows x columns product)? I've not found anything like this in vDSP.h
There is a multiply two vectors (not dot product). Do this on each row of the matrix.
-Scott
_______________________________________________
scitech mailing list | email@hidden
Help/Unsubscribe/Archives: http://www.lists.apple.com/mailman/listinfo/scitech
Do not post admin requests to the list. They will be ignored.
