Re: vDSP Intel performance query
Re: vDSP Intel performance query
- Subject: Re: vDSP Intel performance query
- From: George Taylor <email@hidden>
- Date: Tue, 27 May 2008 15:19:42 -0400
Hi Dave,
If you're using the vDSP_conv function, there is a bug where
vectorized code is not called unless all strides are equal to 1. This
is contrary to Apple's documentation, where they tell you to use a
filter stride of -1. In fact, on Intel Macs, you need to reorganize
things so that your filter stride is +1.
Here's an archived message on the performance list that goes into
more detail:
http://lists.apple.com/archives/coreaudio-api/2007/Feb/msg00071.html
Also I might add that you'll find convolution performance to
generally be slower on Intel Macs than PPC Macs given the same clock
frequency, but once you set filter stride to 1 and move the filter
pointer to the front of the filter buffer, you should see a jump in
performance.
Hope that helps,
George
On May 27, 2008, at 5:25 AM, Dave Malham wrote:
Hi,
Just been playing with the Accelerate framework and, in
particular, vDSP with a view to speeding up some of my Ambisonic
stuff. I have compiled the vDSP examples project and have gotten
results which I find a little odd and would appreciate some
comments/help from you gurus.
On my G4 1.33GHz iBook, I'm typically getting, for the first
2048x256 convolution in the example project, anything between 4 and
8 GFlops performance. However, on our quad core 2.66GHz Mac's in
our studios I'm only getting around 0.8 GFlops - and on my
colleague's Intel power book, this goes down to around 0.3 GFlops.
Both my iBook and the studio machines are running OSX 10.4.11, my
colleague's machine is running Leopard so it's not OS differences.
I've tried compiling natively (ie, not UB) on all machines (no
alterations to the XCode project as supplied) and then UB on all
machines, tried playing with optimisations, etc., etc., with little
effect - am I doing something wrong, or misinterpreting the
results, or is this an expecting result (it certainly wasn't
expected by me...)
Regards
Dave
--
These are my own views and may or may not be shared by my employer
/
*********************************************************************/
/* Dave Malham http://music.york.ac.uk/staff/research/
dave_malham/ */
/* Music Research Centre */
/* Department of Music "http://music.york.ac.uk/"; */
/* The University of York Phone 01904
432448 */
/* Heslington Fax 01904
432450 */
/* York YO10
5DD */
/* UK 'Ambisonics - Component Imaging for
Audio' */
/* "http://www.york.ac.uk/inst/mustech/
3d_audio/" */
/
*********************************************************************/
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Coreaudio-api mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Coreaudio-api mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden