Re: vDSP_conv very slow?
Re: vDSP_conv very slow?
- Subject: Re: vDSP_conv very slow?
- From: Frank Schoep <email@hidden>
- Date: Thu, 8 Feb 2007 10:07:02 +0100
On Feb 8, 2007, at 9:15 AM, Cor Jansen wrote:
...
I finally have a simple FIR filter working.
It uses the vDSP_conv function from the Aceleration framework.
But at a filter size of 2000 it already uses 50% of the cpu time.
I'm using a duo Intel core mac-Mini.
...
So I think the hardware acceleration is not working.
Anybody an idea why not? Is it not supported on a mac Intel?
Are you aligning your variables (vectors)? Check the bottom part of
my humble introduction to vDSP and vecLib for a small explanation:
http://www.ffnn.nl/pages/articles/apple-mac-os-x/vectorizing-with-
vdsp-and-veclib.php
I found a huge speed benefit in almost all vDSP functions when using
aligned variables since scalar code seems to be invoked when they are
not aligned. This especially benefits the FFT functions and possibly
convolution.
Even though there is an "auto-align" compiler option it doesn't seem
to work very well in most cases so manual tuning can give a huge
improvement.
You might also want to try and use Shark (in the CHUD tools) to
profile the code to find hot spots and hot loops. Copying your
samples from one location to another might be done using vDSP_mmov if
I'm not mistaking.
Sincerely,
Frank
_______________________________________________
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