Re: vDSP: performance costs of stride size != 1
Re: vDSP: performance costs of stride size != 1
- Subject: Re: vDSP: performance costs of stride size != 1
- From: Eric Allamanche <email@hidden>
- Date: Fri, 17 Feb 2012 09:42:56 -0800
On Feb 16, 2012, at 12:11 PM, Morgan Packard wrote:
> Hello list,
> I understand it's preferable to keep stride size to 1 when using the vDSP functions of the accelerate framework. Does anyone have any experience (or numbers) relating to performance and stride size of greater than one?
In general, strides arguments unequal to 1 will cause the vDSP functions to take the slow scalar path instead of the vectorized one. There are a few exceptions however, for stride values of -1. This is useful for applying windowing functions in reverse sequence for example.
> I have a single-channel envelope signal that I'd like to apply to an interleaved stereo signal. If I can get much better performance by using a two-channel envelope signal (this keeping the stride to 1), I'll consider reworking my code.
the most efficient way to cope with this is to deinterleave your input signal using vDS_ctoz(), applying your function to each channel and re-interleaving again using vDSP_ztoc().
Eric
_______________________________________________
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