Re: vDSP Intel performance query (Dave Malham)
Re: vDSP Intel performance query (Dave Malham)
- Subject: Re: vDSP Intel performance query (Dave Malham)
- From: Ian Ollmann <email@hidden>
- Date: Tue, 27 May 2008 12:13:57 -0700
The current implementation might not perform well with small filters
on Intel. The code (last I looked at it) was designed to principally
run on problems where the filter is largish, maybe the size of the
signal. This is perhaps not a good fit for the common case. The best
bet is to run Shark on both, and file a performance bug with the Shark
trace attached.
Convolution in general is harder on Intel than it was on AltiVec
because Intel doesn't have a facility to handle misaligned data at
speed in the vector unit. Convolutions are inherently misaligned
problems, even if your data is aligned. For this reason, it seems
unlikely that Intel will ever reach the same performance as PowerPC,
clock for clock, at least not until AVX appears.
If it is a really huge convolution, then you'd be better off
reformulating it as a FFT, provided that you can put up with power of
two size limits.
Ian
_______________________________________________
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