Subject: Re: [linux-audio-dev] exploring LADSPA
From: Steve Harris (S.W.Harris_AT_ecs.soton.ac.uk)
Date: Thu Aug 14 2003 - 11:46:08 EEST
On Wed, Aug 13, 2003 at 09:40:29 -0400, Pete Yadlowsky wrote:
> Yes. I believe double-precision is the standard data type used by most
> floating-pt processors. Single-precision floats must first be converted to
> doubles at each computation, thus actually degrading performance slightly.
This is actually not generally true. In the 387 they are all converted to
long doubles (80 or 96 bit, I forget which) and processing in SSE (and
3DNow and Altivec IIRC) is nativly 4x32 bits wide. SSE2 is nativly 4x64 bits.
You are correct that there is no /processor/ overhead to double v's floats
(in 387, its not true in all systems) the difference comes from memory and
cache effects - most DSP routines are memory bandwidth starved - its
actually quite hard to fill the FPU pipelines.
- Steve
This archive was generated by hypermail 2b28 : Thu Aug 14 2003 - 11:49:15 EEST