Subject: Re: [linux-audio-dev] double <--> float conversions speed benchmark and analysis
From: Benno Senoner (sbenno_AT_gardena.net)
Date: Tue Apr 04 2000 - 18:05:49 EEST
On Tue, 04 Apr 2000, David Benson wrote:
> how much data are you copying? is it being
> copied once or multiple times?
use the source , luke :-)
Anyway I am using input and output buffers
of 1024 elements , but always the same buffers, that
means we are timing CACHED float/double --> float/double assignments
>
> if it is just being copied once, you're probably seeing
> a lot of memory-cache misses. in real situations, the
> buffers tend to be small and reused to minimize latency,
> which has the right cache behavior.
exactly, that is why I decided to use pretty small buffers (1024 x
sizeof(double) = 8Kbytes, which fits nicely in the 2nd level cache)
>
> [am i correct in assuming the double-double/float-float copies
> were normalized to the same number of bytes?]
normalized to the same number of elements not bytes.
that means:
float->float 4 bytes in + 4 bytes out = 8bytes
float->double 4bytes in + 8 bytes out =12 bytes
double->float 8bytes in + 4 bytes out = 12 bytes
double->double 8bytes in + 8 bytes out = 16 bytes
but when doing DSP stuff are interested in the number of operations per second
(FLOPS) , not in the number of bytes moved.
( in the double->double case we achieve the biggest bytes/sec performance)
Benno.
This archive was generated by hypermail 2b28 : Tue Apr 04 2000 - 19:06:52 EEST