bypasses are cool...ev4->ev45 FP speedups

Neal. 01-Sep-1995 0622 +0100 (crook@rdgeng.enet.dec.com)
Fri, 1 Sep 95 07:48:16 MET DST


The differences between the FP units ev4->ev45 *only* affect the floating
point divider.

The new (ev45) divider is *generally* faster, but the latency is
now variable, depending upon the data. The WORST CASE latencies
will be the same, but 'most' divides will be a lot faster, and
'some' will be faster still.

ev4 ev45
single-precision divide with data dependency 34 ??,19,34
double-precision divide with data dependency 63 ??,29,63

(x,y,z) triplet means (min,typical,max) - after all, I *am* a hardware
engineer..

The 'typical' figure seems to tie-in with what DavidMT is seeing.

Now, regarding the bypassing - I find that surprising, but I'm going to
have to go back to the book to check out your theory.

>>If anybody has any comments or if somebody could add some EV5
>>measurements, I'd greatly appreciate it (just holler if you want the
>>source code of the test program).

Well, I tried to persuade my boss that one of the new systems would boost
my productivity, but so success, yet.

Neal.