Wait, what's the context here - which Gambit primitives/features are implemented in terms of these floating point logics you talk about here?

Thanks!


2018-04-17 6:51 GMT+08:00 Bradley Lucier <lucier@math.purdue.edu>:
On 04/16/2018 09:28 AM, Marc Feeley wrote:
>On Apr 13, 2018, at 3:15 PM, Bradley Lucier <lucier@math.purdue.edu> wrote > Bottom line: this page
>
> https://asteroidsathome.net/boinc/cpu_list.php
>
> claims that this cpu gets 4.42 Whetstone GFLOPS per core, and we get a maximum of 5.24 GFLOPS on this fft code.  Not bad.

The code miscalculated the number of operations, the most we get with
strict aliasing and allowing fused-multiply-add is 4.62 GFLOPS:

heine:~/Downloads> gsc -cc-options "-fstrict-aliasing -O3 -save-temps"
-exe fft
heine:~/Downloads> ./fft
<stuff removed>
n = 17; time = .0048208906249999905; GFLOPS = 4.622020853252617; error =
1.9179232434271595e-8

and not allowing fused multiply add gives 4.18 GFLOPS:

heine:~/Downloads> gsc -cc-options "-fstrict-aliasing -O3 -save-temps
-mno-fma" -exe fft
heine:~/Downloads> ./fft
<stuff removed>
n = 17; time = .00532633984375; GFLOPS = 4.183408804856173; error =
1.895030528746498e-8

I'll look into adding something to the Gambit benchmark suite.

Brad
_______________________________________________
Gambit-list mailing list
Gambit-list@iro.umontreal.ca
https://webmail.iro.umontreal.ca/mailman/listinfo/gambit-list