two benchmarks.

[IainS]

A follow-up to http://gcc.gnu.org/ml/fortran/2008-06/msg00163.html.

Anyone else remember BYTE's "Lies, Damn Lies and Benchmarks"     :-) ?

===
I'm interested in possible explanations for what we see here.
===

Two systems, same gcc tree, both bootstrapped.

Intel Core Duo (NOT Core 2).  I'm reasonably sure it's a T2400  
(although I've not taken the lid off to prove this).
1.83 GHz

970mp at 2.5GHz

----

I don't have access to spec - but looking at published data (had to  
find specCPU2000 - for these):

Estimates based on published figures for the same chipsets:

T2400 @ 1.83G   -        1564/1379
970MP @ 2.5G   -         1587/2119

so we'd expect similar INT performance and for the 970 to be 1.5x on FP.
This stacks up with common sense - looking at the architectures and  
bearing in mind that the 970MP is a 64 bit machine.
----

I've run the p05 benchmarks in a variety of configurations:
**** of course, these are not official and were run against an  
experimental gcc

----------------------
I focussed on optimization *without* chip-specific tags.

Observations (P05):

1/ There's a slight disappointment in that a couple of failures come  
up for PPC

2/ The reported geometric mean runtime is spookily within a fraction  
of a percent the same ....

3/ there are huge disparities (both ways) between some tests --
e.g
           ac      2.36       10000     23.12      14  0.0856 (T2400)
           ac      1.46       10000     81.19      10  0.0325  (970MP)

       induct      6.77       10000     85.66      10  0.0498 (T2400)
       induct      7.77       10000     49.73      10  0.0359 (970MP)

4/ there is very little difference in runtime between -O3 and
	-ffast-math -funroll-loops -ftree-loop-linear -ftree-vectorize -O3 - 
march=native -mtune=native -msse3
	on Core duo.

	-- suggesting that the limitation is not from the compiler....

5/ -m64 actually causes a degradation in performance in some areas on  
the 970MP

=====

I've also run DP fortran and c whetstones.

now, I've a feeling that someone said that these tend to test libm  
more than the compiler.

but, the sad news is that rather than 1.5x FP for the 970MP we get  
something like 1.2x for the Code Duo.

dunno if anyone can postulate where the differences come from?

Notes:

I suppressed any processor affinity issues on the Core Duo (by  
running the benchmarks with one core off).
I wasn't able to do that yet with the 970MP - have to be an overnight  
run for that.

ttfn,
Iain


-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: bench-t2400.txt
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20080617/d0ffc093/attachment.txt>
-------------- next part --------------

-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: bench-970mp.txt
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20080617/d0ffc093/attachment-0001.txt>
-------------- next part --------------