[pooma-dev] Good News. Intel's ICC 8.0 Beta looks promising, now.

[Richard Guenther]

Hi Paul / others!

As you have access to KAI CC, I'd be curious to know what numbers you
get for the BlitzLoops / ABCTest benchmarks. Do we really expect the
POOMA-II numbers match the C / C restrict numbers? I know neither
ICC, nor gcc is really there at the moment. For the BlitzLoops with
my patched gcc3.3 I get (-O2 -march=athlon -fomit-frame-pointer
-funroll-loops)

rguenth at phoenix15 BlitzLoops > ./LINUXgcc33/Loop18 --sim-params 1000 3 1
--no-diags --samples 10
              C
N          restrict         C          CppTran       PoomaII
1000       637.74        599.74        441.03        416.92
10000      446.39        412.08        364.48        345.99
100000      94.34         83.25         82.50         81.25
1000000     91.77         74.49         74.39         76.55

which is good once we are memory bandwidth limited. With icc 8.0
(-O3 -xK -tpp6 -ip -restrict)

rguenth at phoenix15 BlitzLoops > ./LINUXICC/Loop18 --sim-params 1000 3 1
--no-diags --samples 10
              C
N          restrict         C          CppTran       PoomaII
1000       857.11        909.79        248.03        187.58
10000      449.74        434.15        230.60        181.60
100000      82.61         83.24         82.04         88.50
1000000     79.18         90.93         80.07         88.65

Again, this looks good for large datasets.

Just for the curious, if I mark both brick engine and brickview engines
data_m member restrict we get the same numbers for Intel icpc, gcc also
doesnt improve with this benchmark.

Richard.

--
Richard Guenther <richard dot guenther at uni-tuebingen dot de>
WWW: http://www.tat.physik.uni-tuebingen.de/~rguenth/