Hi,<div><br></div><div>Keep an eye on the release-4-6 git branch, the AVX 128 (+FMA4) and AVX 256 version of non-bonded kernels will very soon get merged upstream.</div><div><br></div><div>Cheers,<br clear="all">--<br>Szilįrd<br>
<br><br><div class="gmail_quote">On Wed, May 30, 2012 at 4:31 PM, Shun Sakuraba <span dir="ltr"><<a href="mailto:shun.sakuraba@gmail.com" target="_blank">shun.sakuraba@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Dear list,<br>
<br>
I would like to share my trial to port GROMACS 4.5 SSE/SSE2 nonbonded kernel<br>
to AMD's new family 15h chip, "Bulldozer" architecture, performed to measure<br>
the benefit of new instructions. In AMD family 15h new FMA4 instructions<br>
are added; FMA4 is the fused multiplication and addition (subtraction) operations.<br>
FMA4 reduces number of instructions and latency, giving a marginal performance boost.<br>
Also there are XOP instructions, which is useful implementing table interpolation<br>
used in GROMACS.<br>
<br>
As far as I could try, the speedup is only 5% for SP kernels and around 10% for<br>
DP kernels, on gmxbench ( <a href="http://www.gromacs.org/About_Gromacs/Benchmarks" target="_blank">http://www.gromacs.org/About_Gromacs/Benchmarks</a> ).<br>
The source codes and generated kernels are uploaded on<br>
( <a href="https://bitbucket.org/shun.sakuraba/gromacs-fma4-patch" target="_blank">https://bitbucket.org/shun.sakuraba/gromacs-fma4-patch</a> ),<br>
being available for download. I hope this interests some developers.<br>
<br>
gmxbench 3.0, ns/day<br>
vanilla fma4+xop<br>
SP dppc 2.13 2.21<br>
SP lzm.pme 16.15 16.73<br>
SP lzm.cut 23.93 24.79<br>
SP polych2 38.88 38.13<br>
SP villin 88.32 92.71<br>
DP dppc 1.35 1.52<br>
DP lzm.pme 10.28 11.14<br>
DP lzm.cut 15.34 17.09<br>
DP polych2 26.73 29.87<br>
DP villin 48.82 54.92<br>
<br>
Benchmarking results are taken on the following environment:<br>
AMD FX-4100 (2 modules, 4 cores, 3.6GHz / up to 3.9GHz with turbo)<br>
Linux (ArchLinux) 3.1.12-1-ARCH<br>
gcc (GCC) 4.7.0 20120505 (prerelease)<br>
FFTW 3.3.2<br>
GROMACS 4.5.5, used threading parallelization with 4 threads, compiled with<br>
CFLAGS="-O3 -fomit-frame-pointer -finline-functions -Wall -Wno-unused -march=native -funroll-all-loops -std=gnu99 -fexcess-precision=fast".<br>
All benchmarks are taken 3 times and the median is taken.<br>
<br>
Notes:<br>
* In GROMACS, if "-march=native" is replaced with "-msse2" (which is default in GROMACS 4.5.5),<br>
the program runs 5~10% slower than "-march=native".<br>
* I also compiled FFTW with and without -march=native. With default (i.e. without -march=native)<br>
SP/DP runs are ~3% slower in PME.<br>
* This version of FFTW does not use FMA4 SIMD instructions. FMA4 SIMD instruction in FFTW<br>
should increase performance, but I have not tried.<br>
* On NVE simulation, the energy conservation is slightly better than vanilla, possibly because<br>
in IEEE-compliant FMA, intermediate multiplication results are calculated with infinite precision.<br>
<span class="HOEnZb"><font color="#888888"><br>
--<br>
Shun SAKURABA, Ph.D.<br>
Postdoc @ Molecular Modeling & Simulation Group, Japan Atomic Energy Agency<br>
--<br>
gmx-developers mailing list<br>
<a href="mailto:gmx-developers@gromacs.org">gmx-developers@gromacs.org</a><br>
<a href="http://lists.gromacs.org/mailman/listinfo/gmx-developers" target="_blank">http://lists.gromacs.org/mailman/listinfo/gmx-developers</a><br>
Please don't post (un)subscribe requests to the list. Use the<br>
www interface or send it to <a href="mailto:gmx-developers-request@gromacs.org">gmx-developers-request@gromacs.org</a>.<br>
</font></span></blockquote></div><br></div>