<div dir="ltr"><br><div class="gmail_extra"><br><div class="gmail_quote">On Fri, Jun 26, 2015 at 2:20 PM, Tobias Grosser <span dir="ltr"><<a href="mailto:tobias@grosser.es" target="_blank">tobias@grosser.es</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div class="HOEnZb"><div class="h5">On 06/26/2015 08:55 AM, Tobias Grosser wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
On 06/25/2015 10:47 PM, Michael Kruse wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Author: meinersbur<br>
Date: Thu Jun 25 15:47:35 2015<br>
New Revision: 240689<br>
<br>
URL: <a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__llvm.org_viewvc_llvm-2Dproject-3Frev-3D240689-26view-3Drev&d=AwMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=mQ4LZ2PUj9hpadE3cDHZnIdEwhEBrbAstXeMaFoB9tg&m=CO8z2UvKDEAiXAUANT5RTyg36LOAUQTVsRSbzyXMB5g&s=s4A5fIFC8VT2LOGUxyTbuzuDcCEWfR0PS8CyMHuu8Ys&e=" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project?rev=240689&view=rev</a><br>
Log:<br>
Enable ISL's small integer optimization<br>
<br>
Summary:<br>
With small integer optimization (short: sio) enabled, ISL uses 32 bit<br>
integers for its arithmetic and only falls back to a big integer library<br>
(in the case of Polly: IMath) if an operation's result is too large.<br>
This gives a massive performance boost for most application using ISL.<br>
For instance, experiments with ppcg (polyhedral source-to-source<br>
compiler) show speed-ups of 5.8 (compared to plain IMath), respectively<br>
2.7 (compared to GMP).<br>
<br>
In Polly, a smaller fraction of the total compile time is taken by ISL,<br>
but the speed-ups are still very significant. The buildbots measure<br>
compilation speed-up up to 1.8 (oourafft, floyd-warshall, symm). All<br>
Polybench benchmarks compile in at least 9% less time, and about 20%<br>
less on average.<br>
<br>
Detailed Polybench compile time results (median of 10):<br>
correlation -25.51%<br>
covariance -24.82%<br>
2mm -26.64%<br>
3mm -28.69%<br>
atax -13.70%<br>
bicg -10.78%<br>
cholesky -40.67%<br>
doitgen -11.60%<br>
gemm -11.54%<br>
gemver -10.63%<br>
gesummv -11.54%<br>
mvt -9.43%<br>
symm -41.25%<br>
syr2k -14.71%<br>
syrk -14.52%<br>
trisolv -17.65%<br>
trmm -9.78%<br>
durbin -19.32%<br>
dynprog -9.09%<br>
gramschmidt -15.38%<br>
lu -21.77%<br>
floyd-warshall -42.71%<br>
reg_detect -41.17%<br>
adi -36.69%<br>
fdtd-2d -32.61%<br>
fdtd-apml -21.90%<br>
jacobi-1d-imper -9.41%<br>
jacobi-2d-imper -27.65%<br>
seidel-2d -31.00%<br>
</blockquote>
<br>
<br>
Very nice. Here the corresponding performance bot results:<br>
<br>
"clang -O3 -mllvm -polly" BEFORE<br>
vs.<br>
"clang -O3 -mllvm -polly" AFTER<br>
<br>
<a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__llvm.org_perf_db-5Fdefault_v4_nts_27877-3Fcompare-5Fto-3D27869&d=AwMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=mQ4LZ2PUj9hpadE3cDHZnIdEwhEBrbAstXeMaFoB9tg&m=CO8z2UvKDEAiXAUANT5RTyg36LOAUQTVsRSbzyXMB5g&s=y1FU-6AINFxeNjlU1AYnjPxsx-Az7SpnAQo7S0SnLCc&e=" rel="noreferrer" target="_blank">http://llvm.org/perf/db_default/v4/nts/27877?compare_to=27869</a><br>
<br>
"clang -O3"<br>
vs.<br>
"clang -O3 -mllvm -polly" BEFORE<br>
<br>
<a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__llvm.org_perf_db-5Fdefault_v4_nts_27869-3Fcompare-5Fto-3D27876&d=AwMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=mQ4LZ2PUj9hpadE3cDHZnIdEwhEBrbAstXeMaFoB9tg&m=CO8z2UvKDEAiXAUANT5RTyg36LOAUQTVsRSbzyXMB5g&s=rn_a9F1dkQIXNjHgkg0roAhIW_6vmcv1i7MJXJqyUgs&e=" rel="noreferrer" target="_blank">http://llvm.org/perf/db_default/v4/nts/27869?compare_to=27876</a><br>
<br>
"clang -O3"<br>
vs.<br>
"clang -O3 -mllvm -polly" AFTER<br>
<br>
<a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__llvm.org_perf_db-5Fdefault_v4_nts_27877-3Fcompare-5Fto-3D27876&d=AwMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=mQ4LZ2PUj9hpadE3cDHZnIdEwhEBrbAstXeMaFoB9tg&m=CO8z2UvKDEAiXAUANT5RTyg36LOAUQTVsRSbzyXMB5g&s=0jO3XZ15Zt6nVCpvSizjWuS5MbR9KKWvyoEdnn38mTI&e=" rel="noreferrer" target="_blank">http://llvm.org/perf/db_default/v4/nts/27877?compare_to=27876</a><br>
<br>
<br></blockquote></div></div></blockquote><div><br></div><div><br></div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div class="HOEnZb"><div class="h5"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Instead of a 450% slowdown compared to LLVM in terms of compile time,<br>
the largest slowdown is now 200% (0.13 s to 0.44s) and most kernels show<br>
even less slowdown. A larger part of the remaining slowdown (about half<br>
of it) is increased LLVM codegen time due to code versioning.<br>
200% slowdown may sound a lot, but these are indeed the loop kernels we<br>
optimize, meaning their compile-time overall is commonly less than a<br>
second. Compile time impact on entire applications is commonly a lot less.<br></blockquote></div></div></blockquote><div><br></div><div>Great indeed. I guess we can target for a max of 100% slowdown next?</div><div> </div><div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div class="HOEnZb"><div class="h5"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
There is still some headroom for further optimizations, which we are<br>
aiming for. However, for now, thanks again to Michael Kruse for this<br>
great patch and to Pratik Bhatu for the preparations and preliminary<br>
studies that allowed us to understand the performance benefits such an<br>
optimization give.<br>
</blockquote>
<br></div></div>
For the record, besides Pratik's nice prepartions, Michael had already worked on a similar optimization for the gmp backend in the context of his Molly project and as a result also obtained valuable insights (and<br>
experience) that have proven very useful when doing this work.</blockquote><div><br></div><div>I totally agree. Having Michael's prior work and experience was indeed the major factor! </div><div>Merci beaucoup/Danke Michael!</div><div><br></div><div>Best Regards</div><div>Ramakrishna</div></div></div></div>