This is an archive of the discontinued LLVM Phabricator instance.

[AArch64] Reduce vector insert/extract cost for Kryo
ClosedPublic

Authored by mssimpso on Feb 18 2016, 7:28 AM.

Details

Summary

Kryo can perform vector inserts and extracts very efficiently. This patch lowers the default insert/extract cost for Kryo from 3 to 2. Some performance results are given below (positive difference are performance improvements).

Benchmark              Diff (%) 
-------------------    --------
spec2000/eon               1.74
spec2000/gcc               1.17
spec2000/mesa              0.97
spec2000/vortex            0.26
spec2006/dealII            0.76
spec2006/gcc               0.07
spec2006/h264ref           1.63
spec2006/libquantum        2.88
spec2006/milc              0.25
spec2006/namd             -0.17
spec2006/soplex            1.43

Diff Detail

Repository
rL LLVM

Event Timeline

mssimpso updated this revision to Diff 48312.Feb 18 2016, 7:28 AM
mssimpso retitled this revision from to [AArch64] Reduce vector insert/extract cost for Kryo.
mssimpso updated this object.
mcrosier accepted this revision.Feb 18 2016, 7:53 AM
mcrosier edited edge metadata.

LGTM. Thanks, Matt.

This revision is now accepted and ready to land.Feb 18 2016, 7:53 AM
This revision was automatically updated to reflect the committed changes.