Project

General

Profile

Bug #1719

severe performance regression on Piledriver

Added by Szilárd Páll almost 2 years ago. Updated almost 2 years ago.

Status:
Closed
Priority:
Normal
Assignee:
-
Category:
core library
Target version:
-
Affected version - extra info:
5.1-beta1
Affected version:
Difficulty:
uncategorized
Close

Description

I measure 20-30% performance regression in the non-bonded kernels on AMD Piledriver. The bonded kernels seem to have regressed too, but the effect here is only 4-8%.

The former has barely changed in 5.1. The bonded performance has improved in 5.1, but the source of the original issue may have simply been masked by unrelated improvements (in reduction and RB dihedrals).

test_gmx467_1x16_cpu.log View (20 KB) Szilárd Páll, 04/28/2015 03:06 AM

test_gmx50_1x16_cpu.log View (21.7 KB) Szilárd Páll, 04/28/2015 03:06 AM

test_gmx51_1x16_cpu_bdver2.log View (21 KB) Szilárd Páll, 04/28/2015 03:06 AM

topol.tpr - repro input (rnase dodec vsites) (897 KB) Szilárd Páll, 04/28/2015 03:06 AM

Associated revisions

Revision ce772452 (diff)
Added by Berk Hess almost 2 years ago

Re-enabled FMA for AVX-128-FMA

All FMA instructions for AVX-128-FMA were not renamed during a SIMD
macro name change, which effectively removed FMA support for AMD.
Fixes #1719.

Change-Id: I855a379d29742d37c586794a191606a360a89c5f

History

#1 Updated by Berk Hess almost 2 years ago

  • Status changed from New to In Progress

For a water box with PME I get 88 sec in nbnxn kernel with 4.6 and 140 sec with 5.1!
The only difference I found in the log files is that -Wno-strict-overflow -fomit-frame-pointer have been removed. Adding those gives 102 sec, so an enormous difference, but still 15% slower than 4.6.

#2 Updated by Gerrit Code Review Bot almost 2 years ago

Gerrit received a related patchset '1' for Issue #1719.
Uploader: Berk Hess ()
Change-Id: I855a379d29742d37c586794a191606a360a89c5f
Gerrit URL: https://gerrit.gromacs.org/4533

#3 Updated by Mark Abraham almost 2 years ago

  • Status changed from In Progress to Fix uploaded

Did that fix resolve the full issue?

#4 Updated by Szilárd Páll almost 2 years ago

Mark Abraham wrote:

Did that fix resolve the full issue?

Yes, it has (that's what I meant by verify +1).

#5 Updated by Mark Abraham almost 2 years ago

  • Status changed from Fix uploaded to Closed

ok thanks

Also available in: Atom PDF