[AArch64] build cmssw aarch64 with -mno-outline-atomics file #7582

smuzaffar · 2022-01-24T14:51:41Z

cmsbuild · 2022-01-24T14:52:03Z

A new Pull Request was created by @smuzaffar (Malik Shahzad Muzaffar) for branch IB/CMSSW_12_3_X/master.

@cmsbuild, @smuzaffar, @iarspider can you please review it and eventually sign? Thanks.
@perrotta, @dpiparo, @qliphy you are the release manager for this.
cms-bot commands are listed here

smuzaffar · 2022-01-24T14:56:13Z

test parameters:

full_cmssw = true
enable_test = threading
workflow_threading = 4.44,13.0,11634.24,140.56

smuzaffar · 2022-01-24T14:56:29Z

please test for slc7_aarch64_gcc11

cmsbuild · 2022-01-24T20:01:59Z

-1

Failed Tests: Build
Summary: https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-3a3bc3/21960/summary.html
COMMIT: 4b4fa81
CMSSW: CMSSW_12_3_X_2022-01-23-2300/slc7_aarch64_gcc11
Additional Tests: THREADING
User test area: For local testing, you can use /cvmfs/cms-ci.cern.ch/week1/cms-sw/cmsdist/7582/21960/install.sh to create a dev area with all the needed externals and cmssw changes.

The following merge commits were also included on top of IB + this PR after doing git cms-merge-topic:

You can see more details here:
https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-3a3bc3/21960/git-recent-commits.json
https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-3a3bc3/21960/git-merge-result

Build

I found compilation error when building:

Entering library rule at src/FWCore/Integration/test
>> Compiling edm plugin /data/cmsbuild/jenkins_b/workspace/ib-run-pr-tests/CMSSW_12_3_X_2022-01-23-2300/src/FWCore/Integration/test/TestHistoryKeeping.cc
>> Building  edm plugin tmp/slc7_aarch64_gcc11/src/FWCore/Integration/test/TestHistoryKeeping/libTestHistoryKeeping.so
Leaving library rule at src/FWCore/Integration/test
/bin/sh: line 1: 2956474 Illegal instruction     (core dumped) /usr/bin/touch /data/cmsbuild/jenkins_b/workspace/ib-run-pr-tests/CMSSW_12_3_X_2022-01-23-2300/tmp/slc7_aarch64_gcc11/cache/edm_edmPluginRefresh
gmake: *** [lib/slc7_aarch64_gcc11/TestHistoryKeeping.edmplugin] Error 132
Entering library rule at src/FWCore/Integration/test
>> Compiling edm plugin /data/cmsbuild/jenkins_b/workspace/ib-run-pr-tests/CMSSW_12_3_X_2022-01-23-2300/src/FWCore/Integration/test/TestInterProcessProd.cc
>> Building  edm plugin tmp/slc7_aarch64_gcc11/src/FWCore/Integration/test/TestInterProcessProd/libTestInterProcessProd.so
Leaving library rule at src/FWCore/Integration/test
@@@@ Running edmWriteConfigs for TestInterProcessProd

fwyzard · 2022-01-24T20:22:42Z

Building with -mno-outline-atomics will require hardware support for the LSE instructions from the ARMv8.1 ISA.
If our build machines support only the ARMv8.0 ISA, the build is expected to fail at run time :-/

dan131riley · 2022-01-28T14:03:02Z

Building with -mno-outline-atomics will require hardware support for the LSE instructions from the ARMv8.1 ISA.

I thought the point of the outline atomics was to enable use of LSE at runtime, so I expected -mno-outline-atomics to revert to the pre-10.1 behavior that didn't use LSE instructions.

If our build machines support only the ARMv8.0 ISA, the build is expected to fail at run time :-/

Our build machines (at least the ones I'm aware of) have been upgraded to ThunderX2 that support the ARMv8.1 ISA, and since gcc 10.1 have been using the LSE instruction set at runtime. I was wondering if the use of LSE was causing the mysterious failures we've been seeing, so I was hoping that flag would disable LSE.

dan131riley · 2022-01-28T14:07:26Z

/bin/sh: line 1: 2956474 Illegal instruction (core dumped) /usr/bin/touch /data/cmsbuild/jenkins_b/workspace/ib-run-pr-tests/CMSSW_12_3_X_2022-01-23-2300/tmp/slc7_aarch64_gcc11/cache/edm_edmPluginRefresh

This is really weird--/usr/bin/touch crashed? I have no idea how adding that flag could cause that to crash.

fwyzard · 2022-01-28T14:12:14Z

Our build machines (at least the ones I'm aware of) have been upgraded to ThunderX2 that support the ARMv8.1 ISA

Ah, OK.
Do we have interactive access to any of those nodes ?

smuzaffar · 2022-01-28T14:20:37Z

@fwyzard , you can login to those node. I am going to send you the instructions via email

fwyzard · 2022-01-28T14:20:37Z

Building with -mno-outline-atomics will require hardware support for the LSE instructions from the ARMv8.1 ISA.

I thought the point of the outline atomics was to enable use of LSE at runtime, so I expected -mno-outline-atomics to revert to the pre-10.1 behavior that didn't use LSE instructions.

OK, I guess I misread the GCC's documentation:

-moutline-atomics
-mno-outline-atomics

Enable or disable calls to out-of-line helpers to implement atomic operations. These helpers will, at runtime, determine if the LSE instructions from ARMv8.1-A can be used; if not, they will use the load/store-exclusive instructions that are present in the base ARMv8.0 ISA.

This option is only applicable when compiling for the base ARMv8.0 instruction set. If using a later revision, e.g. -march=armv8.1-a or -march=armv8-a+lse, the ARMv8.1-Atomics instructions will be used directly. The same applies when using -mcpu= when the selected cpu supports the lse feature. This option is on by default.

So

with -march=armv8.1-a or later, -moutline-atomics/-mno-outline-atomics has no effect, and LSE are always used
with -march=armv8-a, -moutline-atomics uses helpers to check if LSE are available and use them, or fall back to the old instructions otherwise
with -march=armv8-a, -mno-outline-atomics always uses the old instructions

?

smuzaffar · 2022-01-28T14:25:02Z

please test for slc7_aarch64_gcc11

lets re-run the tests

dan131riley · 2022-01-28T15:10:40Z

with -march=armv8.1-a or later, -moutline-atomics/-mno-outline-atomics has no effect, and LSE are always used

with -march=armv8-a, -moutline-atomics uses helpers to check if LSE are available and use them, or fall back to the old instructions otherwise

with -march=armv8-a, -mno-outline-atomics always uses the old instructions

I think that's correct. We aren't specifying -march or -mcpu so I'd expect it to use the base ARMv8 instructions with -mno-outline-atomics (in fact, I got started on this line of investigation due to a handful crashes in the outline LSE routines).

smuzaffar · 2022-01-28T16:13:49Z

correct, we are not setting -mcpu/arch for aarch64. We have two arm64 nodes

thunderx2: olarm-202
thunderx: olarm-102

but we are currently using olarm-202 only. Let me build all externals and cmssw with -mcpu=armv8-a on olarm-202 and see if that helps.

cmsbuild · 2022-01-28T21:09:32Z

-1

Failed Tests: UnitTests RelVals RelVals-THREADING
Summary: https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-3a3bc3/22060/summary.html
COMMIT: 4b4fa81
CMSSW: CMSSW_12_3_X_2022-01-27-2300/slc7_aarch64_gcc11
Additional Tests: THREADING
User test area: For local testing, you can use /cvmfs/cms-ci.cern.ch/week1/cms-sw/cmsdist/7582/22060/install.sh to create a dev area with all the needed externals and cmssw changes.

The following merge commits were also included on top of IB + this PR after doing git cms-merge-topic:

You can see more details here:
https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-3a3bc3/22060/git-recent-commits.json
https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-3a3bc3/22060/git-merge-result

Unit Tests

I found errors in the following unit tests:

---> test TestFWCoreServicesDriver had ERRORS
---> test testFWCoreUtilities had ERRORS

RelVals

4.53

----- Begin Fatal Exception 28-Jan-2022 21:02:45 CET-----------------------
An exception of category 'Vertex' occurred while
   [0] Processing  Event run: 194533 lumi: 329 event: 462355458 stream: 0
   [1] Running path 'dqmofflineOnPAT_1_step'
   [2] Prefetching for module SingleTopTChannelLeptonDQM_miniAOD/'singleTopElectronMediumDQM_miniAOD'
   [3] Prefetching for module PATMuonSlimmer/'slimmedMuons'
   [4] Prefetching for module PATMuonSelector/'selectedPatMuons'
   [5] Prefetching for module PATMuonProducer/'patMuons'
   [6] Prefetching for module MuonProducer/'muons'
   [7] Prefetching for module PFProducer/'particleFlowTmp'
   [8] Prefetching for module PFBlockProducer/'particleFlowBlock'
   [9] Prefetching for module PFElecTkProducer/'pfTrackElec'
   [10] Prefetching for module PFConversionProducer/'pfConversions'
   [11] Calling method for module ConversionProducer/'allConversions'
Exception Message:
Refitted track not found in list
----- End Fatal Exception -------------------------------------------------

140.56

----- Begin Fatal Exception 28-Jan-2022 21:11:38 CET-----------------------
An exception of category 'Vertex' occurred while
   [0] Processing  Event run: 326479 lumi: 7 event: 1579493 stream: 0
   [1] Running path 'dqmoffline_8_step'
   [2] Prefetching for module SMPDQM/'SMPDQM'
   [3] Prefetching for module MuonProducer/'muons'
   [4] Prefetching for module PFProducer/'particleFlowTmp'
   [5] Prefetching for module PFBlockProducer/'particleFlowBlock'
   [6] Prefetching for module PFElecTkProducer/'pfTrackElec'
   [7] Prefetching for module PFConversionProducer/'pfConversions'
   [8] Calling method for module ConversionProducer/'allConversions'
Exception Message:
Refitted track not found in list
----- End Fatal Exception -------------------------------------------------

136.874

----- Begin Fatal Exception 28-Jan-2022 21:25:32 CET-----------------------
An exception of category 'Vertex' occurred while
   [0] Processing  Event run: 319450 lumi: 76 event: 106007323 stream: 0
   [1] Running path 'dqmoffline_10_step'
   [2] Prefetching for module SMPDQM/'SMPDQM'
   [3] Prefetching for module MuonProducer/'muons'
   [4] Prefetching for module PFProducer/'particleFlowTmp'
   [5] Prefetching for module PFBlockProducer/'particleFlowBlock'
   [6] Prefetching for module PFElecTkProducer/'pfTrackElec'
   [7] Prefetching for module PFConversionProducer/'pfConversions'
   [8] Calling method for module ConversionProducer/'allConversions'
Exception Message:
Refitted track not found in list
----- End Fatal Exception -------------------------------------------------

RelVals-THREADING

13.013.0_QCD_Pt_3000_3500+QCD_Pt_3000_3500INPUT+DIGI+RECO+HARVEST/step3_QCD_Pt_3000_3500+QCD_Pt_3000_3500INPUT+DIGI+RECO+HARVEST.log
4.444.44_RunElectron2012A+RunElectron2012A+HLTD+RECODR1reHLT+HARVESTDR1reHLT/step3_RunElectron2012A+RunElectron2012A+HLTD+RECODR1reHLT+HARVESTDR1reHLT.log
4.534.53_RunPhoton2012B+RunPhoton2012B+HLTD+RECODR1reHLT+HARVESTDR1reHLT/step3_RunPhoton2012B+RunPhoton2012B+HLTD+RECODR1reHLT+HARVESTDR1reHLT.log

Expand to see more relval errors ...

dan131riley · 2022-01-29T15:42:50Z

Refitted track not found in list
----- End Fatal Exception -------------------------------------------------

Well, that kills that theory. Not only does the error reproduce with LSE disabled, but it also shows that the error reproduces in a single-threaded test (which I had not noticed before), which rules out synchronization primitives.

smuzaffar · 2022-01-30T21:39:59Z

@dan131riley , yes looks like outline-atomics is not the issue. I have built gcc with outline atomics disabled ( #7593 ) and still the same error ( #7593 (comment) )

Conflicts: gcc.spec

cmsbuild · 2022-02-03T07:18:41Z

Pull request #7582 was updated.

cmsbuild added externals-pending orp-pending pending-signatures tests-pending labels Jan 24, 2022

cmsbuild added tests-started and removed tests-pending labels Jan 24, 2022

smuzaffar mentioned this pull request Jan 24, 2022

Test aarch64 build with -mno-outline-atomics cms-sw/cmssw#36788

Closed

cmsbuild added tests-rejected and removed tests-started labels Jan 24, 2022

cmsbuild added tests-started and removed tests-rejected labels Jan 28, 2022

cmsbuild added tests-rejected and removed tests-started labels Jan 28, 2022

build aarch64 with -mcpu=armv8-a

9054cf4

smuzaffar and others added 6 commits February 3, 2022 08:13

Update compilation_flags.file

e698645

Update compilation_flags.file

1949e4b

disable gcc outline atomics by default; grpc, root.boot to use armv8a

e3ba7c0

Conflicts: gcc.spec

fix boost flags

b20e693

typo fixed for compilation_flags

74416a9

fix gcc flags for boost

95f309c

smuzaffar force-pushed the smuzaffar-patch-5 branch from 4b4fa81 to 95f309c Compare February 3, 2022 07:18

cmsbuild added tests-pending and removed tests-rejected labels Feb 3, 2022

smuzaffar closed this Feb 3, 2022

smuzaffar deleted the smuzaffar-patch-5 branch February 3, 2022 07:19

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[AArch64] build cmssw aarch64 with -mno-outline-atomics file #7582

[AArch64] build cmssw aarch64 with -mno-outline-atomics file #7582

smuzaffar commented Jan 24, 2022

cmsbuild commented Jan 24, 2022

smuzaffar commented Jan 24, 2022

smuzaffar commented Jan 24, 2022

cmsbuild commented Jan 24, 2022

fwyzard commented Jan 24, 2022

dan131riley commented Jan 28, 2022

dan131riley commented Jan 28, 2022

fwyzard commented Jan 28, 2022

smuzaffar commented Jan 28, 2022

fwyzard commented Jan 28, 2022

smuzaffar commented Jan 28, 2022

dan131riley commented Jan 28, 2022

smuzaffar commented Jan 28, 2022

cmsbuild commented Jan 28, 2022

dan131riley commented Jan 29, 2022

smuzaffar commented Jan 30, 2022 •

edited

Loading

cmsbuild commented Feb 3, 2022

[AArch64] build cmssw aarch64 with -mno-outline-atomics file #7582

[AArch64] build cmssw aarch64 with -mno-outline-atomics file #7582

Conversation

smuzaffar commented Jan 24, 2022

cmsbuild commented Jan 24, 2022

smuzaffar commented Jan 24, 2022

smuzaffar commented Jan 24, 2022

cmsbuild commented Jan 24, 2022

Build

fwyzard commented Jan 24, 2022

dan131riley commented Jan 28, 2022

dan131riley commented Jan 28, 2022

fwyzard commented Jan 28, 2022

smuzaffar commented Jan 28, 2022

fwyzard commented Jan 28, 2022

smuzaffar commented Jan 28, 2022

dan131riley commented Jan 28, 2022

smuzaffar commented Jan 28, 2022

cmsbuild commented Jan 28, 2022

Unit Tests

RelVals

RelVals-THREADING

dan131riley commented Jan 29, 2022

smuzaffar commented Jan 30, 2022 • edited Loading

cmsbuild commented Feb 3, 2022

smuzaffar commented Jan 30, 2022 •

edited

Loading