Debian Bug report logs - #983942
gromacs: ftbfs with -march=x86-64-v3

Reply or subscribe to this bug.

Display info messages

Message #5 received at [email protected] (full text, mbox, reply):

From: Matthias Klose <[email protected]>

To: [email protected]

Subject: gromacs: ftbfs with -march=x86-64-v3

Date: Wed, 03 Mar 2021 15:39:09 +0000

Package: src:gromacs
Version: 2020.5-4
Severity: minor
User: [email protected]
Usertags: ftbfs-x86-64-v3

Starting with LLVM 12 and GCC 11 (and backported to gcc-10 for Debian), 
GCC introduces optimizations for new micro-architecture levels from
the x86-64 psABI.  Filing these bug reports for packages building with
the current level (x86-64), but not with x86-64-v3.

To reproduce these issues, append -march=x86-64-v3 to the compiler
flags (CFLAGS, CXXFLAGS, ...).  You also need hardware supporting these.
The easiest way is to call the dynamic linker in glibc 2.33 like
  /lib/x86_64-linux-gnu/ld-linux-x86-64.so.2 --help
Otherwise look at the flags in /proc/cpuinfo and compare these with
those from table 3.1 in the System V Application Binary Interface doc:
https://gitlab.com/x86-psABIs/x86-64-ABI/-/wikis/x86-64-psABI

Please don't upload a package setting these flags by default.  The
x86-64 baseline is still unchanged.

The full build log can be found at:
http://people.debian.org/~doko/logs/20210228/filtered/gcc10-v3/gromacs_2020.5-4_unstable_gcc10-v3.log
The last lines of the build log are at the end of this report.

[...]
54:   the time step of 2.0e-03 ps.
54:   Maybe you forgot to change the constraints mdp option.
54: 
54: Number of degrees of freedom in T-Coupling group System is 18.00
54: 
54: NOTE 2 [file /<<PKGBUILDDIR>>/build/basic/src/api/cpp/workflow/tests/Testing/Temporary/GmxApiTest_BuildApiWorkflowImpl_input.mdp]:
54:   You are using a plain Coulomb cut-off, which might produce artifacts.
54:   You might want to consider using PME electrostatics.
54: 
54: 
54: 
54: There were 2 notes
54: Determining Verlet buffer for a tolerance of 0.005 kJ/mol/ps at 298 K
54: Calculated rlist for 1x1 atom pair-list as 1.041 nm, buffer size 0.041 nm
54: Set rlist, assuming 4x4 atom pair-list, to 1.041 nm, buffer size 0.041 nm
54: Note that mdrun will redetermine rlist based on the actual pair-list setup
54: This run will generate roughly 0 Mb of data
54: [       OK ] GmxApiTest.BuildApiWorkflowImpl (228 ms)
54: [ RUN      ] GmxApiTest.CreateApiWorkflow
54: Setting the LD random seed to -76234755
54: Generated 331705 of the 331705 non-bonded parameter combinations
54: Generating 1-4 interactions: fudge = 0.5
54: Generated 331705 of the 331705 1-4 parameter combinations
54: Excluding 2 bonded neighbours molecule type 'SOL'
54: Excluding 3 bonded neighbours molecule type 'methane'
54: 
54: NOTE 1 [file spc_and_methane.top, line 33]:
54:   The bond in molecule-type methane between atoms 1 C and 2 H1 has an
54:   estimated oscillational period of 1.1e-02 ps, which is less than 10 times
54:   the time step of 2.0e-03 ps.
54:   Maybe you forgot to change the constraints mdp option.
54: 
54: Number of degrees of freedom in T-Coupling group System is 18.00
54: 
54: NOTE 2 [file /<<PKGBUILDDIR>>/build/basic/src/api/cpp/workflow/tests/Testing/Temporary/GmxApiTest_CreateApiWorkflow_input.mdp]:
54:   You are using a plain Coulomb cut-off, which might produce artifacts.
54:   You might want to consider using PME electrostatics.
54: 
54: 
54: 
54: There were 2 notes
54: Determining Verlet buffer for a tolerance of 0.005 kJ/mol/ps at 298 K
54: Calculated rlist for 1x1 atom pair-list as 1.041 nm, buffer size 0.041 nm
54: Set rlist, assuming 4x4 atom pair-list, to 1.041 nm, buffer size 0.041 nm
54: Note that mdrun will redetermine rlist based on the actual pair-list setup
54: This run will generate roughly 0 Mb of data
54: [       OK ] GmxApiTest.CreateApiWorkflow (215 ms)
54: [----------] 2 tests from GmxApiTest (443 ms total)
54: 
54: [----------] Global test environment tear-down
54: [==========] 2 tests from 1 test case ran. (443 ms total)
54: [  PASSED  ] 2 tests.
54/54 Test #54: GmxapiInternalsMpiTests .............   Passed    0.45 sec

98% tests passed, 1 tests failed out of 54

Label Time Summary:
GTest              = 124.16 sec*proc (54 tests)
IntegrationTest    =  94.42 sec*proc (12 tests)
MpiTest            = 109.82 sec*proc (8 tests)
SlowTest           =  24.25 sec*proc (2 tests)
UnitTest           =   5.49 sec*proc (40 tests)

Total Test time (real) = 124.19 sec

The following tests FAILED:
	 27 - SimdUnitTests (Failed)
Errors while running CTest
make: *** [debian/rules:129: build-basic] Error 1
dpkg-buildpackage: error: debian/rules binary subprocess returned exit status 2

Message #10 received at [email protected] (full text, mbox, reply):

From: Nicholas Breen <[email protected]>

To: Matthias Klose <[email protected]>, [email protected]

Subject: Re: [Debichem-devel] Bug#983942: gromacs: ftbfs with -march=x86-64-v3

Date: Sat, 27 Nov 2021 17:00:39 -0800

On Wed, Mar 03, 2021 at 03:39:09PM +0000, Matthias Klose wrote:
> Package: src:gromacs
> Version: 2020.5-4
> Severity: minor
> User: [email protected]
> Usertags: ftbfs-x86-64-v3
> 
> Starting with LLVM 12 and GCC 11 (and backported to gcc-10 for Debian), 
> GCC introduces optimizations for new micro-architecture levels from
> the x86-64 psABI.  Filing these bug reports for packages building with
> the current level (x86-64), but not with x86-64-v3.

Checking against the newest versions, this is still present -- but it's
also avoidable; the failure only occurs because the build is elsewhere
normally forced to nothing beyond SSE4.1 extensions, as part of the
*upstream* build system, for maximum compatibility.  Building with
-march=x86-64-v3 and with autodetected best-available SIMD flags
(DEB_BUILD_OPTIONS=cpuopt) passes all tests.

If there's a future way to detect an x86-64-v3 build, I think the
sensible fix here is to raise the baseline SIMD choice to match.



-- 
Nicholas Breen
[email protected]

Send a report that this bug log contains spam.