Bug #2974
position-restraints regressiontest fails on intel GPU with OpenCL
Description
Tested on i7-8550U + UHD 620 and i7-8559U + Iris Plus 655. Only reproduces without DD and nonbonded offloaded. The forces seem to be just slightly off.
Attached log and checkforce output.
Related issues
History
#1 Updated by Szilárd Páll over 1 year ago
Target tentative, unless we have a quick solution bump to next release.
#2 Updated by Mark Abraham over 1 year ago
- Target version changed from 2019.3 to 2019.4
I reproduced on my i7-7500U (neo), but without leads, we'll have to bump this for now
#3 Updated by Szilárd Páll over 1 year ago
Mark Abraham wrote:
I reproduced on my i7-7500U (neo), but without leads, we'll have to bump this for now
Sure, I agree.
#4 Updated by Szilárd Páll over 1 year ago
Still failing on both 2019 and master. I have no leads, but it can't be a PME issue as this test does not use PME.
#5 Updated by Mark Abraham over 1 year ago
The logic about whether and how position restraints get run is pretty convoluted, so I could well believe that force contributions could be missed
#6 Updated by Berk Hess over 1 year ago
The differences are very small though, so I wouldn't exclude a too tight tolerance either.
#7 Updated by Paul Bauer over 1 year ago
- Target version changed from 2019.4 to 2019.5
no fix in sight, bumped
#8 Updated by Szilárd Páll about 1 year ago
Something changed that seems to make the same test on the same hardware on the current 2020 branch whereas on 2019 it still fails.
#9 Updated by Paul Bauer about 1 year ago
should I move this to 2020 instead? to indicate it is fixed there?
#10 Updated by Szilárd Páll about 1 year ago
It would be good to know what has changed, even if we don't fix it in 2019, so I'd prefer to not close this yet.
#11 Updated by Szilárd Páll about 1 year ago
- Subject changed from position-restraints regressiontest fails on intel GPU to position-restraints regressiontest fails on intel GPU with OpenCL
#12 Updated by Paul Bauer about 1 year ago
- Target version changed from 2019.5 to 2020
ok, changed to 2020 target
#13 Updated by Paul Bauer about 1 year ago
- Target version changed from 2020 to 2020.1
I would prefer closing this instead
#14 Updated by Paul Bauer 11 months ago
- Target version changed from 2020.1 to 2020.2
still not sure what happens here
#15 Updated by Szilárd Páll 11 months ago
- Related to Bug #3405: intermitted OpenCL regressiontest failures added