GPU result has number of walkers dependency and doesn't match CPU result

[yaoyi92]

Details in the post here.
https://groups.google.com/forum/#!topic/qmcpack/eOE1eIXAgaE

Not 100% sure whether it is a problem for my build. ctest results here in the post
https://groups.google.com/forum/#!topic/qmcpack/1QzNQoceNHs

[yaoyi92]

This seems to be a Volta V100 specific issue. I don't have this problem compiling and running the same simulation on GTX1080.

Need help from someone with experience of Volta V100.

[jtkrogel]

Something does seem to be up with Volta:
https://cdash.qmcpack.org/CDash/index.php?project=QMCPACK&date=2018-09-07.

Compare "Volta-GCC-CUDA-Release" with "GCC-CUDA-Release".

[prckent]

GCC-CUDA-Release runs on a Kepler. This is printed on one on the ~third line of output.

Updated:

Something happened between 7-8 August
Working: https://cdash.qmcpack.org/CDash/testDetails.php?test=2851998&build=26579
Broken: https://cdash.qmcpack.org/CDash/testDetails.php?test=2858752&build=26628
Possibly this was an update to the script or CUDA version, will investigate

[prckent]

To keep this updated: the current belief is that this has hit some kind of edge case (bug) or is a newly surfaced problem on Volta or with recent CUDAs (bug). For whatever reason, our current tests don't show the same problem. @PDoakORNL will try to chase down.

[ye-luo]

Some recent fixes in legacy CUDA probably addressed this issue. Open a new issue if needed.