Without the extra barrier sync towards end of barrierOR function after reading the reduction result, the caller's loop if any is going into infinite loop occasionally which is in turn randoms hangs. This doesn't seem to be an issue on non-nvidia hardware. Hence, we are conditionally adding the extra barrier sync conditionally for nvidia platform.
Fix opencl barrierOR fn in confidence connected opencl kernel
Fix barrierOR fn in confidence connected opencl kernel
9prady9 added a commit that referenced this pull request
Aug 17, 2020
* Fix barrierOR fn in confidence connected opencl kernel Without the extra barrier sync towards end of barrierOR function after reading the reduction result, the caller's loop if any is going into infinite loop occasionally which is in turn randoms hangs. This doesn't seem to be an issue on non-nvidia hardware. Hence, we are conditionally adding the extra barrier sync conditionally for nvidia platform. * Add the hardware check comparison (cherry picked from commit 64855cb )
umar456 pushed a commit that referenced this pull request
Sep 9, 2020
* Fix barrierOR fn in confidence connected opencl kernel Without the extra barrier sync towards end of barrierOR function after reading the reduction result, the caller's loop if any is going into infinite loop occasionally which is in turn randoms hangs. This doesn't seem to be an issue on non-nvidia hardware. Hence, we are conditionally adding the extra barrier sync conditionally for nvidia platform. * Add the hardware check comparison (cherry picked from commit 64855cb )