You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
fixed_width_convert_to_rows() seems to be producing bad results on Pascal GPUs. The errors manifest downstream as described in NVIDIA/spark-rapids/issues/4980.
Salient observations:
This only seems to happen when some of the transposed columns are of a narrow type than 32-bit. E.g. BOOL8, BYTE, SHORT. This might indicate a problem with how the alignment of type is calculated.
When the error occurs, groups/multiples of 32 rows seem to go wrong together. This might indicate a problem synchronizing on Pascal. A hasty/naive attempt at adding __syncthreads() at points of divergence in fixed_width_convert_to_rows() did not fix the problem.
convert_to_rows() does not seem to exhibit this behaviour.
A reduced C++ repro case has proved elusive. But it does seem reproducible from the NVIDIA/spark-rapids window-function tests.
The text was updated successfully, but these errors were encountered:
This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.
Having consulted the team, I think it's best we close this bug for now.
The CUDF JNI code has a workaround for this failure on Pascal. Let's revisit this if we see this on another arch.
fixed_width_convert_to_rows()
seems to be producing bad results on Pascal GPUs. The errors manifest downstream as described in NVIDIA/spark-rapids/issues/4980.Salient observations:
BOOL8
,BYTE
,SHORT
. This might indicate a problem with how the alignment of type is calculated.__syncthreads()
at points of divergence infixed_width_convert_to_rows()
did not fix the problem.convert_to_rows()
does not seem to exhibit this behaviour.A reduced C++ repro case has proved elusive. But it does seem reproducible from the NVIDIA/spark-rapids window-function tests.
The text was updated successfully, but these errors were encountered: