rocPRIM icon indicating copy to clipboard operation
rocPRIM copied to clipboard

Published 20 hours ago • verified publisher icon ROCm

Fix bug in device partition for large problem sizes

Open nolmoonen opened this issue 2 years ago • 2 comments

Previously, a check was performed for the last block of each launch which should have been performed only at the last block across all launches. This gave unintended effects, so the redundant checks are removed. This change has no measurable effects on performance. The bug could be reproduced with sizes 30064767271 and 17179866528 in the LargeInputPartition and LargeInputPartitionThreeWay tests.

nolmoonen avatar Oct 06 '22 14:10 nolmoonen

@stanleytsang-amd I can reproduce the bug reliably on gfx906 and gfx908 when letting get_large_sizes (not get_sizes) return size 30064767271 in test/rocprim/test_device_partition.cpp.

nolmoonen avatar Oct 18 '22 09:10 nolmoonen

Rebased the PR.

nolmoonen avatar Oct 20 '22 08:10 nolmoonen