CUB 1.17.0

[alliepiper]

CUB 1.17.0

Summary

CUB 1.17.0 is the final minor release of the 1.X series. It provides a variety of bug fixes and miscellaneous enhancements, detailed below.

Known Issues

“Run-to-run” Determinism Broken

Several CUB device algorithms are documented to provide deterministic results (per device) for non-associative reduction operators (e.g. floating-point addition). Unfortunately, the implementations of these algorithms contain performance optimizations that violate this guarantee. The DeviceReduce::ReduceByKey and DeviceScan algorithms are known to be affected. We’re currently evaluating the scope and impact of correcting this in a future CUB release. See NVIDIA/cub#471 for details.

Bug Fixes

• Fix select if for mixed types #444: Fixed DeviceSelect to work with discard iterators and mixed input/output types.
• Handle when CMAKE_INSTALL_LIBDIR has nested directories #452: Fixed install issue when CMAKE_INSTALL_LIBDIR contained nested directories. Thanks to @robertmaynard for this contribution.
• Fix segmented sort for SM{61,70} #462: Fixed bug that produced incorrect results from DeviceSegmentedSort on sm_61 and sm_70.
• Normalize flags in select flagged #464: Fixed DeviceSelect::Flagged so that flags are normalized to 0 or 1.
• Fixed computation of offsets in sorting #468: Fixed overflow issues in DeviceRadixSort given num_items close to 2^32. Thanks to @canonizer for this contribution.

Other Enhancements

• Eliminate device synchronization in CDP version of segmented sort #445: Remove device-sync in DeviceSegmentedSort when launched via CDP.
• update contributing.md #449: Fixed invalid link in documentation. Thanks to @kshitij12345 for this contribution.
• update Iterate (BlockDiscontinuity) to use pragma unroll #450: BlockDiscontinuity: Replaced recursive-template loop unrolling with #pragma unroll. Thanks to @kshitij12345 for this contribution.
• Alias deprecated TexRefInputIterator to TexObjInputIterator. #451: Replaced the deprecated TexRefInputIterator implementation with an alias to TexObjInputIterator. This fully removes all usages of the deprecated CUDA texture reference APIs from CUB.
• pragma unroll in block_adjacent_difference::Iterate #456: BlockAdjacentDifference: Replaced recursive-template loop unrolling with #pragma unroll. Thanks to @kshitij12345 for this contribution.
• Alternative approach to 64-bit indexing in adjacent difference #466: cub::DeviceAdjacentDifference API has been updated to use the new OffsetT deduction approach described in Transparent support for 64-bit indexing in device algorithms #212.
• fix doxygen warnings #470: Fix several doxygen-related warnings. Thanks to @karthikeyann for this contribution.

---

This discussion was created from the release CUB 1.17.0.