CUB 2.1.0

[alliepiper]

Breaking Changes

• Deprecate CTK<9 #553: Deprecate the CUB_USE_COOPERATIVE_GROUPS macro, as all supported CTK distributions provide CG. This macro will be removed in a future version of CUB.

New Features

• Adds DeviceBatchMemcpy algorithm and tests #359: Add new DeviceBatchMemcpy algorithm.
• Add DeviceMergeSort::StableSortKeysCopy API #565: Add DeviceMergeSort::StableSortKeysCopy API. Thanks to David Wendt (@davidwendt) for this contribution.
• Sorting policy for H100 #585: Add SM90 tuning policy for DeviceRadixSort. Thanks to Andy Adinets (@canonizer) for this contribution.
• Add option to opt out CDP #586: Introduce a new mechanism to opt-out of compiling CDP support in CUB algorithms by defining CUB_DISABLE_CDP.
• Support reduction for more than 2^31 items #589: Support 64-bit indexing in DeviceReduce.
• Support __{u,}int128_t in radix sort #607: Support 128-bit integers in radix sort.

Bug Fixes

• Fix of CUB linkage #547: Resolve several long-running issues resulting from using multiple versions of CUB within the same process. Adds an inline namespace that encodes CUB version and targeted PTX architectures.
• Fix wrong tid when blockDim.y or z > 1 #562: Fix bug in BlockShuffle resulting from an invalid thread offset. Thanks to @sjfeng1999 for this contribution.
• Fix block radix rank for blocks with non multiple of 32 threads #564: Fix bug in BlockRadixRank when used with blocks that are not a multiple of 32 threads.
• Mask of all ones in SHFL_IDX_SYNC in block radix rank #579: Ensure that all threads in the logical warp participate in the index-shuffle for BlockRadixRank. Thanks to Andy Adinets (@canonizer) for this contribution.
• Fix reordering in CUB member initializer lists #582: Fix reordering in CUB member initializer lists.
• Support reduction for more than 2^31 items #589: Fix DeviceSegmentedSort when used with bool keys.
• Correct install(DIRECTORY ... PATTERN) which matches at end of file name #590: Fix CUB’s CMake install rules. Thanks to Robert Maynard (@robertmaynard) for this contribution.
• Fix overflow in reduce #592: Fix overflow in DeviceReduce.
• Fix RLE when items[0] is NaN #598: Fix DeviceRunLengthEncode when the first item is a NaN.
• Fix warp scan exlusive sum for vector types #611: Fix WarpScanExclusive for vector types.

Other Enhancements

• First version of developer overview #537: Add detailed and expanded version of a [CUB developer overview](https://github.com/NVIDIA/cub/blob/main/docs/developer_overview.md).
• Fix BlockReduceRaking non-commutative docs #549: Fix BlockReduceRaking docs for non-commutative operations. Thanks to Tobias Ribizel (@upsj) for this contribution.
• Optimize Decoupled Look-Back #606: Optimize CUB’s decoupled-lookback implementation.

---

This discussion was created from the release CUB 2.1.0.