Skip to content

v2.4.0

Latest
Compare
Choose a tag to compare
@wmaxey wmaxey released this 23 Apr 21:30
· 183 commits to main since this release
1c009d2

What’s New

We are still hard at work in CCCL on paying down lots of technical debt, improving infrastructure, and various other simplifications as part of the unification of Thrust/CUB/libcu++. In addition to various fixes and documentation improvements, the following notable improvements have been made to Thrust, CUB, and libcudacxx.

Thrust

As part of our kernel consolidation effort, kernels of thrust::unique_by_key, thrust::copy_if, and thrust::partition algorithms are now consolidated in CUB. Kernel consolidation achieves two goals. First, it delivers the latest optimizations of CUB algorithms to Thrust users. Apart from the performance improvements, it introduces support of large problem sizes (64-bit offsets) into Thrust algorithms.

CUB

  • cub::DeviceSelect::UniqueByKey now supports equality operator and large problem sizes.
  • New cub::DeviceFor family of algorithms goes beyond conventional cub::DeviceFor::ForEach. cub::DeviceFor::ForEachCopy can provide you with additional performance benefits from vectorized memory accesses.
  • Many CUB algorithms now support CUDA graph capture mode.

libcudacxx

  • Added new cuda::ptx namespace with wrappers for inline-PTX instructions
  • cuda::std::complex specializations for CUDA types bfloat and half.

What's Changed

New Contributors

Full Changelog: v2.3.2...v2.4.0