Repository navigation
What's New oneCCL 2022.2
- Automatic PCIe switch discovery and topology-aware sub-communicators (node/NUMA/switch/none) with the new CCL_SYCL_P2P_LEVEL control
- oneCCL now selects OFI/libfabric by default instead of IMPI
- Device API for windows registration - LSA
- Native recording enhancements for scaleup collectives
- New onecclScatter and onecclAllToAllV collectives
- Reduce collective with Low-Latency SYCL implementation
- OFI direct algorithms for alltoall, broadcast, scatter and reduce scaleout collectives
- One-shot algorithms for reduce-scatter, allgather and allreduce collectives
- Broadcast scaleup collective with simple protocol for large messages
- Broadcast scaleout collective using the scatter-allgather algorithm
- Support unaligned buffers for broadcast Low-Latency ring implementation
- Simple protocol enhancement to handle unaligned counts correctly
- Support AVG reduction in reduce
- Support custom pre-op for reduce
- Proxy thread framework for scaleout communication of reduce-scatter, allgather, alltoall and allreduce collectives
- Added simple protocol implementation for send/recv
- Use GPU RDMA directly for point-to-point communication when supported, eliminating the host buffer dependency.
- Extend CCL_SYCL_ALLREDUCE_LL to support per-size-range configuration syntax (:-;...) instead of a single algorithm name
- Added onecclCommSplit support for MT (multi-threaded) mode
- Enabled MT mode in the oneCCL benchmark
- Fixed group point-to-point operations for MT mode
- Implemented thread-safe concurrent access to shared resources in MT communicators