Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.15.4

This ccl_2021.15.4-arc branch introduces several enhancements for Intel ARC A and B Series GPU:

• Support for Reduce-Scatter and Point-To-Point in addition to previously enabled Allreduce and Allgather
• Support for 8 bit datatypes (int8, uint8)
• Bug fixes, including removal of previously required setting of IGC_VISAOptions=-activeThreadsOnlyBarrier, which is no longer needed.

The cmake command is the same as before:

make .. -DCMAKE_INSTALL_PREFIX=_install -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DCOMPUTE_BACKEND=dpcpp -DCCL_ENABLE_ARCA=1
An example of the cmake command for Intel ARC B Series GPU :
cmake .. -DCMAKE_INSTALL_PREFIX=_install -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DCOMPUTE_BACKEND=dpcpp -DCCL_ENABLE_ARCB=1

Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.16

What's New 2021.16:

• Added SYCL graph support for Record and Replay for Allgather, Allreduce, Alltoall, ReduceScatter and Broadcast
• Added SYCL-based implementation of ring algorithm for Allgather
• Added SYCL-based implementation for Broadcast
• Added multithread support for Allgather and ReduceScatter scale up impementation
• Added attribute in the communicator to specify blocking operations for CPU
• Bug fixes

Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.15.3

This ccl_2021.15.3-arc branch adds support for Intel ARC A and B Series GPU and some bug fixes.

An example of the cmake command for Intel ARC A Series GPU :
cmake .. -DCMAKE_INSTALL_PREFIX=_install -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DCOMPUTE_BACKEND=dpcpp -DCCL_ENABLE_ARCA=1

An example of the cmake command for Intel ARC B Series GPU :
cmake .. -DCMAKE_INSTALL_PREFIX=_install -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DCOMPUTE_BACKEND=dpcpp -DCCL_ENABLE_ARCB=1

If the system does not have GPU Peer-to-Peer (P2P) support, you will need to add this compiler environment flag (export IGC_VISAOptions=-activeThreadsOnlyBarrier) before compiling. Similarly, on a system without P2P support, add export IGC_VISAOptions=-activeThreadsOnlyBarrier to your command line before running the application.

Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.15.2

What's new:

• Bug fix - Improvement of User Experience based on setting of Environment Variables.

Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.15.1

What's new:

• Bug fixes

Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.15

What's new:

• Support for Average for Allreduce and Reduce-Scatter
• Extend Group API to also support collective operations.
• New split_communicator API with updated parameters.
• Performance optimizations for scaleup for Alltoall

Removals:

• split_communicators is deprecated in 2021.15.0 and will be removed later

Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.14

What's New:

• Optimizations on key-value store support to scale up to 3000 nodes
• New APIs for Allgather, Broadcast and group API calls
• Performance Optimizations for scaleup for Allgather, Allreduce, and Reduce-scatter for scaleup and scaleout
• Performance Optimizations for CPU single node
• Optimizations to reuse Level Zero events.
• Change of the default mechanism for IPC exchange to pidfd

Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.13Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.13.1

What's new:

• Bug fixes

Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.13

What's New:

• Optimizations to limit the memory consumed by oneCCL
• Optimizations to limit the number of file descriptors maintained opened by oneCCL.
• Align the support for in-place for the Allgatherv and Reduce-scatter collectives to follow the same behavior as NCCL.
• In particular, the Allgatherv collective is in place when:
• send_buff == recv_buff + rank_offset, where rank_offset = sum (recv_counts[i]), for all I<rank.
• Reduce-scatter is in-place when recv_buff == send_buff + rank *recv_count.
• When using the environment variable CCL_WORKER_AFFINITY, oneCCL enforces the requirement that the length of the list should be equal to the number of workers.
• Bug fixes.

Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.12

What's New

• Performance improvements for scaleup for all message sizes for AllReduce, Allgather, and Reduce_Scatter.
• Optimizations also include small message sizes that appear in inference apps.
• Performance improvements for scaleout for Allreduce, Reduce, Allgather, and Reduce_Scatter.
• Optimized memory usage of oneCCL.
• Support for PMIx 4.2.6.
• Bug fixes.

Removals

• oneCCL 2021.12 removes support for PMIx 4.2.2