v0.3.1

v0.3.1 - RDMA Refactor and IBVerbs Support

• Add ibverbs RDMA implementation (#11)
• Refactor rdma memory, channel, and buffer abstractions
• Fix fi::Buffer incorrect channels data structure

v0.3.0

v0.3.0 - CMake Refactor and Test Improvements

• Consolidate all CMake find_* calls into cmake/packages.cmake
• Add CPU affinity setup to io_test and queue_test (matching Peer class pattern)
• Update benchmark results section in README

v0.2.10

v0.2.10 - Queue Implementation Benchmarks

• Add GdrQueue and PinnedQueue implementations for GPU-CPU communication
• Add queue type comparison benchmark (cudaMallocManaged, pinned host, GDRCopy)
• Bump EFA installer, NCCL, and NCCL tests versions
• Fix incorrect unit and round-robin benchmark results

v0.2.9

v0.2.9 - IPC Benchmark Improvements

• Add GPU-side verification for IPC data transfers (faster than host-side)
• Refactor IPC functors with configurable grid/block dimensions

v0.2.8

v0.2.8 - CUDA IPC for Intra-node Communication

• Add IPC experiment demonstrating GPU-to-GPU communication via CUDA IPC
• Benchmark NVLink bandwidth scaling with different thread parallelism configurations

v0.2.7

v0.2.7 - CUDA IPC for Intra-node Communication

• Add CUDA IPC support for intra-node GPU-to-GPU communication
• Automatic transport selection: IPC for local ranks, RDMA for remote ranks
• Add OpenIPCHandles() API in SymmetricMemory for IPC handle management
• Extend DeviceContext with ipc_ptrs array for GPU kernel access
• Add GetLocalComm() accessor in MPI wrapper for local node operations

v0.2.6

v0.2.6 - NVSHMEM-like shmem_* API

• Add template shmem_p<T>() / shmem_p_nbi<T>() device put operations
• Add type-specific wrappers: shmem_int_p, shmem_float_p, shmem_double_p, etc.

v0.2.5

v0.2.5 - NVSHMEM-like shmem_* API

• Add NVSHMEM-compatible shmem_* API over EFA/libfabric
• Add shmem_init() / shmem_finalize() for MPI-based bootstrap
• Add shmem_malloc() / shmem_free() for symmetric memory allocation
• Add shmem_int_p() / shmem_int_p_nbi() device put operations
• Add shmem_fence() / shmem_quiet() memory ordering primitives
• Add SHMEM experiment mirroring NVSHMEM docs example
• Note: Current implementation supports inter-node only (1 process per node)

v0.2.4

v0.2.4 - Round-robin channel selection for parallel writes

• Add round-robin write algorithm using per-channel distribution
• Add RoundRobinDMA test to write and all2all experiments
• Add test name display to progress bar (e.g., [RoundRobinDMA] [10.6s] ops=...)
• Add configurable progress bytes/bandwidth parameters to Bench()
• Fix progress bar bandwidth percentage calculation
• Rename gin.cuh to proxy.cuh

v0.2.3

v0.2.3 - GPU-initiated non-blocking interface support

• Add MPSC queue benchmark
• Add non-blocking interface (nbi) for GPU-initiated RDMA via proxy