feat: introduce core infrastructure, OpenMPI backend, and AllReduce example

[Ziminli]

Summary

This PR introduces the initial core infrastructure of InfiniCCL since the initial commit.

It establishes the foundational architecture for device and backend abstraction, compile-time dispatching, runtime support, and a few communication primitives. In addition, it provides an OpenMPI-based implementation of AllReduce along with a complete example program, including simple profiling and validation utilities.

Changes

Core Infrastructure

• Introduce core abstractions:
  • Device, BackendType, and runtime specializations (CPU, NVIDIA, MetaX)
  • Operation base class and dispatching mechanism
  • Compile-time traits and dispatcher utilities
• Add communication-related structures:
  • Communicator and backend-specific instances (e.g., OpenMPI)
• Implement data type system:
  • DataType, TypeMap, and device-aware mappings (fp16, bf16)

OpenMPI Backend Support

• Add OpenMPI-based implementations for:
  • infiniInit()
  • infiniGetRank()
  • infiniGetSize()
  • infiniCommInitAll()
  • infiniCommDestroy()
  • infiniAllReduce()
  • infiniFinalize()
• Introduce OpenMPI-specific mappings and runtime integration

AllReduce Example and Utilities

• Add examples/all_reduce.cc demonstrating distributed AllReduce
• Introduce shared utilities under examples/utils.h:
  • Timer, Metrics, and Validator for profiling and correctness checking
• Enable runtime-based memory management and validation

Build System

• Introduce CMake-based build system
• Implement code generation for include/infiniccl.h and bridge files
• Refine header visibility:
  • expose only include/ as PUBLIC interface
  • keep internal sources PRIVATE
• Link example programs with internal library

Logging

• Add a lightweight logging system:
  • Logger and LOG macro for unified message handling
• Integrate logging into key components

Refactor & Style

• Improve code structure and modularity across components
• Keep consistent formatting, comments, and general style

Known Issues & Future Work

• Additional devices and backends will be supported in the future;
• A unified casting utility will be required and supported in future updates;
• Configure threading model (MPI_THREAD_FUNNELED) in Init to avoid runtime issues since using MPI_THREAD_MULTIPLE hangs;
• The current examples/all_reduce.cc defaults to the Sum reduction operation. However, Prod and Avg have also been tested and verified. Support for configurable reduction types may be added in future updates.

The following have been addressed in #1

• The current AllReduce example has only been tested on a single node (NVIDIA environment). Multi-node and heterogeneous setups still require a complete validation.
• The logging module is temporary and will be replaced with glog in the future.
• In the current test environment, OpenMPI consistently reports the warning: mpirun has exited due to process rank <RANK#> with PID <PID#> on node <IP> exiting improperly message. This issue has been investigated but not fully resolved. It does not appear to impact functionality at this time, but should be addressed in future work;
• The current OpenMPI implementation of the above functions hardcodes MPI_COMM_WORLD within the implementation. This should be made configurable in the future;
• The current OpenMPI implementation of CommInitAll directly relies on OpenMPI-specific environment variables (e.g., OMPI_COMM_WORLD_LOCAL_RANK). This is planned to be abstracted and mapped into InfiniCCL’s own environment variable interface to provide a cleaner and more complete backend abstraction;

Logs & Screenshots

cmake .. && make -j$(nproc) && mpirun -n 4 --mca mtl ^ofi --mca btl tcp,self -x UCX_NET_DEVICES=all --allow-run-as-root ./examples/all_reduce
-- No backend specified. Defaulting to WITH_OMPI=ON
-- Auto-detecting available devices...
-- Auto-detected NVIDIA environment.
-- No MetaX GPU detected
-- InfiniCCL Config: Devices [cpu, nvidia] | Backends [ompi]
-- Configuring done (1.8s)
-- Generating done (0.0s)
-- Build files have been written to: /nfs/lizimin/InfiniCCL/build
[ 20%] Generating InfiniCCL bridge and manifest files for Devices: [cpu;nvidia] Backends: [ompi]...
[ 40%] Building CXX object src/CMakeFiles/infiniccl.dir/comm_bridge.cc.o
[ 60%] Linking CXX shared library libinfiniccl.so
[ 60%] Built target infiniccl
[ 80%] Building CXX object examples/CMakeFiles/all_reduce.dir/all_reduce.cc.o
[100%] Linking CXX executable all_reduce
[100%] Built target all_reduce
[1778167342.070887] [server:418351:0]          parser.c:2305 UCX  WARN  unused environment variable: UCX_ROOT (maybe: UCX_PROTOS?)
[1778167342.070887] [server:418351:0]          parser.c:2305 UCX  WARN  (set UCX_WARN_UNUSED_ENV_VARS=n to suppress this warning)
[1778167342.071891] [server:418349:0]          parser.c:2305 UCX  WARN  unused environment variable: UCX_ROOT (maybe: UCX_PROTOS?)
[1778167342.071891] [server:418349:0]          parser.c:2305 UCX  WARN  (set UCX_WARN_UNUSED_ENV_VARS=n to suppress this warning)
[1778167342.083620] [server:418350:0]          parser.c:2305 UCX  WARN  unused environment variable: UCX_ROOT (maybe: UCX_PROTOS?)
[1778167342.083620] [server:418350:0]          parser.c:2305 UCX  WARN  (set UCX_WARN_UNUSED_ENV_VARS=n to suppress this warning)
[1778167342.083668] [server:418352:0]          parser.c:2305 UCX  WARN  unused environment variable: UCX_ROOT (maybe: UCX_PROTOS?)
[1778167342.083668] [server:418352:0]          parser.c:2305 UCX  WARN  (set UCX_WARN_UNUSED_ENV_VARS=n to suppress this warning)
[Rank 3] Host: server | GPU: nvidia  | Device 3
[Rank 2] Host: server | GPU: nvidia  | Device 2
[Rank 1] Host: server | GPU: nvidia  | Device 1
[Rank 0] Host: server | GPU: nvidia  | Device 0

=== Performing AllReduce on GPU Memory ===
Data size: 1048576 floats (4 MB)
Operation: Sum
Warm-up iterations: 2
Profile iterations: 20

=== AllReduce Results ===
Correct: YES
Expect:  10.00
Actual:  10.00
Time:           12.140 ms
Throughput:     0.48 GB/s (Bus BW)
Alg Bandwidth:  0.32 GB/s
InfiniCCL finalized.