[UIL] Add Unified IR support

[MC952-arch]

PR Category

UIL

PR Types

New Features

PR Description

This PR introduces a transport-transparent “Unified One-Sided IR” device API layer (flagcxDev*) that auto-dispatches between intra-node P2P (IPC/NVLink) and inter-node Net/FIFO paths, along with new unit-test suites and supporting runtime changes (signal/counter slot types, IPC peer pointers for signal buffers, and updated trigger bit packing).

[Copilot]

Pull request overview

This PR introduces a transport-transparent “Unified One-Sided IR” device API layer (flagcxDev*) that auto-dispatches between intra-node P2P (IPC/NVLink) and inter-node Net/FIFO paths, along with new unit-test suites and supporting runtime changes (signal/counter slot types, IPC peer pointers for signal buffers, and updated trigger bit packing).

Changes:

• Add new unified IR C ABI (bindings/ir/flagcx_device_unified_ir*.h) implementing put/get/putValue/signal/wait/read/flush/reset/barrier with P2P-vs-Net dispatch.
• Extend device comm internals and comm traits to support P2P signal delivery via IPC-mapped peer signal-buffer pointers.
• Add new unified intra/inter test binaries and wire them into Makefile + CI.

Reviewed changes

Copilot reviewed 31 out of 31 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
test/unittest/shmem_adaptor/test_nvshmem_adaptor.cu	Update signal/counter slot typedef usage in shmem adaptor tests.
test/unittest/device_api/test_device_ir_unified.cpp	Add legacy 12-combination unified one-sided test driver.
test/unittest/device_api/test_device_ir_unified_intra.cpp	Add unified INTRA+WORLD suite test driver.
test/unittest/device_api/test_device_ir_unified_inter.cpp	Add unified INTER+WORLD suite test driver.
test/unittest/device_api/Makefile	Build/run targets for the new unified test binaries.
test/make.inc	Relax DEVICE_PLATFORM requirement for non-build goals.
test/kernel/nvidia/nvshmem_adaptor_kernel.cu	Update signal/counter slot typedef usage in NVSHMEM adaptor kernels.
test/kernel/nvidia/device_api.cu	Update signal slot typedef usage in device API kernels.
test/kernel/include/device_ir.h	Export unified test kernel launch APIs; include new enums header.
test/kernel/du/device_ir.cu	Update signal/counter slot typedef usage in DU device IR kernels.
test/kernel/du/device_api.cu	Update signal slot typedef usage in DU device API kernels.
flagcx/include/flagcx_kernel_core.h	Expand signalIdx bit-width in trigger packing (8→12 bits).
flagcx/core/proxy.cc	Reduce noisy INFO/WARN logs; add proxy loop exit summary log.
flagcx/adaptor/include/device_api/nvshmem_comm_traits.h	Add P2P signal support hooks; update slot typedefs.
flagcx/adaptor/include/device_api/nccl_comm_traits.h	Add P2P signal support hooks (disabled); update slot typedefs.
flagcx/adaptor/include/device_api/flagcx_device_internal.h	Add signalPeerPtrs + IPC slot tracking to DevComm internal.
flagcx/adaptor/include/device_api/flagcx_device_enums.h	Introduce unified slot typedefs (flagcxDevSignal_t, etc.).
flagcx/adaptor/include/device_api/flagcx_device_core.h	Update Net APIs to use new slot typedefs.
flagcx/adaptor/include/device_api/default_comm_traits.h	Add intraSize + signalPeerPtrs; fix peer pointer indexing; team-rank→world mapping.
flagcx/adaptor/include/device_api/comm_traits.h	Move to unified slot typedefs in action structs.
flagcx/adaptor/flagcx_device.cc	Initialize new signal IPC slot field.
flagcx/adaptor/device_api/default_dev_api_backend.cc	Allocate one-sided buffers more generally; add IPC setup/cleanup for signalPeerPtrs.
bindings/ir/public_symbols.txt	Export new unified IR symbols.
bindings/ir/flagcx_device_wrapper.h	Update wrapper signatures to use new slot typedefs.
bindings/ir/flagcx_device_wrapper_impl.h	Update wrapper impl signatures to use new slot typedefs.
bindings/ir/flagcx_device_unified_ir.h	New: unified transport-transparent device IR public header.
bindings/ir/flagcx_device_unified_ir_impl.h	New: unified IR implementation (P2P dispatch + Net fallback).
bindings/ir/flagcx_device_scalar_ir.h	Deprecate Net-specific APIs in favor of unified APIs; update slot typedefs.
bindings/ir/flagcx_device_scalar_ir_impl.h	Update scalar IR impl to use new slot typedefs; include unified IR impl.
.github/scripts/ci/run_unit_test.sh	Run unified intra/inter test binaries in CI.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 31 out of 32 changed files in this pull request and generated no new comments.

Suppressed comments (7)

test/unittest/device_api/test_device_ir_unified.cpp:135

• hostSend/hostRecv are over-allocated by an extra factor of 12: bufSize is already maxBytes*12 (bytes), so multiplying by 12 again can allocate >1GB at default sizes and OOM CI runs. These buffers only need to cover the device buffer size.

  // Host scratch - allocate 12× buffers for 12 combinations
  float *hostSend = new float[bufSize * 12 / sizeof(float)];
  float *hostRecv = new float[bufSize * 12 / sizeof(float)];

test/unittest/device_api/test_device_ir_unified_intra.cpp:132

• hostSend/hostRecv are over-allocated by an extra factor of 8: bufSize is already maxBytes*8 (bytes), so multiplying by 8 again needlessly inflates host memory usage and may OOM on large -e values.
  test/unittest/device_api/test_device_ir_unified_inter.cpp:136
• hostSend/hostRecv are over-allocated by an extra factor of 8: bufSize is already maxBytes*8 (bytes), so multiplying by 8 again needlessly inflates host memory usage and may OOM on large -e values.

  // Host scratch - allocate 8× buffers for S18's 8 combinations
  float *hostSend = new float[bufSize * 8 / sizeof(float)];
  float *hostRecv = new float[bufSize * 8 / sizeof(float)];

test/unittest/device_api/test_device_ir_unified_intra.cpp:651

• sendBuff/recvBuff are allocated with a non-default allocator (memAllocator) but freed using the default allocator (flagcxMemCCL). If FLAGCX_COMM_TRAITS_SHMEM is enabled this will free with the wrong backend. Pass memAllocator to flagcxMemFree, matching the allocation site.
  test/unittest/device_api/test_device_ir_unified_inter.cpp:685
• sendBuff/recvBuff are allocated with a non-default allocator (memAllocator) but freed using the default allocator (flagcxMemCCL). If FLAGCX_COMM_TRAITS_SHMEM is enabled this will free with the wrong backend. Pass memAllocator to flagcxMemFree, matching the allocation site.

  FLAGCXCHECK(flagcxCommWindowDeregister(comm, sendWin, memAllocator));
  FLAGCXCHECK(flagcxCommWindowDeregister(comm, recvWin, memAllocator));
  FLAGCXCHECK(flagcxMemFree(sendBuff));
  FLAGCXCHECK(flagcxMemFree(recvBuff));

test/unittest/device_api/Makefile:126

• run-mpi-unified-intra omits $(HETERO_ENV), but the unified intra test itself documents FLAGCX_USE_HETERO_COMM=1 as a requirement (for DevComm). As written, local make run-mpi can fail while CI passes because CI injects env separately.

# Unified Intra suite: INTRA + WORLD teams (single-node)
run-mpi-unified-intra: $(BINDIR)/test_device_ir_unified_intra
	@echo "Running unified intra suite tests (INTRA + WORLD, 8-combo)..."
	mpirun -np 8 --allow-run-as-root $(BINDIR)/test_device_ir_unified_intra

bindings/ir/flagcx_device_wrapper.h:328

• The Category 11 header still claims “One-Sided put (16)”, but only 6 put variants are declared in this section now. Please update the section header comment so IR consumers don’t assume the removed combinations are still available.

 * @param coop           Pointer to flagcxCoopAny struct.
 * @param remoteSignal   Remote signal slot (R variants).
 * @param remoteValue    Value to add to remote signal (RSigAdd).
 * @param localCounter   Local counter slot (LCtrInc).
 * ================================================================ */

[Copilot]

Pull request overview

Copilot reviewed 39 out of 40 changed files in this pull request and generated no new comments.

Suppressed comments (3)

test/unittest/device_api/test_device_ir_unified.cpp:14

• This new 670-line test file is not wired into the build: the Makefile only builds test_device_ir_unified_intra and test_device_ir_unified_inter, and no CI script references test_device_ir_unified. As submitted it will never be compiled or run, so it can silently rot. Additionally, this header docstring is stale — it lists S16: Unified Put … S22: Team-resolution, but the actual blocks below are renumbered (S16–S18 are barriers, S19 is team-resolution, S20 Put, S21 Get, S22 Signal, S23 Put+Signal+Wait). Please either add a build/run target for this file (and fix the docstring) or remove it.
  test/unittest/device_api/test_device_ir_unified.cpp:205
• On a single node (nNodes == 1) this block prints "SKIP (single-node)" but still executes allPass &= s17Pass with s17Pass = (hostRes == 1). If launchKernelDevBarrierInterS does not set the result to 1 when there are no inter-node peers, a "skipped" inter-node barrier will fail the whole suite. The allPass accumulation should be gated on nNodes > 1 so a genuinely skipped case cannot flip the result.
  bindings/ir/flagcx_device_scalar_ir_impl.h:529
• This label is incorrect for flagcxDevNetPutS_LCtrInc, which issues a flagcxDevNet_None{} remote action plus a local flagcxDevNet_CounterInc. It should read (None, CtrInc) (matching the declaration in flagcx_device_scalar_ir.h); (SigInc, CtrInc) is the correct label for the following flagcxDevNetPutS_RSigInc_LCtrInc variant.

/* (SigInc, CtrInc) */