Add PE-level IPCQ collective infra + unified ccl_allreduce bench (ADR-0023)

Major changes:

PE-level IPCQ infrastructure:
- New PE_IPCQ component: ring-buffer control plane with 4-direction
  neighbor mapping, head/tail pointers, backpressure (poll/sleep).
- PE_DMA extended with vc_comm channel for IPCQ outbound/inbound DMA,
  including in-flight data snapshot (D9) and op_log recording at
  outbound time for Phase 2 replay correctness.
- IpcqDmaToken piggyback model: data + metadata travel together,
  atomic visibility at receiver (invariant I6).
- Credit return fast path: bottleneck-BW latency, no fabric vc_comm.

Phase 2 data execution (ADR-0020 integration):
- op_log extended: DmaWriteCmd now captures src_space/src_addr for
  Phase 2 dma_write copy; ipcq_copy ops recorded at outbound time.
- DataExecutor replays dma_write + ipcq_copy in t_start order.
- Engine._flush_data_phase: incremental cursor-based replay after
  each engine.wait() so host reads see post-Phase-2 data.
- KernelRunner Phase 1 writes disabled when op_log is active to
  prevent stale data from corrupting the MemoryStore snapshot.

TLContext / kernel API:
- tl.send(dir, src=TensorHandle), tl.recv(dir, shape, dtype),
  tl.recv_async, tl.wait(RecvFuture), copy_to_dst mode.
- TensorHandle operator overloading (add/sub/mul/div) via thread-local
  active TLContext → MathCmd dispatch through PE_MATH.
- PE-local scratch allocator for math output handles.
- tl.load returns space="hbm" handles for correct Phase 2 addressing.
- Additional math functions: maximum, minimum, fma, clamp, softmax, cdiv.

Unified ccl_allreduce bench (PyTorch-compat host code):
- Single benches/ccl_allreduce.py with run() + worker(rank, ws, torch)
  split matching real PyTorch DDP worker pattern.
- torch.distributed facade: init_process_group, get_world_size,
  get_rank, get_backend, all_reduce, barrier — only real PyTorch names.
- AhbmCCLBackend: eager install_ipcq at init, all_reduce dispatches
  kernel via tensor shard metadata (n_elem from shards[0].nbytes).
- world_size derived from topology spec (sips × cubes × pes_per_cube)
  with optional algorithm-level override in ccl.yaml.

Tensor API (PyTorch-compat surface):
- Tensor.numpy(): gather-aware (all shards via VA-based addressing).
- Tensor.copy_(source): scatter from host tensor into sharded target.
- RuntimeContext.from_numpy(arr): host-side staging tensor.
- Tensor.data property fixed to use numpy() (was shards[0]-only).

Algorithm modules moved to src/kernbench/ccl/algorithms/:
- ring_allreduce, mesh_allreduce, tree_allreduce, hello_send.
- Each module exports kernel_args(world_size, n_elem) helper.
- ccl.yaml module paths updated to kernbench.ccl.algorithms.*.

Dead code removed:
- 7 per-variant bench files (ccl_allreduce_{tcm,hbm,sram}, etc.).
- _run_ccl_bench greenlet-per-SIP scheduler.
- benches.loader.is_ccl_bench + run_rank detection.
- benches/ccl/ directory.

Tests:
- New test_ccl_allreduce_matrix.py: 7 parametrized cases
  (ring×3 buffers, ring 8/16, mesh 4, tree 7).
- New test_runtime_api_tensor.py: copy_/numpy/from_numpy unit tests.
- Existing tests updated for new import paths + world_size_override.

Docs:
- Korean ccl-author-guide.md and ADR-0023 paths updated.
- New English versions: ccl-author-guide.en.md, ADR-0023.en.md.

502 tests pass.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-04-12 19:36:59 -07:00
parent ff2c677a9c
commit 998cc85762
60 changed files with 9196 additions and 80 deletions
+76 -7
View File
@@ -51,7 +51,42 @@ class DataExecutor:
self._execute_math(op)
def _execute_memory(self, op: OpRecord) -> None:
"""Memory ops are already handled by Phase 1 MemoryStore. Skip."""
"""Replay memory copy ops in Phase 2 (ADR-0020 + ADR-0023).
- dma_read: no-op (handle already references HBM source).
- dma_write: copy (src_space, src_addr) → (dst_space, dst_addr).
Required because Phase 2 may have just produced new data at the
source addr (e.g. PE_MATH scratch output).
- ipcq_copy: copy across PEs — sender's source → receiver's slot.
Required because the source may be a Phase 2 math output, and
a downstream math op on the receiver reads from the slot.
Legacy entries without src/dst metadata are silently skipped.
"""
p = op.params
if op.op_name == "dma_write" or op.op_name == "ipcq_copy":
src_space = p.get("src_space")
src_addr = p.get("src_addr")
dst_space = p.get("dst_space")
dst_addr = p.get("dst_addr")
if (src_space is None or src_addr is None
or dst_space is None or dst_addr is None):
return
# Prefer the Phase-1-time snapshot (captured at record_end /
# outbound) so we don't read from a source that has since been
# mutated by another op. Fall back to MemoryStore for sources
# that had no Phase 1 data (e.g. math scratch outputs that
# only get populated by Phase 2's math replay).
data = p.get("snapshot")
if data is None:
try:
data = self.store.read(
src_space, src_addr,
shape=p.get("shape"), dtype=p.get("dtype"),
)
except KeyError:
return
self.store.write(dst_space, dst_addr, data)
def _execute_gemm(self, op: OpRecord) -> None:
"""Execute GEMM: out = a @ b."""
@@ -77,18 +112,35 @@ class DataExecutor:
"""Execute math op: unary, binary, or reduction."""
p = op.params
math_op = p.get("op", op.op_name)
space = p.get("addr_space", "tcm")
dtype = p.get("dtype", "f32")
input_addrs = p.get("input_addrs", [])
input_shapes = p.get("input_shapes", [])
# Per-input space/dtype (ADR-0023 CCL accumulation): math ops can
# mix inputs from different MemoryStore spaces (e.g. acc in "hbm",
# recv slot in "tcm"). Fall back to legacy single-space mode when
# the per-input lists are absent.
input_spaces = p.get("input_spaces") or [p.get("addr_space", "tcm")] * len(input_addrs)
input_dtypes = p.get("input_dtypes") or [dtype] * len(input_addrs)
# Per-input data snapshots (ADR-0020 D6): captured at op_log
# record time. Phase 1 has correct values for slot/HBM addrs at
# that moment, which lets Phase 2 sidestep the slot-wraparound
# races where a later round overwrites a slot before this op
# runs in t_start order.
snapshots = p.get("input_snapshots") or [None] * len(input_addrs)
dst_space = p.get("dst_space", p.get("addr_space", "tcm"))
inputs = []
for addr, shape in zip(input_addrs, input_shapes):
inputs.append(self.store.read(space, addr, shape=shape, dtype=dtype))
for addr, shape, space, idtype, snap in zip(
input_addrs, input_shapes, input_spaces, input_dtypes, snapshots
):
if snap is not None:
inputs.append(snap)
else:
inputs.append(self.store.read(space, addr, shape=shape, dtype=idtype))
result = _compute_math(math_op, inputs, p.get("axis"))
if result is not None:
self.store.write(space, p["dst_addr"], result)
self.store.write(dst_space, p["dst_addr"], result)
def verify(self, expected: dict[tuple[str, int], np.ndarray],
rtol: float = 1e-3, atol: float = 1e-3) -> dict[str, bool]:
@@ -146,6 +198,14 @@ def _compute_math(op: str, inputs: list[np.ndarray], axis: int | None) -> np.nda
if op == "min":
return np.min(x, axis=axis, keepdims=True)
# Softmax (numerically stable)
if op == "softmax":
ax = axis if axis is not None else -1
x_max = np.max(x, axis=ax, keepdims=True)
e = np.exp(x - x_max)
s = np.sum(e, axis=ax, keepdims=True)
return e / s
# Binary
if len(inputs) >= 2:
y = inputs[1]
@@ -157,9 +217,18 @@ def _compute_math(op: str, inputs: list[np.ndarray], axis: int | None) -> np.nda
return x * y
if op == "div":
return x / y
if op == "maximum":
return np.maximum(x, y)
if op == "minimum":
return np.minimum(x, y)
# Ternary
if op == "where" and len(inputs) >= 3:
return np.where(inputs[0], inputs[1], inputs[2])
if len(inputs) >= 3:
if op == "where":
return np.where(inputs[0], inputs[1], inputs[2])
if op == "fma":
return inputs[0] * inputs[1] + inputs[2]
if op == "clamp":
return np.minimum(np.maximum(inputs[0], inputs[1]), inputs[2])
return None
+55 -2
View File
@@ -51,8 +51,12 @@ class GraphEngine:
if enable_data:
from kernbench.sim_engine.memory_store import MemoryStore
from kernbench.sim_engine.op_log import OpLogger
self._op_logger = OpLogger()
self._memory_store = MemoryStore()
self._op_logger = OpLogger(memory_store=self._memory_store)
# Cursor for incremental Phase 2 replay (ADR-0020 D6).
# SimPy env.now is monotonic so newly logged records always sort
# to the tail; the cursor remains valid across waits.
self._data_cursor = 0
ctx = ComponentContext(
router=self._router,
@@ -147,11 +151,60 @@ class GraphEngine:
self._env.process(self._process(str(handle), request, event))
return handle
def _flush_data_phase(self) -> None:
"""Replay newly recorded op_log entries through DataExecutor.
ADR-0020 D6 Phase 2: when data tracking is enabled, run DataExecutor
on records added since the last flush so that callers reading
MemoryStore between launches observe correct (compute-replayed)
tensor data.
Cursor-based incremental replay is necessary because Phase 2 is
NOT idempotent across full re-runs: a math op writes a TCM scratch
addr, a later dma_write copies that scratch into HBM[X], and an
even-later math op may then read HBM[X]. Re-running everything
from scratch would let the second pass's first math op read the
already-overwritten HBM[X] instead of the original input.
"""
if self._op_logger is None or self._memory_store is None:
return
records = self._op_logger.records # sorted by t_start (stable)
if self._data_cursor >= len(records):
return
new_records = records[self._data_cursor:]
from kernbench.sim_engine.data_executor import DataExecutor
DataExecutor(new_records, self._memory_store).run()
self._data_cursor = len(records)
def wait(self, handle: RequestHandle) -> None:
key = str(handle)
event = self._events[key]
if not event.triggered:
self._env.run(until=event)
try:
self._env.run(until=event)
except (simpy.core.EmptySchedule, RuntimeError) as exc:
# SimPy raises EmptySchedule directly OR (in newer simpy)
# wraps it as a RuntimeError("No scheduled events left ...").
# Either case while our event is still pending → IPCQ deadlock.
msg = str(exc)
is_deadlock = (
isinstance(exc, simpy.core.EmptySchedule)
or "No scheduled events left" in msg
)
if not is_deadlock:
raise
from kernbench.ccl.diagnostics import IpcqDeadlock, pointer_dump
dump = pointer_dump(self)
if dump.strip():
raise IpcqDeadlock(
"IPCQ deadlock: simulation schedule empty while "
f"request {handle!r} is still pending.\n"
f"Pointer state:\n{dump}"
) from None
raise
# ADR-0020: replay newly logged ops so the caller observes
# post-Phase-2 tensor state from MemoryStore.
self._flush_data_phase()
def get_completion(self, handle: RequestHandle) -> tuple[Completion, Trace | None]:
return self._results[str(handle)]
+84 -1
View File
@@ -29,9 +29,13 @@ class OpLogger:
Records are maintained in t_start stable ordering (insertion order).
"""
def __init__(self) -> None:
def __init__(self, memory_store: Any | None = None) -> None:
self._records: list[OpRecord] = []
self._pending: dict[int, dict[str, Any]] = {} # msg id → partial record
# Optional MemoryStore reference. When set, math op records capture
# input data snapshots at record_end time so Phase 2 replay does
# not depend on slot/scratch addrs surviving until math runs.
self._memory_store = memory_store
@property
def records(self) -> list[OpRecord]:
@@ -53,6 +57,38 @@ class OpLogger:
if pending is None:
return
op_kind, op_name, params = _extract_op_info(msg)
# Snapshot data at record time so Phase 2 replay sidesteps
# downstream mutations of source addrs (e.g. a tl.store that
# overwrites HBM after a load handle was sent, or a slot that
# gets reused on the next ring round).
if self._memory_store is not None:
if op_kind == "math":
snaps: list[Any] = []
for addr, shape, space, idtype in zip(
params.get("input_addrs", []),
params.get("input_shapes", []),
params.get("input_spaces", []),
params.get("input_dtypes", []),
):
try:
arr = self._memory_store.read(
space, addr, shape=shape, dtype=idtype,
)
snaps.append(arr.copy() if hasattr(arr, "copy") else arr)
except Exception:
snaps.append(None)
params["input_snapshots"] = snaps
elif op_name == "dma_write":
try:
arr = self._memory_store.read(
params["src_space"], params["src_addr"],
shape=params.get("shape"), dtype=params.get("dtype"),
)
params["snapshot"] = (
arr.copy() if hasattr(arr, "copy") else arr
)
except Exception:
params["snapshot"] = None
self._records.append(OpRecord(
t_start=pending["t_start"],
t_end=t,
@@ -62,6 +98,45 @@ class OpLogger:
params=params,
))
def record_copy(
self, t_start: float, t_end: float, component_id: str,
src_space: str, src_addr: int,
dst_space: str, dst_addr: int,
shape: tuple[int, ...], dtype: str, nbytes: int,
) -> None:
"""Record a memory copy op for Phase 2 replay (ADR-0023 + ADR-0020).
Used by PE_DMA at outbound (sender) time: the snapshot captures
the source data at the moment the send was issued, so Phase 2
replay does not see later mutations of the source addr (e.g. a
tl.store that runs after the recv at the sender).
For sources whose data is not yet materialized in Phase 1 (math
scratch outputs), the snapshot is None and Phase 2 falls back to
reading from MemoryStore — by which point the corresponding math
op has been replayed and the scratch addr is populated.
"""
snap = None
if self._memory_store is not None:
try:
arr = self._memory_store.read(
src_space, src_addr, shape=shape, dtype=dtype,
)
snap = arr.copy() if hasattr(arr, "copy") else arr
except Exception:
snap = None
self._records.append(OpRecord(
t_start=t_start, t_end=t_end,
component_id=component_id,
op_kind="memory", op_name="ipcq_copy",
params={
"src_space": src_space, "src_addr": src_addr,
"dst_space": dst_space, "dst_addr": dst_addr,
"shape": shape, "dtype": dtype, "nbytes": nbytes,
"snapshot": snap,
},
))
def _extract_op_info(msg: Any) -> tuple[str, str, dict[str, Any]]:
"""Extract op_kind, op_name, params from a data_op message."""
@@ -76,6 +151,11 @@ def _extract_op_info(msg: Any) -> tuple[str, str, dict[str, Any]]:
}
if isinstance(msg, DmaWriteCmd):
return "memory", "dma_write", {
"src_space": getattr(msg.handle, "space", "tcm"),
"src_addr": msg.handle.addr,
"shape": msg.handle.shape,
"dtype": msg.handle.dtype,
"dst_space": "hbm",
"dst_addr": msg.dst_addr,
"nbytes": msg.nbytes,
"handle_id": msg.handle.id,
@@ -96,7 +176,10 @@ def _extract_op_info(msg: Any) -> tuple[str, str, dict[str, Any]]:
return "math", msg.op, {
"input_addrs": [h.addr for h in msg.inputs],
"input_shapes": [h.shape for h in msg.inputs],
"input_spaces": [getattr(h, "space", "tcm") for h in msg.inputs],
"input_dtypes": [h.dtype for h in msg.inputs],
"dst_addr": msg.out.addr,
"dst_space": getattr(msg.out, "space", "tcm"),
"shape_out": msg.out.shape,
"dtype": msg.out.dtype,
"axis": msg.axis,