ywkang
998cc85762
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>
81 lines
3.3 KiB
Python
81 lines
3.3 KiB
Python
"""Ring all-reduce kernel for IPCQ-based PE collective (ADR-0023).
|
|
|
|
Algorithm: 1D ring of N PEs, each PE starts with one tile of data.
|
|
After ``world_size - 1`` rounds, every PE's accumulator holds the sum
|
|
of all PE tiles.
|
|
|
|
Strategy
|
|
--------
|
|
Each PE starts with its own tile in HBM. The kernel:
|
|
1. Loads the local tile into a TensorHandle (the accumulator).
|
|
2. In each of ``world_size - 1`` rounds:
|
|
- Sends the current accumulator/recv slot to the E neighbor.
|
|
- Receives a tile from the W neighbor — the recv handle points
|
|
into the per-direction TCM slot.
|
|
- Adds the received tile to the accumulator using the TensorHandle
|
|
operator overload, which dispatches to ``MathCmd`` (PE_MATH).
|
|
3. Stores the final accumulator back to HBM via tl.store. The store is
|
|
recorded in op_log with both src and dst, so Phase 2 will copy the
|
|
replayed math result from PE-local scratch into HBM.
|
|
|
|
ADR-0020 D3 split: Phase 1 simulates timing only — math results are
|
|
not yet computed, so the accumulator data flowing through Phase 1 may
|
|
be stale. Phase 2's DataExecutor replays math + IPCQ copies + dma_write
|
|
in stable t_start order, producing correct final HBM contents.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
|
|
def kernel_args(world_size: int, n_elem: int) -> tuple:
|
|
"""Return the positional kernel arguments for the ahbm backend.
|
|
|
|
Ring all-reduce takes (n_elem, world_size) after the tensor pointer.
|
|
"""
|
|
return (n_elem, world_size)
|
|
|
|
|
|
def kernel(t_ptr, n_elem, world_size, tl):
|
|
"""Ring all-reduce.
|
|
|
|
Args:
|
|
t_ptr: HBM base address of the column-sharded tensor — all PEs
|
|
share this base. The per-PE slice lives at
|
|
``t_ptr + global_rank * n_elem * 2``.
|
|
n_elem: number of f16 elements per tile.
|
|
world_size: total number of participating ranks (passed by host).
|
|
tl: TLContext (auto-injected, ADR-0022). The kernel derives the
|
|
global rank from ``program_id(axis=0)`` (local PE) and
|
|
``program_id(axis=1)`` (cube id):
|
|
|
|
rank = cube_id * pes_per_cube + local_pe
|
|
"""
|
|
local_pe = tl.program_id(axis=0)
|
|
cube_id = tl.program_id(axis=1)
|
|
pes_per_cube = tl.num_programs(axis=0)
|
|
rank = cube_id * pes_per_cube + local_pe
|
|
nbytes = n_elem * 2 # f16
|
|
|
|
# Each PE reads from its own slice of the shared base address
|
|
pe_addr = t_ptr + rank * nbytes
|
|
|
|
# Load the local tile — handle points at HBM[pe_addr].
|
|
acc = tl.load(pe_addr, shape=(n_elem,), dtype="f16")
|
|
# The ring forwards each received tile to the next neighbor (NOT the
|
|
# cumulative accumulator), so every rank's tile passes through every
|
|
# rank exactly once. The accumulator sums the new arrival each round.
|
|
current = acc
|
|
|
|
for _step in range(world_size - 1):
|
|
tl.send(dir="E", src=current)
|
|
recv = tl.recv(dir="W", shape=(n_elem,), dtype="f16")
|
|
# TensorHandle add → MathCmd → PE_MATH (timing in Phase 1, real
|
|
# numpy in Phase 2 via DataExecutor). The result handle lives at
|
|
# an auto-allocated PE-local scratch addr.
|
|
acc = acc + recv
|
|
current = recv # forward W's tile to E next round
|
|
|
|
# Final result back to this PE's HBM slice. Op_log captures the
|
|
# source (scratch addr) and dst (HBM slice) so Phase 2 copies the
|
|
# accumulated value into HBM for verification.
|
|
tl.store(pe_addr, acc)
|