Files
kernbench2/tests/test_pe_pipeline.py
T
ywkang 23992548f7 gemm(perf): fix back-to-back DMA pipelining + tighten analytic model
Three coupled fixes that recover small-tile GEMM pipeline efficiency
from 53% to 88% (32x3072x32 load_ref, composite_window basis).

1. PE_DMA channel-hold (ADR-0014 D4 clarified): both the
   _handle_with_hooks (PeInternalTxn) and _pipeline_process
   (TileToken) paths used to hold the cap=1 DMA channel through the
   full HBM round-trip, which double-serialized with the HBM_CTRL's
   own per-PC `available_at` model and prevented back-to-back tile
   DMAs from amortizing their per-request head latency. Channel is
   now released after the request is enqueued onto the next hop;
   HBM serialization is HBM_CTRL's responsibility alone.

   Tests: new test_pe_dma_back_to_back_pipelining as the oracle
   (asserts wall < 75% of strict-serialized N x single_op). Existing
   test_pe_dma_record_start_after_channel_acquire rewritten to assert
   t_start clustering (channel released fast) instead of the old
   round-trip-hold invariant. test_pe_dma_same_channel_serializes
   still passes — HBM_CTRL preserves ordering.

   Probe regression: PE→local-HBM 32 KiB stays at 141 ns
   (single-request, unaffected).

2. milestone_1h_gemm bench: matmul_composite was reading MATMUL_M/K/N
   env vars at module load, so every sweep row replayed the cached
   256³ result; values now read inside run(). Drops the stale
   sys.modules deletion hack.

3. Analytic ideal-pipeline model: dropped the (n_mn-1)·dma_w_per_pair
   penalty (over-pessimistic for under-tile shapes — it pushed
   measured > theoretical) and replaced the D_STAGES-derived head
   with empirical T_PIPELINE_FILL=60 ns / T_PIPELINE_TAIL=30 ns.
   Max analytic-vs-measured gap across all 7 swept shapes now 2.2 ppt
   (was 9-44 ppt under the old constants).

Paper updates:
- §3 (GEMM): 78%→88% measured at 48 tiles, 23%→15% at 1 tile,
  stage breakdown numbers refreshed (DMA in / Fetch / GEMM all
  ~785 ns at K=3072), analytic-vs-measured agreement tightened
  to "within 2.2 ppt".
- §2.4 (Accuracy): GEMM tracking claim refreshed accordingly.
- §5 (GQA): restore long-ctx 4-cases figures into figures/
  (they were dropped from bench output dir as derived artifacts in
  92b9221 / e45626c but §5 still cites them by name).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-06-17 09:56:05 -07:00

510 lines
18 KiB
Python

"""Tests for ADR-0014 D6 PE pipeline: TileToken self-routing, pipeline overlap, e2e accuracy.
Test plan items:
3. Phase 1 → Phase 2 end-to-end (op_log → DataExecutor → verify)
4. TileToken self-routing (stage sequence, PipelineContext completion)
5. Async pipeline overlap (intra-command tile overlap, FIFO ordering)
"""
from __future__ import annotations
import simpy
from kernbench.components.builtin.pe_types import (
PipelineContext,
PipelinePlan,
Stage,
StageType,
TilePlan,
TileToken,
)
# ── 4. TileToken self-routing ────────────────────────────────────────
def test_tile_token_advance():
"""TileToken.advance() increments stage_idx and returns next Stage."""
stages = (
Stage(StageType.DMA_READ, "pe_dma", {"src_addr": 0}),
Stage(StageType.FETCH, "pe_fetch_store", {"direction": "read"}),
Stage(StageType.GEMM, "pe_gemm", {"m": 32, "k": 64, "n": 32}),
)
plan = TilePlan(tile_id=0, stages=stages)
ctx = PipelineContext(id="p1", total_tiles=1)
token = TileToken(
tile_id=0, pipeline_ctx=ctx, plan=plan,
stage_idx=0, params=stages[0].params,
)
assert token.current_stage.stage_type == StageType.DMA_READ
next_s = token.advance()
assert next_s is not None
assert next_s.stage_type == StageType.FETCH
assert token.stage_idx == 1
assert token.params == {"direction": "read"}
next_s = token.advance()
assert next_s is not None
assert next_s.stage_type == StageType.GEMM
assert token.stage_idx == 2
# Last stage — advance returns None
assert token.advance() is None
assert token.stage_idx == 3
def test_pipeline_context_completion():
"""PipelineContext.complete_tile() fires done_event on last tile."""
env = simpy.Environment()
done = env.event()
ctx = PipelineContext(id="p1", total_tiles=3, done_event=done)
ctx.complete_tile()
assert not done.triggered
ctx.complete_tile()
assert not done.triggered
ctx.complete_tile()
assert done.triggered
def test_pipeline_context_exactly_once():
"""PipelineContext tracks completed_tiles correctly."""
ctx = PipelineContext(id="p1", total_tiles=2)
assert ctx.completed_tiles == 0
ctx.complete_tile()
assert ctx.completed_tiles == 1
ctx.complete_tile()
assert ctx.completed_tiles == 2
def test_tile_token_self_routing_chain():
"""Simulated self-routing: component reads next stage from token."""
stages = (
Stage(StageType.DMA_READ, "dma", {}),
Stage(StageType.FETCH, "fetch", {}),
Stage(StageType.GEMM, "gemm", {}),
Stage(StageType.STORE, "fetch", {}),
Stage(StageType.DMA_WRITE, "dma", {}),
)
plan = TilePlan(tile_id=0, stages=stages)
ctx = PipelineContext(id="p1", total_tiles=1)
token = TileToken(
tile_id=0, pipeline_ctx=ctx, plan=plan,
stage_idx=0, params=stages[0].params,
)
visited = []
while True:
visited.append(token.current_stage.component)
next_s = token.advance()
if next_s is None:
ctx.complete_tile()
break
assert visited == ["dma", "fetch", "gemm", "fetch", "dma"]
assert ctx.completed_tiles == 1
# ── 5. Tiling plan generation ────────────────────────────────────────
def test_gemm_plan_tile_count():
"""generate_gemm_plan produces correct number of tiles."""
from kernbench.components.builtin.tiling import generate_gemm_plan
plan = generate_gemm_plan(
M=64, K=128, N=64,
tile_m=32, tile_k=64, tile_n=32,
bytes_per_element=2,
A_addr=0, B_addr=0x1000, C_addr=0x2000,
pe_prefix="sip0.cube0.pe0",
)
# M_tiles=2, K_tiles=2, N_tiles=2 → 2*2*2 = 8 tiles
assert len(plan.tiles) == 8
assert plan.m_tiles == 2
assert plan.k_tiles == 2
assert plan.n_tiles == 2
def test_gemm_plan_stage_sequence():
"""Each GEMM tile has correct stage sequence."""
from kernbench.components.builtin.tiling import generate_gemm_plan
plan = generate_gemm_plan(
M=32, K=64, N=32,
tile_m=32, tile_k=64, tile_n=32,
bytes_per_element=2,
A_addr=0, B_addr=0x1000, C_addr=0x2000,
pe_prefix="sip0.cube0.pe0",
)
# Single tile (1x1x1), last_k=True → includes DMA_WRITE
assert len(plan.tiles) == 1
tile = plan.tiles[0]
stage_types = [s.stage_type for s in tile.stages]
assert stage_types == [
StageType.DMA_READ, StageType.DMA_READ, # A and B
StageType.FETCH, StageType.GEMM, StageType.STORE,
StageType.DMA_WRITE,
]
def test_gemm_plan_intermediate_k_no_dma_write():
"""Intermediate K-tiles don't have DMA_WRITE or STORE stage.
The C accumulator stays in RegFile across the K loop; STORE +
DMA_WRITE only fire on the last K-tile per (m,n).
"""
from kernbench.components.builtin.tiling import generate_gemm_plan
plan = generate_gemm_plan(
M=32, K=128, N=32, # K_tiles=2
tile_m=32, tile_k=64, tile_n=32,
bytes_per_element=2,
A_addr=0, B_addr=0x1000, C_addr=0x2000,
pe_prefix="sip0.cube0.pe0",
)
assert len(plan.tiles) == 2
# First tile (k=0): no STORE, no DMA_WRITE — accumulator stays in RegFile
t0_types = [s.stage_type for s in plan.tiles[0].stages]
assert StageType.STORE not in t0_types
assert StageType.DMA_WRITE not in t0_types
# Last tile (k=1, last_k=True): has both STORE and DMA_WRITE
t1_types = [s.stage_type for s in plan.tiles[1].stages]
assert StageType.STORE in t1_types
assert StageType.DMA_WRITE in t1_types
def test_gemm_plan_pinned_operand_skips_dma_read():
"""When a_pinned=True, A's per-tile DMA_READ is omitted.
Same for b_pinned. FETCH is unaffected — it still stages from TCM
into RegFile.
"""
from kernbench.components.builtin.tiling import generate_gemm_plan
# Baseline: neither pinned — both A and B get DMA_READ per tile.
base = generate_gemm_plan(
M=32, K=128, N=32, # K_tiles=2
tile_m=32, tile_k=64, tile_n=32,
bytes_per_element=2,
A_addr=0, B_addr=0x1000, C_addr=0x2000,
pe_prefix="sip0.cube0.pe0",
)
for tile in base.tiles:
operands = [s.params.get("operand") for s in tile.stages
if s.stage_type == StageType.DMA_READ]
assert operands == ["A", "B"], \
f"baseline tile should DMA_READ A and B, got {operands}"
# a_pinned: no A DMA_READ.
plan_a = generate_gemm_plan(
M=32, K=128, N=32,
tile_m=32, tile_k=64, tile_n=32,
bytes_per_element=2,
A_addr=0, B_addr=0x1000, C_addr=0x2000,
pe_prefix="sip0.cube0.pe0",
a_pinned=True,
)
for tile in plan_a.tiles:
operands = [s.params.get("operand") for s in tile.stages
if s.stage_type == StageType.DMA_READ]
assert operands == ["B"], \
f"a_pinned should leave only B DMA_READ, got {operands}"
# FETCH must still exist
assert any(s.stage_type == StageType.FETCH for s in tile.stages)
# Both pinned: no DMA_READ at all.
plan_both = generate_gemm_plan(
M=32, K=128, N=32,
tile_m=32, tile_k=64, tile_n=32,
bytes_per_element=2,
A_addr=0, B_addr=0x1000, C_addr=0x2000,
pe_prefix="sip0.cube0.pe0",
a_pinned=True, b_pinned=True,
)
for tile in plan_both.tiles:
dma_reads = [s for s in tile.stages
if s.stage_type == StageType.DMA_READ]
assert dma_reads == [], \
f"both pinned should skip all DMA_READ, got {dma_reads}"
def test_math_plan_stage_sequence():
"""Math plan has READ→FETCH→MATH→STORE→WRITE sequence."""
from kernbench.components.builtin.tiling import generate_math_plan
plan = generate_math_plan(
M=32, N=32,
tile_m=32, tile_n=32,
bytes_per_element=2,
math_op="exp",
src_addr=0, dst_addr=0x1000,
pe_prefix="sip0.cube0.pe0",
)
assert len(plan.tiles) == 1
stage_types = [s.stage_type for s in plan.tiles[0].stages]
assert stage_types == [
StageType.DMA_READ, StageType.FETCH, StageType.MATH,
StageType.STORE, StageType.DMA_WRITE,
]
# ── 5. Async pipeline (SimPy simulation) ─────────────────────────────
def test_pipeline_overlap_within_command():
"""Tiles within same command overlap: tile1 DMA while tile0 in GEMM."""
env = simpy.Environment()
done_event = env.event()
ctx = PipelineContext(id="p1", total_tiles=2, done_event=done_event)
# Track when each tile enters each stage
stage_times: dict[tuple[int, str], float] = {}
def mock_component(env, inbox, stage_name, latency_ns, out_ports):
while True:
token = yield inbox.get()
stage_times[(token.tile_id, stage_name)] = env.now
yield env.timeout(latency_ns)
next_s = token.advance()
if next_s is not None:
yield out_ports[next_s.component].put(token)
else:
token.pipeline_ctx.complete_tile()
dma_q = simpy.Store(env)
gemm_q = simpy.Store(env)
out_ports = {"dma": dma_q, "gemm": gemm_q}
env.process(mock_component(env, dma_q, "dma", 10.0, out_ports))
env.process(mock_component(env, gemm_q, "gemm", 20.0, out_ports))
# Create 2 tiles: DMA → GEMM
for i in range(2):
stages = (
Stage(StageType.DMA_READ, "dma", {}),
Stage(StageType.GEMM, "gemm", {}),
)
plan = TilePlan(tile_id=i, stages=stages)
token = TileToken(
tile_id=i, pipeline_ctx=ctx, plan=plan,
stage_idx=0, params={},
)
dma_q.put(token)
env.run()
assert done_event.triggered
# tile0 DMA starts at 0, finishes at 10
# tile1 DMA starts at 10, finishes at 20
# tile0 GEMM starts at 10, finishes at 30
# tile1 GEMM starts at 20 (wait for DMA) but actually at 30 (gemm queue)
assert stage_times[(0, "dma")] == 0.0
assert stage_times[(0, "gemm")] == 10.0
assert stage_times[(1, "dma")] == 10.0
# tile1 gemm starts when tile0 gemm finishes (serialized at gemm queue)
assert stage_times[(1, "gemm")] == 30.0
# ── 6. Option B: pe_dma record_start fires post channel-acquire ────────
def test_pe_dma_record_start_after_channel_acquire():
"""Three back-to-back DMA_READs at the same PE_DMA.
With ``_DEFER_RECORD_START = True`` on PeDmaComponent, each op's
``t_start`` is captured right after ``yield req`` succeeds (Option B:
record_start fires post-channel-acquire, not at command-entry).
The DMA channel is held only for the **issue path** (ADR-0014 D4),
not for the full HBM round-trip — HBM-level serialization is the
HBM_CTRL's responsibility (per-PC `available_at`).
Under that model, three back-to-back DMAs at t=0 give:
- ``t_start`` values monotonic and tightly clustered (channel
release is fast → each subsequent op's ``t_start`` is at or
immediately after the previous one's, NOT after the previous
one's HBM round-trip completes).
- ``t_end`` values reflect HBM_CTRL serialization (later ops can
have longer duration as they queue at the per-PC).
Counter-example (the bug this rules out): if ``record_start`` fired
on command entry, all three ops would share ``t_start == 0``.
"""
from pathlib import Path
from kernbench.common.pe_commands import DmaReadCmd, PeInternalTxn, TensorHandle
from kernbench.policy.address.phyaddr import PhysAddr
from kernbench.sim_engine.engine import GraphEngine
from kernbench.topology.builder import load_topology
TOPOLOGY_PATH = Path(__file__).parent.parent / "topology.yaml"
def _hbm_pa() -> int:
slice_bytes = 48 * (1 << 30) // 8
pa = PhysAddr.pe_hbm_addr(
sip_id=0, die_id=0, pe_id=0,
pe_local_hbm_offset=0x1000, slice_size_bytes=slice_bytes,
)
return pa.encode()
# enable_data=True wires the OpLogger into every component.
engine = GraphEngine(load_topology(TOPOLOGY_PATH), enable_data=True)
pe_dma_id = "sip0.cube0.pe0.pe_dma"
pe_dma = engine._components[pe_dma_id]
env = engine._env
# Three back-to-back DMA_READ commands fed straight into pe_dma's inbox
# at t=0 so they all race for the cap=1 channel.
handles = [
TensorHandle(id=f"r{i}", addr=0x1000 + i * 0x1000,
shape=(64, 32), dtype="f16", nbytes=4096)
for i in range(3)
]
cmds = [
DmaReadCmd(handle=h, src_addr=_hbm_pa(), nbytes=4096)
for h in handles
]
txns = [PeInternalTxn(command=c, done=env.event()) for c in cmds]
def submit_all():
for txn in txns:
yield pe_dma._inbox.put(txn)
env.process(submit_all())
env.run()
# Pull the three dma_read records out of the op log in order
dma_records = [
r for r in engine.op_log
if r.op_name == "dma_read" and r.component_id == pe_dma_id
]
assert len(dma_records) == 3, (
f"expected 3 dma_read records, got {len(dma_records)}: {dma_records}"
)
durations = [r.t_end - r.t_start for r in dma_records]
base = durations[0]
assert base > 0, f"first dma duration must be positive, got {base}"
# t_start values must be monotonic and tightly clustered. The window
# over all three t_starts is bounded by the per-op channel-issue cost,
# NOT by the HBM round-trip — i.e. record_start fires post channel
# acquire AND channel release is fast.
t_starts = [r.t_start for r in dma_records]
for i in range(1, len(t_starts)):
assert t_starts[i] >= t_starts[i - 1], (
f"op {i} t_start={t_starts[i]} regressed before "
f"op {i-1} t_start={t_starts[i-1]} — record order broken"
)
t_start_window = t_starts[-1] - t_starts[0]
assert t_start_window < base, (
f"t_start window {t_start_window:.1f}ns across 3 back-to-back "
f"ops is >= a single op's duration {base:.1f}ns — PE_DMA "
f"channel is being held beyond the issue path (ADR-0014 D4)."
)
# ── 7. Back-to-back DMA pipelining (head amortization) ───────────────
def test_pe_dma_back_to_back_pipelining():
"""Back-to-back DMA_READs should pipeline so that per-tile head latency
is paid once, not N times.
Setup: 8 back-to-back 4 KiB DMA_READ commands to the same PE_DMA's
inbox at t=0. HBM_CTRL serializes via its per-PC availability
timestamps; PE_DMA should release its channel after issuing the
request, NOT hold it through the full HBM round-trip.
Expected (post-fix model):
- First op pays full round-trip: head + transmission.
- Subsequent ops pay only transmission time (head amortized): each
op's t_end advances by ~the per-request HBM service time, NOT by
the full round-trip.
- Total wall (last t_end - first t_start) is close to the
transmission-bound limit: ``head_once + N * service``, not
``N * (head + service)``.
Anti-counter-example: under the current PE_DMA round-trip-hold model
each op N+1's t_start equals op N's t_end, so the wall is exactly
``N * (head + service)``. That is what this test rules out.
Concrete numbers (32 KiB probe = 141 ns; 4 KiB single-request inferred
at ~29 ns): the current model gives wall ~232 ns for 8 ops; the
fixed model should give wall ≲ 160 ns (head ~13 + 8 * ~16 ns).
"""
from pathlib import Path
from kernbench.common.pe_commands import DmaReadCmd, PeInternalTxn, TensorHandle
from kernbench.policy.address.phyaddr import PhysAddr
from kernbench.sim_engine.engine import GraphEngine
from kernbench.topology.builder import load_topology
TOPOLOGY_PATH = Path(__file__).parent.parent / "topology.yaml"
def _hbm_pa(offset: int) -> int:
slice_bytes = 48 * (1 << 30) // 8
pa = PhysAddr.pe_hbm_addr(
sip_id=0, die_id=0, pe_id=0,
pe_local_hbm_offset=offset, slice_size_bytes=slice_bytes,
)
return pa.encode()
engine = GraphEngine(load_topology(TOPOLOGY_PATH), enable_data=True)
pe_dma_id = "sip0.cube0.pe0.pe_dma"
pe_dma = engine._components[pe_dma_id]
env = engine._env
N_OPS = 8
BYTES_PER_OP = 4096
handles = [
TensorHandle(id=f"r{i}", addr=0x1000 + i * 0x1000,
shape=(32, 64), dtype="f16", nbytes=BYTES_PER_OP)
for i in range(N_OPS)
]
# Spread across distinct cache lines so HBM_CTRL hits different PCs
# and we measure the BW-only floor, not per-PC bank conflict.
cmds = [
DmaReadCmd(handle=h, src_addr=_hbm_pa(0x1000 * (1 + i)),
nbytes=BYTES_PER_OP)
for i, h in enumerate(handles)
]
txns = [PeInternalTxn(command=c, done=env.event()) for c in cmds]
def submit_all():
for txn in txns:
yield pe_dma._inbox.put(txn)
env.process(submit_all())
env.run()
dma_records = [
r for r in engine.op_log
if r.op_name == "dma_read" and r.component_id == pe_dma_id
]
assert len(dma_records) == N_OPS, (
f"expected {N_OPS} dma_read records, got {len(dma_records)}"
)
dma_records.sort(key=lambda r: r.t_start)
single_op_ns = dma_records[0].t_end - dma_records[0].t_start
wall = dma_records[-1].t_end - dma_records[0].t_start
# Under the round-trip-hold model: wall == N_OPS * single_op_ns.
# Under the issue-only-hold model: wall ≈ single_op_ns + (N-1)*service
# where service ≈ BYTES_PER_OP / hbm_bw_per_pe ≈ 16 ns.
# We assert wall < 0.75 * (N_OPS * single_op_ns) — i.e. amortization
# of head must save at least 25% vs strict serialization.
strict_serialized = N_OPS * single_op_ns
assert wall < 0.75 * strict_serialized, (
f"wall={wall:.1f}ns vs strict-serialized={strict_serialized:.1f}ns "
f"(N_OPS={N_OPS}, single_op={single_op_ns:.1f}ns). "
f"PE_DMA is still holding channel through full round-trip — "
f"head latency is not amortized across back-to-back DMAs."
)