Implement ADR-0021: PE pipeline refactor with token self-routing

Step 1-2: Backup existing code
- builtin/ → builtin_legacy/ (unchanged backup)
- custom/pe_accel/ → custom/pe_accel_legacy/ (unchanged backup)

Step 3-4: New pipeline types and tiling
- pe_types.py: StageType, Stage, TilePlan, PipelinePlan, PipelineContext, TileToken
- tiling.py: generate_gemm_plan, generate_math_plan (ported from pe_accel)

Step 5: Component implementations (ADR-0021 D4-D6)
- PE_SCHEDULER: _feed_loop (singleton FIFO feeder) + plan generation
- PE_FETCH_STORE: new component — TCM ↔ Register File
- PE_GEMM: TileToken pipeline + legacy PeInternalTxn dual-mode
- PE_MATH: TileToken pipeline + legacy dual-mode
- PE_DMA: TileToken pipeline + legacy + fabric Transaction triple-mode
- PE_TCM: TcmRequest handler with dual-channel BW serialization

Step 6: Infrastructure
- topology.yaml: pe_fetch_store component + chaining edges
- components.yaml: pe_fetch_store_v1 registration
- builder.py: PE_COMP_OFFSETS, _add_pe_internal_edges, PE view positions
- Tests: node/edge counts, PE component sets updated

All components handle both TileToken (pipeline) and PeInternalTxn (legacy).
Token self-routing: components read next stage from token.plan, chain via out_port.
366 tests passing.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-04-08 23:35:31 -07:00
parent 161132cdcb
commit b6eb97c49a
40 changed files with 4055 additions and 214 deletions
+176
View File
@@ -0,0 +1,176 @@
"""Tile plan generators for PE pipeline (ADR-0021).
Generates TilePlan with stage sequences for GEMM and Math operations.
Ported from pe_accel tiling.py with stage-based plan structure.
"""
from __future__ import annotations
from math import ceil
from kernbench.components.builtin.pe_types import (
PipelinePlan,
Stage,
StageType,
TilePlan,
)
def generate_gemm_plan(
M: int, K: int, N: int,
tile_m: int, tile_k: int, tile_n: int,
bytes_per_element: int,
A_addr: int, B_addr: int, C_addr: int,
pe_prefix: str,
) -> PipelinePlan:
"""Generate GEMM tile plan: M→N→K order.
Each tile follows stage sequence:
DMA_READ(A) → DMA_READ(B) → FETCH → GEMM → STORE
On last K-tile per (m,n): → DMA_WRITE
Args:
pe_prefix: e.g. "sip0.cube0.pe0" — used to build component IDs.
"""
M_tiles = max(1, ceil(M / tile_m))
K_tiles = max(1, ceil(K / tile_k))
N_tiles = max(1, ceil(N / tile_n))
bpe = bytes_per_element
dma_id = f"{pe_prefix}.pe_dma"
fetch_id = f"{pe_prefix}.pe_fetch_store"
gemm_id = f"{pe_prefix}.pe_gemm"
# math_id = f"{pe_prefix}.pe_math" # for K-accumulation if needed
tiles: list[TilePlan] = []
tile_id = 0
for m in range(M_tiles):
for n in range(N_tiles):
c_addr = C_addr + (m * tile_m * N + n * tile_n) * bpe
for k in range(K_tiles):
last_k = k == K_tiles - 1
a_addr = A_addr + (m * tile_m * K + k * tile_k) * bpe
b_addr = B_addr + (k * tile_k * N + n * tile_n) * bpe
a_bytes = tile_m * tile_k * bpe
b_bytes = tile_k * tile_n * bpe
out_bytes = tile_m * tile_n * bpe
stages: list[Stage] = []
# DMA READ: load A and B tiles from HBM → TCM
stages.append(Stage(
stage_type=StageType.DMA_READ,
component=dma_id,
params={
"src_addr": a_addr, "nbytes": a_bytes,
"operand": "A", "tile_m": tile_m, "tile_k": tile_k,
},
))
stages.append(Stage(
stage_type=StageType.DMA_READ,
component=dma_id,
params={
"src_addr": b_addr, "nbytes": b_bytes,
"operand": "B", "tile_k": tile_k, "tile_n": tile_n,
},
))
# FETCH: TCM → Register File
stages.append(Stage(
stage_type=StageType.FETCH,
component=fetch_id,
params={
"direction": "read",
"nbytes": a_bytes + b_bytes,
},
))
# GEMM: MAC compute
stages.append(Stage(
stage_type=StageType.GEMM,
component=gemm_id,
params={
"m": tile_m, "k": tile_k, "n": tile_n,
"is_last_k": last_k,
},
))
# STORE: Register File → TCM
stages.append(Stage(
stage_type=StageType.STORE,
component=fetch_id,
params={
"direction": "write",
"nbytes": out_bytes,
},
))
# DMA WRITE: TCM → HBM (only on last K-tile)
if last_k:
stages.append(Stage(
stage_type=StageType.DMA_WRITE,
component=dma_id,
params={
"dst_addr": c_addr, "nbytes": out_bytes,
},
))
tiles.append(TilePlan(tile_id=tile_id, stages=tuple(stages)))
tile_id += 1
return PipelinePlan(
tiles=tiles, m_tiles=M_tiles, k_tiles=K_tiles, n_tiles=N_tiles,
)
def generate_math_plan(
M: int, N: int,
tile_m: int, tile_n: int,
bytes_per_element: int,
math_op: str,
src_addr: int, dst_addr: int,
pe_prefix: str,
) -> PipelinePlan:
"""Generate element-wise math tile plan.
Each tile: DMA_READ → FETCH → MATH → STORE → DMA_WRITE
"""
M_tiles = max(1, ceil(M / tile_m))
N_tiles = max(1, ceil(N / tile_n))
bpe = bytes_per_element
dma_id = f"{pe_prefix}.pe_dma"
fetch_id = f"{pe_prefix}.pe_fetch_store"
math_id = f"{pe_prefix}.pe_math"
tiles: list[TilePlan] = []
tile_id = 0
for m in range(M_tiles):
for n in range(N_tiles):
offset = (m * tile_m * N + n * tile_n) * bpe
tile_bytes = tile_m * tile_n * bpe
stages = [
Stage(StageType.DMA_READ, dma_id, {
"src_addr": src_addr + offset, "nbytes": tile_bytes,
}),
Stage(StageType.FETCH, fetch_id, {
"direction": "read", "nbytes": tile_bytes,
}),
Stage(StageType.MATH, math_id, {
"op": math_op, "num_elements": tile_m * tile_n,
}),
Stage(StageType.STORE, fetch_id, {
"direction": "write", "nbytes": tile_bytes,
}),
Stage(StageType.DMA_WRITE, dma_id, {
"dst_addr": dst_addr + offset, "nbytes": tile_bytes,
}),
]
tiles.append(TilePlan(tile_id=tile_id, stages=tuple(stages)))
tile_id += 1
return PipelinePlan(tiles=tiles, m_tiles=M_tiles, n_tiles=N_tiles)