refactor(hw/sim): reorganize CAM top-level tests into per-noise-config suites

Split the monolithic test_cam_basic.py into separate test suites
organized by noise configuration (no_noise, write_noise, read_noise),
with shared utilities extracted to tests/top/utils.py.

- Remove test_cam_basic.py; add no_noise/, write_noise/, read_noise/
  test suites with Makefiles that set noise parameters statically
- Extract helpers (reset_dut, write_rows, query_once, collect_topk,
  etc.) into tests/top/utils.py
- Update hw/sim/Makefile with per-config test targets and a
  test-top-all meta-target
- Update scripts/run_cam_correctness.py to build per-directory
  instead of centrally, removing inline parameter overrides
- Add __init__.py for result_serializer and topk_tracker module tests
- Expand docstrings in test_ref_model_noise.py with architectural
  context and test rationale
This commit is contained in:
2026-05-21 21:19:33 +08:00
parent 5a1d3ea977
commit 424cf6e1d3
16 changed files with 1197 additions and 712 deletions

View File

@@ -0,0 +1,274 @@
# -*- coding: utf-8 -*-
"""
CAM 写入噪声Write Noise集成测试 —— 专用配置。
本文件测试 WRITE_NOISE_EN=1, READ_NOISE_EN=0 配置下,
写入噪声模块的正确性。默认噪声率约 1%NUM=1, DEN=100
=== 测试列表 ===
1. default_noise_reproducible — 固定种子 = 确定性噪声,两次运行结果一致
2. exact_noise_model_match — RTL 存储的哈希与 ref_model.py 的 PRNG 掩码逐位匹配
3. zero_rate_noise — 写入噪声模块连接但 RATE_NUM=0 → 无翻转
4. full_rate_noise — 100% 写入噪声率,与 Python 模型对比
=== 架构背景 ===
写入噪声流水线位置Write Noise → Banked Core Storage → Read Noise → Match Engine
本测试覆盖完整的 cam_top 链路,写入噪声为唯一活跃噪声源。
=== Makefile 子目标 ===
test-zero-rate : make test-zero-rate (WRITE_NOISE_RATE_NUM=0)
test-full-rate : make test-full-rate (WRITE_NOISE_RATE_NUM=1, RATE_DEN=1, SIM_DEBUG)
"""
from __future__ import annotations
import cocotb
import numpy as np
from cocotb.clock import Clock
from cocotb.triggers import RisingEdge
from model.ref_model import (
generate_write_flip_mask,
match_top1,
random_hashes,
)
from tests.top.utils import (
collect_topk,
dut_hash_bits,
dut_num_rows,
get_param,
query_once,
query_topk_once,
reset_dut,
wait_idle,
write_row,
write_rows,
)
# ═══════════════════════════════════════════════════════════════════════════════
# 测试 1默认噪声 ~1%、可复现
# ── 固定种子 → 两次相同写入产生相同结果
# ═══════════════════════════════════════════════════════════════════════════════
@cocotb.test()
async def default_noise_reproducible(dut):
"""可复现性测试:相同种子、相同数据 → 两次独立运行结果一致。
流程:
1. 写入全部行,查询 row 50 → 记录 top1_index 和 top1_score
2. 复位
3. 再次写入相同数据,查询相同行 → 记录结果
4. 断言两次结果完全一致
如果结果不一致,说明 RTL 的 PRNG 状态没有正确复位,
或存在跨运行的状态残留。
"""
cocotb.start_soon(Clock(dut.clk, 10, unit="ns").start())
await reset_dut(dut)
num_rows = dut_num_rows(dut)
hash_bits = dut_hash_bits(dut)
rng = np.random.default_rng(42)
rows = random_hashes(rng, num_rows, width=hash_bits)
await write_rows(dut, rows)
query = rows[min(50, num_rows - 1)]
top1_index_1, top1_score_1, _ = await query_once(dut, query)
await reset_dut(dut)
await write_rows(dut, rows)
top1_index_2, top1_score_2, _ = await query_once(dut, query)
assert top1_index_1 == top1_index_2
assert top1_score_1 == top1_score_2
# ═══════════════════════════════════════════════════════════════════════════════
# 测试 2精确 RTL-vs-模型 PRNG 掩码匹配
# ── RTL 存储的哈希与 ref_model.py 生成的掩码逐位一致
# ═══════════════════════════════════════════════════════════════════════════════
@cocotb.test()
async def exact_noise_model_match(dut):
"""精确噪声模型匹配RTL 的 PRNG 输出必须与 Python 参考模型逐位一致。
测试方法:
1. 用固定 RTL seed 和已知噪声参数,在 Python 中预计算每行的 flip 掩码
2. 预期存储值 = 原始值 XOR flip_mask
3. 写入原始值到 RTL查询预期存储值
4. 断言每行 score = HASH_BITS完全匹配
这验证了 RTL 的 LFSR 实现与 Python 模型的 PRNG 使用相同的
多项式、相同的位宽、相同的种子初始化序列。
"""
if not hasattr(dut, "score_debug_flat"):
dut._log.info("Skipping exact_noise_model_match: requires SIM_DEBUG.")
return
rtol = None
atol = None
cocotb.start_soon(Clock(dut.clk, 10, unit="ns").start())
await reset_dut(dut)
hash_bits = dut_hash_bits(dut)
noise_bits = get_param(dut, "WRITE_NOISE_BITS", 8)
rate_num = get_param(dut, "WRITE_NOISE_RATE_NUM", 1)
rate_den = get_param(dut, "WRITE_NOISE_RATE_DEN", 100)
n_test_rows = 4
rng = np.random.default_rng(99)
rows = random_hashes(rng, n_test_rows, width=hash_bits)
RTL_SEED = 0xB504_F32D_B504_F32D
prng_state = (RTL_SEED << 64) | RTL_SEED
expected_stored = []
for row in rows:
flip, prng_state = generate_write_flip_mask(
prng_state,
hash_bits,
noise_bits,
rate_num,
rate_den,
)
expected_stored.append(row ^ flip)
for idx, val in enumerate(rows):
await write_row(dut, idx, val)
for idx, expected in enumerate(expected_stored):
top1_index, top1_score, score_debug = await query_once(dut, expected)
assert score_debug is not None, (
"score_debug required for mask match verification"
)
assert int(score_debug[idx]) == hash_bits, (
f"Row {idx}: expected stored hash to match model prediction, "
f"score={score_debug[idx]} != {hash_bits}"
)
# ═══════════════════════════════════════════════════════════════════════════════
# 测试 3零噪声率WRITE_NOISE_EN=1, RATE_NUM=0
# ── 噪声模块已连接但翻转概率为 0 → 行为应与无噪声一致
# ═══════════════════════════════════════════════════════════════════════════════
@cocotb.test()
async def zero_rate_noise(dut):
"""零速率噪声WRITE_NOISE_RATE_NUM=0 时不应有任何位被翻转。
这是噪声模块的边界测试——验证 RATE_NUM=0 确实禁用了翻转,
而非产生「默认速率」的噪声。
"""
rate_num = get_param(dut, "WRITE_NOISE_RATE_NUM", 1)
if rate_num != 0:
dut._log.info(
"Skipping zero_rate_noise: requires WRITE_NOISE_RATE_NUM=0."
)
return
cocotb.start_soon(Clock(dut.clk, 10, unit="ns").start())
await reset_dut(dut)
num_rows = dut_num_rows(dut)
hash_bits = dut_hash_bits(dut)
rng = np.random.default_rng(1)
rows = random_hashes(rng, num_rows, width=hash_bits)
query_index = min(123, num_rows - 1)
query = rows[query_index]
await write_rows(dut, rows)
top1_index, top1_score, score_debug = await query_once(dut, query)
expected = match_top1(query, rows, width=hash_bits)
assert top1_index == expected.top1_index
assert top1_score == expected.top1_score
assert top1_index == query_index
assert top1_score == hash_bits
if score_debug is not None:
assert np.array_equal(score_debug, expected.scores)
# ═══════════════════════════════════════════════════════════════════════════════
# 测试 4100% 噪声率RATE_NUM=1, RATE_DEN=1
# ── 每组都翻转 → 精确验证 PRNG 掩码生成
# ═══════════════════════════════════════════════════════════════════════════════
@cocotb.test()
async def full_rate_noise(dut):
"""完全速率噪声:每组 100% 翻转概率。
使用固定 RTL seed (0xB504F32DB504F32D),用 Python 模型预计算
写入全 0 和全 1 行后应存储的哈希值,然后验证 RTL 实际存储的哈希
与模型预测完全一致。
这是最低容忍度的噪声测试——要求 score_debug_flatSIM_DEBUG
且每行的分数必须精确等于 HASH_BITS。
"""
rate_num = get_param(dut, "WRITE_NOISE_RATE_NUM", 1)
rate_den = get_param(dut, "WRITE_NOISE_RATE_DEN", 100)
if rate_num != 1 or rate_den != 1:
dut._log.info(
"Skipping full_rate_noise: requires WRITE_NOISE_RATE_NUM=1, RATE_DEN=1."
)
return
if not hasattr(dut, "score_debug_flat"):
dut._log.info(
"Skipping full_rate_noise: requires SIM_DEBUG (score_debug_flat)."
)
return
cocotb.start_soon(Clock(dut.clk, 10, unit="ns").start())
await reset_dut(dut)
hash_bits = dut_hash_bits(dut)
num_rows = dut_num_rows(dut)
noise_bits = get_param(dut, "WRITE_NOISE_BITS", 8)
all_zero = 0
all_one = (1 << hash_bits) - 1
RTL_SEED = 0xB504_F32D_B504_F32D
prng_state = (RTL_SEED << 64) | RTL_SEED
flip0, prng_state = generate_write_flip_mask(
prng_state,
hash_bits,
noise_bits,
rate_num,
rate_den,
)
expected_row0 = all_zero ^ flip0
flip1, prng_state = generate_write_flip_mask(
prng_state,
hash_bits,
noise_bits,
rate_num,
rate_den,
)
expected_row1 = all_one ^ flip1
rows = [0] * num_rows
rows[0] = all_zero
rows[1] = all_one
await write_rows(dut, rows)
top1_index, top1_score, score_debug = await query_once(dut, expected_row0)
assert score_debug is not None, "score_debug required for full_rate_noise"
assert int(score_debug[0]) == hash_bits, (
f"Row 0: expected exact match, score={score_debug[0]} != {hash_bits}"
)
top1_index, top1_score, score_debug = await query_once(dut, expected_row1)
assert score_debug is not None
assert int(score_debug[1]) == hash_bits, (
f"Row 1: expected exact match, score={score_debug[1]} != {hash_bits}"
)