feat: add hardware retrieval cycle performance measurement

Add cycle-level performance measurement for hardware CAM retrieval benchmarks
to complement existing quality metrics.

- Add query_topk_once_with_latency with accept→first/last cycle timing
- Add QueryTiming dataclass and summarize_query_timings helper
- Integrate cycle performance into benchmark outputs (CSV + Markdown)
- Log RETRIEVAL_PERF_RESULT with cycles/query and queries/cycle
- Update experiment docs with hardware cycle performance section
- Add unit tests for summarize_query_timings and output writers

docs/exps/sw_hw_cam_retrieval_benchmark.md

@@ -1,8 +1,8 @@
 # 软件/硬件 CAM 检索基准实验总结
 
 **日期**：2026-05-27  
-**工作区**：`.workspace/feat_sw_retrieval_benchmark`  
-**目标**：对比同一组 CAM 检索数据在硬件仿真与软件汉明距离检索下的检索质量，并记录软件检索速度基线。
+**工作区**：`/home/sikongjueluo/Projects/Mini-Nav`  
+**目标**：对比同一组 CAM 检索数据在硬件仿真与软件汉明距离检索下的检索质量，并记录软件检索速度基线与硬件仿真周期性能。
 
 ## 1. 实验配置
 
@@ -37,6 +37,12 @@ just remote "python scripts/sw_retrieval_benchmark.py --dataset outputs/cam_retr
 
 硬件噪声扫描数据来自远端已有 Cocotb/Verilator 输出与 `docs/exps/cam_retrieval_noise_sweep_*.md`。
 
+硬件检索周期检测命令：
+
+```bash
+just cam-test-retrieval-artifact outputs/cam_retrieval_benchmark/datasets/cifar10_hash512_rows512_queries128.npz 512
+```
+
 ## 2. 软件检索速度与质量
 
 软件路径只计时 Top-K 匹配阶段，不包含 `.npz` 加载、指标聚合和结果写盘。
@@ -119,7 +125,51 @@ just remote "python scripts/sw_retrieval_benchmark.py --dataset outputs/cam_retr
 - 50% 噪声后检索基本不可用。
 - 与 CIFAR-10 相同，Golden Match@5 在低噪声下也迅速接近 0，说明精确排序比类别命中率更脆弱。
 
-## 5. 指标解释
+## 5. 硬件检索周期性能
+
+本次已将硬件周期检测合并进真实 `.npz` 检索数据集的 Cocotb/Verilator
+benchmark。实现位置：
+
+- `hw/sim/tests/top/utils.py::query_topk_once_with_latency`
+- `hw/sim/benchmarks/retrieval/test_retrieval_benchmark.py::cam_retrieval_benchmark`
+
+周期口径如下：
+
+| 指标 | 含义 |
+|---|---|
+| `accept_to_first_result_cycles` | query 被 `query_valid && query_ready` 接受后，到首个 `result_valid` beat 的周期数。 |
+| `accept_to_last_result_cycles` | query 被接受后，到 `result_last` 断言的周期数，即完整 Top-K 串行结果输出完成。 |
+| `total_query_cycles` | 从拉高 `query_valid` 到 Top-K 输出完成的完整事务周期数。 |
+| `cycles_per_query` | 当前报告中的主指标，等于平均 `accept_to_last_result_cycles`。 |
+| `queries_per_cycle` | `num_queries / total_query_cycles`，用于观察仿真事务吞吐率。 |
+
+选择 `accept_to_last_result_cycles` 作为主 `cycles_per_query` 的原因是：Top-K
+检索只有在串行结果流输出到 `result_last` 后才算完整完成；仅用首个
+`result_valid` 会低估实际 Top-K 检索事务成本。
+
+### CIFAR-10 无噪声硬件周期结果
+
+配置：512 rows × 128 queries × 512-bit hash，`TOPK_K=5`，`LANES=8`，
+`WRITE_NOISE_EN=0`。
+
+| 数据集 | 模式 | 查询数 | accept→first | accept→last / cycles/query | total query cycles | queries/cycle | 状态 |
+|---|---|---:|---:|---:|---:|---:|---|
+| CIFAR-10 | hardware no-noise | 128 | 1027.000000 | 1031.000000 | 132096 | 0.000968992 | pass |
+
+对应日志标记：
+
+```text
+RETRIEVAL_PERF_RESULT mode=no_noise num_queries=128 cycles_per_query=1031.000000 accept_to_first_result_cycles=1027.000000 accept_to_last_result_cycles=1031.000000 total_query_cycles=132096 queries_per_cycle=0.000968992 status=pass
+```
+
+该结果说明：在当前 `NUM_ROWS=512, LANES=8, TOPK_K=5` 的硬件仿真配置下，
+一次完整 Top-K 检索事务约为 **1031 cycles/query**。首个结果 beat 约在
+1027 cycles 后出现，完整 Top-K 输出额外消耗约 4 个周期。
+
+> 注：以上数据来自 Verilator/Cocotb 仿真，不是 FPGA 板上实测。它可用于
+> 架构级周期趋势分析，但不能直接等同于板级频率、吞吐或端到端系统延迟。
+
+## 6. 指标解释
 
 | 指标 | 含义 |
 |---|---|
@@ -130,8 +180,10 @@ just remote "python scripts/sw_retrieval_benchmark.py --dataset outputs/cam_retr
 | Golden Match@K / `exact_match_rate` | Top-K row index 列表是否与参考模型完全一致。该指标比 Hit@K 更严格。 |
 | `ns_per_query` | 软件 Top-K 匹配阶段平均耗时；不含加载和写盘。 |
 | `queries_per_second` | 软件 Top-K 匹配阶段吞吐率。 |
+| `cycles_per_query` | 硬件仿真中一次完整 Top-K 检索事务的平均周期数，当前采用 accept→last 口径。 |
+| `queries_per_cycle` | 硬件仿真中完成 query 数除以总 query 事务周期数。 |
 
-## 6. 当前结论
+## 7. 当前结论
 
 1. **软件汉明距离基线已可作为硬件 CAM 检索的功能参考。**  
    在无噪声条件下，硬件仿真和软件基线的质量指标及 Golden Match 均一致。
@@ -139,8 +191,9 @@ just remote "python scripts/sw_retrieval_benchmark.py --dataset outputs/cam_retr
 2. **当前软件基线速度约为 4.86k queries/s。**  
    该结果来自 Python integer brute-force Hamming scan，数据规模为 512 rows × 128 queries × 512 bits。
 
-3. **硬件检索质量基准目前主要报告质量指标，不报告 cycles/query。**  
-   现有硬件性能基准中已有 cycle 统计逻辑，但尚未和真实 retrieval dataset 的 Top-K benchmark 合并。因此本文不报告硬件 query/s 或 cycles/query。
+3. **硬件检索质量基准现在已经报告 cycles/query。**  
+   在 CIFAR-10 512-row/128-query 无噪声配置下，完整 Top-K 检索为
+   **1031 cycles/query**，首个结果为 **1027 cycles/query**。
 
 4. **写噪声对精确排序影响显著。**  
    即使 Hit@K 保持较高，Golden Match@K 也会快速下降，说明噪声首先破坏精确排序，再进一步破坏类别命中。
@@ -148,13 +201,13 @@ just remote "python scripts/sw_retrieval_benchmark.py --dataset outputs/cam_retr
 5. **CIFAR-100 比 CIFAR-10 更能体现检索难度。**  
    在无噪声下 CIFAR-100 的 Hit@1/Hit@5 分别为 0.695312/0.867188，明显低于 CIFAR-10 的 1.0/1.0，更适合作为后续检索质量对比主数据集。
 
-## 7. 后续建议
+## 8. 后续建议
 
-1. 将 `hw/sim/tests/perf/test_cam_perf.py::query_once_with_latency` 的周期测量逻辑合并到 retrieval benchmark，记录真实数据集上的 `cycles/query`。
+1. 对 CIFAR-100 也运行同样的硬件周期检测，补齐与软件质量表同尺度的硬件周期表。
 2. 在 `docs/exps` 中继续维护：
    - 软件检索速度表；
    - 硬件无噪声一致性表；
    - 硬件噪声鲁棒性表；
-   - 后续硬件 `cycles/query` 表。
+   - 硬件 `cycles/query` 表。
 3. 对软件基线补充 NumPy/PyTorch vectorized Hamming scan，以区分“朴素 Python baseline”和“优化软件 baseline”。
 4. 增加 `NUM_ROWS` sweep：例如 512、1024、2048、4096 rows，观察软件 brute-force scan 的线性增长趋势。

hw/sim/benchmarks/retrieval/test_retrieval_benchmark.py

@@ -20,7 +20,7 @@ from tests.top.utils import (
     dut_lanes,
     dut_num_rows,
     get_param,
-    query_topk_once,
+    query_topk_once_with_latency,
     reset_dut,
     write_rows,
 )
@@ -46,6 +46,13 @@ class RetrievalDataset:
     seed: int
 
 
+@dataclass(frozen=True)
+class QueryTiming:
+    accept_to_first_result_cycles: int
+    accept_to_last_result_cycles: int
+    total_query_cycles: int
+
+
 @dataclass(frozen=True)
 class MetricAccumulator:
     precision_sum: float = 0.0
@@ -203,6 +210,38 @@ def mode_from_params(write_noise_en: int) -> str:
     return "no_noise"
 
 
+def summarize_query_timings(timings: list[QueryTiming]) -> dict[str, float]:
+    if not timings:
+        return {
+            "num_queries": 0,
+            "total_query_cycles": 0,
+            "mean_total_query_cycles": 0.0,
+            "min_total_query_cycles": 0,
+            "max_total_query_cycles": 0,
+            "mean_accept_to_first_result_cycles": 0.0,
+            "mean_accept_to_last_result_cycles": 0.0,
+            "cycles_per_query": 0.0,
+            "queries_per_cycle": 0.0,
+        }
+
+    total_cycles = sum(t.total_query_cycles for t in timings)
+    total_first = sum(t.accept_to_first_result_cycles for t in timings)
+    total_last = sum(t.accept_to_last_result_cycles for t in timings)
+    count = len(timings)
+    mean_last = total_last / float(count)
+    return {
+        "num_queries": count,
+        "total_query_cycles": total_cycles,
+        "mean_total_query_cycles": total_cycles / float(count),
+        "min_total_query_cycles": min(t.total_query_cycles for t in timings),
+        "max_total_query_cycles": max(t.total_query_cycles for t in timings),
+        "mean_accept_to_first_result_cycles": total_first / float(count),
+        "mean_accept_to_last_result_cycles": mean_last,
+        "cycles_per_query": mean_last,
+        "queries_per_cycle": count / float(total_cycles) if total_cycles > 0 else 0.0,
+    }
+
+
 def output_dir_for(mode: str) -> Path:
     run_id = os.environ.get("CAM_RETRIEVAL_RUN_ID")
     if not run_id:
@@ -225,7 +264,9 @@ def write_outputs(out_dir: Path, result: dict) -> None:
         "write_noise_en", "write_noise_rate_num",
         "write_noise_rate_den",
         "num_queries", "k", "macro_precision", "retrieval_recall", "macro_f1",
-        "recall@k", "exact_match_rate", "status",
+        "recall@k", "exact_match_rate", "cycles_per_query",
+        "mean_accept_to_first_result_cycles", "mean_accept_to_last_result_cycles",
+        "mean_total_query_cycles", "total_query_cycles", "queries_per_cycle", "status",
     ]
     with metrics_csv.open("w", newline="", encoding="utf-8") as f:
         writer = csv.DictWriter(f, fieldnames=fieldnames)
@@ -248,6 +289,16 @@ def write_outputs(out_dir: Path, result: dict) -> None:
                 "macro_f1": metrics["macro_f1"],
                 "recall@k": metrics["recall@k"],
                 "exact_match_rate": metrics["exact_match_rate"],
+                "cycles_per_query": result.get("performance", {}).get("cycles_per_query", ""),
+                "mean_accept_to_first_result_cycles": result.get("performance", {}).get(
+                    "mean_accept_to_first_result_cycles", "",
+                ),
+                "mean_accept_to_last_result_cycles": result.get("performance", {}).get(
+                    "mean_accept_to_last_result_cycles", "",
+                ),
+                "mean_total_query_cycles": result.get("performance", {}).get("mean_total_query_cycles", ""),
+                "total_query_cycles": result.get("performance", {}).get("total_query_cycles", ""),
+                "queries_per_cycle": result.get("performance", {}).get("queries_per_cycle", ""),
                 "status": result["status"],
             }
             writer.writerow(row)
@@ -260,6 +311,16 @@ def write_outputs(out_dir: Path, result: dict) -> None:
         f"- status: `{result['status']}`",
         f"- num_queries: `{result['dataset']['num_queries']}`",
         "",
+        "## Hardware performance",
+        "",
+        f"- cycles_per_query: `{result.get('performance', {}).get('cycles_per_query', '')}`",
+        f"- accept_to_first_result_cycles: `{result.get('performance', {}).get('mean_accept_to_first_result_cycles', '')}`",
+        f"- accept_to_last_result_cycles: `{result.get('performance', {}).get('mean_accept_to_last_result_cycles', '')}`",
+        f"- total_query_cycles: `{result.get('performance', {}).get('total_query_cycles', '')}`",
+        f"- queries_per_cycle: `{result.get('performance', {}).get('queries_per_cycle', '')}`",
+        "",
+        "## Retrieval quality",
+        "",
         "| k | macro_precision | retrieval_recall | macro_f1 | recall@k | exact_match_rate |",
         "|---:|---:|---:|---:|---:|---:|",
     ]
@@ -300,9 +361,11 @@ async def cam_retrieval_benchmark(dut):
     await write_rows(dut, dataset.rows)
 
     accumulators = {k: MetricAccumulator() for k in BENCHMARK_KS}
+    timings: list[QueryTiming] = []
 
     for query, query_label in zip(dataset.queries, dataset.query_labels):
-        beats, _, _, _ = await query_topk_once(dut, query)
+        beats, _, _, _, timing = await query_topk_once_with_latency(dut, query)
+        timings.append(QueryTiming(**timing))
         if len(beats) < max(BENCHMARK_KS):
             raise AssertionError(f"Expected at least {max(BENCHMARK_KS)} Top-K beats, got {len(beats)}")
 
@@ -339,6 +402,7 @@ async def cam_retrieval_benchmark(dut):
             "seed": dataset.seed,
         },
         "metrics": {str(k): accumulators[k].as_dict() for k in BENCHMARK_KS},
+        "performance": summarize_query_timings(timings),
     }
 
     out_dir = output_dir_for(mode)
@@ -353,6 +417,21 @@ async def cam_retrieval_benchmark(dut):
             str(out_dir.relative_to(_project_root())),
         )
 
+    performance = result["performance"]
+    dut._log.info(
+        "RETRIEVAL_PERF_RESULT mode=%s num_queries=%d cycles_per_query=%.6f "
+        "accept_to_first_result_cycles=%.6f accept_to_last_result_cycles=%.6f "
+        "total_query_cycles=%d queries_per_cycle=%.9f status=pass output_dir=%s",
+        mode,
+        performance["num_queries"],
+        performance["cycles_per_query"],
+        performance["mean_accept_to_first_result_cycles"],
+        performance["mean_accept_to_last_result_cycles"],
+        performance["total_query_cycles"],
+        performance["queries_per_cycle"],
+        str(out_dir.relative_to(_project_root())),
+    )
+
     if write_noise_en == 0:
         assert result["metrics"]["5"]["exact_match_rate"] == 1.0, (
             f"Expected perfect exact match with no noise, got "

hw/sim/tests/top/utils.py

@@ -15,7 +15,7 @@ CAM 顶层测试的共享辅助函数。
 from __future__ import annotations
 
 import numpy as np
-from cocotb.triggers import RisingEdge
+from cocotb.triggers import ReadOnly, RisingEdge, Timer
 from model.ref_model import (  # noqa: E402
     match_top1,
     unpack_score_debug_flat,
@@ -202,28 +202,73 @@ async def query_topk_once(dut, query, timeout_cycles=None):
       - beats: [(rank, row, score, last), ...]
       - score_debug: np.ndarray 或 None（SIM_DEBUG 模式）
     """
+    beats, top1_index, top1_score, score_debug, _ = await query_topk_once_with_latency(
+        dut, query, timeout_cycles=timeout_cycles,
+    )
+    return beats, top1_index, top1_score, score_debug
+
+
+async def query_topk_once_with_latency(dut, query, timeout_cycles=None):
+    """发起一次查询、收集完整 Top-K 结果流，并返回周期计数。
+
+    返回：(beats, top1_index, top1_score, score_debug, timing)
+
+    ``timing`` 字段：
+      - accept_to_first_result_cycles: query 接受到首个 result_valid beat
+      - accept_to_last_result_cycles: query 接受到 result_last beat（Top-K 完成）
+      - total_query_cycles: 从拉高 query_valid 到 Top-K 完成的总上升沿数
+
+    ``query_ready`` 是组合信号，握手周期在上升沿前采样；结果信号在
+    ReadOnly settled phase 采样，避免重新引入 query_ready 采样时序问题。
+    """
     await wait_idle(dut)
 
     dut.query_hash.value = int(query)
 
     # 等待 query_ready 为高（DUT 已就绪），避免组合逻辑下降沿导致的
-    # valid&&ready 握手丢失问题
+    # valid&&ready 握手丢失问题。
     while not int(dut.query_ready.value):
         await RisingEdge(dut.clk)
 
-    # assert query_valid 覆盖一个上升沿完成握手
+    edge_count = 0
     dut.query_valid.value = 1
+    dut.result_ready.value = 1
+
     await RisingEdge(dut.clk)
+    edge_count += 1
+    q_ready = int(dut.query_ready.value)
+    assert q_ready, "Query accept handshake was missed despite query_ready pre-wait"
+    accept_edge = edge_count
     dut.query_valid.value = 0
 
-    # 若调用者未指定超时，根据 DUT 参数动态估算
     if timeout_cycles is None:
         timeout_cycles = dut_query_timeout_cycles(dut)
 
-    # 消费完整串行结果流
-    beats = await collect_topk(dut, timeout_cycles=timeout_cycles)
+    beats = []
+    first_result_edge = None
+    last_result_edge = None
+
+    for _ in range(timeout_cycles):
+        await RisingEdge(dut.clk)
+        edge_count += 1
+        await ReadOnly()
+        if int(dut.result_valid.value):
+            if first_result_edge is None:
+                first_result_edge = edge_count
+            rank = int(dut.result_rank.value)
+            row = int(dut.result_row.value)
+            score = int(dut.result_score.value)
+            last = int(dut.result_last.value)
+            beats.append((rank, row, score, last))
+            if last:
+                last_result_edge = edge_count
+                await Timer(1, units="step")
+                break
+        await Timer(1, units="step")
+
+    if first_result_edge is None or last_result_edge is None:
+        raise AssertionError("Top-K result stream did not finish")
 
-    # score_debug 在查询完成后可用（需 SIM_DEBUG 编译）
     num_rows = dut_num_rows(dut)
     score_bits = dut_score_bits(dut)
     score_debug = None
@@ -234,7 +279,12 @@ async def query_topk_once(dut, query, timeout_cycles=None):
             score_bits,
         )
 
-    return beats, beats[0][1], beats[0][2], score_debug
+    timing = {
+        "accept_to_first_result_cycles": int(first_result_edge - accept_edge),
+        "accept_to_last_result_cycles": int(last_result_edge - accept_edge),
+        "total_query_cycles": int(edge_count),
+    }
+    return beats, beats[0][1], beats[0][2], score_debug, timing
 
 
 async def query_once(dut, query, timeout_cycles=None):

tests/test_retrieval_benchmark_performance.py

@@ -0,0 +1,102 @@
+from __future__ import annotations
+
+import csv
+import sys
+from pathlib import Path
+
+
+HW_SIM_DIR = Path(__file__).resolve().parents[1] / "hw" / "sim"
+if str(HW_SIM_DIR) not in sys.path:
+    sys.path.insert(0, str(HW_SIM_DIR))
+
+from benchmarks.retrieval.test_retrieval_benchmark import (  # noqa: E402
+    QueryTiming,
+    summarize_query_timings,
+    write_outputs,
+)
+
+
+def test_summarize_query_timings_reports_topk_completion_headline() -> None:
+    summary = summarize_query_timings([
+        QueryTiming(
+            accept_to_first_result_cycles=10,
+            accept_to_last_result_cycles=14,
+            total_query_cycles=16,
+        ),
+        QueryTiming(
+            accept_to_first_result_cycles=12,
+            accept_to_last_result_cycles=18,
+            total_query_cycles=21,
+        ),
+    ])
+
+    assert summary == {
+        "num_queries": 2,
+        "total_query_cycles": 37,
+        "mean_total_query_cycles": 18.5,
+        "min_total_query_cycles": 16,
+        "max_total_query_cycles": 21,
+        "mean_accept_to_first_result_cycles": 11.0,
+        "mean_accept_to_last_result_cycles": 16.0,
+        "cycles_per_query": 16.0,
+        "queries_per_cycle": 2 / 37,
+    }
+
+
+def test_write_outputs_includes_hardware_performance_fields(tmp_path: Path) -> None:
+    result = {
+        "run_id": "test-run",
+        "mode": "no_noise",
+        "status": "pass",
+        "params": {
+            "num_rows": 512,
+            "hash_bits": 512,
+            "lanes": 8,
+            "topk_k": 5,
+            "write_noise_en": 0,
+            "write_noise_rate_num": 0,
+            "write_noise_rate_den": 100,
+        },
+        "dataset": {
+            "num_classes": 10,
+            "positives_per_class": 0,
+            "queries_per_class": 0,
+            "num_queries": 2,
+            "seed": 0,
+        },
+        "metrics": {
+            "1": {
+                "macro_precision": 1.0,
+                "retrieval_recall": 0.5,
+                "macro_f1": 2 / 3,
+                "recall@k": 1.0,
+                "exact_match_rate": 1.0,
+            }
+        },
+        "performance": {
+            "num_queries": 2,
+            "total_query_cycles": 37,
+            "mean_total_query_cycles": 18.5,
+            "min_total_query_cycles": 16,
+            "max_total_query_cycles": 21,
+            "mean_accept_to_first_result_cycles": 11.0,
+            "mean_accept_to_last_result_cycles": 16.0,
+            "cycles_per_query": 16.0,
+            "queries_per_cycle": 2 / 37,
+        },
+    }
+
+    write_outputs(tmp_path, result)
+
+    with (tmp_path / "metrics.csv").open(newline="", encoding="utf-8") as f:
+        row = next(csv.DictReader(f))
+
+    assert row["cycles_per_query"] == "16.0"
+    assert row["mean_accept_to_first_result_cycles"] == "11.0"
+    assert row["mean_accept_to_last_result_cycles"] == "16.0"
+    assert row["queries_per_cycle"] == str(2 / 37)
+
+    summary = (tmp_path / "summary.md").read_text(encoding="utf-8")
+    assert "## Hardware performance" in summary
+    assert "cycles_per_query" in summary
+    assert "accept_to_last_result_cycles" in summary