fix: update benchmark scripts for DU-based metrics

Benchmark output changed from per-instruction to per-dispatch-unit:
  ns_per_insn → ns_per_du, handler_ns → delta_ns,
  insn_count → du_count, ips → du_per_sec,
  --insns → --dus

Scripts accept both old and new field names for backward compatibility.

Author: scc-tw

.github/workflows/benchmark.yml

@@ -102,7 +102,7 @@ jobs:
       # ── Run benchmark ──
       - name: Run benchmark
         shell: bash
-        run: ${{ matrix.bin }} --iterations=11 --insns=500 --policy=debug > bench_raw.json
+        run: ${{ matrix.bin }} --iterations=11 --dus=125 --policy=debug > bench_raw.json
 
       # ── Upload raw result ──
       - name: Upload raw JSON

scripts/bench_compare.py

@@ -46,8 +46,8 @@ def render_markdown(current: dict, baseline: dict | None, platform: str) -> str:
 
     lines.append(f"### {platform}")
     lines.append(f"")
-    lines.append(f"| Opcode | Cat | ns/insn | handler ns | delta |")
-    lines.append(f"|--------|-----|--------:|-----------:|------:|")
+    lines.append(f"| Opcode | Cat | ns/DU | delta ns | vs base |")
+    lines.append(f"|--------|-----|------:|--------:|--------:|")
 
     # Group by category
     from collections import defaultdict
@@ -57,17 +57,18 @@ def render_markdown(current: dict, baseline: dict | None, platform: str) -> str:
 
     for cat_id in sorted(cats.keys()):
         for name, r in cats[cat_id]:
-            ns = r["ns_per_insn"]
-            hns = r["handler_ns"]
+            ns = r.get("ns_per_du", r.get("ns_per_insn", 0))
+            hns = r.get("delta_ns", r.get("handler_ns", 0))
             d = ""
             if name in base:
-                d = delta_str(ns, base[name]["ns_per_insn"])
+                base_ns = base[name].get("ns_per_du", base[name].get("ns_per_insn", 0))
+                d = delta_str(ns, base_ns)
             cat = CAT_NAMES.get(cat_id, "?")
             lines.append(f"| `{name}` | {cat} | {ns:.1f} | {hns:+.1f} | {d} |")
 
-    bl = current.get("baseline_ns_per_insn", 0)
+    bl = current.get("baseline_ns_per_du", current.get("baseline_ns_per_insn", 0))
     lines.append(f"")
-    lines.append(f"_Pipeline baseline (NOP): {bl:.1f} ns/insn "
+    lines.append(f"_Pipeline baseline (NOP): {bl:.1f} ns/DU "
                  f"| Policy: {meta.get('policy','?')} "
                  f"| Build: {meta.get('build_type','?')}_")
     return "\n".join(lines)

scripts/bench_to_chart.py

@@ -15,34 +15,39 @@
 def convert(raw: dict) -> list:
     """Convert our benchmark JSON to github-action-benchmark format."""
     entries = []
-    baseline = raw.get("baseline_ns_per_insn", 0)
+    baseline = raw.get("baseline_ns_per_du",
+                       raw.get("baseline_ns_per_insn", 0))
 
     for r in raw.get("results", []):
         name = r["opcode"]
 
-        # ns_per_insn — total pipeline cost (smaller is better)
+        # ns_per_du — total pipeline cost per dispatch unit (smaller is better)
+        ns_per = r.get("ns_per_du", r.get("ns_per_insn", 0))
+        du_count = r.get("du_count", r.get("insn_count", 0))
         entries.append({
             "name": f"{name} (total)",
-            "unit": "ns/insn",
-            "value": round(r["ns_per_insn"], 2),
+            "unit": "ns/DU",
+            "value": round(ns_per, 2),
             "extra": f"median={r['median_ns']}ns  p95={r['p95_ns']}ns  "
                      f"stddev={r['stddev_ns']:.1f}ns  "
-                     f"iterations={r['iterations']}  insns={r['insn_count']}",
+                     f"iterations={r['iterations']}  DUs={du_count}",
         })
 
-        # handler_ns — handler overhead above baseline (smaller is better)
+        # delta_ns — handler overhead above baseline (smaller is better)
+        delta = r.get("delta_ns", r.get("handler_ns", ns_per - baseline))
         entries.append({
             "name": f"{name} (handler)",
             "unit": "ns",
-            "value": round(r.get("handler_ns", r["ns_per_insn"] - baseline), 2),
+            "value": round(delta, 2),
         })
 
-        # IPS — throughput (bigger is better, separate chart group)
-        if r.get("ips", 0) > 0:
+        # du_per_sec — throughput (bigger is better, separate chart group)
+        throughput = r.get("du_per_sec", r.get("ips", 0))
+        if throughput > 0:
             entries.append({
-                "name": f"{name} (IPS)",
-                "unit": "insn/s",
-                "value": round(r["ips"]),
+                "name": f"{name} (DU/s)",
+                "unit": "DU/s",
+                "value": round(throughput),
             })
 
     return entries