修复 CI benchmark 中的过时硬编码数据

- llvm_cache_compare: SCRATCHV_CONV_PER_MAC 30→12
- dashboard: conv per-MAC ~30→~12, 比值 4.29×→1.71×
- tinyfive_compare: 动态指令 7.77B→3.22B, 比值 4.2x→3.0x

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: k1nsom

scratchv/ci/dashboard.py

@@ -215,8 +215,8 @@ def generate(ld=None, td=None):
         h += f"<tr><td>Static instructions (asm)</td><td class='n'><span class='ll'>{tls['total_static']}</span></td><td class='n'>{tss.get('total_static','—')}</td><td class='n'>{_badge(_ratio(tss.get('total_static',0), tls['total_static']))}</td></tr>"
         h += f"<tr><td>x registers used</td><td class='n'><span class='ll'>{tls.get('x_reg_count','—')}</span></td><td class='n'>{tss.get('x_reg_count','—')}</td><td class='n'>{_badge(_ratio(tss.get('x_reg_count',0), tls.get('x_reg_count',0)))}</td></tr>"
     h += f"<tr><td>Instructions per MAC (kernel body)</td><td class='n'><span class='ll'>{lo.get('total',0)}</span></td><td class='n'>{so.get('total',0)}</td><td class='n'>{_badge(_ratio(so.get('total',0), lo.get('total',0)))}</td></tr>"
-    h += f"<tr><td>Instructions per MAC (full model, conv)</td><td class='n'><span class='ll'>~7</span></td><td class='n'>~30</td><td class='n'><span class='badge r'>4.29×</span></td></tr>"
-    h += f"<tr><td>Instructions per MAC (full model, FC)</td><td class='n'><span class='ll'>~5</span></td><td class='n'>~15</td><td class='n'><span class='badge y'>3.00×</span></td></tr></table>"
+    h += f"<tr><td>Instructions per MAC (full model, conv)</td><td class='n'><span class='ll'>~7</span></td><td class='n'>~12</td><td class='n'><span class='badge y'>1.71×</span></td></tr>"
+    h += f"<tr><td>Instructions per MAC (full model, FC)</td><td class='n'><span class='ll'>~5</span></td><td class='n'>~11</td><td class='n'><span class='badge y'>2.20×</span></td></tr></table>"
     h += f"""<div class="note">Kernel body: {lo.get('total',0)} vs {so.get('total',0)} instructions per MAC iteration. Full model amplification ({_ratio(so.get('total',0), lo.get('total',0))} → ~4.3×) comes from address calculation (no GEP in RV32IM → 3–5 ALU per address), Q16.16 srai shifts, and spill code.</div></div>
 
 <div class="ft">ScratchV CI &nbsp;·&nbsp; LLVM RV64FD baseline &nbsp;·&nbsp; <a href="https://github.com/ScratchV-Compiler/ScratchV" style="color:#94a3b8">GitHub</a></div>

scratchv/standalone/llvm_cache_compare.py

@@ -207,23 +207,23 @@ def compute_layer_dims() -> list[LayerDims]:
 
 # ── ScratchV Q16.16 inner loop analysis ────────────────────────────────────
 
-# ScratchV conv: ~30 instructions per MAC (Q16.16 fixed-point)
+# ScratchV conv: ~12 instructions per MAC (optimized, pointer-walking + no bounds check)
 SCRATCHV_CONV_PER_MAC = {
-    "alu_i": 8.0,
-    "alu_r": 10.0,     # mul + add + srai + addressing
-    "load": 6.0,       # lw (weight + input, multiple due to 32-bit ops)
-    "store": 2.0,      # sw for spills
-    "branch": 2.0,
-    "upper": 1.0,
-    "shift": 1.0,
+    "alu_i": 3.5,      # addi (ptr++, counter++, kh adjust)
+    "alu_r": 3.5,      # mul + add + slt
+    "load": 2.0,       # lw input + lw weight
+    "store": 0.0,      # no spills (optimized)
+    "branch": 1.0,     # bne loop
+    "upper": 0.0,      # constants preloaded in a2-a7
+    "shift": 1.0,      # srai 16
 }
 SCRATCHV_CONV_INSNS_PER_MAC = sum(SCRATCHV_CONV_PER_MAC.values())
 
 SCRATCHV_FC_PER_MAC = {
-    "alu_i": 4.0,
-    "alu_r": 5.0,
-    "load": 4.0,
-    "store": 1.0,
+    "alu_i": 3.0,
+    "alu_r": 4.0,
+    "load": 2.0,
+    "store": 0.5,
     "branch": 1.0,
 }
 SCRATCHV_FC_INSNS_PER_MAC = sum(SCRATCHV_FC_PER_MAC.values())

scratchv/standalone/tinyfive_compare.py

@@ -765,10 +765,10 @@ def generate_tinyfive_report(
     lines.append("")
     lines.append(f"  {'Metric':<30s} {'LLVM float32':>15s} {'ScratchV Q16.16':>16s}  {'Ratio':>8s}")
     lines.append(f"  {'─'*30} {'─'*15} {'─'*16}  {'─'*8}")
-    lines.append(f"  {'Insns per conv MAC':<30s} {'~7':>15s} {'~30':>16s}  {'4.3x':>8s}")
-    lines.append(f"  {'Insns per FC MAC':<30s} {'~5':>15s} {'~15':>16s}  {'3.0x':>8s}")
-    lines.append(f"  {'Total dynamic insns':<30s} {'~1.85B':>15s} {'~7.77B':>16s}  {'4.2x':>8s}")
-    lines.append(f"  {'Total memory ops':<30s} {'~0.53B':>15s} {'~2.08B':>16s}  {'3.9x':>8s}")
+    lines.append(f"  {'Insns per conv MAC':<30s} {'~7':>15s} {'~12':>16s}  {'1.7x':>8s}")
+    lines.append(f"  {'Insns per FC MAC':<30s} {'~5':>15s} {'~11':>16s}  {'2.2x':>8s}")
+    lines.append(f"  {'Total dynamic insns':<30s} {'~1.06B':>15s} {'~3.22B':>16s}  {'3.0x':>8s}")
+    lines.append(f"  {'Total memory ops':<30s} {'~0.53B':>15s} {'~0.55B':>16s}  {'1.0x':>8s}")
     lines.append(f"  {'Total FP/compute ops':<30s} {'~0.52B':>15s} {'—':>16s}  {'—':>8s}")
     lines.append("")
 
@@ -795,7 +795,7 @@ def generate_tinyfive_report(
     lines.append(f"  {'total':<12s} {lv_total:>15d}  {sv_total:>16d}  {sv_total/max(lv_total,1):>7.1f}x")
     lines.append("")
     lines.append(f"  Note: These are per-iteration static counts. The 1.2x difference")
-    lines.append(f"  grows to 4.2x in the full model due to ScratchV's additional:")
+    lines.append(f"  grows to 3.0x in the full model due to ScratchV's additional:")
     lines.append(f"    - Address calculation (no GEP → 3-5 insns per address vs 1)")
     lines.append(f"    - Spill stores (register pressure → sw/lw to stack)")
     lines.append(f"    - Q16.16 shifts (srai for fixed-point normalization)")