AutoFDO / PGO for Fusion Compiler & signoff runtime — comparison + white paper

[sethupathib]

Summary

Adds autofdo-fc-runtime/: a numbers-backed, honesty-gated project for the theme CPU · AutoFDO · up to ~10% kernel latency, mapped to reducing Fusion Compiler / signoff wall time.

Framing (important)

• The published ~10% figure is Linux kernel AutoFDO latency (Neper tcp_rr literature) — not “FC got 10% faster.”
• You cannot AutoFDO Synopsys FC without the vendor.
• Levers: (1) kernel AutoFDO on farm hosts, (2) PGO of owned CAD tools, (3) vendor AE for closed binaries.

What’s in the folder

• src/signoff_proxy.cpp — PD-shaped proxy + GCC PGO harness
• src/dispatch_hot.cpp — mechanism (skewed switch) microbench
• farm/KERNEL_AUTOFDO.md + farm/FC_AB_MEASURE.md — farm cookbook / A/B template
• WHITEAPER.md + PDF, DEFEND_QA.md, mechanism GIF + figures
• results/SUMMARY.txt + CLAIM_GATE.txt from measured runs

Measured on authoring host

• User-space GCC PGO on proxy: ~neutral (CITEABLE=yes_proxy_measured_neutral) — codegen changed (size), wall time did not improve on this cloud CPU
• Cite the literature 10% for kernel; re-run harness on laptop/farm for local numbers

Reproduce

cd autofdo-fc-runtime
sudo apt-get install -y g++
make baseline pgo && ./examples/compare_all.sh
cat results/SUMMARY.txt

Three levers
Literature 10%
autofdo_fc_pipeline.gif

Test plan

• make baseline pgo + ./examples/compare_all.sh — MATCH correctness, CLAIM_GATE written
• make mechanism + ./examples/compare_mechanism.sh
• Figures, GIF, PDF built
• No proprietary FC binary committed

_{To show artifacts inline, enable in settings.}