timing-guide.md

Timing Benchmark Guide

Overview

Timing simulation collects simulated cycle counts for benchmarks, comparing against native M2 execution to measure accuracy.

Target: <20% average error (per issue #141)

Available Benchmarks

Benchmark	Instructions	Status
aha-mont64	1.88M	Ready
crc32	1.57M	Ready
matmult-int	3.85M	Ready
primecount	2.84M	Ready
edn	~1M	Ready

Running Timing Simulation

Quick Test (single benchmark)

cd ~/dev/src/github.com/sarchlab/m2sim
go test -v -run "TestBenchmark.*ahamont64" ./benchmarks/ -timeout 10m

Batch Run (all benchmarks)

./scripts/batch-timing.sh 2>&1 | tee reports/timing-output.log

Warning: Timing simulation is slow. Each benchmark may take 5-10+ minutes. Run overnight or in a dedicated session.

Recommended: Background/Overnight Run

nohup ./scripts/batch-timing.sh > reports/timing-$(date +%Y-%m-%d).log 2>&1 &
echo $! > reports/timing.pid

Check progress:

tail -f reports/timing-$(date +%Y-%m-%d).log

Results Location

• Raw output: reports/embench-timing-YYYY-MM-DD.json
• Logs: reports/timing-YYYY-MM-DD.log

Next Steps After Timing

1. Extract simulated cycles from timing output
2. Compare against native M2 results (if available)
3. Calculate error percentages
4. Tune pipeline parameters (issue width, branch penalty)
5. Re-run timing to verify improvements

Current Accuracy Baseline (Microbenchmarks)

Benchmark	Error
arithmetic	49.3%
dependency	18.9%
branch	51.3%
Average	39.8%

Key tuning targets:

• Arithmetic: 8-wide issue + instruction fusion
• Branch: Reduce misprediction penalty