DAC26 Vibe Coding HLS Tutorial

This repository is a curated tutorial package for showing how hardware specifications can be turned into real, tool-verified HLS implementations through a vibe-coding loop:

specification -> prompt -> LLM candidate -> correctness test -> Vitis HLS report -> accept or reject

The package contains two complete demos, retained prompts and LLM responses, candidate source code, Vitis/Vivado reports, and a curated RTL reference for comparison.

What This Tutorial Demonstrates

Demo	Goal	Main evidence	Result
Systolic array HLS	Improve throughput for a 16x16 matrix-multiply kernel	Vitis HLS latency, interval, estimated GOPS, resource report	0.0754 GOPS -> 41.6614 GOPS, 552.7x throughput improvement
RISC-V HLS step kernel	Reduce controller resources while preserving correctness	Vitis HLS LUT, FF, latency, interval, C simulation	area score 1035 -> 664, 35.8% reduction
DAC2 RTL comparison	Compare conventional RTL handoff against the HLS vibe-coding flow	Vivado OOC reports and curated RTL/docs	Keeps measurement domains explicit and numeric

The point of the demo is not to trust generated code. The point is to show a disciplined engineering loop where every LLM-generated change is checked by tests and synthesis reports before it is accepted.

Start Here

1. Follow the live demo runbook:

   docs/demo_runbook.md
   

2. Review the numeric comparison:

   comparison/DAC2_vs_Vibe_HLS_Numeric_Comparison.md
   

3. Inspect the retained prompt transcripts and measured reports:

   demos/systolic-throughput/
   demos/riscv-resource/
   

Repository Map

Path	Purpose
docs/	Demo runbook, artifact inventory, and upload checklist
demos/systolic-throughput/	Throughput-oriented HLS vibe-coding demo for the systolic array
demos/riscv-resource/	Resource-oriented HLS vibe-coding demo for the RISC-V step kernel
comparison/	DAC2 RTL vs vibe-coding HLS numeric comparison and Vivado reports
reference-models/	Behavioral C reference models, specs, tests, and golden data
rtl-reference/	Curated DAC2 RTL reference, docs, and reports
scripts/	Release-package validation tools

Reproduce the Demo

The fastest replay does not require Vitis or Ollama. It prints the frozen prompt/code/report trajectory from retained artifacts.

cd demos/systolic-throughput
python scripts/vibe_coding_demo.py --compact
python scripts/reproduce_demo.py --mode summary

cd demos/riscv-resource
python scripts/reproduce_resource_demo.py --mode summary

Full HLS reproduction requires:

• Python 3
• AMD Vitis HLS 2024.2
• target device xc7z020-clg400-1
• 10 ns HLS clock constraint

Optional fresh LLM generation requires:

• local Ollama server
• the tested local model was qwen3-coder:30b

Fresh LLM output is not guaranteed to match the frozen trajectory exactly, even at low temperature, because model/runtime versions can change. The retained prompts, responses, candidate sources, and reports are the audit trail for the demonstrated run.

What Is Intentionally Excluded

This is a GitHub review package, not a raw workspace dump. The following are intentionally excluded:

• Vitis/Vivado/OpenLane generated work directories
• presentation files, PPTX files, and speaker notes
• .discovery/, .Xil/, .codex/, .agents/, nested .git/
• zip archives and old downloaded packages
• OpenLane layout sweep intermediate files from DAC2
• Python __pycache__/ and build executables
• internal hardware-spec .docx files unless separately approved

The curated package is small enough for normal GitHub review while preserving the evidence needed to explain and audit the tutorial.

Measurement Boundaries

The systolic-array comparison and RISC-V comparison do not claim that every implementation has identical abstraction, interface, and target assumptions.

• DAC2 RTL uses a conventional RTL handoff path and Vivado out-of-context synthesis.
• The vibe-coding demos use Vitis HLS and accept candidates only after correctness and QoR checks.
• DAC2 layout numbers are ASIC/OpenLane context and should not be directly compared to FPGA LUT/FF/DSP numbers.

Use the comparison as a methodology and evidence discussion, not as a simplistic absolute ranking.