PPO-based Folding Config Generator

[ollycassidy13]

PPO-based Folding Config Generator

Summary

Adds resource-aware PPO folding to automatically select parallelization settings that meet a target FPS while fitting the target FPGA.

How it works

When a board or FPGA part is provided, SetFolding:

• Builds the folding choices for each layer, including PE, SIMD, memory style, resource type, parallel window and pumped compute.
• Uses analytical cycle and resource estimates to evaluate each configuration.
• Uses PPO with deterministic refinement to find a valid configuration.
• Minimizes estimated resource usage when the throughput target is achievable.
• Selects the highest throughput configuration within the device budget when the target cannot be reached.
• Optimizes nested FINNLoop bodies using a shared device resource budget.
• Accounts for FINNLoop iterations, controller overhead, and recurring shuffle stalls in throughput estimates.
• Reports recursive and post-FIFO resource estimates.
• Compares analytical throughput against phase-balanced multi-frame RTL simulation results.

It aims to replace the first 90% of by-hand tuning, jumping users straight to the final tweaks to maximize their design.