NeuroGenesis: A Self-Optimizing Compiler with Transformer IR Encoding and Reinforcement Learning

Dhadi Sai Praneeth Reddy

Previously ATLAS-R-003

PreprintPreprintOpen Access

NeuroGenesis: A Self-Optimizing Compiler with Transformer IR Encoding and Reinforcement Learning

Dhadi Sai Praneeth Reddy¹

¹ Atlas AI Labs

01

Brief

Problem: Compiler optimization passes are hand-written heuristics. Learned approaches such as MLGO and AutoPhase improve on them, but still choose among predefined passes.
Approach: NeuroGenesis encodes compiler IR with a Transformer and trains a reinforcement-learning policy to discover optimization rules autonomously.
Contribution: A 0.28 reduction in execution cost relative to the LLVM baseline (0.00), outperforming MLGO and AutoPhase.

02

Status

Preprint

Idea
Experiments
Drafting
Review
Released (current stage)

03

Abstract

Compiler optimization passes are hand-written heuristics. Learned approaches such as MLGO and AutoPhase improve on them, but still choose among predefined passes. NeuroGenesis encodes compiler IR with a Transformer and trains a reinforcement-learning policy to discover optimization rules autonomously. A 0.28 reduction in execution cost relative to the LLVM baseline (0.00), outperforming MLGO and AutoPhase.

04

Identifiers

05

Key results

0.28
+0.28 vs LLVM
Execution-cost reduction
Relative to the LLVM baseline (0.00).
0.00
LLVM baseline
Reference point for the reduction.
2
Learned optimizers outperformed
MLGO and AutoPhase.

Method	Execution-cost reduction
LLVM (baseline)	0.00
NeuroGenesis	0.28

Table 1. Execution-cost reduction relative to the LLVM baseline (higher is better).

Execution-cost reduction vs. LLVM

Chart 1. Execution-cost reduction vs. LLVM

View as table

LLVM baseline	0
NeuroGenesis	0.28

Methodology and caveats

Reported as reduction in execution cost against the LLVM baseline. Benchmark suite, hardware and run count to be added. MLGO and AutoPhase values are in the paper and will be added to the table once published here.

06

Write-up

Approach

NeuroGenesis treats optimization as something the compiler can learn rather than something engineers must hand-write.

Program IR
Transformer encoder
RL policy
Optimization rules
Optimized program

NeuroGenesis pipeline: IR is encoded by a Transformer, an RL policy proposes optimization rules, and the result is measured by execution cost.

Results

Measured as reduction in execution cost against the LLVM baseline, NeuroGenesis achieves 0.28 (baseline: 0.00) and outperforms the learned-optimizer baselines MLGO and AutoPhase. See Key results above.

07

Cite this work

Reddy, D. S. P. (n.d.). NeuroGenesis: A Self-Optimizing Compiler with Transformer IR Encoding and Reinforcement Learning [Preprint]. Atlas AI Labs. https://dspraneeth07.in/research/neurogenesis

.bib .ris