adu2021 Overview
Autonomous kernel optimization across heterogeneous hardware via agentic synthesis.
Core Technique
Context-Aware Transformation:
# Single universal operator adapts via context
kernel = universal_operator.transform(
operator_spec,
hardware_context=gpu_type,
optimization_history=prior_attempts
)
Retrieval-Augmented Prompting: Hardware-specific constraints guide generation.
Performance
- 1.25-17× speedup on production workloads
- Hours vs weeks of manual optimization
References
- Context-aware kernel generation
- Hardware constraint knowledge bases
- Self-improving kernel optimization