pypto-optimization

What problem does it solve?

PyPTO optimization rules provide a structured approach to tuning tensor operation performance by selecting tile shapes, loop counts, and reduction-axis strategies to maximize efficiency while respecting hardware constraints. We also offer guidance on when to merge reduction axes and how to sequence experiments to find the sweet spot for throughput and latency.

Core Features & Use Cases

• Rule-based tile shaping and auto-tiling guidance for common operators (softmax, logsoftmax, norm, reduction, loss).
• Loop-count sweep strategies with middle-ground optimization to balance task granularity and scheduling overhead.
• Guidance on merging consecutive reduction axes to reduce intermediate tensors and improve kernel performance.

Quick Start

Apply the PyPTO optimization rules to your operator to choose tile shapes and loop counts, then profile the results to select the best configuration.