pypto-optimization

Optimizing tile shapes and loop counts for tensor operators involves applying rule-based tile shaping and middle-ground loop_count sweeps to balance task granularity with scheduling overhead, then profiling to select the best configuration.

Hardware constraints for tensor tile shaping in PyPTO enforce deterministic exploration with limits like prod(tile_shape) <= 16384 and auto_tiles <= 2048 to maximize efficiency while respecting hardware limits.

Merge consecutive reduction axes in tensor operations to reduce intermediate tensors and improve kernel performance, particularly when handling operators with softmax, norm, and loss patterns.

The best way to sequence performance experiments for tensor operator throughput is to apply rule-based auto-tiling guidance, sweep loop counts, and merge reduction axes to find the sweet spot for latency and throughput.

Auto-tiling guidance supports common operators including softmax, logsoftmax, norm, reduction, and loss by providing structured rules for selecting tile shapes and reduction-axis strategies.

Tensor operator performance may drop after selecting tile shapes if loop counts are not optimized with middle-ground sweeps, causing imbalanced task granularity and scheduling overhead, or if reduction axes are not merged properly.