# PolyDL

A polyhedral model based tool to perform loop optimizations for deep learning kernels. The tools especially in the data_reuse_analyzer folder are the based on the techniques published in the publication "PolyDL: Polyhedral Optimizations for Creation of High Performance DL primitives" https://arxiv.org/abs/2006.02230 https://dl.acm.org/doi/abs/10.1145/3433103

data_reuse_analyzer/Main.cpp: Performs data reuse analysis and outputs working set sizes at L1, L2, and L3 caches. The working set sizes become the basis for ranking of code variants.

scripts/PolyRank_pessi.cpp: Performs ranking among various code variants. It considers the cache latencies and L1, L2, L3 working set sizes.

INSTALLATION
sh install_dependencies.sh
sh install.sh

apps/dnn-models: The DNN based ranking python files.

Convolution examples:
apps/experiments/submit_job.sh provides a starting point for running convolution examples through the tool.