Enabling Polyhedral optimizations in TensorFlow through Polly

Annanay Agarwal, Michael Kruse, Brian Retford, Tobias Grosser and Ramakrishna Upadrasta

Presented at US LLVM Developers’ Meeting, 2017


TensorFlow is an Open Source Deep learning Framework by Google. Here, Polly’s SCoP detection was modified to detect the convolution kernel from the LLVM IR generated for tf.conv2d() operation in XLA. The results include reduced cache misses - advanced data locality optimizations like tiling. Also, it performs operator fusion. Also, we got better overall runtime for convolution kernel despite having a greater compilation time.