DeepTune is a prior work of the ProGraML's author. It proposes an automatic feature extraction method for compiler optimization. Specifically, it directly learns valuable feature vectors from the source code of input languages. It starts from source normalizing, where the source codes are re-written to prevent syntactic differences in programs from affecting the learning representation. It then puts the normalized tokens into an LSTM model whose output embeddings are concatenated with auxiliary inputs (if present). The auxiliary inputs are also normalized with BN layer for stabilizing the scale. Finally, dense DNN layers transform the output to a range of [0, 1]. DeepTune targets 2 downstream tasks: 1) device mapping; and 2) thread coarsening. In device mapping, it compares against static mapping and systems built by Grewe et al. DeepTune generally outperforms the other 2 baselines on different devices (3.43x and 1.42 respectively), though in a few of them simple baseline might work better. In thread coarsening, it outperforms the baseline by 16%.
Strength
comparable improvement with only a few NN layers;
conducted the benchmark across various hardware vendors;
Weakness
Ablation study of source normalizing is not conducted;
Should try more challenging and real-world tasks to demonstrate the effectiveness;
DeepTune
https://chriscummins.cc/pub/2017-pact.pdf
Paper Summary
DeepTune is a prior work of the ProGraML's author. It proposes an automatic feature extraction method for compiler optimization. Specifically, it directly learns valuable feature vectors from the source code of input languages. It starts from source normalizing, where the source codes are re-written to prevent syntactic differences in programs from affecting the learning representation. It then puts the normalized tokens into an LSTM model whose output embeddings are concatenated with auxiliary inputs (if present). The auxiliary inputs are also normalized with BN layer for stabilizing the scale. Finally, dense DNN layers transform the output to a range of [0, 1]. DeepTune targets 2 downstream tasks: 1) device mapping; and 2) thread coarsening. In device mapping, it compares against static mapping and systems built by Grewe et al. DeepTune generally outperforms the other 2 baselines on different devices (3.43x and 1.42 respectively), though in a few of them simple baseline might work better. In thread coarsening, it outperforms the baseline by 16%.
Strength
Weakness