The performance of the code generated for the GPU heavily depends on runtime parameters such as the number of thread blocks and threads per thread block. For this reason, a proper choice of such parameters based on both the IR information and runtime information can give a tangible improvement of the performance of the generated code. In this project, we will explore the ML-based approach to this problem which uses machine learning models for making a decision on the runtime parameter values given the IR and runtime features.


Konstantin Sidorov


  • Jon Chesterfield
  • Johannes Doerfert