Skip to content

Enabling of NVidia PTX backend for SYCL & nbnxmKernel performance optimizations

Anton Gorshkov requested to merge anton.gorshkov/gromacs:master into master

Enabled NVidia PTX backend for DPC++, suggested a couple of performance optimizations for nbnxmKernel (mostly for NVidia backend). Performance improvements:

  • RTX2070, benchMEM, kernel - 3.8x, whole bench - 10%;
  • RTX2070, water/0048, kernel - 3.8x, whole bench - 80%;
  • A100, benchMEM, kernel - 2.1x, whole bench - 2%;
  • A100, water/0048, kernel - 2x, whole bench - 19%;
  • GEN9 - just minor improvements (going to work on it later);

Merge request reports