Efficient CUDA schemes

We now explain the principles behind the tiled Map-Reduce CUDA schemes that lie at the heart of the KeOps engine.

Note that this section is extracted from the Chapter 2.2 of Jean Feydy’s PhD thesis.

Efficient CUDA schemes

• A blind spot in the literature
• KeOps Lazy tensors
  • Our focus: exact Map-Reduce computations
  • A generic framework
  • The LazyTensor abstraction
• Structure of the repository
• Map-Reduce schemes
  • The 1D algorithm
  • Performances
  • The 2D algorithm
• Block-sparsity
  • Can we do better?
  • Block-sparse reductions
  • A decent trade-off