The key to a data parallel compiler

Abstract

We present a language-driven strategy for the construction of compilers that are inherently data-parallel in their design and implementation. Using an encoding of the inter-node relationships between nodes in an AST called a Node Coordinate Matrix, we demonstrate how an operator called the Key operator, that applies a function over groupings of array cells grouped and ordered by their keys, when used in conjunction with a Node Coordinate Matrix, permits arbitrary computation over sub-trees of an AST using purely dataparallel array programming techniques. We discuss the use of these techniques when applied to an experimental commercial compiler called Co-dfns, which is a fully data-parallel compiler developed using the techniques discussed here.

Unleashing GPUs for Network Function Virtualization: an open architecture based on Vulkan and Kubernetes

“…Another approach, specific to our application, is refactoring the data structures. Here, the binary decision tree could be transformed into an indexed path matrix [24]. Such a parallel data structure would allow the GPU to also traverse the tree in parallel, avoiding the sequential while loop within the random forest tree traversal that currently limits GPUs performance in our use case.…”