Speaker: Mitesh Sharma
Topic: Ripple - An SPMD Framework for SIMD programming
Description: Modern HPC kernels are dominated by frameworks like CUDA, OpenCL, etc which have inadvertedly popularized SPMD programming models. However, such a model required semantics which are drastically different from ones offered in a conventional CPU. SIMD programming is one such example, where fitting in this SPMD programming framework has proven to be a challenge
Ripple tries to bridge this gap by introducing a lean API based addition into LLVM framework to allow block computations of different dimensions (scalar, vector and tensor computations) to co-exist in the same kernel. This makes it a plausible candidate to bridge the semantic gap between SIMD and SPDM programming models