Software for Incremental Flattening for Nested Data Parallelism
Troels Henriksen, Frederik Thorøe, Martin Elsman, Cosmin E. Oancea · Artifact Digital Object Group · 2019
This artifact contains scripts and programs for reproducing the central experimental figures of the paper (Figures 2, 7 and 8), which show how Futhark programs perform compared to hand-written GPU programs, using different compilation techniques.This artifact requires access to a computer with a modern GPU (a laptop will likely not cut it) with OpenCL and optionally CUDA already set up and working.Apart from that, the software requirements are fairly standard.A Dockerfile is provided that will work on Linux systems with NVIDIA GPUs, assuming NVIDIAs docker runtime has been installed.This is typically not problematic.The artifact is structured around a Makefile with targets for the figures and supporting data.The enclosed README.mdcontains further details and hints about execution and troubleshooting.The artifact is also publicly available at https: //github.com/diku-dk/futhark-ppopp19.