Artifact for Compiling Generalized Histograms for GPU

The experiments in the paper comprise various GPU programs that generally perform their own internal timing. In principle, reproduction should be as easy as compiling and running them. Most of the GPU code is written in ordinary CUDA, but some is written in Futhark, the data-parallel language that is also partially the subject of the paper. An appropriate version of the Futhark compiler is also needed. The paper results have been produced with Futhark 0.15.8.

Each directory of this artifact contains a README.md explaining the use of its contents (or a link to further information).

The artifact is divided into three main parts:

• The library directory contains a CUDA library implementation of our generalized histograms, that can be used by other CUDA programs. It also contains documentation and an example program. It will not be discussed further.

• The prototype directory contains the CUDA prototoype that was used for the design space exploration of Section II and its validation in section IV.A, including reproducing Figure 7.

• The benchmarks directory contains the benchmark programs used for Section IV in the paper.

The artifact is designed as a bunch of Makefiles calling other scripts. Generally, these Makefiles will produce data files with runtime results, and re-running make will not re-run the benchmarks, but only re-report the last results. Use make -B (or make clean) to actually perform a re-run.

System requirements

• A working and properly setup CUDA installation, such that gcc can link with the cuda, nvrtc, and OpenCL libraries without any other options.

• Python 3 with NumPy and Matplotlib.

• Enough LaTeX for Matplotlib's LaTeX backend to work. On most Linux distributions, the "texlive-full" package will suffice.

• make must refer to GNU Make.

• CMake version 3 must be available as either cmake3 or cmake (the tooling will prefer the former, which is what it's called on RHEL 7).

• An approriate version of futhark (0.15.6 or later) must be on the shell PATH. Run make bin/futhark in this directory to unpack a statically linked binary that will be automatically picked up by the other tooling.

Paper environment

We have run the experiments in the paper on a server-class computer with an Intel(R) Xeon(R) CPU E5-2650 v2 CPU, but more importantly an NVIDIA RTX 2080 Ti GPU. The operating system was RHEL 7.7 and we used CUDA 10.1 for the GPU interaction.

Note that some constants (such as the amount of available L2 cache) are tuned for the RTX 2080 Ti GPU, and have not yet been made configurable or self-configuring. Anecdotal evidence suggests that the current configuration is still good on most contemporary NVIDIA hardware.