The artifact for Muller and Hoffmann's POPL 2021 paper "Modeling and Analyzing Evaluation Cost of CUDA Kernels"

RaCUDA is a static analysis tool for finding performance bottlenecks in CUDA kernels. The tool is built on Absynth, a static resource analyzer by Carbonneaux, Hoffmann, Ngo, and Wang. It comes with its rarely installed dependencies. To build it you will still need to have the following core:

• GMP 4.2 or up (with development headers)
• MPFR 2.2 or up (with development headers)
• GSL 1.6 or up (e.g., installed at /usr/local)
• OCaml 4.0 or up (and ocamlbuild)
• A C99 compiler (gcc/clang)
• A C++ compiler (g++/clang++)

The location of most of these dependencies can be specified using the main configure script, see ./configure -help for details.

This repository also contains, in the folder `popl21', the following items specific to our POPL '21 paper:

• coq: Coq scripts
• curesults: Profiling results for GPU executions of kernels.
• experiments: Experiment scripts for reproducing the POPL '21 results
• kernels: The CUDA kernels used as benchmarks in the evaluation. These may overlap with the example kernels in the top level of the repo, but are included here as a record of the benchmarks used in the POPL '21 evaluation
• model-eval: Spreadsheets of the results comparing the evaluation costs obtained by simulating the cost semantics in the paper with those obtained from profiling GPU executions.
• paper-analysis-results: Results from analyzing the kernels in our experiments.
• params: Parameter files for use in analyzing and simulating the benchmark kernels.
• plot: gnuplot files and plots for Figure 7 of the paper.
• simresults: Results from simulating execution of the kernels in our experiments.

./configure
make

The configuration script might fail to find the location of the GSL library. If that is the case, it will fail to build the gsl_caller program. This is however not fatal and many examples can be tested without this program. Similarly, `llvm-config' is not necessary for running any of the CUDA examples and may be missing.

absynth-cuda [-eval] -metric <metric> [-param-file <param file>] <kernel>

The above command runs the tool on the kernel . The common options are:

• -metric <metric> use the supplied resource metric. Supported metrics are:
  • divwarps: number of conditional instructions that cause a warp to diverge.
  • global: number of sectors read from/written to global memory.
  • shared: number of bank conflicts in shared memory.
  • steps: count number of abstract instructions executed.
• -eval: run the tool in evaluation mode. In this mode, the tool does not perform a static resource analysis but rather simulates the execution of the kernel on a GPU and calculates the cost under the given resource metric using the cost model from our POPL '21 paper.
• -param-file <file>: load parameters from . Parameters include block and grid size. For evaluation mode, initial values of any variables that affect control flow must be supplied. Parameter file is optional for analysis mode; default parameters will be used. Parameter files for each kernel from our POPL '21 paper are available in popl21/params (do not use or modify parameter files ending in -temp; these are templates used by the experiment scripts).

There are many examples (including the ones in our paper) in the kernels/ sub-directory.

Parameter files for the kernels used in our POPL '21 paper are available in popl21/params (do not use or modify parameter files ending in -temp; these are templates used by the experiment scripts).