OpenCilk infrastructure

This repo contains tools for building the OpenCilk compiler, runtime, and productivity tools. Specifically, it includes scripts for building OpenCilk from source or building a Docker image of OpenCilk.

Supported systems

OpenCilk has been tested on the following processor architectures:

• Intel x86 processors, Haswell and newer
• AMD x86 processors, Excavator and newer
• Various 64-bit ARM processors, including Apple M1 and M2

The present version has been tested on the following operating systems:

• Recent versions of Ubuntu, including via the Windows Subsystem for Linux v2 (WSL2) on Windows 10
• Recent versions of macOS
• FreeBSD 13
• Recent versions of Fedora

Summary of OpenCilk features

• The cilk_spawn, cilk_sync, and cilk_for keywords are enabled by using the -fopencilk compiler flag and including <cilk/cilk.h>.
• The cilk_scope keyword specifies that all spawns within a given lexical scope are guaranteed to be synced upon exiting that lexical scope. The cilk_scope keyword can also be used as a hint that the runtime system should ensure that Cilk workers are initialized, in order to quiesce performance measurements. Like the other Cilk keywords, cilk_scope is available by using the -fopencilk compiler flag and including <cilk/cilk.h>.
• The compiler is based on LLVM and supports the usual clang options as well as advanced linking features, such as link-time optimization (LTO).
• Both C and C++ are supported, including all standards supported by LLVM 16.
• Support for deterministic parallel random-number generation is available. To enable pedigree support, link the Cilk program with the pedigree library, -lopencilk-pedigrees.
• OpenCilk 2.0 and newer provides language support for reducer hyperobjects. A local or global variable can be made into a reducer by adding cilk_reducer(I, R) to its type, where I and R designate the identity and reduce functions for the reducer.
• Cilksan instrumentation for determinacy-race detection is enabled by using the -fsanitize=cilk compiler flag. Cilksan supports reducers and Pthread mutex locks. In addition, Cilksan offers an API for controlling race detection, which is available by including <cilk/cilksan.h>.
• Cilkscale instrumentation for scalability analysis and profiling is enabled by using the -fcilktool=cilkscale compiler flag. Cilkscale offers an API for analyzing user-specified code regions, which is made available by including <cilk/cilkscale.h>, and includes facilities for benchmarking an application on different numbers of parallel cores and visualizing the results.
• [Beta feature] Cilksan integrates with a custom version of the RR reverse debugger to enable interactive debugging of determinacy races.

OpenCilk is largely compatible with Intel's latest release of Cilk Plus. Unsupported features include:

• The Intel Cilk Plus reducer library.
• Cilk Plus array-slice notation.
• Certain Cilk Plus API functions, such as __cilkrts_set_param().

How to get OpenCilk

Precompiled binaries for OpenCilk are available for some systems here: https://github.com/OpenCilk/opencilk-project/releases/. To install, either download and run the appropriate shell archive (i.e., the .sh file) or unpack the appropriate tarball. A Docker image with OpenCilk installed is available from the same page. Some documentation on how to use the Docker image can be found here: docker.

These precompiled binaries require that standard system header files and libraries are already installed. These header files and libraries can be obtained by installing a modern version of GCC (including g++) on Linux or by installing a modern version of Xcode on macOS.

For other systems, we recommend instructions for downloading and building OpenCilk from source can be found here.

Porting Cilk Plus code to OpenCilk

Reducers: OpenCilk version 2.0 and newer does not support the Intel Cilk Plus reducer library and instead features a new syntax and implementation for reducers. The new reducer implementation allows one to change a local or global variable into a reducer by adding cilk_reducer(I,R) to the variable's type, where I and R designate the identity and reduce functions for the reducer. For example, here is how a simple integer-summation reducer can be implemented using the new reducer syntax:

#include <cilk/cilk.h>

void zero(void *v) {
  *(int *)v = 0;
}

void plus(void *l, void *r) {
  *(int *)l += *(int *)r;
}

int foo(int *A, int n) {
  int cilk_reducer(zero, plus) sum = 0;
  cilk_for (int i = 0; i < n; ++i)
    sum += A[i];
  return sum;
}

To port a Cilk Plus program to OpenCilk, once all uses of unsupported features have been updated, make the following changes to your build process:

• When compiling the program, replace any uses of -fcilkplus with -fopencilk.
• When linking the program, replace any uses of -lcilkrts with -fopencilk.

Useful links

• OpenCilk website: https://www.opencilk.org
• Instructions for building OpenCilk from source: https://www.opencilk.org/doc/users-guide/build-opencilk-from-source/
• Scripts for building a Docker image with OpenCilk: docker
• Link to the OpenCilk infrastructure GitHub repo: https://github.com/OpenCilk/infrastructure
• Some simple demo Cilk programs: https://github.com/OpenCilk/tutorial
• Additional demo Cilk programs: https://github.com/OpenCilk/applications
• Notes for developing on the OpenCilk source: DEVELOPER_NOTES.md

Contact

Bug reports should be posted to the GitHub issue tracker. Other queries and comments should be emailed to contact@opencilk.org.

OpenCilk development team

• Tao B. Schardl, MIT — Director, Chief Architect
• John F. Carr, consultant — Senior Programmer
• Dorothy Curtis, MIT — Project Manager
• Bruce Hoppe, consultant — Documentation Specialist and Outreach Coordinator
• Charles E. Leiserson, MIT — Executive Director
• Tim Kaler, MIT, Research Scientist
• Xuhao Chen, MIT, Research Scientist
• Kyle Singer, MIT, Postdoc

Previous team members

• I-Ting Angelina Lee, WUSTL — Director, Runtime Architect
• Alexandros-Stavros Iliopoulos, MIT, postdoc
• Grace Q. Yin, MIT, intern

Acknowledgments

OpenCilk is supported in part by the National Science Foundation, under grant number CCRI-1925609, in part by the Department of Energy, National Nuclear Security Administration under Award Number DE-NA0003965, and in part by the USAF-MIT AI Accelerator, which is sponsored by United States Air Force Research Laboratory under Cooperative Agreement Number FA8750-19-2-1000.

Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and should not be interpreted as representing the official policies or views, either expressed or implied, of the United states Air Force, the U.S. Government, or the National Science Foundation. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein.