/lomp

Little OpenMP Library

Primary LanguageC++Apache License 2.0Apache-2.0

Little OpenMP* Runtime

LOMP, short for Little OpenMP (runtime), is a small OpenMP runtime implementation that can be used for educational or prototyping purposes. It currently only implements a rather small subset of the OpenMP Application Programming Interface for CPUs (i.e., it has no support for offload to target devices and is also missing many CPU-only features).

LOMP was written to demonstrate the design principles outlined in the book High-Performance Parallel Runtimes.

The library uses the same binary interface as clang/LLVM*, and thus is compatible with several compilers that use that interface. Unless you use a feature that LOMP does not currently support, LOMP can serve as a drop-in replacement for the native OpenMP runtime library of a compatible compiler without requiring re-compilation of your application.

The runtime is mostly written in C++14, though with some features of C++17, and can be compiled for a variety of different architectures. There are no assembler files in the runtime, and the use of inline assembly is restricted to a few features (such as reading the high-resolution clock). For architectures that do not have a code path to access the high-resolution clock via inline assembly, we rely on C++ features to measure time. Atomic operations are all accessed though std::atomic.

As well as the source for the runtime there are also a few micro-benchmarks and some (extremely minimal) sanity tests.

As its name suggest, the LOMP library is significantly smaller than the production LLVM OpenMP runtime library. At the time of the initial check-in, the cloc utility in the source directory shows under 6,000 lines of C++ code (and no assembly code). For comparison, the production LLVM OpenMP runtime has around 63,500 lines of C/C++ code and 1400 lines of assembly code in the CPU part of the library. Of course, this is an unfair comparison, since LOMP is missing many features which are supported by the production runtime. However, it does make the point that if you want somewhere to experiment, or an environment in which to set a student project, LOMP may be an easier codebase to work with!

Supported Target Platforms

The LOMP runtime supports the following target architectures (in parentheses we show the architecture name as reported by the uname command):

  • AMD* Processors (x86_64)
  • Arm* Processors, 32 bit (armv7l)
  • Arm Processors, 64 bit (aarch64)
  • Intel* Processors (x86_64)
  • RISC-V* Processors, 64 bit (riscv64)

The library works with Arm 64-bit processors running macOS (announced as (arm64) by uname there), but an LLVM compiler from at least April 2021 is required, since there was a compiler bug(unrelated to the runtime) which broke OpenMP tasks there as a result of incorrect assumptions about varargs argument passing on that platform. (See this bug for details.)

Supported OpenMP Features

The language supported by LOMP is restricted to a small subset of the OpenMP API for shared-memory multi-threading. Supported OpenMP features are

  • the parallel construct, including the data-sharing clauses shared, private, firstprivate, and lastprivate;
  • the master and single constructs;
  • the barrier construct;
  • reductions (though not yet up a tree at the barrier);
  • worksharing constructs for (C/C++) and do (Fortran);
  • scheduling types static, dynamic, auto ,and guided, along with the additional monotonic and nonmonotonic qualifiers;
  • the critical construct;
  • the flush construct;
  • the task construct, including the data-sharing clauses shared, private, firstprivate, and if;
  • the taskwait construct; and
  • the taskgroup construct.

The supported OpenMP runtime routines are:

  • omp_get_thread_num() and omp_get_num_threads();
  • omp_set_num_threads() but only if performed before the runtime is initialized or if it is setting the same value as is already in use;
  • omp_get_max_threads() and omp_in_parallel();
  • omp_set_schedule() and omp_get_schedule();
  • omp_init_lock(), omp_set_lock(), omp_unset_lock(), and omp_destroy_lock(); and
  • omp_get_wtime().

Important Things Which Are Not Yet Supported

Since this is a small and relatively simple runtime (at least for now), there are few restrictions and many things which have not yet been implemented. A, possibly incomplete, list is:

  • nested parallelism;
  • changing the number of threads;
  • nested locks;
  • more elaborate tasking features such as task dependences and taskloop;
  • parsing many of the OpenMP-mandated environment variables (beyond OMP_NUM_THREADS), and support for their related internal control variables;
  • explicitly controlling thread affinity;
  • the OMPT and OMPD profiling and debugging interfaces;
  • ordered loops;
  • the teams and distribute constructs;
  • the cancellation constructs;
  • offloading to accelerator devices, such as GPUs; and
  • probably other things which we haven't noticed!

The runtime is also, of course, limited in the language it can support by the compiler. There are therefore some OpenMP API version 5.1 features which are not yet implemented since there is no compiler support for them yet.

If you would like to contribute any features to LOMP, please see below.

How to Build (and Install)

Here are some, hopefully useful, remarks about how you can set up LOMP on your system. The instructions come without any warranty, and may be wrong, or incomplete.

Software Versions

To build the LOMP library, you need the following software environment:

  • CMake, minimum version 3.13.0
  • A clang-compatible compiler, one of:
    • clang, minimum version 10.0.0
    • AOCC, minimum version 2.2.0
    • Intel Next-gen Compiler (from Intel oneAPI), minimum version 2020.0
  • libnuma, minimum version 2.0 (optional)
  • Python, minimum version 3.x (optional; required for the micro-benchmarks but not the library itself)

Other versions of software tools may work, but we have not tested them with our code.

While the LOMP library itself compiles fine with the GNU Compiler Collection (GCC), we have not implemented all of the entry points that GCC requires for its OpenMP support. So, while you can compile the LOMP runtime code with GCC, you will need a clang-compatible compiler to generate code to exercise the LOMP library that you have built.

The micro-benchmarks (in the directory microBM) should work with any OpenMP implementation.

Building LOMP

Building LOMP follows the usual process of building a CMake-based project. Here are the steps needed:

  • Checkout LOMP from the project website:

    • via SSH: git clone git@github.com:parallel-runtimes/lomp.git
    • via HTTPS: git clone https://github.com/parallel-runtimes/lomp.git
  • Create a build directory for an out-of-tree build: mkdir lomp_build

  • Go to that new directory and run cmake there (before doing this, check the CMake Configuration Options section below; you will probably need to add some options to set the appropriate compiler).

    • cd lomp_build
    • cmake ../lomp
  • Compile LOMP (depending on which build system you asked cmake to create build files for)

    • with GNU Make: make
    • with Ninja: ninja
  • Run one of the compiled examples from the examples folder, e.g., Hello World:

    $ ./examples/hello_world
    Before parallel region
    =======================================
    Hello World: I am thread 6, and my secrets are 42.000000 and 21
    Hello World: I am thread 4, and my secrets are 42.000000 and 21
    Hello World: I am thread 1, and my secrets are 42.000000 and 21
    Hello World: I am thread 2, and my secrets are 42.000000 and 21
    Hello World: I am thread 0, and my secrets are 42.000000 and 21
    Hello World: I am thread 5, and my secrets are 42.000000 and 21
    Hello World: I am thread 3, and my secrets are 42.000000 and 21
    Hello World: I am thread 7, and my secrets are 42.000000 and 21
    =======================================
    After parallel region
    $
    
  • You can also run the (tiny) test suite using the following commands (please accept a few failed tests for v0.1):

    • with GNU Make: make test
    • with Ninja: ninja test

The default build configuration is "Release" mode, which enables compiler optimizations. Please see below for how to change this default.

To use LOMP with an existing code, once you have built the library, you should be able to use LD_LIBRARY_PATH on Linux (or DYLD_LIBRARY_PATH on macOS) to place its directory before the system one where the production OpenMP library lives so that LOMP is used without needing to recompile your executable. If you also set LOMP_DEBUG=1 you should see some output that proves that you are using the library you expect. (Of course, the ldd command on Linux can also show you that.)

If you compiled the Hello World example that comes with LOMP using one of the supported OpenMP compilers and if LOMP has been compiled in $HOME/build_lomp, the following will dynamically bind LOMP to your compiled code:

$ export LD_LIBRARY_PATH=$HOME/build_lomp/src/:$LD_LIBRARY_PATH
$ LOMP_DEBUG=1 OMP_NUM_THREADS=4 ./a.out
Before parallel region
=======================================
LOMP:runtime version 0.1 (SO version 1) compiled at 19:26:59 on Jan 28 2021
from Git commit 0abcdef for x86_64 by LLVM:11:0:0
LOMP:with configuration -mrtm;-mcx16;DEBUG=10;LOMP_GNU_SUPPORT=1;LOMP_HAVE_RTM=1;LOMP_HAVE_CMPXCHG16B=1
Hello World: I am thread 1, and my secrets are 42.000000 and 21
Hello World: I am thread 2, and my secrets are 42.000000 and 21
Hello World: I am thread 0, and my secrets are 42.000000 and 21
Hello World: I am thread 3, and my secrets are 42.000000 and 21
=======================================
After parallel region
$

By using the export statement you will make LOMP your default OpenMP runtime for processes started from this shell. If you didn't want that, remember to reset LD_LIBRARY_PATH.

CMake Configuration Options

The following options can be set using the cmake command line interface:

  • -G Ninja: Sets the build system to Ninja (the default is GNU Make).
  • -DCMAKE_C_COMPILER=xyz: Set the C compiler to be xyz, the GNU Compiler Collection (GCC) is the default on most systems, but we want clang. (If you have clang in your path you can use -DCMAKE_C_COMPILER=clang).
  • -DCMAKE_CXX_COMPILER=xyz: Set the C++ compiler to be xyz, GCC's g++ is the default on most systems, but we want clang++. (If you have clang++ in your path you can use -DCMAKE_CXX_COMPILER=clang++.)
  • -DLOMP_SERIAL=[on|off]: on builds a serial version of the library with all entrypoints, but with only one thread at runtime. The default is off.
  • -DLOMP_BUILD_EXAMPLES=[on|off]: on builds the examples, off does not. The default is on.
  • -DLOMP_SHARED_LIB=[on|off]: on builds LOMP as a shared library, off builds a static library instead. The default is on.
  • -DCMAKE_INSTALL_PREFIX=<path>: Define the path that will be used to install LOMP. The default is /usr/local on Linux* systems.
  • -DCMAKE_BUILD_TYPE=[release|debug|relwithdebinfo]: release builds with optimization; debug builds without optimization and with debug information; relwithdebinfo builds with optimizations and debug information. The default is release.
  • -DCMAKE_VERBOSE_MAKEFILE=[on|off]: on shows compiler invocation, while off does not. The default is off.
  • -DLOMP_ARM64_ARCHITECTURE=arch: selects the ISA version for ARM64-based processors; the default is armv8.1, use armv8-a for Raspberry Pi 3 and 4, or armv7-a for Raspberry Pi 2.
  • -DLOMP_BUILD_MICROBM=[on|off]: on builds the micro-benchmarks, off does not, the default is on.
  • -DLOMP_GNU_SUPPORT=[on|off]: on builds GCC entry points for libgomp; off does not build these entry points; the default is off. This option is for the brave and will likely produce errors, as most these entry points have not yet been implemented.
  • -DLOMP_ICC_SUPPORT=[on|off]: on builds entry points for the Intel classic compiler, off does not build these entry points; the default is off. This option is for the brave and will likely produce errors, as most of these entry points have not been implemented.
  • -DLOMP_MICROBM_WITH_LOMP=[on|off]: on links the micro-benchmarks with LOMP, off links them against the native OpenMP runtime of the compiler being used; the default is off.
  • -DLOMP_WARN_ARCH_FEATURES=[on|off]: on emits a warning if a dummy function is used for an unsupported architectural feature; off does not; the default is on.
  • -DLOMP_MAX_THREADS=[n]: set the maximum number of threads supported by LOMP to n. The default is 256 threads.
  • -DLOMP_WARN_API_STUBS=[on|off]: on emits warnings about entry point stubs; off does not; the default is on.

Installing LOMP

LOMP supports to be installed using the install target. The location is determined via the -DCMAKE_INSTALL_PREFIX=<path> configuration option for CMake. After a successful build, the following will install LOMP

  • with GNU Make: make install
  • with Ninja: ninja install

Environment Variables

The LOMP runtime library supports various environment variables that control its behavior:

  • OMP_NUM_THREADS: set the number of threads, see the OpenMP specification for details.
  • OMP_SCHEDULE: set the loop schedule for loops that were compiled with the runtime schedule, see the OpenMP specification for details.
  • LOMP_LOCK_KIND: This environment variable controls the lock implementation that LOMP uses for OpenMP locks. The default is to use the C++ std::mutex lock. The values it supports are:
    • TTAS: use test and test-and-set lock.
    • MCS: use a fair, scalable lock based on the ideas of Mellor-Crummey and Scott.
    • cxx: use the std::mutex lock of C++. This is the default.
    • pthread: use the pthread_mutex (this requires that the C++ std::thread implementation is based on pthreads, which may not be true on all platforms).
    • speculative: use a speculative lock that internally uses TTAS as the fallback lock (this needs support for speculative execution in hardware).
  • LOMP_BARRIER_KIND: Controls which barrier implementation from LOMP's barrier zoo is used. There are too many to list here, so "Use the source, Luke!"
  • LOMP_DEBUG: Enable printing of debugging messages. The higher the level, the more debugging output will appear on the screen, higher numbers will also enable the lower-number levels. Useful levels are:
    • 0: show no debug messages. This is the default.
    • 1: print the library's name, target, and compilation information.
    • 2: print informational messages.
    • 10: print more details.
    • 20: print debugging messages for LOMP threading subsystems.
    • 30: print debugging message for memory allocations.
    • 40: print debugging messages for barriers.
    • 50: print debugging messages for loop scheduling.
    • 60: print debugging messages for lock implementations.
    • 1000: print debugging messages for internal function invocations.
  • LOMP_TRACE: Enable LOMP's internal tracing facility, setting the debug level to the value specified for LOMP_TRACE. See LOMP_DEBUG for the supported levels.

Note that when debugging the library it is often convenient to change the order of the debug tags, so as only to print information from the subsystem of interest, so the values for LOMP_DEBUG may change over time.

Micro-Benchmarks

The micro-benchmarks are in the microBM directory. These were used to measure hardware properties shown in the book. You can use them to measure the properties of your own machines. Each benchmark can be invoked by an appropriate Python script which will run all of the available measurements, or the ones which you request and write appropriately named files containing the results.

To use the scripts, ensure that your current directory is the microBM directory in the appropriate build, then execute the relevant Python script from the microBM source directory, e.g.

    $ cd lomp_build/microBM
    $ python ~/lomp/microBM/runAtomics.py
    Arch (may be wrong!):
    Model:
    Cores:  8
    Running  OMP_NUM_THREADS=8 KMP_HW_SUBSET=1T KMP_AFFINITY='compact,granularity=fine' ./atomics Ie > AtomicsIe_Mac-mini_2021-01-28_1.res
    ./atomics Ie
    ........
    Running  OMP_NUM_THREADS=8 KMP_HW_SUBSET=1T KMP_AFFINITY='compact,granularity=fine' ./atomics If > AtomicsIf_Mac-mini_2021-01-28_1.res
    ./atomics If
    ........
    Running  OMP_NUM_THREADS=8 KMP_HW_SUBSET=1T KMP_AFFINITY='compact,granularity=fine' ./atomics Ii > AtomicsIi_Mac-mini_2021-01-28_1.res
    ./atomics Ii
    ........
    Running  OMP_NUM_THREADS=8 KMP_HW_SUBSET=1T KMP_AFFINITY='compact,granularity=fine' ./atomics It > AtomicsIt_Mac-mini_2021-01-28_1.res
    ./atomics It
    ........
    $

The output files can be converted into web pages containing tabulated results and plots by using the plot.py script in the scripts directory and feeding it the various files for a specific measurement, and should also be easy to read. If you feel the desperate need to use a spreadsheet, the toCSV.py script in the scripts directory will convert them into a comma-separated file format which can be read into whichever spreadsheet you suffer.

Bugs

Please submit bug reports or other feedback via GitHub issues by filling in the issue templates that are provided.

Contributions

Contributions are welcome, as is other feedback. We hope that this runtime will provide a useful environment for experimentation with new OpenMP features (such as different loop schedules, or barrier implementations), while remaining simple enough to be easy to use in a university course.

If you want to contribute (for fame and glory), please fork the repository and submit a pull request with the changes you would like to make

License

The LOMP runtime is licensed under the "Apache 2.0 License with LLVM exceptions" license. Any contributions must use that license to be acceptable.

Reference

If you use LOMP in your research and publish a paper, please use the following in your citation:

Jim Cownie and Michael Klemm, Little OpenMP Runtime, https://github.com/parallel-runtimes/lomp, February 2021.

Contributors

Trademarks

Trademarks and registered names are marked with an asterisk (*) at their first use where we have recognized them. Other names and trademarks may be the property of others.