/mapmap_cpu

A high-performance general-purpose MRF MAP solver, heavily exploiting SIMD instructions.

Primary LanguageC++BSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

mapMAP MRF MAP Solver Build Status

Change Log

Compared with the development version from our HPG paper (cf. below).

  • v1.5 (6/25/2018):
    • Added tech report for new tree selection algorithm (from v1.2) in doc/.
    • Added novel envelopes for supermodular cost function types ("Antipotts", "LinearPeak"). A tech report will follow.
    • Several bugfixes.
  • v1.4 (1/10/2018):
    • Deterministic solver path with user-provided seed.
    • Several bugfixes and smaller improvements.
  • v1.3 (10/18/2017):
    • Envelope optimization for Potts, TruncatedLinear, TruncatedQuadratic.
    • Ability to have individual cost functions per edge.
    • Removed UNARY/PAIRWISE template parameters from solver, hiding these internally.
    • Improved multilevel performance, even in the case of individual costs.
    • Added GTest for automatic built instead of a hard dependency.
  • v1.2 (5/29/2017):
    • Introduced a new, multicoloring-based tree selection algorithm - lock-free.
  • v1.1 (4/12/2017):
    • Tuned the tree growing implementation for early termination and
    • option for relaxing the maximality requirement.
  • v1.0 (2/8/2017):
    • Stable release.
    • Logging callbacks for use as library.
    • Clean, documented interface and documentation.
    • Added a demo for correct usage.
  • beta (12/6/2016):
    • Initial release, mirrors functionality outlined in the paper.
    • Automated vectorization (compile-time detection) for float/double.
    • Supporting scalar/SSE2-4/AVX/AVX2.
    • Added cost function instances.
    • Added unit tests.

Overview

CPU-implementation of out massively-parallel, generic, MRF MAP solver named mapMAP posing minimal assumptions to the input, allowing rapid solution of a large class of MRF problems.

mapMAP's algorithmic foundation and parallelization concept has been presented at High Performance Graphics 2016 in Dublin, Ireland. For a reprint and further information, please refer to our project page (see below).

Currently, this code implements the following modules and features:

  • Customizable (parallel) performance
    • Change cost type between float and double
    • Templated SIMD width (1, 4, 8 for float; 1, 2, 4 for double)
    • Automatically setting SIMD width at compile time
    • Supports SSE4/AVX/AVX2, autodetected during build
    • Automatically using linear-time optimization for certain submodular cost functions
    • Novel linear-time optimization for certain supermodular cost functions
    • Two algorithms for parallel tree sampling
  • Extensible interfaces for all components, providing user hooks
    • Cost functions (unary and pairwise)
    • Termination criteria
    • Node grouping criteria for the multilevel module
    • Choice of two (parallel) coordinate selection algorithms
    • Use of heuristics, thereby modifying the solver's structure.
    • User hooks for logging intermediate results.
  • Solver modules
    • Acyclic (BCD) descent
    • Spanning tree descent
    • Multilevel solving
  • Finding and exploiting connected components in the topology
  • Test suite for each individual module

In the future, we will add:

  • Capability to process label costs as outlined in the paper

For the license and terms of usage, please see "License, Terms of usage & Reference".

Prerequisites

  • CMake building system (>= 3.0.2)
  • C++11 compatible compiler (e.g. gcc-5, MSVC 13, icc 17)
  • Intel TBB (>= 4.4, see Webpage)

The code has been tested (and compiles without issues) on an Ubuntu 16.04 system with an AVX-compliant Intel i7-3930K CPU with 64 GB RAM and using gcc/g++ 5.4.0 and Intel TBB (v2017u3). The latter is licensed under the 3BSD-compatible Apache 2.0 licence (see ASF legal FAQ). Please make sure to use an C++11-comptabile compiler and activate the necessary options. If you are a Ubuntu user, please install the packages libtbb2 libtbb-dev (see also Travis CI-script). Google Test will automatically be downloaded and built.

The provided FindTBB.cmake is taken from justusc and licensed under the MIT license.

Quickstart

The following instructions are provided for linux; the Windows workflow should be somewhat similar, though GUI-based.

Step-by-step instructions:

  1. git clone https://github.com/dthuerck/mapmap_cpu
  2. cd mapmap_cpu && mkdir build && cd build && cmake ..
  3. ccmake . and configure the following options (if you want to...):
  • CMAKE_C_COMPILER - command for your C-compiler, e.g. gcc-5
  • CMAKE_CXX_COMPILER - command for your C++-compiler, e.g. g++-5
  • TBB_INCLUDE_DIRS - path containing the tbb/ folder with include files, e.g. /usr/include
  • TBB_LIBRARY - path containing the TBB library files, e.g. /lib
  • BUILD_MEMSAVE - determines if the dynamic programming should allocate memory as needed (ON), saving memory but causing slightly longer execution times or preallocate the whole table (OFF)
  • BUILD_DEMO - decides whether the demo from the wiki is built as mapmap_demo
  • BUILD_TEST - decides whether the test suite is built as mapmap_test
  1. Configure and generate the Makefile (press c and g from ccmake).
  2. Build the project using make (or make -j for parallel build).
  3. Depending on your configuration, you can now run mapmap_test and/or mapmap_demo (assuming you activated BUILD_TEST and BUILD_DEMO).

Using mapMAP as a library in your own projects

mapMAP is implemented as a templated, header only library. A simple

#include "mapmap/full.h"

will do the trick. Remember that in order to work you need to compile your whole project with C++11 support. All functions and classes are organized in the namespace mapmap.

For the users of GCC, we recommend the following options for the best performance:

-std=c++11 -Wall -march=native -O2 -flto -mfpmath=sse -funroll-loops

As a good starting point, we recommend studying mapmap_demo.cc closely, which is mostly self-explanatory.

Documentation

For extended documentation on building, using and extending mapMAP, please see the integrated wiki.

License, Terms of Usage & Reference

Our program is licensed under the liberal BSD 3-Clause license included as LICENSE.txt file.

If you decide to use our code or code based on this project in your application, please make sure to cite our HPG 2016 paper:

@inproceedings{Thuerck2016MRF,
    title = {A Fast, Massively Parallel Solver for Large, Irregular Pairwise {M}arkov Random Fields},
    author = {Thuerck, Daniel and Waechter, Michael and Widmer, Sven and von Buelow, Max and Seemann, Patrick and Pfetsch, Marc E. and Goesele, Michael},
    booktitle = {Proceedings of High Performance Graphics 2016},
    year = {2016},
}

A PDF reprint is available here: PDF reprint and PDF Supplementary Material.

Contact

For any trouble with building, using or extending this software, please use the project's integrated issue tracker. We'll be happy to help you there or discuss feature requests.