search

dthuerck / mapmap_cpu

A high-performance general-purpose MRF MAP solver, heavily exploiting SIMD instructions.
BSD 3-Clause "New" or "Revised" License
102 stars 51 forks source link
algorithm computer-vision dynamic-programming feedback-vertex-set markov-random-field massively-parallel mrf simd simd-programming vectorization

readme

mapMAP MRF MAP Solver Build Status tipi.build

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:

What it lacks:

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

In addition to a "classical", CMake-based workflow, the great folks at tipi.build contributed (optional) support for their C++ build and dependency manager. With tipi, building mapmap or using it as a library in your own projects just happens by a snap of your finger.

Prerequisites

The code has been tested (and compiles without issues) on an Ubuntu 16.04 system with an AVX2-compliant Intel CPU using gcc/g++ 9.3.0 and Intel OneTBB (2021.5.0). 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 click on APT on the OneTBB page linked above. 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.

The classical way

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_DIR - path containing TBBConfig.cmake
    • 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
  4. Configure and generate the Makefile (press c and g from ccmake).
  5. Build the project using make (or make -j for parallel build).
  6. Depending on your configuration, you can now run mapmap_test and/or mapmap_demo (assuming you activated BUILD_TEST and BUILD_DEMO).

Using tipi.build

All prerequisites are provided by tipi.build and the .tipi/deps file, to compile this project simply run : 

tipi . -t <platform>

Where platform is either one of linux, windows, macos or any of the supported environments.

Using mapMAP as a library in your own projects

The classical way

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.

Using tipi.build

mapMAP can be easily used with the tipi.build dependency manager, by adding the following to a .tipi/deps:

{
    "dthuerck/mapmap_cpu": { }
}

Documentation

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


An example to start with is available in [example-mapmap_cpu](https://github.com/tipi-deps/example-mapmap_cpu) (change the target name appropriately to `linux` or `macos` or `windows`):

```bash
tipi . -t <target>

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.

Change Log

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