N-point correlation function estimation
npoint library

Contact: Bill March (march@gatech.edu)

Additional authors: 
Dongryeol Lee (drselee@gmail.com)
Marat Dukhan (maratek@gmail.com)
Kenneth Czechowski (kent.czechowski@gmail.com)
Thomas Benson (Thomas.Benson@gtri.gatech.edu)

References: 
Original tree-based npcf estimation: 
Gray & Moore, NIPS 2000.
Efficient jackknife resampling and multi-matcher algorithms:
March, Connolly, & Gray, SIGKDD 2012.
Optimized base case kernels:
March, et al., Supercomputing 2012.


============================================================================
Dependencies: 
This code requires the MLPACK machine learning library, available at
mlpack.org for space-partitioning trees and general I/O functionality.

CMake (version 2.8 or higher) is required to build the code.

============================================================================
Building: 

For source code in $SRC_DIR, you may build the code in a directory $BUILD_DIR
by:
cd $BUILD_DIR
cmake -D DEBUG=OFF -D PROFILE=OFF -D MLPACK_INCLUDE_DIR=$MLPACK_INCLUDE_DIR
-D MLPACK_LIBDIR=$MLPACK_LIB_DIR $SRC_DIR
make

In order to support all of the optimized base case kernels, use gcc 4.6.


============================================================================
Executables:

${n}_point main computes the raw correlation counts for a single matcher 
(set of distance constraints) on a single node.  distributed_${n}_point_main
does the same using MPI for inter-node communication. Thread parallelism is 
currently supported through creating multiple MPI processes per node. 

distributed_angle_main (and it's serial version, angle_3pt_main) compute
3pcf raw correlation counts for an angle matcher. See the files for a 
description of the format for angle matchers.

distributed_multi_matcher_main (and multi_matcher_main) compute npcf raw 
correlation counts for multi-matchers.  See the files for a description of 
the format for multi matchers.

============================================================================
Example use: 

data.csv: 
Input data points, contained in a cubic region of size 100 on each side. 
Arguments a, b, c specify the region size.

lower_matcher.csv:
0.0 0.9 0.9
0.9 0.0 0.9
0.9 0.9 0.0

upper_matcher.csv:
0.0 1.0 1.0
1.0 0.0 1.0
1.0 1.0 0.0

./3_point_main -v -d data.csv -R 1000 -l lower_matcher.csv -u upper_matcher.csv
-a 100 -b 100 -c 100 -x 2 -y 2 -z 2

This call will split the data in data.csv (contained in a cube of side length
100), into 8 equal sized jackknife resampling regions.  It will compute (and 
output to cout) the DDD, DDR, DRR, and RRR raw correlation counts for the data 
and 1000 uniformly distributed random points. The algorithm will count triples 
where each pairwise distance is between 0.9 and 1.0 (in the same units as the 
input data).

Leaving the arguments x,y,z unspecified will not perform any resampling. 
Setting -R to 0 (or leaving it unspecified) will only compute counts for the 
data. The -v argument prints additional execution and timing info.