This repo contains a set of codes to calculate correlation functions and other clustering statistics in a cosmological box (co-moving XYZ) or on a mock (RA, DEC, CZ). Read the documentation on corrfunc.rtfd.io.
- Fast Theory pair-counting is 7x faster than
SciPy cKDTree
, and at least 2x faster than all existing public codes. - OpenMP Parallel All pair-counting codes can be done in parallel (with strong scaling efficiency >~ 95% up to 10 cores)
- Python Extensions Python extensions allow you to do the compute-heavy bits using C while retaining all of the user-friendliness of python.
- Weights All correlation functions now support arbitrary, user-specified weights for individual points
- Modular The code is written in a modular fashion and is easily extensible to compute arbitrary clustering statistics.
- Future-proof As I get access to newer instruction-sets, the codes will get updated to use the latest and greatest CPU features.
If you use the codes for your analysis, please star this repo -- that helps us keep track of the number of users.
Please see this gist for some benchmarks with current codes. If you have a pair-counter that you would like to compare, please add in a corresponding function and update the timings.
make >= 3.80
- OpenMP capable compiler like
icc
,gcc>=4.6
orclang >= 3.7
. If not available, please disableUSE_OMP
option option intheory.options
andmocks.options
. You might need to ask your sys-admin for system-wide installs of the compiler; if you prefer to install your own thenconda install gcc
(MAC/linux) or(sudo) port install gcc5
(on MAC) should work. gsl >= 2.4
. Use eitherconda install -c conda-forge gsl
(MAC/linux) or(sudo) port install gsl
(MAC) to installgsl
if necessary.python >= 2.7
orpython>=3.4
for compiling the C extensions.numpy>=1.7
for compiling the C extensions.
$ git clone https://github.com/manodeep/Corrfunc/
$ make
$ make install
$ python setup.py install (--user)
$ make tests
Assuming you have gcc
in your PATH
, make
and make install
should compile and install the C libraries + python extensions within the source directory. If you would like to install the python C extensions in your environment, then python setup.py install (--user)
should be sufficient. If you are primarily interested in the python
interface, you can condense all of the steps by using python setup.py install CC=yourcompiler (--user)
after git clone
.
- If python and/or numpy are not available, then the C extensions will not be compiled.
make install
simply copies files into thelib/bin/include
sub-directories. You do not needroot
permissions- Default compiler on MAC is set to
clang
, if you want to specify a different compiler, you will have to callmake CC=yourcompiler
,make install CC=yourcompiler
,make tests CC=yourcompiler
etc. If you want to permanently change the default compiler, then please edit the common.mk file in the base directory. - If you are directly using
python setup.py install CC=yourcompiler (--user)
, please run amake distclean
beforehand (especially if switching compilers)
The python package is directly installable via pip install Corrfunc
. However, in that case you will lose the ability to recompile the code according to your needs. Installing via pip
is not recommended, please open an install issue on this repo first; doing so helps improve the code-base and saves future users from running into similar install issues.
If compilation went smoothly, please run make tests
to ensure the code is working correctly. Depending on the hardware and compilation options, the tests might take more than a few minutes. Note that the tests are exhaustive and not traditional unit tests.
While I have tried to ensure that the package compiles and runs out of the box, cross-platform compatibility turns out to be incredibly hard. If you run into any issues during compilation and you have all of the pre-requisites, please see the FAQ or email the Corrfunc mailing list. Also, feel free to create a new issue with the Installation
label.
All codes that work on cosmological boxes with co-moving positions are located in the theory
directory. The various clustering measures are:
DD
-- Measures auto/cross-correlations between two boxes. The boxes do not need to be cubes.xi
-- Measures 3-d auto-correlation in a cubic cosmological box. Assumes PERIODIC boundary conditions.wp
-- Measures auto 2-d point projected correlation function in a cubic cosmological box. Assumes PERIODIC boundary conditions.DDrppi
-- Measures the auto/cross correlation function between two boxes. The boxes do not need to be cubes.DDsmu
-- Measures the auto/cross correlation function between two boxes. The boxes do not need to be cubes.vpf
-- Measures the void probability function + counts-in-cells.
All codes that work on mock catalogs (RA, DEC, CZ) are located in the mocks
directory. The various clustering measures are:
DDrppi_mocks
-- The standard auto/cross correlation between two data sets. The outputs, DD, DR and RR can be combined usingwprp
to produce the Landy-Szalay estimator for wp(rp).DDsmu_mocks
-- The standard auto/cross correlation between two data sets. The outputs, DD, DR and RR can be combined using the python utilityconvert_3d_counts_to_cf
to produce the Landy-Szalay estimator for xi(s, mu).DDtheta_mocks
-- Computes angular correlation function between two data sets. The outputs fromDDtheta_mocks
need to be combined withwtheta
to get the full omega(theta)vpf_mocks
-- Computes the void probability function on mocks.
If you plan to use the command-line, then you will have to specify the code runtime options at compile-time. For theory routines, these options are in the file theory.options while for the mocks, these options are in file mocks.options.
Note All options can be specified at runtime if you use the python interface or the static libraries. Each one of the following Makefile
option has a corresponding entry for the runtime libraries.
Theory (in theory.options)
PERIODIC
(ignored in case of wp/xi) -- switches periodic boundary conditions on/off. Enabled by default.OUTPUT_RPAVG
-- switches on output of<rp>
in eachrp
bin. Can be a massive performance hit (~ 2.2x in case of wp). Disabled by default.DOUBLE_PREC
-- switches on calculations in double precision. Disabled by default (i.e., calculations are performed in single precision by default).
Mocks (in mocks.options)
OUTPUT_RPAVG
-- switches on output of<rp>
in eachrp
bin forDDrppi_mocks
. Enabled by default.OUTPUT_THETAAVG
-- switches on output of in each theta bin. Can be extremely slow (~5x) depending on compiler, and CPU capabilities. Disabled by default.DOUBLE_PREC
-- switches on calculations in double precision. Disabled by default (i.e., calculations are performed in single precision by default).LINK_IN_DEC
-- creates binning in declination forDDtheta
. Please check that for your desired limits\theta
, this binning does not produce incorrect results (due to numerical precision). Generally speaking, if your\thetamax
(the max.\theta
to consider pairs within) is too small (probaly less than 1 degree), then you should check with and without this option. Errors are typically sub-percent level.LINK_IN_RA
-- creates binning in RA once binning in DEC has been enabled. Same numerical issues asLINK_IN_DEC
FAST_DIVIDE
-- Disabled by default. Divisions are slow but requiredDD(r_p,\pi)
. Enabling this option, replaces the divisions with a reciprocal followed by a Newton-Raphson. The code will run ~20% faster at the expense of some numerical precision. Please check that the loss of precision is not important for your use-case.FAST_ACOS
-- Relevant only whenOUTPUT_THETAAVG
is enabled. Disabled by default. Anarccos
is required to calculate<\theta>
. In absence of vectorizedarccos
(intel compiler,icc
provides one via intel Short Vector Math Library), this calculation is extremely slow. However, we can approximatearccos
using polynomials (with Remez Algorithm). The approximations are taken from implementations released by Geometric Tools. Depending on the level of accuracy desired, this implementation offast acos
can be tweaked in the file utils/fast_acos.h. An alternate, less accurate implementation is already present in that file. Please check that the loss of precision is not important for your use-case.COMOVING_DIST
-- Currently there is no support inCorrfunc
for different cosmologies. However, for the mocks routines like,DDrppi_mocks
andvpf_mocks
, cosmology parameters are required to convert between redshift and co-moving distance. BothDDrppi_mocks
andvpf_mocks
expects to receive aredshift
array as input; however, with this option enabled, theredshift
array will be assumed to contain already converted co-moving distances. So, if you have redshifts and want to use an arbitrary cosmology, then convert the redshifts into co-moving distances, enable this option, and pass the co-moving distance array into the routines.
Read the documentation on corrfunc.rtfd.io.
Navigate to the correct directory. Make sure that the options, set in either theory.options or mocks.options in the root directory are what you want. If not, edit those two files (and possibly common.mk), and recompile. Then, you can use the command-line executables in each individual subdirectory corresponding to the clustering measure you are interested in. For example, if you want to compute the full 3-D correlation function, \xi(r)
, then run the executable theory/xi/xi
. If you run executables without any arguments, the program will output a message with all the required arguments.
Look under the run_correlations.c and run_correlations_mocks.c to see examples of calling the C API directly. If you run the executables, run_correlations
and run_correlations_mocks
, the output will also show how to call the command-line interface for the various clustering measures.
If all went well, the codes can be directly called from python
. Please see call_correlation_functions.py and call_correlation_functions_mocks.py for examples on how to use the C extensions directly. Here are a few examples:
from __future__ import print_function
import os.path as path
import numpy as np
import Corrfunc
from Corrfunc.theory import wp
# Setup the problem for wp
boxsize = 500.0
pimax = 40.0
nthreads = 4
# Create a fake data-set.
Npts = 100000
x = np.float32(np.random.random(Npts))
y = np.float32(np.random.random(Npts))
z = np.float32(np.random.random(Npts))
x *= boxsize
y *= boxsize
z *= boxsize
# Setup the bins
rmin = 0.1
rmax = 20.0
nbins = 20
# Create the bins
rbins = np.logspace(np.log10(0.1), np.log10(rmax), nbins + 1)
# Call wp
wp_results = wp(boxsize, pimax, nthreads, rbins, x, y, z, verbose=True, output_rpavg=True)
# Print the results
print("#############################################################################")
print("## rmin rmax rpavg wp npairs")
print("#############################################################################")
print(wp_results)
USE_OMP
-- uses OpenMP parallelization. Scaling is great for DD (close to perfect scaling up to 12 threads in our tests) and okay (runtime becomes constant ~6-8 threads in our tests) forDDrppi
andwp
. Enabled by default. TheMakefile
will compare the CC variable with known OpenMP enabled compilers and set compile options accordingly. Set in common.mk by default.
Optimization for your architecture
- The values of
bin_refine_factor
and/orzbin_refine_factor
in thecountpairs\_\*.c
files control the cache-misses, and consequently, the runtime. In my trial-and-error methods, I have seen any values larger than 3 are always slower. But some different combination of 1/2 for(z)bin_refine_factor
might be faster on your platform. - If you have AVX2/AVX-512/KNC, you will need to add a new kernel within the
*_kernels.c
and edit the runtime dispatch code to call this new kernel.
Corrfunc was designed by Manodeep Sinha and is currently maintained by Lehman Garrison and Manodeep Sinha
If you use the code, please cite using the ascl entry as indexed by ADS. The BibTex entry for the code is
@misc{2017ascl.soft03003S,
author = {{Sinha}, M. and {Garrison}, L.},
title = "{Corrfunc: Blazing fast correlation functions on the CPU}",
keywords = {Software},
howpublished = {Astrophysics Source Code Library},
year = 2017,
month = mar,
archivePrefix = "ascl",
eprint = {1703.003},
adsurl = {http://adsabs.harvard.edu/abs/2017ascl.soft03003S},
adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}
If you have questions or comments about the package, please do so on the mailing list: https://groups.google.com/forum/#!forum/corrfunc
Corrfunc is released under the MIT license. Basically, do what you want with the code, including using it in commercial application.
- documentation (http://corrfunc.rtfd.io/)
- version control (https://github.com/manodeep/Corrfunc)