Blazing fast correlation functions on the CPU


Keywords
correlation, functions, simulations, surveys, galaxies, astrophysics, avx, avx2, avx512, c, correlation-functions, cosmology, intrinsics, large-scale-structure, openmp, pair-counting, python, simd, sse42
License
MIT
Install
pip install Corrfunc==2.5.2

Documentation

Corrfunc logo

Latest Release PyPI Release MIT License Build Status GitHub Actions Status Documentation Status Open Issues

Core Infrastructure Best Practices Status Fair Software (EU) Compliance

Corrfunc Paper I Corrfunc Paper II

Description

This repo contains a suite of codes to calculate correlation functions and other clustering statistics for simulated galaxies in a cosmological box (co-moving XYZ) and on observed galaxies with on-sky positions (RA, DEC, CZ). Read the documentation on corrfunc.rtfd.io.

Why Should You Use it

  1. Fast Theory pair-counting is 7x faster than SciPy cKDTree, and at least 2x faster than all existing public codes.
  2. OpenMP Parallel All pair-counting codes can be done in parallel (with strong scaling efficiency >~ 95% up to 10 cores)
  3. Python Extensions Python extensions allow you to do the compute-heavy bits using C while retaining all of the user-friendliness of Python.
  4. Weights All correlation functions now support arbitrary, user-specified weights for individual points
  5. Modular The code is written in a modular fashion and is easily extensible to compute arbitrary clustering statistics.
  6. Future-proof As we 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.

Benchmark against Existing Codes

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.

Installation

Pre-requisites

  1. make >= 3.80
  2. OpenMP capable compiler like icc, gcc>=4.6 or clang >= 3.7. If not available, please disable USE_OMP option option in theory.options and mocks.options. On a HPC cluster, consult the cluster documentation for how to load a compiler (often module load gcc or similar). If you are using Corrfunc with Anaconda Python, then conda install gcc (MAC/linux) should work. On MAC, (sudo) port install gcc5 is also an option.
  3. gsl >= 2.4. On an HPC cluster, consult the cluster documentation (often module load gsl will work). With Anaconda Python, use conda install -c conda-forge gsl (MAC/linux). On MAC, you can use (sudo) port install gsl (MAC) if necessary.
  4. python >= 2.7 or python>=3.4 for compiling the CPython extensions.
  5. numpy>=1.7 for compiling the CPython extensions.
$ git clone https://github.com/manodeep/Corrfunc.git
$ cd Corrfunc
$ make
$ make install
$ python -m pip install . [--user]

$ make tests  # run the C tests
$ python -m pip install pytest
$ python -m pytest  # run the Python 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 CPython extensions in your environment, then python -m pip install . [--user] should be sufficient. If you are primarily interested in the Python interface, you can condense all of the steps by using python -m pip install . [--user] --install-option="CC=yourcompiler" after git clone [...] and cd Corrfunc.

Compilation Notes

  • If Python and/or numpy are not available, then the CPython extensions will not be compiled.
  • make install simply copies files into the lib/bin/include sub-directories. You do not need root permissions
  • Default compiler on MAC is set to clang, if you want to specify a different compiler, you will have to call make 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 -m pip install . [--user] --install-option="CC=yourcompiler", please run a make distclean beforehand (especially if switching compilers)
  • Please note that Corrfunc is compiling with optimizations for the architecture it is compiled on. That is, it uses gcc -march=native or similar. For this reason, please try to compile Corrfunc on the architecture it will be run on (usually this is only a concern in heterogeneous compute environments, like an HPC cluster with multiple node types). In many cases, you can compile on a more capable architecture (e.g. with AVX-512 support) then run on a less capable architecture (e.g. with only AVX2), because the runtime dispatch will select the appropriate kernel. But the non-kernel elements of Corrfunc may emit AVX-512 instructions due to -march=native. If an Illegal instruction error occurs, then you'll need to recompile on the target architecture.

Installation notes

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.

For Python tests, please run python -m pip install pytest and python -m pytest from the Corrfunc root dir.

While we 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.

Method 2: pip installation

The Python package is directly installable via python -m pip install Corrfunc. However, in that case you will lose the ability to recompile the code. This usually fine if you are only using the Python interface and are on a single machine, like a laptop. For usage on a cluster or other environment with multiple CPU architectures, you may find it more useful to use the Source Installation method above in case you need to compile for a different architecture later.

Testing a pip-installed Corrfunc

You can check that a pip-installed Corrfunc is working with:

$ python -m pytest --pyargs Corrfunc

The pip installation does not include all of the test data contained in the main repo, since it would total over 100 MB and the tests that generate on-the-fly data are similarly exhaustive. pytest will mark tests where the data files are not availabe as "skipped". If you would like to run the data-based tests, please use the Source Installation method.

OpenMP on OSX

Automatically detecting OpenMP support from the compiler and the runtime is a bit tricky. If you run into any issues compiling (or running) with OpenMP, please refer to the FAQ for potential solutions.

Clustering Measures on simulated galaxies

Input data

The input galaxies (or any discrete distribution of points) are derived from a simulation. For instance, the galaxies could be a result of an Halo Occupation Distribution (HOD) model, a Subhalo Abundance matching (SHAM) model, a Semi-Empirical model (SEM), or a Semi-Analytic model (SAM) etc. The input set of points can also be the dark matter halos, or the dark matter particles from a cosmological simulation. The input set of points are expected to have positions specified in Cartesian XYZ.

Types of available clustering statistics

All codes that work on cosmological boxes with co-moving positions are located in the theory directory. The various clustering measures are:

  1. DD -- Measures auto/cross-correlations between two boxes. The boxes do not need to be cubes.
  2. xi -- Measures 3-d auto-correlation in a cubic cosmological box. Assumes PERIODIC boundary conditions.
  3. wp -- Measures auto 2-d point projected correlation function in a cubic cosmological box. Assumes PERIODIC boundary conditions.
  4. DDrppi -- Measures the auto/cross correlation function between two boxes. The boxes do not need to be cubes.
  5. DDsmu -- Measures the auto/cross correlation function between two boxes. The boxes do not need to be cubes.
  6. vpf -- Measures the void probability function + counts-in-cells.

Clustering measures on observed galaxies

Input data

The input galaxies are typically observed galaxies coming from a large-scale galaxy survey. In addition, simulated galaxies that have been projected onto the sky (i.e., where observational systematics have been incorporated and on-sky positions have been generated) can also be used. We generically refer to both these kinds of galaxies as "mocks".

The input galaxies are expected to have positions specified in spherical co-ordinates with at least right ascension (RA) and declination (DEC). For spatial correlation functions, an approximate "co-moving" distance (speed of light multiplied by redshift, CZ) is also required.

Types of available clustering statistics

All codes that work on mock catalogs (RA, DEC, CZ) are located in the mocks directory. The various clustering measures are:

  1. DDrppi_mocks -- The standard auto/cross correlation between two data sets. The outputs, DD, DR and RR can be combined using wprp to produce the Landy-Szalay estimator for wp(rp).
  2. DDsmu_mocks -- The standard auto/cross correlation between two data sets. The outputs, DD, DR and RR can be combined using the Python utility convert_3d_counts_to_cf to produce the Landy-Szalay estimator for xi(s, mu).
  3. DDtheta_mocks -- Computes angular correlation function between two data sets. The outputs from DDtheta_mocks need to be combined with wtheta to get the full omega(theta)
  4. vpf_mocks -- Computes the void probability function on mocks.

Science options

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)

  1. PERIODIC (ignored in case of wp/xi) -- switches periodic boundary conditions on/off. Enabled by default.
  2. OUTPUT_RPAVG -- switches on output of <rp> in each rp bin. Can be a massive performance hit (~ 2.2x in case of wp). Disabled by default.

Mocks (in mocks.options)

  1. OUTPUT_RPAVG -- switches on output of <rp> in each rp bin for DDrppi_mocks. Enabled by default.
  2. OUTPUT_THETAAVG -- switches on output of in each theta bin. Can be extremely slow (~5x) depending on compiler, and CPU capabilities. Disabled by default.
  3. LINK_IN_DEC -- creates binning in declination for DDtheta_mocks. 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.
  4. LINK_IN_RA -- creates binning in RA once binning in DEC has been enabled for DDtheta_mocks. Same numerical issues as LINK_IN_DEC
  5. FAST_ACOS -- Relevant only when OUTPUT_THETAAVG is enabled for DDtheta_mocks. Disabled by default. An arccos is required to calculate <\theta>. In absence of vectorized arccos (intel compiler, icc provides one via intel Short Vector Math Library), this calculation is extremely slow. However, we can approximate arccos using polynomials (with Remez Algorithm). The approximations are taken from implementations released by Geometric Tools. Depending on the level of accuracy desired, this implementation of fast 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.
  6. COMOVING_DIST -- Currently there is no support in Corrfunc for different cosmologies. However, for the mocks routines like, DDrppi_mocks and vpf_mocks, cosmology parameters are required to convert between redshift and co-moving distance. Both DDrppi_mocks and vpf_mocks expects to receive a redshift array as input; however, with this option enabled, the redshift 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.

Common Code options for both Mocks and Theory

  1. DOUBLE_PREC -- switches on calculations in double precision. Calculations are performed in double precision when enabled. This option is disabled by default in theory and enabled by default in the mocks routines.
  2. 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) for DDrppi and wp. Enabled by default. The Makefile will compare the CC variable with known OpenMP enabled compilers and set compile options accordingly. Set in common.mk by default.
  3. ENABLE_MIN_SEP_OPT -- uses some further optimisations based on the minimum separation between pairs of cells. Enabled by default.
  4. COPY_PARTICLES -- whether or not to create a copy of the particle positions (and weights, if supplied). Enabled by default (copies of the particle arrays are created)
  5. FAST_DIVIDE -- Disabled by default. Divisions are slow but required DDrppi_mocks(r_p,\pi), DDsmu_mocks(s, \mu) and DD(s, \mu). 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.

Optimization for your architecture

  1. The values of bin_refine_factor and/or zbin_refine_factor in the countpairs\_\*.c files control the cache-misses, and consequently, the runtime. In trial-and-error methods, Manodeep has seen any values larger than 3 are generally slower for theory routines but can be faster for mocks. But some different combination of 1/2 for (z)bin_refine_factor might be faster on your platform.
  2. If you are using the angular correlation function and need thetaavg, you might benefit from using the INTEL MKL library. The vectorized trigonometric functions provided by MKL can provide significant speedup.

Running the codes

Read the documentation on corrfunc.rtfd.io.

Using the command-line interface

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.

Calling from C

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.

Calling from Python

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 CPython 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)

Author & Maintainers

Corrfunc was designed and implemented by Manodeep Sinha, with contributions from Lehman Garrison, Nick Hand, and Arnaud de Mattia. Corrfunc is currently maintained by Manodeep Sinha and Lehman Garrison.

Citing

If you use Corrfunc for research, please cite using the MNRAS code paper with the following bibtex entry:

@ARTICLE{2020MNRAS.491.3022S,
    author = {{Sinha}, Manodeep and {Garrison}, Lehman H.},
    title = "{CORRFUNC - a suite of blazing fast correlation functions on
    the CPU}",
    journal = {\mnras},
    keywords = {methods: numerical, galaxies: general, galaxies:
    haloes, dark matter, large-scale structure of Universe, cosmology:
    theory},
    year = "2020",
    month = "Jan",
    volume = {491},
    number = {2},
    pages = {3022-3041},
    doi = {10.1093/mnras/stz3157},
    adsurl =
    {https://ui.adsabs.harvard.edu/abs/2020MNRAS.491.3022S},
    adsnote = {Provided by the SAO/NASA
    Astrophysics Data System}
}

If you are using Corrfunc v2.3.0 or later, and you benefit from the enhanced vectorised kernels, then please additionally cite this paper:

@InProceedings{10.1007/978-981-13-7729-7_1,
    author="Sinha, Manodeep and Garrison, Lehman",
    editor="Majumdar, Amit and Arora, Ritu",
    title="CORRFUNC: Blazing Fast Correlation Functions with AVX512F SIMD Intrinsics",
    booktitle="Software Challenges to Exascale Computing",
    year="2019",
    publisher="Springer Singapore",
    address="Singapore",
    pages="3--20",
    isbn="978-981-13-7729-7",
    url={https://doi.org/10.1007/978-981-13-7729-7_1}
}

Mailing list

If you have questions or comments about the package, please do so on the mailing list: https://groups.google.com/forum/#!forum/corrfunc

LICENSE

Corrfunc is released under the MIT license. Basically, do what you want with the code, including using it in commercial application.

Project URLs