johnh2o2 / cudastet

GPU algorithms for Stetson J variability statistic

CudaStet

Provides both CPU and GPU functions for the Stetson J statistic (Stetson 1996).

Requires:

• CUDA-enabled device
• CUDA Toolkit >= 7.5
• A suitable C++ compiler (e.g. g++)

To install:

1. Clone this repository

2. Edit the Makefile accordingly

   • ARCH -- the Compute Capability number for your device, e.g. 5.2 becomes 52
   • REAL_TYPE -- you can switch between double and single precision by editing this Makefile variable to either double or float.
   • CUDA_VERSION -- make sure you have the CUDA Toolkit installed; this code was developed using version 7.5
   • The BLOCK_SIZE is also something to play around with but is optional to change.

3. Run make to generate a binary that you can run for testing. This is mainly for my own debugging purposes, but you can inspect the code to get an idea of how to call the stetson_j_cpu and stetson_j_gpu functions.

To use the compiled binary

usage: ./CudaStet <filename> <{s, l}> <skiprows>

  filename      : path to datafile with either a list of filenames
                  or a path to a single file containing 'x y yerr' 
                  on each line
  s             : single file
  l             : list of files
  skiprows      : number of rows to skip in the datafile

To call stetson_j_{gpu, cpu}:

Arguments (in order):

• real_type *x : values for independent variable
• real_type *y : values for dependent variable
• real_type *err : values for uncertainty of dependent variable
• weight_type WEIGHTING : Must be one of
  • CONSTANT : all pairs of observations are weighted equally
  • EXP : pairs of observations are exponentially suppressed by their distance in x (uses exp(-|t1 - t2| / median(dt))). See Zhang et. al. 2003 for a real-world application of this weighting scheme. Note however that Zhang et. al. 2003 used the mean function instead of the median.
• int N : number of datapoints

Notes

• Be careful with single precision. If you're using the exponential weighting scheme, the accuracy for ~50,000 datapoints is about 10^(-3) for both the CPU and GPU variants. For a constant weighting scheme, however, the CPU variant is very inaccurate (off by factors of 100 or more) at 50,000 datapoints, while the GPU variant remains accurate to a factor of 10^(-3).

Goals for the future

• Python bindings with Swig
• Addition of Stetson K and L variability indices
• Adding configure script to simplify the install process

Timing

Some tests run on an Ubuntu 14.04 desktop with an i7-5930K overclocked to 4.5GHz, and a 980 Ti graphics card. For these timing tests, CudaStet was compiled with single precision and using the -O3 optimization flag for g++ and --use_fast_math for nvcc.

Read as: N dt

Where N is the number of data points, dt is the execution time in seconds.

timing : (CPU, EXPONENTIAL)
10         9.000e-06 
1009       3.612e-02 
2008       9.482e-02 
3007       1.367e-01 
4006       1.922e-01 
5005       2.986e-01 
6004       4.134e-01 
7003       4.819e-01 
8002       6.265e-01 
9001       7.888e-01 
timing : (GPU, EXPONENTIAL)
10         1.648e-01 
1009       1.351e-03 
2008       2.490e-03 
3007       3.450e-03 
4006       4.739e-03 
5005       5.822e-03 
6004       7.672e-03 
7003       8.494e-03 
8002       9.271e-03 
9001       1.066e-02 
timing : (CPU, CONSTANT)
10         1.000e-06 
1009       5.004e-03 
2008       1.877e-02 
3007       3.833e-02 
4006       6.095e-02 
5005       9.584e-02 
6004       1.385e-01 
7003       1.869e-01 
8002       2.532e-01 
9001       2.829e-01 
timing : (GPU, CONSTANT)
10         1.740e-04 
1009       9.020e-04 
2008       1.727e-03 
3007       3.200e-03 
4006       4.346e-03 
5005       4.290e-03 
6004       5.602e-03 
7003       7.210e-03 
8002       7.446e-03 
9001       8.685e-03