This repo contains the CUDA implementation of five gamma generator kernels and benchmarking code for the M.Sc. thesis : Gamma Random Number Generation on GPUs using CUDA

The best performing kernels on all GPUs tested are cheng1977 and marsaglia_tsang

For usage on GPUs I recommend using either cheng1977 or marsaglia_tsang, the best performing kernel can depend on your GPU type, but they should outperform all other kernels for all values of alpha larger than 1. More information about the the theory of RNGs and the gamma distribution and further references can be found in the thesis.

The following gamma generators were selected as potential candidates for effective generators on GPU architectures and implemented in CUDA (gamma_generators.cuh):

marsaglia_tsang - The generator (without squeeze step) from:

• George Marsaglia and Wai Wan Tsang. A simple method for generating gamma variables. ACM Trans. Math. Softw. 26, 3 (Sept. 2000): pp. 363–372. https://doi.org/10.1145/358407.358414

cheng1977 - The generator (GA) from:

• R. C. H. Cheng. The Generation of Gamma Variables with Non-Integral Shape Parameter. Journal of the Royal Statistical Society. Series C (Applied Statistics) 26, no. 1 (1977): pp. 71–75. https://doi.org/10.2307/2346871

GKM1, GKM2, GKM3 - The generators with corresponding names from:

• R. C. H. Cheng and G. M. Feast. Some Simple Gamma Variate Generators. Journal of the Royal Statistical Society. Series C (Applied Statistics) 28.3 (1979): pp. 290–295. https://doi.org/10.2307/2347200.

GC - The generator (GA) from:

• J.H. Ahrens and U. Dieter. Computer methods for sampling from gamma, beta, poisson and bionomial distributions. *ComputingÄ 12.3 (1974): pp. 223–246. https://doi.org/10.1007/BF02293108.

best1978 - The generator (XG) from:

• D. J. Best. Letters to the Editors. Journal of the Royal Statistical Society: Series C (Applied Statistics) 27.2 (1978): pp. 181–182. https://doi.org/10.1111/j.1467-9876.1978.tb01041.x.

The build system used is cmake and the target gamma_kernel_benchmark correspond to the benchmark executable. An example of how to build and execute the code, and analyze the benchmark output can be found in the google COLAB notebook: gammaRNG.

The kernels are written to be used inlined and to be compiled with the O3 optimization flag. If you would like to use the kernel with lower optimization or run out of registers (and are unable to inline), then I highly recommend you to split up the kernels into:

1. a setup step (that initialize the constants before the do loop).
2. a rejection loop (coresponding to the do loop).