Opened 11 years ago
#6135 new Patches
Significant speed-up in gamma rng by switching from Knuth method to Marsaglia and Tsang (2000)
| Reported by: | Owned by: | No-Maintainer | |
|---|---|---|---|
| Milestone: | To Be Determined | Component: | random |
| Version: | Boost Development Trunk | Severity: | Optimization |
| Keywords: | gamma, rng | Cc: |
Description
When compared to the GNU Scientific Library (GSL), the generation of Gamma-distributed random numbers is very slow in Boost. A simple test, generating 100000000 random numbers, reveals
GSL : 22.3233 MSamples/s
Boost : 5.41275 MSamples/s
where MSamples/s is million samples per second. The difference is due to Boost using the Knuth method, whereas the GSL implements a newer method by Marsaglia and Tsang (2000).
With the attached patch, which implements the newer method, the performance improves to
GSL : 22.7472 MSamples/s
Boost : 14.5018 MSamples/s
The remaining difference between Boost and GSL can be traced back to GSL using Zigguraut rather than Box-Muller to generate Normal random numbers. At some point I might submit a patch for normal_distribution.hpp that implements the Zigguraut algorithm.
The attached patch exactly follows gsl_ran_gamma of GSL v1.15 (gsl-1.15/randdist/gamma.c)
