This is an automated email from the ASF dual-hosted git repository.
aherbert pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/commons-rng.git
The following commit(s) were added to refs/heads/master by this push:
new 7a45d6ff Update PHILOX_4X64 Gaussian sampler performance
7a45d6ff is described below
commit 7a45d6ff18b92bfc56e32d546e2ffc5f3e189a1d
Author: Alex Herbert <[email protected]>
AuthorDate: Tue Apr 21 13:31:13 2026 +0100
Update PHILOX_4X64 Gaussian sampler performance
---
src/site/apt/userguide/rng.apt | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/src/site/apt/userguide/rng.apt b/src/site/apt/userguide/rng.apt
index d974eba1..051b5a02 100644
--- a/src/site/apt/userguide/rng.apt
+++ b/src/site/apt/userguide/rng.apt
@@ -975,7 +975,7 @@ double[] coordinate = sampler.sample();
The <<<RandomSource.JDK>>> generator uses thread-safe (synchronized)
<<<int>>> generation which has a performance overhead (see the <<<int>>>
generation results). Note that the output will be low quality and this
generator should not be used. See the {{{a5._Quality}Quality}} section for
details. Multi-threaded applications should use a generator for each thread.
The speed of <<<boolean>>> generation is related to the base implementation
that caches the 32-bit or 64-bit output from the generator. In these results
the 32-bit generators have the better performance. These timings are relative
and all implements are very fast. A RNG to compute boolean samples should be
chosen based on the {{{a5._Quality}quality}} of the output.
-
+
The <<<RandomSource.PHILOX_4X64>>> generator uses multiply high methods from
<<<java.lang.Math>>> if available. The <<<multiplyHigh>>> (JDK 9+) and
<<<unsignedMultiplyHigh>>> (JDK 18+) significantly increase performance if the
128-bit product of two 64-bit factors is supported by hardware instructions.
These results are on a platform with supported hardware.
@@ -1102,7 +1102,7 @@ double[] coordinate = sampler.sample();
*-----------------------+---------:---------:---------:---------:
| PHILOX_4X32 | 1.08207 | 1.07704 | 0.38838 | 0.37289 |
*-----------------------+---------:---------:---------:---------:
-| PHILOX_4X64 | 1.12164 | 1.37654 | 0.52689 | 0.51213 |
+| PHILOX_4X64 | 0.91379 | 0.91941 | 0.26007 | 0.24542 |
*-----------------------+---------:---------:---------:---------:
@@ -1110,6 +1110,8 @@ double[] coordinate = sampler.sample();
The reference <<<java.util.Random>>> nextGaussian() method uses synchronized
method calls per sample. The <<<RandomSource.JDK>>> RNG will use synchronized
method calls when generating numbers for the
<<<BoxMullerNormalizedGaussianSampler>>> but the calls to obtain the samples
are not synchronized, hence the observed difference. All the other RNGs are not
synchronized.
+ The <<<RandomSource.PHILOX_4X64>>> generator uses multiply high methods from
<<<java.lang.Math>>> if available. The <<<multiplyHigh>>> (JDK 9+) and
<<<unsignedMultiplyHigh>>> (JDK 18+) significantly increase performance if the
128-bit product of two 64-bit factors is supported by hardware instructions.
These results are on a platform with supported hardware.
+
5. Quality
