[SYSTEMDS-3556] Counter based random number generator by ichbinstudent · Pull Request #2186 · apache/systemds

ichbinstudent · 2025-01-23T11:49:24Z

Summary

Adding counter based RNG which improves random quality and speed

Details

The PRNG uses Philox4x64_10 to generate batches of random double values with a 64 bit randomness. The algorithm was tested for correctness by comparing its output for certain counters and keys with an existing implementation in c.
This reference implementation called openRAND was tested by the authors using various statistical methods described here.

Advantages

Quality: Instead of 32 bits in java.util.Random, Philox4_64 produces 64 bits of randomness. While java.util.Random has a period of only (2^48), while the period of Philox4_64 is 2^256 - 1.
Speed: While the Java version of Philox4x64_10 is only about half as fast as java.util.Random, there is a Cuda kernel version available producing the exact same sequence of random numbers. This means that the Cuda and Java versions can be used interchangeably. If a system has support for Cuda, the kernel ist used, if not, the Java version can be used as a fallback. The kernel version is around 200 times faster than java.util.Random, and even faster, if the results are not copied to the CPU but kept in the GPU's memory.
Parallelisation: When using state based PRNGs, it is impossible to generate the same random matrix when changing the block size. With counter based PRNGs it is possible to change the block size but still compute the same random matrix by using the global index (row * row_size + col) as the counter.

…hbinstudent@users.noreply.github.com> Signed-off-by: chris-1187 <christian.munz@posteo.de>

…boxmuller

codecov · 2025-01-23T19:44:44Z

Codecov Report

Attention: Patch coverage is 86.25954% with 18 lines in your changes missing coverage. Please review.

Project coverage is 72.47%. Comparing base (9484f11) to head (fe8a986).
Report is 117 commits behind head on main.

Files with missing lines	Patch %	Lines
...ysds/runtime/util/PhiloxUniformCBPRNGenerator.java	90.58%	4 Missing and 4 partials ⚠️
...he/sysds/runtime/matrix/data/LibMatrixDatagen.java	68.42%	3 Missing and 3 partials ⚠️
...sysds/runtime/util/PhiloxNormalCBPRNGenerator.java	80.00%	2 Missing and 2 partials ⚠️

Additional details and impacted files

@@             Coverage Diff              @@
##               main    #2186      +/-   ##
============================================
+ Coverage     71.84%   72.47%   +0.63%     
- Complexity    44596    45465     +869     
============================================
  Files          1448     1472      +24     
  Lines        168967   170995    +2028     
  Branches      32934    33331     +397     
============================================
+ Hits         121393   123937    +2544     
+ Misses        38244    37641     -603     
- Partials       9330     9417      +87

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

Baunsgaard

Many good things here, please address the minor comments, and describe the performance differences in the PR.

Next steps, is to integrate it into the compiler as options, i would suggest to do it via arguments to random "CB_uniform" could be the argument for instance.

src/main/java/org/apache/sysds/runtime/matrix/data/LibMatrixDatagen.java

src/main/java/org/apache/sysds/runtime/matrix/data/RandomMatrixGenerator.java

src/test/java/org/apache/sysds/test/component/matrix/LibMatrixDatagenTest.java

Baunsgaard · 2025-01-26T13:17:34Z

If you resolve some of the comments, please feel free to mark them as such.

Signed-off-by: chris-1187 <christian.munz@posteo.de>

Baunsgaard · 2025-01-28T22:53:30Z

src/main/java/org/apache/sysds/runtime/matrix/data/LibMatrixDatagen.java

+		if (valuePRNG instanceof PRNGenerator) {
+			rngStream = Stream.generate(() -> min + (range * ((PRNGenerator) valuePRNG).nextDouble())).iterator();
+		} else if (valuePRNG instanceof CounterBasedPRNGenerator) {
+			rngStream = Arrays.stream(((CounterBasedPRNGenerator)valuePRNG).getDoubles(ctr, blockrows * blockcols)).map(i -> min + (range * i)).iterator();


yup the stream looks nice, however, since it modifies old code, we need to be careful that the new code returns the same values as the previous.

…ame.

scripts/staging/cuda-counter-based-prng/readme.md

scripts/staging/cuda-counter-based-prng/PhiloxJNvrtcExample.java

scripts/staging/cuda-counter-based-prng/philox_kernel.ptx

Signed-off-by: chris-1187 <christian.munz@posteo.de>

Removed because can be compiled from the provided sources

ichbinstudent · 2025-02-14T20:12:14Z

Ok, deleted the compiled kernel file.

This commit adds a new random generator for SystemDS, that improves the speed at witch we generate random matrices. The algorithm uses Philox4x64_10 to generate batches of random double values with 64 bit randomness. The implementation is based of an implementation in openRAND, and verified on various statistical methods. - Quality: Instead of 32 bits in java.util.Random, Philox4_64 produces 64 bits of randomness. While java.util.Random has a period of only (2^48), while the period of Philox4_64 is 2^256 - 1. - Speed: While the Java version of Philox4x64_10 is only about half as fast as java.util.Random, there is a Cuda kernel version available producing the exact same sequence of random numbers. This means that the Cuda and Java versions can be used interchangeably. If a system has support for Cuda, the kernel is used, if not, the Java version can be used as a fallback. The kernel version is around 200 times faster than java.util.Random, and even faster, if the results are not copied to the CPU but kept in the GPU's memory. - Parallelisation: When using state based PRNGs, it is impossible to generate the same random matrix when changing the block size. With counter based PRNGs it is possible to change the block size but still compute the same random matrix by using the global index (row * row_size + col) as the counter. Co-authored-by: ichbinstudent <45435943+ichbinstudent@users.noreply.github.com> Co-authored-by: chris-1187 <christian.munz@posteo.de> Closes apache#2186

Baunsgaard · 2025-05-13T09:47:53Z

Hi @ichbinstudent and @chris-1187

I will take this PR over and get it merged. See #2260

Thanks for the contribution.

This commit adds a new random generator for SystemDS, that improves the speed at witch we generate random matrices. The algorithm uses Philox4x64_10 to generate batches of random double values with 64 bit randomness. The implementation is based of an implementation in openRAND, and verified on various statistical methods. - Quality: Instead of 32 bits in java.util.Random, Philox4_64 produces 64 bits of randomness. While java.util.Random has a period of only (2^48), while the period of Philox4_64 is 2^256 - 1. - Speed: While the Java version of Philox4x64_10 is only about half as fast as java.util.Random, there is a Cuda kernel version available producing the exact same sequence of random numbers. This means that the Cuda and Java versions can be used interchangeably. If a system has support for Cuda, the kernel is used, if not, the Java version can be used as a fallback. The kernel version is around 200 times faster than java.util.Random, and even faster, if the results are not copied to the CPU but kept in the GPU's memory. - Parallelisation: When using state based PRNGs, it is impossible to generate the same random matrix when changing the block size. With counter based PRNGs it is possible to change the block size but still compute the same random matrix by using the global index (row * row_size + col) as the counter. Co-authored-by: ichbinstudent <45435943+ichbinstudent@users.noreply.github.com> Co-authored-by: chris-1187 <christian.munz@posteo.de> Closes apache#2186

This commit adds a new random generator for SystemDS, that improves the speed at witch we generate random matrices. The algorithm uses Philox4x64_10 to generate batches of random double values with 64 bit randomness. The implementation is based of an implementation in openRAND, and verified on various statistical methods. - Quality: Instead of 32 bits in java.util.Random, Philox4_64 produces 64 bits of randomness. While java.util.Random has a period of only (2^48), while the period of Philox4_64 is 2^256 - 1. - Speed: While the Java version of Philox4x64_10 is only about half as fast as java.util.Random, there is a Cuda kernel version available producing the exact same sequence of random numbers. This means that the Cuda and Java versions can be used interchangeably. If a system has support for Cuda, the kernel is used, if not, the Java version can be used as a fallback. The kernel version is around 200 times faster than java.util.Random, and even faster, if the results are not copied to the CPU but kept in the GPU's memory. - Parallelisation: When using state based PRNGs, it is impossible to generate the same random matrix when changing the block size. With counter based PRNGs it is possible to change the block size but still compute the same random matrix by using the global index (row * row_size + col) as the counter. Closes apache#2186 Closes apache#2260 Co-authored-by: ichbinstudent <45435943+ichbinstudent@users.noreply.github.com> Co-authored-by: chris-1187 <christian.munz@posteo.de>

chris-1187 and others added 2 commits January 20, 2025 19:02

Counter-based Philox RNG - Co-authored-by: ichbinstudent <45435943+ic…

3233829

…hbinstudent@users.noreply.github.com> Signed-off-by: chris-1187 <christian.munz@posteo.de>

Added unittests and added counter based normal distributed rng using …

fc01ff5

…boxmuller

Baunsgaard reviewed Jan 24, 2025

View reviewed changes

ichbinstudent added 4 commits January 26, 2025 13:50

fix imports

3a6df20

Improve unit test and remove unnecessary tests

75f224f

Add missing license

30ac4e2

refactor genFullyDense

e54c9f3

ichbinstudent and others added 3 commits January 26, 2025 14:21

Removed wildcard import

8dc6975

Removed wildcard import

931f317

RandCounterBased DML

79c58c4

Signed-off-by: chris-1187 <christian.munz@posteo.de>

Baunsgaard reviewed Jan 28, 2025

View reviewed changes

ichbinstudent added 2 commits January 29, 2025 21:57

Added test to make sure values generated when using streams are the s…

a8d63ff

…ame.

Add instructions on how to use the cuda version to the staging folder

543f3b2

Baunsgaard reviewed Feb 3, 2025

View reviewed changes

scripts/staging/cuda-counter-based-prng/readme.md Outdated Show resolved Hide resolved

Baunsgaard reviewed Feb 3, 2025

View reviewed changes

scripts/staging/cuda-counter-based-prng/PhiloxJNvrtcExample.java Show resolved Hide resolved

Baunsgaard reviewed Feb 3, 2025

View reviewed changes

scripts/staging/cuda-counter-based-prng/philox_kernel.ptx Outdated Show resolved Hide resolved

Added licenses, shortened readme.md

fc578c6

Signed-off-by: chris-1187 <christian.munz@posteo.de>

Baunsgaard changed the title ~~SYSTEMDS-3556 Counter based random number generator~~ [SYSTEMDS-3556] Counter based random number generator Feb 5, 2025

Delete scripts/staging/cuda-counter-based-prng/philox_kernel.ptx

fe8a986

Removed because can be compiled from the provided sources

Baunsgaard mentioned this pull request May 13, 2025

[SYSTEMDS-3556] Counter-based random number generator #2260

Merged

Baunsgaard closed this May 13, 2025

github-project-automation bot moved this from In Progress to Done in SystemDS PR Queue May 13, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[SYSTEMDS-3556] Counter based random number generator#2186

[SYSTEMDS-3556] Counter based random number generator#2186
ichbinstudent wants to merge 13 commits intoapache:mainfrom
chris-1187:counter-based-rng

ichbinstudent commented Jan 23, 2025

Uh oh!

codecov bot commented Jan 23, 2025 •

edited

Loading

Uh oh!

Baunsgaard left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Baunsgaard commented Jan 26, 2025

Uh oh!

Baunsgaard Jan 28, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ichbinstudent commented Feb 14, 2025

Uh oh!

Baunsgaard commented May 13, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

ichbinstudent commented Jan 23, 2025

Summary

Details

Advantages

Uh oh!

codecov bot commented Jan 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Baunsgaard left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Baunsgaard commented Jan 26, 2025

Uh oh!

Baunsgaard Jan 28, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ichbinstudent commented Feb 14, 2025

Uh oh!

Baunsgaard commented May 13, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

codecov bot commented Jan 23, 2025 •

edited

Loading