This replaces the current Mersenne Twister with a new pseudorandom number generator known as Small Fast Counter 32. SFC32 seemed to be the best all around PRNG that I could find that worked well across the targets I can test on (HTML5 Electron and Firefox, Windows...yeah still limited here, sorry).
SFC32 was always faster or generated more numbers in a shorter time period than the current implementation in a few various test cases that I tried out which are all obviously contrived and more about trying to identify a noticeable difference rather than trying to extract an exact value. Those cases were the following:
Out of a list of 50k entities, select an entity and create a new Vector2 with a new random X and new random Y float position for it
Generate an array of 50k ints
Generate as many ints as possible in 10 seconds
For case 1 and 2, this was done over a period of 10 seconds 60 times per second. This was using a timetask, however these tests weren't stressful enough to trigger a slowdown which would mess with the final timings (I made sure to check for that). For case 3, I basically just ended up getting bottlenecked somewhere across all targets for both implementations. Dunno why, don't really care because SFC32 still wins.
For every case, the values are rough and eyeballed, but still show a trend. They were each also run once before hand before measurements were taken.
Test Case 1 (Vectors): Cumulative time spent
Impl
Windows
Electron
Firefox
SFC32
0.382s
0.997s
0.999s
MT
0.952s
1.396s
1.757s
Test Case 2 (Int Array): Cumulative time spent
Impl
Windows
Electron
Firefox
SFC32
0.126s
0.156s
0.485s
MT
0.427s
0.575s
0.911s
Test Case 3 (Shit ton of ints over 10 seconds): Total count
Impl
Windows
Electron
Firefox
SFC32
956000000?
1943000000?
1070000000?
MT
490000000?
630000000?
540000000?
SFC32 also comes with some bonuses like being more cryptographically secure than the Mersenne Twister according to some smart people who aren't me, and it is a public domain algorithm. You can verify that here: http://pracrand.sourceforge.net/
The implementation I used was derived from here, and slightly modified: https://github.com/bryc/code/blob/master/jshash/PRNGs.md#sfc32. Removing a few of the bitwise ORs resulted in no change to the output but slightly increased speed. The rest of them needed to remain, as removing them resulted in a slight speed decrease for the HTML5 targets.
Hopefully, this performance holds up across all targets that I am unable to test.
This replaces the current Mersenne Twister with a new pseudorandom number generator known as Small Fast Counter 32. SFC32 seemed to be the best all around PRNG that I could find that worked well across the targets I can test on (HTML5 Electron and Firefox, Windows...yeah still limited here, sorry).
SFC32 was always faster or generated more numbers in a shorter time period than the current implementation in a few various test cases that I tried out which are all obviously contrived and more about trying to identify a noticeable difference rather than trying to extract an exact value. Those cases were the following:
For case 1 and 2, this was done over a period of 10 seconds 60 times per second. This was using a timetask, however these tests weren't stressful enough to trigger a slowdown which would mess with the final timings (I made sure to check for that). For case 3, I basically just ended up getting bottlenecked somewhere across all targets for both implementations. Dunno why, don't really care because SFC32 still wins.
For every case, the values are rough and eyeballed, but still show a trend. They were each also run once before hand before measurements were taken.
SFC32 also comes with some bonuses like being more cryptographically secure than the Mersenne Twister according to some smart people who aren't me, and it is a public domain algorithm. You can verify that here: http://pracrand.sourceforge.net/
The implementation I used was derived from here, and slightly modified: https://github.com/bryc/code/blob/master/jshash/PRNGs.md#sfc32. Removing a few of the bitwise ORs resulted in no change to the output but slightly increased speed. The rest of them needed to remain, as removing them resulted in a slight speed decrease for the HTML5 targets.
Hopefully, this performance holds up across all targets that I am unable to test.
Thanks for reading my blog, have a nice day.