[static_h] Switch Math.random() to faster 64-bit seeded implementation. (#1175) #1180
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…k#1175) Summary: The `std::minstd_rand` implementation has its `seed()` value type defined as `result_type` which is aliased to `unsigned int`, a 32-bit value on supported platforms. Using 32-bit seeds and results can lead to real-world [birthday problem](https://en.wikipedia.org/wiki/Birthday_problem) collisions. As per Purujit's recommendation, we update the implementation of `Math.random()` to utilize a 64-bit LCG that can accept a 64-bit seed and also generate a 64-bit result. There is a lot of [analysis](https://prng.di.unimi.it/) and [blog posts](https://v8.dev/blog/math-random) around the best way to implement random number generators. This PR attempts to deliver two improvements on the current implementation: * Having 64-bit seeded values and result types instead of 32-bit. * Ideally, being faster or at least the same speed. * (optional) having better randomness properties. Addressing facebook#1169, we benchmark the `xoroshiro128+` (used by browsers); `randomDevice` (cryptographically secure); `mt19937_64` (complex but fast) and `lcg64` (simple & fast) unsigned 64-bit integer PRNGs against three different uniform random distribution implementations: 1. **`std::uniform_real_distribution()`**: standard library implementation, but slow as it requires multiple random numbers to be generated for a single [0, 1) value; plus it has known bugs where it can in fact return 1 because of rounding. 2. **Bit Twiddle Approach 1**: use the `(0x3FFL << 52 | uint64_t(x) >>> 12) - 1.0` from Java and Firefox that pairs a fixed exponent with 52 random mantissa bits to generate a double between [1, 2) and then subtracting 1. 3. **Bit Twiddle Approach 2**: use the `(uint64_t(x) >> 11) * 0x1.0p-53` approach which generates a double between [0, 1) directly with 53 bits worth of possible output (twice as many values as the 52 bits from Approach 1). The bit twiddling approaches are described [here](https://prng.di.unimi.it/#:~:text=Generating%20uniform%20doubles%20in%20the%20unit%20interval). Given the large number of platforms and architectures that Hermes is compiled to, we make copious use of `static_assert`s to make sure that we don't encounter unexpected rug-pulls or changing bit-widths when upgrading compilers; or adding new targets. These should of course be elided after compilation and so will not affect runtime performance. ### Alternatives Considered - Instead of manually bit packing a 64-bit seed, an implementation leveraging C++11's `seed_seq` was explored that would ideally provide even more entropy, however the author of [this article](https://www.pcg-random.org/posts/cpp-seeding-surprises.html) asserts that 64 bits of seed seq data does not necessarily produce 64 bits of output. As such, we keep it simple with a single seed value so it is easier to reason with in the future. Pull Request resolved: facebook#1175 Test Plan: We compile the following `benchmark.js` to byte-code using `../bin/hermes --emit-binary -fstatic-builtins -O ./benchmark.js -out ./benchmark.out.bin`: ``` const n_iterations = 10_000_000; let sum = 0, min = 2, max = -2; const rndFunction = Math.random; for (const i = 0; i < n_iterations; i++) { const rnd = rndFunction(); if (rnd < min) min = rnd; if (rnd > max) max = rnd; sum += rnd; } print(JSON.stringify({min, avg: sum / n_iterations, max})); ``` And then we execute this bytecode on each of the implementation's resulting `hvm` binary (I'm using `hyperfine`) on my local development machine (Apple MacBook Air M2 with ARM64): | Command | Mean [s] | Min [s] | Max [s] | Relative | |:---|---:|---:|---:|---:| | `./hvm-main ./benchmark.out.bin` | 8.371 ± 0.065 | 8.310 | 8.508 | 1.05 ± 0.01 | | `./hvm-xoroshiro128plus-twiddlebits ./benchmark.out.bin` | 8.075 ± 0.023 | 8.040 | 8.127 | 1.01 ± 0.00 | | **`./hvm-xoroshiro128plus-twiddlebits2 ./benchmark.out.bin`** | 8.019 ± 0.033 | 7.975 | 8.068 | 1.00 ± 0.01 | | `./hvm-lcg64-urd ./benchmark.out.bin` | 8.288 ± 0.041 | 8.232 | 8.369 | 1.04 ± 0.01 | | `./hvm-lcg64-twiddlebits ./benchmark.out.bin` | 7.988 ± 0.030 | 7.949 | 8.039 | 1.00 | | `./hvm-lcg64-twiddlebits2 ./benchmark.out.bin` | 8.010 ± 0.029 | 7.965 | 8.058 | 1.00 ± 0.01 | | `./hvm-mt19937_64-urd ./benchmark.out.bin` | 8.109 ± 0.047 | 8.058 | 8.190 | 1.02 ± 0.01 | | `./hvm-mt19937_64-twiddlebits ./benchmark.out.bin` | 8.135 ± 0.055 | 8.063 | 8.262 | 1.02 ± 0.01 | | `./hvm-mt19937_64-twiddlebits2 ./benchmark.out.bin` | 8.143 ± 0.039 | 8.067 | 8.187 | 1.02 ± 0.01 | | `./hvm-mt19937_64-urd-hoisted ./benchmark.out.bin` | 8.028 ± 0.029 | 7.977 | 8.082 | 1.01 ± 0.01 | | `./hvm-randomDevice-urd ./benchmark.out.bin` | 9.859 ± 0.036 | 9.800 | 9.918 | 1.23 ± 0.01 | | `./hvm-randomDevice-twiddlebits ./benchmark.out.bin` | 9.514 ± 0.028 | 9.463 | 9.550 | 1.19 ± 0.01 | | `./hvm-randomDevice-twiddlebits2 ./benchmark.out.bin` | 9.558 ± 0.033 | 9.498 | 9.610 | 1.20 ± 0.01 | Although `xoroshiro128plus-twiddlebits2` appears optimal, neildhar preferred `mt19937_64-urd` as it's less invasive a change and has less complexity to maintain. Reviewed By: avp Differential Revision: D50792073 Pulled By: neildhar fbshipit-source-id: 9a5a829dd32adf8986b912e4f424b48a6f937bd0
facebook-github-bot
added
the
CLA Signed
|
@tmikov has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
This was picked up by the periodic sync we do from main to static_h, so this PR can be closed. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Cherry-picking #1175 for
static_h.