Simpler faster (illustration, for fun)

[lemire]

This replaces a 25-minute benchmark with a 19 seconds one, and it drastically improves the performance in some cases. I have not updated the README, and I image you won't want to merge the PR as is, but the performance boost is significant for some cases.

On an Ice Lake server (with 512-bit registers)....

This PR:

Benchmark               (testFile)   Mode  Cnt      Score   Error  Units
BenchJDK.jdk         /twitter.json  thrpt         952.397          ops/s
BenchJDK.scalar      /twitter.json  thrpt        2234.691          ops/s
BenchJDK.vector_128  /twitter.json  thrpt       14630.812          ops/s
BenchJDK.vector_256  /twitter.json  thrpt       17697.810          ops/s
BenchJDK.vector_512  /twitter.json  thrpt       31678.039          ops/s

In GB/s, it is...

jdk                  0.6 GB/s
scalar             1.4 GB/s
vector_128     9 GB/s
vector_256     11 GB/s
vector_512     20 GB/s

These are reasonable results !!!

Previous results (same machine):

Benchmark                          (branchPrediction)              (testFile)   Mode  Cnt       Score        Error  Units
BenchJDK.jdk                                      N/A           /twitter.json  thrpt    5    1730.981 ±     10.969  ops/s
BenchJDK.vector_128                               N/A           /twitter.json  thrpt    5    5312.508 ±     59.222  ops/s
BenchJDK.vector_256                               N/A           /twitter.json  thrpt    5    8186.286 ±     47.844  ops/s
BenchJDK.vector_512                               N/A           /twitter.json  thrpt    5   31273.521 ±    227.835  ops/s

So, in my tests, I triple the performance for vector_128 and double it for vector_256.

An interesting addition in this PR is a scalar validation function, which allows you to see directly the benefits of the SIMD operations.

This PR also retains just one benchmark file, because, again, I want quick benchmarks as I do not like to wait several minutes for the results. Ultimately, one would want the input file to be a parameter but I could not get it to work using the instructions in the README (the files are seemingly never found).

The important point here is that the performance can be made better.

[lemire]

On my ARM-based laptop, the scalar trick does not pay (since it is limited to 128-bit).... but you can still see the massive performance gain compared to scalar with vector_128. In that instance, the JDK 'string creation' trick is faster, likely due to the fact that we have some good optimization in the standard library.

BenchJDK.jdk         /twitter.json  thrpt        3857.591          ops/s
BenchJDK.scalar      /twitter.json  thrpt        1977.185          ops/s
BenchJDK.vector_128  /twitter.json  thrpt       19250.184          ops/s
BenchJDK.vector_256  /twitter.json  thrpt          42.616          ops/s
BenchJDK.vector_512  /twitter.json  thrpt          26.642          ops/s

In terms of GB/s, we have

jdk            2.5 GB/s
vector_128  12 GB/s

Again, fairly reasonable results.

[lemire]

My results are several times better than the ones you report in the README...

My CPUs are not special though they are recent.... still you should be getting better results....

[lemire]

I am going to close this as it was just for fun.

[AugustNagro]

Hey I'd love to merge this MR.. this whole repo is just for fun, at least until the vector API stabilizes.

This replaces a 25-minute benchmark with a 19 seconds one

Yeah totally agree it can be shortened up. My machine requires a bit longer it, but I can explore and adjust that in a separate MR.

An interesting addition in this PR is a scalar validation function, which allows you to see directly the benefits of the SIMD operations.

Awesome, thanks. Those are excellent additions.

drastically improves the performance in some cases

I ran the the new benchmark with

• openjdk 22.0.1+8-16
• i9-10940X (cascade lake, has slow avx-512)
• @Warmup(iterations = 3, time = 5).

And got these results without the code-change in Utf8.java:

Benchmark               (testFile)   Mode  Cnt     Score   Error  Units
BenchJDK.jdk         /twitter.json  thrpt       1541.310          ops/s
BenchJDK.scalar      /twitter.json  thrpt       2519.318          ops/s
BenchJDK.vector_128  /twitter.json  thrpt       5802.873          ops/s
BenchJDK.vector_256  /twitter.json  thrpt       7525.890          ops/s
BenchJDK.vector_512  /twitter.json  thrpt       4861.439          ops/s

With your changes:

Benchmark               (testFile)   Mode  Cnt      Score   Error  Units
BenchJDK.jdk         /twitter.json  thrpt        1519.961          ops/s
BenchJDK.scalar      /twitter.json  thrpt        2704.991          ops/s
BenchJDK.vector_128  /twitter.json  thrpt       15619.757          ops/s
BenchJDK.vector_256  /twitter.json  thrpt       18113.341          ops/s
BenchJDK.vector_512  /twitter.json  thrpt        4739.419          ops/s

Epic! If you re-open I'd be happy to approve & merge.

[lemire]

Re-opened as requested.

[AugustNagro]

Merged; I'll update the readme as well.

I used the OneBranchTooMany benchmark to identify an issue where a branch would cause the vector intrinsics to fail (causing them to box).

https://mail.openjdk.org/pipermail/panama-dev/2020-December/011650.html

https://mail.openjdk.org/pipermail/panama-dev/2020-December/011651.html

It has been fixed in openjdk now, so I'll remove it.