Add blurhash benchmark

[tenderlove]

This is a benchmark for a pure Ruby version of the Blurhash gem:

https://github.com/Gargron/blurhash

I think we're trying to find a benchmark that would measure the performance of setbyte, but I'm not 100% convinced this is the right benchmark. It only calls setbyte a handful of times per iteration. I think this benchmark might be more suitable for floating point math benchmarks?

Regardless, I think we should add this benchmark and try to speed it up. The C extension version of this code is 10x faster than when YJIT is enabled:

$ ruby -I lib bench.rb
ruby 3.4.0dev (2024-01-12T21:13:42Z master 206388b19e) [arm64-darwin23]
Warming up --------------------------------------
                ruby     1.000 i/100ms
                cext     7.000 i/100ms
Calculating -------------------------------------
                ruby      3.589 (± 0.0%) i/s -     18.000 in   5.015422s
                cext     78.361 (± 0.0%) i/s -    392.000 in   5.002508s
$ ruby --yjit -I lib bench.rb
ruby 3.4.0dev (2024-01-12T21:13:42Z master 206388b19e) +YJIT [arm64-darwin23]
Warming up --------------------------------------
                ruby     1.000 i/100ms
                cext     7.000 i/100ms
Calculating -------------------------------------
                ruby      6.596 (± 0.0%) i/s -     33.000 in   5.002780s
                cext     78.332 (± 0.0%) i/s -    392.000 in   5.004399s

[maximecb]

The benchmark looks good 👍

One thing that's a bit confusing to me is that based on stats:

./run_once.sh --yjit-stats benchmarks/blurhash/benchmark.rb
...
total_exits:                    0
Top-20 most frequent C calls (97.2% of C calls):
               Float#/:  3,498,099 (19.4%)
               Float#*:  3,001,055 (16.7%)
             Integer#*:  2,996,283 (16.7%)
               Float#+:  2,996,251 (16.7%)
              Float#**:  1,498,162 ( 8.3%)
          Integer#to_f:  1,498,155 ( 8.3%)
              Float#<=:  1,498,147 ( 8.3%)
            Module#cos:    501,782 ( 2.8%)
           Symbol#to_s:        445 ( 0.0%)
    Symbol#start_with?:        402 ( 0.0%)
           Array#fetch:         40 ( 0.0%)
          Symbol#match:         33 ( 0.0%)
        String#setbyte:         28 ( 0.0%)

The setbyte function is only called 28 times. I guess it's only used to encode a short hash at the end, and not actually the whole image?

If anything, this benchmark makes me think we ought to pay more attention to floating-point support 😅 Still a good benchmark and inclined to merge it 👍

[tenderlove]

The setbyte function is only called 28 times. I guess it's only used to encode a short hash at the end, and not actually the whole image?

Ya, as I said in the description I think this is probably not a good benchmark for setbyte but rather floating point math. We should probably do some work on floating point math, but I'll look for a better setbyte benchmark

[maximecb]

No stress, probably best to look for relevant benchmarks, pure-Ruby code rather than benchmarks specifically using setbyte.

[eregon]

I think we're trying to find a benchmark that would measure the performance of setbyte,

FWIW I recall https://github.com/oracle/truffleruby/issues/2336#issuecomment-826722788. It actually uses String#[]= and not setbyte (on a BINARY or US-ASCII String from a quick look). But as mentioned in that comment I think using a String like a byte buffer is rather inefficient (notably it can cause needless coderange computations).