optimize to_byte_idx

[CeleritasCelery]

I got a chance to optimize to_byte_idx. There are a few things I changed here:

small string fast path

added a fast path for small strings. We already had this chars::count. This resulted in about 10% speedup on the trival strings.

fixed chunking bug

I found a performance bug where max_round_length was being divided by MAX_ACC instead of SIZE. This made the fast path way shorter than needed. For example if your char_idx was 1000, then the old logic would only use of the fast path for the first 3 chunks (48 bytes). And if char_idx was below 255 then it would skip the fast path entirely. Fixing this 1 line resulted in up to 40% speedup on some benchmarks.

loop unrolling

This is the same approach that was used last time, with unrolling the loop for faster performance. Now the routine goes brrrr and I am getting over 50GB/s on my arm machine.

benchmark text file path

lastly I updated the benchmarks so that they can be run from any directory instead of just the crate root.

benchmarks

Everything looks like it improved except for jp_0102 which shows a 8% performance regression on both arm and x86. I can't seem to explain the regression.

arm benchmarks

chars::to_byte_idx/en_0001
                        time:   [931.41 ps 931.82 ps 932.45 ps]
                        thrpt:  [1022.8 MiB/s 1023.5 MiB/s 1023.9 MiB/s]
                 change:
                        time:   [-40.121% -40.032% -39.932%] (p = 0.00 < 0.05)
                        thrpt:  [+66.477% +66.755% +67.003%]
                        Performance has improved.
chars::to_byte_idx/en_0010
                        time:   [4.4668 ns 4.4751 ns 4.4839 ns]
                        thrpt:  [2.0771 GiB/s 2.0811 GiB/s 2.0850 GiB/s]
                 change:
                        time:   [-15.638% -15.402% -15.148%] (p = 0.00 < 0.05)
                        thrpt:  [+17.853% +18.206% +18.536%]
                        Performance has improved.
chars::to_byte_idx/en_0100
                        time:   [4.8657 ns 4.8773 ns 4.8899 ns]
                        thrpt:  [19.046 GiB/s 19.095 GiB/s 19.140 GiB/s]
                 change:
                        time:   [-20.208% -20.005% -19.790%] (p = 0.00 < 0.05)
                        thrpt:  [+24.673% +25.007% +25.326%]
                        Performance has improved.
chars::to_byte_idx/en_1000
                        time:   [21.431 ns 21.468 ns 21.512 ns]
                        thrpt:  [43.292 GiB/s 43.383 GiB/s 43.457 GiB/s]
                 change:
                        time:   [-53.628% -53.540% -53.445%] (p = 0.00 < 0.05)
                        thrpt:  [+114.80% +115.24% +115.65%]
                        Performance has improved.
chars::to_byte_idx/en_10000
                        time:   [182.21 ns 182.28 ns 182.38 ns]
                        thrpt:  [51.065 GiB/s 51.093 GiB/s 51.114 GiB/s]
                 change:
                        time:   [-59.001% -58.937% -58.866%] (p = 0.00 < 0.05)
                        thrpt:  [+143.11% +143.53% +143.91%]
                        Performance has improved.
chars::to_byte_idx/jp_0003
                        time:   [2.1751 ns 2.1762 ns 2.1777 ns]
                        thrpt:  [1.2830 GiB/s 1.2839 GiB/s 1.2845 GiB/s]
                 change:
                        time:   [-12.671% -12.526% -12.387%] (p = 0.00 < 0.05)
                        thrpt:  [+14.138% +14.319% +14.509%]
                        Performance has improved.
chars::to_byte_idx/jp_0102
                        time:   [7.3852 ns 7.4068 ns 7.4302 ns]
                        thrpt:  [12.785 GiB/s 12.825 GiB/s 12.863 GiB/s]
                 change:
                        time:   [+7.8921% +8.2350% +8.5959%] (p = 0.00 < 0.05)
                        thrpt:  [-7.9155% -7.6084% -7.3148%]
                        Performance has regressed.
chars::to_byte_idx/jp_1001
                        time:   [36.486 ns 36.526 ns 36.573 ns]
                        thrpt:  [25.490 GiB/s 25.523 GiB/s 25.551 GiB/s]
                 change:
                        time:   [-21.254% -21.135% -21.003%] (p = 0.00 < 0.05)
                        thrpt:  [+26.587% +26.799% +26.990%]
                        Performance has improved.
chars::to_byte_idx/jp_10000
                        time:   [348.39 ns 348.75 ns 349.11 ns]
                        thrpt:  [26.704 GiB/s 26.732 GiB/s 26.759 GiB/s]
                 change:
                        time:   [-20.138% -20.019% -19.894%] (p = 0.00 < 0.05)
                        thrpt:  [+24.834% +25.029% +25.216%]
                        Performance has improved.

x86 benchmarks

note that this is with a really old Intel Core 2 Duo E8500 on linux

chars::to_byte_idx/en_0001
                        time:   [2.2377 ns 2.2380 ns 2.2384 ns]
                        thrpt:  [426.06 MiB/s 426.13 MiB/s 426.18 MiB/s]
                 change:
                        time:   [-45.754% -45.708% -45.643%] (p = 0.00 < 0.05)
                        thrpt:  [+83.968% +84.188% +84.347%]
                        Performance has improved.
chars::to_byte_idx/en_0010
                        time:   [19.928 ns 19.954 ns 19.974 ns]
                        thrpt:  [477.45 MiB/s 477.93 MiB/s 478.56 MiB/s]
                 change:
                        time:   [-9.9976% -9.8883% -9.7874%] (p = 0.00 < 0.05)
                        thrpt:  [+10.849% +10.973% +11.108%]
                        Performance has improved.
chars::to_byte_idx/en_0100
                        time:   [18.047 ns 18.057 ns 18.068 ns]
                        thrpt:  [5.1545 GiB/s 5.1576 GiB/s 5.1606 GiB/s]
                 change:
                        time:   [-23.535% -23.478% -23.407%] (p = 0.00 < 0.05)
                        thrpt:  [+30.560% +30.681% +30.779%]
                        Performance has improved.
chars::to_byte_idx/en_1000
                        time:   [72.238 ns 72.262 ns 72.285 ns]
                        thrpt:  [12.884 GiB/s 12.888 GiB/s 12.892 GiB/s]
                 change:
                        time:   [-57.904% -57.856% -57.805%] (p = 0.00 < 0.05)
                        thrpt:  [+136.99% +137.28% +137.55%]
                        Performance has improved.
chars::to_byte_idx/en_10000
                        time:   [578.62 ns 578.99 ns 579.39 ns]
                        thrpt:  [16.074 GiB/s 16.085 GiB/s 16.096 GiB/s]
                 change:
                        time:   [-62.480% -62.423% -62.367%] (p = 0.00 < 0.05)
                        thrpt:  [+165.73% +166.12% +166.52%]
                        Performance has improved.
chars::to_byte_idx/jp_0003
                        time:   [4.8644 ns 4.8660 ns 4.8677 ns]
                        thrpt:  [587.75 MiB/s 587.96 MiB/s 588.16 MiB/s]
                 change:
                        time:   [-26.968% -26.915% -26.850%] (p = 0.00 < 0.05)
                        thrpt:  [+36.705% +36.828% +36.926%]
                        Performance has improved.
chars::to_byte_idx/jp_0102
                        time:   [27.801 ns 27.806 ns 27.812 ns]
                        thrpt:  [3.4156 GiB/s 3.4163 GiB/s 3.4170 GiB/s]
                 change:
                        time:   [+5.3434% +5.4355% +5.5065%] (p = 0.00 < 0.05)
                        thrpt:  [-5.2191% -5.1553% -5.0723%]
                        Performance has regressed.
chars::to_byte_idx/jp_1001
                        time:   [139.85 ns 139.88 ns 139.91 ns]
                        thrpt:  [6.6634 GiB/s 6.6648 GiB/s 6.6662 GiB/s]
                 change:
                        time:   [-20.610% -20.588% -20.561%] (p = 0.00 < 0.05)
                        thrpt:  [+25.883% +25.925% +25.961%]
                        Performance has improved.
chars::to_byte_idx/jp_10000
                        time:   [1.2295 µs 1.2297 µs 1.2299 µs]
                        thrpt:  [7.5799 GiB/s 7.5812 GiB/s 7.5823 GiB/s]
                 change:
                        time:   [-22.340% -22.256% -22.178%] (p = 0.00 < 0.05)
                        thrpt:  [+28.498% +28.627% +28.767%]
                        Performance has improved.

[cessen]

Thanks for this! I'm currently on an extended vacation, so unless this is urgent I'll get to it some time in April. Sorry for the delay!

[cessen]

Sorry for the delay on this! As I mentioned in the other PR, I ended up moving, etc., and it's taken me a while to settle in and get back into rhythm.

Over-all this looks great. Just a few nits.

[CeleritasCelery]

Done. Glad you are back!

[cessen]

Thanks!