Skip intermediate coord representation

[michaelkirk]

We can just build up the coords directly rather than building up arrays and then converting them to Coords.

One more little perf boost. I tried this change previously, but it was causing a surprising regression in the encode benchmarks (even though this only touches a decode code path).

My suspicion is that it was the same perf cliff in #48. I really don't know, but I'm imaging some incidental code layout thing that I was on the edge of. Now that that we've dived off that cliff anyway, this seems like strictly a (small) improvement.

$ cargo bench --bench="*" --  --baseline=sha-6872788
   Compiling polyline v0.10.2 (/Users/mkirk/src/georust/polyline)
    Finished `bench` profile [optimized] target(s) in 0.94s
     Running benches/benchmarks.rs
(target/release/deps/benchmarks-8b91c88197106af3)
encode 10_000 coordinates at precision 1e-5
                        time:   [180.56 µs 180.85 µs 181.16 µs]
                        change: [-1.3886% -1.0790% -0.7773%] (p = 0.00 < 0.05)
                        Change within noise threshold.
Found 5 outliers among 100 measurements (5.00%)
  2 (2.00%) low mild
  3 (3.00%) high mild

encode 10_000 coordinates at precision 1e-6
                        time:   [219.72 µs 220.82 µs 221.84 µs]
                        change: [+0.5301% +1.0508% +1.5979%] (p = 0.00 < 0.05)
                        Change within noise threshold.

decode 10_000 coordinates at precision 1e-5
                        time:   [80.937 µs 81.828 µs 82.713 µs]
                        change: [-4.9284% -3.9764% -2.9753%] (p = 0.00 < 0.05)
                        Performance has improved.

decode 10_000 coordinates at precision 1e-6
                        time:   [97.706 µs 98.786 µs 99.825 µs]
                        change: [-3.1947% -2.2649% -1.3617%] (p = 0.00 < 0.05)
                        Performance has improved.

[urschrei]

critcmp main decode (decode being this PR):

group                                          decode                                 main
-----                                          ------                                 ----
decode 10_000 coordinates at precision 1e-5    1.01     71.0±1.74µs        ? ?/sec    1.00     70.6±2.11µs        ? ?/sec
decode 10_000 coordinates at precision 1e-6    1.00     86.3±1.87µs        ? ?/sec    1.01     86.8±2.23µs        ? ?/sec
encode 10_000 coordinates at precision 1e-5    1.00     83.8±1.58µs        ? ?/sec    1.00     83.5±0.70µs        ? ?/sec
encode 10_000 coordinates at precision 1e-6    1.05    100.9±1.26µs        ? ?/sec    1.00     96.4±2.14µs        ? ?/sec

So a 5 % hit on my machine on M2 arm64 on the last benchmark (which is NBD, to be clear)

[michaelkirk]

Huh wild that editing the decode path would slow down the encode benchmarks, but that seems to be the name of the game.

I just ran the benchmarks on my x86_64 linux machine and see no change good or bad.

So, I'm inclined to say this is "probably no different" from a performance perspective, and slightly clearer from a code readability perspective, so I'm still inclined to merge it if you're OK with it.