Validate output of decoding

[michaelkirk]

update: merged! Depends on #43, so please review that first.

Supersedes #40

FIXES #39 and #37 with less performance regression.

$ cargo bench --bench="*" -- --baseline=sha-4cc9ff0
    Finished `bench` profile [optimized] target(s) in 0.03s
     Running benches/benchmarks.rs (target/release/deps/benchmarks-8b91c88197106af3)
encode 10_000 coordinates at precision 1e-5
                        time:   [106.15 µs 106.34 µs 106.53 µs]
                        change: [+0.3976% +0.6082% +0.8269%] (p = 0.00 < 0.05)
                        Change within noise threshold.

# I'm assuming this is spurious since I haven't touched the encode codepath but 🤷 
encode 10_000 coordinates at precision 1e-6
                        time:   [129.93 µs 130.66 µs 131.30 µs]
                        change: [+1.3900% +2.1013% +2.8347%] (p = 0.00 < 0.05)
                        Performance has regressed.

decode 10_000 coordinates at precision 1e-5
                        time:   [88.225 µs 89.156 µs 90.035 µs]
                        change: [+6.7607% +8.0710% +9.4026%] (p = 0.00 < 0.05)
                        Performance has regressed.

decode 10_000 coordinates at precision 1e-6
                        time:   [111.01 µs 111.91 µs 112.82 µs]
                        change: [+18.163% +19.518% +20.903%] (p = 0.00 < 0.05)
                        Performance has regressed.

@mattiZed - Thanks for the inspiration - I've incorporated some of your code into 0d827dd, but I've omitted the LUT stuff for now. I wasn't able to convince myself that that part was an improvement yet.

I will note: some of this code is super touchy. Tiny, seemingly irrelevant changes, would result in 50% swings in bench performance. Though my benchmark setup is crude, I've run the benchmarks back and forth enough to convince myself that this is not just noise.

[urschrei]

I'm surprised that this is touchy – there's nothing obviously "weird" in the new code. I'm not at my machine, and in any case it's a similar setup to yours iirc, so I can't confirm.

[michaelkirk]

Here's an example of the touchiness:

One tiny change to interpolate an existing var into the err message. We're not ever hitting this branch at runtime, so presumably it has some outsized effects on... locality... or register use.. ? 🤷 (surely not a compiler bug... right???).

Whatever it is, I can reliably reproduce it with this change. And there were a couple other small changes which you'd think innocuous, but tanked performance. So I had to be very careful while making these changes.

[17:15:26] mulberry arm64 ~/src/georust/polyline (mkirk/correctness)
$ GIT_PAGER=less git diff       
diff --git a/src/lib.rs b/src/lib.rs
index 09f9cea..4307337 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -151,7 +151,7 @@ fn decode_next(
     let mut result = 0;
     while let Some((idx, mut byte)) = chars.next() {
         if byte < 63 || (shift > 64 - 5) {
-            return Err(format!("Cannot decode character at index {idx}"));
+            return Err(format!("Cannot decode character at index {idx}, byte: {byte}"));
         }
         byte -= 63;
         result |= ((byte & 0x1f) as u64) << shift;
[17:15:29] mulberry arm64 ~/src/georust/polyline (mkirk/correctness)
$ cargo bench --bench="*" -- --baseline=sha-4cc9ff0

   Compiling polyline v0.10.2 (/Users/mkirk/src/georust/polyline)
    Finished `bench` profile [optimized] target(s) in 0.91s
     Running benches/benchmarks.rs (target/release/deps/benchmarks-8b91c88197106af3)
encode 10_000 coordinates at precision 1e-5
                        time:   [105.82 µs 106.10 µs 106.39 µs]
                        change: [+0.0992% +0.3137% +0.5416%] (p = 0.00 < 0.05)
                        Change within noise threshold.
Found 8 outliers among 100 measurements (8.00%)
  8 (8.00%) high mild

encode 10_000 coordinates at precision 1e-6
                        time:   [131.05 µs 131.50 µs 131.92 µs]
                        change: [+2.3540% +2.9801% +3.6276%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 3 outliers among 100 measurements (3.00%)
  3 (3.00%) low mild

# +75% vs. +10% 🔥🔥🔥 
decode 10_000 coordinates at precision 1e-5
                        time:   [143.63 µs 144.36 µs 145.13 µs]
                        change: [+73.357% +75.179% +77.021%] (p = 0.00 < 0.05)
                        Performance has regressed.

# +67% vs. +20% 🔥🔥🔥 
decode 10_000 coordinates at precision 1e-6
                        time:   [157.42 µs 158.13 µs 158.81 µs]
                        change: [+65.843% +67.560% +69.225%] (p = 0.00 < 0.05)
                        Performance has regressed.

I really don't understand it, but I don't understand a lot of things.