Performance improvements by refactoring penalty and interleaveBytes

[Pjb518]

Includes a significant refactor of penalty that is much more efficient.

Also makes interleaveBytes marginally more memory and CPU efficient by calculating the resulting array size of and creating the resulting array up front to avoid dynamic resizing. The new implementation also avoids unnecessary bounding checks.

Here are some old benchmarks:

======== encode/ascii ========
encode/paulmillr x 1,998 ops/sec @ 500μs/op
encode/qrcode-generator x 2,928 ops/sec @ 341μs/op ± 1.13% (313μs..3ms)
encode/nuintun x 1,923 ops/sec @ 519μs/op ± 4.14% (418μs..9ms)
======== encode/gif ========
encode/paulmillr x 2,006 ops/sec @ 498μs/op ± 4.44% (440μs..15ms)
encode/qrcode-generator x 1,927 ops/sec @ 518μs/op
encode/nuintun x 2,324 ops/sec @ 430μs/op
======== encode: big ========
encode/paulmillr x 124 ops/sec @ 8ms/op
encode/qrcode-generator x 131 ops/sec @ 7ms/op
encode/nuintun x 166 ops/sec @ 6ms/op
======== decode ========
decode/paulmillr x 106 ops/sec @ 9ms/op ± 3.26% (8ms..24ms)
decode/jsqr x 35 ops/sec @ 28ms/op ± 3.29% (26ms..43ms)
decode/nuintun x 34 ops/sec @ 28ms/op ± 6.43% (26ms..48ms)
decode/instascan x 80 ops/sec @ 12ms/op ± 35.89% (7ms..193ms)

And here is the new version:

======== encode/ascii ========
encode/paulmillr x 3,430 ops/sec @ 291μs/op
encode/qrcode-generator x 2,995 ops/sec @ 333μs/op
encode/nuintun x 2,279 ops/sec @ 438μs/op
======== encode/gif ========
encode/paulmillr x 3,331 ops/sec @ 300μs/op
encode/qrcode-generator x 1,949 ops/sec @ 513μs/op
encode/nuintun x 2,295 ops/sec @ 435μs/op
======== encode: big ========
encode/paulmillr x 179 ops/sec @ 5ms/op
encode/qrcode-generator x 127 ops/sec @ 7ms/op ± 2.12% (7ms..17ms)
encode/nuintun x 158 ops/sec @ 6ms/op ± 1.67% (5ms..11ms)
======== decode ========
decode/paulmillr x 99 ops/sec @ 10ms/op ± 5.79% (8ms..29ms)
decode/jsqr x 34 ops/sec @ 28ms/op ± 4.48% (27ms..45ms)
decode/nuintun x 33 ops/sec @ 29ms/op ± 6.60% (26ms..49ms)
decode/instascan x 78 ops/sec @ 12ms/op ± 39.12% (7ms..211ms)

[paulmillr]

interleaveBytes change is great.

Not sure about the loop unroll.

Why did decode got slower? It's much more important vs encode. Is there any way to make it as fast?

[Pjb518]

Why did decode got slower? It's much more important vs encode. Is there any way to make it as fast?

I imagine it's just variance in the benchmark. None of these changes touch decoding.

There are definitely ways to make it faster. I think significantly, but I'd have to actually try some options.

interleaveBytes change is great.

Not sure about the loop unroll.

Almost all of the performance gains are from the changes to penalty, just FYI. The impact of the interleaveBytes changes are minimal.

[paulmillr]

Thank you.

[paulmillr]

The PR is live in 0.5.3.