Reduce cost of bounds checks in transforms

[kornelski]

According to llvm-mca estimate, the 4x3 transform function

before:

uOps Per Cycle: 2.70
IPC: 1.97
Block RThroughput: 159.3

after:

uOps Per Cycle: 3.48
IPC: 2.69
Block RThroughput: 279.8

[jrmuizel]

Was the motivation for this? Avoiding the casts?

[kornelski]

That's the natural size for it, as read from the input.

I hoped it would also explain to LLVM that grid_size.pow(4) can't overflow, but unfortunately that didn't have any effect.

[jrmuizel]

Is this code showing up in profiles for you? It should only be used when building a lookup table during transform creation. The actual color transformation should be using a fast path.

[jrmuizel]

The first 3 commits are queued for landing. https://bugzilla.mozilla.org/show_bug.cgi?id=1947889

The other 3 seem ok but I'd like to understand the motivation a little better before landing them.

[kornelski]

Yes, on small images (<640px) the time to build the lookup table takes more time than the transformation itself. Reduction in bounds checks also made the code a bit smaller.

[jrmuizel]

I've landed the grid size patch. The other two commits cause:

---- gtest::gtest::v4_output stdout ----
thread 'gtest::gtest::v4_output' panicked at src/chain.rs:1109:5:
assertion failed: false

when running cargo test --all-features

[jrmuizel]

I'll make it so that it runs tests in github actions when I get a chance.