core,render: Improve performance of "round to nearest, ties to even" float operations

[kjarosh]

This patch improves performance of ecma_conversions::round_to_even() and matrix::round_to_i32():

1. by using f64::round_ties_even()/f32::round_ties_even(), which have been stable since 1.77.0, instead of a custom algorithm; and
2. by removing an unnecessary comparison to i32::MIN, as casting a float to an integer automatically saturates values smaller than the minimum integer value to the minimum value of the integer type.

A primitive benchmark shows around 20% faster execution of these operations. In the case of matrix::round_to_i32() we can potentially expect noticeable improvements in overall performance as this operation is quite commonly used.

Platform	Operation	Type	Improvement
amd64	ecma_conversions::round_to_even()	f64	~21%
WASM amd64	ecma_conversions::round_to_even()	f64	~20%
amd64	matrix::round_to_i32()	f32	~22%
WASM amd64	matrix::round_to_i32()	f32	~23%

Those functions have been covered by tests to ensure the behavior hasn't changed.

[adrian17]

LGTM.
Out of curiosity, did you find any content that's particularly slow due to these, or is this PR more inspired by round_ties_even being stabilized?

[kjarosh]

I remember analyzing performance of some SWF game where the results showed that sometimes matrix::round_to_i32 took a noticeable amount of time. I tried optimizing matrix operations using SIMD intrinsics (round_ties_even was unstable at the time), but that was extremely complicated, and I basically would have had to implement portable SIMD.

Today I noticed that round_ties_even has been stabilized so that was a no-brainer optimization to do :)

[torokati44]

and I basically would have had to implement portable SIMD.

We use the wide crate for this over in https://github.com/ruffle-rs/h263-rs/. It maps neatly to intrinsics on supported platforms, and expands to scalar stuff otherwise.

EDIT: That is, in the colorspace conversion part (the yuv package). In the decoding itself, the most critical part was the IDCT, which thankfully gets autovectorized anyway.