Merge pull request #686 from ludovic111/perf/sprite-blend-round
Blend animated sprite frames without a roundf call per byte on x86
This commit is contained in:
@@ -3507,13 +3507,39 @@ fn mix_frames(current: &mut [u8], next: &[u8], millis: u32) {
|
||||
let t = millis as f32 / 1000.0;
|
||||
for (c, n) in current.iter_mut().zip(next) {
|
||||
let (a, b) = (f32::from(*c) / 255.0, f32::from(*n) / 255.0);
|
||||
*c = ((a * (1.0 - t) + b * t) * 255.0).round() as u8;
|
||||
*c = round_unorm8((a * (1.0 - t) + b * t) * 255.0);
|
||||
}
|
||||
}
|
||||
/// `x.round() as u8` for `x >= 0`, without the `roundf` library call that
|
||||
/// x86-64 makes for every `f32::round` (it has no rounding instruction before
|
||||
/// SSE4.1): the fraction `x - trunc(x)` is exact below 2^23, so rounding half
|
||||
/// away from zero is adding one when it is at least a half.
|
||||
#[inline]
|
||||
fn round_unorm8(x: f32) -> u8 {
|
||||
let whole = x as u8;
|
||||
whole.saturating_add(u8::from(x - f32::from(whole) >= 0.5))
|
||||
}
|
||||
#[cfg(test)]
|
||||
mod sprite_upload_tests {
|
||||
use super::*;
|
||||
#[test]
|
||||
fn blended_frames_round_like_f32_round() {
|
||||
for millis in (0..1000).step_by(7).chain([999]) {
|
||||
let t = millis as f32 / 1000.0;
|
||||
for a in 0..=255u8 {
|
||||
for b in 0..=255u8 {
|
||||
let x = (f32::from(a) / 255.0 * (1.0 - t) + f32::from(b) / 255.0 * t) * 255.0;
|
||||
assert_eq!(round_unorm8(x), x.round() as u8, "{a} {b} {millis}");
|
||||
}
|
||||
}
|
||||
}
|
||||
for x in [
|
||||
0.0, 0.49999997, 0.5, 1.5, 2.5, 254.49998, 254.5, 255.0, 255.00002, 300.0,
|
||||
] {
|
||||
assert_eq!(round_unorm8(x), x.round() as u8, "{x}");
|
||||
}
|
||||
}
|
||||
#[test]
|
||||
fn real_rail_mips_do_not_change_with_larger_array_layers() {
|
||||
let root = std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
|
||||
.join("../../.local/horizon-client/26.3");
|
||||
|
||||
Reference in New Issue
Block a user