mirror of
https://github.com/SpaceManiac/SpacemanDMM.git
synced 2025-12-23 05:36:47 +00:00
Optimize map rendering (#242)
It is much faster now, way less terrible blitting code. --- SpaceManiac's benchmarks indicate about 9% speedup when rendering MetaStation with `--release` on.
This commit is contained in:
Pieter-Jan Briers
GitHub
parent
0b95746cb8
commit
e07801070e
5 changed files with 109 additions and 57 deletions
15
Cargo.lock
generated
15
Cargo.lock
generated
|
|
@ -189,6 +189,20 @@ name = "bytemuck"
|
|||
version = "1.5.0"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "5a4bad0c5981acc24bc09e532f35160f952e35422603f0563cd7a73c2c2e65a0"
|
||||
dependencies = [
|
||||
"bytemuck_derive",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "bytemuck_derive"
|
||||
version = "1.0.1"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "8e215f8c2f9f79cb53c8335e687ffd07d5bfcb6fe5fc80723762d0be46e7cc54"
|
||||
dependencies = [
|
||||
"proc-macro2 1.0.24",
|
||||
"quote 1.0.8",
|
||||
"syn",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "byteorder"
|
||||
|
|
@ -475,6 +489,7 @@ name = "dmm-tools"
|
|||
version = "0.1.0"
|
||||
dependencies = [
|
||||
"bumpalo",
|
||||
"bytemuck",
|
||||
"dreammaker",
|
||||
"gfx_core",
|
||||
"inflate",
|
||||
|
|
|
|||
|
|
@ -14,6 +14,10 @@ rand = "0.7.0"
|
|||
dreammaker = { path = "../dreammaker" }
|
||||
lodepng = "3.0.0"
|
||||
|
||||
[dependencies.bytemuck]
|
||||
version = "1.5"
|
||||
features = ["derive"]
|
||||
|
||||
[dependencies.bumpalo]
|
||||
version = "3.0.0"
|
||||
features = ["collections"]
|
||||
|
|
|
|||
139
src/tools/dmi.rs
139
src/tools/dmi.rs
|
|
@ -4,11 +4,13 @@
|
|||
|
||||
use std::io;
|
||||
use std::path::Path;
|
||||
use bytemuck::Pod;
|
||||
|
||||
use ndarray::Array3;
|
||||
use lodepng::{self, RGBA, Decoder, ColorType};
|
||||
use ndarray::Array2;
|
||||
|
||||
pub use dm::dmi::*;
|
||||
use std::ops::{Index, IndexMut};
|
||||
|
||||
type Rect = (u32, u32, u32, u32);
|
||||
|
||||
|
|
@ -49,6 +51,43 @@ impl IconFile {
|
|||
}
|
||||
}
|
||||
|
||||
#[derive(Default, Clone, Copy, Pod, Zeroable, Eq, PartialEq)]
|
||||
#[repr(C)]
|
||||
pub struct Rgba8 {
|
||||
pub r: u8,
|
||||
pub g: u8,
|
||||
pub b: u8,
|
||||
pub a: u8,
|
||||
}
|
||||
|
||||
impl Rgba8 {
|
||||
pub fn new(r: u8, g: u8, b: u8, a: u8) -> Rgba8 {
|
||||
Rgba8 { r, g, b, a }
|
||||
}
|
||||
|
||||
pub fn as_bytes(&self) -> &[u8; 4] {
|
||||
bytemuck::cast_ref(self)
|
||||
}
|
||||
|
||||
pub fn as_bytes_mut(&mut self) -> &mut [u8; 4] {
|
||||
bytemuck::cast_mut(self)
|
||||
}
|
||||
}
|
||||
|
||||
impl Index<u8> for Rgba8 {
|
||||
type Output = u8;
|
||||
|
||||
fn index(&self, index: u8) -> &Self::Output {
|
||||
&self.as_bytes()[index as usize]
|
||||
}
|
||||
}
|
||||
|
||||
impl IndexMut<u8> for Rgba8 {
|
||||
fn index_mut(&mut self, index: u8) -> &mut Self::Output {
|
||||
&mut self.as_bytes_mut()[index as usize]
|
||||
}
|
||||
}
|
||||
|
||||
// ----------------------------------------------------------------------------
|
||||
// Image manipulation
|
||||
|
||||
|
|
@ -56,7 +95,7 @@ impl IconFile {
|
|||
pub struct Image {
|
||||
pub width: u32,
|
||||
pub height: u32,
|
||||
pub data: Array3<u8>,
|
||||
pub data: Array2<Rgba8>,
|
||||
}
|
||||
|
||||
impl Image {
|
||||
|
|
@ -64,7 +103,9 @@ impl Image {
|
|||
Image {
|
||||
width,
|
||||
height,
|
||||
data: Array3::zeros((height as usize, width as usize, 4)),
|
||||
data: {
|
||||
Array2::default((width as usize, height as usize))
|
||||
},
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -72,16 +113,12 @@ impl Image {
|
|||
Image {
|
||||
width: bitmap.width as u32,
|
||||
height: bitmap.height as u32,
|
||||
data: Array3::from_shape_fn((bitmap.height, bitmap.width, 4), |(y, x, c)| {
|
||||
let rgba = bitmap.buffer[y * bitmap.width + x];
|
||||
match c {
|
||||
0 => rgba.r,
|
||||
1 => rgba.g,
|
||||
2 => rgba.b,
|
||||
3 => rgba.a,
|
||||
_ => unreachable!(),
|
||||
}
|
||||
}),
|
||||
data: {
|
||||
let cast_input = bytemuck::cast_slice(bitmap.buffer.as_slice());
|
||||
let mut arr = Array2::default((bitmap.width, bitmap.height));
|
||||
arr.as_slice_mut().unwrap().copy_from_slice(cast_input);
|
||||
arr
|
||||
},
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -105,7 +142,7 @@ impl Image {
|
|||
Ok(Image::from_rgba(bitmap))
|
||||
}
|
||||
|
||||
#[cfg(feature="png")]
|
||||
#[cfg(feature = "png")]
|
||||
pub fn to_file(&self, path: &Path) -> io::Result<()> {
|
||||
use std::fs::File;
|
||||
|
||||
|
|
@ -115,51 +152,49 @@ impl Image {
|
|||
let mut writer = encoder.write_header()?;
|
||||
// TODO: metadata with write_chunk()
|
||||
|
||||
writer.write_image_data(self.data.as_slice().unwrap())?;
|
||||
writer.write_image_data(bytemuck::cast_slice(self.data.as_slice().unwrap()))?;
|
||||
Ok(())
|
||||
}
|
||||
|
||||
pub fn composite(&mut self, other: &Image, pos: (u32, u32), crop: Rect, color: [u8; 4]) {
|
||||
use ndarray::Axis;
|
||||
|
||||
let mut destination = self.data.slice_mut(s![
|
||||
pos.1 as isize..(pos.1 + crop.3) as isize,
|
||||
pos.0 as isize..(pos.0 + crop.2) as isize,
|
||||
..
|
||||
]);
|
||||
let source = other.data.slice(s![
|
||||
crop.1 as isize..(crop.1 + crop.3) as isize,
|
||||
crop.0 as isize..(crop.0 + crop.2) as isize,
|
||||
..
|
||||
]);
|
||||
|
||||
// loop over each [r, g, b, a] available in the relevant area
|
||||
for (mut dest, orig_src) in destination.lanes_mut(Axis(2)).into_iter().zip(source.lanes(Axis(2))) {
|
||||
macro_rules! tint {
|
||||
($i:expr) => {
|
||||
mul255(
|
||||
*orig_src.get($i).unwrap_or(&255),
|
||||
*color.get($i).unwrap_or(&255),
|
||||
)
|
||||
};
|
||||
}
|
||||
let src = [tint!(0), tint!(1), tint!(2), tint!(3)];
|
||||
|
||||
// out_A = src_A + dst_A (1 - src_A)
|
||||
// out_RGB = (src_RGB src_A + dst_RGB dst_A (1 - src_A)) / out_A
|
||||
let out_a = src[3] + mul255(dest[3], 255 - src[3]);
|
||||
if out_a != 0 {
|
||||
for i in 0..3 {
|
||||
dest[i] = ((src[i] as u32 * src[3] as u32
|
||||
+ dest[i] as u32 * dest[3] as u32 * (255 - src[3] as u32) / 255)
|
||||
/ out_a as u32) as u8;
|
||||
let other_dat = other.data.as_slice().unwrap();
|
||||
let self_dat = self.data.as_slice_mut().unwrap();
|
||||
let mut sy = crop.1;
|
||||
for y in pos.1..(pos.1 + crop.3) {
|
||||
let mut sx = crop.0;
|
||||
for x in pos.0..(pos.0 + crop.2) {
|
||||
let src = other_dat[(sy * other.width + sx) as usize];
|
||||
macro_rules! tint {
|
||||
($i:expr) => {
|
||||
mul255(
|
||||
src[$i],
|
||||
color[$i],
|
||||
)
|
||||
};
|
||||
}
|
||||
} else {
|
||||
for i in 0..3 {
|
||||
dest[i] = 0;
|
||||
let mut dst = &mut self_dat[(y * self.width + x) as usize];
|
||||
let src_tint = Rgba8::new(tint!(0), tint!(1), tint!(2), tint!(3));
|
||||
|
||||
// out_A = src_A + dst_A (1 - src_A)
|
||||
// out_RGB = (src_RGB src_A + dst_RGB dst_A (1 - src_A)) / out_A
|
||||
let out_a = src_tint.a + mul255(dst.a, 255 - src_tint.a);
|
||||
if out_a != 0 {
|
||||
for i in 0..3 {
|
||||
dst[i] = ((src_tint[i] as u32 * src_tint.a as u32
|
||||
+ dst[i] as u32 * dst.a as u32 * (255 - src_tint.a as u32) / 255)
|
||||
/ out_a as u32) as u8;
|
||||
}
|
||||
} else {
|
||||
for i in 0..3 {
|
||||
dst[i] = 0;
|
||||
}
|
||||
}
|
||||
dst.a = out_a as u8;
|
||||
|
||||
sx += 1;
|
||||
}
|
||||
dest[3] = out_a as u8;
|
||||
|
||||
sy += 1;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -49,13 +49,11 @@ fn all_same(icon_file: &IconFile, states: &[&State]) -> bool {
|
|||
let rect2 = icon_file.rect_of_index(start_index + i);
|
||||
let slice1 = icon_file.image.data.slice(s![
|
||||
rect1.1 as isize..(rect1.1 + rect1.3) as isize,
|
||||
rect1.0 as isize..(rect1.0 + rect1.2) as isize,
|
||||
..
|
||||
rect1.0 as isize..(rect1.0 + rect1.2) as isize
|
||||
]);
|
||||
let slice2 = icon_file.image.data.slice(s![
|
||||
rect2.1 as isize..(rect2.1 + rect2.3) as isize,
|
||||
rect2.0 as isize..(rect2.0 + rect2.2) as isize,
|
||||
..
|
||||
rect2.0 as isize..(rect2.0 + rect2.2) as isize
|
||||
]);
|
||||
if slice1 != slice2 {
|
||||
return false;
|
||||
|
|
|
|||
|
|
@ -7,7 +7,7 @@ extern crate dreammaker as dm;
|
|||
extern crate lodepng;
|
||||
extern crate inflate;
|
||||
|
||||
#[macro_use] extern crate ndarray;
|
||||
#[macro_use] extern crate bytemuck;
|
||||
extern crate linked_hash_map;
|
||||
extern crate rand;
|
||||
extern crate bumpalo;
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue