Loop unrolling and Area method

Cargo.toml

@@ -2,7 +2,7 @@ workspace = { members = ["app"] }
 
 [package]
 name = "pic-scale-safe"
-version = "0.1.2"
+version = "0.1.3"
 edition = "2021"
 description = "Fast and safe performance image scaling"
 readme = "README.md"

README.md

@@ -12,8 +12,6 @@ let img = ImageReader::open("./assets/nasa-4928x3279.png")
 let dimensions = img.dimensions();
 let transient = img.to_rgb8();
 
-let start = Instant::now();
-
 let src_size = ImageSize::new(dimensions.0 as usize, dimensions.1 as usize);
 let dst_size = ImageSize::new(dimensions.0 as usize / 4, dimensions.1 as usize / 4);
 
@@ -30,28 +28,32 @@ cargo bench --bench resize_rgba --manifest-path ./app/Cargo.toml
 |                           | Lanczos3 | Bilinear |
 |---------------------------|:--------:|:--------:|
 | image(aarch64)            |  121.19  |  48.89   |
-| pic-scale(aarch64)        |  26.90   |  15.13   |
-| fir(aarch64)              |  25.93   |  11.30   |
+| pic-scale(aarch64)        |  11.89   |   8.92   |
+| fir(aarch64)              |  25.89   |  11.30   |
 | image(x86)                |  192.52  |  88.63   |
-| pic-scale(x86)            |  49.79   |  35.98   |
-| pic-scale(x86-cpu-native) |  27.21   |  20.48   |
+| pic-scale(x86)            |  25.50   |  18.37   |
 | fir(x86)                  |  42.89   |  24.13   |
-| fir(x86-cpu-native)       |  41.17   |  23.62   |
+| image(x86-cpu-native)     |  205.64  |  89.02   |
+| pic-scale(x86-cpu-native) |  14.39   |  11.31   |
+| fir(x86-cpu-native)       |  41.21   |  22.77   |
 
 Example comparison time for downscale RGB 4928x3279 `8 bit` image in 4 times.
 
 ```bash
 cargo bench --bench resize_rgb --manifest-path ./app/Cargo.toml
 ```
 
-|                    | Lanczos3 | Bilinear |
-|--------------------|:--------:|:--------:|
-| image(aarch64)     |  123.85  |  51.30   |
-| pic-scale(aarch64) |  31.73   |  18.20   |
-| fir(aarch64)       |  24.04   |  11.37   |
-| image(x86)         |  201.52  |  90.82   |
-| pic-scale(x86)     |  34.54   |  25.05   |
-| fir(x86)           |  41.97   |  25.21   |
+|                           | Lanczos3 | Bilinear |
+|---------------------------|:--------:|:--------:|
+| image(aarch64)            |  123.85  |  51.30   |
+| pic-scale(aarch64)        |  17.23   |  12.32   |
+| fir(aarch64)              |  23.61   |  10.53   |
+| image(x86)                |  201.52  |  90.82   |
+| pic-scale(x86)            |  27.17   |  21.08   |
+| fir(x86)                  |  41.97   |  24.39   |
+| image(x86-cpu-native)     |  184.57  |  84.69   |
+| pic-scale(x86-cpu-native) |  20.96   |  15.16   |
+| fir(x86-cpu-native)       |  41.49   |  20.38   |
 
 Example comparison time for downscale RGBA 4928x3279 `16 bit` image in 4 times.
 
@@ -61,40 +63,51 @@ cargo bench --bench resize_rgba_u16 --manifest-path ./app/Cargo.toml
 
 |                           | Lanczos3 | Bilinear |
 |---------------------------|:--------:|:--------:|
-| image(aarch64)            |  123.27  |  52.91   |
-| pic-scale(aarch64)        |  28.041  |  18.89   |
-| fir(aarch64)              |  149.87  |  50.08   |
-| image(x86)                |  196.28  |  194.75  |
-| pic-scale(x86)            |  59.89   |  57.99   |
-| pic-scale(x86-cpu-native) |  44.07   |  57.99   |
-| fir(x86)                  |  52.73   |  28.35   |
+| image(aarch64)            |  262.32  |  76.91   |
+| pic-scale(aarch64)        |  15.49   |  11.38   |
+| fir(aarch64)              |  141.78  |  50.08   |
+| image(x86)                |  196.28  |  107.78  |
+| pic-scale(x86)            |  57.48   |  50.85   |
+| fir(x86)                  |  139.56  |  58.48   |
+| image(x86-cpu-native)     |  192.85  |  102.05  |
+| pic-scale(x86-cpu-native) |  39.60   |  46.44   |
+| fir(x86-cpu-native)       |  101.48  |  52.58   |
 
 Example comparison time for downscale RGB 4928x3279 `16 bit` image in 4 times.
 
 ```bash
 cargo bench --bench resize_rgb_u16 --manifest-path ./app/Cargo.toml
 ```
 
-|                    | Lanczos3 | Bilinear |
-|--------------------|:--------:|:--------:|
-| image(aarch64)     |  130.45  |  61.06   |
-| pic-scale(aarch64) |  36.10   |  23.80   |
-| fir(aarch64)       |  122.01  |  43.36   |
+|                           | Lanczos3 | Bilinear |
+|---------------------------|:--------:|:--------:|
+| image(aarch64)            |  130.45  |  57.38   |
+| pic-scale(aarch64)        |  16.17   |  12.11   |
+| fir(aarch64)              |  110.06  |  42.04   |
+| image(x86)                |  204.10  |  148.34  |
+| pic-scale(x86)            |  43.21   |  145.73  |
+| fir(x86)                  |  210.28  |  51.29   |
+| image(x86-cpu-native)     |  190.21  |  98.42   |
+| pic-scale(x86-cpu-native) |  33.48   |  28.50   |
+| fir(x86-cpu-native)       |  72.88   |  45.17   |
 
 Example comparison time for downscale RGBA 4928x3279 `f32` image in 4 times.
 
 ```bash
 cargo bench --bench resize_rgba_f32 --manifest-path ./app/Cargo.toml
 ```
 
-|                    | Lanczos3 | Bilinear |
-|--------------------|:--------:|:--------:|
-| image(aarch64)     |  100.16  |  51.21   |
-| pic-scale(aarch64) |  43.04   |  19.16   |
-| fir(aarch64)       |  114.35  |  37.75   |
-| image(x86)         |  164.04  |  98.90   |
-| pic-scale(x86)     |  57.39   |  43.84   |
-| fir(x86)           |  60.30   |  29.92   |
+|                           | Lanczos3 | Bilinear |
+|---------------------------|:--------:|:--------:|
+| image(aarch64)            |  100.16  |  50.09   |
+| pic-scale(aarch64)        |  14.07   |  11.18   |
+| fir(aarch64)              |  105.30  |  37.75   |
+| image(x86)                |  208.25  |  107.84  |
+| pic-scale(x86)            |  33.55   |  28.97   |
+| fir(x86)                  |  92.38   |  74.12   |
+| image(x86-cpu-native)     |  162.83  |  108.54  |
+| pic-scale(x86-cpu-native) |  33.13   |  28.54   |
+| fir(x86-cpu-native)       |  56.65   |  59.96   |
 
 This project is licensed under either of
 

app/benches/resize_rgb/main.rs

@@ -57,7 +57,7 @@ pub fn criterion_benchmark(c: &mut Criterion) {
         })
     });
 
-    c.bench_function("Image RGB: Bilinear", |b| {
+    c.bench_function("Image RGB: Lanczos3", |b| {
         b.iter(|| {
             _ = dyn_image.clone().resize_exact(
                 dimensions.0 / 4,

app/src/main.rs

@@ -31,13 +31,15 @@ mod image_wrapper;
 use fast_image_resize::images::Image;
 use fast_image_resize::{CpuExtensions, FilterType, PixelType, ResizeAlg, ResizeOptions, Resizer};
 use image::{
-    DynamicImage, EncodableLayout, GenericImageView, ImageBuffer, ImageFormat, ImageReader, Rgb,
-    RgbImage,
+    imageops, DynamicImage, EncodableLayout, GenericImageView, ImageBuffer, ImageFormat,
+    ImageReader, Rgb, RgbImage,
 };
 use pic_scale_safe::{
-    resize_floating_point, resize_rgb16, resize_rgb8, resize_rgba16, resize_rgba8, ImageSize,
+    premultiply_rgba8, resize_fixed_point, resize_floating_point, resize_rgb16, resize_rgb8,
+    resize_rgb_f32, resize_rgba16, resize_rgba8, unpremultiply_rgba8, ImageSize,
     ResamplingFunction,
 };
+use std::ops::{BitXor, Shr};
 use std::time::Instant;
 
 fn main() {
@@ -46,44 +48,55 @@ fn main() {
         .decode()
         .unwrap();
     let dimensions = img.dimensions();
-    let transient = img.to_rgb8();
+    let transient = img.to_rgba8();
 
     let mut working_store = transient.to_vec();
 
     let start = Instant::now();
 
     let src_size = ImageSize::new(dimensions.0 as usize, dimensions.1 as usize);
-    let dst_size = ImageSize::new(dimensions.0 as usize * 4, dimensions.1 as usize * 4);
+    let dst_size = ImageSize::new(dimensions.0 as usize / 2, dimensions.1 as usize / 2);
 
-    let mut resized = resize_rgb8(
+    let start_mul = Instant::now();
+
+    premultiply_rgba8(&mut working_store);
+
+    println!("Alpha mul time {:?}", start_mul.elapsed());
+
+    let mut resized = resize_rgba8(
         &working_store,
         src_size,
         dst_size,
-        ResamplingFunction::MitchellNetravalli,
+        ResamplingFunction::Lanczos3,
     )
     .unwrap();
 
+    // unpremultiply_rgba8(&mut resized);
+
     println!("Working time {:?}", start.elapsed());
 
     // let rgba_image = DynamicImage::ImageRgb16(ImageBuffer::<Rgb<u16>, Vec<u16>>::from_vec(dimensions.0 * 4, dimensions.1 / 4, resized).unwrap());
     // rgba_image.save_with_format("converted.png", ImageFormat::Png).unwrap();
 
-    // let shifted = resized.iter().map(|&x| (x >> 8) as u8).collect::<Vec<_>>();
+    // let shifted = resized
+    //     .iter()
+    //     .map(|&x| (x * 255.) as u8)
+    //     .collect::<Vec<_>>();
 
     image::save_buffer(
-        "converted.jpg",
+        "converted.png",
         &resized,
         dst_size.width as u32,
         dst_size.height as u32,
-        image::ColorType::Rgb8,
+        image::ColorType::Rgba8,
     )
     .unwrap();
 
     // let mut src_bytes = transient.as_bytes().to_vec();
     // let pixel_type: PixelType = PixelType::U8x3;
     // let src_image =
     //     Image::from_slice_u8(dimensions.0, dimensions.1, &mut src_bytes, pixel_type).unwrap();
-    // let mut dst_image = Image::new(dimensions.0 * 4, dimensions.1 * 4, pixel_type);
+    // let mut dst_image = Image::new(dimensions.0 / 8, dimensions.1 / 8, pixel_type);
     //
     // let mut resizer = Resizer::new();
     // unsafe {
@@ -97,7 +110,7 @@ fn main() {
     //         &src_image,
     //         &mut dst_image,
     //         &ResizeOptions::new()
-    //             .resize_alg(ResizeAlg::Convolution(FilterType::Mitchell))
+    //             .resize_alg(ResizeAlg::Convolution(FilterType::Bilinear))
     //             .use_alpha(false),
     //     )
     //     .unwrap();
@@ -109,13 +122,13 @@ fn main() {
     // // let rgba_image = DynamicImage::ImageRgb8(RgbImage::from_raw(dst_image.width() as u32, dst_image.height() as u32, dst_image.buffer().to_vec()).unwrap());
     // // rgba_image.save_with_format("fast_image.png", ImageFormat::Png).unwrap();
     // image::save_buffer(
-    //     "fast_image.jpg",
+    //     "fast_image.png",
     //     dst_image.buffer(),
     //     dst_image.width(),
     //     dst_image.height(),
     //     image::ColorType::Rgb8,
     // )
-    //     .unwrap();
+    // .unwrap();
 }
 
 fn u8_to_u16(u8_buffer: &[u8]) -> &[u16] {

command.ps1

@@ -1,2 +1,2 @@
-# $env:RUSTFLAGS = "-C target-cpu=native"
-cargo bench --bench resize_rgba_u16 --manifest-path ./app/Cargo.toml
+$env:RUSTFLAGS = "-C target-cpu=native"
+cargo bench --bench resize_rgba_f32 --manifest-path ./app/Cargo.toml