Improvements on x86

.cargo/cargo.toml

@@ -0,0 +1,5 @@
+[env]
+OPENCV_INCLUDE_PATHS= { value = "C:\\vcpkg\\installed\\x64-windows-static-md\\include", relative = false }
+OPENCV_LINK_PATHS= { value = "C:\\vcpkg\\installed\\x64-windows-static-md\\lib", relative = false }
+OPENCV_LINK_LIBS="opencv_core4.lib,opencv_imgproc4.lib,zlib.lib"
+OPENCV_MSVC_CRT="static"

Cargo.toml

@@ -18,4 +18,4 @@ exclude = ["*.jpg", "*.png"]
 [dependencies]
 colorutils-rs = "0.5.12"
 num-traits = "0.2.19"
-rayon = "1.10.0"
+rayon = "1.10.0"

README.md

@@ -51,6 +51,21 @@ M3 Pro, NEON dilation RGBA image 2731x4096 with specified kernel size
 | FM     | 48.25ms | 81.12ms | 111.79ms | 274.01ms | 515.54ms |
 | OpenCV | 30.22ms | 72.63ms | 138.69ms | 555.51ms | 1.19s    |
 
+SSE dilation RGB image 2731x4096 with specified kernel size
+
+| SE     | 9x9     | 15x15    | 21x21    | 41x41    | 61x61  |
+|--------|---------|----------|----------|----------|--------|
+| FM     | 84.19ms | 186.53ms | 254.70ms | 673.45ms | 1.37s  |
+| OpenCV | 28.61ms | 62.43ms  | 114.80ms | 428.87ms | 1.16ms |
+
+SSE dilation RGBA image 2731x4096 with specified kernel size
+
+| SE     | 9x9      | 15x15    | 21x21    | 41x41    | 61x61  |
+|--------|----------|----------|----------|----------|--------|
+| FM     | 109.37ms | 229.11ms | 329.31ms | 981.48ms | 2.05ms |
+| OpenCV | 39.20ms  | 76.09ms  | 149.12ms | 569.36ms | 1.33s  |
+
+
 This project is licensed under either of
 
 - BSD-3-Clause License (see [LICENSE](LICENSE.md))

app/Cargo.toml

@@ -7,7 +7,7 @@ edition = "2021"
 image = "0.25.2"
 fast_morphology = {path = "../"}
 imageproc = "0.25.0"
-opencv = {version = "0.93.0", features = ["clang-runtime", "imgproc"]}
+opencv = {version = "0.93.0", features = ["imgproc"]}
 
 [dev-dependencies]
 criterion = {version = "0.5.1", features = ["html_reports"]}

app/benches/dilation/main.rs

@@ -29,10 +29,12 @@
 
 use criterion::{criterion_group, criterion_main, Criterion};
 use fast_morphology::{
-    dilate_rgb, dilate_rgba, BorderMode, ImageSize, KernelShape, MorphologyThreadingPolicy,
+    dilate, dilate_rgb, dilate_rgba, BorderMode, ImageSize, KernelShape, MorphologyThreadingPolicy,
 };
 use image::{EncodableLayout, GenericImageView, ImageReader};
-use opencv::core::{Mat, MatTrait, Point, Scalar, BORDER_REPLICATE, CV_8U, CV_8UC3, CV_8UC4};
+use opencv::core::{
+    Mat, MatTrait, Point, Scalar, BORDER_REPLICATE, CV_8U, CV_8UC1, CV_8UC3, CV_8UC4,
+};
 use opencv::imgproc;
 
 pub fn circle_se(radius: usize) -> Vec<u8> {
@@ -145,6 +147,88 @@ fn exec_bench_rgb(c: &mut Criterion, size: usize) {
     );
 }
 
+fn exec_bench_gray(c: &mut Criterion, size: usize) {
+    let img = ImageReader::open("../assets/fruits.jpg")
+        .unwrap()
+        .decode()
+        .unwrap();
+    let dimensions = img.dimensions();
+    let rgb_image = img.to_luma8();
+    let rgb_bytes = rgb_image.as_bytes();
+
+    let radius_size_7 = size;
+    let se_size_15 = radius_size_7 * 2 + 1;
+    let structuring_element_15 = circle_se(radius_size_7);
+
+    let mut kernel_15 = Mat::new_rows_cols_with_default(
+        se_size_15 as i32,
+        se_size_15 as i32,
+        CV_8U,
+        Scalar::new(0., 0., 0., 0.),
+    )
+    .unwrap();
+    unsafe {
+        for (index, &byte) in structuring_element_15.iter().enumerate() {
+            kernel_15.data_mut().add(index).write(byte);
+        }
+    }
+
+    c.bench_function(
+        format!("FM, Gray Image dilation: SE {}x{}", se_size_15, se_size_15).as_str(),
+        |b| {
+            b.iter(|| {
+                let mut dst_image = vec![0u8; dimensions.0 as usize * dimensions.1 as usize];
+                dilate(
+                    rgb_bytes,
+                    &mut dst_image,
+                    ImageSize::new(dimensions.0 as usize, dimensions.1 as usize),
+                    &structuring_element_15,
+                    KernelShape::new(se_size_15, se_size_15),
+                    BorderMode::default(),
+                    MorphologyThreadingPolicy::Adaptive,
+                )
+                .unwrap();
+            })
+        },
+    );
+
+    let mut mat = Mat::new_rows_cols_with_default(
+        dimensions.1 as i32,
+        dimensions.0 as i32,
+        CV_8UC1,
+        Scalar::new(0., 0., 0., 0.),
+    )
+    .unwrap();
+    unsafe {
+        for (index, &byte) in rgb_bytes.iter().enumerate() {
+            mat.data_mut().add(index).write(byte);
+        }
+    }
+
+    c.bench_function(
+        format!(
+            "OpenCV, Gray Image dilation: SE {}x{}",
+            se_size_15, se_size_15
+        )
+        .as_str(),
+        |b| {
+            b.iter(|| {
+                let mut dst_mat = Mat::default();
+                imgproc::dilate(
+                    &mat,
+                    &mut dst_mat,
+                    &kernel_15,
+                    Point::new(-1, -1),
+                    1,
+                    BORDER_REPLICATE,
+                    Scalar::new(0., 0., 0., 0.),
+                )
+                .unwrap();
+            })
+        },
+    );
+}
+
 fn exec_bench_rgba(c: &mut Criterion, size: usize) {
     let img = ImageReader::open("../assets/fruits.jpg")
         .unwrap()
@@ -229,20 +313,26 @@ fn exec_bench_rgba(c: &mut Criterion, size: usize) {
 
 pub fn criterion_benchmark(c: &mut Criterion) {
     opencv::core::set_use_opencl(false).expect("Failed to disable OpenCL");
-    opencv::core::set_use_ipp(true).expect("Failed to disable IPP");
+    opencv::core::set_use_ipp(false).expect("Failed to disable IPP");
     opencv::core::set_use_optimized(false).expect("Failed to disable opts");
 
     exec_bench_rgb(c, 4);
-    exec_bench_rgb(c, 7);
-    exec_bench_rgb(c, 10);
-    exec_bench_rgb(c, 20);
-    exec_bench_rgb(c, 30);
-
-    exec_bench_rgba(c, 4);
-    exec_bench_rgba(c, 7);
-    exec_bench_rgba(c, 10);
-    exec_bench_rgba(c, 20);
-    exec_bench_rgba(c, 30);
+    // exec_bench_rgb(c, 7);
+    // exec_bench_rgb(c, 10);
+    // exec_bench_rgb(c, 20);
+    // exec_bench_rgb(c, 30);
+    //
+    // exec_bench_rgba(c, 4);
+    // exec_bench_rgba(c, 7);
+    // exec_bench_rgba(c, 10);
+    // exec_bench_rgba(c, 20);
+    // exec_bench_rgba(c, 30);
+
+    exec_bench_gray(c, 4);
+    // exec_bench_gray(c, 7);
+    // exec_bench_gray(c, 10);
+    // exec_bench_gray(c, 20);
+    // exec_bench_gray(c, 30);
 }
 
 criterion_group!(benches, criterion_benchmark);

app/src/main.rs

@@ -66,12 +66,12 @@ fn gaussian_kernel(size: usize, sigma: f32) -> Vec<Vec<f32>> {
 }
 
 fn main() {
-    let radius_size = 5;
+    let radius_size = 15;
     let mut structuring_element = circle_se(radius_size);
 
-    // opencv::core::set_use_opencl(false).expect("Failed to disable OpenCL");
-    // opencv::core::set_use_ipp(true).expect("Failed to disable IPP");
-    // opencv::core::set_use_optimized(false).expect("Failed to disable opts");
+    opencv::core::set_use_opencl(false).expect("Failed to disable OpenCL");
+    opencv::core::set_use_ipp(false).expect("Failed to disable IPP");
+    opencv::core::set_use_optimized(false).expect("Failed to disable opts");
 
     let se_size = radius_size * 2 + 1;
     let full_size = se_size;
@@ -80,7 +80,7 @@ fn main() {
         println!("{:?}", Vec::from(elements));
     }
 
-    let img = ImageReader::open("./assets/ebelhard.jpg")
+    let img = ImageReader::open("./assets/fruits.jpg")
         .unwrap()
         .decode()
         .unwrap();
@@ -180,46 +180,46 @@ fn main() {
 
     println!("rgb exec time {:?}", exec_time.elapsed());
 
-    // let mut mat = Mat::new_rows_cols_with_default(
-    //     dimensions.1 as i32,
-    //     dimensions.0 as i32,
-    //     CV_8UC3,
-    //     Scalar::new(0., 0., 0., 0.),
-    // )
-    // .unwrap();
-    // unsafe {
-    //     for (index, &byte) in saved_origin.iter().enumerate() {
-    //         mat.data_mut().add(index).write(byte);
-    //     }
-    // }
-    // let mut kernel = Mat::new_rows_cols_with_default(
-    //     full_size as i32,
-    //     full_size as i32,
-    //     CV_8U,
-    //     Scalar::new(0., 0., 0., 0.),
-    // )
-    // .unwrap();
-    // unsafe {
-    //     for (index, &byte) in structuring_element.iter().enumerate() {
-    //         kernel.data_mut().add(index).write(byte);
-    //     }
-    // }
+    let mut mat = Mat::new_rows_cols_with_default(
+        dimensions.1 as i32,
+        dimensions.0 as i32,
+        CV_8UC3,
+        Scalar::new(0., 0., 0., 0.),
+    )
+    .unwrap();
+    unsafe {
+        for (index, &byte) in saved_origin.iter().enumerate() {
+            mat.data_mut().add(index).write(byte);
+        }
+    }
+    let mut kernel = Mat::new_rows_cols_with_default(
+        full_size as i32,
+        full_size as i32,
+        CV_8U,
+        Scalar::new(0., 0., 0., 0.),
+    )
+    .unwrap();
+    unsafe {
+        for (index, &byte) in structuring_element.iter().enumerate() {
+            kernel.data_mut().add(index).write(byte);
+        }
+    }
 
     let exec_time = Instant::now();
 
-    // let mut dst_mat = Mat::default();
-    // imgproc::dilate(
-    //     &mat,
-    //     &mut dst_mat,
-    //     &kernel,
-    //     Point::new(-1, -1),
-    //     1,
-    //     BORDER_REPLICATE,
-    //     Scalar::new(0., 0., 0., 0.),
-    // )
-    // .unwrap();
+    let mut dst_mat = Mat::default();
+    imgproc::dilate(
+        &mat,
+        &mut dst_mat,
+        &kernel,
+        Point::new(-1, -1),
+        1,
+        BORDER_REPLICATE,
+        Scalar::new(0., 0., 0., 0.),
+    )
+    .unwrap();
 
-    // let open_cv_bytes = dst_mat.data_bytes().unwrap();
+    let open_cv_bytes = dst_mat.data_bytes().unwrap();
 
     println!("opencv exec time {:?}", exec_time.elapsed());
 
@@ -241,12 +241,12 @@ fn main() {
     )
     .unwrap();
 
-    // image::save_buffer(
-    //     "converted_opencv.png",
-    //     &open_cv_bytes,
-    //     dimensions.0,
-    //     dimensions.1,
-    //     image::ColorType::Rgb8,
-    // )
-    // .unwrap();
+    image::save_buffer(
+        "converted_opencv.png",
+        &open_cv_bytes,
+        dimensions.0,
+        dimensions.1,
+        image::ColorType::Rgb8,
+    )
+    .unwrap();
 }

src/arena_roi.rs

@@ -111,6 +111,7 @@ unsafe fn copy_row_sse(dst: &mut [u8], src: &[u8], start: usize, stride: usize)
 }
 
 /// Copies ROI from one image to another
+#[allow(clippy::type_complexity)]
 pub fn copy_roi(arena: &mut [u8], roi: &[u8], arena_stride: usize, stride: usize, height: usize) {
     let mut dst = arena;
     let mut src = roi;

src/flat_se.rs

@@ -54,10 +54,7 @@ pub(crate) struct AnalyzedSe {
 }
 
 impl AnalyzedSe {
-    pub fn new(
-        original_se: Vec<u8>,
-        left_front: FlatSe,
-    ) -> AnalyzedSe {
+    pub fn new(original_se: Vec<u8>, left_front: FlatSe) -> AnalyzedSe {
         let is_empty =
             left_front.element_offsets.is_empty() && left_front.element_offsets.is_empty();
         AnalyzedSe {

src/lib.rs

@@ -29,6 +29,7 @@
 extern crate core;
 
 mod arena;
+mod arena_roi;
 mod border_mode;
 mod filter;
 mod filter_op_declare;
@@ -46,7 +47,6 @@ mod se_scan;
 mod structuring_element;
 mod thread_policy;
 mod unsafe_slice;
-mod arena_roi;
 
 pub use border_mode::BorderMode;
 pub use img_size::ImageSize;

src/ops/mod.rs

@@ -37,6 +37,7 @@ mod morph_rows_rgba_4_op;
 #[cfg(all(target_arch = "aarch64", target_feature = "neon"))]
 pub mod neon;
 mod op;
+mod smart_allocator;
 #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
 pub mod sse;
 mod utils;