Removing rayon from most paths

Cargo.toml

@@ -2,7 +2,7 @@ workspace = { members = ["fuzz", "app/accelerate", "app"] }
 
 [package]
 name = "libblur"
-version = "0.19.2"
+version = "0.19.3"
 edition = "2021"
 description = "Fast image blurring in pure Rust"
 readme = "./README.md"
@@ -28,6 +28,7 @@ image = { version = "0.25", optional = true, default-features = false }
 rustfft = { version = "6.3", optional = true }
 fast_transpose = { version = "0.2.5", optional = true }
 num-complex = "0.4"
+novtb = "^0.1.4"
 
 [features]
 default = ["avx", "sse", "rdm", "neon"]

README.md

@@ -118,14 +118,14 @@ Example comparison time for blurring image 3000x4000 RGB 8-bit in multithreaded
 
 |         | time(NEON) | time(SSE) |
 |---------|:----------:|:---------:|
-| libblur |   5.11ms   |  5.30ms   |
+| libblur |   4.95ms   |  5.30ms   |
 | OpenCV  |   8.43ms   |  10.36ms  |
 
 Example comparison time for blurring image 2828x4242 RGBA 8-bit in multithreaded mode with 77 radius.
 
 |         | time(NEON) | time(SSE) |
 |---------|:----------:|:---------:|
-| libblur |   5.31ms   |  5.48ms   |
+| libblur |   4.97ms   |  5.48ms   |
 | OpenCV  |   8.00ms   |  8.55ms   |
 
 ### Fast gaussian

app/Cargo.toml

@@ -7,7 +7,7 @@ edition = "2021"
 colorutils-rs = "0.7.0"
 half = "2.4.1"
 image = "0.25.5"
-libblur = { path = "../", features = ["image", "fft", "rdm", "neon", "sse", "avx", "nightly_avx512", "nightly_fcma"], default-features = false }
+libblur = { path = "../", features = ["image", "fft", "rdm", "neon", "sse", "avx", "nightly_avx512"], default-features = false }
 accelerate = { path = "accelerate" }
 rayon = "1.10.0"
 fast_transpose = "0.2.5"

app/benches/gauss/main.rs

@@ -147,7 +147,7 @@ pub fn criterion_benchmark(c: &mut Criterion) {
         })
     });
 
-    c.bench_function("RGB f32 gauss blur edge clamp: rad 51", |b| {
+    c.bench_function("RGB f32 gauss blur edge clamp: rad 25", |b| {
         let mut dst_bytes = BlurImageMut::default();
         let src_bytes = img
             .as_bytes()

app/src/main.rs

@@ -33,17 +33,18 @@ mod split;
 use image::{EncodableLayout, GenericImageView, ImageReader};
 use libblur::{
     bilateral_filter, complex_gaussian_kernel, fast_bilateral_filter, fast_bilateral_filter_u16,
-    filter_1d_complex, filter_1d_complex_fixed_point, filter_2d_rgba_fft, gaussian_kernel_1d,
-    lens_kernel, sigma_size, AnisotropicRadius, BilateralBlurParams, BlurImage, BlurImageMut,
-    EdgeMode, FastBlurChannels, KernelShape, Scalar, ThreadingPolicy, TransferFunction,
+    filter_1d_complex, filter_1d_complex_fixed_point, filter_2d_rgba_fft, gaussian_blur,
+    gaussian_kernel_1d, lens_kernel, sigma_size, AnisotropicRadius, BilateralBlurParams, BlurImage,
+    BlurImageMut, BoxBlurParameters, CLTParameters, ConvolutionMode, EdgeMode, FastBlurChannels,
+    GaussianBlurParams, KernelShape, Scalar, ThreadingPolicy, TransferFunction,
 };
 use num_complex::Complex;
+use rayon::prelude::{IntoParallelIterator, ParallelIterator};
 use std::any::Any;
 use std::fs;
 use std::fs::File;
 use std::io::{BufReader, Read};
 use std::time::Instant;
-use rayon::prelude::{IntoParallelIterator, ParallelIterator};
 
 #[allow(dead_code)]
 fn f32_to_f16(bytes: Vec<f32>) -> Vec<u16> {
@@ -119,10 +120,22 @@ fn main() {
     // let bokeh = generate_complex_bokeh_kernel(35, 30.);
     let start_time = Instant::now();
     // let gaussian_kernel = gaussian_kernel_1d(31, sigma_size(31.)).iter().map(|&x| Complex::new(x, 0.0)).collect::<Vec<Complex<f32>>>();
-    let gaussian_kernel = complex_gaussian_kernel(51., 0.75, 25.);
+    let gaussian_kernel = complex_gaussian_kernel(51., 0.75, 5.);
+
+    let mut dst_image = cvt.clone_as_mut();
+
+    // gaussian_blur(
+    //     &cvt,
+    //     &mut dst_image,
+    //     GaussianBlurParams::new_from_kernel(15.),
+    //     EdgeMode::Clamp,
+    //     ThreadingPolicy::Single,
+    //     ConvolutionMode::FixedPoint,
+    // )
+    //     .unwrap();
 
-    // filter_2d_rgba_fft::<u16, f32, f32>(
-    //     &image,
+    // filter_2d_rgba_fft::<u8, f32, f32>(
+    //     &cvt,
     //     &mut dst_image,
     //     &motion,
     //     KernelShape::new(151, 151),
@@ -142,44 +155,42 @@ fn main() {
 
     // }
 
-    // libblur::gaussian_blur(
-    //     &image,
+    libblur::bilateral_filter(
+        &cvt,
+        &mut dst_image,
+        BilateralBlurParams {
+            kernel_size: 15,
+            spatial_sigma: 5.,
+            range_sigma: 5.,
+        },
+        EdgeMode::Clamp,
+        Scalar::default(),
+        ThreadingPolicy::Single,
+    )
+    .unwrap();
+
+    // libblur::stack_blur(
     //     &mut dst_image,
-    //     GaussianBlurParams {
-    //         x_kernel: 7,
-    //         x_sigma: 0.,
-    //         y_kernel: 9,
-    //         y_sigma: 0.,
-    //     },
-    //     EdgeMode::Clamp,
+    //     AnisotropicRadius::create(35, 35),
     //     ThreadingPolicy::Single,
-    //     ConvolutionMode::FixedPoint,
-    // )
-    // .unwrap();
-
-    // libblur::fast_gaussian_next_f32(
-    //     &mut dst_image,
-    //     AnisotropicRadius::create(8, 35),
-    //     ThreadingPolicy::Single,
-    //     EdgeMode::Clamp,
     // )
     // .unwrap();
 
     // libblur::motion_blur(
-    //     &image,
+    //     &cvt,
     //     &mut dst_image,
     //     35.,
-    //     15,
+    //     21,
     //     EdgeMode::Clamp,
     //     Scalar::new(0.0, 0.0, 0.0, 0.0),
-    //     ThreadingPolicy::Single,
+    //     ThreadingPolicy::Adaptive,
     // )
     // .unwrap();
 
     // let j_dag = dst_image.to_immutable_ref();
     // let gamma = j_dag.gamma8(TransferFunction::Srgb, true).unwrap();
 
-   /* dst_bytes = dst_image
+    dst_bytes = dst_image
         .data
         .borrow_mut()
         .iter()
@@ -234,5 +245,5 @@ fn main() {
             },
         )
         .unwrap();
-    }*/
+    }
 }

fuzz/bilateral/bilateral.rs

@@ -44,6 +44,7 @@ pub struct SrcImage {
     pub channels: u8,
     pub kernel_size: u8,
     pub angle: f32,
+    pub multi_threading: bool,
 }
 
 fuzz_target!(|data: SrcImage| {
@@ -71,6 +72,7 @@ fuzz_target!(|data: SrcImage| {
         data.angle,
         channels,
         edge_mode,
+        data.multi_threading,
     );
 });
 
@@ -81,6 +83,7 @@ fn fuzz_8bit(
     sigma: f32,
     channels: FastBlurChannels,
     edge_mode: EdgeMode,
+    multi_threading: bool,
 ) {
     if width == 0 || height == 0 || radius == 0 {
         return;
@@ -101,7 +104,11 @@ fn fuzz_8bit(
         },
         edge_mode,
         Scalar::new(0.0, 0.0, 0.0, 0.0),
-        ThreadingPolicy::Single,
+        if multi_threading {
+            ThreadingPolicy::Adaptive
+        } else {
+            ThreadingPolicy::Single
+        },
     )
     .unwrap();
 }

fuzz/fast_gaussian/fast_gaussian.rs

@@ -44,6 +44,7 @@ pub struct SrcImage {
     pub x_radius: u8,
     pub y_radius: u8,
     pub plane: u8,
+    pub threading: bool,
 }
 
 fuzz_target!(|data: SrcImage| {
@@ -69,6 +70,11 @@ fuzz_target!(|data: SrcImage| {
         plane_match,
         edge_mode,
         data.value,
+        if data.threading {
+            ThreadingPolicy::Adaptive
+        } else {
+            ThreadingPolicy::Single
+        },
     );
 });
 
@@ -80,6 +86,7 @@ fn fuzz_image(
     channels: FastBlurChannels,
     edge_mode: EdgeMode,
     data: u8,
+    threading_policy: ThreadingPolicy,
 ) {
     if width == 0 || height == 0 {
         return;
@@ -91,7 +98,7 @@ fn fuzz_image(
     fast_gaussian(
         &mut dst_image,
         AnisotropicRadius::create(x_radius as u32, y_radius as u32),
-        ThreadingPolicy::Single,
+        threading_policy,
         edge_mode,
     )
     .unwrap();

fuzz/fast_gaussian_next/fast_gaussian_next.rs

@@ -45,6 +45,7 @@ pub struct SrcImage {
     pub x_radius: u8,
     pub y_radius: u8,
     pub plane: u8,
+    pub threading: bool,
 }
 
 fuzz_target!(|data: SrcImage| {
@@ -70,6 +71,11 @@ fuzz_target!(|data: SrcImage| {
         plane_match,
         edge_mode,
         data.value,
+        if data.threading {
+            ThreadingPolicy::Adaptive
+        } else {
+            ThreadingPolicy::Single
+        },
     );
 });
 
@@ -81,6 +87,7 @@ fn fuzz_image(
     channels: FastBlurChannels,
     edge_mode: EdgeMode,
     value: u8,
+    threading_policy: ThreadingPolicy,
 ) {
     if width == 0 || height == 0 {
         return;
@@ -91,7 +98,7 @@ fn fuzz_image(
     fast_gaussian_next(
         &mut dst_image,
         AnisotropicRadius::create(x_radius as u32, y_radius as u32),
-        ThreadingPolicy::Single,
+        threading_policy,
         edge_mode,
     )
     .unwrap();