AVX-512 base added

Author: awxkee

Cargo.toml

@@ -27,4 +27,5 @@ libc = "0.2.158"
 
 [features]
 default = ["colorspaces"]
-colorspaces = ["dep:colorutils-rs"]
+colorspaces = ["dep:colorutils-rs"]
+nightly_avx512 = []

README.md

@@ -44,25 +44,27 @@ Despite all implementation are fast, not all the paths are implemented using SIM
 
 `~` - Partially implemented
 
-|                | NEON | SSE | AVX2 | WASM | 
-|----------------|------|-----|------|------| 
-| RGBA (8 bit)   | x    | x   | x    | ~    | 
-| RGB (8 bit)    | x    | x   | ~    | ~    | 
-| Plane (8 bit)  | x    | x   | ~    | ~    | 
-| RGBA (8+ bit)  | x    | x   | ~    | -    | 
-| RGB (8+ bit)   | x    | x   | ~    | -    | 
-| Plane (8+ bit) | ~    | ~   | ~    | -    | 
-| RGBA (f32)     | x    | x   | x    | -    | 
-| RGB (f32)      | x    | x   | ~    | -    | 
-| Plane (f32)    | x    | x   | ~    | -    | 
-| RGBA (f16)     | x    | x   | x    | -    | 
-| RGB (f16)      | x    | ~   | ~    | -    | 
-| Plane (f16)    | ~    | ~   | ~    | -    |
-| AR30/RA30      | x    | -   | -    | -    |
+|                | NEON | SSE | AVX2 | AVX-512 | WASM | 
+|----------------|------|-----|------|---------|------| 
+| RGBA (8 bit)   | x    | x   | x    | ~       | ~    | 
+| RGB (8 bit)    | x    | x   | ~    | ~       | ~    | 
+| Plane (8 bit)  | x    | x   | ~    | ~       | ~    | 
+| RGBA (8+ bit)  | x    | x   | ~    | -       | -    | 
+| RGB (8+ bit)   | x    | x   | ~    | -       | -    | 
+| Plane (8+ bit) | ~    | ~   | ~    | -       | -    | 
+| RGBA (f32)     | x    | x   | x    | -       | -    | 
+| RGB (f32)      | x    | x   | ~    | -       | -    | 
+| Plane (f32)    | x    | x   | ~    | -       | -    | 
+| RGBA (f16)     | x    | x   | x    | -       | -    | 
+| RGB (f16)      | x    | ~   | ~    | -       | -    | 
+| Plane (f16)    | ~    | ~   | ~    | -       | -    |
+| AR30/RA30      | x    | -   | -    | -       | -    |
 
 #### Features
 
-To enable support of `f16` the feature `half` should be activated.
+Features: 
+ -  To enable support of `f16` the feature `half` should be activated.
+ -  `nightly_avx512` activates AVX-512 feature set and requires `nightly` compiler channel 
 
 #### Target features with runtime dispatch
 
@@ -72,6 +74,8 @@ For x86 and aarch64 NEON runtime dispatch is used.
 
 `avx2`, `fma`, `sse4.1`, `f16c` will be detected automatically if available, no additional actions need, and called the best path.
 
+`avx512` requires feature `nightly_avx512` and requires `nightly` compiler channel, runtime detection if it is available then will be used.
+
 `fullfp16` NEON target detection performed in runtime, when available best the best paths for *f16* images are available on ARM.
 
 WASM `simd128` target feature activating is mandatory in build flags.

app/Cargo.toml

@@ -6,7 +6,7 @@ edition = "2021"
 [dependencies]
 image = { version = "0.25.5", features = ["default"] }
 #image = { path= "../../../RustroverProjects/image", features = ["default", "avif", "avif-native"] }
-pic-scale = { path = "..", features = ["half"], default-features = true }
+pic-scale = { path = "..", features = ["half", "nightly_avx512"], default-features = true }
 fast_image_resize = { version = "5.0.0", features = [] }
 half = { version = "2.4.1", default-features = true }
 

app/src/main.rs

@@ -42,7 +42,7 @@ fn resize_plane(
 
 fn main() {
     // test_fast_image();
-    let img = ImageReader::open("./assets/asset_4.png")
+    let img = ImageReader::open("./assets/nasa-4928x3279-rgba.png")
         .unwrap()
         .decode()
         .unwrap();
@@ -59,7 +59,7 @@ fn main() {
 
     //
     let store =
-        ImageStore::<u16, 4>::from_slice(&choke, dimensions.0 as usize, dimensions.1 as usize)
+        ImageStore::<u8, 4>::from_slice(&bytes, dimensions.0 as usize, dimensions.1 as usize)
             .unwrap();
 
     let dst_size = ImageSize::new(dimensions.0 as usize / 4, dimensions.1 as usize / 4);
@@ -75,15 +75,13 @@ fn main() {
     //     )
     //     .unwrap();
 
-    let mut dst_store = ImageStoreMut::<u16, 4>::alloc_with_depth(
+    let mut dst_store = ImageStoreMut::<u8, 4>::alloc_with_depth(
         dimensions.0 as usize,
         dimensions.1 as usize / 2,
         10,
     );
 
-    scaler
-        .resize_rgba_u16(&store, &mut dst_store, true)
-        .unwrap();
+    scaler.resize_rgba(&store, &mut dst_store, true).unwrap();
 
     let elapsed_time = start_time.elapsed();
     // Print the elapsed time in milliseconds
@@ -160,13 +158,13 @@ fn main() {
     //     .map(|&x| (x * 255f32) as u8)
     //     .collect();
 
-    let dst: Vec<u8> = dst_store
-        .as_bytes()
-        .iter()
-        .map(|&x| (x >> 2) as u8)
-        .collect();
+    // let dst: Vec<u8> = dst_store
+    //     .as_bytes()
+    //     .iter()
+    //     .map(|&x| (x >> 2) as u8)
+    //     .collect();
 
-    // let dst = dst_store.as_bytes();
+    let dst = dst_store.as_bytes();
     // let dst = resized;
     // image::save_buffer(
     //     "converted.png",

src/alpha_handle_u8.rs

@@ -137,6 +137,13 @@ pub(crate) fn premultiply_alpha_rgba(
         if is_x86_feature_detected!("avx2") {
             _dispatcher = avx_premultiply_alpha_rgba;
         }
+        #[cfg(feature = "nightly_avx512")]
+        if std::arch::is_x86_feature_detected!("avx512f")
+            && std::arch::is_x86_feature_detected!("avx512bw")
+        {
+            use crate::avx512::avx512_premultiply_alpha_rgba;
+            _dispatcher = avx512_premultiply_alpha_rgba;
+        }
     }
     #[cfg(all(target_arch = "wasm32", target_feature = "simd128"))]
     {
@@ -159,15 +166,22 @@ pub(crate) fn unpremultiply_alpha_rgba(
     }
     #[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
     {
-        if is_x86_feature_detected!("sse4.1") {
+        if std::arch::is_x86_feature_detected!("sse4.1") {
             _dispatcher = sse_unpremultiply_alpha_rgba;
         }
     }
     #[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
     {
-        if is_x86_feature_detected!("avx2") {
+        if std::arch::is_x86_feature_detected!("avx2") {
             _dispatcher = avx_unpremultiply_alpha_rgba;
         }
+        #[cfg(feature = "nightly_avx512")]
+        if std::arch::is_x86_feature_detected!("avx512f")
+            && std::arch::is_x86_feature_detected!("avx512bw")
+        {
+            use crate::avx512::avx512_unpremultiply_alpha_rgba;
+            _dispatcher = avx512_unpremultiply_alpha_rgba;
+        }
     }
     #[cfg(all(target_arch = "wasm32", target_feature = "simd128"))]
     {

src/avx2/alpha_u16.rs

@@ -45,8 +45,8 @@ unsafe fn _mm256_scale_by_alpha(px: __m256i, low_low_a: __m256, low_high_a: __m2
     let low_px = _mm256_cvtepi32_ps(_mm256_unpacklo_epi16(px, zeros));
     let high_px = _mm256_cvtepi32_ps(_mm256_unpackhi_epi16(px, zeros));
 
-    let new_ll = _mm256_cvtps_epi32(_mm256_round_ps::<0x02>(_mm256_mul_ps(low_px, low_low_a)));
-    let new_lh = _mm256_cvtps_epi32(_mm256_round_ps::<0x02>(_mm256_mul_ps(high_px, low_high_a)));
+    let new_ll = _mm256_cvtps_epi32(_mm256_round_ps::<0x00>(_mm256_mul_ps(low_px, low_low_a)));
+    let new_lh = _mm256_cvtps_epi32(_mm256_round_ps::<0x00>(_mm256_mul_ps(high_px, low_high_a)));
 
     _mm256_packus_epi32(new_ll, new_lh)
 }
@@ -110,8 +110,7 @@ trait Avx2PremultiplyExecutor {
 struct Avx2PremultiplyExecutorDefault<const BIT_DEPTH: usize> {}
 
 impl<const BIT_DEPTH: usize> Avx2PremultiplyExecutorDefault<BIT_DEPTH> {
-    #[inline]
-    #[target_feature(enable = "avx2")]
+    #[inline(always)]
     unsafe fn premultiply_chunk(&self, dst: &mut [u16], src: &[u16]) {
         let src_ptr = src.as_ptr();
         let lane0 = _mm256_loadu_si256(src_ptr as *const __m256i);
@@ -203,8 +202,7 @@ impl<const BIT_DEPTH: usize> Avx2PremultiplyExecutor for Avx2PremultiplyExecutor
 struct Avx2PremultiplyExecutorAnyBit {}
 
 impl Avx2PremultiplyExecutorAnyBit {
-    #[inline]
-    #[target_feature(enable = "avx2")]
+    #[inline(always)]
     unsafe fn premultiply_chunk(&self, dst: &mut [u16], src: &[u16], scale: __m256) {
         let src_ptr = src.as_ptr();
         let lane0 = _mm256_loadu_si256(src_ptr as *const __m256i);