Improve SSE, improve Stackblur

awxkee · Jun 11, 2024 · beb2c4f · beb2c4f
1 parent 01ec35e
commit beb2c4f
Show file tree

Hide file tree

Showing 8 changed files with 348 additions and 197 deletions.
diff --git a/Cargo.lock b/Cargo.lock
diff --git a/src/lib/Cargo.toml b/src/lib/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "libblur"
-version = "0.10.6"
+version = "0.10.7"
 edition = "2021"
 description = "High performance blur in pure rust"
 readme = "../../README.md"
@@ -11,7 +11,7 @@ documentation = "https://github.com/awxkee/libblur"
 categories = ["multimedia::images", "multimedia::video", "algorithms"]
 homepage = "https://github.com/awxkee/libblur"
 repository = "https://github.com/awxkee/libblur.git"
-exclude = ["*.jpg", "../../assets/*"]
+exclude = ["*.jpg", "../../assets/*", "*.png"]
 
 [lib]
 name = "libblur"

diff --git a/src/lib/fast_gaussian_next_sse.rs b/src/lib/fast_gaussian_next_sse.rs
@@ -72,12 +72,23 @@ pub mod sse_support {
                         unsafe { _mm_packus_epi16(prepared_u16, prepared_u16) };
 
                     let pixel = unsafe { _mm_extract_epi32::<0>(prepared_u8) };
-                    let bits = pixel.to_le_bytes();
 
-                    unsafe {
-                        bytes.write(current_y + current_px, bits[0]);
-                        bytes.write(current_y + current_px + 1, bits[1]);
-                        bytes.write(current_y + current_px + 2, bits[2]);
+                    let bytes_offset = current_y + current_px;
+
+                    if CHANNELS_COUNT == 4 {
+                        unsafe {
+                            let dst_ptr =
+                                (bytes.slice.as_ptr() as *mut u8).add(bytes_offset) as *mut i32;
+                            *dst_ptr = pixel;
+                        }
+                    } else {
+                        let bits = pixel.to_le_bytes();
+
+                        unsafe {
+                            bytes.write(bytes_offset, bits[0]);
+                            bytes.write(bytes_offset + 1, bits[1]);
+                            bytes.write(bytes_offset + 2, bits[2]);
+                        }
                     }
 
                     let d_arr_index_1 = ((y + radius_64) & 1023) as usize;
@@ -177,12 +188,23 @@ pub mod sse_support {
                         unsafe { _mm_packus_epi16(prepared_u16, prepared_u16) };
 
                     let pixel = unsafe { _mm_extract_epi32::<0>(prepared_u8) };
-                    let bits = pixel.to_le_bytes();
 
-                    unsafe {
-                        bytes.write(current_y + current_px, bits[0]);
-                        bytes.write(current_y + current_px + 1, bits[1]);
-                        bytes.write(current_y + current_px + 2, bits[2]);
+                    let bytes_offset = current_y + current_px;
+
+                    if CHANNELS_COUNT == 4 {
+                        unsafe {
+                            let dst_ptr =
+                                (bytes.slice.as_ptr() as *mut u8).add(bytes_offset) as *mut i32;
+                            *dst_ptr = pixel;
+                        }
+                    } else {
+                        let bits = pixel.to_le_bytes();
+
+                        unsafe {
+                            bytes.write(bytes_offset, bits[0]);
+                            bytes.write(bytes_offset + 1, bits[1]);
+                            bytes.write(bytes_offset + 2, bits[2]);
+                        }
                     }
 
                     let d_arr_index_1 = ((x + radius_64) & 1023) as usize;

diff --git a/src/lib/fast_gaussian_sse.rs b/src/lib/fast_gaussian_sse.rs
@@ -71,12 +71,23 @@ pub mod sse_support {
                     let prepared_u8 =
                         unsafe { _mm_packus_epi16(prepared_u16, prepared_u16) };
                     let pixel = unsafe { _mm_extract_epi32::<0>(prepared_u8) };
-                    let bits = pixel.to_le_bytes();
 
-                    unsafe {
-                        bytes.write(current_y + current_px, bits[0]);
-                        bytes.write(current_y + current_px + 1, bits[1]);
-                        bytes.write(current_y + current_px + 2, bits[2]);
+                    let bytes_offset = current_y + current_px;
+
+                    if CHANNELS_COUNT == 4 {
+                        unsafe {
+                            let dst_ptr =
+                                (bytes.slice.as_ptr() as *mut u8).add(bytes_offset) as *mut i32;
+                            *dst_ptr = pixel;
+                        }
+                    } else {
+                        let bits = pixel.to_le_bytes();
+
+                        unsafe {
+                            bytes.write(current_y + current_px, bits[0]);
+                            bytes.write(current_y + current_px + 1, bits[1]);
+                            bytes.write(current_y + current_px + 2, bits[2]);
+                        }
                     }
 
                     let arr_index = ((x - radius_64) & 1023) as usize;
@@ -155,12 +166,23 @@ pub mod sse_support {
                         unsafe { _mm_packus_epi16(prepared_u16, prepared_u16) };
 
                     let pixel = unsafe { _mm_extract_epi32::<0>(prepared_u8) };
-                    let bits = pixel.to_le_bytes();
 
-                    unsafe {
-                        bytes.write(current_y + current_px, bits[0]);
-                        bytes.write(current_y + current_px + 1, bits[1]);
-                        bytes.write(current_y + current_px + 2, bits[2]);
+                    let bytes_offset = current_y + current_px;
+
+                    if CHANNELS_COUNT == 4 {
+                        unsafe {
+                            let dst_ptr =
+                                (bytes.slice.as_ptr() as *mut u8).add(bytes_offset) as *mut i32;
+                            *dst_ptr = pixel;
+                        }
+                    } else {
+                        let bits = pixel.to_le_bytes();
+
+                        unsafe {
+                            bytes.write(bytes_offset, bits[0]);
+                            bytes.write(bytes_offset + 1, bits[1]);
+                            bytes.write(bytes_offset + 2, bits[2]);
+                        }
                     }
 
                     let arr_index = ((y - radius_64) & 1023) as usize;