Improve gaussian speed

awxkee · Jun 10, 2024 · 76c96d0 · 76c96d0
1 parent 1a708bc
commit 76c96d0
Show file tree

Hide file tree

Showing 5 changed files with 29 additions and 31 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 = "0.10.4"
+version = "0.10.5"
 edition = "2021"
 description = "High performance blur in pure rust"
 readme = "../../README.md"

diff --git a/src/lib/box_blur_sse.rs b/src/lib/box_blur_sse.rs
@@ -30,14 +30,14 @@
     target_feature = "sse4.1"
 ))]
 pub mod sse_support {
-    use crate::mul_table::{
-        MUL_TABLE_DOUBLE, MUL_TABLE_TWICE_RAD, SHR_TABLE_DOUBLE, SHR_TABLE_TWICE_RAD,
-    };
     #[cfg(target_arch = "x86")]
     use std::arch::x86::*;
     #[cfg(target_arch = "x86_64")]
     use std::arch::x86_64::*;
 
+    use crate::mul_table::{
+        MUL_TABLE_TWICE_RAD, SHR_TABLE_TWICE_RAD,
+    };
     use crate::sse_utils::sse_utils::load_u8_s32_fast;
     use crate::unsafe_slice::UnsafeSlice;
 
@@ -61,7 +61,7 @@ pub mod sse_support {
         let mul_value = MUL_TABLE_TWICE_RAD[radius as usize];
         let shr_value = SHR_TABLE_TWICE_RAD[radius as usize];
         let v_mul_value = unsafe { _mm_set1_epi32(mul_value) };
-        let v_shr_value = unsafe { _mm_set1_epi32(shr_value) };
+        let v_shr_value = unsafe { _mm_setr_epi32(shr_value,0,0,0) };
 
         let kernel_size = radius * 2 + 1;
         let edge_count = (kernel_size / 2) + 1;
@@ -117,7 +117,7 @@ pub mod sse_support {
                 }
 
                 let scale_store =
-                    unsafe { _mm_sra_epi32(_mm_mullo_epi32(store, v_mul_value), v_shr_value) };
+                    unsafe { _mm_srl_epi32(_mm_mullo_epi32(store, v_mul_value), v_shr_value) };
                 let px_16 = unsafe { _mm_packus_epi32(scale_store, scale_store) };
                 let px_8 = unsafe { _mm_packus_epi16(px_16, px_16) };
                 let pixel = unsafe { _mm_extract_epi32::<0>(px_8) };
@@ -157,7 +157,7 @@ pub mod sse_support {
         let mul_value = MUL_TABLE_TWICE_RAD[radius as usize];
         let shr_value = SHR_TABLE_TWICE_RAD[radius as usize];
         let v_mul_value = unsafe { _mm_set1_epi32(mul_value) };
-        let v_shr_value = unsafe { _mm_set1_epi32(shr_value) };
+        let v_shr_value = unsafe { _mm_setr_epi32(shr_value,0,0,0) };
 
         let kernel_size = radius * 2 + 1;
         let edge_count = (kernel_size / 2) + 1;
@@ -226,9 +226,9 @@ pub mod sse_support {
                 }
 
                 let scale_store_0 =
-                    unsafe { _mm_sra_epi32(_mm_mullo_epi32(store_0, v_mul_value), v_shr_value) };
+                    unsafe { _mm_srl_epi32(_mm_mullo_epi32(store_0, v_mul_value), v_shr_value) };
                 let scale_store_1 =
-                    unsafe { _mm_sra_epi32(_mm_mullo_epi32(store_1, v_mul_value), v_shr_value) };
+                    unsafe { _mm_srl_epi32(_mm_mullo_epi32(store_1, v_mul_value), v_shr_value) };
 
                 if CHANNELS == 3 {
                     let px_16 = unsafe { _mm_packus_epi32(scale_store_0, scale_store_0) };
@@ -322,7 +322,7 @@ pub mod sse_support {
                 }
 
                 let scale_store =
-                    unsafe { _mm_sra_epi32(_mm_mullo_epi32(store, v_mul_value), v_shr_value) };
+                    unsafe { _mm_srl_epi32(_mm_mullo_epi32(store, v_mul_value), v_shr_value) };
                 let px_16 = unsafe { _mm_packus_epi32(scale_store, scale_store) };
                 let px_8 = unsafe { _mm_packus_epi16(px_16, px_16) };
 
@@ -348,8 +348,8 @@ pub mod sse_support {
     target_feature = "sse4.1"
 )))]
 pub mod sse_support {
-    use crate::unsafe_slice::UnsafeSlice;
     use crate::FastBlurChannels;
+    use crate::unsafe_slice::UnsafeSlice;
 
     #[allow(dead_code)]
     pub(crate) fn box_blur_horizontal_pass_sse(

diff --git a/src/lib/fast_gaussian_sse.rs b/src/lib/fast_gaussian_sse.rs
@@ -52,7 +52,7 @@ pub mod sse_support {
         let mul_value = MUL_TABLE_DOUBLE[radius as usize];
         let shr_value = SHR_TABLE_DOUBLE[radius as usize];
         let v_mul_value = unsafe { _mm_set1_epi32(mul_value) };
-        let v_shr_value = unsafe { _mm_set1_epi32(shr_value) };
+        let v_shr_value = unsafe { _mm_setr_epi32(shr_value,0,0,0) };
         for y in start..std::cmp::min(height, end) {
             let mut diffs = unsafe { _mm_set1_epi32(0) };
             let mut summs = unsafe { _mm_set1_epi32(initial_sum) };
@@ -64,9 +64,8 @@ pub mod sse_support {
                 if x >= 0 {
                     let current_px = ((std::cmp::max(x, 0) as u32) * CHANNELS_COUNT as u32) as usize;
 
-                    const ROUNDING_FLAGS: i32 = _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC;
                     let prepared_px_s32 = unsafe {
-                        _mm_sra_epi32(_mm_mullo_epi32(summs, v_mul_value), v_shr_value)
+                        _mm_srl_epi32(_mm_mullo_epi32(summs, v_mul_value), v_shr_value)
                     };
                     let prepared_u16 = unsafe { _mm_packus_epi32(prepared_px_s32, prepared_px_s32) };
                     let prepared_u8 =
@@ -137,7 +136,7 @@ pub mod sse_support {
         let mul_value = MUL_TABLE_DOUBLE[radius as usize];
         let shr_value = SHR_TABLE_DOUBLE[radius as usize];
         let v_mul_value = unsafe { _mm_set1_epi32(mul_value) };
-        let v_shr_value = unsafe { _mm_set1_epi32(shr_value) };
+        let v_shr_value = unsafe { _mm_setr_epi32(shr_value,0,0,0) };
         for x in start..std::cmp::min(width, end) {
             let mut diffs = unsafe { _mm_set1_epi32(0) };
             let mut summs = unsafe { _mm_set1_epi32(initial_sum) };
@@ -148,9 +147,8 @@ pub mod sse_support {
 
                 if y >= 0 {
                     let current_px = ((std::cmp::max(x, 0)) * CHANNELS_COUNT as u32) as usize;
-                    const ROUNDING_FLAGS: i32 = _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC;
                     let prepared_px_s32 = unsafe {
-                        _mm_sra_epi32(_mm_mullo_epi32(summs, v_mul_value), v_shr_value)
+                        _mm_srl_epi32(_mm_mullo_epi32(summs, v_mul_value), v_shr_value)
                     };
                     let prepared_u16 = unsafe { _mm_packus_epi32(prepared_px_s32, prepared_px_s32) };
                     let prepared_u8 =

diff --git a/src/main.rs b/src/main.rs
@@ -43,23 +43,23 @@ fn main() {
     //     stride as u32,
     //     dimensions.0,
     //     dimensions.1,
-    //     319,
+    //     32,
     //     FastBlurChannels::Channels3,
     //     ThreadingPolicy::Adaptive,
     // );
 
-    // libblur::gaussian_box_blur(
-    //     &bytes,
-    //     stride as u32,
-    //     &mut dst_bytes,
-    //     stride as u32,
-    //     dimensions.0,
-    //     dimensions.1,
-    //     32,
-    //     FastBlurChannels::Channels3,
-    //     ThreadingPolicy::Single,
-    // );
-    // bytes = dst_bytes;
+    libblur::gaussian_box_blur(
+        &bytes,
+        stride as u32,
+        &mut dst_bytes,
+        stride as u32,
+        dimensions.0,
+        dimensions.1,
+        35,
+        FastBlurChannels::Channels3,
+        ThreadingPolicy::Single,
+    );
+    bytes = dst_bytes;
     // libblur::gaussian_blur(
     //     &bytes,
     //     stride as u32,