Dot 8 RGB8

src/avx2/mod.rs

@@ -34,6 +34,8 @@ mod alpha_u16;
 mod alpha_u8;
 mod check_alpha;
 mod rgb_u8;
+#[cfg(feature = "nightly_avx512")]
+mod rgb_u8_dot_i8;
 #[cfg(feature = "half")]
 mod rgba_f16;
 mod rgba_f32;
@@ -57,6 +59,10 @@ pub(crate) use check_alpha::{
     avx_has_non_constant_cap_alpha_rgba16, avx_has_non_constant_cap_alpha_rgba8,
 };
 pub(crate) use rgb_u8::{convolve_horizontal_rgb_avx_row_one, convolve_horizontal_rgb_avx_rows_4};
+#[cfg(feature = "nightly_avx512")]
+pub(crate) use rgb_u8_dot_i8::{
+    convolve_horizontal_rgb_avx_row_i8_one, convolve_horizontal_rgb_avx_rows_4_i8,
+};
 #[cfg(feature = "half")]
 pub(crate) use rgba_f16::{
     convolve_horizontal_rgba_avx_row_one_f16, convolve_horizontal_rgba_avx_rows_4_f16,

src/avx2/rgb_u8.rs

@@ -47,7 +47,7 @@ pub(crate) fn convolve_horizontal_rgb_avx_rows_4(
         #[cfg(feature = "nightly_avx512")]
         {
             if std::arch::is_x86_feature_detected!("avxvnni") {
-                return convolve_horizontal_rgb_sse_rows_4_vnni(
+                return convolve_horizontal_rgb_avx_rows_4_vnni(
                     src,
                     src_stride,
                     dst,
@@ -56,19 +56,19 @@ pub(crate) fn convolve_horizontal_rgb_avx_rows_4(
                 );
             }
         }
-        convolve_horizontal_rgb_sse_rows_4_reg(src, src_stride, dst, dst_stride, filter_weights);
+        convolve_horizontal_rgb_avx_rows_4_reg(src, src_stride, dst, dst_stride, filter_weights);
     }
 }
 
 #[target_feature(enable = "avx2")]
-unsafe fn convolve_horizontal_rgb_sse_rows_4_reg(
+unsafe fn convolve_horizontal_rgb_avx_rows_4_reg(
     src: &[u8],
     src_stride: usize,
     dst: &mut [u8],
     dst_stride: usize,
     filter_weights: &FilterWeights<i16>,
 ) {
-    convolve_horizontal_rgb_sse_rows_4_impl::<false>(
+    convolve_horizontal_rgb_avx_rows_4_impl::<false>(
         src,
         src_stride,
         dst,
@@ -79,14 +79,14 @@ unsafe fn convolve_horizontal_rgb_sse_rows_4_reg(
 
 #[cfg(feature = "nightly_avx512")]
 #[target_feature(enable = "avx2", enable = "avxvnni")]
-unsafe fn convolve_horizontal_rgb_sse_rows_4_vnni(
+unsafe fn convolve_horizontal_rgb_avx_rows_4_vnni(
     src: &[u8],
     src_stride: usize,
     dst: &mut [u8],
     dst_stride: usize,
     filter_weights: &FilterWeights<i16>,
 ) {
-    convolve_horizontal_rgb_sse_rows_4_impl::<true>(
+    convolve_horizontal_rgb_avx_rows_4_impl::<true>(
         src,
         src_stride,
         dst,
@@ -167,7 +167,7 @@ unsafe fn make_second_4(pixel: __m128i, pixel2: __m128i, shuf: __m256i) -> __m25
 }
 
 #[inline(always)]
-unsafe fn convolve_horizontal_rgb_sse_rows_4_impl<const HAS_DOT: bool>(
+unsafe fn convolve_horizontal_rgb_avx_rows_4_impl<const HAS_DOT: bool>(
     src: &[u8],
     src_stride: usize,
     dst: &mut [u8],