Rounding and bugfixes

README.md

@@ -12,6 +12,7 @@ Supports:
 - [x] YCgCo-R
 - [x] YUY2
 - [x] Identity ( GBR )
+- [x] Sharp YUV
 
 # SIMD
 
@@ -65,11 +66,11 @@ yuv422_to_rgb(&y_plane, y_stride,
 
 ```rust
 rgb_to_ycgco420(&mut y_plane, y_stride,
-              &mut cg_plane, cg_width,
-              &mut cg_plane, cg_width,
-              &rgb, rgb_stride,
-              width, height, 
-              YuvRange::TV);
+                &mut cg_plane, cg_width,
+                &mut cg_plane, cg_width,
+                &rgb, rgb_stride,
+                width, height, 
+                YuvRange::TV);
 ```
 
 ### YCgCo to RGB

src/avx2/yuv_nv_to_rgba.rs

@@ -56,7 +56,7 @@ pub unsafe fn avx2_yuv_nv_to_rgba_row<
     let rounding_const = _mm256_set1_epi16(1 << 5);
 
     while cx + 32 < width {
-        let y_values = _mm256_subs_epi8(
+        let y_values = _mm256_subs_epu8(
             _mm256_loadu_si256(y_ptr.add(y_offset + cx) as *const __m256i),
             y_corr,
         );

src/avx2/yuv_to_rgba.rs

@@ -54,7 +54,7 @@ pub unsafe fn avx2_yuv_to_rgba_row<const DESTINATION_CHANNELS: u8, const SAMPLIN
     let rounding_const = _mm256_set1_epi16(1 << 5);
 
     while cx + 32 < width {
-        let y_values = _mm256_subs_epi8(
+        let y_values = _mm256_subs_epu8(
             _mm256_loadu_si256(y_ptr.add(y_offset + cx) as *const __m256i),
             y_corr,
         );

src/avx2/yuv_to_rgba_alpha.rs

@@ -57,7 +57,7 @@ pub unsafe fn avx2_yuv_to_rgba_alpha<const DESTINATION_CHANNELS: u8, const SAMPL
     let rounding_const = _mm256_set1_epi16(1 << 5);
 
     while cx + 32 < width {
-        let y_values = _mm256_subs_epi8(
+        let y_values = _mm256_subs_epu8(
             _mm256_loadu_si256(y_ptr.add(y_offset + cx) as *const __m256i),
             y_corr,
         );

src/avx512bw/yuv_nv_to_rgba.rs

@@ -58,7 +58,7 @@ pub unsafe fn avx512_yuv_nv_to_rgba<
     let v_alpha = _mm512_set1_epi8(255u8 as i8);
 
     while cx + 32 < width {
-        let y_values = _mm512_subs_epi8(
+        let y_values = _mm512_subs_epu8(
             _mm512_loadu_si512(y_ptr.add(y_offset + cx) as *const i32),
             y_corr,
         );

src/avx512bw/yuv_to_rgba.rs

@@ -53,7 +53,7 @@ pub unsafe fn avx512_yuv_to_rgba<const DESTINATION_CHANNELS: u8, const SAMPLING:
     let v_alpha = _mm512_set1_epi8(255u8 as i8);
 
     while cx + 64 < width {
-        let y_values = _mm512_subs_epi8(
+        let y_values = _mm512_subs_epu8(
             _mm512_loadu_si512(y_ptr.add(y_offset + cx) as *const i32),
             y_corr,
         );

src/avx512bw/yuv_to_rgba_alpha.rs

@@ -55,7 +55,7 @@ pub unsafe fn avx512_yuv_to_rgba_alpha<const DESTINATION_CHANNELS: u8, const SAM
     let v_g_coeff_2 = _mm512_set1_epi16(-1 * transform.g_coeff_2 as i16);
 
     while cx + 64 < width {
-        let y_values = _mm512_subs_epi8(
+        let y_values = _mm512_subs_epu8(
             _mm512_loadu_si512(y_ptr.add(y_offset + cx) as *const i32),
             y_corr,
         );

src/rgb_to_yuv_p16.rs

@@ -66,9 +66,10 @@ fn rgbx_to_yuv_impl<
         kr_kb.kb,
     );
     let transform = transform_precise.to_integers(8);
-    let precision_scale = (1 << 8) as f32;
-    let bias_y = ((range.bias_y as f32 + 0.5f32) * precision_scale) as i32;
-    let bias_uv = ((range.bias_uv as f32 + 0.5f32) * precision_scale) as i32;
+    const PRECISION: i32 = 8;
+    const ROUNDING_CONST_BIAS: i32 = 1 << (PRECISION - 1);
+    let bias_y = range.bias_y as i32 * (1 << PRECISION) + ROUNDING_CONST_BIAS;
+    let bias_uv = range.bias_uv as i32 * (1 << PRECISION) + ROUNDING_CONST_BIAS;
 
     let iterator_step = match chroma_subsampling {
         YuvChromaSample::YUV420 => 2usize,
@@ -152,9 +153,9 @@ fn rgbx_to_yuv_impl<
             let r = unsafe { src.add(src_chans.get_r_channel_offset()).read_unaligned() } as i32;
             let g = unsafe { src.add(src_chans.get_g_channel_offset()).read_unaligned() } as i32;
             let b = unsafe { src.add(src_chans.get_b_channel_offset()).read_unaligned() } as i32;
-            let y_0 = (r * transform.yr + g * transform.yg + b * transform.yb + bias_y) >> 8;
-            let cb = (r * transform.cb_r + g * transform.cb_g + b * transform.cb_b + bias_uv) >> 8;
-            let cr = (r * transform.cr_r + g * transform.cr_g + b * transform.cr_b + bias_uv) >> 8;
+            let y_0 = (r * transform.yr + g * transform.yg + b * transform.yb + bias_y) >> PRECISION;
+            let cb = (r * transform.cb_r + g * transform.cb_g + b * transform.cb_b + bias_uv) >> PRECISION;
+            let cr = (r * transform.cr_r + g * transform.cr_g + b * transform.cr_b + bias_uv) >> PRECISION;
             unsafe {
                 y_st_ptr.add(x).write_unaligned(transform_integer::<
                     ENDIANNESS,
@@ -199,7 +200,7 @@ fn rgbx_to_yuv_impl<
                             unsafe { src.add(src_chans.get_b_channel_offset()).read_unaligned() }
                                 as i32;
                         let y_1 =
-                            (r * transform.yr + g * transform.yg + b * transform.yb + bias_y) >> 8;
+                            (r * transform.yr + g * transform.yg + b * transform.yb + bias_y) >> PRECISION;
                         unsafe {
                             y_st_ptr.add(x + 1).write_unaligned(transform_integer::<
                                 ENDIANNESS,

src/rgba_to_nv.rs

@@ -38,10 +38,11 @@ fn rgbx_to_nv<const ORIGIN_CHANNELS: u8, const UV_ORDER: u8, const SAMPLING: u8>
         kr_kb.kr,
         kr_kb.kb,
     );
-    let transform = transform_precise.to_integers(8);
-    const ROUNDING_CONST_BIAS: i32 = 1 << 7;
-    let bias_y = range.bias_y as i32 * (1 << 8) + ROUNDING_CONST_BIAS;
-    let bias_uv = range.bias_uv as i32 * (1 << 8) + ROUNDING_CONST_BIAS;
+    const PRECISION: i32 = 8;
+    let transform = transform_precise.to_integers(PRECISION as u32);
+    const ROUNDING_CONST_BIAS: i32 = 1 << (PRECISION - 1);
+    let bias_y = range.bias_y as i32 * (1 << PRECISION) + ROUNDING_CONST_BIAS;
+    let bias_uv = range.bias_uv as i32 * (1 << PRECISION) + ROUNDING_CONST_BIAS;
 
     let iterator_step = match chroma_subsampling {
         YuvChromaSample::YUV420 => 2usize,
@@ -135,9 +136,12 @@ fn rgbx_to_nv<const ORIGIN_CHANNELS: u8, const UV_ORDER: u8, const SAMPLING: u8>
                 as i32;
             let b = unsafe { *source_slice.get_unchecked(source_channels.get_b_channel_offset()) }
                 as i32;
-            let y_0 = (r * transform.yr + g * transform.yg + b * transform.yb + bias_y) >> 8;
-            let cb = (r * transform.cb_r + g * transform.cb_g + b * transform.cb_b + bias_uv) >> 8;
-            let cr = (r * transform.cr_r + g * transform.cr_g + b * transform.cr_b + bias_uv) >> 8;
+            let y_0 =
+                (r * transform.yr + g * transform.yg + b * transform.yb + bias_y) >> PRECISION;
+            let cb = (r * transform.cb_r + g * transform.cb_g + b * transform.cb_b + bias_uv)
+                >> PRECISION;
+            let cr = (r * transform.cr_r + g * transform.cr_g + b * transform.cr_b + bias_uv)
+                >> PRECISION;
             unsafe {
                 *y_plane.get_unchecked_mut(y_offset + x) = y_0.clamp(i_bias_y, i_cap_y) as u8;
             }
@@ -164,8 +168,8 @@ fn rgbx_to_nv<const ORIGIN_CHANNELS: u8, const UV_ORDER: u8, const SAMPLING: u8>
                         let b = unsafe {
                             *source_slice.get_unchecked(source_channels.get_b_channel_offset())
                         } as i32;
-                        let y_1 =
-                            (r * transform.yr + g * transform.yg + b * transform.yb + bias_y) >> 8;
+                        let y_1 = (r * transform.yr + g * transform.yg + b * transform.yb + bias_y)
+                            >> PRECISION;
                         unsafe {
                             *y_plane.get_unchecked_mut(y_offset + next_x) =
                                 y_1.clamp(i_bias_y, i_cap_y) as u8;

src/sse/rgba_to_yuv.rs

@@ -45,10 +45,11 @@ pub unsafe fn sse_rgba_to_yuv_row<const ORIGIN_CHANNELS: u8, const SAMPLING: u8>
 
     let mut cx = start_cx;
     let mut uv_x = start_ux;
+    const PRECISION: i32 = 8;
 
-    const ROUNDING_CONST_BIAS: i32 = 1 << 7;
-    let bias_y = range.bias_y as i32 * (1 << 8) + ROUNDING_CONST_BIAS;
-    let bias_uv = range.bias_uv as i32 * (1 << 8) + ROUNDING_CONST_BIAS;
+    const ROUNDING_CONST_BIAS: i32 = 1 << (PRECISION - 1);
+    let bias_y = range.bias_y as i32 * (1 << PRECISION) + ROUNDING_CONST_BIAS;
+    let bias_uv = range.bias_uv as i32 * (1 << PRECISION) + ROUNDING_CONST_BIAS;
 
     let zeros = _mm_setzero_si128();
 

src/sse/yuv_nv_to_rgba.rs

@@ -62,7 +62,7 @@ pub unsafe fn sse_yuv_nv_to_rgba<
     let distribute_shuffle = _mm_setr_epi8(0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7);
 
     while cx + 16 < width {
-        let y_values = _mm_subs_epi8(
+        let y_values = _mm_subs_epu8(
             _mm_loadu_si128(y_ptr.add(y_offset + cx) as *const __m128i),
             y_corr,
         );

src/sse/yuv_to_rgba.rs

@@ -49,7 +49,6 @@ pub unsafe fn sse_yuv_to_rgba_row<const DESTINATION_CHANNELS: u8, const SAMPLING
     let v_luma_coeff = _mm_set1_epi16(transform.y_coef as i16);
     let v_cr_coeff = _mm_set1_epi16(transform.cr_coef as i16);
     let v_cb_coeff = _mm_set1_epi16(transform.cb_coef as i16);
-    let v_min_values = _mm_setzero_si128();
     let v_g_coeff_1 = _mm_set1_epi16(-1 * transform.g_coeff_1 as i16);
     let v_g_coeff_2 = _mm_set1_epi16(-1 * transform.g_coeff_2 as i16);
     let v_alpha = _mm_set1_epi8(255u8 as i8);
@@ -58,7 +57,7 @@ pub unsafe fn sse_yuv_to_rgba_row<const DESTINATION_CHANNELS: u8, const SAMPLING
     let zeros = _mm_setzero_si128();
 
     while cx + 16 < width {
-        let y_values = _mm_subs_epi8(
+        let y_values = _mm_subs_epu8(
             _mm_loadu_si128(y_ptr.add(y_offset + cx) as *const __m128i),
             y_corr,
         );
@@ -96,14 +95,14 @@ pub unsafe fn sse_yuv_to_rgba_row<const DESTINATION_CHANNELS: u8, const SAMPLING
         let r_high = _mm_srai_epi16::<6>(_mm_adds_epi16(
             _mm_max_epi16(
                 _mm_adds_epi16(y_high, _mm_mullo_epi16(v_high, v_cr_coeff)),
-                v_min_values,
+                zeros,
             ),
             rounding_const,
         ));
         let b_high = _mm_srai_epi16::<6>(_mm_adds_epi16(
             _mm_max_epi16(
                 _mm_adds_epi16(y_high, _mm_mullo_epi16(u_high, v_cb_coeff)),
-                v_min_values,
+                zeros,
             ),
             rounding_const,
         ));
@@ -116,7 +115,7 @@ pub unsafe fn sse_yuv_to_rgba_row<const DESTINATION_CHANNELS: u8, const SAMPLING
                         _mm_mullo_epi16(u_high, v_g_coeff_2),
                     ),
                 ),
-                v_min_values,
+                zeros,
             ),
             rounding_const,
         ));
@@ -128,14 +127,14 @@ pub unsafe fn sse_yuv_to_rgba_row<const DESTINATION_CHANNELS: u8, const SAMPLING
         let r_low = _mm_srai_epi16::<6>(_mm_adds_epi16(
             _mm_max_epi16(
                 _mm_adds_epi16(y_low, _mm_mullo_epi16(v_low, v_cr_coeff)),
-                v_min_values,
+                zeros,
             ),
             rounding_const,
         ));
         let b_low = _mm_srai_epi16::<6>(_mm_adds_epi16(
             _mm_max_epi16(
                 _mm_adds_epi16(y_low, _mm_mullo_epi16(u_low, v_cb_coeff)),
-                v_min_values,
+                zeros,
             ),
             rounding_const,
         ));
@@ -148,7 +147,7 @@ pub unsafe fn sse_yuv_to_rgba_row<const DESTINATION_CHANNELS: u8, const SAMPLING
                         _mm_mullo_epi16(u_low, v_g_coeff_2),
                     ),
                 ),
-                v_min_values,
+                zeros,
             ),
             rounding_const,
         ));
@@ -230,14 +229,14 @@ pub unsafe fn sse_yuv_to_rgba_row<const DESTINATION_CHANNELS: u8, const SAMPLING
         let r_low = _mm_srai_epi16::<6>(_mm_adds_epi16(
             _mm_max_epi16(
                 _mm_adds_epi16(y_low, _mm_mullo_epi16(v_low, v_cr_coeff)),
-                v_min_values,
+                zeros,
             ),
             rounding_const,
         ));
         let b_low = _mm_srai_epi16::<6>(_mm_adds_epi16(
             _mm_max_epi16(
                 _mm_adds_epi16(y_low, _mm_mullo_epi16(u_low, v_cb_coeff)),
-                v_min_values,
+                zeros,
             ),
             rounding_const,
         ));
@@ -250,7 +249,7 @@ pub unsafe fn sse_yuv_to_rgba_row<const DESTINATION_CHANNELS: u8, const SAMPLING
                         _mm_mullo_epi16(u_low, v_g_coeff_2),
                     ),
                 ),
-                v_min_values,
+                zeros,
             ),
             rounding_const,
         ));

src/sse/yuv_to_rgba_alpha.rs

@@ -60,7 +60,7 @@ pub unsafe fn sse_yuv_to_rgba_alpha_row<const DESTINATION_CHANNELS: u8, const SA
     let zeros = _mm_setzero_si128();
 
     while cx + 16 < width {
-        let y_values = _mm_subs_epi8(
+        let y_values = _mm_subs_epu8(
             _mm_loadu_si128(y_ptr.add(y_offset + cx) as *const __m128i),
             y_corr,
         );