+add AVX2 optimizations of function BgrToYuv422pV2.

ermig1979 · Sep 18, 2023 · 657e54c · 657e54c
1 parent 49cf217
commit 657e54c
Show file tree

Hide file tree

Showing 6 changed files with 78 additions and 12 deletions.
diff --git a/docs/2023.html b/docs/2023.html
@@ -38,7 +38,7 @@ <h4>Algorithms</h4>
 <h5>New features</h5>
 <ul>
  <li>Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function BgrToYuv420pV2.</li>
- <li>Base implementation, SSE4.1 optimizations of function BgrToYuv422pV2.</li>
+ <li>Base implementation, SSE4.1, AVX2 optimizations of function BgrToYuv422pV2.</li>
 </ul>
 <ul>
  <li>Error in AVX-512BW optimizations of function SynetSoftmaxLayerForward.</li>

diff --git a/src/Simd/SimdAvx2.h b/src/Simd/SimdAvx2.h
@@ -153,6 +153,9 @@ namespace Simd
 
         void BgrToYuv422p(const uint8_t * bgr, size_t width, size_t height, size_t bgrStride, uint8_t * y, size_t yStride, uint8_t * u, size_t uStride, uint8_t * v, size_t vStride);
 
+        void BgrToYuv422pV2(const uint8_t* bgr, size_t bgrStride, size_t width, size_t height,
+            uint8_t* y, size_t yStride, uint8_t* u, size_t uStride, uint8_t* v, size_t vStride, SimdYuvType yuvType);
+
         void BgrToYuv444p(const uint8_t * bgr, size_t width, size_t height, size_t bgrStride, uint8_t * y, size_t yStride, uint8_t * u, size_t uStride, uint8_t * v, size_t vStride);
 
         void Binarization(const uint8_t * src, size_t srcStride, size_t width, size_t height,

diff --git a/src/Simd/SimdAvx2BgrToYuv.cpp b/src/Simd/SimdAvx2BgrToYuv.cpp
@@ -309,6 +309,69 @@ namespace Simd
             default:
                 assert(0);
             }
+#endif
+        }
+
+        //-------------------------------------------------------------------------------------------------
+
+        template <class T> SIMD_INLINE void BgrToYuv422pV2(const uint8_t* bgr, uint8_t* y, uint8_t* u, uint8_t* v)
+        {
+            __m256i blue[2], green[2], red[2];
+
+            LoadBgr<false>((__m256i*)bgr + 0, blue[0], green[0], red[0]);
+            _mm256_storeu_si256((__m256i*)y + 0, BgrToY8<T>(blue[0], green[0], red[0]));
+
+            LoadBgr<false>((__m256i*)bgr + 3, blue[1], green[1], red[1]);
+            _mm256_storeu_si256((__m256i*)y + 1, BgrToY8<T>(blue[1], green[1], red[1]));
+
+            Average16(blue[0]);
+            Average16(blue[1]);
+            Average16(green[0]);
+            Average16(green[1]);
+            Average16(red[0]);
+            Average16(red[1]);
+
+            _mm256_storeu_si256((__m256i*)u, PackI16ToU8(BgrToU16<T>(blue[0], green[0], red[0]), BgrToU16<T>(blue[1], green[1], red[1])));
+            _mm256_storeu_si256((__m256i*)v, PackI16ToU8(BgrToV16<T>(blue[0], green[0], red[0]), BgrToV16<T>(blue[1], green[1], red[1])));
+        }
+
+        template <class T>  void BgrToYuv422pV2(const uint8_t* bgr, size_t bgrStride, size_t width, size_t height, uint8_t* y, size_t yStride,
+            uint8_t* u, size_t uStride, uint8_t* v, size_t vStride)
+        {
+            assert((width % 2 == 0) && (width >= DA));
+
+            size_t widthDA = AlignLo(width, DA);
+            for (size_t row = 0; row < height; row += 1)
+            {
+                for (size_t colUV = 0, colY = 0, colBgr = 0; colY < widthDA; colY += DA, colUV += A, colBgr += A * 6)
+                    BgrToYuv422pV2<T>(bgr + colBgr, y + colY, u + colUV, v + colUV);
+                if (width != widthDA)
+                {
+                    size_t colY = width - DA;
+                    BgrToYuv422pV2<T>(bgr + colY * 3, y + colY, u + colY / 2, v + colY / 2);
+                }
+                y += yStride;
+                u += uStride;
+                v += vStride;
+                bgr += bgrStride;
+            }
+        }
+
+        void BgrToYuv422pV2(const uint8_t* bgr, size_t bgrStride, size_t width, size_t height, uint8_t* y, size_t yStride,
+            uint8_t* u, size_t uStride, uint8_t* v, size_t vStride, SimdYuvType yuvType)
+        {
+#if defined(SIMD_X86_ENABLE) && defined(NDEBUG) && defined(_MSC_VER) && _MSC_VER <= 1900
+            Base::BgrToYuv422pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);
+#else
+            switch (yuvType)
+            {
+            case SimdYuvBt601: BgrToYuv422pV2<Base::Bt601>(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride); break;
+            case SimdYuvBt709: BgrToYuv422pV2<Base::Bt709>(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride); break;
+            case SimdYuvBt2020: BgrToYuv422pV2<Base::Bt2020>(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride); break;
+            case SimdYuvTrect871: BgrToYuv422pV2<Base::Trect871>(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride); break;
+            default:
+                assert(0);
+            }
 #endif
         }
     }

diff --git a/src/Simd/SimdLib.cpp b/src/Simd/SimdLib.cpp
@@ -1659,11 +1659,11 @@ SIMD_API void SimdBgrToYuv422pV2(const uint8_t* bgr, size_t bgrStride, size_t wi
 //        Avx512bw::BgrToYuv422pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);
 //    else
 //#endif
-//#ifdef SIMD_AVX2_ENABLE
-//    if (Avx2::Enable && width >= Avx2::DA)
-//        Avx2::BgrToYuv422pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);
-//    else
-//#endif
+#ifdef SIMD_AVX2_ENABLE
+    if (Avx2::Enable && width >= Avx2::DA)
+        Avx2::BgrToYuv422pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);
+    else
+#endif
 #ifdef SIMD_SSE41_ENABLE
     if (Sse41::Enable && width >= Sse41::DA)
         Sse41::BgrToYuv422pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);

diff --git a/src/Simd/SimdSse41BgrToYuv.cpp b/src/Simd/SimdSse41BgrToYuv.cpp
@@ -352,7 +352,7 @@ namespace Simd
             uint8_t* u, size_t uStride, uint8_t* v, size_t vStride, SimdYuvType yuvType)
         {
 #if defined(SIMD_X86_ENABLE) && defined(NDEBUG) && defined(_MSC_VER) && _MSC_VER <= 1900
-            Base::BgrToYuv420pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);
+            Base::BgrToYuv422pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);
 #else
             switch (yuvType)
             {

diff --git a/src/Test/TestAnyToYuv.cpp b/src/Test/TestAnyToYuv.cpp
@@ -539,11 +539,11 @@ namespace Test
             result = result && AnyToYuvV2AutoTest(View::Bgr24, 2, 1, FUNC_YUV2(Simd::Sse41::BgrToYuv422pV2), FUNC_YUV2(SimdBgrToYuv422pV2));
 #endif 
 
-//#ifdef SIMD_AVX2_ENABLE
-//        if (Simd::Avx2::Enable)
-//            result = result && AnyToYuvV2AutoTest(View::Bgr24, 2, 1, FUNC_YUV2(Simd::Avx2::BgrToYuv422pV2), FUNC_YUV2(SimdBgrToYuv422pV2));
-//#endif 
-//
+#ifdef SIMD_AVX2_ENABLE
+        if (Simd::Avx2::Enable)
+            result = result && AnyToYuvV2AutoTest(View::Bgr24, 2, 1, FUNC_YUV2(Simd::Avx2::BgrToYuv422pV2), FUNC_YUV2(SimdBgrToYuv422pV2));
+#endif 
+
 //#ifdef SIMD_AVX512BW_ENABLE
 //        if (Simd::Avx512bw::Enable)
 //            result = result && AnyToYuvV2AutoTest(View::Bgr24, 2, 1, FUNC_YUV2(Simd::Avx512bw::BgrToYuv422pV2), FUNC_YUV2(SimdBgrToYuv422pV2));