+add SSE4.1 optimizations of function BgrToYuv420pV2.

ermig1979 · Sep 14, 2023 · ab58297 · ab58297
1 parent 70a6eb4
commit ab58297
Show file tree

Hide file tree

Showing 6 changed files with 89 additions and 12 deletions.
diff --git a/docs/2023.html b/docs/2023.html
@@ -37,7 +37,7 @@ <h3 id="R130">October X, 2023 (version X.X.130)</h3>
 <h4>Algorithms</h4>
 <h5>New features</h5>
 <ul>
- <li>Base implementation of function BgrToYuv420pV2.</li>
+ <li>Base implementation, SSE4.1 optimizations of function BgrToYuv420pV2.</li>
 </ul>
 <ul>
  <li>Error in AVX-512BW optimizations of function SynetSoftmaxLayerForward.</li>

diff --git a/src/Simd/SimdBaseBgrToYuv.cpp b/src/Simd/SimdBaseBgrToYuv.cpp
@@ -124,7 +124,7 @@ namespace Simd
             uint8_t* y1 = y0 + yStride;
 
             y0[0] = BgrToY<YuvType>(bgr0[0], bgr0[1], bgr0[2]);
-            y0[1] = BgrToY<YuvType>(bgr0[2], bgr0[4], bgr0[5]);
+            y0[1] = BgrToY<YuvType>(bgr0[3], bgr0[4], bgr0[5]);
             y1[0] = BgrToY<YuvType>(bgr1[0], bgr1[1], bgr1[2]);
             y1[1] = BgrToY<YuvType>(bgr1[3], bgr1[4], bgr1[5]);
 

diff --git a/src/Simd/SimdLib.cpp b/src/Simd/SimdLib.cpp
@@ -1606,11 +1606,11 @@ SIMD_API void SimdBgrToYuv420pV2(const uint8_t* bgr, size_t bgrStride, size_t wi
 //        Avx2::BgrToYuv420pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);
 //    else
 //#endif
-//#ifdef SIMD_SSE41_ENABLE
-//    if (Sse41::Enable && width >= Sse41::DA)
-//        Sse41::BgrToYuv420pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);
-//    else
-//#endif
+#ifdef SIMD_SSE41_ENABLE
+    if (Sse41::Enable && width >= Sse41::DA)
+        Sse41::BgrToYuv420pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);
+    else
+#endif
 //#ifdef SIMD_NEON_ENABLE
 //    if (Neon::Enable && width >= Neon::DA)
 //        Neon::BgrToYuv420pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);

diff --git a/src/Simd/SimdSse41.h b/src/Simd/SimdSse41.h
@@ -149,6 +149,9 @@ namespace Simd
 
         void BgrToYuv420p(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 BgrToYuv420pV2(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 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 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);

diff --git a/src/Simd/SimdSse41BgrToYuv.cpp b/src/Simd/SimdSse41BgrToYuv.cpp
@@ -24,6 +24,8 @@
 #include "Simd/SimdMemory.h"
 #include "Simd/SimdStore.h"
 #include "Simd/SimdConversion.h"
+#include "Simd/SimdYuvToBgr.h"
+#include "Simd/SimdBase.h"
 
 namespace Simd
 {
@@ -228,6 +230,78 @@ namespace Simd
             else
                 BgrToYuv444p<false>(bgr, width, height, bgrStride, y, yStride, u, uStride, v, vStride);
         }
+
+        //-------------------------------------------------------------------------------------------------
+
+        template <class T> SIMD_INLINE void BgrToYuv420pV2(const uint8_t* bgr0, size_t bgrStride, uint8_t* y0, size_t yStride, uint8_t* u, uint8_t* v)
+        {
+            const uint8_t* bgr1 = bgr0 + bgrStride;
+            uint8_t* y1 = y0 + yStride;
+
+            __m128i blue[2][2], green[2][2], red[2][2];
+
+            LoadBgr<false>((__m128i*)bgr0 + 0, blue[0][0], green[0][0], red[0][0]);
+            _mm_storeu_si128((__m128i*)y0 + 0, BgrToY8<T>(blue[0][0], green[0][0], red[0][0]));
+
+            LoadBgr<false>((__m128i*)bgr0 + 3, blue[0][1], green[0][1], red[0][1]);
+            _mm_storeu_si128((__m128i*)y0 + 1, BgrToY8<T>(blue[0][1], green[0][1], red[0][1]));
+
+            LoadBgr<false>((__m128i*)bgr1 + 0, blue[1][0], green[1][0], red[1][0]);
+            _mm_storeu_si128((__m128i*)y1 + 0, BgrToY8<T>(blue[1][0], green[1][0], red[1][0]));
+
+            LoadBgr<false>((__m128i*)bgr1 + 3, blue[1][1], green[1][1], red[1][1]);
+            _mm_storeu_si128((__m128i*)y1 + 1, BgrToY8<T>(blue[1][1], green[1][1], red[1][1]));
+
+            blue[0][0] = Average16(blue[0][0], blue[1][0]);
+            blue[0][1] = Average16(blue[0][1], blue[1][1]);
+            green[0][0] = Average16(green[0][0], green[1][0]);
+            green[0][1] = Average16(green[0][1], green[1][1]);
+            red[0][0] = Average16(red[0][0], red[1][0]);
+            red[0][1] = Average16(red[0][1], red[1][1]);
+
+            _mm_storeu_si128((__m128i*)u, _mm_packus_epi16(BgrToU16<T>(blue[0][0], green[0][0], red[0][0]), BgrToU16<T>(blue[0][1], green[0][1], red[0][1])));
+            _mm_storeu_si128((__m128i*)v, _mm_packus_epi16(BgrToV16<T>(blue[0][0], green[0][0], red[0][0]), BgrToV16<T>(blue[0][1], green[0][1], red[0][1])));
+        }
+
+        template <class T>  void BgrToYuv420pV2(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) && (height % 2 == 0) && (width >= DA) && (height >= 2));
+
+            size_t widthDA = AlignLo(width, DA);
+            for (size_t row = 0; row < height; row += 2)
+            {
+                for (size_t colUV = 0, colY = 0, colBgr = 0; colY < widthDA; colY += DA, colUV += A, colBgr += A * 6)
+                    BgrToYuv420pV2<T>(bgr + colBgr, bgrStride, y + colY, yStride, u + colUV, v + colUV);
+                if (width != widthDA)
+                {
+                    size_t colY = width - DA;
+                    BgrToYuv420pV2<T>(bgr + colY * 3, bgrStride, y + colY, yStride, u + colY / 2, v + colY / 2);
+                }
+                y += 2 * yStride;
+                u += uStride;
+                v += vStride;
+                bgr += 2 * bgrStride;
+            }
+        }
+
+        void BgrToYuv420pV2(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::BgrToYuv420pV2(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride, yuvType);
+#else
+            switch (yuvType)
+            {
+            case SimdYuvBt601: BgrToYuv420pV2<Base::Bt601>(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride); break;
+            case SimdYuvBt709: BgrToYuv420pV2<Base::Bt709>(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride); break;
+            case SimdYuvBt2020: BgrToYuv420pV2<Base::Bt2020>(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride); break;
+            case SimdYuvTrect871: BgrToYuv420pV2<Base::Trect871>(bgr, bgrStride, width, height, y, yStride, u, uStride, v, vStride); break;
+            default:
+                assert(0);
+            }
+#endif
+        }
     }
 #endif
 }
diff --git a/src/Test/TestAnyToYuv.cpp b/src/Test/TestAnyToYuv.cpp
@@ -505,11 +505,11 @@ namespace Test
 
         result = result && AnyToYuvV2AutoTest(View::Bgr24, 2, 2, FUNC_YUV2(Simd::Base::BgrToYuv420pV2), FUNC_YUV2(SimdBgrToYuv420pV2));
 
-//#ifdef SIMD_SSE41_ENABLE
-//        if (Simd::Sse41::Enable)
-//            result = result && AnyToYuvV2AutoTest(View::Bgr24, 2, 2, FUNC_YUV2(Simd::Sse41::BgrToYuv420pV2), FUNC_YUV2(SimdBgrToYuv420pV2));
-//#endif 
-//
+#ifdef SIMD_SSE41_ENABLE
+        if (Simd::Sse41::Enable)
+            result = result && AnyToYuvV2AutoTest(View::Bgr24, 2, 2, FUNC_YUV2(Simd::Sse41::BgrToYuv420pV2), FUNC_YUV2(SimdBgrToYuv420pV2));
+#endif 
+
 //#ifdef SIMD_AVX2_ENABLE
 //        if (Simd::Avx2::Enable)
 //            result = result && AnyToYuvV2AutoTest(View::Bgr24, 2, 2, FUNC_YUV2(Simd::Avx2::BgrToYuv420pV2), FUNC_YUV2(SimdBgrToYuv420pV2));