JVET-R0208: Rounding offset fix for ALF virtual boundary processing

source/Lib/CommonLib/AdaptiveLoopFilter.cpp

@@ -1262,7 +1262,11 @@ void AdaptiveLoopFilter::filterBlk(AlfClassifier **classifier, const PelUnitBuf
           }
           else
           {
+#if JVET_R0208_ALF_VB_ROUNDING_FIX
+            sum = (sum + (1 << ((shift + 3) - 1))) >> (shift + 3);
+#else
             sum = (sum + offset) >> (shift + 3);
+#endif
           }
           sum += curr;
           pRec1[jj] = ClipPel( sum, clpRng );

source/Lib/CommonLib/TypeDef.h

@@ -53,6 +53,7 @@
 //########### place macros to be removed in next cycle below this line ###############
 
 #define JVET_Q0471_CHROMA_QT_SPLIT                        1 // JVET-Q0471: Chroma QT split
+#define JVET_R0208_ALF_VB_ROUNDING_FIX                    1 // JVET-R0208: Rounding offset fix for ALF virtual boundary processing
 
 
 //########### place macros to be be kept below this line ###############

source/Lib/CommonLib/x86/AdaptiveLoopFilterX86.h

@@ -313,6 +313,9 @@ static void simdFilter5x5Blk(AlfClassifier **classifier, const PelUnitBuf &recDs
 
   constexpr int SHIFT = AdaptiveLoopFilter::m_NUM_BITS - 1;
   constexpr int ROUND = 1 << (SHIFT - 1);
+#if JVET_R0208_ALF_VB_ROUNDING_FIX
+  const __m128i mmOffset1 = _mm_set1_epi32((1 << ((SHIFT + 3) - 1)) - ROUND);
+#endif
 
   const size_t width  = blk.width;
   const size_t height = blk.height;
@@ -425,8 +428,13 @@ static void simdFilter5x5Blk(AlfClassifier **classifier, const PelUnitBuf &recDs
         }
         else
         {
+#if JVET_R0208_ALF_VB_ROUNDING_FIX
+          accumA = _mm_srai_epi32(_mm_add_epi32(accumA, mmOffset1), SHIFT + 3);
+          accumB = _mm_srai_epi32(_mm_add_epi32(accumB, mmOffset1), SHIFT + 3);
+#else
           accumA = _mm_srai_epi32(accumA, SHIFT + 3);
           accumB = _mm_srai_epi32(accumB, SHIFT + 3);
+#endif
         }
         accumA = _mm_packs_epi32(accumA, accumB);
         accumA = _mm_add_epi16(accumA, cur);
@@ -507,6 +515,9 @@ static void simdFilter7x7Blk(AlfClassifier **classifier, const PelUnitBuf &recDs
   Pel *      dst = dstBuffer.buf + blkDst.y * dstStride + blkDst.x;
 
   const __m128i mmOffset = _mm_set1_epi32(ROUND);
+#if JVET_R0208_ALF_VB_ROUNDING_FIX
+  const __m128i mmOffset1 = _mm_set1_epi32((1 << ((SHIFT + 3) - 1)) - ROUND);
+#endif
   const __m128i mmMin = _mm_set1_epi16( clpRng.min );
   const __m128i mmMax = _mm_set1_epi16( clpRng.max );
 
@@ -655,8 +666,13 @@ static void simdFilter7x7Blk(AlfClassifier **classifier, const PelUnitBuf &recDs
         }
         else
         {
+#if JVET_R0208_ALF_VB_ROUNDING_FIX
+          accumA = _mm_srai_epi32(_mm_add_epi32(accumA, mmOffset1), SHIFT + 3);
+          accumB = _mm_srai_epi32(_mm_add_epi32(accumB, mmOffset1), SHIFT + 3);
+#else
           accumA = _mm_srai_epi32(accumA, SHIFT + 3);
           accumB = _mm_srai_epi32(accumB, SHIFT + 3);
+#endif 
         }
         accumA = _mm_packs_epi32(accumA, accumB);
         accumA = _mm_add_epi16(accumA, cur);