Ticket #1320: fix_1320_rdpcm_accumulator.patch

source/Lib/TLibCommon/TComTrQuant.cpp

@@ -1630 +1630 @@
   UInt uiX = 0;
   UInt uiY = 0;
 
+#if FIX_1320_RDPCM_ACCUMULATOR
+        UInt &majorAxis             = (mode == RDPCM_VER) ? uiX      : uiY;
+        UInt &minorAxis             = (mode == RDPCM_VER) ? uiY      : uiX;
+  const UInt  majorAxisLimit        = (mode == RDPCM_VER) ? uiWidth  : uiHeight;
+  const UInt  minorAxisLimit        = (mode == RDPCM_VER) ? uiHeight : uiWidth;
+  static const TCoeff pelMin=(Int) std::numeric_limits<Pel>::min();
+  static const TCoeff pelMax=(Int) std::numeric_limits<Pel>::max();
+#else
         UInt &majorAxis             = (mode == RDPCM_HOR) ? uiX      : uiY;
         UInt &minorAxis             = (mode == RDPCM_HOR) ? uiY      : uiX;
   const UInt  majorAxisLimit        = (mode == RDPCM_HOR) ? uiWidth  : uiHeight;
   const UInt  minorAxisLimit        = (mode == RDPCM_HOR) ? uiHeight : uiWidth;
   const UInt  referenceSampleOffset = (mode == RDPCM_HOR) ? 1        : uiWidth;
+#endif
 
-  const Bool bUseHalfRoundingPoint = (mode != RDPCM_OFF);
+  const Bool bUseHalfRoundingPoint  = (mode != RDPCM_OFF);
 
   uiAbsSum = 0;
 
   for ( majorAxis = 0; majorAxis < majorAxisLimit; majorAxis++ )
   {
+#if FIX_1320_RDPCM_ACCUMULATOR
+    TCoeff accumulatorValue = 0; // 32-bit accumulator
+#endif
     for ( minorAxis = 0; minorAxis < minorAxisLimit; minorAxis++ )
     {
       const UInt sampleIndex      = (uiY * uiWidth) + uiX;
       const UInt coefficientIndex = (rotateResidual ? (uiSizeMinus1-sampleIndex) : sampleIndex);
       const Pel  currentSample    = pcResidual[(uiY * uiStride) + uiX];
+#if FIX_1320_RDPCM_ACCUMULATOR
+      const TCoeff encoderSideDelta = TCoeff(currentSample) - accumulatorValue;
+#else
       const Pel  referenceSample  = ((mode != RDPCM_OFF) && (majorAxis > 0)) ? reconstructedResi[sampleIndex - referenceSampleOffset] : 0;
 
       const Pel  encoderSideDelta = currentSample - referenceSample;
+#endif
 
       Pel reconstructedDelta;
       if ( bLossless )
@@ -1665 +1681 @@
 
       uiAbsSum += abs(pcCoeff[coefficientIndex]);
 
+#if FIX_1320_RDPCM_ACCUMULATOR
+      if (mode == RDPCM_OFF) { reconstructedResi[sampleIndex] = reconstructedDelta; }
+      else
+      {
+        accumulatorValue += reconstructedDelta;
+        reconstructedResi[sampleIndex] = (Pel) Clip3<TCoeff>(pelMin, pelMax, accumulatorValue);
+      }
+#else
       reconstructedResi[sampleIndex] = reconstructedDelta + referenceSample;
+#endif
     }
   }
 }
@@ -1764 +1789 @@
       rdpcmMode = RDPCMMode(pcCU->getExplicitRdpcmMode( compID, uiAbsPartIdx ));
     }
 
+#if FIX_1320_RDPCM_ACCUMULATOR
+    static const TCoeff pelMin=(Int) std::numeric_limits<Pel>::min();
+    static const TCoeff pelMax=(Int) std::numeric_limits<Pel>::max();
+#endif
     if (rdpcmMode == RDPCM_VER)
     {
+#if FIX_1320_RDPCM_ACCUMULATOR
+      for( UInt uiX = 0; uiX < uiWidth; uiX++ )
+      {
+        Pel *pcCurResidual = pcResidual+uiX;
+        TCoeff accumulator = *pcCurResidual; // 32-bit accumulator
+        pcCurResidual+=uiStride;
+        for( UInt uiY = 1; uiY < uiHeight; uiY++, pcCurResidual+=uiStride )
+        {
+          accumulator += *(pcCurResidual);
+          *pcCurResidual = (Pel)Clip3<TCoeff>(pelMin, pelMax, accumulator);
+        }
+      }
+#else
       pcResidual += uiStride; //start from row 1
 
       for( UInt uiY = 1; uiY < uiHeight; uiY++ )
@@ -1776 +1818 @@
         }
         pcResidual += uiStride;
       }
+#endif
     }
     else if (rdpcmMode == RDPCM_HOR)
     {
+#if FIX_1320_RDPCM_ACCUMULATOR
+      for( UInt uiY = 0; uiY < uiHeight; uiY++ )
+      {
+        Pel *pcCurResidual = pcResidual+uiY*uiStride;
+        TCoeff accumulator = *pcCurResidual;
+        pcCurResidual++;
+        for( UInt uiX = 1; uiX < uiWidth; uiX++, pcCurResidual++ )
+        {
+          accumulator += *(pcCurResidual);
+          *pcCurResidual = (Pel)Clip3<TCoeff>(pelMin, pelMax, accumulator);
+        }
+      }
+#else
       for( UInt uiY = 0; uiY < uiHeight; uiY++ )
       {
         for( UInt uiX = 1; uiX < uiWidth; uiX++ )
@@ -1787 +1843 @@
         }
         pcResidual += uiStride;
       }
+#endif
     }
   }
 }
@@ -3141 +3198 @@
   }
 }
 
+#if FIX_1320_RDPCM_ACCUMULATOR
+Void TComTrQuant::transformSkipQuantOneSample(TComTU &rTu, const ComponentID compID, const TCoeff resiDiff, TCoeff* pcCoeff, const UInt uiPos, const QpParam &cQP, const Bool bUseHalfRoundingPoint)
+#else
 Void TComTrQuant::transformSkipQuantOneSample(TComTU &rTu, const ComponentID compID, const Pel resiDiff, TCoeff* pcCoeff, const UInt uiPos, const QpParam &cQP, const Bool bUseHalfRoundingPoint)
+#endif
 {
         TComDataCU    *pcCU                           = rTu.getCU();
   const UInt           uiAbsPartIdx                   = rTu.GetAbsPartIdxTU();

source/Lib/TLibCommon/TypeDef.h

@@ -235 +235 @@
 
 #define MAX_NUM_LONG_TERM_REF_PICS                       33
 
+#define FIX_1320_RDPCM_ACCUMULATOR                        1 ///< Fix for accumulator size in RDPCM.
 
 // ====================================================================================================================
 // RExt control settings

source/Lib/TLibCommon/TComTrQuant.h

@@ -194 +194 @@
   Int*    getSliceNSamples(){ return m_sliceNsamples ;}
   Double* getSliceSumC()    { return m_sliceSumC; }
 #endif
+#if FIX_1320_RDPCM_ACCUMULATOR
+  Void transformSkipQuantOneSample(TComTU &rTu, const ComponentID compID, const TCoeff resiDiff, TCoeff* pcCoeff, const UInt uiPos, const QpParam &cQP, const Bool bUseHalfRoundingPoint);
+#else
   Void transformSkipQuantOneSample(TComTU &rTu, const ComponentID compID, const Pel resiDiff, TCoeff* pcCoeff, const UInt uiPos, const QpParam &cQP, const Bool bUseHalfRoundingPoint);
+#endif
   Void invTrSkipDeQuantOneSample(TComTU &rTu, ComponentID compID, TCoeff pcCoeff, Pel &reconSample, const QpParam &cQP, UInt uiPos );
 
 protected: