JVET-P0530: align chroma weights with luma weights for TPM blending

source/Lib/CommonLib/InterpolationFilter.cpp

@@ -697,33 +697,69 @@ void InterpolationFilter::xWeightedTriangleBlk( const PredictionUnit &pu, const
   const int32_t offsetDefault     = (1<<(shiftDefault-1)) + IF_INTERNAL_OFFS;
   const int32_t shiftWeighted     = std::max<int>(2, (IF_INTERNAL_PREC - clipbd)) + log2WeightBase;
   const int32_t offsetWeighted    = (1 << (shiftWeighted - 1)) + (IF_INTERNAL_OFFS << log2WeightBase);
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+  int32_t logSubWidthC = getChannelTypeScaleX(CHANNEL_TYPE_CHROMA, pu.chromaFormat);
+  int32_t logSubHeightC = getChannelTypeScaleY(CHANNEL_TYPE_CHROMA, pu.chromaFormat);
 
+  int32_t stepX = (compIdx == 0) ? 1 : (1 << logSubWidthC);
+  int32_t stepY = (compIdx == 0) ? 1 : (1 << logSubHeightC);
+
+  int32_t widthY = (compIdx == 0) ? width : (width << logSubWidthC);
+  int32_t heightY = (compIdx == 0) ? height : (height << logSubHeightC);
+
+  int32_t ratioWH = (widthY > heightY) ? (widthY / heightY) : 1;
+  int32_t ratioHW = (widthY > heightY) ? 1 : (heightY / widthY);
+
+  int32_t weightedLength = 7;
+  int32_t weightedStartPos = (splitDir == 0) ? (0 - (weightedLength >> 1) * ratioWH) : (widthY - ((weightedLength + 1) >> 1) * ratioWH);
+#else
   const int32_t ratioWH           = (width > height) ? (width / height) : 1;
   const int32_t ratioHW           = (width > height) ? 1 : (height / width);
 
   const bool    longWeight        = (compIdx == COMPONENT_Y);
   const int32_t weightedLength    = longWeight ? 7 : 3;
         int32_t weightedStartPos  = ( splitDir == 0 ) ? ( 0 - (weightedLength >> 1) * ratioWH ) : ( width - ((weightedLength + 1) >> 1) * ratioWH );
+#endif       
         int32_t weightedEndPos    = weightedStartPos + weightedLength * ratioWH - 1;
         int32_t weightedPosoffset = ( splitDir == 0 ) ? ratioWH : -ratioWH;
 
         Pel     tmpPelWeighted;
         int32_t weightIdx;
         int32_t x, y, tmpX, tmpY, tmpWeightedStart, tmpWeightedEnd;
-
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+  for (y = 0; y < heightY; y += ratioHW)
+  {
+    if (y % stepY != 0)
+    {
+      weightedStartPos += weightedPosoffset;
+      weightedEndPos += weightedPosoffset;
+      continue;
+    }
+    for (tmpY = ratioHW; tmpY > 0; tmpY -= stepY)
+    {
+      for (x = 0; x < weightedStartPos; x++)
+      {
+        if (x % stepX != 0)
+          continue;
+#else
   for( y = 0; y < height; y+= ratioHW )
   {
     for( tmpY = ratioHW; tmpY > 0; tmpY-- )
     {
       for( x = 0; x < weightedStartPos; x++ )
       {
+#endif
         *dst++ = ClipPel( rightShift( (splitDir == 0 ? *src1 : *src0) + offsetDefault, shiftDefault), clipRng );
         src0++;
         src1++;
       }
 
       tmpWeightedStart = std::max((int32_t)0, weightedStartPos);
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+      tmpWeightedEnd = std::min(weightedEndPos, (int32_t)(widthY - 1));
+#else
       tmpWeightedEnd   = std::min(weightedEndPos, (int32_t)(width - 1));
+#endif
       weightIdx        = 1;
       if( weightedStartPos < 0 )
       {
@@ -731,17 +767,36 @@ void InterpolationFilter::xWeightedTriangleBlk( const PredictionUnit &pu, const
       }
       for( x = tmpWeightedStart; x <= tmpWeightedEnd; x+= ratioWH )
       {
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+        if (x % stepX != 0)
+        {
+          weightIdx++;
+          continue;
+        }
+
+        for (tmpX = ratioWH; tmpX > 0; tmpX -= stepX)
+        {
+          tmpPelWeighted = Clip3(1, 7, weightIdx);
+#else
         for( tmpX = ratioWH; tmpX > 0; tmpX-- )
         {
           tmpPelWeighted = Clip3( 1, 7, longWeight ? weightIdx : (weightIdx * 2));
+#endif
           tmpPelWeighted = splitDir ? ( 8 - tmpPelWeighted ) : tmpPelWeighted;
           *dst++         = ClipPel( rightShift( (tmpPelWeighted*(*src0++) + ((8 - tmpPelWeighted) * (*src1++)) + offsetWeighted), shiftWeighted ), clipRng );
         }
         weightIdx ++;
       }
 
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+      for (x = weightedEndPos + 1; x < widthY; x++)
+      {
+        if (x % stepX != 0)
+          continue;
+#else
       for( x = weightedEndPos + 1; x < width; x++ )
       {
+#endif
         *dst++ = ClipPel( rightShift( (splitDir == 0 ? *src0 : *src1) + offsetDefault, shiftDefault ), clipRng );
         src0++;
         src1++;

source/Lib/CommonLib/Rom.cpp

@@ -434,15 +434,32 @@ void initROM()
       g_triangleWeights[0][1][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
       g_triangleWeights[1][0][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
       g_triangleWeights[1][1][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
-
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+      g_triangleWeights[2][0][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
+      g_triangleWeights[2][1][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
+      g_triangleWeights[3][0][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
+      g_triangleWeights[3][1][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
+      int nCbR_422 = (nCbW * 2 > nCbH) ? (nCbW * 2) / nCbH : nCbH / (nCbW * 2);
+#endif
       for (int y = 0; y < nCbH; y++)
       {
         for (int x = 0; x < nCbW; x++)
         {
           g_triangleWeights[0][0][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, x / nCbR - y + 4) : Clip3(0, 8, x - y / nCbR + 4);
           g_triangleWeights[0][1][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, nCbH - 1 - x / nCbR - y + 4) : Clip3(0, 8, nCbW - 1 - x - y / nCbR + 4);
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+          g_triangleWeights[CHROMA_420][0][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, (x * 2) / nCbR - (y * 2) + 4) : Clip3(0, 8, (x * 2) - (y * 2) / nCbR + 4);
+          g_triangleWeights[CHROMA_420][1][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, nCbH * 2 - 1 - (x * 2) / nCbR - (y * 2) + 4) : Clip3(0, 8, nCbW * 2 - 1 - (x * 2) - (y * 2) / nCbR + 4);
+
+          g_triangleWeights[CHROMA_422][0][idxH][idxW][y*nCbW + x] = (nCbW * 2 > nCbH) ? Clip3(0, 8, (x * 2) / nCbR_422 - y + 4) : Clip3(0, 8, (x * 2) - y / nCbR_422 + 4);
+          g_triangleWeights[CHROMA_422][1][idxH][idxW][y*nCbW + x] = (nCbW * 2 > nCbH) ? Clip3(0, 8, nCbH - 1 - (x * 2) / nCbR_422 - y + 4) : Clip3(0, 8, nCbW * 2 - 1 - (x * 2) - y / nCbR_422 + 4);
+
+          g_triangleWeights[CHROMA_444][0][idxH][idxW][y*nCbW + x] = g_triangleWeights[0][0][idxH][idxW][y*nCbW + x];
+          g_triangleWeights[CHROMA_444][1][idxH][idxW][y*nCbW + x] = g_triangleWeights[0][1][idxH][idxW][y*nCbW + x];
+#else
           g_triangleWeights[1][0][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 4, x / nCbR - y + 2) * 2 : Clip3(0, 4, x - y / nCbR + 2) * 2;
           g_triangleWeights[1][1][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 4, nCbH - 1 - x / nCbR - y + 2) * 2 : Clip3(0, 4, nCbW - 1 - x - y / nCbR + 2) * 2;
+#endif
         }
       }
     }
@@ -486,6 +503,16 @@ void destroyROM()
       g_triangleWeights[0][1][idxH][idxW] = nullptr;
       g_triangleWeights[1][0][idxH][idxW] = nullptr;
       g_triangleWeights[1][1][idxH][idxW] = nullptr;
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+      delete[] g_triangleWeights[2][0][idxH][idxW];
+      delete[] g_triangleWeights[2][1][idxH][idxW];
+      delete[] g_triangleWeights[3][0][idxH][idxW];
+      delete[] g_triangleWeights[3][1][idxH][idxW];
+      g_triangleWeights[2][0][idxH][idxW] = nullptr;
+      g_triangleWeights[2][1][idxH][idxW] = nullptr;
+      g_triangleWeights[3][0][idxH][idxW] = nullptr;
+      g_triangleWeights[3][1][idxH][idxW] = nullptr;
+#endif
     }
   }
 }
@@ -728,7 +755,11 @@ const uint32_t g_scalingListSizeX[SCALING_LIST_SIZE_NUM] = { 1, 2,  4,  8,  16,
 
 
 uint8_t g_triangleMvStorage[TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 1][MAX_CU_DEPTH - MIN_CU_LOG2 + 1][MAX_CU_SIZE >> MIN_CU_LOG2][MAX_CU_SIZE >> MIN_CU_LOG2];
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+int16_t *g_triangleWeights[4][TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 2][MAX_CU_DEPTH - MIN_CU_LOG2 + 2];
+#else
 int16_t *g_triangleWeights[2][TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 2][MAX_CU_DEPTH - MIN_CU_LOG2 + 2];
+#endif
 Mv   g_reusedUniMVs[32][32][8][8][2][33];
 bool g_isReusedUniMVsFilled[32][32][8][8];
 

source/Lib/CommonLib/Rom.h

@@ -211,8 +211,11 @@ constexpr uint8_t g_tbMax[257] = { 0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3,
 
 extern       uint8_t g_triangleMvStorage[TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 1][MAX_CU_DEPTH - MIN_CU_LOG2 + 1][MAX_CU_SIZE >> MIN_CU_LOG2][MAX_CU_SIZE >> MIN_CU_LOG2];
 // 7-tap/3-tap, direction, 2/4/8/16/32/64/128
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+extern int16_t *g_triangleWeights[4][TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 2][MAX_CU_DEPTH - MIN_CU_LOG2 + 2];
+#else
 extern int16_t *g_triangleWeights[2][TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 2][MAX_CU_DEPTH - MIN_CU_LOG2 + 2];
-
+#endif
 extern bool g_mctsDecCheckEnabled;
 
 class  Mv;

source/Lib/CommonLib/TypeDef.h

@@ -50,6 +50,8 @@
 #include <assert.h>
 #include <cassert>
 
+#define JVET_P0530_TPM_WEIGHT_ALIGN                       1 // JVET-P0530: align chroma weights with luma weights for TPM blending
+
 #define JVET_P0491_BDOFPROF_MVD_RANGE                     1 // JVET-P0491: clip the MVD in BDOF/PROF to [-31 31]
 
 #define JVET_P0460_PLT_TS_MIN_QP                          1 // JVET-P0460: Use TS min QP for Palette Escape mode

source/Lib/CommonLib/x86/InterpolationFilterX86.h

@@ -1233,10 +1233,15 @@ void xWeightedTriangleBlk_SSE(const PredictionUnit &pu, const uint32_t width, co
   const ClpRng  clpRng = pu.cu->slice->clpRngs().comp[compIdx];
   const int32_t shiftWeighted = std::max<int>(2, (IF_INTERNAL_PREC - clpRng.bd)) + log2WeightBase;
   const int32_t offsetWeighted = (1 << (shiftWeighted - 1)) + (IF_INTERNAL_OFFS << log2WeightBase);
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+  int wIdx = (compIdx == COMPONENT_Y) ? 0 : pu.cs->sps->getChromaFormatIdc();
+  int16_t *weight = g_triangleWeights[wIdx][splitDir][log2Height][log2Width];
+#else
   const bool    longWeight = (compIdx == COMPONENT_Y);
   const bool    shortWeight = !longWeight;
 
   int16_t *weight = g_triangleWeights[shortWeight][splitDir][log2Height][log2Width];
+#endif
 
   const __m128i mmEight = _mm_set1_epi16(8);
   const __m128i mmOffset = _mm_set1_epi32(offsetWeighted);