Merge pull request #818 from sijchen/md_merge133

[Encoder MD] Add Inter Psub16 MD for screen content coding

codec/encoder/core/inc/md.h

@@ -66,6 +66,15 @@ namespace WelsSVCEnc {
 
 #define NO_BEST_FRAC_PIX   1 // REFINE_ME_NO_BEST_HALF_PIXEL + ME_NO_BEST_QUAR_PIXEL
 
+//for vaa constants
+#define MBVAASIGN_FLAT       15
+#define MBVAASIGN_HOR1      3
+#define MBVAASIGN_HOR2      12
+#define MBVAASIGN_VER1       5
+#define MBVAASIGN_VER2       10
+#define MBVAASIGN_CMPX1    6
+#define MBVAASIGN_CMPX2    9
+
 extern const int32_t g_kiQpCostTable[52];
 extern const int8_t g_kiMapModeI16x16[7];
 //extern const int8_t g_kiMapModeI4x4[14];

codec/encoder/core/inc/svc_base_layer_md.h

@@ -66,6 +66,8 @@ int32_t WelsMdP8x8 (SWelsFuncPtrList* pFunc, SDqLayer* pCurDqLayer, SWelsMD* pWe
 /*static*/  void WelsMdInterInit (sWelsEncCtx* pEncCtx, SSlice* pSlice, SMB* pCurMb, const int32_t kiSliceFirstMbXY);
 /*static*/ void WelsMdInterFinePartition (void* pEnc, void* pMd, SSlice* pSlice, SMB* pCurMb, int32_t bestCost);
 /*static*/ void WelsMdInterFinePartitionVaa (void* pEnc, void* pMd, SSlice* pSlice, SMB* pCurMb, int32_t bestCost);
+/*static*/ void WelsMdInterFinePartitionVaaOnScreen (void* pEnc, void* pMd, SSlice* pSlice, SMB* pCurMb,
+    int32_t bestCost);
 void WelsMdInterMbRefinement (sWelsEncCtx* pEncCtx, SWelsMD* pWelsMd, SMB* pCurMb, SMbCache* pMbCache);
 bool WelsMdFirstIntraMode (void* pEnc, void* pMd, SMB* pCurMb, SMbCache* pMbCache);
 //bool svc_md_first_intra_mode_constrained(void* pEnc, void* pMd, SMB* pCurMb, SMbCache *pMbCache);

codec/encoder/core/src/encoder_ext.cpp

@@ -2470,6 +2470,9 @@ void PreprocessSliceCoding (sWelsEncCtx* pCtx) {
       pFeatureSearchPreparation->iHighFreMbCount = 0;
 
       if (P_SLICE == pCtx->eSliceType) {
+        //MD related func pointers
+        pFuncList->pfInterFineMd = WelsMdInterFinePartitionVaaOnScreen;
+
         //calculate bFMESwitchFlag
         SVAAFrameInfoExt* pVaaExt		= static_cast<SVAAFrameInfoExt*> (pCtx->pVaa);
         const int32_t kiMbSize = pCurLayer->iMbHeight * pCurLayer->iMbWidth;

codec/encoder/core/src/svc_base_layer_md.cpp

@@ -1868,7 +1868,86 @@ void WelsMdIntraSecondaryModesEnc (sWelsEncCtx* pEncCtx, SWelsMD* pWelsMd, SMB*
   pCurMb->pSadCost[0] = 0;
 }
 
+//
+//func pointer of inter MD for sub16x16 INTER MD for screen content coding
+//
+static inline void MergeSub16Me (const SWelsME& sSrcMe0, const SWelsME& sSrcMe1, SWelsME* pTarMe) {
+  memcpy (pTarMe, &sSrcMe0, sizeof (sSrcMe0)); // confirmed_safe_unsafe_usage
 
+  pTarMe->uiSadCost = sSrcMe0.uiSadCost + sSrcMe1.uiSadCost;//not precise cost since MVD cost is not the same
+  pTarMe->uiSatdCost = sSrcMe0.uiSatdCost + sSrcMe1.uiSatdCost;//not precise cost since MVD cost is not the same
+}
+static inline bool IsSameMv (const SMVUnitXY& sMv0, const SMVUnitXY& sMv1) {
+  return ((sMv0.iMvX == sMv1.iMvX) && (sMv0.iMvY == sMv1.iMvY));
+}
+static inline int32_t Mvd (const SMVUnitXY& sMv, const SMVUnitXY& sMvp) {
+  return (WELS_ABS (sMv.iMvX - sMvp.iMvX) + WELS_ABS (sMv.iMvY - sMvp.iMvY));
+}
+bool TryModeMerge (SMbCache* pMbCache, SWelsMD* pWelsMd, SMB* pCurMb) {
+  SWelsME* pMe8x8 = & (pWelsMd->sMe.sMe8x8[0]);
+  const bool bSameMv16x8_0 = IsSameMv (pMe8x8[0].sMv, pMe8x8[1].sMv);
+  const bool bSameMv16x8_1 = IsSameMv (pMe8x8[2].sMv, pMe8x8[3].sMv);
+
+  const bool bSameMv8x16_0 = IsSameMv (pMe8x8[0].sMv, pMe8x8[2].sMv);
+  const bool bSameMv8x16_1 = IsSameMv (pMe8x8[1].sMv, pMe8x8[3].sMv);
+  //need to consider iRefIdx when multi ref is available
+  const bool bSameRefIdx16x8_0 = true; //pMe8x8[0].iRefIdx == pMe8x8[1].iRefIdx;
+  const bool bSameRefIdx16x8_1 = true; //pMe8x8[2].iRefIdx == pMe8x8[3].iRefIdx;
+  const bool bSameRefIdx8x16_0 = true; //pMe8x8[0].iRefIdx == pMe8x8[2].iRefIdx;
+  const bool bSameRefIdx8x16_1 = true; //pMe8x8[1].iRefIdx == pMe8x8[3].iRefIdx;
+  const int32_t iSameMv = (bSameMv16x8_0 << 7) | (bSameRefIdx16x8_0 << 6) | (bSameMv16x8_1 << 5) |
+                          (bSameRefIdx16x8_1 << 4) |
+                          (bSameMv8x16_0 << 3) | (bSameRefIdx8x16_0 << 2) | (bSameMv8x16_1 << 1) | (bSameRefIdx8x16_1);
+
+  switch (iSameMv) {
+  case 0xF0:
+    pCurMb->uiMbType = MB_TYPE_16x8;
+    MergeSub16Me (pMe8x8[0], pMe8x8[1], & (pWelsMd->sMe.sMe16x8[0]));
+    MergeSub16Me (pMe8x8[2], pMe8x8[3], & (pWelsMd->sMe.sMe16x8[1]));
+    PredInter16x8Mv (pMbCache, 0, 0, & (pWelsMd->sMe.sMe16x8[0].sMvp));
+    PredInter16x8Mv (pMbCache, 8, 0, & (pWelsMd->sMe.sMe16x8[1].sMvp));
+    break;
+  case 0x0F:
+    pCurMb->uiMbType = MB_TYPE_8x16;
+    MergeSub16Me (pMe8x8[0], pMe8x8[2], & (pWelsMd->sMe.sMe8x16[0]));
+    MergeSub16Me (pMe8x8[1], pMe8x8[3], & (pWelsMd->sMe.sMe8x16[1]));
+    PredInter8x16Mv (pMbCache, 0, 0, & (pWelsMd->sMe.sMe8x16[0].sMvp));
+    PredInter8x16Mv (pMbCache, 4, 0, & (pWelsMd->sMe.sMe8x16[1].sMvp));
+    break;
+  case 0xFF:
+  //MERGE_16x16
+  //from test results of multiple sequences show that using the following 0x0F to merge 16x16
+  //for some seq there is BR saving some loss
+  //on the whole the BR will increase little bit
+  //to save complexity we decided not to merge 16x16 at present (10/12/2012)
+  default:
+    break;
+  }
+  return (MB_TYPE_8x8 != pCurMb->uiMbType);
+}
+
+
+void WelsMdInterFinePartitionVaaOnScreen (void* pEnc, void* pMd, SSlice* pSlice, SMB* pCurMb, int32_t iBestCost) {
+  sWelsEncCtx* pEncCtx = (sWelsEncCtx*)pEnc;
+  SWelsMD* pWelsMd = (SWelsMD*)pMd;
+  SMbCache* pMbCache = &pSlice->sMbCacheInfo;
+  SDqLayer* pCurDqLayer = pEncCtx->pCurDqLayer;
+  int32_t iCostP8x8;
+  uint8_t uiMbSign = pEncCtx->pFuncList->pfGetMbSignFromInterVaa (&pEncCtx->pVaa->sVaaCalcInfo.pSad8x8[pCurMb->iMbXY][0]);
+
+  if (MBVAASIGN_FLAT == uiMbSign) {
+    return;
+  }
+
+  iCostP8x8 = WelsMdP8x8 (pEncCtx->pFuncList, pCurDqLayer, pWelsMd, pSlice);
+  if (iCostP8x8 < iBestCost) {
+    iBestCost = iCostP8x8;
+    pCurMb->uiMbType = MB_TYPE_8x8;
+
+    TryModeMerge (pMbCache, pWelsMd, pCurMb);
+  }
+  pWelsMd->iCostLuma = iBestCost;
+}
 
 
 } // namespace WelsSVCEnc