shithub: openh264

Download patch

ref: 3022d4f6559596e0a6a1fae4d638f3c02f7d2a9e
parent: b304687197cf59ed24ee6fc4756f7f91b72f613f
parent: 63926a3d10ca35db6ecaa35a6271abc43acbe06f
author: huili2 <[email protected]>
date: Thu Mar 5 04:39:31 EST 2015 

Merge pull request #1832 from sijchen/imp_mc

[Encoder] replace conditional judgements with faster operations


--- a/codec/encoder/core/inc/md.h
+++ b/codec/encoder/core/inc/md.h
@@ -122,6 +122,7 @@
 uint8_t* pQuarPixBest;
 uint8_t* pQuarPixTmp;
 
+PCopyFunc pfCopyBlockByMode;
 } SMeRefinePointer;
 
 void FillNeighborCacheIntra (SMbCache* pMbCache, SMB* pCurMb, int32_t iMbWidth/*, bool constrained_intra_pred_flag*/);
--- a/codec/encoder/core/inc/svc_enc_frame.h
+++ b/codec/encoder/core/inc/svc_enc_frame.h
@@ -93,6 +93,8 @@
 int16_t					iMbHeight;		// MB height of this picture, equal to pSps.iMbHeight;
 
 bool					bBaseLayerAvailableFlag;	// whether base layer is available for prediction?
+bool					bSatdInMdFlag; // whether SATD is calculated in ME and integer-pel MD
+
 uint8_t					iLoopFilterDisableIdc;	// 0: on, 1: off, 2: on except for slice boundaries
 int8_t					iLoopFilterAlphaC0Offset;// AlphaOffset: valid range [-6, 6], default 0
 int8_t					iLoopFilterBetaOffset;	// BetaOffset:	valid range [-6, 6], default 0
--- a/codec/encoder/core/src/encoder_ext.cpp
+++ b/codec/encoder/core/src/encoder_ext.cpp
@@ -3046,6 +3046,10 @@
       pCurLayer->pFeatureSearchPreparation->uiFMEGoodFrameCount = FMESWITCH_DEFAULT_GOODFRAME_NUM;
     }
   }
+
+  // update some layer dependent variable to save judgements in mb-level
+  pCurLayer->bSatdInMdFlag = ((pFuncList->sSampleDealingFuncs.pfMeCost == pFuncList->sSampleDealingFuncs.pfSampleSatd)
+                              && (pFuncList->sSampleDealingFuncs.pfMdCost == pFuncList->sSampleDealingFuncs.pfSampleSatd));
 }
 
 /*!
--- a/codec/encoder/core/src/md.cpp
+++ b/codec/encoder/core/src/md.cpp
@@ -598,8 +598,7 @@
   int32_t iCurCost;
   int32_t iBestHalfPix;
 
-  if ((pFunc->sSampleDealingFuncs.pfMeCost == pFunc->sSampleDealingFuncs.pfSampleSatd)
-      && (pFunc->sSampleDealingFuncs.pfMdCost == pFunc->sSampleDealingFuncs.pfSampleSatd)) {
+  if (pEncCtx->pCurDqLayer->bSatdInMdFlag) {
     iBestCost = pMe->uSadPredISatd.uiSatd + COST_MVD (pMe->pMvdCost, iMvx - pMe->sMvp.iMvX, iMvy - pMe->sMvp.iMvY);
   } else {
     iBestCost = pFunc->sSampleDealingFuncs.pfMeCost[pMe->uiBlockSize] (pEncData, kiStrideEnc, pRef, kiStrideRef) +
@@ -765,17 +764,8 @@
     pBestPredInter = pRef;
     iInterBlk4Stride = kiStrideRef;
   }
-  if (MB_WIDTH_LUMA == iWidth && MB_HEIGHT_LUMA == iHeight) { //P16x16
-    pFunc->pfCopy16x16NotAligned (pMemPredInterMb, MB_WIDTH_LUMA, pBestPredInter,
-                                  iInterBlk4Stride);	// dst can be align with 16 bytes, but not sure at pSrc, 12/29/2011
-  } else if (MB_WIDTH_LUMA == iWidth && MB_HEIGHT_CHROMA == iHeight) { //P16x8
-    pFunc->pfCopy16x8NotAligned (pMemPredInterMb, MB_WIDTH_LUMA, pBestPredInter,
-                                 iInterBlk4Stride);	// dst can be align with 16 bytes, but not sure at pSrc, 12/29/2011
-  } else if (MB_WIDTH_CHROMA == iWidth && MB_HEIGHT_LUMA == iHeight) { //P8x16
-    pFunc->pfCopy8x16Aligned (pMemPredInterMb, MB_WIDTH_LUMA, pBestPredInter, iInterBlk4Stride);
-  } else { //P8x8
-    pFunc->pfCopy8x8Aligned (pMemPredInterMb, MB_WIDTH_LUMA, pBestPredInter, iInterBlk4Stride);
-  }
+  pMeRefine->pfCopyBlockByMode (pMemPredInterMb, MB_WIDTH_LUMA, pBestPredInter,
+                                iInterBlk4Stride);
 }
 
 void InitBlkStrideWithRef (int32_t* pBlkStride, const int32_t kiStrideRef) {
--- a/codec/encoder/core/src/svc_base_layer_md.cpp
+++ b/codec/encoder/core/src/svc_base_layer_md.cpp
@@ -1148,7 +1148,8 @@
   }
 }
 
-void WelsMdInterFinePartitionVaa (sWelsEncCtx* pEncCtx, SWelsMD* pWelsMd, SSlice* pSlice, SMB* pCurMb, int32_t iBestCost) {
+void WelsMdInterFinePartitionVaa (sWelsEncCtx* pEncCtx, SWelsMD* pWelsMd, SSlice* pSlice, SMB* pCurMb,
+                                  int32_t iBestCost) {
   SDqLayer* pCurDqLayer = pEncCtx->pCurDqLayer;
 //	SMbCache *pMbCache = &pSlice->sMbCacheInfo;
   int32_t iCostP8x16, iCostP16x8, iCostP8x8;
@@ -1428,6 +1429,7 @@
 
 void WelsMdInterMbRefinement (sWelsEncCtx* pEncCtx, SWelsMD* pWelsMd, SMB* pCurMb, SMbCache* pMbCache) {
   SDqLayer* pCurDqLayer = pEncCtx->pCurDqLayer;
+  SWelsFuncPtrList* pFunc = pEncCtx->pFuncList;
   uint8_t* pTmpRefCb, *pTmpRefCr, *pTmpDstCb, *pTmpDstCr;
   int32_t iMvStride, iRefBlk4Stride, iDstBlk4Stride;
   SMVUnitXY* pMv;
@@ -1448,6 +1450,8 @@
   case MB_TYPE_16x16:
     //luma
     InitMeRefinePointer (&sMeRefine, pMbCache, 0);
+    sMeRefine.pfCopyBlockByMode =
+      pFunc->pfCopy16x16NotAligned; // dst can be align with 16 bytes, but not sure at pSrc, 12/29/2011
     MeRefineFracPixel (pEncCtx, pDstLuma, &pWelsMd->sMe.sMe16x16, &sMeRefine, 16, 16);
     UpdateP16x16MotionInfo (pMbCache, pCurMb, pWelsMd->uiRef, &pWelsMd->sMe.sMe16x16.sMv);
 
@@ -1474,6 +1478,8 @@
 
   case MB_TYPE_16x8:
     iPixStride = 0;
+    sMeRefine.pfCopyBlockByMode =
+      pFunc->pfCopy16x8NotAligned; // dst can be align with 16 bytes, but not sure at pSrc, 12/29/2011
     for (i = 0; i < 2; i++) {
       //luma
       iIdx = i << 3;
@@ -1503,6 +1509,7 @@
 
   case MB_TYPE_8x16:
     iPixStride = 0;
+    sMeRefine.pfCopyBlockByMode = pFunc->pfCopy8x16Aligned;
     for (i = 0; i < 2; i++) {
       //luma
       iIdx = i << 2;
@@ -1528,8 +1535,8 @@
       pEncCtx->pFuncList->sMcFuncs.pMcChromaFunc (pTmpRefCr, iLineSizeRefUV, pTmpDstCr, 8, pMv->iMvX, pMv->iMvY, 4, 8); //Cr
     }
     break;
-
   case MB_TYPE_8x8:
+    sMeRefine.pfCopyBlockByMode = pFunc->pfCopy8x8Aligned;
     for (i = 0; i < 4; i++) {
       int32_t iBlk8Idx = i << 2; //0, 4, 8, 12
       int32_t	iBlk4X, iBlk4Y;
@@ -1558,8 +1565,10 @@
       pTmpDstCb = pDstCb + iDstBlk4Stride;
       pTmpRefCr = pRefCr + iRefBlk4Stride;
       pTmpDstCr = pDstCr + iDstBlk4Stride;
-      pEncCtx->pFuncList->sMcFuncs.pMcChromaFunc (pTmpRefCb + iMvStride, iLineSizeRefUV, pTmpDstCb, 8, pMv->iMvX, pMv->iMvY, 4, 4); //Cb
-      pEncCtx->pFuncList->sMcFuncs.pMcChromaFunc (pTmpRefCr + iMvStride, iLineSizeRefUV, pTmpDstCr, 8, pMv->iMvX, pMv->iMvY, 4, 4); //Cr
+      pEncCtx->pFuncList->sMcFuncs.pMcChromaFunc (pTmpRefCb + iMvStride, iLineSizeRefUV, pTmpDstCb, 8, pMv->iMvX, pMv->iMvY,
+          4, 4); //Cb
+      pEncCtx->pFuncList->sMcFuncs.pMcChromaFunc (pTmpRefCr + iMvStride, iLineSizeRefUV, pTmpDstCr, 8, pMv->iMvX, pMv->iMvY,
+          4, 4); //Cr
 
     }
     break;