shithub: openh264

Info  •  Files  •  Log  •  Branches

Merge pull request #557 from sijchen/me_refactor4444

[Encoder Refactor] Expand MVD table

--- a/codec/encoder/core/inc/encoder_context.h

+++ b/codec/encoder/core/inc/encoder_context.h

@@ -116,6 +116,7 @@

   int32_t*					pSadCostMb;

   /* MVD cost tables for Inter MB */

+  int32_t              iMvRange;

   uint16_t*					pMvdCostTableInter; //[52];	// adaptive to spatial layers

   SMVUnitXY*

   pMvUnitBlock4x4;	// (*pMvUnitBlock4x4[2])[MB_BLOCK4x4_NUM];	    // for store each 4x4 blocks' mv unit, the two swap after different d layer

--- a/codec/encoder/core/inc/slice.h

+++ b/codec/encoder/core/inc/slice.h

@@ -163,8 +163,8 @@

   SSliceHeaderExt	sSliceHeaderExt;

-  SMVUnitXY	sMvMin;

-  SMVUnitXY	sMvMax;

+  SMVUnitXY	sMvStartMin;

+  SMVUnitXY	sMvStartMax;

   SMVUnitXY	sMvc[5];

   uint8_t		uiMvcNum;

   uint8_t		sScaleShift;

--- a/codec/encoder/core/inc/svc_motion_estimate.h

+++ b/codec/encoder/core/inc/svc_motion_estimate.h

@@ -45,9 +45,14 @@

 #include "wels_func_ptr_def.h"

 namespace WelsSVCEnc {

-#define MV_RANGE (64)

+#define CAMERA_STARTMV_RANGE (64)

 #define	ITERATIVE_TIMES	(16)

-#define	BASE_MV_MB_NMB	((2*(MV_RANGE+ITERATIVE_TIMES)/MB_WIDTH_LUMA)-1)

+#define CAMERA_MV_RANGE (CAMERA_STARTMV_RANGE+ITERATIVE_TIMES)

+#define CAMERA_MVD_RANGE  ((CAMERA_MV_RANGE+1)<<1) //mvd=mv_range*2;

+#define	BASE_MV_MB_NMB	((2*CAMERA_MV_RANGE/MB_WIDTH_LUMA)-1)

+#define CAMERA_HIGHLAYER_MVD_RANGE (243)//mvd range;

+#define EXPANDED_MV_RANGE (504) //=512-8 rather than 511 to sacrifice same edge point but save complexity in assemblys

+#define EXPANDED_MVD_RANGE ((504+1)<<1)

 union SadPredISatdUnit {

 uint32_t	uiSadPred;

--- a/codec/encoder/core/src/encoder_ext.cpp

+++ b/codec/encoder/core/src/encoder_ext.cpp

@@ -1201,8 +1201,10 @@

   int32_t iResult					= 0;

   float	fCompressRatioThr		= .5f;

   const int32_t kiNumDependencyLayers	= pParam->iSpatialLayerNum;

-  const uint32_t kuiMvdInterTableSize	= (kiNumDependencyLayers == 1 ? (1 + (648 << 1)) : (1 + (972 << 1)));

-  const uint32_t kuiMvdCacheAlginedSize	= kuiMvdInterTableSize * sizeof (uint16_t);

+  (*ppCtx)->iMvRange = pParam->iUsageType?EXPANDED_MV_RANGE:CAMERA_STARTMV_RANGE;

+  const int32_t kiMvdRange = (pParam->iUsageType?EXPANDED_MVD_RANGE:((kiNumDependencyLayers == 1)?CAMERA_MVD_RANGE:CAMERA_HIGHLAYER_MVD_RANGE));

+  const uint32_t kuiMvdInterTableSize	=  1 + (kiMvdRange << 3);//intepel*4=qpel;  qpel_mv_range*2=(+/-);

+  const uint32_t kuiMvdCacheAlignedSize	= kuiMvdInterTableSize * sizeof (uint16_t);

   int32_t iVclLayersBsSizeCount		= 0;

   int32_t iNonVclLayersBsSizeCount	= 0;

 #if defined(MT_ENABLED)

@@ -1401,7 +1403,7 @@

     return 1;

   }

-  (*ppCtx)->pMvdCostTableInter = (uint16_t*)pMa->WelsMallocz (52 * kuiMvdCacheAlginedSize, "pMvdCostTableInter");

+  (*ppCtx)->pMvdCostTableInter = (uint16_t*)pMa->WelsMallocz (52 * kuiMvdCacheAlignedSize, "pMvdCostTableInter");

   WELS_VERIFY_RETURN_PROC_IF (1, (NULL == (*ppCtx)->pMvdCostTableInter), FreeMemorySvc (ppCtx))

   MvdCostInit ((*ppCtx)->pMvdCostTableInter, kuiMvdInterTableSize);  //should put to a better place?

--- a/codec/encoder/core/src/svc_base_layer_md.cpp

+++ b/codec/encoder/core/src/svc_base_layer_md.cpp

@@ -359,7 +359,7 @@

   ST32 (&pCurMb->sP16x16Mv, 0);

   ST32 (&pCurLayer->pDecPic->sMvList[kiMbXY], 0);

-  SetMvWithinMvRange( kiMbWidth, kiMbHeight, kiMbX, kiMbY, MV_RANGE, &(pSlice->sMvMin), &(pSlice->sMvMax));

+  SetMvWithinMvRange( kiMbWidth, kiMbHeight, kiMbX, kiMbY, CAMERA_STARTMV_RANGE, &(pSlice->sMvStartMin), &(pSlice->sMvStartMax));

 }

 int32_t WelsMdI16x16 (SWelsFuncPtrList* pFunc, SDqLayer* pCurDqLayer, SMbCache* pMbCache, int32_t iLambda) {

--- a/codec/encoder/core/src/svc_motion_estimate.cpp

+++ b/codec/encoder/core/src/svc_motion_estimate.cpp

@@ -106,15 +106,15 @@

   uint32_t i;

   const uint32_t kuiMvcNum		= pSlice->uiMvcNum;

   const SMVUnitXY* kpMvcList	= &pSlice->sMvc[0];

-  const SMVUnitXY ksMvMin		= pSlice->sMvMin;

-  const SMVUnitXY ksMvMax		= pSlice->sMvMax;

+  const SMVUnitXY ksMvStartMin		= pSlice->sMvStartMin;

+  const SMVUnitXY ksMvStartMax		= pSlice->sMvStartMax;

   const SMVUnitXY ksMvp		= pMe->sMvp;

   SMVUnitXY sMv;

   //  Step 1: Initial point prediction

   // init with sMvp

-  sMv.iMvX	= WELS_CLIP3 ((2 + ksMvp.iMvX) >> 2, ksMvMin.iMvX, ksMvMax.iMvX);

-  sMv.iMvY	= WELS_CLIP3 ((2 + ksMvp.iMvY) >> 2, ksMvMin.iMvY, ksMvMax.iMvY);

+  sMv.iMvX	= WELS_CLIP3 ((2 + ksMvp.iMvX) >> 2, ksMvStartMin.iMvX, ksMvStartMax.iMvX);

+  sMv.iMvY	= WELS_CLIP3 ((2 + ksMvp.iMvY) >> 2, ksMvStartMin.iMvY, ksMvStartMax.iMvY);

   pRefMb = &pMe->pRefMb[sMv.iMvY * iStrideRef + sMv.iMvX];

@@ -123,8 +123,8 @@

   for (i = 0; i < kuiMvcNum; i++) {

     //clipping here is essential since some pOut-of-range MVC may happen here (i.e., refer to baseMV)

-    iMvc0 = WELS_CLIP3 ((2 + kpMvcList[i].iMvX) >> 2, ksMvMin.iMvX, ksMvMax.iMvX);

-    iMvc1 = WELS_CLIP3 ((2 + kpMvcList[i].iMvY) >> 2, ksMvMin.iMvY, ksMvMax.iMvY);

+    iMvc0 = WELS_CLIP3 ((2 + kpMvcList[i].iMvX) >> 2, ksMvStartMin.iMvX, ksMvStartMax.iMvX);

+    iMvc1 = WELS_CLIP3 ((2 + kpMvcList[i].iMvY) >> 2, ksMvStartMin.iMvY, ksMvStartMax.iMvY);

     if (((iMvc0 - sMv.iMvX) || (iMvc1 - sMv.iMvY))) {

       pFref2 = &pMe->pRefMb[iMvc1 * iStrideRef + iMvc0];