ref: 6ebacd0cbfe3796e7ce80281c24aaf15d2d6ec48
dir: /codec/encoder/core/src/svc_motion_estimate.cpp/
/*!
* \copy
* Copyright (c) 2009-2013, Cisco Systems
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
*
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
*
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in
* the documentation and/or other materials provided with the
* distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
* FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
* COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
* ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*
*
* \file svc motion estimate.c
*
* \brief Interfaces introduced in svc mb motion estimation
*
* \date 08/11/2009 Created
*
*************************************************************************************
*/
#include "cpu_core.h"
#include "ls_defines.h"
#include "svc_motion_estimate.h"
#include "wels_transpose_matrix.h"
namespace WelsEnc {
const int32_t QStepx16ByQp[52] = { /* save QStep<<4 for int32_t */
10, 11, 13, 14, 16, 18, /* 0~5 */
20, 22, 26, 28, 32, 36, /* 6~11 */
40, 44, 52, 56, 64, 72, /* 12~17 */
80, 88, 104, 112, 128, 144, /* 18~23 */
160, 176, 208, 224, 256, 288, /* 24~29 */
320, 352, 416, 448, 512, 576, /* 30~35 */
640, 704, 832, 896, 1024, 1152, /* 36~41 */
1280, 1408, 1664, 1792, 2048, 2304, /* 42~47 */
2560, 2816, 3328, 3584 /* 48~51 */
};
static inline void UpdateMeResults (const SMVUnitXY ksBestMv, const uint32_t kiBestSadCost, uint8_t* pRef,
SWelsME* pMe) {
pMe->sMv = ksBestMv;
pMe->pRefMb = pRef;
pMe->uiSadCost = kiBestSadCost;
}
static inline void MeEndIntepelSearch (SWelsME* pMe) {
/* -> qpel mv */
pMe->sMv.iMvX *= (1 << 2);
pMe->sMv.iMvY *= (1 << 2);
pMe->uiSatdCost = pMe->uiSadCost;
}
void WelsInitMeFunc (SWelsFuncPtrList* pFuncList, uint32_t uiCpuFlag, bool bScreenContent) {
pFuncList->pfUpdateFMESwitch = UpdateFMESwitchNull;
if (!bScreenContent) {
pFuncList->pfCheckDirectionalMv = CheckDirectionalMvFalse;
pFuncList->pfCalculateBlockFeatureOfFrame[0] =
pFuncList->pfCalculateBlockFeatureOfFrame[1] = NULL;
pFuncList->pfCalculateSingleBlockFeature[0] =
pFuncList->pfCalculateSingleBlockFeature[1] = NULL;
} else {
pFuncList->pfCheckDirectionalMv = CheckDirectionalMv;
//for cross serarch
pFuncList->pfVerticalFullSearch = LineFullSearch_c;
pFuncList->pfHorizontalFullSearch = LineFullSearch_c;
#if defined (X86_ASM)
if (uiCpuFlag & WELS_CPU_SSE41) {
pFuncList->pfSampleSadHor8[0] = SampleSad8x8Hor8_sse41;
pFuncList->pfSampleSadHor8[1] = SampleSad16x16Hor8_sse41;
pFuncList->pfVerticalFullSearch = VerticalFullSearchUsingSSE41;
pFuncList->pfHorizontalFullSearch = HorizontalFullSearchUsingSSE41;
}
#endif
//for feature search
pFuncList->pfInitializeHashforFeature = InitializeHashforFeature_c;
pFuncList->pfFillQpelLocationByFeatureValue = FillQpelLocationByFeatureValue_c;
pFuncList->pfCalculateBlockFeatureOfFrame[0] = SumOf8x8BlockOfFrame_c;
pFuncList->pfCalculateBlockFeatureOfFrame[1] = SumOf16x16BlockOfFrame_c;
//TODO: it is possible to differentiate width that is times of 8, so as to accelerate the speed when width is times of 8?
pFuncList->pfCalculateSingleBlockFeature[0] = SumOf8x8SingleBlock_c;
pFuncList->pfCalculateSingleBlockFeature[1] = SumOf16x16SingleBlock_c;
#if defined (X86_ASM)
if (uiCpuFlag & WELS_CPU_SSE2) {
//for feature search
pFuncList->pfInitializeHashforFeature = InitializeHashforFeature_sse2;
pFuncList->pfFillQpelLocationByFeatureValue = FillQpelLocationByFeatureValue_sse2;
pFuncList->pfCalculateBlockFeatureOfFrame[0] = SumOf8x8BlockOfFrame_sse2;
pFuncList->pfCalculateBlockFeatureOfFrame[1] = SumOf16x16BlockOfFrame_sse2;
//TODO: it is possible to differentiate width that is times of 8, so as to accelerate the speed when width is times of 8?
pFuncList->pfCalculateSingleBlockFeature[0] = SumOf8x8SingleBlock_sse2;
pFuncList->pfCalculateSingleBlockFeature[1] = SumOf16x16SingleBlock_sse2;
}
if (uiCpuFlag & WELS_CPU_SSE41) {
//for feature search
pFuncList->pfCalculateBlockFeatureOfFrame[0] = SumOf8x8BlockOfFrame_sse4;
pFuncList->pfCalculateBlockFeatureOfFrame[1] = SumOf16x16BlockOfFrame_sse4;
}
#endif
#if defined (HAVE_NEON)
if (uiCpuFlag & WELS_CPU_NEON) {
//for feature search
pFuncList->pfInitializeHashforFeature = InitializeHashforFeature_neon;
pFuncList->pfFillQpelLocationByFeatureValue = FillQpelLocationByFeatureValue_neon;
pFuncList->pfCalculateBlockFeatureOfFrame[0] = SumOf8x8BlockOfFrame_neon;
pFuncList->pfCalculateBlockFeatureOfFrame[1] = SumOf16x16BlockOfFrame_neon;
//TODO: it is possible to differentiate width that is times of 8, so as to accelerate the speed when width is times of 8?
pFuncList->pfCalculateSingleBlockFeature[0] = SumOf8x8SingleBlock_neon;
pFuncList->pfCalculateSingleBlockFeature[1] = SumOf16x16SingleBlock_neon;
}
#endif
#if defined (HAVE_NEON_AARCH64)
if (uiCpuFlag & WELS_CPU_NEON) {
//for feature search
pFuncList->pfInitializeHashforFeature = InitializeHashforFeature_AArch64_neon;
pFuncList->pfFillQpelLocationByFeatureValue = FillQpelLocationByFeatureValue_AArch64_neon;
pFuncList->pfCalculateBlockFeatureOfFrame[0] = SumOf8x8BlockOfFrame_AArch64_neon;
pFuncList->pfCalculateBlockFeatureOfFrame[1] = SumOf16x16BlockOfFrame_AArch64_neon;
//TODO: it is possible to differentiate width that is times of 8, so as to accelerate the speed when width is times of 8?
pFuncList->pfCalculateSingleBlockFeature[0] = SumOf8x8SingleBlock_AArch64_neon;
pFuncList->pfCalculateSingleBlockFeature[1] = SumOf16x16SingleBlock_AArch64_neon;
}
#endif
}
}
/*!
* \brief BL mb motion estimate search
*
* \param enc Wels encoder context
* \param pMe Wels me information
*
* \return NONE
*/
void WelsMotionEstimateSearch (SWelsFuncPtrList* pFuncList, SDqLayer* pCurDqLayer, SWelsME* pMe, SSlice* pSlice) {
const int32_t kiStrideEnc = pCurDqLayer->iEncStride[0];
const int32_t kiStrideRef = pCurDqLayer->pRefPic->iLineSize[0];
// Step 1: Initial point prediction
if (!WelsMotionEstimateInitialPoint (pFuncList, pMe, pSlice, kiStrideEnc, kiStrideRef)) {
pFuncList->pfSearchMethod[pMe->uiBlockSize] (pFuncList, pMe, pSlice, kiStrideEnc, kiStrideRef);
MeEndIntepelSearch (pMe);
}
pFuncList->pfCalculateSatd (pFuncList->sSampleDealingFuncs.pfSampleSatd[pMe->uiBlockSize], pMe, kiStrideEnc,
kiStrideRef);
}
void WelsMotionEstimateSearchStatic (SWelsFuncPtrList* pFuncList, SDqLayer* pCurDqLayer, SWelsME* pMe, SSlice* pLpslice) {
const int32_t kiStrideEnc = pCurDqLayer->iEncStride[0];
const int32_t kiStrideRef = pCurDqLayer->pRefPic->iLineSize[0];
pMe->sMv.iMvX = pMe->sMv.iMvY = 0;
pMe->uiSadCost =
pFuncList->sSampleDealingFuncs.pfSampleSad[pMe->uiBlockSize] (pMe->pEncMb, kiStrideEnc, pMe->pRefMb, kiStrideRef) ;
pMe->uiSadCost += COST_MVD (pMe->pMvdCost, - pMe->sMvp.iMvX, - pMe->sMvp.iMvY);
MeEndIntepelSearch (pMe);
pFuncList->pfCalculateSatd (pFuncList->sSampleDealingFuncs.pfSampleSatd[pMe->uiBlockSize], pMe, kiStrideEnc,
kiStrideRef);
}
void WelsMotionEstimateSearchScrolled (SWelsFuncPtrList* pFuncList, SDqLayer* pCurDqLayer, SWelsME* pMe, SSlice* pSlice) {
const int32_t kiStrideEnc = pCurDqLayer->iEncStride[0];
const int32_t kiStrideRef = pCurDqLayer->pRefPic->iLineSize[0];
pMe->sMv = pMe->sDirectionalMv;
pMe->pRefMb = pMe->pColoRefMb + pMe->sMv.iMvY * kiStrideRef + pMe->sMv.iMvX;
pMe->uiSadCost =
pFuncList->sSampleDealingFuncs.pfSampleSad[pMe->uiBlockSize] (pMe->pEncMb, kiStrideEnc, pMe->pRefMb, kiStrideRef)
+ COST_MVD (pMe->pMvdCost, (pMe->sMv.iMvX * (1 << 2)) - pMe->sMvp.iMvX, (pMe->sMv.iMvY * (1 << 2)) - pMe->sMvp.iMvY);
MeEndIntepelSearch (pMe);
pFuncList->pfCalculateSatd (pFuncList->sSampleDealingFuncs.pfSampleSatd[pMe->uiBlockSize], pMe, kiStrideEnc,
kiStrideRef);
}
/*!
* \brief EL mb motion estimate initial point testing
*
* \param pix_pFuncList SSampleDealingFunc
* \param pMe Wels me information
* \param mv_range search range in motion estimate
* \param point the best match point in motion estimation
*
* \return NONE
*/
bool WelsMotionEstimateInitialPoint (SWelsFuncPtrList* pFuncList, SWelsME* pMe, SSlice* pSlice, int32_t iStrideEnc,
int32_t iStrideRef) {
PSampleSadSatdCostFunc pSad = pFuncList->sSampleDealingFuncs.pfSampleSad[pMe->uiBlockSize];
const uint16_t* kpMvdCost = pMe->pMvdCost;
uint8_t* const kpEncMb = pMe->pEncMb;
int16_t iMvc0, iMvc1;
int32_t iSadCost;
int32_t iBestSadCost;
uint8_t* pRefMb;
uint8_t* pFref2;
uint32_t i;
const uint32_t kuiMvcNum = pSlice->uiMvcNum;
const SMVUnitXY* kpMvcList = &pSlice->sMvc[0];
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, ksMvStartMin.iMvX, ksMvStartMax.iMvX);
sMv.iMvY = WELS_CLIP3 ((2 + ksMvp.iMvY) >> 2, ksMvStartMin.iMvY, ksMvStartMax.iMvY);
pRefMb = &pMe->pRefMb[sMv.iMvY * iStrideRef + sMv.iMvX];
iBestSadCost = pSad (kpEncMb, iStrideEnc, pRefMb, iStrideRef);
iBestSadCost += COST_MVD (kpMvdCost, ((sMv.iMvX) * (1 << 2)) - ksMvp.iMvX, ((sMv.iMvY) * (1 << 2)) - ksMvp.iMvY);
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, 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];
iSadCost = pSad (kpEncMb, iStrideEnc, pFref2, iStrideRef) +
COST_MVD (kpMvdCost, (iMvc0 * (1 << 2)) - ksMvp.iMvX, (iMvc1 * (1 << 2)) - ksMvp.iMvY);
if (iSadCost < iBestSadCost) {
sMv.iMvX = iMvc0;
sMv.iMvY = iMvc1;
pRefMb = pFref2;
iBestSadCost = iSadCost;
}
}
}
if (pFuncList->pfCheckDirectionalMv
(pSad, pMe, ksMvStartMin, ksMvStartMax, iStrideEnc, iStrideRef, iSadCost)) {
sMv = pMe->sDirectionalMv;
pRefMb = &pMe->pColoRefMb[sMv.iMvY * iStrideRef + sMv.iMvX];
iBestSadCost = iSadCost;
}
UpdateMeResults (sMv, iBestSadCost, pRefMb, pMe);
if (iBestSadCost < static_cast<int32_t> (pMe->uSadPredISatd.uiSadPred)) {
//Initial point early Stop
MeEndIntepelSearch (pMe);
return true;
}
return false;
}
void CalculateSatdCost (PSampleSadSatdCostFunc pSatd, SWelsME* pMe,
const int32_t kiEncStride, const int32_t kiRefStride) {
pMe->uSadPredISatd.uiSatd = pSatd (pMe->pEncMb, kiEncStride, pMe->pRefMb, kiRefStride);
pMe->uiSatdCost = pMe->uSadPredISatd.uiSatd + COST_MVD (pMe->pMvdCost, pMe->sMv.iMvX - pMe->sMvp.iMvX,
pMe->sMv.iMvY - pMe->sMvp.iMvY);
}
void NotCalculateSatdCost (PSampleSadSatdCostFunc pSatd, SWelsME* pMe,
const int32_t kiEncStride, const int32_t kiRefStride) {
}
/////////////////////////
// Diamond Search Basics
/////////////////////////
bool WelsMeSadCostSelect (int32_t* iSadCost, const uint16_t* kpMvdCost, int32_t* pBestCost, const int32_t kiDx,
const int32_t kiDy, int32_t* pIx, int32_t* pIy) {
int32_t iTempSadCost[4];
int32_t iInputSadCost = *pBestCost;
iTempSadCost[0] = iSadCost[0] + COST_MVD (kpMvdCost, kiDx, kiDy - 4);
iTempSadCost[1] = iSadCost[1] + COST_MVD (kpMvdCost, kiDx, kiDy + 4);
iTempSadCost[2] = iSadCost[2] + COST_MVD (kpMvdCost, kiDx - 4, kiDy);
iTempSadCost[3] = iSadCost[3] + COST_MVD (kpMvdCost, kiDx + 4, kiDy);
if (iTempSadCost[0] < *pBestCost) {
*pBestCost = iTempSadCost[0];
*pIx = 0;
*pIy = 1;
}
if (iTempSadCost[1] < *pBestCost) {
*pBestCost = iTempSadCost[1];
*pIx = 0;
*pIy = -1;
}
if (iTempSadCost[2] < *pBestCost) {
*pBestCost = iTempSadCost[2];
*pIx = 1;
*pIy = 0;
}
if (iTempSadCost[3] < *pBestCost) {
*pBestCost = iTempSadCost[3];
*pIx = -1;
*pIy = 0;
}
return (*pBestCost == iInputSadCost);
}
void WelsDiamondSearch (SWelsFuncPtrList* pFuncList, SWelsME* pMe, SSlice* pSlice,
const int32_t kiStrideEnc, const int32_t kiStrideRef) {
PSample4SadCostFunc pSad = pFuncList->sSampleDealingFuncs.pfSample4Sad[pMe->uiBlockSize];
uint8_t* pFref = pMe->pRefMb;
uint8_t* const kpEncMb = pMe->pEncMb;
const uint16_t* kpMvdCost = pMe->pMvdCost;
const SMVUnitXY ksMvStartMin = pSlice->sMvStartMin;
const SMVUnitXY ksMvStartMax = pSlice->sMvStartMax;
int32_t iMvDx = ((pMe->sMv.iMvX) * (1 << 2)) - pMe->sMvp.iMvX;
int32_t iMvDy = ((pMe->sMv.iMvY) * (1 << 2)) - pMe->sMvp.iMvY;
uint8_t* pRefMb = pFref;
int32_t iBestCost = (pMe->uiSadCost);
int32_t iTimeThreshold = ITERATIVE_TIMES;
ENFORCE_STACK_ALIGN_1D (int32_t, iSadCosts, 4, 16)
while (iTimeThreshold--) {
pMe->sMv.iMvX = (iMvDx + pMe->sMvp.iMvX) >> 2;
pMe->sMv.iMvY = (iMvDy + pMe->sMvp.iMvY) >> 2;
if (!CheckMvInRange (pMe->sMv, ksMvStartMin, ksMvStartMax))
continue;
pSad (kpEncMb, kiStrideEnc, pRefMb, kiStrideRef, &iSadCosts[0]);
int32_t iX, iY;
const bool kbIsBestCostWorse = WelsMeSadCostSelect (iSadCosts, kpMvdCost, &iBestCost, iMvDx, iMvDy, &iX, &iY);
if (kbIsBestCostWorse)
break;
iMvDx -= (iX * (1 << 2)) ;
iMvDy -= (iY * (1 << 2)) ;
pRefMb -= (iX + iY * kiStrideRef);
}
/* integer-pel mv */
pMe->sMv.iMvX = (iMvDx + pMe->sMvp.iMvX) >> 2;
pMe->sMv.iMvY = (iMvDy + pMe->sMvp.iMvY) >> 2;
pMe->uiSatdCost = pMe->uiSadCost = (iBestCost);
pMe->pRefMb = pRefMb;
}
/////////////////////////
// DirectionalMv Basics
/////////////////////////
bool CheckDirectionalMv (PSampleSadSatdCostFunc pSad, SWelsME* pMe,
const SMVUnitXY ksMinMv, const SMVUnitXY ksMaxMv, const int32_t kiEncStride, const int32_t kiRefStride,
int32_t& iBestSadCost) {
const int16_t kiMvX = pMe->sDirectionalMv.iMvX;
const int16_t kiMvY = pMe->sDirectionalMv.iMvY;
//Check MV from scrolling detection
if ((BLOCK_16x16 != pMe->uiBlockSize) //scrolled_MV with P16x16 is checked SKIP checking function
&& (kiMvX | kiMvY) //(0,0) checked in ordinary initial point checking
&& CheckMvInRange (pMe->sDirectionalMv, ksMinMv, ksMaxMv)) {
uint8_t* pRef = &pMe->pColoRefMb[kiMvY * kiRefStride + kiMvX];
uint32_t uiCurrentSadCost = pSad (pMe->pEncMb, kiEncStride, pRef, kiRefStride) +
COST_MVD (pMe->pMvdCost, (kiMvX * (1 << 2)) - pMe->sMvp.iMvX, (kiMvY * (1 << 2)) - pMe->sMvp.iMvY);
if (uiCurrentSadCost < pMe->uiSadCost) {
iBestSadCost = uiCurrentSadCost;
return true;
}
}
return false;
}
bool CheckDirectionalMvFalse (PSampleSadSatdCostFunc pSad, SWelsME* vpMe,
const SMVUnitXY ksMinMv, const SMVUnitXY ksMaxMv, const int32_t kiEncStride, const int32_t kiRefStride,
int32_t& iBestSadCost) {
return false;
}
/////////////////////////
// Cross Search Basics
/////////////////////////
#if defined (X86_ASM)
void CalcMvdCostx8_c (uint16_t* pMvdCost, const int32_t kiStartMv, uint16_t* pMvdTable, const uint16_t kiFixedCost) {
uint16_t* pBaseCost = pMvdCost;
const int32_t kiOffset = (kiStartMv * (1 << 2));
uint16_t* pMvd = pMvdTable + kiOffset;
for (int32_t i = 0; i < 8; ++ i) {
pBaseCost[i] = ((*pMvd) + kiFixedCost);
pMvd += 4;
}
}
void VerticalFullSearchUsingSSE41 (SWelsFuncPtrList* pFuncList, SWelsME* pMe,
uint16_t* pMvdTable,
const int32_t kiEncStride, const int32_t kiRefStride,
const int16_t kiMinMv, const int16_t kiMaxMv,
const bool bVerticalSearch) {
uint8_t* kpEncMb = pMe->pEncMb;
const int32_t kiCurMeBlockPix = pMe->iCurMeBlockPixY;
uint8_t* pRef = &pMe->pColoRefMb[kiMinMv * kiRefStride];
const int32_t kiCurMeBlockPixY = pMe->iCurMeBlockPixY;
int32_t iMinPos = kiCurMeBlockPixY + kiMinMv;
int32_t iMaxPos = kiCurMeBlockPixY + kiMaxMv;
int32_t iFixedMvd = * (pMvdTable - pMe->sMvp.iMvX);
uint16_t* pMvdCost = & (pMvdTable[ (kiMinMv * (1 << 2)) - pMe->sMvp.iMvY]);
int16_t iStartMv = 0;
const int32_t kIsBlock16x16 = pMe->uiBlockSize == BLOCK_16x16;
const int32_t kiEdgeBlocks = kIsBlock16x16 ? 16 : 8;
PSampleSadHor8Func pSampleSadHor8 = pFuncList->pfSampleSadHor8[kIsBlock16x16];
PSampleSadSatdCostFunc pSad = pFuncList->sSampleDealingFuncs.pfSampleSad[pMe->uiBlockSize];
PTransposeMatrixBlockFunc TransposeMatrixBlock = kIsBlock16x16 ? TransposeMatrixBlock16x16_sse2 :
TransposeMatrixBlock8x8_mmx;
PTransposeMatrixBlocksFunc TransposeMatrixBlocks = kIsBlock16x16 ? TransposeMatrixBlocksx16_sse2 :
TransposeMatrixBlocksx8_mmx;
const int32_t kiDiff = iMaxPos - iMinPos;
const int32_t kiRowNum = WELS_ALIGN ((kiDiff - kiEdgeBlocks + 1), kiEdgeBlocks);
const int32_t kiBlocksNum = kIsBlock16x16 ? (kiRowNum >> 4) : (kiRowNum >> 3);
int32_t iCountLoop8 = (kiRowNum - kiEdgeBlocks) >> 3;
const int32_t kiRemainingVectors = kiDiff - (iCountLoop8 << 3);
const int32_t kiMatrixStride = MAX_VERTICAL_MV_RANGE;
ENFORCE_STACK_ALIGN_2D (uint8_t, uiMatrixRef, 16, kiMatrixStride, 16); // transpose matrix result for ref
ENFORCE_STACK_ALIGN_2D (uint8_t, uiMatrixEnc, 16, 16, 16); // transpose matrix result for enc
assert (kiRowNum <= kiMatrixStride); // make sure effective memory
TransposeMatrixBlock (&uiMatrixEnc[0][0], 16, kpEncMb, kiEncStride);
TransposeMatrixBlocks (&uiMatrixRef[0][0], kiMatrixStride, pRef, kiRefStride, kiBlocksNum);
ENFORCE_STACK_ALIGN_1D (uint16_t, uiBaseCost, 8, 16);
int32_t iTargetPos = iMinPos;
int16_t iBestPos = pMe->sMv.iMvX;
uint32_t uiBestCost = pMe->uiSadCost;
uint32_t uiCostMin;
int32_t iIndexMinPos;
kpEncMb = &uiMatrixEnc[0][0];
pRef = &uiMatrixRef[0][0];
while (iCountLoop8 > 0) {
CalcMvdCostx8_c (uiBaseCost, iStartMv, pMvdCost, iFixedMvd);
uiCostMin = pSampleSadHor8 (kpEncMb, 16, pRef, kiMatrixStride, uiBaseCost, &iIndexMinPos);
if (uiCostMin < uiBestCost) {
uiBestCost = uiCostMin;
iBestPos = iTargetPos + iIndexMinPos;
}
iTargetPos += 8;
pRef += 8;
iStartMv += 8;
-- iCountLoop8;
}
if (kiRemainingVectors > 0) {
kpEncMb = pMe->pEncMb;
pRef = &pMe->pColoRefMb[ (iTargetPos - kiCurMeBlockPix) * kiRefStride];
while (iTargetPos < iMaxPos) {
const uint16_t uiMvdCost = pMvdCost[iStartMv * (1 << 2)];
uint32_t uiSadCost = pSad (kpEncMb, kiEncStride, pRef, kiRefStride) + (iFixedMvd + uiMvdCost);
if (uiSadCost < uiBestCost) {
uiBestCost = uiSadCost;
iBestPos = iTargetPos;
}
iStartMv++;
pRef += kiRefStride;
++iTargetPos;
}
}
if (uiBestCost < pMe->uiSadCost) {
SMVUnitXY sBestMv;
sBestMv.iMvX = 0;
sBestMv.iMvY = iBestPos - kiCurMeBlockPix;
UpdateMeResults (sBestMv, uiBestCost, &pMe->pColoRefMb[sBestMv.iMvY * kiRefStride], pMe);
}
}
void HorizontalFullSearchUsingSSE41 (SWelsFuncPtrList* pFuncList, SWelsME* pMe,
uint16_t* pMvdTable,
const int32_t kiEncStride, const int32_t kiRefStride,
const int16_t kiMinMv, const int16_t kiMaxMv,
const bool bVerticalSearch) {
uint8_t* kpEncMb = pMe->pEncMb;
const int32_t iCurMeBlockPixX = pMe->iCurMeBlockPixX;
int32_t iMinPos = iCurMeBlockPixX + kiMinMv;
int32_t iMaxPos = iCurMeBlockPixX + kiMaxMv;
int32_t iFixedMvd = * (pMvdTable - pMe->sMvp.iMvY);
uint16_t* pMvdCost = & (pMvdTable[ (kiMinMv * (1 << 2)) - pMe->sMvp.iMvX]);
int16_t iStartMv = 0;
uint8_t* pRef = &pMe->pColoRefMb[kiMinMv];
const int32_t kIsBlock16x16 = pMe->uiBlockSize == BLOCK_16x16;
PSampleSadHor8Func pSampleSadHor8 = pFuncList->pfSampleSadHor8[kIsBlock16x16];
PSampleSadSatdCostFunc pSad = pFuncList->sSampleDealingFuncs.pfSampleSad[pMe->uiBlockSize];
ENFORCE_STACK_ALIGN_1D (uint16_t, uiBaseCost, 8, 16);
const int32_t kiNumVector = iMaxPos - iMinPos;
int32_t iCountLoop8 = kiNumVector >> 3;
const int32_t kiRemainingLoop8 = kiNumVector & 7;
int32_t iTargetPos = iMinPos;
int16_t iBestPos = pMe->sMv.iMvX;
uint32_t uiBestCost = pMe->uiSadCost;
uint32_t uiCostMin;
int32_t iIndexMinPos;
while (iCountLoop8 > 0) {
CalcMvdCostx8_c (uiBaseCost, iStartMv, pMvdCost, iFixedMvd);
uiCostMin = pSampleSadHor8 (kpEncMb, kiEncStride, pRef, kiRefStride, uiBaseCost, &iIndexMinPos);
if (uiCostMin < uiBestCost) {
uiBestCost = uiCostMin;
iBestPos = iTargetPos + iIndexMinPos;
}
iTargetPos += 8;
pRef += 8;
iStartMv += 8;
-- iCountLoop8;
}
if (kiRemainingLoop8 > 0) {
while (iTargetPos < iMaxPos) {
const uint16_t uiMvdCost = pMvdCost[iStartMv * (1 << 2)];
uint32_t uiSadCost = pSad (kpEncMb, kiEncStride, pRef, kiRefStride) + (iFixedMvd + uiMvdCost);
if (uiSadCost < uiBestCost) {
uiBestCost = uiSadCost;
iBestPos = iTargetPos;
}
iStartMv++;
++pRef;
++iTargetPos;
}
}
if (uiBestCost < pMe->uiSadCost) {
SMVUnitXY sBestMv;
sBestMv.iMvX = iBestPos - iCurMeBlockPixX;
sBestMv.iMvY = 0;
UpdateMeResults (sBestMv, uiBestCost, &pMe->pColoRefMb[sBestMv.iMvX], pMe);
}
}
#endif
void LineFullSearch_c (SWelsFuncPtrList* pFuncList, SWelsME* pMe,
uint16_t* pMvdTable,
const int32_t kiEncStride, const int32_t kiRefStride,
const int16_t iMinMv, const int16_t iMaxMv,
const bool bVerticalSearch) {
PSampleSadSatdCostFunc pSad = pFuncList->sSampleDealingFuncs.pfSampleSad[pMe->uiBlockSize];
const int32_t kiCurMeBlockPixX = pMe->iCurMeBlockPixX;
const int32_t kiCurMeBlockPixY = pMe->iCurMeBlockPixY;
int32_t iMinPos, iMaxPos;
int32_t iFixedMvd;
int32_t iCurMeBlockPix;
int32_t iStride;
uint16_t* pMvdCost;
if (bVerticalSearch) {
iMinPos = kiCurMeBlockPixY + iMinMv;
iMaxPos = kiCurMeBlockPixY + iMaxMv;
iFixedMvd = * (pMvdTable - pMe->sMvp.iMvX);
iCurMeBlockPix = pMe->iCurMeBlockPixY;
iStride = kiRefStride;
pMvdCost = & (pMvdTable[ (iMinMv * (1 << 2)) - pMe->sMvp.iMvY]);
} else {
iMinPos = kiCurMeBlockPixX + iMinMv;
iMaxPos = kiCurMeBlockPixX + iMaxMv;
iFixedMvd = * (pMvdTable - pMe->sMvp.iMvY);
iCurMeBlockPix = pMe->iCurMeBlockPixX;
iStride = 1;
pMvdCost = & (pMvdTable[ (iMinMv * (1 << 2)) - pMe->sMvp.iMvX]);
}
uint8_t* pRef = &pMe->pColoRefMb[ iMinMv * iStride];
uint32_t uiBestCost = 0xFFFFFFFF;
int32_t iBestPos = 0;
for (int32_t iTargetPos = iMinPos; iTargetPos < iMaxPos; ++ iTargetPos) {
uint8_t* const kpEncMb = pMe->pEncMb;
uint32_t uiSadCost = pSad (kpEncMb, kiEncStride, pRef, kiRefStride) + (iFixedMvd + *pMvdCost);
if (uiSadCost < uiBestCost) {
uiBestCost = uiSadCost;
iBestPos = iTargetPos;
}
pRef += iStride;
pMvdCost += 4;
}
if (uiBestCost < pMe->uiSadCost) {
SMVUnitXY sBestMv;
sBestMv.iMvX = bVerticalSearch ? 0 : (iBestPos - iCurMeBlockPix);
sBestMv.iMvY = bVerticalSearch ? (iBestPos - iCurMeBlockPix) : 0;
UpdateMeResults (sBestMv, uiBestCost, &pMe->pColoRefMb[sBestMv.iMvY * kiRefStride + sBestMv.iMvX], pMe);
}
}
void WelsMotionCrossSearch (SWelsFuncPtrList* pFuncList, SWelsME* pMe, SSlice* pSlice,
const int32_t kiEncStride, const int32_t kiRefStride) {
PLineFullSearchFunc pfVerticalFullSearchFunc = pFuncList->pfVerticalFullSearch;
PLineFullSearchFunc pfHorizontalFullSearchFunc = pFuncList->pfHorizontalFullSearch;
//vertical search
pfVerticalFullSearchFunc (pFuncList, pMe,
pMe->pMvdCost,
kiEncStride, kiRefStride,
pSlice->sMvStartMin.iMvY,
pSlice->sMvStartMax.iMvY, true);
//horizontal search
if (pMe->uiSadCost >= pMe->uiSadCostThreshold) {
pfHorizontalFullSearchFunc (pFuncList, pMe,
pMe->pMvdCost,
kiEncStride, kiRefStride,
pSlice->sMvStartMin.iMvX,
pSlice->sMvStartMax.iMvX,
false);
}
}
/////////////////////////
// Feature Search Basics
/////////////////////////
//memory related
int32_t RequestFeatureSearchPreparation (CMemoryAlign* pMa, const int32_t kiFrameWidth, const int32_t kiFrameHeight,
const int32_t iNeedFeatureStorage,
SFeatureSearchPreparation* pFeatureSearchPreparation) {
const int32_t kiFeatureStrategyIndex = iNeedFeatureStorage >> 16;
const bool bFme8x8 = ((iNeedFeatureStorage & 0x0000FF & ME_FME) == ME_FME);
const int32_t kiMarginSize = bFme8x8 ? 8 : 16;
const int32_t kiFrameSize = (kiFrameWidth - kiMarginSize) * (kiFrameHeight - kiMarginSize);
int32_t iListOfFeatureOfBlock;
if (0 == kiFeatureStrategyIndex) {
iListOfFeatureOfBlock = sizeof (uint16_t) * kiFrameSize;
} else {
iListOfFeatureOfBlock = sizeof (uint16_t) * kiFrameSize +
(kiFrameWidth - kiMarginSize) * sizeof (uint32_t) + kiFrameWidth * 8 * sizeof (uint8_t);
}
pFeatureSearchPreparation->pFeatureOfBlock =
(uint16_t*)pMa->WelsMalloc (iListOfFeatureOfBlock, "pFeatureOfBlock");
WELS_VERIFY_RETURN_IF (ENC_RETURN_MEMALLOCERR, NULL == (pFeatureSearchPreparation->pFeatureOfBlock))
pFeatureSearchPreparation->uiFeatureStrategyIndex = kiFeatureStrategyIndex;
pFeatureSearchPreparation->bFMESwitchFlag = true;
pFeatureSearchPreparation->uiFMEGoodFrameCount = FMESWITCH_DEFAULT_GOODFRAME_NUM;
pFeatureSearchPreparation->iHighFreMbCount = 0;
return ENC_RETURN_SUCCESS;
}
int32_t ReleaseFeatureSearchPreparation (CMemoryAlign* pMa, uint16_t*& pFeatureOfBlock) {
if (pMa && pFeatureOfBlock) {
pMa->WelsFree (pFeatureOfBlock, "pFeatureOfBlock");
pFeatureOfBlock = NULL;
return ENC_RETURN_SUCCESS;
}
return ENC_RETURN_UNEXPECTED;
}
int32_t RequestScreenBlockFeatureStorage (CMemoryAlign* pMa, const int32_t kiFrameWidth, const int32_t kiFrameHeight,
const int32_t iNeedFeatureStorage,
SScreenBlockFeatureStorage* pScreenBlockFeatureStorage) {
const int32_t kiFeatureStrategyIndex = iNeedFeatureStorage >> 16;
const int32_t kiMe8x8FME = iNeedFeatureStorage & 0x0000FF & ME_FME;
const int32_t kiMe16x16FME = ((iNeedFeatureStorage & 0x00FF00) >> 8) & ME_FME;
if ((kiMe8x8FME == ME_FME) && (kiMe16x16FME == ME_FME)) {
return ENC_RETURN_UNSUPPORTED_PARA;
//the following memory allocation cannot support when FME at both size
}
const bool bIsBlock8x8 = (kiMe8x8FME == ME_FME);
const int32_t kiMarginSize = bIsBlock8x8 ? 8 : 16;
const int32_t kiFrameSize = (kiFrameWidth - kiMarginSize) * (kiFrameHeight - kiMarginSize);
const int32_t kiListSize = (0 == kiFeatureStrategyIndex) ? (bIsBlock8x8 ? LIST_SIZE_SUM_8x8 : LIST_SIZE_SUM_16x16) :
256;
pScreenBlockFeatureStorage->pTimesOfFeatureValue = (uint32_t*)pMa->WelsMalloc (kiListSize * sizeof (uint32_t),
"pScreenBlockFeatureStorage->pTimesOfFeatureValue");
WELS_VERIFY_RETURN_IF (ENC_RETURN_MEMALLOCERR, NULL == pScreenBlockFeatureStorage->pTimesOfFeatureValue)
pScreenBlockFeatureStorage->pLocationOfFeature = (uint16_t**)pMa->WelsMalloc (kiListSize * sizeof (uint16_t*),
"pScreenBlockFeatureStorage->pLocationOfFeature");
WELS_VERIFY_RETURN_IF (ENC_RETURN_MEMALLOCERR, NULL == pScreenBlockFeatureStorage->pLocationOfFeature)
pScreenBlockFeatureStorage->pLocationPointer = (uint16_t*)pMa->WelsMalloc (2 * kiFrameSize * sizeof (uint16_t),
"pScreenBlockFeatureStorage->pLocationPointer");
WELS_VERIFY_RETURN_IF (ENC_RETURN_MEMALLOCERR, NULL == pScreenBlockFeatureStorage->pLocationPointer)
// uint16_t* pFeatureValuePointerList[WELS_MAX (LIST_SIZE_SUM_16x16, LIST_SIZE_MSE_16x16)] = {0};
pScreenBlockFeatureStorage->pFeatureValuePointerList = (uint16_t**)pMa->WelsMalloc (WELS_MAX (LIST_SIZE_SUM_16x16,
LIST_SIZE_MSE_16x16) * sizeof (uint16_t*),
"pScreenBlockFeatureStorage->pFeatureValuePointerList");
WELS_VERIFY_RETURN_IF (ENC_RETURN_MEMALLOCERR, NULL == pScreenBlockFeatureStorage->pFeatureValuePointerList)
pScreenBlockFeatureStorage->pFeatureOfBlockPointer = NULL;
pScreenBlockFeatureStorage->iIs16x16 = !bIsBlock8x8;
pScreenBlockFeatureStorage->uiFeatureStrategyIndex = kiFeatureStrategyIndex;
pScreenBlockFeatureStorage->iActualListSize = kiListSize;
WelsSetMemMultiplebytes_c (pScreenBlockFeatureStorage->uiSadCostThreshold, UINT_MAX, BLOCK_SIZE_ALL, sizeof(uint32_t));
pScreenBlockFeatureStorage->bRefBlockFeatureCalculated = false;
return ENC_RETURN_SUCCESS;
}
int32_t ReleaseScreenBlockFeatureStorage (CMemoryAlign* pMa, SScreenBlockFeatureStorage* pScreenBlockFeatureStorage) {
if (pMa && pScreenBlockFeatureStorage) {
if (pScreenBlockFeatureStorage->pTimesOfFeatureValue) {
pMa->WelsFree (pScreenBlockFeatureStorage->pTimesOfFeatureValue, "pScreenBlockFeatureStorage->pTimesOfFeatureValue");
pScreenBlockFeatureStorage->pTimesOfFeatureValue = NULL;
}
if (pScreenBlockFeatureStorage->pLocationOfFeature) {
pMa->WelsFree (pScreenBlockFeatureStorage->pLocationOfFeature, "pScreenBlockFeatureStorage->pLocationOfFeature");
pScreenBlockFeatureStorage->pLocationOfFeature = NULL;
}
if (pScreenBlockFeatureStorage->pLocationPointer) {
pMa->WelsFree (pScreenBlockFeatureStorage->pLocationPointer, "pScreenBlockFeatureStorage->pLocationPointer");
pScreenBlockFeatureStorage->pLocationPointer = NULL;
}
if (pScreenBlockFeatureStorage->pFeatureValuePointerList) {
pMa->WelsFree (pScreenBlockFeatureStorage->pFeatureValuePointerList,
"pScreenBlockFeatureStorage->pFeatureValuePointerList");
pScreenBlockFeatureStorage->pFeatureValuePointerList = NULL;
}
return ENC_RETURN_SUCCESS;
}
return ENC_RETURN_UNEXPECTED;
}
//preprocess related
int32_t SumOf8x8SingleBlock_c (uint8_t* pRef, const int32_t kiRefStride) {
int32_t iSum = 0, i;
for (i = 0; i < 8; i++) {
iSum += pRef[0] + pRef[1] + pRef[2] + pRef[3];
iSum += pRef[4] + pRef[5] + pRef[6] + pRef[7];
pRef += kiRefStride;
}
return iSum;
}
int32_t SumOf16x16SingleBlock_c (uint8_t* pRef, const int32_t kiRefStride) {
int32_t iSum = 0, i;
for (i = 0; i < 16; i++) {
iSum += pRef[0] + pRef[1] + pRef[2] + pRef[3];
iSum += pRef[4] + pRef[5] + pRef[6] + pRef[7];
iSum += pRef[8] + pRef[9] + pRef[10] + pRef[11];
iSum += pRef[12] + pRef[13] + pRef[14] + pRef[15];
pRef += kiRefStride;
}
return iSum;
}
void SumOf8x8BlockOfFrame_c (uint8_t* pRefPicture, const int32_t kiWidth, const int32_t kiHeight,
const int32_t kiRefStride,
uint16_t* pFeatureOfBlock, uint32_t pTimesOfFeatureValue[]) {
int32_t x, y;
uint8_t* pRef;
uint16_t* pBuffer;
int32_t iSum;
for (y = 0; y < kiHeight; y++) {
pRef = pRefPicture + kiRefStride * y;
pBuffer = pFeatureOfBlock + kiWidth * y;
for (x = 0; x < kiWidth; x++) {
iSum = SumOf8x8SingleBlock_c (pRef + x, kiRefStride);
pBuffer[x] = iSum;
pTimesOfFeatureValue[iSum]++;
}
}
}
void SumOf16x16BlockOfFrame_c (uint8_t* pRefPicture, const int32_t kiWidth, const int32_t kiHeight,
const int32_t kiRefStride,
uint16_t* pFeatureOfBlock, uint32_t pTimesOfFeatureValue[]) {
//TODO: this is similar to SumOf8x8BlockOfFrame_c expect the calling of single block func, refactor-able?
int32_t x, y;
uint8_t* pRef;
uint16_t* pBuffer;
int32_t iSum;
for (y = 0; y < kiHeight; y++) {
pRef = pRefPicture + kiRefStride * y;
pBuffer = pFeatureOfBlock + kiWidth * y;
for (x = 0; x < kiWidth; x++) {
iSum = SumOf16x16SingleBlock_c (pRef + x, kiRefStride);
pBuffer[x] = iSum;
pTimesOfFeatureValue[iSum]++;
}
}
}
void InitializeHashforFeature_c (uint32_t* pTimesOfFeatureValue, uint16_t* pBuf, const int32_t kiListSize,
uint16_t** pLocationOfFeature, uint16_t** pFeatureValuePointerList) {
//assign location pointer
uint16_t* pBufPos = pBuf;
for (int32_t i = 0 ; i < kiListSize; ++i) {
pLocationOfFeature[i] =
pFeatureValuePointerList[i] = pBufPos;
pBufPos += (pTimesOfFeatureValue[i] << 1);
}
}
void FillQpelLocationByFeatureValue_c (uint16_t* pFeatureOfBlock, const int32_t kiWidth, const int32_t kiHeight,
uint16_t** pFeatureValuePointerList) {
//assign each pixel's position
uint16_t* pSrcPointer = pFeatureOfBlock;
int32_t iQpelY = 0;
for (int32_t y = 0; y < kiHeight; y++) {
for (int32_t x = 0; x < kiWidth; x++) {
uint16_t uiFeature = pSrcPointer[x];
pFeatureValuePointerList[uiFeature][0] = x << 2;
pFeatureValuePointerList[uiFeature][1] = iQpelY;
pFeatureValuePointerList[uiFeature] += 2;
}
iQpelY += 4;
pSrcPointer += kiWidth;
}
}
bool CalculateFeatureOfBlock (SWelsFuncPtrList* pFunc, SPicture* pRef,
SScreenBlockFeatureStorage* pScreenBlockFeatureStorage) {
uint16_t* pFeatureOfBlock = pScreenBlockFeatureStorage->pFeatureOfBlockPointer;
uint32_t* pTimesOfFeatureValue = pScreenBlockFeatureStorage->pTimesOfFeatureValue;
uint16_t** pLocationOfFeature = pScreenBlockFeatureStorage->pLocationOfFeature;
uint16_t* pBuf = pScreenBlockFeatureStorage->pLocationPointer;
if (NULL == pFeatureOfBlock || NULL == pTimesOfFeatureValue || NULL == pLocationOfFeature || NULL == pBuf
|| NULL == pRef->pData[0]) {
return false;
}
uint8_t* pRefData = pRef->pData[0];
const int32_t iRefStride = pRef->iLineSize[0];
int32_t iIs16x16 = pScreenBlockFeatureStorage->iIs16x16;
const int32_t iEdgeDiscard = (iIs16x16 ? 16 : 8); //this is to save complexity of padding on pRef
const int32_t iWidth = pRef->iWidthInPixel - iEdgeDiscard;
const int32_t kiHeight = pRef->iHeightInPixel - iEdgeDiscard;
const int32_t kiActualListSize = pScreenBlockFeatureStorage->iActualListSize;
memset (pTimesOfFeatureValue, 0, sizeof (int32_t)*kiActualListSize);
(pFunc->pfCalculateBlockFeatureOfFrame[iIs16x16]) (pRefData, iWidth, kiHeight, iRefStride, pFeatureOfBlock,
pTimesOfFeatureValue);
//assign pLocationOfFeature pointer
pFunc->pfInitializeHashforFeature (pTimesOfFeatureValue, pBuf, kiActualListSize,
pLocationOfFeature, pScreenBlockFeatureStorage->pFeatureValuePointerList);
//assign each pixel's pLocationOfFeature
pFunc->pfFillQpelLocationByFeatureValue (pFeatureOfBlock, iWidth, kiHeight,
pScreenBlockFeatureStorage->pFeatureValuePointerList);
return true;
}
void PerformFMEPreprocess (SWelsFuncPtrList* pFunc, SPicture* pRef, uint16_t* pFeatureOfBlock,
SScreenBlockFeatureStorage* pScreenBlockFeatureStorage) {
pScreenBlockFeatureStorage->pFeatureOfBlockPointer = pFeatureOfBlock;
pScreenBlockFeatureStorage->bRefBlockFeatureCalculated = CalculateFeatureOfBlock (pFunc, pRef,
pScreenBlockFeatureStorage);
if (pScreenBlockFeatureStorage->bRefBlockFeatureCalculated) {
uint32_t uiRefPictureAvgQstepx16 = QStepx16ByQp[WelsMedian (0, pRef->iFrameAverageQp, 51)];
uint32_t uiSadCostThreshold16x16 = ((30 * (uiRefPictureAvgQstepx16 + 160)) >> 3);
pScreenBlockFeatureStorage->uiSadCostThreshold[BLOCK_16x16] = uiSadCostThreshold16x16;
pScreenBlockFeatureStorage->uiSadCostThreshold[BLOCK_8x8] = (uiSadCostThreshold16x16 >> 2);
pScreenBlockFeatureStorage->uiSadCostThreshold[BLOCK_16x8]
= pScreenBlockFeatureStorage->uiSadCostThreshold[BLOCK_8x16]
= pScreenBlockFeatureStorage->uiSadCostThreshold[BLOCK_4x4] = UINT_MAX;
}
}
//search related
bool SetFeatureSearchIn (SWelsFuncPtrList* pFunc, const SWelsME& sMe,
const SSlice* pSlice, SScreenBlockFeatureStorage* pRefFeatureStorage,
const int32_t kiEncStride, const int32_t kiRefStride,
SFeatureSearchIn* pFeatureSearchIn) {
pFeatureSearchIn->pSad = pFunc->sSampleDealingFuncs.pfSampleSad[sMe.uiBlockSize];
pFeatureSearchIn->iFeatureOfCurrent = pFunc->pfCalculateSingleBlockFeature[BLOCK_16x16 == sMe.uiBlockSize] (sMe.pEncMb,
kiEncStride);
pFeatureSearchIn->pEnc = sMe.pEncMb;
pFeatureSearchIn->pColoRef = sMe.pColoRefMb;
pFeatureSearchIn->iEncStride = kiEncStride;
pFeatureSearchIn->iRefStride = kiRefStride;
pFeatureSearchIn->uiSadCostThresh = sMe.uiSadCostThreshold;
pFeatureSearchIn->iCurPixX = sMe.iCurMeBlockPixX;
pFeatureSearchIn->iCurPixXQpel = (pFeatureSearchIn->iCurPixX << 2);
pFeatureSearchIn->iCurPixY = sMe.iCurMeBlockPixY;
pFeatureSearchIn->iCurPixYQpel = (pFeatureSearchIn->iCurPixY << 2);
pFeatureSearchIn->pTimesOfFeature = pRefFeatureStorage->pTimesOfFeatureValue;
pFeatureSearchIn->pQpelLocationOfFeature = pRefFeatureStorage->pLocationOfFeature;
pFeatureSearchIn->pMvdCostX = sMe.pMvdCost - pFeatureSearchIn->iCurPixXQpel - sMe.sMvp.iMvX;
pFeatureSearchIn->pMvdCostY = sMe.pMvdCost - pFeatureSearchIn->iCurPixYQpel - sMe.sMvp.iMvY;
pFeatureSearchIn->iMinQpelX = pFeatureSearchIn->iCurPixXQpel + ((pSlice->sMvStartMin.iMvX) * (1 << 2));
pFeatureSearchIn->iMinQpelY = pFeatureSearchIn->iCurPixYQpel + ((pSlice->sMvStartMin.iMvY) * (1 << 2));
pFeatureSearchIn->iMaxQpelX = pFeatureSearchIn->iCurPixXQpel + ((pSlice->sMvStartMax.iMvX) * (1 << 2));
pFeatureSearchIn->iMaxQpelY = pFeatureSearchIn->iCurPixYQpel + ((pSlice->sMvStartMax.iMvY) * (1 << 2));
if (NULL == pFeatureSearchIn->pSad || NULL == pFeatureSearchIn->pTimesOfFeature
|| NULL == pFeatureSearchIn->pQpelLocationOfFeature) {
return false;
}
return true;
}
void SaveFeatureSearchOut (const SMVUnitXY sBestMv, const uint32_t uiBestSadCost, uint8_t* pRef,
SFeatureSearchOut* pFeatureSearchOut) {
pFeatureSearchOut->sBestMv = sBestMv;
pFeatureSearchOut->uiBestSadCost = uiBestSadCost;
pFeatureSearchOut->pBestRef = pRef;
}
bool FeatureSearchOne (SFeatureSearchIn& sFeatureSearchIn, const int32_t iFeatureDifference,
const uint32_t kuiExpectedSearchTimes,
SFeatureSearchOut* pFeatureSearchOut) {
const int32_t iFeatureOfRef = (sFeatureSearchIn.iFeatureOfCurrent + iFeatureDifference);
if (iFeatureOfRef < 0 || iFeatureOfRef >= LIST_SIZE)
return true;
PSampleSadSatdCostFunc pSad = sFeatureSearchIn.pSad;
uint8_t* pEnc = sFeatureSearchIn.pEnc;
uint8_t* pColoRef = sFeatureSearchIn.pColoRef;
const int32_t iEncStride = sFeatureSearchIn.iEncStride;
const int32_t iRefStride = sFeatureSearchIn.iRefStride;
const uint16_t uiSadCostThresh = sFeatureSearchIn.uiSadCostThresh;
const int32_t iCurPixX = sFeatureSearchIn.iCurPixX;
const int32_t iCurPixY = sFeatureSearchIn.iCurPixY;
const int32_t iCurPixXQpel = sFeatureSearchIn.iCurPixXQpel;
const int32_t iCurPixYQpel = sFeatureSearchIn.iCurPixYQpel;
const int32_t iMinQpelX = sFeatureSearchIn.iMinQpelX;
const int32_t iMinQpelY = sFeatureSearchIn.iMinQpelY;
const int32_t iMaxQpelX = sFeatureSearchIn.iMaxQpelX;
const int32_t iMaxQpelY = sFeatureSearchIn.iMaxQpelY;
const int32_t iSearchTimes = WELS_MIN (sFeatureSearchIn.pTimesOfFeature[iFeatureOfRef], kuiExpectedSearchTimes);
const int32_t iSearchTimesx2 = (iSearchTimes << 1);
const uint16_t* pQpelPosition = sFeatureSearchIn.pQpelLocationOfFeature[iFeatureOfRef];
SMVUnitXY sBestMv;
uint32_t uiBestCost, uiTmpCost;
uint8_t* pBestRef, *pCurRef;
int32_t iQpelX, iQpelY;
int32_t iIntepelX, iIntepelY;
int32_t i;
sBestMv.iMvX = pFeatureSearchOut->sBestMv.iMvX;
sBestMv.iMvY = pFeatureSearchOut->sBestMv.iMvY;
uiBestCost = pFeatureSearchOut->uiBestSadCost;
pBestRef = pFeatureSearchOut->pBestRef;
for (i = 0; i < iSearchTimesx2; i += 2) {
iQpelX = pQpelPosition[i];
iQpelY = pQpelPosition[i + 1];
if ((iQpelX > iMaxQpelX) || (iQpelX < iMinQpelX)
|| (iQpelY > iMaxQpelY) || (iQpelY < iMinQpelY)
|| (iQpelX == iCurPixXQpel) || (iQpelY == iCurPixYQpel))
continue;
uiTmpCost = sFeatureSearchIn.pMvdCostX[ iQpelX ] + sFeatureSearchIn.pMvdCostY[ iQpelY ];
if (uiTmpCost + iFeatureDifference >= uiBestCost)
continue;
iIntepelX = (iQpelX >> 2) - iCurPixX;
iIntepelY = (iQpelY >> 2) - iCurPixY;
pCurRef = &pColoRef[iIntepelX + iIntepelY * iRefStride];
uiTmpCost += pSad (pEnc, iEncStride, pCurRef, iRefStride);
if (uiTmpCost < uiBestCost) {
sBestMv.iMvX = iIntepelX;
sBestMv.iMvY = iIntepelY;
uiBestCost = uiTmpCost;
pBestRef = pCurRef;
if (uiBestCost < uiSadCostThresh)
break;
}
}
SaveFeatureSearchOut (sBestMv, uiBestCost, pBestRef, pFeatureSearchOut);
return (i < iSearchTimesx2);
}
void MotionEstimateFeatureFullSearch (SFeatureSearchIn& sFeatureSearchIn,
const uint32_t kuiMaxSearchPoint,
SWelsME* pMe) {
SFeatureSearchOut sFeatureSearchOut = { { 0 } };//TODO: this can be refactored and removed
sFeatureSearchOut.uiBestSadCost = pMe->uiSadCost;
sFeatureSearchOut.sBestMv = pMe->sMv;
sFeatureSearchOut.pBestRef = pMe->pRefMb;
int32_t iFeatureDifference = 0;//TODO: change it according to computational-complexity setting when needed
FeatureSearchOne (sFeatureSearchIn, iFeatureDifference, kuiMaxSearchPoint, &sFeatureSearchOut);
if (sFeatureSearchOut.uiBestSadCost < pMe->uiSadCost) { //TODO: this may be refactored and removed
UpdateMeResults (sFeatureSearchOut.sBestMv,
sFeatureSearchOut.uiBestSadCost, sFeatureSearchOut.pBestRef,
pMe);
}
}
//switch related
static uint32_t CountFMECostDown (const SDqLayer* pCurLayer) {
uint32_t uiCostDownSum = 0;
const int32_t kiSliceCount = GetCurrentSliceNum (pCurLayer);
if (kiSliceCount >= 1) {
int32_t iSliceIndex = 0;
SSlice* pSlice = &pCurLayer->sLayerInfo.pSliceInLayer[iSliceIndex];
while (iSliceIndex < kiSliceCount) {
uiCostDownSum += pSlice->uiSliceFMECostDown;
++ pSlice;
++ iSliceIndex;
}
}
return uiCostDownSum;
}
#define FMESWITCH_MBAVERCOSTSAVING_THRESHOLD (2) //empirically set.
#define FMESWITCH_GOODFRAMECOUNT_MAX (5) //empirically set.
static void UpdateFMEGoodFrameCount (const uint32_t iAvMBNormalizedRDcostDown, uint8_t& uiFMEGoodFrameCount) {
//this strategy may be changed, here the number is derived from empirical-numbers
// uiFMEGoodFrameCount lies in [0,FMESWITCH_GOODFRAMECOUNT_MAX]
if (iAvMBNormalizedRDcostDown > FMESWITCH_MBAVERCOSTSAVING_THRESHOLD) {
if (uiFMEGoodFrameCount < FMESWITCH_GOODFRAMECOUNT_MAX)
++ uiFMEGoodFrameCount;
} else {
if (uiFMEGoodFrameCount > 0)
-- uiFMEGoodFrameCount;
}
}
void UpdateFMESwitch (SDqLayer* pCurLayer) {
const uint32_t iFMECost = CountFMECostDown (pCurLayer);
const uint32_t iAvMBNormalizedRDcostDown = iFMECost / (pCurLayer->iMbWidth * pCurLayer->iMbHeight);
UpdateFMEGoodFrameCount (iAvMBNormalizedRDcostDown, pCurLayer->pFeatureSearchPreparation->uiFMEGoodFrameCount);
}
void UpdateFMESwitchNull (SDqLayer* pCurLayer) {
}
/////////////////////////
// Search function options
/////////////////////////
void WelsDiamondCrossSearch (SWelsFuncPtrList* pFunc, SWelsME* pMe, SSlice* pSlice, const int32_t kiEncStride,
const int32_t kiRefStride) {
// Step 1: diamond search
WelsDiamondSearch (pFunc, pMe, pSlice, kiEncStride, kiRefStride);
// Step 2: CROSS search
pMe->uiSadCostThreshold = pMe->pRefFeatureStorage->uiSadCostThreshold[pMe->uiBlockSize];
if (pMe->uiSadCost >= pMe->uiSadCostThreshold) {
WelsMotionCrossSearch (pFunc, pMe, pSlice, kiEncStride, kiRefStride);
}
}
void WelsDiamondCrossFeatureSearch (SWelsFuncPtrList* pFunc, SWelsME* pMe, SSlice* pSlice, const int32_t kiEncStride,
const int32_t kiRefStride) {
// Step 1: diamond search + cross
WelsDiamondCrossSearch (pFunc, pMe, pSlice, kiEncStride, kiRefStride);
// Step 2: FeatureSearch
if (pMe->uiSadCost >= pMe->uiSadCostThreshold) {
pSlice->uiSliceFMECostDown += pMe->uiSadCost;
uint32_t uiMaxSearchPoint = INT_MAX;//TODO: change it according to computational-complexity setting
SFeatureSearchIn sFeatureSearchIn = {0};
if (SetFeatureSearchIn (pFunc, *pMe, pSlice, pMe->pRefFeatureStorage,
kiEncStride, kiRefStride,
&sFeatureSearchIn)) {
MotionEstimateFeatureFullSearch (sFeatureSearchIn, uiMaxSearchPoint, pMe);
}
pSlice->uiSliceFMECostDown -= pMe->uiSadCost;
}
}
} // namespace WelsEnc