generic-library/openh264
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Remove UBSAN warnings about negative left shift

Browse Source
This commit is contained in:
Haibo Zhu 2015-10-12 22:51:45 -07:00
parent fcab199083
commit 6e052edcef
12 changed files with 71 additions and 71 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
codec/common/src/deblocking_common.cpp
View File

@ -21,14 +21,14 @@ void DeblockLumaLt4_c (uint8_t* pPix, int32_t iStrideX, int32_t iStrideY, int32_
bool bDetaP2P0 = WELS_ABS (p2 - p0) < iBeta;
bool bDetaQ2Q0 = WELS_ABS (q2 - q0) < iBeta;
if (bDetaP2P0) {
pPix[-2 * iStrideX] = p1 + WELS_CLIP3 ((p2 + ((p0 + q0 + 1) >> 1) - (p1 << 1)) >> 1, -iTc0, iTc0);
pPix[-2 * iStrideX] = p1 + WELS_CLIP3 ((p2 + ((p0 + q0 + 1) >> 1) - (p1 * (1 << 1))) >> 1, -iTc0, iTc0);
iTc++;
}
if (bDetaQ2Q0) {
pPix[iStrideX] = q1 + WELS_CLIP3 ((q2 + ((p0 + q0 + 1) >> 1) - (q1 << 1)) >> 1, -iTc0, iTc0);
pPix[iStrideX] = q1 + WELS_CLIP3 ((q2 + ((p0 + q0 + 1) >> 1) - (q1 * (1 << 1))) >> 1, -iTc0, iTc0);
iTc++;
}
int32_t iDeta = WELS_CLIP3 ((((q0 - p0) << 2) + (p1 - q1) + 4) >> 3, -iTc, iTc);
int32_t iDeta = WELS_CLIP3 ((((q0 - p0) * (1 << 2)) + (p1 - q1) + 4) >> 3, -iTc, iTc);
pPix[-iStrideX] = WelsClip1 (p0 + iDeta); /* p0' */
pPix[0] = WelsClip1 (q0 - iDeta); /* q0' */
}
@ -56,23 +56,23 @@ void DeblockLumaEq4_c (uint8_t* pPix, int32_t iStrideX, int32_t iStrideY, int32_
bool bDetaQ2Q0 = WELS_ABS (q2 - q0) < iBeta;
if (bDetaP2P0) {
const int32_t p3 = pPix[-4 * iStrideX];
pPix[-iStrideX] = (p2 + (p1 << 1) + (p0 << 1) + (q0 << 1) + q1 + 4) >> 3; //p0
pPix[-iStrideX] = (p2 + (p1 * (1 << 1)) + (p0 * (1 << 1)) + (q0 * (1 << 1)) + q1 + 4) >> 3; //p0
pPix[-2 * iStrideX] = (p2 + p1 + p0 + q0 + 2) >> 2; //p1
pPix[-3 * iStrideX] = ((p3 << 1) + p2 + (p2 << 1) + p1 + p0 + q0 + 4) >> 3; //p2
pPix[-3 * iStrideX] = ((p3 * (1 << 1)) + p2 + (p2 * (1 << 1)) + p1 + p0 + q0 + 4) >> 3; //p2
} else {
pPix[-1 * iStrideX] = ((p1 << 1) + p0 + q1 + 2) >> 2; //p0
pPix[-1 * iStrideX] = ((p1 * (1 << 1)) + p0 + q1 + 2) >> 2; //p0
}
if (bDetaQ2Q0) {
const int32_t q3 = pPix[3 * iStrideX];
pPix[0] = (p1 + (p0 << 1) + (q0 << 1) + (q1 << 1) + q2 + 4) >> 3; //q0
pPix[0] = (p1 + (p0 * (1 << 1)) + (q0 * (1 << 1)) + (q1 * (1 << 1)) + q2 + 4) >> 3; //q0
pPix[iStrideX] = (p0 + q0 + q1 + q2 + 2) >> 2; //q1
pPix[2 * iStrideX] = ((q3 << 1) + q2 + (q2 << 1) + q1 + q0 + p0 + 4) >> 3; //q2
pPix[2 * iStrideX] = ((q3 * (1 << 1)) + q2 + (q2 * (1 << 1)) + q1 + q0 + p0 + 4) >> 3; //q2
} else {
pPix[0] = ((q1 << 1) + q0 + p1 + 2) >> 2; //q0
pPix[0] = ((q1 * (1 << 1)) + q0 + p1 + 2) >> 2; //q0
}
} else {
pPix[-iStrideX] = ((p1 << 1) + p0 + q1 + 2) >> 2; //p0
pPix[ 0] = ((q1 << 1) + q0 + p1 + 2) >> 2; //q0
pPix[-iStrideX] = ((p1 * (1 << 1)) + p0 + q1 + 2) >> 2; //p0
pPix[ 0] = ((q1 * (1 << 1)) + q0 + p1 + 2) >> 2; //q0
}
}
pPix += iStrideY;
@ -107,7 +107,7 @@ void DeblockChromaLt4_c (uint8_t* pPixCb, uint8_t* pPixCr, int32_t iStrideX, int
bDetaP1P0 = WELS_ABS (p1 - p0) < iBeta;
bDetaQ1Q0 = WELS_ABS (q1 - q0) < iBeta;
if (bDetaP0Q0 && bDetaP1P0 && bDetaQ1Q0) {
iDeta = WELS_CLIP3 ((((q0 - p0) << 2) + (p1 - q1) + 4) >> 3, -iTc0, iTc0);
iDeta = WELS_CLIP3 ((((q0 - p0) * (1 << 2)) + (p1 - q1) + 4) >> 3, -iTc0, iTc0);
pPixCb[-iStrideX] = WelsClip1 (p0 + iDeta); /* p0' */
pPixCb[0] = WelsClip1 (q0 - iDeta); /* q0' */
}
@ -123,7 +123,7 @@ void DeblockChromaLt4_c (uint8_t* pPixCb, uint8_t* pPixCr, int32_t iStrideX, int
bDetaQ1Q0 = WELS_ABS (q1 - q0) < iBeta;
if (bDetaP0Q0 && bDetaP1P0 && bDetaQ1Q0) {
iDeta = WELS_CLIP3 ((((q0 - p0) << 2) + (p1 - q1) + 4) >> 3, -iTc0, iTc0);
iDeta = WELS_CLIP3 ((((q0 - p0) * (1 << 2)) + (p1 - q1) + 4) >> 3, -iTc0, iTc0);
pPixCr[-iStrideX] = WelsClip1 (p0 + iDeta); /* p0' */
pPixCr[0] = WelsClip1 (q0 - iDeta); /* q0' */
}
@ -146,8 +146,8 @@ void DeblockChromaEq4_c (uint8_t* pPixCb, uint8_t* pPixCr, int32_t iStrideX, int
bDetaP1P0 = WELS_ABS (p1 - p0) < iBeta;
bDetaQ1Q0 = WELS_ABS (q1 - q0) < iBeta;
if (bDetaP0Q0 && bDetaP1P0 && bDetaQ1Q0) {
pPixCb[-iStrideX] = ((p1 << 1) + p0 + q1 + 2) >> 2; /* p0' */
pPixCb[0] = ((q1 << 1) + q0 + p1 + 2) >> 2; /* q0' */
pPixCb[-iStrideX] = ((p1 * (1 << 1)) + p0 + q1 + 2) >> 2; /* p0' */
pPixCb[0] = ((q1 * (1 << 1)) + q0 + p1 + 2) >> 2; /* q0' */
}
//cr
@ -159,8 +159,8 @@ void DeblockChromaEq4_c (uint8_t* pPixCb, uint8_t* pPixCr, int32_t iStrideX, int
bDetaP1P0 = WELS_ABS (p1 - p0) < iBeta;
bDetaQ1Q0 = WELS_ABS (q1 - q0) < iBeta;
if (bDetaP0Q0 && bDetaP1P0 && bDetaQ1Q0) {
pPixCr[-iStrideX] = ((p1 << 1) + p0 + q1 + 2) >> 2; /* p0' */
pPixCr[0] = ((q1 << 1) + q0 + p1 + 2) >> 2; /* q0' */
pPixCr[-iStrideX] = ((p1 * (1 << 1)) + p0 + q1 + 2) >> 2; /* p0' */
pPixCr[0] = ((q1 * (1 << 1)) + q0 + p1 + 2) >> 2; /* q0' */
}
pPixCr += iStrideY;
pPixCb += iStrideY;
@ -198,7 +198,7 @@ void DeblockChromaLt42_c (uint8_t* pPixCbCr, int32_t iStrideX, int32_t iStrideY,
bDetaP1P0 = WELS_ABS (p1 - p0) < iBeta;
bDetaQ1Q0 = WELS_ABS (q1 - q0) < iBeta;
if (bDetaP0Q0 && bDetaP1P0 && bDetaQ1Q0) {
iDeta = WELS_CLIP3 ((((q0 - p0) << 2) + (p1 - q1) + 4) >> 3, -iTc0, iTc0);
iDeta = WELS_CLIP3 ((((q0 - p0) * (1 << 2)) + (p1 - q1) + 4) >> 3, -iTc0, iTc0);
pPixCbCr[-iStrideX] = WelsClip1 (p0 + iDeta); /* p0' */
pPixCbCr[0] = WelsClip1 (q0 - iDeta); /* q0' */
}
@ -221,8 +221,8 @@ void DeblockChromaEq42_c (uint8_t* pPixCbCr, int32_t iStrideX, int32_t iStrideY,
bDetaP1P0 = WELS_ABS (p1 - p0) < iBeta;
bDetaQ1Q0 = WELS_ABS (q1 - q0) < iBeta;
if (bDetaP0Q0 && bDetaP1P0 && bDetaQ1Q0) {
pPixCbCr[-iStrideX] = ((p1 << 1) + p0 + q1 + 2) >> 2; /* p0' */
pPixCbCr[0] = ((q1 << 1) + q0 + p1 + 2) >> 2; /* q0' */
pPixCbCr[-iStrideX] = ((p1 * (1 << 1)) + p0 + q1 + 2) >> 2; /* p0' */
pPixCbCr[0] = ((q1 * (1 << 1)) + q0 + p1 + 2) >> 2; /* q0' */
}
pPixCbCr += iStrideY;

20
codec/decoder/core/src/error_concealment.cpp
View File

@ -224,23 +224,23 @@ void DoMbECMvCopy (PWelsDecoderContext pCtx, PPicture pDec, PPicture pRef, int32
iPicHeightBottomLimit = (pMCRefMem->iPicHeight - pCtx->sFrameCrop.iTopOffset * 2);
}
// further make sure no need to expand picture
int32_t iMinLeftOffset = (iPicWidthLeftLimit + 2) << 2;
int32_t iMaxRightOffset = ((iPicWidthRightLimit - 19) << 2);
int32_t iMinTopOffset = (iPicHeightTopLimit + 2) << 2;
int32_t iMaxBottomOffset = ((iPicHeightBottomLimit - 19) << 2);
int32_t iMinLeftOffset = (iPicWidthLeftLimit + 2) * (1 << 2);
int32_t iMaxRightOffset = ((iPicWidthRightLimit - 19) * (1 << 2));
int32_t iMinTopOffset = (iPicHeightTopLimit + 2) * (1 << 2);
int32_t iMaxBottomOffset = ((iPicHeightBottomLimit - 19) * (1 << 2));
if (iFullMVx < iMinLeftOffset) {
iFullMVx = (iFullMVx >> 2) << 2;
iFullMVx = (iFullMVx >> 2) * (1 << 2);
iFullMVx = WELS_MAX (iPicWidthLeftLimit, iFullMVx);
} else if (iFullMVx > iMaxRightOffset) {
iFullMVx = (iFullMVx >> 2) << 2;
iFullMVx = WELS_MIN (((iPicWidthRightLimit - 17) << 2), iFullMVx);
iFullMVx = (iFullMVx >> 2) * (1 << 2);
iFullMVx = WELS_MIN (((iPicWidthRightLimit - 17) * (1 << 2)), iFullMVx);
}
if (iFullMVy < iMinTopOffset) {
iFullMVy = (iFullMVy >> 2) << 2;
iFullMVy = (iFullMVy >> 2) * (1 << 2);
iFullMVy = WELS_MAX (iPicHeightTopLimit, iFullMVy);
} else if (iFullMVy > iMaxBottomOffset) {
iFullMVy = (iFullMVy >> 2) << 2;
iFullMVy = WELS_MIN (((iPicHeightBottomLimit - 17) << 2), iFullMVy);
iFullMVy = (iFullMVy >> 2) * (1 << 2);
iFullMVy = WELS_MIN (((iPicHeightBottomLimit - 17) * (1 << 2)), iFullMVy);
}
iMVs[0] = iFullMVx - (iMbXInPix << 2);
iMVs[1] = iFullMVy - (iMbYInPix << 2);

2
codec/decoder/core/src/parse_mb_syn_cabac.cpp
View File

@ -889,7 +889,7 @@ int32_t ParseResidualBlockCabac8x8 (PWelsNeighAvail pNeighAvail, uint8_t* pNonZe
do {
if (pSignificantMap[j] != 0) {
i = pScanTable[ j ];
sTCoeff[i] = uiQp >= 36 ? ((pSignificantMap[j] * pDeQuantMul[i]) << (uiQp / 6 - 6)) : ((
sTCoeff[i] = uiQp >= 36 ? ((pSignificantMap[j] * pDeQuantMul[i]) * (1 << (uiQp / 6 - 6))) : ((
pSignificantMap[j] * pDeQuantMul[i] + (1 << (5 - uiQp / 6))) >> (6 - uiQp / 6));
}
++j;

2
codec/decoder/core/src/parse_mb_syn_cavlc.cpp
View File

@ -981,7 +981,7 @@ int32_t WelsResidualBlockCavlc8x8 (SVlcTable* pVlcTable, uint8_t* pNonZeroCountC
iCoeffNum += iRun[i] + 1; //FIXME add 1 earlier ?
j = (iCoeffNum << 2) + iIdx4x4;
j = kpZigzagTable[ j ];
pTCoeff[j] = uiQp >= 36 ? ((iLevel[i] * kpDequantCoeff[j]) << (uiQp / 6 - 6))
pTCoeff[j] = uiQp >= 36 ? ((iLevel[i] * kpDequantCoeff[j]) * (1 << (uiQp / 6 - 6)))
: ((iLevel[i] * kpDequantCoeff[j] + (1 << (5 - uiQp / 6))) >> (6 - uiQp / 6));
}

4
codec/decoder/core/src/rec_mb.cpp
View File

@ -243,8 +243,8 @@ void BaseMC (sMCRefMember* pMCRefMem, int32_t iXOffset, int32_t iYOffset, SMcFun
int32_t iBlkWidth, int32_t iBlkHeight, int16_t iMVs[2]) {
int32_t iFullMVx = (iXOffset << 2) + iMVs[0]; //quarter pixel
int32_t iFullMVy = (iYOffset << 2) + iMVs[1];
iFullMVx = WELS_CLIP3 (iFullMVx, ((-PADDING_LENGTH + 2) << 2), ((pMCRefMem->iPicWidth + PADDING_LENGTH - 19) << 2));
iFullMVy = WELS_CLIP3 (iFullMVy, ((-PADDING_LENGTH + 2) << 2), ((pMCRefMem->iPicHeight + PADDING_LENGTH - 19) << 2));
iFullMVx = WELS_CLIP3 (iFullMVx, ((-PADDING_LENGTH + 2) * (1 << 2)), ((pMCRefMem->iPicWidth + PADDING_LENGTH - 19) * (1 << 2)));
iFullMVy = WELS_CLIP3 (iFullMVy, ((-PADDING_LENGTH + 2) * (1 << 2)), ((pMCRefMem->iPicHeight + PADDING_LENGTH - 19) * (1 << 2)));
int32_t iSrcPixOffsetLuma = (iFullMVx >> 2) + (iFullMVy >> 2) * pMCRefMem->iSrcLineLuma;
int32_t iSrcPixOffsetChroma = (iFullMVx >> 3) + (iFullMVy >> 3) * pMCRefMem->iSrcLineChroma;

4
codec/encoder/core/inc/param_svc.h
View File

@ -215,8 +215,8 @@ typedef struct TagWelsSvcCodingParam: SEncParamExt {
SUsedPicRect.iLeft = 0;
SUsedPicRect.iTop = 0;
SUsedPicRect.iWidth = ((iPicWidth >> 1) << 1);
SUsedPicRect.iHeight = ((iPicHeight >> 1) << 1);
SUsedPicRect.iWidth = ((iPicWidth >> 1) * (1 << 1));
SUsedPicRect.iHeight = ((iPicHeight >> 1) * (1 << 1));
iRCMode = pCodingParam.iRCMode; // rc mode

8
codec/encoder/core/inc/svc_motion_estimate.h
View File

@ -336,10 +336,10 @@ inline void SetMvWithinIntegerMvRange (const int32_t kiMbWidth, const int32_t ki
const int32_t kiMbY,
const int32_t kiMaxMvRange,
SMVUnitXY* pMvMin, SMVUnitXY* pMvMax) {
pMvMin->iMvX = WELS_MAX (-1 * ((kiMbX + 1) << 4) + INTPEL_NEEDED_MARGIN, -1 * kiMaxMvRange);
pMvMin->iMvY = WELS_MAX (-1 * ((kiMbY + 1) << 4) + INTPEL_NEEDED_MARGIN, -1 * kiMaxMvRange);
pMvMax->iMvX = WELS_MIN (((kiMbWidth - kiMbX) << 4) - INTPEL_NEEDED_MARGIN, kiMaxMvRange);
pMvMax->iMvY = WELS_MIN (((kiMbHeight - kiMbY) << 4) - INTPEL_NEEDED_MARGIN, kiMaxMvRange);
pMvMin->iMvX = WELS_MAX (-1 * ((kiMbX + 1) * (1 << 4)) + INTPEL_NEEDED_MARGIN, -1 * kiMaxMvRange);
pMvMin->iMvY = WELS_MAX (-1 * ((kiMbY + 1) * (1 << 4)) + INTPEL_NEEDED_MARGIN, -1 * kiMaxMvRange);
pMvMax->iMvX = WELS_MIN (((kiMbWidth - kiMbX) * (1 << 4)) - INTPEL_NEEDED_MARGIN, kiMaxMvRange);
pMvMax->iMvY = WELS_MIN (((kiMbHeight - kiMbY) * (1 << 4)) - INTPEL_NEEDED_MARGIN, kiMaxMvRange);
}
inline bool CheckMvInRange (const SMVUnitXY ksCurrentMv, const SMVUnitXY ksMinMv, const SMVUnitXY ksMaxMv) {

8
codec/encoder/core/src/encode_mb_aux.cpp
View File

@ -333,8 +333,8 @@ void WelsDctT4_c (int16_t* pDct, uint8_t* pPixel1, int32_t iStride1, uint8_t* pP
pDct[i ] = s[0] + s[1];
pDct[kiI2] = s[0] - s[1];
pDct[kiI1] = (s[3] << 1) + s[2];
pDct[kiI3] = s[3] - (s[2] << 1);
pDct[kiI1] = (s[3] * (1 << 1)) + s[2];
pDct[kiI3] = s[3] - (s[2] * (1 << 1));
}
/* vertical transform */
@ -350,8 +350,8 @@ void WelsDctT4_c (int16_t* pDct, uint8_t* pPixel1, int32_t iStride1, uint8_t* pP
pDct[i ] = s[0] + s[1];
pDct[kiI8 ] = s[0] - s[1];
pDct[kiI4 ] = (s[3] << 1) + s[2];
pDct[kiI12] = s[3] - (s[2] << 1);
pDct[kiI4 ] = (s[3] * (1 << 1)) + s[2];
pDct[kiI12] = s[3] - (s[2] * (1 << 1));
}
}

2
codec/encoder/core/src/set_mb_syn_cavlc.cpp
View File

@ -165,7 +165,7 @@ int32_t WriteBlockResidualCavlc (SWelsFuncPtrList* pFuncList, int16_t* pCoffLev
for (i = iTrailingOnes; i < iTotalCoeffs; i++) {
int32_t iVal = iLevel[i];
iLevelCode = (iVal - 1) << 1;
iLevelCode = (iVal - 1) * (1 << 1);
uiSign = (iLevelCode >> 31);
iLevelCode = (iLevelCode ^ uiSign) + (uiSign << 1);
iLevelCode -= ((i == iTrailingOnes) && (iTrailingOnes < 3)) << 1;

4
codec/encoder/core/src/svc_mode_decision.cpp
View File

@ -133,8 +133,8 @@ void SetMvBaseEnhancelayer (SWelsMD* pMd, SMB* pCurMb, const SMB* kpRefMb) {
SMVUnitXY sMv;
int32_t iRefMbPartIdx = ((pCurMb->iMbY & 0x01) << 1) + (pCurMb->iMbX & 0x01); //may be need modified
int32_t iScan4RefPartIdx = g_kuiMbCountScan4Idx[ (iRefMbPartIdx << 2)];
sMv.iMvX = kpRefMb->sMv[iScan4RefPartIdx].iMvX << 1;
sMv.iMvY = kpRefMb->sMv[iScan4RefPartIdx].iMvY << 1;
sMv.iMvX = kpRefMb->sMv[iScan4RefPartIdx].iMvX * (1 << 1);
sMv.iMvY = kpRefMb->sMv[iScan4RefPartIdx].iMvY * (1 << 1);
pMd->sMe.sMe16x16.sMvBase = sMv;

42
codec/encoder/core/src/svc_motion_estimate.cpp
View File

@ -65,8 +65,8 @@ static inline void UpdateMeResults (const SMVUnitXY ksBestMv, const uint32_t kiB
}
static inline void MeEndIntepelSearch (SWelsME* pMe) {
/* -> qpel mv */
pMe->sMv.iMvX <<= 2;
pMe->sMv.iMvY <<= 2;
pMe->sMv.iMvX *= (1 << 2);
pMe->sMv.iMvY *= (1 << 2);
pMe->uiSatdCost = pMe->uiSadCost;
}
@ -194,7 +194,7 @@ void WelsMotionEstimateSearchScrolled (SWelsFuncPtrList* pFuncList, SDqLayer* pC
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 << 2) - pMe->sMvp.iMvX, (pMe->sMv.iMvY << 2) - pMe->sMvp.iMvY);
+ 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);
@ -235,7 +235,7 @@ bool WelsMotionEstimateInitialPoint (SWelsFuncPtrList* pFuncList, SWelsME* pMe,
pRefMb = &pMe->pRefMb[sMv.iMvY * iStrideRef + sMv.iMvX];
iBestSadCost = pSad (kpEncMb, iStrideEnc, pRefMb, iStrideRef);
iBestSadCost += COST_MVD (kpMvdCost, ((sMv.iMvX) << 2) - ksMvp.iMvX, ((sMv.iMvY) << 2) - ksMvp.iMvY);
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)
@ -246,7 +246,7 @@ bool WelsMotionEstimateInitialPoint (SWelsFuncPtrList* pFuncList, SWelsME* pMe,
pFref2 = &pMe->pRefMb[iMvc1 * iStrideRef + iMvc0];
iSadCost = pSad (kpEncMb, iStrideEnc, pFref2, iStrideRef) +
COST_MVD (kpMvdCost, (iMvc0 << 2) - ksMvp.iMvX, (iMvc1 << 2) - ksMvp.iMvY);
COST_MVD (kpMvdCost, (iMvc0 * (1 << 2)) - ksMvp.iMvX, (iMvc1 * (1 << 2)) - ksMvp.iMvY);
if (iSadCost < iBestSadCost) {
sMv.iMvX = iMvc0;
@ -333,8 +333,8 @@ void WelsDiamondSearch (SWelsFuncPtrList* pFuncList, SWelsME* pMe, SSlice* pSlic
const SMVUnitXY ksMvStartMin = pSlice->sMvStartMin;
const SMVUnitXY ksMvStartMax = pSlice->sMvStartMax;
int32_t iMvDx = ((pMe->sMv.iMvX) << 2) - pMe->sMvp.iMvX;
int32_t iMvDy = ((pMe->sMv.iMvY) << 2) - pMe->sMvp.iMvY;
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);
@ -355,8 +355,8 @@ void WelsDiamondSearch (SWelsFuncPtrList* pFuncList, SWelsME* pMe, SSlice* pSlic
if (kbIsBestCostWorse)
break;
iMvDx -= iX << 2 ;
iMvDy -= iY << 2 ;
iMvDx -= (iX * (1 << 2)) ;
iMvDy -= (iY * (1 << 2)) ;
pRefMb -= (iX + iY * kiStrideRef);
@ -384,7 +384,7 @@ bool CheckDirectionalMv (PSampleSadSatdCostFunc pSad, SWelsME* pMe,
&& 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 << 2) - pMe->sMvp.iMvX, (kiMvY << 2) - pMe->sMvp.iMvY);
COST_MVD (pMe->pMvdCost, (kiMvX * (1 << 2)) - pMe->sMvp.iMvX, (kiMvY * (1 << 2)) - pMe->sMvp.iMvY);
if (uiCurrentSadCost < pMe->uiSadCost) {
iBestSadCost = uiCurrentSadCost;
return true;
@ -405,7 +405,7 @@ bool CheckDirectionalMvFalse (PSampleSadSatdCostFunc pSad, SWelsME* vpMe,
#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 << 2);
const int32_t kiOffset = (kiStartMv * (1 << 2));
uint16_t* pMvd = pMvdTable + kiOffset;
for (int32_t i = 0; i < 8; ++ i) {
pBaseCost[i] = ((*pMvd) + kiFixedCost);
@ -426,7 +426,7 @@ void VerticalFullSearchUsingSSE41 (SWelsFuncPtrList* pFuncList, SWelsME* pMe,
int32_t iMinPos = kiCurMeBlockPixY + kiMinMv;
int32_t iMaxPos = kiCurMeBlockPixY + kiMaxMv;
int32_t iFixedMvd = * (pMvdTable - pMe->sMvp.iMvX);
uint16_t* pMvdCost = & (pMvdTable[ (kiMinMv << 2) - pMe->sMvp.iMvY]);
uint16_t* pMvdCost = & (pMvdTable[ (kiMinMv * (1 << 2)) - pMe->sMvp.iMvY]);
int16_t iStartMv = 0;
@ -476,7 +476,7 @@ void VerticalFullSearchUsingSSE41 (SWelsFuncPtrList* pFuncList, SWelsME* pMe,
kpEncMb = pMe->pEncMb;
pRef = &pMe->pColoRefMb[ (iTargetPos - kiCurMeBlockPix) * kiRefStride];
while (iTargetPos < iMaxPos) {
const uint16_t uiMvdCost = pMvdCost[iStartMv << 2];
const uint16_t uiMvdCost = pMvdCost[iStartMv * (1 << 2)];
uint32_t uiSadCost = pSad (kpEncMb, kiEncStride, pRef, kiRefStride) + (iFixedMvd + uiMvdCost);
if (uiSadCost < uiBestCost) {
uiBestCost = uiSadCost;
@ -506,7 +506,7 @@ void HorizontalFullSearchUsingSSE41 (SWelsFuncPtrList* pFuncList, SWelsME* pMe,
int32_t iMinPos = iCurMeBlockPixX + kiMinMv;
int32_t iMaxPos = iCurMeBlockPixX + kiMaxMv;
int32_t iFixedMvd = * (pMvdTable - pMe->sMvp.iMvY);
uint16_t* pMvdCost = & (pMvdTable[ (kiMinMv << 2) - pMe->sMvp.iMvX]);
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;
@ -536,7 +536,7 @@ void HorizontalFullSearchUsingSSE41 (SWelsFuncPtrList* pFuncList, SWelsME* pMe,
}
if (kiRemainingLoop8 > 0) {
while (iTargetPos < iMaxPos) {
const uint16_t uiMvdCost = pMvdCost[iStartMv << 2];
const uint16_t uiMvdCost = pMvdCost[iStartMv * (1 << 2)];
uint32_t uiSadCost = pSad (kpEncMb, kiEncStride, pRef, kiRefStride) + (iFixedMvd + uiMvdCost);
if (uiSadCost < uiBestCost) {
uiBestCost = uiSadCost;
@ -575,14 +575,14 @@ void LineFullSearch_c (SWelsFuncPtrList* pFuncList, SWelsME* pMe,
iFixedMvd = * (pMvdTable - pMe->sMvp.iMvX);
iCurMeBlockPix = pMe->iCurMeBlockPixY;
iStride = kiRefStride;
pMvdCost = & (pMvdTable[ (iMinMv << 2) - pMe->sMvp.iMvY]);
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 << 2) - pMe->sMvp.iMvX]);
pMvdCost = & (pMvdTable[ (iMinMv * (1 << 2)) - pMe->sMvp.iMvX]);
}
uint8_t* pRef = &pMe->pColoRefMb[ iMinMv * iStride];
uint32_t uiBestCost = 0xFFFFFFFF;
@ -906,10 +906,10 @@ bool SetFeatureSearchIn (SWelsFuncPtrList* pFunc, const SWelsME& sMe,
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) << 2);
pFeatureSearchIn->iMinQpelY = pFeatureSearchIn->iCurPixYQpel + ((pSlice->sMvStartMin.iMvY) << 2);
pFeatureSearchIn->iMaxQpelX = pFeatureSearchIn->iCurPixXQpel + ((pSlice->sMvStartMax.iMvX) << 2);
pFeatureSearchIn->iMaxQpelY = pFeatureSearchIn->iCurPixYQpel + ((pSlice->sMvStartMax.iMvY) << 2);
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) {

6
test/decoder/DecUT_DeblockCommon.cpp
View File

@ -122,7 +122,7 @@ void anchor_DeblockingLumaNormal (uint8_t* pPix, int32_t iStrideX, int32_t iStri
iTc++;
}
// 8-467,468,469
iDelta = WELS_CLIP3 (((((q[0] - p[0]) << 2) + (p[1] - q[1]) + 4) >> 3), -1 * iTc, iTc);
iDelta = WELS_CLIP3 (((((q[0] - p[0]) * (1 << 2)) + (p[1] - q[1]) + 4) >> 3), -1 * iTc, iTc);
pPix[iStrideX * -1] = WELS_CLIP3 ((p[0] + iDelta), 0, 255);
pPix[0] = WELS_CLIP3 ((q[0] - iDelta), 0, 255);
}
@ -194,7 +194,7 @@ void anchor_DeblockingChromaNormal (uint8_t* pPixCb, uint8_t* pPixCr, int32_t iS
// filterSampleFlag, 8-460
if (abs (p[0] - q[0]) < iAlpha && abs (p[1] - p[0]) < iBeta && abs (q[1] - q[0]) < iBeta) {
// 8-467, 468, 469
iDelta = WELS_CLIP3 (((((q[0] - p[0]) << 2) + (p[1] - q[1]) + 4) >> 3), -1 * iTc, iTc);
iDelta = WELS_CLIP3 (((((q[0] - p[0]) * (1 << 2)) + (p[1] - q[1]) + 4) >> 3), -1 * iTc, iTc);
pPixCb[iStrideX * -1] = WELS_CLIP3 ((p[0] + iDelta), 0, 255);
pPixCb[iStrideX * 0 ] = WELS_CLIP3 ((q[0] - iDelta), 0, 255);
}
@ -209,7 +209,7 @@ void anchor_DeblockingChromaNormal (uint8_t* pPixCb, uint8_t* pPixCr, int32_t iS
// filterSampleFlag, 8-460
if (abs (p[0] - q[0]) < iAlpha && abs (p[1] - p[0]) < iBeta && abs (q[1] - q[0]) < iBeta) {
// 8-467, 468, 469
iDelta = WELS_CLIP3 (((((q[0] - p[0]) << 2) + (p[1] - q[1]) + 4) >> 3), -1 * iTc, iTc);
iDelta = WELS_CLIP3 (((((q[0] - p[0]) * (1 << 2)) + (p[1] - q[1]) + 4) >> 3), -1 * iTc, iTc);
pPixCr[iStrideX * -1] = WELS_CLIP3 ((p[0] + iDelta), 0, 255);
pPixCr[iStrideX * 0 ] = WELS_CLIP3 ((q[0] - iDelta), 0, 255);
}