int32_t NeedDynamicAdjust (SSlice* pSliceInLayer, const int32_t iSliceNum) {
if ( NULL == pSliceInLayer )
return false;
uint32_t uiTotalConsume = 0;
int32_t iSliceIdx = 0;
int32_t iNeedAdj = false;
WelsEmms();
while (iSliceIdx < iSliceNum) {
uiTotalConsume += pSliceInLayer[iSliceIdx].uiSliceConsumeTime;
iSliceIdx ++;
}
if (uiTotalConsume == 0) {
MT_TRACE_LOG (NULL, WELS_LOG_DEBUG,
"[MT] NeedDynamicAdjust(), herein do no adjust due first picture, iCountSliceNum= %d",
iSliceNum);
return false;
}
iSliceIdx = 0;
float fThr = EPSN; // threshold for various cores cases
float fRmse = .0f; // root mean square error of pSlice consume ratios
const float kfMeanRatio = 1.0f / iSliceNum;
do {
const float fRatio = 1.0f * pSliceInLayer[iSliceIdx].uiSliceConsumeTime / uiTotalConsume;
const float fDiffRatio = fRatio - kfMeanRatio;
fRmse += (fDiffRatio * fDiffRatio);
++ iSliceIdx;
} while (iSliceIdx + 1 < iSliceNum);
fRmse = sqrtf (fRmse / iSliceNum);
if (iSliceNum >= 8) {
fThr += THRESHOLD_RMSE_CORE8;
} else if (iSliceNum >= 4) {
fThr += THRESHOLD_RMSE_CORE4;
} else if (iSliceNum >= 2) {
fThr += THRESHOLD_RMSE_CORE2;
} else
fThr = 1.0f;
if (fRmse > fThr)
iNeedAdj = true;
MT_TRACE_LOG (NULL, WELS_LOG_DEBUG,
"[MT] NeedDynamicAdjust(), herein adjustment decision is made (iNeedAdj= %d) by: fRmse of pSlice complexity ratios %.6f, the corresponding threshold %.6f, iCountSliceNum %d",
iNeedAdj, fRmse, fThr, iSliceNum);
return iNeedAdj;
}
void CalcSliceComplexRatio (void* pRatio, SDqLayer* pCurDq, uint32_t* pSliceConsume) {
SSliceCtx* pSliceCtx = &pCurDq->sSliceEncCtx;
int32_t* pRatioList = (int32_t*)pRatio;
int32_t iAvI[MAX_SLICES_NUM];
int32_t iSumAv = 0;
uint32_t* pSliceTime = (uint32_t*)pSliceConsume;
int32_t* pCountMbInSlice = (int32_t*)pSliceCtx->pCountMbNumInSlice;
const int32_t kiSliceCount = pSliceCtx->iSliceNumInFrame;
int32_t iSliceIdx = 0;
WelsEmms();
while (iSliceIdx < kiSliceCount) {
iAvI[iSliceIdx] = WELS_DIV_ROUND (INT_MULTIPLY * pCountMbInSlice[iSliceIdx], pSliceTime[iSliceIdx]);
MT_TRACE_LOG (NULL, WELS_LOG_DEBUG, "[MT] CalcSliceComplexRatio(), pSliceConsumeTime[%d]= %d us, slice_run= %d",
iSliceIdx,
pSliceTime[iSliceIdx], pCountMbInSlice[iSliceIdx]);
iSumAv += iAvI[iSliceIdx];
++ iSliceIdx;
}
while (-- iSliceIdx >= 0) {
pRatioList[iSliceIdx] = WELS_DIV_ROUND (INT_MULTIPLY * iAvI[iSliceIdx], iSumAv);
}
}
int32_t CreateSliceThreads (sWelsEncCtx* pCtx) {
const int32_t kiThreadCount = pCtx->pSvcParam->iCountThreadsNum;
int32_t iIdx = 0;
while (iIdx < kiThreadCount) {
if (WelsThreadCreate (&pCtx->pSliceThreading->pThreadHandles[iIdx], CodingSliceThreadProc,
&pCtx->pSliceThreading->pThreadPEncCtx[iIdx], 0)) {
return 1;
}
++ iIdx;
}
MT_TRACE_LOG (pCtx, WELS_LOG_INFO, "CreateSliceThreads() exit..");
return 0;
}
void CalcSliceComplexRatio (SDqLayer* pCurDq) {
SSliceCtx* pSliceCtx = &pCurDq->sSliceEncCtx;
SSlice* pSliceInLayer = pCurDq->sLayerInfo.pSliceInLayer;
int32_t iSumAv = 0;
const int32_t kiSliceCount = pSliceCtx->iSliceNumInFrame;
int32_t iSliceIdx = 0;
int32_t iAvI[MAX_SLICES_NUM];
WelsEmms();
while (iSliceIdx < kiSliceCount) {
iAvI[iSliceIdx] = WELS_DIV_ROUND (INT_MULTIPLY * pSliceInLayer[iSliceIdx].iCountMbNumInSlice,
pSliceInLayer[iSliceIdx].uiSliceConsumeTime);
MT_TRACE_LOG (NULL, WELS_LOG_DEBUG, "[MT] CalcSliceComplexRatio(), uiSliceConsumeTime[%d]= %d us, slice_run= %d",
iSliceIdx,
pSliceInLayer[iSliceIdx].uiSliceConsumeTime, pSliceInLayer[iSliceIdx].iCountMbNumInSlice);
iSumAv += iAvI[iSliceIdx];
++ iSliceIdx;
}
while (-- iSliceIdx >= 0) {
pSliceInLayer[iSliceIdx].iSliceComplexRatio = WELS_DIV_ROUND (INT_MULTIPLY * iAvI[iSliceIdx], iSumAv);
}
}
void DynamicAdjustSlicing (sWelsEncCtx* pCtx,
SDqLayer* pCurDqLayer,
void* pComplexRatio,
int32_t iCurDid) {
SSliceCtx* pSliceCtx = pCurDqLayer->pSliceEncCtx;
const int32_t kiCountSliceNum = pSliceCtx->iSliceNumInFrame;
const int32_t kiCountNumMb = pSliceCtx->iMbNumInFrame;
int32_t iMinimalMbNum = pSliceCtx->iMbWidth; // in theory we need only 1 SMB, here let it as one SMB row required
int32_t iMaximalMbNum = 0; // dynamically assign later
int32_t* pSliceComplexRatio = (int32_t*)pComplexRatio;
int32_t iMbNumLeft = kiCountNumMb;
int32_t iRunLen[MAX_THREADS_NUM] = {0};
int32_t iSliceIdx = 0;
int32_t iNumMbInEachGom = 0;
SWelsSvcRc* pWelsSvcRc = &pCtx->pWelsSvcRc[iCurDid];
if (pCtx->pSvcParam->iRCMode != RC_OFF_MODE) {
iNumMbInEachGom = pWelsSvcRc->iNumberMbGom;
if (iNumMbInEachGom <= 0) {
WelsLog (& (pCtx->sLogCtx), WELS_LOG_ERROR,
"[MT] DynamicAdjustSlicing(), invalid iNumMbInEachGom= %d from RC, iDid= %d, iCountNumMb= %d", iNumMbInEachGom,
iCurDid, kiCountNumMb);
return;
}
// do not adjust in case no extra iNumMbInEachGom based left for slicing adjustment,
// extra MB of non integrated GOM assigned at the last pSlice in default, keep up on early initial result.
if (iNumMbInEachGom * kiCountSliceNum >= kiCountNumMb) {
return;
}
iMinimalMbNum = iNumMbInEachGom;
}
if (kiCountSliceNum < 2 || (kiCountSliceNum & 0x01)) // we need suppose uiSliceNum is even for multiple threading
return;
iMaximalMbNum = kiCountNumMb - (kiCountSliceNum - 1) * iMinimalMbNum;
WelsEmms();
MT_TRACE_LOG (pCtx, WELS_LOG_DEBUG, "[MT] DynamicAdjustSlicing(), iDid= %d, iCountNumMb= %d", iCurDid, kiCountNumMb);
iSliceIdx = 0;
while (iSliceIdx + 1 < kiCountSliceNum) {
int32_t iNumMbAssigning = WELS_DIV_ROUND (kiCountNumMb * pSliceComplexRatio[iSliceIdx], INT_MULTIPLY);
// GOM boundary aligned
if (pCtx->pSvcParam->iRCMode != RC_OFF_MODE) {
iNumMbAssigning = iNumMbAssigning / iNumMbInEachGom * iNumMbInEachGom;
}
// make sure one GOM at least in each pSlice for safe
if (iNumMbAssigning < iMinimalMbNum)
iNumMbAssigning = iMinimalMbNum;
else if (iNumMbAssigning > iMaximalMbNum)
iNumMbAssigning = iMaximalMbNum;
assert (iNumMbAssigning > 0);
iMbNumLeft -= iNumMbAssigning;
if (iMbNumLeft <= 0) { // error due to we can not support slice_skip now yet, do not adjust this time
assert (0);
return;
}
iRunLen[iSliceIdx] = iNumMbAssigning;
MT_TRACE_LOG (pCtx, WELS_LOG_DEBUG,
"[MT] DynamicAdjustSlicing(), uiSliceIdx= %d, pSliceComplexRatio= %.2f, slice_run_org= %d, slice_run_adj= %d",
iSliceIdx, pSliceComplexRatio[iSliceIdx] * 1.0f / INT_MULTIPLY, pSliceCtx->pCountMbNumInSlice[iSliceIdx],
iNumMbAssigning);
++ iSliceIdx;
iMaximalMbNum = iMbNumLeft - (kiCountSliceNum - iSliceIdx - 1) * iMinimalMbNum; // get maximal num_mb in left parts
}
iRunLen[iSliceIdx] = iMbNumLeft;
MT_TRACE_LOG (pCtx, WELS_LOG_DEBUG,
"[MT] DynamicAdjustSlicing(), iSliceIdx= %d, pSliceComplexRatio= %.2f, slice_run_org= %d, slice_run_adj= %d",
iSliceIdx, pSliceComplexRatio[iSliceIdx] * 1.0f / INT_MULTIPLY, pSliceCtx->pCountMbNumInSlice[iSliceIdx], iMbNumLeft);
if (DynamicAdjustSlicePEncCtxAll (pSliceCtx, iRunLen) == 0) {
const int32_t kiThreadNum = pCtx->pSvcParam->iCountThreadsNum;
int32_t iThreadIdx = 0;
do {
WelsEventSignal (&pCtx->pSliceThreading->pUpdateMbListEvent[iThreadIdx]);
WelsEventSignal (&pCtx->pSliceThreading->pThreadMasterEvent[iThreadIdx]);
++ iThreadIdx;
} while (iThreadIdx < kiThreadNum);
WelsMultipleEventsWaitAllBlocking (kiThreadNum, &pCtx->pSliceThreading->pFinUpdateMbListEvent[0]);
}
}
// thread process for coding one pSlice
WELS_THREAD_ROUTINE_TYPE CodingSliceThreadProc (void* arg) {
SSliceThreadPrivateData* pPrivateData = (SSliceThreadPrivateData*)arg;
sWelsEncCtx* pEncPEncCtx = NULL;
SDqLayer* pCurDq = NULL;
SSlice* pSlice = NULL;
SWelsSliceBs* pSliceBs = NULL;
WELS_EVENT pEventsList[3];
int32_t iEventCount = 0;
WELS_THREAD_ERROR_CODE iWaitRet = WELS_THREAD_ERROR_GENERAL;
uint32_t uiThrdRet = 0;
int32_t iSliceSize = 0;
int32_t iSliceIdx = -1;
int32_t iThreadIdx = -1;
int32_t iEventIdx = -1;
bool bNeedPrefix = false;
EWelsNalUnitType eNalType = NAL_UNIT_UNSPEC_0;
EWelsNalRefIdc eNalRefIdc = NRI_PRI_LOWEST;
int32_t iReturn = ENC_RETURN_SUCCESS;
if (NULL == pPrivateData)
WELS_THREAD_ROUTINE_RETURN (1);
pEncPEncCtx = (sWelsEncCtx*)pPrivateData->pWelsPEncCtx;
iThreadIdx = pPrivateData->iThreadIndex;
iEventIdx = iThreadIdx;
pEventsList[iEventCount++] = pEncPEncCtx->pSliceThreading->pReadySliceCodingEvent[iEventIdx];
pEventsList[iEventCount++] = pEncPEncCtx->pSliceThreading->pExitEncodeEvent[iEventIdx];
pEventsList[iEventCount++] = pEncPEncCtx->pSliceThreading->pUpdateMbListEvent[iEventIdx];
WelsThreadSetName ("OpenH264Enc_CodingSliceThreadProc");
do {
MT_TRACE_LOG (pEncPEncCtx, WELS_LOG_INFO,
"[MT] CodingSliceThreadProc(), try to call WelsMultipleEventsWaitSingleBlocking(pEventsList= %p %p %p), pEncPEncCtx= %p!",
pEventsList[0], pEventsList[1], pEventsList[1], (void*)pEncPEncCtx);
iWaitRet = WelsMultipleEventsWaitSingleBlocking (iEventCount,
&pEventsList[0],
&pEncPEncCtx->pSliceThreading->pThreadMasterEvent[iEventIdx]); // blocking until at least one event is signalled
if (WELS_THREAD_ERROR_WAIT_OBJECT_0 == iWaitRet) { // start pSlice coding signal waited
//int iLayerIndex = pEncPEncCtx->pOut->iLayerBsIndex;
//SFrameBSInfo* pFrameBsInfo = pPrivateData->pFrameBsInfo;
//SLayerBSInfo* pLbi = &pFrameBsInfo->sLayerInfo [iLayerIndex];
const int32_t kiCurDid = pEncPEncCtx->uiDependencyId;
SWelsSvcCodingParam* pCodingParam = pEncPEncCtx->pSvcParam;
SSpatialLayerConfig* pParamD = &pCodingParam->sSpatialLayers[kiCurDid];
pCurDq = pEncPEncCtx->pCurDqLayer;
eNalType = pEncPEncCtx->eNalType;
eNalRefIdc = pEncPEncCtx->eNalPriority;
bNeedPrefix = pEncPEncCtx->bNeedPrefixNalFlag;
if (pParamD->sSliceArgument.uiSliceMode != SM_SIZELIMITED_SLICE) {
int64_t iSliceStart = 0;
bool bDsaFlag = false;
iSliceIdx = pPrivateData->iSliceIndex;
pSlice = &pCurDq->sLayerInfo.pSliceInLayer[iSliceIdx];
pSliceBs = &pEncPEncCtx->pSliceBs[iSliceIdx];
bDsaFlag = ((pParamD->sSliceArgument.uiSliceMode == SM_FIXEDSLCNUM_SLICE) &&
pCodingParam->iMultipleThreadIdc > 1 &&
pCodingParam->iMultipleThreadIdc >= pParamD->sSliceArgument.uiSliceNum);
if (bDsaFlag)
iSliceStart = WelsTime();
pSliceBs->uiBsPos = 0;
pSliceBs->iNalIndex = 0;
assert ((void*) (&pSliceBs->sBsWrite) == (void*)pSlice->pSliceBsa);
InitBits (&pSliceBs->sBsWrite, pSliceBs->pBsBuffer, pSliceBs->uiSize);
#if MT_DEBUG_BS_WR
pSliceBs->bSliceCodedFlag = false;
#endif//MT_DEBUG_BS_WR
if (bNeedPrefix) {
if (eNalRefIdc != NRI_PRI_LOWEST) {
WelsLoadNalForSlice (pSliceBs, NAL_UNIT_PREFIX, eNalRefIdc);
WelsWriteSVCPrefixNal (&pSliceBs->sBsWrite, eNalRefIdc, (NAL_UNIT_CODED_SLICE_IDR == eNalType));
WelsUnloadNalForSlice (pSliceBs);
} else { // No Prefix NAL Unit RBSP syntax here, but need add NAL Unit Header extension
WelsLoadNalForSlice (pSliceBs, NAL_UNIT_PREFIX, eNalRefIdc);
// No need write any syntax of prefix NAL Unit RBSP here
WelsUnloadNalForSlice (pSliceBs);
}
}
WelsLoadNalForSlice (pSliceBs, eNalType, eNalRefIdc);
iReturn = WelsCodeOneSlice (pEncPEncCtx, iSliceIdx, eNalType);
if (ENC_RETURN_SUCCESS != iReturn) {
uiThrdRet = iReturn;
WELS_THREAD_SIGNAL_AND_BREAK (pEncPEncCtx->pSliceThreading->pSliceCodedEvent,
pEncPEncCtx->pSliceThreading->pSliceCodedMasterEvent,
iEventIdx);
}
WelsUnloadNalForSlice (pSliceBs);
int32_t iLeftBufferSize = (iSliceIdx > 0) ?
(pSliceBs->uiSize - pSliceBs->uiBsPos)
: (pEncPEncCtx->iFrameBsSize - pEncPEncCtx->iPosBsBuffer);
iReturn = WriteSliceBs (pEncPEncCtx, pSliceBs->pBs,
&pSliceBs->iNalLen[0],
iLeftBufferSize,
iSliceIdx, iSliceSize);
if (ENC_RETURN_SUCCESS != iReturn) {
uiThrdRet = iReturn;
WELS_THREAD_SIGNAL_AND_BREAK (pEncPEncCtx->pSliceThreading->pSliceCodedEvent,
pEncPEncCtx->pSliceThreading->pSliceCodedMasterEvent,
iEventIdx);
}
if (0 == iSliceIdx) {
pEncPEncCtx->iPosBsBuffer += iSliceSize;
}
pEncPEncCtx->pFuncList->pfDeblocking.pfDeblockingFilterSlice (pCurDq, pEncPEncCtx->pFuncList, iSliceIdx);
if (bDsaFlag) {
pEncPEncCtx->pCurDqLayer->sLayerInfo.pSliceInLayer[iSliceIdx].uiSliceConsumeTime = (uint32_t) (
WelsTime() - iSliceStart);
MT_TRACE_LOG (& (pEncPEncCtx->sLogCtx), WELS_LOG_INFO,
"[MT] CodingSliceThreadProc(), coding_idx %d, uiSliceIdx %d, uiSliceConsumeTime %d, iSliceSize %d, iFirstMbInSlice %d, count_num_mb_in_slice %d",
pEncPEncCtx->iCodingIndex, iSliceIdx,
pEncPEncCtx->pCurDqLayer->sLayerInfo.pSliceInLayer[iSliceIdx].uiSliceConsumeTime, iSliceSize,
pCurDq->sLayerInfo.pSliceInLayer[iSliceIdx].sSliceHeaderExt.sSliceHeader.iFirstMbInSlice,
pCurDq->sLayerInfo.pSliceInLayer[iSliceIdx].iCountMbNumInSlice);
}
#if defined(SLICE_INFO_OUTPUT)
fprintf (stderr,
"@pSlice=%-6d sliceType:%c idc:%d size:%-6d\n",
iSliceIdx,
(pEncPEncCtx->eSliceType == P_SLICE ? 'P' : 'I'),
eNalRefIdc,
iSliceSize
);
#endif//SLICE_INFO_OUTPUT
#if MT_DEBUG_BS_WR
pSliceBs->bSliceCodedFlag = true;
#endif//MT_DEBUG_BS_WR
WelsEventSignal (
&pEncPEncCtx->pSliceThreading->pSliceCodedEvent[iEventIdx]); // mean finished coding current pSlice
WelsEventSignal (
&pEncPEncCtx->pSliceThreading->pSliceCodedMasterEvent);
} else { // for SM_SIZELIMITED_SLICE parallelization
SSliceCtx* pSliceCtx = &pCurDq->sSliceEncCtx;
const int32_t kiPartitionId = iThreadIdx;
const int32_t kiSliceIdxStep = pEncPEncCtx->iActiveThreadsNum;
const int32_t kiFirstMbInPartition = pPrivateData->iStartMbIndex; // inclusive
const int32_t kiEndMbInPartition = pPrivateData->iEndMbIndex; // exclusive
int32_t iAnyMbLeftInPartition = kiEndMbInPartition - kiFirstMbInPartition;
iSliceIdx = pPrivateData->iSliceIndex;
SSliceHeaderExt* pStartSliceHeaderExt = &pCurDq->sLayerInfo.pSliceInLayer[iSliceIdx].sSliceHeaderExt;
pStartSliceHeaderExt->sSliceHeader.iFirstMbInSlice = kiFirstMbInPartition;
pCurDq->pNumSliceCodedOfPartition[kiPartitionId] =
1; // one pSlice per partition intialized, dynamic slicing inside
pCurDq->pLastMbIdxOfPartition[kiPartitionId] = kiEndMbInPartition - 1;
pCurDq->pLastCodedMbIdxOfPartition[kiPartitionId] = 0;
while (iAnyMbLeftInPartition > 0) {
if (iSliceIdx >= pSliceCtx->iMaxSliceNumConstraint) {
// TODO: need exception handler for not large enough of MAX_SLICES_NUM related memory usage
// No idea about its solution due MAX_SLICES_NUM is fixed lenght in relevent pData structure
uiThrdRet = 1;
WELS_THREAD_SIGNAL_AND_BREAK (pEncPEncCtx->pSliceThreading->pSliceCodedEvent,
pEncPEncCtx->pSliceThreading->pSliceCodedMasterEvent,
iEventIdx);
}
SetOneSliceBsBufferUnderMultithread (pEncPEncCtx, kiPartitionId, iSliceIdx);
pSlice = &pCurDq->sLayerInfo.pSliceInLayer[iSliceIdx];
pSliceBs = &pEncPEncCtx->pSliceBs[iSliceIdx];
pSliceBs->uiBsPos = 0;
pSliceBs->iNalIndex = 0;
InitBits (&pSliceBs->sBsWrite, pSliceBs->pBsBuffer, pSliceBs->uiSize);
if (bNeedPrefix) {
if (eNalRefIdc != NRI_PRI_LOWEST) {
WelsLoadNalForSlice (pSliceBs, NAL_UNIT_PREFIX, eNalRefIdc);
WelsWriteSVCPrefixNal (&pSliceBs->sBsWrite, eNalRefIdc, (NAL_UNIT_CODED_SLICE_IDR == eNalType));
WelsUnloadNalForSlice (pSliceBs);
} else { // No Prefix NAL Unit RBSP syntax here, but need add NAL Unit Header extension
WelsLoadNalForSlice (pSliceBs, NAL_UNIT_PREFIX, eNalRefIdc);
// No need write any syntax of prefix NAL Unit RBSP here
WelsUnloadNalForSlice (pSliceBs);
}
}
WelsLoadNalForSlice (pSliceBs, eNalType, eNalRefIdc);
iReturn = WelsCodeOneSlice (pEncPEncCtx, iSliceIdx, eNalType);
if (ENC_RETURN_SUCCESS != iReturn) {
uiThrdRet = iReturn;
WELS_THREAD_SIGNAL_AND_BREAK (pEncPEncCtx->pSliceThreading->pSliceCodedEvent,
pEncPEncCtx->pSliceThreading->pSliceCodedMasterEvent,
iEventIdx);
}
WelsUnloadNalForSlice (pSliceBs);
int32_t iLeftBufferSize = (iSliceIdx > 0) ? (pSliceBs->uiSize - pSliceBs->uiBsPos) : (pEncPEncCtx->iFrameBsSize - pEncPEncCtx->iPosBsBuffer);
iReturn = WriteSliceBs (pEncPEncCtx, pSliceBs->pBs, &pSliceBs->iNalLen[0],
iLeftBufferSize,
iSliceIdx, iSliceSize);
if (ENC_RETURN_SUCCESS != iReturn) {
uiThrdRet = iReturn;
WELS_THREAD_SIGNAL_AND_BREAK (pEncPEncCtx->pSliceThreading->pSliceCodedEvent,
pEncPEncCtx->pSliceThreading->pSliceCodedMasterEvent,
iEventIdx);
}
if (0 == iSliceIdx) {
pEncPEncCtx->iPosBsBuffer += iSliceSize;
}
pEncPEncCtx->pFuncList->pfDeblocking.pfDeblockingFilterSlice (pCurDq, pEncPEncCtx->pFuncList, iSliceIdx);
#if defined(SLICE_INFO_OUTPUT)
fprintf (stderr,
"@pSlice=%-6d sliceType:%c idc:%d size:%-6d\n",
iSliceIdx,
(pEncPEncCtx->eSliceType == P_SLICE ? 'P' : 'I'),
eNalRefIdc,
iSliceSize
);
#endif//SLICE_INFO_OUTPUT
MT_TRACE_LOG (pEncPEncCtx, WELS_LOG_INFO,
"[MT] CodingSliceThreadProc(), coding_idx %d, iPartitionId %d, uiSliceIdx %d, iSliceSize %d, count_mb_slice %d, iEndMbInPartition %d, pCurDq->pLastCodedMbIdxOfPartition[%d] %d\n",
pEncPEncCtx->iCodingIndex, kiPartitionId, iSliceIdx, iSliceSize,
pCurDq->sLayerInfo.pSliceInLayer[iSliceIdx].iCountMbNumInSlice,
kiEndMbInPartition, kiPartitionId, pCurDq->pLastCodedMbIdxOfPartition[kiPartitionId]);
iAnyMbLeftInPartition = kiEndMbInPartition - (1 + pCurDq->pLastCodedMbIdxOfPartition[kiPartitionId]);
iSliceIdx += kiSliceIdxStep;
}
if (uiThrdRet) { // any exception??
WELS_THREAD_SIGNAL_AND_BREAK (pEncPEncCtx->pSliceThreading->pSliceCodedEvent,
pEncPEncCtx->pSliceThreading->pSliceCodedMasterEvent,
iEventIdx);
}
WelsEventSignal (&pEncPEncCtx->pSliceThreading->pSliceCodedEvent[iEventIdx]); // mean finished coding current pSlice
WelsEventSignal (&pEncPEncCtx->pSliceThreading->pSliceCodedMasterEvent);
}
} else if (WELS_THREAD_ERROR_WAIT_OBJECT_0 + 1 == iWaitRet) { // exit thread signal
uiThrdRet = 0;
break;
} else if (WELS_THREAD_ERROR_WAIT_OBJECT_0 + 2 == iWaitRet) { // update pMb list singal
iSliceIdx =
iEventIdx; // pPrivateData->iSliceIndex; old threads can not be terminated, pPrivateData is not correct for applicable
pCurDq = pEncPEncCtx->pCurDqLayer;
UpdateMbListNeighborParallel (pCurDq, pCurDq->sMbDataP, iSliceIdx);
WelsEventSignal (
&pEncPEncCtx->pSliceThreading->pFinUpdateMbListEvent[iEventIdx]); // mean finished update pMb list for this pSlice
} else { // WELS_THREAD_ERROR_WAIT_TIMEOUT, or WELS_THREAD_ERROR_WAIT_FAILED
WelsLog (& (pEncPEncCtx->sLogCtx), WELS_LOG_WARNING,
"[MT] CodingSliceThreadProc(), waiting pReadySliceCodingEvent[%d] failed(%d) and thread%d terminated!", iEventIdx,
iWaitRet, iThreadIdx);
uiThrdRet = 1;
break;
}
} while (1);
//sync multi-threading error
WelsMutexLock (&pEncPEncCtx->mutexEncoderError);
if (uiThrdRet) pEncPEncCtx->iEncoderError |= uiThrdRet;
WelsMutexUnlock (&pEncPEncCtx->mutexEncoderError);
WELS_THREAD_ROUTINE_RETURN (uiThrdRet);
}
int32_t RequestMtResource (sWelsEncCtx** ppCtx, SWelsSvcCodingParam* pCodingParam, const int32_t iCountBsLen,
const int32_t iMaxSliceBufferSize, bool bDynamicSlice) {
CMemoryAlign* pMa = NULL;
SWelsSvcCodingParam* pPara = NULL;
SSliceThreading* pSmt = NULL;
int32_t iNumSpatialLayers = 0;
int32_t iThreadNum = 0;
int32_t iIdx = 0;
int16_t iMaxSliceNum = 1;
int32_t iReturn = ENC_RETURN_SUCCESS;
bool bWillUseTaskManage = false;
if (NULL == ppCtx || NULL == pCodingParam || NULL == *ppCtx || iCountBsLen <= 0)
return 1;
pMa = (*ppCtx)->pMemAlign;
pPara = pCodingParam;
iNumSpatialLayers = pPara->iSpatialLayerNum;
iThreadNum = pPara->iCountThreadsNum;
iMaxSliceNum = (*ppCtx)->iMaxSliceCount;
pSmt = (SSliceThreading*)pMa->WelsMalloc (sizeof (SSliceThreading), "SSliceThreading");
WELS_VERIFY_RETURN_PROC_IF (1, (NULL == pSmt), FreeMemorySvc (ppCtx))
(*ppCtx)->pSliceThreading = pSmt;
pSmt->pThreadPEncCtx = (SSliceThreadPrivateData*)pMa->WelsMalloc (sizeof (SSliceThreadPrivateData) * iThreadNum,
"pThreadPEncCtx");
WELS_VERIFY_RETURN_PROC_IF (1, (NULL == pSmt->pThreadPEncCtx), FreeMemorySvc (ppCtx))
#ifdef _WIN32
// Dummy event namespace, the windows events don't actually use this
WelsSnprintf (pSmt->eventNamespace, sizeof (pSmt->eventNamespace), "%p", (void*) *ppCtx);
#else
WelsSnprintf (pSmt->eventNamespace, sizeof (pSmt->eventNamespace), "%p%x", (void*) *ppCtx, getpid());
#endif//!_WIN32
iIdx = 0;
while (iIdx < iNumSpatialLayers) {
SSliceArgument* pSliceArgument = &pPara->sSpatialLayers[iIdx].sSliceArgument;
if (pSliceArgument->uiSliceMode == SM_FIXEDSLCNUM_SLICE || pSliceArgument->uiSliceMode == SM_RASTER_SLICE) {
bWillUseTaskManage = true;
}
++ iIdx;
}
#ifdef MT_DEBUG
// file handle for MT debug
pSmt->pFSliceDiff = NULL;
if (pSmt->pFSliceDiff) {
fclose (pSmt->pFSliceDiff);
pSmt->pFSliceDiff = NULL;
}
pSmt->pFSliceDiff = fopen ("slice_time.txt", "wt+");
#endif//MT_DEBUG
MT_TRACE_LOG (*ppCtx, WELS_LOG_INFO, "encpEncCtx= 0x%p", (void*) *ppCtx);
char name[SEM_NAME_MAX] = {0};
WELS_GCC_UNUSED WELS_THREAD_ERROR_CODE err = 0;
iIdx = 0;
while (iIdx < iThreadNum) {
pSmt->pThreadPEncCtx[iIdx].pWelsPEncCtx = (void*) *ppCtx;
pSmt->pThreadPEncCtx[iIdx].iSliceIndex = iIdx;
pSmt->pThreadPEncCtx[iIdx].iThreadIndex = iIdx;
pSmt->pThreadHandles[iIdx] = 0;
WelsSnprintf (name, SEM_NAME_MAX, "ee%d%s", iIdx, pSmt->eventNamespace);
err = WelsEventOpen (&pSmt->pExitEncodeEvent[iIdx], name);
MT_TRACE_LOG (*ppCtx, WELS_LOG_INFO, "[MT] Open pExitEncodeEvent%d named(%s) ret%d err%d", iIdx, name, err, errno);
WelsSnprintf (name, SEM_NAME_MAX, "tm%d%s", iIdx, pSmt->eventNamespace);
err = WelsEventOpen (&pSmt->pThreadMasterEvent[iIdx], name);
MT_TRACE_LOG (*ppCtx, WELS_LOG_INFO, "[MT] Open pThreadMasterEvent%d named(%s) ret%d err%d", iIdx, name, err, errno);
// length of semaphore name should be system constrained at least on mac 10.7
WelsSnprintf (name, SEM_NAME_MAX, "ud%d%s", iIdx, pSmt->eventNamespace);
err = WelsEventOpen (&pSmt->pUpdateMbListEvent[iIdx], name);
MT_TRACE_LOG (*ppCtx, WELS_LOG_INFO, "[MT] Open pUpdateMbListEvent%d named(%s) ret%d err%d", iIdx, name, err, errno);
WelsSnprintf (name, SEM_NAME_MAX, "fu%d%s", iIdx, pSmt->eventNamespace);
err = WelsEventOpen (&pSmt->pFinUpdateMbListEvent[iIdx], name);
MT_TRACE_LOG (*ppCtx, WELS_LOG_INFO, "[MT] Open pFinUpdateMbListEvent%d named(%s) ret%d err%d", iIdx, name, err,
errno);
WelsSnprintf (name, SEM_NAME_MAX, "sc%d%s", iIdx, pSmt->eventNamespace);
err = WelsEventOpen (&pSmt->pSliceCodedEvent[iIdx], name);
MT_TRACE_LOG (*ppCtx, WELS_LOG_INFO, "[MT] Open pSliceCodedEvent%d named(%s) ret%d err%d", iIdx, name, err, errno);
WelsSnprintf (name, SEM_NAME_MAX, "rc%d%s", iIdx, pSmt->eventNamespace);
err = WelsEventOpen (&pSmt->pReadySliceCodingEvent[iIdx], name);
MT_TRACE_LOG (*ppCtx, WELS_LOG_INFO, "[MT] Open pReadySliceCodingEvent%d = 0x%p named(%s) ret%d err%d", iIdx,
(void*)pSmt->pReadySliceCodingEvent[iIdx], name, err, errno);
pSmt->pThreadBsBuffer[iIdx] = (uint8_t*)pMa->WelsMalloc (iCountBsLen, "pSmt->pThreadBsBuffer");
WELS_VERIFY_RETURN_PROC_IF (1, (NULL == pSmt->pThreadBsBuffer[iIdx]), FreeMemorySvc (ppCtx))
++ iIdx;
}
for (; iIdx < MAX_THREADS_NUM; iIdx++) {
pSmt->pThreadBsBuffer[iIdx] = NULL;
}
//previous conflict
WelsSnprintf (name, SEM_NAME_MAX, "scm%s", pSmt->eventNamespace);
err = WelsEventOpen (&pSmt->pSliceCodedMasterEvent, name);
MT_TRACE_LOG (*ppCtx, WELS_LOG_INFO, "[MT] Open pSliceCodedMasterEvent named(%s) ret%d err%d", name, err, errno);
//previous conflict ends
iReturn = SetMultiSliceBuffer (ppCtx, pMa, pSmt, iMaxSliceNum,
iMaxSliceBufferSize,
iCountBsLen,
bDynamicSlice);
WELS_VERIFY_RETURN_PROC_IF (iReturn, (ENC_RETURN_SUCCESS != iReturn), FreeMemorySvc (ppCtx))
iReturn = WelsMutexInit (&pSmt->mutexSliceNumUpdate);
WELS_VERIFY_RETURN_PROC_IF (1, (WELS_THREAD_ERROR_OK != iReturn), FreeMemorySvc (ppCtx))
if (bWillUseTaskManage) {
(*ppCtx)->pTaskManage = IWelsTaskManage::CreateTaskManage (*ppCtx, iNumSpatialLayers, bDynamicSlice);
WELS_VERIFY_RETURN_PROC_IF (iReturn, (NULL == (*ppCtx)->pTaskManage), FreeMemorySvc (ppCtx))
}
memset (&pSmt->bThreadBsBufferUsage, 0, MAX_THREADS_NUM * sizeof (bool));
iReturn = WelsMutexInit (&pSmt->mutexThreadBsBufferUsage);
WELS_VERIFY_RETURN_PROC_IF (1, (WELS_THREAD_ERROR_OK != iReturn), FreeMemorySvc (ppCtx))
iReturn = WelsMutexInit (& (*ppCtx)->mutexEncoderError);
WELS_VERIFY_RETURN_PROC_IF (1, (WELS_THREAD_ERROR_OK != iReturn), FreeMemorySvc (ppCtx))
MT_TRACE_LOG (*ppCtx, WELS_LOG_INFO, "RequestMtResource(), iThreadNum=%d, iCountSliceNum= %d",
pPara->iCountThreadsNum,
iMaxSliceNum);
return 0;
}