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 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 (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]);
}
}
