ref: 159ca4461fbb24c985976c331a222f1c73d91a15
parent: bee0d7d23030ab8c493640bb9fef3d73670f9320
parent: 43ca6db9f516ae6fd69df64ede749eb4f1ae248c
author: huili2 <huili2@cisco.com>
date: Fri Jul 3 07:10:05 EDT 2015
Merge pull request #2021 from sijchen/mt32 [Encoder] put pSliceBsBuffer to thread buffer rather than per slice, so as …
--- a/codec/encoder/core/inc/mt_defs.h
+++ b/codec/encoder/core/inc/mt_defs.h
@@ -87,6 +87,7 @@
FILE* pFSliceDiff; // file handle for debug
#endif//MT_DEBUG
+uint8_t* pThreadBsBuffer[MAX_THREADS_NUM]; //actual memory for slice buffer
} SSliceThreading;
#endif//MULTIPLE_THREADING_DEFINES_H__
--- a/codec/encoder/core/inc/slice_multi_threading.h
+++ b/codec/encoder/core/inc/slice_multi_threading.h
@@ -100,6 +100,7 @@
void TrackSliceConsumeTime (sWelsEncCtx* pCtx, int32_t* pDidList, const int32_t kiSpatialNum);
#endif//defined(MT_DEBUG)
+void SetOneSliceBsBufferUnderMultithread(sWelsEncCtx* pCtx, const int32_t kiThreadIdx, const int32_t iSliceIdx);
}
#endif//SVC_SLICE_MULTIPLE_THREADING_H__
--- a/codec/encoder/core/src/encoder_ext.cpp
+++ b/codec/encoder/core/src/encoder_ext.cpp
@@ -4061,9 +4061,9 @@
// pick up succeeding slice for threading
// thread_id equal to iEventId per implementation here
pCtx->pSliceThreading->pThreadPEncCtx[iEventId].iSliceIndex = iIndexOfSliceToBeCoded;
+ SetOneSliceBsBufferUnderMultithread(pCtx, iEventId, iIndexOfSliceToBeCoded);
WelsEventSignal (&pCtx->pSliceThreading->pReadySliceCodingEvent[iEventId]);
WelsEventSignal (&pCtx->pSliceThreading->pThreadMasterEvent[iEventId]);
-
++ iIndexOfSliceToBeCoded;
} else { // no other slices left for coding-- iNumThreadsRunning;
--- a/codec/encoder/core/src/slice_multi_threading.cpp
+++ b/codec/encoder/core/src/slice_multi_threading.cpp
@@ -289,20 +289,50 @@
}
}
+int32_t SetMultiSliceBuffer (sWelsEncCtx** ppCtx, CMemoryAlign* pMa, SSliceThreading* pSmt,
+ int32_t iMaxSliceNum, int32_t iSlice1Len, int32_t iSlice0Len, bool bDynamicSlice) {+ (*ppCtx)->pSliceBs = (SWelsSliceBs*)pMa->WelsMalloc (sizeof (SWelsSliceBs) * iMaxSliceNum, "pSliceBs");
+ if (NULL == (*ppCtx)->pSliceBs) {+ return ENC_RETURN_MEMALLOCERR;
+ }
+ if (iSlice0Len <= 0) {+ return ENC_RETURN_UNEXPECTED;
+ }
+ //slice 0
+ (*ppCtx)->pSliceBs[0].uiSize = iSlice0Len;
+ (*ppCtx)->pSliceBs[0].pBs = (*ppCtx)->pFrameBs;
+ (*ppCtx)->pSliceBs[0].uiBsPos = 0;
+ (*ppCtx)->pSliceBs[0].pBsBuffer = pSmt->pThreadBsBuffer[0];
+ if ((iMaxSliceNum == 1) && (!bDynamicSlice)) {+ return ENC_RETURN_SUCCESS;
+ }
+ //slice >0
+ if (iSlice1Len <= 0) {+ return ENC_RETURN_UNEXPECTED;
+ }
+ for (int32_t k = 1; k < iMaxSliceNum; k++) {+ (*ppCtx)->pSliceBs[k].uiSize = iSlice1Len;
+ (*ppCtx)->pSliceBs[k].pBs = (*ppCtx)->pSliceBs[k - 1].pBs + (*ppCtx)->pSliceBs[k - 1].uiSize;
+ }
+ if ((*ppCtx)->iFrameBsSize < (iSlice0Len + (iMaxSliceNum - 1)*iSlice1Len)) {+ return ENC_RETURN_MEMALLOCERR;
+ }
+ return ENC_RETURN_SUCCESS;
+}
+
int32_t RequestMtResource (sWelsEncCtx** ppCtx, SWelsSvcCodingParam* pCodingParam, const int32_t iCountBsLen,
const int32_t iTargetSpatialBsSize) {CMemoryAlign* pMa = NULL;
SWelsSvcCodingParam* pPara = NULL;
SSliceThreading* pSmt = NULL;
- SWelsSliceBs* pSliceB = NULL;
- uint8_t* pBsBase = NULL;
int32_t iNumSpatialLayers = 0;
int32_t iThreadNum = 0;
int32_t iIdx = 0;
- int32_t iSliceBsBufferSize = 0;
int16_t iMaxSliceNum = 1;
int32_t iReturn = ENC_RETURN_SUCCESS;
+ bool bDynamicSlice = false;
+ uint32_t uiMaxSliceSizeConstraint = 0;
if (NULL == ppCtx || NULL == pCodingParam || NULL == *ppCtx || iCountBsLen <= 0)
return 1;
@@ -342,6 +372,13 @@
pSmt->pSliceConsumeTime[iIdx] = NULL;
pSmt->pSliceComplexRatio[iIdx] = NULL;
}
+
+ if (pMso->uiSliceMode == SM_DYN_SLICE) {+ bDynamicSlice = true;
+ if (uiMaxSliceSizeConstraint < pMso->sSliceArgument.uiSliceSizeConstraint) {+ uiMaxSliceSizeConstraint = pMso->sSliceArgument.uiSliceSizeConstraint;
+ }
+ }
++ iIdx;
}
// NULL for pSliceConsumeTime[iIdx]: iIdx from iNumSpatialLayers to MAX_DEPENDENCY_LAYERS
@@ -391,38 +428,26 @@
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 (iTargetSpatialBsSize, "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;
+ }
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);
- (*ppCtx)->pSliceBs = (SWelsSliceBs*)pMa->WelsMalloc (sizeof (SWelsSliceBs) * iMaxSliceNum, "pSliceBs");
- WELS_VERIFY_RETURN_PROC_IF (1, (NULL == (*ppCtx)->pSliceBs), FreeMemorySvc (ppCtx))
+ iReturn = SetMultiSliceBuffer (ppCtx, pMa, pSmt, iMaxSliceNum,
+ iTargetSpatialBsSize, //TODO: may use uiMaxSliceSizeConstraint<<1 when bDynamicSlice, but need more twist
+ iCountBsLen,
+ bDynamicSlice);
+ WELS_VERIFY_RETURN_PROC_IF (iReturn, (ENC_RETURN_SUCCESS != iReturn), FreeMemorySvc (ppCtx))
- pBsBase = (*ppCtx)->pFrameBs + iCountBsLen;
- pSliceB = (*ppCtx)->pSliceBs;
- iSliceBsBufferSize = iTargetSpatialBsSize;
- iIdx = 0;
- while (iIdx < iMaxSliceNum) {- pSliceB->pBsBuffer = (uint8_t*)pMa->WelsMalloc (iSliceBsBufferSize, "pSliceB->pBsBuffer");
-
- WELS_VERIFY_RETURN_PROC_IF (1, (NULL == pSliceB->pBsBuffer), FreeMemorySvc (ppCtx))
- pSliceB->uiSize = iSliceBsBufferSize;
-
- if (iIdx > 0) {- pSliceB->pBs = pBsBase;
- pSliceB->uiBsPos = 0;
- pBsBase += iSliceBsBufferSize;
- } else {- pSliceB->pBs = NULL;
- pSliceB->uiBsPos = 0;
- }
- ++ pSliceB;
- ++ iIdx;
- }
-
iReturn = WelsMutexInit (&pSmt->mutexSliceNumUpdate);
WELS_VERIFY_RETURN_PROC_IF (1, (WELS_THREAD_ERROR_OK != iReturn), FreeMemorySvc (ppCtx))
@@ -486,14 +511,19 @@
pSmt->pThreadPEncCtx = NULL;
}
+ for (int i = 0; i < MAX_THREADS_NUM; i++) {+ if (pSmt->pThreadBsBuffer[i]) {+ pMa->WelsFree (pSmt->pThreadBsBuffer[i], "pSmt->pThreadBsBuffer");
+ pSmt->pThreadBsBuffer[i] = NULL;
+ }
+ }
+
pSliceB = (*ppCtx)->pSliceBs;
iIdx = 0;
while (pSliceB != NULL && iIdx < uiSliceNum) {- if (pSliceB->pBsBuffer) {- pMa->WelsFree (pSliceB->pBsBuffer, "pSliceB->pBsBuffer");
- pSliceB->pBsBuffer = NULL;
- pSliceB->uiSize = 0;
- }
+ pSliceB->pBsBuffer = NULL;
+ pSliceB->uiSize = 0;
+ pSliceB->uiBsPos = 0;
++ iIdx;
++ pSliceB;
}
@@ -863,6 +893,7 @@
iEventIdx);
}
+ SetOneSliceBsBufferUnderMultithread (pEncPEncCtx, kiPartitionId, iSliceIdx);
pSlice = &pCurDq->sLayerInfo.pSliceInLayer[iSliceIdx];
pSliceBs = &pEncPEncCtx->pSliceBs[iSliceIdx];
@@ -1028,6 +1059,7 @@
while (iIdx < kiEventCnt) {pPriData[iIdx].pLayerBs = pLbi;
pPriData[iIdx].iSliceIndex = iIdx;
+ SetOneSliceBsBufferUnderMultithread (pCtx, iIdx, iIdx);
if (pEventsList[iIdx])
WelsEventSignal (&pEventsList[iIdx]);
if (pMasterEventsList[iIdx])
@@ -1178,5 +1210,9 @@
}
#endif//#if defined(MT_DEBUG)
+void SetOneSliceBsBufferUnderMultithread (sWelsEncCtx* pCtx, const int32_t kiThreadIdx, const int32_t iSliceIdx) {+ pCtx->pSliceBs[iSliceIdx].pBsBuffer = pCtx->pSliceThreading->pThreadBsBuffer[kiThreadIdx];
+ pCtx->pSliceBs[iSliceIdx].uiBsPos = 0;
+}
}