void InitDecFuncs (PWelsDecoderContext pCtx, uint32_t uiCpuFlag) {
WelsBlockFuncInit (&pCtx->sBlockFunc, uiCpuFlag);
InitPredFunc (pCtx, uiCpuFlag);
InitMcFunc (& (pCtx->sMcFunc), uiCpuFlag);
InitExpandPictureFunc (& (pCtx->sExpandPicFunc), uiCpuFlag);
DeblockingInit (&pCtx->sDeblockingFunc, uiCpuFlag);
}
/*!
* \brief initialize function pointers that potentially used in Wels encoding
* \param pEncCtx sWelsEncCtx*
* \return successful - 0; otherwise none 0 for failed
*/
int32_t InitFunctionPointers (SWelsFuncPtrList* pFuncList, SWelsSvcCodingParam* pParam, uint32_t uiCpuFlag) {
int32_t iReturn = ENC_RETURN_SUCCESS;
/* Functionality utilization of CPU instructions dependency */
pFuncList->pfSetMemZeroSize8 = WelsSetMemZero_c; // confirmed_safe_unsafe_usage
pFuncList->pfSetMemZeroSize64Aligned16 = WelsSetMemZero_c; // confirmed_safe_unsafe_usage
pFuncList->pfSetMemZeroSize64 = WelsSetMemZero_c; // confirmed_safe_unsafe_usage
#if defined(X86_ASM)
if (uiCpuFlag & WELS_CPU_MMXEXT) {
pFuncList->pfSetMemZeroSize8 = WelsSetMemZeroSize8_mmx; // confirmed_safe_unsafe_usage
pFuncList->pfSetMemZeroSize64Aligned16 = WelsSetMemZeroSize64_mmx; // confirmed_safe_unsafe_usage
pFuncList->pfSetMemZeroSize64 = WelsSetMemZeroSize64_mmx; // confirmed_safe_unsafe_usage
}
if (uiCpuFlag & WELS_CPU_SSE2) {
pFuncList->pfSetMemZeroSize64Aligned16 = WelsSetMemZeroAligned64_sse2; // confirmed_safe_unsafe_usage
}
#endif//X86_ASM
#if defined(HAVE_NEON)
if (uiCpuFlag & WELS_CPU_NEON) {
pFuncList->pfSetMemZeroSize8 = WelsSetMemZero_neon;
pFuncList->pfSetMemZeroSize64Aligned16 = WelsSetMemZero_neon;
pFuncList->pfSetMemZeroSize64 = WelsSetMemZero_neon;
}
#endif
InitExpandPictureFunc (pFuncList, uiCpuFlag);
/* Intra_Prediction_fn*/
WelsInitFillingPredFuncs (uiCpuFlag);
WelsInitIntraPredFuncs (pFuncList, uiCpuFlag);
/* ME func */
WelsInitMeFunc(pFuncList, uiCpuFlag, SCREEN_CONTENT_REAL_TIME==pParam->iUsageType);
/* sad, satd, average */
WelsInitSampleSadFunc (pFuncList, uiCpuFlag);
//
WelsInitBGDFunc (pFuncList, pParam->bEnableBackgroundDetection);
// for pfGetVarianceFromIntraVaa function ptr adaptive by CPU features, 6/7/2010
InitIntraAnalysisVaaInfo (pFuncList, uiCpuFlag);
/* Motion compensation */
/*init pixel average function*/
/*get one column or row pixel when refinement*/
WelsInitMcFuncs (pFuncList, uiCpuFlag);
InitCoeffFunc (uiCpuFlag);
WelsInitEncodingFuncs (pFuncList, uiCpuFlag);
WelsInitReconstructionFuncs (pFuncList, uiCpuFlag);
DeblockingInit (&pFuncList->pfDeblocking, uiCpuFlag);
WelsBlockFuncInit (&pFuncList->pfSetNZCZero, uiCpuFlag);
InitFillNeighborCacheInterFunc (pFuncList, pParam->bEnableBackgroundDetection);
InitRefListMgrFunc(pFuncList,pParam->iUsageType);
return iReturn;
}
void AssignFuncPointerForRec (PWelsDecoderContext pCtx) {
pCtx->pGetI16x16LumaPredFunc[I16_PRED_V ] = WelsI16x16LumaPredV_c;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_H ] = WelsI16x16LumaPredH_c;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_DC ] = WelsI16x16LumaPredDc_c;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_P ] = WelsI16x16LumaPredPlane_c;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_DC_L ] = WelsI16x16LumaPredDcLeft_c;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_DC_T ] = WelsI16x16LumaPredDcTop_c;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_DC_128] = WelsI16x16LumaPredDcNA_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_V ] = WelsI4x4LumaPredV_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_H ] = WelsI4x4LumaPredH_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DC ] = WelsI4x4LumaPredDc_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DC_L ] = WelsI4x4LumaPredDcLeft_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DC_T ] = WelsI4x4LumaPredDcTop_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DC_128] = WelsI4x4LumaPredDcNA_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DDL ] = WelsI4x4LumaPredDDL_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DDL_TOP] = WelsI4x4LumaPredDDLTop_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DDR ] = WelsI4x4LumaPredDDR_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_VL ] = WelsI4x4LumaPredVL_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_VL_TOP] = WelsI4x4LumaPredVLTop_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_VR ] = WelsI4x4LumaPredVR_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_HU ] = WelsI4x4LumaPredHU_c;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_HD ] = WelsI4x4LumaPredHD_c;
pCtx->pGetIChromaPredFunc[C_PRED_DC ] = WelsIChromaPredDc_c;
pCtx->pGetIChromaPredFunc[C_PRED_H ] = WelsIChromaPredH_c;
pCtx->pGetIChromaPredFunc[C_PRED_V ] = WelsIChromaPredV_c;
pCtx->pGetIChromaPredFunc[C_PRED_P ] = WelsIChromaPredPlane_c;
pCtx->pGetIChromaPredFunc[C_PRED_DC_L ] = WelsIChromaPredDcLeft_c;
pCtx->pGetIChromaPredFunc[C_PRED_DC_T ] = WelsIChromaPredDcTop_c;
pCtx->pGetIChromaPredFunc[C_PRED_DC_128] = WelsIChromaPredDcNA_c;
pCtx->pIdctResAddPredFunc = IdctResAddPred_c;
#if defined(HAVE_NEON)
if (pCtx->uiCpuFlag & WELS_CPU_NEON) {
pCtx->pIdctResAddPredFunc = IdctResAddPred_neon;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_DC] = WelsDecoderI16x16LumaPredDc_neon;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_P] = WelsDecoderI16x16LumaPredPlane_neon;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_H] = WelsDecoderI16x16LumaPredH_neon;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_V] = WelsDecoderI16x16LumaPredV_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_V ] = WelsDecoderI4x4LumaPredV_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_H ] = WelsDecoderI4x4LumaPredH_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DDL ] = WelsDecoderI4x4LumaPredDDL_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DDR ] = WelsDecoderI4x4LumaPredDDR_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_VL ] = WelsDecoderI4x4LumaPredVL_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_VR ] = WelsDecoderI4x4LumaPredVR_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_HU ] = WelsDecoderI4x4LumaPredHU_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_HD ] = WelsDecoderI4x4LumaPredHD_neon;
pCtx->pGetIChromaPredFunc[C_PRED_H] = WelsDecoderIChromaPredH_neon;
pCtx->pGetIChromaPredFunc[C_PRED_V] = WelsDecoderIChromaPredV_neon;
pCtx->pGetIChromaPredFunc[C_PRED_P ] = WelsDecoderIChromaPredPlane_neon;
pCtx->pGetIChromaPredFunc[C_PRED_DC] = WelsDecoderIChromaPredDc_neon;
}
#endif//HAVE_NEON
#if defined(HAVE_NEON_AARCH64)
if (pCtx->uiCpuFlag & WELS_CPU_NEON) {
pCtx->pIdctResAddPredFunc = IdctResAddPred_AArch64_neon;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_DC] = WelsDecoderI16x16LumaPredDc_AArch64_neon;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_P] = WelsDecoderI16x16LumaPredPlane_AArch64_neon;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_H] = WelsDecoderI16x16LumaPredH_AArch64_neon;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_V] = WelsDecoderI16x16LumaPredV_AArch64_neon;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_DC_L] = WelsDecoderI16x16LumaPredDcLeft_AArch64_neon;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_DC_T] = WelsDecoderI16x16LumaPredDcTop_AArch64_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_H ] = WelsDecoderI4x4LumaPredH_AArch64_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DDL ] = WelsDecoderI4x4LumaPredDDL_AArch64_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DDL_TOP] = WelsDecoderI4x4LumaPredDDLTop_AArch64_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_VL ] = WelsDecoderI4x4LumaPredVL_AArch64_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_VL_TOP ] = WelsDecoderI4x4LumaPredVLTop_AArch64_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_VR ] = WelsDecoderI4x4LumaPredVR_AArch64_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_HU ] = WelsDecoderI4x4LumaPredHU_AArch64_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_HD ] = WelsDecoderI4x4LumaPredHD_AArch64_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DC ] = WelsDecoderI4x4LumaPredDc_AArch64_neon;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DC_T ] = WelsDecoderI4x4LumaPredDcTop_AArch64_neon;
pCtx->pGetIChromaPredFunc[C_PRED_H] = WelsDecoderIChromaPredH_AArch64_neon;
pCtx->pGetIChromaPredFunc[C_PRED_V] = WelsDecoderIChromaPredV_AArch64_neon;
pCtx->pGetIChromaPredFunc[C_PRED_P ] = WelsDecoderIChromaPredPlane_AArch64_neon;
pCtx->pGetIChromaPredFunc[C_PRED_DC] = WelsDecoderIChromaPredDc_AArch64_neon;
pCtx->pGetIChromaPredFunc[C_PRED_DC_T] = WelsDecoderIChromaPredDcTop_AArch64_neon;
}
#endif//HAVE_NEON_AARCH64
#if defined(X86_ASM)
if (pCtx->uiCpuFlag & WELS_CPU_MMXEXT) {
pCtx->pIdctResAddPredFunc = IdctResAddPred_mmx;
/////////mmx code opt---
pCtx->pGetIChromaPredFunc[C_PRED_H] = WelsDecoderIChromaPredH_mmx;
pCtx->pGetIChromaPredFunc[C_PRED_V] = WelsDecoderIChromaPredV_mmx;
pCtx->pGetIChromaPredFunc[C_PRED_DC_L ] = WelsDecoderIChromaPredDcLeft_mmx;
pCtx->pGetIChromaPredFunc[C_PRED_DC_128] = WelsDecoderIChromaPredDcNA_mmx;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DDR] = WelsDecoderI4x4LumaPredDDR_mmx;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_HD ] = WelsDecoderI4x4LumaPredHD_mmx;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_HU ] = WelsDecoderI4x4LumaPredHU_mmx;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_VR ] = WelsDecoderI4x4LumaPredVR_mmx;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_DDL] = WelsDecoderI4x4LumaPredDDL_mmx;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_VL ] = WelsDecoderI4x4LumaPredVL_mmx;
}
if (pCtx->uiCpuFlag & WELS_CPU_SSE2) {
/////////sse2 code opt---
pCtx->pGetI16x16LumaPredFunc[I16_PRED_DC] = WelsDecoderI16x16LumaPredDc_sse2;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_P] = WelsDecoderI16x16LumaPredPlane_sse2;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_H] = WelsDecoderI16x16LumaPredH_sse2;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_V] = WelsDecoderI16x16LumaPredV_sse2;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_DC_T ] = WelsDecoderI16x16LumaPredDcTop_sse2;
pCtx->pGetI16x16LumaPredFunc[I16_PRED_DC_128] = WelsDecoderI16x16LumaPredDcNA_sse2;
pCtx->pGetIChromaPredFunc[C_PRED_P ] = WelsDecoderIChromaPredPlane_sse2;
pCtx->pGetIChromaPredFunc[C_PRED_DC] = WelsDecoderIChromaPredDc_sse2;
pCtx->pGetIChromaPredFunc[C_PRED_DC_T] = WelsDecoderIChromaPredDcTop_sse2;
pCtx->pGetI4x4LumaPredFunc[I4_PRED_H] = WelsDecoderI4x4LumaPredH_sse2;
}
#endif
DeblockingInit (&pCtx->sDeblockingFunc, pCtx->uiCpuFlag);
WelsBlockFuncInit (&pCtx->sBlockFunc, pCtx->uiCpuFlag);
}
/*!
* \brief initialize function pointers that potentially used in Wels encoding
* \param pEncCtx sWelsEncCtx*
* \return successful - 0; otherwise none 0 for failed
*/
int32_t InitFunctionPointers (sWelsEncCtx* pEncCtx, SWelsSvcCodingParam* pParam, uint32_t uiCpuFlag) {
int32_t iReturn = ENC_RETURN_SUCCESS;
SWelsFuncPtrList* pFuncList = pEncCtx->pFuncList;
bool bScreenContent = (SCREEN_CONTENT_REAL_TIME == pParam->iUsageType);
/* Functionality utilization of CPU instructions dependency */
pFuncList->pfSetMemZeroSize8 = WelsSetMemZero_c; // confirmed_safe_unsafe_usage
pFuncList->pfSetMemZeroSize64Aligned16 = WelsSetMemZero_c; // confirmed_safe_unsafe_usage
pFuncList->pfSetMemZeroSize64 = WelsSetMemZero_c; // confirmed_safe_unsafe_usage
#if defined(X86_ASM)
if (uiCpuFlag & WELS_CPU_MMXEXT) {
pFuncList->pfSetMemZeroSize8 = WelsSetMemZeroSize8_mmx; // confirmed_safe_unsafe_usage
pFuncList->pfSetMemZeroSize64Aligned16 = WelsSetMemZeroSize64_mmx; // confirmed_safe_unsafe_usage
pFuncList->pfSetMemZeroSize64 = WelsSetMemZeroSize64_mmx; // confirmed_safe_unsafe_usage
}
if (uiCpuFlag & WELS_CPU_SSE2) {
pFuncList->pfSetMemZeroSize64Aligned16 = WelsSetMemZeroAligned64_sse2; // confirmed_safe_unsafe_usage
}
#endif//X86_ASM
#if defined(HAVE_NEON)
if (uiCpuFlag & WELS_CPU_NEON) {
pFuncList->pfSetMemZeroSize8 = WelsSetMemZero_neon;
pFuncList->pfSetMemZeroSize64Aligned16 = WelsSetMemZero_neon;
pFuncList->pfSetMemZeroSize64 = WelsSetMemZero_neon;
}
#endif
#if defined(HAVE_NEON_AARCH64)
if (uiCpuFlag & WELS_CPU_NEON) {
pFuncList->pfSetMemZeroSize8 = WelsSetMemZero_AArch64_neon;
pFuncList->pfSetMemZeroSize64Aligned16 = WelsSetMemZero_AArch64_neon;
pFuncList->pfSetMemZeroSize64 = WelsSetMemZero_AArch64_neon;
}
#endif
InitExpandPictureFunc (& (pFuncList->sExpandPicFunc), uiCpuFlag);
/* Intra_Prediction_fn*/
WelsInitIntraPredFuncs (pFuncList, uiCpuFlag);
/* ME func */
WelsInitMeFunc (pFuncList, uiCpuFlag, bScreenContent);
/* sad, satd, average */
WelsInitSampleSadFunc (pFuncList, uiCpuFlag);
//
WelsInitBGDFunc (pFuncList, pParam->bEnableBackgroundDetection);
WelsInitSCDPskipFunc (pFuncList, bScreenContent && (pParam->bEnableSceneChangeDetect));
// for pfGetVarianceFromIntraVaa function ptr adaptive by CPU features, 6/7/2010
InitIntraAnalysisVaaInfo (pFuncList, uiCpuFlag);
/* Motion compensation */
/*init pixel average function*/
/*get one column or row pixel when refinement*/
InitMcFunc (&pFuncList->sMcFuncs, uiCpuFlag);
InitCoeffFunc (pFuncList, uiCpuFlag, pParam->iEntropyCodingModeFlag);
WelsInitEncodingFuncs (pFuncList, uiCpuFlag);
WelsInitReconstructionFuncs (pFuncList, uiCpuFlag);
DeblockingInit (&pFuncList->pfDeblocking, uiCpuFlag);
WelsBlockFuncInit (&pFuncList->pfSetNZCZero, uiCpuFlag);
InitFillNeighborCacheInterFunc (pFuncList, pParam->bEnableBackgroundDetection);
pFuncList->pParametersetStrategy = IWelsParametersetStrategy::CreateParametersetStrategy (pParam->eSpsPpsIdStrategy,
pParam->bSimulcastAVC, pParam->iSpatialLayerNum);
WELS_VERIFY_RETURN_IF (ENC_RETURN_MEMALLOCERR, (NULL == pFuncList->pParametersetStrategy))
return iReturn;
}
