--- trunk/xvidcore/src/encoder.c 2002/03/16 15:55:12 36 +++ trunk/xvidcore/src/encoder.c 2002/03/29 07:03:24 84 @@ -16,6 +16,7 @@ #include "bitstream/mbcoding.h" #include "quant/adapt_quant.h" #include "quant/quant_matrix.h" +#include "utils/mem_align.h" #define ENC_CHECK(X) if(!(X)) return XVID_ERR_FORMAT @@ -81,7 +82,7 @@ pParam->bitrate = 900000; if (pParam->rc_buffersize <= 0) - pParam->rc_buffersize = pParam->bitrate * pParam->fbase; + pParam->rc_buffersize = 16; if ((pParam->min_quantizer <= 0) || (pParam->min_quantizer > 31)) pParam->min_quantizer = 1; @@ -95,7 +96,7 @@ if (pParam->max_quantizer < pParam->min_quantizer) pParam->max_quantizer = pParam->min_quantizer; - if ((pEnc = (Encoder *) malloc(sizeof(Encoder))) == NULL) + if ((pEnc = (Encoder *) xvid_malloc(sizeof(Encoder), CACHE_LINE)) == NULL) return XVID_ERR_MEMORY; /* Fill members of Encoder structure */ @@ -120,55 +121,38 @@ pEnc->iFrameNum = 0; pEnc->iMaxKeyInterval = pParam->max_key_interval; - if (image_create(&pEnc->sCurrent, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0) - { - free(pEnc); - return XVID_ERR_MEMORY; - } - - if (image_create(&pEnc->sReference, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0) - { - image_destroy(&pEnc->sCurrent, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - free(pEnc); - return XVID_ERR_MEMORY; - } - - if (image_create(&pEnc->vInterH, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0) - { - image_destroy(&pEnc->sCurrent, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - image_destroy(&pEnc->sReference, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - free(pEnc); - return XVID_ERR_MEMORY; - } - - if (image_create(&pEnc->vInterV, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0) - { - image_destroy(&pEnc->sCurrent, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - image_destroy(&pEnc->sReference, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - image_destroy(&pEnc->vInterH, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - free(pEnc); - return XVID_ERR_MEMORY; - } + /* try to allocate memory */ - if (image_create(&pEnc->vInterHV, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0) - { - image_destroy(&pEnc->sCurrent, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - image_destroy(&pEnc->sReference, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - image_destroy(&pEnc->vInterH, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - image_destroy(&pEnc->vInterV, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - free(pEnc); - return XVID_ERR_MEMORY; - } - - pEnc->pMBs = malloc(sizeof(MACROBLOCK) * pEnc->mbParam.mb_width * pEnc->mbParam.mb_height); - if (pEnc->pMBs == NULL) + pEnc->sCurrent.y = pEnc->sCurrent.u = pEnc->sCurrent.v = NULL; + pEnc->sReference.y = pEnc->sReference.u = pEnc->sReference.v = NULL; + pEnc->vInterH.y = pEnc->vInterH.u = pEnc->vInterH.v = NULL; + pEnc->vInterV.y = pEnc->vInterV.u = pEnc->vInterV.v = NULL; + pEnc->vInterVf.y = pEnc->vInterVf.u = pEnc->vInterVf.v = NULL; + pEnc->vInterHV.y = pEnc->vInterHV.u = pEnc->vInterHV.v = NULL; + pEnc->vInterHVf.y = pEnc->vInterHVf.u = pEnc->vInterHVf.v = NULL; + + pEnc->pMBs = NULL; + + if (image_create(&pEnc->sCurrent, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0 || + image_create(&pEnc->sReference, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0 || + image_create(&pEnc->vInterH, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0 || + image_create(&pEnc->vInterV, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0 || + image_create(&pEnc->vInterVf, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0 || + image_create(&pEnc->vInterHV, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0 || + image_create(&pEnc->vInterHVf, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height) < 0 || + (pEnc->pMBs = xvid_malloc(sizeof(MACROBLOCK) * pEnc->mbParam.mb_width * pEnc->mbParam.mb_height, CACHE_LINE)) == NULL) { image_destroy(&pEnc->sCurrent, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); image_destroy(&pEnc->sReference, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); image_destroy(&pEnc->vInterH, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); image_destroy(&pEnc->vInterV, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); + image_destroy(&pEnc->vInterVf, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); image_destroy(&pEnc->vInterHV, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - free(pEnc); + image_destroy(&pEnc->vInterHVf, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); + if (pEnc) + { + xvid_free(pEnc); + } return XVID_ERR_MEMORY; } @@ -182,8 +166,8 @@ if (pParam->bitrate) { - RateControlInit(pParam->bitrate, pParam->rc_buffersize, pParam->fbase, pParam->width, - pParam->height, pParam->max_quantizer, pParam->min_quantizer); + RateControlInit(pParam->bitrate, pParam->rc_buffersize, pParam->fbase * 100 / pParam->fincr, + pParam->max_quantizer, pParam->min_quantizer); } create_vlc_tables(); @@ -199,13 +183,13 @@ ENC_CHECK(pEnc->sCurrent.y); ENC_CHECK(pEnc->sReference.y); - free(pEnc->pMBs); + xvid_free(pEnc->pMBs); image_destroy(&pEnc->sCurrent, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); image_destroy(&pEnc->sReference, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); image_destroy(&pEnc->vInterH, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); image_destroy(&pEnc->vInterV, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); image_destroy(&pEnc->vInterHV, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height); - free(pEnc); + xvid_free(pEnc); destroy_vlc_tables(); @@ -237,6 +221,8 @@ } stop_conv_timer(); + EMMS(); + BitstreamInit(&bs, pFrame->bitstream, 0); if (pFrame->quant == 0) @@ -250,11 +236,16 @@ if ((pEnc->mbParam.global_flags & XVID_LUMIMASKING) > 0) { - int * temp_dquants = (int *) malloc(pEnc->mbParam.mb_width * pEnc->mbParam.mb_height * sizeof(int)); + int * temp_dquants = (int *) xvid_malloc(pEnc->mbParam.mb_width * pEnc->mbParam.mb_height * sizeof(int), CACHE_LINE); - pEnc->mbParam.quant = adaptive_quantization(pEnc->sCurrent.y, pEnc->mbParam.width, - temp_dquants, pFrame->quant, pFrame->quant, - 2*pFrame->quant, pEnc->mbParam.mb_width, pEnc->mbParam.mb_height); + pEnc->mbParam.quant = adaptive_quantization(pEnc->sCurrent.y, + pEnc->mbParam.width, + temp_dquants, + pEnc->mbParam.quant, + pEnc->mbParam.quant, + 2*pEnc->mbParam.quant, + pEnc->mbParam.mb_width, + pEnc->mbParam.mb_height); for (y = 0; y < pEnc->mbParam.mb_height; y++) for (x = 0; x < pEnc->mbParam.mb_width; x++) @@ -262,7 +253,7 @@ MACROBLOCK *pMB = &pEnc->pMBs[x + y * pEnc->mbParam.mb_width]; pMB->dquant = iDQtab[(temp_dquants[y * pEnc->mbParam.mb_width + x] + 2)]; } - free(temp_dquants); + xvid_free(temp_dquants); } if(pEnc->mbParam.global_flags & XVID_H263QUANT) { @@ -273,6 +264,8 @@ else if(pEnc->mbParam.global_flags & XVID_MPEGQUANT) { int ret1, ret2; + ret1 = ret2 = 0; + if(pEnc->mbParam.quant_type != MPEG4_QUANT) write_vol_header = 1; @@ -323,6 +316,8 @@ pResult->ublks = pEnc->sStat.ublks; } + EMMS(); + if (pFrame->quant == 0) { RateControlUpdate(pEnc->mbParam.quant, pFrame->length, pFrame->intra); @@ -359,18 +354,18 @@ static int FrameCodeI(Encoder * pEnc, Bitstream * bs, uint32_t *pBits) { - int16_t dct_codes[6][64]; - int16_t qcoeff[6][64]; + + DECLARE_ALIGNED_MATRIX(dct_codes, 6, 64, int16_t, CACHE_LINE); + DECLARE_ALIGNED_MATRIX(qcoeff, 6, 64, int16_t, CACHE_LINE); + uint16_t x, y; pEnc->iFrameNum = 0; pEnc->mbParam.rounding_type = 1; pEnc->mbParam.coding_type = I_VOP; - BitstreamWriteVolHeader(bs, pEnc->mbParam.width, pEnc->mbParam.height, pEnc->mbParam.quant_type); - BitstreamWriteVopHeader(bs, I_VOP, pEnc->mbParam.rounding_type, - pEnc->mbParam.quant, - pEnc->mbParam.fixed_code); + BitstreamWriteVolHeader(bs, &pEnc->mbParam); + BitstreamWriteVopHeader(bs, &pEnc->mbParam); *pBits = BitstreamPos(bs); @@ -385,7 +380,7 @@ CodeIntraMB(pEnc, pMB); - MBTransQuantIntra(&pEnc->mbParam, x, y, dct_codes, qcoeff, &pEnc->sCurrent); + MBTransQuantIntra(&pEnc->mbParam, pMB, x, y, dct_codes, qcoeff, &pEnc->sCurrent); start_timer(); MBPrediction(&pEnc->mbParam, x, y, pEnc->mbParam.mb_width, qcoeff, pEnc->pMBs); @@ -413,8 +408,10 @@ static int FrameCodeP(Encoder * pEnc, Bitstream * bs, uint32_t *pBits, bool force_inter, bool vol_header) { float fSigma; - int16_t dct_codes[6][64]; - int16_t qcoeff[6][64]; + + DECLARE_ALIGNED_MATRIX(dct_codes, 6, 64, int16_t, CACHE_LINE); + DECLARE_ALIGNED_MATRIX(qcoeff, 6, 64, int16_t, CACHE_LINE); + int iLimit; uint32_t x, y; int iSearchRange; @@ -423,7 +420,14 @@ IMAGE *pCurrent = &pEnc->sCurrent; IMAGE *pRef = &pEnc->sReference; - image_setedges(pRef,pEnc->mbParam.edged_width, pEnc->mbParam.edged_height, pEnc->mbParam.width, pEnc->mbParam.height); + start_timer(); + image_setedges(pRef, + pEnc->mbParam.edged_width, + pEnc->mbParam.edged_height, + pEnc->mbParam.width, + pEnc->mbParam.height, + pEnc->mbParam.global_flags & XVID_INTERLACING); + stop_edges_timer(); pEnc->mbParam.rounding_type = 1 - pEnc->mbParam.rounding_type; @@ -432,18 +436,49 @@ else iLimit = pEnc->mbParam.mb_width * pEnc->mbParam.mb_height + 1; - if ((pEnc->mbParam.global_flags & XVID_HALFPEL) > 0) { + if ((pEnc->mbParam.global_flags & XVID_HALFPEL) > 0) + { + IMAGE *vInterV = NULL; + IMAGE *vInterVf = NULL; + IMAGE *vInterHV = NULL; + IMAGE *vInterHVf = NULL; + + // interpolate fields together if field ME is used + if (pEnc->mbParam.global_flags & XVID_INTERLACING && + pEnc->mbParam.global_flags & XVID_FIELDME) + { + vInterVf = &pEnc->vInterVf; + vInterHVf = &pEnc->vInterHVf; + } + + // perform normal interpolation, unless only field-based ME is allowed + if (!(pEnc->mbParam.global_flags & XVID_INTERLACING) || + !(pEnc->mbParam.global_flags & XVID_FIELDMEONLY)) + { + vInterV = &pEnc->vInterV; + vInterHV = &pEnc->vInterHV; + } + start_timer(); - image_interpolate(pRef, &pEnc->vInterH, &pEnc->vInterV, &pEnc->vInterHV, - pEnc->mbParam.edged_width, pEnc->mbParam.edged_height, - pEnc->mbParam.rounding_type); + image_interpolate(pRef, + &pEnc->vInterH, + vInterV, vInterVf, + vInterHV, vInterHVf, + pEnc->mbParam.edged_width, + pEnc->mbParam.edged_height, + pEnc->mbParam.rounding_type); stop_inter_timer(); } start_timer(); - bIntra = MotionEstimation(pEnc->pMBs, &pEnc->mbParam, &pEnc->sReference, - &pEnc->vInterH, &pEnc->vInterV, - &pEnc->vInterHV, &pEnc->sCurrent, iLimit); + bIntra = MotionEstimation(pEnc->pMBs, + &pEnc->mbParam, + &pEnc->sReference, + &pEnc->vInterH, + &pEnc->vInterV, &pEnc->vInterVf, + &pEnc->vInterHV, &pEnc->vInterHVf, + &pEnc->sCurrent, + iLimit); stop_motion_timer(); if (bIntra == 1) @@ -452,11 +487,9 @@ pEnc->mbParam.coding_type = P_VOP; if(vol_header) - BitstreamWriteVolHeader(bs, pEnc->mbParam.width, pEnc->mbParam.height, pEnc->mbParam.quant_type); + BitstreamWriteVolHeader(bs, &pEnc->mbParam); - BitstreamWriteVopHeader(bs, P_VOP, pEnc->mbParam.rounding_type, - pEnc->mbParam.quant, - pEnc->mbParam.fixed_code); + BitstreamWriteVopHeader(bs, &pEnc->mbParam); *pBits = BitstreamPos(bs); @@ -476,9 +509,14 @@ if (!bIntra) { start_timer(); - MBMotionCompensation(pMB, x, y, &pEnc->sReference, - &pEnc->vInterH, &pEnc->vInterV, - &pEnc->vInterHV, &pEnc->sCurrent, dct_codes, + MBMotionCompensation(pMB, + x, y, + &pEnc->sReference, + &pEnc->vInterH, + &pEnc->vInterV, &pEnc->vInterVf, + &pEnc->vInterHV, &pEnc->vInterHVf, + &pEnc->sCurrent, + dct_codes, pEnc->mbParam.width, pEnc->mbParam.height, pEnc->mbParam.edged_width, @@ -495,12 +533,14 @@ } pMB->quant = pEnc->mbParam.quant; - pMB->cbp = MBTransQuantInter(&pEnc->mbParam, x, y, dct_codes, qcoeff, pCurrent); + pMB->field_pred = 0; + + pMB->cbp = MBTransQuantInter(&pEnc->mbParam, pMB, x, y, dct_codes, qcoeff, pCurrent); } else { CodeIntraMB(pEnc, pMB); - MBTransQuantIntra(&pEnc->mbParam, x, y, dct_codes, qcoeff, pCurrent); + MBTransQuantIntra(&pEnc->mbParam, pMB, x, y, dct_codes, qcoeff, pCurrent); } start_timer();