--- mbtransquant.c 2002/06/12 20:38:41 1.7 +++ mbtransquant.c 2003/03/22 14:04:48 1.22 @@ -42,10 +42,10 @@ * * * Revision history: * * * - * 29.03.2002 interlacing speedup - used transfer strides instead of - * manual field-to-frame conversion - * 26.03.2002 interlacing support - moved transfers outside loops - * 22.12.2001 get_dc_scaler() moved to common.h + * 29.03.2002 interlacing speedup - used transfer strides instead of * + * manual field-to-frame conversion * + * 26.03.2002 interlacing support - moved transfers outside loops * + * 22.12.2001 get_dc_scaler() moved to common.h * * 19.11.2001 introduced coefficient thresholding (Isibaar) * * 17.11.2001 initial version * * * @@ -65,159 +65,214 @@ #include "../quant/quant_h263.h" #include "../encoder.h" -#define MIN(X, Y) ((X)<(Y)?(X):(Y)) -#define MAX(X, Y) ((X)>(Y)?(X):(Y)) +#include "../image/reduced.h" -#define TOOSMALL_LIMIT 1 /* skip blocks having a coefficient sum below this value */ +MBFIELDTEST_PTR MBFieldTest; -/* this isnt pretty, but its better than 20 ifdefs */ +#define TOOSMALL_LIMIT 1 /* skip blocks having a coefficient sum below this value */ + +static __inline void +MBfDCT(int16_t data[6 * 64]) +{ + start_timer(); + fdct(&data[0 * 64]); + fdct(&data[1 * 64]); + fdct(&data[2 * 64]); + fdct(&data[3 * 64]); + fdct(&data[4 * 64]); + fdct(&data[5 * 64]); + stop_dct_timer(); +} + + +static __inline uint32_t +QuantizeInterBlock( int16_t qcoeff[64], + const int16_t data[64], + const uint32_t iQuant, + const uint32_t quant_type) +{ + uint32_t sum; + + start_timer(); + if (quant_type == H263_QUANT) + sum = quant_inter(qcoeff, data, iQuant); + else + sum = quant4_inter(qcoeff, data, iQuant); + + stop_quant_timer(); + return sum; +} void -MBTransQuantIntra(const MBParam * pParam, - FRAMEINFO * frame, - MACROBLOCK * pMB, - const uint32_t x_pos, - const uint32_t y_pos, - int16_t data[6 * 64], - int16_t qcoeff[6 * 64]) +MBTransQuantIntra(const MBParam * const pParam, + FRAMEINFO * const frame, + MACROBLOCK * const pMB, + const uint32_t x_pos, + const uint32_t y_pos, + int16_t data[6 * 64], + int16_t qcoeff[6 * 64]) { uint32_t stride = pParam->edged_width; - uint32_t stride2 = stride / 2; - uint32_t next_block = stride * 8; - uint32_t i; - uint32_t iQuant = frame->quant; + const uint32_t stride2 = stride / 2; + uint32_t next_block = stride * ((frame->global_flags & XVID_REDUCED)?16:8); + int i; + const uint32_t iQuant = pMB->quant; uint8_t *pY_Cur, *pU_Cur, *pV_Cur; - IMAGE *pCurrent = &frame->image; - - pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4); - pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3); - pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3); + const IMAGE * const pCurrent = &frame->image; start_timer(); - transfer_8to16copy(&data[0 * 64], pY_Cur, stride); - transfer_8to16copy(&data[1 * 64], pY_Cur + 8, stride); - transfer_8to16copy(&data[2 * 64], pY_Cur + next_block, stride); - transfer_8to16copy(&data[3 * 64], pY_Cur + next_block + 8, stride); - transfer_8to16copy(&data[4 * 64], pU_Cur, stride2); - transfer_8to16copy(&data[5 * 64], pV_Cur, stride2); + if ((frame->global_flags & XVID_REDUCED)) + { + pY_Cur = pCurrent->y + (y_pos << 5) * stride + (x_pos << 5); + pU_Cur = pCurrent->u + (y_pos << 4) * stride2 + (x_pos << 4); + pV_Cur = pCurrent->v + (y_pos << 4) * stride2 + (x_pos << 4); + + filter_18x18_to_8x8(&data[0 * 64], pY_Cur, stride); + filter_18x18_to_8x8(&data[1 * 64], pY_Cur + 16, stride); + filter_18x18_to_8x8(&data[2 * 64], pY_Cur + next_block, stride); + filter_18x18_to_8x8(&data[3 * 64], pY_Cur + next_block + 16, stride); + filter_18x18_to_8x8(&data[4 * 64], pU_Cur, stride2); + filter_18x18_to_8x8(&data[5 * 64], pV_Cur, stride2); + } else { + pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4); + pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3); + pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3); + + transfer_8to16copy(&data[0 * 64], pY_Cur, stride); + transfer_8to16copy(&data[1 * 64], pY_Cur + 8, stride); + transfer_8to16copy(&data[2 * 64], pY_Cur + next_block, stride); + transfer_8to16copy(&data[3 * 64], pY_Cur + next_block + 8, stride); + transfer_8to16copy(&data[4 * 64], pU_Cur, stride2); + transfer_8to16copy(&data[5 * 64], pV_Cur, stride2); + } stop_transfer_timer(); + /* XXX: rrv+interlacing is buggy */ start_timer(); pMB->field_dct = 0; - if ((frame->global_flags & XVID_INTERLACING)) { + if ((frame->global_flags & XVID_INTERLACING) && + (x_pos>0) && (x_posmb_width-1) && + (y_pos>0) && (y_posmb_height-1)) { pMB->field_dct = MBDecideFieldDCT(data); } stop_interlacing_timer(); + MBfDCT(data); + for (i = 0; i < 6; i++) { - uint32_t iDcScaler = get_dc_scaler(iQuant, i < 4); + const uint32_t iDcScaler = get_dc_scaler(iQuant, i < 4); start_timer(); - fdct(&data[i * 64]); - stop_dct_timer(); - - if (pParam->m_quant_type == H263_QUANT) { - start_timer(); + if (pParam->m_quant_type == H263_QUANT) quant_intra(&qcoeff[i * 64], &data[i * 64], iQuant, iDcScaler); - stop_quant_timer(); + else + quant4_intra(&qcoeff[i * 64], &data[i * 64], iQuant, iDcScaler); + stop_quant_timer(); + /* speedup: dont decode when encoding only ivops */ + if (pParam->iMaxKeyInterval != 1 || pParam->max_bframes > 0) + { start_timer(); - dequant_intra(&data[i * 64], &qcoeff[i * 64], iQuant, iDcScaler); + if (pParam->m_quant_type == H263_QUANT) + dequant_intra(&data[i * 64], &qcoeff[i * 64], iQuant, iDcScaler); + else + dequant4_intra(&data[i * 64], &qcoeff[i * 64], iQuant, iDcScaler); stop_iquant_timer(); - } else { - start_timer(); - quant4_intra(&qcoeff[i * 64], &data[i * 64], iQuant, iDcScaler); - stop_quant_timer(); start_timer(); - dequant4_intra(&data[i * 64], &qcoeff[i * 64], iQuant, iDcScaler); - stop_iquant_timer(); + idct(&data[i * 64]); + stop_idct_timer(); } - - start_timer(); - idct(&data[i * 64]); - stop_idct_timer(); } - if (pMB->field_dct) { - next_block = stride; - stride *= 2; - } + /* speedup: dont decode when encoding only ivops */ + if (pParam->iMaxKeyInterval != 1 || pParam->max_bframes > 0) + { + + if (pMB->field_dct) { + next_block = stride; + stride *= 2; + } - start_timer(); - transfer_16to8copy(pY_Cur, &data[0 * 64], stride); - transfer_16to8copy(pY_Cur + 8, &data[1 * 64], stride); - transfer_16to8copy(pY_Cur + next_block, &data[2 * 64], stride); - transfer_16to8copy(pY_Cur + next_block + 8, &data[3 * 64], stride); - transfer_16to8copy(pU_Cur, &data[4 * 64], stride2); - transfer_16to8copy(pV_Cur, &data[5 * 64], stride2); - stop_transfer_timer(); + start_timer(); + if ((frame->global_flags & XVID_REDUCED)) { + copy_upsampled_8x8_16to8(pY_Cur, &data[0 * 64], stride); + copy_upsampled_8x8_16to8(pY_Cur + 16, &data[1 * 64], stride); + copy_upsampled_8x8_16to8(pY_Cur + next_block, &data[2 * 64], stride); + copy_upsampled_8x8_16to8(pY_Cur + next_block + 16, &data[3 * 64], stride); + copy_upsampled_8x8_16to8(pU_Cur, &data[4 * 64], stride2); + copy_upsampled_8x8_16to8(pV_Cur, &data[5 * 64], stride2); + } else { + transfer_16to8copy(pY_Cur, &data[0 * 64], stride); + transfer_16to8copy(pY_Cur + 8, &data[1 * 64], stride); + transfer_16to8copy(pY_Cur + next_block, &data[2 * 64], stride); + transfer_16to8copy(pY_Cur + next_block + 8, &data[3 * 64], stride); + transfer_16to8copy(pU_Cur, &data[4 * 64], stride2); + transfer_16to8copy(pV_Cur, &data[5 * 64], stride2); + } + stop_transfer_timer(); + } } - uint8_t -MBTransQuantInter(const MBParam * pParam, - FRAMEINFO * frame, - MACROBLOCK * pMB, - const uint32_t x_pos, - const uint32_t y_pos, - int16_t data[6 * 64], - int16_t qcoeff[6 * 64]) +MBTransQuantInter(const MBParam * const pParam, + FRAMEINFO * const frame, + MACROBLOCK * const pMB, + const uint32_t x_pos, + const uint32_t y_pos, + int16_t data[6 * 64], + int16_t qcoeff[6 * 64]) { - uint32_t stride = pParam->edged_width; - uint32_t stride2 = stride / 2; - uint32_t next_block = stride * 8; - uint32_t i; - uint32_t iQuant = frame->quant; + const uint32_t stride2 = stride / 2; + uint32_t next_block = stride * ((frame->global_flags & XVID_REDUCED)?16:8); + int i; + const uint32_t iQuant = pMB->quant; uint8_t *pY_Cur, *pU_Cur, *pV_Cur; - uint8_t cbp = 0; + int cbp = 0; uint32_t sum; - IMAGE *pCurrent = &frame->image; + const IMAGE * const pCurrent = &frame->image; - pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4); - pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3); - pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3); + if ((frame->global_flags & XVID_REDUCED)) { + pY_Cur = pCurrent->y + (y_pos << 5) * stride + (x_pos << 5); + pU_Cur = pCurrent->u + (y_pos << 4) * stride2 + (x_pos << 4); + pV_Cur = pCurrent->v + (y_pos << 4) * stride2 + (x_pos << 4); + } else { + pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4); + pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3); + pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3); + } start_timer(); pMB->field_dct = 0; - if ((frame->global_flags & XVID_INTERLACING)) { + if ((frame->global_flags & XVID_INTERLACING) && + (x_pos>0) && (x_posmb_width-1) && + (y_pos>0) && (y_posmb_height-1)) { pMB->field_dct = MBDecideFieldDCT(data); } stop_interlacing_timer(); + MBfDCT(data); + for (i = 0; i < 6; i++) { - /* + const uint32_t limit = TOOSMALL_LIMIT + ((iQuant == 1) ? 1 : 0); + /* * no need to transfer 8->16-bit - * (this is performed already in motion compensation) + * (this is performed already in motion compensation) */ - start_timer(); - fdct(&data[i * 64]); - stop_dct_timer(); - if (pParam->m_quant_type == 0) { - start_timer(); - sum = quant_inter(&qcoeff[i * 64], &data[i * 64], iQuant); - stop_quant_timer(); - } else { - start_timer(); - sum = quant4_inter(&qcoeff[i * 64], &data[i * 64], iQuant); - stop_quant_timer(); - } + sum = QuantizeInterBlock(&qcoeff[i * 64], &data[i * 64], iQuant, pParam->m_quant_type); - if (sum >= TOOSMALL_LIMIT) { // skip block ? + if (sum >= limit) { - if (pParam->m_quant_type == H263_QUANT) { - start_timer(); + start_timer(); + if (pParam->m_quant_type == H263_QUANT) dequant_inter(&data[i * 64], &qcoeff[i * 64], iQuant); - stop_iquant_timer(); - } else { - start_timer(); + else dequant4_inter(&data[i * 64], &qcoeff[i * 64], iQuant); - stop_iquant_timer(); - } + stop_iquant_timer(); cbp |= 1 << (5 - i); @@ -233,32 +288,97 @@ } start_timer(); - if (cbp & 32) - transfer_16to8add(pY_Cur, &data[0 * 64], stride); - if (cbp & 16) - transfer_16to8add(pY_Cur + 8, &data[1 * 64], stride); - if (cbp & 8) - transfer_16to8add(pY_Cur + next_block, &data[2 * 64], stride); - if (cbp & 4) - transfer_16to8add(pY_Cur + next_block + 8, &data[3 * 64], stride); - if (cbp & 2) - transfer_16to8add(pU_Cur, &data[4 * 64], stride2); - if (cbp & 1) - transfer_16to8add(pV_Cur, &data[5 * 64], stride2); + if ((frame->global_flags & XVID_REDUCED)) { + if (cbp & 32) + add_upsampled_8x8_16to8(pY_Cur, &data[0 * 64], stride); + if (cbp & 16) + add_upsampled_8x8_16to8(pY_Cur + 16, &data[1 * 64], stride); + if (cbp & 8) + add_upsampled_8x8_16to8(pY_Cur + next_block, &data[2 * 64], stride); + if (cbp & 4) + add_upsampled_8x8_16to8(pY_Cur + 16 + next_block, &data[3 * 64], stride); + if (cbp & 2) + add_upsampled_8x8_16to8(pU_Cur, &data[4 * 64], stride2); + if (cbp & 1) + add_upsampled_8x8_16to8(pV_Cur, &data[5 * 64], stride2); + } else { + if (cbp & 32) + transfer_16to8add(pY_Cur, &data[0 * 64], stride); + if (cbp & 16) + transfer_16to8add(pY_Cur + 8, &data[1 * 64], stride); + if (cbp & 8) + transfer_16to8add(pY_Cur + next_block, &data[2 * 64], stride); + if (cbp & 4) + transfer_16to8add(pY_Cur + next_block + 8, &data[3 * 64], stride); + if (cbp & 2) + transfer_16to8add(pU_Cur, &data[4 * 64], stride2); + if (cbp & 1) + transfer_16to8add(pV_Cur, &data[5 * 64], stride2); + } stop_transfer_timer(); - return cbp; - + return (uint8_t) cbp; } +uint8_t +MBTransQuantInterBVOP(const MBParam * pParam, + FRAMEINFO * frame, + MACROBLOCK * pMB, + int16_t data[6 * 64], + int16_t qcoeff[6 * 64]) +{ + int cbp = 0; + int i; -/* if sum(diff between field lines) < sum(diff between frame lines), use field dct */ +/* there is no MBTrans for Inter block, that's done in motion compensation already */ + start_timer(); + pMB->field_dct = 0; + if ((frame->global_flags & XVID_INTERLACING)) { + pMB->field_dct = MBDecideFieldDCT(data); + } + stop_interlacing_timer(); -uint32_t + MBfDCT(data); + + for (i = 0; i < 6; i++) { + int codedecision = 0; + + int sum = QuantizeInterBlock(&qcoeff[i * 64], &data[i * 64], pMB->quant, pParam->m_quant_type); + + if ((sum > 2) || (qcoeff[i*64+1] != 0) || (qcoeff[i*64+8] != 0) ) codedecision = 1; + else { + if (pMB->mode == MODE_DIRECT || pMB->mode == MODE_DIRECT_NO4V) { + // dark blocks prevention for direct mode + if ( (qcoeff[i*64] < -1) || (qcoeff[i*64] > 0) ) codedecision = 1; + } else + if (qcoeff[i*64] != 0) codedecision = 1; // not direct mode + } + + if (codedecision) cbp |= 1 << (5 - i); + } + +/* we don't have to DeQuant, iDCT and Transfer back data for B-frames if we don't reconstruct this frame */ +/* warning: reconstruction not supported yet */ + return (uint8_t) cbp; +} + +/* permute block and return field dct choice */ + +static uint32_t MBDecideFieldDCT(int16_t data[6 * 64]) { + const uint32_t field = MBFieldTest(data); + if (field) MBFrameToField(data); + return field; +} + +/* if sum(diff between field lines) < sum(diff between frame lines), use field dct */ + +uint32_t +MBFieldTest_c(int16_t data[6 * 64]) +{ const uint8_t blocks[] = { 0 * 64, 0 * 64, 0 * 64, 0 * 64, 2 * 64, 2 * 64, 2 * 64, 2 * 64 }; const uint8_t lines[] = { 0, 16, 32, 48, 0, 16, 32, 48 }; @@ -292,11 +412,7 @@ } } - if (frame > field) { - MBFrameToField(data); - } - - return (frame > field); + return (frame >= (field + 350)); }