Use the fast quantizer for inter mode selection and the
regular quantizer for the rest of the encode for good quality,
speed 1. Both performance and quality were improved. The
quality gains will make up for the quality loss mentioned in
I9dc089007ca08129fb6c11fe7692777ebb8647b0.
Change-Id: Ia90bc9cf326a7c65d60d31fa32f6465ab6984d21
void vp8_fast_quantize_b_neon(BLOCK *b, BLOCKD *d)
{
void vp8_fast_quantize_b_neon(BLOCK *b, BLOCKD *d)
{
- d->eob = vp8_fast_quantize_b_neon_func(b->coeff, b->zbin, d->qcoeff, d->dqcoeff, d->dequant, vp8_rvsplus1_default_zig_zag1d, b->round, b->quant);
+ d->eob = vp8_fast_quantize_b_neon_func(b->coeff, b->zbin, d->qcoeff, d->dqcoeff, d->dequant, vp8_rvsplus1_default_zig_zag1d, b->round, b->quant_fast);
// 16 Y blocks, 4 U blocks, 4 V blocks each with 16 entries
short *quant;
// 16 Y blocks, 4 U blocks, 4 V blocks each with 16 entries
short *quant;
short *quant_shift;
short *zbin;
short *zrun_zbin_boost;
short *quant_shift;
short *zbin;
short *zrun_zbin_boost;
{
// dc values
quant_val = vp8_dc_quant(Q, cpi->common.y1dc_delta_q);
{
// dc values
quant_val = vp8_dc_quant(Q, cpi->common.y1dc_delta_q);
+ cpi->Y1quant_fast[Q][0] = (1 << 16) / quant_val;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->Y1quant[Q] + 0,
cpi->Y1quant_shift[Q] + 0, quant_val);
cpi->Y1zbin[Q][0] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->Y1quant[Q] + 0,
cpi->Y1quant_shift[Q] + 0, quant_val);
cpi->Y1zbin[Q][0] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
cpi->zrun_zbin_boost_y1[Q][0] = (quant_val * zbin_boost[0]) >> 7;
quant_val = vp8_dc2quant(Q, cpi->common.y2dc_delta_q);
cpi->zrun_zbin_boost_y1[Q][0] = (quant_val * zbin_boost[0]) >> 7;
quant_val = vp8_dc2quant(Q, cpi->common.y2dc_delta_q);
+ cpi->Y2quant_fast[Q][0] = (1 << 16) / quant_val;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->Y2quant[Q] + 0,
cpi->Y2quant_shift[Q] + 0, quant_val);
cpi->Y2zbin[Q][0] = ((qzbin_factors_y2[Q] * quant_val) + 64) >> 7;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->Y2quant[Q] + 0,
cpi->Y2quant_shift[Q] + 0, quant_val);
cpi->Y2zbin[Q][0] = ((qzbin_factors_y2[Q] * quant_val) + 64) >> 7;
cpi->zrun_zbin_boost_y2[Q][0] = (quant_val * zbin_boost[0]) >> 7;
quant_val = vp8_dc_uv_quant(Q, cpi->common.uvdc_delta_q);
cpi->zrun_zbin_boost_y2[Q][0] = (quant_val * zbin_boost[0]) >> 7;
quant_val = vp8_dc_uv_quant(Q, cpi->common.uvdc_delta_q);
+ cpi->UVquant_fast[Q][0] = (1 << 16) / quant_val;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->UVquant[Q] + 0,
cpi->UVquant_shift[Q] + 0, quant_val);
cpi->UVzbin[Q][0] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->UVquant[Q] + 0,
cpi->UVquant_shift[Q] + 0, quant_val);
cpi->UVzbin[Q][0] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;;
int rc = vp8_default_zig_zag1d[i];
quant_val = vp8_ac_yquant(Q);
int rc = vp8_default_zig_zag1d[i];
quant_val = vp8_ac_yquant(Q);
+ cpi->Y1quant_fast[Q][rc] = (1 << 16) / quant_val;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->Y1quant[Q] + rc,
cpi->Y1quant_shift[Q] + rc, quant_val);
cpi->Y1zbin[Q][rc] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->Y1quant[Q] + rc,
cpi->Y1quant_shift[Q] + rc, quant_val);
cpi->Y1zbin[Q][rc] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
cpi->zrun_zbin_boost_y1[Q][i] = (quant_val * zbin_boost[i]) >> 7;
quant_val = vp8_ac2quant(Q, cpi->common.y2ac_delta_q);
cpi->zrun_zbin_boost_y1[Q][i] = (quant_val * zbin_boost[i]) >> 7;
quant_val = vp8_ac2quant(Q, cpi->common.y2ac_delta_q);
+ cpi->Y2quant_fast[Q][rc] = (1 << 16) / quant_val;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->Y2quant[Q] + rc,
cpi->Y2quant_shift[Q] + rc, quant_val);
cpi->Y2zbin[Q][rc] = ((qzbin_factors_y2[Q] * quant_val) + 64) >> 7;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->Y2quant[Q] + rc,
cpi->Y2quant_shift[Q] + rc, quant_val);
cpi->Y2zbin[Q][rc] = ((qzbin_factors_y2[Q] * quant_val) + 64) >> 7;
cpi->zrun_zbin_boost_y2[Q][i] = (quant_val * zbin_boost[i]) >> 7;
quant_val = vp8_ac_uv_quant(Q, cpi->common.uvac_delta_q);
cpi->zrun_zbin_boost_y2[Q][i] = (quant_val * zbin_boost[i]) >> 7;
quant_val = vp8_ac_uv_quant(Q, cpi->common.uvac_delta_q);
+ cpi->UVquant_fast[Q][rc] = (1 << 16) / quant_val;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->UVquant[Q] + rc,
cpi->UVquant_shift[Q] + rc, quant_val);
cpi->UVzbin[Q][rc] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
vp8cx_invert_quant(cpi->sf.improved_quant, cpi->UVquant[Q] + rc,
cpi->UVquant_shift[Q] + rc, quant_val);
cpi->UVzbin[Q][rc] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
for (i = 0; i < 16; i++)
{
x->block[i].quant = cpi->Y1quant[QIndex];
for (i = 0; i < 16; i++)
{
x->block[i].quant = cpi->Y1quant[QIndex];
+ x->block[i].quant_fast = cpi->Y1quant_fast[QIndex];
x->block[i].quant_shift = cpi->Y1quant_shift[QIndex];
x->block[i].zbin = cpi->Y1zbin[QIndex];
x->block[i].round = cpi->Y1round[QIndex];
x->block[i].quant_shift = cpi->Y1quant_shift[QIndex];
x->block[i].zbin = cpi->Y1zbin[QIndex];
x->block[i].round = cpi->Y1round[QIndex];
for (i = 16; i < 24; i++)
{
x->block[i].quant = cpi->UVquant[QIndex];
for (i = 16; i < 24; i++)
{
x->block[i].quant = cpi->UVquant[QIndex];
+ x->block[i].quant_fast = cpi->UVquant_fast[QIndex];
x->block[i].quant_shift = cpi->UVquant_shift[QIndex];
x->block[i].zbin = cpi->UVzbin[QIndex];
x->block[i].round = cpi->UVround[QIndex];
x->block[i].quant_shift = cpi->UVquant_shift[QIndex];
x->block[i].zbin = cpi->UVzbin[QIndex];
x->block[i].round = cpi->UVround[QIndex];
// Y2
zbin_extra = (cpi->common.Y2dequant[QIndex][1] * ((cpi->zbin_over_quant / 2) + cpi->zbin_mode_boost)) >> 7;
// Y2
zbin_extra = (cpi->common.Y2dequant[QIndex][1] * ((cpi->zbin_over_quant / 2) + cpi->zbin_mode_boost)) >> 7;
+ x->block[24].quant_fast = cpi->Y2quant_fast[QIndex];
x->block[24].quant = cpi->Y2quant[QIndex];
x->block[24].quant_shift = cpi->Y2quant_shift[QIndex];
x->block[24].zbin = cpi->Y2zbin[QIndex];
x->block[24].quant = cpi->Y2quant[QIndex];
x->block[24].quant_shift = cpi->Y2quant_shift[QIndex];
x->block[24].zbin = cpi->Y2zbin[QIndex];
+ /* Are we using the fast quantizer for the mode selection? */
+ if(cpi->sf.use_fastquant_for_pick)
+ cpi->mb.quantize_b = QUANTIZE_INVOKE(&cpi->rtcd.quantize, fastquantb);
+
inter_error = vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate, &distortion, &intra_error);
inter_error = vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate, &distortion, &intra_error);
+
+ /* switch back to the regular quantizer for the encode */
+ if (cpi->sf.improved_quant)
+ {
+ cpi->mb.quantize_b = QUANTIZE_INVOKE(&cpi->rtcd.quantize, quantb);
+ }
+
sf->max_fs_radius = 32;
sf->iterative_sub_pixel = 1;
sf->optimize_coefficients = 1;
sf->max_fs_radius = 32;
sf->iterative_sub_pixel = 1;
sf->optimize_coefficients = 1;
+ sf->use_fastquant_for_pick = 0;
sf->first_step = 0;
sf->max_step_search_steps = MAX_MVSEARCH_STEPS;
sf->first_step = 0;
sf->max_step_search_steps = MAX_MVSEARCH_STEPS;
cpi->mode_check_freq[THR_SPLITG] = 4;
cpi->mode_check_freq[THR_SPLITA] = 4;
cpi->mode_check_freq[THR_SPLITG] = 4;
cpi->mode_check_freq[THR_SPLITA] = 4;
- cpi->mode_check_freq[THR_SPLITMV] = 0;
+ cpi->mode_check_freq[THR_SPLITMV] = 2;
sf->thresh_mult[THR_TM ] = 1500;
sf->thresh_mult[THR_V_PRED ] = 1500;
sf->thresh_mult[THR_TM ] = 1500;
sf->thresh_mult[THR_V_PRED ] = 1500;
sf->thresh_mult[THR_SPLITA ] = 20000;
}
sf->thresh_mult[THR_SPLITA ] = 20000;
}
- sf->improved_quant = 0;
- sf->improved_dct = 0;
+ sf->use_fastquant_for_pick = 1;
sf->first_step = 1;
sf->max_step_search_steps = MAX_MVSEARCH_STEPS;
sf->first_step = 1;
sf->max_step_search_steps = MAX_MVSEARCH_STEPS;
+ sf->use_fastquant_for_pick = 0;
+
cpi->mode_check_freq[THR_SPLITG] = 15;
cpi->mode_check_freq[THR_SPLITA] = 15;
cpi->mode_check_freq[THR_SPLITMV] = 7;
cpi->mode_check_freq[THR_SPLITG] = 15;
cpi->mode_check_freq[THR_SPLITA] = 15;
cpi->mode_check_freq[THR_SPLITMV] = 7;
sf->thresh_mult[THR_SPLITA ] = 50000;
}
sf->thresh_mult[THR_SPLITA ] = 50000;
}
+ sf->first_step = 1;
+
+ sf->improved_quant = 0;
+ sf->improved_dct = 0;
+
// Only do recode loop on key frames, golden frames and
// alt ref frames
sf->recode_loop = 2;
// Only do recode loop on key frames, golden frames and
// alt ref frames
sf->recode_loop = 2;
int first_step;
int optimize_coefficients;
int first_step;
int optimize_coefficients;
+ int use_fastquant_for_pick;
+
} SPEED_FEATURES;
typedef struct
} SPEED_FEATURES;
typedef struct
DECLARE_ALIGNED(16, short, zrun_zbin_boost_y1[QINDEX_RANGE][16]);
DECLARE_ALIGNED(16, short, zrun_zbin_boost_y2[QINDEX_RANGE][16]);
DECLARE_ALIGNED(16, short, zrun_zbin_boost_uv[QINDEX_RANGE][16]);
DECLARE_ALIGNED(16, short, zrun_zbin_boost_y1[QINDEX_RANGE][16]);
DECLARE_ALIGNED(16, short, zrun_zbin_boost_y2[QINDEX_RANGE][16]);
DECLARE_ALIGNED(16, short, zrun_zbin_boost_uv[QINDEX_RANGE][16]);
+ DECLARE_ALIGNED(16, short, Y1quant_fast[QINDEX_RANGE][16]);
+ DECLARE_ALIGNED(16, short, Y2quant_fast[QINDEX_RANGE][16]);
+ DECLARE_ALIGNED(16, short, UVquant_fast[QINDEX_RANGE][16]);
short *coeff_ptr = b->coeff;
short *zbin_ptr = b->zbin;
short *round_ptr = b->round;
short *coeff_ptr = b->coeff;
short *zbin_ptr = b->zbin;
short *round_ptr = b->round;
- short *quant_ptr = b->quant;
+ short *quant_ptr = b->quant_fast;
short *quant_shift_ptr = b->quant_shift;
short *qcoeff_ptr = d->qcoeff;
short *dqcoeff_ptr = d->dqcoeff;
short *quant_shift_ptr = b->quant_shift;
short *qcoeff_ptr = d->qcoeff;
short *dqcoeff_ptr = d->dqcoeff;
int x, y, z, sz;
short *coeff_ptr = b->coeff;
short *round_ptr = b->round;
int x, y, z, sz;
short *coeff_ptr = b->coeff;
short *round_ptr = b->round;
- short *quant_ptr = b->quant;
+ short *quant_ptr = b->quant_fast;
short *qcoeff_ptr = d->qcoeff;
short *dqcoeff_ptr = d->dqcoeff;
short *dequant_ptr = d->dequant;
short *qcoeff_ptr = d->qcoeff;
short *dqcoeff_ptr = d->dqcoeff;
short *dequant_ptr = d->dequant;
short *coeff_ptr = b->coeff;
short *zbin_ptr = b->zbin;
short *round_ptr = b->round;
short *coeff_ptr = b->coeff;
short *zbin_ptr = b->zbin;
short *round_ptr = b->round;
- short *quant_ptr = b->quant;
+ short *quant_ptr = b->quant_fast;
short *qcoeff_ptr = d->qcoeff;
short *dqcoeff_ptr = d->dqcoeff;
short *dequant_ptr = d->dequant;
short *qcoeff_ptr = d->qcoeff;
short *dqcoeff_ptr = d->dqcoeff;
short *dequant_ptr = d->dequant;
short *scan_mask = vp8_default_zig_zag_mask;//d->scan_order_mask_ptr;
short *coeff_ptr = b->coeff;
short *round_ptr = b->round;
short *scan_mask = vp8_default_zig_zag_mask;//d->scan_order_mask_ptr;
short *coeff_ptr = b->coeff;
short *round_ptr = b->round;
- short *quant_ptr = b->quant;
+ short *quant_ptr = b->quant_fast;
short *qcoeff_ptr = d->qcoeff;
short *dqcoeff_ptr = d->dqcoeff;
short *dequant_ptr = d->dequant;
short *qcoeff_ptr = d->qcoeff;
short *dqcoeff_ptr = d->dqcoeff;
short *dequant_ptr = d->dequant;
d->qcoeff,
d->dequant,
b->round,
d->qcoeff,
d->dequant,
b->round,