diff options
| -rw-r--r-- | vp8/vp8_cx_iface.c | 2 | ||||
| -rw-r--r-- | vp9/encoder/vp9_encodemb.c | 83 | ||||
| -rw-r--r-- | vp9/encoder/vp9_mcomp.c | 126 | ||||
| -rw-r--r-- | vp9/encoder/vp9_onyx_if.c | 18 | ||||
| -rw-r--r-- | vp9/encoder/vp9_onyx_int.h | 5 | ||||
| -rw-r--r-- | vp9/encoder/vp9_rdopt.c | 36 | ||||
| -rw-r--r-- | vp9/encoder/vp9_speed_features.c | 4 | ||||
| -rw-r--r-- | vp9/vp9_cx_iface.c | 14 | ||||
| -rw-r--r-- | vpx/vpx_encoder.h | 13 | ||||
| -rw-r--r-- | vpxenc.c | 18 |
10 files changed, 149 insertions, 170 deletions
diff --git a/vp8/vp8_cx_iface.c b/vp8/vp8_cx_iface.c index 4c896b1d1..6ca6087b2 100644 --- a/vp8/vp8_cx_iface.c +++ b/vp8/vp8_cx_iface.c @@ -1235,6 +1235,8 @@ static vpx_codec_enc_cfg_map_t vp8e_usage_cfg_map[] = 0, /* rc_dropframe_thresh */ 0, /* rc_resize_allowed */ + 1, /* rc_scaled_width */ + 1, /* rc_scaled_height */ 60, /* rc_resize_down_thresold */ 30, /* rc_resize_up_thresold */ diff --git a/vp9/encoder/vp9_encodemb.c b/vp9/encoder/vp9_encodemb.c index 2bea93f0f..b0c014eef 100644 --- a/vp9/encoder/vp9_encodemb.c +++ b/vp9/encoder/vp9_encodemb.c @@ -63,24 +63,17 @@ void vp9_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) { } #define RDTRUNC(RM, DM, R, D) ((128 + (R) * (RM)) & 0xFF) -typedef struct vp9_token_state vp9_token_state; -struct vp9_token_state { +typedef struct vp9_token_state { int rate; int error; int next; signed char token; short qc; -}; +} vp9_token_state; // TODO(jimbankoski): experiment to find optimal RD numbers. -#define Y1_RD_MULT 4 -#define UV_RD_MULT 2 - -static const int plane_rd_mult[4] = { - Y1_RD_MULT, - UV_RD_MULT, -}; +static const int plane_rd_mult[PLANE_TYPES] = { 4, 2 }; #define UPDATE_RD_COST()\ {\ @@ -108,56 +101,53 @@ static int trellis_get_coeff_context(const int16_t *scan, static int optimize_b(MACROBLOCK *mb, int plane, int block, BLOCK_SIZE plane_bsize, TX_SIZE tx_size, int ctx) { MACROBLOCKD *const xd = &mb->e_mbd; - struct macroblock_plane *p = &mb->plane[plane]; - struct macroblockd_plane *pd = &xd->plane[plane]; + struct macroblock_plane *const p = &mb->plane[plane]; + struct macroblockd_plane *const pd = &xd->plane[plane]; const int ref = is_inter_block(&xd->mi[0]->mbmi); vp9_token_state tokens[1025][2]; unsigned best_index[1025][2]; - const int16_t *coeff = BLOCK_OFFSET(mb->plane[plane].coeff, block); - int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block); - int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block); - int eob = p->eobs[block], final_eob, sz = 0; - const int i0 = 0; - int rc, x, next, i; - int64_t rdmult, rddiv, rd_cost0, rd_cost1; - int rate0, rate1, error0, error1, t0, t1; - int best, band, pt; - PLANE_TYPE type = pd->plane_type; - int err_mult = plane_rd_mult[type]; + uint8_t token_cache[1024]; + const int16_t *const coeff = BLOCK_OFFSET(mb->plane[plane].coeff, block); + int16_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block); + int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block); + const int eob = p->eobs[block]; + const PLANE_TYPE type = pd->plane_type; const int default_eob = 16 << (tx_size << 1); const int mul = 1 + (tx_size == TX_32X32); - uint8_t token_cache[1024]; const int16_t *dequant_ptr = pd->dequant; const uint8_t *const band_translate = get_band_translate(tx_size); - const scan_order *so = get_scan(xd, tx_size, type, block); - const int16_t *scan = so->scan; - const int16_t *nb = so->neighbors; + const scan_order *const so = get_scan(xd, tx_size, type, block); + const int16_t *const scan = so->scan; + const int16_t *const nb = so->neighbors; + int next = eob, sz = 0; + int64_t rdmult = mb->rdmult * plane_rd_mult[type], rddiv = mb->rddiv; + int64_t rd_cost0, rd_cost1; + int rate0, rate1, error0, error1, t0, t1; + int best, band, pt, i, final_eob; assert((!type && !plane) || (type && plane)); assert(eob <= default_eob); /* Now set up a Viterbi trellis to evaluate alternative roundings. */ - rdmult = mb->rdmult * err_mult; - if (!is_inter_block(&mb->e_mbd.mi[0]->mbmi)) + if (!ref) rdmult = (rdmult * 9) >> 4; - rddiv = mb->rddiv; + /* Initialize the sentinel node of the trellis. */ tokens[eob][0].rate = 0; tokens[eob][0].error = 0; tokens[eob][0].next = default_eob; tokens[eob][0].token = EOB_TOKEN; tokens[eob][0].qc = 0; - *(tokens[eob] + 1) = *(tokens[eob] + 0); - next = eob; + tokens[eob][1] = tokens[eob][0]; + for (i = 0; i < eob; i++) - token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[ - qcoeff[scan[i]]].token]; + token_cache[scan[i]] = + vp9_pt_energy_class[vp9_dct_value_tokens_ptr[qcoeff[scan[i]]].token]; - for (i = eob; i-- > i0;) { + for (i = eob; i-- > 0;) { int base_bits, d2, dx; - - rc = scan[i]; - x = qcoeff[rc]; + const int rc = scan[i]; + int x = qcoeff[rc]; /* Only add a trellis state for non-zero coefficients. */ if (x) { int shortcut = 0; @@ -179,7 +169,7 @@ static int optimize_b(MACROBLOCK *mb, int plane, int block, UPDATE_RD_COST(); /* And pick the best. */ best = rd_cost1 < rd_cost0; - base_bits = *(vp9_dct_value_cost_ptr + x); + base_bits = vp9_dct_value_cost_ptr[x]; dx = mul * (dqcoeff[rc] - coeff[rc]); d2 = dx * dx; tokens[i][0].rate = base_bits + (best ? rate1 : rate0); @@ -193,9 +183,9 @@ static int optimize_b(MACROBLOCK *mb, int plane, int block, rate0 = tokens[next][0].rate; rate1 = tokens[next][1].rate; - if ((abs(x)*dequant_ptr[rc != 0] > abs(coeff[rc]) * mul) && - (abs(x)*dequant_ptr[rc != 0] < abs(coeff[rc]) * mul + - dequant_ptr[rc != 0])) + if ((abs(x) * dequant_ptr[rc != 0] > abs(coeff[rc]) * mul) && + (abs(x) * dequant_ptr[rc != 0] < abs(coeff[rc]) * mul + + dequant_ptr[rc != 0])) shortcut = 1; else shortcut = 0; @@ -232,7 +222,7 @@ static int optimize_b(MACROBLOCK *mb, int plane, int block, UPDATE_RD_COST(); /* And pick the best. */ best = rd_cost1 < rd_cost0; - base_bits = *(vp9_dct_value_cost_ptr + x); + base_bits = vp9_dct_value_cost_ptr[x]; if (shortcut) { dx -= (dequant_ptr[rc != 0] + sz) ^ sz; @@ -281,15 +271,16 @@ static int optimize_b(MACROBLOCK *mb, int plane, int block, rate1 += mb->token_costs[tx_size][type][ref][band][0][ctx][t1]; UPDATE_RD_COST(); best = rd_cost1 < rd_cost0; - final_eob = i0 - 1; + final_eob = -1; vpx_memset(qcoeff, 0, sizeof(*qcoeff) * (16 << (tx_size * 2))); vpx_memset(dqcoeff, 0, sizeof(*dqcoeff) * (16 << (tx_size * 2))); for (i = next; i < eob; i = next) { - x = tokens[i][best].qc; + const int x = tokens[i][best].qc; + const int rc = scan[i]; if (x) { final_eob = i; } - rc = scan[i]; + qcoeff[rc] = x; dqcoeff[rc] = (x * dequant_ptr[rc != 0]) / mul; diff --git a/vp9/encoder/vp9_mcomp.c b/vp9/encoder/vp9_mcomp.c index f5511ffa3..8a7901172 100644 --- a/vp9/encoder/vp9_mcomp.c +++ b/vp9/encoder/vp9_mcomp.c @@ -66,7 +66,7 @@ int vp9_init_search_range(VP9_COMP *cpi, int size) { } static INLINE int mv_cost(const MV *mv, - const int *joint_cost, int *comp_cost[2]) { + const int *joint_cost, int *const comp_cost[2]) { return joint_cost[vp9_get_mv_joint(mv)] + comp_cost[0][mv->row] + comp_cost[1][mv->col]; } @@ -90,14 +90,13 @@ static int mv_err_cost(const MV *mv, const MV *ref, return 0; } -static int mvsad_err_cost(const MV *mv, const MV *ref, - const int *mvjsadcost, int *mvsadcost[2], +static int mvsad_err_cost(const MACROBLOCK *x, const MV *mv, const MV *ref, int error_per_bit) { - if (mvsadcost) { + if (x->nmvsadcost) { const MV diff = { mv->row - ref->row, mv->col - ref->col }; - return ROUND_POWER_OF_TWO(mv_cost(&diff, mvjsadcost, mvsadcost) * - error_per_bit, 8); + return ROUND_POWER_OF_TWO(mv_cost(&diff, x->nmvjointsadcost, + x->nmvsadcost) * error_per_bit, 8); } return 0; } @@ -478,8 +477,7 @@ static INLINE int is_mv_in(const MACROBLOCK *x, const MV *mv) { {\ if (thissad < bestsad) {\ if (use_mvcost) \ - thissad += mvsad_err_cost(&this_mv, &fcenter_mv, \ - mvjsadcost, mvsadcost, sad_per_bit);\ + thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);\ if (thissad < bestsad) {\ bestsad = thissad;\ best_site = i;\ @@ -519,9 +517,6 @@ static int vp9_pattern_search(const MACROBLOCK *x, int k = -1; const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3}; int best_init_s = search_param_to_steps[search_param]; - const int *const mvjsadcost = x->nmvjointsadcost; - int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; - // adjust ref_mv to make sure it is within MV range clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max); br = ref_mv->row; @@ -530,8 +525,8 @@ static int vp9_pattern_search(const MACROBLOCK *x, // Work out the start point for the search bestsad = vfp->sdf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv), in_what->stride, - 0x7fffffff) + mvsad_err_cost(ref_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); + 0x7fffffff) + mvsad_err_cost(x, ref_mv, &fcenter_mv, + sad_per_bit); // Search all possible scales upto the search param around the center point // pick the scale of the point that is best as the starting scale of @@ -891,10 +886,6 @@ int vp9_full_range_search_c(const MACROBLOCK *x, MV *ref_mv, MV *best_mv, unsigned int thissad; const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3}; - - const int *mvjsadcost = x->nmvjointsadcost; - int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; - int tr, tc; int best_tr = 0; int best_tc = 0; @@ -916,8 +907,7 @@ int vp9_full_range_search_c(const MACROBLOCK *x, MV *ref_mv, MV *best_mv, // Check the starting position bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff) - + mvsad_err_cost(best_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); + + mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit); start_row = MAX(-range, x->mv_row_min - ref_row); start_col = MAX(-range, x->mv_col_min - ref_col); @@ -938,8 +928,7 @@ int vp9_full_range_search_c(const MACROBLOCK *x, MV *ref_mv, MV *best_mv, if (sad_array[i] < bestsad) { const MV this_mv = {ref_row + tr, ref_col + tc + i}; thissad = sad_array[i] + - mvsad_err_cost(&this_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); + mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); if (thissad < bestsad) { bestsad = thissad; best_tr = tr; @@ -955,8 +944,7 @@ int vp9_full_range_search_c(const MACROBLOCK *x, MV *ref_mv, MV *best_mv, if (thissad < bestsad) { const MV this_mv = {ref_row + tr, ref_col + tc + i}; - thissad += mvsad_err_cost(&this_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); + thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); if (thissad < bestsad) { bestsad = thissad; @@ -988,8 +976,6 @@ int vp9_diamond_search_sad_c(const MACROBLOCK *x, const search_site *const ss = &x->ss[search_param * x->searches_per_step]; const int tot_steps = (x->ss_count / x->searches_per_step) - search_param; const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3}; - const int *mvjsadcost = x->nmvjointsadcost; - int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; const uint8_t *best_address, *in_what_ref; int best_sad = INT_MAX; int best_site = 0; @@ -1005,7 +991,7 @@ int vp9_diamond_search_sad_c(const MACROBLOCK *x, // Check the starting position best_sad = fn_ptr->sdf(what->buf, what->stride, best_address, in_what->stride, 0x7fffffff) + - mvsad_err_cost(best_mv, &fcenter_mv, mvjsadcost, mvsadcost, sad_per_bit); + mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit); i = 1; @@ -1018,8 +1004,7 @@ int vp9_diamond_search_sad_c(const MACROBLOCK *x, best_address + ss[i].offset, in_what->stride, best_sad); if (sad < best_sad) { - sad += mvsad_err_cost(&mv, &fcenter_mv, mvjsadcost, mvsadcost, - sad_per_bit); + sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit); if (sad < best_sad) { best_sad = sad; best_site = i; @@ -1044,8 +1029,7 @@ int vp9_diamond_search_sad_c(const MACROBLOCK *x, best_address + ss[best_site].offset, in_what->stride, best_sad); if (sad < best_sad) { - sad += mvsad_err_cost(&this_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); + sad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); if (sad < best_sad) { best_sad = sad; best_mv->row += ss[best_site].mv.row; @@ -1095,10 +1079,6 @@ int vp9_diamond_search_sadx4(const MACROBLOCK *x, const int tot_steps = (x->ss_count / x->searches_per_step) - search_param; const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3}; - - const int *mvjsadcost = x->nmvjointsadcost; - int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; - clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max); ref_row = ref_mv->row; ref_col = ref_mv->col; @@ -1112,8 +1092,7 @@ int vp9_diamond_search_sadx4(const MACROBLOCK *x, // Check the starting position bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff) - + mvsad_err_cost(best_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); + + mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit); i = 1; @@ -1146,9 +1125,8 @@ int vp9_diamond_search_sadx4(const MACROBLOCK *x, if (sad_array[t] < bestsad) { const MV this_mv = {best_mv->row + ss[i].mv.row, best_mv->col + ss[i].mv.col}; - sad_array[t] += mvsad_err_cost(&this_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); - + sad_array[t] += mvsad_err_cost(x, &this_mv, &fcenter_mv, + sad_per_bit); if (sad_array[t] < bestsad) { bestsad = sad_array[t]; best_site = i; @@ -1168,9 +1146,7 @@ int vp9_diamond_search_sadx4(const MACROBLOCK *x, in_what_stride, bestsad); if (thissad < bestsad) { - thissad += mvsad_err_cost(&this_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); - + thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); if (thissad < bestsad) { bestsad = thissad; best_site = i; @@ -1289,12 +1265,10 @@ int vp9_full_search_sad_c(const MACROBLOCK *x, const MV *ref_mv, const int row_max = MIN(ref_mv->row + distance, x->mv_row_max); const int col_min = MAX(ref_mv->col - distance, x->mv_col_min); const int col_max = MIN(ref_mv->col + distance, x->mv_col_max); - const int *mvjsadcost = x->nmvjointsadcost; - int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3}; int best_sad = fn_ptr->sdf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv), in_what->stride, 0x7fffffff) + - mvsad_err_cost(ref_mv, &fcenter_mv, mvjsadcost, mvsadcost, sad_per_bit); + mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit); *best_mv = *ref_mv; for (r = row_min; r < row_max; ++r) { @@ -1302,9 +1276,7 @@ int vp9_full_search_sad_c(const MACROBLOCK *x, const MV *ref_mv, const MV mv = {r, c}; const int sad = fn_ptr->sdf(what->buf, what->stride, get_buf_from_mv(in_what, &mv), in_what->stride, best_sad) + - mvsad_err_cost(&mv, &fcenter_mv, mvjsadcost, mvsadcost, - sad_per_bit); - + mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit); if (sad < best_sad) { best_sad = sad; *best_mv = mv; @@ -1338,8 +1310,6 @@ int vp9_full_search_sadx3(const MACROBLOCK *x, const MV *ref_mv, const int col_max = MIN(ref_col + distance, x->mv_col_max); unsigned int sad_array[3]; const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3}; - const int *mvjsadcost = x->nmvjointsadcost; - int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; // Work out the mid point for the search const uint8_t *bestaddress = &in_what[ref_row * in_what_stride + ref_col]; @@ -1350,8 +1320,7 @@ int vp9_full_search_sadx3(const MACROBLOCK *x, const MV *ref_mv, // Baseline value at the centre bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride, 0x7fffffff) - + mvsad_err_cost(best_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); + + mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit); for (r = row_min; r < row_max; r++) { const uint8_t *check_here = &in_what[r * in_what_stride + col_min]; @@ -1368,9 +1337,7 @@ int vp9_full_search_sadx3(const MACROBLOCK *x, const MV *ref_mv, if (thissad < bestsad) { this_mv.col = c; - thissad += mvsad_err_cost(&this_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); - + thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); if (thissad < bestsad) { bestsad = thissad; best_mv->row = r; @@ -1388,9 +1355,7 @@ int vp9_full_search_sadx3(const MACROBLOCK *x, const MV *ref_mv, if (thissad < bestsad) { this_mv.col = c; - thissad += mvsad_err_cost(&this_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); - + thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); if (thissad < bestsad) { bestsad = thissad; best_mv->row = r; @@ -1430,9 +1395,6 @@ int vp9_full_search_sadx8(const MACROBLOCK *x, const MV *ref_mv, unsigned int sad_array[3]; const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3}; - const int *mvjsadcost = x->nmvjointsadcost; - int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; - // Work out the mid point for the search const uint8_t *bestaddress = &in_what[ref_row * in_what_stride + ref_col]; @@ -1442,8 +1404,7 @@ int vp9_full_search_sadx8(const MACROBLOCK *x, const MV *ref_mv, // Baseline value at the center bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride, 0x7fffffff) - + mvsad_err_cost(best_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); + + mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit); for (r = row_min; r < row_max; r++) { const uint8_t *check_here = &in_what[r * in_what_stride + col_min]; @@ -1460,9 +1421,7 @@ int vp9_full_search_sadx8(const MACROBLOCK *x, const MV *ref_mv, if (thissad < bestsad) { this_mv.col = c; - thissad += mvsad_err_cost(&this_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); - + thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); if (thissad < bestsad) { bestsad = thissad; best_mv->row = r; @@ -1485,9 +1444,7 @@ int vp9_full_search_sadx8(const MACROBLOCK *x, const MV *ref_mv, if (thissad < bestsad) { this_mv.col = c; - thissad += mvsad_err_cost(&this_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); - + thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); if (thissad < bestsad) { bestsad = thissad; best_mv->row = r; @@ -1506,9 +1463,7 @@ int vp9_full_search_sadx8(const MACROBLOCK *x, const MV *ref_mv, if (thissad < bestsad) { this_mv.col = c; - thissad += mvsad_err_cost(&this_mv, &fcenter_mv, - mvjsadcost, mvsadcost, sad_per_bit); - + thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); if (thissad < bestsad) { bestsad = thissad; best_mv->row = r; @@ -1533,13 +1488,10 @@ int vp9_refining_search_sad_c(const MACROBLOCK *x, const struct buf_2d *const what = &x->plane[0].src; const struct buf_2d *const in_what = &xd->plane[0].pre[0]; const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3}; - const int *mvjsadcost = x->nmvjointsadcost; - int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; - unsigned int best_sad = fn_ptr->sdf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv), in_what->stride, 0x7fffffff) + - mvsad_err_cost(ref_mv, &fcenter_mv, mvjsadcost, mvsadcost, error_per_bit); + mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit); int i, j; for (i = 0; i < search_range; i++) { @@ -1552,8 +1504,7 @@ int vp9_refining_search_sad_c(const MACROBLOCK *x, unsigned int sad = fn_ptr->sdf(what->buf, what->stride, get_buf_from_mv(in_what, &mv), in_what->stride, best_sad); if (sad < best_sad) { - sad += mvsad_err_cost(&mv, &fcenter_mv, mvjsadcost, mvsadcost, - error_per_bit); + sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit); if (sad < best_sad) { best_sad = sad; best_site = j; @@ -1582,12 +1533,10 @@ int vp9_refining_search_sadx4(const MACROBLOCK *x, const struct buf_2d *const what = &x->plane[0].src; const struct buf_2d *const in_what = &xd->plane[0].pre[0]; const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3}; - const int *mvjsadcost = x->nmvjointsadcost; - int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; const uint8_t *best_address = get_buf_from_mv(in_what, ref_mv); unsigned int best_sad = fn_ptr->sdf(what->buf, what->stride, best_address, in_what->stride, 0x7fffffff) + - mvsad_err_cost(ref_mv, &fcenter_mv, mvjsadcost, mvsadcost, error_per_bit); + mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit); int i, j; for (i = 0; i < search_range; i++) { @@ -1612,9 +1561,7 @@ int vp9_refining_search_sadx4(const MACROBLOCK *x, if (sads[j] < best_sad) { const MV mv = {ref_mv->row + neighbors[j].row, ref_mv->col + neighbors[j].col}; - sads[j] += mvsad_err_cost(&mv, &fcenter_mv, - mvjsadcost, mvsadcost, error_per_bit); - + sads[j] += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit); if (sads[j] < best_sad) { best_sad = sads[j]; best_site = j; @@ -1631,9 +1578,7 @@ int vp9_refining_search_sadx4(const MACROBLOCK *x, get_buf_from_mv(in_what, &mv), in_what->stride, best_sad); if (sad < best_sad) { - sad += mvsad_err_cost(&mv, &fcenter_mv, - mvjsadcost, mvsadcost, error_per_bit); - + sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit); if (sad < best_sad) { best_sad = sad; best_site = j; @@ -1669,12 +1614,10 @@ int vp9_refining_search_8p_c(const MACROBLOCK *x, const struct buf_2d *const what = &x->plane[0].src; const struct buf_2d *const in_what = &xd->plane[0].pre[0]; const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3}; - const int *mvjsadcost = x->nmvjointsadcost; - int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; unsigned int best_sad = fn_ptr->sdaf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv), in_what->stride, second_pred, 0x7fffffff) + - mvsad_err_cost(ref_mv, &fcenter_mv, mvjsadcost, mvsadcost, error_per_bit); + mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit); int i, j; for (i = 0; i < search_range; ++i) { @@ -1689,8 +1632,7 @@ int vp9_refining_search_8p_c(const MACROBLOCK *x, get_buf_from_mv(in_what, &mv), in_what->stride, second_pred, best_sad); if (sad < best_sad) { - sad += mvsad_err_cost(&mv, &fcenter_mv, - mvjsadcost, mvsadcost, error_per_bit); + sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit); if (sad < best_sad) { best_sad = sad; best_site = j; diff --git a/vp9/encoder/vp9_onyx_if.c b/vp9/encoder/vp9_onyx_if.c index dc9e4dadb..4b2aa8735 100644 --- a/vp9/encoder/vp9_onyx_if.c +++ b/vp9/encoder/vp9_onyx_if.c @@ -815,10 +815,7 @@ void vp9_change_config(struct VP9_COMP *cpi, const VP9_CONFIG *oxcf) { cm->display_width = cpi->oxcf.width; cm->display_height = cpi->oxcf.height; - // VP8 sharpness level mapping 0-7 (vs 0-10 in general VPx dialogs) - cpi->oxcf.sharpness = MIN(7, cpi->oxcf.sharpness); - - cpi->common.lf.sharpness_level = cpi->oxcf.sharpness; + cm->lf.sharpness_level = cpi->oxcf.sharpness; if (cpi->initial_width) { // Increasing the size of the frame beyond the first seen frame, or some @@ -838,10 +835,6 @@ void vp9_change_config(struct VP9_COMP *cpi, const VP9_CONFIG *oxcf) { cpi->speed = abs(cpi->oxcf.cpu_used); - // Limit on lag buffers as these are not currently dynamically allocated. - if (cpi->oxcf.lag_in_frames > MAX_LAG_BUFFERS) - cpi->oxcf.lag_in_frames = MAX_LAG_BUFFERS; - #if CONFIG_MULTIPLE_ARF vp9_zero(cpi->alt_ref_source); #else @@ -3004,6 +2997,15 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags, cm->frame_flags = *frame_flags; + if (cpi->pass == 2 && + cm->current_video_frame == 0 && + cpi->oxcf.allow_spatial_resampling && + cpi->oxcf.end_usage == USAGE_LOCAL_FILE_PLAYBACK) { + // Internal scaling is triggered on the first frame. + vp9_set_size_literal(cpi, cpi->oxcf.scaled_frame_width, + cpi->oxcf.scaled_frame_height); + } + // Reset the frame pointers to the current frame size vp9_realloc_frame_buffer(get_frame_new_buffer(cm), cm->width, cm->height, diff --git a/vp9/encoder/vp9_onyx_int.h b/vp9/encoder/vp9_onyx_int.h index 9ad6ae49e..ad56e2a85 100644 --- a/vp9/encoder/vp9_onyx_int.h +++ b/vp9/encoder/vp9_onyx_int.h @@ -232,6 +232,11 @@ typedef struct VP9_CONFIG { int lossless; AQ_MODE aq_mode; // Adaptive Quantization mode + // Internal frame size scaling. + int allow_spatial_resampling; + int scaled_frame_width; + int scaled_frame_height; + // Enable feature to reduce the frame quantization every x frames. int frame_periodic_boost; diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c index 1bf8621e7..7ef21fa5d 100644 --- a/vp9/encoder/vp9_rdopt.c +++ b/vp9/encoder/vp9_rdopt.c @@ -1677,14 +1677,14 @@ static INLINE int mv_has_subpel(const MV *mv) { static int check_best_zero_mv( const VP9_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES], int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], - int disable_inter_mode_mask, int this_mode, int ref_frame, - int second_ref_frame) { + int disable_inter_mode_mask, int this_mode, + const MV_REFERENCE_FRAME ref_frames[2]) { if (!(disable_inter_mode_mask & (1 << INTER_OFFSET(ZEROMV))) && (this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) && - frame_mv[this_mode][ref_frame].as_int == 0 && - (second_ref_frame == NONE || - frame_mv[this_mode][second_ref_frame].as_int == 0)) { - int rfc = mode_context[ref_frame]; + frame_mv[this_mode][ref_frames[0]].as_int == 0 && + (ref_frames[1] == NONE || + frame_mv[this_mode][ref_frames[1]].as_int == 0)) { + int rfc = mode_context[ref_frames[0]]; int c1 = cost_mv_ref(cpi, NEARMV, rfc); int c2 = cost_mv_ref(cpi, NEARESTMV, rfc); int c3 = cost_mv_ref(cpi, ZEROMV, rfc); @@ -1695,15 +1695,15 @@ static int check_best_zero_mv( if (c2 > c3) return 0; } else { assert(this_mode == ZEROMV); - if (second_ref_frame == NONE) { - if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frame].as_int == 0) || - (c3 >= c1 && frame_mv[NEARMV][ref_frame].as_int == 0)) + if (ref_frames[1] == NONE) { + if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) || + (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0)) return 0; } else { - if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frame].as_int == 0 && - frame_mv[NEARESTMV][second_ref_frame].as_int == 0) || - (c3 >= c1 && frame_mv[NEARMV][ref_frame].as_int == 0 && - frame_mv[NEARMV][second_ref_frame].as_int == 0)) + if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 && + frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) || + (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 && + frame_mv[NEARMV][ref_frames[1]].as_int == 0)) return 0; } } @@ -1781,8 +1781,7 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x, if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv, disable_inter_mode_mask, - this_mode, mbmi->ref_frame[0], - mbmi->ref_frame[1])) + this_mode, mbmi->ref_frame)) continue; vpx_memcpy(orig_pre, pd->pre, sizeof(orig_pre)); @@ -3379,11 +3378,12 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, } } else { if (x->in_active_map && - !vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) + !vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) { + const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame}; if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv, - disable_inter_mode_mask, this_mode, ref_frame, - second_ref_frame)) + disable_inter_mode_mask, this_mode, ref_frames)) continue; + } } mbmi->mode = this_mode; diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c index adad800b6..f1fefe4ef 100644 --- a/vp9/encoder/vp9_speed_features.c +++ b/vp9/encoder/vp9_speed_features.c @@ -73,8 +73,8 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm, } if (speed >= 2) { - sf->tx_size_search_method = vp9_frame_is_boosted(cpi) ? USE_FULL_RD - : USE_LARGESTALL; + sf->tx_size_search_method = frame_is_intra_only(cm) ? USE_FULL_RD + : USE_LARGESTALL; if (MIN(cm->width, cm->height) >= 720) sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT diff --git a/vp9/vp9_cx_iface.c b/vp9/vp9_cx_iface.c index 0623ad132..967431c0f 100644 --- a/vp9/vp9_cx_iface.c +++ b/vp9/vp9_cx_iface.c @@ -168,6 +168,11 @@ static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx, RANGE_CHECK_HI(cfg, rc_resize_down_thresh, 100); RANGE_CHECK(cfg, g_pass, VPX_RC_ONE_PASS, VPX_RC_LAST_PASS); + if (cfg->rc_resize_allowed == 1) { + RANGE_CHECK(cfg, rc_scaled_width, 1, cfg->g_w); + RANGE_CHECK(cfg, rc_scaled_height, 1, cfg->g_h); + } + RANGE_CHECK(cfg, ss_number_layers, 1, VPX_SS_MAX_LAYERS); RANGE_CHECK(cfg, ts_number_layers, 1, VPX_TS_MAX_LAYERS); if (cfg->ts_number_layers > 1) { @@ -332,6 +337,10 @@ static vpx_codec_err_t set_encoder_config( oxcf->under_shoot_pct = cfg->rc_undershoot_pct; oxcf->over_shoot_pct = cfg->rc_overshoot_pct; + oxcf->allow_spatial_resampling = cfg->rc_resize_allowed; + oxcf->scaled_frame_width = cfg->rc_scaled_width; + oxcf->scaled_frame_height = cfg->rc_scaled_height; + oxcf->maximum_buffer_size = cfg->rc_buf_sz; oxcf->starting_buffer_level = cfg->rc_buf_initial_sz; oxcf->optimal_buffer_level = cfg->rc_buf_optimal_sz; @@ -410,6 +419,9 @@ static vpx_codec_err_t set_encoder_config( printf("fixed_q: %d\n", oxcf->fixed_q); printf("worst_allowed_q: %d\n", oxcf->worst_allowed_q); printf("best_allowed_q: %d\n", oxcf->best_allowed_q); + printf("allow_spatial_resampling: %d\n", oxcf->allow_spatial_resampling); + printf("scaled_frame_width: %d\n", oxcf->scaled_frame_width); + printf("scaled_frame_height: %d\n", oxcf->scaled_frame_height); printf("two_pass_vbrbias: %d\n", oxcf->two_pass_vbrbias); printf("two_pass_vbrmin_section: %d\n", oxcf->two_pass_vbrmin_section); printf("two_pass_vbrmax_section: %d\n", oxcf->two_pass_vbrmax_section); @@ -1128,6 +1140,8 @@ static vpx_codec_enc_cfg_map_t encoder_usage_cfg_map[] = { 0, // rc_dropframe_thresh 0, // rc_resize_allowed + 1, // rc_scaled_width + 1, // rc_scaled_height 60, // rc_resize_down_thresold 30, // rc_resize_up_thresold diff --git a/vpx/vpx_encoder.h b/vpx/vpx_encoder.h index 2c882c19a..571ad3fc5 100644 --- a/vpx/vpx_encoder.h +++ b/vpx/vpx_encoder.h @@ -396,6 +396,19 @@ extern "C" { */ unsigned int rc_resize_allowed; + /*!\brief Internal coded frame width. + * + * If spatial resampling is enabled this specifies the width of the + * encoded frame. + */ + unsigned int rc_scaled_width; + + /*!\brief Internal coded frame height. + * + * If spatial resampling is enabled this specifies the height of the + * encoded frame. + */ + unsigned int rc_scaled_height; /*!\brief Spatial resampling up watermark. * @@ -235,6 +235,10 @@ static const arg_def_t dropframe_thresh = ARG_DEF(NULL, "drop-frame", 1, "Temporal resampling threshold (buf %)"); static const arg_def_t resize_allowed = ARG_DEF(NULL, "resize-allowed", 1, "Spatial resampling enabled (bool)"); +static const arg_def_t resize_width = ARG_DEF(NULL, "resize-width", 1, + "Width of encoded frame"); +static const arg_def_t resize_height = ARG_DEF(NULL, "resize-height", 1, + "Height of encoded frame"); static const arg_def_t resize_up_thresh = ARG_DEF(NULL, "resize-up", 1, "Upscale threshold (buf %)"); static const arg_def_t resize_down_thresh = ARG_DEF(NULL, "resize-down", 1, @@ -265,10 +269,10 @@ static const arg_def_t buf_initial_sz = ARG_DEF(NULL, "buf-initial-sz", 1, static const arg_def_t buf_optimal_sz = ARG_DEF(NULL, "buf-optimal-sz", 1, "Client optimal buffer size (ms)"); static const arg_def_t *rc_args[] = { - &dropframe_thresh, &resize_allowed, &resize_up_thresh, &resize_down_thresh, - &end_usage, &target_bitrate, &min_quantizer, &max_quantizer, - &undershoot_pct, &overshoot_pct, &buf_sz, &buf_initial_sz, &buf_optimal_sz, - NULL + &dropframe_thresh, &resize_allowed, &resize_width, &resize_height, + &resize_up_thresh, &resize_down_thresh, &end_usage, &target_bitrate, + &min_quantizer, &max_quantizer, &undershoot_pct, &overshoot_pct, &buf_sz, + &buf_initial_sz, &buf_optimal_sz, NULL }; @@ -883,6 +887,10 @@ static int parse_stream_params(struct VpxEncoderConfig *global, config->cfg.rc_dropframe_thresh = arg_parse_uint(&arg); } else if (arg_match(&arg, &resize_allowed, argi)) { config->cfg.rc_resize_allowed = arg_parse_uint(&arg); + } else if (arg_match(&arg, &resize_width, argi)) { + config->cfg.rc_scaled_width = arg_parse_uint(&arg); + } else if (arg_match(&arg, &resize_height, argi)) { + config->cfg.rc_scaled_height = arg_parse_uint(&arg); } else if (arg_match(&arg, &resize_up_thresh, argi)) { config->cfg.rc_resize_up_thresh = arg_parse_uint(&arg); } else if (arg_match(&arg, &resize_down_thresh, argi)) { @@ -1067,6 +1075,8 @@ static void show_stream_config(struct stream_state *stream, SHOW(g_lag_in_frames); SHOW(rc_dropframe_thresh); SHOW(rc_resize_allowed); + SHOW(rc_scaled_width); + SHOW(rc_scaled_height); SHOW(rc_resize_up_thresh); SHOW(rc_resize_down_thresh); SHOW(rc_end_usage); |