Enable greedy version of av1_optimize_b().
This has been found to be better than the original version in both ways:
(1) Better compression: lowres -0.229, midres -0.147
(2) Faster too in my quick test over 5 different clips with 30
frames: 2.7% to 10.5% faster.
Change-Id: I4d46e0915d6e4b8e7bfc03d0c8b88cbe3351ca20
diff --git a/av1/encoder/encodemb.c b/av1/encoder/encodemb.c
index b0945ed..246b656 100644
--- a/av1/encoder/encodemb.c
+++ b/av1/encoder/encodemb.c
@@ -133,9 +133,6 @@
}
#if !CONFIG_LV_MAP
-#define USE_GREEDY_OPTIMIZE_B 0
-
-#if USE_GREEDY_OPTIMIZE_B
typedef struct av1_token_state_greedy {
int16_t token;
@@ -454,354 +451,6 @@
return final_eob;
}
-#else // USE_GREEDY_OPTIMIZE_B
-
-typedef struct av1_token_state_org {
- int64_t error;
- int rate;
- int16_t next;
- int16_t token;
- tran_low_t qc;
- tran_low_t dqc;
- uint8_t best_index;
-} av1_token_state_org;
-
-static int optimize_b_org(const AV1_COMMON *cm, MACROBLOCK *mb, int plane,
- int block, TX_SIZE tx_size, int ctx) {
- MACROBLOCKD *const xd = &mb->e_mbd;
- 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);
- av1_token_state_org tokens[MAX_TX_SQUARE + 1][2];
- uint8_t token_cache[MAX_TX_SQUARE];
- const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
- tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
- tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
- const int eob = p->eobs[block];
- const PLANE_TYPE plane_type = pd->plane_type;
- const int default_eob = tx_size_2d[tx_size];
- const int16_t *const dequant_ptr = pd->dequant;
- const uint8_t *const band_translate = get_band_translate(tx_size);
- TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
- const SCAN_ORDER *const scan_order =
- get_scan(cm, tx_size, tx_type, is_inter_block(&xd->mi[0]->mbmi));
- const int16_t *const scan = scan_order->scan;
- const int16_t *const nb = scan_order->neighbors;
- int dqv;
- const int shift = av1_get_tx_scale(tx_size);
-#if CONFIG_AOM_QM
- int seg_id = xd->mi[0]->mbmi.segment_id;
- const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!ref][tx_size];
-#endif
-#if CONFIG_NEW_QUANT
- int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
- const dequant_val_type_nuq *dequant_val = pd->dequant_val_nuq[dq];
-#endif // CONFIG_NEW_QUANT
- int next = eob, sz = 0;
- const int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;
- const int64_t rddiv = mb->rddiv;
- int64_t rd_cost0, rd_cost1;
- int rate0, rate1;
- int64_t error0, error1;
- int16_t t0, t1;
- int best, band = (eob < default_eob) ? band_translate[eob]
- : band_translate[eob - 1];
- int pt, i, final_eob;
- const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
- unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
- mb->token_costs[txsize_sqr_map[tx_size]][plane_type][ref];
- const uint16_t *band_counts = &band_count_table[tx_size][band];
- uint16_t band_left = eob - band_cum_count_table[tx_size][band] + 1;
- int shortcut = 0;
- int next_shortcut = 0;
-
-#if CONFIG_EXT_DELTA_Q
- const int qindex = cm->seg.enabled
- ? av1_get_qindex(&cm->seg, xd->mi[0]->mbmi.segment_id,
- cm->base_qindex)
- : cm->base_qindex;
- assert(qindex > 0);
- (void)qindex;
-#else
- assert(mb->qindex > 0);
-#endif
-
- token_costs += band;
-
- assert((!plane_type && !plane) || (plane_type && plane));
- assert(eob <= default_eob);
-
- /* Now set up a Viterbi trellis to evaluate alternative roundings. */
- /* 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];
-
- for (i = 0; i < eob; i++) {
- const int rc = scan[i];
- tokens[i][0].rate = av1_get_token_cost(qcoeff[rc], &t0, cat6_bits);
- tokens[i][0].token = t0;
- token_cache[rc] = av1_pt_energy_class[t0];
- }
-
- for (i = eob; i-- > 0;) {
- int base_bits, dx;
- int64_t d2;
- const int rc = scan[i];
- int x = qcoeff[rc];
-#if CONFIG_AOM_QM
- int iwt = iqmatrix[rc];
- dqv = dequant_ptr[rc != 0];
- dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
-#else
- dqv = dequant_ptr[rc != 0];
-#endif
- next_shortcut = shortcut;
-
- /* Only add a trellis state for non-zero coefficients. */
- if (UNLIKELY(x)) {
- error0 = tokens[next][0].error;
- error1 = tokens[next][1].error;
- /* Evaluate the first possibility for this state. */
- rate0 = tokens[next][0].rate;
- rate1 = tokens[next][1].rate;
-
- if (next_shortcut) {
- /* Consider both possible successor states. */
- if (next < default_eob) {
- pt = get_coef_context(nb, token_cache, i + 1);
- rate0 +=
- get_token_bit_costs(*token_costs, 0, pt, tokens[next][0].token);
- rate1 +=
- get_token_bit_costs(*token_costs, 0, pt, tokens[next][1].token);
- }
- UPDATE_RD_COST();
- /* And pick the best. */
- best = rd_cost1 < rd_cost0;
- } else {
- if (next < default_eob) {
- pt = get_coef_context(nb, token_cache, i + 1);
- rate0 +=
- get_token_bit_costs(*token_costs, 0, pt, tokens[next][0].token);
- }
- best = 0;
- }
-
- dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
-#if CONFIG_HIGHBITDEPTH
- if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
- dx >>= xd->bd - 8;
- }
-#endif // CONFIG_HIGHBITDEPTH
- d2 = (int64_t)dx * dx;
- tokens[i][0].rate += (best ? rate1 : rate0);
- tokens[i][0].error = d2 + (best ? error1 : error0);
- tokens[i][0].next = next;
- tokens[i][0].qc = x;
- tokens[i][0].dqc = dqcoeff[rc];
- tokens[i][0].best_index = best;
-
- /* Evaluate the second possibility for this state. */
- rate0 = tokens[next][0].rate;
- rate1 = tokens[next][1].rate;
-
- // The threshold of 3 is empirically obtained.
- if (UNLIKELY(abs(x) > 3)) {
- shortcut = 0;
- } else {
-#if CONFIG_NEW_QUANT
- shortcut = ((av1_dequant_abscoeff_nuq(abs(x), dqv,
- dequant_val[band_translate[i]]) >
- (abs(coeff[rc]) << shift)) &&
- (av1_dequant_abscoeff_nuq(abs(x) - 1, dqv,
- dequant_val[band_translate[i]]) <
- (abs(coeff[rc]) << shift)));
-#else // CONFIG_NEW_QUANT
-#if CONFIG_AOM_QM
- if ((abs(x) * dequant_ptr[rc != 0] * iwt >
- ((abs(coeff[rc]) << shift) << AOM_QM_BITS)) &&
- (abs(x) * dequant_ptr[rc != 0] * iwt <
- (((abs(coeff[rc]) << shift) + dequant_ptr[rc != 0])
- << AOM_QM_BITS)))
-#else
- if ((abs(x) * dequant_ptr[rc != 0] > (abs(coeff[rc]) << shift)) &&
- (abs(x) * dequant_ptr[rc != 0] <
- (abs(coeff[rc]) << shift) + dequant_ptr[rc != 0]))
-#endif // CONFIG_AOM_QM
- shortcut = 1;
- else
- shortcut = 0;
-#endif // CONFIG_NEW_QUANT
- }
-
- if (shortcut) {
- sz = -(x < 0);
- x -= 2 * sz + 1;
- } else {
- tokens[i][1] = tokens[i][0];
- next = i;
-
- if (UNLIKELY(!(--band_left))) {
- --band_counts;
- band_left = *band_counts;
- --token_costs;
- }
- continue;
- }
-
- /* Consider both possible successor states. */
- if (!x) {
- /* If we reduced this coefficient to zero, check to see if
- * we need to move the EOB back here.
- */
- t0 = tokens[next][0].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
- t1 = tokens[next][1].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
- base_bits = 0;
- } else {
- base_bits = av1_get_token_cost(x, &t0, cat6_bits);
- t1 = t0;
- }
-
- if (next_shortcut) {
- if (LIKELY(next < default_eob)) {
- if (t0 != EOB_TOKEN) {
- token_cache[rc] = av1_pt_energy_class[t0];
- pt = get_coef_context(nb, token_cache, i + 1);
- rate0 += get_token_bit_costs(*token_costs, !x, pt,
- tokens[next][0].token);
- }
- if (t1 != EOB_TOKEN) {
- token_cache[rc] = av1_pt_energy_class[t1];
- pt = get_coef_context(nb, token_cache, i + 1);
- rate1 += get_token_bit_costs(*token_costs, !x, pt,
- tokens[next][1].token);
- }
- }
-
- UPDATE_RD_COST();
- /* And pick the best. */
- best = rd_cost1 < rd_cost0;
- } else {
- // The two states in next stage are identical.
- if (next < default_eob && t0 != EOB_TOKEN) {
- token_cache[rc] = av1_pt_energy_class[t0];
- pt = get_coef_context(nb, token_cache, i + 1);
- rate0 +=
- get_token_bit_costs(*token_costs, !x, pt, tokens[next][0].token);
- }
- best = 0;
- }
-
-#if CONFIG_NEW_QUANT
- dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
- (coeff[rc] << shift);
-#if CONFIG_HIGHBITDEPTH
- if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
- dx >>= xd->bd - 8;
- }
-#endif // CONFIG_HIGHBITDEPTH
-#else // CONFIG_NEW_QUANT
-#if CONFIG_HIGHBITDEPTH
- if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
- dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
- } else {
- dx -= (dqv + sz) ^ sz;
- }
-#else
- dx -= (dqv + sz) ^ sz;
-#endif // CONFIG_HIGHBITDEPTH
-#endif // CONFIG_NEW_QUANT
- d2 = (int64_t)dx * dx;
-
- tokens[i][1].rate = base_bits + (best ? rate1 : rate0);
- tokens[i][1].error = d2 + (best ? error1 : error0);
- tokens[i][1].next = next;
- tokens[i][1].token = best ? t1 : t0;
- tokens[i][1].qc = x;
-
- if (x) {
-#if CONFIG_NEW_QUANT
- tokens[i][1].dqc = av1_dequant_abscoeff_nuq(
- abs(x), dqv, dequant_val[band_translate[i]]);
- tokens[i][1].dqc = shift ? ROUND_POWER_OF_TWO(tokens[i][1].dqc, shift)
- : tokens[i][1].dqc;
- if (sz) tokens[i][1].dqc = -tokens[i][1].dqc;
-#else
- if (x < 0)
- tokens[i][1].dqc = -((-x * dqv) >> shift);
- else
- tokens[i][1].dqc = (x * dqv) >> shift;
-#endif // CONFIG_NEW_QUANT
- } else {
- tokens[i][1].dqc = 0;
- }
-
- tokens[i][1].best_index = best;
- /* Finally, make this the new head of the trellis. */
- next = i;
- } else {
- /* There's no choice to make for a zero coefficient, so we don't
- * add a new trellis node, but we do need to update the costs.
- */
- t0 = tokens[next][0].token;
- t1 = tokens[next][1].token;
- pt = get_coef_context(nb, token_cache, i + 1);
- /* Update the cost of each path if we're past the EOB token. */
- if (t0 != EOB_TOKEN) {
- tokens[next][0].rate += get_token_bit_costs(*token_costs, 1, pt, t0);
- tokens[next][0].token = ZERO_TOKEN;
- }
- if (t1 != EOB_TOKEN) {
- tokens[next][1].rate += get_token_bit_costs(*token_costs, 1, pt, t1);
- tokens[next][1].token = ZERO_TOKEN;
- }
- tokens[i][0].best_index = tokens[i][1].best_index = 0;
- shortcut = (tokens[next][0].rate != tokens[next][1].rate);
- /* Don't update next, because we didn't add a new node. */
- }
-
- if (UNLIKELY(!(--band_left))) {
- --band_counts;
- band_left = *band_counts;
- --token_costs;
- }
- }
-
- /* Now pick the best path through the whole trellis. */
- rate0 = tokens[next][0].rate;
- rate1 = tokens[next][1].rate;
- error0 = tokens[next][0].error;
- error1 = tokens[next][1].error;
- t0 = tokens[next][0].token;
- t1 = tokens[next][1].token;
- rate0 += get_token_bit_costs(*token_costs, 0, ctx, t0);
- rate1 += get_token_bit_costs(*token_costs, 0, ctx, t1);
- UPDATE_RD_COST();
- best = rd_cost1 < rd_cost0;
-
- final_eob = -1;
-
- for (i = next; i < eob; i = next) {
- const int x = tokens[i][best].qc;
- const int rc = scan[i];
- if (x) final_eob = i;
- qcoeff[rc] = x;
- dqcoeff[rc] = tokens[i][best].dqc;
-
- next = tokens[i][best].next;
- best = tokens[i][best].best_index;
- }
- final_eob++;
-
- mb->plane[plane].eobs[block] = final_eob;
- assert(final_eob <= default_eob);
- return final_eob;
-}
-
-#endif // USE_GREEDY_OPTIMIZE_B
#endif // !CONFIG_LV_MAP
int av1_optimize_b(const AV1_COMMON *cm, MACROBLOCK *mb, int plane, int block,
@@ -827,13 +476,8 @@
int ctx = get_entropy_context(tx_size, a, l);
#else
int ctx = combine_entropy_contexts(*a, *l);
-#endif
-
-#if USE_GREEDY_OPTIMIZE_B
+#endif // CONFIG_VAR_TX
return optimize_b_greedy(cm, mb, plane, block, tx_size, ctx);
-#else // USE_GREEDY_OPTIMIZE_B
- return optimize_b_org(cm, mb, plane, block, tx_size, ctx);
-#endif // USE_GREEDY_OPTIMIZE_B
#else // !CONFIG_LV_MAP
TXB_CTX txb_ctx;
get_txb_ctx(plane_bsize, tx_size, plane, a, l, &txb_ctx);