ref: 25435257000505b8af36bd4d46744c3f65266d0a
parent: 908e1218cdac2d849a635486c63c155194810d45
parent: a4c3e2e4f26b5f76ca75aa994c9466644bae6bdd
author: Dmitry Kovalev <dkovalev@google.com>
date: Mon Apr 14 06:32:58 EDT 2014
Merge "Cleaning up optimize_b()."
--- a/vp9/encoder/vp9_encodemb.c
+++ b/vp9/encoder/vp9_encodemb.c
@@ -63,25 +63,18 @@
}
#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[PLANE_TYPES] = { 4, 2 };-static const int plane_rd_mult[4] = {- Y1_RD_MULT,
- UV_RD_MULT,
-};
-
#define UPDATE_RD_COST()\
{\rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);\
@@ -108,39 +101,37 @@
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;
@@ -147,17 +138,16 @@
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 @@
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 @@
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 @@
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 @@
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;