Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(8)

Issues Search
    My Issues | Starred     Open | Closed | All

Unified Diff: source/libvpx/vp9/encoder/vp9_rdopt.c

Issue 23600008: libvpx: Pull from upstream (Closed) Base URL: svn://chrome-svn/chrome/trunk/deps/third_party/libvpx/
Patch Set: Created 7 years, 4 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View side-by-side diff with in-line comments
Download patch
« no previous file with comments | « source/libvpx/vp9/encoder/vp9_rdopt.h ('k') | source/libvpx/vp9/encoder/vp9_segmentation.c » ('j') | no next file with comments »
Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
Index: source/libvpx/vp9/encoder/vp9_rdopt.c
===================================================================
--- source/libvpx/vp9/encoder/vp9_rdopt.c (revision 219822)
+++ source/libvpx/vp9/encoder/vp9_rdopt.c (working copy)
@@ -8,7 +8,6 @@
* be found in the AUTHORS file in the root of the source tree.
*/
-
#include <stdio.h>
#include <math.h>
#include <limits.h>
@@ -54,6 +53,8 @@
const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
{NEARESTMV, LAST_FRAME, NONE},
+ {DC_PRED, INTRA_FRAME, NONE},
+
{NEARESTMV, ALTREF_FRAME, NONE},
{NEARESTMV, GOLDEN_FRAME, NONE},
{NEWMV, LAST_FRAME, NONE},
@@ -61,8 +62,6 @@
{NEARMV, LAST_FRAME, NONE},
{NEARESTMV, GOLDEN_FRAME, ALTREF_FRAME},
- {DC_PRED, INTRA_FRAME, NONE},
-
{NEWMV, GOLDEN_FRAME, NONE},
{NEWMV, ALTREF_FRAME, NONE},
{NEARMV, ALTREF_FRAME, NONE},
@@ -91,7 +90,7 @@
{H_PRED, INTRA_FRAME, NONE},
{V_PRED, INTRA_FRAME, NONE},
{D135_PRED, INTRA_FRAME, NONE},
- {D27_PRED, INTRA_FRAME, NONE},
+ {D207_PRED, INTRA_FRAME, NONE},
{D153_PRED, INTRA_FRAME, NONE},
{D63_PRED, INTRA_FRAME, NONE},
{D117_PRED, INTRA_FRAME, NONE},
@@ -102,14 +101,13 @@
// certain modes are assumed to be based on 8x8 blocks.
// This table is used to correct for blocks size.
// The factors here are << 2 (2 = x0.5, 32 = x8 etc).
-static int rd_thresh_block_size_factor[BLOCK_SIZE_TYPES] =
+static int rd_thresh_block_size_factor[BLOCK_SIZES] =
{2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32};
-#define BASE_RD_THRESH_FREQ_FACT 16
-#define MAX_RD_THRESH_FREQ_FACT 32
-#define MAX_RD_THRESH_FREQ_INC 1
+#define MAX_RD_THRESH_FACT 64
+#define RD_THRESH_INC 1
-static void fill_token_costs(vp9_coeff_count (*c)[BLOCK_TYPES][2],
+static void fill_token_costs(vp9_coeff_cost *c,
vp9_coeff_probs_model (*p)[BLOCK_TYPES]) {
int i, j, k, l;
TX_SIZE t;
@@ -120,12 +118,12 @@
for (l = 0; l < PREV_COEF_CONTEXTS; l++) {
vp9_prob probs[ENTROPY_NODES];
vp9_model_to_full_probs(p[t][i][j][k][l], probs);
- vp9_cost_tokens((int *)c[t][i][j][0][k][l], probs,
+ vp9_cost_tokens((int *)c[t][i][j][k][0][l], probs,
vp9_coef_tree);
- vp9_cost_tokens_skip((int *)c[t][i][j][1][k][l], probs,
+ vp9_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs,
vp9_coef_tree);
- assert(c[t][i][j][0][k][l][DCT_EOB_TOKEN] ==
- c[t][i][j][1][k][l][DCT_EOB_TOKEN]);
+ assert(c[t][i][j][k][0][l][DCT_EOB_TOKEN] ==
+ c[t][i][j][k][1][l][DCT_EOB_TOKEN]);
}
}
@@ -199,7 +197,7 @@
cpi->RDDIV = 1;
cpi->RDMULT /= 100;
- for (bsize = 0; bsize < BLOCK_SIZE_TYPES; ++bsize) {
+ for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
for (i = 0; i < MAX_MODES; ++i) {
// Threshold here seem unecessarily harsh but fine given actual
// range of values used for cpi->sf.thresh_mult[]
@@ -213,18 +211,12 @@
} else {
cpi->rd_threshes[bsize][i] = INT_MAX;
}
- cpi->rd_baseline_thresh[bsize][i] = cpi->rd_threshes[bsize][i];
-
- if (cpi->sf.adaptive_rd_thresh)
- cpi->rd_thresh_freq_fact[bsize][i] = MAX_RD_THRESH_FREQ_FACT;
- else
- cpi->rd_thresh_freq_fact[bsize][i] = BASE_RD_THRESH_FREQ_FACT;
}
}
} else {
cpi->RDDIV = 100;
- for (bsize = 0; bsize < BLOCK_SIZE_TYPES; ++bsize) {
+ for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
for (i = 0; i < MAX_MODES; i++) {
// Threshold here seem unecessarily harsh but fine given actual
// range of values used for cpi->sf.thresh_mult[]
@@ -237,12 +229,6 @@
} else {
cpi->rd_threshes[bsize][i] = INT_MAX;
}
- cpi->rd_baseline_thresh[bsize][i] = cpi->rd_threshes[bsize][i];
-
- if (cpi->sf.adaptive_rd_thresh)
- cpi->rd_thresh_freq_fact[bsize][i] = MAX_RD_THRESH_FREQ_FACT;
- else
- cpi->rd_thresh_freq_fact[bsize][i] = BASE_RD_THRESH_FREQ_FACT;
}
}
}
@@ -277,16 +263,6 @@
}
}
-static INLINE BLOCK_SIZE_TYPE get_block_size(int bwl, int bhl) {
- return bsize_from_dim_lookup[bwl][bhl];
-}
-
-static BLOCK_SIZE_TYPE get_plane_block_size(BLOCK_SIZE_TYPE bsize,
- struct macroblockd_plane *pd) {
- return get_block_size(plane_block_width_log2by4(bsize, pd),
- plane_block_height_log2by4(bsize, pd));
-}
-
static INLINE void linear_interpolate2(double x, int ntab, int inv_step,
const double *tab1, const double *tab2,
double *v1, double *v2) {
@@ -388,7 +364,7 @@
vp9_clear_system_state();
}
-static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE_TYPE bsize,
+static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
MACROBLOCK *x, MACROBLOCKD *xd,
int *out_rate_sum, int64_t *out_dist_sum) {
// Note our transform coeffs are 8 times an orthogonal transform.
@@ -399,18 +375,14 @@
for (i = 0; i < MAX_MB_PLANE; ++i) {
struct macroblock_plane *const p = &x->plane[i];
struct macroblockd_plane *const pd = &xd->plane[i];
-
- // TODO(dkovalev) the same code in get_plane_block_size
- const int bwl = plane_block_width_log2by4(bsize, pd);
- const int bhl = plane_block_height_log2by4(bsize, pd);
- const BLOCK_SIZE_TYPE bs = get_block_size(bwl, bhl);
+ const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
unsigned int sse;
int rate;
int64_t dist;
(void) cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride,
pd->dst.buf, pd->dst.stride, &sse);
// sse works better than var, since there is no dc prediction used
- model_rd_from_var_lapndz(sse, 16 << (bwl + bhl),
+ model_rd_from_var_lapndz(sse, 1 << num_pels_log2_lookup[bs],
pd->dequant[1] >> 3, &rate, &dist);
rate_sum += rate;
@@ -421,81 +393,52 @@
*out_dist_sum = dist_sum << 4;
}
-static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE_TYPE bsize,
- MACROBLOCK *x, MACROBLOCKD *xd,
- int *out_rate_sum, int64_t *out_dist_sum) {
- // Note our transform coeffs are 8 times an orthogonal transform.
- // Hence quantizer step is also 8 times. To get effective quantizer
- // we need to divide by 8 before sending to modeling function.
- struct macroblock_plane *const p = &x->plane[0];
- struct macroblockd_plane *const pd = &xd->plane[0];
-
- // TODO(dkovalev) the same code in get_plane_block_size
- const int bwl = plane_block_width_log2by4(bsize, pd);
- const int bhl = plane_block_height_log2by4(bsize, pd);
- const BLOCK_SIZE_TYPE bs = get_block_size(bwl, bhl);
- unsigned int sse;
- int rate;
- int64_t dist;
- (void) cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride,
- pd->dst.buf, pd->dst.stride, &sse);
- // sse works better than var, since there is no dc prediction used
- model_rd_from_var_lapndz(sse, 16 << (bwl + bhl),
- pd->dequant[1] >> 3, &rate, &dist);
-
- *out_rate_sum = rate;
- *out_dist_sum = dist << 4;
-}
-
-static void model_rd_for_sb_y_tx(VP9_COMP *cpi, BLOCK_SIZE_TYPE bsize,
+static void model_rd_for_sb_y_tx(VP9_COMP *cpi, BLOCK_SIZE bsize,
TX_SIZE tx_size,
MACROBLOCK *x, MACROBLOCKD *xd,
int *out_rate_sum, int64_t *out_dist_sum,
int *out_skip) {
- int t = 4, j, k;
- BLOCK_SIZE_TYPE bs = BLOCK_SIZE_AB4X4;
+ int j, k;
+ BLOCK_SIZE bs;
struct macroblock_plane *const p = &x->plane[0];
struct macroblockd_plane *const pd = &xd->plane[0];
- const int width = plane_block_width(bsize, pd);
- const int height = plane_block_height(bsize, pd);
+ const int width = 4 << num_4x4_blocks_wide_lookup[bsize];
+ const int height = 4 << num_4x4_blocks_high_lookup[bsize];
int rate_sum = 0;
int64_t dist_sum = 0;
+ const int t = 4 << tx_size;
if (tx_size == TX_4X4) {
bs = BLOCK_4X4;
- t = 4;
} else if (tx_size == TX_8X8) {
bs = BLOCK_8X8;
- t = 8;
} else if (tx_size == TX_16X16) {
bs = BLOCK_16X16;
- t = 16;
} else if (tx_size == TX_32X32) {
bs = BLOCK_32X32;
- t = 32;
} else {
assert(0);
}
+
*out_skip = 1;
for (j = 0; j < height; j += t) {
for (k = 0; k < width; k += t) {
int rate;
int64_t dist;
unsigned int sse;
- (void) cpi->fn_ptr[bs].vf(p->src.buf + j * p->src.stride + k,
- p->src.stride,
- pd->dst.buf + j * pd->dst.stride + k,
- pd->dst.stride, &sse);
+ cpi->fn_ptr[bs].vf(&p->src.buf[j * p->src.stride + k], p->src.stride,
+ &pd->dst.buf[j * pd->dst.stride + k], pd->dst.stride,
+ &sse);
// sse works better than var, since there is no dc prediction used
- model_rd_from_var_lapndz(sse, t * t, pd->dequant[1] >> 3,
- &rate, &dist);
+ model_rd_from_var_lapndz(sse, t * t, pd->dequant[1] >> 3, &rate, &dist);
rate_sum += rate;
dist_sum += dist;
*out_skip &= (rate < 1024);
}
}
+
*out_rate_sum = rate_sum;
- *out_dist_sum = (dist_sum << 4);
+ *out_dist_sum = dist_sum << 4;
}
int64_t vp9_block_error_c(int16_t *coeff, int16_t *dqcoeff,
@@ -513,52 +456,55 @@
return error;
}
-static const int16_t band_counts[TX_SIZE_MAX_SB][8] = {
- { 1, 2, 3, 4, 3, 16 - 13 },
- { 1, 2, 3, 4, 11, 64 - 21 },
- { 1, 2, 3, 4, 11, 256 - 21 },
- { 1, 2, 3, 4, 11, 1024 - 21 },
+/* The trailing '0' is a terminator which is used inside cost_coeffs() to
+ * decide whether to include cost of a trailing EOB node or not (i.e. we
+ * can skip this if the last coefficient in this transform block, e.g. the
+ * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
+ * were non-zero). */
+static const int16_t band_counts[TX_SIZES][8] = {
+ { 1, 2, 3, 4, 3, 16 - 13, 0 },
+ { 1, 2, 3, 4, 11, 64 - 21, 0 },
+ { 1, 2, 3, 4, 11, 256 - 21, 0 },
+ { 1, 2, 3, 4, 11, 1024 - 21, 0 },
};
-static INLINE int cost_coeffs(VP9_COMMON *const cm, MACROBLOCK *mb,
- int plane, int block, PLANE_TYPE type,
+static INLINE int cost_coeffs(MACROBLOCK *mb,
+ int plane, int block,
ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
TX_SIZE tx_size,
const int16_t *scan, const int16_t *nb) {
MACROBLOCKD *const xd = &mb->e_mbd;
MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi;
- int pt, c, cost;
- const int16_t *band_count = band_counts[tx_size];
- const int eob = xd->plane[plane].eobs[block];
- const int16_t *qcoeff_ptr = BLOCK_OFFSET(xd->plane[plane].qcoeff, block, 16);
+ struct macroblockd_plane *pd = &xd->plane[plane];
+ const PLANE_TYPE type = pd->plane_type;
+ const int16_t *band_count = &band_counts[tx_size][1];
+ const int eob = pd->eobs[block];
+ const int16_t *const qcoeff_ptr = BLOCK_OFFSET(pd->qcoeff, block);
const int ref = mbmi->ref_frame[0] != INTRA_FRAME;
- unsigned int (*token_costs)[COEF_BANDS][PREV_COEF_CONTEXTS]
- [MAX_ENTROPY_TOKENS] = mb->token_costs[tx_size][type][ref];
- ENTROPY_CONTEXT above_ec = !!*A, left_ec = !!*L;
+ unsigned int (*token_costs)[2][PREV_COEF_CONTEXTS][MAX_ENTROPY_TOKENS] =
+ mb->token_costs[tx_size][type][ref];
+ const ENTROPY_CONTEXT above_ec = !!*A, left_ec = !!*L;
uint8_t token_cache[1024];
+ int pt = combine_entropy_contexts(above_ec, left_ec);
+ int c, cost;
// Check for consistency of tx_size with mode info
- assert((!type && !plane) || (type && plane));
- if (type == PLANE_TYPE_Y_WITH_DC) {
- assert(xd->mode_info_context->mbmi.txfm_size == tx_size);
- } else {
- assert(tx_size == get_uv_tx_size(mbmi));
- }
+ assert(type == PLANE_TYPE_Y_WITH_DC ? mbmi->txfm_size == tx_size
+ : get_uv_tx_size(mbmi) == tx_size);
- pt = combine_entropy_contexts(above_ec, left_ec);
-
if (eob == 0) {
// single eob token
cost = token_costs[0][0][pt][DCT_EOB_TOKEN];
c = 0;
} else {
- int v, prev_t, band = 1, band_left = band_count[1];
+ int band_left = *band_count++;
// dc token
- v = qcoeff_ptr[0];
- prev_t = vp9_dct_value_tokens_ptr[v].token;
- cost = token_costs[0][0][pt][prev_t] + vp9_dct_value_cost_ptr[v];
+ int v = qcoeff_ptr[0];
+ int prev_t = vp9_dct_value_tokens_ptr[v].token;
+ cost = (*token_costs)[0][pt][prev_t] + vp9_dct_value_cost_ptr[v];
token_cache[0] = vp9_pt_energy_class[prev_t];
+ ++token_costs;
// ac tokens
for (c = 1; c < eob; c++) {
@@ -568,29 +514,29 @@
v = qcoeff_ptr[rc];
t = vp9_dct_value_tokens_ptr[v].token;
pt = get_coef_context(nb, token_cache, c);
- cost += token_costs[!prev_t][band][pt][t] + vp9_dct_value_cost_ptr[v];
+ cost += (*token_costs)[!prev_t][pt][t] + vp9_dct_value_cost_ptr[v];
token_cache[rc] = vp9_pt_energy_class[t];
prev_t = t;
if (!--band_left) {
- band_left = band_count[++band];
+ band_left = *band_count++;
+ ++token_costs;
}
}
// eob token
- if (band < 6) {
+ if (band_left) {
pt = get_coef_context(nb, token_cache, c);
- cost += token_costs[0][band][pt][DCT_EOB_TOKEN];
+ cost += (*token_costs)[0][pt][DCT_EOB_TOKEN];
}
}
// is eob first coefficient;
- *A = *L = c > 0;
+ *A = *L = (c > 0);
return cost;
}
struct rdcost_block_args {
- VP9_COMMON *cm;
MACROBLOCK *x;
ENTROPY_CONTEXT t_above[16];
ENTROPY_CONTEXT t_left[16];
@@ -605,17 +551,17 @@
const int16_t *scan, *nb;
};
-static void dist_block(int plane, int block, BLOCK_SIZE_TYPE bsize,
- int ss_txfrm_size, void *arg) {
+static void dist_block(int plane, int block, TX_SIZE tx_size, void *arg) {
+ const int ss_txfrm_size = tx_size << 1;
struct rdcost_block_args* args = arg;
MACROBLOCK* const x = args->x;
MACROBLOCKD* const xd = &x->e_mbd;
- struct macroblock_plane *const p = &x->plane[0];
- struct macroblockd_plane *const pd = &xd->plane[0];
+ struct macroblock_plane *const p = &x->plane[plane];
+ struct macroblockd_plane *const pd = &xd->plane[plane];
int64_t this_sse;
int shift = args->tx_size == TX_32X32 ? 0 : 2;
- int16_t *const coeff = BLOCK_OFFSET(p->coeff, block, 16);
- int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block, 16);
+ int16_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+ int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
args->dist += vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
&this_sse) >> shift;
args->sse += this_sse >> shift;
@@ -630,119 +576,25 @@
}
}
-static void rate_block(int plane, int block, BLOCK_SIZE_TYPE bsize,
- int ss_txfrm_size, void *arg) {
+static void rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
+ TX_SIZE tx_size, void *arg) {
struct rdcost_block_args* args = arg;
+
int x_idx, y_idx;
- MACROBLOCKD * const xd = &args->x->e_mbd;
+ txfrm_block_to_raster_xy(plane_bsize, args->tx_size, block, &x_idx, &y_idx);
- txfrm_block_to_raster_xy(xd, bsize, plane, block, args->tx_size * 2, &x_idx,
- &y_idx);
-
- args->rate += cost_coeffs(args->cm, args->x, plane, block,
- xd->plane[plane].plane_type, args->t_above + x_idx,
+ args->rate += cost_coeffs(args->x, plane, block,
+ args->t_above + x_idx,
args->t_left + y_idx, args->tx_size,
args->scan, args->nb);
}
-// FIXME(jingning): need to make the rd test of chroma components consistent
-// with that of luma component. this function should be deprecated afterwards.
-static int rdcost_plane(VP9_COMMON * const cm, MACROBLOCK *x, int plane,
- BLOCK_SIZE_TYPE bsize, TX_SIZE tx_size) {
- MACROBLOCKD * const xd = &x->e_mbd;
- const int bwl = plane_block_width_log2by4(bsize, &xd->plane[plane]);
- const int bhl = plane_block_height_log2by4(bsize, &xd->plane[plane]);
- const int bw = 1 << bwl, bh = 1 << bhl;
- int i;
- struct rdcost_block_args args = { cm, x, { 0 }, { 0 }, tx_size, bw, bh,
- 0, 0, 0, INT64_MAX, 0 };
-
- switch (tx_size) {
- case TX_4X4:
- vpx_memcpy(&args.t_above, xd->plane[plane].above_context,
- sizeof(ENTROPY_CONTEXT) * bw);
- vpx_memcpy(&args.t_left, xd->plane[plane].left_context,
- sizeof(ENTROPY_CONTEXT) * bh);
- args.scan = vp9_default_scan_4x4;
- args.nb = vp9_default_scan_4x4_neighbors;
- break;
- case TX_8X8:
- for (i = 0; i < bw; i += 2)
- args.t_above[i] = !!*(uint16_t *)&xd->plane[plane].above_context[i];
- for (i = 0; i < bh; i += 2)
- args.t_left[i] = !!*(uint16_t *)&xd->plane[plane].left_context[i];
- args.scan = vp9_default_scan_8x8;
- args.nb = vp9_default_scan_8x8_neighbors;
- break;
- case TX_16X16:
- for (i = 0; i < bw; i += 4)
- args.t_above[i] = !!*(uint32_t *)&xd->plane[plane].above_context[i];
- for (i = 0; i < bh; i += 4)
- args.t_left[i] = !!*(uint32_t *)&xd->plane[plane].left_context[i];
- args.scan = vp9_default_scan_16x16;
- args.nb = vp9_default_scan_16x16_neighbors;
- break;
- case TX_32X32:
- for (i = 0; i < bw; i += 8)
- args.t_above[i] = !!*(uint64_t *)&xd->plane[plane].above_context[i];
- for (i = 0; i < bh; i += 8)
- args.t_left[i] = !!*(uint64_t *)&xd->plane[plane].left_context[i];
- args.scan = vp9_default_scan_32x32;
- args.nb = vp9_default_scan_32x32_neighbors;
- break;
- default:
- assert(0);
- }
-
- foreach_transformed_block_in_plane(xd, bsize, plane, rate_block, &args);
- return args.rate;
-}
-
-static int rdcost_uv(VP9_COMMON *const cm, MACROBLOCK *x,
- BLOCK_SIZE_TYPE bsize, TX_SIZE tx_size) {
- int cost = 0, plane;
-
- for (plane = 1; plane < MAX_MB_PLANE; plane++) {
- cost += rdcost_plane(cm, x, plane, bsize, tx_size);
- }
- return cost;
-}
-
-static int block_error_sby(MACROBLOCK *x, BLOCK_SIZE_TYPE bsize,
- int shift, int64_t *sse) {
- struct macroblockd_plane *p = &x->e_mbd.plane[0];
- const int bwl = plane_block_width_log2by4(bsize, p);
- const int bhl = plane_block_height_log2by4(bsize, p);
- int64_t e = vp9_block_error(x->plane[0].coeff, x->e_mbd.plane[0].dqcoeff,
- 16 << (bwl + bhl), sse) >> shift;
- *sse >>= shift;
- return e;
-}
-
-static int64_t block_error_sbuv(MACROBLOCK *x, BLOCK_SIZE_TYPE bsize,
- int shift, int64_t *sse) {
- int64_t sum = 0, this_sse;
- int plane;
-
- *sse = 0;
- for (plane = 1; plane < MAX_MB_PLANE; plane++) {
- struct macroblockd_plane *p = &x->e_mbd.plane[plane];
- const int bwl = plane_block_width_log2by4(bsize, p);
- const int bhl = plane_block_height_log2by4(bsize, p);
- sum += vp9_block_error(x->plane[plane].coeff, x->e_mbd.plane[plane].dqcoeff,
- 16 << (bwl + bhl), &this_sse);
- *sse += this_sse;
- }
- *sse >>= shift;
- return sum >> shift;
-}
-
-static void block_yrd_txfm(int plane, int block, BLOCK_SIZE_TYPE bsize,
- int ss_txfrm_size, void *arg) {
+static void block_yrd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
+ TX_SIZE tx_size, void *arg) {
struct rdcost_block_args *args = arg;
MACROBLOCK *const x = args->x;
MACROBLOCKD *const xd = &x->e_mbd;
- struct encode_b_args encode_args = {args->cm, x, NULL};
+ struct encode_b_args encode_args = {x, NULL};
int64_t rd1, rd2, rd;
if (args->skip)
@@ -758,58 +610,61 @@
return;
}
- if (xd->mode_info_context->mbmi.ref_frame[0] == INTRA_FRAME)
- encode_block_intra(plane, block, bsize, ss_txfrm_size, &encode_args);
+ if (!is_inter_block(&xd->mode_info_context->mbmi))
+ vp9_encode_block_intra(plane, block, plane_bsize, tx_size, &encode_args);
else
- xform_quant(plane, block, bsize, ss_txfrm_size, &encode_args);
+ vp9_xform_quant(plane, block, plane_bsize, tx_size, &encode_args);
- dist_block(plane, block, bsize, ss_txfrm_size, args);
- rate_block(plane, block, bsize, ss_txfrm_size, args);
+ dist_block(plane, block, tx_size, args);
+ rate_block(plane, block, plane_bsize, tx_size, args);
}
-static void super_block_yrd_for_txfm(VP9_COMMON *const cm, MACROBLOCK *x,
- int *rate, int64_t *distortion,
- int *skippable, int64_t *sse,
- int64_t ref_best_rd,
- BLOCK_SIZE_TYPE bsize, TX_SIZE tx_size) {
+static void txfm_rd_in_plane(MACROBLOCK *x,
+ int *rate, int64_t *distortion,
+ int *skippable, int64_t *sse,
+ int64_t ref_best_rd, int plane,
+ BLOCK_SIZE bsize, TX_SIZE tx_size) {
MACROBLOCKD *const xd = &x->e_mbd;
- struct macroblockd_plane *const pd = &xd->plane[0];
- const int bwl = plane_block_width_log2by4(bsize, pd);
- const int bhl = plane_block_height_log2by4(bsize, pd);
- const int bw = 1 << bwl, bh = 1 << bhl;
+ struct macroblockd_plane *const pd = &xd->plane[plane];
+ const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
+ const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bs];
+ const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bs];
int i;
- struct rdcost_block_args args = { cm, x, { 0 }, { 0 }, tx_size, bw, bh,
+ struct rdcost_block_args args = { x, { 0 }, { 0 }, tx_size,
+ num_4x4_blocks_wide, num_4x4_blocks_high,
0, 0, 0, ref_best_rd, 0 };
- xd->mode_info_context->mbmi.txfm_size = tx_size;
+ if (plane == 0)
+ xd->mode_info_context->mbmi.txfm_size = tx_size;
+
switch (tx_size) {
case TX_4X4:
vpx_memcpy(&args.t_above, pd->above_context,
- sizeof(ENTROPY_CONTEXT) * bw);
+ sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide);
vpx_memcpy(&args.t_left, pd->left_context,
- sizeof(ENTROPY_CONTEXT) * bh);
- get_scan_nb_4x4(get_tx_type_4x4(PLANE_TYPE_Y_WITH_DC, xd, 0),
+ sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high);
+ get_scan_nb_4x4(get_tx_type_4x4(pd->plane_type, xd, 0),
&args.scan, &args.nb);
break;
case TX_8X8:
- for (i = 0; i < bw; i += 2)
+ for (i = 0; i < num_4x4_blocks_wide; i += 2)
args.t_above[i] = !!*(uint16_t *)&pd->above_context[i];
- for (i = 0; i < bh; i += 2)
+ for (i = 0; i < num_4x4_blocks_high; i += 2)
args.t_left[i] = !!*(uint16_t *)&pd->left_context[i];
- get_scan_nb_8x8(get_tx_type_8x8(PLANE_TYPE_Y_WITH_DC, xd),
+ get_scan_nb_8x8(get_tx_type_8x8(pd->plane_type, xd),
&args.scan, &args.nb);
break;
case TX_16X16:
- for (i = 0; i < bw; i += 4)
+ for (i = 0; i < num_4x4_blocks_wide; i += 4)
args.t_above[i] = !!*(uint32_t *)&pd->above_context[i];
- for (i = 0; i < bh; i += 4)
+ for (i = 0; i < num_4x4_blocks_high; i += 4)
args.t_left[i] = !!*(uint32_t *)&pd->left_context[i];
- get_scan_nb_16x16(get_tx_type_16x16(PLANE_TYPE_Y_WITH_DC, xd),
+ get_scan_nb_16x16(get_tx_type_16x16(pd->plane_type, xd),
&args.scan, &args.nb);
break;
case TX_32X32:
- for (i = 0; i < bw; i += 8)
+ for (i = 0; i < num_4x4_blocks_wide; i += 8)
args.t_above[i] = !!*(uint64_t *)&pd->above_context[i];
- for (i = 0; i < bh; i += 8)
+ for (i = 0; i < num_4x4_blocks_high; i += 8)
args.t_left[i] = !!*(uint64_t *)&pd->left_context[i];
args.scan = vp9_default_scan_32x32;
args.nb = vp9_default_scan_32x32_neighbors;
@@ -818,20 +673,19 @@
assert(0);
}
- foreach_transformed_block_in_plane(xd, bsize, 0, block_yrd_txfm, &args);
+ foreach_transformed_block_in_plane(xd, bsize, plane, block_yrd_txfm, &args);
*distortion = args.dist;
*rate = args.rate;
*sse = args.sse;
- *skippable = vp9_sby_is_skippable(xd, bsize) && (!args.skip);
+ *skippable = vp9_is_skippable_in_plane(xd, bsize, plane) && (!args.skip);
}
static void choose_largest_txfm_size(VP9_COMP *cpi, MACROBLOCK *x,
int *rate, int64_t *distortion,
int *skip, int64_t *sse,
int64_t ref_best_rd,
- BLOCK_SIZE_TYPE bs) {
- const TX_SIZE max_txfm_size = TX_32X32
- - (bs < BLOCK_SIZE_SB32X32) - (bs < BLOCK_SIZE_MB16X16);
+ BLOCK_SIZE bs) {
+ const TX_SIZE max_txfm_size = max_txsize_lookup[bs];
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi;
@@ -849,9 +703,9 @@
} else {
mbmi->txfm_size = TX_4X4;
}
- super_block_yrd_for_txfm(cm, x, rate, distortion, skip,
- &sse[mbmi->txfm_size], ref_best_rd, bs,
- mbmi->txfm_size);
+ txfm_rd_in_plane(x, rate, distortion, skip,
+ &sse[mbmi->txfm_size], ref_best_rd, 0, bs,
+ mbmi->txfm_size);
cpi->txfm_stepdown_count[0]++;
}
@@ -859,25 +713,24 @@
int (*r)[2], int *rate,
int64_t *d, int64_t *distortion,
int *s, int *skip,
- int64_t txfm_cache[NB_TXFM_MODES],
- BLOCK_SIZE_TYPE bs) {
- const TX_SIZE max_txfm_size = TX_32X32
- - (bs < BLOCK_SIZE_SB32X32) - (bs < BLOCK_SIZE_MB16X16);
+ int64_t tx_cache[TX_MODES],
+ BLOCK_SIZE bs) {
+ const TX_SIZE max_tx_size = max_txsize_lookup[bs];
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi;
vp9_prob skip_prob = vp9_get_pred_prob_mbskip(cm, xd);
- int64_t rd[TX_SIZE_MAX_SB][2];
+ int64_t rd[TX_SIZES][2];
int n, m;
int s0, s1;
const vp9_prob *tx_probs = get_tx_probs2(xd, &cm->fc.tx_probs);
- for (n = TX_4X4; n <= max_txfm_size; n++) {
+ for (n = TX_4X4; n <= max_tx_size; n++) {
r[n][1] = r[n][0];
if (r[n][0] == INT_MAX)
continue;
- for (m = 0; m <= n - (n == max_txfm_size); m++) {
+ for (m = 0; m <= n - (n == max_tx_size); m++) {
if (m == n)
r[n][1] += vp9_cost_zero(tx_probs[m]);
else
@@ -889,7 +742,7 @@
s0 = vp9_cost_bit(skip_prob, 0);
s1 = vp9_cost_bit(skip_prob, 1);
- for (n = TX_4X4; n <= max_txfm_size; n++) {
+ for (n = TX_4X4; n <= max_tx_size; n++) {
if (d[n] == INT64_MAX) {
rd[n][0] = rd[n][1] = INT64_MAX;
continue;
@@ -902,13 +755,13 @@
}
}
- if (max_txfm_size == TX_32X32 &&
+ if (max_tx_size == TX_32X32 &&
(cm->tx_mode == ALLOW_32X32 ||
(cm->tx_mode == TX_MODE_SELECT &&
rd[TX_32X32][1] < rd[TX_16X16][1] && rd[TX_32X32][1] < rd[TX_8X8][1] &&
rd[TX_32X32][1] < rd[TX_4X4][1]))) {
mbmi->txfm_size = TX_32X32;
- } else if (max_txfm_size >= TX_16X16 &&
+ } else if (max_tx_size >= TX_16X16 &&
(cm->tx_mode == ALLOW_16X16 ||
cm->tx_mode == ALLOW_32X32 ||
(cm->tx_mode == TX_MODE_SELECT &&
@@ -928,34 +781,34 @@
*rate = r[mbmi->txfm_size][cm->tx_mode == TX_MODE_SELECT];
*skip = s[mbmi->txfm_size];
- txfm_cache[ONLY_4X4] = rd[TX_4X4][0];
- txfm_cache[ALLOW_8X8] = rd[TX_8X8][0];
- txfm_cache[ALLOW_16X16] = rd[MIN(max_txfm_size, TX_16X16)][0];
- txfm_cache[ALLOW_32X32] = rd[MIN(max_txfm_size, TX_32X32)][0];
- if (max_txfm_size == TX_32X32 &&
+ tx_cache[ONLY_4X4] = rd[TX_4X4][0];
+ tx_cache[ALLOW_8X8] = rd[TX_8X8][0];
+ tx_cache[ALLOW_16X16] = rd[MIN(max_tx_size, TX_16X16)][0];
+ tx_cache[ALLOW_32X32] = rd[MIN(max_tx_size, TX_32X32)][0];
+ if (max_tx_size == TX_32X32 &&
rd[TX_32X32][1] < rd[TX_16X16][1] && rd[TX_32X32][1] < rd[TX_8X8][1] &&
rd[TX_32X32][1] < rd[TX_4X4][1])
- txfm_cache[TX_MODE_SELECT] = rd[TX_32X32][1];
- else if (max_txfm_size >= TX_16X16 &&
+ tx_cache[TX_MODE_SELECT] = rd[TX_32X32][1];
+ else if (max_tx_size >= TX_16X16 &&
rd[TX_16X16][1] < rd[TX_8X8][1] && rd[TX_16X16][1] < rd[TX_4X4][1])
- txfm_cache[TX_MODE_SELECT] = rd[TX_16X16][1];
+ tx_cache[TX_MODE_SELECT] = rd[TX_16X16][1];
else
- txfm_cache[TX_MODE_SELECT] = rd[TX_4X4][1] < rd[TX_8X8][1] ?
+ tx_cache[TX_MODE_SELECT] = rd[TX_4X4][1] < rd[TX_8X8][1] ?
rd[TX_4X4][1] : rd[TX_8X8][1];
- if (max_txfm_size == TX_32X32 &&
+ if (max_tx_size == TX_32X32 &&
rd[TX_32X32][1] < rd[TX_16X16][1] &&
rd[TX_32X32][1] < rd[TX_8X8][1] &&
rd[TX_32X32][1] < rd[TX_4X4][1]) {
cpi->txfm_stepdown_count[0]++;
- } else if (max_txfm_size >= TX_16X16 &&
+ } else if (max_tx_size >= TX_16X16 &&
rd[TX_16X16][1] < rd[TX_8X8][1] &&
rd[TX_16X16][1] < rd[TX_4X4][1]) {
- cpi->txfm_stepdown_count[max_txfm_size - TX_16X16]++;
+ cpi->txfm_stepdown_count[max_tx_size - TX_16X16]++;
} else if (rd[TX_8X8][1] < rd[TX_4X4][1]) {
- cpi->txfm_stepdown_count[max_txfm_size - TX_8X8]++;
+ cpi->txfm_stepdown_count[max_tx_size - TX_8X8]++;
} else {
- cpi->txfm_stepdown_count[max_txfm_size - TX_4X4]++;
+ cpi->txfm_stepdown_count[max_tx_size - TX_4X4]++;
}
}
@@ -964,19 +817,17 @@
int64_t *d, int64_t *distortion,
int *s, int *skip, int64_t *sse,
int64_t ref_best_rd,
- BLOCK_SIZE_TYPE bs,
- int *model_used) {
- const TX_SIZE max_txfm_size = TX_32X32
- - (bs < BLOCK_SIZE_SB32X32) - (bs < BLOCK_SIZE_MB16X16);
+ BLOCK_SIZE bs) {
+ const TX_SIZE max_txfm_size = max_txsize_lookup[bs];
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi;
vp9_prob skip_prob = vp9_get_pred_prob_mbskip(cm, xd);
- int64_t rd[TX_SIZE_MAX_SB][2];
+ int64_t rd[TX_SIZES][2];
int n, m;
int s0, s1;
- double scale_rd[TX_SIZE_MAX_SB] = {1.73, 1.44, 1.20, 1.00};
- // double scale_r[TX_SIZE_MAX_SB] = {2.82, 2.00, 1.41, 1.00};
+ double scale_rd[TX_SIZES] = {1.73, 1.44, 1.20, 1.00};
+ // double scale_r[TX_SIZES] = {2.82, 2.00, 1.41, 1.00};
const vp9_prob *tx_probs = get_tx_probs2(xd, &cm->fc.tx_probs);
@@ -1034,17 +885,10 @@
mbmi->txfm_size = TX_4X4;
}
- if (model_used[mbmi->txfm_size]) {
- // Actually encode using the chosen mode if a model was used, but do not
- // update the r, d costs
- super_block_yrd_for_txfm(cm, x, rate, distortion, skip,
- &sse[mbmi->txfm_size], ref_best_rd,
- bs, mbmi->txfm_size);
- } else {
- *distortion = d[mbmi->txfm_size];
- *rate = r[mbmi->txfm_size][cm->tx_mode == TX_MODE_SELECT];
- *skip = s[mbmi->txfm_size];
- }
+ // Actually encode using the chosen mode if a model was used, but do not
+ // update the r, d costs
+ txfm_rd_in_plane(x, rate, distortion, skip, &sse[mbmi->txfm_size],
+ ref_best_rd, 0, bs, mbmi->txfm_size);
if (max_txfm_size == TX_32X32 &&
rd[TX_32X32][1] <= rd[TX_16X16][1] &&
@@ -1064,12 +908,11 @@
static void super_block_yrd(VP9_COMP *cpi,
MACROBLOCK *x, int *rate, int64_t *distortion,
- int *skip, int64_t *psse, BLOCK_SIZE_TYPE bs,
- int64_t txfm_cache[NB_TXFM_MODES],
+ int *skip, int64_t *psse, BLOCK_SIZE bs,
+ int64_t txfm_cache[TX_MODES],
int64_t ref_best_rd) {
- VP9_COMMON *const cm = &cpi->common;
- int r[TX_SIZE_MAX_SB][2], s[TX_SIZE_MAX_SB];
- int64_t d[TX_SIZE_MAX_SB], sse[TX_SIZE_MAX_SB];
+ int r[TX_SIZES][2], s[TX_SIZES];
+ int64_t d[TX_SIZES], sse[TX_SIZES];
MACROBLOCKD *xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi;
@@ -1080,7 +923,7 @@
if (cpi->sf.tx_size_search_method == USE_LARGESTALL ||
(cpi->sf.tx_size_search_method != USE_FULL_RD &&
mbmi->ref_frame[0] == INTRA_FRAME)) {
- vpx_memset(txfm_cache, 0, NB_TXFM_MODES * sizeof(int64_t));
+ vpx_memset(txfm_cache, 0, TX_MODES * sizeof(int64_t));
choose_largest_txfm_size(cpi, x, rate, distortion, skip, sse,
ref_best_rd, bs);
if (psse)
@@ -1090,56 +933,32 @@
if (cpi->sf.tx_size_search_method == USE_LARGESTINTRA_MODELINTER &&
mbmi->ref_frame[0] > INTRA_FRAME) {
- int model_used[TX_SIZE_MAX_SB] = {1, 1, 1, 1};
- if (bs >= BLOCK_SIZE_SB32X32) {
- if (model_used[TX_32X32]) {
- model_rd_for_sb_y_tx(cpi, bs, TX_32X32, x, xd,
- &r[TX_32X32][0], &d[TX_32X32], &s[TX_32X32]);
- } else {
- super_block_yrd_for_txfm(cm, x, &r[TX_32X32][0], &d[TX_32X32],
- &s[TX_32X32], &sse[TX_32X32], INT64_MAX,
- bs, TX_32X32);
- }
- }
- if (bs >= BLOCK_SIZE_MB16X16) {
- if (model_used[TX_16X16]) {
- model_rd_for_sb_y_tx(cpi, bs, TX_16X16, x, xd,
- &r[TX_16X16][0], &d[TX_16X16], &s[TX_16X16]);
- } else {
- super_block_yrd_for_txfm(cm, x, &r[TX_16X16][0], &d[TX_16X16],
- &s[TX_16X16], &sse[TX_16X16], INT64_MAX,
- bs, TX_16X16);
- }
- }
- if (model_used[TX_8X8]) {
- model_rd_for_sb_y_tx(cpi, bs, TX_8X8, x, xd,
- &r[TX_8X8][0], &d[TX_8X8], &s[TX_8X8]);
- } else {
- super_block_yrd_for_txfm(cm, x, &r[TX_8X8][0], &d[TX_8X8], &s[TX_8X8],
- &sse[TX_8X8], INT64_MAX, bs, TX_8X8);
- }
- if (model_used[TX_4X4]) {
- model_rd_for_sb_y_tx(cpi, bs, TX_4X4, x, xd,
- &r[TX_4X4][0], &d[TX_4X4], &s[TX_4X4]);
- } else {
- super_block_yrd_for_txfm(cm, x, &r[TX_4X4][0], &d[TX_4X4], &s[TX_4X4],
- &sse[TX_4X4], INT64_MAX, bs, TX_4X4);
- }
+ if (bs >= BLOCK_32X32)
+ model_rd_for_sb_y_tx(cpi, bs, TX_32X32, x, xd,
+ &r[TX_32X32][0], &d[TX_32X32], &s[TX_32X32]);
+ if (bs >= BLOCK_16X16)
+ model_rd_for_sb_y_tx(cpi, bs, TX_16X16, x, xd,
+ &r[TX_16X16][0], &d[TX_16X16], &s[TX_16X16]);
+
+ model_rd_for_sb_y_tx(cpi, bs, TX_8X8, x, xd,
+ &r[TX_8X8][0], &d[TX_8X8], &s[TX_8X8]);
+
+ model_rd_for_sb_y_tx(cpi, bs, TX_4X4, x, xd,
+ &r[TX_4X4][0], &d[TX_4X4], &s[TX_4X4]);
+
choose_txfm_size_from_modelrd(cpi, x, r, rate, d, distortion, s,
- skip, sse, ref_best_rd, bs, model_used);
+ skip, sse, ref_best_rd, bs);
} else {
- if (bs >= BLOCK_SIZE_SB32X32)
- super_block_yrd_for_txfm(cm, x, &r[TX_32X32][0], &d[TX_32X32],
- &s[TX_32X32], &sse[TX_32X32], ref_best_rd,
- bs, TX_32X32);
- if (bs >= BLOCK_SIZE_MB16X16)
- super_block_yrd_for_txfm(cm, x, &r[TX_16X16][0], &d[TX_16X16],
- &s[TX_16X16], &sse[TX_16X16], ref_best_rd,
- bs, TX_16X16);
- super_block_yrd_for_txfm(cm, x, &r[TX_8X8][0], &d[TX_8X8], &s[TX_8X8],
- &sse[TX_8X8], ref_best_rd, bs, TX_8X8);
- super_block_yrd_for_txfm(cm, x, &r[TX_4X4][0], &d[TX_4X4], &s[TX_4X4],
- &sse[TX_4X4], ref_best_rd, bs, TX_4X4);
+ if (bs >= BLOCK_32X32)
+ txfm_rd_in_plane(x, &r[TX_32X32][0], &d[TX_32X32], &s[TX_32X32],
+ &sse[TX_32X32], ref_best_rd, 0, bs, TX_32X32);
+ if (bs >= BLOCK_16X16)
+ txfm_rd_in_plane(x, &r[TX_16X16][0], &d[TX_16X16], &s[TX_16X16],
+ &sse[TX_16X16], ref_best_rd, 0, bs, TX_16X16);
+ txfm_rd_in_plane(x, &r[TX_8X8][0], &d[TX_8X8], &s[TX_8X8],
+ &sse[TX_8X8], ref_best_rd, 0, bs, TX_8X8);
+ txfm_rd_in_plane(x, &r[TX_4X4][0], &d[TX_4X4], &s[TX_4X4],
+ &sse[TX_4X4], ref_best_rd, 0, bs, TX_4X4);
choose_txfm_size_from_rd(cpi, x, r, rate, d, distortion, s,
skip, txfm_cache, bs);
}
@@ -1157,7 +976,7 @@
best_intra_mode != V_PRED &&
best_intra_mode != D45_PRED)
return 1;
- if (mode == D27_PRED &&
+ if (mode == D207_PRED &&
best_intra_mode != H_PRED &&
best_intra_mode != D45_PRED)
return 1;
@@ -1174,28 +993,29 @@
ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
int *bestrate, int *bestratey,
int64_t *bestdistortion,
- BLOCK_SIZE_TYPE bsize) {
+ BLOCK_SIZE bsize, int64_t rd_thresh) {
MB_PREDICTION_MODE mode;
MACROBLOCKD *xd = &x->e_mbd;
- int64_t best_rd = INT64_MAX;
+ int64_t best_rd = rd_thresh;
int rate = 0;
int64_t distortion;
- VP9_COMMON *const cm = &cpi->common;
struct macroblock_plane *p = &x->plane[0];
struct macroblockd_plane *pd = &xd->plane[0];
const int src_stride = p->src.stride;
const int dst_stride = pd->dst.stride;
- uint8_t *src, *dst;
+ uint8_t *src_init = raster_block_offset_uint8(BLOCK_8X8, ib,
+ p->src.buf, src_stride);
+ uint8_t *dst_init = raster_block_offset_uint8(BLOCK_8X8, ib,
+ pd->dst.buf, dst_stride);
int16_t *src_diff, *coeff;
ENTROPY_CONTEXT ta[2], tempa[2];
ENTROPY_CONTEXT tl[2], templ[2];
TX_TYPE tx_type = DCT_DCT;
- TX_TYPE best_tx_type = DCT_DCT;
- int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
- int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
+ const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
+ const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
int idx, idy, block;
- DECLARE_ALIGNED(16, int16_t, best_dqcoeff[4][16]);
+ uint8_t best_dst[8 * 8];
assert(ib < 4);
@@ -1206,6 +1026,10 @@
for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
int64_t this_rd;
int ratey = 0;
+
+ if (!(cpi->sf.intra_y_mode_mask & (1 << mode)))
+ continue;
+
// Only do the oblique modes if the best so far is
// one of the neighboring directional modes
if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
@@ -1223,17 +1047,14 @@
for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
int64_t ssz;
const int16_t *scan;
+ uint8_t *src = src_init + idx * 4 + idy * 4 * src_stride;
+ uint8_t *dst = dst_init + idx * 4 + idy * 4 * dst_stride;
block = ib + idy * 2 + idx;
xd->mode_info_context->bmi[block].as_mode = mode;
- src = raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, block,
- p->src.buf, src_stride);
- src_diff = raster_block_offset_int16(xd, BLOCK_SIZE_SB8X8, 0, block,
- p->src_diff);
- coeff = BLOCK_OFFSET(x->plane[0].coeff, block, 16);
- dst = raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, block,
- pd->dst.buf, dst_stride);
- vp9_predict_intra_block(xd, block, b_width_log2(BLOCK_SIZE_SB8X8),
+ src_diff = raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
+ coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
+ vp9_predict_intra_block(xd, block, 1,
TX_4X4, mode,
x->skip_encode ? src : dst,
x->skip_encode ? src_stride : dst_stride,
@@ -1252,18 +1073,19 @@
}
scan = get_scan_4x4(get_tx_type_4x4(PLANE_TYPE_Y_WITH_DC, xd, block));
- ratey += cost_coeffs(cm, x, 0, block, PLANE_TYPE_Y_WITH_DC,
+ ratey += cost_coeffs(x, 0, block,
tempa + idx, templ + idy, TX_4X4, scan,
vp9_get_coef_neighbors_handle(scan));
- distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff,
- block, 16),
+ distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
16, &ssz) >> 2;
+ if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
+ goto next;
if (tx_type != DCT_DCT)
- vp9_short_iht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block, 16),
+ vp9_short_iht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block),
dst, pd->dst.stride, tx_type);
else
- xd->inv_txm4x4_add(BLOCK_OFFSET(pd->dqcoeff, block, 16),
+ xd->inv_txm4x4_add(BLOCK_OFFSET(pd->dqcoeff, block),
dst, pd->dst.stride);
}
}
@@ -1277,61 +1099,40 @@
*bestdistortion = distortion;
best_rd = this_rd;
*best_mode = mode;
- best_tx_type = tx_type;
vpx_memcpy(a, tempa, sizeof(tempa));
vpx_memcpy(l, templ, sizeof(templ));
- for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
- for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
- block = ib + idy * 2 + idx;
- vpx_memcpy(best_dqcoeff[idy * 2 + idx],
- BLOCK_OFFSET(pd->dqcoeff, block, 16),
- sizeof(best_dqcoeff[0]));
- }
- }
+ for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
+ vpx_memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
+ num_4x4_blocks_wide * 4);
}
+ next:
+ {}
}
- if (x->skip_encode)
+ if (best_rd >= rd_thresh || x->skip_encode)
return best_rd;
- for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
- for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
- block = ib + idy * 2 + idx;
- xd->mode_info_context->bmi[block].as_mode = *best_mode;
- src = raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, block,
- p->src.buf, src_stride);
- dst = raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, block,
- pd->dst.buf, dst_stride);
+ for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
+ vpx_memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
+ num_4x4_blocks_wide * 4);
- vp9_predict_intra_block(xd, block, b_width_log2(BLOCK_SIZE_SB8X8), TX_4X4,
- *best_mode,
- x->skip_encode ? src : dst,
- x->skip_encode ? src_stride : dst_stride,
- dst, dst_stride);
- // inverse transform
- if (best_tx_type != DCT_DCT)
- vp9_short_iht4x4_add(best_dqcoeff[idy * 2 + idx], dst,
- dst_stride, best_tx_type);
- else
- xd->inv_txm4x4_add(best_dqcoeff[idy * 2 + idx], dst,
- dst_stride);
- }
- }
-
return best_rd;
}
-static int64_t rd_pick_intra4x4mby_modes(VP9_COMP *cpi, MACROBLOCK *mb,
- int *Rate, int *rate_y,
- int64_t *Distortion, int64_t best_rd) {
+static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP * const cpi,
+ MACROBLOCK * const mb,
+ int * const rate,
+ int * const rate_y,
+ int64_t * const distortion,
+ int64_t best_rd) {
int i, j;
MACROBLOCKD *const xd = &mb->e_mbd;
- BLOCK_SIZE_TYPE bsize = xd->mode_info_context->mbmi.sb_type;
- int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
- int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
+ const BLOCK_SIZE bsize = xd->mode_info_context->mbmi.sb_type;
+ const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
+ const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
int idx, idy;
int cost = 0;
- int64_t distortion = 0;
+ int64_t total_distortion = 0;
int tot_rate_y = 0;
int64_t total_rd = 0;
ENTROPY_CONTEXT t_above[4], t_left[4];
@@ -1343,14 +1144,14 @@
bmode_costs = mb->mbmode_cost;
+ // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
const int mis = xd->mode_info_stride;
- MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode);
- int UNINITIALIZED_IS_SAFE(r), UNINITIALIZED_IS_SAFE(ry);
- int64_t UNINITIALIZED_IS_SAFE(d);
+ MB_PREDICTION_MODE best_mode = DC_PRED;
+ int r = INT_MAX, ry = INT_MAX;
+ int64_t d = INT64_MAX, this_rd = INT64_MAX;
i = idy * 2 + idx;
-
if (cpi->common.frame_type == KEY_FRAME) {
const MB_PREDICTION_MODE A = above_block_mode(mic, i, mis);
const MB_PREDICTION_MODE L = (xd->left_available || idx) ?
@@ -1359,11 +1160,15 @@
bmode_costs = mb->y_mode_costs[A][L];
}
- total_rd += rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs,
- t_above + idx, t_left + idy,
- &r, &ry, &d, bsize);
+ this_rd = rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs,
+ t_above + idx, t_left + idy, &r, &ry, &d,
+ bsize, best_rd - total_rd);
+ if (this_rd >= best_rd - total_rd)
+ return INT64_MAX;
+
+ total_rd += this_rd;
cost += r;
- distortion += d;
+ total_distortion += d;
tot_rate_y += ry;
mic->bmi[i].as_mode = best_mode;
@@ -1377,40 +1182,42 @@
}
}
- *Rate = cost;
+ *rate = cost;
*rate_y = tot_rate_y;
- *Distortion = distortion;
+ *distortion = total_distortion;
xd->mode_info_context->mbmi.mode = mic->bmi[3].as_mode;
- return RDCOST(mb->rdmult, mb->rddiv, cost, distortion);
+ return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
}
static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
int *rate, int *rate_tokenonly,
int64_t *distortion, int *skippable,
- BLOCK_SIZE_TYPE bsize,
- int64_t txfm_cache[NB_TXFM_MODES],
+ BLOCK_SIZE bsize,
+ int64_t tx_cache[TX_MODES],
int64_t best_rd) {
MB_PREDICTION_MODE mode;
- MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(mode_selected);
+ MB_PREDICTION_MODE mode_selected = DC_PRED;
MACROBLOCKD *const xd = &x->e_mbd;
+ MODE_INFO *const mic = xd->mode_info_context;
int this_rate, this_rate_tokenonly, s;
int64_t this_distortion, this_rd;
- TX_SIZE UNINITIALIZED_IS_SAFE(best_tx);
+ TX_SIZE best_tx = TX_4X4;
int i;
int *bmode_costs = x->mbmode_cost;
- if (cpi->sf.tx_size_search_method == USE_FULL_RD) {
- for (i = 0; i < NB_TXFM_MODES; i++)
- txfm_cache[i] = INT64_MAX;
- }
+ if (cpi->sf.tx_size_search_method == USE_FULL_RD)
+ for (i = 0; i < TX_MODES; i++)
+ tx_cache[i] = INT64_MAX;
- /* Y Search for 32x32 intra prediction mode */
+ /* Y Search for intra prediction mode */
for (mode = DC_PRED; mode <= TM_PRED; mode++) {
- int64_t local_txfm_cache[NB_TXFM_MODES];
- MODE_INFO *const mic = xd->mode_info_context;
+ int64_t local_tx_cache[TX_MODES];
const int mis = xd->mode_info_stride;
+ if (!(cpi->sf.intra_y_mode_mask & (1 << mode)))
+ continue;
+
if (cpi->common.frame_type == KEY_FRAME) {
const MB_PREDICTION_MODE A = above_block_mode(mic, 0, mis);
const MB_PREDICTION_MODE L = xd->left_available ?
@@ -1418,10 +1225,10 @@
bmode_costs = x->y_mode_costs[A][L];
}
- x->e_mbd.mode_info_context->mbmi.mode = mode;
+ mic->mbmi.mode = mode;
super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s, NULL,
- bsize, local_txfm_cache, best_rd);
+ bsize, local_tx_cache, best_rd);
if (this_rate_tokenonly == INT_MAX)
continue;
@@ -1432,7 +1239,7 @@
if (this_rd < best_rd) {
mode_selected = mode;
best_rd = this_rd;
- best_tx = x->e_mbd.mode_info_context->mbmi.txfm_size;
+ best_tx = mic->mbmi.txfm_size;
*rate = this_rate;
*rate_tokenonly = this_rate_tokenonly;
*distortion = this_distortion;
@@ -1440,71 +1247,87 @@
}
if (cpi->sf.tx_size_search_method == USE_FULL_RD && this_rd < INT64_MAX) {
- for (i = 0; i < NB_TXFM_MODES; i++) {
- int64_t adj_rd = this_rd + local_txfm_cache[i] -
- local_txfm_cache[cpi->common.tx_mode];
- if (adj_rd < txfm_cache[i]) {
- txfm_cache[i] = adj_rd;
+ for (i = 0; i < TX_MODES; i++) {
+ const int64_t adj_rd = this_rd + local_tx_cache[i] -
+ local_tx_cache[cpi->common.tx_mode];
+ if (adj_rd < tx_cache[i]) {
+ tx_cache[i] = adj_rd;
}
}
}
}
- x->e_mbd.mode_info_context->mbmi.mode = mode_selected;
- x->e_mbd.mode_info_context->mbmi.txfm_size = best_tx;
+ mic->mbmi.mode = mode_selected;
+ mic->mbmi.txfm_size = best_tx;
return best_rd;
}
-static void super_block_uvrd_for_txfm(VP9_COMMON *const cm, MACROBLOCK *x,
- int *rate, int64_t *distortion,
- int *skippable, int64_t *sse,
- BLOCK_SIZE_TYPE bsize,
- TX_SIZE uv_tx_size) {
- MACROBLOCKD *const xd = &x->e_mbd;
- int64_t dummy;
- if (xd->mode_info_context->mbmi.ref_frame[0] == INTRA_FRAME)
- vp9_encode_intra_block_uv(cm, x, bsize);
- else
- vp9_xform_quant_sbuv(cm, x, bsize);
-
- *distortion = block_error_sbuv(x, bsize, uv_tx_size == TX_32X32 ? 0 : 2,
- sse ? sse : &dummy);
- *rate = rdcost_uv(cm, x, bsize, uv_tx_size);
- *skippable = vp9_sbuv_is_skippable(xd, bsize);
-}
-
static void super_block_uvrd(VP9_COMMON *const cm, MACROBLOCK *x,
int *rate, int64_t *distortion, int *skippable,
- int64_t *sse, BLOCK_SIZE_TYPE bsize) {
+ int64_t *sse, BLOCK_SIZE bsize,
+ int64_t ref_best_rd) {
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi;
TX_SIZE uv_txfm_size = get_uv_tx_size(mbmi);
+ int plane;
+ int pnrate = 0, pnskip = 1;
+ int64_t pndist = 0, pnsse = 0;
- if (mbmi->ref_frame[0] > INTRA_FRAME)
+ if (ref_best_rd < 0)
+ goto term;
+
+ if (is_inter_block(mbmi))
vp9_subtract_sbuv(x, bsize);
- super_block_uvrd_for_txfm(cm, x, rate, distortion, skippable, sse, bsize,
- uv_txfm_size);
+ *rate = 0;
+ *distortion = 0;
+ *sse = 0;
+ *skippable = 1;
+
+ for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
+ txfm_rd_in_plane(x, &pnrate, &pndist, &pnskip, &pnsse,
+ ref_best_rd, plane, bsize, uv_txfm_size);
+ if (pnrate == INT_MAX)
+ goto term;
+ *rate += pnrate;
+ *distortion += pndist;
+ *sse += pnsse;
+ *skippable &= pnskip;
+ }
+ return;
+
+ term:
+ *rate = INT_MAX;
+ *distortion = INT64_MAX;
+ *sse = INT64_MAX;
+ *skippable = 0;
+ return;
}
static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
int *rate, int *rate_tokenonly,
int64_t *distortion, int *skippable,
- BLOCK_SIZE_TYPE bsize) {
+ BLOCK_SIZE bsize) {
MB_PREDICTION_MODE mode;
- MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(mode_selected);
+ MB_PREDICTION_MODE mode_selected = DC_PRED;
int64_t best_rd = INT64_MAX, this_rd;
int this_rate_tokenonly, this_rate, s;
- int64_t this_distortion;
+ int64_t this_distortion, this_sse;
- MB_PREDICTION_MODE last_mode = bsize <= BLOCK_SIZE_SB8X8 ?
- TM_PRED : cpi->sf.last_chroma_intra_mode;
+ // int mode_mask = (bsize <= BLOCK_8X8)
+ // ? ALL_INTRA_MODES : cpi->sf.intra_uv_mode_mask;
- for (mode = DC_PRED; mode <= last_mode; mode++) {
+ for (mode = DC_PRED; mode <= TM_PRED; mode++) {
+ // if (!(mode_mask & (1 << mode)))
+ if (!(cpi->sf.intra_uv_mode_mask & (1 << mode)))
+ continue;
+
x->e_mbd.mode_info_context->mbmi.uv_mode = mode;
super_block_uvrd(&cpi->common, x, &this_rate_tokenonly,
- &this_distortion, &s, NULL, bsize);
+ &this_distortion, &s, &this_sse, bsize, best_rd);
+ if (this_rate_tokenonly == INT_MAX)
+ continue;
this_rate = this_rate_tokenonly +
x->intra_uv_mode_cost[cpi->common.frame_type][mode];
this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
@@ -1527,22 +1350,21 @@
static int64_t rd_sbuv_dcpred(VP9_COMP *cpi, MACROBLOCK *x,
int *rate, int *rate_tokenonly,
int64_t *distortion, int *skippable,
- BLOCK_SIZE_TYPE bsize) {
+ BLOCK_SIZE bsize) {
int64_t this_rd;
+ int64_t this_sse;
x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED;
super_block_uvrd(&cpi->common, x, rate_tokenonly,
- distortion, skippable, NULL, bsize);
+ distortion, skippable, &this_sse, bsize, INT64_MAX);
*rate = *rate_tokenonly +
x->intra_uv_mode_cost[cpi->common.frame_type][DC_PRED];
this_rd = RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
- x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED;
-
return this_rd;
}
-static void choose_intra_uv_mode(VP9_COMP *cpi, BLOCK_SIZE_TYPE bsize,
+static void choose_intra_uv_mode(VP9_COMP *cpi, BLOCK_SIZE bsize,
int *rate_uv, int *rate_uv_tokenonly,
int64_t *dist_uv, int *skip_uv,
MB_PREDICTION_MODE *mode_uv) {
@@ -1552,15 +1374,13 @@
// appropriate speed flag is set.
if (cpi->sf.use_uv_intra_rd_estimate) {
rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
- (bsize < BLOCK_SIZE_SB8X8) ? BLOCK_SIZE_SB8X8 :
- bsize);
+ bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
// Else do a proper rd search for each possible transform size that may
// be considered in the main rd loop.
} else {
rd_pick_intra_sbuv_mode(cpi, x,
rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
- (bsize < BLOCK_SIZE_SB8X8) ? BLOCK_SIZE_SB8X8
- : bsize);
+ bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
}
*mode_uv = x->e_mbd.mode_info_context->mbmi.uv_mode;
}
@@ -1572,7 +1392,7 @@
const int segment_id = xd->mode_info_context->mbmi.segment_id;
// Don't account for mode here if segment skip is enabled.
- if (!vp9_segfeature_active(&xd->seg, segment_id, SEG_LVL_SKIP)) {
+ if (!vp9_segfeature_active(&cpi->common.seg, segment_id, SEG_LVL_SKIP)) {
assert(is_inter_mode(mode));
return x->inter_mode_cost[mode_context][mode - NEARESTMV];
} else {
@@ -1586,13 +1406,13 @@
}
static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
- BLOCK_SIZE_TYPE bsize,
+ BLOCK_SIZE bsize,
int_mv *frame_mv,
int mi_row, int mi_col,
int_mv single_newmv[MAX_REF_FRAMES],
int *rate_mv);
static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
- BLOCK_SIZE_TYPE bsize,
+ BLOCK_SIZE bsize,
int mi_row, int mi_col,
int_mv *tmp_mv, int *rate_mv);
@@ -1606,11 +1426,11 @@
int *mvjcost, int *mvcost[2], VP9_COMP *cpi) {
MACROBLOCKD *const xd = &x->e_mbd;
MODE_INFO *const mic = xd->mode_info_context;
- MB_MODE_INFO * mbmi = &mic->mbmi;
+ MB_MODE_INFO *mbmi = &mic->mbmi;
int cost = 0, thismvcost = 0;
int idx, idy;
- int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
- int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
+ const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
+ const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
/* We have to be careful retrieving previously-encoded motion vectors.
Ones from this macroblock have to be pulled from the BLOCKD array
@@ -1623,12 +1443,11 @@
case NEWMV:
this_mv->as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
thismvcost = vp9_mv_bit_cost(this_mv, best_ref_mv, mvjcost, mvcost,
- 102, xd->allow_high_precision_mv);
+ 102);
if (mbmi->ref_frame[1] > 0) {
this_second_mv->as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
thismvcost += vp9_mv_bit_cost(this_second_mv, second_best_ref_mv,
- mvjcost, mvcost, 102,
- xd->allow_high_precision_mv);
+ mvjcost, mvcost, 102);
}
break;
case NEARESTMV:
@@ -1653,7 +1472,7 @@
}
cost = cost_mv_ref(cpi, this_mode,
- mbmi->mb_mode_context[mbmi->ref_frame[0]]);
+ mbmi->mode_context[mbmi->ref_frame[0]]);
mic->bmi[i].as_mv[0].as_int = this_mv->as_int;
if (mbmi->ref_frame[1] > 0)
@@ -1678,53 +1497,46 @@
ENTROPY_CONTEXT *ta,
ENTROPY_CONTEXT *tl) {
int k;
- VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *xd = &x->e_mbd;
- BLOCK_SIZE_TYPE bsize = xd->mode_info_context->mbmi.sb_type;
- const int width = plane_block_width(bsize, &xd->plane[0]);
- const int height = plane_block_height(bsize, &xd->plane[0]);
+ struct macroblockd_plane *const pd = &xd->plane[0];
+ MODE_INFO *const mi = xd->mode_info_context;
+ const BLOCK_SIZE bsize = mi->mbmi.sb_type;
+ const int width = plane_block_width(bsize, pd);
+ const int height = plane_block_height(bsize, pd);
int idx, idy;
const int src_stride = x->plane[0].src.stride;
- uint8_t* const src = raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i,
+ uint8_t* const src = raster_block_offset_uint8(BLOCK_8X8, i,
x->plane[0].src.buf,
src_stride);
- int16_t* src_diff = raster_block_offset_int16(xd, BLOCK_SIZE_SB8X8, 0, i,
+ int16_t* src_diff = raster_block_offset_int16(BLOCK_8X8, i,
x->plane[0].src_diff);
- int16_t* coeff = BLOCK_OFFSET(x->plane[0].coeff, 16, i);
- uint8_t* const pre = raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i,
- xd->plane[0].pre[0].buf,
- xd->plane[0].pre[0].stride);
- uint8_t* const dst = raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i,
- xd->plane[0].dst.buf,
- xd->plane[0].dst.stride);
+ int16_t* coeff = BLOCK_OFFSET(x->plane[0].coeff, i);
+ uint8_t* const pre = raster_block_offset_uint8(BLOCK_8X8, i,
+ pd->pre[0].buf,
+ pd->pre[0].stride);
+ uint8_t* const dst = raster_block_offset_uint8(BLOCK_8X8, i,
+ pd->dst.buf, pd->dst.stride);
int64_t thisdistortion = 0, thissse = 0;
int thisrate = 0;
- vp9_build_inter_predictor(pre,
- xd->plane[0].pre[0].stride,
- dst,
- xd->plane[0].dst.stride,
- &xd->mode_info_context->bmi[i].as_mv[0],
+ vp9_build_inter_predictor(pre, pd->pre[0].stride,
+ dst, pd->dst.stride,
+ &mi->bmi[i].as_mv[0].as_mv,
&xd->scale_factor[0],
- width, height, 0, &xd->subpix,
- MV_PRECISION_Q3);
+ width, height, 0, &xd->subpix, MV_PRECISION_Q3);
- if (xd->mode_info_context->mbmi.ref_frame[1] > 0) {
+ if (mi->mbmi.ref_frame[1] > 0) {
uint8_t* const second_pre =
- raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i,
- xd->plane[0].pre[1].buf,
- xd->plane[0].pre[1].stride);
- vp9_build_inter_predictor(second_pre, xd->plane[0].pre[1].stride,
- dst, xd->plane[0].dst.stride,
- &xd->mode_info_context->bmi[i].as_mv[1],
+ raster_block_offset_uint8(BLOCK_8X8, 0, pd->pre[1].buf, pd->pre[1].stride);
+ vp9_build_inter_predictor(second_pre, pd->pre[1].stride,
+ dst, pd->dst.stride,
+ &mi->bmi[i].as_mv[1].as_mv,
&xd->scale_factor[1],
- width, height, 1,
- &xd->subpix, MV_PRECISION_Q3);
+ width, height, 1, &xd->subpix, MV_PRECISION_Q3);
}
- vp9_subtract_block(height, width, src_diff, 8,
- src, src_stride,
- dst, xd->plane[0].dst.stride);
+ vp9_subtract_block(height, width, src_diff, 8, src, src_stride,
+ dst, pd->dst.stride);
k = i;
for (idy = 0; idy < height / 4; ++idy) {
@@ -1732,16 +1544,15 @@
int64_t ssz, rd, rd1, rd2;
k += (idy * 2 + idx);
- src_diff = raster_block_offset_int16(xd, BLOCK_SIZE_SB8X8, 0, k,
+ src_diff = raster_block_offset_int16(BLOCK_8X8, k,
x->plane[0].src_diff);
- coeff = BLOCK_OFFSET(x->plane[0].coeff, 16, k);
+ coeff = BLOCK_OFFSET(x->plane[0].coeff, k);
x->fwd_txm4x4(src_diff, coeff, 16);
x->quantize_b_4x4(x, k, DCT_DCT, 16);
- thisdistortion += vp9_block_error(coeff,
- BLOCK_OFFSET(xd->plane[0].dqcoeff,
- k, 16), 16, &ssz);
+ thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
+ 16, &ssz);
thissse += ssz;
- thisrate += cost_coeffs(cm, x, 0, k, PLANE_TYPE_Y_WITH_DC,
+ thisrate += cost_coeffs(x, 0, k,
ta + (k & 1),
tl + (k >> 1), TX_4X4,
vp9_default_scan_4x4,
@@ -1782,7 +1593,7 @@
int64_t sse;
int segment_yrate;
MB_PREDICTION_MODE modes[4];
- SEG_RDSTAT rdstat[4][VP9_INTER_MODES];
+ SEG_RDSTAT rdstat[4][INTER_MODES];
int mvthresh;
} BEST_SEG_INFO;
@@ -1796,21 +1607,18 @@
}
static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
- MB_MODE_INFO *mbmi = &x->e_mbd.mode_info_context->mbmi;
- x->plane[0].src.buf =
- raster_block_offset_uint8(&x->e_mbd, BLOCK_SIZE_SB8X8, 0, i,
- x->plane[0].src.buf,
- x->plane[0].src.stride);
- assert(((intptr_t)x->e_mbd.plane[0].pre[0].buf & 0x7) == 0);
- x->e_mbd.plane[0].pre[0].buf =
- raster_block_offset_uint8(&x->e_mbd, BLOCK_SIZE_SB8X8, 0, i,
- x->e_mbd.plane[0].pre[0].buf,
- x->e_mbd.plane[0].pre[0].stride);
+ MB_MODE_INFO *const mbmi = &x->e_mbd.mode_info_context->mbmi;
+ struct macroblock_plane *const p = &x->plane[0];
+ struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
+
+ p->src.buf = raster_block_offset_uint8(BLOCK_8X8, i, p->src.buf,
+ p->src.stride);
+ assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
+ pd->pre[0].buf = raster_block_offset_uint8(BLOCK_8X8, i, pd->pre[0].buf,
+ pd->pre[0].stride);
if (mbmi->ref_frame[1])
- x->e_mbd.plane[0].pre[1].buf =
- raster_block_offset_uint8(&x->e_mbd, BLOCK_SIZE_SB8X8, 0, i,
- x->e_mbd.plane[0].pre[1].buf,
- x->e_mbd.plane[0].pre[1].stride);
+ pd->pre[1].buf = raster_block_offset_uint8(BLOCK_8X8, i, pd->pre[1].buf,
+ pd->pre[1].stride);
}
static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
@@ -1835,9 +1643,9 @@
int64_t this_segment_rd = 0;
int label_mv_thresh;
int segmentyrate = 0;
- BLOCK_SIZE_TYPE bsize = mbmi->sb_type;
- int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
- int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
+ const BLOCK_SIZE bsize = mbmi->sb_type;
+ const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
+ const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
vp9_variance_fn_ptr_t *v_fn_ptr;
ENTROPY_CONTEXT t_above[2], t_left[2];
BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
@@ -1871,12 +1679,12 @@
vp9_append_sub8x8_mvs_for_idx(&cpi->common, &x->e_mbd,
&frame_mv[NEARESTMV][mbmi->ref_frame[0]],
&frame_mv[NEARMV][mbmi->ref_frame[0]],
- i, 0);
+ i, 0, mi_row, mi_col);
if (mbmi->ref_frame[1] > 0)
vp9_append_sub8x8_mvs_for_idx(&cpi->common, &x->e_mbd,
&frame_mv[NEARESTMV][mbmi->ref_frame[1]],
&frame_mv[NEARMV][mbmi->ref_frame[1]],
- i, 1);
+ i, 1, mi_row, mi_col);
// search for the best motion vector on this segment
for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
@@ -1892,7 +1700,7 @@
frame_mv[this_mode][mbmi->ref_frame[0]].as_int == 0 &&
(mbmi->ref_frame[1] <= 0 ||
frame_mv[this_mode][mbmi->ref_frame[1]].as_int == 0)) {
- int rfc = mbmi->mb_mode_context[mbmi->ref_frame[0]];
+ int rfc = mbmi->mode_context[mbmi->ref_frame[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);
@@ -1937,6 +1745,7 @@
int thissme, bestsme = INT_MAX;
int sadpb = x->sadperbit4;
int_mv mvp_full;
+ int max_mv;
/* Is the best so far sufficiently good that we cant justify doing
* and new motion search. */
@@ -1953,19 +1762,16 @@
x->e_mbd.mode_info_context->bmi[i - 2].as_mv[0].as_int;
}
}
+ if (i == 0)
+ max_mv = x->max_mv_context[mbmi->ref_frame[0]];
+ else
+ max_mv = MAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
if (cpi->sf.auto_mv_step_size && cpi->common.show_frame) {
// Take wtd average of the step_params based on the last frame's
// max mv magnitude and the best ref mvs of the current block for
// the given reference.
- if (i == 0)
- step_param = (vp9_init_search_range(
- cpi, x->max_mv_context[mbmi->ref_frame[0]]) +
- cpi->mv_step_param) >> 1;
- else
- step_param = (vp9_init_search_range(
- cpi, MAX(abs(bsi->mvp.as_mv.row),
- abs(bsi->mvp.as_mv.col)) >> 3) +
- cpi->mv_step_param) >> 1;
+ step_param = (vp9_init_search_range(cpi, max_mv) +
+ cpi->mv_step_param) >> 1;
} else {
step_param = cpi->mv_step_param;
}
@@ -1977,14 +1783,31 @@
// adjust src pointer for this block
mi_buf_shift(x, i);
- bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param,
- sadpb, further_steps, 0, v_fn_ptr,
- bsi->ref_mv, &mode_mv[NEWMV]);
+ if (cpi->sf.search_method == HEX) {
+ bestsme = vp9_hex_search(x, &mvp_full,
+ step_param,
+ sadpb, 1, v_fn_ptr, 1,
+ bsi->ref_mv, &mode_mv[NEWMV]);
+ } else if (cpi->sf.search_method == SQUARE) {
+ bestsme = vp9_square_search(x, &mvp_full,
+ step_param,
+ sadpb, 1, v_fn_ptr, 1,
+ bsi->ref_mv, &mode_mv[NEWMV]);
+ } else if (cpi->sf.search_method == BIGDIA) {
+ bestsme = vp9_bigdia_search(x, &mvp_full,
+ step_param,
+ sadpb, 1, v_fn_ptr, 1,
+ bsi->ref_mv, &mode_mv[NEWMV]);
+ } else {
+ bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param,
+ sadpb, further_steps, 0, v_fn_ptr,
+ bsi->ref_mv, &mode_mv[NEWMV]);
+ }
// Should we do a full search (best quality only)
if (cpi->compressor_speed == 0) {
/* Check if mvp_full is within the range. */
- clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max,
+ clamp_mv(&mvp_full.as_mv, x->mv_col_min, x->mv_col_max,
x->mv_row_min, x->mv_row_max);
thissme = cpi->full_search_sad(x, &mvp_full,
@@ -2009,6 +1832,7 @@
unsigned int sse;
cpi->find_fractional_mv_step(x, &mode_mv[NEWMV],
bsi->ref_mv, x->errorperbit, v_fn_ptr,
+ 0, cpi->sf.subpel_iters_per_step,
x->nmvjointcost, x->mvcost,
&distortion, &sse);
@@ -2021,7 +1845,7 @@
}
if (mbmi->ref_frame[1] > 0 && this_mode == NEWMV &&
- mbmi->interp_filter == vp9_switchable_interp[0]) {
+ mbmi->interp_filter == EIGHTTAP) {
if (seg_mvs[i][mbmi->ref_frame[1]].as_int == INVALID_MV ||
seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV)
continue;
@@ -2132,7 +1956,7 @@
if (best_rd == INT64_MAX) {
int iy, midx;
for (iy = i + 1; iy < 4; ++iy)
- for (midx = 0; midx < VP9_INTER_MODES; ++midx)
+ for (midx = 0; midx < INTER_MODES; ++midx)
bsi->rdstat[iy][midx].brdcost = INT64_MAX;
bsi->segment_rd = INT64_MAX;
return;
@@ -2156,7 +1980,7 @@
if (this_segment_rd > bsi->segment_rd) {
int iy, midx;
for (iy = i + 1; iy < 4; ++iy)
- for (midx = 0; midx < VP9_INTER_MODES; ++midx)
+ for (midx = 0; midx < INTER_MODES; ++midx)
bsi->rdstat[iy][midx].brdcost = INT64_MAX;
bsi->segment_rd = INT64_MAX;
return;
@@ -2204,7 +2028,7 @@
MB_MODE_INFO *mbmi = &mi->mbmi;
int mode_idx;
- vpx_memset(bsi, 0, sizeof(*bsi));
+ vp9_zero(*bsi);
bsi->segment_rd = best_rd;
bsi->ref_mv = best_ref_mv;
@@ -2235,7 +2059,7 @@
*returntotrate = bsi->r;
*returndistortion = bsi->d;
*returnyrate = bsi->segment_yrate;
- *skippable = vp9_sby_is_skippable(&x->e_mbd, BLOCK_SIZE_SB8X8);
+ *skippable = vp9_is_skippable_in_plane(&x->e_mbd, BLOCK_8X8, 0);
*psse = bsi->sse;
mbmi->mode = bsi->modes[3];
@@ -2244,7 +2068,7 @@
static void mv_pred(VP9_COMP *cpi, MACROBLOCK *x,
uint8_t *ref_y_buffer, int ref_y_stride,
- int ref_frame, BLOCK_SIZE_TYPE block_size ) {
+ int ref_frame, BLOCK_SIZE block_size ) {
MACROBLOCKD *xd = &x->e_mbd;
MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi;
int_mv this_mv;
@@ -2297,7 +2121,7 @@
vp9_prob *comp_mode_p) {
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &cpi->mb.e_mbd;
- int seg_ref_active = vp9_segfeature_active(&xd->seg, segment_id,
+ int seg_ref_active = vp9_segfeature_active(&cm->seg, segment_id,
SEG_LVL_REF_FRAME);
if (seg_ref_active) {
vpx_memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
@@ -2358,8 +2182,8 @@
int_mv *ref_mv,
int_mv *second_ref_mv,
int64_t comp_pred_diff[NB_PREDICTION_TYPES],
- int64_t txfm_size_diff[NB_TXFM_MODES],
- int64_t best_filter_diff[VP9_SWITCHABLE_FILTERS + 1]) {
+ int64_t tx_size_diff[TX_MODES],
+ int64_t best_filter_diff[SWITCHABLE_FILTERS + 1]) {
MACROBLOCKD *const xd = &x->e_mbd;
// Take a snapshot of the coding context so it can be
@@ -2380,9 +2204,9 @@
// FIXME(rbultje) does this memcpy the whole array? I believe sizeof()
// doesn't actually work this way
- memcpy(ctx->txfm_rd_diff, txfm_size_diff, sizeof(ctx->txfm_rd_diff));
+ memcpy(ctx->tx_rd_diff, tx_size_diff, sizeof(ctx->tx_rd_diff));
memcpy(ctx->best_filter_diff, best_filter_diff,
- sizeof(*best_filter_diff) * (VP9_SWITCHABLE_FILTERS + 1));
+ sizeof(*best_filter_diff) * (SWITCHABLE_FILTERS + 1));
}
static void setup_pred_block(const MACROBLOCKD *xd,
@@ -2413,7 +2237,7 @@
static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
int idx, MV_REFERENCE_FRAME frame_type,
- BLOCK_SIZE_TYPE block_size,
+ BLOCK_SIZE block_size,
int mi_row, int mi_col,
int_mv frame_nearest_mv[MAX_REF_FRAMES],
int_mv frame_near_mv[MAX_REF_FRAMES],
@@ -2429,10 +2253,10 @@
scale[frame_type].x_offset_q4 =
ROUND_POWER_OF_TWO(mi_col * MI_SIZE * scale[frame_type].x_scale_fp,
- VP9_REF_SCALE_SHIFT) & 0xf;
+ REF_SCALE_SHIFT) & 0xf;
scale[frame_type].y_offset_q4 =
ROUND_POWER_OF_TWO(mi_row * MI_SIZE * scale[frame_type].y_scale_fp,
- VP9_REF_SCALE_SHIFT) & 0xf;
+ REF_SCALE_SHIFT) & 0xf;
// TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
// use the UV scaling factors.
@@ -2443,8 +2267,7 @@
vp9_find_mv_refs(&cpi->common, xd, xd->mode_info_context,
xd->prev_mode_info_context,
frame_type,
- mbmi->ref_mvs[frame_type],
- cpi->common.ref_frame_sign_bias);
+ mbmi->ref_mvs[frame_type], mi_row, mi_col);
// Candidate refinement carried out at encoder and decoder
vp9_find_best_ref_mvs(xd,
@@ -2455,8 +2278,7 @@
// Further refinement that is encode side only to test the top few candidates
// in full and choose the best as the centre point for subsequent searches.
// The current implementation doesn't support scaling.
- if (scale[frame_type].x_scale_fp == VP9_REF_NO_SCALE &&
- scale[frame_type].y_scale_fp == VP9_REF_NO_SCALE)
+ if (!vp9_is_scaled(&scale[frame_type]))
mv_pred(cpi, x, yv12_mb[frame_type][0].buf, yv12->y_stride,
frame_type, block_size);
}
@@ -2469,17 +2291,16 @@
return scaled_ref_frame;
}
-static INLINE int get_switchable_rate(VP9_COMMON *cm, MACROBLOCK *x) {
- MACROBLOCKD *xd = &x->e_mbd;
- MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi;
-
- const int c = vp9_get_pred_context_switchable_interp(xd);
- const int m = vp9_switchable_interp_map[mbmi->interp_filter];
- return SWITCHABLE_INTERP_RATE_FACTOR * x->switchable_interp_costs[c][m];
+static INLINE int get_switchable_rate(const MACROBLOCK *x) {
+ const MACROBLOCKD *const xd = &x->e_mbd;
+ const MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi;
+ const int ctx = vp9_get_pred_context_switchable_interp(xd);
+ return SWITCHABLE_INTERP_RATE_FACTOR *
+ x->switchable_interp_costs[ctx][mbmi->interp_filter];
}
static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
- BLOCK_SIZE_TYPE bsize,
+ BLOCK_SIZE bsize,
int mi_row, int mi_col,
int_mv *tmp_mv, int *rate_mv) {
MACROBLOCKD *xd = &x->e_mbd;
@@ -2492,7 +2313,7 @@
int_mv mvp_full;
int ref = mbmi->ref_frame[0];
int_mv ref_mv = mbmi->ref_mvs[ref][0];
- const BLOCK_SIZE_TYPE block_size = get_plane_block_size(bsize, &xd->plane[0]);
+ const BLOCK_SIZE block_size = get_plane_block_size(bsize, &xd->plane[0]);
int tmp_col_min = x->mv_col_min;
int tmp_col_max = x->mv_col_max;
@@ -2512,7 +2333,7 @@
setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
}
- vp9_clamp_mv_min_max(x, &ref_mv);
+ vp9_clamp_mv_min_max(x, &ref_mv.as_mv);
// Adjust search parameters based on small partitions' result.
if (x->fast_ms) {
@@ -2554,10 +2375,30 @@
// Further step/diamond searches as necessary
further_steps = (cpi->sf.max_step_search_steps - 1) - step_param;
- bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param,
- sadpb, further_steps, 1,
- &cpi->fn_ptr[block_size],
- &ref_mv, tmp_mv);
+ if (cpi->sf.search_method == HEX) {
+ bestsme = vp9_hex_search(x, &mvp_full,
+ step_param,
+ sadpb, 1,
+ &cpi->fn_ptr[block_size], 1,
+ &ref_mv, tmp_mv);
+ } else if (cpi->sf.search_method == SQUARE) {
+ bestsme = vp9_square_search(x, &mvp_full,
+ step_param,
+ sadpb, 1,
+ &cpi->fn_ptr[block_size], 1,
+ &ref_mv, tmp_mv);
+ } else if (cpi->sf.search_method == BIGDIA) {
+ bestsme = vp9_bigdia_search(x, &mvp_full,
+ step_param,
+ sadpb, 1,
+ &cpi->fn_ptr[block_size], 1,
+ &ref_mv, tmp_mv);
+ } else {
+ bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param,
+ sadpb, further_steps, 1,
+ &cpi->fn_ptr[block_size],
+ &ref_mv, tmp_mv);
+ }
x->mv_col_min = tmp_col_min;
x->mv_col_max = tmp_col_max;
@@ -2565,17 +2406,18 @@
x->mv_row_max = tmp_row_max;
if (bestsme < INT_MAX) {
- int dis; /* TODO: use dis in distortion calculation later. */
+ int dis; /* TODO: use dis in distortion calculation later. */
unsigned int sse;
cpi->find_fractional_mv_step(x, tmp_mv, &ref_mv,
x->errorperbit,
&cpi->fn_ptr[block_size],
+ 0, cpi->sf.subpel_iters_per_step,
x->nmvjointcost, x->mvcost,
&dis, &sse);
}
*rate_mv = vp9_mv_bit_cost(tmp_mv, &ref_mv,
x->nmvjointcost, x->mvcost,
- 96, xd->allow_high_precision_mv);
+ 96);
if (scaled_ref_frame) {
int i;
for (i = 0; i < MAX_MB_PLANE; i++)
@@ -2584,7 +2426,7 @@
}
static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
- BLOCK_SIZE_TYPE bsize,
+ BLOCK_SIZE bsize,
int_mv *frame_mv,
int mi_row, int mi_col,
int_mv single_newmv[MAX_REF_FRAMES],
@@ -2595,7 +2437,7 @@
int refs[2] = { mbmi->ref_frame[0],
(mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) };
int_mv ref_mv[2];
- const BLOCK_SIZE_TYPE block_size = get_plane_block_size(bsize, &xd->plane[0]);
+ const BLOCK_SIZE block_size = get_plane_block_size(bsize, &xd->plane[0]);
int ite;
// Prediction buffer from second frame.
uint8_t *second_pred = vpx_memalign(16, pw * ph * sizeof(uint8_t));
@@ -2663,7 +2505,7 @@
vp9_build_inter_predictor(ref_yv12[!id].buf,
ref_yv12[!id].stride,
second_pred, pw,
- &frame_mv[refs[!id]],
+ &frame_mv[refs[!id]].as_mv,
&xd->scale_factor[!id],
pw, ph, 0,
&xd->subpix, MV_PRECISION_Q3);
@@ -2671,7 +2513,7 @@
// Compound motion search on first ref frame.
if (id)
xd->plane[0].pre[0] = ref_yv12[id];
- vp9_clamp_mv_min_max(x, &ref_mv[id]);
+ vp9_clamp_mv_min_max(x, &ref_mv[id].as_mv);
// Use mv result from single mode as mvp.
tmp_mv.as_int = frame_mv[refs[id]].as_int;
@@ -2696,13 +2538,15 @@
int dis; /* TODO: use dis in distortion calculation later. */
unsigned int sse;
- bestsme = vp9_find_best_sub_pixel_comp(x, &tmp_mv,
- &ref_mv[id],
- x->errorperbit,
- &cpi->fn_ptr[block_size],
- x->nmvjointcost, x->mvcost,
- &dis, &sse, second_pred,
- pw, ph);
+ bestsme = cpi->find_fractional_mv_step_comp(
+ x, &tmp_mv,
+ &ref_mv[id],
+ x->errorperbit,
+ &cpi->fn_ptr[block_size],
+ 0, cpi->sf.subpel_iters_per_step,
+ x->nmvjointcost, x->mvcost,
+ &dis, &sse, second_pred,
+ pw, ph);
}
if (id)
@@ -2730,18 +2574,16 @@
}
*rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]],
&mbmi->ref_mvs[refs[0]][0],
- x->nmvjointcost, x->mvcost, 96,
- x->e_mbd.allow_high_precision_mv);
+ x->nmvjointcost, x->mvcost, 96);
*rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]],
&mbmi->ref_mvs[refs[1]][0],
- x->nmvjointcost, x->mvcost, 96,
- x->e_mbd.allow_high_precision_mv);
+ x->nmvjointcost, x->mvcost, 96);
vpx_free(second_pred);
}
static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
- BLOCK_SIZE_TYPE bsize,
+ BLOCK_SIZE bsize,
int64_t txfm_cache[],
int *rate2, int64_t *distortion,
int *skippable,
@@ -2752,7 +2594,8 @@
int_mv (*mode_mv)[MAX_REF_FRAMES],
int mi_row, int mi_col,
int_mv single_newmv[MAX_REF_FRAMES],
- int64_t *psse, int64_t ref_best_rd) {
+ int64_t *psse,
+ const int64_t ref_best_rd) {
VP9_COMMON *cm = &cpi->common;
MACROBLOCKD *xd = &x->e_mbd;
MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi;
@@ -2767,7 +2610,6 @@
int64_t this_rd = 0;
DECLARE_ALIGNED_ARRAY(16, uint8_t, tmp_buf, MAX_MB_PLANE * 64 * 64);
int pred_exists = 0;
- int interpolating_intpel_seen = 0;
int intpel_mv;
int64_t rd, best_rd = INT64_MAX;
int best_needs_copy = 0;
@@ -2775,54 +2617,44 @@
int orig_dst_stride[MAX_MB_PLANE];
int rs = 0;
- switch (this_mode) {
+ if (this_mode == NEWMV) {
int rate_mv;
- case NEWMV:
- if (is_comp_pred) {
- // Initialize mv using single prediction mode result.
- frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
- frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
+ if (is_comp_pred) {
+ // Initialize mv using single prediction mode result.
+ frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
+ frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
- if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
- joint_motion_search(cpi, x, bsize, frame_mv,
- mi_row, mi_col, single_newmv, &rate_mv);
- } else {
- rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]],
- &mbmi->ref_mvs[refs[0]][0],
- x->nmvjointcost, x->mvcost, 96,
- x->e_mbd.allow_high_precision_mv);
- rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]],
- &mbmi->ref_mvs[refs[1]][0],
- x->nmvjointcost, x->mvcost, 96,
- x->e_mbd.allow_high_precision_mv);
- }
- if (frame_mv[refs[0]].as_int == INVALID_MV ||
- frame_mv[refs[1]].as_int == INVALID_MV)
- return INT64_MAX;
- *rate2 += rate_mv;
+ if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
+ joint_motion_search(cpi, x, bsize, frame_mv,
+ mi_row, mi_col, single_newmv, &rate_mv);
} else {
- int_mv tmp_mv;
- single_motion_search(cpi, x, bsize, mi_row, mi_col,
- &tmp_mv, &rate_mv);
- *rate2 += rate_mv;
- frame_mv[refs[0]].as_int =
- xd->mode_info_context->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
- single_newmv[refs[0]].as_int = tmp_mv.as_int;
+ rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]],
+ &mbmi->ref_mvs[refs[0]][0],
+ x->nmvjointcost, x->mvcost, 96);
+ rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]],
+ &mbmi->ref_mvs[refs[1]][0],
+ x->nmvjointcost, x->mvcost, 96);
}
- break;
- case NEARMV:
- case NEARESTMV:
- case ZEROMV:
- default:
- break;
+ if (frame_mv[refs[0]].as_int == INVALID_MV ||
+ frame_mv[refs[1]].as_int == INVALID_MV)
+ return INT64_MAX;
+ *rate2 += rate_mv;
+ } else {
+ int_mv tmp_mv;
+ single_motion_search(cpi, x, bsize, mi_row, mi_col, &tmp_mv, &rate_mv);
+ *rate2 += rate_mv;
+ frame_mv[refs[0]].as_int =
+ xd->mode_info_context->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
+ single_newmv[refs[0]].as_int = tmp_mv.as_int;
+ }
}
// if we're near/nearest and mv == 0,0, compare to zeromv
if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
frame_mv[refs[0]].as_int == 0 &&
- !vp9_segfeature_active(&xd->seg, mbmi->segment_id, SEG_LVL_SKIP) &&
+ !vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP) &&
(num_refs == 1 || frame_mv[refs[1]].as_int == 0)) {
- int rfc = mbmi->mb_mode_context[mbmi->ref_frame[0]];
+ int rfc = mbmi->mode_context[mbmi->ref_frame[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);
@@ -2856,10 +2688,8 @@
for (i = 0; i < num_refs; ++i) {
cur_mv[i] = frame_mv[refs[i]];
// Clip "next_nearest" so that it does not extend to far out of image
- if (this_mode == NEWMV)
- assert(!clamp_mv2(&cur_mv[i], xd));
- else
- clamp_mv2(&cur_mv[i], xd);
+ if (this_mode != NEWMV)
+ clamp_mv2(&cur_mv[i].as_mv, xd);
if (mv_check_bounds(x, &cur_mv[i]))
return INT64_MAX;
@@ -2881,7 +2711,7 @@
* words if you present them in that order, the second one is always known
* if the first is known */
*rate2 += cost_mv_ref(cpi, this_mode,
- mbmi->mb_mode_context[mbmi->ref_frame[0]]);
+ mbmi->mode_context[mbmi->ref_frame[0]]);
if (!(*mode_excluded)) {
if (is_comp_pred) {
@@ -2892,7 +2722,6 @@
}
pred_exists = 0;
- interpolating_intpel_seen = 0;
// Are all MVs integer pel for Y and UV
intpel_mv = (mbmi->mv[0].as_mv.row & 15) == 0 &&
(mbmi->mv[0].as_mv.col & 15) == 0;
@@ -2901,105 +2730,104 @@
(mbmi->mv[1].as_mv.col & 15) == 0;
// Search for best switchable filter by checking the variance of
// pred error irrespective of whether the filter will be used
- *best_filter = EIGHTTAP;
- if (cpi->sf.use_8tap_always) {
+ if (cm->mcomp_filter_type != BILINEAR) {
*best_filter = EIGHTTAP;
- vp9_zero(cpi->rd_filter_cache);
- } else {
- int i, newbest;
- int tmp_rate_sum = 0;
- int64_t tmp_dist_sum = 0;
+ if (x->source_variance <
+ cpi->sf.disable_filter_search_var_thresh) {
+ *best_filter = EIGHTTAP;
+ vp9_zero(cpi->rd_filter_cache);
+ } else {
+ int i, newbest;
+ int tmp_rate_sum = 0;
+ int64_t tmp_dist_sum = 0;
- cpi->rd_filter_cache[VP9_SWITCHABLE_FILTERS] = INT64_MAX;
- for (i = 0; i < VP9_SWITCHABLE_FILTERS; ++i) {
- int j;
- int64_t rs_rd;
- const INTERPOLATIONFILTERTYPE filter = vp9_switchable_interp[i];
- const int is_intpel_interp = intpel_mv;
- mbmi->interp_filter = filter;
- vp9_setup_interp_filters(xd, mbmi->interp_filter, cm);
- rs = get_switchable_rate(cm, x);
- rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
+ cpi->rd_filter_cache[SWITCHABLE_FILTERS] = INT64_MAX;
+ for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
+ int j;
+ int64_t rs_rd;
+ mbmi->interp_filter = i;
+ vp9_setup_interp_filters(xd, mbmi->interp_filter, cm);
+ rs = get_switchable_rate(x);
+ rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
- if (interpolating_intpel_seen && is_intpel_interp) {
- cpi->rd_filter_cache[i] = RDCOST(x->rdmult, x->rddiv,
- tmp_rate_sum, tmp_dist_sum);
- cpi->rd_filter_cache[VP9_SWITCHABLE_FILTERS] =
- MIN(cpi->rd_filter_cache[VP9_SWITCHABLE_FILTERS],
- cpi->rd_filter_cache[i] + rs_rd);
- rd = cpi->rd_filter_cache[i];
- if (cm->mcomp_filter_type == SWITCHABLE)
- rd += rs_rd;
- } else {
- int rate_sum = 0;
- int64_t dist_sum = 0;
- if ((cm->mcomp_filter_type == SWITCHABLE &&
- (!i || best_needs_copy)) ||
- (cm->mcomp_filter_type != SWITCHABLE &&
- (cm->mcomp_filter_type == mbmi->interp_filter ||
- (!interpolating_intpel_seen && is_intpel_interp)))) {
- for (j = 0; j < MAX_MB_PLANE; j++) {
- xd->plane[j].dst.buf = orig_dst[j];
- xd->plane[j].dst.stride = orig_dst_stride[j];
- }
+ if (i > 0 && intpel_mv) {
+ cpi->rd_filter_cache[i] = RDCOST(x->rdmult, x->rddiv,
+ tmp_rate_sum, tmp_dist_sum);
+ cpi->rd_filter_cache[SWITCHABLE_FILTERS] =
+ MIN(cpi->rd_filter_cache[SWITCHABLE_FILTERS],
+ cpi->rd_filter_cache[i] + rs_rd);
+ rd = cpi->rd_filter_cache[i];
+ if (cm->mcomp_filter_type == SWITCHABLE)
+ rd += rs_rd;
} else {
- for (j = 0; j < MAX_MB_PLANE; j++) {
- xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
- xd->plane[j].dst.stride = 64;
+ int rate_sum = 0;
+ int64_t dist_sum = 0;
+ if ((cm->mcomp_filter_type == SWITCHABLE &&
+ (!i || best_needs_copy)) ||
+ (cm->mcomp_filter_type != SWITCHABLE &&
+ (cm->mcomp_filter_type == mbmi->interp_filter ||
+ (i == 0 && intpel_mv)))) {
+ for (j = 0; j < MAX_MB_PLANE; j++) {
+ xd->plane[j].dst.buf = orig_dst[j];
+ xd->plane[j].dst.stride = orig_dst_stride[j];
+ }
+ } else {
+ for (j = 0; j < MAX_MB_PLANE; j++) {
+ xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
+ xd->plane[j].dst.stride = 64;
+ }
}
+ vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
+ model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum);
+ cpi->rd_filter_cache[i] = RDCOST(x->rdmult, x->rddiv,
+ rate_sum, dist_sum);
+ cpi->rd_filter_cache[SWITCHABLE_FILTERS] =
+ MIN(cpi->rd_filter_cache[SWITCHABLE_FILTERS],
+ cpi->rd_filter_cache[i] + rs_rd);
+ rd = cpi->rd_filter_cache[i];
+ if (cm->mcomp_filter_type == SWITCHABLE)
+ rd += rs_rd;
+ if (i == 0 && intpel_mv) {
+ tmp_rate_sum = rate_sum;
+ tmp_dist_sum = dist_sum;
+ }
}
- vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
- model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum);
- cpi->rd_filter_cache[i] = RDCOST(x->rdmult, x->rddiv,
- rate_sum, dist_sum);
- cpi->rd_filter_cache[VP9_SWITCHABLE_FILTERS] =
- MIN(cpi->rd_filter_cache[VP9_SWITCHABLE_FILTERS],
- cpi->rd_filter_cache[i] + rs_rd);
- rd = cpi->rd_filter_cache[i];
- if (cm->mcomp_filter_type == SWITCHABLE)
- rd += rs_rd;
- if (!interpolating_intpel_seen && is_intpel_interp) {
- tmp_rate_sum = rate_sum;
- tmp_dist_sum = dist_sum;
- }
- }
- if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
- if (rd / 2 > ref_best_rd) {
- for (i = 0; i < MAX_MB_PLANE; i++) {
- xd->plane[i].dst.buf = orig_dst[i];
- xd->plane[i].dst.stride = orig_dst_stride[i];
+ if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
+ if (rd / 2 > ref_best_rd) {
+ for (i = 0; i < MAX_MB_PLANE; i++) {
+ xd->plane[i].dst.buf = orig_dst[i];
+ xd->plane[i].dst.stride = orig_dst_stride[i];
+ }
+ return INT64_MAX;
}
- return INT64_MAX;
}
- }
- newbest = i == 0 || rd < best_rd;
+ newbest = i == 0 || rd < best_rd;
- if (newbest) {
- best_rd = rd;
- *best_filter = mbmi->interp_filter;
- if (cm->mcomp_filter_type == SWITCHABLE && i &&
- !(interpolating_intpel_seen && is_intpel_interp))
- best_needs_copy = !best_needs_copy;
+ if (newbest) {
+ best_rd = rd;
+ *best_filter = mbmi->interp_filter;
+ if (cm->mcomp_filter_type == SWITCHABLE && i && !intpel_mv)
+ best_needs_copy = !best_needs_copy;
+ }
+
+ if ((cm->mcomp_filter_type == SWITCHABLE && newbest) ||
+ (cm->mcomp_filter_type != SWITCHABLE &&
+ cm->mcomp_filter_type == mbmi->interp_filter)) {
+ pred_exists = 1;
+ }
}
- if ((cm->mcomp_filter_type == SWITCHABLE && newbest) ||
- (cm->mcomp_filter_type != SWITCHABLE &&
- cm->mcomp_filter_type == mbmi->interp_filter)) {
- pred_exists = 1;
+ for (i = 0; i < MAX_MB_PLANE; i++) {
+ xd->plane[i].dst.buf = orig_dst[i];
+ xd->plane[i].dst.stride = orig_dst_stride[i];
}
- interpolating_intpel_seen |= is_intpel_interp;
}
-
- for (i = 0; i < MAX_MB_PLANE; i++) {
- xd->plane[i].dst.buf = orig_dst[i];
- xd->plane[i].dst.stride = orig_dst_stride[i];
- }
}
- // Set the appripriate filter
+ // Set the appropriate filter
mbmi->interp_filter = cm->mcomp_filter_type != SWITCHABLE ?
cm->mcomp_filter_type : *best_filter;
vp9_setup_interp_filters(xd, mbmi->interp_filter, cm);
- rs = (cm->mcomp_filter_type == SWITCHABLE ? get_switchable_rate(cm, x) : 0);
+ rs = cm->mcomp_filter_type == SWITCHABLE ? get_switchable_rate(x) : 0;
if (pred_exists) {
if (best_needs_copy) {
@@ -3033,55 +2861,81 @@
}
if (cpi->common.mcomp_filter_type == SWITCHABLE)
- *rate2 += get_switchable_rate(cm, x);
+ *rate2 += get_switchable_rate(x);
- if (cpi->active_map_enabled && x->active_ptr[0] == 0)
- x->skip = 1;
- else if (x->encode_breakout) {
- const BLOCK_SIZE_TYPE y_size = get_plane_block_size(bsize, &xd->plane[0]);
- const BLOCK_SIZE_TYPE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
+ if (!is_comp_pred) {
+ if (cpi->active_map_enabled && x->active_ptr[0] == 0)
+ x->skip = 1;
+ else if (x->encode_breakout) {
+ const BLOCK_SIZE y_size = get_plane_block_size(bsize, &xd->plane[0]);
+ const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
+ unsigned int var, sse;
+ // Skipping threshold for ac.
+ unsigned int thresh_ac;
+ // The encode_breakout input
+ unsigned int encode_breakout = x->encode_breakout << 4;
- unsigned int var, sse;
- int threshold = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1] >> 4);
+ // Calculate threshold according to dequant value.
+ thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
+ // Set a maximum for threshold to avoid big PSNR loss in low bitrate case.
+ if (thresh_ac > 36000)
+ thresh_ac = 36000;
- if (threshold < x->encode_breakout)
- threshold = x->encode_breakout;
+ // Use encode_breakout input if it is bigger than internal threshold.
+ if (thresh_ac < encode_breakout)
+ thresh_ac = encode_breakout;
- var = cpi->fn_ptr[y_size].vf(x->plane[0].src.buf, x->plane[0].src.stride,
- xd->plane[0].dst.buf, xd->plane[0].dst.stride,
- &sse);
+ var = cpi->fn_ptr[y_size].vf(x->plane[0].src.buf, x->plane[0].src.stride,
+ xd->plane[0].dst.buf,
+ xd->plane[0].dst.stride, &sse);
- if ((int)sse < threshold) {
- unsigned int q2dc = xd->plane[0].dequant[0];
- // If there is no codeable 2nd order dc
- // or a very small uniform pixel change change
- if ((sse - var < q2dc * q2dc >> 4) ||
- (sse / 2 > var && sse - var < 64)) {
- // Check u and v to make sure skip is ok
- int sse2;
- unsigned int sse2u, sse2v;
- var = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
- x->plane[1].src.stride,
- xd->plane[1].dst.buf,
- xd->plane[1].dst.stride, &sse2u);
- var = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
- x->plane[2].src.stride,
- xd->plane[2].dst.buf,
- xd->plane[2].dst.stride, &sse2v);
- sse2 = sse2u + sse2v;
+ // Adjust threshold according to partition size.
+ thresh_ac >>= 8 - (b_width_log2_lookup[bsize] +
+ b_height_log2_lookup[bsize]);
- if (sse2 * 2 < threshold) {
- x->skip = 1;
- *distortion = sse + sse2;
- *rate2 = 500;
+ // Y skipping condition checking
+ if (sse < thresh_ac || sse == 0) {
+ // Skipping threshold for dc
+ unsigned int thresh_dc;
- // for best yrd calculation
- *rate_uv = 0;
- *distortion_uv = sse2;
+ thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
- *disable_skip = 1;
- this_rd = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
+ // dc skipping checking
+ if ((sse - var) < thresh_dc || sse == var) {
+ unsigned int sse_u, sse_v;
+ unsigned int var_u, var_v;
+
+ var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
+ x->plane[1].src.stride,
+ xd->plane[1].dst.buf,
+ xd->plane[1].dst.stride, &sse_u);
+
+ // U skipping condition checking
+ if ((sse_u * 4 < thresh_ac || sse_u == 0) &&
+ (sse_u - var_u < thresh_dc || sse_u == var_u)) {
+ var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
+ x->plane[2].src.stride,
+ xd->plane[2].dst.buf,
+ xd->plane[2].dst.stride, &sse_v);
+
+ // V skipping condition checking
+ if ((sse_v * 4 < thresh_ac || sse_v == 0) &&
+ (sse_v - var_v < thresh_dc || sse_v == var_v)) {
+ x->skip = 1;
+
+ *rate2 = 500;
+ *rate_uv = 0;
+
+ // Scaling factor for SSE from spatial domain to frequency domain
+ // is 16. Adjust distortion accordingly.
+ *distortion_uv = (sse_u + sse_v) << 4;
+ *distortion = (sse << 4) + *distortion_uv;
+
+ *disable_skip = 1;
+ this_rd = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
+ }
+ }
}
}
}
@@ -3089,7 +2943,8 @@
if (!x->skip) {
int skippable_y, skippable_uv;
- int64_t sseuv = INT_MAX;
+ int64_t sseuv = INT64_MAX;
+ int64_t rdcosty = INT64_MAX;
// Y cost and distortion
super_block_yrd(cpi, x, rate_y, distortion_y, &skippable_y, psse,
@@ -3108,9 +2963,21 @@
*rate2 += *rate_y;
*distortion += *distortion_y;
- super_block_uvrd(cm, x, rate_uv, distortion_uv,
- &skippable_uv, &sseuv, bsize);
+ rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
+ rdcosty = MIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
+ super_block_uvrd(cm, x, rate_uv, distortion_uv, &skippable_uv, &sseuv,
+ bsize, ref_best_rd - rdcosty);
+ if (*rate_uv == INT_MAX) {
+ *rate2 = INT_MAX;
+ *distortion = INT64_MAX;
+ for (i = 0; i < MAX_MB_PLANE; i++) {
+ xd->plane[i].dst.buf = orig_dst[i];
+ xd->plane[i].dst.stride = orig_dst_stride[i];
+ }
+ return INT64_MAX;
+ }
+
*psse += sseuv;
*rate2 += *rate_uv;
*distortion += *distortion_uv;
@@ -3127,21 +2994,19 @@
void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
int *returnrate, int64_t *returndist,
- BLOCK_SIZE_TYPE bsize,
+ BLOCK_SIZE bsize,
PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
int y_skip = 0, uv_skip;
- int64_t dist_y = 0, dist_uv = 0, txfm_cache[NB_TXFM_MODES];
-
+ int64_t dist_y = 0, dist_uv = 0, tx_cache[TX_MODES] = { 0 };
x->skip_encode = 0;
- vpx_memset(&txfm_cache, 0, sizeof(txfm_cache));
ctx->skip = 0;
xd->mode_info_context->mbmi.ref_frame[0] = INTRA_FRAME;
- if (bsize >= BLOCK_SIZE_SB8X8) {
+ if (bsize >= BLOCK_8X8) {
if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly,
- &dist_y, &y_skip, bsize, txfm_cache,
+ &dist_y, &y_skip, bsize, tx_cache,
best_rd) >= best_rd) {
*returnrate = INT_MAX;
return;
@@ -3150,30 +3015,28 @@
&dist_uv, &uv_skip, bsize);
} else {
y_skip = 0;
- if (rd_pick_intra4x4mby_modes(cpi, x, &rate_y, &rate_y_tokenonly,
- &dist_y, best_rd) >= best_rd) {
+ if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
+ &dist_y, best_rd) >= best_rd) {
*returnrate = INT_MAX;
return;
}
rd_pick_intra_sbuv_mode(cpi, x, &rate_uv, &rate_uv_tokenonly,
- &dist_uv, &uv_skip, BLOCK_SIZE_SB8X8);
+ &dist_uv, &uv_skip, BLOCK_8X8);
}
if (y_skip && uv_skip) {
*returnrate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
vp9_cost_bit(vp9_get_pred_prob_mbskip(cm, xd), 1);
- *returndist = dist_y + (dist_uv >> 2);
- memset(ctx->txfm_rd_diff, 0, sizeof(ctx->txfm_rd_diff));
+ *returndist = dist_y + dist_uv;
+ vp9_zero(ctx->tx_rd_diff);
} else {
int i;
*returnrate = rate_y + rate_uv +
vp9_cost_bit(vp9_get_pred_prob_mbskip(cm, xd), 0);
- *returndist = dist_y + (dist_uv >> 2);
- if (cpi->sf.tx_size_search_method == USE_FULL_RD) {
- for (i = 0; i < NB_TXFM_MODES; i++) {
- ctx->txfm_rd_diff[i] = txfm_cache[i] - txfm_cache[cm->tx_mode];
- }
- }
+ *returndist = dist_y + dist_uv;
+ if (cpi->sf.tx_size_search_method == USE_FULL_RD)
+ for (i = 0; i < TX_MODES; i++)
+ ctx->tx_rd_diff[i] = tx_cache[i] - tx_cache[cm->tx_mode];
}
ctx->mic = *xd->mode_info_context;
@@ -3183,20 +3046,21 @@
int mi_row, int mi_col,
int *returnrate,
int64_t *returndistortion,
- BLOCK_SIZE_TYPE bsize,
+ BLOCK_SIZE bsize,
PICK_MODE_CONTEXT *ctx,
int64_t best_rd_so_far) {
VP9_COMMON *cm = &cpi->common;
MACROBLOCKD *xd = &x->e_mbd;
MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi;
- const BLOCK_SIZE_TYPE block_size = get_plane_block_size(bsize, &xd->plane[0]);
+ const struct segmentation *seg = &cm->seg;
+ const BLOCK_SIZE block_size = get_plane_block_size(bsize, &xd->plane[0]);
MB_PREDICTION_MODE this_mode;
- MV_REFERENCE_FRAME ref_frame;
+ MV_REFERENCE_FRAME ref_frame, second_ref_frame;
unsigned char segment_id = xd->mode_info_context->mbmi.segment_id;
int comp_pred, i;
int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
struct buf_2d yv12_mb[4][MAX_MB_PLANE];
- int_mv single_newmv[MAX_REF_FRAMES];
+ int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
VP9_ALT_FLAG };
int idx_list[4] = {0,
@@ -3205,13 +3069,13 @@
cpi->alt_fb_idx};
int64_t best_rd = best_rd_so_far;
int64_t best_yrd = best_rd_so_far; // FIXME(rbultje) more precise
- int64_t best_txfm_rd[NB_TXFM_MODES];
- int64_t best_txfm_diff[NB_TXFM_MODES];
+ int64_t best_tx_rd[TX_MODES];
+ int64_t best_tx_diff[TX_MODES];
int64_t best_pred_diff[NB_PREDICTION_TYPES];
int64_t best_pred_rd[NB_PREDICTION_TYPES];
- int64_t best_filter_rd[VP9_SWITCHABLE_FILTERS + 1];
- int64_t best_filter_diff[VP9_SWITCHABLE_FILTERS + 1];
- MB_MODE_INFO best_mbmode;
+ int64_t best_filter_rd[SWITCHABLE_FILTERS + 1];
+ int64_t best_filter_diff[SWITCHABLE_FILTERS + 1];
+ MB_MODE_INFO best_mbmode = { 0 };
int j;
int mode_index, best_mode_index = 0;
unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
@@ -3222,10 +3086,10 @@
// MB_PREDICTION_MODE best_inter_mode = ZEROMV;
MV_REFERENCE_FRAME best_inter_ref_frame = LAST_FRAME;
INTERPOLATIONFILTERTYPE tmp_best_filter = SWITCHABLE;
- int rate_uv_intra[TX_SIZE_MAX_SB], rate_uv_tokenonly[TX_SIZE_MAX_SB];
- int64_t dist_uv[TX_SIZE_MAX_SB];
- int skip_uv[TX_SIZE_MAX_SB];
- MB_PREDICTION_MODE mode_uv[TX_SIZE_MAX_SB];
+ int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
+ int64_t dist_uv[TX_SIZES];
+ int skip_uv[TX_SIZES];
+ MB_PREDICTION_MODE mode_uv[TX_SIZES];
struct scale_factors scale_factor[4];
unsigned int ref_frame_mask = 0;
unsigned int mode_mask = 0;
@@ -3254,24 +3118,22 @@
ctx->frames_with_high_error = 0;
ctx->modes_with_high_error = 0;
- xd->mode_info_context->mbmi.segment_id = segment_id;
estimate_ref_frame_costs(cpi, segment_id, ref_costs_single, ref_costs_comp,
&comp_mode_p);
- vpx_memset(&best_mbmode, 0, sizeof(best_mbmode));
- vpx_memset(&single_newmv, 0, sizeof(single_newmv));
for (i = 0; i < NB_PREDICTION_TYPES; ++i)
best_pred_rd[i] = INT64_MAX;
- for (i = 0; i < NB_TXFM_MODES; i++)
- best_txfm_rd[i] = INT64_MAX;
- for (i = 0; i <= VP9_SWITCHABLE_FILTERS; i++)
+ for (i = 0; i < TX_MODES; i++)
+ best_tx_rd[i] = INT64_MAX;
+ for (i = 0; i <= SWITCHABLE_FILTERS; i++)
best_filter_rd[i] = INT64_MAX;
- for (i = 0; i < TX_SIZE_MAX_SB; i++)
+ for (i = 0; i < TX_SIZES; i++)
rate_uv_intra[i] = INT_MAX;
*returnrate = INT_MAX;
- // Create a mask set to 1 for each frame used by a smaller resolution.
+ // Create a mask set to 1 for each reference frame used by a smaller
+ // resolution.
if (cpi->sf.use_avoid_tested_higherror) {
switch (block_size) {
case BLOCK_64X64:
@@ -3320,101 +3182,97 @@
int compmode_cost = 0;
int rate2 = 0, rate_y = 0, rate_uv = 0;
int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
- int skippable;
- int64_t txfm_cache[NB_TXFM_MODES];
+ int skippable = 0;
+ int64_t tx_cache[TX_MODES];
int i;
int this_skip2 = 0;
int64_t total_sse = INT_MAX;
int early_term = 0;
- for (i = 0; i < NB_TXFM_MODES; ++i)
- txfm_cache[i] = INT64_MAX;
+ for (i = 0; i < TX_MODES; ++i)
+ tx_cache[i] = INT64_MAX;
+ x->skip = 0;
this_mode = vp9_mode_order[mode_index].mode;
ref_frame = vp9_mode_order[mode_index].ref_frame;
+ second_ref_frame = vp9_mode_order[mode_index].second_ref_frame;
- // Slip modes that have been masked off but always consider first mode.
- if ( mode_index && (bsize > cpi->sf.unused_mode_skip_lvl) &&
+ // Skip modes that have been masked off but always consider first mode.
+ if (mode_index && (bsize > cpi->sf.unused_mode_skip_lvl) &&
(cpi->unused_mode_skip_mask & (1 << mode_index)) )
continue;
- // Skip if the current refernce frame has been masked off
+ // Skip if the current reference frame has been masked off
if (cpi->sf.reference_masking && !cpi->set_ref_frame_mask &&
(cpi->ref_frame_mask & (1 << ref_frame)))
continue;
// Test best rd so far against threshold for trying this mode.
if ((best_rd < ((cpi->rd_threshes[bsize][mode_index] *
- cpi->rd_thresh_freq_fact[bsize][mode_index]) >> 4)) ||
+ cpi->rd_thresh_freq_fact[bsize][mode_index]) >> 5)) ||
cpi->rd_threshes[bsize][mode_index] == INT_MAX)
continue;
// Do not allow compound prediction if the segment level reference
// frame feature is in use as in this case there can only be one reference.
- if ((vp9_mode_order[mode_index].second_ref_frame > INTRA_FRAME) &&
- vp9_segfeature_active(&xd->seg, segment_id, SEG_LVL_REF_FRAME))
+ if ((second_ref_frame > INTRA_FRAME) &&
+ vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
continue;
- x->skip = 0;
-
// Skip some checking based on small partitions' result.
if (x->fast_ms > 1 && !ref_frame)
continue;
if (x->fast_ms > 2 && ref_frame != x->subblock_ref)
continue;
- if (cpi->sf.use_avoid_tested_higherror && bsize >= BLOCK_SIZE_SB8X8) {
+ if (cpi->sf.use_avoid_tested_higherror && bsize >= BLOCK_8X8) {
if (!(ref_frame_mask & (1 << ref_frame))) {
continue;
}
if (!(mode_mask & (1 << this_mode))) {
continue;
}
- if (vp9_mode_order[mode_index].second_ref_frame != NONE
- && !(ref_frame_mask
- & (1 << vp9_mode_order[mode_index].second_ref_frame))) {
+ if (second_ref_frame != NONE
+ && !(ref_frame_mask & (1 << second_ref_frame))) {
continue;
}
}
mbmi->ref_frame[0] = ref_frame;
- mbmi->ref_frame[1] = vp9_mode_order[mode_index].second_ref_frame;
+ mbmi->ref_frame[1] = second_ref_frame;
if (!(ref_frame == INTRA_FRAME
|| (cpi->ref_frame_flags & flag_list[ref_frame]))) {
continue;
}
- if (!(mbmi->ref_frame[1] == NONE
- || (cpi->ref_frame_flags & flag_list[mbmi->ref_frame[1]]))) {
+ if (!(second_ref_frame == NONE
+ || (cpi->ref_frame_flags & flag_list[second_ref_frame]))) {
continue;
}
- comp_pred = mbmi->ref_frame[1] > INTRA_FRAME;
+ comp_pred = second_ref_frame > INTRA_FRAME;
if (comp_pred) {
if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA)
if (vp9_mode_order[best_mode_index].ref_frame == INTRA_FRAME)
continue;
if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_REFMISMATCH)
- if (vp9_mode_order[mode_index].ref_frame != best_inter_ref_frame &&
- vp9_mode_order[mode_index].second_ref_frame != best_inter_ref_frame)
+ if (ref_frame != best_inter_ref_frame &&
+ second_ref_frame != best_inter_ref_frame)
continue;
}
// TODO(jingning, jkoleszar): scaling reference frame not supported for
// SPLITMV.
- if (mbmi->ref_frame[0] > 0 &&
- (scale_factor[mbmi->ref_frame[0]].x_scale_fp != VP9_REF_NO_SCALE ||
- scale_factor[mbmi->ref_frame[0]].y_scale_fp != VP9_REF_NO_SCALE) &&
+ if (ref_frame > 0 &&
+ vp9_is_scaled(&scale_factor[ref_frame]) &&
this_mode == SPLITMV)
continue;
- if (mbmi->ref_frame[1] > 0 &&
- (scale_factor[mbmi->ref_frame[1]].x_scale_fp != VP9_REF_NO_SCALE ||
- scale_factor[mbmi->ref_frame[1]].y_scale_fp != VP9_REF_NO_SCALE) &&
+ if (second_ref_frame > 0 &&
+ vp9_is_scaled(&scale_factor[second_ref_frame]) &&
this_mode == SPLITMV)
continue;
- set_scale_factors(xd, mbmi->ref_frame[0], mbmi->ref_frame[1],
- scale_factor);
+ set_scale_factors(xd, ref_frame, second_ref_frame, scale_factor);
mbmi->mode = this_mode;
mbmi->uv_mode = DC_PRED;
@@ -3423,54 +3281,51 @@
mbmi->interp_filter = cm->mcomp_filter_type;
vp9_setup_interp_filters(xd, mbmi->interp_filter, &cpi->common);
- if (bsize >= BLOCK_SIZE_SB8X8 &&
+ if (bsize >= BLOCK_8X8 &&
(this_mode == I4X4_PRED || this_mode == SPLITMV))
continue;
- if (bsize < BLOCK_SIZE_SB8X8 &&
+ if (bsize < BLOCK_8X8 &&
!(this_mode == I4X4_PRED || this_mode == SPLITMV))
continue;
if (comp_pred) {
- if (!(cpi->ref_frame_flags & flag_list[mbmi->ref_frame[1]]))
+ if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
continue;
- set_scale_factors(xd, mbmi->ref_frame[0], mbmi->ref_frame[1],
- scale_factor);
+ set_scale_factors(xd, ref_frame, second_ref_frame, scale_factor);
mode_excluded = mode_excluded
? mode_excluded
: cm->comp_pred_mode == SINGLE_PREDICTION_ONLY;
} else {
- // mbmi->ref_frame[1] = vp9_mode_order[mode_index].ref_frame[1];
- if (ref_frame != INTRA_FRAME) {
- if (mbmi->ref_frame[1] != INTRA_FRAME)
- mode_excluded =
- mode_excluded ?
- mode_excluded : cm->comp_pred_mode == COMP_PREDICTION_ONLY;
+ if (ref_frame != INTRA_FRAME && second_ref_frame != INTRA_FRAME) {
+ mode_excluded =
+ mode_excluded ?
+ mode_excluded : cm->comp_pred_mode == COMP_PREDICTION_ONLY;
}
}
- // Select predictors
+ // Select prediction reference frames.
for (i = 0; i < MAX_MB_PLANE; i++) {
xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
if (comp_pred)
- xd->plane[i].pre[1] = yv12_mb[mbmi->ref_frame[1]][i];
+ xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
}
// If the segment reference frame feature is enabled....
// then do nothing if the current ref frame is not allowed..
- if (vp9_segfeature_active(&xd->seg, segment_id, SEG_LVL_REF_FRAME) &&
- vp9_get_segdata(&xd->seg, segment_id, SEG_LVL_REF_FRAME) !=
+ if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
+ vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) !=
(int)ref_frame) {
continue;
// If the segment skip feature is enabled....
// then do nothing if the current mode is not allowed..
- } else if (vp9_segfeature_active(&xd->seg, segment_id, SEG_LVL_SKIP) &&
+ } else if (vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP) &&
(this_mode != ZEROMV && ref_frame != INTRA_FRAME)) {
continue;
// Disable this drop out case if the ref frame
// segment level feature is enabled for this segment. This is to
// prevent the possibility that we end up unable to pick any mode.
- } else if (!vp9_segfeature_active(&xd->seg, segment_id,
+ } else if (!vp9_segfeature_active(seg, segment_id,
SEG_LVL_REF_FRAME)) {
// Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
// unless ARNR filtering is enabled in which case we want
@@ -3506,9 +3361,11 @@
continue;
*/
+ // I4X4_PRED is only considered for block sizes less than 8x8.
mbmi->txfm_size = TX_4X4;
- rd_pick_intra4x4mby_modes(cpi, x, &rate, &rate_y,
- &distortion_y, INT64_MAX);
+ if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y,
+ &distortion_y, best_rd) >= best_rd)
+ continue;
rate2 += rate;
rate2 += intra_cost_penalty;
distortion2 += distortion_y;
@@ -3524,11 +3381,21 @@
distortion2 += dist_uv[TX_4X4];
distortion_uv = dist_uv[TX_4X4];
mbmi->uv_mode = mode_uv[TX_4X4];
- txfm_cache[ONLY_4X4] = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
- for (i = 0; i < NB_TXFM_MODES; ++i)
- txfm_cache[i] = txfm_cache[ONLY_4X4];
+ tx_cache[ONLY_4X4] = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+ for (i = 0; i < TX_MODES; ++i)
+ tx_cache[i] = tx_cache[ONLY_4X4];
} else if (ref_frame == INTRA_FRAME) {
TX_SIZE uv_tx;
+ // Disable intra modes other than DC_PRED for blocks with low variance
+ // Threshold for intra skipping based on source variance
+ // TODO(debargha): Specialize the threshold for super block sizes
+ static const int skip_intra_var_thresh[BLOCK_SIZES] = {
+ 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+ };
+ if ((cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
+ this_mode != DC_PRED &&
+ x->source_variance < skip_intra_var_thresh[mbmi->sb_type])
+ continue;
// Only search the oblique modes if the best so far is
// one of the neighboring directional modes
if ((cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
@@ -3541,7 +3408,7 @@
continue;
}
super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable, NULL,
- bsize, txfm_cache, best_rd);
+ bsize, tx_cache, best_rd);
if (rate_y == INT_MAX)
continue;
@@ -3564,7 +3431,7 @@
rate2 += intra_cost_penalty;
distortion2 = distortion_y + distortion_uv;
} else if (this_mode == SPLITMV) {
- const int is_comp_pred = mbmi->ref_frame[1] > 0;
+ const int is_comp_pred = second_ref_frame > 0;
int rate;
int64_t distortion;
int64_t this_rd_thresh;
@@ -3574,11 +3441,11 @@
int tmp_best_skippable = 0;
int switchable_filter_index;
int_mv *second_ref = is_comp_pred ?
- &mbmi->ref_mvs[mbmi->ref_frame[1]][0] : NULL;
+ &mbmi->ref_mvs[second_ref_frame][0] : NULL;
union b_mode_info tmp_best_bmodes[16];
MB_MODE_INFO tmp_best_mbmode;
PARTITION_INFO tmp_best_partition;
- BEST_SEG_INFO bsi[VP9_SWITCHABLE_FILTERS];
+ BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
int pred_exists = 0;
int uv_skippable;
if (is_comp_pred) {
@@ -3586,81 +3453,89 @@
if (vp9_mode_order[best_mode_index].ref_frame == INTRA_FRAME)
continue;
if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_REFMISMATCH)
- if (vp9_mode_order[mode_index].ref_frame != best_inter_ref_frame &&
- vp9_mode_order[mode_index].second_ref_frame !=
- best_inter_ref_frame)
+ if (ref_frame != best_inter_ref_frame &&
+ second_ref_frame != best_inter_ref_frame)
continue;
}
- this_rd_thresh = (mbmi->ref_frame[0] == LAST_FRAME) ?
+ this_rd_thresh = (ref_frame == LAST_FRAME) ?
cpi->rd_threshes[bsize][THR_NEWMV] :
cpi->rd_threshes[bsize][THR_NEWA];
- this_rd_thresh = (mbmi->ref_frame[0] == GOLDEN_FRAME) ?
+ this_rd_thresh = (ref_frame == GOLDEN_FRAME) ?
cpi->rd_threshes[bsize][THR_NEWG] : this_rd_thresh;
xd->mode_info_context->mbmi.txfm_size = TX_4X4;
- cpi->rd_filter_cache[VP9_SWITCHABLE_FILTERS] = INT64_MAX;
- for (switchable_filter_index = 0;
- switchable_filter_index < VP9_SWITCHABLE_FILTERS;
- ++switchable_filter_index) {
- int newbest, rs;
- int64_t rs_rd;
- mbmi->interp_filter =
- vp9_switchable_interp[switchable_filter_index];
- vp9_setup_interp_filters(xd, mbmi->interp_filter, &cpi->common);
+ cpi->rd_filter_cache[SWITCHABLE_FILTERS] = INT64_MAX;
+ if (cm->mcomp_filter_type != BILINEAR) {
+ tmp_best_filter = EIGHTTAP;
+ if (x->source_variance <
+ cpi->sf.disable_filter_search_var_thresh) {
+ tmp_best_filter = EIGHTTAP;
+ vp9_zero(cpi->rd_filter_cache);
+ } else {
+ for (switchable_filter_index = 0;
+ switchable_filter_index < SWITCHABLE_FILTERS;
+ ++switchable_filter_index) {
+ int newbest, rs;
+ int64_t rs_rd;
+ mbmi->interp_filter = switchable_filter_index;
+ vp9_setup_interp_filters(xd, mbmi->interp_filter, &cpi->common);
- tmp_rd = rd_pick_best_mbsegmentation(cpi, x,
- &mbmi->ref_mvs[mbmi->ref_frame[0]][0],
- second_ref,
- best_yrd,
- &rate, &rate_y, &distortion,
- &skippable, &total_sse,
- (int)this_rd_thresh, seg_mvs,
- bsi, switchable_filter_index,
- mi_row, mi_col);
+ tmp_rd = rd_pick_best_mbsegmentation(cpi, x,
+ &mbmi->ref_mvs[ref_frame][0],
+ second_ref,
+ best_yrd,
+ &rate, &rate_y, &distortion,
+ &skippable, &total_sse,
+ (int)this_rd_thresh, seg_mvs,
+ bsi, switchable_filter_index,
+ mi_row, mi_col);
- if (tmp_rd == INT64_MAX)
- continue;
- cpi->rd_filter_cache[switchable_filter_index] = tmp_rd;
- rs = get_switchable_rate(cm, x);
- rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
- cpi->rd_filter_cache[VP9_SWITCHABLE_FILTERS] =
- MIN(cpi->rd_filter_cache[VP9_SWITCHABLE_FILTERS], tmp_rd + rs_rd);
- if (cm->mcomp_filter_type == SWITCHABLE)
- tmp_rd += rs_rd;
+ if (tmp_rd == INT64_MAX)
+ continue;
+ cpi->rd_filter_cache[switchable_filter_index] = tmp_rd;
+ rs = get_switchable_rate(x);
+ rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
+ cpi->rd_filter_cache[SWITCHABLE_FILTERS] =
+ MIN(cpi->rd_filter_cache[SWITCHABLE_FILTERS],
+ tmp_rd + rs_rd);
+ if (cm->mcomp_filter_type == SWITCHABLE)
+ tmp_rd += rs_rd;
- newbest = (tmp_rd < tmp_best_rd);
- if (newbest) {
- tmp_best_filter = mbmi->interp_filter;
- tmp_best_rd = tmp_rd;
- }
- if ((newbest && cm->mcomp_filter_type == SWITCHABLE) ||
- (mbmi->interp_filter == cm->mcomp_filter_type &&
- cm->mcomp_filter_type != SWITCHABLE)) {
- tmp_best_rdu = tmp_rd;
- tmp_best_rate = rate;
- tmp_best_ratey = rate_y;
- tmp_best_distortion = distortion;
- tmp_best_sse = total_sse;
- tmp_best_skippable = skippable;
- tmp_best_mbmode = *mbmi;
- tmp_best_partition = *x->partition_info;
- for (i = 0; i < 4; i++)
- tmp_best_bmodes[i] = xd->mode_info_context->bmi[i];
- pred_exists = 1;
- if (switchable_filter_index == 0 &&
- cpi->sf.use_rd_breakout &&
- best_rd < INT64_MAX) {
- if (tmp_best_rdu / 2 > best_rd) {
- // skip searching the other filters if the first is
- // already substantially larger than the best so far
+ newbest = (tmp_rd < tmp_best_rd);
+ if (newbest) {
tmp_best_filter = mbmi->interp_filter;
- tmp_best_rdu = INT64_MAX;
- break;
+ tmp_best_rd = tmp_rd;
}
- }
+ if ((newbest && cm->mcomp_filter_type == SWITCHABLE) ||
+ (mbmi->interp_filter == cm->mcomp_filter_type &&
+ cm->mcomp_filter_type != SWITCHABLE)) {
+ tmp_best_rdu = tmp_rd;
+ tmp_best_rate = rate;
+ tmp_best_ratey = rate_y;
+ tmp_best_distortion = distortion;
+ tmp_best_sse = total_sse;
+ tmp_best_skippable = skippable;
+ tmp_best_mbmode = *mbmi;
+ tmp_best_partition = *x->partition_info;
+ for (i = 0; i < 4; i++)
+ tmp_best_bmodes[i] = xd->mode_info_context->bmi[i];
+ pred_exists = 1;
+ if (switchable_filter_index == 0 &&
+ cpi->sf.use_rd_breakout &&
+ best_rd < INT64_MAX) {
+ if (tmp_best_rdu / 2 > best_rd) {
+ // skip searching the other filters if the first is
+ // already substantially larger than the best so far
+ tmp_best_filter = mbmi->interp_filter;
+ tmp_best_rdu = INT64_MAX;
+ break;
+ }
+ }
+ }
+ } // switchable_filter_index loop
}
- } // switchable_filter_index loop
+ }
if (tmp_best_rdu == INT64_MAX)
continue;
@@ -3672,7 +3547,7 @@
// Handles the special case when a filter that is not in the
// switchable list (bilinear, 6-tap) is indicated at the frame level
tmp_rd = rd_pick_best_mbsegmentation(cpi, x,
- &mbmi->ref_mvs[mbmi->ref_frame[0]][0],
+ &mbmi->ref_mvs[ref_frame][0],
second_ref,
best_yrd,
&rate, &rate_y, &distortion,
@@ -3684,7 +3559,7 @@
continue;
} else {
if (cpi->common.mcomp_filter_type == SWITCHABLE) {
- int rs = get_switchable_rate(cm, x);
+ int rs = get_switchable_rate(x);
tmp_best_rdu -= RDCOST(x->rdmult, x->rddiv, rs, 0);
}
tmp_rd = tmp_best_rdu;
@@ -3703,7 +3578,7 @@
distortion2 += distortion;
if (cpi->common.mcomp_filter_type == SWITCHABLE)
- rate2 += get_switchable_rate(cm, x);
+ rate2 += get_switchable_rate(x);
if (!mode_excluded) {
if (is_comp_pred)
@@ -3713,30 +3588,32 @@
}
compmode_cost = vp9_cost_bit(comp_mode_p, is_comp_pred);
- if (RDCOST(x->rdmult, x->rddiv, rate2, distortion2) <
- best_rd) {
+ tmp_best_rdu = best_rd -
+ MIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
+ RDCOST(x->rdmult, x->rddiv, 0, total_sse));
+
+ if (tmp_best_rdu > 0) {
// If even the 'Y' rd value of split is higher than best so far
// then dont bother looking at UV
vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col,
- BLOCK_SIZE_SB8X8);
- vp9_subtract_sbuv(x, BLOCK_SIZE_SB8X8);
- super_block_uvrd_for_txfm(cm, x, &rate_uv, &distortion_uv,
- &uv_skippable, &uv_sse,
- BLOCK_SIZE_SB8X8, TX_4X4);
+ BLOCK_8X8);
+ super_block_uvrd(cm, x, &rate_uv, &distortion_uv, &uv_skippable,
+ &uv_sse, BLOCK_8X8, tmp_best_rdu);
+ if (rate_uv == INT_MAX)
+ continue;
rate2 += rate_uv;
distortion2 += distortion_uv;
skippable = skippable && uv_skippable;
total_sse += uv_sse;
- txfm_cache[ONLY_4X4] = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
- for (i = 0; i < NB_TXFM_MODES; ++i)
- txfm_cache[i] = txfm_cache[ONLY_4X4];
+ tx_cache[ONLY_4X4] = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+ for (i = 0; i < TX_MODES; ++i)
+ tx_cache[i] = tx_cache[ONLY_4X4];
}
} else {
- compmode_cost = vp9_cost_bit(comp_mode_p,
- mbmi->ref_frame[1] > INTRA_FRAME);
+ compmode_cost = vp9_cost_bit(comp_mode_p, second_ref_frame > INTRA_FRAME);
this_rd = handle_inter_mode(cpi, x, bsize,
- txfm_cache,
+ tx_cache,
&rate2, &distortion2, &skippable,
&rate_y, &distortion_y,
&rate_uv, &distortion_uv,
@@ -3754,10 +3631,10 @@
// Estimate the reference frame signaling cost and add it
// to the rolling cost variable.
- if (mbmi->ref_frame[1] > INTRA_FRAME) {
- rate2 += ref_costs_comp[mbmi->ref_frame[0]];
+ if (second_ref_frame > INTRA_FRAME) {
+ rate2 += ref_costs_comp[ref_frame];
} else {
- rate2 += ref_costs_single[mbmi->ref_frame[0]];
+ rate2 += ref_costs_single[ref_frame];
}
if (!disable_skip) {
@@ -3766,10 +3643,10 @@
// necessary adjustment for rate. Ignore if skip is coded at
// segment level as the cost wont have been added in.
// Is Mb level skip allowed (i.e. not coded at segment level).
- const int mb_skip_allowed = !vp9_segfeature_active(&xd->seg, segment_id,
+ const int mb_skip_allowed = !vp9_segfeature_active(seg, segment_id,
SEG_LVL_SKIP);
- if (skippable && bsize >= BLOCK_SIZE_SB8X8) {
+ if (skippable && bsize >= BLOCK_8X8) {
// Back out the coefficient coding costs
rate2 -= (rate_y + rate_uv);
// for best yrd calculation
@@ -3787,8 +3664,7 @@
rate2 += prob_skip_cost;
}
}
- } else if (mb_skip_allowed && ref_frame != INTRA_FRAME &&
- !xd->lossless) {
+ } else if (mb_skip_allowed && ref_frame != INTRA_FRAME && !xd->lossless) {
if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
// Add in the cost of the no skip flag.
@@ -3835,10 +3711,10 @@
// best_inter_mode = xd->mode_info_context->mbmi.mode;
}
- if (!disable_skip && mbmi->ref_frame[0] == INTRA_FRAME) {
+ if (!disable_skip && ref_frame == INTRA_FRAME) {
for (i = 0; i < NB_PREDICTION_TYPES; ++i)
best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
- for (i = 0; i <= VP9_SWITCHABLE_FILTERS; i++)
+ for (i = 0; i <= SWITCHABLE_FILTERS; i++)
best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
}
@@ -3848,9 +3724,9 @@
|| distortion2 < mode_distortions[this_mode]) {
mode_distortions[this_mode] = distortion2;
}
- if (frame_distortions[mbmi->ref_frame[0]] == -1
- || distortion2 < frame_distortions[mbmi->ref_frame[0]]) {
- frame_distortions[mbmi->ref_frame[0]] = distortion2;
+ if (frame_distortions[ref_frame] == -1
+ || distortion2 < frame_distortions[ref_frame]) {
+ frame_distortions[ref_frame] = distortion2;
}
}
@@ -3858,8 +3734,6 @@
if (this_rd < best_rd || x->skip) {
if (!mode_excluded) {
// Note index of best mode so far
- const int qstep = xd->plane[0].dequant[1];
-
best_mode_index = mode_index;
if (ref_frame == INTRA_FRAME) {
@@ -3882,37 +3756,24 @@
// TODO(debargha): enhance this test with a better distortion prediction
// based on qp, activity mask and history
- if (cpi->sf.mode_search_skip_flags & FLAG_EARLY_TERMINATE)
- if (ref_frame > INTRA_FRAME && distortion2 * 4 < qstep * qstep)
+ if (cpi->sf.mode_search_skip_flags & FLAG_EARLY_TERMINATE) {
+ const int qstep = xd->plane[0].dequant[1];
+ // TODO(debargha): Enhance this by specializing for each mode_index
+ int scale = 4;
+ if (x->source_variance < UINT_MAX) {
+ const int var_adjust = (x->source_variance < 16);
+ scale -= var_adjust;
+ }
+ if (ref_frame > INTRA_FRAME &&
+ distortion2 * scale < qstep * qstep) {
early_term = 1;
+ }
+ }
}
-#if 0
- // Testing this mode gave rise to an improvement in best error score.
- // Lower threshold a bit for next time
- cpi->rd_thresh_mult[mode_index] =
- (cpi->rd_thresh_mult[mode_index] >= (MIN_THRESHMULT + 2)) ?
- cpi->rd_thresh_mult[mode_index] - 2 : MIN_THRESHMULT;
- cpi->rd_threshes[mode_index] =
- (cpi->rd_baseline_thresh[mode_index] >> 7)
- * cpi->rd_thresh_mult[mode_index];
-#endif
- } else {
- // If the mode did not help improve the best error case then
- // raise the threshold for testing that mode next time around.
-#if 0
- cpi->rd_thresh_mult[mode_index] += 4;
-
- if (cpi->rd_thresh_mult[mode_index] > MAX_THRESHMULT)
- cpi->rd_thresh_mult[mode_index] = MAX_THRESHMULT;
-
- cpi->rd_threshes[mode_index] =
- (cpi->rd_baseline_thresh[mode_index] >> 7)
- * cpi->rd_thresh_mult[mode_index];
-#endif
}
/* keep record of best compound/single-only prediction */
- if (!disable_skip && mbmi->ref_frame[0] != INTRA_FRAME) {
+ if (!disable_skip && ref_frame != INTRA_FRAME) {
int single_rd, hybrid_rd, single_rate, hybrid_rate;
if (cpi->common.comp_pred_mode == HYBRID_PREDICTION) {
@@ -3926,10 +3787,10 @@
single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
- if (mbmi->ref_frame[1] <= INTRA_FRAME &&
+ if (second_ref_frame <= INTRA_FRAME &&
single_rd < best_pred_rd[SINGLE_PREDICTION_ONLY]) {
best_pred_rd[SINGLE_PREDICTION_ONLY] = single_rd;
- } else if (mbmi->ref_frame[1] > INTRA_FRAME &&
+ } else if (second_ref_frame > INTRA_FRAME &&
single_rd < best_pred_rd[COMP_PREDICTION_ONLY]) {
best_pred_rd[COMP_PREDICTION_ONLY] = single_rd;
}
@@ -3938,12 +3799,11 @@
}
/* keep record of best filter type */
- if (!mode_excluded && !disable_skip && mbmi->ref_frame[0] != INTRA_FRAME &&
+ if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
cm->mcomp_filter_type != BILINEAR) {
int64_t ref = cpi->rd_filter_cache[cm->mcomp_filter_type == SWITCHABLE ?
- VP9_SWITCHABLE_FILTERS :
- vp9_switchable_interp_map[cm->mcomp_filter_type]];
- for (i = 0; i <= VP9_SWITCHABLE_FILTERS; i++) {
+ SWITCHABLE_FILTERS : cm->mcomp_filter_type];
+ for (i = 0; i <= SWITCHABLE_FILTERS; i++) {
int64_t adj_rd;
// In cases of poor prediction, filter_cache[] can contain really big
// values, which actually are bigger than this_rd itself. This can
@@ -3958,34 +3818,35 @@
}
/* keep record of best txfm size */
- if (bsize < BLOCK_SIZE_SB32X32) {
- if (bsize < BLOCK_SIZE_MB16X16) {
+ if (bsize < BLOCK_32X32) {
+ if (bsize < BLOCK_16X16) {
if (this_mode == SPLITMV || this_mode == I4X4_PRED)
- txfm_cache[ALLOW_8X8] = txfm_cache[ONLY_4X4];
- txfm_cache[ALLOW_16X16] = txfm_cache[ALLOW_8X8];
+ tx_cache[ALLOW_8X8] = tx_cache[ONLY_4X4];
+ tx_cache[ALLOW_16X16] = tx_cache[ALLOW_8X8];
}
- txfm_cache[ALLOW_32X32] = txfm_cache[ALLOW_16X16];
+ tx_cache[ALLOW_32X32] = tx_cache[ALLOW_16X16];
}
if (!mode_excluded && this_rd != INT64_MAX) {
- for (i = 0; i < NB_TXFM_MODES; i++) {
+ for (i = 0; i < TX_MODES; i++) {
int64_t adj_rd = INT64_MAX;
if (this_mode != I4X4_PRED) {
- adj_rd = this_rd + txfm_cache[i] - txfm_cache[cm->tx_mode];
+ adj_rd = this_rd + tx_cache[i] - tx_cache[cm->tx_mode];
} else {
adj_rd = this_rd;
}
- if (adj_rd < best_txfm_rd[i])
- best_txfm_rd[i] = adj_rd;
+ if (adj_rd < best_tx_rd[i])
+ best_tx_rd[i] = adj_rd;
}
}
if (early_term)
break;
- if (x->skip && !mode_excluded)
+ if (x->skip && !comp_pred)
break;
}
+
if (best_rd >= best_rd_so_far)
return INT64_MAX;
@@ -3998,8 +3859,7 @@
&rate_uv_tokenonly[uv_tx_size],
&dist_uv[uv_tx_size],
&skip_uv[uv_tx_size],
- (bsize < BLOCK_SIZE_SB8X8) ? BLOCK_SIZE_SB8X8
- : bsize);
+ bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
}
}
@@ -4034,7 +3894,7 @@
}
}
- if (best_rd == INT64_MAX && bsize < BLOCK_SIZE_SB8X8) {
+ if (best_rd == INT64_MAX && bsize < BLOCK_8X8) {
*returnrate = INT_MAX;
*returndistortion = INT_MAX;
return best_rd;
@@ -4044,7 +3904,7 @@
(cm->mcomp_filter_type == best_mbmode.interp_filter) ||
(best_mbmode.ref_frame[0] == INTRA_FRAME));
- // Updating rd_thresh_freq_fact[] here means that the differnt
+ // Updating rd_thresh_freq_fact[] here means that the different
// partition/block sizes are handled independently based on the best
// choice for the current partition. It may well be better to keep a scaled
// best rd so far value and update rd_thresh_freq_fact based on the mode/size
@@ -4052,44 +3912,30 @@
if (cpi->sf.adaptive_rd_thresh) {
for (mode_index = 0; mode_index < MAX_MODES; ++mode_index) {
if (mode_index == best_mode_index) {
- cpi->rd_thresh_freq_fact[bsize][mode_index] = BASE_RD_THRESH_FREQ_FACT;
+ cpi->rd_thresh_freq_fact[bsize][mode_index] -=
+ (cpi->rd_thresh_freq_fact[bsize][mode_index] >> 3);
} else {
- cpi->rd_thresh_freq_fact[bsize][mode_index] += MAX_RD_THRESH_FREQ_INC;
+ cpi->rd_thresh_freq_fact[bsize][mode_index] += RD_THRESH_INC;
if (cpi->rd_thresh_freq_fact[bsize][mode_index] >
- (cpi->sf.adaptive_rd_thresh * MAX_RD_THRESH_FREQ_FACT)) {
+ (cpi->sf.adaptive_rd_thresh * MAX_RD_THRESH_FACT)) {
cpi->rd_thresh_freq_fact[bsize][mode_index] =
- cpi->sf.adaptive_rd_thresh * MAX_RD_THRESH_FREQ_FACT;
+ cpi->sf.adaptive_rd_thresh * MAX_RD_THRESH_FACT;
}
}
}
}
- // TODO(rbultje) integrate with RD trd_thresh_freq_facthresholding
-#if 0
- // Reduce the activation RD thresholds for the best choice mode
- if ((cpi->rd_baseline_thresh[best_mode_index] > 0) &&
- (cpi->rd_baseline_thresh[best_mode_index] < (INT_MAX >> 2))) {
- int best_adjustment = (cpi->rd_thresh_mult[best_mode_index] >> 2);
-
- cpi->rd_thresh_mult[best_mode_index] =
- (cpi->rd_thresh_mult[best_mode_index] >= (MIN_THRESHMULT + best_adjustment)) ?
- cpi->rd_thresh_mult[best_mode_index] - best_adjustment : MIN_THRESHMULT;
- cpi->rd_threshes[best_mode_index] =
- (cpi->rd_baseline_thresh[best_mode_index] >> 7) * cpi->rd_thresh_mult[best_mode_index];
- }
-#endif
-
// macroblock modes
*mbmi = best_mbmode;
x->skip |= best_skip2;
if (best_mbmode.ref_frame[0] == INTRA_FRAME &&
- best_mbmode.sb_type < BLOCK_SIZE_SB8X8) {
+ best_mbmode.sb_type < BLOCK_8X8) {
for (i = 0; i < 4; i++)
xd->mode_info_context->bmi[i].as_mode = best_bmodes[i].as_mode;
}
if (best_mbmode.ref_frame[0] != INTRA_FRAME &&
- best_mbmode.sb_type < BLOCK_SIZE_SB8X8) {
+ best_mbmode.sb_type < BLOCK_8X8) {
for (i = 0; i < 4; i++)
xd->mode_info_context->bmi[i].as_mv[0].as_int =
best_bmodes[i].as_mv[0].as_int;
@@ -4113,27 +3959,27 @@
}
if (!x->skip) {
- for (i = 0; i <= VP9_SWITCHABLE_FILTERS; i++) {
+ for (i = 0; i <= SWITCHABLE_FILTERS; i++) {
if (best_filter_rd[i] == INT64_MAX)
best_filter_diff[i] = 0;
else
best_filter_diff[i] = best_rd - best_filter_rd[i];
}
if (cm->mcomp_filter_type == SWITCHABLE)
- assert(best_filter_diff[VP9_SWITCHABLE_FILTERS] == 0);
+ assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
} else {
vpx_memset(best_filter_diff, 0, sizeof(best_filter_diff));
}
if (!x->skip) {
- for (i = 0; i < NB_TXFM_MODES; i++) {
- if (best_txfm_rd[i] == INT64_MAX)
- best_txfm_diff[i] = 0;
+ for (i = 0; i < TX_MODES; i++) {
+ if (best_tx_rd[i] == INT64_MAX)
+ best_tx_diff[i] = 0;
else
- best_txfm_diff[i] = best_rd - best_txfm_rd[i];
+ best_tx_diff[i] = best_rd - best_tx_rd[i];
}
} else {
- vpx_memset(best_txfm_diff, 0, sizeof(best_txfm_diff));
+ vpx_memset(best_tx_diff, 0, sizeof(best_tx_diff));
}
set_scale_factors(xd, mbmi->ref_frame[0], mbmi->ref_frame[1],
@@ -4143,7 +3989,7 @@
&mbmi->ref_mvs[mbmi->ref_frame[0]][0],
&mbmi->ref_mvs[mbmi->ref_frame[1] < 0 ? 0 :
mbmi->ref_frame[1]][0],
- best_pred_diff, best_txfm_diff, best_filter_diff);
+ best_pred_diff, best_tx_diff, best_filter_diff);
return best_rd;
}
« no previous file with comments | « source/libvpx/vp9/encoder/vp9_rdopt.h ('k') | source/libvpx/vp9/encoder/vp9_segmentation.c » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698