Line data Source code
1 : /*
2 : * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3 : *
4 : * Use of this source code is governed by a BSD-style license
5 : * that can be found in the LICENSE file in the root of the source
6 : * tree. An additional intellectual property rights grant can be found
7 : * in the file PATENTS. All contributing project authors may
8 : * be found in the AUTHORS file in the root of the source tree.
9 : */
10 :
11 : #include <assert.h>
12 : #include <math.h>
13 : #include <stdio.h>
14 :
15 : #include "./vp9_rtcd.h"
16 :
17 : #include "vpx_dsp/vpx_dsp_common.h"
18 : #include "vpx_mem/vpx_mem.h"
19 : #include "vpx_ports/bitops.h"
20 : #include "vpx_ports/mem.h"
21 : #include "vpx_ports/system_state.h"
22 :
23 : #include "vp9/common/vp9_common.h"
24 : #include "vp9/common/vp9_entropy.h"
25 : #include "vp9/common/vp9_entropymode.h"
26 : #include "vp9/common/vp9_mvref_common.h"
27 : #include "vp9/common/vp9_pred_common.h"
28 : #include "vp9/common/vp9_quant_common.h"
29 : #include "vp9/common/vp9_reconinter.h"
30 : #include "vp9/common/vp9_reconintra.h"
31 : #include "vp9/common/vp9_seg_common.h"
32 :
33 : #include "vp9/encoder/vp9_cost.h"
34 : #include "vp9/encoder/vp9_encodemb.h"
35 : #include "vp9/encoder/vp9_encodemv.h"
36 : #include "vp9/encoder/vp9_encoder.h"
37 : #include "vp9/encoder/vp9_mcomp.h"
38 : #include "vp9/encoder/vp9_quantize.h"
39 : #include "vp9/encoder/vp9_ratectrl.h"
40 : #include "vp9/encoder/vp9_rd.h"
41 : #include "vp9/encoder/vp9_tokenize.h"
42 :
43 : #define RD_THRESH_POW 1.25
44 :
45 : // Factor to weigh the rate for switchable interp filters.
46 : #define SWITCHABLE_INTERP_RATE_FACTOR 1
47 :
48 0 : void vp9_rd_cost_reset(RD_COST *rd_cost) {
49 0 : rd_cost->rate = INT_MAX;
50 0 : rd_cost->dist = INT64_MAX;
51 0 : rd_cost->rdcost = INT64_MAX;
52 0 : }
53 :
54 0 : void vp9_rd_cost_init(RD_COST *rd_cost) {
55 0 : rd_cost->rate = 0;
56 0 : rd_cost->dist = 0;
57 0 : rd_cost->rdcost = 0;
58 0 : }
59 :
60 : // The baseline rd thresholds for breaking out of the rd loop for
61 : // certain modes are assumed to be based on 8x8 blocks.
62 : // This table is used to correct for block size.
63 : // The factors here are << 2 (2 = x0.5, 32 = x8 etc).
64 : static const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES] = {
65 : 2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32
66 : };
67 :
68 0 : static void fill_mode_costs(VP9_COMP *cpi) {
69 0 : const FRAME_CONTEXT *const fc = cpi->common.fc;
70 : int i, j;
71 :
72 0 : for (i = 0; i < INTRA_MODES; ++i)
73 0 : for (j = 0; j < INTRA_MODES; ++j)
74 0 : vp9_cost_tokens(cpi->y_mode_costs[i][j], vp9_kf_y_mode_prob[i][j],
75 : vp9_intra_mode_tree);
76 :
77 0 : vp9_cost_tokens(cpi->mbmode_cost, fc->y_mode_prob[1], vp9_intra_mode_tree);
78 0 : for (i = 0; i < INTRA_MODES; ++i) {
79 0 : vp9_cost_tokens(cpi->intra_uv_mode_cost[KEY_FRAME][i],
80 0 : vp9_kf_uv_mode_prob[i], vp9_intra_mode_tree);
81 0 : vp9_cost_tokens(cpi->intra_uv_mode_cost[INTER_FRAME][i],
82 0 : fc->uv_mode_prob[i], vp9_intra_mode_tree);
83 : }
84 :
85 0 : for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
86 0 : vp9_cost_tokens(cpi->switchable_interp_costs[i],
87 0 : fc->switchable_interp_prob[i], vp9_switchable_interp_tree);
88 0 : }
89 :
90 0 : static void fill_token_costs(vp9_coeff_cost *c,
91 : vp9_coeff_probs_model (*p)[PLANE_TYPES]) {
92 : int i, j, k, l;
93 : TX_SIZE t;
94 0 : for (t = TX_4X4; t <= TX_32X32; ++t)
95 0 : for (i = 0; i < PLANE_TYPES; ++i)
96 0 : for (j = 0; j < REF_TYPES; ++j)
97 0 : for (k = 0; k < COEF_BANDS; ++k)
98 0 : for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
99 : vpx_prob probs[ENTROPY_NODES];
100 0 : vp9_model_to_full_probs(p[t][i][j][k][l], probs);
101 0 : vp9_cost_tokens((int *)c[t][i][j][k][0][l], probs, vp9_coef_tree);
102 0 : vp9_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs,
103 : vp9_coef_tree);
104 0 : assert(c[t][i][j][k][0][l][EOB_TOKEN] ==
105 : c[t][i][j][k][1][l][EOB_TOKEN]);
106 : }
107 0 : }
108 :
109 : // Values are now correlated to quantizer.
110 : static int sad_per_bit16lut_8[QINDEX_RANGE];
111 : static int sad_per_bit4lut_8[QINDEX_RANGE];
112 :
113 : #if CONFIG_VP9_HIGHBITDEPTH
114 : static int sad_per_bit16lut_10[QINDEX_RANGE];
115 : static int sad_per_bit4lut_10[QINDEX_RANGE];
116 : static int sad_per_bit16lut_12[QINDEX_RANGE];
117 : static int sad_per_bit4lut_12[QINDEX_RANGE];
118 : #endif
119 :
120 0 : static void init_me_luts_bd(int *bit16lut, int *bit4lut, int range,
121 : vpx_bit_depth_t bit_depth) {
122 : int i;
123 : // Initialize the sad lut tables using a formulaic calculation for now.
124 : // This is to make it easier to resolve the impact of experimental changes
125 : // to the quantizer tables.
126 0 : for (i = 0; i < range; i++) {
127 0 : const double q = vp9_convert_qindex_to_q(i, bit_depth);
128 0 : bit16lut[i] = (int)(0.0418 * q + 2.4107);
129 0 : bit4lut[i] = (int)(0.063 * q + 2.742);
130 : }
131 0 : }
132 :
133 0 : void vp9_init_me_luts(void) {
134 0 : init_me_luts_bd(sad_per_bit16lut_8, sad_per_bit4lut_8, QINDEX_RANGE,
135 : VPX_BITS_8);
136 : #if CONFIG_VP9_HIGHBITDEPTH
137 : init_me_luts_bd(sad_per_bit16lut_10, sad_per_bit4lut_10, QINDEX_RANGE,
138 : VPX_BITS_10);
139 : init_me_luts_bd(sad_per_bit16lut_12, sad_per_bit4lut_12, QINDEX_RANGE,
140 : VPX_BITS_12);
141 : #endif
142 0 : }
143 :
144 : static const int rd_boost_factor[16] = { 64, 32, 32, 32, 24, 16, 12, 12,
145 : 8, 8, 4, 4, 2, 2, 1, 0 };
146 : static const int rd_frame_type_factor[FRAME_UPDATE_TYPES] = { 128, 144, 128,
147 : 128, 144 };
148 :
149 0 : int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex) {
150 0 : const int64_t q = vp9_dc_quant(qindex, 0, cpi->common.bit_depth);
151 : #if CONFIG_VP9_HIGHBITDEPTH
152 : int64_t rdmult = 0;
153 : switch (cpi->common.bit_depth) {
154 : case VPX_BITS_8: rdmult = 88 * q * q / 24; break;
155 : case VPX_BITS_10: rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 4); break;
156 : case VPX_BITS_12: rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 8); break;
157 : default:
158 : assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
159 : return -1;
160 : }
161 : #else
162 0 : int64_t rdmult = 88 * q * q / 24;
163 : #endif // CONFIG_VP9_HIGHBITDEPTH
164 0 : if (cpi->oxcf.pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
165 0 : const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
166 0 : const FRAME_UPDATE_TYPE frame_type = gf_group->update_type[gf_group->index];
167 0 : const int boost_index = VPXMIN(15, (cpi->rc.gfu_boost / 100));
168 :
169 0 : rdmult = (rdmult * rd_frame_type_factor[frame_type]) >> 7;
170 0 : rdmult += ((rdmult * rd_boost_factor[boost_index]) >> 7);
171 : }
172 0 : if (rdmult < 1) rdmult = 1;
173 0 : return (int)rdmult;
174 : }
175 :
176 0 : static int compute_rd_thresh_factor(int qindex, vpx_bit_depth_t bit_depth) {
177 : double q;
178 : #if CONFIG_VP9_HIGHBITDEPTH
179 : switch (bit_depth) {
180 : case VPX_BITS_8: q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0; break;
181 : case VPX_BITS_10: q = vp9_dc_quant(qindex, 0, VPX_BITS_10) / 16.0; break;
182 : case VPX_BITS_12: q = vp9_dc_quant(qindex, 0, VPX_BITS_12) / 64.0; break;
183 : default:
184 : assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
185 : return -1;
186 : }
187 : #else
188 : (void)bit_depth;
189 0 : q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0;
190 : #endif // CONFIG_VP9_HIGHBITDEPTH
191 : // TODO(debargha): Adjust the function below.
192 0 : return VPXMAX((int)(pow(q, RD_THRESH_POW) * 5.12), 8);
193 : }
194 :
195 0 : void vp9_initialize_me_consts(VP9_COMP *cpi, MACROBLOCK *x, int qindex) {
196 : #if CONFIG_VP9_HIGHBITDEPTH
197 : switch (cpi->common.bit_depth) {
198 : case VPX_BITS_8:
199 : x->sadperbit16 = sad_per_bit16lut_8[qindex];
200 : x->sadperbit4 = sad_per_bit4lut_8[qindex];
201 : break;
202 : case VPX_BITS_10:
203 : x->sadperbit16 = sad_per_bit16lut_10[qindex];
204 : x->sadperbit4 = sad_per_bit4lut_10[qindex];
205 : break;
206 : case VPX_BITS_12:
207 : x->sadperbit16 = sad_per_bit16lut_12[qindex];
208 : x->sadperbit4 = sad_per_bit4lut_12[qindex];
209 : break;
210 : default:
211 : assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
212 : }
213 : #else
214 : (void)cpi;
215 0 : x->sadperbit16 = sad_per_bit16lut_8[qindex];
216 0 : x->sadperbit4 = sad_per_bit4lut_8[qindex];
217 : #endif // CONFIG_VP9_HIGHBITDEPTH
218 0 : }
219 :
220 0 : static void set_block_thresholds(const VP9_COMMON *cm, RD_OPT *rd) {
221 : int i, bsize, segment_id;
222 :
223 0 : for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) {
224 0 : const int qindex =
225 0 : clamp(vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex) +
226 0 : cm->y_dc_delta_q,
227 : 0, MAXQ);
228 0 : const int q = compute_rd_thresh_factor(qindex, cm->bit_depth);
229 :
230 0 : for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
231 : // Threshold here seems unnecessarily harsh but fine given actual
232 : // range of values used for cpi->sf.thresh_mult[].
233 0 : const int t = q * rd_thresh_block_size_factor[bsize];
234 0 : const int thresh_max = INT_MAX / t;
235 :
236 0 : if (bsize >= BLOCK_8X8) {
237 0 : for (i = 0; i < MAX_MODES; ++i)
238 0 : rd->threshes[segment_id][bsize][i] = rd->thresh_mult[i] < thresh_max
239 0 : ? rd->thresh_mult[i] * t / 4
240 0 : : INT_MAX;
241 : } else {
242 0 : for (i = 0; i < MAX_REFS; ++i)
243 0 : rd->threshes[segment_id][bsize][i] =
244 0 : rd->thresh_mult_sub8x8[i] < thresh_max
245 0 : ? rd->thresh_mult_sub8x8[i] * t / 4
246 0 : : INT_MAX;
247 : }
248 : }
249 : }
250 0 : }
251 :
252 0 : void vp9_initialize_rd_consts(VP9_COMP *cpi) {
253 0 : VP9_COMMON *const cm = &cpi->common;
254 0 : MACROBLOCK *const x = &cpi->td.mb;
255 0 : MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
256 0 : RD_OPT *const rd = &cpi->rd;
257 : int i;
258 :
259 0 : vpx_clear_system_state();
260 :
261 0 : rd->RDDIV = RDDIV_BITS; // In bits (to multiply D by 128).
262 0 : rd->RDMULT = vp9_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);
263 :
264 0 : set_error_per_bit(x, rd->RDMULT);
265 :
266 0 : x->select_tx_size = (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
267 0 : cm->frame_type != KEY_FRAME)
268 : ? 0
269 0 : : 1;
270 :
271 0 : set_block_thresholds(cm, rd);
272 0 : set_partition_probs(cm, xd);
273 :
274 0 : if (cpi->oxcf.pass == 1) {
275 0 : if (!frame_is_intra_only(cm))
276 0 : vp9_build_nmv_cost_table(
277 0 : x->nmvjointcost,
278 0 : cm->allow_high_precision_mv ? x->nmvcost_hp : x->nmvcost,
279 0 : &cm->fc->nmvc, cm->allow_high_precision_mv);
280 : } else {
281 0 : if (!cpi->sf.use_nonrd_pick_mode || cm->frame_type == KEY_FRAME)
282 0 : fill_token_costs(x->token_costs, cm->fc->coef_probs);
283 :
284 0 : if (cpi->sf.partition_search_type != VAR_BASED_PARTITION ||
285 0 : cm->frame_type == KEY_FRAME) {
286 0 : for (i = 0; i < PARTITION_CONTEXTS; ++i)
287 0 : vp9_cost_tokens(cpi->partition_cost[i], get_partition_probs(xd, i),
288 : vp9_partition_tree);
289 : }
290 :
291 0 : if (!cpi->sf.use_nonrd_pick_mode || (cm->current_video_frame & 0x07) == 1 ||
292 0 : cm->frame_type == KEY_FRAME) {
293 0 : fill_mode_costs(cpi);
294 :
295 0 : if (!frame_is_intra_only(cm)) {
296 0 : vp9_build_nmv_cost_table(
297 0 : x->nmvjointcost,
298 0 : cm->allow_high_precision_mv ? x->nmvcost_hp : x->nmvcost,
299 0 : &cm->fc->nmvc, cm->allow_high_precision_mv);
300 :
301 0 : for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
302 0 : vp9_cost_tokens((int *)cpi->inter_mode_cost[i],
303 0 : cm->fc->inter_mode_probs[i], vp9_inter_mode_tree);
304 : }
305 : }
306 : }
307 0 : }
308 :
309 0 : static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) {
310 : // NOTE: The tables below must be of the same size.
311 :
312 : // The functions described below are sampled at the four most significant
313 : // bits of x^2 + 8 / 256.
314 :
315 : // Normalized rate:
316 : // This table models the rate for a Laplacian source with given variance
317 : // when quantized with a uniform quantizer with given stepsize. The
318 : // closed form expression is:
319 : // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
320 : // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
321 : // and H(x) is the binary entropy function.
322 : static const int rate_tab_q10[] = {
323 : 65536, 6086, 5574, 5275, 5063, 4899, 4764, 4651, 4553, 4389, 4255, 4142,
324 : 4044, 3958, 3881, 3811, 3748, 3635, 3538, 3453, 3376, 3307, 3244, 3186,
325 : 3133, 3037, 2952, 2877, 2809, 2747, 2690, 2638, 2589, 2501, 2423, 2353,
326 : 2290, 2232, 2179, 2130, 2084, 2001, 1928, 1862, 1802, 1748, 1698, 1651,
327 : 1608, 1530, 1460, 1398, 1342, 1290, 1243, 1199, 1159, 1086, 1021, 963,
328 : 911, 864, 821, 781, 745, 680, 623, 574, 530, 490, 455, 424,
329 : 395, 345, 304, 269, 239, 213, 190, 171, 154, 126, 104, 87,
330 : 73, 61, 52, 44, 38, 28, 21, 16, 12, 10, 8, 6,
331 : 5, 3, 2, 1, 1, 1, 0, 0,
332 : };
333 :
334 : // Normalized distortion:
335 : // This table models the normalized distortion for a Laplacian source
336 : // with given variance when quantized with a uniform quantizer
337 : // with given stepsize. The closed form expression is:
338 : // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
339 : // where x = qpstep / sqrt(variance).
340 : // Note the actual distortion is Dn * variance.
341 : static const int dist_tab_q10[] = {
342 : 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 4, 5,
343 : 5, 6, 7, 7, 8, 9, 11, 12, 13, 15, 16, 17,
344 : 18, 21, 24, 26, 29, 31, 34, 36, 39, 44, 49, 54,
345 : 59, 64, 69, 73, 78, 88, 97, 106, 115, 124, 133, 142,
346 : 151, 167, 184, 200, 215, 231, 245, 260, 274, 301, 327, 351,
347 : 375, 397, 418, 439, 458, 495, 528, 559, 587, 613, 637, 659,
348 : 680, 717, 749, 777, 801, 823, 842, 859, 874, 899, 919, 936,
349 : 949, 960, 969, 977, 983, 994, 1001, 1006, 1010, 1013, 1015, 1017,
350 : 1018, 1020, 1022, 1022, 1023, 1023, 1023, 1024,
351 : };
352 : static const int xsq_iq_q10[] = {
353 : 0, 4, 8, 12, 16, 20, 24, 28, 32,
354 : 40, 48, 56, 64, 72, 80, 88, 96, 112,
355 : 128, 144, 160, 176, 192, 208, 224, 256, 288,
356 : 320, 352, 384, 416, 448, 480, 544, 608, 672,
357 : 736, 800, 864, 928, 992, 1120, 1248, 1376, 1504,
358 : 1632, 1760, 1888, 2016, 2272, 2528, 2784, 3040, 3296,
359 : 3552, 3808, 4064, 4576, 5088, 5600, 6112, 6624, 7136,
360 : 7648, 8160, 9184, 10208, 11232, 12256, 13280, 14304, 15328,
361 : 16352, 18400, 20448, 22496, 24544, 26592, 28640, 30688, 32736,
362 : 36832, 40928, 45024, 49120, 53216, 57312, 61408, 65504, 73696,
363 : 81888, 90080, 98272, 106464, 114656, 122848, 131040, 147424, 163808,
364 : 180192, 196576, 212960, 229344, 245728,
365 : };
366 0 : const int tmp = (xsq_q10 >> 2) + 8;
367 0 : const int k = get_msb(tmp) - 3;
368 0 : const int xq = (k << 3) + ((tmp >> k) & 0x7);
369 0 : const int one_q10 = 1 << 10;
370 0 : const int a_q10 = ((xsq_q10 - xsq_iq_q10[xq]) << 10) >> (2 + k);
371 0 : const int b_q10 = one_q10 - a_q10;
372 0 : *r_q10 = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10;
373 0 : *d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10;
374 0 : }
375 :
376 0 : void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n_log2,
377 : unsigned int qstep, int *rate,
378 : int64_t *dist) {
379 : // This function models the rate and distortion for a Laplacian
380 : // source with given variance when quantized with a uniform quantizer
381 : // with given stepsize. The closed form expressions are in:
382 : // Hang and Chen, "Source Model for transform video coder and its
383 : // application - Part I: Fundamental Theory", IEEE Trans. Circ.
384 : // Sys. for Video Tech., April 1997.
385 0 : if (var == 0) {
386 0 : *rate = 0;
387 0 : *dist = 0;
388 : } else {
389 : int d_q10, r_q10;
390 : static const uint32_t MAX_XSQ_Q10 = 245727;
391 0 : const uint64_t xsq_q10_64 =
392 0 : (((uint64_t)qstep * qstep << (n_log2 + 10)) + (var >> 1)) / var;
393 0 : const int xsq_q10 = (int)VPXMIN(xsq_q10_64, MAX_XSQ_Q10);
394 0 : model_rd_norm(xsq_q10, &r_q10, &d_q10);
395 0 : *rate = ROUND_POWER_OF_TWO(r_q10 << n_log2, 10 - VP9_PROB_COST_SHIFT);
396 0 : *dist = (var * (int64_t)d_q10 + 512) >> 10;
397 : }
398 0 : }
399 :
400 0 : void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
401 : const struct macroblockd_plane *pd,
402 : ENTROPY_CONTEXT t_above[16],
403 : ENTROPY_CONTEXT t_left[16]) {
404 0 : const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
405 0 : const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
406 0 : const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
407 0 : const ENTROPY_CONTEXT *const above = pd->above_context;
408 0 : const ENTROPY_CONTEXT *const left = pd->left_context;
409 :
410 : int i;
411 0 : switch (tx_size) {
412 : case TX_4X4:
413 0 : memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
414 0 : memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
415 0 : break;
416 : case TX_8X8:
417 0 : for (i = 0; i < num_4x4_w; i += 2)
418 0 : t_above[i] = !!*(const uint16_t *)&above[i];
419 0 : for (i = 0; i < num_4x4_h; i += 2)
420 0 : t_left[i] = !!*(const uint16_t *)&left[i];
421 0 : break;
422 : case TX_16X16:
423 0 : for (i = 0; i < num_4x4_w; i += 4)
424 0 : t_above[i] = !!*(const uint32_t *)&above[i];
425 0 : for (i = 0; i < num_4x4_h; i += 4)
426 0 : t_left[i] = !!*(const uint32_t *)&left[i];
427 0 : break;
428 : case TX_32X32:
429 0 : for (i = 0; i < num_4x4_w; i += 8)
430 0 : t_above[i] = !!*(const uint64_t *)&above[i];
431 0 : for (i = 0; i < num_4x4_h; i += 8)
432 0 : t_left[i] = !!*(const uint64_t *)&left[i];
433 0 : break;
434 0 : default: assert(0 && "Invalid transform size."); break;
435 : }
436 0 : }
437 :
438 0 : void vp9_mv_pred(VP9_COMP *cpi, MACROBLOCK *x, uint8_t *ref_y_buffer,
439 : int ref_y_stride, int ref_frame, BLOCK_SIZE block_size) {
440 : int i;
441 0 : int zero_seen = 0;
442 0 : int best_index = 0;
443 0 : int best_sad = INT_MAX;
444 0 : int this_sad = INT_MAX;
445 0 : int max_mv = 0;
446 : int near_same_nearest;
447 0 : uint8_t *src_y_ptr = x->plane[0].src.buf;
448 : uint8_t *ref_y_ptr;
449 0 : const int num_mv_refs =
450 : MAX_MV_REF_CANDIDATES +
451 0 : (cpi->sf.adaptive_motion_search && block_size < x->max_partition_size);
452 :
453 : MV pred_mv[3];
454 0 : pred_mv[0] = x->mbmi_ext->ref_mvs[ref_frame][0].as_mv;
455 0 : pred_mv[1] = x->mbmi_ext->ref_mvs[ref_frame][1].as_mv;
456 0 : pred_mv[2] = x->pred_mv[ref_frame];
457 0 : assert(num_mv_refs <= (int)(sizeof(pred_mv) / sizeof(pred_mv[0])));
458 :
459 0 : near_same_nearest = x->mbmi_ext->ref_mvs[ref_frame][0].as_int ==
460 0 : x->mbmi_ext->ref_mvs[ref_frame][1].as_int;
461 : // Get the sad for each candidate reference mv.
462 0 : for (i = 0; i < num_mv_refs; ++i) {
463 0 : const MV *this_mv = &pred_mv[i];
464 : int fp_row, fp_col;
465 :
466 0 : if (i == 1 && near_same_nearest) continue;
467 0 : fp_row = (this_mv->row + 3 + (this_mv->row >= 0)) >> 3;
468 0 : fp_col = (this_mv->col + 3 + (this_mv->col >= 0)) >> 3;
469 0 : max_mv = VPXMAX(max_mv, VPXMAX(abs(this_mv->row), abs(this_mv->col)) >> 3);
470 :
471 0 : if (fp_row == 0 && fp_col == 0 && zero_seen) continue;
472 0 : zero_seen |= (fp_row == 0 && fp_col == 0);
473 :
474 0 : ref_y_ptr = &ref_y_buffer[ref_y_stride * fp_row + fp_col];
475 : // Find sad for current vector.
476 0 : this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride,
477 : ref_y_ptr, ref_y_stride);
478 : // Note if it is the best so far.
479 0 : if (this_sad < best_sad) {
480 0 : best_sad = this_sad;
481 0 : best_index = i;
482 : }
483 : }
484 :
485 : // Note the index of the mv that worked best in the reference list.
486 0 : x->mv_best_ref_index[ref_frame] = best_index;
487 0 : x->max_mv_context[ref_frame] = max_mv;
488 0 : x->pred_mv_sad[ref_frame] = best_sad;
489 0 : }
490 :
491 0 : void vp9_setup_pred_block(const MACROBLOCKD *xd,
492 : struct buf_2d dst[MAX_MB_PLANE],
493 : const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col,
494 : const struct scale_factors *scale,
495 : const struct scale_factors *scale_uv) {
496 : int i;
497 :
498 0 : dst[0].buf = src->y_buffer;
499 0 : dst[0].stride = src->y_stride;
500 0 : dst[1].buf = src->u_buffer;
501 0 : dst[2].buf = src->v_buffer;
502 0 : dst[1].stride = dst[2].stride = src->uv_stride;
503 :
504 0 : for (i = 0; i < MAX_MB_PLANE; ++i) {
505 0 : setup_pred_plane(dst + i, dst[i].buf, dst[i].stride, mi_row, mi_col,
506 : i ? scale_uv : scale, xd->plane[i].subsampling_x,
507 : xd->plane[i].subsampling_y);
508 : }
509 0 : }
510 :
511 0 : int vp9_raster_block_offset(BLOCK_SIZE plane_bsize, int raster_block,
512 : int stride) {
513 0 : const int bw = b_width_log2_lookup[plane_bsize];
514 0 : const int y = 4 * (raster_block >> bw);
515 0 : const int x = 4 * (raster_block & ((1 << bw) - 1));
516 0 : return y * stride + x;
517 : }
518 :
519 0 : int16_t *vp9_raster_block_offset_int16(BLOCK_SIZE plane_bsize, int raster_block,
520 : int16_t *base) {
521 0 : const int stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
522 0 : return base + vp9_raster_block_offset(plane_bsize, raster_block, stride);
523 : }
524 :
525 0 : YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const VP9_COMP *cpi,
526 : int ref_frame) {
527 0 : const VP9_COMMON *const cm = &cpi->common;
528 0 : const int scaled_idx = cpi->scaled_ref_idx[ref_frame - 1];
529 0 : const int ref_idx = get_ref_frame_buf_idx(cpi, ref_frame);
530 0 : return (scaled_idx != ref_idx && scaled_idx != INVALID_IDX)
531 0 : ? &cm->buffer_pool->frame_bufs[scaled_idx].buf
532 0 : : NULL;
533 : }
534 :
535 0 : int vp9_get_switchable_rate(const VP9_COMP *cpi, const MACROBLOCKD *const xd) {
536 0 : const MODE_INFO *const mi = xd->mi[0];
537 0 : const int ctx = get_pred_context_switchable_interp(xd);
538 0 : return SWITCHABLE_INTERP_RATE_FACTOR *
539 0 : cpi->switchable_interp_costs[ctx][mi->interp_filter];
540 : }
541 :
542 0 : void vp9_set_rd_speed_thresholds(VP9_COMP *cpi) {
543 : int i;
544 0 : RD_OPT *const rd = &cpi->rd;
545 0 : SPEED_FEATURES *const sf = &cpi->sf;
546 :
547 : // Set baseline threshold values.
548 0 : for (i = 0; i < MAX_MODES; ++i)
549 0 : rd->thresh_mult[i] = cpi->oxcf.mode == BEST ? -500 : 0;
550 :
551 0 : if (sf->adaptive_rd_thresh) {
552 0 : rd->thresh_mult[THR_NEARESTMV] = 300;
553 0 : rd->thresh_mult[THR_NEARESTG] = 300;
554 0 : rd->thresh_mult[THR_NEARESTA] = 300;
555 : } else {
556 0 : rd->thresh_mult[THR_NEARESTMV] = 0;
557 0 : rd->thresh_mult[THR_NEARESTG] = 0;
558 0 : rd->thresh_mult[THR_NEARESTA] = 0;
559 : }
560 :
561 0 : rd->thresh_mult[THR_DC] += 1000;
562 :
563 0 : rd->thresh_mult[THR_NEWMV] += 1000;
564 0 : rd->thresh_mult[THR_NEWA] += 1000;
565 0 : rd->thresh_mult[THR_NEWG] += 1000;
566 :
567 0 : rd->thresh_mult[THR_NEARMV] += 1000;
568 0 : rd->thresh_mult[THR_NEARA] += 1000;
569 0 : rd->thresh_mult[THR_COMP_NEARESTLA] += 1000;
570 0 : rd->thresh_mult[THR_COMP_NEARESTGA] += 1000;
571 :
572 0 : rd->thresh_mult[THR_TM] += 1000;
573 :
574 0 : rd->thresh_mult[THR_COMP_NEARLA] += 1500;
575 0 : rd->thresh_mult[THR_COMP_NEWLA] += 2000;
576 0 : rd->thresh_mult[THR_NEARG] += 1000;
577 0 : rd->thresh_mult[THR_COMP_NEARGA] += 1500;
578 0 : rd->thresh_mult[THR_COMP_NEWGA] += 2000;
579 :
580 0 : rd->thresh_mult[THR_ZEROMV] += 2000;
581 0 : rd->thresh_mult[THR_ZEROG] += 2000;
582 0 : rd->thresh_mult[THR_ZEROA] += 2000;
583 0 : rd->thresh_mult[THR_COMP_ZEROLA] += 2500;
584 0 : rd->thresh_mult[THR_COMP_ZEROGA] += 2500;
585 :
586 0 : rd->thresh_mult[THR_H_PRED] += 2000;
587 0 : rd->thresh_mult[THR_V_PRED] += 2000;
588 0 : rd->thresh_mult[THR_D45_PRED] += 2500;
589 0 : rd->thresh_mult[THR_D135_PRED] += 2500;
590 0 : rd->thresh_mult[THR_D117_PRED] += 2500;
591 0 : rd->thresh_mult[THR_D153_PRED] += 2500;
592 0 : rd->thresh_mult[THR_D207_PRED] += 2500;
593 0 : rd->thresh_mult[THR_D63_PRED] += 2500;
594 0 : }
595 :
596 0 : void vp9_set_rd_speed_thresholds_sub8x8(VP9_COMP *cpi) {
597 : static const int thresh_mult[2][MAX_REFS] = {
598 : { 2500, 2500, 2500, 4500, 4500, 2500 },
599 : { 2000, 2000, 2000, 4000, 4000, 2000 }
600 : };
601 0 : RD_OPT *const rd = &cpi->rd;
602 0 : const int idx = cpi->oxcf.mode == BEST;
603 0 : memcpy(rd->thresh_mult_sub8x8, thresh_mult[idx], sizeof(thresh_mult[idx]));
604 0 : }
605 :
606 0 : void vp9_update_rd_thresh_fact(int (*factor_buf)[MAX_MODES], int rd_thresh,
607 : int bsize, int best_mode_index) {
608 0 : if (rd_thresh > 0) {
609 0 : const int top_mode = bsize < BLOCK_8X8 ? MAX_REFS : MAX_MODES;
610 : int mode;
611 0 : for (mode = 0; mode < top_mode; ++mode) {
612 0 : const BLOCK_SIZE min_size = VPXMAX(bsize - 1, BLOCK_4X4);
613 0 : const BLOCK_SIZE max_size = VPXMIN(bsize + 2, BLOCK_64X64);
614 : BLOCK_SIZE bs;
615 0 : for (bs = min_size; bs <= max_size; ++bs) {
616 0 : int *const fact = &factor_buf[bs][mode];
617 0 : if (mode == best_mode_index) {
618 0 : *fact -= (*fact >> 4);
619 : } else {
620 0 : *fact = VPXMIN(*fact + RD_THRESH_INC, rd_thresh * RD_THRESH_MAX_FACT);
621 : }
622 : }
623 : }
624 : }
625 0 : }
626 :
627 0 : int vp9_get_intra_cost_penalty(int qindex, int qdelta,
628 : vpx_bit_depth_t bit_depth) {
629 0 : const int q = vp9_dc_quant(qindex, qdelta, bit_depth);
630 : #if CONFIG_VP9_HIGHBITDEPTH
631 : switch (bit_depth) {
632 : case VPX_BITS_8: return 20 * q;
633 : case VPX_BITS_10: return 5 * q;
634 : case VPX_BITS_12: return ROUND_POWER_OF_TWO(5 * q, 2);
635 : default:
636 : assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
637 : return -1;
638 : }
639 : #else
640 0 : return 20 * q;
641 : #endif // CONFIG_VP9_HIGHBITDEPTH
642 : }
|