OLD | NEW |
(Empty) | |
| 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_mem/vpx_mem.h" |
| 18 |
| 19 #include "vp9/common/vp9_common.h" |
| 20 #include "vp9/common/vp9_entropy.h" |
| 21 #include "vp9/common/vp9_entropymode.h" |
| 22 #include "vp9/common/vp9_mvref_common.h" |
| 23 #include "vp9/common/vp9_pred_common.h" |
| 24 #include "vp9/common/vp9_quant_common.h" |
| 25 #include "vp9/common/vp9_reconinter.h" |
| 26 #include "vp9/common/vp9_reconintra.h" |
| 27 #include "vp9/common/vp9_seg_common.h" |
| 28 #include "vp9/common/vp9_systemdependent.h" |
| 29 |
| 30 #include "vp9/encoder/vp9_cost.h" |
| 31 #include "vp9/encoder/vp9_encodemb.h" |
| 32 #include "vp9/encoder/vp9_encodemv.h" |
| 33 #include "vp9/encoder/vp9_encoder.h" |
| 34 #include "vp9/encoder/vp9_mcomp.h" |
| 35 #include "vp9/encoder/vp9_quantize.h" |
| 36 #include "vp9/encoder/vp9_ratectrl.h" |
| 37 #include "vp9/encoder/vp9_rd.h" |
| 38 #include "vp9/encoder/vp9_tokenize.h" |
| 39 #include "vp9/encoder/vp9_variance.h" |
| 40 |
| 41 #define RD_THRESH_POW 1.25 |
| 42 #define RD_MULT_EPB_RATIO 64 |
| 43 |
| 44 // Factor to weigh the rate for switchable interp filters. |
| 45 #define SWITCHABLE_INTERP_RATE_FACTOR 1 |
| 46 |
| 47 // The baseline rd thresholds for breaking out of the rd loop for |
| 48 // certain modes are assumed to be based on 8x8 blocks. |
| 49 // This table is used to correct for block size. |
| 50 // The factors here are << 2 (2 = x0.5, 32 = x8 etc). |
| 51 static const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES] = { |
| 52 2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32 |
| 53 }; |
| 54 |
| 55 static void fill_mode_costs(VP9_COMP *cpi) { |
| 56 const FRAME_CONTEXT *const fc = &cpi->common.fc; |
| 57 int i, j; |
| 58 |
| 59 for (i = 0; i < INTRA_MODES; ++i) |
| 60 for (j = 0; j < INTRA_MODES; ++j) |
| 61 vp9_cost_tokens(cpi->y_mode_costs[i][j], vp9_kf_y_mode_prob[i][j], |
| 62 vp9_intra_mode_tree); |
| 63 |
| 64 vp9_cost_tokens(cpi->mbmode_cost, fc->y_mode_prob[1], vp9_intra_mode_tree); |
| 65 vp9_cost_tokens(cpi->intra_uv_mode_cost[KEY_FRAME], |
| 66 vp9_kf_uv_mode_prob[TM_PRED], vp9_intra_mode_tree); |
| 67 vp9_cost_tokens(cpi->intra_uv_mode_cost[INTER_FRAME], |
| 68 fc->uv_mode_prob[TM_PRED], vp9_intra_mode_tree); |
| 69 |
| 70 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) |
| 71 vp9_cost_tokens(cpi->switchable_interp_costs[i], |
| 72 fc->switchable_interp_prob[i], vp9_switchable_interp_tree); |
| 73 } |
| 74 |
| 75 static void fill_token_costs(vp9_coeff_cost *c, |
| 76 vp9_coeff_probs_model (*p)[PLANE_TYPES]) { |
| 77 int i, j, k, l; |
| 78 TX_SIZE t; |
| 79 for (t = TX_4X4; t <= TX_32X32; ++t) |
| 80 for (i = 0; i < PLANE_TYPES; ++i) |
| 81 for (j = 0; j < REF_TYPES; ++j) |
| 82 for (k = 0; k < COEF_BANDS; ++k) |
| 83 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) { |
| 84 vp9_prob probs[ENTROPY_NODES]; |
| 85 vp9_model_to_full_probs(p[t][i][j][k][l], probs); |
| 86 vp9_cost_tokens((int *)c[t][i][j][k][0][l], probs, |
| 87 vp9_coef_tree); |
| 88 vp9_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs, |
| 89 vp9_coef_tree); |
| 90 assert(c[t][i][j][k][0][l][EOB_TOKEN] == |
| 91 c[t][i][j][k][1][l][EOB_TOKEN]); |
| 92 } |
| 93 } |
| 94 |
| 95 static const uint8_t rd_iifactor[32] = { |
| 96 4, 4, 3, 2, 1, 0, 0, 0, |
| 97 0, 0, 0, 0, 0, 0, 0, 0, |
| 98 0, 0, 0, 0, 0, 0, 0, 0, |
| 99 0, 0, 0, 0, 0, 0, 0, 0, |
| 100 }; |
| 101 |
| 102 // Values are now correlated to quantizer. |
| 103 static int sad_per_bit16lut[QINDEX_RANGE]; |
| 104 static int sad_per_bit4lut[QINDEX_RANGE]; |
| 105 |
| 106 void vp9_init_me_luts() { |
| 107 int i; |
| 108 |
| 109 // Initialize the sad lut tables using a formulaic calculation for now. |
| 110 // This is to make it easier to resolve the impact of experimental changes |
| 111 // to the quantizer tables. |
| 112 for (i = 0; i < QINDEX_RANGE; ++i) { |
| 113 const double q = vp9_convert_qindex_to_q(i); |
| 114 sad_per_bit16lut[i] = (int)(0.0418 * q + 2.4107); |
| 115 sad_per_bit4lut[i] = (int)(0.063 * q + 2.742); |
| 116 } |
| 117 } |
| 118 |
| 119 int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex) { |
| 120 const int q = vp9_dc_quant(qindex, 0); |
| 121 // TODO(debargha): Adjust the function below. |
| 122 int rdmult = 88 * q * q / 25; |
| 123 if (cpi->pass == 2 && (cpi->common.frame_type != KEY_FRAME)) { |
| 124 if (cpi->twopass.next_iiratio > 31) |
| 125 rdmult += (rdmult * rd_iifactor[31]) >> 4; |
| 126 else |
| 127 rdmult += (rdmult * rd_iifactor[cpi->twopass.next_iiratio]) >> 4; |
| 128 } |
| 129 return rdmult; |
| 130 } |
| 131 |
| 132 static int compute_rd_thresh_factor(int qindex) { |
| 133 // TODO(debargha): Adjust the function below. |
| 134 const int q = (int)(pow(vp9_dc_quant(qindex, 0) / 4.0, RD_THRESH_POW) * 5.12); |
| 135 return MAX(q, 8); |
| 136 } |
| 137 |
| 138 void vp9_initialize_me_consts(VP9_COMP *cpi, int qindex) { |
| 139 cpi->mb.sadperbit16 = sad_per_bit16lut[qindex]; |
| 140 cpi->mb.sadperbit4 = sad_per_bit4lut[qindex]; |
| 141 } |
| 142 |
| 143 static void set_block_thresholds(const VP9_COMMON *cm, RD_OPT *rd) { |
| 144 int i, bsize, segment_id; |
| 145 |
| 146 for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) { |
| 147 const int qindex = |
| 148 clamp(vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex) + |
| 149 cm->y_dc_delta_q, |
| 150 0, MAXQ); |
| 151 const int q = compute_rd_thresh_factor(qindex); |
| 152 |
| 153 for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) { |
| 154 // Threshold here seems unnecessarily harsh but fine given actual |
| 155 // range of values used for cpi->sf.thresh_mult[]. |
| 156 const int t = q * rd_thresh_block_size_factor[bsize]; |
| 157 const int thresh_max = INT_MAX / t; |
| 158 |
| 159 if (bsize >= BLOCK_8X8) { |
| 160 for (i = 0; i < MAX_MODES; ++i) |
| 161 rd->threshes[segment_id][bsize][i] = |
| 162 rd->thresh_mult[i] < thresh_max |
| 163 ? rd->thresh_mult[i] * t / 4 |
| 164 : INT_MAX; |
| 165 } else { |
| 166 for (i = 0; i < MAX_REFS; ++i) |
| 167 rd->threshes[segment_id][bsize][i] = |
| 168 rd->thresh_mult_sub8x8[i] < thresh_max |
| 169 ? rd->thresh_mult_sub8x8[i] * t / 4 |
| 170 : INT_MAX; |
| 171 } |
| 172 } |
| 173 } |
| 174 } |
| 175 |
| 176 void vp9_initialize_rd_consts(VP9_COMP *cpi) { |
| 177 VP9_COMMON *const cm = &cpi->common; |
| 178 MACROBLOCK *const x = &cpi->mb; |
| 179 RD_OPT *const rd = &cpi->rd; |
| 180 int i; |
| 181 |
| 182 vp9_clear_system_state(); |
| 183 |
| 184 rd->RDDIV = RDDIV_BITS; // In bits (to multiply D by 128). |
| 185 rd->RDMULT = vp9_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q); |
| 186 |
| 187 x->errorperbit = rd->RDMULT / RD_MULT_EPB_RATIO; |
| 188 x->errorperbit += (x->errorperbit == 0); |
| 189 |
| 190 x->select_tx_size = (cpi->sf.tx_size_search_method == USE_LARGESTALL && |
| 191 cm->frame_type != KEY_FRAME) ? 0 : 1; |
| 192 |
| 193 set_block_thresholds(cm, rd); |
| 194 |
| 195 if (!cpi->sf.use_nonrd_pick_mode || cm->frame_type == KEY_FRAME) { |
| 196 fill_token_costs(x->token_costs, cm->fc.coef_probs); |
| 197 |
| 198 for (i = 0; i < PARTITION_CONTEXTS; ++i) |
| 199 vp9_cost_tokens(cpi->partition_cost[i], get_partition_probs(cm, i), |
| 200 vp9_partition_tree); |
| 201 } |
| 202 |
| 203 if (!cpi->sf.use_nonrd_pick_mode || (cm->current_video_frame & 0x07) == 1 || |
| 204 cm->frame_type == KEY_FRAME) { |
| 205 fill_mode_costs(cpi); |
| 206 |
| 207 if (!frame_is_intra_only(cm)) { |
| 208 vp9_build_nmv_cost_table(x->nmvjointcost, |
| 209 cm->allow_high_precision_mv ? x->nmvcost_hp |
| 210 : x->nmvcost, |
| 211 &cm->fc.nmvc, cm->allow_high_precision_mv); |
| 212 |
| 213 for (i = 0; i < INTER_MODE_CONTEXTS; ++i) |
| 214 vp9_cost_tokens((int *)cpi->inter_mode_cost[i], |
| 215 cm->fc.inter_mode_probs[i], vp9_inter_mode_tree); |
| 216 } |
| 217 } |
| 218 } |
| 219 |
| 220 static const int MAX_XSQ_Q10 = 245727; |
| 221 |
| 222 static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) { |
| 223 // NOTE: The tables below must be of the same size. |
| 224 |
| 225 // The functions described below are sampled at the four most significant |
| 226 // bits of x^2 + 8 / 256. |
| 227 |
| 228 // Normalized rate: |
| 229 // This table models the rate for a Laplacian source with given variance |
| 230 // when quantized with a uniform quantizer with given stepsize. The |
| 231 // closed form expression is: |
| 232 // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)], |
| 233 // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance), |
| 234 // and H(x) is the binary entropy function. |
| 235 static const int rate_tab_q10[] = { |
| 236 65536, 6086, 5574, 5275, 5063, 4899, 4764, 4651, |
| 237 4553, 4389, 4255, 4142, 4044, 3958, 3881, 3811, |
| 238 3748, 3635, 3538, 3453, 3376, 3307, 3244, 3186, |
| 239 3133, 3037, 2952, 2877, 2809, 2747, 2690, 2638, |
| 240 2589, 2501, 2423, 2353, 2290, 2232, 2179, 2130, |
| 241 2084, 2001, 1928, 1862, 1802, 1748, 1698, 1651, |
| 242 1608, 1530, 1460, 1398, 1342, 1290, 1243, 1199, |
| 243 1159, 1086, 1021, 963, 911, 864, 821, 781, |
| 244 745, 680, 623, 574, 530, 490, 455, 424, |
| 245 395, 345, 304, 269, 239, 213, 190, 171, |
| 246 154, 126, 104, 87, 73, 61, 52, 44, |
| 247 38, 28, 21, 16, 12, 10, 8, 6, |
| 248 5, 3, 2, 1, 1, 1, 0, 0, |
| 249 }; |
| 250 // Normalized distortion: |
| 251 // This table models the normalized distortion for a Laplacian source |
| 252 // with given variance when quantized with a uniform quantizer |
| 253 // with given stepsize. The closed form expression is: |
| 254 // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2)) |
| 255 // where x = qpstep / sqrt(variance). |
| 256 // Note the actual distortion is Dn * variance. |
| 257 static const int dist_tab_q10[] = { |
| 258 0, 0, 1, 1, 1, 2, 2, 2, |
| 259 3, 3, 4, 5, 5, 6, 7, 7, |
| 260 8, 9, 11, 12, 13, 15, 16, 17, |
| 261 18, 21, 24, 26, 29, 31, 34, 36, |
| 262 39, 44, 49, 54, 59, 64, 69, 73, |
| 263 78, 88, 97, 106, 115, 124, 133, 142, |
| 264 151, 167, 184, 200, 215, 231, 245, 260, |
| 265 274, 301, 327, 351, 375, 397, 418, 439, |
| 266 458, 495, 528, 559, 587, 613, 637, 659, |
| 267 680, 717, 749, 777, 801, 823, 842, 859, |
| 268 874, 899, 919, 936, 949, 960, 969, 977, |
| 269 983, 994, 1001, 1006, 1010, 1013, 1015, 1017, |
| 270 1018, 1020, 1022, 1022, 1023, 1023, 1023, 1024, |
| 271 }; |
| 272 static const int xsq_iq_q10[] = { |
| 273 0, 4, 8, 12, 16, 20, 24, 28, |
| 274 32, 40, 48, 56, 64, 72, 80, 88, |
| 275 96, 112, 128, 144, 160, 176, 192, 208, |
| 276 224, 256, 288, 320, 352, 384, 416, 448, |
| 277 480, 544, 608, 672, 736, 800, 864, 928, |
| 278 992, 1120, 1248, 1376, 1504, 1632, 1760, 1888, |
| 279 2016, 2272, 2528, 2784, 3040, 3296, 3552, 3808, |
| 280 4064, 4576, 5088, 5600, 6112, 6624, 7136, 7648, |
| 281 8160, 9184, 10208, 11232, 12256, 13280, 14304, 15328, |
| 282 16352, 18400, 20448, 22496, 24544, 26592, 28640, 30688, |
| 283 32736, 36832, 40928, 45024, 49120, 53216, 57312, 61408, |
| 284 65504, 73696, 81888, 90080, 98272, 106464, 114656, 122848, |
| 285 131040, 147424, 163808, 180192, 196576, 212960, 229344, 245728, |
| 286 }; |
| 287 const int tmp = (xsq_q10 >> 2) + 8; |
| 288 const int k = get_msb(tmp) - 3; |
| 289 const int xq = (k << 3) + ((tmp >> k) & 0x7); |
| 290 const int one_q10 = 1 << 10; |
| 291 const int a_q10 = ((xsq_q10 - xsq_iq_q10[xq]) << 10) >> (2 + k); |
| 292 const int b_q10 = one_q10 - a_q10; |
| 293 *r_q10 = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10; |
| 294 *d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10; |
| 295 } |
| 296 |
| 297 void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n, |
| 298 unsigned int qstep, int *rate, |
| 299 int64_t *dist) { |
| 300 // This function models the rate and distortion for a Laplacian |
| 301 // source with given variance when quantized with a uniform quantizer |
| 302 // with given stepsize. The closed form expressions are in: |
| 303 // Hang and Chen, "Source Model for transform video coder and its |
| 304 // application - Part I: Fundamental Theory", IEEE Trans. Circ. |
| 305 // Sys. for Video Tech., April 1997. |
| 306 if (var == 0) { |
| 307 *rate = 0; |
| 308 *dist = 0; |
| 309 } else { |
| 310 int d_q10, r_q10; |
| 311 const uint64_t xsq_q10_64 = |
| 312 ((((uint64_t)qstep * qstep * n) << 10) + (var >> 1)) / var; |
| 313 const int xsq_q10 = xsq_q10_64 > MAX_XSQ_Q10 ? |
| 314 MAX_XSQ_Q10 : (int)xsq_q10_64; |
| 315 model_rd_norm(xsq_q10, &r_q10, &d_q10); |
| 316 *rate = (n * r_q10 + 2) >> 2; |
| 317 *dist = (var * (int64_t)d_q10 + 512) >> 10; |
| 318 } |
| 319 } |
| 320 |
| 321 void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size, |
| 322 const struct macroblockd_plane *pd, |
| 323 ENTROPY_CONTEXT t_above[16], |
| 324 ENTROPY_CONTEXT t_left[16]) { |
| 325 const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd); |
| 326 const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize]; |
| 327 const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize]; |
| 328 const ENTROPY_CONTEXT *const above = pd->above_context; |
| 329 const ENTROPY_CONTEXT *const left = pd->left_context; |
| 330 |
| 331 int i; |
| 332 switch (tx_size) { |
| 333 case TX_4X4: |
| 334 vpx_memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w); |
| 335 vpx_memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h); |
| 336 break; |
| 337 case TX_8X8: |
| 338 for (i = 0; i < num_4x4_w; i += 2) |
| 339 t_above[i] = !!*(const uint16_t *)&above[i]; |
| 340 for (i = 0; i < num_4x4_h; i += 2) |
| 341 t_left[i] = !!*(const uint16_t *)&left[i]; |
| 342 break; |
| 343 case TX_16X16: |
| 344 for (i = 0; i < num_4x4_w; i += 4) |
| 345 t_above[i] = !!*(const uint32_t *)&above[i]; |
| 346 for (i = 0; i < num_4x4_h; i += 4) |
| 347 t_left[i] = !!*(const uint32_t *)&left[i]; |
| 348 break; |
| 349 case TX_32X32: |
| 350 for (i = 0; i < num_4x4_w; i += 8) |
| 351 t_above[i] = !!*(const uint64_t *)&above[i]; |
| 352 for (i = 0; i < num_4x4_h; i += 8) |
| 353 t_left[i] = !!*(const uint64_t *)&left[i]; |
| 354 break; |
| 355 default: |
| 356 assert(0 && "Invalid transform size."); |
| 357 } |
| 358 } |
| 359 |
| 360 void vp9_mv_pred(VP9_COMP *cpi, MACROBLOCK *x, |
| 361 uint8_t *ref_y_buffer, int ref_y_stride, |
| 362 int ref_frame, BLOCK_SIZE block_size) { |
| 363 MACROBLOCKD *xd = &x->e_mbd; |
| 364 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi; |
| 365 int_mv this_mv; |
| 366 int i; |
| 367 int zero_seen = 0; |
| 368 int best_index = 0; |
| 369 int best_sad = INT_MAX; |
| 370 int this_sad = INT_MAX; |
| 371 int max_mv = 0; |
| 372 |
| 373 uint8_t *src_y_ptr = x->plane[0].src.buf; |
| 374 uint8_t *ref_y_ptr; |
| 375 int row_offset, col_offset; |
| 376 int num_mv_refs = MAX_MV_REF_CANDIDATES + |
| 377 (cpi->sf.adaptive_motion_search && |
| 378 cpi->common.show_frame && |
| 379 block_size < cpi->sf.max_partition_size); |
| 380 |
| 381 MV pred_mv[3]; |
| 382 pred_mv[0] = mbmi->ref_mvs[ref_frame][0].as_mv; |
| 383 pred_mv[1] = mbmi->ref_mvs[ref_frame][1].as_mv; |
| 384 pred_mv[2] = x->pred_mv[ref_frame]; |
| 385 |
| 386 // Get the sad for each candidate reference mv. |
| 387 for (i = 0; i < num_mv_refs; ++i) { |
| 388 this_mv.as_mv = pred_mv[i]; |
| 389 |
| 390 max_mv = MAX(max_mv, |
| 391 MAX(abs(this_mv.as_mv.row), abs(this_mv.as_mv.col)) >> 3); |
| 392 // Only need to check zero mv once. |
| 393 if (!this_mv.as_int && zero_seen) |
| 394 continue; |
| 395 |
| 396 zero_seen = zero_seen || !this_mv.as_int; |
| 397 |
| 398 row_offset = this_mv.as_mv.row >> 3; |
| 399 col_offset = this_mv.as_mv.col >> 3; |
| 400 ref_y_ptr = ref_y_buffer + (ref_y_stride * row_offset) + col_offset; |
| 401 |
| 402 // Find sad for current vector. |
| 403 this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride, |
| 404 ref_y_ptr, ref_y_stride); |
| 405 |
| 406 // Note if it is the best so far. |
| 407 if (this_sad < best_sad) { |
| 408 best_sad = this_sad; |
| 409 best_index = i; |
| 410 } |
| 411 } |
| 412 |
| 413 // Note the index of the mv that worked best in the reference list. |
| 414 x->mv_best_ref_index[ref_frame] = best_index; |
| 415 x->max_mv_context[ref_frame] = max_mv; |
| 416 x->pred_mv_sad[ref_frame] = best_sad; |
| 417 } |
| 418 |
| 419 void vp9_setup_pred_block(const MACROBLOCKD *xd, |
| 420 struct buf_2d dst[MAX_MB_PLANE], |
| 421 const YV12_BUFFER_CONFIG *src, |
| 422 int mi_row, int mi_col, |
| 423 const struct scale_factors *scale, |
| 424 const struct scale_factors *scale_uv) { |
| 425 int i; |
| 426 |
| 427 dst[0].buf = src->y_buffer; |
| 428 dst[0].stride = src->y_stride; |
| 429 dst[1].buf = src->u_buffer; |
| 430 dst[2].buf = src->v_buffer; |
| 431 dst[1].stride = dst[2].stride = src->uv_stride; |
| 432 #if CONFIG_ALPHA |
| 433 dst[3].buf = src->alpha_buffer; |
| 434 dst[3].stride = src->alpha_stride; |
| 435 #endif |
| 436 |
| 437 for (i = 0; i < MAX_MB_PLANE; ++i) { |
| 438 setup_pred_plane(dst + i, dst[i].buf, dst[i].stride, mi_row, mi_col, |
| 439 i ? scale_uv : scale, |
| 440 xd->plane[i].subsampling_x, xd->plane[i].subsampling_y); |
| 441 } |
| 442 } |
| 443 |
| 444 const YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const VP9_COMP *cpi, |
| 445 int ref_frame) { |
| 446 const VP9_COMMON *const cm = &cpi->common; |
| 447 const int ref_idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)]; |
| 448 const int scaled_idx = cpi->scaled_ref_idx[ref_frame - 1]; |
| 449 return (scaled_idx != ref_idx) ? &cm->frame_bufs[scaled_idx].buf : NULL; |
| 450 } |
| 451 |
| 452 int vp9_get_switchable_rate(const VP9_COMP *cpi) { |
| 453 const MACROBLOCKD *const xd = &cpi->mb.e_mbd; |
| 454 const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi; |
| 455 const int ctx = vp9_get_pred_context_switchable_interp(xd); |
| 456 return SWITCHABLE_INTERP_RATE_FACTOR * |
| 457 cpi->switchable_interp_costs[ctx][mbmi->interp_filter]; |
| 458 } |
| 459 |
| 460 void vp9_set_rd_speed_thresholds(VP9_COMP *cpi) { |
| 461 int i; |
| 462 RD_OPT *const rd = &cpi->rd; |
| 463 SPEED_FEATURES *const sf = &cpi->sf; |
| 464 |
| 465 // Set baseline threshold values. |
| 466 for (i = 0; i < MAX_MODES; ++i) |
| 467 rd->thresh_mult[i] = is_best_mode(cpi->oxcf.mode) ? -500 : 0; |
| 468 |
| 469 rd->thresh_mult[THR_NEARESTMV] = 0; |
| 470 rd->thresh_mult[THR_NEARESTG] = 0; |
| 471 rd->thresh_mult[THR_NEARESTA] = 0; |
| 472 |
| 473 rd->thresh_mult[THR_DC] += 1000; |
| 474 |
| 475 rd->thresh_mult[THR_NEWMV] += 1000; |
| 476 rd->thresh_mult[THR_NEWA] += 1000; |
| 477 rd->thresh_mult[THR_NEWG] += 1000; |
| 478 |
| 479 // Adjust threshold only in real time mode, which only uses last |
| 480 // reference frame. |
| 481 rd->thresh_mult[THR_NEWMV] += sf->elevate_newmv_thresh; |
| 482 |
| 483 rd->thresh_mult[THR_NEARMV] += 1000; |
| 484 rd->thresh_mult[THR_NEARA] += 1000; |
| 485 rd->thresh_mult[THR_COMP_NEARESTLA] += 1000; |
| 486 rd->thresh_mult[THR_COMP_NEARESTGA] += 1000; |
| 487 |
| 488 rd->thresh_mult[THR_TM] += 1000; |
| 489 |
| 490 rd->thresh_mult[THR_COMP_NEARLA] += 1500; |
| 491 rd->thresh_mult[THR_COMP_NEWLA] += 2000; |
| 492 rd->thresh_mult[THR_NEARG] += 1000; |
| 493 rd->thresh_mult[THR_COMP_NEARGA] += 1500; |
| 494 rd->thresh_mult[THR_COMP_NEWGA] += 2000; |
| 495 |
| 496 rd->thresh_mult[THR_ZEROMV] += 2000; |
| 497 rd->thresh_mult[THR_ZEROG] += 2000; |
| 498 rd->thresh_mult[THR_ZEROA] += 2000; |
| 499 rd->thresh_mult[THR_COMP_ZEROLA] += 2500; |
| 500 rd->thresh_mult[THR_COMP_ZEROGA] += 2500; |
| 501 |
| 502 rd->thresh_mult[THR_H_PRED] += 2000; |
| 503 rd->thresh_mult[THR_V_PRED] += 2000; |
| 504 rd->thresh_mult[THR_D45_PRED ] += 2500; |
| 505 rd->thresh_mult[THR_D135_PRED] += 2500; |
| 506 rd->thresh_mult[THR_D117_PRED] += 2500; |
| 507 rd->thresh_mult[THR_D153_PRED] += 2500; |
| 508 rd->thresh_mult[THR_D207_PRED] += 2500; |
| 509 rd->thresh_mult[THR_D63_PRED] += 2500; |
| 510 |
| 511 // Disable frame modes if flags not set. |
| 512 if (!(cpi->ref_frame_flags & VP9_LAST_FLAG)) { |
| 513 rd->thresh_mult[THR_NEWMV ] = INT_MAX; |
| 514 rd->thresh_mult[THR_NEARESTMV] = INT_MAX; |
| 515 rd->thresh_mult[THR_ZEROMV ] = INT_MAX; |
| 516 rd->thresh_mult[THR_NEARMV ] = INT_MAX; |
| 517 } |
| 518 if (!(cpi->ref_frame_flags & VP9_GOLD_FLAG)) { |
| 519 rd->thresh_mult[THR_NEARESTG ] = INT_MAX; |
| 520 rd->thresh_mult[THR_ZEROG ] = INT_MAX; |
| 521 rd->thresh_mult[THR_NEARG ] = INT_MAX; |
| 522 rd->thresh_mult[THR_NEWG ] = INT_MAX; |
| 523 } |
| 524 if (!(cpi->ref_frame_flags & VP9_ALT_FLAG)) { |
| 525 rd->thresh_mult[THR_NEARESTA ] = INT_MAX; |
| 526 rd->thresh_mult[THR_ZEROA ] = INT_MAX; |
| 527 rd->thresh_mult[THR_NEARA ] = INT_MAX; |
| 528 rd->thresh_mult[THR_NEWA ] = INT_MAX; |
| 529 } |
| 530 |
| 531 if ((cpi->ref_frame_flags & (VP9_LAST_FLAG | VP9_ALT_FLAG)) != |
| 532 (VP9_LAST_FLAG | VP9_ALT_FLAG)) { |
| 533 rd->thresh_mult[THR_COMP_ZEROLA ] = INT_MAX; |
| 534 rd->thresh_mult[THR_COMP_NEARESTLA] = INT_MAX; |
| 535 rd->thresh_mult[THR_COMP_NEARLA ] = INT_MAX; |
| 536 rd->thresh_mult[THR_COMP_NEWLA ] = INT_MAX; |
| 537 } |
| 538 if ((cpi->ref_frame_flags & (VP9_GOLD_FLAG | VP9_ALT_FLAG)) != |
| 539 (VP9_GOLD_FLAG | VP9_ALT_FLAG)) { |
| 540 rd->thresh_mult[THR_COMP_ZEROGA ] = INT_MAX; |
| 541 rd->thresh_mult[THR_COMP_NEARESTGA] = INT_MAX; |
| 542 rd->thresh_mult[THR_COMP_NEARGA ] = INT_MAX; |
| 543 rd->thresh_mult[THR_COMP_NEWGA ] = INT_MAX; |
| 544 } |
| 545 } |
| 546 |
| 547 void vp9_set_rd_speed_thresholds_sub8x8(VP9_COMP *cpi) { |
| 548 const SPEED_FEATURES *const sf = &cpi->sf; |
| 549 RD_OPT *const rd = &cpi->rd; |
| 550 int i; |
| 551 |
| 552 for (i = 0; i < MAX_REFS; ++i) |
| 553 rd->thresh_mult_sub8x8[i] = is_best_mode(cpi->oxcf.mode) ? -500 : 0; |
| 554 |
| 555 rd->thresh_mult_sub8x8[THR_LAST] += 2500; |
| 556 rd->thresh_mult_sub8x8[THR_GOLD] += 2500; |
| 557 rd->thresh_mult_sub8x8[THR_ALTR] += 2500; |
| 558 rd->thresh_mult_sub8x8[THR_INTRA] += 2500; |
| 559 rd->thresh_mult_sub8x8[THR_COMP_LA] += 4500; |
| 560 rd->thresh_mult_sub8x8[THR_COMP_GA] += 4500; |
| 561 |
| 562 // Check for masked out split cases. |
| 563 for (i = 0; i < MAX_REFS; ++i) |
| 564 if (sf->disable_split_mask & (1 << i)) |
| 565 rd->thresh_mult_sub8x8[i] = INT_MAX; |
| 566 |
| 567 // Disable mode test if frame flag is not set. |
| 568 if (!(cpi->ref_frame_flags & VP9_LAST_FLAG)) |
| 569 rd->thresh_mult_sub8x8[THR_LAST] = INT_MAX; |
| 570 if (!(cpi->ref_frame_flags & VP9_GOLD_FLAG)) |
| 571 rd->thresh_mult_sub8x8[THR_GOLD] = INT_MAX; |
| 572 if (!(cpi->ref_frame_flags & VP9_ALT_FLAG)) |
| 573 rd->thresh_mult_sub8x8[THR_ALTR] = INT_MAX; |
| 574 if ((cpi->ref_frame_flags & (VP9_LAST_FLAG | VP9_ALT_FLAG)) != |
| 575 (VP9_LAST_FLAG | VP9_ALT_FLAG)) |
| 576 rd->thresh_mult_sub8x8[THR_COMP_LA] = INT_MAX; |
| 577 if ((cpi->ref_frame_flags & (VP9_GOLD_FLAG | VP9_ALT_FLAG)) != |
| 578 (VP9_GOLD_FLAG | VP9_ALT_FLAG)) |
| 579 rd->thresh_mult_sub8x8[THR_COMP_GA] = INT_MAX; |
| 580 } |
OLD | NEW |