/src/libvpx/vp9/encoder/vp9_rd.c
Line | Count | Source (jump to first uncovered line) |
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 | 44.7M | #define SWITCHABLE_INTERP_RATE_FACTOR 1 |
47 | | |
48 | 4.40M | void vp9_rd_cost_reset(RD_COST *rd_cost) { |
49 | 4.40M | rd_cost->rate = INT_MAX; |
50 | 4.40M | rd_cost->dist = INT64_MAX; |
51 | 4.40M | rd_cost->rdcost = INT64_MAX; |
52 | 4.40M | } |
53 | | |
54 | 6.82M | void vp9_rd_cost_init(RD_COST *rd_cost) { |
55 | 6.82M | rd_cost->rate = 0; |
56 | 6.82M | rd_cost->dist = 0; |
57 | 6.82M | rd_cost->rdcost = 0; |
58 | 6.82M | } |
59 | | |
60 | 21.0M | int64_t vp9_calculate_rd_cost(int mult, int div, int rate, int64_t dist) { |
61 | 21.0M | assert(mult >= 0); |
62 | 21.0M | assert(div > 0); |
63 | 21.0M | if (rate >= 0 && dist >= 0) { |
64 | 20.5M | return RDCOST(mult, div, rate, dist); |
65 | 20.5M | } |
66 | 543k | if (rate >= 0 && dist < 0) { |
67 | 68.1k | return RDCOST_NEG_D(mult, div, rate, -dist); |
68 | 68.1k | } |
69 | 475k | if (rate < 0 && dist >= 0) { |
70 | 475k | return RDCOST_NEG_R(mult, div, -rate, dist); |
71 | 475k | } |
72 | 13 | return -RDCOST(mult, div, -rate, -dist); |
73 | 475k | } |
74 | | |
75 | 13.3M | void vp9_rd_cost_update(int mult, int div, RD_COST *rd_cost) { |
76 | 13.3M | if (rd_cost->rate < INT_MAX && rd_cost->dist < INT64_MAX) { |
77 | 12.6M | rd_cost->rdcost = |
78 | 12.6M | vp9_calculate_rd_cost(mult, div, rd_cost->rate, rd_cost->dist); |
79 | 12.6M | } else { |
80 | 658k | vp9_rd_cost_reset(rd_cost); |
81 | 658k | } |
82 | 13.3M | } |
83 | | |
84 | | // The baseline rd thresholds for breaking out of the rd loop for |
85 | | // certain modes are assumed to be based on 8x8 blocks. |
86 | | // This table is used to correct for block size. |
87 | | // The factors here are << 2 (2 = x0.5, 32 = x8 etc). |
88 | | static const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES] = { |
89 | | 2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32 |
90 | | }; |
91 | | |
92 | 42.2k | static void fill_mode_costs(VP9_COMP *cpi) { |
93 | 42.2k | const FRAME_CONTEXT *const fc = cpi->common.fc; |
94 | 42.2k | int i, j; |
95 | | |
96 | 465k | for (i = 0; i < INTRA_MODES; ++i) { |
97 | 4.65M | for (j = 0; j < INTRA_MODES; ++j) { |
98 | 4.22M | vp9_cost_tokens(cpi->y_mode_costs[i][j], vp9_kf_y_mode_prob[i][j], |
99 | 4.22M | vp9_intra_mode_tree); |
100 | 4.22M | } |
101 | 422k | } |
102 | | |
103 | 42.2k | vp9_cost_tokens(cpi->mbmode_cost, fc->y_mode_prob[1], vp9_intra_mode_tree); |
104 | 465k | for (i = 0; i < INTRA_MODES; ++i) { |
105 | 422k | vp9_cost_tokens(cpi->intra_uv_mode_cost[KEY_FRAME][i], |
106 | 422k | vp9_kf_uv_mode_prob[i], vp9_intra_mode_tree); |
107 | 422k | vp9_cost_tokens(cpi->intra_uv_mode_cost[INTER_FRAME][i], |
108 | 422k | fc->uv_mode_prob[i], vp9_intra_mode_tree); |
109 | 422k | } |
110 | | |
111 | 211k | for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) { |
112 | 169k | vp9_cost_tokens(cpi->switchable_interp_costs[i], |
113 | 169k | fc->switchable_interp_prob[i], vp9_switchable_interp_tree); |
114 | 169k | } |
115 | | |
116 | 169k | for (i = TX_8X8; i < TX_SIZES; ++i) { |
117 | 380k | for (j = 0; j < TX_SIZE_CONTEXTS; ++j) { |
118 | 253k | const vpx_prob *tx_probs = get_tx_probs(i, j, &fc->tx_probs); |
119 | 253k | int k; |
120 | 1.01M | for (k = 0; k <= i; ++k) { |
121 | 761k | int cost = 0; |
122 | 761k | int m; |
123 | 2.11M | for (m = 0; m <= k - (k == i); ++m) { |
124 | 1.35M | if (m == k) |
125 | 507k | cost += vp9_cost_zero(tx_probs[m]); |
126 | 845k | else |
127 | 845k | cost += vp9_cost_one(tx_probs[m]); |
128 | 1.35M | } |
129 | 761k | cpi->tx_size_cost[i - 1][j][k] = cost; |
130 | 761k | } |
131 | 253k | } |
132 | 126k | } |
133 | 42.2k | } |
134 | | |
135 | | static void fill_token_costs(vp9_coeff_cost *c, |
136 | 42.2k | vp9_coeff_probs_model (*p)[PLANE_TYPES]) { |
137 | 42.2k | int i, j, k, l; |
138 | 42.2k | TX_SIZE t; |
139 | 211k | for (t = TX_4X4; t <= TX_32X32; ++t) |
140 | 507k | for (i = 0; i < PLANE_TYPES; ++i) |
141 | 1.01M | for (j = 0; j < REF_TYPES; ++j) |
142 | 4.73M | for (k = 0; k < COEF_BANDS; ++k) |
143 | 26.3M | for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) { |
144 | 22.3M | vpx_prob probs[ENTROPY_NODES]; |
145 | 22.3M | vp9_model_to_full_probs(p[t][i][j][k][l], probs); |
146 | 22.3M | vp9_cost_tokens((int *)c[t][i][j][k][0][l], probs, vp9_coef_tree); |
147 | 22.3M | vp9_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs, |
148 | 22.3M | vp9_coef_tree); |
149 | 22.3M | assert(c[t][i][j][k][0][l][EOB_TOKEN] == |
150 | 22.3M | c[t][i][j][k][1][l][EOB_TOKEN]); |
151 | 22.3M | } |
152 | 42.2k | } |
153 | | |
154 | | // Values are now correlated to quantizer. |
155 | | static int sad_per_bit16lut_8[QINDEX_RANGE]; |
156 | | static int sad_per_bit4lut_8[QINDEX_RANGE]; |
157 | | |
158 | | #if CONFIG_VP9_HIGHBITDEPTH |
159 | | static int sad_per_bit16lut_10[QINDEX_RANGE]; |
160 | | static int sad_per_bit4lut_10[QINDEX_RANGE]; |
161 | | static int sad_per_bit16lut_12[QINDEX_RANGE]; |
162 | | static int sad_per_bit4lut_12[QINDEX_RANGE]; |
163 | | #endif |
164 | | |
165 | | static void init_me_luts_bd(int *bit16lut, int *bit4lut, int range, |
166 | 3 | vpx_bit_depth_t bit_depth) { |
167 | 3 | int i; |
168 | | // Initialize the sad lut tables using a formulaic calculation for now. |
169 | | // This is to make it easier to resolve the impact of experimental changes |
170 | | // to the quantizer tables. |
171 | 771 | for (i = 0; i < range; i++) { |
172 | 768 | const double q = vp9_convert_qindex_to_q(i, bit_depth); |
173 | 768 | bit16lut[i] = (int)(0.0418 * q + 2.4107); |
174 | 768 | bit4lut[i] = (int)(0.063 * q + 2.742); |
175 | 768 | } |
176 | 3 | } |
177 | | |
178 | 1 | void vp9_init_me_luts(void) { |
179 | 1 | init_me_luts_bd(sad_per_bit16lut_8, sad_per_bit4lut_8, QINDEX_RANGE, |
180 | 1 | VPX_BITS_8); |
181 | 1 | #if CONFIG_VP9_HIGHBITDEPTH |
182 | 1 | init_me_luts_bd(sad_per_bit16lut_10, sad_per_bit4lut_10, QINDEX_RANGE, |
183 | 1 | VPX_BITS_10); |
184 | 1 | init_me_luts_bd(sad_per_bit16lut_12, sad_per_bit4lut_12, QINDEX_RANGE, |
185 | 1 | VPX_BITS_12); |
186 | 1 | #endif |
187 | 1 | } |
188 | | |
189 | | static const int rd_boost_factor[16] = { 64, 32, 32, 32, 24, 16, 12, 12, |
190 | | 8, 8, 4, 4, 2, 2, 1, 0 }; |
191 | | |
192 | | // Note that the element below for frame type "USE_BUF_FRAME", which indicates |
193 | | // that the show frame flag is set, should not be used as no real frame |
194 | | // is encoded so we should not reach here. However, a dummy value |
195 | | // is inserted here to make sure the data structure has the right number |
196 | | // of values assigned. |
197 | | static const int rd_frame_type_factor[FRAME_UPDATE_TYPES] = { 128, 144, 128, |
198 | | 128, 144, 144 }; |
199 | | |
200 | | // Configure Vizier RD parameters. |
201 | | // Later this function will use passed in command line values. |
202 | 2.98k | void vp9_init_rd_parameters(VP9_COMP *cpi) { |
203 | 2.98k | RD_CONTROL *const rdc = &cpi->rd_ctrl; |
204 | | |
205 | | // When |use_vizier_rc_params| is 1, we expect the rd parameters have been |
206 | | // initialized by the pass in values. |
207 | | // Be careful that parameters below are only initialized to 1, if we do not |
208 | | // pass values to them. It is desired to take care of each parameter when |
209 | | // using |use_vizier_rc_params|. |
210 | 2.98k | if (cpi->twopass.use_vizier_rc_params) return; |
211 | | |
212 | | // Make sure this function is floating point safe. |
213 | 2.98k | vpx_clear_system_state(); |
214 | | |
215 | 2.98k | rdc->rd_mult_inter_qp_fac = 1.0; |
216 | 2.98k | rdc->rd_mult_arf_qp_fac = 1.0; |
217 | 2.98k | rdc->rd_mult_key_qp_fac = 1.0; |
218 | 2.98k | } |
219 | | |
220 | | // Returns the default rd multiplier for inter frames for a given qindex. |
221 | | // The function here is a first pass estimate based on data from |
222 | | // a previous Vizer run |
223 | 5.66M | static double def_inter_rd_multiplier(int qindex) { |
224 | 5.66M | return 4.15 + (0.001 * (double)qindex); |
225 | 5.66M | } |
226 | | |
227 | | // Returns the default rd multiplier for ARF/Golden Frames for a given qindex. |
228 | | // The function here is a first pass estimate based on data from |
229 | | // a previous Vizer run |
230 | 394k | static double def_arf_rd_multiplier(int qindex) { |
231 | 394k | return 4.25 + (0.001 * (double)qindex); |
232 | 394k | } |
233 | | |
234 | | // Returns the default rd multiplier for key frames for a given qindex. |
235 | | // The function here is a first pass estimate based on data from |
236 | | // a previous Vizer run |
237 | 2.87M | static double def_kf_rd_multiplier(int qindex) { |
238 | 2.87M | return 4.35 + (0.001 * (double)qindex); |
239 | 2.87M | } |
240 | | |
241 | 8.93M | int vp9_compute_rd_mult_based_on_qindex(const VP9_COMP *cpi, int qindex) { |
242 | 8.93M | const RD_CONTROL *rdc = &cpi->rd_ctrl; |
243 | 8.93M | const int q = vp9_dc_quant(qindex, 0, cpi->common.bit_depth); |
244 | | // largest dc_quant is 21387, therefore rdmult should fit in int32_t |
245 | 8.93M | int rdmult = q * q; |
246 | | |
247 | 8.93M | if (cpi->ext_ratectrl.ready && |
248 | 8.93M | (cpi->ext_ratectrl.funcs.rc_type & VPX_RC_RDMULT) != 0 && |
249 | 8.93M | cpi->ext_ratectrl.ext_rdmult != VPX_DEFAULT_RDMULT) { |
250 | 0 | return cpi->ext_ratectrl.ext_rdmult; |
251 | 0 | } |
252 | | |
253 | | // Make sure this function is floating point safe. |
254 | 8.93M | vpx_clear_system_state(); |
255 | | |
256 | 8.93M | if (cpi->common.frame_type == KEY_FRAME) { |
257 | 2.87M | double def_rd_q_mult = def_kf_rd_multiplier(qindex); |
258 | 2.87M | rdmult = (int)((double)rdmult * def_rd_q_mult * rdc->rd_mult_key_qp_fac); |
259 | 6.06M | } else if (!cpi->rc.is_src_frame_alt_ref && |
260 | 6.06M | (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) { |
261 | 394k | double def_rd_q_mult = def_arf_rd_multiplier(qindex); |
262 | 394k | rdmult = (int)((double)rdmult * def_rd_q_mult * rdc->rd_mult_arf_qp_fac); |
263 | 5.66M | } else { |
264 | 5.66M | double def_rd_q_mult = def_inter_rd_multiplier(qindex); |
265 | 5.66M | rdmult = (int)((double)rdmult * def_rd_q_mult * rdc->rd_mult_inter_qp_fac); |
266 | 5.66M | } |
267 | | |
268 | 8.93M | #if CONFIG_VP9_HIGHBITDEPTH |
269 | 8.93M | switch (cpi->common.bit_depth) { |
270 | 0 | case VPX_BITS_10: rdmult = ROUND_POWER_OF_TWO(rdmult, 4); break; |
271 | 0 | case VPX_BITS_12: rdmult = ROUND_POWER_OF_TWO(rdmult, 8); break; |
272 | 8.93M | default: break; |
273 | 8.93M | } |
274 | 8.93M | #endif // CONFIG_VP9_HIGHBITDEPTH |
275 | 8.93M | return rdmult > 0 ? rdmult : 1; |
276 | 8.93M | } |
277 | | |
278 | 8.93M | static int modulate_rdmult(const VP9_COMP *cpi, int rdmult) { |
279 | 8.93M | int64_t rdmult_64 = rdmult; |
280 | 8.93M | if (cpi->oxcf.pass == 2 && (cpi->common.frame_type != KEY_FRAME)) { |
281 | 0 | const GF_GROUP *const gf_group = &cpi->twopass.gf_group; |
282 | 0 | const FRAME_UPDATE_TYPE frame_type = gf_group->update_type[gf_group->index]; |
283 | 0 | const int gfu_boost = cpi->multi_layer_arf |
284 | 0 | ? gf_group->gfu_boost[gf_group->index] |
285 | 0 | : cpi->rc.gfu_boost; |
286 | 0 | const int boost_index = VPXMIN(15, (gfu_boost / 100)); |
287 | |
|
288 | 0 | rdmult_64 = (rdmult_64 * rd_frame_type_factor[frame_type]) >> 7; |
289 | 0 | rdmult_64 += ((rdmult_64 * rd_boost_factor[boost_index]) >> 7); |
290 | 0 | } |
291 | 8.93M | return (int)rdmult_64; |
292 | 8.93M | } |
293 | | |
294 | 8.93M | int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex) { |
295 | 8.93M | int rdmult = vp9_compute_rd_mult_based_on_qindex(cpi, qindex); |
296 | 8.93M | if (cpi->ext_ratectrl.ready && |
297 | 8.93M | (cpi->ext_ratectrl.funcs.rc_type & VPX_RC_RDMULT) != 0 && |
298 | 8.93M | cpi->ext_ratectrl.ext_rdmult != VPX_DEFAULT_RDMULT) { |
299 | 0 | return cpi->ext_ratectrl.ext_rdmult; |
300 | 0 | } |
301 | 8.93M | return modulate_rdmult(cpi, rdmult); |
302 | 8.93M | } |
303 | | |
304 | 0 | int vp9_get_adaptive_rdmult(const VP9_COMP *cpi, double beta) { |
305 | 0 | int rdmult = |
306 | 0 | vp9_compute_rd_mult_based_on_qindex(cpi, cpi->common.base_qindex); |
307 | 0 | rdmult = (int)((double)rdmult / beta); |
308 | 0 | rdmult = rdmult > 0 ? rdmult : 1; |
309 | 0 | return modulate_rdmult(cpi, rdmult); |
310 | 0 | } |
311 | | |
312 | 338k | static int compute_rd_thresh_factor(int qindex, vpx_bit_depth_t bit_depth) { |
313 | 338k | double q; |
314 | 338k | #if CONFIG_VP9_HIGHBITDEPTH |
315 | 338k | switch (bit_depth) { |
316 | 338k | case VPX_BITS_8: q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0; break; |
317 | 0 | case VPX_BITS_10: q = vp9_dc_quant(qindex, 0, VPX_BITS_10) / 16.0; break; |
318 | 0 | default: |
319 | 0 | assert(bit_depth == VPX_BITS_12); |
320 | 0 | q = vp9_dc_quant(qindex, 0, VPX_BITS_12) / 64.0; |
321 | 0 | break; |
322 | 338k | } |
323 | | #else |
324 | | (void)bit_depth; |
325 | | q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0; |
326 | | #endif // CONFIG_VP9_HIGHBITDEPTH |
327 | | // TODO(debargha): Adjust the function below. |
328 | 338k | return VPXMAX((int)(pow(q, RD_THRESH_POW) * 5.12), 8); |
329 | 338k | } |
330 | | |
331 | 8.93M | void vp9_initialize_me_consts(VP9_COMP *cpi, MACROBLOCK *x, int qindex) { |
332 | 8.93M | #if CONFIG_VP9_HIGHBITDEPTH |
333 | 8.93M | switch (cpi->common.bit_depth) { |
334 | 8.93M | case VPX_BITS_8: |
335 | 8.93M | x->sadperbit16 = sad_per_bit16lut_8[qindex]; |
336 | 8.93M | x->sadperbit4 = sad_per_bit4lut_8[qindex]; |
337 | 8.93M | break; |
338 | 0 | case VPX_BITS_10: |
339 | 0 | x->sadperbit16 = sad_per_bit16lut_10[qindex]; |
340 | 0 | x->sadperbit4 = sad_per_bit4lut_10[qindex]; |
341 | 0 | break; |
342 | 0 | default: |
343 | 0 | assert(cpi->common.bit_depth == VPX_BITS_12); |
344 | 0 | x->sadperbit16 = sad_per_bit16lut_12[qindex]; |
345 | 0 | x->sadperbit4 = sad_per_bit4lut_12[qindex]; |
346 | 0 | break; |
347 | 8.93M | } |
348 | | #else |
349 | | (void)cpi; |
350 | | x->sadperbit16 = sad_per_bit16lut_8[qindex]; |
351 | | x->sadperbit4 = sad_per_bit4lut_8[qindex]; |
352 | | #endif // CONFIG_VP9_HIGHBITDEPTH |
353 | 8.93M | } |
354 | | |
355 | 42.2k | static void set_block_thresholds(const VP9_COMMON *cm, RD_OPT *rd) { |
356 | 42.2k | int i, bsize, segment_id; |
357 | | |
358 | 380k | for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) { |
359 | 338k | const int qindex = |
360 | 338k | clamp(vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex) + |
361 | 338k | cm->y_dc_delta_q, |
362 | 338k | 0, MAXQ); |
363 | 338k | const int q = compute_rd_thresh_factor(qindex, cm->bit_depth); |
364 | | |
365 | 4.73M | for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) { |
366 | | // Threshold here seems unnecessarily harsh but fine given actual |
367 | | // range of values used for cpi->sf.thresh_mult[]. |
368 | 4.39M | const int t = q * rd_thresh_block_size_factor[bsize]; |
369 | 4.39M | const int thresh_max = INT_MAX / t; |
370 | | |
371 | 4.39M | if (bsize >= BLOCK_8X8) { |
372 | 104M | for (i = 0; i < MAX_MODES; ++i) |
373 | 101M | rd->threshes[segment_id][bsize][i] = rd->thresh_mult[i] < thresh_max |
374 | 101M | ? rd->thresh_mult[i] * t / 4 |
375 | 101M | : INT_MAX; |
376 | 3.38M | } else { |
377 | 7.10M | for (i = 0; i < MAX_REFS; ++i) |
378 | 6.08M | rd->threshes[segment_id][bsize][i] = |
379 | 6.08M | rd->thresh_mult_sub8x8[i] < thresh_max |
380 | 6.08M | ? rd->thresh_mult_sub8x8[i] * t / 4 |
381 | 6.08M | : INT_MAX; |
382 | 1.01M | } |
383 | 4.39M | } |
384 | 338k | } |
385 | 42.2k | } |
386 | | |
387 | 34.5k | void vp9_build_inter_mode_cost(VP9_COMP *cpi) { |
388 | 34.5k | const VP9_COMMON *const cm = &cpi->common; |
389 | 34.5k | int i; |
390 | 276k | for (i = 0; i < INTER_MODE_CONTEXTS; ++i) { |
391 | 241k | vp9_cost_tokens((int *)cpi->inter_mode_cost[i], cm->fc->inter_mode_probs[i], |
392 | 241k | vp9_inter_mode_tree); |
393 | 241k | } |
394 | 34.5k | } |
395 | | |
396 | 42.2k | void vp9_initialize_rd_consts(VP9_COMP *cpi) { |
397 | 42.2k | VP9_COMMON *const cm = &cpi->common; |
398 | 42.2k | MACROBLOCK *const x = &cpi->td.mb; |
399 | 42.2k | MACROBLOCKD *const xd = &cpi->td.mb.e_mbd; |
400 | 42.2k | RD_OPT *const rd = &cpi->rd; |
401 | 42.2k | int i; |
402 | | |
403 | 42.2k | vpx_clear_system_state(); |
404 | | |
405 | 42.2k | rd->RDDIV = RDDIV_BITS; // In bits (to multiply D by 128). |
406 | 42.2k | rd->RDMULT = vp9_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q); |
407 | | |
408 | 42.2k | set_error_per_bit(x, rd->RDMULT); |
409 | | |
410 | 42.2k | x->select_tx_size = (cpi->sf.tx_size_search_method == USE_LARGESTALL && |
411 | 42.2k | cm->frame_type != KEY_FRAME) |
412 | 42.2k | ? 0 |
413 | 42.2k | : 1; |
414 | | |
415 | 42.2k | set_block_thresholds(cm, rd); |
416 | 42.2k | set_partition_probs(cm, xd); |
417 | | |
418 | 42.2k | if (cpi->oxcf.pass == 1) { |
419 | 0 | if (!frame_is_intra_only(cm)) |
420 | 0 | vp9_build_nmv_cost_table( |
421 | 0 | x->nmvjointcost, |
422 | 0 | cm->allow_high_precision_mv ? x->nmvcost_hp : x->nmvcost, |
423 | 0 | &cm->fc->nmvc, cm->allow_high_precision_mv); |
424 | 42.2k | } else { |
425 | 42.2k | if (!cpi->sf.use_nonrd_pick_mode || cm->frame_type == KEY_FRAME) |
426 | 42.2k | fill_token_costs(x->token_costs, cm->fc->coef_probs); |
427 | | |
428 | 42.2k | if (cpi->sf.partition_search_type != VAR_BASED_PARTITION || |
429 | 42.2k | cm->frame_type == KEY_FRAME) { |
430 | 718k | for (i = 0; i < PARTITION_CONTEXTS; ++i) |
431 | 676k | vp9_cost_tokens(cpi->partition_cost[i], get_partition_probs(xd, i), |
432 | 676k | vp9_partition_tree); |
433 | 42.2k | } |
434 | | |
435 | 42.2k | if (!cpi->sf.use_nonrd_pick_mode || (cm->current_video_frame & 0x07) == 1 || |
436 | 42.2k | cm->frame_type == KEY_FRAME) { |
437 | 42.2k | fill_mode_costs(cpi); |
438 | | |
439 | 42.2k | if (!frame_is_intra_only(cm)) { |
440 | 34.5k | vp9_build_nmv_cost_table( |
441 | 34.5k | x->nmvjointcost, |
442 | 34.5k | cm->allow_high_precision_mv ? x->nmvcost_hp : x->nmvcost, |
443 | 34.5k | &cm->fc->nmvc, cm->allow_high_precision_mv); |
444 | 34.5k | vp9_build_inter_mode_cost(cpi); |
445 | 34.5k | } |
446 | 42.2k | } |
447 | 42.2k | } |
448 | 42.2k | } |
449 | | |
450 | | // NOTE: The tables below must be of the same size. |
451 | | |
452 | | // The functions described below are sampled at the four most significant |
453 | | // bits of x^2 + 8 / 256. |
454 | | |
455 | | // Normalized rate: |
456 | | // This table models the rate for a Laplacian source with given variance |
457 | | // when quantized with a uniform quantizer with given stepsize. The |
458 | | // closed form expression is: |
459 | | // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)], |
460 | | // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance), |
461 | | // and H(x) is the binary entropy function. |
462 | | static const int rate_tab_q10[] = { |
463 | | 65536, 6086, 5574, 5275, 5063, 4899, 4764, 4651, 4553, 4389, 4255, 4142, 4044, |
464 | | 3958, 3881, 3811, 3748, 3635, 3538, 3453, 3376, 3307, 3244, 3186, 3133, 3037, |
465 | | 2952, 2877, 2809, 2747, 2690, 2638, 2589, 2501, 2423, 2353, 2290, 2232, 2179, |
466 | | 2130, 2084, 2001, 1928, 1862, 1802, 1748, 1698, 1651, 1608, 1530, 1460, 1398, |
467 | | 1342, 1290, 1243, 1199, 1159, 1086, 1021, 963, 911, 864, 821, 781, 745, |
468 | | 680, 623, 574, 530, 490, 455, 424, 395, 345, 304, 269, 239, 213, |
469 | | 190, 171, 154, 126, 104, 87, 73, 61, 52, 44, 38, 28, 21, |
470 | | 16, 12, 10, 8, 6, 5, 3, 2, 1, 1, 1, 0, 0, |
471 | | }; |
472 | | |
473 | | // Normalized distortion: |
474 | | // This table models the normalized distortion for a Laplacian source |
475 | | // with given variance when quantized with a uniform quantizer |
476 | | // with given stepsize. The closed form expression is: |
477 | | // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2)) |
478 | | // where x = qpstep / sqrt(variance). |
479 | | // Note the actual distortion is Dn * variance. |
480 | | static const int dist_tab_q10[] = { |
481 | | 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 4, 5, 5, |
482 | | 6, 7, 7, 8, 9, 11, 12, 13, 15, 16, 17, 18, 21, |
483 | | 24, 26, 29, 31, 34, 36, 39, 44, 49, 54, 59, 64, 69, |
484 | | 73, 78, 88, 97, 106, 115, 124, 133, 142, 151, 167, 184, 200, |
485 | | 215, 231, 245, 260, 274, 301, 327, 351, 375, 397, 418, 439, 458, |
486 | | 495, 528, 559, 587, 613, 637, 659, 680, 717, 749, 777, 801, 823, |
487 | | 842, 859, 874, 899, 919, 936, 949, 960, 969, 977, 983, 994, 1001, |
488 | | 1006, 1010, 1013, 1015, 1017, 1018, 1020, 1022, 1022, 1023, 1023, 1023, 1024, |
489 | | }; |
490 | | static const int xsq_iq_q10[] = { |
491 | | 0, 4, 8, 12, 16, 20, 24, 28, 32, |
492 | | 40, 48, 56, 64, 72, 80, 88, 96, 112, |
493 | | 128, 144, 160, 176, 192, 208, 224, 256, 288, |
494 | | 320, 352, 384, 416, 448, 480, 544, 608, 672, |
495 | | 736, 800, 864, 928, 992, 1120, 1248, 1376, 1504, |
496 | | 1632, 1760, 1888, 2016, 2272, 2528, 2784, 3040, 3296, |
497 | | 3552, 3808, 4064, 4576, 5088, 5600, 6112, 6624, 7136, |
498 | | 7648, 8160, 9184, 10208, 11232, 12256, 13280, 14304, 15328, |
499 | | 16352, 18400, 20448, 22496, 24544, 26592, 28640, 30688, 32736, |
500 | | 36832, 40928, 45024, 49120, 53216, 57312, 61408, 65504, 73696, |
501 | | 81888, 90080, 98272, 106464, 114656, 122848, 131040, 147424, 163808, |
502 | | 180192, 196576, 212960, 229344, 245728, |
503 | | }; |
504 | | |
505 | 83.4M | static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) { |
506 | 83.4M | const int tmp = (xsq_q10 >> 2) + 8; |
507 | 83.4M | const int k = get_msb(tmp) - 3; |
508 | 83.4M | const int xq = (k << 3) + ((tmp >> k) & 0x7); |
509 | 83.4M | const int one_q10 = 1 << 10; |
510 | 83.4M | const int a_q10 = ((xsq_q10 - xsq_iq_q10[xq]) << 10) >> (2 + k); |
511 | 83.4M | const int b_q10 = one_q10 - a_q10; |
512 | 83.4M | *r_q10 = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10; |
513 | 83.4M | *d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10; |
514 | 83.4M | } |
515 | | |
516 | | static const uint32_t MAX_XSQ_Q10 = 245727; |
517 | | |
518 | | void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n_log2, |
519 | | unsigned int qstep, int *rate, |
520 | 85.7M | int64_t *dist) { |
521 | | // This function models the rate and distortion for a Laplacian |
522 | | // source with given variance when quantized with a uniform quantizer |
523 | | // with given stepsize. The closed form expressions are in: |
524 | | // Hang and Chen, "Source Model for transform video coder and its |
525 | | // application - Part I: Fundamental Theory", IEEE Trans. Circ. |
526 | | // Sys. for Video Tech., April 1997. |
527 | 85.7M | if (var == 0) { |
528 | 2.35M | *rate = 0; |
529 | 2.35M | *dist = 0; |
530 | 83.4M | } else { |
531 | 83.4M | int d_q10, r_q10; |
532 | 83.4M | const uint64_t xsq_q10_64 = |
533 | 83.4M | (((uint64_t)qstep * qstep << (n_log2 + 10)) + (var >> 1)) / var; |
534 | 83.4M | const int xsq_q10 = (int)VPXMIN(xsq_q10_64, MAX_XSQ_Q10); |
535 | 83.4M | model_rd_norm(xsq_q10, &r_q10, &d_q10); |
536 | 83.4M | *rate = ROUND_POWER_OF_TWO(r_q10 << n_log2, 10 - VP9_PROB_COST_SHIFT); |
537 | 83.4M | *dist = (var * (int64_t)d_q10 + 512) >> 10; |
538 | 83.4M | } |
539 | 85.7M | } |
540 | | |
541 | | // Disable gcc 12.2 false positive warning. |
542 | | // warning: writing 1 byte into a region of size 0 [-Wstringop-overflow=] |
543 | | #if defined(__GNUC__) && !defined(__clang__) |
544 | | #pragma GCC diagnostic push |
545 | | #pragma GCC diagnostic ignored "-Wstringop-overflow" |
546 | | #endif |
547 | | void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size, |
548 | | const struct macroblockd_plane *pd, |
549 | | ENTROPY_CONTEXT t_above[16], |
550 | 202M | ENTROPY_CONTEXT t_left[16]) { |
551 | 202M | const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd); |
552 | 202M | const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize]; |
553 | 202M | const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize]; |
554 | 202M | const ENTROPY_CONTEXT *const above = pd->above_context; |
555 | 202M | const ENTROPY_CONTEXT *const left = pd->left_context; |
556 | | |
557 | 202M | int i; |
558 | 202M | switch (tx_size) { |
559 | 139M | case TX_4X4: |
560 | 139M | memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w); |
561 | 139M | memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h); |
562 | 139M | break; |
563 | 49.8M | case TX_8X8: |
564 | 110M | for (i = 0; i < num_4x4_w; i += 2) |
565 | 60.4M | t_above[i] = !!*(const uint16_t *)&above[i]; |
566 | 110M | for (i = 0; i < num_4x4_h; i += 2) |
567 | 60.2M | t_left[i] = !!*(const uint16_t *)&left[i]; |
568 | 49.8M | break; |
569 | 10.4M | case TX_16X16: |
570 | 23.2M | for (i = 0; i < num_4x4_w; i += 4) |
571 | 12.8M | t_above[i] = !!*(const uint32_t *)&above[i]; |
572 | 23.1M | for (i = 0; i < num_4x4_h; i += 4) |
573 | 12.7M | t_left[i] = !!*(const uint32_t *)&left[i]; |
574 | 10.4M | break; |
575 | 2.32M | default: |
576 | 2.32M | assert(tx_size == TX_32X32); |
577 | 5.30M | for (i = 0; i < num_4x4_w; i += 8) |
578 | 2.98M | t_above[i] = !!*(const uint64_t *)&above[i]; |
579 | 5.25M | for (i = 0; i < num_4x4_h; i += 8) |
580 | 2.93M | t_left[i] = !!*(const uint64_t *)&left[i]; |
581 | 2.32M | break; |
582 | 202M | } |
583 | 202M | } |
584 | | #if defined(__GNUC__) && !defined(__clang__) |
585 | | #pragma GCC diagnostic pop |
586 | | #endif |
587 | | |
588 | | void vp9_mv_pred(VP9_COMP *cpi, MACROBLOCK *x, uint8_t *ref_y_buffer, |
589 | 7.08M | int ref_y_stride, int ref_frame, BLOCK_SIZE block_size) { |
590 | 7.08M | int i; |
591 | 7.08M | int zero_seen = 0; |
592 | 7.08M | int best_index = 0; |
593 | 7.08M | int best_sad = INT_MAX; |
594 | 7.08M | int this_sad = INT_MAX; |
595 | 7.08M | int max_mv = 0; |
596 | 7.08M | int near_same_nearest; |
597 | 7.08M | uint8_t *src_y_ptr = x->plane[0].src.buf; |
598 | 7.08M | uint8_t *ref_y_ptr; |
599 | 7.08M | const int num_mv_refs = |
600 | 7.08M | MAX_MV_REF_CANDIDATES + (block_size < x->max_partition_size); |
601 | | |
602 | 7.08M | MV pred_mv[3]; |
603 | 7.08M | pred_mv[0] = x->mbmi_ext->ref_mvs[ref_frame][0].as_mv; |
604 | 7.08M | pred_mv[1] = x->mbmi_ext->ref_mvs[ref_frame][1].as_mv; |
605 | 7.08M | pred_mv[2] = x->pred_mv[ref_frame]; |
606 | 7.08M | assert(num_mv_refs <= (int)(sizeof(pred_mv) / sizeof(pred_mv[0]))); |
607 | | |
608 | 7.08M | near_same_nearest = x->mbmi_ext->ref_mvs[ref_frame][0].as_int == |
609 | 7.08M | x->mbmi_ext->ref_mvs[ref_frame][1].as_int; |
610 | | |
611 | | // Get the sad for each candidate reference mv. |
612 | 28.2M | for (i = 0; i < num_mv_refs; ++i) { |
613 | 21.1M | const MV *this_mv = &pred_mv[i]; |
614 | 21.1M | int fp_row, fp_col; |
615 | 21.1M | if (this_mv->row == INT16_MAX || this_mv->col == INT16_MAX) continue; |
616 | 20.5M | if (i == 1 && near_same_nearest) continue; |
617 | 18.0M | fp_row = (this_mv->row + 3 + (this_mv->row >= 0)) >> 3; |
618 | 18.0M | fp_col = (this_mv->col + 3 + (this_mv->col >= 0)) >> 3; |
619 | 18.0M | max_mv = VPXMAX(max_mv, VPXMAX(abs(this_mv->row), abs(this_mv->col)) >> 3); |
620 | | |
621 | 18.0M | if (fp_row == 0 && fp_col == 0 && zero_seen) continue; |
622 | 17.6M | zero_seen |= (fp_row == 0 && fp_col == 0); |
623 | | |
624 | 17.6M | ref_y_ptr = &ref_y_buffer[ref_y_stride * fp_row + fp_col]; |
625 | | // Find sad for current vector. |
626 | 17.6M | this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride, |
627 | 17.6M | ref_y_ptr, ref_y_stride); |
628 | | // Note if it is the best so far. |
629 | 17.6M | if (this_sad < best_sad) { |
630 | 11.9M | best_sad = this_sad; |
631 | 11.9M | best_index = i; |
632 | 11.9M | } |
633 | 17.6M | } |
634 | | |
635 | | // Note the index of the mv that worked best in the reference list. |
636 | 7.08M | x->mv_best_ref_index[ref_frame] = best_index; |
637 | 7.08M | x->max_mv_context[ref_frame] = max_mv; |
638 | 7.08M | x->pred_mv_sad[ref_frame] = best_sad; |
639 | 7.08M | } |
640 | | |
641 | | void vp9_setup_pred_block(const MACROBLOCKD *xd, |
642 | | struct buf_2d dst[MAX_MB_PLANE], |
643 | | const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col, |
644 | | const struct scale_factors *scale, |
645 | 13.3M | const struct scale_factors *scale_uv) { |
646 | 13.3M | int i; |
647 | | |
648 | 13.3M | dst[0].buf = src->y_buffer; |
649 | 13.3M | dst[0].stride = src->y_stride; |
650 | 13.3M | dst[1].buf = src->u_buffer; |
651 | 13.3M | dst[2].buf = src->v_buffer; |
652 | 13.3M | dst[1].stride = dst[2].stride = src->uv_stride; |
653 | | |
654 | 53.4M | for (i = 0; i < MAX_MB_PLANE; ++i) { |
655 | 40.0M | setup_pred_plane(dst + i, dst[i].buf, dst[i].stride, mi_row, mi_col, |
656 | 40.0M | i ? scale_uv : scale, xd->plane[i].subsampling_x, |
657 | 40.0M | xd->plane[i].subsampling_y); |
658 | 40.0M | } |
659 | 13.3M | } |
660 | | |
661 | | int vp9_raster_block_offset(BLOCK_SIZE plane_bsize, int raster_block, |
662 | 414M | int stride) { |
663 | 414M | const int bw = b_width_log2_lookup[plane_bsize]; |
664 | 414M | const int y = 4 * (raster_block >> bw); |
665 | 414M | const int x = 4 * (raster_block & ((1 << bw) - 1)); |
666 | 414M | return y * stride + x; |
667 | 414M | } |
668 | | |
669 | | int16_t *vp9_raster_block_offset_int16(BLOCK_SIZE plane_bsize, int raster_block, |
670 | 269M | int16_t *base) { |
671 | 269M | const int stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize]; |
672 | 269M | return base + vp9_raster_block_offset(plane_bsize, raster_block, stride); |
673 | 269M | } |
674 | | |
675 | | YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const VP9_COMP *cpi, |
676 | 13.7M | int ref_frame) { |
677 | 13.7M | const VP9_COMMON *const cm = &cpi->common; |
678 | 13.7M | const int scaled_idx = cpi->scaled_ref_idx[ref_frame - 1]; |
679 | 13.7M | const int ref_idx = get_ref_frame_buf_idx(cpi, ref_frame); |
680 | 13.7M | assert(ref_frame >= LAST_FRAME && ref_frame <= ALTREF_FRAME); |
681 | 13.7M | return (scaled_idx != ref_idx && scaled_idx != INVALID_IDX) |
682 | 13.7M | ? &cm->buffer_pool->frame_bufs[scaled_idx].buf |
683 | 13.7M | : NULL; |
684 | 13.7M | } |
685 | | |
686 | 44.7M | int vp9_get_switchable_rate(const VP9_COMP *cpi, const MACROBLOCKD *const xd) { |
687 | 44.7M | const MODE_INFO *const mi = xd->mi[0]; |
688 | 44.7M | const int ctx = get_pred_context_switchable_interp(xd); |
689 | 44.7M | return SWITCHABLE_INTERP_RATE_FACTOR * |
690 | 44.7M | cpi->switchable_interp_costs[ctx][mi->interp_filter]; |
691 | 44.7M | } |
692 | | |
693 | 42.2k | void vp9_set_rd_speed_thresholds(VP9_COMP *cpi) { |
694 | 42.2k | int i; |
695 | 42.2k | RD_OPT *const rd = &cpi->rd; |
696 | 42.2k | SPEED_FEATURES *const sf = &cpi->sf; |
697 | | |
698 | | // Set baseline threshold values. |
699 | 1.31M | for (i = 0; i < MAX_MODES; ++i) |
700 | 1.26M | rd->thresh_mult[i] = cpi->oxcf.mode == BEST ? -500 : 0; |
701 | | |
702 | 42.2k | if (sf->adaptive_rd_thresh) { |
703 | 42.2k | rd->thresh_mult[THR_NEARESTMV] = 300; |
704 | 42.2k | rd->thresh_mult[THR_NEARESTG] = 300; |
705 | 42.2k | rd->thresh_mult[THR_NEARESTA] = 300; |
706 | 42.2k | } else { |
707 | 0 | rd->thresh_mult[THR_NEARESTMV] = 0; |
708 | 0 | rd->thresh_mult[THR_NEARESTG] = 0; |
709 | 0 | rd->thresh_mult[THR_NEARESTA] = 0; |
710 | 0 | } |
711 | | |
712 | 42.2k | rd->thresh_mult[THR_DC] += 1000; |
713 | | |
714 | 42.2k | rd->thresh_mult[THR_NEWMV] += 1000; |
715 | 42.2k | rd->thresh_mult[THR_NEWA] += 1000; |
716 | 42.2k | rd->thresh_mult[THR_NEWG] += 1000; |
717 | | |
718 | 42.2k | rd->thresh_mult[THR_NEARMV] += 1000; |
719 | 42.2k | rd->thresh_mult[THR_NEARA] += 1000; |
720 | 42.2k | rd->thresh_mult[THR_COMP_NEARESTLA] += 1000; |
721 | 42.2k | rd->thresh_mult[THR_COMP_NEARESTGA] += 1000; |
722 | | |
723 | 42.2k | rd->thresh_mult[THR_TM] += 1000; |
724 | | |
725 | 42.2k | rd->thresh_mult[THR_COMP_NEARLA] += 1500; |
726 | 42.2k | rd->thresh_mult[THR_COMP_NEWLA] += 2000; |
727 | 42.2k | rd->thresh_mult[THR_NEARG] += 1000; |
728 | 42.2k | rd->thresh_mult[THR_COMP_NEARGA] += 1500; |
729 | 42.2k | rd->thresh_mult[THR_COMP_NEWGA] += 2000; |
730 | | |
731 | 42.2k | rd->thresh_mult[THR_ZEROMV] += 2000; |
732 | 42.2k | rd->thresh_mult[THR_ZEROG] += 2000; |
733 | 42.2k | rd->thresh_mult[THR_ZEROA] += 2000; |
734 | 42.2k | rd->thresh_mult[THR_COMP_ZEROLA] += 2500; |
735 | 42.2k | rd->thresh_mult[THR_COMP_ZEROGA] += 2500; |
736 | | |
737 | 42.2k | rd->thresh_mult[THR_H_PRED] += 2000; |
738 | 42.2k | rd->thresh_mult[THR_V_PRED] += 2000; |
739 | 42.2k | rd->thresh_mult[THR_D45_PRED] += 2500; |
740 | 42.2k | rd->thresh_mult[THR_D135_PRED] += 2500; |
741 | 42.2k | rd->thresh_mult[THR_D117_PRED] += 2500; |
742 | 42.2k | rd->thresh_mult[THR_D153_PRED] += 2500; |
743 | 42.2k | rd->thresh_mult[THR_D207_PRED] += 2500; |
744 | 42.2k | rd->thresh_mult[THR_D63_PRED] += 2500; |
745 | 42.2k | } |
746 | | |
747 | 42.2k | void vp9_set_rd_speed_thresholds_sub8x8(VP9_COMP *cpi) { |
748 | 42.2k | static const int thresh_mult[2][MAX_REFS] = { |
749 | 42.2k | { 2500, 2500, 2500, 4500, 4500, 2500 }, |
750 | 42.2k | { 2000, 2000, 2000, 4000, 4000, 2000 } |
751 | 42.2k | }; |
752 | 42.2k | RD_OPT *const rd = &cpi->rd; |
753 | 42.2k | const int idx = cpi->oxcf.mode == BEST; |
754 | 42.2k | memcpy(rd->thresh_mult_sub8x8, thresh_mult[idx], sizeof(thresh_mult[idx])); |
755 | 42.2k | } |
756 | | |
757 | | void vp9_update_rd_thresh_fact(int (*factor_buf)[MAX_MODES], int rd_thresh, |
758 | 3.75M | int bsize, int best_mode_index) { |
759 | 3.75M | if (rd_thresh > 0) { |
760 | 3.75M | const int top_mode = bsize < BLOCK_8X8 ? MAX_REFS : MAX_MODES; |
761 | 3.75M | int mode; |
762 | 87.7M | for (mode = 0; mode < top_mode; ++mode) { |
763 | 84.0M | const BLOCK_SIZE min_size = VPXMAX(bsize - 1, BLOCK_4X4); |
764 | 84.0M | const BLOCK_SIZE max_size = VPXMIN(bsize + 2, BLOCK_64X64); |
765 | 84.0M | BLOCK_SIZE bs; |
766 | 413M | for (bs = min_size; bs <= max_size; ++bs) { |
767 | 329M | int *const fact = &factor_buf[bs][mode]; |
768 | 329M | if (mode == best_mode_index) { |
769 | 14.0M | *fact -= (*fact >> 4); |
770 | 315M | } else { |
771 | 315M | *fact = VPXMIN(*fact + RD_THRESH_INC, rd_thresh * RD_THRESH_MAX_FACT); |
772 | 315M | } |
773 | 329M | } |
774 | 84.0M | } |
775 | 3.75M | } |
776 | 3.75M | } |
777 | | |
778 | | int vp9_get_intra_cost_penalty(const VP9_COMP *const cpi, BLOCK_SIZE bsize, |
779 | 5.99M | int qindex, int qdelta) { |
780 | | // Reduce the intra cost penalty for small blocks (<=16x16). |
781 | 5.99M | int reduction_fac = |
782 | 5.99M | (bsize <= BLOCK_16X16) ? ((bsize <= BLOCK_8X8) ? 4 : 2) : 0; |
783 | | |
784 | 5.99M | if (cpi->noise_estimate.enabled && cpi->noise_estimate.level == kHigh) |
785 | | // Don't reduce intra cost penalty if estimated noise level is high. |
786 | 0 | reduction_fac = 0; |
787 | | |
788 | | // Always use VPX_BITS_8 as input here because the penalty is applied |
789 | | // to rate not distortion so we want a consistent penalty for all bit |
790 | | // depths. If the actual bit depth were passed in here then the value |
791 | | // retured by vp9_dc_quant() would scale with the bit depth and we would |
792 | | // then need to apply inverse scaling to correct back to a bit depth |
793 | | // independent rate penalty. |
794 | 5.99M | return (20 * vp9_dc_quant(qindex, qdelta, VPX_BITS_8)) >> reduction_fac; |
795 | 5.99M | } |