/work/svt-av1/Source/Lib/Codec/mcomp.c
Line | Count | Source |
1 | | /* |
2 | | * Copyright (c) 2016, Alliance for Open Media. All rights reserved |
3 | | * |
4 | | * This source code is subject to the terms of the BSD 2 Clause License and |
5 | | * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License |
6 | | * was not distributed with this source code in the LICENSE file, you can |
7 | | * obtain it at www.aomedia.org/license/software. If the Alliance for Open |
8 | | * Media Patent License 1.0 was not distributed with this source code in the |
9 | | * PATENTS file, you can obtain it at www.aomedia.org/license/patent. |
10 | | */ |
11 | | |
12 | | #include <limits.h> |
13 | | #include <math.h> |
14 | | #include <stdio.h> |
15 | | #include "mcomp.h" |
16 | | #include "mv.h" |
17 | | #include "av1_common.h" |
18 | | #include "coding_unit.h" |
19 | | #include "block_structures.h" |
20 | | #include "av1me.h" |
21 | | #include "aom_dsp_rtcd.h" |
22 | | #include "rd_cost.h" |
23 | | |
24 | | // ============================================================================= |
25 | | // Subpixel Motion Search: Translational |
26 | | // ============================================================================= |
27 | 0 | #define INIT_SUBPEL_STEP_SIZE (4) |
28 | | |
29 | | /* |
30 | | * To avoid the penalty for crossing cache-line read, preload the reference |
31 | | * area in a small buffer, which is aligned to make sure there won't be crossing |
32 | | * cache-line read while reading from this buffer. This reduced the cpu |
33 | | * cycles spent on reading ref data in sub-pixel filter functions. |
34 | | * TODO: Currently, since sub-pixel search range here is -3 ~ 3, copy 22 rows x |
35 | | * 32 cols area that is enough for 16x16 macroblock. Later, for SPLITMV, we |
36 | | * could reduce the area. |
37 | | */ |
38 | | |
39 | | // Returns the subpel offset used by various subpel variance functions [m]sv[a]f |
40 | 0 | static INLINE int svt_get_subpel_part(int x) { |
41 | 0 | return x & 7; |
42 | 0 | } |
43 | | |
44 | | // Gets the address of the ref buffer at subpel location (r, c), rounded to the |
45 | | // nearest fullpel precision toward - \infty |
46 | | |
47 | 0 | static INLINE const uint8_t* svt_get_buf_from_mv(const struct svt_buf_2d* buf, const Mv mv) { |
48 | 0 | const int offset = (mv.y >> 3) * buf->stride + (mv.x >> 3); |
49 | 0 | return &buf->buf[offset]; |
50 | 0 | } |
51 | | |
52 | | // Calculates the variance of prediction residue. |
53 | | static int svt_upsampled_pref_error(MacroBlockD* xd, const struct AV1Common* const cm, Mv this_mv, |
54 | 0 | const SUBPEL_SEARCH_VAR_PARAMS* var_params) { |
55 | 0 | const AomVarianceFnPtr* vfp = var_params->vfp; |
56 | 0 | const SUBPEL_SEARCH_TYPE subpel_search_type = var_params->subpel_search_type; |
57 | |
|
58 | 0 | const MSBuffers* ms_buffers = &var_params->ms_buffers; |
59 | 0 | const uint8_t* src = ms_buffers->src->buf; |
60 | 0 | const uint8_t* ref = svt_get_buf_from_mv(ms_buffers->ref, this_mv); |
61 | 0 | const int src_stride = ms_buffers->src->stride; |
62 | 0 | const int ref_stride = ms_buffers->ref->stride; |
63 | 0 | const int w = var_params->w; |
64 | 0 | const int h = var_params->h; |
65 | 0 | const int mi_row = xd->mi_row; |
66 | 0 | const int mi_col = xd->mi_col; |
67 | 0 | const int subpel_x_q3 = svt_get_subpel_part(this_mv.x); |
68 | 0 | const int subpel_y_q3 = svt_get_subpel_part(this_mv.y); |
69 | |
|
70 | 0 | unsigned int besterr; |
71 | 0 | { |
72 | 0 | DECLARE_ALIGNED(16, uint8_t, pred[MAX_SB_SQUARE]); |
73 | |
|
74 | 0 | { |
75 | 0 | svt_aom_upsampled_pred( |
76 | 0 | xd, cm, mi_row, mi_col, pred, w, h, subpel_x_q3, subpel_y_q3, ref, ref_stride, subpel_search_type); |
77 | 0 | } |
78 | 0 | unsigned int sse; |
79 | 0 | besterr = vfp->vf(pred, w, src, src_stride, &sse); |
80 | 0 | } |
81 | |
|
82 | 0 | return besterr; |
83 | 0 | } |
84 | | |
85 | | // Estimates the variance of prediction residue using bilinear filter for fast |
86 | | // search. |
87 | 0 | static INLINE int svt_estimated_pref_error(Mv this_mv, const SUBPEL_SEARCH_VAR_PARAMS* var_params) { |
88 | 0 | const AomVarianceFnPtr* vfp = var_params->vfp; |
89 | |
|
90 | 0 | const MSBuffers* ms_buffers = &var_params->ms_buffers; |
91 | 0 | const uint8_t* src = ms_buffers->src->buf; |
92 | 0 | const uint8_t* ref = svt_get_buf_from_mv(ms_buffers->ref, this_mv); |
93 | 0 | const int src_stride = ms_buffers->src->stride; |
94 | 0 | const int ref_stride = ms_buffers->ref->stride; |
95 | |
|
96 | 0 | const int subpel_x_q3 = svt_get_subpel_part(this_mv.x); |
97 | 0 | const int subpel_y_q3 = svt_get_subpel_part(this_mv.y); |
98 | |
|
99 | 0 | unsigned int sse; |
100 | 0 | return vfp->svf(ref, ref_stride, subpel_x_q3, subpel_y_q3, src, src_stride, &sse); |
101 | 0 | } |
102 | | |
103 | | // Estimates whether this_mv is better than best_mv. This function incorporates |
104 | | // both prediction error and residue into account. It is suffixed "fast" because |
105 | | // it uses bilinear filter to estimate the prediction. |
106 | | static INLINE unsigned int svt_check_better_fast(Mv this_mv, Mv* best_mv, const SubpelMvLimits* mv_limits, |
107 | | const SUBPEL_SEARCH_VAR_PARAMS* var_params, |
108 | | const svt_mv_cost_param* mv_cost_params, unsigned int* besterr, |
109 | 0 | int* distortion) { |
110 | 0 | if (svt_av1_is_subpelmv_in_range(mv_limits, this_mv)) { |
111 | 0 | unsigned int cost; |
112 | 0 | int thismse; |
113 | 0 | cost = svt_aom_fp_mv_err_cost(this_mv, mv_cost_params); |
114 | 0 | if (mv_cost_params->mv_cost_type == MV_COST_OPT) { |
115 | 0 | int64_t bestcost = *distortion + cost; |
116 | 0 | if (bestcost > (((int64_t)*besterr * (int64_t)mv_cost_params->early_exit_th) / 1000)) { |
117 | 0 | return (uint32_t)bestcost; |
118 | 0 | } |
119 | 0 | } |
120 | 0 | thismse = svt_estimated_pref_error(this_mv, var_params); |
121 | 0 | cost += thismse; |
122 | 0 | int weight = 100; |
123 | 0 | if (var_params->bias_fp && best_mv->x % 8 == 0 && best_mv->y % 8 == 0) { |
124 | 0 | weight = var_params->bias_fp; |
125 | 0 | } |
126 | 0 | if ((((uint64_t)cost * weight) / 100) < *besterr) { |
127 | 0 | *besterr = cost; |
128 | 0 | *best_mv = this_mv; |
129 | 0 | *distortion = thismse; |
130 | 0 | } |
131 | 0 | return cost; |
132 | 0 | } |
133 | 0 | return INT_MAX; |
134 | 0 | } |
135 | | |
136 | | // Checks whether this_mv is better than best_mv. This function incorporates |
137 | | // both prediction error and residue into account. |
138 | | static AOM_FORCE_INLINE unsigned int svt_check_better(MacroBlockD* xd, const struct AV1Common* const cm, Mv this_mv, |
139 | | Mv* best_mv, const SubpelMvLimits* mv_limits, |
140 | | const SUBPEL_SEARCH_VAR_PARAMS* var_params, |
141 | 0 | const svt_mv_cost_param* mv_cost_params, unsigned int* besterr) { |
142 | 0 | unsigned int cost; |
143 | 0 | if (svt_av1_is_subpelmv_in_range(mv_limits, this_mv)) { |
144 | 0 | int thismse; |
145 | 0 | thismse = svt_upsampled_pref_error(xd, cm, this_mv, var_params); |
146 | 0 | cost = svt_aom_fp_mv_err_cost(this_mv, mv_cost_params); |
147 | 0 | cost += thismse; |
148 | 0 | int weight = 100; |
149 | 0 | if (var_params->bias_fp && best_mv->x % 8 == 0 && best_mv->y % 8 == 0) { |
150 | 0 | weight = var_params->bias_fp; |
151 | 0 | } |
152 | 0 | if ((((uint64_t)cost * weight) / 100) < *besterr) { |
153 | 0 | *besterr = cost; |
154 | 0 | *best_mv = this_mv; |
155 | 0 | } |
156 | 0 | } else { |
157 | 0 | cost = INT_MAX; |
158 | 0 | } |
159 | 0 | return cost; |
160 | 0 | } |
161 | | |
162 | | static INLINE Mv get_best_diag_step(int step_size, unsigned int left_cost, unsigned int right_cost, |
163 | | unsigned int up_cost, unsigned int down_cost) { |
164 | | const Mv diag_step = { |
165 | | {left_cost <= right_cost ? -step_size : step_size, up_cost <= down_cost ? -step_size : step_size}}; |
166 | | |
167 | | return diag_step; |
168 | | } |
169 | | |
170 | | static AOM_FORCE_INLINE Mv svt_first_level_check(MacroBlockD* xd, const struct AV1Common* const cm, const Mv this_mv, |
171 | | Mv* best_mv, const int hstep, const SubpelMvLimits* mv_limits, |
172 | | const SUBPEL_SEARCH_VAR_PARAMS* var_params, |
173 | 0 | const svt_mv_cost_param* mv_cost_params, unsigned int* besterr) { |
174 | 0 | const Mv left_mv = {{this_mv.x - hstep, this_mv.y}}; |
175 | 0 | const Mv right_mv = {{this_mv.x + hstep, this_mv.y}}; |
176 | 0 | const Mv top_mv = {{this_mv.x, this_mv.y - hstep}}; |
177 | 0 | const Mv bottom_mv = {{this_mv.x, this_mv.y + hstep}}; |
178 | |
|
179 | 0 | const unsigned int left = svt_check_better( |
180 | 0 | xd, cm, left_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr); |
181 | 0 | const unsigned int right = svt_check_better( |
182 | 0 | xd, cm, right_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr); |
183 | 0 | const unsigned int up = svt_check_better(xd, cm, top_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr); |
184 | 0 | const unsigned int down = svt_check_better( |
185 | 0 | xd, cm, bottom_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr); |
186 | |
|
187 | 0 | const Mv diag_step = get_best_diag_step(hstep, left, right, up, down); |
188 | 0 | const Mv diag_mv = {{this_mv.x + diag_step.x, this_mv.y + diag_step.y}}; |
189 | | |
190 | | // Check the diagonal direction with the best mv |
191 | 0 | svt_check_better(xd, cm, diag_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr); |
192 | |
|
193 | 0 | return diag_step; |
194 | 0 | } |
195 | | |
196 | | // A newer version of second level check that gives better quality. |
197 | | // TODO(chiyotsai@google.com): evaluate this on subpel_search_types different |
198 | | // from av1_find_best_sub_pixel_tree |
199 | | static AOM_FORCE_INLINE void svt_second_level_check_v2(MacroBlockD* xd, const struct AV1Common* const cm, |
200 | | const Mv this_mv, Mv diag_step, Mv* best_mv, |
201 | | const SubpelMvLimits* mv_limits, |
202 | | const SUBPEL_SEARCH_VAR_PARAMS* var_params, |
203 | 0 | const svt_mv_cost_param* mv_cost_params, unsigned int* besterr) { |
204 | 0 | assert(best_mv->y == this_mv.y + diag_step.y || best_mv->x == this_mv.x + diag_step.x); |
205 | 0 | if (CHECK_MV_EQUAL(this_mv, *best_mv)) { |
206 | 0 | return; |
207 | 0 | } else if (this_mv.y == best_mv->y) { |
208 | | // Search away from diagonal step since diagonal search did not provide any |
209 | | // improvement |
210 | 0 | diag_step.y *= -1; |
211 | 0 | } else if (this_mv.x == best_mv->x) { |
212 | 0 | diag_step.x *= -1; |
213 | 0 | } |
214 | | |
215 | 0 | const Mv row_bias_mv = {{best_mv->x, best_mv->y + diag_step.y}}; |
216 | 0 | const Mv col_bias_mv = {{best_mv->x + diag_step.x, best_mv->y}}; |
217 | 0 | const Mv diag_bias_mv = {{best_mv->x + diag_step.x, best_mv->y + diag_step.y}}; |
218 | | |
219 | | // svt_check_better only lowers *besterr on an improvement, so a drop across |
220 | | // the two probes is equivalent to "a better mv was found". |
221 | 0 | const unsigned int besterr_before = *besterr; |
222 | 0 | svt_check_better(xd, cm, row_bias_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr); |
223 | 0 | svt_check_better(xd, cm, col_bias_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr); |
224 | | |
225 | | // Do an additional search if the second iteration gives a better mv |
226 | 0 | if (*besterr < besterr_before) { |
227 | 0 | svt_check_better(xd, cm, diag_bias_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr); |
228 | 0 | } |
229 | 0 | } |
230 | | |
231 | | // Gets the error at the beginning when the mv has fullpel precision |
232 | | static unsigned int svt_upsampled_setup_center_error(const Mv bestmv, const SUBPEL_SEARCH_VAR_PARAMS* var_params, |
233 | | const svt_mv_cost_param* mv_cost_params, |
234 | 0 | unsigned int* distortion) { |
235 | 0 | const MSBuffers* ms_buffers = &var_params->ms_buffers; |
236 | 0 | const uint8_t* ref = svt_get_buf_from_mv(ms_buffers->ref, bestmv); |
237 | 0 | *distortion = var_params->vfp->vf( |
238 | 0 | ref, ms_buffers->ref->stride, ms_buffers->src->buf, ms_buffers->src->stride, distortion); |
239 | 0 | return *distortion + svt_aom_fp_mv_err_cost(bestmv, mv_cost_params); |
240 | 0 | } |
241 | | |
242 | | // Searches the four cardinal direction for a better mv, then follows up with a |
243 | | // search in the best quadrant. This uses bilinear filter to speed up the |
244 | | // calculation. |
245 | | static AOM_FORCE_INLINE Mv first_level_check_fast(const Mv this_mv, Mv* best_mv, int hstep, |
246 | | const SubpelMvLimits* mv_limits, |
247 | | const SUBPEL_SEARCH_VAR_PARAMS* var_params, |
248 | | const svt_mv_cost_param* mv_cost_params, unsigned int* besterr, |
249 | | unsigned int orgerr, int* distortion) { |
250 | | // Check the four cardinal directions |
251 | | const Mv left_mv = {{this_mv.x - hstep, this_mv.y}}; |
252 | | const unsigned int left = svt_check_better_fast( |
253 | | left_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
254 | | |
255 | | const Mv right_mv = {{this_mv.x + hstep, this_mv.y}}; |
256 | | const unsigned int right = svt_check_better_fast( |
257 | | right_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
258 | | |
259 | | const Mv top_mv = {{this_mv.x, this_mv.y - hstep}}; |
260 | | const unsigned int up = svt_check_better_fast( |
261 | | top_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
262 | | |
263 | | const Mv bottom_mv = {{this_mv.x, this_mv.y + hstep}}; |
264 | | const unsigned int down = svt_check_better_fast( |
265 | | bottom_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
266 | | |
267 | | const Mv diag_step = get_best_diag_step(hstep, left, right, up, down); |
268 | | const Mv diag_mv = {{this_mv.x + diag_step.x, this_mv.y + diag_step.y}}; |
269 | | if (*besterr >= orgerr) { |
270 | | return diag_step; |
271 | | } |
272 | | // Check the diagonal direction with the best mv |
273 | | svt_check_better_fast(diag_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
274 | | |
275 | | return diag_step; |
276 | | } |
277 | | |
278 | | // Performs a following up search after first_level_check_fast is called. This |
279 | | // performs two extra chess pattern searches in the best quadrant. |
280 | | static AOM_FORCE_INLINE void second_level_check_fast(const Mv this_mv, const Mv diag_step, Mv* best_mv, int hstep, |
281 | | const SubpelMvLimits* mv_limits, |
282 | | const SUBPEL_SEARCH_VAR_PARAMS* var_params, |
283 | | const svt_mv_cost_param* mv_cost_params, unsigned int* besterr, |
284 | | int* distortion) { |
285 | | assert(diag_step.y == hstep || diag_step.y == -hstep); |
286 | | assert(diag_step.x == hstep || diag_step.x == -hstep); |
287 | | const int tr = this_mv.y; |
288 | | const int tc = this_mv.x; |
289 | | const int br = best_mv->y; |
290 | | const int bc = best_mv->x; |
291 | | if (tr != br && tc != bc) { |
292 | | assert(diag_step.x == bc - tc); |
293 | | assert(diag_step.y == br - tr); |
294 | | const Mv chess_mv_1 = {{bc + diag_step.x, br}}; |
295 | | const Mv chess_mv_2 = {{bc, br + diag_step.y}}; |
296 | | svt_check_better_fast(chess_mv_1, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
297 | | svt_check_better_fast(chess_mv_2, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
298 | | } else if (tr == br && tc != bc) { |
299 | | assert(diag_step.x == bc - tc); |
300 | | // Continue searching in the best direction |
301 | | const Mv bottom_long_mv = {{bc + diag_step.x, br + hstep}}; |
302 | | const Mv top_long_mv = {{bc + diag_step.x, br - hstep}}; |
303 | | svt_check_better_fast(bottom_long_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
304 | | svt_check_better_fast(top_long_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
305 | | |
306 | | // Search in the direction opposite of the best quadrant |
307 | | const Mv rev_mv = {{bc, br - diag_step.y}}; |
308 | | svt_check_better_fast(rev_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
309 | | } else if (tr != br && tc == bc) { |
310 | | assert(diag_step.y == br - tr); |
311 | | // Continue searching in the best direction |
312 | | const Mv right_long_mv = {{bc + hstep, br + diag_step.y}}; |
313 | | const Mv left_long_mv = {{bc - hstep, br + diag_step.y}}; |
314 | | svt_check_better_fast(right_long_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
315 | | svt_check_better_fast(left_long_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
316 | | |
317 | | // Search in the direction opposite of the best quadrant |
318 | | const Mv rev_mv = {{bc - diag_step.x, br}}; |
319 | | svt_check_better_fast(rev_mv, best_mv, mv_limits, var_params, mv_cost_params, besterr, distortion); |
320 | | } |
321 | | } |
322 | | |
323 | | // Combines first level check and second level check when applicable. This first |
324 | | // searches the four cardinal directions, and perform several |
325 | | // diagonal/chess-pattern searches in the best quadrant. |
326 | | static AOM_FORCE_INLINE void two_level_checks_fast(const Mv this_mv, Mv* best_mv, int hstep, |
327 | | const SubpelMvLimits* mv_limits, |
328 | | const SUBPEL_SEARCH_VAR_PARAMS* var_params, |
329 | | const svt_mv_cost_param* mv_cost_params, unsigned int* besterr, |
330 | | unsigned int orgerr, int* distortion, int iters) { |
331 | | const Mv diag_step = first_level_check_fast( |
332 | | this_mv, best_mv, hstep, mv_limits, var_params, mv_cost_params, besterr, orgerr, distortion); |
333 | | if (*besterr < orgerr && iters > 1) { |
334 | | second_level_check_fast( |
335 | | this_mv, diag_step, best_mv, hstep, mv_limits, var_params, mv_cost_params, besterr, distortion); |
336 | | } |
337 | | } |
338 | | |
339 | | extern const uint8_t svt_aom_eb_av1_var_offs[MAX_SB_SIZE]; |
340 | | |
341 | | int svt_av1_find_best_sub_pixel_tree_pruned(void* ictx, MacroBlockD* xd, const struct AV1Common* const cm, |
342 | | SUBPEL_MOTION_SEARCH_PARAMS* ms_params, Mv start_mv, Mv* bestmv, |
343 | 0 | BlockSize bsize) { |
344 | 0 | const int allow_hp = ms_params->allow_hp; |
345 | 0 | const int forced_stop = ms_params->forced_stop; |
346 | 0 | const int iters_per_step = ms_params->iters_per_step; |
347 | 0 | const SubpelMvLimits* mv_limits = &ms_params->mv_limits; |
348 | 0 | const svt_mv_cost_param* mv_cost_params = &ms_params->mv_cost_params; |
349 | 0 | const SUBPEL_SEARCH_VAR_PARAMS* var_params = &ms_params->var_params; |
350 | 0 | int hstep = INIT_SUBPEL_STEP_SIZE; // Step size, initialized to 4/8=1/2 pel |
351 | 0 | unsigned int besterr; |
352 | 0 | unsigned int org_error; |
353 | 0 | int distortion_storage = 0; |
354 | 0 | int* distortion = &distortion_storage; |
355 | 0 | *bestmv = start_mv; |
356 | |
|
357 | 0 | (void)xd; |
358 | 0 | (void)cm; |
359 | 0 | besterr = svt_upsampled_setup_center_error(*bestmv, var_params, mv_cost_params, (unsigned int*)distortion); |
360 | |
|
361 | 0 | if (ictx != NULL && ms_params->search_stage == SPEL_ME) { |
362 | 0 | ModeDecisionContext* ctx = (ModeDecisionContext*)ictx; |
363 | 0 | ctx->fp_me_dist[ms_params->list_idx][ms_params->ref_idx] = besterr; |
364 | 0 | } |
365 | |
|
366 | 0 | const uint32_t th_normalizer = var_params->w * var_params->h * ms_params->abs_th_mult; |
367 | 0 | if (besterr < th_normalizer) { |
368 | 0 | return besterr; |
369 | 0 | } |
370 | | // How many steps to take. A round of 0 means fullpel search only, 1 means |
371 | | // half-pel, and so on. |
372 | 0 | const int round = AOMMIN(FULL_PEL - forced_stop, 3 - !allow_hp); |
373 | | |
374 | | // If forced_stop is FULL_PEL, return. |
375 | 0 | if (!round) { |
376 | 0 | return besterr; |
377 | 0 | } |
378 | | // Exit subpel search if the variance of the full-pel predicted samples is low (i.e. where likely interpolation will not modify the integer samples) |
379 | 0 | if (ms_params->pred_variance_th) { |
380 | 0 | const MSBuffers* ms_buffers = &var_params->ms_buffers; |
381 | 0 | const uint8_t* ref = svt_get_buf_from_mv(ms_buffers->ref, *bestmv); |
382 | 0 | unsigned int sse; |
383 | 0 | const unsigned int var = var_params->vfp->vf(ref, ms_buffers->ref->stride, svt_aom_eb_av1_var_offs, 0, &sse); |
384 | 0 | int block_var = ROUND_POWER_OF_TWO(var, eb_num_pels_log2_lookup[bsize]); |
385 | |
|
386 | 0 | if (block_var < ms_params->pred_variance_th) { |
387 | 0 | return besterr; |
388 | 0 | } |
389 | 0 | } |
390 | 0 | if (ms_params->skip_diag_refinement >= 4) { |
391 | 0 | org_error = 0; |
392 | 0 | } else { |
393 | 0 | unsigned int demo = ms_params->skip_diag_refinement >= 2 |
394 | 0 | ? ((var_params->w >= 64 || var_params->h >= 64) ? 2 : 1) |
395 | 0 | : 1; |
396 | 0 | org_error = ms_params->skip_diag_refinement ? besterr / demo : INT_MAX; |
397 | 0 | } |
398 | 0 | for (int iter = 0; iter < round; ++iter) { |
399 | 0 | unsigned int prev_besterr = besterr; |
400 | 0 | two_level_checks_fast(start_mv, |
401 | 0 | bestmv, |
402 | 0 | hstep, |
403 | 0 | mv_limits, |
404 | 0 | var_params, |
405 | 0 | mv_cost_params, |
406 | 0 | &besterr, |
407 | 0 | org_error, |
408 | 0 | distortion, |
409 | 0 | iters_per_step); |
410 | 0 | hstep >>= 1; |
411 | 0 | start_mv = *bestmv; |
412 | 0 | if (ms_params->skip_diag_refinement && iter < QUARTER_PEL) { |
413 | 0 | org_error = MIN(org_error, besterr); |
414 | 0 | } |
415 | 0 | int32_t deviation = (((int64_t)MAX(besterr, 1) - (int64_t)MAX(prev_besterr, 1)) * 100) / |
416 | 0 | (int64_t)MAX(prev_besterr, 1); |
417 | 0 | if (deviation >= ms_params->round_dev_th) { |
418 | 0 | return besterr; |
419 | 0 | } |
420 | 0 | } |
421 | 0 | return besterr; |
422 | 0 | } |
423 | | |
424 | | int svt_av1_find_best_sub_pixel_tree(void* ictx, MacroBlockD* xd, const struct AV1Common* const cm, |
425 | 0 | SUBPEL_MOTION_SEARCH_PARAMS* ms_params, Mv start_mv, Mv* bestmv, BlockSize bsize) { |
426 | 0 | ModeDecisionContext* ctx = (ModeDecisionContext*)ictx; |
427 | 0 | const int allow_hp = ms_params->allow_hp; |
428 | 0 | const int forced_stop = ms_params->forced_stop; |
429 | 0 | const int iters_per_step = ms_params->iters_per_step; |
430 | |
|
431 | 0 | svt_mv_cost_param* mv_cost_params = &ms_params->mv_cost_params; |
432 | 0 | const SUBPEL_SEARCH_VAR_PARAMS* var_params = &ms_params->var_params; |
433 | 0 | const SubpelMvLimits* mv_limits = &ms_params->mv_limits; |
434 | | |
435 | | // How many steps to take. A round of 0 means fullpel search only, 1 means |
436 | | // half-pel, and so on. |
437 | 0 | int round = AOMMIN(FULL_PEL - forced_stop, 3 - !allow_hp); |
438 | 0 | int hstep = INIT_SUBPEL_STEP_SIZE; // Step size, initialized to 4/8=1/2 pel |
439 | |
|
440 | 0 | unsigned int besterr; |
441 | 0 | int distortion_storage = 0; |
442 | 0 | int* distortion = &distortion_storage; |
443 | |
|
444 | 0 | *bestmv = start_mv; |
445 | 0 | besterr = svt_upsampled_setup_center_error(*bestmv, var_params, mv_cost_params, (unsigned int*)distortion); |
446 | 0 | if (ctx != NULL && ms_params->search_stage == SPEL_ME) { |
447 | 0 | ctx->fp_me_dist[ms_params->list_idx][ms_params->ref_idx] = besterr; |
448 | 0 | if (ctx->pd_pass == PD_PASS_1 && ctx->md_subpel_me_ctrls.mvp_th) { |
449 | 0 | unsigned int best_mvperr = ctx->best_fp_mvp_dist[ms_params->list_idx][ms_params->ref_idx]; |
450 | 0 | int best_mvp_idx = ctx->best_fp_mvp_idx[ms_params->list_idx][ms_params->ref_idx]; |
451 | 0 | const int mvp_err = best_mvperr + 1; |
452 | 0 | const int me_err = besterr + 1; |
453 | 0 | const int32_t deviation = ((me_err - mvp_err) * 100) / me_err; |
454 | 0 | if (deviation >= ctx->md_subpel_me_ctrls.mvp_th) { |
455 | 0 | round = 1; |
456 | 0 | } else if (ABS(bestmv->x - ctx->mvp_array[ms_params->list_idx][ms_params->ref_idx][best_mvp_idx].x) > |
457 | 0 | ctx->md_subpel_me_ctrls.hp_mv_th || |
458 | 0 | ABS(bestmv->y - ctx->mvp_array[ms_params->list_idx][ms_params->ref_idx][best_mvp_idx].y) > |
459 | 0 | ctx->md_subpel_me_ctrls.hp_mv_th) { |
460 | 0 | round = MIN(round, 2); |
461 | 0 | } |
462 | 0 | } |
463 | 0 | } |
464 | 0 | const uint32_t th_normalizer = var_params->w * var_params->h * ms_params->abs_th_mult; |
465 | 0 | if (besterr < th_normalizer) { |
466 | 0 | return besterr; |
467 | 0 | } |
468 | | |
469 | | // If forced_stop is FULL_PEL, return. |
470 | 0 | if (!round) { |
471 | 0 | return besterr; |
472 | 0 | } |
473 | | // Exit subpel search if the variance of the full-pel predicted samples is low (i.e. where likely interpolation will not modify the integer samples) |
474 | 0 | if (ms_params->pred_variance_th) { |
475 | 0 | const MSBuffers* ms_buffers = &var_params->ms_buffers; |
476 | 0 | const uint8_t* ref = svt_get_buf_from_mv(ms_buffers->ref, *bestmv); |
477 | 0 | unsigned int sse; |
478 | 0 | const unsigned int var = var_params->vfp->vf(ref, ms_buffers->ref->stride, svt_aom_eb_av1_var_offs, 0, &sse); |
479 | 0 | int block_var = ROUND_POWER_OF_TWO(var, eb_num_pels_log2_lookup[bsize]); |
480 | |
|
481 | 0 | if (block_var < ms_params->pred_variance_th) { |
482 | 0 | return besterr; |
483 | 0 | } |
484 | 0 | } |
485 | 0 | for (int iter = 0; iter < round; ++iter) { |
486 | 0 | Mv iter_center_mv = *bestmv; |
487 | 0 | Mv diag_step; |
488 | 0 | diag_step = svt_first_level_check( |
489 | 0 | xd, cm, iter_center_mv, bestmv, hstep, mv_limits, var_params, mv_cost_params, &besterr); |
490 | | |
491 | | // Check diagonal sub-pixel position |
492 | 0 | if (!CHECK_MV_EQUAL(iter_center_mv, *bestmv) && iters_per_step > 1) { |
493 | 0 | svt_second_level_check_v2( |
494 | 0 | xd, cm, iter_center_mv, diag_step, bestmv, mv_limits, var_params, mv_cost_params, &besterr); |
495 | 0 | } |
496 | |
|
497 | 0 | hstep >>= 1; |
498 | 0 | } |
499 | |
|
500 | 0 | return besterr; |
501 | 0 | } |
502 | | |
503 | | // ============================================================================= |
504 | | // SVT Functions |
505 | | // ============================================================================= |