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 <limits.h>
13 #include <math.h>
14 #include <stdio.h>
15
16 #include "./vpx_config.h"
17 #include "./vpx_dsp_rtcd.h"
18
19 #include "vpx_dsp/vpx_dsp_common.h"
20 #include "vpx_mem/vpx_mem.h"
21 #include "vpx_ports/mem.h"
22
23 #include "vp9/common/vp9_common.h"
24 #include "vp9/common/vp9_mvref_common.h"
25 #include "vp9/common/vp9_reconinter.h"
26
27 #include "vp9/encoder/vp9_encoder.h"
28 #include "vp9/encoder/vp9_mcomp.h"
29
30 // #define NEW_DIAMOND_SEARCH
31
get_buf_from_mv(const struct buf_2d * buf,const MV * mv)32 static INLINE const uint8_t *get_buf_from_mv(const struct buf_2d *buf,
33 const MV *mv) {
34 return &buf->buf[mv->row * buf->stride + mv->col];
35 }
36
vp9_set_mv_search_range(MvLimits * mv_limits,const MV * mv)37 void vp9_set_mv_search_range(MvLimits *mv_limits, const MV *mv) {
38 int col_min = (mv->col >> 3) - MAX_FULL_PEL_VAL + (mv->col & 7 ? 1 : 0);
39 int row_min = (mv->row >> 3) - MAX_FULL_PEL_VAL + (mv->row & 7 ? 1 : 0);
40 int col_max = (mv->col >> 3) + MAX_FULL_PEL_VAL;
41 int row_max = (mv->row >> 3) + MAX_FULL_PEL_VAL;
42
43 col_min = VPXMAX(col_min, (MV_LOW >> 3) + 1);
44 row_min = VPXMAX(row_min, (MV_LOW >> 3) + 1);
45 col_max = VPXMIN(col_max, (MV_UPP >> 3) - 1);
46 row_max = VPXMIN(row_max, (MV_UPP >> 3) - 1);
47
48 // Get intersection of UMV window and valid MV window to reduce # of checks
49 // in diamond search.
50 if (mv_limits->col_min < col_min) mv_limits->col_min = col_min;
51 if (mv_limits->col_max > col_max) mv_limits->col_max = col_max;
52 if (mv_limits->row_min < row_min) mv_limits->row_min = row_min;
53 if (mv_limits->row_max > row_max) mv_limits->row_max = row_max;
54 }
55
vp9_set_subpel_mv_search_range(MvLimits * subpel_mv_limits,const MvLimits * umv_window_limits,const MV * ref_mv)56 void vp9_set_subpel_mv_search_range(MvLimits *subpel_mv_limits,
57 const MvLimits *umv_window_limits,
58 const MV *ref_mv) {
59 subpel_mv_limits->col_min = VPXMAX(umv_window_limits->col_min * 8,
60 ref_mv->col - MAX_FULL_PEL_VAL * 8);
61 subpel_mv_limits->col_max = VPXMIN(umv_window_limits->col_max * 8,
62 ref_mv->col + MAX_FULL_PEL_VAL * 8);
63 subpel_mv_limits->row_min = VPXMAX(umv_window_limits->row_min * 8,
64 ref_mv->row - MAX_FULL_PEL_VAL * 8);
65 subpel_mv_limits->row_max = VPXMIN(umv_window_limits->row_max * 8,
66 ref_mv->row + MAX_FULL_PEL_VAL * 8);
67
68 subpel_mv_limits->col_min = VPXMAX(MV_LOW + 1, subpel_mv_limits->col_min);
69 subpel_mv_limits->col_max = VPXMIN(MV_UPP - 1, subpel_mv_limits->col_max);
70 subpel_mv_limits->row_min = VPXMAX(MV_LOW + 1, subpel_mv_limits->row_min);
71 subpel_mv_limits->row_max = VPXMIN(MV_UPP - 1, subpel_mv_limits->row_max);
72 }
73
vp9_init_search_range(int size)74 int vp9_init_search_range(int size) {
75 int sr = 0;
76 // Minimum search size no matter what the passed in value.
77 size = VPXMAX(16, size);
78
79 while ((size << sr) < MAX_FULL_PEL_VAL) sr++;
80
81 sr = VPXMIN(sr, MAX_MVSEARCH_STEPS - 2);
82 return sr;
83 }
84
mv_cost(const MV * mv,const int * joint_cost,int * const comp_cost[2])85 static INLINE int mv_cost(const MV *mv, const int *joint_cost,
86 int *const comp_cost[2]) {
87 assert(mv->row >= -MV_MAX && mv->row < MV_MAX);
88 assert(mv->col >= -MV_MAX && mv->col < MV_MAX);
89 return joint_cost[vp9_get_mv_joint(mv)] + comp_cost[0][mv->row] +
90 comp_cost[1][mv->col];
91 }
92
vp9_mv_bit_cost(const MV * mv,const MV * ref,const int * mvjcost,int * mvcost[2],int weight)93 int vp9_mv_bit_cost(const MV *mv, const MV *ref, const int *mvjcost,
94 int *mvcost[2], int weight) {
95 const MV diff = { mv->row - ref->row, mv->col - ref->col };
96 return ROUND_POWER_OF_TWO(mv_cost(&diff, mvjcost, mvcost) * weight, 7);
97 }
98
99 #define PIXEL_TRANSFORM_ERROR_SCALE 4
mv_err_cost(const MV * mv,const MV * ref,const int * mvjcost,int * mvcost[2],int error_per_bit)100 static int mv_err_cost(const MV *mv, const MV *ref, const int *mvjcost,
101 int *mvcost[2], int error_per_bit) {
102 if (mvcost) {
103 const MV diff = { mv->row - ref->row, mv->col - ref->col };
104 return (int)ROUND64_POWER_OF_TWO(
105 (int64_t)mv_cost(&diff, mvjcost, mvcost) * error_per_bit,
106 RDDIV_BITS + VP9_PROB_COST_SHIFT - RD_EPB_SHIFT +
107 PIXEL_TRANSFORM_ERROR_SCALE);
108 }
109 return 0;
110 }
111
mvsad_err_cost(const MACROBLOCK * x,const MV * mv,const MV * ref,int sad_per_bit)112 static int mvsad_err_cost(const MACROBLOCK *x, const MV *mv, const MV *ref,
113 int sad_per_bit) {
114 const MV diff = { mv->row - ref->row, mv->col - ref->col };
115 return ROUND_POWER_OF_TWO(
116 (unsigned)mv_cost(&diff, x->nmvjointsadcost, x->nmvsadcost) * sad_per_bit,
117 VP9_PROB_COST_SHIFT);
118 }
119
vp9_init_dsmotion_compensation(search_site_config * cfg,int stride)120 void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride) {
121 int len;
122 int ss_count = 0;
123
124 for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
125 // Generate offsets for 4 search sites per step.
126 const MV ss_mvs[] = { { -len, 0 }, { len, 0 }, { 0, -len }, { 0, len } };
127 int i;
128 for (i = 0; i < 4; ++i, ++ss_count) {
129 cfg->ss_mv[ss_count] = ss_mvs[i];
130 cfg->ss_os[ss_count] = ss_mvs[i].row * stride + ss_mvs[i].col;
131 }
132 }
133
134 cfg->searches_per_step = 4;
135 cfg->total_steps = ss_count / cfg->searches_per_step;
136 }
137
vp9_init3smotion_compensation(search_site_config * cfg,int stride)138 void vp9_init3smotion_compensation(search_site_config *cfg, int stride) {
139 int len;
140 int ss_count = 0;
141
142 for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
143 // Generate offsets for 8 search sites per step.
144 const MV ss_mvs[8] = { { -len, 0 }, { len, 0 }, { 0, -len },
145 { 0, len }, { -len, -len }, { -len, len },
146 { len, -len }, { len, len } };
147 int i;
148 for (i = 0; i < 8; ++i, ++ss_count) {
149 cfg->ss_mv[ss_count] = ss_mvs[i];
150 cfg->ss_os[ss_count] = ss_mvs[i].row * stride + ss_mvs[i].col;
151 }
152 }
153
154 cfg->searches_per_step = 8;
155 cfg->total_steps = ss_count / cfg->searches_per_step;
156 }
157
158 // convert motion vector component to offset for sv[a]f calc
sp(int x)159 static INLINE int sp(int x) { return x & 7; }
160
pre(const uint8_t * buf,int stride,int r,int c)161 static INLINE const uint8_t *pre(const uint8_t *buf, int stride, int r, int c) {
162 return &buf[(r >> 3) * stride + (c >> 3)];
163 }
164
165 #if CONFIG_VP9_HIGHBITDEPTH
166 /* checks if (r, c) has better score than previous best */
167 #define CHECK_BETTER(v, r, c) \
168 if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
169 int64_t tmpmse; \
170 const MV mv = { r, c }; \
171 const MV ref_mv = { rr, rc }; \
172 if (second_pred == NULL) { \
173 thismse = vfp->svf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z, \
174 src_stride, &sse); \
175 } else { \
176 thismse = vfp->svaf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z, \
177 src_stride, &sse, second_pred); \
178 } \
179 tmpmse = thismse; \
180 tmpmse += mv_err_cost(&mv, &ref_mv, mvjcost, mvcost, error_per_bit); \
181 if (tmpmse >= INT_MAX) { \
182 v = INT_MAX; \
183 } else if ((v = (uint32_t)tmpmse) < besterr) { \
184 besterr = v; \
185 br = r; \
186 bc = c; \
187 *distortion = thismse; \
188 *sse1 = sse; \
189 } \
190 } else { \
191 v = INT_MAX; \
192 }
193 #else
194 /* checks if (r, c) has better score than previous best */
195 #define CHECK_BETTER(v, r, c) \
196 if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
197 const MV mv = { r, c }; \
198 const MV ref_mv = { rr, rc }; \
199 if (second_pred == NULL) \
200 thismse = vfp->svf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z, \
201 src_stride, &sse); \
202 else \
203 thismse = vfp->svaf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z, \
204 src_stride, &sse, second_pred); \
205 if ((v = mv_err_cost(&mv, &ref_mv, mvjcost, mvcost, error_per_bit) + \
206 thismse) < besterr) { \
207 besterr = v; \
208 br = r; \
209 bc = c; \
210 *distortion = thismse; \
211 *sse1 = sse; \
212 } \
213 } else { \
214 v = INT_MAX; \
215 }
216
217 #endif
218 #define FIRST_LEVEL_CHECKS \
219 { \
220 unsigned int left, right, up, down, diag; \
221 CHECK_BETTER(left, tr, tc - hstep); \
222 CHECK_BETTER(right, tr, tc + hstep); \
223 CHECK_BETTER(up, tr - hstep, tc); \
224 CHECK_BETTER(down, tr + hstep, tc); \
225 whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2); \
226 switch (whichdir) { \
227 case 0: CHECK_BETTER(diag, tr - hstep, tc - hstep); break; \
228 case 1: CHECK_BETTER(diag, tr - hstep, tc + hstep); break; \
229 case 2: CHECK_BETTER(diag, tr + hstep, tc - hstep); break; \
230 case 3: CHECK_BETTER(diag, tr + hstep, tc + hstep); break; \
231 } \
232 }
233
234 #define SECOND_LEVEL_CHECKS \
235 { \
236 int kr, kc; \
237 unsigned int second; \
238 if (tr != br && tc != bc) { \
239 kr = br - tr; \
240 kc = bc - tc; \
241 CHECK_BETTER(second, tr + kr, tc + 2 * kc); \
242 CHECK_BETTER(second, tr + 2 * kr, tc + kc); \
243 } else if (tr == br && tc != bc) { \
244 kc = bc - tc; \
245 CHECK_BETTER(second, tr + hstep, tc + 2 * kc); \
246 CHECK_BETTER(second, tr - hstep, tc + 2 * kc); \
247 switch (whichdir) { \
248 case 0: \
249 case 1: CHECK_BETTER(second, tr + hstep, tc + kc); break; \
250 case 2: \
251 case 3: CHECK_BETTER(second, tr - hstep, tc + kc); break; \
252 } \
253 } else if (tr != br && tc == bc) { \
254 kr = br - tr; \
255 CHECK_BETTER(second, tr + 2 * kr, tc + hstep); \
256 CHECK_BETTER(second, tr + 2 * kr, tc - hstep); \
257 switch (whichdir) { \
258 case 0: \
259 case 2: CHECK_BETTER(second, tr + kr, tc + hstep); break; \
260 case 1: \
261 case 3: CHECK_BETTER(second, tr + kr, tc - hstep); break; \
262 } \
263 } \
264 }
265
266 #define SETUP_SUBPEL_SEARCH \
267 const uint8_t *const z = x->plane[0].src.buf; \
268 const int src_stride = x->plane[0].src.stride; \
269 const MACROBLOCKD *xd = &x->e_mbd; \
270 unsigned int besterr = UINT_MAX; \
271 unsigned int sse; \
272 unsigned int whichdir; \
273 int thismse; \
274 const unsigned int halfiters = iters_per_step; \
275 const unsigned int quarteriters = iters_per_step; \
276 const unsigned int eighthiters = iters_per_step; \
277 const int y_stride = xd->plane[0].pre[0].stride; \
278 const int offset = bestmv->row * y_stride + bestmv->col; \
279 const uint8_t *const y = xd->plane[0].pre[0].buf; \
280 \
281 int rr = ref_mv->row; \
282 int rc = ref_mv->col; \
283 int br = bestmv->row * 8; \
284 int bc = bestmv->col * 8; \
285 int hstep = 4; \
286 int minc, maxc, minr, maxr; \
287 int tr = br; \
288 int tc = bc; \
289 MvLimits subpel_mv_limits; \
290 \
291 vp9_set_subpel_mv_search_range(&subpel_mv_limits, &x->mv_limits, ref_mv); \
292 minc = subpel_mv_limits.col_min; \
293 maxc = subpel_mv_limits.col_max; \
294 minr = subpel_mv_limits.row_min; \
295 maxr = subpel_mv_limits.row_max; \
296 \
297 bestmv->row *= 8; \
298 bestmv->col *= 8;
299
setup_center_error(const MACROBLOCKD * xd,const MV * bestmv,const MV * ref_mv,int error_per_bit,const vp9_variance_fn_ptr_t * vfp,const uint8_t * const src,const int src_stride,const uint8_t * const y,int y_stride,const uint8_t * second_pred,int w,int h,int offset,int * mvjcost,int * mvcost[2],uint32_t * sse1,uint32_t * distortion)300 static unsigned int setup_center_error(
301 const MACROBLOCKD *xd, const MV *bestmv, const MV *ref_mv,
302 int error_per_bit, const vp9_variance_fn_ptr_t *vfp,
303 const uint8_t *const src, const int src_stride, const uint8_t *const y,
304 int y_stride, const uint8_t *second_pred, int w, int h, int offset,
305 int *mvjcost, int *mvcost[2], uint32_t *sse1, uint32_t *distortion) {
306 #if CONFIG_VP9_HIGHBITDEPTH
307 uint64_t besterr;
308 if (second_pred != NULL) {
309 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
310 DECLARE_ALIGNED(16, uint16_t, comp_pred16[64 * 64]);
311 vpx_highbd_comp_avg_pred(comp_pred16, CONVERT_TO_SHORTPTR(second_pred), w,
312 h, CONVERT_TO_SHORTPTR(y + offset), y_stride);
313 besterr =
314 vfp->vf(CONVERT_TO_BYTEPTR(comp_pred16), w, src, src_stride, sse1);
315 } else {
316 DECLARE_ALIGNED(16, uint8_t, comp_pred[64 * 64]);
317 vpx_comp_avg_pred(comp_pred, second_pred, w, h, y + offset, y_stride);
318 besterr = vfp->vf(comp_pred, w, src, src_stride, sse1);
319 }
320 } else {
321 besterr = vfp->vf(y + offset, y_stride, src, src_stride, sse1);
322 }
323 *distortion = (uint32_t)besterr;
324 besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
325 if (besterr >= UINT_MAX) return UINT_MAX;
326 return (uint32_t)besterr;
327 #else
328 uint32_t besterr;
329 (void)xd;
330 if (second_pred != NULL) {
331 DECLARE_ALIGNED(16, uint8_t, comp_pred[64 * 64]);
332 vpx_comp_avg_pred(comp_pred, second_pred, w, h, y + offset, y_stride);
333 besterr = vfp->vf(comp_pred, w, src, src_stride, sse1);
334 } else {
335 besterr = vfp->vf(y + offset, y_stride, src, src_stride, sse1);
336 }
337 *distortion = besterr;
338 besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
339 return besterr;
340 #endif // CONFIG_VP9_HIGHBITDEPTH
341 }
342
divide_and_round(const int64_t n,const int64_t d)343 static INLINE int64_t divide_and_round(const int64_t n, const int64_t d) {
344 return ((n < 0) ^ (d < 0)) ? ((n - d / 2) / d) : ((n + d / 2) / d);
345 }
346
is_cost_list_wellbehaved(int * cost_list)347 static INLINE int is_cost_list_wellbehaved(int *cost_list) {
348 return cost_list[0] < cost_list[1] && cost_list[0] < cost_list[2] &&
349 cost_list[0] < cost_list[3] && cost_list[0] < cost_list[4];
350 }
351
352 // Returns surface minima estimate at given precision in 1/2^n bits.
353 // Assume a model for the cost surface: S = A(x - x0)^2 + B(y - y0)^2 + C
354 // For a given set of costs S0, S1, S2, S3, S4 at points
355 // (y, x) = (0, 0), (0, -1), (1, 0), (0, 1) and (-1, 0) respectively,
356 // the solution for the location of the minima (x0, y0) is given by:
357 // x0 = 1/2 (S1 - S3)/(S1 + S3 - 2*S0),
358 // y0 = 1/2 (S4 - S2)/(S4 + S2 - 2*S0).
359 // The code below is an integerized version of that.
get_cost_surf_min(int * cost_list,int * ir,int * ic,int bits)360 static void get_cost_surf_min(int *cost_list, int *ir, int *ic, int bits) {
361 const int64_t x0 = (int64_t)cost_list[1] - cost_list[3];
362 const int64_t y0 = cost_list[1] - 2 * (int64_t)cost_list[0] + cost_list[3];
363 const int64_t x1 = (int64_t)cost_list[4] - cost_list[2];
364 const int64_t y1 = cost_list[4] - 2 * (int64_t)cost_list[0] + cost_list[2];
365 const int b = 1 << (bits - 1);
366 *ic = (int)divide_and_round(x0 * b, y0);
367 *ir = (int)divide_and_round(x1 * b, y1);
368 }
369
vp9_skip_sub_pixel_tree(const MACROBLOCK * x,MV * bestmv,const MV * ref_mv,int allow_hp,int error_per_bit,const vp9_variance_fn_ptr_t * vfp,int forced_stop,int iters_per_step,int * cost_list,int * mvjcost,int * mvcost[2],uint32_t * distortion,uint32_t * sse1,const uint8_t * second_pred,int w,int h,int use_accurate_subpel_search)370 uint32_t vp9_skip_sub_pixel_tree(
371 const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
372 int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
373 int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
374 uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
375 int h, int use_accurate_subpel_search) {
376 SETUP_SUBPEL_SEARCH;
377 besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp, z,
378 src_stride, y, y_stride, second_pred, w, h,
379 offset, mvjcost, mvcost, sse1, distortion);
380 (void)halfiters;
381 (void)quarteriters;
382 (void)eighthiters;
383 (void)whichdir;
384 (void)allow_hp;
385 (void)forced_stop;
386 (void)hstep;
387 (void)rr;
388 (void)rc;
389 (void)minr;
390 (void)minc;
391 (void)maxr;
392 (void)maxc;
393 (void)tr;
394 (void)tc;
395 (void)sse;
396 (void)thismse;
397 (void)cost_list;
398 (void)use_accurate_subpel_search;
399
400 return besterr;
401 }
402
vp9_find_best_sub_pixel_tree_pruned_evenmore(const MACROBLOCK * x,MV * bestmv,const MV * ref_mv,int allow_hp,int error_per_bit,const vp9_variance_fn_ptr_t * vfp,int forced_stop,int iters_per_step,int * cost_list,int * mvjcost,int * mvcost[2],uint32_t * distortion,uint32_t * sse1,const uint8_t * second_pred,int w,int h,int use_accurate_subpel_search)403 uint32_t vp9_find_best_sub_pixel_tree_pruned_evenmore(
404 const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
405 int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
406 int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
407 uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
408 int h, int use_accurate_subpel_search) {
409 SETUP_SUBPEL_SEARCH;
410 besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp, z,
411 src_stride, y, y_stride, second_pred, w, h,
412 offset, mvjcost, mvcost, sse1, distortion);
413 (void)halfiters;
414 (void)quarteriters;
415 (void)eighthiters;
416 (void)whichdir;
417 (void)allow_hp;
418 (void)forced_stop;
419 (void)hstep;
420 (void)use_accurate_subpel_search;
421
422 if (cost_list && cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
423 cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
424 cost_list[4] != INT_MAX && is_cost_list_wellbehaved(cost_list)) {
425 int ir, ic;
426 unsigned int minpt = INT_MAX;
427 get_cost_surf_min(cost_list, &ir, &ic, 2);
428 if (ir != 0 || ic != 0) {
429 CHECK_BETTER(minpt, tr + 2 * ir, tc + 2 * ic);
430 }
431 } else {
432 FIRST_LEVEL_CHECKS;
433 if (halfiters > 1) {
434 SECOND_LEVEL_CHECKS;
435 }
436
437 tr = br;
438 tc = bc;
439
440 // Each subsequent iteration checks at least one point in common with
441 // the last iteration could be 2 ( if diag selected) 1/4 pel
442 // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
443 if (forced_stop != 2) {
444 hstep >>= 1;
445 FIRST_LEVEL_CHECKS;
446 if (quarteriters > 1) {
447 SECOND_LEVEL_CHECKS;
448 }
449 }
450 }
451
452 tr = br;
453 tc = bc;
454
455 if (allow_hp && use_mv_hp(ref_mv) && forced_stop == 0) {
456 hstep >>= 1;
457 FIRST_LEVEL_CHECKS;
458 if (eighthiters > 1) {
459 SECOND_LEVEL_CHECKS;
460 }
461 }
462
463 bestmv->row = br;
464 bestmv->col = bc;
465
466 return besterr;
467 }
468
vp9_find_best_sub_pixel_tree_pruned_more(const MACROBLOCK * x,MV * bestmv,const MV * ref_mv,int allow_hp,int error_per_bit,const vp9_variance_fn_ptr_t * vfp,int forced_stop,int iters_per_step,int * cost_list,int * mvjcost,int * mvcost[2],uint32_t * distortion,uint32_t * sse1,const uint8_t * second_pred,int w,int h,int use_accurate_subpel_search)469 uint32_t vp9_find_best_sub_pixel_tree_pruned_more(
470 const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
471 int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
472 int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
473 uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
474 int h, int use_accurate_subpel_search) {
475 SETUP_SUBPEL_SEARCH;
476 (void)use_accurate_subpel_search;
477
478 besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp, z,
479 src_stride, y, y_stride, second_pred, w, h,
480 offset, mvjcost, mvcost, sse1, distortion);
481 if (cost_list && cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
482 cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
483 cost_list[4] != INT_MAX && is_cost_list_wellbehaved(cost_list)) {
484 unsigned int minpt;
485 int ir, ic;
486 get_cost_surf_min(cost_list, &ir, &ic, 1);
487 if (ir != 0 || ic != 0) {
488 CHECK_BETTER(minpt, tr + ir * hstep, tc + ic * hstep);
489 }
490 } else {
491 FIRST_LEVEL_CHECKS;
492 if (halfiters > 1) {
493 SECOND_LEVEL_CHECKS;
494 }
495 }
496
497 // Each subsequent iteration checks at least one point in common with
498 // the last iteration could be 2 ( if diag selected) 1/4 pel
499
500 // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
501 if (forced_stop != 2) {
502 tr = br;
503 tc = bc;
504 hstep >>= 1;
505 FIRST_LEVEL_CHECKS;
506 if (quarteriters > 1) {
507 SECOND_LEVEL_CHECKS;
508 }
509 }
510
511 if (allow_hp && use_mv_hp(ref_mv) && forced_stop == 0) {
512 tr = br;
513 tc = bc;
514 hstep >>= 1;
515 FIRST_LEVEL_CHECKS;
516 if (eighthiters > 1) {
517 SECOND_LEVEL_CHECKS;
518 }
519 }
520 // These lines insure static analysis doesn't warn that
521 // tr and tc aren't used after the above point.
522 (void)tr;
523 (void)tc;
524
525 bestmv->row = br;
526 bestmv->col = bc;
527
528 return besterr;
529 }
530
vp9_find_best_sub_pixel_tree_pruned(const MACROBLOCK * x,MV * bestmv,const MV * ref_mv,int allow_hp,int error_per_bit,const vp9_variance_fn_ptr_t * vfp,int forced_stop,int iters_per_step,int * cost_list,int * mvjcost,int * mvcost[2],uint32_t * distortion,uint32_t * sse1,const uint8_t * second_pred,int w,int h,int use_accurate_subpel_search)531 uint32_t vp9_find_best_sub_pixel_tree_pruned(
532 const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
533 int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
534 int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
535 uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
536 int h, int use_accurate_subpel_search) {
537 SETUP_SUBPEL_SEARCH;
538 (void)use_accurate_subpel_search;
539
540 besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp, z,
541 src_stride, y, y_stride, second_pred, w, h,
542 offset, mvjcost, mvcost, sse1, distortion);
543 if (cost_list && cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
544 cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
545 cost_list[4] != INT_MAX) {
546 unsigned int left, right, up, down, diag;
547 whichdir = (cost_list[1] < cost_list[3] ? 0 : 1) +
548 (cost_list[2] < cost_list[4] ? 0 : 2);
549 switch (whichdir) {
550 case 0:
551 CHECK_BETTER(left, tr, tc - hstep);
552 CHECK_BETTER(down, tr + hstep, tc);
553 CHECK_BETTER(diag, tr + hstep, tc - hstep);
554 break;
555 case 1:
556 CHECK_BETTER(right, tr, tc + hstep);
557 CHECK_BETTER(down, tr + hstep, tc);
558 CHECK_BETTER(diag, tr + hstep, tc + hstep);
559 break;
560 case 2:
561 CHECK_BETTER(left, tr, tc - hstep);
562 CHECK_BETTER(up, tr - hstep, tc);
563 CHECK_BETTER(diag, tr - hstep, tc - hstep);
564 break;
565 case 3:
566 CHECK_BETTER(right, tr, tc + hstep);
567 CHECK_BETTER(up, tr - hstep, tc);
568 CHECK_BETTER(diag, tr - hstep, tc + hstep);
569 break;
570 }
571 } else {
572 FIRST_LEVEL_CHECKS;
573 if (halfiters > 1) {
574 SECOND_LEVEL_CHECKS;
575 }
576 }
577
578 tr = br;
579 tc = bc;
580
581 // Each subsequent iteration checks at least one point in common with
582 // the last iteration could be 2 ( if diag selected) 1/4 pel
583
584 // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
585 if (forced_stop != 2) {
586 hstep >>= 1;
587 FIRST_LEVEL_CHECKS;
588 if (quarteriters > 1) {
589 SECOND_LEVEL_CHECKS;
590 }
591 tr = br;
592 tc = bc;
593 }
594
595 if (allow_hp && use_mv_hp(ref_mv) && forced_stop == 0) {
596 hstep >>= 1;
597 FIRST_LEVEL_CHECKS;
598 if (eighthiters > 1) {
599 SECOND_LEVEL_CHECKS;
600 }
601 tr = br;
602 tc = bc;
603 }
604 // These lines insure static analysis doesn't warn that
605 // tr and tc aren't used after the above point.
606 (void)tr;
607 (void)tc;
608
609 bestmv->row = br;
610 bestmv->col = bc;
611
612 return besterr;
613 }
614
615 /* clang-format off */
616 static const MV search_step_table[12] = {
617 // left, right, up, down
618 { 0, -4 }, { 0, 4 }, { -4, 0 }, { 4, 0 },
619 { 0, -2 }, { 0, 2 }, { -2, 0 }, { 2, 0 },
620 { 0, -1 }, { 0, 1 }, { -1, 0 }, { 1, 0 }
621 };
622 /* clang-format on */
623
accurate_sub_pel_search(const MACROBLOCKD * xd,const MV * this_mv,const struct scale_factors * sf,const InterpKernel * kernel,const vp9_variance_fn_ptr_t * vfp,const uint8_t * const src_address,const int src_stride,const uint8_t * const pre_address,int y_stride,const uint8_t * second_pred,int w,int h,uint32_t * sse)624 static int accurate_sub_pel_search(
625 const MACROBLOCKD *xd, const MV *this_mv, const struct scale_factors *sf,
626 const InterpKernel *kernel, const vp9_variance_fn_ptr_t *vfp,
627 const uint8_t *const src_address, const int src_stride,
628 const uint8_t *const pre_address, int y_stride, const uint8_t *second_pred,
629 int w, int h, uint32_t *sse) {
630 #if CONFIG_VP9_HIGHBITDEPTH
631 uint64_t besterr;
632 assert(sf->x_step_q4 == 16 && sf->y_step_q4 == 16);
633 assert(w != 0 && h != 0);
634 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
635 DECLARE_ALIGNED(16, uint16_t, pred16[64 * 64]);
636 vp9_highbd_build_inter_predictor(CONVERT_TO_SHORTPTR(pre_address), y_stride,
637 pred16, w, this_mv, sf, w, h, 0, kernel,
638 MV_PRECISION_Q3, 0, 0, xd->bd);
639 if (second_pred != NULL) {
640 DECLARE_ALIGNED(16, uint16_t, comp_pred16[64 * 64]);
641 vpx_highbd_comp_avg_pred(comp_pred16, CONVERT_TO_SHORTPTR(second_pred), w,
642 h, pred16, w);
643 besterr = vfp->vf(CONVERT_TO_BYTEPTR(comp_pred16), w, src_address,
644 src_stride, sse);
645 } else {
646 besterr =
647 vfp->vf(CONVERT_TO_BYTEPTR(pred16), w, src_address, src_stride, sse);
648 }
649 } else {
650 DECLARE_ALIGNED(16, uint8_t, pred[64 * 64]);
651 vp9_build_inter_predictor(pre_address, y_stride, pred, w, this_mv, sf, w, h,
652 0, kernel, MV_PRECISION_Q3, 0, 0);
653 if (second_pred != NULL) {
654 DECLARE_ALIGNED(16, uint8_t, comp_pred[64 * 64]);
655 vpx_comp_avg_pred(comp_pred, second_pred, w, h, pred, w);
656 besterr = vfp->vf(comp_pred, w, src_address, src_stride, sse);
657 } else {
658 besterr = vfp->vf(pred, w, src_address, src_stride, sse);
659 }
660 }
661 if (besterr >= UINT_MAX) return UINT_MAX;
662 return (int)besterr;
663 #else
664 int besterr;
665 DECLARE_ALIGNED(16, uint8_t, pred[64 * 64]);
666 assert(sf->x_step_q4 == 16 && sf->y_step_q4 == 16);
667 assert(w != 0 && h != 0);
668 (void)xd;
669
670 vp9_build_inter_predictor(pre_address, y_stride, pred, w, this_mv, sf, w, h,
671 0, kernel, MV_PRECISION_Q3, 0, 0);
672 if (second_pred != NULL) {
673 DECLARE_ALIGNED(16, uint8_t, comp_pred[64 * 64]);
674 vpx_comp_avg_pred(comp_pred, second_pred, w, h, pred, w);
675 besterr = vfp->vf(comp_pred, w, src_address, src_stride, sse);
676 } else {
677 besterr = vfp->vf(pred, w, src_address, src_stride, sse);
678 }
679 return besterr;
680 #endif // CONFIG_VP9_HIGHBITDEPTH
681 }
682
683 // TODO(yunqing): this part can be further refactored.
684 #if CONFIG_VP9_HIGHBITDEPTH
685 /* checks if (r, c) has better score than previous best */
686 #define CHECK_BETTER1(v, r, c) \
687 if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
688 int64_t tmpmse; \
689 const MV mv = { r, c }; \
690 const MV ref_mv = { rr, rc }; \
691 thismse = \
692 accurate_sub_pel_search(xd, &mv, x->me_sf, kernel, vfp, z, src_stride, \
693 y, y_stride, second_pred, w, h, &sse); \
694 tmpmse = thismse; \
695 tmpmse += mv_err_cost(&mv, &ref_mv, mvjcost, mvcost, error_per_bit); \
696 if (tmpmse >= INT_MAX) { \
697 v = INT_MAX; \
698 } else if ((v = (uint32_t)tmpmse) < besterr) { \
699 besterr = v; \
700 br = r; \
701 bc = c; \
702 *distortion = thismse; \
703 *sse1 = sse; \
704 } \
705 } else { \
706 v = INT_MAX; \
707 }
708 #else
709 /* checks if (r, c) has better score than previous best */
710 #define CHECK_BETTER1(v, r, c) \
711 if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
712 const MV mv = { r, c }; \
713 const MV ref_mv = { rr, rc }; \
714 thismse = \
715 accurate_sub_pel_search(xd, &mv, x->me_sf, kernel, vfp, z, src_stride, \
716 y, y_stride, second_pred, w, h, &sse); \
717 if ((v = mv_err_cost(&mv, &ref_mv, mvjcost, mvcost, error_per_bit) + \
718 thismse) < besterr) { \
719 besterr = v; \
720 br = r; \
721 bc = c; \
722 *distortion = thismse; \
723 *sse1 = sse; \
724 } \
725 } else { \
726 v = INT_MAX; \
727 }
728
729 #endif
730
vp9_find_best_sub_pixel_tree(const MACROBLOCK * x,MV * bestmv,const MV * ref_mv,int allow_hp,int error_per_bit,const vp9_variance_fn_ptr_t * vfp,int forced_stop,int iters_per_step,int * cost_list,int * mvjcost,int * mvcost[2],uint32_t * distortion,uint32_t * sse1,const uint8_t * second_pred,int w,int h,int use_accurate_subpel_search)731 uint32_t vp9_find_best_sub_pixel_tree(
732 const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
733 int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
734 int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
735 uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
736 int h, int use_accurate_subpel_search) {
737 const uint8_t *const z = x->plane[0].src.buf;
738 const uint8_t *const src_address = z;
739 const int src_stride = x->plane[0].src.stride;
740 const MACROBLOCKD *xd = &x->e_mbd;
741 unsigned int besterr = UINT_MAX;
742 unsigned int sse;
743 int thismse;
744 const int y_stride = xd->plane[0].pre[0].stride;
745 const int offset = bestmv->row * y_stride + bestmv->col;
746 const uint8_t *const y = xd->plane[0].pre[0].buf;
747
748 int rr = ref_mv->row;
749 int rc = ref_mv->col;
750 int br = bestmv->row * 8;
751 int bc = bestmv->col * 8;
752 int hstep = 4;
753 int iter, round = 3 - forced_stop;
754
755 int minc, maxc, minr, maxr;
756 int tr = br;
757 int tc = bc;
758 const MV *search_step = search_step_table;
759 int idx, best_idx = -1;
760 unsigned int cost_array[5];
761 int kr, kc;
762 MvLimits subpel_mv_limits;
763
764 // TODO(yunqing): need to add 4-tap filter optimization to speed up the
765 // encoder.
766 const InterpKernel *kernel =
767 (use_accurate_subpel_search > 0)
768 ? ((use_accurate_subpel_search == USE_4_TAPS)
769 ? vp9_filter_kernels[FOURTAP]
770 : ((use_accurate_subpel_search == USE_8_TAPS)
771 ? vp9_filter_kernels[EIGHTTAP]
772 : vp9_filter_kernels[EIGHTTAP_SHARP]))
773 : vp9_filter_kernels[BILINEAR];
774
775 vp9_set_subpel_mv_search_range(&subpel_mv_limits, &x->mv_limits, ref_mv);
776 minc = subpel_mv_limits.col_min;
777 maxc = subpel_mv_limits.col_max;
778 minr = subpel_mv_limits.row_min;
779 maxr = subpel_mv_limits.row_max;
780
781 if (!(allow_hp && use_mv_hp(ref_mv)))
782 if (round == 3) round = 2;
783
784 bestmv->row *= 8;
785 bestmv->col *= 8;
786
787 besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp, z,
788 src_stride, y, y_stride, second_pred, w, h,
789 offset, mvjcost, mvcost, sse1, distortion);
790
791 (void)cost_list; // to silence compiler warning
792
793 for (iter = 0; iter < round; ++iter) {
794 // Check vertical and horizontal sub-pixel positions.
795 for (idx = 0; idx < 4; ++idx) {
796 tr = br + search_step[idx].row;
797 tc = bc + search_step[idx].col;
798 if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
799 MV this_mv;
800 this_mv.row = tr;
801 this_mv.col = tc;
802
803 if (use_accurate_subpel_search) {
804 thismse = accurate_sub_pel_search(xd, &this_mv, x->me_sf, kernel, vfp,
805 src_address, src_stride, y,
806 y_stride, second_pred, w, h, &sse);
807 } else {
808 const uint8_t *const pre_address =
809 y + (tr >> 3) * y_stride + (tc >> 3);
810 if (second_pred == NULL)
811 thismse = vfp->svf(pre_address, y_stride, sp(tc), sp(tr),
812 src_address, src_stride, &sse);
813 else
814 thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
815 src_address, src_stride, &sse, second_pred);
816 }
817
818 cost_array[idx] = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost,
819 mvcost, error_per_bit);
820
821 if (cost_array[idx] < besterr) {
822 best_idx = idx;
823 besterr = cost_array[idx];
824 *distortion = thismse;
825 *sse1 = sse;
826 }
827 } else {
828 cost_array[idx] = UINT_MAX;
829 }
830 }
831
832 // Check diagonal sub-pixel position
833 kc = (cost_array[0] <= cost_array[1] ? -hstep : hstep);
834 kr = (cost_array[2] <= cost_array[3] ? -hstep : hstep);
835
836 tc = bc + kc;
837 tr = br + kr;
838 if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
839 MV this_mv = { tr, tc };
840 if (use_accurate_subpel_search) {
841 thismse = accurate_sub_pel_search(xd, &this_mv, x->me_sf, kernel, vfp,
842 src_address, src_stride, y, y_stride,
843 second_pred, w, h, &sse);
844 } else {
845 const uint8_t *const pre_address = y + (tr >> 3) * y_stride + (tc >> 3);
846 if (second_pred == NULL)
847 thismse = vfp->svf(pre_address, y_stride, sp(tc), sp(tr), src_address,
848 src_stride, &sse);
849 else
850 thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
851 src_address, src_stride, &sse, second_pred);
852 }
853
854 cost_array[4] = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost,
855 error_per_bit);
856
857 if (cost_array[4] < besterr) {
858 best_idx = 4;
859 besterr = cost_array[4];
860 *distortion = thismse;
861 *sse1 = sse;
862 }
863 } else {
864 cost_array[idx] = UINT_MAX;
865 }
866
867 if (best_idx < 4 && best_idx >= 0) {
868 br += search_step[best_idx].row;
869 bc += search_step[best_idx].col;
870 } else if (best_idx == 4) {
871 br = tr;
872 bc = tc;
873 }
874
875 if (iters_per_step > 0 && best_idx != -1) {
876 unsigned int second;
877 const int br0 = br;
878 const int bc0 = bc;
879 assert(tr == br || tc == bc);
880
881 if (tr == br && tc != bc) {
882 kc = bc - tc;
883 if (iters_per_step == 1) {
884 if (use_accurate_subpel_search) {
885 CHECK_BETTER1(second, br0, bc0 + kc);
886 } else {
887 CHECK_BETTER(second, br0, bc0 + kc);
888 }
889 }
890 } else if (tr != br && tc == bc) {
891 kr = br - tr;
892 if (iters_per_step == 1) {
893 if (use_accurate_subpel_search) {
894 CHECK_BETTER1(second, br0 + kr, bc0);
895 } else {
896 CHECK_BETTER(second, br0 + kr, bc0);
897 }
898 }
899 }
900
901 if (iters_per_step > 1) {
902 if (use_accurate_subpel_search) {
903 CHECK_BETTER1(second, br0 + kr, bc0);
904 CHECK_BETTER1(second, br0, bc0 + kc);
905 if (br0 != br || bc0 != bc) {
906 CHECK_BETTER1(second, br0 + kr, bc0 + kc);
907 }
908 } else {
909 CHECK_BETTER(second, br0 + kr, bc0);
910 CHECK_BETTER(second, br0, bc0 + kc);
911 if (br0 != br || bc0 != bc) {
912 CHECK_BETTER(second, br0 + kr, bc0 + kc);
913 }
914 }
915 }
916 }
917
918 search_step += 4;
919 hstep >>= 1;
920 best_idx = -1;
921 }
922
923 // Each subsequent iteration checks at least one point in common with
924 // the last iteration could be 2 ( if diag selected) 1/4 pel
925
926 // These lines insure static analysis doesn't warn that
927 // tr and tc aren't used after the above point.
928 (void)tr;
929 (void)tc;
930
931 bestmv->row = br;
932 bestmv->col = bc;
933
934 return besterr;
935 }
936
937 #undef CHECK_BETTER
938 #undef CHECK_BETTER1
939
check_bounds(const MvLimits * mv_limits,int row,int col,int range)940 static INLINE int check_bounds(const MvLimits *mv_limits, int row, int col,
941 int range) {
942 return ((row - range) >= mv_limits->row_min) &
943 ((row + range) <= mv_limits->row_max) &
944 ((col - range) >= mv_limits->col_min) &
945 ((col + range) <= mv_limits->col_max);
946 }
947
is_mv_in(const MvLimits * mv_limits,const MV * mv)948 static INLINE int is_mv_in(const MvLimits *mv_limits, const MV *mv) {
949 return (mv->col >= mv_limits->col_min) && (mv->col <= mv_limits->col_max) &&
950 (mv->row >= mv_limits->row_min) && (mv->row <= mv_limits->row_max);
951 }
952
953 #define CHECK_BETTER \
954 { \
955 if (thissad < bestsad) { \
956 if (use_mvcost) \
957 thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); \
958 if (thissad < bestsad) { \
959 bestsad = thissad; \
960 best_site = i; \
961 } \
962 } \
963 }
964
965 #define MAX_PATTERN_SCALES 11
966 #define MAX_PATTERN_CANDIDATES 8 // max number of canddiates per scale
967 #define PATTERN_CANDIDATES_REF 3 // number of refinement candidates
968
969 // Calculate and return a sad+mvcost list around an integer best pel.
calc_int_cost_list(const MACROBLOCK * x,const MV * ref_mv,int sadpb,const vp9_variance_fn_ptr_t * fn_ptr,const MV * best_mv,int * cost_list)970 static INLINE void calc_int_cost_list(const MACROBLOCK *x, const MV *ref_mv,
971 int sadpb,
972 const vp9_variance_fn_ptr_t *fn_ptr,
973 const MV *best_mv, int *cost_list) {
974 static const MV neighbors[4] = { { 0, -1 }, { 1, 0 }, { 0, 1 }, { -1, 0 } };
975 const struct buf_2d *const what = &x->plane[0].src;
976 const struct buf_2d *const in_what = &x->e_mbd.plane[0].pre[0];
977 const MV fcenter_mv = { ref_mv->row >> 3, ref_mv->col >> 3 };
978 int br = best_mv->row;
979 int bc = best_mv->col;
980 MV this_mv;
981 int i;
982 unsigned int sse;
983
984 this_mv.row = br;
985 this_mv.col = bc;
986 cost_list[0] =
987 fn_ptr->vf(what->buf, what->stride, get_buf_from_mv(in_what, &this_mv),
988 in_what->stride, &sse) +
989 mvsad_err_cost(x, &this_mv, &fcenter_mv, sadpb);
990 if (check_bounds(&x->mv_limits, br, bc, 1)) {
991 for (i = 0; i < 4; i++) {
992 const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
993 cost_list[i + 1] = fn_ptr->vf(what->buf, what->stride,
994 get_buf_from_mv(in_what, &this_mv),
995 in_what->stride, &sse) +
996 mv_err_cost(&this_mv, &fcenter_mv, x->nmvjointcost,
997 x->mvcost, x->errorperbit);
998 }
999 } else {
1000 for (i = 0; i < 4; i++) {
1001 const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
1002 if (!is_mv_in(&x->mv_limits, &this_mv))
1003 cost_list[i + 1] = INT_MAX;
1004 else
1005 cost_list[i + 1] = fn_ptr->vf(what->buf, what->stride,
1006 get_buf_from_mv(in_what, &this_mv),
1007 in_what->stride, &sse) +
1008 mv_err_cost(&this_mv, &fcenter_mv, x->nmvjointcost,
1009 x->mvcost, x->errorperbit);
1010 }
1011 }
1012 }
1013
1014 // Generic pattern search function that searches over multiple scales.
1015 // Each scale can have a different number of candidates and shape of
1016 // candidates as indicated in the num_candidates and candidates arrays
1017 // passed into this function
1018 //
vp9_pattern_search(const MACROBLOCK * x,MV * ref_mv,int search_param,int sad_per_bit,int do_init_search,int * cost_list,const vp9_variance_fn_ptr_t * vfp,int use_mvcost,const MV * center_mv,MV * best_mv,const int num_candidates[MAX_PATTERN_SCALES],const MV candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES])1019 static int vp9_pattern_search(
1020 const MACROBLOCK *x, MV *ref_mv, int search_param, int sad_per_bit,
1021 int do_init_search, int *cost_list, const vp9_variance_fn_ptr_t *vfp,
1022 int use_mvcost, const MV *center_mv, MV *best_mv,
1023 const int num_candidates[MAX_PATTERN_SCALES],
1024 const MV candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES]) {
1025 const MACROBLOCKD *const xd = &x->e_mbd;
1026 static const int search_param_to_steps[MAX_MVSEARCH_STEPS] = {
1027 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0,
1028 };
1029 int i, s, t;
1030 const struct buf_2d *const what = &x->plane[0].src;
1031 const struct buf_2d *const in_what = &xd->plane[0].pre[0];
1032 int br, bc;
1033 int bestsad = INT_MAX;
1034 int thissad;
1035 int k = -1;
1036 const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
1037 int best_init_s = search_param_to_steps[search_param];
1038 // adjust ref_mv to make sure it is within MV range
1039 clamp_mv(ref_mv, x->mv_limits.col_min, x->mv_limits.col_max,
1040 x->mv_limits.row_min, x->mv_limits.row_max);
1041 br = ref_mv->row;
1042 bc = ref_mv->col;
1043
1044 // Work out the start point for the search
1045 bestsad = vfp->sdf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv),
1046 in_what->stride) +
1047 mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
1048
1049 // Search all possible scales upto the search param around the center point
1050 // pick the scale of the point that is best as the starting scale of
1051 // further steps around it.
1052 if (do_init_search) {
1053 s = best_init_s;
1054 best_init_s = -1;
1055 for (t = 0; t <= s; ++t) {
1056 int best_site = -1;
1057 if (check_bounds(&x->mv_limits, br, bc, 1 << t)) {
1058 for (i = 0; i < num_candidates[t]; i++) {
1059 const MV this_mv = { br + candidates[t][i].row,
1060 bc + candidates[t][i].col };
1061 thissad =
1062 vfp->sdf(what->buf, what->stride,
1063 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1064 CHECK_BETTER
1065 }
1066 } else {
1067 for (i = 0; i < num_candidates[t]; i++) {
1068 const MV this_mv = { br + candidates[t][i].row,
1069 bc + candidates[t][i].col };
1070 if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1071 thissad =
1072 vfp->sdf(what->buf, what->stride,
1073 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1074 CHECK_BETTER
1075 }
1076 }
1077 if (best_site == -1) {
1078 continue;
1079 } else {
1080 best_init_s = t;
1081 k = best_site;
1082 }
1083 }
1084 if (best_init_s != -1) {
1085 br += candidates[best_init_s][k].row;
1086 bc += candidates[best_init_s][k].col;
1087 }
1088 }
1089
1090 // If the center point is still the best, just skip this and move to
1091 // the refinement step.
1092 if (best_init_s != -1) {
1093 int best_site = -1;
1094 s = best_init_s;
1095
1096 do {
1097 // No need to search all 6 points the 1st time if initial search was used
1098 if (!do_init_search || s != best_init_s) {
1099 if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1100 for (i = 0; i < num_candidates[s]; i++) {
1101 const MV this_mv = { br + candidates[s][i].row,
1102 bc + candidates[s][i].col };
1103 thissad =
1104 vfp->sdf(what->buf, what->stride,
1105 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1106 CHECK_BETTER
1107 }
1108 } else {
1109 for (i = 0; i < num_candidates[s]; i++) {
1110 const MV this_mv = { br + candidates[s][i].row,
1111 bc + candidates[s][i].col };
1112 if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1113 thissad =
1114 vfp->sdf(what->buf, what->stride,
1115 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1116 CHECK_BETTER
1117 }
1118 }
1119
1120 if (best_site == -1) {
1121 continue;
1122 } else {
1123 br += candidates[s][best_site].row;
1124 bc += candidates[s][best_site].col;
1125 k = best_site;
1126 }
1127 }
1128
1129 do {
1130 int next_chkpts_indices[PATTERN_CANDIDATES_REF];
1131 best_site = -1;
1132 next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
1133 next_chkpts_indices[1] = k;
1134 next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
1135
1136 if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1137 for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1138 const MV this_mv = {
1139 br + candidates[s][next_chkpts_indices[i]].row,
1140 bc + candidates[s][next_chkpts_indices[i]].col
1141 };
1142 thissad =
1143 vfp->sdf(what->buf, what->stride,
1144 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1145 CHECK_BETTER
1146 }
1147 } else {
1148 for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1149 const MV this_mv = {
1150 br + candidates[s][next_chkpts_indices[i]].row,
1151 bc + candidates[s][next_chkpts_indices[i]].col
1152 };
1153 if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1154 thissad =
1155 vfp->sdf(what->buf, what->stride,
1156 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1157 CHECK_BETTER
1158 }
1159 }
1160
1161 if (best_site != -1) {
1162 k = next_chkpts_indices[best_site];
1163 br += candidates[s][k].row;
1164 bc += candidates[s][k].col;
1165 }
1166 } while (best_site != -1);
1167 } while (s--);
1168 }
1169
1170 // Returns the one-away integer pel sad values around the best as follows:
1171 // cost_list[0]: cost at the best integer pel
1172 // cost_list[1]: cost at delta {0, -1} (left) from the best integer pel
1173 // cost_list[2]: cost at delta { 1, 0} (bottom) from the best integer pel
1174 // cost_list[3]: cost at delta { 0, 1} (right) from the best integer pel
1175 // cost_list[4]: cost at delta {-1, 0} (top) from the best integer pel
1176 if (cost_list) {
1177 const MV best_mv = { br, bc };
1178 calc_int_cost_list(x, &fcenter_mv, sad_per_bit, vfp, &best_mv, cost_list);
1179 }
1180 best_mv->row = br;
1181 best_mv->col = bc;
1182 return bestsad;
1183 }
1184
1185 // A specialized function where the smallest scale search candidates
1186 // are 4 1-away neighbors, and cost_list is non-null
1187 // TODO(debargha): Merge this function with the one above. Also remove
1188 // use_mvcost option since it is always 1, to save unnecessary branches.
vp9_pattern_search_sad(const MACROBLOCK * x,MV * ref_mv,int search_param,int sad_per_bit,int do_init_search,int * cost_list,const vp9_variance_fn_ptr_t * vfp,int use_mvcost,const MV * center_mv,MV * best_mv,const int num_candidates[MAX_PATTERN_SCALES],const MV candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES])1189 static int vp9_pattern_search_sad(
1190 const MACROBLOCK *x, MV *ref_mv, int search_param, int sad_per_bit,
1191 int do_init_search, int *cost_list, const vp9_variance_fn_ptr_t *vfp,
1192 int use_mvcost, const MV *center_mv, MV *best_mv,
1193 const int num_candidates[MAX_PATTERN_SCALES],
1194 const MV candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES]) {
1195 const MACROBLOCKD *const xd = &x->e_mbd;
1196 static const int search_param_to_steps[MAX_MVSEARCH_STEPS] = {
1197 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0,
1198 };
1199 int i, s, t;
1200 const struct buf_2d *const what = &x->plane[0].src;
1201 const struct buf_2d *const in_what = &xd->plane[0].pre[0];
1202 int br, bc;
1203 int bestsad = INT_MAX;
1204 int thissad;
1205 int k = -1;
1206 const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
1207 int best_init_s = search_param_to_steps[search_param];
1208 // adjust ref_mv to make sure it is within MV range
1209 clamp_mv(ref_mv, x->mv_limits.col_min, x->mv_limits.col_max,
1210 x->mv_limits.row_min, x->mv_limits.row_max);
1211 br = ref_mv->row;
1212 bc = ref_mv->col;
1213 if (cost_list != NULL) {
1214 cost_list[0] = cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] =
1215 INT_MAX;
1216 }
1217
1218 // Work out the start point for the search
1219 bestsad = vfp->sdf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv),
1220 in_what->stride) +
1221 mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
1222
1223 // Search all possible scales upto the search param around the center point
1224 // pick the scale of the point that is best as the starting scale of
1225 // further steps around it.
1226 if (do_init_search) {
1227 s = best_init_s;
1228 best_init_s = -1;
1229 for (t = 0; t <= s; ++t) {
1230 int best_site = -1;
1231 if (check_bounds(&x->mv_limits, br, bc, 1 << t)) {
1232 for (i = 0; i < num_candidates[t]; i++) {
1233 const MV this_mv = { br + candidates[t][i].row,
1234 bc + candidates[t][i].col };
1235 thissad =
1236 vfp->sdf(what->buf, what->stride,
1237 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1238 CHECK_BETTER
1239 }
1240 } else {
1241 for (i = 0; i < num_candidates[t]; i++) {
1242 const MV this_mv = { br + candidates[t][i].row,
1243 bc + candidates[t][i].col };
1244 if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1245 thissad =
1246 vfp->sdf(what->buf, what->stride,
1247 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1248 CHECK_BETTER
1249 }
1250 }
1251 if (best_site == -1) {
1252 continue;
1253 } else {
1254 best_init_s = t;
1255 k = best_site;
1256 }
1257 }
1258 if (best_init_s != -1) {
1259 br += candidates[best_init_s][k].row;
1260 bc += candidates[best_init_s][k].col;
1261 }
1262 }
1263
1264 // If the center point is still the best, just skip this and move to
1265 // the refinement step.
1266 if (best_init_s != -1) {
1267 int do_sad = (num_candidates[0] == 4 && cost_list != NULL);
1268 int best_site = -1;
1269 s = best_init_s;
1270
1271 for (; s >= do_sad; s--) {
1272 if (!do_init_search || s != best_init_s) {
1273 if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1274 for (i = 0; i < num_candidates[s]; i++) {
1275 const MV this_mv = { br + candidates[s][i].row,
1276 bc + candidates[s][i].col };
1277 thissad =
1278 vfp->sdf(what->buf, what->stride,
1279 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1280 CHECK_BETTER
1281 }
1282 } else {
1283 for (i = 0; i < num_candidates[s]; i++) {
1284 const MV this_mv = { br + candidates[s][i].row,
1285 bc + candidates[s][i].col };
1286 if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1287 thissad =
1288 vfp->sdf(what->buf, what->stride,
1289 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1290 CHECK_BETTER
1291 }
1292 }
1293
1294 if (best_site == -1) {
1295 continue;
1296 } else {
1297 br += candidates[s][best_site].row;
1298 bc += candidates[s][best_site].col;
1299 k = best_site;
1300 }
1301 }
1302
1303 do {
1304 int next_chkpts_indices[PATTERN_CANDIDATES_REF];
1305 best_site = -1;
1306 next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
1307 next_chkpts_indices[1] = k;
1308 next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
1309
1310 if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1311 for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1312 const MV this_mv = {
1313 br + candidates[s][next_chkpts_indices[i]].row,
1314 bc + candidates[s][next_chkpts_indices[i]].col
1315 };
1316 thissad =
1317 vfp->sdf(what->buf, what->stride,
1318 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1319 CHECK_BETTER
1320 }
1321 } else {
1322 for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1323 const MV this_mv = {
1324 br + candidates[s][next_chkpts_indices[i]].row,
1325 bc + candidates[s][next_chkpts_indices[i]].col
1326 };
1327 if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1328 thissad =
1329 vfp->sdf(what->buf, what->stride,
1330 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1331 CHECK_BETTER
1332 }
1333 }
1334
1335 if (best_site != -1) {
1336 k = next_chkpts_indices[best_site];
1337 br += candidates[s][k].row;
1338 bc += candidates[s][k].col;
1339 }
1340 } while (best_site != -1);
1341 }
1342
1343 // Note: If we enter the if below, then cost_list must be non-NULL.
1344 if (s == 0) {
1345 cost_list[0] = bestsad;
1346 if (!do_init_search || s != best_init_s) {
1347 if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1348 for (i = 0; i < num_candidates[s]; i++) {
1349 const MV this_mv = { br + candidates[s][i].row,
1350 bc + candidates[s][i].col };
1351 cost_list[i + 1] = thissad =
1352 vfp->sdf(what->buf, what->stride,
1353 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1354 CHECK_BETTER
1355 }
1356 } else {
1357 for (i = 0; i < num_candidates[s]; i++) {
1358 const MV this_mv = { br + candidates[s][i].row,
1359 bc + candidates[s][i].col };
1360 if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1361 cost_list[i + 1] = thissad =
1362 vfp->sdf(what->buf, what->stride,
1363 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1364 CHECK_BETTER
1365 }
1366 }
1367
1368 if (best_site != -1) {
1369 br += candidates[s][best_site].row;
1370 bc += candidates[s][best_site].col;
1371 k = best_site;
1372 }
1373 }
1374 while (best_site != -1) {
1375 int next_chkpts_indices[PATTERN_CANDIDATES_REF];
1376 best_site = -1;
1377 next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
1378 next_chkpts_indices[1] = k;
1379 next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
1380 cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] = INT_MAX;
1381 cost_list[((k + 2) % 4) + 1] = cost_list[0];
1382 cost_list[0] = bestsad;
1383
1384 if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1385 for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1386 const MV this_mv = {
1387 br + candidates[s][next_chkpts_indices[i]].row,
1388 bc + candidates[s][next_chkpts_indices[i]].col
1389 };
1390 cost_list[next_chkpts_indices[i] + 1] = thissad =
1391 vfp->sdf(what->buf, what->stride,
1392 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1393 CHECK_BETTER
1394 }
1395 } else {
1396 for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1397 const MV this_mv = {
1398 br + candidates[s][next_chkpts_indices[i]].row,
1399 bc + candidates[s][next_chkpts_indices[i]].col
1400 };
1401 if (!is_mv_in(&x->mv_limits, &this_mv)) {
1402 cost_list[next_chkpts_indices[i] + 1] = INT_MAX;
1403 continue;
1404 }
1405 cost_list[next_chkpts_indices[i] + 1] = thissad =
1406 vfp->sdf(what->buf, what->stride,
1407 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1408 CHECK_BETTER
1409 }
1410 }
1411
1412 if (best_site != -1) {
1413 k = next_chkpts_indices[best_site];
1414 br += candidates[s][k].row;
1415 bc += candidates[s][k].col;
1416 }
1417 }
1418 }
1419 }
1420
1421 // Returns the one-away integer pel sad values around the best as follows:
1422 // cost_list[0]: sad at the best integer pel
1423 // cost_list[1]: sad at delta {0, -1} (left) from the best integer pel
1424 // cost_list[2]: sad at delta { 1, 0} (bottom) from the best integer pel
1425 // cost_list[3]: sad at delta { 0, 1} (right) from the best integer pel
1426 // cost_list[4]: sad at delta {-1, 0} (top) from the best integer pel
1427 if (cost_list) {
1428 static const MV neighbors[4] = { { 0, -1 }, { 1, 0 }, { 0, 1 }, { -1, 0 } };
1429 if (cost_list[0] == INT_MAX) {
1430 cost_list[0] = bestsad;
1431 if (check_bounds(&x->mv_limits, br, bc, 1)) {
1432 for (i = 0; i < 4; i++) {
1433 const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
1434 cost_list[i + 1] =
1435 vfp->sdf(what->buf, what->stride,
1436 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1437 }
1438 } else {
1439 for (i = 0; i < 4; i++) {
1440 const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
1441 if (!is_mv_in(&x->mv_limits, &this_mv))
1442 cost_list[i + 1] = INT_MAX;
1443 else
1444 cost_list[i + 1] =
1445 vfp->sdf(what->buf, what->stride,
1446 get_buf_from_mv(in_what, &this_mv), in_what->stride);
1447 }
1448 }
1449 } else {
1450 if (use_mvcost) {
1451 for (i = 0; i < 4; i++) {
1452 const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
1453 if (cost_list[i + 1] != INT_MAX) {
1454 cost_list[i + 1] +=
1455 mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
1456 }
1457 }
1458 }
1459 }
1460 }
1461 best_mv->row = br;
1462 best_mv->col = bc;
1463 return bestsad;
1464 }
1465
vp9_get_mvpred_var(const MACROBLOCK * x,const MV * best_mv,const MV * center_mv,const vp9_variance_fn_ptr_t * vfp,int use_mvcost)1466 int vp9_get_mvpred_var(const MACROBLOCK *x, const MV *best_mv,
1467 const MV *center_mv, const vp9_variance_fn_ptr_t *vfp,
1468 int use_mvcost) {
1469 const MACROBLOCKD *const xd = &x->e_mbd;
1470 const struct buf_2d *const what = &x->plane[0].src;
1471 const struct buf_2d *const in_what = &xd->plane[0].pre[0];
1472 const MV mv = { best_mv->row * 8, best_mv->col * 8 };
1473 uint32_t unused;
1474 #if CONFIG_VP9_HIGHBITDEPTH
1475 uint64_t err =
1476 vfp->vf(what->buf, what->stride, get_buf_from_mv(in_what, best_mv),
1477 in_what->stride, &unused);
1478 err += (use_mvcost ? mv_err_cost(&mv, center_mv, x->nmvjointcost, x->mvcost,
1479 x->errorperbit)
1480 : 0);
1481 if (err >= INT_MAX) return INT_MAX;
1482 return (int)err;
1483 #else
1484 return vfp->vf(what->buf, what->stride, get_buf_from_mv(in_what, best_mv),
1485 in_what->stride, &unused) +
1486 (use_mvcost ? mv_err_cost(&mv, center_mv, x->nmvjointcost, x->mvcost,
1487 x->errorperbit)
1488 : 0);
1489 #endif
1490 }
1491
vp9_get_mvpred_av_var(const MACROBLOCK * x,const MV * best_mv,const MV * center_mv,const uint8_t * second_pred,const vp9_variance_fn_ptr_t * vfp,int use_mvcost)1492 int vp9_get_mvpred_av_var(const MACROBLOCK *x, const MV *best_mv,
1493 const MV *center_mv, const uint8_t *second_pred,
1494 const vp9_variance_fn_ptr_t *vfp, int use_mvcost) {
1495 const MACROBLOCKD *const xd = &x->e_mbd;
1496 const struct buf_2d *const what = &x->plane[0].src;
1497 const struct buf_2d *const in_what = &xd->plane[0].pre[0];
1498 const MV mv = { best_mv->row * 8, best_mv->col * 8 };
1499 unsigned int unused;
1500
1501 return vfp->svaf(get_buf_from_mv(in_what, best_mv), in_what->stride, 0, 0,
1502 what->buf, what->stride, &unused, second_pred) +
1503 (use_mvcost ? mv_err_cost(&mv, center_mv, x->nmvjointcost, x->mvcost,
1504 x->errorperbit)
1505 : 0);
1506 }
1507
hex_search(const MACROBLOCK * x,MV * ref_mv,int search_param,int sad_per_bit,int do_init_search,int * cost_list,const vp9_variance_fn_ptr_t * vfp,int use_mvcost,const MV * center_mv,MV * best_mv)1508 static int hex_search(const MACROBLOCK *x, MV *ref_mv, int search_param,
1509 int sad_per_bit, int do_init_search, int *cost_list,
1510 const vp9_variance_fn_ptr_t *vfp, int use_mvcost,
1511 const MV *center_mv, MV *best_mv) {
1512 // First scale has 8-closest points, the rest have 6 points in hex shape
1513 // at increasing scales
1514 static const int hex_num_candidates[MAX_PATTERN_SCALES] = { 8, 6, 6, 6, 6, 6,
1515 6, 6, 6, 6, 6 };
1516 // Note that the largest candidate step at each scale is 2^scale
1517 /* clang-format off */
1518 static const MV hex_candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES] = {
1519 { { -1, -1 }, { 0, -1 }, { 1, -1 }, { 1, 0 }, { 1, 1 }, { 0, 1 }, { -1, 1 },
1520 { -1, 0 } },
1521 { { -1, -2 }, { 1, -2 }, { 2, 0 }, { 1, 2 }, { -1, 2 }, { -2, 0 } },
1522 { { -2, -4 }, { 2, -4 }, { 4, 0 }, { 2, 4 }, { -2, 4 }, { -4, 0 } },
1523 { { -4, -8 }, { 4, -8 }, { 8, 0 }, { 4, 8 }, { -4, 8 }, { -8, 0 } },
1524 { { -8, -16 }, { 8, -16 }, { 16, 0 }, { 8, 16 }, { -8, 16 }, { -16, 0 } },
1525 { { -16, -32 }, { 16, -32 }, { 32, 0 }, { 16, 32 }, { -16, 32 },
1526 { -32, 0 } },
1527 { { -32, -64 }, { 32, -64 }, { 64, 0 }, { 32, 64 }, { -32, 64 },
1528 { -64, 0 } },
1529 { { -64, -128 }, { 64, -128 }, { 128, 0 }, { 64, 128 }, { -64, 128 },
1530 { -128, 0 } },
1531 { { -128, -256 }, { 128, -256 }, { 256, 0 }, { 128, 256 }, { -128, 256 },
1532 { -256, 0 } },
1533 { { -256, -512 }, { 256, -512 }, { 512, 0 }, { 256, 512 }, { -256, 512 },
1534 { -512, 0 } },
1535 { { -512, -1024 }, { 512, -1024 }, { 1024, 0 }, { 512, 1024 },
1536 { -512, 1024 }, { -1024, 0 } }
1537 };
1538 /* clang-format on */
1539 return vp9_pattern_search(
1540 x, ref_mv, search_param, sad_per_bit, do_init_search, cost_list, vfp,
1541 use_mvcost, center_mv, best_mv, hex_num_candidates, hex_candidates);
1542 }
1543
bigdia_search(const MACROBLOCK * x,MV * ref_mv,int search_param,int sad_per_bit,int do_init_search,int * cost_list,const vp9_variance_fn_ptr_t * vfp,int use_mvcost,const MV * center_mv,MV * best_mv)1544 static int bigdia_search(const MACROBLOCK *x, MV *ref_mv, int search_param,
1545 int sad_per_bit, int do_init_search, int *cost_list,
1546 const vp9_variance_fn_ptr_t *vfp, int use_mvcost,
1547 const MV *center_mv, MV *best_mv) {
1548 // First scale has 4-closest points, the rest have 8 points in diamond
1549 // shape at increasing scales
1550 static const int bigdia_num_candidates[MAX_PATTERN_SCALES] = {
1551 4, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1552 };
1553 // Note that the largest candidate step at each scale is 2^scale
1554 /* clang-format off */
1555 static const MV
1556 bigdia_candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES] = {
1557 { { 0, -1 }, { 1, 0 }, { 0, 1 }, { -1, 0 } },
1558 { { -1, -1 }, { 0, -2 }, { 1, -1 }, { 2, 0 }, { 1, 1 }, { 0, 2 },
1559 { -1, 1 }, { -2, 0 } },
1560 { { -2, -2 }, { 0, -4 }, { 2, -2 }, { 4, 0 }, { 2, 2 }, { 0, 4 },
1561 { -2, 2 }, { -4, 0 } },
1562 { { -4, -4 }, { 0, -8 }, { 4, -4 }, { 8, 0 }, { 4, 4 }, { 0, 8 },
1563 { -4, 4 }, { -8, 0 } },
1564 { { -8, -8 }, { 0, -16 }, { 8, -8 }, { 16, 0 }, { 8, 8 }, { 0, 16 },
1565 { -8, 8 }, { -16, 0 } },
1566 { { -16, -16 }, { 0, -32 }, { 16, -16 }, { 32, 0 }, { 16, 16 },
1567 { 0, 32 }, { -16, 16 }, { -32, 0 } },
1568 { { -32, -32 }, { 0, -64 }, { 32, -32 }, { 64, 0 }, { 32, 32 },
1569 { 0, 64 }, { -32, 32 }, { -64, 0 } },
1570 { { -64, -64 }, { 0, -128 }, { 64, -64 }, { 128, 0 }, { 64, 64 },
1571 { 0, 128 }, { -64, 64 }, { -128, 0 } },
1572 { { -128, -128 }, { 0, -256 }, { 128, -128 }, { 256, 0 }, { 128, 128 },
1573 { 0, 256 }, { -128, 128 }, { -256, 0 } },
1574 { { -256, -256 }, { 0, -512 }, { 256, -256 }, { 512, 0 }, { 256, 256 },
1575 { 0, 512 }, { -256, 256 }, { -512, 0 } },
1576 { { -512, -512 }, { 0, -1024 }, { 512, -512 }, { 1024, 0 },
1577 { 512, 512 }, { 0, 1024 }, { -512, 512 }, { -1024, 0 } }
1578 };
1579 /* clang-format on */
1580 return vp9_pattern_search_sad(
1581 x, ref_mv, search_param, sad_per_bit, do_init_search, cost_list, vfp,
1582 use_mvcost, center_mv, best_mv, bigdia_num_candidates, bigdia_candidates);
1583 }
1584
square_search(const MACROBLOCK * x,MV * ref_mv,int search_param,int sad_per_bit,int do_init_search,int * cost_list,const vp9_variance_fn_ptr_t * vfp,int use_mvcost,const MV * center_mv,MV * best_mv)1585 static int square_search(const MACROBLOCK *x, MV *ref_mv, int search_param,
1586 int sad_per_bit, int do_init_search, int *cost_list,
1587 const vp9_variance_fn_ptr_t *vfp, int use_mvcost,
1588 const MV *center_mv, MV *best_mv) {
1589 // All scales have 8 closest points in square shape
1590 static const int square_num_candidates[MAX_PATTERN_SCALES] = {
1591 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1592 };
1593 // Note that the largest candidate step at each scale is 2^scale
1594 /* clang-format off */
1595 static const MV
1596 square_candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES] = {
1597 { { -1, -1 }, { 0, -1 }, { 1, -1 }, { 1, 0 }, { 1, 1 }, { 0, 1 },
1598 { -1, 1 }, { -1, 0 } },
1599 { { -2, -2 }, { 0, -2 }, { 2, -2 }, { 2, 0 }, { 2, 2 }, { 0, 2 },
1600 { -2, 2 }, { -2, 0 } },
1601 { { -4, -4 }, { 0, -4 }, { 4, -4 }, { 4, 0 }, { 4, 4 }, { 0, 4 },
1602 { -4, 4 }, { -4, 0 } },
1603 { { -8, -8 }, { 0, -8 }, { 8, -8 }, { 8, 0 }, { 8, 8 }, { 0, 8 },
1604 { -8, 8 }, { -8, 0 } },
1605 { { -16, -16 }, { 0, -16 }, { 16, -16 }, { 16, 0 }, { 16, 16 },
1606 { 0, 16 }, { -16, 16 }, { -16, 0 } },
1607 { { -32, -32 }, { 0, -32 }, { 32, -32 }, { 32, 0 }, { 32, 32 },
1608 { 0, 32 }, { -32, 32 }, { -32, 0 } },
1609 { { -64, -64 }, { 0, -64 }, { 64, -64 }, { 64, 0 }, { 64, 64 },
1610 { 0, 64 }, { -64, 64 }, { -64, 0 } },
1611 { { -128, -128 }, { 0, -128 }, { 128, -128 }, { 128, 0 }, { 128, 128 },
1612 { 0, 128 }, { -128, 128 }, { -128, 0 } },
1613 { { -256, -256 }, { 0, -256 }, { 256, -256 }, { 256, 0 }, { 256, 256 },
1614 { 0, 256 }, { -256, 256 }, { -256, 0 } },
1615 { { -512, -512 }, { 0, -512 }, { 512, -512 }, { 512, 0 }, { 512, 512 },
1616 { 0, 512 }, { -512, 512 }, { -512, 0 } },
1617 { { -1024, -1024 }, { 0, -1024 }, { 1024, -1024 }, { 1024, 0 },
1618 { 1024, 1024 }, { 0, 1024 }, { -1024, 1024 }, { -1024, 0 } }
1619 };
1620 /* clang-format on */
1621 return vp9_pattern_search(
1622 x, ref_mv, search_param, sad_per_bit, do_init_search, cost_list, vfp,
1623 use_mvcost, center_mv, best_mv, square_num_candidates, square_candidates);
1624 }
1625
fast_hex_search(const MACROBLOCK * x,MV * ref_mv,int search_param,int sad_per_bit,int do_init_search,int * cost_list,const vp9_variance_fn_ptr_t * vfp,int use_mvcost,const MV * center_mv,MV * best_mv)1626 static int fast_hex_search(const MACROBLOCK *x, MV *ref_mv, int search_param,
1627 int sad_per_bit,
1628 int do_init_search, // must be zero for fast_hex
1629 int *cost_list, const vp9_variance_fn_ptr_t *vfp,
1630 int use_mvcost, const MV *center_mv, MV *best_mv) {
1631 return hex_search(x, ref_mv, VPXMAX(MAX_MVSEARCH_STEPS - 2, search_param),
1632 sad_per_bit, do_init_search, cost_list, vfp, use_mvcost,
1633 center_mv, best_mv);
1634 }
1635
fast_dia_search(const MACROBLOCK * x,MV * ref_mv,int search_param,int sad_per_bit,int do_init_search,int * cost_list,const vp9_variance_fn_ptr_t * vfp,int use_mvcost,const MV * center_mv,MV * best_mv)1636 static int fast_dia_search(const MACROBLOCK *x, MV *ref_mv, int search_param,
1637 int sad_per_bit, int do_init_search, int *cost_list,
1638 const vp9_variance_fn_ptr_t *vfp, int use_mvcost,
1639 const MV *center_mv, MV *best_mv) {
1640 return bigdia_search(x, ref_mv, VPXMAX(MAX_MVSEARCH_STEPS - 2, search_param),
1641 sad_per_bit, do_init_search, cost_list, vfp, use_mvcost,
1642 center_mv, best_mv);
1643 }
1644
1645 #undef CHECK_BETTER
1646
1647 // Exhuastive motion search around a given centre position with a given
1648 // step size.
exhuastive_mesh_search(const MACROBLOCK * x,MV * ref_mv,MV * best_mv,int range,int step,int sad_per_bit,const vp9_variance_fn_ptr_t * fn_ptr,const MV * center_mv)1649 static int exhuastive_mesh_search(const MACROBLOCK *x, MV *ref_mv, MV *best_mv,
1650 int range, int step, int sad_per_bit,
1651 const vp9_variance_fn_ptr_t *fn_ptr,
1652 const MV *center_mv) {
1653 const MACROBLOCKD *const xd = &x->e_mbd;
1654 const struct buf_2d *const what = &x->plane[0].src;
1655 const struct buf_2d *const in_what = &xd->plane[0].pre[0];
1656 MV fcenter_mv = { center_mv->row, center_mv->col };
1657 unsigned int best_sad = INT_MAX;
1658 int r, c, i;
1659 int start_col, end_col, start_row, end_row;
1660 int col_step = (step > 1) ? step : 4;
1661
1662 assert(step >= 1);
1663
1664 clamp_mv(&fcenter_mv, x->mv_limits.col_min, x->mv_limits.col_max,
1665 x->mv_limits.row_min, x->mv_limits.row_max);
1666 *best_mv = fcenter_mv;
1667 best_sad =
1668 fn_ptr->sdf(what->buf, what->stride,
1669 get_buf_from_mv(in_what, &fcenter_mv), in_what->stride) +
1670 mvsad_err_cost(x, &fcenter_mv, ref_mv, sad_per_bit);
1671 start_row = VPXMAX(-range, x->mv_limits.row_min - fcenter_mv.row);
1672 start_col = VPXMAX(-range, x->mv_limits.col_min - fcenter_mv.col);
1673 end_row = VPXMIN(range, x->mv_limits.row_max - fcenter_mv.row);
1674 end_col = VPXMIN(range, x->mv_limits.col_max - fcenter_mv.col);
1675
1676 for (r = start_row; r <= end_row; r += step) {
1677 for (c = start_col; c <= end_col; c += col_step) {
1678 // Step > 1 means we are not checking every location in this pass.
1679 if (step > 1) {
1680 const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c };
1681 unsigned int sad =
1682 fn_ptr->sdf(what->buf, what->stride, get_buf_from_mv(in_what, &mv),
1683 in_what->stride);
1684 if (sad < best_sad) {
1685 sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
1686 if (sad < best_sad) {
1687 best_sad = sad;
1688 *best_mv = mv;
1689 }
1690 }
1691 } else {
1692 // 4 sads in a single call if we are checking every location
1693 if (c + 3 <= end_col) {
1694 unsigned int sads[4];
1695 const uint8_t *addrs[4];
1696 for (i = 0; i < 4; ++i) {
1697 const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c + i };
1698 addrs[i] = get_buf_from_mv(in_what, &mv);
1699 }
1700 fn_ptr->sdx4df(what->buf, what->stride, addrs, in_what->stride, sads);
1701
1702 for (i = 0; i < 4; ++i) {
1703 if (sads[i] < best_sad) {
1704 const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c + i };
1705 const unsigned int sad =
1706 sads[i] + mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
1707 if (sad < best_sad) {
1708 best_sad = sad;
1709 *best_mv = mv;
1710 }
1711 }
1712 }
1713 } else {
1714 for (i = 0; i < end_col - c; ++i) {
1715 const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c + i };
1716 unsigned int sad =
1717 fn_ptr->sdf(what->buf, what->stride,
1718 get_buf_from_mv(in_what, &mv), in_what->stride);
1719 if (sad < best_sad) {
1720 sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
1721 if (sad < best_sad) {
1722 best_sad = sad;
1723 *best_mv = mv;
1724 }
1725 }
1726 }
1727 }
1728 }
1729 }
1730 }
1731
1732 return best_sad;
1733 }
1734
1735 #if CONFIG_NON_GREEDY_MV
vp9_nb_mvs_inconsistency(const MV * mv,const int_mv * nb_mvs,int mv_num)1736 double vp9_nb_mvs_inconsistency(const MV *mv, const int_mv *nb_mvs,
1737 int mv_num) {
1738 int i;
1739 int update = 0;
1740 double best_cost = 0;
1741 vpx_clear_system_state();
1742 for (i = 0; i < mv_num; ++i) {
1743 if (nb_mvs[i].as_int != INVALID_MV) {
1744 MV nb_mv = nb_mvs[i].as_mv;
1745 const double row_diff = mv->row - nb_mv.row;
1746 const double col_diff = mv->col - nb_mv.col;
1747 double cost = row_diff * row_diff + col_diff * col_diff;
1748 cost = log2(1 + cost);
1749 if (update == 0) {
1750 best_cost = cost;
1751 update = 1;
1752 } else {
1753 best_cost = cost < best_cost ? cost : best_cost;
1754 }
1755 }
1756 }
1757 return best_cost;
1758 }
1759
vp9_diamond_search_sad_new(const MACROBLOCK * x,const search_site_config * cfg,const MV * init_full_mv,MV * best_full_mv,double * best_mv_dist,double * best_mv_cost,int search_param,double lambda,int * num00,const vp9_variance_fn_ptr_t * fn_ptr,const int_mv * nb_full_mvs,int full_mv_num)1760 double vp9_diamond_search_sad_new(const MACROBLOCK *x,
1761 const search_site_config *cfg,
1762 const MV *init_full_mv, MV *best_full_mv,
1763 double *best_mv_dist, double *best_mv_cost,
1764 int search_param, double lambda, int *num00,
1765 const vp9_variance_fn_ptr_t *fn_ptr,
1766 const int_mv *nb_full_mvs, int full_mv_num) {
1767 int i, j, step;
1768
1769 const MACROBLOCKD *const xd = &x->e_mbd;
1770 uint8_t *what = x->plane[0].src.buf;
1771 const int what_stride = x->plane[0].src.stride;
1772 const uint8_t *in_what;
1773 const int in_what_stride = xd->plane[0].pre[0].stride;
1774 const uint8_t *best_address;
1775
1776 double bestsad;
1777 int best_site = -1;
1778 int last_site = -1;
1779
1780 // search_param determines the length of the initial step and hence the number
1781 // of iterations.
1782 // 0 = initial step (MAX_FIRST_STEP) pel
1783 // 1 = (MAX_FIRST_STEP/2) pel,
1784 // 2 = (MAX_FIRST_STEP/4) pel...
1785 // const search_site *ss = &cfg->ss[search_param * cfg->searches_per_step];
1786 const MV *ss_mv = &cfg->ss_mv[search_param * cfg->searches_per_step];
1787 const intptr_t *ss_os = &cfg->ss_os[search_param * cfg->searches_per_step];
1788 const int tot_steps = cfg->total_steps - search_param;
1789 vpx_clear_system_state();
1790
1791 *best_full_mv = *init_full_mv;
1792 clamp_mv(best_full_mv, x->mv_limits.col_min, x->mv_limits.col_max,
1793 x->mv_limits.row_min, x->mv_limits.row_max);
1794 *num00 = 0;
1795
1796 // Work out the start point for the search
1797 in_what = xd->plane[0].pre[0].buf + best_full_mv->row * in_what_stride +
1798 best_full_mv->col;
1799 best_address = in_what;
1800
1801 // Check the starting position
1802 *best_mv_dist = fn_ptr->sdf(what, what_stride, in_what, in_what_stride);
1803 *best_mv_cost =
1804 vp9_nb_mvs_inconsistency(best_full_mv, nb_full_mvs, full_mv_num);
1805 bestsad = (*best_mv_dist) + lambda * (*best_mv_cost);
1806
1807 i = 0;
1808
1809 for (step = 0; step < tot_steps; step++) {
1810 int all_in = 1, t;
1811
1812 // All_in is true if every one of the points we are checking are within
1813 // the bounds of the image.
1814 all_in &= ((best_full_mv->row + ss_mv[i].row) > x->mv_limits.row_min);
1815 all_in &= ((best_full_mv->row + ss_mv[i + 1].row) < x->mv_limits.row_max);
1816 all_in &= ((best_full_mv->col + ss_mv[i + 2].col) > x->mv_limits.col_min);
1817 all_in &= ((best_full_mv->col + ss_mv[i + 3].col) < x->mv_limits.col_max);
1818
1819 // If all the pixels are within the bounds we don't check whether the
1820 // search point is valid in this loop, otherwise we check each point
1821 // for validity..
1822 if (all_in) {
1823 unsigned int sad_array[4];
1824
1825 for (j = 0; j < cfg->searches_per_step; j += 4) {
1826 unsigned char const *block_offset[4];
1827
1828 for (t = 0; t < 4; t++) block_offset[t] = ss_os[i + t] + best_address;
1829
1830 fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride,
1831 sad_array);
1832
1833 for (t = 0; t < 4; t++, i++) {
1834 const MV this_mv = { best_full_mv->row + ss_mv[i].row,
1835 best_full_mv->col + ss_mv[i].col };
1836 const double mv_dist = sad_array[t];
1837 const double mv_cost =
1838 vp9_nb_mvs_inconsistency(&this_mv, nb_full_mvs, full_mv_num);
1839 double thissad = mv_dist + lambda * mv_cost;
1840 if (thissad < bestsad) {
1841 bestsad = thissad;
1842 *best_mv_dist = mv_dist;
1843 *best_mv_cost = mv_cost;
1844 best_site = i;
1845 }
1846 }
1847 }
1848 } else {
1849 for (j = 0; j < cfg->searches_per_step; j++) {
1850 // Trap illegal vectors
1851 const MV this_mv = { best_full_mv->row + ss_mv[i].row,
1852 best_full_mv->col + ss_mv[i].col };
1853
1854 if (is_mv_in(&x->mv_limits, &this_mv)) {
1855 const uint8_t *const check_here = ss_os[i] + best_address;
1856 const double mv_dist =
1857 fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
1858 const double mv_cost =
1859 vp9_nb_mvs_inconsistency(&this_mv, nb_full_mvs, full_mv_num);
1860 double thissad = mv_dist + lambda * mv_cost;
1861 if (thissad < bestsad) {
1862 bestsad = thissad;
1863 *best_mv_dist = mv_dist;
1864 *best_mv_cost = mv_cost;
1865 best_site = i;
1866 }
1867 }
1868 i++;
1869 }
1870 }
1871 if (best_site != last_site) {
1872 best_full_mv->row += ss_mv[best_site].row;
1873 best_full_mv->col += ss_mv[best_site].col;
1874 best_address += ss_os[best_site];
1875 last_site = best_site;
1876 } else if (best_address == in_what) {
1877 (*num00)++;
1878 }
1879 }
1880 return bestsad;
1881 }
1882
vp9_prepare_nb_full_mvs(const TplDepFrame * tpl_frame,int mi_row,int mi_col,int rf_idx,BLOCK_SIZE bsize,int_mv * nb_full_mvs)1883 void vp9_prepare_nb_full_mvs(const TplDepFrame *tpl_frame, int mi_row,
1884 int mi_col, int rf_idx, BLOCK_SIZE bsize,
1885 int_mv *nb_full_mvs) {
1886 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
1887 const int mi_height = num_8x8_blocks_high_lookup[bsize];
1888 const int dirs[NB_MVS_NUM][2] = { { -1, 0 }, { 0, -1 }, { 1, 0 }, { 0, 1 } };
1889 int i;
1890 for (i = 0; i < NB_MVS_NUM; ++i) {
1891 int r = dirs[i][0] * mi_height;
1892 int c = dirs[i][1] * mi_width;
1893 if (mi_row + r >= 0 && mi_row + r < tpl_frame->mi_rows && mi_col + c >= 0 &&
1894 mi_col + c < tpl_frame->mi_cols) {
1895 const TplDepStats *tpl_ptr =
1896 &tpl_frame
1897 ->tpl_stats_ptr[(mi_row + r) * tpl_frame->stride + mi_col + c];
1898 int_mv *mv =
1899 get_pyramid_mv(tpl_frame, rf_idx, bsize, mi_row + r, mi_col + c);
1900 if (tpl_ptr->ready[rf_idx]) {
1901 nb_full_mvs[i].as_mv = get_full_mv(&mv->as_mv);
1902 } else {
1903 nb_full_mvs[i].as_int = INVALID_MV;
1904 }
1905 } else {
1906 nb_full_mvs[i].as_int = INVALID_MV;
1907 }
1908 }
1909 }
1910 #endif // CONFIG_NON_GREEDY_MV
1911
vp9_diamond_search_sad_c(const MACROBLOCK * x,const search_site_config * cfg,MV * ref_mv,MV * best_mv,int search_param,int sad_per_bit,int * num00,const vp9_variance_fn_ptr_t * fn_ptr,const MV * center_mv)1912 int vp9_diamond_search_sad_c(const MACROBLOCK *x, const search_site_config *cfg,
1913 MV *ref_mv, MV *best_mv, int search_param,
1914 int sad_per_bit, int *num00,
1915 const vp9_variance_fn_ptr_t *fn_ptr,
1916 const MV *center_mv) {
1917 int i, j, step;
1918
1919 const MACROBLOCKD *const xd = &x->e_mbd;
1920 uint8_t *what = x->plane[0].src.buf;
1921 const int what_stride = x->plane[0].src.stride;
1922 const uint8_t *in_what;
1923 const int in_what_stride = xd->plane[0].pre[0].stride;
1924 const uint8_t *best_address;
1925
1926 unsigned int bestsad = INT_MAX;
1927 int best_site = -1;
1928 int last_site = -1;
1929
1930 int ref_row;
1931 int ref_col;
1932
1933 // search_param determines the length of the initial step and hence the number
1934 // of iterations.
1935 // 0 = initial step (MAX_FIRST_STEP) pel
1936 // 1 = (MAX_FIRST_STEP/2) pel,
1937 // 2 = (MAX_FIRST_STEP/4) pel...
1938 // const search_site *ss = &cfg->ss[search_param * cfg->searches_per_step];
1939 const MV *ss_mv = &cfg->ss_mv[search_param * cfg->searches_per_step];
1940 const intptr_t *ss_os = &cfg->ss_os[search_param * cfg->searches_per_step];
1941 const int tot_steps = cfg->total_steps - search_param;
1942
1943 const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
1944 clamp_mv(ref_mv, x->mv_limits.col_min, x->mv_limits.col_max,
1945 x->mv_limits.row_min, x->mv_limits.row_max);
1946 ref_row = ref_mv->row;
1947 ref_col = ref_mv->col;
1948 *num00 = 0;
1949 best_mv->row = ref_row;
1950 best_mv->col = ref_col;
1951
1952 // Work out the start point for the search
1953 in_what = xd->plane[0].pre[0].buf + ref_row * in_what_stride + ref_col;
1954 best_address = in_what;
1955
1956 // Check the starting position
1957 bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride) +
1958 mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit);
1959
1960 i = 0;
1961
1962 for (step = 0; step < tot_steps; step++) {
1963 int all_in = 1, t;
1964
1965 // All_in is true if every one of the points we are checking are within
1966 // the bounds of the image.
1967 all_in &= ((best_mv->row + ss_mv[i].row) > x->mv_limits.row_min);
1968 all_in &= ((best_mv->row + ss_mv[i + 1].row) < x->mv_limits.row_max);
1969 all_in &= ((best_mv->col + ss_mv[i + 2].col) > x->mv_limits.col_min);
1970 all_in &= ((best_mv->col + ss_mv[i + 3].col) < x->mv_limits.col_max);
1971
1972 // If all the pixels are within the bounds we don't check whether the
1973 // search point is valid in this loop, otherwise we check each point
1974 // for validity..
1975 if (all_in) {
1976 unsigned int sad_array[4];
1977
1978 for (j = 0; j < cfg->searches_per_step; j += 4) {
1979 unsigned char const *block_offset[4];
1980
1981 for (t = 0; t < 4; t++) block_offset[t] = ss_os[i + t] + best_address;
1982
1983 fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride,
1984 sad_array);
1985
1986 for (t = 0; t < 4; t++, i++) {
1987 if (sad_array[t] < bestsad) {
1988 const MV this_mv = { best_mv->row + ss_mv[i].row,
1989 best_mv->col + ss_mv[i].col };
1990 sad_array[t] +=
1991 mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
1992 if (sad_array[t] < bestsad) {
1993 bestsad = sad_array[t];
1994 best_site = i;
1995 }
1996 }
1997 }
1998 }
1999 } else {
2000 for (j = 0; j < cfg->searches_per_step; j++) {
2001 // Trap illegal vectors
2002 const MV this_mv = { best_mv->row + ss_mv[i].row,
2003 best_mv->col + ss_mv[i].col };
2004
2005 if (is_mv_in(&x->mv_limits, &this_mv)) {
2006 const uint8_t *const check_here = ss_os[i] + best_address;
2007 unsigned int thissad =
2008 fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
2009
2010 if (thissad < bestsad) {
2011 thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
2012 if (thissad < bestsad) {
2013 bestsad = thissad;
2014 best_site = i;
2015 }
2016 }
2017 }
2018 i++;
2019 }
2020 }
2021 if (best_site != last_site) {
2022 best_mv->row += ss_mv[best_site].row;
2023 best_mv->col += ss_mv[best_site].col;
2024 best_address += ss_os[best_site];
2025 last_site = best_site;
2026 #if defined(NEW_DIAMOND_SEARCH)
2027 while (1) {
2028 const MV this_mv = { best_mv->row + ss_mv[best_site].row,
2029 best_mv->col + ss_mv[best_site].col };
2030 if (is_mv_in(&x->mv_limits, &this_mv)) {
2031 const uint8_t *const check_here = ss_os[best_site] + best_address;
2032 unsigned int thissad =
2033 fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
2034 if (thissad < bestsad) {
2035 thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
2036 if (thissad < bestsad) {
2037 bestsad = thissad;
2038 best_mv->row += ss_mv[best_site].row;
2039 best_mv->col += ss_mv[best_site].col;
2040 best_address += ss_os[best_site];
2041 continue;
2042 }
2043 }
2044 }
2045 break;
2046 }
2047 #endif
2048 } else if (best_address == in_what) {
2049 (*num00)++;
2050 }
2051 }
2052 return bestsad;
2053 }
2054
vector_match(int16_t * ref,int16_t * src,int bwl)2055 static int vector_match(int16_t *ref, int16_t *src, int bwl) {
2056 int best_sad = INT_MAX;
2057 int this_sad;
2058 int d;
2059 int center, offset = 0;
2060 int bw = 4 << bwl; // redundant variable, to be changed in the experiments.
2061 for (d = 0; d <= bw; d += 16) {
2062 this_sad = vpx_vector_var(&ref[d], src, bwl);
2063 if (this_sad < best_sad) {
2064 best_sad = this_sad;
2065 offset = d;
2066 }
2067 }
2068 center = offset;
2069
2070 for (d = -8; d <= 8; d += 16) {
2071 int this_pos = offset + d;
2072 // check limit
2073 if (this_pos < 0 || this_pos > bw) continue;
2074 this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
2075 if (this_sad < best_sad) {
2076 best_sad = this_sad;
2077 center = this_pos;
2078 }
2079 }
2080 offset = center;
2081
2082 for (d = -4; d <= 4; d += 8) {
2083 int this_pos = offset + d;
2084 // check limit
2085 if (this_pos < 0 || this_pos > bw) continue;
2086 this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
2087 if (this_sad < best_sad) {
2088 best_sad = this_sad;
2089 center = this_pos;
2090 }
2091 }
2092 offset = center;
2093
2094 for (d = -2; d <= 2; d += 4) {
2095 int this_pos = offset + d;
2096 // check limit
2097 if (this_pos < 0 || this_pos > bw) continue;
2098 this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
2099 if (this_sad < best_sad) {
2100 best_sad = this_sad;
2101 center = this_pos;
2102 }
2103 }
2104 offset = center;
2105
2106 for (d = -1; d <= 1; d += 2) {
2107 int this_pos = offset + d;
2108 // check limit
2109 if (this_pos < 0 || this_pos > bw) continue;
2110 this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
2111 if (this_sad < best_sad) {
2112 best_sad = this_sad;
2113 center = this_pos;
2114 }
2115 }
2116
2117 return (center - (bw >> 1));
2118 }
2119
2120 static const MV search_pos[4] = {
2121 { -1, 0 },
2122 { 0, -1 },
2123 { 0, 1 },
2124 { 1, 0 },
2125 };
2126
vp9_int_pro_motion_estimation(const VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int mi_row,int mi_col,const MV * ref_mv)2127 unsigned int vp9_int_pro_motion_estimation(const VP9_COMP *cpi, MACROBLOCK *x,
2128 BLOCK_SIZE bsize, int mi_row,
2129 int mi_col, const MV *ref_mv) {
2130 MACROBLOCKD *xd = &x->e_mbd;
2131 MODE_INFO *mi = xd->mi[0];
2132 struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
2133 DECLARE_ALIGNED(16, int16_t, hbuf[128]);
2134 DECLARE_ALIGNED(16, int16_t, vbuf[128]);
2135 DECLARE_ALIGNED(16, int16_t, src_hbuf[64]);
2136 DECLARE_ALIGNED(16, int16_t, src_vbuf[64]);
2137 int idx;
2138 const int bw = 4 << b_width_log2_lookup[bsize];
2139 const int bh = 4 << b_height_log2_lookup[bsize];
2140 const int search_width = bw << 1;
2141 const int search_height = bh << 1;
2142 const int src_stride = x->plane[0].src.stride;
2143 const int ref_stride = xd->plane[0].pre[0].stride;
2144 uint8_t const *ref_buf, *src_buf;
2145 MV *tmp_mv = &xd->mi[0]->mv[0].as_mv;
2146 unsigned int best_sad, tmp_sad, this_sad[4];
2147 MV this_mv;
2148 const int norm_factor = 3 + (bw >> 5);
2149 const YV12_BUFFER_CONFIG *scaled_ref_frame =
2150 vp9_get_scaled_ref_frame(cpi, mi->ref_frame[0]);
2151 MvLimits subpel_mv_limits;
2152
2153 if (scaled_ref_frame) {
2154 int i;
2155 // Swap out the reference frame for a version that's been scaled to
2156 // match the resolution of the current frame, allowing the existing
2157 // motion search code to be used without additional modifications.
2158 for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
2159 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2160 }
2161
2162 #if CONFIG_VP9_HIGHBITDEPTH
2163 // TODO(jingning): Implement integral projection functions for high bit-depth
2164 // setting and remove this part of code.
2165 if (xd->bd != 8) {
2166 unsigned int this_sad;
2167 tmp_mv->row = 0;
2168 tmp_mv->col = 0;
2169 this_sad = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf, src_stride,
2170 xd->plane[0].pre[0].buf, ref_stride);
2171
2172 if (scaled_ref_frame) {
2173 int i;
2174 for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
2175 }
2176 return this_sad;
2177 }
2178 #endif
2179
2180 // Set up prediction 1-D reference set
2181 ref_buf = xd->plane[0].pre[0].buf - (bw >> 1);
2182 for (idx = 0; idx < search_width; idx += 16) {
2183 vpx_int_pro_row(&hbuf[idx], ref_buf, ref_stride, bh);
2184 ref_buf += 16;
2185 }
2186
2187 ref_buf = xd->plane[0].pre[0].buf - (bh >> 1) * ref_stride;
2188 for (idx = 0; idx < search_height; ++idx) {
2189 vbuf[idx] = vpx_int_pro_col(ref_buf, bw) >> norm_factor;
2190 ref_buf += ref_stride;
2191 }
2192
2193 // Set up src 1-D reference set
2194 for (idx = 0; idx < bw; idx += 16) {
2195 src_buf = x->plane[0].src.buf + idx;
2196 vpx_int_pro_row(&src_hbuf[idx], src_buf, src_stride, bh);
2197 }
2198
2199 src_buf = x->plane[0].src.buf;
2200 for (idx = 0; idx < bh; ++idx) {
2201 src_vbuf[idx] = vpx_int_pro_col(src_buf, bw) >> norm_factor;
2202 src_buf += src_stride;
2203 }
2204
2205 // Find the best match per 1-D search
2206 tmp_mv->col = vector_match(hbuf, src_hbuf, b_width_log2_lookup[bsize]);
2207 tmp_mv->row = vector_match(vbuf, src_vbuf, b_height_log2_lookup[bsize]);
2208
2209 this_mv = *tmp_mv;
2210 src_buf = x->plane[0].src.buf;
2211 ref_buf = xd->plane[0].pre[0].buf + this_mv.row * ref_stride + this_mv.col;
2212 best_sad = cpi->fn_ptr[bsize].sdf(src_buf, src_stride, ref_buf, ref_stride);
2213
2214 {
2215 const uint8_t *const pos[4] = {
2216 ref_buf - ref_stride,
2217 ref_buf - 1,
2218 ref_buf + 1,
2219 ref_buf + ref_stride,
2220 };
2221
2222 cpi->fn_ptr[bsize].sdx4df(src_buf, src_stride, pos, ref_stride, this_sad);
2223 }
2224
2225 for (idx = 0; idx < 4; ++idx) {
2226 if (this_sad[idx] < best_sad) {
2227 best_sad = this_sad[idx];
2228 tmp_mv->row = search_pos[idx].row + this_mv.row;
2229 tmp_mv->col = search_pos[idx].col + this_mv.col;
2230 }
2231 }
2232
2233 if (this_sad[0] < this_sad[3])
2234 this_mv.row -= 1;
2235 else
2236 this_mv.row += 1;
2237
2238 if (this_sad[1] < this_sad[2])
2239 this_mv.col -= 1;
2240 else
2241 this_mv.col += 1;
2242
2243 ref_buf = xd->plane[0].pre[0].buf + this_mv.row * ref_stride + this_mv.col;
2244
2245 tmp_sad = cpi->fn_ptr[bsize].sdf(src_buf, src_stride, ref_buf, ref_stride);
2246 if (best_sad > tmp_sad) {
2247 *tmp_mv = this_mv;
2248 best_sad = tmp_sad;
2249 }
2250
2251 tmp_mv->row *= 8;
2252 tmp_mv->col *= 8;
2253
2254 vp9_set_subpel_mv_search_range(&subpel_mv_limits, &x->mv_limits, ref_mv);
2255 clamp_mv(tmp_mv, subpel_mv_limits.col_min, subpel_mv_limits.col_max,
2256 subpel_mv_limits.row_min, subpel_mv_limits.row_max);
2257
2258 if (scaled_ref_frame) {
2259 int i;
2260 for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
2261 }
2262
2263 return best_sad;
2264 }
2265
2266 #if CONFIG_NON_GREEDY_MV
2267 // Runs sequence of diamond searches in smaller steps for RD.
2268 /* do_refine: If last step (1-away) of n-step search doesn't pick the center
2269 point as the best match, we will do a final 1-away diamond
2270 refining search */
vp9_full_pixel_diamond_new(const VP9_COMP * cpi,MACROBLOCK * x,MV * mvp_full,int step_param,double lambda,int do_refine,const vp9_variance_fn_ptr_t * fn_ptr,const int_mv * nb_full_mvs,int full_mv_num,MV * best_mv,double * best_mv_dist,double * best_mv_cost)2271 double vp9_full_pixel_diamond_new(const VP9_COMP *cpi, MACROBLOCK *x,
2272 MV *mvp_full, int step_param, double lambda,
2273 int do_refine,
2274 const vp9_variance_fn_ptr_t *fn_ptr,
2275 const int_mv *nb_full_mvs, int full_mv_num,
2276 MV *best_mv, double *best_mv_dist,
2277 double *best_mv_cost) {
2278 int n, num00 = 0;
2279 double thissme;
2280 double bestsme;
2281 const int further_steps = MAX_MVSEARCH_STEPS - 1 - step_param;
2282 vpx_clear_system_state();
2283 bestsme = vp9_diamond_search_sad_new(
2284 x, &cpi->ss_cfg, mvp_full, best_mv, best_mv_dist, best_mv_cost,
2285 step_param, lambda, &n, fn_ptr, nb_full_mvs, full_mv_num);
2286
2287 // If there won't be more n-step search, check to see if refining search is
2288 // needed.
2289 if (n > further_steps) do_refine = 0;
2290
2291 while (n < further_steps) {
2292 ++n;
2293 if (num00) {
2294 num00--;
2295 } else {
2296 MV temp_mv;
2297 double mv_dist;
2298 double mv_cost;
2299 thissme = vp9_diamond_search_sad_new(
2300 x, &cpi->ss_cfg, mvp_full, &temp_mv, &mv_dist, &mv_cost,
2301 step_param + n, lambda, &num00, fn_ptr, nb_full_mvs, full_mv_num);
2302 // check to see if refining search is needed.
2303 if (num00 > further_steps - n) do_refine = 0;
2304
2305 if (thissme < bestsme) {
2306 bestsme = thissme;
2307 *best_mv = temp_mv;
2308 *best_mv_dist = mv_dist;
2309 *best_mv_cost = mv_cost;
2310 }
2311 }
2312 }
2313
2314 // final 1-away diamond refining search
2315 if (do_refine) {
2316 const int search_range = 8;
2317 MV temp_mv = *best_mv;
2318 double mv_dist;
2319 double mv_cost;
2320 thissme = vp9_refining_search_sad_new(x, &temp_mv, &mv_dist, &mv_cost,
2321 lambda, search_range, fn_ptr,
2322 nb_full_mvs, full_mv_num);
2323 if (thissme < bestsme) {
2324 bestsme = thissme;
2325 *best_mv = temp_mv;
2326 *best_mv_dist = mv_dist;
2327 *best_mv_cost = mv_cost;
2328 }
2329 }
2330 return bestsme;
2331 }
2332 #endif // CONFIG_NON_GREEDY_MV
2333
2334 // Runs sequence of diamond searches in smaller steps for RD.
2335 /* do_refine: If last step (1-away) of n-step search doesn't pick the center
2336 point as the best match, we will do a final 1-away diamond
2337 refining search */
full_pixel_diamond(const VP9_COMP * cpi,MACROBLOCK * x,MV * mvp_full,int step_param,int sadpb,int further_steps,int do_refine,int * cost_list,const vp9_variance_fn_ptr_t * fn_ptr,const MV * ref_mv,MV * dst_mv)2338 static int full_pixel_diamond(const VP9_COMP *cpi, MACROBLOCK *x, MV *mvp_full,
2339 int step_param, int sadpb, int further_steps,
2340 int do_refine, int *cost_list,
2341 const vp9_variance_fn_ptr_t *fn_ptr,
2342 const MV *ref_mv, MV *dst_mv) {
2343 MV temp_mv;
2344 int thissme, n, num00 = 0;
2345 int bestsme = cpi->diamond_search_sad(x, &cpi->ss_cfg, mvp_full, &temp_mv,
2346 step_param, sadpb, &n, fn_ptr, ref_mv);
2347 if (bestsme < INT_MAX)
2348 bestsme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
2349 *dst_mv = temp_mv;
2350
2351 // If there won't be more n-step search, check to see if refining search is
2352 // needed.
2353 if (n > further_steps) do_refine = 0;
2354
2355 while (n < further_steps) {
2356 ++n;
2357
2358 if (num00) {
2359 num00--;
2360 } else {
2361 thissme = cpi->diamond_search_sad(x, &cpi->ss_cfg, mvp_full, &temp_mv,
2362 step_param + n, sadpb, &num00, fn_ptr,
2363 ref_mv);
2364 if (thissme < INT_MAX)
2365 thissme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
2366
2367 // check to see if refining search is needed.
2368 if (num00 > further_steps - n) do_refine = 0;
2369
2370 if (thissme < bestsme) {
2371 bestsme = thissme;
2372 *dst_mv = temp_mv;
2373 }
2374 }
2375 }
2376
2377 // final 1-away diamond refining search
2378 if (do_refine) {
2379 const int search_range = 8;
2380 MV best_mv = *dst_mv;
2381 thissme = vp9_refining_search_sad(x, &best_mv, sadpb, search_range, fn_ptr,
2382 ref_mv);
2383 if (thissme < INT_MAX)
2384 thissme = vp9_get_mvpred_var(x, &best_mv, ref_mv, fn_ptr, 1);
2385 if (thissme < bestsme) {
2386 bestsme = thissme;
2387 *dst_mv = best_mv;
2388 }
2389 }
2390
2391 // Return cost list.
2392 if (cost_list) {
2393 calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, dst_mv, cost_list);
2394 }
2395 return bestsme;
2396 }
2397
2398 #define MIN_RANGE 7
2399 #define MAX_RANGE 256
2400 #define MIN_INTERVAL 1
2401 // Runs an limited range exhaustive mesh search using a pattern set
2402 // according to the encode speed profile.
full_pixel_exhaustive(VP9_COMP * cpi,MACROBLOCK * x,MV * centre_mv_full,int sadpb,int * cost_list,const vp9_variance_fn_ptr_t * fn_ptr,const MV * ref_mv,MV * dst_mv)2403 static int full_pixel_exhaustive(VP9_COMP *cpi, MACROBLOCK *x,
2404 MV *centre_mv_full, int sadpb, int *cost_list,
2405 const vp9_variance_fn_ptr_t *fn_ptr,
2406 const MV *ref_mv, MV *dst_mv) {
2407 const SPEED_FEATURES *const sf = &cpi->sf;
2408 MV temp_mv = { centre_mv_full->row, centre_mv_full->col };
2409 MV f_ref_mv = { ref_mv->row >> 3, ref_mv->col >> 3 };
2410 int bestsme;
2411 int i;
2412 int interval = sf->mesh_patterns[0].interval;
2413 int range = sf->mesh_patterns[0].range;
2414 int baseline_interval_divisor;
2415
2416 // Trap illegal values for interval and range for this function.
2417 if ((range < MIN_RANGE) || (range > MAX_RANGE) || (interval < MIN_INTERVAL) ||
2418 (interval > range))
2419 return INT_MAX;
2420
2421 baseline_interval_divisor = range / interval;
2422
2423 // Check size of proposed first range against magnitude of the centre
2424 // value used as a starting point.
2425 range = VPXMAX(range, (5 * VPXMAX(abs(temp_mv.row), abs(temp_mv.col))) / 4);
2426 range = VPXMIN(range, MAX_RANGE);
2427 interval = VPXMAX(interval, range / baseline_interval_divisor);
2428
2429 // initial search
2430 bestsme = exhuastive_mesh_search(x, &f_ref_mv, &temp_mv, range, interval,
2431 sadpb, fn_ptr, &temp_mv);
2432
2433 if ((interval > MIN_INTERVAL) && (range > MIN_RANGE)) {
2434 // Progressive searches with range and step size decreasing each time
2435 // till we reach a step size of 1. Then break out.
2436 for (i = 1; i < MAX_MESH_STEP; ++i) {
2437 // First pass with coarser step and longer range
2438 bestsme = exhuastive_mesh_search(
2439 x, &f_ref_mv, &temp_mv, sf->mesh_patterns[i].range,
2440 sf->mesh_patterns[i].interval, sadpb, fn_ptr, &temp_mv);
2441
2442 if (sf->mesh_patterns[i].interval == 1) break;
2443 }
2444 }
2445
2446 if (bestsme < INT_MAX)
2447 bestsme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
2448 *dst_mv = temp_mv;
2449
2450 // Return cost list.
2451 if (cost_list) {
2452 calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, dst_mv, cost_list);
2453 }
2454 return bestsme;
2455 }
2456
2457 #if CONFIG_NON_GREEDY_MV
vp9_refining_search_sad_new(const MACROBLOCK * x,MV * best_full_mv,double * best_mv_dist,double * best_mv_cost,double lambda,int search_range,const vp9_variance_fn_ptr_t * fn_ptr,const int_mv * nb_full_mvs,int full_mv_num)2458 double vp9_refining_search_sad_new(const MACROBLOCK *x, MV *best_full_mv,
2459 double *best_mv_dist, double *best_mv_cost,
2460 double lambda, int search_range,
2461 const vp9_variance_fn_ptr_t *fn_ptr,
2462 const int_mv *nb_full_mvs, int full_mv_num) {
2463 const MACROBLOCKD *const xd = &x->e_mbd;
2464 const MV neighbors[4] = { { -1, 0 }, { 0, -1 }, { 0, 1 }, { 1, 0 } };
2465 const struct buf_2d *const what = &x->plane[0].src;
2466 const struct buf_2d *const in_what = &xd->plane[0].pre[0];
2467 const uint8_t *best_address = get_buf_from_mv(in_what, best_full_mv);
2468 double best_sad;
2469 int i, j;
2470 vpx_clear_system_state();
2471 *best_mv_dist =
2472 fn_ptr->sdf(what->buf, what->stride, best_address, in_what->stride);
2473 *best_mv_cost =
2474 vp9_nb_mvs_inconsistency(best_full_mv, nb_full_mvs, full_mv_num);
2475 best_sad = (*best_mv_dist) + lambda * (*best_mv_cost);
2476
2477 for (i = 0; i < search_range; i++) {
2478 int best_site = -1;
2479 const int all_in = ((best_full_mv->row - 1) > x->mv_limits.row_min) &
2480 ((best_full_mv->row + 1) < x->mv_limits.row_max) &
2481 ((best_full_mv->col - 1) > x->mv_limits.col_min) &
2482 ((best_full_mv->col + 1) < x->mv_limits.col_max);
2483
2484 if (all_in) {
2485 unsigned int sads[4];
2486 const uint8_t *const positions[4] = { best_address - in_what->stride,
2487 best_address - 1, best_address + 1,
2488 best_address + in_what->stride };
2489
2490 fn_ptr->sdx4df(what->buf, what->stride, positions, in_what->stride, sads);
2491
2492 for (j = 0; j < 4; ++j) {
2493 const MV mv = { best_full_mv->row + neighbors[j].row,
2494 best_full_mv->col + neighbors[j].col };
2495 const double mv_dist = sads[j];
2496 const double mv_cost =
2497 vp9_nb_mvs_inconsistency(&mv, nb_full_mvs, full_mv_num);
2498 const double thissad = mv_dist + lambda * mv_cost;
2499 if (thissad < best_sad) {
2500 best_sad = thissad;
2501 *best_mv_dist = mv_dist;
2502 *best_mv_cost = mv_cost;
2503 best_site = j;
2504 }
2505 }
2506 } else {
2507 for (j = 0; j < 4; ++j) {
2508 const MV mv = { best_full_mv->row + neighbors[j].row,
2509 best_full_mv->col + neighbors[j].col };
2510
2511 if (is_mv_in(&x->mv_limits, &mv)) {
2512 const double mv_dist =
2513 fn_ptr->sdf(what->buf, what->stride,
2514 get_buf_from_mv(in_what, &mv), in_what->stride);
2515 const double mv_cost =
2516 vp9_nb_mvs_inconsistency(&mv, nb_full_mvs, full_mv_num);
2517 const double thissad = mv_dist + lambda * mv_cost;
2518 if (thissad < best_sad) {
2519 best_sad = thissad;
2520 *best_mv_dist = mv_dist;
2521 *best_mv_cost = mv_cost;
2522 best_site = j;
2523 }
2524 }
2525 }
2526 }
2527
2528 if (best_site == -1) {
2529 break;
2530 } else {
2531 best_full_mv->row += neighbors[best_site].row;
2532 best_full_mv->col += neighbors[best_site].col;
2533 best_address = get_buf_from_mv(in_what, best_full_mv);
2534 }
2535 }
2536
2537 return best_sad;
2538 }
2539 #endif // CONFIG_NON_GREEDY_MV
2540
vp9_refining_search_sad(const MACROBLOCK * x,MV * ref_mv,int error_per_bit,int search_range,const vp9_variance_fn_ptr_t * fn_ptr,const MV * center_mv)2541 int vp9_refining_search_sad(const MACROBLOCK *x, MV *ref_mv, int error_per_bit,
2542 int search_range,
2543 const vp9_variance_fn_ptr_t *fn_ptr,
2544 const MV *center_mv) {
2545 const MACROBLOCKD *const xd = &x->e_mbd;
2546 const MV neighbors[4] = { { -1, 0 }, { 0, -1 }, { 0, 1 }, { 1, 0 } };
2547 const struct buf_2d *const what = &x->plane[0].src;
2548 const struct buf_2d *const in_what = &xd->plane[0].pre[0];
2549 const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
2550 const uint8_t *best_address = get_buf_from_mv(in_what, ref_mv);
2551 unsigned int best_sad =
2552 fn_ptr->sdf(what->buf, what->stride, best_address, in_what->stride) +
2553 mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
2554 int i, j;
2555
2556 for (i = 0; i < search_range; i++) {
2557 int best_site = -1;
2558 const int all_in = ((ref_mv->row - 1) > x->mv_limits.row_min) &
2559 ((ref_mv->row + 1) < x->mv_limits.row_max) &
2560 ((ref_mv->col - 1) > x->mv_limits.col_min) &
2561 ((ref_mv->col + 1) < x->mv_limits.col_max);
2562
2563 if (all_in) {
2564 unsigned int sads[4];
2565 const uint8_t *const positions[4] = { best_address - in_what->stride,
2566 best_address - 1, best_address + 1,
2567 best_address + in_what->stride };
2568
2569 fn_ptr->sdx4df(what->buf, what->stride, positions, in_what->stride, sads);
2570
2571 for (j = 0; j < 4; ++j) {
2572 if (sads[j] < best_sad) {
2573 const MV mv = { ref_mv->row + neighbors[j].row,
2574 ref_mv->col + neighbors[j].col };
2575 sads[j] += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
2576 if (sads[j] < best_sad) {
2577 best_sad = sads[j];
2578 best_site = j;
2579 }
2580 }
2581 }
2582 } else {
2583 for (j = 0; j < 4; ++j) {
2584 const MV mv = { ref_mv->row + neighbors[j].row,
2585 ref_mv->col + neighbors[j].col };
2586
2587 if (is_mv_in(&x->mv_limits, &mv)) {
2588 unsigned int sad =
2589 fn_ptr->sdf(what->buf, what->stride,
2590 get_buf_from_mv(in_what, &mv), in_what->stride);
2591 if (sad < best_sad) {
2592 sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
2593 if (sad < best_sad) {
2594 best_sad = sad;
2595 best_site = j;
2596 }
2597 }
2598 }
2599 }
2600 }
2601
2602 if (best_site == -1) {
2603 break;
2604 } else {
2605 ref_mv->row += neighbors[best_site].row;
2606 ref_mv->col += neighbors[best_site].col;
2607 best_address = get_buf_from_mv(in_what, ref_mv);
2608 }
2609 }
2610
2611 return best_sad;
2612 }
2613
2614 // This function is called when we do joint motion search in comp_inter_inter
2615 // mode.
vp9_refining_search_8p_c(const MACROBLOCK * x,MV * ref_mv,int error_per_bit,int search_range,const vp9_variance_fn_ptr_t * fn_ptr,const MV * center_mv,const uint8_t * second_pred)2616 int vp9_refining_search_8p_c(const MACROBLOCK *x, MV *ref_mv, int error_per_bit,
2617 int search_range,
2618 const vp9_variance_fn_ptr_t *fn_ptr,
2619 const MV *center_mv, const uint8_t *second_pred) {
2620 const MV neighbors[8] = { { -1, 0 }, { 0, -1 }, { 0, 1 }, { 1, 0 },
2621 { -1, -1 }, { 1, -1 }, { -1, 1 }, { 1, 1 } };
2622 const MACROBLOCKD *const xd = &x->e_mbd;
2623 const struct buf_2d *const what = &x->plane[0].src;
2624 const struct buf_2d *const in_what = &xd->plane[0].pre[0];
2625 const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
2626 unsigned int best_sad = INT_MAX;
2627 int i, j;
2628 clamp_mv(ref_mv, x->mv_limits.col_min, x->mv_limits.col_max,
2629 x->mv_limits.row_min, x->mv_limits.row_max);
2630 best_sad =
2631 fn_ptr->sdaf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv),
2632 in_what->stride, second_pred) +
2633 mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
2634
2635 for (i = 0; i < search_range; ++i) {
2636 int best_site = -1;
2637
2638 for (j = 0; j < 8; ++j) {
2639 const MV mv = { ref_mv->row + neighbors[j].row,
2640 ref_mv->col + neighbors[j].col };
2641
2642 if (is_mv_in(&x->mv_limits, &mv)) {
2643 unsigned int sad =
2644 fn_ptr->sdaf(what->buf, what->stride, get_buf_from_mv(in_what, &mv),
2645 in_what->stride, second_pred);
2646 if (sad < best_sad) {
2647 sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
2648 if (sad < best_sad) {
2649 best_sad = sad;
2650 best_site = j;
2651 }
2652 }
2653 }
2654 }
2655
2656 if (best_site == -1) {
2657 break;
2658 } else {
2659 ref_mv->row += neighbors[best_site].row;
2660 ref_mv->col += neighbors[best_site].col;
2661 }
2662 }
2663 return best_sad;
2664 }
2665
vp9_full_pixel_search(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,MV * mvp_full,int step_param,int search_method,int error_per_bit,int * cost_list,const MV * ref_mv,MV * tmp_mv,int var_max,int rd)2666 int vp9_full_pixel_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
2667 MV *mvp_full, int step_param, int search_method,
2668 int error_per_bit, int *cost_list, const MV *ref_mv,
2669 MV *tmp_mv, int var_max, int rd) {
2670 const SPEED_FEATURES *const sf = &cpi->sf;
2671 const SEARCH_METHODS method = (SEARCH_METHODS)search_method;
2672 vp9_variance_fn_ptr_t *fn_ptr = &cpi->fn_ptr[bsize];
2673 int var = 0;
2674 int run_exhaustive_search = 0;
2675
2676 if (cost_list) {
2677 cost_list[0] = INT_MAX;
2678 cost_list[1] = INT_MAX;
2679 cost_list[2] = INT_MAX;
2680 cost_list[3] = INT_MAX;
2681 cost_list[4] = INT_MAX;
2682 }
2683
2684 switch (method) {
2685 case FAST_DIAMOND:
2686 var = fast_dia_search(x, mvp_full, step_param, error_per_bit, 0,
2687 cost_list, fn_ptr, 1, ref_mv, tmp_mv);
2688 break;
2689 case FAST_HEX:
2690 var = fast_hex_search(x, mvp_full, step_param, error_per_bit, 0,
2691 cost_list, fn_ptr, 1, ref_mv, tmp_mv);
2692 break;
2693 case HEX:
2694 var = hex_search(x, mvp_full, step_param, error_per_bit, 1, cost_list,
2695 fn_ptr, 1, ref_mv, tmp_mv);
2696 break;
2697 case SQUARE:
2698 var = square_search(x, mvp_full, step_param, error_per_bit, 1, cost_list,
2699 fn_ptr, 1, ref_mv, tmp_mv);
2700 break;
2701 case BIGDIA:
2702 var = bigdia_search(x, mvp_full, step_param, error_per_bit, 1, cost_list,
2703 fn_ptr, 1, ref_mv, tmp_mv);
2704 break;
2705 case NSTEP:
2706 case MESH:
2707 var = full_pixel_diamond(cpi, x, mvp_full, step_param, error_per_bit,
2708 MAX_MVSEARCH_STEPS - 1 - step_param, 1,
2709 cost_list, fn_ptr, ref_mv, tmp_mv);
2710 break;
2711 default: assert(0 && "Unknown search method");
2712 }
2713
2714 if (method == NSTEP) {
2715 if (sf->exhaustive_searches_thresh < INT_MAX &&
2716 !cpi->rc.is_src_frame_alt_ref) {
2717 const int64_t exhuastive_thr =
2718 sf->exhaustive_searches_thresh >>
2719 (8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]));
2720 if (var > exhuastive_thr) run_exhaustive_search = 1;
2721 }
2722 } else if (method == MESH) {
2723 run_exhaustive_search = 1;
2724 }
2725
2726 if (run_exhaustive_search) {
2727 int var_ex;
2728 MV tmp_mv_ex;
2729 var_ex = full_pixel_exhaustive(cpi, x, tmp_mv, error_per_bit, cost_list,
2730 fn_ptr, ref_mv, &tmp_mv_ex);
2731 if (var_ex < var) {
2732 var = var_ex;
2733 *tmp_mv = tmp_mv_ex;
2734 }
2735 }
2736
2737 if (method != NSTEP && method != MESH && rd && var < var_max)
2738 var = vp9_get_mvpred_var(x, tmp_mv, ref_mv, fn_ptr, 1);
2739
2740 return var;
2741 }
2742
2743 // Note(yunqingwang): The following 2 functions are only used in the motion
2744 // vector unit test, which return extreme motion vectors allowed by the MV
2745 // limits.
2746 #define COMMON_MV_TEST \
2747 SETUP_SUBPEL_SEARCH; \
2748 \
2749 (void)error_per_bit; \
2750 (void)vfp; \
2751 (void)z; \
2752 (void)src_stride; \
2753 (void)y; \
2754 (void)y_stride; \
2755 (void)second_pred; \
2756 (void)w; \
2757 (void)h; \
2758 (void)offset; \
2759 (void)mvjcost; \
2760 (void)mvcost; \
2761 (void)sse1; \
2762 (void)distortion; \
2763 \
2764 (void)halfiters; \
2765 (void)quarteriters; \
2766 (void)eighthiters; \
2767 (void)whichdir; \
2768 (void)allow_hp; \
2769 (void)forced_stop; \
2770 (void)hstep; \
2771 (void)rr; \
2772 (void)rc; \
2773 \
2774 (void)tr; \
2775 (void)tc; \
2776 (void)sse; \
2777 (void)thismse; \
2778 (void)cost_list; \
2779 (void)use_accurate_subpel_search;
2780
2781 // Return the maximum MV.
vp9_return_max_sub_pixel_mv(const MACROBLOCK * x,MV * bestmv,const MV * ref_mv,int allow_hp,int error_per_bit,const vp9_variance_fn_ptr_t * vfp,int forced_stop,int iters_per_step,int * cost_list,int * mvjcost,int * mvcost[2],uint32_t * distortion,uint32_t * sse1,const uint8_t * second_pred,int w,int h,int use_accurate_subpel_search)2782 uint32_t vp9_return_max_sub_pixel_mv(
2783 const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
2784 int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
2785 int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
2786 uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
2787 int h, int use_accurate_subpel_search) {
2788 COMMON_MV_TEST;
2789
2790 (void)minr;
2791 (void)minc;
2792
2793 bestmv->row = maxr;
2794 bestmv->col = maxc;
2795 besterr = 0;
2796
2797 // In the sub-pel motion search, if hp is not used, then the last bit of mv
2798 // has to be 0.
2799 lower_mv_precision(bestmv, allow_hp && use_mv_hp(ref_mv));
2800
2801 return besterr;
2802 }
2803 // Return the minimum MV.
vp9_return_min_sub_pixel_mv(const MACROBLOCK * x,MV * bestmv,const MV * ref_mv,int allow_hp,int error_per_bit,const vp9_variance_fn_ptr_t * vfp,int forced_stop,int iters_per_step,int * cost_list,int * mvjcost,int * mvcost[2],uint32_t * distortion,uint32_t * sse1,const uint8_t * second_pred,int w,int h,int use_accurate_subpel_search)2804 uint32_t vp9_return_min_sub_pixel_mv(
2805 const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
2806 int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
2807 int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
2808 uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
2809 int h, int use_accurate_subpel_search) {
2810 COMMON_MV_TEST;
2811
2812 (void)maxr;
2813 (void)maxc;
2814
2815 bestmv->row = minr;
2816 bestmv->col = minc;
2817 besterr = 0;
2818
2819 // In the sub-pel motion search, if hp is not used, then the last bit of mv
2820 // has to be 0.
2821 lower_mv_precision(bestmv, allow_hp && use_mv_hp(ref_mv));
2822
2823 return besterr;
2824 }
2825