• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
vp9_set_mv_search_range(MvLimits * mv_limits,const MV * mv)32 void vp9_set_mv_search_range(MvLimits *mv_limits, const MV *mv) {
33   int col_min = (mv->col >> 3) - MAX_FULL_PEL_VAL + (mv->col & 7 ? 1 : 0);
34   int row_min = (mv->row >> 3) - MAX_FULL_PEL_VAL + (mv->row & 7 ? 1 : 0);
35   int col_max = (mv->col >> 3) + MAX_FULL_PEL_VAL;
36   int row_max = (mv->row >> 3) + MAX_FULL_PEL_VAL;
37 
38   col_min = VPXMAX(col_min, (MV_LOW >> 3) + 1);
39   row_min = VPXMAX(row_min, (MV_LOW >> 3) + 1);
40   col_max = VPXMIN(col_max, (MV_UPP >> 3) - 1);
41   row_max = VPXMIN(row_max, (MV_UPP >> 3) - 1);
42 
43   // Get intersection of UMV window and valid MV window to reduce # of checks
44   // in diamond search.
45   if (mv_limits->col_min < col_min) mv_limits->col_min = col_min;
46   if (mv_limits->col_max > col_max) mv_limits->col_max = col_max;
47   if (mv_limits->row_min < row_min) mv_limits->row_min = row_min;
48   if (mv_limits->row_max > row_max) mv_limits->row_max = row_max;
49 }
50 
vp9_set_subpel_mv_search_range(MvLimits * subpel_mv_limits,const MvLimits * umv_window_limits,const MV * ref_mv)51 void vp9_set_subpel_mv_search_range(MvLimits *subpel_mv_limits,
52                                     const MvLimits *umv_window_limits,
53                                     const MV *ref_mv) {
54   subpel_mv_limits->col_min = VPXMAX(umv_window_limits->col_min * 8,
55                                      ref_mv->col - MAX_FULL_PEL_VAL * 8);
56   subpel_mv_limits->col_max = VPXMIN(umv_window_limits->col_max * 8,
57                                      ref_mv->col + MAX_FULL_PEL_VAL * 8);
58   subpel_mv_limits->row_min = VPXMAX(umv_window_limits->row_min * 8,
59                                      ref_mv->row - MAX_FULL_PEL_VAL * 8);
60   subpel_mv_limits->row_max = VPXMIN(umv_window_limits->row_max * 8,
61                                      ref_mv->row + MAX_FULL_PEL_VAL * 8);
62 
63   subpel_mv_limits->col_min = VPXMAX(MV_LOW + 1, subpel_mv_limits->col_min);
64   subpel_mv_limits->col_max = VPXMIN(MV_UPP - 1, subpel_mv_limits->col_max);
65   subpel_mv_limits->row_min = VPXMAX(MV_LOW + 1, subpel_mv_limits->row_min);
66   subpel_mv_limits->row_max = VPXMIN(MV_UPP - 1, subpel_mv_limits->row_max);
67 }
68 
vp9_init_search_range(int size)69 int vp9_init_search_range(int size) {
70   int sr = 0;
71   // Minimum search size no matter what the passed in value.
72   size = VPXMAX(16, size);
73 
74   while ((size << sr) < MAX_FULL_PEL_VAL) sr++;
75 
76   sr = VPXMIN(sr, MAX_MVSEARCH_STEPS - 2);
77   return sr;
78 }
79 
mv_cost(const MV * mv,const int * joint_cost,int * const comp_cost[2])80 static INLINE int mv_cost(const MV *mv, const int *joint_cost,
81                           int *const comp_cost[2]) {
82   assert(mv->row >= -MV_MAX && mv->row < MV_MAX);
83   assert(mv->col >= -MV_MAX && mv->col < MV_MAX);
84   return joint_cost[vp9_get_mv_joint(mv)] + comp_cost[0][mv->row] +
85          comp_cost[1][mv->col];
86 }
87 
vp9_mv_bit_cost(const MV * mv,const MV * ref,const int * mvjcost,int * mvcost[2],int weight)88 int vp9_mv_bit_cost(const MV *mv, const MV *ref, const int *mvjcost,
89                     int *mvcost[2], int weight) {
90   const MV diff = { mv->row - ref->row, mv->col - ref->col };
91   return ROUND_POWER_OF_TWO(mv_cost(&diff, mvjcost, mvcost) * weight, 7);
92 }
93 
94 #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)95 static int mv_err_cost(const MV *mv, const MV *ref, const int *mvjcost,
96                        int *mvcost[2], int error_per_bit) {
97   if (mvcost) {
98     const MV diff = { mv->row - ref->row, mv->col - ref->col };
99     return (int)ROUND64_POWER_OF_TWO(
100         (int64_t)mv_cost(&diff, mvjcost, mvcost) * error_per_bit,
101         RDDIV_BITS + VP9_PROB_COST_SHIFT - RD_EPB_SHIFT +
102             PIXEL_TRANSFORM_ERROR_SCALE);
103   }
104   return 0;
105 }
106 
mvsad_err_cost(const MACROBLOCK * x,const MV * mv,const MV * ref,int sad_per_bit)107 static int mvsad_err_cost(const MACROBLOCK *x, const MV *mv, const MV *ref,
108                           int sad_per_bit) {
109   const MV diff = { mv->row - ref->row, mv->col - ref->col };
110   return ROUND_POWER_OF_TWO(
111       (unsigned)mv_cost(&diff, x->nmvjointsadcost, x->nmvsadcost) * sad_per_bit,
112       VP9_PROB_COST_SHIFT);
113 }
114 
vp9_init_dsmotion_compensation(search_site_config * cfg,int stride)115 void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride) {
116   int len;
117   int ss_count = 0;
118 
119   for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
120     // Generate offsets for 4 search sites per step.
121     const MV ss_mvs[] = { { -len, 0 }, { len, 0 }, { 0, -len }, { 0, len } };
122     int i;
123     for (i = 0; i < 4; ++i, ++ss_count) {
124       cfg->ss_mv[ss_count] = ss_mvs[i];
125       cfg->ss_os[ss_count] = ss_mvs[i].row * stride + ss_mvs[i].col;
126     }
127   }
128 
129   cfg->searches_per_step = 4;
130   cfg->total_steps = ss_count / cfg->searches_per_step;
131 }
132 
vp9_init3smotion_compensation(search_site_config * cfg,int stride)133 void vp9_init3smotion_compensation(search_site_config *cfg, int stride) {
134   int len;
135   int ss_count = 0;
136 
137   for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
138     // Generate offsets for 8 search sites per step.
139     const MV ss_mvs[8] = { { -len, 0 },   { len, 0 },     { 0, -len },
140                            { 0, len },    { -len, -len }, { -len, len },
141                            { len, -len }, { len, len } };
142     int i;
143     for (i = 0; i < 8; ++i, ++ss_count) {
144       cfg->ss_mv[ss_count] = ss_mvs[i];
145       cfg->ss_os[ss_count] = ss_mvs[i].row * stride + ss_mvs[i].col;
146     }
147   }
148 
149   cfg->searches_per_step = 8;
150   cfg->total_steps = ss_count / cfg->searches_per_step;
151 }
152 
153 // convert motion vector component to offset for sv[a]f calc
sp(int x)154 static INLINE int sp(int x) { return x & 7; }
155 
pre(const uint8_t * buf,int stride,int r,int c)156 static INLINE const uint8_t *pre(const uint8_t *buf, int stride, int r, int c) {
157   return &buf[(r >> 3) * stride + (c >> 3)];
158 }
159 
160 #if CONFIG_VP9_HIGHBITDEPTH
161 /* checks if (r, c) has better score than previous best */
162 #define CHECK_BETTER(v, r, c)                                                \
163   if (c >= minc && c <= maxc && r >= minr && r <= maxr) {                    \
164     int64_t tmpmse;                                                          \
165     const MV mv = { r, c };                                                  \
166     const MV ref_mv = { rr, rc };                                            \
167     if (second_pred == NULL) {                                               \
168       thismse = vfp->svf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z,  \
169                          src_stride, &sse);                                  \
170     } else {                                                                 \
171       thismse = vfp->svaf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z, \
172                           src_stride, &sse, second_pred);                    \
173     }                                                                        \
174     tmpmse = thismse;                                                        \
175     tmpmse += mv_err_cost(&mv, &ref_mv, mvjcost, mvcost, error_per_bit);     \
176     if (tmpmse >= INT_MAX) {                                                 \
177       v = INT_MAX;                                                           \
178     } else if ((v = (uint32_t)tmpmse) < besterr) {                           \
179       besterr = v;                                                           \
180       br = r;                                                                \
181       bc = c;                                                                \
182       *distortion = thismse;                                                 \
183       *sse1 = sse;                                                           \
184     }                                                                        \
185   } else {                                                                   \
186     v = INT_MAX;                                                             \
187   }
188 #else
189 /* checks if (r, c) has better score than previous best */
190 #define CHECK_BETTER(v, r, c)                                                \
191   if (c >= minc && c <= maxc && r >= minr && r <= maxr) {                    \
192     const MV mv = { r, c };                                                  \
193     const MV ref_mv = { rr, rc };                                            \
194     if (second_pred == NULL)                                                 \
195       thismse = vfp->svf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z,  \
196                          src_stride, &sse);                                  \
197     else                                                                     \
198       thismse = vfp->svaf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z, \
199                           src_stride, &sse, second_pred);                    \
200     if ((v = mv_err_cost(&mv, &ref_mv, mvjcost, mvcost, error_per_bit) +     \
201              thismse) < besterr) {                                           \
202       besterr = v;                                                           \
203       br = r;                                                                \
204       bc = c;                                                                \
205       *distortion = thismse;                                                 \
206       *sse1 = sse;                                                           \
207     }                                                                        \
208   } else {                                                                   \
209     v = INT_MAX;                                                             \
210   }
211 
212 #endif
213 #define FIRST_LEVEL_CHECKS                                       \
214   {                                                              \
215     unsigned int left, right, up, down, diag;                    \
216     CHECK_BETTER(left, tr, tc - hstep);                          \
217     CHECK_BETTER(right, tr, tc + hstep);                         \
218     CHECK_BETTER(up, tr - hstep, tc);                            \
219     CHECK_BETTER(down, tr + hstep, tc);                          \
220     whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);     \
221     switch (whichdir) {                                          \
222       case 0: CHECK_BETTER(diag, tr - hstep, tc - hstep); break; \
223       case 1: CHECK_BETTER(diag, tr - hstep, tc + hstep); break; \
224       case 2: CHECK_BETTER(diag, tr + hstep, tc - hstep); break; \
225       case 3: CHECK_BETTER(diag, tr + hstep, tc + hstep); break; \
226     }                                                            \
227   }
228 
229 #define SECOND_LEVEL_CHECKS                                       \
230   {                                                               \
231     int kr, kc;                                                   \
232     unsigned int second;                                          \
233     if (tr != br && tc != bc) {                                   \
234       kr = br - tr;                                               \
235       kc = bc - tc;                                               \
236       CHECK_BETTER(second, tr + kr, tc + 2 * kc);                 \
237       CHECK_BETTER(second, tr + 2 * kr, tc + kc);                 \
238     } else if (tr == br && tc != bc) {                            \
239       kc = bc - tc;                                               \
240       CHECK_BETTER(second, tr + hstep, tc + 2 * kc);              \
241       CHECK_BETTER(second, tr - hstep, tc + 2 * kc);              \
242       switch (whichdir) {                                         \
243         case 0:                                                   \
244         case 1: CHECK_BETTER(second, tr + hstep, tc + kc); break; \
245         case 2:                                                   \
246         case 3: CHECK_BETTER(second, tr - hstep, tc + kc); break; \
247       }                                                           \
248     } else if (tr != br && tc == bc) {                            \
249       kr = br - tr;                                               \
250       CHECK_BETTER(second, tr + 2 * kr, tc + hstep);              \
251       CHECK_BETTER(second, tr + 2 * kr, tc - hstep);              \
252       switch (whichdir) {                                         \
253         case 0:                                                   \
254         case 2: CHECK_BETTER(second, tr + kr, tc + hstep); break; \
255         case 1:                                                   \
256         case 3: CHECK_BETTER(second, tr + kr, tc - hstep); break; \
257       }                                                           \
258     }                                                             \
259   }
260 
261 #define SETUP_SUBPEL_SEARCH                                                 \
262   const uint8_t *const z = x->plane[0].src.buf;                             \
263   const int src_stride = x->plane[0].src.stride;                            \
264   const MACROBLOCKD *xd = &x->e_mbd;                                        \
265   unsigned int besterr = UINT_MAX;                                          \
266   unsigned int sse;                                                         \
267   unsigned int whichdir;                                                    \
268   int thismse;                                                              \
269   const unsigned int halfiters = iters_per_step;                            \
270   const unsigned int quarteriters = iters_per_step;                         \
271   const unsigned int eighthiters = iters_per_step;                          \
272   const int y_stride = xd->plane[0].pre[0].stride;                          \
273   const int offset = bestmv->row * y_stride + bestmv->col;                  \
274   const uint8_t *const y = xd->plane[0].pre[0].buf;                         \
275                                                                             \
276   int rr = ref_mv->row;                                                     \
277   int rc = ref_mv->col;                                                     \
278   int br = bestmv->row * 8;                                                 \
279   int bc = bestmv->col * 8;                                                 \
280   int hstep = 4;                                                            \
281   int minc, maxc, minr, maxr;                                               \
282   int tr = br;                                                              \
283   int tc = bc;                                                              \
284   MvLimits subpel_mv_limits;                                                \
285                                                                             \
286   vp9_set_subpel_mv_search_range(&subpel_mv_limits, &x->mv_limits, ref_mv); \
287   minc = subpel_mv_limits.col_min;                                          \
288   maxc = subpel_mv_limits.col_max;                                          \
289   minr = subpel_mv_limits.row_min;                                          \
290   maxr = subpel_mv_limits.row_max;                                          \
291                                                                             \
292   bestmv->row *= 8;                                                         \
293   bestmv->col *= 8;
294 
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)295 static unsigned int setup_center_error(
296     const MACROBLOCKD *xd, const MV *bestmv, const MV *ref_mv,
297     int error_per_bit, const vp9_variance_fn_ptr_t *vfp,
298     const uint8_t *const src, const int src_stride, const uint8_t *const y,
299     int y_stride, const uint8_t *second_pred, int w, int h, int offset,
300     int *mvjcost, int *mvcost[2], uint32_t *sse1, uint32_t *distortion) {
301 #if CONFIG_VP9_HIGHBITDEPTH
302   uint64_t besterr;
303   if (second_pred != NULL) {
304     if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
305       DECLARE_ALIGNED(16, uint16_t, comp_pred16[64 * 64]);
306       vpx_highbd_comp_avg_pred(comp_pred16, CONVERT_TO_SHORTPTR(second_pred), w,
307                                h, CONVERT_TO_SHORTPTR(y + offset), y_stride);
308       besterr =
309           vfp->vf(CONVERT_TO_BYTEPTR(comp_pred16), w, src, src_stride, sse1);
310     } else {
311       DECLARE_ALIGNED(16, uint8_t, comp_pred[64 * 64]);
312       vpx_comp_avg_pred(comp_pred, second_pred, w, h, y + offset, y_stride);
313       besterr = vfp->vf(comp_pred, w, src, src_stride, sse1);
314     }
315   } else {
316     besterr = vfp->vf(y + offset, y_stride, src, src_stride, sse1);
317   }
318   *distortion = (uint32_t)besterr;
319   besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
320   if (besterr >= UINT_MAX) return UINT_MAX;
321   return (uint32_t)besterr;
322 #else
323   uint32_t besterr;
324   (void)xd;
325   if (second_pred != NULL) {
326     DECLARE_ALIGNED(16, uint8_t, comp_pred[64 * 64]);
327     vpx_comp_avg_pred(comp_pred, second_pred, w, h, y + offset, y_stride);
328     besterr = vfp->vf(comp_pred, w, src, src_stride, sse1);
329   } else {
330     besterr = vfp->vf(y + offset, y_stride, src, src_stride, sse1);
331   }
332   *distortion = besterr;
333   besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
334   return besterr;
335 #endif  // CONFIG_VP9_HIGHBITDEPTH
336 }
337 
divide_and_round(const int64_t n,const int64_t d)338 static INLINE int64_t divide_and_round(const int64_t n, const int64_t d) {
339   return ((n < 0) ^ (d < 0)) ? ((n - d / 2) / d) : ((n + d / 2) / d);
340 }
341 
is_cost_list_wellbehaved(int * cost_list)342 static INLINE int is_cost_list_wellbehaved(int *cost_list) {
343   return cost_list[0] < cost_list[1] && cost_list[0] < cost_list[2] &&
344          cost_list[0] < cost_list[3] && cost_list[0] < cost_list[4];
345 }
346 
347 // Returns surface minima estimate at given precision in 1/2^n bits.
348 // Assume a model for the cost surface: S = A(x - x0)^2 + B(y - y0)^2 + C
349 // For a given set of costs S0, S1, S2, S3, S4 at points
350 // (y, x) = (0, 0), (0, -1), (1, 0), (0, 1) and (-1, 0) respectively,
351 // the solution for the location of the minima (x0, y0) is given by:
352 // x0 = 1/2 (S1 - S3)/(S1 + S3 - 2*S0),
353 // y0 = 1/2 (S4 - S2)/(S4 + S2 - 2*S0).
354 // The code below is an integerized version of that.
get_cost_surf_min(int * cost_list,int * ir,int * ic,int bits)355 static void get_cost_surf_min(int *cost_list, int *ir, int *ic, int bits) {
356   const int64_t x0 = (int64_t)cost_list[1] - cost_list[3];
357   const int64_t y0 = cost_list[1] - 2 * (int64_t)cost_list[0] + cost_list[3];
358   const int64_t x1 = (int64_t)cost_list[4] - cost_list[2];
359   const int64_t y1 = cost_list[4] - 2 * (int64_t)cost_list[0] + cost_list[2];
360   const int b = 1 << (bits - 1);
361   *ic = (int)divide_and_round(x0 * b, y0);
362   *ir = (int)divide_and_round(x1 * b, y1);
363 }
364 
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)365 uint32_t vp9_skip_sub_pixel_tree(
366     const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
367     int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
368     int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
369     uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
370     int h, int use_accurate_subpel_search) {
371   SETUP_SUBPEL_SEARCH;
372   besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp, z,
373                                src_stride, y, y_stride, second_pred, w, h,
374                                offset, mvjcost, mvcost, sse1, distortion);
375   (void)halfiters;
376   (void)quarteriters;
377   (void)eighthiters;
378   (void)whichdir;
379   (void)allow_hp;
380   (void)forced_stop;
381   (void)hstep;
382   (void)rr;
383   (void)rc;
384   (void)minr;
385   (void)minc;
386   (void)maxr;
387   (void)maxc;
388   (void)tr;
389   (void)tc;
390   (void)sse;
391   (void)thismse;
392   (void)cost_list;
393   (void)use_accurate_subpel_search;
394 
395   return besterr;
396 }
397 
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)398 uint32_t vp9_find_best_sub_pixel_tree_pruned_evenmore(
399     const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
400     int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
401     int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
402     uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
403     int h, int use_accurate_subpel_search) {
404   SETUP_SUBPEL_SEARCH;
405   besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp, z,
406                                src_stride, y, y_stride, second_pred, w, h,
407                                offset, mvjcost, mvcost, sse1, distortion);
408   (void)halfiters;
409   (void)quarteriters;
410   (void)eighthiters;
411   (void)whichdir;
412   (void)allow_hp;
413   (void)forced_stop;
414   (void)hstep;
415   (void)use_accurate_subpel_search;
416 
417   if (cost_list && cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
418       cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
419       cost_list[4] != INT_MAX && is_cost_list_wellbehaved(cost_list)) {
420     int ir, ic;
421     unsigned int minpt = INT_MAX;
422     get_cost_surf_min(cost_list, &ir, &ic, 2);
423     if (ir != 0 || ic != 0) {
424       CHECK_BETTER(minpt, tr + 2 * ir, tc + 2 * ic);
425     }
426   } else {
427     FIRST_LEVEL_CHECKS;
428     if (halfiters > 1) {
429       SECOND_LEVEL_CHECKS;
430     }
431 
432     tr = br;
433     tc = bc;
434 
435     // Each subsequent iteration checks at least one point in common with
436     // the last iteration could be 2 ( if diag selected) 1/4 pel
437     // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
438     if (forced_stop != 2) {
439       hstep >>= 1;
440       FIRST_LEVEL_CHECKS;
441       if (quarteriters > 1) {
442         SECOND_LEVEL_CHECKS;
443       }
444     }
445   }
446 
447   tr = br;
448   tc = bc;
449 
450   if (allow_hp && use_mv_hp(ref_mv) && forced_stop == 0) {
451     hstep >>= 1;
452     FIRST_LEVEL_CHECKS;
453     if (eighthiters > 1) {
454       SECOND_LEVEL_CHECKS;
455     }
456   }
457 
458   bestmv->row = br;
459   bestmv->col = bc;
460 
461   return besterr;
462 }
463 
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)464 uint32_t vp9_find_best_sub_pixel_tree_pruned_more(
465     const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
466     int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
467     int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
468     uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
469     int h, int use_accurate_subpel_search) {
470   SETUP_SUBPEL_SEARCH;
471   (void)use_accurate_subpel_search;
472 
473   besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp, z,
474                                src_stride, y, y_stride, second_pred, w, h,
475                                offset, mvjcost, mvcost, sse1, distortion);
476   if (cost_list && cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
477       cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
478       cost_list[4] != INT_MAX && is_cost_list_wellbehaved(cost_list)) {
479     unsigned int minpt;
480     int ir, ic;
481     get_cost_surf_min(cost_list, &ir, &ic, 1);
482     if (ir != 0 || ic != 0) {
483       CHECK_BETTER(minpt, tr + ir * hstep, tc + ic * hstep);
484     }
485   } else {
486     FIRST_LEVEL_CHECKS;
487     if (halfiters > 1) {
488       SECOND_LEVEL_CHECKS;
489     }
490   }
491 
492   // Each subsequent iteration checks at least one point in common with
493   // the last iteration could be 2 ( if diag selected) 1/4 pel
494 
495   // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
496   if (forced_stop != 2) {
497     tr = br;
498     tc = bc;
499     hstep >>= 1;
500     FIRST_LEVEL_CHECKS;
501     if (quarteriters > 1) {
502       SECOND_LEVEL_CHECKS;
503     }
504   }
505 
506   if (allow_hp && use_mv_hp(ref_mv) && forced_stop == 0) {
507     tr = br;
508     tc = bc;
509     hstep >>= 1;
510     FIRST_LEVEL_CHECKS;
511     if (eighthiters > 1) {
512       SECOND_LEVEL_CHECKS;
513     }
514   }
515   // These lines insure static analysis doesn't warn that
516   // tr and tc aren't used after the above point.
517   (void)tr;
518   (void)tc;
519 
520   bestmv->row = br;
521   bestmv->col = bc;
522 
523   return besterr;
524 }
525 
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)526 uint32_t vp9_find_best_sub_pixel_tree_pruned(
527     const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
528     int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
529     int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
530     uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
531     int h, int use_accurate_subpel_search) {
532   SETUP_SUBPEL_SEARCH;
533   (void)use_accurate_subpel_search;
534 
535   besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp, z,
536                                src_stride, y, y_stride, second_pred, w, h,
537                                offset, mvjcost, mvcost, sse1, distortion);
538   if (cost_list && cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
539       cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
540       cost_list[4] != INT_MAX) {
541     unsigned int left, right, up, down, diag;
542     whichdir = (cost_list[1] < cost_list[3] ? 0 : 1) +
543                (cost_list[2] < cost_list[4] ? 0 : 2);
544     switch (whichdir) {
545       case 0:
546         CHECK_BETTER(left, tr, tc - hstep);
547         CHECK_BETTER(down, tr + hstep, tc);
548         CHECK_BETTER(diag, tr + hstep, tc - hstep);
549         break;
550       case 1:
551         CHECK_BETTER(right, tr, tc + hstep);
552         CHECK_BETTER(down, tr + hstep, tc);
553         CHECK_BETTER(diag, tr + hstep, tc + hstep);
554         break;
555       case 2:
556         CHECK_BETTER(left, tr, tc - hstep);
557         CHECK_BETTER(up, tr - hstep, tc);
558         CHECK_BETTER(diag, tr - hstep, tc - hstep);
559         break;
560       case 3:
561         CHECK_BETTER(right, tr, tc + hstep);
562         CHECK_BETTER(up, tr - hstep, tc);
563         CHECK_BETTER(diag, tr - hstep, tc + hstep);
564         break;
565     }
566   } else {
567     FIRST_LEVEL_CHECKS;
568     if (halfiters > 1) {
569       SECOND_LEVEL_CHECKS;
570     }
571   }
572 
573   tr = br;
574   tc = bc;
575 
576   // Each subsequent iteration checks at least one point in common with
577   // the last iteration could be 2 ( if diag selected) 1/4 pel
578 
579   // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
580   if (forced_stop != 2) {
581     hstep >>= 1;
582     FIRST_LEVEL_CHECKS;
583     if (quarteriters > 1) {
584       SECOND_LEVEL_CHECKS;
585     }
586     tr = br;
587     tc = bc;
588   }
589 
590   if (allow_hp && use_mv_hp(ref_mv) && forced_stop == 0) {
591     hstep >>= 1;
592     FIRST_LEVEL_CHECKS;
593     if (eighthiters > 1) {
594       SECOND_LEVEL_CHECKS;
595     }
596     tr = br;
597     tc = bc;
598   }
599   // These lines insure static analysis doesn't warn that
600   // tr and tc aren't used after the above point.
601   (void)tr;
602   (void)tc;
603 
604   bestmv->row = br;
605   bestmv->col = bc;
606 
607   return besterr;
608 }
609 
610 /* clang-format off */
611 static const MV search_step_table[12] = {
612   // left, right, up, down
613   { 0, -4 }, { 0, 4 }, { -4, 0 }, { 4, 0 },
614   { 0, -2 }, { 0, 2 }, { -2, 0 }, { 2, 0 },
615   { 0, -1 }, { 0, 1 }, { -1, 0 }, { 1, 0 }
616 };
617 /* clang-format on */
618 
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)619 static int accurate_sub_pel_search(
620     const MACROBLOCKD *xd, const MV *this_mv, const struct scale_factors *sf,
621     const InterpKernel *kernel, const vp9_variance_fn_ptr_t *vfp,
622     const uint8_t *const src_address, const int src_stride,
623     const uint8_t *const pre_address, int y_stride, const uint8_t *second_pred,
624     int w, int h, uint32_t *sse) {
625 #if CONFIG_VP9_HIGHBITDEPTH
626   uint64_t besterr;
627   assert(sf->x_step_q4 == 16 && sf->y_step_q4 == 16);
628   assert(w != 0 && h != 0);
629   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
630     DECLARE_ALIGNED(16, uint16_t, pred16[64 * 64]);
631     vp9_highbd_build_inter_predictor(CONVERT_TO_SHORTPTR(pre_address), y_stride,
632                                      pred16, w, this_mv, sf, w, h, 0, kernel,
633                                      MV_PRECISION_Q3, 0, 0, xd->bd);
634     if (second_pred != NULL) {
635       DECLARE_ALIGNED(16, uint16_t, comp_pred16[64 * 64]);
636       vpx_highbd_comp_avg_pred(comp_pred16, CONVERT_TO_SHORTPTR(second_pred), w,
637                                h, pred16, w);
638       besterr = vfp->vf(CONVERT_TO_BYTEPTR(comp_pred16), w, src_address,
639                         src_stride, sse);
640     } else {
641       besterr =
642           vfp->vf(CONVERT_TO_BYTEPTR(pred16), w, src_address, src_stride, sse);
643     }
644   } else {
645     DECLARE_ALIGNED(16, uint8_t, pred[64 * 64]);
646     vp9_build_inter_predictor(pre_address, y_stride, pred, w, this_mv, sf, w, h,
647                               0, kernel, MV_PRECISION_Q3, 0, 0);
648     if (second_pred != NULL) {
649       DECLARE_ALIGNED(16, uint8_t, comp_pred[64 * 64]);
650       vpx_comp_avg_pred(comp_pred, second_pred, w, h, pred, w);
651       besterr = vfp->vf(comp_pred, w, src_address, src_stride, sse);
652     } else {
653       besterr = vfp->vf(pred, w, src_address, src_stride, sse);
654     }
655   }
656   if (besterr >= UINT_MAX) return UINT_MAX;
657   return (int)besterr;
658 #else
659   int besterr;
660   DECLARE_ALIGNED(16, uint8_t, pred[64 * 64]);
661   assert(sf->x_step_q4 == 16 && sf->y_step_q4 == 16);
662   assert(w != 0 && h != 0);
663   (void)xd;
664 
665   vp9_build_inter_predictor(pre_address, y_stride, pred, w, this_mv, sf, w, h,
666                             0, kernel, MV_PRECISION_Q3, 0, 0);
667   if (second_pred != NULL) {
668     DECLARE_ALIGNED(16, uint8_t, comp_pred[64 * 64]);
669     vpx_comp_avg_pred(comp_pred, second_pred, w, h, pred, w);
670     besterr = vfp->vf(comp_pred, w, src_address, src_stride, sse);
671   } else {
672     besterr = vfp->vf(pred, w, src_address, src_stride, sse);
673   }
674   return besterr;
675 #endif  // CONFIG_VP9_HIGHBITDEPTH
676 }
677 
678 // TODO(yunqing): this part can be further refactored.
679 #if CONFIG_VP9_HIGHBITDEPTH
680 /* checks if (r, c) has better score than previous best */
681 #define CHECK_BETTER1(v, r, c)                                                 \
682   if (c >= minc && c <= maxc && r >= minr && r <= maxr) {                      \
683     int64_t tmpmse;                                                            \
684     const MV mv = { r, c };                                                    \
685     const MV ref_mv = { rr, rc };                                              \
686     thismse =                                                                  \
687         accurate_sub_pel_search(xd, &mv, x->me_sf, kernel, vfp, z, src_stride, \
688                                 y, y_stride, second_pred, w, h, &sse);         \
689     tmpmse = thismse;                                                          \
690     tmpmse += mv_err_cost(&mv, &ref_mv, mvjcost, mvcost, error_per_bit);       \
691     if (tmpmse >= INT_MAX) {                                                   \
692       v = INT_MAX;                                                             \
693     } else if ((v = (uint32_t)tmpmse) < besterr) {                             \
694       besterr = v;                                                             \
695       br = r;                                                                  \
696       bc = c;                                                                  \
697       *distortion = thismse;                                                   \
698       *sse1 = sse;                                                             \
699     }                                                                          \
700   } else {                                                                     \
701     v = INT_MAX;                                                               \
702   }
703 #else
704 /* checks if (r, c) has better score than previous best */
705 #define CHECK_BETTER1(v, r, c)                                                 \
706   if (c >= minc && c <= maxc && r >= minr && r <= maxr) {                      \
707     const MV mv = { r, c };                                                    \
708     const MV ref_mv = { rr, rc };                                              \
709     thismse =                                                                  \
710         accurate_sub_pel_search(xd, &mv, x->me_sf, kernel, vfp, z, src_stride, \
711                                 y, y_stride, second_pred, w, h, &sse);         \
712     if ((v = mv_err_cost(&mv, &ref_mv, mvjcost, mvcost, error_per_bit) +       \
713              thismse) < besterr) {                                             \
714       besterr = v;                                                             \
715       br = r;                                                                  \
716       bc = c;                                                                  \
717       *distortion = thismse;                                                   \
718       *sse1 = sse;                                                             \
719     }                                                                          \
720   } else {                                                                     \
721     v = INT_MAX;                                                               \
722   }
723 
724 #endif
725 
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)726 uint32_t vp9_find_best_sub_pixel_tree(
727     const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
728     int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
729     int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
730     uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
731     int h, int use_accurate_subpel_search) {
732   const uint8_t *const z = x->plane[0].src.buf;
733   const uint8_t *const src_address = z;
734   const int src_stride = x->plane[0].src.stride;
735   const MACROBLOCKD *xd = &x->e_mbd;
736   unsigned int besterr = UINT_MAX;
737   unsigned int sse;
738   int thismse;
739   const int y_stride = xd->plane[0].pre[0].stride;
740   const int offset = bestmv->row * y_stride + bestmv->col;
741   const uint8_t *const y = xd->plane[0].pre[0].buf;
742 
743   int rr = ref_mv->row;
744   int rc = ref_mv->col;
745   int br = bestmv->row * 8;
746   int bc = bestmv->col * 8;
747   int hstep = 4;
748   int iter, round = 3 - forced_stop;
749 
750   int minc, maxc, minr, maxr;
751   int tr = br;
752   int tc = bc;
753   const MV *search_step = search_step_table;
754   int idx, best_idx = -1;
755   unsigned int cost_array[5];
756   int kr, kc;
757   MvLimits subpel_mv_limits;
758 
759   // TODO(yunqing): need to add 4-tap filter optimization to speed up the
760   // encoder.
761   const InterpKernel *kernel =
762       (use_accurate_subpel_search > 0)
763           ? ((use_accurate_subpel_search == USE_4_TAPS)
764                  ? vp9_filter_kernels[FOURTAP]
765                  : ((use_accurate_subpel_search == USE_8_TAPS)
766                         ? vp9_filter_kernels[EIGHTTAP]
767                         : vp9_filter_kernels[EIGHTTAP_SHARP]))
768           : vp9_filter_kernels[BILINEAR];
769 
770   vp9_set_subpel_mv_search_range(&subpel_mv_limits, &x->mv_limits, ref_mv);
771   minc = subpel_mv_limits.col_min;
772   maxc = subpel_mv_limits.col_max;
773   minr = subpel_mv_limits.row_min;
774   maxr = subpel_mv_limits.row_max;
775 
776   if (!(allow_hp && use_mv_hp(ref_mv)))
777     if (round == 3) round = 2;
778 
779   bestmv->row *= 8;
780   bestmv->col *= 8;
781 
782   besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp, z,
783                                src_stride, y, y_stride, second_pred, w, h,
784                                offset, mvjcost, mvcost, sse1, distortion);
785 
786   (void)cost_list;  // to silence compiler warning
787 
788   for (iter = 0; iter < round; ++iter) {
789     // Check vertical and horizontal sub-pixel positions.
790     for (idx = 0; idx < 4; ++idx) {
791       tr = br + search_step[idx].row;
792       tc = bc + search_step[idx].col;
793       if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
794         MV this_mv;
795         this_mv.row = tr;
796         this_mv.col = tc;
797 
798         if (use_accurate_subpel_search) {
799           thismse = accurate_sub_pel_search(xd, &this_mv, x->me_sf, kernel, vfp,
800                                             src_address, src_stride, y,
801                                             y_stride, second_pred, w, h, &sse);
802         } else {
803           const uint8_t *const pre_address =
804               y + (tr >> 3) * y_stride + (tc >> 3);
805           if (second_pred == NULL)
806             thismse = vfp->svf(pre_address, y_stride, sp(tc), sp(tr),
807                                src_address, src_stride, &sse);
808           else
809             thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
810                                 src_address, src_stride, &sse, second_pred);
811         }
812 
813         cost_array[idx] = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost,
814                                                 mvcost, error_per_bit);
815 
816         if (cost_array[idx] < besterr) {
817           best_idx = idx;
818           besterr = cost_array[idx];
819           *distortion = thismse;
820           *sse1 = sse;
821         }
822       } else {
823         cost_array[idx] = UINT_MAX;
824       }
825     }
826 
827     // Check diagonal sub-pixel position
828     kc = (cost_array[0] <= cost_array[1] ? -hstep : hstep);
829     kr = (cost_array[2] <= cost_array[3] ? -hstep : hstep);
830 
831     tc = bc + kc;
832     tr = br + kr;
833     if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
834       MV this_mv = { tr, tc };
835       if (use_accurate_subpel_search) {
836         thismse = accurate_sub_pel_search(xd, &this_mv, x->me_sf, kernel, vfp,
837                                           src_address, src_stride, y, y_stride,
838                                           second_pred, w, h, &sse);
839       } else {
840         const uint8_t *const pre_address = y + (tr >> 3) * y_stride + (tc >> 3);
841         if (second_pred == NULL)
842           thismse = vfp->svf(pre_address, y_stride, sp(tc), sp(tr), src_address,
843                              src_stride, &sse);
844         else
845           thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
846                               src_address, src_stride, &sse, second_pred);
847       }
848 
849       cost_array[4] = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost,
850                                             error_per_bit);
851 
852       if (cost_array[4] < besterr) {
853         best_idx = 4;
854         besterr = cost_array[4];
855         *distortion = thismse;
856         *sse1 = sse;
857       }
858     } else {
859       cost_array[idx] = UINT_MAX;
860     }
861 
862     if (best_idx < 4 && best_idx >= 0) {
863       br += search_step[best_idx].row;
864       bc += search_step[best_idx].col;
865     } else if (best_idx == 4) {
866       br = tr;
867       bc = tc;
868     }
869 
870     if (iters_per_step > 0 && best_idx != -1) {
871       unsigned int second;
872       const int br0 = br;
873       const int bc0 = bc;
874       assert(tr == br || tc == bc);
875 
876       if (tr == br && tc != bc) {
877         kc = bc - tc;
878         if (iters_per_step == 1) {
879           if (use_accurate_subpel_search) {
880             CHECK_BETTER1(second, br0, bc0 + kc);
881           } else {
882             CHECK_BETTER(second, br0, bc0 + kc);
883           }
884         }
885       } else if (tr != br && tc == bc) {
886         kr = br - tr;
887         if (iters_per_step == 1) {
888           if (use_accurate_subpel_search) {
889             CHECK_BETTER1(second, br0 + kr, bc0);
890           } else {
891             CHECK_BETTER(second, br0 + kr, bc0);
892           }
893         }
894       }
895 
896       if (iters_per_step > 1) {
897         if (use_accurate_subpel_search) {
898           CHECK_BETTER1(second, br0 + kr, bc0);
899           CHECK_BETTER1(second, br0, bc0 + kc);
900           if (br0 != br || bc0 != bc) {
901             CHECK_BETTER1(second, br0 + kr, bc0 + kc);
902           }
903         } else {
904           CHECK_BETTER(second, br0 + kr, bc0);
905           CHECK_BETTER(second, br0, bc0 + kc);
906           if (br0 != br || bc0 != bc) {
907             CHECK_BETTER(second, br0 + kr, bc0 + kc);
908           }
909         }
910       }
911     }
912 
913     search_step += 4;
914     hstep >>= 1;
915     best_idx = -1;
916   }
917 
918   // Each subsequent iteration checks at least one point in common with
919   // the last iteration could be 2 ( if diag selected) 1/4 pel
920 
921   // These lines insure static analysis doesn't warn that
922   // tr and tc aren't used after the above point.
923   (void)tr;
924   (void)tc;
925 
926   bestmv->row = br;
927   bestmv->col = bc;
928 
929   return besterr;
930 }
931 
932 #undef CHECK_BETTER
933 #undef CHECK_BETTER1
934 
check_bounds(const MvLimits * mv_limits,int row,int col,int range)935 static INLINE int check_bounds(const MvLimits *mv_limits, int row, int col,
936                                int range) {
937   return ((row - range) >= mv_limits->row_min) &
938          ((row + range) <= mv_limits->row_max) &
939          ((col - range) >= mv_limits->col_min) &
940          ((col + range) <= mv_limits->col_max);
941 }
942 
is_mv_in(const MvLimits * mv_limits,const MV * mv)943 static INLINE int is_mv_in(const MvLimits *mv_limits, const MV *mv) {
944   return (mv->col >= mv_limits->col_min) && (mv->col <= mv_limits->col_max) &&
945          (mv->row >= mv_limits->row_min) && (mv->row <= mv_limits->row_max);
946 }
947 
948 #define CHECK_BETTER                                                      \
949   {                                                                       \
950     if (thissad < bestsad) {                                              \
951       if (use_mvcost)                                                     \
952         thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); \
953       if (thissad < bestsad) {                                            \
954         bestsad = thissad;                                                \
955         best_site = i;                                                    \
956       }                                                                   \
957     }                                                                     \
958   }
959 
960 #define MAX_PATTERN_SCALES 11
961 #define MAX_PATTERN_CANDIDATES 8  // max number of canddiates per scale
962 #define PATTERN_CANDIDATES_REF 3  // number of refinement candidates
963 
964 // 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)965 static INLINE void calc_int_cost_list(const MACROBLOCK *x, const MV *ref_mv,
966                                       int sadpb,
967                                       const vp9_variance_fn_ptr_t *fn_ptr,
968                                       const MV *best_mv, int *cost_list) {
969   static const MV neighbors[4] = { { 0, -1 }, { 1, 0 }, { 0, 1 }, { -1, 0 } };
970   const struct buf_2d *const what = &x->plane[0].src;
971   const struct buf_2d *const in_what = &x->e_mbd.plane[0].pre[0];
972   const MV fcenter_mv = { ref_mv->row >> 3, ref_mv->col >> 3 };
973   int br = best_mv->row;
974   int bc = best_mv->col;
975   MV this_mv;
976   int i;
977   unsigned int sse;
978 
979   this_mv.row = br;
980   this_mv.col = bc;
981   cost_list[0] =
982       fn_ptr->vf(what->buf, what->stride, get_buf_from_mv(in_what, &this_mv),
983                  in_what->stride, &sse) +
984       mvsad_err_cost(x, &this_mv, &fcenter_mv, sadpb);
985   if (check_bounds(&x->mv_limits, br, bc, 1)) {
986     for (i = 0; i < 4; i++) {
987       const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
988       cost_list[i + 1] = fn_ptr->vf(what->buf, what->stride,
989                                     get_buf_from_mv(in_what, &this_mv),
990                                     in_what->stride, &sse) +
991                          mv_err_cost(&this_mv, &fcenter_mv, x->nmvjointcost,
992                                      x->mvcost, x->errorperbit);
993     }
994   } else {
995     for (i = 0; i < 4; i++) {
996       const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
997       if (!is_mv_in(&x->mv_limits, &this_mv))
998         cost_list[i + 1] = INT_MAX;
999       else
1000         cost_list[i + 1] = fn_ptr->vf(what->buf, what->stride,
1001                                       get_buf_from_mv(in_what, &this_mv),
1002                                       in_what->stride, &sse) +
1003                            mv_err_cost(&this_mv, &fcenter_mv, x->nmvjointcost,
1004                                        x->mvcost, x->errorperbit);
1005     }
1006   }
1007 }
1008 
1009 // Generic pattern search function that searches over multiple scales.
1010 // Each scale can have a different number of candidates and shape of
1011 // candidates as indicated in the num_candidates and candidates arrays
1012 // passed into this function
1013 //
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])1014 static int vp9_pattern_search(
1015     const MACROBLOCK *x, MV *ref_mv, int search_param, int sad_per_bit,
1016     int do_init_search, int *cost_list, const vp9_variance_fn_ptr_t *vfp,
1017     int use_mvcost, const MV *center_mv, MV *best_mv,
1018     const int num_candidates[MAX_PATTERN_SCALES],
1019     const MV candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES]) {
1020   const MACROBLOCKD *const xd = &x->e_mbd;
1021   static const int search_param_to_steps[MAX_MVSEARCH_STEPS] = {
1022     10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0,
1023   };
1024   int i, s, t;
1025   const struct buf_2d *const what = &x->plane[0].src;
1026   const struct buf_2d *const in_what = &xd->plane[0].pre[0];
1027   int br, bc;
1028   int bestsad = INT_MAX;
1029   int thissad;
1030   int k = -1;
1031   const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
1032   int best_init_s = search_param_to_steps[search_param];
1033   // adjust ref_mv to make sure it is within MV range
1034   clamp_mv(ref_mv, x->mv_limits.col_min, x->mv_limits.col_max,
1035            x->mv_limits.row_min, x->mv_limits.row_max);
1036   br = ref_mv->row;
1037   bc = ref_mv->col;
1038 
1039   // Work out the start point for the search
1040   bestsad = vfp->sdf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv),
1041                      in_what->stride) +
1042             mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
1043 
1044   // Search all possible scales upto the search param around the center point
1045   // pick the scale of the point that is best as the starting scale of
1046   // further steps around it.
1047   if (do_init_search) {
1048     s = best_init_s;
1049     best_init_s = -1;
1050     for (t = 0; t <= s; ++t) {
1051       int best_site = -1;
1052       if (check_bounds(&x->mv_limits, br, bc, 1 << t)) {
1053         for (i = 0; i < num_candidates[t]; i++) {
1054           const MV this_mv = { br + candidates[t][i].row,
1055                                bc + candidates[t][i].col };
1056           thissad =
1057               vfp->sdf(what->buf, what->stride,
1058                        get_buf_from_mv(in_what, &this_mv), in_what->stride);
1059           CHECK_BETTER
1060         }
1061       } else {
1062         for (i = 0; i < num_candidates[t]; i++) {
1063           const MV this_mv = { br + candidates[t][i].row,
1064                                bc + candidates[t][i].col };
1065           if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1066           thissad =
1067               vfp->sdf(what->buf, what->stride,
1068                        get_buf_from_mv(in_what, &this_mv), in_what->stride);
1069           CHECK_BETTER
1070         }
1071       }
1072       if (best_site == -1) {
1073         continue;
1074       } else {
1075         best_init_s = t;
1076         k = best_site;
1077       }
1078     }
1079     if (best_init_s != -1) {
1080       br += candidates[best_init_s][k].row;
1081       bc += candidates[best_init_s][k].col;
1082     }
1083   }
1084 
1085   // If the center point is still the best, just skip this and move to
1086   // the refinement step.
1087   if (best_init_s != -1) {
1088     int best_site = -1;
1089     s = best_init_s;
1090 
1091     do {
1092       // No need to search all 6 points the 1st time if initial search was used
1093       if (!do_init_search || s != best_init_s) {
1094         if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1095           for (i = 0; i < num_candidates[s]; i++) {
1096             const MV this_mv = { br + candidates[s][i].row,
1097                                  bc + candidates[s][i].col };
1098             thissad =
1099                 vfp->sdf(what->buf, what->stride,
1100                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1101             CHECK_BETTER
1102           }
1103         } else {
1104           for (i = 0; i < num_candidates[s]; i++) {
1105             const MV this_mv = { br + candidates[s][i].row,
1106                                  bc + candidates[s][i].col };
1107             if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1108             thissad =
1109                 vfp->sdf(what->buf, what->stride,
1110                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1111             CHECK_BETTER
1112           }
1113         }
1114 
1115         if (best_site == -1) {
1116           continue;
1117         } else {
1118           br += candidates[s][best_site].row;
1119           bc += candidates[s][best_site].col;
1120           k = best_site;
1121         }
1122       }
1123 
1124       do {
1125         int next_chkpts_indices[PATTERN_CANDIDATES_REF];
1126         best_site = -1;
1127         next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
1128         next_chkpts_indices[1] = k;
1129         next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
1130 
1131         if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1132           for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1133             const MV this_mv = {
1134               br + candidates[s][next_chkpts_indices[i]].row,
1135               bc + candidates[s][next_chkpts_indices[i]].col
1136             };
1137             thissad =
1138                 vfp->sdf(what->buf, what->stride,
1139                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1140             CHECK_BETTER
1141           }
1142         } else {
1143           for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1144             const MV this_mv = {
1145               br + candidates[s][next_chkpts_indices[i]].row,
1146               bc + candidates[s][next_chkpts_indices[i]].col
1147             };
1148             if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1149             thissad =
1150                 vfp->sdf(what->buf, what->stride,
1151                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1152             CHECK_BETTER
1153           }
1154         }
1155 
1156         if (best_site != -1) {
1157           k = next_chkpts_indices[best_site];
1158           br += candidates[s][k].row;
1159           bc += candidates[s][k].col;
1160         }
1161       } while (best_site != -1);
1162     } while (s--);
1163   }
1164 
1165   // Returns the one-away integer pel sad values around the best as follows:
1166   // cost_list[0]: cost at the best integer pel
1167   // cost_list[1]: cost at delta {0, -1} (left)   from the best integer pel
1168   // cost_list[2]: cost at delta { 1, 0} (bottom) from the best integer pel
1169   // cost_list[3]: cost at delta { 0, 1} (right)  from the best integer pel
1170   // cost_list[4]: cost at delta {-1, 0} (top)    from the best integer pel
1171   if (cost_list) {
1172     const MV best_mv = { br, bc };
1173     calc_int_cost_list(x, &fcenter_mv, sad_per_bit, vfp, &best_mv, cost_list);
1174   }
1175   best_mv->row = br;
1176   best_mv->col = bc;
1177   return bestsad;
1178 }
1179 
1180 // A specialized function where the smallest scale search candidates
1181 // are 4 1-away neighbors, and cost_list is non-null
1182 // TODO(debargha): Merge this function with the one above. Also remove
1183 // 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])1184 static int vp9_pattern_search_sad(
1185     const MACROBLOCK *x, MV *ref_mv, int search_param, int sad_per_bit,
1186     int do_init_search, int *cost_list, const vp9_variance_fn_ptr_t *vfp,
1187     int use_mvcost, const MV *center_mv, MV *best_mv,
1188     const int num_candidates[MAX_PATTERN_SCALES],
1189     const MV candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES]) {
1190   const MACROBLOCKD *const xd = &x->e_mbd;
1191   static const int search_param_to_steps[MAX_MVSEARCH_STEPS] = {
1192     10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0,
1193   };
1194   int i, s, t;
1195   const struct buf_2d *const what = &x->plane[0].src;
1196   const struct buf_2d *const in_what = &xd->plane[0].pre[0];
1197   int br, bc;
1198   int bestsad = INT_MAX;
1199   int thissad;
1200   int k = -1;
1201   const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
1202   int best_init_s = search_param_to_steps[search_param];
1203   // adjust ref_mv to make sure it is within MV range
1204   clamp_mv(ref_mv, x->mv_limits.col_min, x->mv_limits.col_max,
1205            x->mv_limits.row_min, x->mv_limits.row_max);
1206   br = ref_mv->row;
1207   bc = ref_mv->col;
1208   if (cost_list != NULL) {
1209     cost_list[0] = cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] =
1210         INT_MAX;
1211   }
1212 
1213   // Work out the start point for the search
1214   bestsad = vfp->sdf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv),
1215                      in_what->stride) +
1216             mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
1217 
1218   // Search all possible scales upto the search param around the center point
1219   // pick the scale of the point that is best as the starting scale of
1220   // further steps around it.
1221   if (do_init_search) {
1222     s = best_init_s;
1223     best_init_s = -1;
1224     for (t = 0; t <= s; ++t) {
1225       int best_site = -1;
1226       if (check_bounds(&x->mv_limits, br, bc, 1 << t)) {
1227         for (i = 0; i < num_candidates[t]; i++) {
1228           const MV this_mv = { br + candidates[t][i].row,
1229                                bc + candidates[t][i].col };
1230           thissad =
1231               vfp->sdf(what->buf, what->stride,
1232                        get_buf_from_mv(in_what, &this_mv), in_what->stride);
1233           CHECK_BETTER
1234         }
1235       } else {
1236         for (i = 0; i < num_candidates[t]; i++) {
1237           const MV this_mv = { br + candidates[t][i].row,
1238                                bc + candidates[t][i].col };
1239           if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1240           thissad =
1241               vfp->sdf(what->buf, what->stride,
1242                        get_buf_from_mv(in_what, &this_mv), in_what->stride);
1243           CHECK_BETTER
1244         }
1245       }
1246       if (best_site == -1) {
1247         continue;
1248       } else {
1249         best_init_s = t;
1250         k = best_site;
1251       }
1252     }
1253     if (best_init_s != -1) {
1254       br += candidates[best_init_s][k].row;
1255       bc += candidates[best_init_s][k].col;
1256     }
1257   }
1258 
1259   // If the center point is still the best, just skip this and move to
1260   // the refinement step.
1261   if (best_init_s != -1) {
1262     int do_sad = (num_candidates[0] == 4 && cost_list != NULL);
1263     int best_site = -1;
1264     s = best_init_s;
1265 
1266     for (; s >= do_sad; s--) {
1267       if (!do_init_search || s != best_init_s) {
1268         if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1269           for (i = 0; i < num_candidates[s]; i++) {
1270             const MV this_mv = { br + candidates[s][i].row,
1271                                  bc + candidates[s][i].col };
1272             thissad =
1273                 vfp->sdf(what->buf, what->stride,
1274                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1275             CHECK_BETTER
1276           }
1277         } else {
1278           for (i = 0; i < num_candidates[s]; i++) {
1279             const MV this_mv = { br + candidates[s][i].row,
1280                                  bc + candidates[s][i].col };
1281             if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1282             thissad =
1283                 vfp->sdf(what->buf, what->stride,
1284                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1285             CHECK_BETTER
1286           }
1287         }
1288 
1289         if (best_site == -1) {
1290           continue;
1291         } else {
1292           br += candidates[s][best_site].row;
1293           bc += candidates[s][best_site].col;
1294           k = best_site;
1295         }
1296       }
1297 
1298       do {
1299         int next_chkpts_indices[PATTERN_CANDIDATES_REF];
1300         best_site = -1;
1301         next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
1302         next_chkpts_indices[1] = k;
1303         next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
1304 
1305         if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1306           for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1307             const MV this_mv = {
1308               br + candidates[s][next_chkpts_indices[i]].row,
1309               bc + candidates[s][next_chkpts_indices[i]].col
1310             };
1311             thissad =
1312                 vfp->sdf(what->buf, what->stride,
1313                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1314             CHECK_BETTER
1315           }
1316         } else {
1317           for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1318             const MV this_mv = {
1319               br + candidates[s][next_chkpts_indices[i]].row,
1320               bc + candidates[s][next_chkpts_indices[i]].col
1321             };
1322             if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1323             thissad =
1324                 vfp->sdf(what->buf, what->stride,
1325                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1326             CHECK_BETTER
1327           }
1328         }
1329 
1330         if (best_site != -1) {
1331           k = next_chkpts_indices[best_site];
1332           br += candidates[s][k].row;
1333           bc += candidates[s][k].col;
1334         }
1335       } while (best_site != -1);
1336     }
1337 
1338     // Note: If we enter the if below, then cost_list must be non-NULL.
1339     if (s == 0) {
1340       cost_list[0] = bestsad;
1341       if (!do_init_search || s != best_init_s) {
1342         if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1343           for (i = 0; i < num_candidates[s]; i++) {
1344             const MV this_mv = { br + candidates[s][i].row,
1345                                  bc + candidates[s][i].col };
1346             cost_list[i + 1] = thissad =
1347                 vfp->sdf(what->buf, what->stride,
1348                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1349             CHECK_BETTER
1350           }
1351         } else {
1352           for (i = 0; i < num_candidates[s]; i++) {
1353             const MV this_mv = { br + candidates[s][i].row,
1354                                  bc + candidates[s][i].col };
1355             if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
1356             cost_list[i + 1] = thissad =
1357                 vfp->sdf(what->buf, what->stride,
1358                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1359             CHECK_BETTER
1360           }
1361         }
1362 
1363         if (best_site != -1) {
1364           br += candidates[s][best_site].row;
1365           bc += candidates[s][best_site].col;
1366           k = best_site;
1367         }
1368       }
1369       while (best_site != -1) {
1370         int next_chkpts_indices[PATTERN_CANDIDATES_REF];
1371         best_site = -1;
1372         next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
1373         next_chkpts_indices[1] = k;
1374         next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
1375         cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] = INT_MAX;
1376         cost_list[((k + 2) % 4) + 1] = cost_list[0];
1377         cost_list[0] = bestsad;
1378 
1379         if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
1380           for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1381             const MV this_mv = {
1382               br + candidates[s][next_chkpts_indices[i]].row,
1383               bc + candidates[s][next_chkpts_indices[i]].col
1384             };
1385             cost_list[next_chkpts_indices[i] + 1] = thissad =
1386                 vfp->sdf(what->buf, what->stride,
1387                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1388             CHECK_BETTER
1389           }
1390         } else {
1391           for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
1392             const MV this_mv = {
1393               br + candidates[s][next_chkpts_indices[i]].row,
1394               bc + candidates[s][next_chkpts_indices[i]].col
1395             };
1396             if (!is_mv_in(&x->mv_limits, &this_mv)) {
1397               cost_list[next_chkpts_indices[i] + 1] = INT_MAX;
1398               continue;
1399             }
1400             cost_list[next_chkpts_indices[i] + 1] = thissad =
1401                 vfp->sdf(what->buf, what->stride,
1402                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1403             CHECK_BETTER
1404           }
1405         }
1406 
1407         if (best_site != -1) {
1408           k = next_chkpts_indices[best_site];
1409           br += candidates[s][k].row;
1410           bc += candidates[s][k].col;
1411         }
1412       }
1413     }
1414   }
1415 
1416   // Returns the one-away integer pel sad values around the best as follows:
1417   // cost_list[0]: sad at the best integer pel
1418   // cost_list[1]: sad at delta {0, -1} (left)   from the best integer pel
1419   // cost_list[2]: sad at delta { 1, 0} (bottom) from the best integer pel
1420   // cost_list[3]: sad at delta { 0, 1} (right)  from the best integer pel
1421   // cost_list[4]: sad at delta {-1, 0} (top)    from the best integer pel
1422   if (cost_list) {
1423     static const MV neighbors[4] = { { 0, -1 }, { 1, 0 }, { 0, 1 }, { -1, 0 } };
1424     if (cost_list[0] == INT_MAX) {
1425       cost_list[0] = bestsad;
1426       if (check_bounds(&x->mv_limits, br, bc, 1)) {
1427         for (i = 0; i < 4; i++) {
1428           const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
1429           cost_list[i + 1] =
1430               vfp->sdf(what->buf, what->stride,
1431                        get_buf_from_mv(in_what, &this_mv), in_what->stride);
1432         }
1433       } else {
1434         for (i = 0; i < 4; i++) {
1435           const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
1436           if (!is_mv_in(&x->mv_limits, &this_mv))
1437             cost_list[i + 1] = INT_MAX;
1438           else
1439             cost_list[i + 1] =
1440                 vfp->sdf(what->buf, what->stride,
1441                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
1442         }
1443       }
1444     } else {
1445       if (use_mvcost) {
1446         for (i = 0; i < 4; i++) {
1447           const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
1448           if (cost_list[i + 1] != INT_MAX) {
1449             cost_list[i + 1] +=
1450                 mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
1451           }
1452         }
1453       }
1454     }
1455   }
1456   best_mv->row = br;
1457   best_mv->col = bc;
1458   return bestsad;
1459 }
1460 
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)1461 int vp9_get_mvpred_var(const MACROBLOCK *x, const MV *best_mv,
1462                        const MV *center_mv, const vp9_variance_fn_ptr_t *vfp,
1463                        int use_mvcost) {
1464   const MACROBLOCKD *const xd = &x->e_mbd;
1465   const struct buf_2d *const what = &x->plane[0].src;
1466   const struct buf_2d *const in_what = &xd->plane[0].pre[0];
1467   const MV mv = { best_mv->row * 8, best_mv->col * 8 };
1468   uint32_t unused;
1469 #if CONFIG_VP9_HIGHBITDEPTH
1470   uint64_t err =
1471       vfp->vf(what->buf, what->stride, get_buf_from_mv(in_what, best_mv),
1472               in_what->stride, &unused);
1473   err += (use_mvcost ? mv_err_cost(&mv, center_mv, x->nmvjointcost, x->mvcost,
1474                                    x->errorperbit)
1475                      : 0);
1476   if (err >= INT_MAX) return INT_MAX;
1477   return (int)err;
1478 #else
1479   return vfp->vf(what->buf, what->stride, get_buf_from_mv(in_what, best_mv),
1480                  in_what->stride, &unused) +
1481          (use_mvcost ? mv_err_cost(&mv, center_mv, x->nmvjointcost, x->mvcost,
1482                                    x->errorperbit)
1483                      : 0);
1484 #endif
1485 }
1486 
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)1487 int vp9_get_mvpred_av_var(const MACROBLOCK *x, const MV *best_mv,
1488                           const MV *center_mv, const uint8_t *second_pred,
1489                           const vp9_variance_fn_ptr_t *vfp, int use_mvcost) {
1490   const MACROBLOCKD *const xd = &x->e_mbd;
1491   const struct buf_2d *const what = &x->plane[0].src;
1492   const struct buf_2d *const in_what = &xd->plane[0].pre[0];
1493   const MV mv = { best_mv->row * 8, best_mv->col * 8 };
1494   unsigned int unused;
1495 
1496   return vfp->svaf(get_buf_from_mv(in_what, best_mv), in_what->stride, 0, 0,
1497                    what->buf, what->stride, &unused, second_pred) +
1498          (use_mvcost ? mv_err_cost(&mv, center_mv, x->nmvjointcost, x->mvcost,
1499                                    x->errorperbit)
1500                      : 0);
1501 }
1502 
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)1503 static int hex_search(const MACROBLOCK *x, MV *ref_mv, int search_param,
1504                       int sad_per_bit, int do_init_search, int *cost_list,
1505                       const vp9_variance_fn_ptr_t *vfp, int use_mvcost,
1506                       const MV *center_mv, MV *best_mv) {
1507   // First scale has 8-closest points, the rest have 6 points in hex shape
1508   // at increasing scales
1509   static const int hex_num_candidates[MAX_PATTERN_SCALES] = { 8, 6, 6, 6, 6, 6,
1510                                                               6, 6, 6, 6, 6 };
1511   // Note that the largest candidate step at each scale is 2^scale
1512   /* clang-format off */
1513   static const MV hex_candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES] = {
1514     { { -1, -1 }, { 0, -1 }, { 1, -1 }, { 1, 0 }, { 1, 1 }, { 0, 1 }, { -1, 1 },
1515       { -1, 0 } },
1516     { { -1, -2 }, { 1, -2 }, { 2, 0 }, { 1, 2 }, { -1, 2 }, { -2, 0 } },
1517     { { -2, -4 }, { 2, -4 }, { 4, 0 }, { 2, 4 }, { -2, 4 }, { -4, 0 } },
1518     { { -4, -8 }, { 4, -8 }, { 8, 0 }, { 4, 8 }, { -4, 8 }, { -8, 0 } },
1519     { { -8, -16 }, { 8, -16 }, { 16, 0 }, { 8, 16 }, { -8, 16 }, { -16, 0 } },
1520     { { -16, -32 }, { 16, -32 }, { 32, 0 }, { 16, 32 }, { -16, 32 },
1521       { -32, 0 } },
1522     { { -32, -64 }, { 32, -64 }, { 64, 0 }, { 32, 64 }, { -32, 64 },
1523       { -64, 0 } },
1524     { { -64, -128 }, { 64, -128 }, { 128, 0 }, { 64, 128 }, { -64, 128 },
1525       { -128, 0 } },
1526     { { -128, -256 }, { 128, -256 }, { 256, 0 }, { 128, 256 }, { -128, 256 },
1527       { -256, 0 } },
1528     { { -256, -512 }, { 256, -512 }, { 512, 0 }, { 256, 512 }, { -256, 512 },
1529       { -512, 0 } },
1530     { { -512, -1024 }, { 512, -1024 }, { 1024, 0 }, { 512, 1024 },
1531       { -512, 1024 }, { -1024, 0 } }
1532   };
1533   /* clang-format on */
1534   return vp9_pattern_search(
1535       x, ref_mv, search_param, sad_per_bit, do_init_search, cost_list, vfp,
1536       use_mvcost, center_mv, best_mv, hex_num_candidates, hex_candidates);
1537 }
1538 
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)1539 static int bigdia_search(const MACROBLOCK *x, MV *ref_mv, int search_param,
1540                          int sad_per_bit, int do_init_search, int *cost_list,
1541                          const vp9_variance_fn_ptr_t *vfp, int use_mvcost,
1542                          const MV *center_mv, MV *best_mv) {
1543   // First scale has 4-closest points, the rest have 8 points in diamond
1544   // shape at increasing scales
1545   static const int bigdia_num_candidates[MAX_PATTERN_SCALES] = {
1546     4, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1547   };
1548   // Note that the largest candidate step at each scale is 2^scale
1549   /* clang-format off */
1550   static const MV
1551       bigdia_candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES] = {
1552         { { 0, -1 }, { 1, 0 }, { 0, 1 }, { -1, 0 } },
1553         { { -1, -1 }, { 0, -2 }, { 1, -1 }, { 2, 0 }, { 1, 1 }, { 0, 2 },
1554           { -1, 1 }, { -2, 0 } },
1555         { { -2, -2 }, { 0, -4 }, { 2, -2 }, { 4, 0 }, { 2, 2 }, { 0, 4 },
1556           { -2, 2 }, { -4, 0 } },
1557         { { -4, -4 }, { 0, -8 }, { 4, -4 }, { 8, 0 }, { 4, 4 }, { 0, 8 },
1558           { -4, 4 }, { -8, 0 } },
1559         { { -8, -8 }, { 0, -16 }, { 8, -8 }, { 16, 0 }, { 8, 8 }, { 0, 16 },
1560           { -8, 8 }, { -16, 0 } },
1561         { { -16, -16 }, { 0, -32 }, { 16, -16 }, { 32, 0 }, { 16, 16 },
1562           { 0, 32 }, { -16, 16 }, { -32, 0 } },
1563         { { -32, -32 }, { 0, -64 }, { 32, -32 }, { 64, 0 }, { 32, 32 },
1564           { 0, 64 }, { -32, 32 }, { -64, 0 } },
1565         { { -64, -64 }, { 0, -128 }, { 64, -64 }, { 128, 0 }, { 64, 64 },
1566           { 0, 128 }, { -64, 64 }, { -128, 0 } },
1567         { { -128, -128 }, { 0, -256 }, { 128, -128 }, { 256, 0 }, { 128, 128 },
1568           { 0, 256 }, { -128, 128 }, { -256, 0 } },
1569         { { -256, -256 }, { 0, -512 }, { 256, -256 }, { 512, 0 }, { 256, 256 },
1570           { 0, 512 }, { -256, 256 }, { -512, 0 } },
1571         { { -512, -512 }, { 0, -1024 }, { 512, -512 }, { 1024, 0 },
1572           { 512, 512 }, { 0, 1024 }, { -512, 512 }, { -1024, 0 } }
1573       };
1574   /* clang-format on */
1575   return vp9_pattern_search_sad(
1576       x, ref_mv, search_param, sad_per_bit, do_init_search, cost_list, vfp,
1577       use_mvcost, center_mv, best_mv, bigdia_num_candidates, bigdia_candidates);
1578 }
1579 
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)1580 static int square_search(const MACROBLOCK *x, MV *ref_mv, int search_param,
1581                          int sad_per_bit, int do_init_search, int *cost_list,
1582                          const vp9_variance_fn_ptr_t *vfp, int use_mvcost,
1583                          const MV *center_mv, MV *best_mv) {
1584   // All scales have 8 closest points in square shape
1585   static const int square_num_candidates[MAX_PATTERN_SCALES] = {
1586     8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
1587   };
1588   // Note that the largest candidate step at each scale is 2^scale
1589   /* clang-format off */
1590   static const MV
1591       square_candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES] = {
1592         { { -1, -1 }, { 0, -1 }, { 1, -1 }, { 1, 0 }, { 1, 1 }, { 0, 1 },
1593           { -1, 1 }, { -1, 0 } },
1594         { { -2, -2 }, { 0, -2 }, { 2, -2 }, { 2, 0 }, { 2, 2 }, { 0, 2 },
1595           { -2, 2 }, { -2, 0 } },
1596         { { -4, -4 }, { 0, -4 }, { 4, -4 }, { 4, 0 }, { 4, 4 }, { 0, 4 },
1597           { -4, 4 }, { -4, 0 } },
1598         { { -8, -8 }, { 0, -8 }, { 8, -8 }, { 8, 0 }, { 8, 8 }, { 0, 8 },
1599           { -8, 8 }, { -8, 0 } },
1600         { { -16, -16 }, { 0, -16 }, { 16, -16 }, { 16, 0 }, { 16, 16 },
1601           { 0, 16 }, { -16, 16 }, { -16, 0 } },
1602         { { -32, -32 }, { 0, -32 }, { 32, -32 }, { 32, 0 }, { 32, 32 },
1603           { 0, 32 }, { -32, 32 }, { -32, 0 } },
1604         { { -64, -64 }, { 0, -64 }, { 64, -64 }, { 64, 0 }, { 64, 64 },
1605           { 0, 64 }, { -64, 64 }, { -64, 0 } },
1606         { { -128, -128 }, { 0, -128 }, { 128, -128 }, { 128, 0 }, { 128, 128 },
1607           { 0, 128 }, { -128, 128 }, { -128, 0 } },
1608         { { -256, -256 }, { 0, -256 }, { 256, -256 }, { 256, 0 }, { 256, 256 },
1609           { 0, 256 }, { -256, 256 }, { -256, 0 } },
1610         { { -512, -512 }, { 0, -512 }, { 512, -512 }, { 512, 0 }, { 512, 512 },
1611           { 0, 512 }, { -512, 512 }, { -512, 0 } },
1612         { { -1024, -1024 }, { 0, -1024 }, { 1024, -1024 }, { 1024, 0 },
1613           { 1024, 1024 }, { 0, 1024 }, { -1024, 1024 }, { -1024, 0 } }
1614       };
1615   /* clang-format on */
1616   return vp9_pattern_search(
1617       x, ref_mv, search_param, sad_per_bit, do_init_search, cost_list, vfp,
1618       use_mvcost, center_mv, best_mv, square_num_candidates, square_candidates);
1619 }
1620 
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)1621 static int fast_hex_search(const MACROBLOCK *x, MV *ref_mv, int search_param,
1622                            int sad_per_bit,
1623                            int do_init_search,  // must be zero for fast_hex
1624                            int *cost_list, const vp9_variance_fn_ptr_t *vfp,
1625                            int use_mvcost, const MV *center_mv, MV *best_mv) {
1626   return hex_search(x, ref_mv, VPXMAX(MAX_MVSEARCH_STEPS - 2, search_param),
1627                     sad_per_bit, do_init_search, cost_list, vfp, use_mvcost,
1628                     center_mv, best_mv);
1629 }
1630 
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)1631 static int fast_dia_search(const MACROBLOCK *x, MV *ref_mv, int search_param,
1632                            int sad_per_bit, int do_init_search, int *cost_list,
1633                            const vp9_variance_fn_ptr_t *vfp, int use_mvcost,
1634                            const MV *center_mv, MV *best_mv) {
1635   return bigdia_search(x, ref_mv, VPXMAX(MAX_MVSEARCH_STEPS - 2, search_param),
1636                        sad_per_bit, do_init_search, cost_list, vfp, use_mvcost,
1637                        center_mv, best_mv);
1638 }
1639 
1640 #undef CHECK_BETTER
1641 
1642 // Exhuastive motion search around a given centre position with a given
1643 // step size.
exhaustive_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)1644 static int exhaustive_mesh_search(const MACROBLOCK *x, MV *ref_mv, MV *best_mv,
1645                                   int range, int step, int sad_per_bit,
1646                                   const vp9_variance_fn_ptr_t *fn_ptr,
1647                                   const MV *center_mv) {
1648   const MACROBLOCKD *const xd = &x->e_mbd;
1649   const struct buf_2d *const what = &x->plane[0].src;
1650   const struct buf_2d *const in_what = &xd->plane[0].pre[0];
1651   MV fcenter_mv = { center_mv->row, center_mv->col };
1652   unsigned int best_sad = INT_MAX;
1653   int r, c, i;
1654   int start_col, end_col, start_row, end_row;
1655   int col_step = (step > 1) ? step : 4;
1656 
1657   assert(step >= 1);
1658 
1659   clamp_mv(&fcenter_mv, x->mv_limits.col_min, x->mv_limits.col_max,
1660            x->mv_limits.row_min, x->mv_limits.row_max);
1661   *best_mv = fcenter_mv;
1662   best_sad =
1663       fn_ptr->sdf(what->buf, what->stride,
1664                   get_buf_from_mv(in_what, &fcenter_mv), in_what->stride) +
1665       mvsad_err_cost(x, &fcenter_mv, ref_mv, sad_per_bit);
1666   start_row = VPXMAX(-range, x->mv_limits.row_min - fcenter_mv.row);
1667   start_col = VPXMAX(-range, x->mv_limits.col_min - fcenter_mv.col);
1668   end_row = VPXMIN(range, x->mv_limits.row_max - fcenter_mv.row);
1669   end_col = VPXMIN(range, x->mv_limits.col_max - fcenter_mv.col);
1670 
1671   for (r = start_row; r <= end_row; r += step) {
1672     for (c = start_col; c <= end_col; c += col_step) {
1673       // Step > 1 means we are not checking every location in this pass.
1674       if (step > 1) {
1675         const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c };
1676         unsigned int sad =
1677             fn_ptr->sdf(what->buf, what->stride, get_buf_from_mv(in_what, &mv),
1678                         in_what->stride);
1679         if (sad < best_sad) {
1680           sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
1681           if (sad < best_sad) {
1682             best_sad = sad;
1683             *best_mv = mv;
1684           }
1685         }
1686       } else {
1687         // 4 sads in a single call if we are checking every location
1688         if (c + 3 <= end_col) {
1689           unsigned int sads[4];
1690           const uint8_t *addrs[4];
1691           for (i = 0; i < 4; ++i) {
1692             const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c + i };
1693             addrs[i] = get_buf_from_mv(in_what, &mv);
1694           }
1695           fn_ptr->sdx4df(what->buf, what->stride, addrs, in_what->stride, sads);
1696 
1697           for (i = 0; i < 4; ++i) {
1698             if (sads[i] < best_sad) {
1699               const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c + i };
1700               const unsigned int sad =
1701                   sads[i] + mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
1702               if (sad < best_sad) {
1703                 best_sad = sad;
1704                 *best_mv = mv;
1705               }
1706             }
1707           }
1708         } else {
1709           for (i = 0; i < end_col - c; ++i) {
1710             const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c + i };
1711             unsigned int sad =
1712                 fn_ptr->sdf(what->buf, what->stride,
1713                             get_buf_from_mv(in_what, &mv), in_what->stride);
1714             if (sad < best_sad) {
1715               sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
1716               if (sad < best_sad) {
1717                 best_sad = sad;
1718                 *best_mv = mv;
1719               }
1720             }
1721           }
1722         }
1723       }
1724     }
1725   }
1726 
1727   return best_sad;
1728 }
1729 
1730 #define MIN_RANGE 7
1731 #define MAX_RANGE 256
1732 #define MIN_INTERVAL 1
1733 #if CONFIG_NON_GREEDY_MV
exhaustive_mesh_search_multi_step(MV * best_mv,const MV * center_mv,int range,int step,const struct buf_2d * src,const struct buf_2d * pre,int lambda,const int_mv * nb_full_mvs,int full_mv_num,const MvLimits * mv_limits,const vp9_variance_fn_ptr_t * fn_ptr)1734 static int64_t exhaustive_mesh_search_multi_step(
1735     MV *best_mv, const MV *center_mv, int range, int step,
1736     const struct buf_2d *src, const struct buf_2d *pre, int lambda,
1737     const int_mv *nb_full_mvs, int full_mv_num, const MvLimits *mv_limits,
1738     const vp9_variance_fn_ptr_t *fn_ptr) {
1739   int64_t best_sad;
1740   int r, c;
1741   int start_col, end_col, start_row, end_row;
1742   *best_mv = *center_mv;
1743   best_sad =
1744       ((int64_t)fn_ptr->sdf(src->buf, src->stride,
1745                             get_buf_from_mv(pre, center_mv), pre->stride)
1746        << LOG2_PRECISION) +
1747       lambda * vp9_nb_mvs_inconsistency(best_mv, nb_full_mvs, full_mv_num);
1748   start_row = VPXMAX(center_mv->row - range, mv_limits->row_min);
1749   start_col = VPXMAX(center_mv->col - range, mv_limits->col_min);
1750   end_row = VPXMIN(center_mv->row + range, mv_limits->row_max);
1751   end_col = VPXMIN(center_mv->col + range, mv_limits->col_max);
1752   for (r = start_row; r <= end_row; r += step) {
1753     for (c = start_col; c <= end_col; c += step) {
1754       const MV mv = { r, c };
1755       int64_t sad = (int64_t)fn_ptr->sdf(src->buf, src->stride,
1756                                          get_buf_from_mv(pre, &mv), pre->stride)
1757                     << LOG2_PRECISION;
1758       if (sad < best_sad) {
1759         sad += lambda * vp9_nb_mvs_inconsistency(&mv, nb_full_mvs, full_mv_num);
1760         if (sad < best_sad) {
1761           best_sad = sad;
1762           *best_mv = mv;
1763         }
1764       }
1765     }
1766   }
1767   return best_sad;
1768 }
1769 
exhaustive_mesh_search_single_step(MV * best_mv,const MV * center_mv,int range,const struct buf_2d * src,const struct buf_2d * pre,int lambda,const int_mv * nb_full_mvs,int full_mv_num,const MvLimits * mv_limits,const vp9_variance_fn_ptr_t * fn_ptr)1770 static int64_t exhaustive_mesh_search_single_step(
1771     MV *best_mv, const MV *center_mv, int range, const struct buf_2d *src,
1772     const struct buf_2d *pre, int lambda, const int_mv *nb_full_mvs,
1773     int full_mv_num, const MvLimits *mv_limits,
1774     const vp9_variance_fn_ptr_t *fn_ptr) {
1775   int64_t best_sad;
1776   int r, c, i;
1777   int start_col, end_col, start_row, end_row;
1778 
1779   *best_mv = *center_mv;
1780   best_sad =
1781       ((int64_t)fn_ptr->sdf(src->buf, src->stride,
1782                             get_buf_from_mv(pre, center_mv), pre->stride)
1783        << LOG2_PRECISION) +
1784       lambda * vp9_nb_mvs_inconsistency(best_mv, nb_full_mvs, full_mv_num);
1785   start_row = VPXMAX(center_mv->row - range, mv_limits->row_min);
1786   start_col = VPXMAX(center_mv->col - range, mv_limits->col_min);
1787   end_row = VPXMIN(center_mv->row + range, mv_limits->row_max);
1788   end_col = VPXMIN(center_mv->col + range, mv_limits->col_max);
1789   for (r = start_row; r <= end_row; r += 1) {
1790     c = start_col;
1791     // sdx8f may not be available some block size
1792     if (fn_ptr->sdx8f) {
1793       while (c + 7 <= end_col) {
1794         unsigned int sads[8];
1795         const MV mv = { r, c };
1796         const uint8_t *buf = get_buf_from_mv(pre, &mv);
1797         fn_ptr->sdx8f(src->buf, src->stride, buf, pre->stride, sads);
1798 
1799         for (i = 0; i < 8; ++i) {
1800           int64_t sad = (int64_t)sads[i] << LOG2_PRECISION;
1801           if (sad < best_sad) {
1802             const MV mv = { r, c + i };
1803             sad += lambda *
1804                    vp9_nb_mvs_inconsistency(&mv, nb_full_mvs, full_mv_num);
1805             if (sad < best_sad) {
1806               best_sad = sad;
1807               *best_mv = mv;
1808             }
1809           }
1810         }
1811         c += 8;
1812       }
1813     }
1814     while (c + 3 <= end_col) {
1815       unsigned int sads[4];
1816       const uint8_t *addrs[4];
1817       for (i = 0; i < 4; ++i) {
1818         const MV mv = { r, c + i };
1819         addrs[i] = get_buf_from_mv(pre, &mv);
1820       }
1821       fn_ptr->sdx4df(src->buf, src->stride, addrs, pre->stride, sads);
1822 
1823       for (i = 0; i < 4; ++i) {
1824         int64_t sad = (int64_t)sads[i] << LOG2_PRECISION;
1825         if (sad < best_sad) {
1826           const MV mv = { r, c + i };
1827           sad +=
1828               lambda * vp9_nb_mvs_inconsistency(&mv, nb_full_mvs, full_mv_num);
1829           if (sad < best_sad) {
1830             best_sad = sad;
1831             *best_mv = mv;
1832           }
1833         }
1834       }
1835       c += 4;
1836     }
1837     while (c <= end_col) {
1838       const MV mv = { r, c };
1839       int64_t sad = (int64_t)fn_ptr->sdf(src->buf, src->stride,
1840                                          get_buf_from_mv(pre, &mv), pre->stride)
1841                     << LOG2_PRECISION;
1842       if (sad < best_sad) {
1843         sad += lambda * vp9_nb_mvs_inconsistency(&mv, nb_full_mvs, full_mv_num);
1844         if (sad < best_sad) {
1845           best_sad = sad;
1846           *best_mv = mv;
1847         }
1848       }
1849       c += 1;
1850     }
1851   }
1852   return best_sad;
1853 }
1854 
exhaustive_mesh_search_new(const MACROBLOCK * x,MV * best_mv,int range,int step,const vp9_variance_fn_ptr_t * fn_ptr,const MV * center_mv,int lambda,const int_mv * nb_full_mvs,int full_mv_num)1855 static int64_t exhaustive_mesh_search_new(const MACROBLOCK *x, MV *best_mv,
1856                                           int range, int step,
1857                                           const vp9_variance_fn_ptr_t *fn_ptr,
1858                                           const MV *center_mv, int lambda,
1859                                           const int_mv *nb_full_mvs,
1860                                           int full_mv_num) {
1861   const MACROBLOCKD *const xd = &x->e_mbd;
1862   const struct buf_2d *src = &x->plane[0].src;
1863   const struct buf_2d *pre = &xd->plane[0].pre[0];
1864   assert(step >= 1);
1865   assert(is_mv_in(&x->mv_limits, center_mv));
1866   if (step == 1) {
1867     return exhaustive_mesh_search_single_step(
1868         best_mv, center_mv, range, src, pre, lambda, nb_full_mvs, full_mv_num,
1869         &x->mv_limits, fn_ptr);
1870   }
1871   return exhaustive_mesh_search_multi_step(best_mv, center_mv, range, step, src,
1872                                            pre, lambda, nb_full_mvs,
1873                                            full_mv_num, &x->mv_limits, fn_ptr);
1874 }
1875 
full_pixel_exhaustive_new(const VP9_COMP * cpi,MACROBLOCK * x,MV * centre_mv_full,const vp9_variance_fn_ptr_t * fn_ptr,MV * dst_mv,int lambda,const int_mv * nb_full_mvs,int full_mv_num)1876 static int64_t full_pixel_exhaustive_new(const VP9_COMP *cpi, MACROBLOCK *x,
1877                                          MV *centre_mv_full,
1878                                          const vp9_variance_fn_ptr_t *fn_ptr,
1879                                          MV *dst_mv, int lambda,
1880                                          const int_mv *nb_full_mvs,
1881                                          int full_mv_num) {
1882   const SPEED_FEATURES *const sf = &cpi->sf;
1883   MV temp_mv = { centre_mv_full->row, centre_mv_full->col };
1884   int64_t bestsme;
1885   int i;
1886   int interval = sf->mesh_patterns[0].interval;
1887   int range = sf->mesh_patterns[0].range;
1888   int baseline_interval_divisor;
1889 
1890   // Trap illegal values for interval and range for this function.
1891   if ((range < MIN_RANGE) || (range > MAX_RANGE) || (interval < MIN_INTERVAL) ||
1892       (interval > range)) {
1893     printf("ERROR: invalid range\n");
1894     assert(0);
1895   }
1896 
1897   baseline_interval_divisor = range / interval;
1898 
1899   // Check size of proposed first range against magnitude of the centre
1900   // value used as a starting point.
1901   range = VPXMAX(range, (5 * VPXMAX(abs(temp_mv.row), abs(temp_mv.col))) / 4);
1902   range = VPXMIN(range, MAX_RANGE);
1903   interval = VPXMAX(interval, range / baseline_interval_divisor);
1904 
1905   // initial search
1906   bestsme =
1907       exhaustive_mesh_search_new(x, &temp_mv, range, interval, fn_ptr, &temp_mv,
1908                                  lambda, nb_full_mvs, full_mv_num);
1909 
1910   if ((interval > MIN_INTERVAL) && (range > MIN_RANGE)) {
1911     // Progressive searches with range and step size decreasing each time
1912     // till we reach a step size of 1. Then break out.
1913     for (i = 1; i < MAX_MESH_STEP; ++i) {
1914       // First pass with coarser step and longer range
1915       bestsme = exhaustive_mesh_search_new(
1916           x, &temp_mv, sf->mesh_patterns[i].range,
1917           sf->mesh_patterns[i].interval, fn_ptr, &temp_mv, lambda, nb_full_mvs,
1918           full_mv_num);
1919 
1920       if (sf->mesh_patterns[i].interval == 1) break;
1921     }
1922   }
1923 
1924   *dst_mv = temp_mv;
1925 
1926   return bestsme;
1927 }
1928 
diamond_search_sad_new(const MACROBLOCK * x,const search_site_config * cfg,const MV * init_full_mv,MV * best_full_mv,int search_param,int lambda,int * num00,const vp9_variance_fn_ptr_t * fn_ptr,const int_mv * nb_full_mvs,int full_mv_num)1929 static int64_t diamond_search_sad_new(const MACROBLOCK *x,
1930                                       const search_site_config *cfg,
1931                                       const MV *init_full_mv, MV *best_full_mv,
1932                                       int search_param, int lambda, int *num00,
1933                                       const vp9_variance_fn_ptr_t *fn_ptr,
1934                                       const int_mv *nb_full_mvs,
1935                                       int full_mv_num) {
1936   int i, j, step;
1937 
1938   const MACROBLOCKD *const xd = &x->e_mbd;
1939   uint8_t *what = x->plane[0].src.buf;
1940   const int what_stride = x->plane[0].src.stride;
1941   const uint8_t *in_what;
1942   const int in_what_stride = xd->plane[0].pre[0].stride;
1943   const uint8_t *best_address;
1944 
1945   int64_t bestsad;
1946   int best_site = -1;
1947   int last_site = -1;
1948 
1949   // search_param determines the length of the initial step and hence the number
1950   // of iterations.
1951   // 0 = initial step (MAX_FIRST_STEP) pel
1952   // 1 = (MAX_FIRST_STEP/2) pel,
1953   // 2 = (MAX_FIRST_STEP/4) pel...
1954   //  const search_site *ss = &cfg->ss[search_param * cfg->searches_per_step];
1955   const MV *ss_mv = &cfg->ss_mv[search_param * cfg->searches_per_step];
1956   const intptr_t *ss_os = &cfg->ss_os[search_param * cfg->searches_per_step];
1957   const int tot_steps = cfg->total_steps - search_param;
1958   vpx_clear_system_state();
1959 
1960   *best_full_mv = *init_full_mv;
1961   clamp_mv(best_full_mv, x->mv_limits.col_min, x->mv_limits.col_max,
1962            x->mv_limits.row_min, x->mv_limits.row_max);
1963   *num00 = 0;
1964 
1965   // Work out the start point for the search
1966   in_what = xd->plane[0].pre[0].buf + best_full_mv->row * in_what_stride +
1967             best_full_mv->col;
1968   best_address = in_what;
1969 
1970   // Check the starting position
1971   {
1972     const int64_t mv_dist =
1973         (int64_t)fn_ptr->sdf(what, what_stride, in_what, in_what_stride)
1974         << LOG2_PRECISION;
1975     const int64_t mv_cost =
1976         vp9_nb_mvs_inconsistency(best_full_mv, nb_full_mvs, full_mv_num);
1977     bestsad = mv_dist + lambda * mv_cost;
1978   }
1979 
1980   i = 0;
1981 
1982   for (step = 0; step < tot_steps; step++) {
1983     int all_in = 1, t;
1984 
1985     // All_in is true if every one of the points we are checking are within
1986     // the bounds of the image.
1987     all_in &= ((best_full_mv->row + ss_mv[i].row) > x->mv_limits.row_min);
1988     all_in &= ((best_full_mv->row + ss_mv[i + 1].row) < x->mv_limits.row_max);
1989     all_in &= ((best_full_mv->col + ss_mv[i + 2].col) > x->mv_limits.col_min);
1990     all_in &= ((best_full_mv->col + ss_mv[i + 3].col) < x->mv_limits.col_max);
1991 
1992     // If all the pixels are within the bounds we don't check whether the
1993     // search point is valid in this loop,  otherwise we check each point
1994     // for validity..
1995     if (all_in) {
1996       unsigned int sad_array[4];
1997 
1998       for (j = 0; j < cfg->searches_per_step; j += 4) {
1999         unsigned char const *block_offset[4];
2000 
2001         for (t = 0; t < 4; t++) block_offset[t] = ss_os[i + t] + best_address;
2002 
2003         fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride,
2004                        sad_array);
2005 
2006         for (t = 0; t < 4; t++, i++) {
2007           const int64_t mv_dist = (int64_t)sad_array[t] << LOG2_PRECISION;
2008           if (mv_dist < bestsad) {
2009             const MV this_mv = { best_full_mv->row + ss_mv[i].row,
2010                                  best_full_mv->col + ss_mv[i].col };
2011             const int64_t mv_cost =
2012                 vp9_nb_mvs_inconsistency(&this_mv, nb_full_mvs, full_mv_num);
2013             const int64_t thissad = mv_dist + lambda * mv_cost;
2014             if (thissad < bestsad) {
2015               bestsad = thissad;
2016               best_site = i;
2017             }
2018           }
2019         }
2020       }
2021     } else {
2022       for (j = 0; j < cfg->searches_per_step; j++) {
2023         // Trap illegal vectors
2024         const MV this_mv = { best_full_mv->row + ss_mv[i].row,
2025                              best_full_mv->col + ss_mv[i].col };
2026 
2027         if (is_mv_in(&x->mv_limits, &this_mv)) {
2028           const uint8_t *const check_here = ss_os[i] + best_address;
2029           const int64_t mv_dist =
2030               (int64_t)fn_ptr->sdf(what, what_stride, check_here,
2031                                    in_what_stride)
2032               << LOG2_PRECISION;
2033           if (mv_dist < bestsad) {
2034             const int64_t mv_cost =
2035                 vp9_nb_mvs_inconsistency(&this_mv, nb_full_mvs, full_mv_num);
2036             const int64_t thissad = mv_dist + lambda * mv_cost;
2037             if (thissad < bestsad) {
2038               bestsad = thissad;
2039               best_site = i;
2040             }
2041           }
2042         }
2043         i++;
2044       }
2045     }
2046     if (best_site != last_site) {
2047       best_full_mv->row += ss_mv[best_site].row;
2048       best_full_mv->col += ss_mv[best_site].col;
2049       best_address += ss_os[best_site];
2050       last_site = best_site;
2051     } else if (best_address == in_what) {
2052       (*num00)++;
2053     }
2054   }
2055   return bestsad;
2056 }
2057 
vp9_prepare_nb_full_mvs(const MotionField * motion_field,int mi_row,int mi_col,int_mv * nb_full_mvs)2058 int vp9_prepare_nb_full_mvs(const MotionField *motion_field, int mi_row,
2059                             int mi_col, int_mv *nb_full_mvs) {
2060   const int mi_width = num_8x8_blocks_wide_lookup[motion_field->bsize];
2061   const int mi_height = num_8x8_blocks_high_lookup[motion_field->bsize];
2062   const int dirs[NB_MVS_NUM][2] = { { -1, 0 }, { 0, -1 }, { 1, 0 }, { 0, 1 } };
2063   int nb_full_mv_num = 0;
2064   int i;
2065   assert(mi_row % mi_height == 0);
2066   assert(mi_col % mi_width == 0);
2067   for (i = 0; i < NB_MVS_NUM; ++i) {
2068     int r = dirs[i][0];
2069     int c = dirs[i][1];
2070     int brow = mi_row / mi_height + r;
2071     int bcol = mi_col / mi_width + c;
2072     if (brow >= 0 && brow < motion_field->block_rows && bcol >= 0 &&
2073         bcol < motion_field->block_cols) {
2074       if (vp9_motion_field_is_mv_set(motion_field, brow, bcol)) {
2075         int_mv mv = vp9_motion_field_get_mv(motion_field, brow, bcol);
2076         nb_full_mvs[nb_full_mv_num].as_mv = get_full_mv(&mv.as_mv);
2077         ++nb_full_mv_num;
2078       }
2079     }
2080   }
2081   return nb_full_mv_num;
2082 }
2083 #endif  // CONFIG_NON_GREEDY_MV
2084 
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)2085 int vp9_diamond_search_sad_c(const MACROBLOCK *x, const search_site_config *cfg,
2086                              MV *ref_mv, MV *best_mv, int search_param,
2087                              int sad_per_bit, int *num00,
2088                              const vp9_variance_fn_ptr_t *fn_ptr,
2089                              const MV *center_mv) {
2090   int i, j, step;
2091 
2092   const MACROBLOCKD *const xd = &x->e_mbd;
2093   uint8_t *what = x->plane[0].src.buf;
2094   const int what_stride = x->plane[0].src.stride;
2095   const uint8_t *in_what;
2096   const int in_what_stride = xd->plane[0].pre[0].stride;
2097   const uint8_t *best_address;
2098 
2099   unsigned int bestsad = INT_MAX;
2100   int best_site = -1;
2101   int last_site = -1;
2102 
2103   int ref_row;
2104   int ref_col;
2105 
2106   // search_param determines the length of the initial step and hence the number
2107   // of iterations.
2108   // 0 = initial step (MAX_FIRST_STEP) pel
2109   // 1 = (MAX_FIRST_STEP/2) pel,
2110   // 2 = (MAX_FIRST_STEP/4) pel...
2111   //  const search_site *ss = &cfg->ss[search_param * cfg->searches_per_step];
2112   const MV *ss_mv = &cfg->ss_mv[search_param * cfg->searches_per_step];
2113   const intptr_t *ss_os = &cfg->ss_os[search_param * cfg->searches_per_step];
2114   const int tot_steps = cfg->total_steps - search_param;
2115 
2116   const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
2117   clamp_mv(ref_mv, x->mv_limits.col_min, x->mv_limits.col_max,
2118            x->mv_limits.row_min, x->mv_limits.row_max);
2119   ref_row = ref_mv->row;
2120   ref_col = ref_mv->col;
2121   *num00 = 0;
2122   best_mv->row = ref_row;
2123   best_mv->col = ref_col;
2124 
2125   // Work out the start point for the search
2126   in_what = xd->plane[0].pre[0].buf + ref_row * in_what_stride + ref_col;
2127   best_address = in_what;
2128 
2129   // Check the starting position
2130   bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride) +
2131             mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit);
2132 
2133   i = 0;
2134 
2135   for (step = 0; step < tot_steps; step++) {
2136     int all_in = 1, t;
2137 
2138     // All_in is true if every one of the points we are checking are within
2139     // the bounds of the image.
2140     all_in &= ((best_mv->row + ss_mv[i].row) > x->mv_limits.row_min);
2141     all_in &= ((best_mv->row + ss_mv[i + 1].row) < x->mv_limits.row_max);
2142     all_in &= ((best_mv->col + ss_mv[i + 2].col) > x->mv_limits.col_min);
2143     all_in &= ((best_mv->col + ss_mv[i + 3].col) < x->mv_limits.col_max);
2144 
2145     // If all the pixels are within the bounds we don't check whether the
2146     // search point is valid in this loop,  otherwise we check each point
2147     // for validity..
2148     if (all_in) {
2149       unsigned int sad_array[4];
2150 
2151       for (j = 0; j < cfg->searches_per_step; j += 4) {
2152         unsigned char const *block_offset[4];
2153 
2154         for (t = 0; t < 4; t++) block_offset[t] = ss_os[i + t] + best_address;
2155 
2156         fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride,
2157                        sad_array);
2158 
2159         for (t = 0; t < 4; t++, i++) {
2160           if (sad_array[t] < bestsad) {
2161             const MV this_mv = { best_mv->row + ss_mv[i].row,
2162                                  best_mv->col + ss_mv[i].col };
2163             sad_array[t] +=
2164                 mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
2165             if (sad_array[t] < bestsad) {
2166               bestsad = sad_array[t];
2167               best_site = i;
2168             }
2169           }
2170         }
2171       }
2172     } else {
2173       for (j = 0; j < cfg->searches_per_step; j++) {
2174         // Trap illegal vectors
2175         const MV this_mv = { best_mv->row + ss_mv[i].row,
2176                              best_mv->col + ss_mv[i].col };
2177 
2178         if (is_mv_in(&x->mv_limits, &this_mv)) {
2179           const uint8_t *const check_here = ss_os[i] + best_address;
2180           unsigned int thissad =
2181               fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
2182 
2183           if (thissad < bestsad) {
2184             thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
2185             if (thissad < bestsad) {
2186               bestsad = thissad;
2187               best_site = i;
2188             }
2189           }
2190         }
2191         i++;
2192       }
2193     }
2194     if (best_site != last_site) {
2195       best_mv->row += ss_mv[best_site].row;
2196       best_mv->col += ss_mv[best_site].col;
2197       best_address += ss_os[best_site];
2198       last_site = best_site;
2199 #if defined(NEW_DIAMOND_SEARCH)
2200       while (1) {
2201         const MV this_mv = { best_mv->row + ss_mv[best_site].row,
2202                              best_mv->col + ss_mv[best_site].col };
2203         if (is_mv_in(&x->mv_limits, &this_mv)) {
2204           const uint8_t *const check_here = ss_os[best_site] + best_address;
2205           unsigned int thissad =
2206               fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
2207           if (thissad < bestsad) {
2208             thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
2209             if (thissad < bestsad) {
2210               bestsad = thissad;
2211               best_mv->row += ss_mv[best_site].row;
2212               best_mv->col += ss_mv[best_site].col;
2213               best_address += ss_os[best_site];
2214               continue;
2215             }
2216           }
2217         }
2218         break;
2219       }
2220 #endif
2221     } else if (best_address == in_what) {
2222       (*num00)++;
2223     }
2224   }
2225   return bestsad;
2226 }
2227 
vector_match(int16_t * ref,int16_t * src,int bwl)2228 static int vector_match(int16_t *ref, int16_t *src, int bwl) {
2229   int best_sad = INT_MAX;
2230   int this_sad;
2231   int d;
2232   int center, offset = 0;
2233   int bw = 4 << bwl;  // redundant variable, to be changed in the experiments.
2234   for (d = 0; d <= bw; d += 16) {
2235     this_sad = vpx_vector_var(&ref[d], src, bwl);
2236     if (this_sad < best_sad) {
2237       best_sad = this_sad;
2238       offset = d;
2239     }
2240   }
2241   center = offset;
2242 
2243   for (d = -8; d <= 8; d += 16) {
2244     int this_pos = offset + d;
2245     // check limit
2246     if (this_pos < 0 || this_pos > bw) continue;
2247     this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
2248     if (this_sad < best_sad) {
2249       best_sad = this_sad;
2250       center = this_pos;
2251     }
2252   }
2253   offset = center;
2254 
2255   for (d = -4; d <= 4; d += 8) {
2256     int this_pos = offset + d;
2257     // check limit
2258     if (this_pos < 0 || this_pos > bw) continue;
2259     this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
2260     if (this_sad < best_sad) {
2261       best_sad = this_sad;
2262       center = this_pos;
2263     }
2264   }
2265   offset = center;
2266 
2267   for (d = -2; d <= 2; d += 4) {
2268     int this_pos = offset + d;
2269     // check limit
2270     if (this_pos < 0 || this_pos > bw) continue;
2271     this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
2272     if (this_sad < best_sad) {
2273       best_sad = this_sad;
2274       center = this_pos;
2275     }
2276   }
2277   offset = center;
2278 
2279   for (d = -1; d <= 1; d += 2) {
2280     int this_pos = offset + d;
2281     // check limit
2282     if (this_pos < 0 || this_pos > bw) continue;
2283     this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
2284     if (this_sad < best_sad) {
2285       best_sad = this_sad;
2286       center = this_pos;
2287     }
2288   }
2289 
2290   return (center - (bw >> 1));
2291 }
2292 
2293 static const MV search_pos[4] = {
2294   { -1, 0 },
2295   { 0, -1 },
2296   { 0, 1 },
2297   { 1, 0 },
2298 };
2299 
vp9_int_pro_motion_estimation(const VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int mi_row,int mi_col,const MV * ref_mv)2300 unsigned int vp9_int_pro_motion_estimation(const VP9_COMP *cpi, MACROBLOCK *x,
2301                                            BLOCK_SIZE bsize, int mi_row,
2302                                            int mi_col, const MV *ref_mv) {
2303   MACROBLOCKD *xd = &x->e_mbd;
2304   MODE_INFO *mi = xd->mi[0];
2305   struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
2306   DECLARE_ALIGNED(16, int16_t, hbuf[128]);
2307   DECLARE_ALIGNED(16, int16_t, vbuf[128]);
2308   DECLARE_ALIGNED(16, int16_t, src_hbuf[64]);
2309   DECLARE_ALIGNED(16, int16_t, src_vbuf[64]);
2310   int idx;
2311   const int bw = 4 << b_width_log2_lookup[bsize];
2312   const int bh = 4 << b_height_log2_lookup[bsize];
2313   const int search_width = bw << 1;
2314   const int search_height = bh << 1;
2315   const int src_stride = x->plane[0].src.stride;
2316   const int ref_stride = xd->plane[0].pre[0].stride;
2317   uint8_t const *ref_buf, *src_buf;
2318   MV *tmp_mv = &xd->mi[0]->mv[0].as_mv;
2319   unsigned int best_sad, tmp_sad, this_sad[4];
2320   MV this_mv;
2321   const int norm_factor = 3 + (bw >> 5);
2322   const YV12_BUFFER_CONFIG *scaled_ref_frame =
2323       vp9_get_scaled_ref_frame(cpi, mi->ref_frame[0]);
2324   MvLimits subpel_mv_limits;
2325 
2326   if (scaled_ref_frame) {
2327     int i;
2328     // Swap out the reference frame for a version that's been scaled to
2329     // match the resolution of the current frame, allowing the existing
2330     // motion search code to be used without additional modifications.
2331     for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
2332     vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2333   }
2334 
2335 #if CONFIG_VP9_HIGHBITDEPTH
2336   // TODO(jingning): Implement integral projection functions for high bit-depth
2337   // setting and remove this part of code.
2338   if (xd->bd != 8) {
2339     unsigned int this_sad;
2340     tmp_mv->row = 0;
2341     tmp_mv->col = 0;
2342     this_sad = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf, src_stride,
2343                                       xd->plane[0].pre[0].buf, ref_stride);
2344 
2345     if (scaled_ref_frame) {
2346       int i;
2347       for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
2348     }
2349     return this_sad;
2350   }
2351 #endif
2352 
2353   // Set up prediction 1-D reference set
2354   ref_buf = xd->plane[0].pre[0].buf - (bw >> 1);
2355   for (idx = 0; idx < search_width; idx += 16) {
2356     vpx_int_pro_row(&hbuf[idx], ref_buf, ref_stride, bh);
2357     ref_buf += 16;
2358   }
2359 
2360   ref_buf = xd->plane[0].pre[0].buf - (bh >> 1) * ref_stride;
2361   for (idx = 0; idx < search_height; ++idx) {
2362     vbuf[idx] = vpx_int_pro_col(ref_buf, bw) >> norm_factor;
2363     ref_buf += ref_stride;
2364   }
2365 
2366   // Set up src 1-D reference set
2367   for (idx = 0; idx < bw; idx += 16) {
2368     src_buf = x->plane[0].src.buf + idx;
2369     vpx_int_pro_row(&src_hbuf[idx], src_buf, src_stride, bh);
2370   }
2371 
2372   src_buf = x->plane[0].src.buf;
2373   for (idx = 0; idx < bh; ++idx) {
2374     src_vbuf[idx] = vpx_int_pro_col(src_buf, bw) >> norm_factor;
2375     src_buf += src_stride;
2376   }
2377 
2378   // Find the best match per 1-D search
2379   tmp_mv->col = vector_match(hbuf, src_hbuf, b_width_log2_lookup[bsize]);
2380   tmp_mv->row = vector_match(vbuf, src_vbuf, b_height_log2_lookup[bsize]);
2381 
2382   this_mv = *tmp_mv;
2383   src_buf = x->plane[0].src.buf;
2384   ref_buf = xd->plane[0].pre[0].buf + this_mv.row * ref_stride + this_mv.col;
2385   best_sad = cpi->fn_ptr[bsize].sdf(src_buf, src_stride, ref_buf, ref_stride);
2386 
2387   {
2388     const uint8_t *const pos[4] = {
2389       ref_buf - ref_stride,
2390       ref_buf - 1,
2391       ref_buf + 1,
2392       ref_buf + ref_stride,
2393     };
2394 
2395     cpi->fn_ptr[bsize].sdx4df(src_buf, src_stride, pos, ref_stride, this_sad);
2396   }
2397 
2398   for (idx = 0; idx < 4; ++idx) {
2399     if (this_sad[idx] < best_sad) {
2400       best_sad = this_sad[idx];
2401       tmp_mv->row = search_pos[idx].row + this_mv.row;
2402       tmp_mv->col = search_pos[idx].col + this_mv.col;
2403     }
2404   }
2405 
2406   if (this_sad[0] < this_sad[3])
2407     this_mv.row -= 1;
2408   else
2409     this_mv.row += 1;
2410 
2411   if (this_sad[1] < this_sad[2])
2412     this_mv.col -= 1;
2413   else
2414     this_mv.col += 1;
2415 
2416   ref_buf = xd->plane[0].pre[0].buf + this_mv.row * ref_stride + this_mv.col;
2417 
2418   tmp_sad = cpi->fn_ptr[bsize].sdf(src_buf, src_stride, ref_buf, ref_stride);
2419   if (best_sad > tmp_sad) {
2420     *tmp_mv = this_mv;
2421     best_sad = tmp_sad;
2422   }
2423 
2424   tmp_mv->row *= 8;
2425   tmp_mv->col *= 8;
2426 
2427   vp9_set_subpel_mv_search_range(&subpel_mv_limits, &x->mv_limits, ref_mv);
2428   clamp_mv(tmp_mv, subpel_mv_limits.col_min, subpel_mv_limits.col_max,
2429            subpel_mv_limits.row_min, subpel_mv_limits.row_max);
2430 
2431   if (scaled_ref_frame) {
2432     int i;
2433     for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
2434   }
2435 
2436   return best_sad;
2437 }
2438 
get_exhaustive_threshold(int exhaustive_searches_thresh,BLOCK_SIZE bsize)2439 static int get_exhaustive_threshold(int exhaustive_searches_thresh,
2440                                     BLOCK_SIZE bsize) {
2441   return exhaustive_searches_thresh >>
2442          (8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]));
2443 }
2444 
2445 #if CONFIG_NON_GREEDY_MV
2446 // Runs sequence of diamond searches in smaller steps for RD.
2447 /* do_refine: If last step (1-away) of n-step search doesn't pick the center
2448               point as the best match, we will do a final 1-away diamond
2449               refining search  */
vp9_full_pixel_diamond_new(const VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,MV * mvp_full,int step_param,int lambda,int do_refine,const int_mv * nb_full_mvs,int full_mv_num,MV * best_mv)2450 int vp9_full_pixel_diamond_new(const VP9_COMP *cpi, MACROBLOCK *x,
2451                                BLOCK_SIZE bsize, MV *mvp_full, int step_param,
2452                                int lambda, int do_refine,
2453                                const int_mv *nb_full_mvs, int full_mv_num,
2454                                MV *best_mv) {
2455   const vp9_variance_fn_ptr_t *fn_ptr = &cpi->fn_ptr[bsize];
2456   const SPEED_FEATURES *const sf = &cpi->sf;
2457   int n, num00 = 0;
2458   int thissme;
2459   int bestsme;
2460   const int further_steps = MAX_MVSEARCH_STEPS - 1 - step_param;
2461   const MV center_mv = { 0, 0 };
2462   vpx_clear_system_state();
2463   diamond_search_sad_new(x, &cpi->ss_cfg, mvp_full, best_mv, step_param, lambda,
2464                          &n, fn_ptr, nb_full_mvs, full_mv_num);
2465 
2466   bestsme = vp9_get_mvpred_var(x, best_mv, &center_mv, fn_ptr, 0);
2467 
2468   // If there won't be more n-step search, check to see if refining search is
2469   // needed.
2470   if (n > further_steps) do_refine = 0;
2471 
2472   while (n < further_steps) {
2473     ++n;
2474     if (num00) {
2475       num00--;
2476     } else {
2477       MV temp_mv;
2478       diamond_search_sad_new(x, &cpi->ss_cfg, mvp_full, &temp_mv,
2479                              step_param + n, lambda, &num00, fn_ptr,
2480                              nb_full_mvs, full_mv_num);
2481       thissme = vp9_get_mvpred_var(x, &temp_mv, &center_mv, fn_ptr, 0);
2482       // check to see if refining search is needed.
2483       if (num00 > further_steps - n) do_refine = 0;
2484 
2485       if (thissme < bestsme) {
2486         bestsme = thissme;
2487         *best_mv = temp_mv;
2488       }
2489     }
2490   }
2491 
2492   // final 1-away diamond refining search
2493   if (do_refine) {
2494     const int search_range = 8;
2495     MV temp_mv = *best_mv;
2496     vp9_refining_search_sad_new(x, &temp_mv, lambda, search_range, fn_ptr,
2497                                 nb_full_mvs, full_mv_num);
2498     thissme = vp9_get_mvpred_var(x, &temp_mv, &center_mv, fn_ptr, 0);
2499     if (thissme < bestsme) {
2500       bestsme = thissme;
2501       *best_mv = temp_mv;
2502     }
2503   }
2504 
2505   if (sf->exhaustive_searches_thresh < INT_MAX &&
2506       !cpi->rc.is_src_frame_alt_ref) {
2507     const int64_t exhaustive_thr =
2508         get_exhaustive_threshold(sf->exhaustive_searches_thresh, bsize);
2509     if (bestsme > exhaustive_thr) {
2510       full_pixel_exhaustive_new(cpi, x, best_mv, fn_ptr, best_mv, lambda,
2511                                 nb_full_mvs, full_mv_num);
2512       bestsme = vp9_get_mvpred_var(x, best_mv, &center_mv, fn_ptr, 0);
2513     }
2514   }
2515   return bestsme;
2516 }
2517 #endif  // CONFIG_NON_GREEDY_MV
2518 
2519 // Runs sequence of diamond searches in smaller steps for RD.
2520 /* do_refine: If last step (1-away) of n-step search doesn't pick the center
2521               point as the best match, we will do a final 1-away diamond
2522               refining search  */
full_pixel_diamond(const VP9_COMP * const cpi,const MACROBLOCK * const 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)2523 static int full_pixel_diamond(const VP9_COMP *const cpi,
2524                               const MACROBLOCK *const x, MV *mvp_full,
2525                               int step_param, int sadpb, int further_steps,
2526                               int do_refine, int *cost_list,
2527                               const vp9_variance_fn_ptr_t *fn_ptr,
2528                               const MV *ref_mv, MV *dst_mv) {
2529   MV temp_mv;
2530   int thissme, n, num00 = 0;
2531   int bestsme = cpi->diamond_search_sad(x, &cpi->ss_cfg, mvp_full, &temp_mv,
2532                                         step_param, sadpb, &n, fn_ptr, ref_mv);
2533   if (bestsme < INT_MAX)
2534     bestsme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
2535   *dst_mv = temp_mv;
2536 
2537   // If there won't be more n-step search, check to see if refining search is
2538   // needed.
2539   if (n > further_steps) do_refine = 0;
2540 
2541   while (n < further_steps) {
2542     ++n;
2543 
2544     if (num00) {
2545       num00--;
2546     } else {
2547       thissme = cpi->diamond_search_sad(x, &cpi->ss_cfg, mvp_full, &temp_mv,
2548                                         step_param + n, sadpb, &num00, fn_ptr,
2549                                         ref_mv);
2550       if (thissme < INT_MAX)
2551         thissme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
2552 
2553       // check to see if refining search is needed.
2554       if (num00 > further_steps - n) do_refine = 0;
2555 
2556       if (thissme < bestsme) {
2557         bestsme = thissme;
2558         *dst_mv = temp_mv;
2559       }
2560     }
2561   }
2562 
2563   // final 1-away diamond refining search
2564   if (do_refine) {
2565     const int search_range = 8;
2566     MV best_mv = *dst_mv;
2567     thissme = vp9_refining_search_sad(x, &best_mv, sadpb, search_range, fn_ptr,
2568                                       ref_mv);
2569     if (thissme < INT_MAX)
2570       thissme = vp9_get_mvpred_var(x, &best_mv, ref_mv, fn_ptr, 1);
2571     if (thissme < bestsme) {
2572       bestsme = thissme;
2573       *dst_mv = best_mv;
2574     }
2575   }
2576 
2577   // Return cost list.
2578   if (cost_list) {
2579     calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, dst_mv, cost_list);
2580   }
2581   return bestsme;
2582 }
2583 
2584 // Runs an limited range exhaustive mesh search using a pattern set
2585 // according to the encode speed profile.
full_pixel_exhaustive(const VP9_COMP * const cpi,const MACROBLOCK * const 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)2586 static int full_pixel_exhaustive(const VP9_COMP *const cpi,
2587                                  const MACROBLOCK *const x, MV *centre_mv_full,
2588                                  int sadpb, int *cost_list,
2589                                  const vp9_variance_fn_ptr_t *fn_ptr,
2590                                  const MV *ref_mv, MV *dst_mv) {
2591   const SPEED_FEATURES *const sf = &cpi->sf;
2592   MV temp_mv = { centre_mv_full->row, centre_mv_full->col };
2593   MV f_ref_mv = { ref_mv->row >> 3, ref_mv->col >> 3 };
2594   int bestsme;
2595   int i;
2596   int interval = sf->mesh_patterns[0].interval;
2597   int range = sf->mesh_patterns[0].range;
2598   int baseline_interval_divisor;
2599 
2600   // Trap illegal values for interval and range for this function.
2601   if ((range < MIN_RANGE) || (range > MAX_RANGE) || (interval < MIN_INTERVAL) ||
2602       (interval > range))
2603     return INT_MAX;
2604 
2605   baseline_interval_divisor = range / interval;
2606 
2607   // Check size of proposed first range against magnitude of the centre
2608   // value used as a starting point.
2609   range = VPXMAX(range, (5 * VPXMAX(abs(temp_mv.row), abs(temp_mv.col))) / 4);
2610   range = VPXMIN(range, MAX_RANGE);
2611   interval = VPXMAX(interval, range / baseline_interval_divisor);
2612 
2613   // initial search
2614   bestsme = exhaustive_mesh_search(x, &f_ref_mv, &temp_mv, range, interval,
2615                                    sadpb, fn_ptr, &temp_mv);
2616 
2617   if ((interval > MIN_INTERVAL) && (range > MIN_RANGE)) {
2618     // Progressive searches with range and step size decreasing each time
2619     // till we reach a step size of 1. Then break out.
2620     for (i = 1; i < MAX_MESH_STEP; ++i) {
2621       // First pass with coarser step and longer range
2622       bestsme = exhaustive_mesh_search(
2623           x, &f_ref_mv, &temp_mv, sf->mesh_patterns[i].range,
2624           sf->mesh_patterns[i].interval, sadpb, fn_ptr, &temp_mv);
2625 
2626       if (sf->mesh_patterns[i].interval == 1) break;
2627     }
2628   }
2629 
2630   if (bestsme < INT_MAX)
2631     bestsme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
2632   *dst_mv = temp_mv;
2633 
2634   // Return cost list.
2635   if (cost_list) {
2636     calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, dst_mv, cost_list);
2637   }
2638   return bestsme;
2639 }
2640 
2641 #if CONFIG_NON_GREEDY_MV
vp9_refining_search_sad_new(const MACROBLOCK * x,MV * best_full_mv,int lambda,int search_range,const vp9_variance_fn_ptr_t * fn_ptr,const int_mv * nb_full_mvs,int full_mv_num)2642 int64_t vp9_refining_search_sad_new(const MACROBLOCK *x, MV *best_full_mv,
2643                                     int lambda, int search_range,
2644                                     const vp9_variance_fn_ptr_t *fn_ptr,
2645                                     const int_mv *nb_full_mvs,
2646                                     int full_mv_num) {
2647   const MACROBLOCKD *const xd = &x->e_mbd;
2648   const MV neighbors[4] = { { -1, 0 }, { 0, -1 }, { 0, 1 }, { 1, 0 } };
2649   const struct buf_2d *const what = &x->plane[0].src;
2650   const struct buf_2d *const in_what = &xd->plane[0].pre[0];
2651   const uint8_t *best_address = get_buf_from_mv(in_what, best_full_mv);
2652   int64_t best_sad;
2653   int i, j;
2654   vpx_clear_system_state();
2655   {
2656     const int64_t mv_dist = (int64_t)fn_ptr->sdf(what->buf, what->stride,
2657                                                  best_address, in_what->stride)
2658                             << LOG2_PRECISION;
2659     const int64_t mv_cost =
2660         vp9_nb_mvs_inconsistency(best_full_mv, nb_full_mvs, full_mv_num);
2661     best_sad = mv_dist + lambda * mv_cost;
2662   }
2663 
2664   for (i = 0; i < search_range; i++) {
2665     int best_site = -1;
2666     const int all_in = ((best_full_mv->row - 1) > x->mv_limits.row_min) &
2667                        ((best_full_mv->row + 1) < x->mv_limits.row_max) &
2668                        ((best_full_mv->col - 1) > x->mv_limits.col_min) &
2669                        ((best_full_mv->col + 1) < x->mv_limits.col_max);
2670 
2671     if (all_in) {
2672       unsigned int sads[4];
2673       const uint8_t *const positions[4] = { best_address - in_what->stride,
2674                                             best_address - 1, best_address + 1,
2675                                             best_address + in_what->stride };
2676 
2677       fn_ptr->sdx4df(what->buf, what->stride, positions, in_what->stride, sads);
2678 
2679       for (j = 0; j < 4; ++j) {
2680         const MV mv = { best_full_mv->row + neighbors[j].row,
2681                         best_full_mv->col + neighbors[j].col };
2682         const int64_t mv_dist = (int64_t)sads[j] << LOG2_PRECISION;
2683         const int64_t mv_cost =
2684             vp9_nb_mvs_inconsistency(&mv, nb_full_mvs, full_mv_num);
2685         const int64_t thissad = mv_dist + lambda * mv_cost;
2686         if (thissad < best_sad) {
2687           best_sad = thissad;
2688           best_site = j;
2689         }
2690       }
2691     } else {
2692       for (j = 0; j < 4; ++j) {
2693         const MV mv = { best_full_mv->row + neighbors[j].row,
2694                         best_full_mv->col + neighbors[j].col };
2695 
2696         if (is_mv_in(&x->mv_limits, &mv)) {
2697           const int64_t mv_dist =
2698               (int64_t)fn_ptr->sdf(what->buf, what->stride,
2699                                    get_buf_from_mv(in_what, &mv),
2700                                    in_what->stride)
2701               << LOG2_PRECISION;
2702           const int64_t mv_cost =
2703               vp9_nb_mvs_inconsistency(&mv, nb_full_mvs, full_mv_num);
2704           const int64_t thissad = mv_dist + lambda * mv_cost;
2705           if (thissad < best_sad) {
2706             best_sad = thissad;
2707             best_site = j;
2708           }
2709         }
2710       }
2711     }
2712 
2713     if (best_site == -1) {
2714       break;
2715     } else {
2716       best_full_mv->row += neighbors[best_site].row;
2717       best_full_mv->col += neighbors[best_site].col;
2718       best_address = get_buf_from_mv(in_what, best_full_mv);
2719     }
2720   }
2721 
2722   return best_sad;
2723 }
2724 #endif  // CONFIG_NON_GREEDY_MV
2725 
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)2726 int vp9_refining_search_sad(const MACROBLOCK *x, MV *ref_mv, int error_per_bit,
2727                             int search_range,
2728                             const vp9_variance_fn_ptr_t *fn_ptr,
2729                             const MV *center_mv) {
2730   const MACROBLOCKD *const xd = &x->e_mbd;
2731   const MV neighbors[4] = { { -1, 0 }, { 0, -1 }, { 0, 1 }, { 1, 0 } };
2732   const struct buf_2d *const what = &x->plane[0].src;
2733   const struct buf_2d *const in_what = &xd->plane[0].pre[0];
2734   const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
2735   const uint8_t *best_address = get_buf_from_mv(in_what, ref_mv);
2736   unsigned int best_sad =
2737       fn_ptr->sdf(what->buf, what->stride, best_address, in_what->stride) +
2738       mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
2739   int i, j;
2740 
2741   for (i = 0; i < search_range; i++) {
2742     int best_site = -1;
2743     const int all_in = ((ref_mv->row - 1) > x->mv_limits.row_min) &
2744                        ((ref_mv->row + 1) < x->mv_limits.row_max) &
2745                        ((ref_mv->col - 1) > x->mv_limits.col_min) &
2746                        ((ref_mv->col + 1) < x->mv_limits.col_max);
2747 
2748     if (all_in) {
2749       unsigned int sads[4];
2750       const uint8_t *const positions[4] = { best_address - in_what->stride,
2751                                             best_address - 1, best_address + 1,
2752                                             best_address + in_what->stride };
2753 
2754       fn_ptr->sdx4df(what->buf, what->stride, positions, in_what->stride, sads);
2755 
2756       for (j = 0; j < 4; ++j) {
2757         if (sads[j] < best_sad) {
2758           const MV mv = { ref_mv->row + neighbors[j].row,
2759                           ref_mv->col + neighbors[j].col };
2760           sads[j] += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
2761           if (sads[j] < best_sad) {
2762             best_sad = sads[j];
2763             best_site = j;
2764           }
2765         }
2766       }
2767     } else {
2768       for (j = 0; j < 4; ++j) {
2769         const MV mv = { ref_mv->row + neighbors[j].row,
2770                         ref_mv->col + neighbors[j].col };
2771 
2772         if (is_mv_in(&x->mv_limits, &mv)) {
2773           unsigned int sad =
2774               fn_ptr->sdf(what->buf, what->stride,
2775                           get_buf_from_mv(in_what, &mv), in_what->stride);
2776           if (sad < best_sad) {
2777             sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
2778             if (sad < best_sad) {
2779               best_sad = sad;
2780               best_site = j;
2781             }
2782           }
2783         }
2784       }
2785     }
2786 
2787     if (best_site == -1) {
2788       break;
2789     } else {
2790       ref_mv->row += neighbors[best_site].row;
2791       ref_mv->col += neighbors[best_site].col;
2792       best_address = get_buf_from_mv(in_what, ref_mv);
2793     }
2794   }
2795 
2796   return best_sad;
2797 }
2798 
2799 // This function is called when we do joint motion search in comp_inter_inter
2800 // 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)2801 int vp9_refining_search_8p_c(const MACROBLOCK *x, MV *ref_mv, int error_per_bit,
2802                              int search_range,
2803                              const vp9_variance_fn_ptr_t *fn_ptr,
2804                              const MV *center_mv, const uint8_t *second_pred) {
2805   const MV neighbors[8] = { { -1, 0 },  { 0, -1 }, { 0, 1 },  { 1, 0 },
2806                             { -1, -1 }, { 1, -1 }, { -1, 1 }, { 1, 1 } };
2807   const MACROBLOCKD *const xd = &x->e_mbd;
2808   const struct buf_2d *const what = &x->plane[0].src;
2809   const struct buf_2d *const in_what = &xd->plane[0].pre[0];
2810   const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
2811   unsigned int best_sad = INT_MAX;
2812   int i, j;
2813   clamp_mv(ref_mv, x->mv_limits.col_min, x->mv_limits.col_max,
2814            x->mv_limits.row_min, x->mv_limits.row_max);
2815   best_sad =
2816       fn_ptr->sdaf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv),
2817                    in_what->stride, second_pred) +
2818       mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
2819 
2820   for (i = 0; i < search_range; ++i) {
2821     int best_site = -1;
2822 
2823     for (j = 0; j < 8; ++j) {
2824       const MV mv = { ref_mv->row + neighbors[j].row,
2825                       ref_mv->col + neighbors[j].col };
2826 
2827       if (is_mv_in(&x->mv_limits, &mv)) {
2828         unsigned int sad =
2829             fn_ptr->sdaf(what->buf, what->stride, get_buf_from_mv(in_what, &mv),
2830                          in_what->stride, second_pred);
2831         if (sad < best_sad) {
2832           sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
2833           if (sad < best_sad) {
2834             best_sad = sad;
2835             best_site = j;
2836           }
2837         }
2838       }
2839     }
2840 
2841     if (best_site == -1) {
2842       break;
2843     } else {
2844       ref_mv->row += neighbors[best_site].row;
2845       ref_mv->col += neighbors[best_site].col;
2846     }
2847   }
2848   return best_sad;
2849 }
2850 
vp9_full_pixel_search(const VP9_COMP * const cpi,const MACROBLOCK * const 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)2851 int vp9_full_pixel_search(const VP9_COMP *const cpi, const MACROBLOCK *const x,
2852                           BLOCK_SIZE bsize, MV *mvp_full, int step_param,
2853                           int search_method, int error_per_bit, int *cost_list,
2854                           const MV *ref_mv, MV *tmp_mv, int var_max, int rd) {
2855   const SPEED_FEATURES *const sf = &cpi->sf;
2856   const SEARCH_METHODS method = (SEARCH_METHODS)search_method;
2857   const vp9_variance_fn_ptr_t *fn_ptr = &cpi->fn_ptr[bsize];
2858   int var = 0;
2859   int run_exhaustive_search = 0;
2860 
2861   if (cost_list) {
2862     cost_list[0] = INT_MAX;
2863     cost_list[1] = INT_MAX;
2864     cost_list[2] = INT_MAX;
2865     cost_list[3] = INT_MAX;
2866     cost_list[4] = INT_MAX;
2867   }
2868 
2869   switch (method) {
2870     case FAST_DIAMOND:
2871       var = fast_dia_search(x, mvp_full, step_param, error_per_bit, 0,
2872                             cost_list, fn_ptr, 1, ref_mv, tmp_mv);
2873       break;
2874     case FAST_HEX:
2875       var = fast_hex_search(x, mvp_full, step_param, error_per_bit, 0,
2876                             cost_list, fn_ptr, 1, ref_mv, tmp_mv);
2877       break;
2878     case HEX:
2879       var = hex_search(x, mvp_full, step_param, error_per_bit, 1, cost_list,
2880                        fn_ptr, 1, ref_mv, tmp_mv);
2881       break;
2882     case SQUARE:
2883       var = square_search(x, mvp_full, step_param, error_per_bit, 1, cost_list,
2884                           fn_ptr, 1, ref_mv, tmp_mv);
2885       break;
2886     case BIGDIA:
2887       var = bigdia_search(x, mvp_full, step_param, error_per_bit, 1, cost_list,
2888                           fn_ptr, 1, ref_mv, tmp_mv);
2889       break;
2890     case NSTEP:
2891     case MESH:
2892       var = full_pixel_diamond(cpi, x, mvp_full, step_param, error_per_bit,
2893                                MAX_MVSEARCH_STEPS - 1 - step_param, 1,
2894                                cost_list, fn_ptr, ref_mv, tmp_mv);
2895       break;
2896     default: assert(0 && "Unknown search method");
2897   }
2898 
2899   if (method == NSTEP) {
2900     if (sf->exhaustive_searches_thresh < INT_MAX &&
2901         !cpi->rc.is_src_frame_alt_ref) {
2902       const int64_t exhaustive_thr =
2903           get_exhaustive_threshold(sf->exhaustive_searches_thresh, bsize);
2904       if (var > exhaustive_thr) {
2905         run_exhaustive_search = 1;
2906       }
2907     }
2908   } else if (method == MESH) {
2909     run_exhaustive_search = 1;
2910   }
2911 
2912   if (run_exhaustive_search) {
2913     int var_ex;
2914     MV tmp_mv_ex;
2915     var_ex = full_pixel_exhaustive(cpi, x, tmp_mv, error_per_bit, cost_list,
2916                                    fn_ptr, ref_mv, &tmp_mv_ex);
2917     if (var_ex < var) {
2918       var = var_ex;
2919       *tmp_mv = tmp_mv_ex;
2920     }
2921   }
2922 
2923   if (method != NSTEP && method != MESH && rd && var < var_max)
2924     var = vp9_get_mvpred_var(x, tmp_mv, ref_mv, fn_ptr, 1);
2925 
2926   return var;
2927 }
2928 
2929 // Note(yunqingwang): The following 2 functions are only used in the motion
2930 // vector unit test, which return extreme motion vectors allowed by the MV
2931 // limits.
2932 #define COMMON_MV_TEST \
2933   SETUP_SUBPEL_SEARCH; \
2934                        \
2935   (void)error_per_bit; \
2936   (void)vfp;           \
2937   (void)z;             \
2938   (void)src_stride;    \
2939   (void)y;             \
2940   (void)y_stride;      \
2941   (void)second_pred;   \
2942   (void)w;             \
2943   (void)h;             \
2944   (void)offset;        \
2945   (void)mvjcost;       \
2946   (void)mvcost;        \
2947   (void)sse1;          \
2948   (void)distortion;    \
2949                        \
2950   (void)halfiters;     \
2951   (void)quarteriters;  \
2952   (void)eighthiters;   \
2953   (void)whichdir;      \
2954   (void)allow_hp;      \
2955   (void)forced_stop;   \
2956   (void)hstep;         \
2957   (void)rr;            \
2958   (void)rc;            \
2959                        \
2960   (void)tr;            \
2961   (void)tc;            \
2962   (void)sse;           \
2963   (void)thismse;       \
2964   (void)cost_list;     \
2965   (void)use_accurate_subpel_search;
2966 
2967 // 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)2968 uint32_t vp9_return_max_sub_pixel_mv(
2969     const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
2970     int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
2971     int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
2972     uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
2973     int h, int use_accurate_subpel_search) {
2974   COMMON_MV_TEST;
2975 
2976   (void)minr;
2977   (void)minc;
2978 
2979   bestmv->row = maxr;
2980   bestmv->col = maxc;
2981   besterr = 0;
2982 
2983   // In the sub-pel motion search, if hp is not used, then the last bit of mv
2984   // has to be 0.
2985   lower_mv_precision(bestmv, allow_hp && use_mv_hp(ref_mv));
2986 
2987   return besterr;
2988 }
2989 // 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)2990 uint32_t vp9_return_min_sub_pixel_mv(
2991     const MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
2992     int error_per_bit, const vp9_variance_fn_ptr_t *vfp, int forced_stop,
2993     int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
2994     uint32_t *distortion, uint32_t *sse1, const uint8_t *second_pred, int w,
2995     int h, int use_accurate_subpel_search) {
2996   COMMON_MV_TEST;
2997 
2998   (void)maxr;
2999   (void)maxc;
3000 
3001   bestmv->row = minr;
3002   bestmv->col = minc;
3003   besterr = 0;
3004 
3005   // In the sub-pel motion search, if hp is not used, then the last bit of mv
3006   // has to be 0.
3007   lower_mv_precision(bestmv, allow_hp && use_mv_hp(ref_mv));
3008 
3009   return besterr;
3010 }
3011