1 /*
2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #ifndef AOM_AV1_COMMON_MV_H_
13 #define AOM_AV1_COMMON_MV_H_
14
15 #include <stdlib.h>
16
17 #include "av1/common/common.h"
18 #include "av1/common/common_data.h"
19 #include "aom_dsp/aom_filter.h"
20 #include "aom_dsp/flow_estimation/flow_estimation.h"
21
22 #ifdef __cplusplus
23 extern "C" {
24 #endif
25
26 #define INVALID_MV 0x80008000
27 #define INVALID_MV_ROW_COL -32768
28 #define GET_MV_RAWPEL(x) (((x) + 3 + ((x) >= 0)) >> 3)
29 #define GET_MV_SUBPEL(x) ((x)*8)
30
31 #define MARK_MV_INVALID(mv) \
32 do { \
33 ((int_mv *)(mv))->as_int = INVALID_MV; \
34 } while (0)
35 #define CHECK_MV_EQUAL(x, y) (((x).row == (y).row) && ((x).col == (y).col))
36
37 // The motion vector in units of full pixel
38 typedef struct fullpel_mv {
39 int16_t row;
40 int16_t col;
41 } FULLPEL_MV;
42
43 // The motion vector in units of 1/8-pel
44 typedef struct mv {
45 int16_t row;
46 int16_t col;
47 } MV;
48
49 static const MV kZeroMv = { 0, 0 };
50 static const FULLPEL_MV kZeroFullMv = { 0, 0 };
51
52 typedef union int_mv {
53 uint32_t as_int;
54 MV as_mv;
55 FULLPEL_MV as_fullmv;
56 } int_mv; /* facilitates faster equality tests and copies */
57
58 typedef struct mv32 {
59 int32_t row;
60 int32_t col;
61 } MV32;
62
63 // The mv limit for fullpel mvs
64 typedef struct {
65 int col_min;
66 int col_max;
67 int row_min;
68 int row_max;
69 } FullMvLimits;
70
71 // The mv limit for subpel mvs
72 typedef struct {
73 int col_min;
74 int col_max;
75 int row_min;
76 int row_max;
77 } SubpelMvLimits;
78
get_fullmv_from_mv(const MV * subpel_mv)79 static AOM_INLINE FULLPEL_MV get_fullmv_from_mv(const MV *subpel_mv) {
80 const FULLPEL_MV full_mv = { (int16_t)GET_MV_RAWPEL(subpel_mv->row),
81 (int16_t)GET_MV_RAWPEL(subpel_mv->col) };
82 return full_mv;
83 }
84
get_mv_from_fullmv(const FULLPEL_MV * full_mv)85 static AOM_INLINE MV get_mv_from_fullmv(const FULLPEL_MV *full_mv) {
86 const MV subpel_mv = { (int16_t)GET_MV_SUBPEL(full_mv->row),
87 (int16_t)GET_MV_SUBPEL(full_mv->col) };
88 return subpel_mv;
89 }
90
convert_fullmv_to_mv(int_mv * mv)91 static AOM_INLINE void convert_fullmv_to_mv(int_mv *mv) {
92 mv->as_mv = get_mv_from_fullmv(&mv->as_fullmv);
93 }
94
95 // Bits of precision used for the model
96 #define WARPEDMODEL_PREC_BITS 16
97 #define WARPEDMODEL_ROW3HOMO_PREC_BITS 16
98
99 #define WARPEDMODEL_TRANS_CLAMP (128 << WARPEDMODEL_PREC_BITS)
100 #define WARPEDMODEL_NONDIAGAFFINE_CLAMP (1 << (WARPEDMODEL_PREC_BITS - 3))
101 #define WARPEDMODEL_ROW3HOMO_CLAMP (1 << (WARPEDMODEL_PREC_BITS - 2))
102
103 // Bits of subpel precision for warped interpolation
104 #define WARPEDPIXEL_PREC_BITS 6
105 #define WARPEDPIXEL_PREC_SHIFTS (1 << WARPEDPIXEL_PREC_BITS)
106
107 #define WARP_PARAM_REDUCE_BITS 6
108
109 #define WARPEDDIFF_PREC_BITS (WARPEDMODEL_PREC_BITS - WARPEDPIXEL_PREC_BITS)
110
111 // Number of types used for global motion (must be >= 3 and <= TRANS_TYPES)
112 // The following can be useful:
113 // GLOBAL_TRANS_TYPES 3 - up to rotation-zoom
114 // GLOBAL_TRANS_TYPES 4 - up to affine
115 // GLOBAL_TRANS_TYPES 6 - up to hor/ver trapezoids
116 // GLOBAL_TRANS_TYPES 7 - up to full homography
117 #define GLOBAL_TRANS_TYPES 4
118
119 typedef struct {
120 int global_warp_allowed;
121 int local_warp_allowed;
122 } WarpTypesAllowed;
123
124 // The order of values in the wmmat matrix below is best described
125 // by the homography:
126 // [x' (m2 m3 m0 [x
127 // z . y' = m4 m5 m1 * y
128 // 1] m6 m7 1) 1]
129 typedef struct {
130 int32_t wmmat[6];
131 int16_t alpha, beta, gamma, delta;
132 TransformationType wmtype;
133 int8_t invalid;
134 } WarpedMotionParams;
135
136 /* clang-format off */
137 static const WarpedMotionParams default_warp_params = {
138 { 0, 0, (1 << WARPEDMODEL_PREC_BITS), 0, 0, (1 << WARPEDMODEL_PREC_BITS) },
139 0, 0, 0, 0,
140 IDENTITY,
141 0,
142 };
143 /* clang-format on */
144
145 // The following constants describe the various precisions
146 // of different parameters in the global motion experiment.
147 //
148 // Given the general homography:
149 // [x' (a b c [x
150 // z . y' = d e f * y
151 // 1] g h i) 1]
152 //
153 // Constants using the name ALPHA here are related to parameters
154 // a, b, d, e. Constants using the name TRANS are related
155 // to parameters c and f.
156 //
157 // Anything ending in PREC_BITS is the number of bits of precision
158 // to maintain when converting from double to integer.
159 //
160 // The ABS parameters are used to create an upper and lower bound
161 // for each parameter. In other words, after a parameter is integerized
162 // it is clamped between -(1 << ABS_XXX_BITS) and (1 << ABS_XXX_BITS).
163 //
164 // XXX_PREC_DIFF and XXX_DECODE_FACTOR
165 // are computed once here to prevent repetitive
166 // computation on the decoder side. These are
167 // to allow the global motion parameters to be encoded in a lower
168 // precision than the warped model precision. This means that they
169 // need to be changed to warped precision when they are decoded.
170 //
171 // XX_MIN, XX_MAX are also computed to avoid repeated computation
172
173 #define SUBEXPFIN_K 3
174 #define GM_TRANS_PREC_BITS 6
175 #define GM_ABS_TRANS_BITS 12
176 #define GM_ABS_TRANS_ONLY_BITS (GM_ABS_TRANS_BITS - GM_TRANS_PREC_BITS + 3)
177 #define GM_TRANS_PREC_DIFF (WARPEDMODEL_PREC_BITS - GM_TRANS_PREC_BITS)
178 #define GM_TRANS_ONLY_PREC_DIFF (WARPEDMODEL_PREC_BITS - 3)
179 #define GM_TRANS_DECODE_FACTOR (1 << GM_TRANS_PREC_DIFF)
180 #define GM_TRANS_ONLY_DECODE_FACTOR (1 << GM_TRANS_ONLY_PREC_DIFF)
181
182 #define GM_ALPHA_PREC_BITS 15
183 #define GM_ABS_ALPHA_BITS 12
184 #define GM_ALPHA_PREC_DIFF (WARPEDMODEL_PREC_BITS - GM_ALPHA_PREC_BITS)
185 #define GM_ALPHA_DECODE_FACTOR (1 << GM_ALPHA_PREC_DIFF)
186
187 #define GM_ROW3HOMO_PREC_BITS 16
188 #define GM_ABS_ROW3HOMO_BITS 11
189 #define GM_ROW3HOMO_PREC_DIFF \
190 (WARPEDMODEL_ROW3HOMO_PREC_BITS - GM_ROW3HOMO_PREC_BITS)
191 #define GM_ROW3HOMO_DECODE_FACTOR (1 << GM_ROW3HOMO_PREC_DIFF)
192
193 #define GM_TRANS_MAX (1 << GM_ABS_TRANS_BITS)
194 #define GM_ALPHA_MAX (1 << GM_ABS_ALPHA_BITS)
195 #define GM_ROW3HOMO_MAX (1 << GM_ABS_ROW3HOMO_BITS)
196
197 #define GM_TRANS_MIN -GM_TRANS_MAX
198 #define GM_ALPHA_MIN -GM_ALPHA_MAX
199 #define GM_ROW3HOMO_MIN -GM_ROW3HOMO_MAX
200
block_center_x(int mi_col,BLOCK_SIZE bs)201 static INLINE int block_center_x(int mi_col, BLOCK_SIZE bs) {
202 const int bw = block_size_wide[bs];
203 return mi_col * MI_SIZE + bw / 2 - 1;
204 }
205
block_center_y(int mi_row,BLOCK_SIZE bs)206 static INLINE int block_center_y(int mi_row, BLOCK_SIZE bs) {
207 const int bh = block_size_high[bs];
208 return mi_row * MI_SIZE + bh / 2 - 1;
209 }
210
convert_to_trans_prec(int allow_hp,int coor)211 static INLINE int convert_to_trans_prec(int allow_hp, int coor) {
212 if (allow_hp)
213 return ROUND_POWER_OF_TWO_SIGNED(coor, WARPEDMODEL_PREC_BITS - 3);
214 else
215 return ROUND_POWER_OF_TWO_SIGNED(coor, WARPEDMODEL_PREC_BITS - 2) * 2;
216 }
integer_mv_precision(MV * mv)217 static INLINE void integer_mv_precision(MV *mv) {
218 int mod = (mv->row % 8);
219 if (mod != 0) {
220 mv->row -= mod;
221 if (abs(mod) > 4) {
222 if (mod > 0) {
223 mv->row += 8;
224 } else {
225 mv->row -= 8;
226 }
227 }
228 }
229
230 mod = (mv->col % 8);
231 if (mod != 0) {
232 mv->col -= mod;
233 if (abs(mod) > 4) {
234 if (mod > 0) {
235 mv->col += 8;
236 } else {
237 mv->col -= 8;
238 }
239 }
240 }
241 }
242 // Convert a global motion vector into a motion vector at the centre of the
243 // given block.
244 //
245 // The resulting motion vector will have three fractional bits of precision. If
246 // allow_hp is zero, the bottom bit will always be zero. If CONFIG_AMVR and
247 // is_integer is true, the bottom three bits will be zero (so the motion vector
248 // represents an integer)
gm_get_motion_vector(const WarpedMotionParams * gm,int allow_hp,BLOCK_SIZE bsize,int mi_col,int mi_row,int is_integer)249 static INLINE int_mv gm_get_motion_vector(const WarpedMotionParams *gm,
250 int allow_hp, BLOCK_SIZE bsize,
251 int mi_col, int mi_row,
252 int is_integer) {
253 int_mv res;
254
255 if (gm->wmtype == IDENTITY) {
256 res.as_int = 0;
257 return res;
258 }
259
260 const int32_t *mat = gm->wmmat;
261 int x, y, tx, ty;
262
263 if (gm->wmtype == TRANSLATION) {
264 // All global motion vectors are stored with WARPEDMODEL_PREC_BITS (16)
265 // bits of fractional precision. The offset for a translation is stored in
266 // entries 0 and 1. For translations, all but the top three (two if
267 // cm->features.allow_high_precision_mv is false) fractional bits are always
268 // zero.
269 //
270 // After the right shifts, there are 3 fractional bits of precision. If
271 // allow_hp is false, the bottom bit is always zero (so we don't need a
272 // call to convert_to_trans_prec here)
273 //
274 // Note: There is an AV1 specification bug here:
275 //
276 // gm->wmmat[0] is supposed to be the horizontal translation, and so should
277 // go into res.as_mv.col, and gm->wmmat[1] is supposed to be the vertical
278 // translation and so should go into res.as_mv.row
279 //
280 // However, in the spec, these assignments are accidentally reversed, and so
281 // we must keep this incorrect logic to match the spec.
282 //
283 // See also: https://crbug.com/aomedia/3328
284 res.as_mv.row = gm->wmmat[0] >> GM_TRANS_ONLY_PREC_DIFF;
285 res.as_mv.col = gm->wmmat[1] >> GM_TRANS_ONLY_PREC_DIFF;
286 assert(IMPLIES(1 & (res.as_mv.row | res.as_mv.col), allow_hp));
287 if (is_integer) {
288 integer_mv_precision(&res.as_mv);
289 }
290 return res;
291 }
292
293 x = block_center_x(mi_col, bsize);
294 y = block_center_y(mi_row, bsize);
295
296 if (gm->wmtype == ROTZOOM) {
297 assert(gm->wmmat[5] == gm->wmmat[2]);
298 assert(gm->wmmat[4] == -gm->wmmat[3]);
299 }
300
301 const int xc =
302 (mat[2] - (1 << WARPEDMODEL_PREC_BITS)) * x + mat[3] * y + mat[0];
303 const int yc =
304 mat[4] * x + (mat[5] - (1 << WARPEDMODEL_PREC_BITS)) * y + mat[1];
305 tx = convert_to_trans_prec(allow_hp, xc);
306 ty = convert_to_trans_prec(allow_hp, yc);
307
308 res.as_mv.row = ty;
309 res.as_mv.col = tx;
310
311 if (is_integer) {
312 integer_mv_precision(&res.as_mv);
313 }
314 return res;
315 }
316
get_wmtype(const WarpedMotionParams * gm)317 static INLINE TransformationType get_wmtype(const WarpedMotionParams *gm) {
318 if (gm->wmmat[5] == (1 << WARPEDMODEL_PREC_BITS) && !gm->wmmat[4] &&
319 gm->wmmat[2] == (1 << WARPEDMODEL_PREC_BITS) && !gm->wmmat[3]) {
320 return ((!gm->wmmat[1] && !gm->wmmat[0]) ? IDENTITY : TRANSLATION);
321 }
322 if (gm->wmmat[2] == gm->wmmat[5] && gm->wmmat[3] == -gm->wmmat[4])
323 return ROTZOOM;
324 else
325 return AFFINE;
326 }
327
328 typedef struct candidate_mv {
329 int_mv this_mv;
330 int_mv comp_mv;
331 } CANDIDATE_MV;
332
is_zero_mv(const MV * mv)333 static INLINE int is_zero_mv(const MV *mv) {
334 return *((const uint32_t *)mv) == 0;
335 }
336
is_equal_mv(const MV * a,const MV * b)337 static INLINE int is_equal_mv(const MV *a, const MV *b) {
338 return *((const uint32_t *)a) == *((const uint32_t *)b);
339 }
340
clamp_mv(MV * mv,const SubpelMvLimits * mv_limits)341 static INLINE void clamp_mv(MV *mv, const SubpelMvLimits *mv_limits) {
342 mv->col = clamp(mv->col, mv_limits->col_min, mv_limits->col_max);
343 mv->row = clamp(mv->row, mv_limits->row_min, mv_limits->row_max);
344 }
345
clamp_fullmv(FULLPEL_MV * mv,const FullMvLimits * mv_limits)346 static INLINE void clamp_fullmv(FULLPEL_MV *mv, const FullMvLimits *mv_limits) {
347 mv->col = clamp(mv->col, mv_limits->col_min, mv_limits->col_max);
348 mv->row = clamp(mv->row, mv_limits->row_min, mv_limits->row_max);
349 }
350
351 #ifdef __cplusplus
352 } // extern "C"
353 #endif
354
355 #endif // AOM_AV1_COMMON_MV_H_
356