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_ENCODER_BLOCK_H_
13 #define AOM_AV1_ENCODER_BLOCK_H_
14
15 #include "av1/common/entropymv.h"
16 #include "av1/common/entropy.h"
17 #include "av1/common/mvref_common.h"
18 #include "av1/encoder/hash.h"
19 #if CONFIG_DIST_8X8
20 #include "aom/aomcx.h"
21 #endif
22
23 #ifdef __cplusplus
24 extern "C" {
25 #endif
26
27 typedef struct {
28 unsigned int sse;
29 int sum;
30 unsigned int var;
31 } DIFF;
32
33 typedef struct macroblock_plane {
34 DECLARE_ALIGNED(16, int16_t, src_diff[MAX_SB_SQUARE]);
35 tran_low_t *qcoeff;
36 tran_low_t *coeff;
37 uint16_t *eobs;
38 uint8_t *txb_entropy_ctx;
39 struct buf_2d src;
40
41 // Quantizer setings
42 // These are used/accessed only in the quantization process
43 // RDO does not / must not depend on any of these values
44 // All values below share the coefficient scale/shift used in TX
45 const int16_t *quant_fp_QTX;
46 const int16_t *round_fp_QTX;
47 const int16_t *quant_QTX;
48 const int16_t *quant_shift_QTX;
49 const int16_t *zbin_QTX;
50 const int16_t *round_QTX;
51 const int16_t *dequant_QTX;
52 } MACROBLOCK_PLANE;
53
54 typedef struct {
55 int txb_skip_cost[TXB_SKIP_CONTEXTS][2];
56 int base_eob_cost[SIG_COEF_CONTEXTS_EOB][3];
57 int base_cost[SIG_COEF_CONTEXTS][4];
58 int eob_extra_cost[EOB_COEF_CONTEXTS][2];
59 int dc_sign_cost[DC_SIGN_CONTEXTS][2];
60 int lps_cost[LEVEL_CONTEXTS][COEFF_BASE_RANGE + 1];
61 } LV_MAP_COEFF_COST;
62
63 typedef struct {
64 int eob_cost[2][11];
65 } LV_MAP_EOB_COST;
66
67 typedef struct {
68 tran_low_t tcoeff[MAX_MB_PLANE][MAX_SB_SQUARE];
69 uint16_t eobs[MAX_MB_PLANE][MAX_SB_SQUARE / (TX_SIZE_W_MIN * TX_SIZE_H_MIN)];
70 uint8_t txb_skip_ctx[MAX_MB_PLANE]
71 [MAX_SB_SQUARE / (TX_SIZE_W_MIN * TX_SIZE_H_MIN)];
72 int dc_sign_ctx[MAX_MB_PLANE]
73 [MAX_SB_SQUARE / (TX_SIZE_W_MIN * TX_SIZE_H_MIN)];
74 } CB_COEFF_BUFFER;
75
76 typedef struct {
77 int16_t mode_context[MODE_CTX_REF_FRAMES];
78 // TODO(angiebird): Reduce the buffer size according to sb_type
79 tran_low_t *tcoeff[MAX_MB_PLANE];
80 uint16_t *eobs[MAX_MB_PLANE];
81 uint8_t *txb_skip_ctx[MAX_MB_PLANE];
82 int *dc_sign_ctx[MAX_MB_PLANE];
83 uint8_t ref_mv_count[MODE_CTX_REF_FRAMES];
84 CANDIDATE_MV ref_mv_stack[MODE_CTX_REF_FRAMES][MAX_REF_MV_STACK_SIZE];
85 int_mv global_mvs[REF_FRAMES];
86 int16_t compound_mode_context[MODE_CTX_REF_FRAMES];
87 } MB_MODE_INFO_EXT;
88
89 typedef struct {
90 int col_min;
91 int col_max;
92 int row_min;
93 int row_max;
94 } MvLimits;
95
96 typedef struct {
97 uint8_t best_palette_color_map[MAX_PALETTE_SQUARE];
98 int kmeans_data_buf[2 * MAX_PALETTE_SQUARE];
99 } PALETTE_BUFFER;
100
101 typedef struct {
102 TX_SIZE tx_size;
103 TX_SIZE inter_tx_size[INTER_TX_SIZE_BUF_LEN];
104 uint8_t blk_skip[MAX_MIB_SIZE * MAX_MIB_SIZE];
105 TX_TYPE txk_type[TXK_TYPE_BUF_LEN];
106 RD_STATS rd_stats;
107 uint32_t hash_value;
108 } MB_RD_INFO;
109
110 #define RD_RECORD_BUFFER_LEN 8
111 typedef struct {
112 MB_RD_INFO tx_rd_info[RD_RECORD_BUFFER_LEN]; // Circular buffer.
113 int index_start;
114 int num;
115 CRC32C crc_calculator; // Hash function.
116 } MB_RD_RECORD;
117
118 typedef struct {
119 int64_t dist;
120 int64_t sse;
121 int rate;
122 uint16_t eob;
123 TX_TYPE tx_type;
124 uint16_t entropy_context;
125 uint8_t txb_entropy_ctx;
126 uint8_t valid;
127 uint8_t fast; // This is not being used now.
128 } TXB_RD_INFO;
129
130 #define TX_SIZE_RD_RECORD_BUFFER_LEN 256
131 typedef struct {
132 uint32_t hash_vals[TX_SIZE_RD_RECORD_BUFFER_LEN];
133 TXB_RD_INFO tx_rd_info[TX_SIZE_RD_RECORD_BUFFER_LEN];
134 int index_start;
135 int num;
136 } TXB_RD_RECORD;
137
138 typedef struct tx_size_rd_info_node {
139 TXB_RD_INFO *rd_info_array; // Points to array of size TX_TYPES.
140 struct tx_size_rd_info_node *children[4];
141 } TXB_RD_INFO_NODE;
142
143 // Region size for mode decision sampling in the first pass of partition
144 // search(two_pass_partition_search speed feature), in units of mi size(4).
145 // Used by the mode_pruning_based_on_two_pass_partition_search speed feature.
146 #define FIRST_PARTITION_PASS_SAMPLE_REGION 8
147 #define FIRST_PARTITION_PASS_SAMPLE_REGION_LOG2 3
148 #define FIRST_PARTITION_PASS_STATS_TABLES \
149 (MAX_MIB_SIZE >> FIRST_PARTITION_PASS_SAMPLE_REGION_LOG2) * \
150 (MAX_MIB_SIZE >> FIRST_PARTITION_PASS_SAMPLE_REGION_LOG2)
151 #define FIRST_PARTITION_PASS_STATS_STRIDE \
152 (MAX_MIB_SIZE_LOG2 - FIRST_PARTITION_PASS_SAMPLE_REGION_LOG2)
153
av1_first_partition_pass_stats_index(int mi_row,int mi_col)154 static INLINE int av1_first_partition_pass_stats_index(int mi_row, int mi_col) {
155 const int row =
156 (mi_row & MAX_MIB_MASK) >> FIRST_PARTITION_PASS_SAMPLE_REGION_LOG2;
157 const int col =
158 (mi_col & MAX_MIB_MASK) >> FIRST_PARTITION_PASS_SAMPLE_REGION_LOG2;
159 return (row << FIRST_PARTITION_PASS_STATS_STRIDE) + col;
160 }
161
162 typedef struct {
163 uint8_t ref0_counts[REF_FRAMES]; // Counters for ref_frame[0].
164 uint8_t ref1_counts[REF_FRAMES]; // Counters for ref_frame[1].
165 int sample_counts; // Number of samples collected.
166 } FIRST_PARTITION_PASS_STATS;
167
168 #define MAX_INTERP_FILTER_STATS 64
169 typedef struct {
170 InterpFilters filters;
171 int_mv mv[2];
172 int8_t ref_frames[2];
173 COMPOUND_TYPE comp_type;
174 } INTERPOLATION_FILTER_STATS;
175
176 typedef struct macroblock MACROBLOCK;
177 struct macroblock {
178 struct macroblock_plane plane[MAX_MB_PLANE];
179
180 // Determine if one would go with reduced complexity transform block
181 // search model to select prediction modes, or full complexity model
182 // to select transform kernel.
183 int rd_model;
184
185 // Indicate if the encoder is running in the first pass partition search.
186 // In that case, apply certain speed features therein to reduce the overhead
187 // cost in the first pass search.
188 int cb_partition_scan;
189
190 FIRST_PARTITION_PASS_STATS
191 first_partition_pass_stats[FIRST_PARTITION_PASS_STATS_TABLES];
192
193 // [comp_idx][saved stat_idx]
194 INTERPOLATION_FILTER_STATS interp_filter_stats[2][MAX_INTERP_FILTER_STATS];
195 int interp_filter_stats_idx[2];
196
197 // Activate constrained coding block partition search range.
198 int use_cb_search_range;
199
200 // Inter macroblock RD search info.
201 MB_RD_RECORD mb_rd_record;
202
203 // Inter transform block RD search info. for square TX sizes.
204 TXB_RD_RECORD txb_rd_record_8X8[(MAX_MIB_SIZE >> 1) * (MAX_MIB_SIZE >> 1)];
205 TXB_RD_RECORD txb_rd_record_16X16[(MAX_MIB_SIZE >> 2) * (MAX_MIB_SIZE >> 2)];
206 TXB_RD_RECORD txb_rd_record_32X32[(MAX_MIB_SIZE >> 3) * (MAX_MIB_SIZE >> 3)];
207 TXB_RD_RECORD txb_rd_record_64X64[(MAX_MIB_SIZE >> 4) * (MAX_MIB_SIZE >> 4)];
208
209 // Intra transform block RD search info. for square TX sizes.
210 TXB_RD_RECORD txb_rd_record_intra;
211
212 MACROBLOCKD e_mbd;
213 MB_MODE_INFO_EXT *mbmi_ext;
214 int skip_block;
215 int qindex;
216
217 // The equivalent error at the current rdmult of one whole bit (not one
218 // bitcost unit).
219 int errorperbit;
220 // The equivalend SAD error of one (whole) bit at the current quantizer
221 // for large blocks.
222 int sadperbit16;
223 // The equivalend SAD error of one (whole) bit at the current quantizer
224 // for sub-8x8 blocks.
225 int sadperbit4;
226 int rdmult;
227 int mb_energy;
228 int sb_energy_level;
229 int *m_search_count_ptr;
230 int *ex_search_count_ptr;
231
232 unsigned int txb_split_count;
233
234 // These are set to their default values at the beginning, and then adjusted
235 // further in the encoding process.
236 BLOCK_SIZE min_partition_size;
237 BLOCK_SIZE max_partition_size;
238
239 unsigned int max_mv_context[REF_FRAMES];
240 unsigned int source_variance;
241 unsigned int pred_sse[REF_FRAMES];
242 int pred_mv_sad[REF_FRAMES];
243
244 int *nmvjointcost;
245 int nmv_vec_cost[MV_JOINTS];
246 int *nmvcost[2];
247 int *nmvcost_hp[2];
248 int **mv_cost_stack;
249 int **mvcost;
250
251 int32_t *wsrc_buf;
252 int32_t *mask_buf;
253 uint8_t *above_pred_buf;
254 uint8_t *left_pred_buf;
255
256 PALETTE_BUFFER *palette_buffer;
257
258 CONV_BUF_TYPE *tmp_conv_dst;
259 uint8_t *tmp_obmc_bufs[2];
260
261 // buffer for hash value calculation of a block
262 // used only in av1_get_block_hash_value()
263 // [first hash/second hash]
264 // [two buffers used ping-pong]
265 uint32_t *hash_value_buffer[2][2];
266
267 CRC_CALCULATOR crc_calculator1;
268 CRC_CALCULATOR crc_calculator2;
269 int g_crc_initialized;
270
271 // These define limits to motion vector components to prevent them
272 // from extending outside the UMV borders
273 MvLimits mv_limits;
274
275 uint8_t blk_skip[MAX_MIB_SIZE * MAX_MIB_SIZE];
276
277 int skip;
278 int skip_chroma_rd;
279 int skip_cost[SKIP_CONTEXTS][2];
280
281 int skip_mode; // 0: off; 1: on
282 int skip_mode_cost[SKIP_CONTEXTS][2];
283
284 int compound_idx;
285
286 LV_MAP_COEFF_COST coeff_costs[TX_SIZES][PLANE_TYPES];
287 LV_MAP_EOB_COST eob_costs[7][2];
288 uint16_t cb_offset;
289
290 // mode costs
291 int intra_inter_cost[INTRA_INTER_CONTEXTS][2];
292
293 int mbmode_cost[BLOCK_SIZE_GROUPS][INTRA_MODES];
294 int newmv_mode_cost[NEWMV_MODE_CONTEXTS][2];
295 int zeromv_mode_cost[GLOBALMV_MODE_CONTEXTS][2];
296 int refmv_mode_cost[REFMV_MODE_CONTEXTS][2];
297 int drl_mode_cost0[DRL_MODE_CONTEXTS][2];
298
299 int comp_inter_cost[COMP_INTER_CONTEXTS][2];
300 int single_ref_cost[REF_CONTEXTS][SINGLE_REFS - 1][2];
301 int comp_ref_type_cost[COMP_REF_TYPE_CONTEXTS]
302 [CDF_SIZE(COMP_REFERENCE_TYPES)];
303 int uni_comp_ref_cost[UNI_COMP_REF_CONTEXTS][UNIDIR_COMP_REFS - 1]
304 [CDF_SIZE(2)];
305 // Cost for signaling ref_frame[0] (LAST_FRAME, LAST2_FRAME, LAST3_FRAME or
306 // GOLDEN_FRAME) in bidir-comp mode.
307 int comp_ref_cost[REF_CONTEXTS][FWD_REFS - 1][2];
308 // Cost for signaling ref_frame[1] (ALTREF_FRAME, ALTREF2_FRAME, or
309 // BWDREF_FRAME) in bidir-comp mode.
310 int comp_bwdref_cost[REF_CONTEXTS][BWD_REFS - 1][2];
311 int inter_compound_mode_cost[INTER_MODE_CONTEXTS][INTER_COMPOUND_MODES];
312 int compound_type_cost[BLOCK_SIZES_ALL][COMPOUND_TYPES - 1];
313 int wedge_idx_cost[BLOCK_SIZES_ALL][16];
314 int interintra_cost[BLOCK_SIZE_GROUPS][2];
315 int wedge_interintra_cost[BLOCK_SIZES_ALL][2];
316 int interintra_mode_cost[BLOCK_SIZE_GROUPS][INTERINTRA_MODES];
317 int motion_mode_cost[BLOCK_SIZES_ALL][MOTION_MODES];
318 int motion_mode_cost1[BLOCK_SIZES_ALL][2];
319 int intra_uv_mode_cost[CFL_ALLOWED_TYPES][INTRA_MODES][UV_INTRA_MODES];
320 int y_mode_costs[INTRA_MODES][INTRA_MODES][INTRA_MODES];
321 int filter_intra_cost[BLOCK_SIZES_ALL][2];
322 int filter_intra_mode_cost[FILTER_INTRA_MODES];
323 int switchable_interp_costs[SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS];
324 int partition_cost[PARTITION_CONTEXTS][EXT_PARTITION_TYPES];
325 int palette_y_size_cost[PALATTE_BSIZE_CTXS][PALETTE_SIZES];
326 int palette_uv_size_cost[PALATTE_BSIZE_CTXS][PALETTE_SIZES];
327 int palette_y_color_cost[PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS]
328 [PALETTE_COLORS];
329 int palette_uv_color_cost[PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS]
330 [PALETTE_COLORS];
331 int palette_y_mode_cost[PALATTE_BSIZE_CTXS][PALETTE_Y_MODE_CONTEXTS][2];
332 int palette_uv_mode_cost[PALETTE_UV_MODE_CONTEXTS][2];
333 // The rate associated with each alpha codeword
334 int cfl_cost[CFL_JOINT_SIGNS][CFL_PRED_PLANES][CFL_ALPHABET_SIZE];
335 int tx_size_cost[TX_SIZES - 1][TX_SIZE_CONTEXTS][TX_SIZES];
336 int txfm_partition_cost[TXFM_PARTITION_CONTEXTS][2];
337 int inter_tx_type_costs[EXT_TX_SETS_INTER][EXT_TX_SIZES][TX_TYPES];
338 int intra_tx_type_costs[EXT_TX_SETS_INTRA][EXT_TX_SIZES][INTRA_MODES]
339 [TX_TYPES];
340 int angle_delta_cost[DIRECTIONAL_MODES][2 * MAX_ANGLE_DELTA + 1];
341 int switchable_restore_cost[RESTORE_SWITCHABLE_TYPES];
342 int wiener_restore_cost[2];
343 int sgrproj_restore_cost[2];
344 int intrabc_cost[2];
345
346 // Used to store sub partition's choices.
347 MV pred_mv[REF_FRAMES];
348
349 // Store the best motion vector during motion search
350 int_mv best_mv;
351 // Store the second best motion vector during full-pixel motion search
352 int_mv second_best_mv;
353
354 // use default transform and skip transform type search for intra modes
355 int use_default_intra_tx_type;
356 // use default transform and skip transform type search for inter modes
357 int use_default_inter_tx_type;
358 #if CONFIG_DIST_8X8
359 int using_dist_8x8;
360 aom_tune_metric tune_metric;
361 #endif // CONFIG_DIST_8X8
362 int comp_idx_cost[COMP_INDEX_CONTEXTS][2];
363 int comp_group_idx_cost[COMP_GROUP_IDX_CONTEXTS][2];
364 // Bit flags for pruning tx type search, tx split, etc.
365 int tx_search_prune[EXT_TX_SET_TYPES];
366 int must_find_valid_partition;
367 int tx_split_prune_flag; // Flag to skip tx split RD search.
368 int recalc_luma_mc_data; // Flag to indicate recalculation of MC data during
369 // interpolation filter search
370 };
371
is_rect_tx_allowed_bsize(BLOCK_SIZE bsize)372 static INLINE int is_rect_tx_allowed_bsize(BLOCK_SIZE bsize) {
373 static const char LUT[BLOCK_SIZES_ALL] = {
374 0, // BLOCK_4X4
375 1, // BLOCK_4X8
376 1, // BLOCK_8X4
377 0, // BLOCK_8X8
378 1, // BLOCK_8X16
379 1, // BLOCK_16X8
380 0, // BLOCK_16X16
381 1, // BLOCK_16X32
382 1, // BLOCK_32X16
383 0, // BLOCK_32X32
384 1, // BLOCK_32X64
385 1, // BLOCK_64X32
386 0, // BLOCK_64X64
387 0, // BLOCK_64X128
388 0, // BLOCK_128X64
389 0, // BLOCK_128X128
390 1, // BLOCK_4X16
391 1, // BLOCK_16X4
392 1, // BLOCK_8X32
393 1, // BLOCK_32X8
394 1, // BLOCK_16X64
395 1, // BLOCK_64X16
396 };
397
398 return LUT[bsize];
399 }
400
is_rect_tx_allowed(const MACROBLOCKD * xd,const MB_MODE_INFO * mbmi)401 static INLINE int is_rect_tx_allowed(const MACROBLOCKD *xd,
402 const MB_MODE_INFO *mbmi) {
403 return is_rect_tx_allowed_bsize(mbmi->sb_type) &&
404 !xd->lossless[mbmi->segment_id];
405 }
406
tx_size_to_depth(TX_SIZE tx_size,BLOCK_SIZE bsize)407 static INLINE int tx_size_to_depth(TX_SIZE tx_size, BLOCK_SIZE bsize) {
408 TX_SIZE ctx_size = max_txsize_rect_lookup[bsize];
409 int depth = 0;
410 while (tx_size != ctx_size) {
411 depth++;
412 ctx_size = sub_tx_size_map[ctx_size];
413 assert(depth <= MAX_TX_DEPTH);
414 }
415 return depth;
416 }
417
set_blk_skip(MACROBLOCK * x,int plane,int blk_idx,int skip)418 static INLINE void set_blk_skip(MACROBLOCK *x, int plane, int blk_idx,
419 int skip) {
420 if (skip)
421 x->blk_skip[blk_idx] |= 1UL << plane;
422 else
423 x->blk_skip[blk_idx] &= ~(1UL << plane);
424 #ifndef NDEBUG
425 // Set chroma planes to uninitialized states when luma is set to check if
426 // it will be set later
427 if (plane == 0) {
428 x->blk_skip[blk_idx] |= 1UL << (1 + 4);
429 x->blk_skip[blk_idx] |= 1UL << (2 + 4);
430 }
431
432 // Clear the initialization checking bit
433 x->blk_skip[blk_idx] &= ~(1UL << (plane + 4));
434 #endif
435 }
436
is_blk_skip(MACROBLOCK * x,int plane,int blk_idx)437 static INLINE int is_blk_skip(MACROBLOCK *x, int plane, int blk_idx) {
438 #ifndef NDEBUG
439 // Check if this is initialized
440 assert(!(x->blk_skip[blk_idx] & (1UL << (plane + 4))));
441
442 // The magic number is 0x77, this is to test if there is garbage data
443 assert((x->blk_skip[blk_idx] & 0x88) == 0);
444 #endif
445 return (x->blk_skip[blk_idx] >> plane) & 1;
446 }
447
448 #ifdef __cplusplus
449 } // extern "C"
450 #endif
451
452 #endif // AOM_AV1_ENCODER_BLOCK_H_
453