vp9_speed_features.h 16 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483
  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. #ifndef VP9_ENCODER_VP9_SPEED_FEATURES_H_
  11. #define VP9_ENCODER_VP9_SPEED_FEATURES_H_
  12. #include "vp9/common/vp9_enums.h"
  13. #ifdef __cplusplus
  14. extern "C" {
  15. #endif
  16. enum {
  17. INTRA_ALL = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED) | (1 << D45_PRED) |
  18. (1 << D135_PRED) | (1 << D117_PRED) | (1 << D153_PRED) |
  19. (1 << D207_PRED) | (1 << D63_PRED) | (1 << TM_PRED),
  20. INTRA_DC = (1 << DC_PRED),
  21. INTRA_DC_TM = (1 << DC_PRED) | (1 << TM_PRED),
  22. INTRA_DC_H_V = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED),
  23. INTRA_DC_TM_H_V =
  24. (1 << DC_PRED) | (1 << TM_PRED) | (1 << V_PRED) | (1 << H_PRED)
  25. };
  26. enum {
  27. INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) | (1 << NEWMV),
  28. INTER_NEAREST = (1 << NEARESTMV),
  29. INTER_NEAREST_NEW = (1 << NEARESTMV) | (1 << NEWMV),
  30. INTER_NEAREST_ZERO = (1 << NEARESTMV) | (1 << ZEROMV),
  31. INTER_NEAREST_NEW_ZERO = (1 << NEARESTMV) | (1 << ZEROMV) | (1 << NEWMV),
  32. INTER_NEAREST_NEAR_NEW = (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV),
  33. INTER_NEAREST_NEAR_ZERO = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV),
  34. };
  35. enum {
  36. DISABLE_ALL_INTER_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
  37. (1 << THR_ALTR) | (1 << THR_GOLD) | (1 << THR_LAST),
  38. DISABLE_ALL_SPLIT = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
  39. DISABLE_COMPOUND_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
  40. LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
  41. (1 << THR_ALTR) | (1 << THR_GOLD)
  42. };
  43. typedef enum {
  44. DIAMOND = 0,
  45. NSTEP = 1,
  46. HEX = 2,
  47. BIGDIA = 3,
  48. SQUARE = 4,
  49. FAST_HEX = 5,
  50. FAST_DIAMOND = 6
  51. } SEARCH_METHODS;
  52. typedef enum {
  53. // No recode.
  54. DISALLOW_RECODE = 0,
  55. // Allow recode for KF and exceeding maximum frame bandwidth.
  56. ALLOW_RECODE_KFMAXBW = 1,
  57. // Allow recode only for KF/ARF/GF frames.
  58. ALLOW_RECODE_KFARFGF = 2,
  59. // Allow recode for ARF/GF/KF and first normal frame in each group.
  60. ALLOW_RECODE_FIRST = 3,
  61. // Allow recode for all frames based on bitrate constraints.
  62. ALLOW_RECODE = 4,
  63. } RECODE_LOOP_TYPE;
  64. typedef enum {
  65. SUBPEL_TREE = 0,
  66. SUBPEL_TREE_PRUNED = 1, // Prunes 1/2-pel searches
  67. SUBPEL_TREE_PRUNED_MORE = 2, // Prunes 1/2-pel searches more aggressively
  68. SUBPEL_TREE_PRUNED_EVENMORE = 3, // Prunes 1/2- and 1/4-pel searches
  69. // Other methods to come
  70. } SUBPEL_SEARCH_METHODS;
  71. typedef enum {
  72. NO_MOTION_THRESHOLD = 0,
  73. LOW_MOTION_THRESHOLD = 7
  74. } MOTION_THRESHOLD;
  75. typedef enum {
  76. USE_FULL_RD = 0,
  77. USE_LARGESTALL,
  78. USE_TX_8X8
  79. } TX_SIZE_SEARCH_METHOD;
  80. typedef enum {
  81. NOT_IN_USE = 0,
  82. RELAXED_NEIGHBORING_MIN_MAX = 1,
  83. STRICT_NEIGHBORING_MIN_MAX = 2
  84. } AUTO_MIN_MAX_MODE;
  85. typedef enum {
  86. // Try the full image with different values.
  87. LPF_PICK_FROM_FULL_IMAGE,
  88. // Try a small portion of the image with different values.
  89. LPF_PICK_FROM_SUBIMAGE,
  90. // Estimate the level based on quantizer and frame type
  91. LPF_PICK_FROM_Q,
  92. // Pick 0 to disable LPF if LPF was enabled last frame
  93. LPF_PICK_MINIMAL_LPF
  94. } LPF_PICK_METHOD;
  95. typedef enum {
  96. // Terminate search early based on distortion so far compared to
  97. // qp step, distortion in the neighborhood of the frame, etc.
  98. FLAG_EARLY_TERMINATE = 1 << 0,
  99. // Skips comp inter modes if the best so far is an intra mode.
  100. FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
  101. // Skips oblique intra modes if the best so far is an inter mode.
  102. FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
  103. // Skips oblique intra modes at angles 27, 63, 117, 153 if the best
  104. // intra so far is not one of the neighboring directions.
  105. FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
  106. // Skips intra modes other than DC_PRED if the source variance is small
  107. FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
  108. } MODE_SEARCH_SKIP_LOGIC;
  109. typedef enum {
  110. FLAG_SKIP_EIGHTTAP = 1 << EIGHTTAP,
  111. FLAG_SKIP_EIGHTTAP_SMOOTH = 1 << EIGHTTAP_SMOOTH,
  112. FLAG_SKIP_EIGHTTAP_SHARP = 1 << EIGHTTAP_SHARP,
  113. } INTERP_FILTER_MASK;
  114. typedef enum {
  115. // Search partitions using RD/NONRD criterion
  116. SEARCH_PARTITION,
  117. // Always use a fixed size partition
  118. FIXED_PARTITION,
  119. REFERENCE_PARTITION,
  120. // Use an arbitrary partitioning scheme based on source variance within
  121. // a 64X64 SB
  122. VAR_BASED_PARTITION,
  123. // Use non-fixed partitions based on source variance
  124. SOURCE_VAR_BASED_PARTITION
  125. } PARTITION_SEARCH_TYPE;
  126. typedef enum {
  127. // Does a dry run to see if any of the contexts need to be updated or not,
  128. // before the final run.
  129. TWO_LOOP = 0,
  130. // No dry run, also only half the coef contexts and bands are updated.
  131. // The rest are not updated at all.
  132. ONE_LOOP_REDUCED = 1
  133. } FAST_COEFF_UPDATE;
  134. typedef struct MV_SPEED_FEATURES {
  135. // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
  136. SEARCH_METHODS search_method;
  137. // This parameter controls which step in the n-step process we start at.
  138. // It's changed adaptively based on circumstances.
  139. int reduce_first_step_size;
  140. // If this is set to 1, we limit the motion search range to 2 times the
  141. // largest motion vector found in the last frame.
  142. int auto_mv_step_size;
  143. // Subpel_search_method can only be subpel_tree which does a subpixel
  144. // logarithmic search that keeps stepping at 1/2 pixel units until
  145. // you stop getting a gain, and then goes on to 1/4 and repeats
  146. // the same process. Along the way it skips many diagonals.
  147. SUBPEL_SEARCH_METHODS subpel_search_method;
  148. // Maximum number of steps in logarithmic subpel search before giving up.
  149. int subpel_iters_per_step;
  150. // Control when to stop subpel search:
  151. // 0: Full subpel search.
  152. // 1: Stop at quarter pixel.
  153. // 2: Stop at half pixel.
  154. // 3: Stop at full pixel.
  155. int subpel_force_stop;
  156. // This variable sets the step_param used in full pel motion search.
  157. int fullpel_search_step_param;
  158. } MV_SPEED_FEATURES;
  159. #define MAX_MESH_STEP 4
  160. typedef struct MESH_PATTERN {
  161. int range;
  162. int interval;
  163. } MESH_PATTERN;
  164. typedef struct SPEED_FEATURES {
  165. MV_SPEED_FEATURES mv;
  166. // Frame level coding parameter update
  167. int frame_parameter_update;
  168. RECODE_LOOP_TYPE recode_loop;
  169. // Trellis (dynamic programming) optimization of quantized values (+1, 0).
  170. int optimize_coefficients;
  171. // Always set to 0. If on it enables 0 cost background transmission
  172. // (except for the initial transmission of the segmentation). The feature is
  173. // disabled because the addition of very large block sizes make the
  174. // backgrounds very to cheap to encode, and the segmentation we have
  175. // adds overhead.
  176. int static_segmentation;
  177. // If 1 we iterate finding a best reference for 2 ref frames together - via
  178. // a log search that iterates 4 times (check around mv for last for best
  179. // error of combined predictor then check around mv for alt). If 0 we
  180. // we just use the best motion vector found for each frame by itself.
  181. BLOCK_SIZE comp_inter_joint_search_thresh;
  182. // This variable is used to cap the maximum number of times we skip testing a
  183. // mode to be evaluated. A high value means we will be faster.
  184. int adaptive_rd_thresh;
  185. // Enables skipping the reconstruction step (idct, recon) in the
  186. // intermediate steps assuming the last frame didn't have too many intra
  187. // blocks and the q is less than a threshold.
  188. int skip_encode_sb;
  189. int skip_encode_frame;
  190. // Speed feature to allow or disallow skipping of recode at block
  191. // level within a frame.
  192. int allow_skip_recode;
  193. // Coefficient probability model approximation step size
  194. int coeff_prob_appx_step;
  195. // Enable uniform quantizer followed by trellis coefficient optimization
  196. int allow_quant_coeff_opt;
  197. double quant_opt_thresh;
  198. // Use transform domain distortion. Use pixel domain distortion in speed 0
  199. // and certain situations in higher speed to improve the RD model precision.
  200. int allow_txfm_domain_distortion;
  201. double tx_domain_thresh;
  202. // The threshold is to determine how slow the motino is, it is used when
  203. // use_lastframe_partitioning is set to LAST_FRAME_PARTITION_LOW_MOTION
  204. MOTION_THRESHOLD lf_motion_threshold;
  205. // Determine which method we use to determine transform size. We can choose
  206. // between options like full rd, largest for prediction size, largest
  207. // for intra and model coefs for the rest.
  208. TX_SIZE_SEARCH_METHOD tx_size_search_method;
  209. // Low precision 32x32 fdct keeps everything in 16 bits and thus is less
  210. // precise but significantly faster than the non lp version.
  211. int use_lp32x32fdct;
  212. // After looking at the first set of modes (set by index here), skip
  213. // checking modes for reference frames that don't match the reference frame
  214. // of the best so far.
  215. int mode_skip_start;
  216. // TODO(JBB): Remove this.
  217. int reference_masking;
  218. PARTITION_SEARCH_TYPE partition_search_type;
  219. // Used if partition_search_type = FIXED_SIZE_PARTITION
  220. BLOCK_SIZE always_this_block_size;
  221. // Skip rectangular partition test when partition type none gives better
  222. // rd than partition type split.
  223. int less_rectangular_check;
  224. // Disable testing non square partitions. (eg 16x32)
  225. int use_square_partition_only;
  226. BLOCK_SIZE use_square_only_threshold;
  227. // Sets min and max partition sizes for this 64x64 region based on the
  228. // same 64x64 in last encoded frame, and the left and above neighbor.
  229. AUTO_MIN_MAX_MODE auto_min_max_partition_size;
  230. // Ensures the rd based auto partition search will always
  231. // go down at least to the specified level.
  232. BLOCK_SIZE rd_auto_partition_min_limit;
  233. // Min and max partition size we enable (block_size) as per auto
  234. // min max, but also used by adjust partitioning, and pick_partitioning.
  235. BLOCK_SIZE default_min_partition_size;
  236. BLOCK_SIZE default_max_partition_size;
  237. // Whether or not we allow partitions one smaller or one greater than the last
  238. // frame's partitioning. Only used if use_lastframe_partitioning is set.
  239. int adjust_partitioning_from_last_frame;
  240. // How frequently we re do the partitioning from scratch. Only used if
  241. // use_lastframe_partitioning is set.
  242. int last_partitioning_redo_frequency;
  243. // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable
  244. // it always, to allow it for only Last frame and Intra, disable it for all
  245. // inter modes or to enable it always.
  246. int disable_split_mask;
  247. // TODO(jingning): combine the related motion search speed features
  248. // This allows us to use motion search at other sizes as a starting
  249. // point for this motion search and limits the search range around it.
  250. int adaptive_motion_search;
  251. // Flag for allowing some use of exhaustive searches;
  252. int allow_exhaustive_searches;
  253. // Threshold for allowing exhaistive motion search.
  254. int exhaustive_searches_thresh;
  255. // Maximum number of exhaustive searches for a frame.
  256. int max_exaustive_pct;
  257. // Pattern to be used for any exhaustive mesh searches.
  258. MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
  259. int schedule_mode_search;
  260. // Allows sub 8x8 modes to use the prediction filter that was determined
  261. // best for 8x8 mode. If set to 0 we always re check all the filters for
  262. // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter
  263. // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
  264. int adaptive_pred_interp_filter;
  265. // Adaptive prediction mode search
  266. int adaptive_mode_search;
  267. // Chessboard pattern prediction filter type search
  268. int cb_pred_filter_search;
  269. int cb_partition_search;
  270. int motion_field_mode_search;
  271. int alt_ref_search_fp;
  272. // Fast quantization process path
  273. int use_quant_fp;
  274. // Use finer quantizer in every other few frames that run variable block
  275. // partition type search.
  276. int force_frame_boost;
  277. // Maximally allowed base quantization index fluctuation.
  278. int max_delta_qindex;
  279. // Implements various heuristics to skip searching modes
  280. // The heuristics selected are based on flags
  281. // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
  282. unsigned int mode_search_skip_flags;
  283. // A source variance threshold below which filter search is disabled
  284. // Choose a very large value (UINT_MAX) to use 8-tap always
  285. unsigned int disable_filter_search_var_thresh;
  286. // These bit masks allow you to enable or disable intra modes for each
  287. // transform size separately.
  288. int intra_y_mode_mask[TX_SIZES];
  289. int intra_uv_mode_mask[TX_SIZES];
  290. // These bit masks allow you to enable or disable intra modes for each
  291. // prediction block size separately.
  292. int intra_y_mode_bsize_mask[BLOCK_SIZES];
  293. // This variable enables an early break out of mode testing if the model for
  294. // rd built from the prediction signal indicates a value that's much
  295. // higher than the best rd we've seen so far.
  296. int use_rd_breakout;
  297. // This enables us to use an estimate for intra rd based on dc mode rather
  298. // than choosing an actual uv mode in the stage of encoding before the actual
  299. // final encode.
  300. int use_uv_intra_rd_estimate;
  301. // This feature controls how the loop filter level is determined.
  302. LPF_PICK_METHOD lpf_pick;
  303. // This feature limits the number of coefficients updates we actually do
  304. // by only looking at counts from 1/2 the bands.
  305. FAST_COEFF_UPDATE use_fast_coef_updates;
  306. // This flag controls the use of non-RD mode decision.
  307. int use_nonrd_pick_mode;
  308. // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
  309. // modes are used in order from LSB to MSB for each BLOCK_SIZE.
  310. int inter_mode_mask[BLOCK_SIZES];
  311. // This feature controls whether we do the expensive context update and
  312. // calculation in the rd coefficient costing loop.
  313. int use_fast_coef_costing;
  314. // This feature controls the tolerence vs target used in deciding whether to
  315. // recode a frame. It has no meaning if recode is disabled.
  316. int recode_tolerance_low;
  317. int recode_tolerance_high;
  318. // This variable controls the maximum block size where intra blocks can be
  319. // used in inter frames.
  320. // TODO(aconverse): Fold this into one of the other many mode skips
  321. BLOCK_SIZE max_intra_bsize;
  322. // The frequency that we check if SOURCE_VAR_BASED_PARTITION or
  323. // FIXED_PARTITION search type should be used.
  324. int search_type_check_frequency;
  325. // When partition is pre-set, the inter prediction result from pick_inter_mode
  326. // can be reused in final block encoding process. It is enabled only for real-
  327. // time mode speed 6.
  328. int reuse_inter_pred_sby;
  329. // This variable sets the encode_breakout threshold. Currently, it is only
  330. // enabled in real time mode.
  331. int encode_breakout_thresh;
  332. // default interp filter choice
  333. INTERP_FILTER default_interp_filter;
  334. // Early termination in transform size search, which only applies while
  335. // tx_size_search_method is USE_FULL_RD.
  336. int tx_size_search_breakout;
  337. // adaptive interp_filter search to allow skip of certain filter types.
  338. int adaptive_interp_filter_search;
  339. // mask for skip evaluation of certain interp_filter type.
  340. INTERP_FILTER_MASK interp_filter_search_mask;
  341. // Partition search early breakout thresholds.
  342. int64_t partition_search_breakout_dist_thr;
  343. int partition_search_breakout_rate_thr;
  344. // Allow skipping partition search for still image frame
  345. int allow_partition_search_skip;
  346. // Fast approximation of vp9_model_rd_from_var_lapndz
  347. int simple_model_rd_from_var;
  348. // Skip a number of expensive mode evaluations for blocks with zero source
  349. // variance.
  350. int short_circuit_flat_blocks;
  351. // Skip a number of expensive mode evaluations for blocks with very low
  352. // temporal variance.
  353. // 1: Skip golden non-zeromv and ALL INTRA for bsize >= 32x32.
  354. // 2: Skip golden non-zeromv and newmv-last for bsize >= 16x16, skip ALL
  355. // INTRA for bsize >= 32x32 and vert/horz INTRA for bsize 16x16, 16x32 and
  356. // 32x16.
  357. int short_circuit_low_temp_var;
  358. // Limits the rd-threshold update for early exit for the newmv-last mode,
  359. // for non-rd mode.
  360. int limit_newmv_early_exit;
  361. // Adds a bias against golden reference, for non-rd mode.
  362. int bias_golden;
  363. // Bias to use base mv and skip 1/4 subpel search when use base mv in
  364. // enhancement layer.
  365. int base_mv_aggressive;
  366. } SPEED_FEATURES;
  367. struct VP9_COMP;
  368. void vp9_set_speed_features_framesize_independent(struct VP9_COMP *cpi);
  369. void vp9_set_speed_features_framesize_dependent(struct VP9_COMP *cpi);
  370. #ifdef __cplusplus
  371. } // extern "C"
  372. #endif
  373. #endif // VP9_ENCODER_VP9_SPEED_FEATURES_H_