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_RESTORATION_H_
13 #define AOM_AV1_COMMON_RESTORATION_H_
14
15 #include "aom_ports/mem.h"
16 #include "config/aom_config.h"
17
18 #include "av1/common/blockd.h"
19 #include "av1/common/enums.h"
20
21 #ifdef __cplusplus
22 extern "C" {
23 #endif
24
25 /*! @file */
26
27 /*!\cond */
28
29 // Border for Loop restoration buffer
30 #define AOM_RESTORATION_FRAME_BORDER 32
31 #define CLIP(x, lo, hi) ((x) < (lo) ? (lo) : (x) > (hi) ? (hi) : (x))
32 #define RINT(x) ((x) < 0 ? (int)((x)-0.5) : (int)((x) + 0.5))
33
34 #define RESTORATION_PROC_UNIT_SIZE 64
35
36 // Filter tile grid offset upwards compared to the superblock grid
37 #define RESTORATION_UNIT_OFFSET 8
38
39 #define SGRPROJ_BORDER_VERT 3 // Vertical border used for Sgr
40 #define SGRPROJ_BORDER_HORZ 3 // Horizontal border used for Sgr
41
42 #define WIENER_BORDER_VERT 2 // Vertical border used for Wiener
43 #define WIENER_HALFWIN 3
44 #define WIENER_BORDER_HORZ (WIENER_HALFWIN) // Horizontal border for Wiener
45
46 // RESTORATION_BORDER_VERT determines line buffer requirement for LR.
47 // Should be set at the max of SGRPROJ_BORDER_VERT and WIENER_BORDER_VERT.
48 // Note the line buffer needed is twice the value of this macro.
49 #if SGRPROJ_BORDER_VERT >= WIENER_BORDER_VERT
50 #define RESTORATION_BORDER_VERT (SGRPROJ_BORDER_VERT)
51 #else
52 #define RESTORATION_BORDER_VERT (WIENER_BORDER_VERT)
53 #endif // SGRPROJ_BORDER_VERT >= WIENER_BORDER_VERT
54
55 #if SGRPROJ_BORDER_HORZ >= WIENER_BORDER_HORZ
56 #define RESTORATION_BORDER_HORZ (SGRPROJ_BORDER_HORZ)
57 #else
58 #define RESTORATION_BORDER_HORZ (WIENER_BORDER_HORZ)
59 #endif // SGRPROJ_BORDER_VERT >= WIENER_BORDER_VERT
60
61 // How many border pixels do we need for each processing unit?
62 #define RESTORATION_BORDER 3
63
64 // How many rows of deblocked pixels do we save above/below each processing
65 // stripe?
66 #define RESTORATION_CTX_VERT 2
67
68 // Additional pixels to the left and right in above/below buffers
69 // It is RESTORATION_BORDER_HORZ rounded up to get nicer buffer alignment
70 #define RESTORATION_EXTRA_HORZ 4
71
72 // Pad up to 20 more (may be much less is needed)
73 #define RESTORATION_PADDING 20
74 #define RESTORATION_PROC_UNIT_PELS \
75 ((RESTORATION_PROC_UNIT_SIZE + RESTORATION_BORDER_HORZ * 2 + \
76 RESTORATION_PADDING) * \
77 (RESTORATION_PROC_UNIT_SIZE + RESTORATION_BORDER_VERT * 2 + \
78 RESTORATION_PADDING))
79
80 #define RESTORATION_UNITSIZE_MAX 256
81 #define RESTORATION_UNITPELS_HORZ_MAX \
82 (RESTORATION_UNITSIZE_MAX * 3 / 2 + 2 * RESTORATION_BORDER_HORZ + 16)
83 #define RESTORATION_UNITPELS_VERT_MAX \
84 ((RESTORATION_UNITSIZE_MAX * 3 / 2 + 2 * RESTORATION_BORDER_VERT + \
85 RESTORATION_UNIT_OFFSET))
86 #define RESTORATION_UNITPELS_MAX \
87 (RESTORATION_UNITPELS_HORZ_MAX * RESTORATION_UNITPELS_VERT_MAX)
88
89 // Two 32-bit buffers needed for the restored versions from two filters
90 // TODO(debargha, rupert): Refactor to not need the large tilesize to be stored
91 // on the decoder side.
92 #define SGRPROJ_TMPBUF_SIZE (RESTORATION_UNITPELS_MAX * 2 * sizeof(int32_t))
93
94 #define SGRPROJ_EXTBUF_SIZE (0)
95 #define SGRPROJ_PARAMS_BITS 4
96 #define SGRPROJ_PARAMS (1 << SGRPROJ_PARAMS_BITS)
97
98 // Precision bits for projection
99 #define SGRPROJ_PRJ_BITS 7
100 // Restoration precision bits generated higher than source before projection
101 #define SGRPROJ_RST_BITS 4
102 // Internal precision bits for core selfguided_restoration
103 #define SGRPROJ_SGR_BITS 8
104 #define SGRPROJ_SGR (1 << SGRPROJ_SGR_BITS)
105
106 #define SGRPROJ_PRJ_MIN0 (-(1 << SGRPROJ_PRJ_BITS) * 3 / 4)
107 #define SGRPROJ_PRJ_MAX0 (SGRPROJ_PRJ_MIN0 + (1 << SGRPROJ_PRJ_BITS) - 1)
108 #define SGRPROJ_PRJ_MIN1 (-(1 << SGRPROJ_PRJ_BITS) / 4)
109 #define SGRPROJ_PRJ_MAX1 (SGRPROJ_PRJ_MIN1 + (1 << SGRPROJ_PRJ_BITS) - 1)
110
111 #define SGRPROJ_PRJ_SUBEXP_K 4
112
113 #define SGRPROJ_BITS (SGRPROJ_PRJ_BITS * 2 + SGRPROJ_PARAMS_BITS)
114
115 #define MAX_RADIUS 2 // Only 1, 2, 3 allowed
116 #define MAX_NELEM ((2 * MAX_RADIUS + 1) * (2 * MAX_RADIUS + 1))
117 #define SGRPROJ_MTABLE_BITS 20
118 #define SGRPROJ_RECIP_BITS 12
119
120 #define WIENER_HALFWIN1 (WIENER_HALFWIN + 1)
121 #define WIENER_WIN (2 * WIENER_HALFWIN + 1)
122 #define WIENER_WIN2 ((WIENER_WIN) * (WIENER_WIN))
123 #define WIENER_TMPBUF_SIZE (0)
124 #define WIENER_EXTBUF_SIZE (0)
125
126 // If WIENER_WIN_CHROMA == WIENER_WIN - 2, that implies 5x5 filters are used for
127 // chroma. To use 7x7 for chroma set WIENER_WIN_CHROMA to WIENER_WIN.
128 #define WIENER_WIN_CHROMA (WIENER_WIN - 2)
129 #define WIENER_WIN_REDUCED (WIENER_WIN - 2)
130 #define WIENER_WIN2_CHROMA ((WIENER_WIN_CHROMA) * (WIENER_WIN_CHROMA))
131 #define WIENER_STATS_DOWNSAMPLE_FACTOR 4
132
133 #define WIENER_FILT_PREC_BITS 7
134 #define WIENER_FILT_STEP (1 << WIENER_FILT_PREC_BITS)
135
136 // Central values for the taps
137 #define WIENER_FILT_TAP0_MIDV (3)
138 #define WIENER_FILT_TAP1_MIDV (-7)
139 #define WIENER_FILT_TAP2_MIDV (15)
140 #define WIENER_FILT_TAP3_MIDV \
141 (WIENER_FILT_STEP - 2 * (WIENER_FILT_TAP0_MIDV + WIENER_FILT_TAP1_MIDV + \
142 WIENER_FILT_TAP2_MIDV))
143
144 #define WIENER_FILT_TAP0_BITS 4
145 #define WIENER_FILT_TAP1_BITS 5
146 #define WIENER_FILT_TAP2_BITS 6
147
148 #define WIENER_FILT_BITS \
149 ((WIENER_FILT_TAP0_BITS + WIENER_FILT_TAP1_BITS + WIENER_FILT_TAP2_BITS) * 2)
150
151 #define WIENER_FILT_TAP0_MINV \
152 (WIENER_FILT_TAP0_MIDV - (1 << WIENER_FILT_TAP0_BITS) / 2)
153 #define WIENER_FILT_TAP1_MINV \
154 (WIENER_FILT_TAP1_MIDV - (1 << WIENER_FILT_TAP1_BITS) / 2)
155 #define WIENER_FILT_TAP2_MINV \
156 (WIENER_FILT_TAP2_MIDV - (1 << WIENER_FILT_TAP2_BITS) / 2)
157
158 #define WIENER_FILT_TAP0_MAXV \
159 (WIENER_FILT_TAP0_MIDV - 1 + (1 << WIENER_FILT_TAP0_BITS) / 2)
160 #define WIENER_FILT_TAP1_MAXV \
161 (WIENER_FILT_TAP1_MIDV - 1 + (1 << WIENER_FILT_TAP1_BITS) / 2)
162 #define WIENER_FILT_TAP2_MAXV \
163 (WIENER_FILT_TAP2_MIDV - 1 + (1 << WIENER_FILT_TAP2_BITS) / 2)
164
165 #define WIENER_FILT_TAP0_SUBEXP_K 1
166 #define WIENER_FILT_TAP1_SUBEXP_K 2
167 #define WIENER_FILT_TAP2_SUBEXP_K 3
168
169 // Max of SGRPROJ_TMPBUF_SIZE, DOMAINTXFMRF_TMPBUF_SIZE, WIENER_TMPBUF_SIZE
170 #define RESTORATION_TMPBUF_SIZE (SGRPROJ_TMPBUF_SIZE)
171
172 // Max of SGRPROJ_EXTBUF_SIZE, WIENER_EXTBUF_SIZE
173 #define RESTORATION_EXTBUF_SIZE (WIENER_EXTBUF_SIZE)
174
175 // Check the assumptions of the existing code
176 #if SUBPEL_TAPS != WIENER_WIN + 1
177 #error "Wiener filter currently only works if SUBPEL_TAPS == WIENER_WIN + 1"
178 #endif
179 #if WIENER_FILT_PREC_BITS != 7
180 #error "Wiener filter currently only works if WIENER_FILT_PREC_BITS == 7"
181 #endif
182
183 #define LR_TILE_ROW 0
184 #define LR_TILE_COL 0
185 #define LR_TILE_COLS 1
186
187 typedef struct {
188 int r[2]; // radii
189 int s[2]; // sgr parameters for r[0] and r[1], based on GenSgrprojVtable()
190 } sgr_params_type;
191 /*!\endcond */
192
193 /*!\brief Parameters related to Restoration Unit Info */
194 typedef struct {
195 /*!
196 * restoration type
197 */
198 RestorationType restoration_type;
199
200 /*!
201 * Wiener filter parameters if restoration_type indicates Wiener
202 */
203 WienerInfo wiener_info;
204
205 /*!
206 * Sgrproj filter parameters if restoration_type indicates Sgrproj
207 */
208 SgrprojInfo sgrproj_info;
209 } RestorationUnitInfo;
210
211 /*!\cond */
212
213 // A restoration line buffer needs space for two lines plus a horizontal filter
214 // margin of RESTORATION_EXTRA_HORZ on each side.
215 #define RESTORATION_LINEBUFFER_WIDTH \
216 (RESTORATION_UNITSIZE_MAX * 3 / 2 + 2 * RESTORATION_EXTRA_HORZ)
217
218 // Similarly, the column buffers (used when we're at a vertical tile edge
219 // that we can't filter across) need space for one processing unit's worth
220 // of pixels, plus the top/bottom border width
221 #define RESTORATION_COLBUFFER_HEIGHT \
222 (RESTORATION_PROC_UNIT_SIZE + 2 * RESTORATION_BORDER)
223
224 typedef struct {
225 // Temporary buffers to save/restore 3 lines above/below the restoration
226 // stripe.
227 uint16_t tmp_save_above[RESTORATION_BORDER][RESTORATION_LINEBUFFER_WIDTH];
228 uint16_t tmp_save_below[RESTORATION_BORDER][RESTORATION_LINEBUFFER_WIDTH];
229 } RestorationLineBuffers;
230 /*!\endcond */
231
232 /*!\brief Parameters related to Restoration Stripe boundaries */
233 typedef struct {
234 /*!
235 * stripe boundary above
236 */
237 uint8_t *stripe_boundary_above;
238
239 /*!
240 * stripe boundary below
241 */
242 uint8_t *stripe_boundary_below;
243
244 /*!
245 * strides for stripe boundaries above and below
246 */
247 int stripe_boundary_stride;
248
249 /*!
250 * size of stripe boundaries above and below
251 */
252 int stripe_boundary_size;
253 } RestorationStripeBoundaries;
254
255 /*!\brief Parameters related to Restoration Info */
256 typedef struct {
257 /*!
258 * Restoration type for frame
259 */
260 RestorationType frame_restoration_type;
261
262 /*!
263 * Restoration unit size
264 */
265 int restoration_unit_size;
266
267 /**
268 * \name Fields allocated and initialised by av1_alloc_restoration_struct.
269 * (horz_)units_per_tile give the number of restoration units in
270 * (one row of) the largest tile in the frame.
271 */
272 /**@{*/
273 /*!
274 * Number of units per tile for the largest tile in the frame
275 */
276 int units_per_tile;
277
278 /*!
279 * Number of vertical units per tile
280 */
281 int vert_units_per_tile;
282
283 /*!
284 * Number of horizontal units per tile for the largest tile in the frame
285 */
286 int horz_units_per_tile;
287 /**@}*/
288
289 /*!
290 * List of info for units in tile.
291 * The data in unit_info is laid out with units_per_tile entries for each
292 * tile, which have stride horz_units_per_tile.
293 * Even if there are tiles of different sizes, the data in unit_info is
294 * laid out as if all tiles are of full size.
295 */
296 RestorationUnitInfo *unit_info;
297
298 /*!
299 * Restoration Stripe boundary info
300 */
301 RestorationStripeBoundaries boundaries;
302
303 /*!
304 * Whether optimized lr can be used for speed.
305 * That includes cases of no cdef and no superres, or if fast trial runs
306 * are used on the encoder side.
307 */
308 int optimized_lr;
309 } RestorationInfo;
310
311 /*!\cond */
312
set_default_sgrproj(SgrprojInfo * sgrproj_info)313 static INLINE void set_default_sgrproj(SgrprojInfo *sgrproj_info) {
314 sgrproj_info->xqd[0] = (SGRPROJ_PRJ_MIN0 + SGRPROJ_PRJ_MAX0) / 2;
315 sgrproj_info->xqd[1] = (SGRPROJ_PRJ_MIN1 + SGRPROJ_PRJ_MAX1) / 2;
316 }
317
set_default_wiener(WienerInfo * wiener_info)318 static INLINE void set_default_wiener(WienerInfo *wiener_info) {
319 wiener_info->vfilter[0] = wiener_info->hfilter[0] = WIENER_FILT_TAP0_MIDV;
320 wiener_info->vfilter[1] = wiener_info->hfilter[1] = WIENER_FILT_TAP1_MIDV;
321 wiener_info->vfilter[2] = wiener_info->hfilter[2] = WIENER_FILT_TAP2_MIDV;
322 wiener_info->vfilter[WIENER_HALFWIN] = wiener_info->hfilter[WIENER_HALFWIN] =
323 -2 *
324 (WIENER_FILT_TAP2_MIDV + WIENER_FILT_TAP1_MIDV + WIENER_FILT_TAP0_MIDV);
325 wiener_info->vfilter[4] = wiener_info->hfilter[4] = WIENER_FILT_TAP2_MIDV;
326 wiener_info->vfilter[5] = wiener_info->hfilter[5] = WIENER_FILT_TAP1_MIDV;
327 wiener_info->vfilter[6] = wiener_info->hfilter[6] = WIENER_FILT_TAP0_MIDV;
328 }
329
330 typedef struct {
331 int h_start, h_end, v_start, v_end;
332 } RestorationTileLimits;
333
334 typedef void (*rest_unit_visitor_t)(const RestorationTileLimits *limits,
335 const PixelRect *tile_rect,
336 int rest_unit_idx, void *priv,
337 int32_t *tmpbuf,
338 RestorationLineBuffers *rlbs);
339
340 typedef struct FilterFrameCtxt {
341 const RestorationInfo *rsi;
342 int tile_stripe0;
343 int ss_x, ss_y;
344 int highbd, bit_depth;
345 uint8_t *data8, *dst8;
346 int data_stride, dst_stride;
347 PixelRect tile_rect;
348 } FilterFrameCtxt;
349
350 typedef struct AV1LrStruct {
351 rest_unit_visitor_t on_rest_unit;
352 FilterFrameCtxt ctxt[MAX_MB_PLANE];
353 YV12_BUFFER_CONFIG *frame;
354 YV12_BUFFER_CONFIG *dst;
355 } AV1LrStruct;
356
357 extern const sgr_params_type av1_sgr_params[SGRPROJ_PARAMS];
358 extern int sgrproj_mtable[SGRPROJ_PARAMS][2];
359 extern const int32_t av1_x_by_xplus1[256];
360 extern const int32_t av1_one_by_x[MAX_NELEM];
361
362 void av1_alloc_restoration_struct(struct AV1Common *cm, RestorationInfo *rsi,
363 int is_uv);
364 void av1_free_restoration_struct(RestorationInfo *rst_info);
365
366 void av1_extend_frame(uint8_t *data, int width, int height, int stride,
367 int border_horz, int border_vert, int highbd);
368 void av1_decode_xq(const int *xqd, int *xq, const sgr_params_type *params);
369
370 /*!\endcond */
371
372 /*!\brief Function for applying loop restoration filter to a single unit.
373 *
374 * \ingroup in_loop_restoration
375 * This function applies the loop restoration filter to a single
376 * loop restoration unit.
377 *
378 * \param[in] limits Limits of the unit
379 * \param[in] rui The parameters to use for this unit and its
380 * coefficients
381 * \param[in] rsb Deblocked pixels to use for stripe boundaries
382 * \param[in] rlbs Space to use as a scratch buffer
383 * \param[in] tile_rect Limits of the tile containing this unit
384 * \param[in] tile_stripe0 Index of the first stripe in this tile
385 * \param[in] ss_x Horizontal subsampling for plane
386 * \param[in] ss_y Vertical subsampling for plane
387 * \param[in] highbd Whether high bitdepth pipeline is used
388 * \param[in] bit_depth Bit-depth of the video
389 * \param[in] data8 Frame data (pointing at the top-left corner of
390 * the frame, not the restoration unit).
391 * \param[in] stride Stride of \c data8
392 * \param[out] dst8 Buffer where the results will be written. Like
393 * \c data8, \c dst8 should point at the top-left
394 * corner of the frame
395 * \param[in] dst_stride Stride of \c dst8
396 * \param[in] tmpbuf Scratch buffer used by the sgrproj filter which
397 * should be at least SGRPROJ_TMPBUF_SIZE big.
398 * \param[in] optimized_lr Whether to use fast optimized Loop Restoration
399 *
400 * \remark Nothing is returned. Instead, the filtered unit is output in
401 * \c dst8 at the proper restoration unit offset.
402 */
403 void av1_loop_restoration_filter_unit(
404 const RestorationTileLimits *limits, const RestorationUnitInfo *rui,
405 const RestorationStripeBoundaries *rsb, RestorationLineBuffers *rlbs,
406 const PixelRect *tile_rect, int tile_stripe0, int ss_x, int ss_y,
407 int highbd, int bit_depth, uint8_t *data8, int stride, uint8_t *dst8,
408 int dst_stride, int32_t *tmpbuf, int optimized_lr);
409
410 /*!\brief Function for applying loop restoration filter to a frame
411 *
412 * \ingroup in_loop_restoration
413 * This function applies the loop restoration filter to a frame.
414 *
415 * \param[in, out] frame Compressed frame buffer
416 * \param[in, out] cm Pointer to top level common structure
417 * \param[in] optimized_lr Whether to use fast optimized Loop Restoration
418 * \param[in] lr_ctxt Loop restoration context
419 *
420 * \remark Nothing is returned. Instead, the filtered frame is output in
421 * \c frame.
422 */
423 void av1_loop_restoration_filter_frame(YV12_BUFFER_CONFIG *frame,
424 struct AV1Common *cm, int optimized_lr,
425 void *lr_ctxt);
426 /*!\cond */
427
428 void av1_loop_restoration_precal();
429
430 typedef void (*rest_tile_start_visitor_t)(int tile_row, int tile_col,
431 void *priv);
432 struct AV1LrSyncData;
433
434 typedef void (*sync_read_fn_t)(void *const lr_sync, int r, int c, int plane);
435
436 typedef void (*sync_write_fn_t)(void *const lr_sync, int r, int c,
437 const int sb_cols, int plane);
438
439 // Call on_rest_unit for each loop restoration unit in the plane.
440 void av1_foreach_rest_unit_in_plane(const struct AV1Common *cm, int plane,
441 rest_unit_visitor_t on_rest_unit,
442 void *priv, PixelRect *tile_rect,
443 int32_t *tmpbuf,
444 RestorationLineBuffers *rlbs);
445
446 // Return 1 iff the block at mi_row, mi_col with size bsize is a
447 // top-level superblock containing the top-left corner of at least one
448 // loop restoration unit.
449 //
450 // If the block is a top-level superblock, the function writes to
451 // *rcol0, *rcol1, *rrow0, *rrow1. The rectangle of restoration unit
452 // indices given by [*rcol0, *rcol1) x [*rrow0, *rrow1) are relative
453 // to the current tile, whose starting index is returned as
454 // *tile_tl_idx.
455 int av1_loop_restoration_corners_in_sb(const struct AV1Common *cm, int plane,
456 int mi_row, int mi_col, BLOCK_SIZE bsize,
457 int *rcol0, int *rcol1, int *rrow0,
458 int *rrow1);
459
460 void av1_loop_restoration_save_boundary_lines(const YV12_BUFFER_CONFIG *frame,
461 struct AV1Common *cm,
462 int after_cdef);
463 void av1_loop_restoration_filter_frame_init(AV1LrStruct *lr_ctxt,
464 YV12_BUFFER_CONFIG *frame,
465 struct AV1Common *cm,
466 int optimized_lr, int num_planes);
467 void av1_loop_restoration_copy_planes(AV1LrStruct *loop_rest_ctxt,
468 struct AV1Common *cm, int num_planes);
469 void av1_foreach_rest_unit_in_row(
470 RestorationTileLimits *limits, const PixelRect *tile_rect,
471 rest_unit_visitor_t on_rest_unit, int row_number, int unit_size,
472 int unit_idx0, int hunits_per_tile, int vunits_per_tile, int plane,
473 void *priv, int32_t *tmpbuf, RestorationLineBuffers *rlbs,
474 sync_read_fn_t on_sync_read, sync_write_fn_t on_sync_write,
475 struct AV1LrSyncData *const lr_sync);
476 PixelRect av1_whole_frame_rect(const struct AV1Common *cm, int is_uv);
477 int av1_lr_count_units_in_tile(int unit_size, int tile_size);
478 void av1_lr_sync_read_dummy(void *const lr_sync, int r, int c, int plane);
479 void av1_lr_sync_write_dummy(void *const lr_sync, int r, int c,
480 const int sb_cols, int plane);
481
482 /*!\endcond */
483
484 #ifdef __cplusplus
485 } // extern "C"
486 #endif
487
488 #endif // AOM_AV1_COMMON_RESTORATION_H_
489