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 // Border for Loop restoration buffer
26 #define AOM_RESTORATION_FRAME_BORDER 32
27 #define CLIP(x, lo, hi) ((x) < (lo) ? (lo) : (x) > (hi) ? (hi) : (x))
28 #define RINT(x) ((x) < 0 ? (int)((x)-0.5) : (int)((x) + 0.5))
29
30 #define RESTORATION_PROC_UNIT_SIZE 64
31
32 // Filter tile grid offset upwards compared to the superblock grid
33 #define RESTORATION_UNIT_OFFSET 8
34
35 #define SGRPROJ_BORDER_VERT 3 // Vertical border used for Sgr
36 #define SGRPROJ_BORDER_HORZ 3 // Horizontal border used for Sgr
37
38 #define WIENER_BORDER_VERT 2 // Vertical border used for Wiener
39 #define WIENER_HALFWIN 3
40 #define WIENER_BORDER_HORZ (WIENER_HALFWIN) // Horizontal border for Wiener
41
42 // RESTORATION_BORDER_VERT determines line buffer requirement for LR.
43 // Should be set at the max of SGRPROJ_BORDER_VERT and WIENER_BORDER_VERT.
44 // Note the line buffer needed is twice the value of this macro.
45 #if SGRPROJ_BORDER_VERT >= WIENER_BORDER_VERT
46 #define RESTORATION_BORDER_VERT (SGRPROJ_BORDER_VERT)
47 #else
48 #define RESTORATION_BORDER_VERT (WIENER_BORDER_VERT)
49 #endif // SGRPROJ_BORDER_VERT >= WIENER_BORDER_VERT
50
51 #if SGRPROJ_BORDER_HORZ >= WIENER_BORDER_HORZ
52 #define RESTORATION_BORDER_HORZ (SGRPROJ_BORDER_HORZ)
53 #else
54 #define RESTORATION_BORDER_HORZ (WIENER_BORDER_HORZ)
55 #endif // SGRPROJ_BORDER_VERT >= WIENER_BORDER_VERT
56
57 // How many border pixels do we need for each processing unit?
58 #define RESTORATION_BORDER 3
59
60 // How many rows of deblocked pixels do we save above/below each processing
61 // stripe?
62 #define RESTORATION_CTX_VERT 2
63
64 // Additional pixels to the left and right in above/below buffers
65 // It is RESTORATION_BORDER_HORZ rounded up to get nicer buffer alignment
66 #define RESTORATION_EXTRA_HORZ 4
67
68 // Pad up to 20 more (may be much less is needed)
69 #define RESTORATION_PADDING 20
70 #define RESTORATION_PROC_UNIT_PELS \
71 ((RESTORATION_PROC_UNIT_SIZE + RESTORATION_BORDER_HORZ * 2 + \
72 RESTORATION_PADDING) * \
73 (RESTORATION_PROC_UNIT_SIZE + RESTORATION_BORDER_VERT * 2 + \
74 RESTORATION_PADDING))
75
76 #define RESTORATION_UNITSIZE_MAX 256
77 #define RESTORATION_UNITPELS_HORZ_MAX \
78 (RESTORATION_UNITSIZE_MAX * 3 / 2 + 2 * RESTORATION_BORDER_HORZ + 16)
79 #define RESTORATION_UNITPELS_VERT_MAX \
80 ((RESTORATION_UNITSIZE_MAX * 3 / 2 + 2 * RESTORATION_BORDER_VERT + \
81 RESTORATION_UNIT_OFFSET))
82 #define RESTORATION_UNITPELS_MAX \
83 (RESTORATION_UNITPELS_HORZ_MAX * RESTORATION_UNITPELS_VERT_MAX)
84
85 // Two 32-bit buffers needed for the restored versions from two filters
86 // TODO(debargha, rupert): Refactor to not need the large tilesize to be stored
87 // on the decoder side.
88 #define SGRPROJ_TMPBUF_SIZE (RESTORATION_UNITPELS_MAX * 2 * sizeof(int32_t))
89
90 #define SGRPROJ_EXTBUF_SIZE (0)
91 #define SGRPROJ_PARAMS_BITS 4
92 #define SGRPROJ_PARAMS (1 << SGRPROJ_PARAMS_BITS)
93
94 // Precision bits for projection
95 #define SGRPROJ_PRJ_BITS 7
96 // Restoration precision bits generated higher than source before projection
97 #define SGRPROJ_RST_BITS 4
98 // Internal precision bits for core selfguided_restoration
99 #define SGRPROJ_SGR_BITS 8
100 #define SGRPROJ_SGR (1 << SGRPROJ_SGR_BITS)
101
102 #define SGRPROJ_PRJ_MIN0 (-(1 << SGRPROJ_PRJ_BITS) * 3 / 4)
103 #define SGRPROJ_PRJ_MAX0 (SGRPROJ_PRJ_MIN0 + (1 << SGRPROJ_PRJ_BITS) - 1)
104 #define SGRPROJ_PRJ_MIN1 (-(1 << SGRPROJ_PRJ_BITS) / 4)
105 #define SGRPROJ_PRJ_MAX1 (SGRPROJ_PRJ_MIN1 + (1 << SGRPROJ_PRJ_BITS) - 1)
106
107 #define SGRPROJ_PRJ_SUBEXP_K 4
108
109 #define SGRPROJ_BITS (SGRPROJ_PRJ_BITS * 2 + SGRPROJ_PARAMS_BITS)
110
111 #define MAX_RADIUS 2 // Only 1, 2, 3 allowed
112 #define MAX_NELEM ((2 * MAX_RADIUS + 1) * (2 * MAX_RADIUS + 1))
113 #define SGRPROJ_MTABLE_BITS 20
114 #define SGRPROJ_RECIP_BITS 12
115
116 #define WIENER_HALFWIN1 (WIENER_HALFWIN + 1)
117 #define WIENER_WIN (2 * WIENER_HALFWIN + 1)
118 #define WIENER_WIN2 ((WIENER_WIN) * (WIENER_WIN))
119 #define WIENER_TMPBUF_SIZE (0)
120 #define WIENER_EXTBUF_SIZE (0)
121
122 // If WIENER_WIN_CHROMA == WIENER_WIN - 2, that implies 5x5 filters are used for
123 // chroma. To use 7x7 for chroma set WIENER_WIN_CHROMA to WIENER_WIN.
124 #define WIENER_WIN_CHROMA (WIENER_WIN - 2)
125 #define WIENER_WIN2_CHROMA ((WIENER_WIN_CHROMA) * (WIENER_WIN_CHROMA))
126
127 #define WIENER_FILT_PREC_BITS 7
128 #define WIENER_FILT_STEP (1 << WIENER_FILT_PREC_BITS)
129
130 // Central values for the taps
131 #define WIENER_FILT_TAP0_MIDV (3)
132 #define WIENER_FILT_TAP1_MIDV (-7)
133 #define WIENER_FILT_TAP2_MIDV (15)
134 #define WIENER_FILT_TAP3_MIDV \
135 (WIENER_FILT_STEP - 2 * (WIENER_FILT_TAP0_MIDV + WIENER_FILT_TAP1_MIDV + \
136 WIENER_FILT_TAP2_MIDV))
137
138 #define WIENER_FILT_TAP0_BITS 4
139 #define WIENER_FILT_TAP1_BITS 5
140 #define WIENER_FILT_TAP2_BITS 6
141
142 #define WIENER_FILT_BITS \
143 ((WIENER_FILT_TAP0_BITS + WIENER_FILT_TAP1_BITS + WIENER_FILT_TAP2_BITS) * 2)
144
145 #define WIENER_FILT_TAP0_MINV \
146 (WIENER_FILT_TAP0_MIDV - (1 << WIENER_FILT_TAP0_BITS) / 2)
147 #define WIENER_FILT_TAP1_MINV \
148 (WIENER_FILT_TAP1_MIDV - (1 << WIENER_FILT_TAP1_BITS) / 2)
149 #define WIENER_FILT_TAP2_MINV \
150 (WIENER_FILT_TAP2_MIDV - (1 << WIENER_FILT_TAP2_BITS) / 2)
151
152 #define WIENER_FILT_TAP0_MAXV \
153 (WIENER_FILT_TAP0_MIDV - 1 + (1 << WIENER_FILT_TAP0_BITS) / 2)
154 #define WIENER_FILT_TAP1_MAXV \
155 (WIENER_FILT_TAP1_MIDV - 1 + (1 << WIENER_FILT_TAP1_BITS) / 2)
156 #define WIENER_FILT_TAP2_MAXV \
157 (WIENER_FILT_TAP2_MIDV - 1 + (1 << WIENER_FILT_TAP2_BITS) / 2)
158
159 #define WIENER_FILT_TAP0_SUBEXP_K 1
160 #define WIENER_FILT_TAP1_SUBEXP_K 2
161 #define WIENER_FILT_TAP2_SUBEXP_K 3
162
163 // Max of SGRPROJ_TMPBUF_SIZE, DOMAINTXFMRF_TMPBUF_SIZE, WIENER_TMPBUF_SIZE
164 #define RESTORATION_TMPBUF_SIZE (SGRPROJ_TMPBUF_SIZE)
165
166 // Max of SGRPROJ_EXTBUF_SIZE, WIENER_EXTBUF_SIZE
167 #define RESTORATION_EXTBUF_SIZE (WIENER_EXTBUF_SIZE)
168
169 // Check the assumptions of the existing code
170 #if SUBPEL_TAPS != WIENER_WIN + 1
171 #error "Wiener filter currently only works if SUBPEL_TAPS == WIENER_WIN + 1"
172 #endif
173 #if WIENER_FILT_PREC_BITS != 7
174 #error "Wiener filter currently only works if WIENER_FILT_PREC_BITS == 7"
175 #endif
176
177 #define LR_TILE_ROW 0
178 #define LR_TILE_COL 0
179 #define LR_TILE_COLS 1
180
181 typedef struct {
182 int r[2]; // radii
183 int s[2]; // sgr parameters for r[0] and r[1], based on GenSgrprojVtable()
184 } sgr_params_type;
185
186 typedef struct {
187 RestorationType restoration_type;
188 WienerInfo wiener_info;
189 SgrprojInfo sgrproj_info;
190 } RestorationUnitInfo;
191
192 // A restoration line buffer needs space for two lines plus a horizontal filter
193 // margin of RESTORATION_EXTRA_HORZ on each side.
194 #define RESTORATION_LINEBUFFER_WIDTH \
195 (RESTORATION_UNITSIZE_MAX * 3 / 2 + 2 * RESTORATION_EXTRA_HORZ)
196
197 // Similarly, the column buffers (used when we're at a vertical tile edge
198 // that we can't filter across) need space for one processing unit's worth
199 // of pixels, plus the top/bottom border width
200 #define RESTORATION_COLBUFFER_HEIGHT \
201 (RESTORATION_PROC_UNIT_SIZE + 2 * RESTORATION_BORDER)
202
203 typedef struct {
204 // Temporary buffers to save/restore 3 lines above/below the restoration
205 // stripe.
206 uint16_t tmp_save_above[RESTORATION_BORDER][RESTORATION_LINEBUFFER_WIDTH];
207 uint16_t tmp_save_below[RESTORATION_BORDER][RESTORATION_LINEBUFFER_WIDTH];
208 } RestorationLineBuffers;
209
210 typedef struct {
211 uint8_t *stripe_boundary_above;
212 uint8_t *stripe_boundary_below;
213 int stripe_boundary_stride;
214 int stripe_boundary_size;
215 } RestorationStripeBoundaries;
216
217 typedef struct {
218 RestorationType frame_restoration_type;
219 int restoration_unit_size;
220
221 // Fields below here are allocated and initialised by
222 // av1_alloc_restoration_struct. (horz_)units_per_tile give the number of
223 // restoration units in (one row of) the largest tile in the frame. The data
224 // in unit_info is laid out with units_per_tile entries for each tile, which
225 // have stride horz_units_per_tile.
226 //
227 // Even if there are tiles of different sizes, the data in unit_info is laid
228 // out as if all tiles are of full size.
229 int units_per_tile;
230 int vert_units_per_tile, horz_units_per_tile;
231 RestorationUnitInfo *unit_info;
232 RestorationStripeBoundaries boundaries;
233 int optimized_lr;
234 } RestorationInfo;
235
set_default_sgrproj(SgrprojInfo * sgrproj_info)236 static INLINE void set_default_sgrproj(SgrprojInfo *sgrproj_info) {
237 sgrproj_info->xqd[0] = (SGRPROJ_PRJ_MIN0 + SGRPROJ_PRJ_MAX0) / 2;
238 sgrproj_info->xqd[1] = (SGRPROJ_PRJ_MIN1 + SGRPROJ_PRJ_MAX1) / 2;
239 }
240
set_default_wiener(WienerInfo * wiener_info)241 static INLINE void set_default_wiener(WienerInfo *wiener_info) {
242 wiener_info->vfilter[0] = wiener_info->hfilter[0] = WIENER_FILT_TAP0_MIDV;
243 wiener_info->vfilter[1] = wiener_info->hfilter[1] = WIENER_FILT_TAP1_MIDV;
244 wiener_info->vfilter[2] = wiener_info->hfilter[2] = WIENER_FILT_TAP2_MIDV;
245 wiener_info->vfilter[WIENER_HALFWIN] = wiener_info->hfilter[WIENER_HALFWIN] =
246 -2 *
247 (WIENER_FILT_TAP2_MIDV + WIENER_FILT_TAP1_MIDV + WIENER_FILT_TAP0_MIDV);
248 wiener_info->vfilter[4] = wiener_info->hfilter[4] = WIENER_FILT_TAP2_MIDV;
249 wiener_info->vfilter[5] = wiener_info->hfilter[5] = WIENER_FILT_TAP1_MIDV;
250 wiener_info->vfilter[6] = wiener_info->hfilter[6] = WIENER_FILT_TAP0_MIDV;
251 }
252
253 typedef struct {
254 int h_start, h_end, v_start, v_end;
255 } RestorationTileLimits;
256
257 typedef void (*rest_unit_visitor_t)(const RestorationTileLimits *limits,
258 const AV1PixelRect *tile_rect,
259 int rest_unit_idx, void *priv,
260 int32_t *tmpbuf,
261 RestorationLineBuffers *rlbs);
262
263 typedef struct FilterFrameCtxt {
264 const RestorationInfo *rsi;
265 int tile_stripe0;
266 int ss_x, ss_y;
267 int highbd, bit_depth;
268 uint8_t *data8, *dst8;
269 int data_stride, dst_stride;
270 AV1PixelRect tile_rect;
271 } FilterFrameCtxt;
272
273 typedef struct AV1LrStruct {
274 rest_unit_visitor_t on_rest_unit;
275 FilterFrameCtxt ctxt[MAX_MB_PLANE];
276 YV12_BUFFER_CONFIG *frame;
277 YV12_BUFFER_CONFIG *dst;
278 } AV1LrStruct;
279
280 extern const sgr_params_type sgr_params[SGRPROJ_PARAMS];
281 extern int sgrproj_mtable[SGRPROJ_PARAMS][2];
282 extern const int32_t x_by_xplus1[256];
283 extern const int32_t one_by_x[MAX_NELEM];
284
285 void av1_alloc_restoration_struct(struct AV1Common *cm, RestorationInfo *rsi,
286 int is_uv);
287 void av1_free_restoration_struct(RestorationInfo *rst_info);
288
289 void extend_frame(uint8_t *data, int width, int height, int stride,
290 int border_horz, int border_vert, int highbd);
291 void decode_xq(const int *xqd, int *xq, const sgr_params_type *params);
292
293 // Filter a single loop restoration unit.
294 //
295 // limits is the limits of the unit. rui gives the mode to use for this unit
296 // and its coefficients. If striped loop restoration is enabled, rsb contains
297 // deblocked pixels to use for stripe boundaries; rlbs is just some space to
298 // use as a scratch buffer. tile_rect gives the limits of the tile containing
299 // this unit. tile_stripe0 is the index of the first stripe in this tile.
300 //
301 // ss_x and ss_y are flags which should be 1 if this is a plane with
302 // horizontal/vertical subsampling, respectively. highbd is a flag which should
303 // be 1 in high bit depth mode, in which case bit_depth is the bit depth.
304 //
305 // data8 is the frame data (pointing at the top-left corner of the frame, not
306 // the restoration unit) and stride is its stride. dst8 is the buffer where the
307 // results will be written and has stride dst_stride. Like data8, dst8 should
308 // point at the top-left corner of the frame.
309 //
310 // Finally tmpbuf is a scratch buffer used by the sgrproj filter which should
311 // be at least SGRPROJ_TMPBUF_SIZE big.
312 void av1_loop_restoration_filter_unit(
313 const RestorationTileLimits *limits, const RestorationUnitInfo *rui,
314 const RestorationStripeBoundaries *rsb, RestorationLineBuffers *rlbs,
315 const AV1PixelRect *tile_rect, int tile_stripe0, int ss_x, int ss_y,
316 int highbd, int bit_depth, uint8_t *data8, int stride, uint8_t *dst8,
317 int dst_stride, int32_t *tmpbuf, int optimized_lr);
318
319 void av1_loop_restoration_filter_frame(YV12_BUFFER_CONFIG *frame,
320 struct AV1Common *cm, int optimized_lr,
321 void *lr_ctxt);
322 void av1_loop_restoration_precal();
323
324 typedef void (*rest_tile_start_visitor_t)(int tile_row, int tile_col,
325 void *priv);
326 struct AV1LrSyncData;
327
328 typedef void (*sync_read_fn_t)(void *const lr_sync, int r, int c, int plane);
329
330 typedef void (*sync_write_fn_t)(void *const lr_sync, int r, int c,
331 const int sb_cols, int plane);
332
333 // Call on_rest_unit for each loop restoration unit in the plane.
334 void av1_foreach_rest_unit_in_plane(const struct AV1Common *cm, int plane,
335 rest_unit_visitor_t on_rest_unit,
336 void *priv, AV1PixelRect *tile_rect,
337 int32_t *tmpbuf,
338 RestorationLineBuffers *rlbs);
339
340 // Return 1 iff the block at mi_row, mi_col with size bsize is a
341 // top-level superblock containing the top-left corner of at least one
342 // loop restoration unit.
343 //
344 // If the block is a top-level superblock, the function writes to
345 // *rcol0, *rcol1, *rrow0, *rrow1. The rectangle of restoration unit
346 // indices given by [*rcol0, *rcol1) x [*rrow0, *rrow1) are relative
347 // to the current tile, whose starting index is returned as
348 // *tile_tl_idx.
349 int av1_loop_restoration_corners_in_sb(const struct AV1Common *cm, int plane,
350 int mi_row, int mi_col, BLOCK_SIZE bsize,
351 int *rcol0, int *rcol1, int *rrow0,
352 int *rrow1);
353
354 void av1_loop_restoration_save_boundary_lines(const YV12_BUFFER_CONFIG *frame,
355 struct AV1Common *cm,
356 int after_cdef);
357 void av1_loop_restoration_filter_frame_init(AV1LrStruct *lr_ctxt,
358 YV12_BUFFER_CONFIG *frame,
359 struct AV1Common *cm,
360 int optimized_lr, int num_planes);
361 void av1_loop_restoration_copy_planes(AV1LrStruct *loop_rest_ctxt,
362 struct AV1Common *cm, int num_planes);
363 void av1_foreach_rest_unit_in_row(
364 RestorationTileLimits *limits, const AV1PixelRect *tile_rect,
365 rest_unit_visitor_t on_rest_unit, int row_number, int unit_size,
366 int unit_idx0, int hunits_per_tile, int vunits_per_tile, int plane,
367 void *priv, int32_t *tmpbuf, RestorationLineBuffers *rlbs,
368 sync_read_fn_t on_sync_read, sync_write_fn_t on_sync_write,
369 struct AV1LrSyncData *const lr_sync);
370 AV1PixelRect av1_whole_frame_rect(const struct AV1Common *cm, int is_uv);
371 int av1_lr_count_units_in_tile(int unit_size, int tile_size);
372 void av1_lr_sync_read_dummy(void *const lr_sync, int r, int c, int plane);
373 void av1_lr_sync_write_dummy(void *const lr_sync, int r, int c,
374 const int sb_cols, int plane);
375 #ifdef __cplusplus
376 } // extern "C"
377 #endif
378
379 #endif // AOM_AV1_COMMON_RESTORATION_H_
380