/src/aom/av1/encoder/pickrst.c
Line | Count | Source (jump to first uncovered line) |
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 | | #include <assert.h> |
13 | | #include <float.h> |
14 | | #include <limits.h> |
15 | | #include <math.h> |
16 | | |
17 | | #include "config/aom_scale_rtcd.h" |
18 | | #include "config/av1_rtcd.h" |
19 | | |
20 | | #include "aom_dsp/aom_dsp_common.h" |
21 | | #include "aom_dsp/binary_codes_writer.h" |
22 | | #include "aom_dsp/mathutils.h" |
23 | | #include "aom_dsp/psnr.h" |
24 | | #include "aom_mem/aom_mem.h" |
25 | | #include "aom_ports/mem.h" |
26 | | #include "av1/common/av1_common_int.h" |
27 | | #include "av1/common/quant_common.h" |
28 | | #include "av1/common/restoration.h" |
29 | | |
30 | | #include "av1/encoder/av1_quantize.h" |
31 | | #include "av1/encoder/encoder.h" |
32 | | #include "av1/encoder/picklpf.h" |
33 | | #include "av1/encoder/pickrst.h" |
34 | | |
35 | | // When set to RESTORE_WIENER or RESTORE_SGRPROJ only those are allowed. |
36 | | // When set to RESTORE_TYPES we allow switchable. |
37 | | static const RestorationType force_restore_type = RESTORE_TYPES; |
38 | | |
39 | | // Number of Wiener iterations |
40 | 0 | #define NUM_WIENER_ITERS 5 |
41 | | |
42 | | // Penalty factor for use of dual sgr |
43 | 0 | #define DUAL_SGR_PENALTY_MULT 0.01 |
44 | | |
45 | | // Working precision for Wiener filter coefficients |
46 | 0 | #define WIENER_TAP_SCALE_FACTOR ((int64_t)1 << 16) |
47 | | |
48 | 0 | #define SGRPROJ_EP_GRP1_START_IDX 0 |
49 | 0 | #define SGRPROJ_EP_GRP1_END_IDX 9 |
50 | 0 | #define SGRPROJ_EP_GRP1_SEARCH_COUNT 4 |
51 | 0 | #define SGRPROJ_EP_GRP2_3_SEARCH_COUNT 2 |
52 | | static const int sgproj_ep_grp1_seed[SGRPROJ_EP_GRP1_SEARCH_COUNT] = { 0, 3, 6, |
53 | | 9 }; |
54 | | static const int sgproj_ep_grp2_3[SGRPROJ_EP_GRP2_3_SEARCH_COUNT][14] = { |
55 | | { 10, 10, 11, 11, 12, 12, 13, 13, 13, 13, -1, -1, -1, -1 }, |
56 | | { 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15 } |
57 | | }; |
58 | | |
59 | | typedef int64_t (*sse_extractor_type)(const YV12_BUFFER_CONFIG *a, |
60 | | const YV12_BUFFER_CONFIG *b); |
61 | | typedef int64_t (*sse_part_extractor_type)(const YV12_BUFFER_CONFIG *a, |
62 | | const YV12_BUFFER_CONFIG *b, |
63 | | int hstart, int width, int vstart, |
64 | | int height); |
65 | | typedef uint64_t (*var_part_extractor_type)(const YV12_BUFFER_CONFIG *a, |
66 | | int hstart, int width, int vstart, |
67 | | int height); |
68 | | |
69 | | #if CONFIG_AV1_HIGHBITDEPTH |
70 | | #define NUM_EXTRACTORS (3 * (1 + 1)) |
71 | | #else |
72 | | #define NUM_EXTRACTORS 3 |
73 | | #endif |
74 | | static const sse_part_extractor_type sse_part_extractors[NUM_EXTRACTORS] = { |
75 | | aom_get_y_sse_part, aom_get_u_sse_part, |
76 | | aom_get_v_sse_part, |
77 | | #if CONFIG_AV1_HIGHBITDEPTH |
78 | | aom_highbd_get_y_sse_part, aom_highbd_get_u_sse_part, |
79 | | aom_highbd_get_v_sse_part, |
80 | | #endif |
81 | | }; |
82 | | static const var_part_extractor_type var_part_extractors[NUM_EXTRACTORS] = { |
83 | | aom_get_y_var, aom_get_u_var, aom_get_v_var, |
84 | | #if CONFIG_AV1_HIGHBITDEPTH |
85 | | aom_highbd_get_y_var, aom_highbd_get_u_var, aom_highbd_get_v_var, |
86 | | #endif |
87 | | }; |
88 | | |
89 | | static int64_t sse_restoration_unit(const RestorationTileLimits *limits, |
90 | | const YV12_BUFFER_CONFIG *src, |
91 | | const YV12_BUFFER_CONFIG *dst, int plane, |
92 | 0 | int highbd) { |
93 | 0 | return sse_part_extractors[3 * highbd + plane]( |
94 | 0 | src, dst, limits->h_start, limits->h_end - limits->h_start, |
95 | 0 | limits->v_start, limits->v_end - limits->v_start); |
96 | 0 | } |
97 | | |
98 | | static uint64_t var_restoration_unit(const RestorationTileLimits *limits, |
99 | | const YV12_BUFFER_CONFIG *src, int plane, |
100 | 0 | int highbd) { |
101 | 0 | return var_part_extractors[3 * highbd + plane]( |
102 | 0 | src, limits->h_start, limits->h_end - limits->h_start, limits->v_start, |
103 | 0 | limits->v_end - limits->v_start); |
104 | 0 | } |
105 | | |
106 | | typedef struct { |
107 | | // The best coefficients for Wiener or Sgrproj restoration |
108 | | WienerInfo wiener; |
109 | | SgrprojInfo sgrproj; |
110 | | |
111 | | // The sum of squared errors for this rtype. |
112 | | int64_t sse[RESTORE_SWITCHABLE_TYPES]; |
113 | | |
114 | | // The rtype to use for this unit given a frame rtype as |
115 | | // index. Indices: WIENER, SGRPROJ, SWITCHABLE. |
116 | | RestorationType best_rtype[RESTORE_TYPES - 1]; |
117 | | |
118 | | // This flag will be set based on the speed feature |
119 | | // 'prune_sgr_based_on_wiener'. 0 implies no pruning and 1 implies pruning. |
120 | | uint8_t skip_sgr_eval; |
121 | | } RestUnitSearchInfo; |
122 | | |
123 | | typedef struct { |
124 | | const YV12_BUFFER_CONFIG *src; |
125 | | YV12_BUFFER_CONFIG *dst; |
126 | | |
127 | | const AV1_COMMON *cm; |
128 | | const MACROBLOCK *x; |
129 | | int plane; |
130 | | int plane_width; |
131 | | int plane_height; |
132 | | RestUnitSearchInfo *rusi; |
133 | | |
134 | | // Speed features |
135 | | const LOOP_FILTER_SPEED_FEATURES *lpf_sf; |
136 | | |
137 | | uint8_t *dgd_buffer; |
138 | | int dgd_stride; |
139 | | const uint8_t *src_buffer; |
140 | | int src_stride; |
141 | | |
142 | | // sse and bits are initialised by reset_rsc in search_rest_type |
143 | | int64_t sse; |
144 | | int64_t bits; |
145 | | int tile_y0, tile_stripe0; |
146 | | |
147 | | // sgrproj and wiener are initialised by rsc_on_tile when starting the first |
148 | | // tile in the frame. |
149 | | SgrprojInfo sgrproj; |
150 | | WienerInfo wiener; |
151 | | AV1PixelRect tile_rect; |
152 | | } RestSearchCtxt; |
153 | | |
154 | 0 | static AOM_INLINE void rsc_on_tile(void *priv) { |
155 | 0 | RestSearchCtxt *rsc = (RestSearchCtxt *)priv; |
156 | 0 | set_default_sgrproj(&rsc->sgrproj); |
157 | 0 | set_default_wiener(&rsc->wiener); |
158 | 0 | rsc->tile_stripe0 = 0; |
159 | 0 | } |
160 | | |
161 | 0 | static AOM_INLINE void reset_rsc(RestSearchCtxt *rsc) { |
162 | 0 | rsc->sse = 0; |
163 | 0 | rsc->bits = 0; |
164 | 0 | } |
165 | | |
166 | | static AOM_INLINE void init_rsc(const YV12_BUFFER_CONFIG *src, |
167 | | const AV1_COMMON *cm, const MACROBLOCK *x, |
168 | | const LOOP_FILTER_SPEED_FEATURES *lpf_sf, |
169 | | int plane, RestUnitSearchInfo *rusi, |
170 | 0 | YV12_BUFFER_CONFIG *dst, RestSearchCtxt *rsc) { |
171 | 0 | rsc->src = src; |
172 | 0 | rsc->dst = dst; |
173 | 0 | rsc->cm = cm; |
174 | 0 | rsc->x = x; |
175 | 0 | rsc->plane = plane; |
176 | 0 | rsc->rusi = rusi; |
177 | 0 | rsc->lpf_sf = lpf_sf; |
178 | |
|
179 | 0 | const YV12_BUFFER_CONFIG *dgd = &cm->cur_frame->buf; |
180 | 0 | const int is_uv = plane != AOM_PLANE_Y; |
181 | 0 | rsc->plane_width = src->crop_widths[is_uv]; |
182 | 0 | rsc->plane_height = src->crop_heights[is_uv]; |
183 | 0 | rsc->src_buffer = src->buffers[plane]; |
184 | 0 | rsc->src_stride = src->strides[is_uv]; |
185 | 0 | rsc->dgd_buffer = dgd->buffers[plane]; |
186 | 0 | rsc->dgd_stride = dgd->strides[is_uv]; |
187 | 0 | rsc->tile_rect = av1_whole_frame_rect(cm, is_uv); |
188 | 0 | assert(src->crop_widths[is_uv] == dgd->crop_widths[is_uv]); |
189 | 0 | assert(src->crop_heights[is_uv] == dgd->crop_heights[is_uv]); |
190 | 0 | } |
191 | | |
192 | | static int64_t try_restoration_unit(const RestSearchCtxt *rsc, |
193 | | const RestorationTileLimits *limits, |
194 | | const AV1PixelRect *tile_rect, |
195 | 0 | const RestorationUnitInfo *rui) { |
196 | 0 | const AV1_COMMON *const cm = rsc->cm; |
197 | 0 | const int plane = rsc->plane; |
198 | 0 | const int is_uv = plane > 0; |
199 | 0 | const RestorationInfo *rsi = &cm->rst_info[plane]; |
200 | 0 | RestorationLineBuffers rlbs; |
201 | 0 | const int bit_depth = cm->seq_params->bit_depth; |
202 | 0 | const int highbd = cm->seq_params->use_highbitdepth; |
203 | |
|
204 | 0 | const YV12_BUFFER_CONFIG *fts = &cm->cur_frame->buf; |
205 | | // TODO(yunqing): For now, only use optimized LR filter in decoder. Can be |
206 | | // also used in encoder. |
207 | 0 | const int optimized_lr = 0; |
208 | |
|
209 | 0 | av1_loop_restoration_filter_unit( |
210 | 0 | limits, rui, &rsi->boundaries, &rlbs, tile_rect, rsc->tile_stripe0, |
211 | 0 | is_uv && cm->seq_params->subsampling_x, |
212 | 0 | is_uv && cm->seq_params->subsampling_y, highbd, bit_depth, |
213 | 0 | fts->buffers[plane], fts->strides[is_uv], rsc->dst->buffers[plane], |
214 | 0 | rsc->dst->strides[is_uv], cm->rst_tmpbuf, optimized_lr); |
215 | |
|
216 | 0 | return sse_restoration_unit(limits, rsc->src, rsc->dst, plane, highbd); |
217 | 0 | } |
218 | | |
219 | | int64_t av1_lowbd_pixel_proj_error_c(const uint8_t *src8, int width, int height, |
220 | | int src_stride, const uint8_t *dat8, |
221 | | int dat_stride, int32_t *flt0, |
222 | | int flt0_stride, int32_t *flt1, |
223 | | int flt1_stride, int xq[2], |
224 | 0 | const sgr_params_type *params) { |
225 | 0 | int i, j; |
226 | 0 | const uint8_t *src = src8; |
227 | 0 | const uint8_t *dat = dat8; |
228 | 0 | int64_t err = 0; |
229 | 0 | if (params->r[0] > 0 && params->r[1] > 0) { |
230 | 0 | for (i = 0; i < height; ++i) { |
231 | 0 | for (j = 0; j < width; ++j) { |
232 | 0 | assert(flt1[j] < (1 << 15) && flt1[j] > -(1 << 15)); |
233 | 0 | assert(flt0[j] < (1 << 15) && flt0[j] > -(1 << 15)); |
234 | 0 | const int32_t u = (int32_t)(dat[j] << SGRPROJ_RST_BITS); |
235 | 0 | int32_t v = u << SGRPROJ_PRJ_BITS; |
236 | 0 | v += xq[0] * (flt0[j] - u) + xq[1] * (flt1[j] - u); |
237 | 0 | const int32_t e = |
238 | 0 | ROUND_POWER_OF_TWO(v, SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) - src[j]; |
239 | 0 | err += ((int64_t)e * e); |
240 | 0 | } |
241 | 0 | dat += dat_stride; |
242 | 0 | src += src_stride; |
243 | 0 | flt0 += flt0_stride; |
244 | 0 | flt1 += flt1_stride; |
245 | 0 | } |
246 | 0 | } else if (params->r[0] > 0) { |
247 | 0 | for (i = 0; i < height; ++i) { |
248 | 0 | for (j = 0; j < width; ++j) { |
249 | 0 | assert(flt0[j] < (1 << 15) && flt0[j] > -(1 << 15)); |
250 | 0 | const int32_t u = (int32_t)(dat[j] << SGRPROJ_RST_BITS); |
251 | 0 | int32_t v = u << SGRPROJ_PRJ_BITS; |
252 | 0 | v += xq[0] * (flt0[j] - u); |
253 | 0 | const int32_t e = |
254 | 0 | ROUND_POWER_OF_TWO(v, SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) - src[j]; |
255 | 0 | err += ((int64_t)e * e); |
256 | 0 | } |
257 | 0 | dat += dat_stride; |
258 | 0 | src += src_stride; |
259 | 0 | flt0 += flt0_stride; |
260 | 0 | } |
261 | 0 | } else if (params->r[1] > 0) { |
262 | 0 | for (i = 0; i < height; ++i) { |
263 | 0 | for (j = 0; j < width; ++j) { |
264 | 0 | assert(flt1[j] < (1 << 15) && flt1[j] > -(1 << 15)); |
265 | 0 | const int32_t u = (int32_t)(dat[j] << SGRPROJ_RST_BITS); |
266 | 0 | int32_t v = u << SGRPROJ_PRJ_BITS; |
267 | 0 | v += xq[1] * (flt1[j] - u); |
268 | 0 | const int32_t e = |
269 | 0 | ROUND_POWER_OF_TWO(v, SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) - src[j]; |
270 | 0 | err += ((int64_t)e * e); |
271 | 0 | } |
272 | 0 | dat += dat_stride; |
273 | 0 | src += src_stride; |
274 | 0 | flt1 += flt1_stride; |
275 | 0 | } |
276 | 0 | } else { |
277 | 0 | for (i = 0; i < height; ++i) { |
278 | 0 | for (j = 0; j < width; ++j) { |
279 | 0 | const int32_t e = (int32_t)(dat[j]) - src[j]; |
280 | 0 | err += ((int64_t)e * e); |
281 | 0 | } |
282 | 0 | dat += dat_stride; |
283 | 0 | src += src_stride; |
284 | 0 | } |
285 | 0 | } |
286 | |
|
287 | 0 | return err; |
288 | 0 | } |
289 | | |
290 | | #if CONFIG_AV1_HIGHBITDEPTH |
291 | | int64_t av1_highbd_pixel_proj_error_c(const uint8_t *src8, int width, |
292 | | int height, int src_stride, |
293 | | const uint8_t *dat8, int dat_stride, |
294 | | int32_t *flt0, int flt0_stride, |
295 | | int32_t *flt1, int flt1_stride, int xq[2], |
296 | 0 | const sgr_params_type *params) { |
297 | 0 | const uint16_t *src = CONVERT_TO_SHORTPTR(src8); |
298 | 0 | const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8); |
299 | 0 | int i, j; |
300 | 0 | int64_t err = 0; |
301 | 0 | const int32_t half = 1 << (SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS - 1); |
302 | 0 | if (params->r[0] > 0 && params->r[1] > 0) { |
303 | 0 | int xq0 = xq[0]; |
304 | 0 | int xq1 = xq[1]; |
305 | 0 | for (i = 0; i < height; ++i) { |
306 | 0 | for (j = 0; j < width; ++j) { |
307 | 0 | const int32_t d = dat[j]; |
308 | 0 | const int32_t s = src[j]; |
309 | 0 | const int32_t u = (int32_t)(d << SGRPROJ_RST_BITS); |
310 | 0 | int32_t v0 = flt0[j] - u; |
311 | 0 | int32_t v1 = flt1[j] - u; |
312 | 0 | int32_t v = half; |
313 | 0 | v += xq0 * v0; |
314 | 0 | v += xq1 * v1; |
315 | 0 | const int32_t e = (v >> (SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS)) + d - s; |
316 | 0 | err += ((int64_t)e * e); |
317 | 0 | } |
318 | 0 | dat += dat_stride; |
319 | 0 | flt0 += flt0_stride; |
320 | 0 | flt1 += flt1_stride; |
321 | 0 | src += src_stride; |
322 | 0 | } |
323 | 0 | } else if (params->r[0] > 0 || params->r[1] > 0) { |
324 | 0 | int exq; |
325 | 0 | int32_t *flt; |
326 | 0 | int flt_stride; |
327 | 0 | if (params->r[0] > 0) { |
328 | 0 | exq = xq[0]; |
329 | 0 | flt = flt0; |
330 | 0 | flt_stride = flt0_stride; |
331 | 0 | } else { |
332 | 0 | exq = xq[1]; |
333 | 0 | flt = flt1; |
334 | 0 | flt_stride = flt1_stride; |
335 | 0 | } |
336 | 0 | for (i = 0; i < height; ++i) { |
337 | 0 | for (j = 0; j < width; ++j) { |
338 | 0 | const int32_t d = dat[j]; |
339 | 0 | const int32_t s = src[j]; |
340 | 0 | const int32_t u = (int32_t)(d << SGRPROJ_RST_BITS); |
341 | 0 | int32_t v = half; |
342 | 0 | v += exq * (flt[j] - u); |
343 | 0 | const int32_t e = (v >> (SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS)) + d - s; |
344 | 0 | err += ((int64_t)e * e); |
345 | 0 | } |
346 | 0 | dat += dat_stride; |
347 | 0 | flt += flt_stride; |
348 | 0 | src += src_stride; |
349 | 0 | } |
350 | 0 | } else { |
351 | 0 | for (i = 0; i < height; ++i) { |
352 | 0 | for (j = 0; j < width; ++j) { |
353 | 0 | const int32_t d = dat[j]; |
354 | 0 | const int32_t s = src[j]; |
355 | 0 | const int32_t e = d - s; |
356 | 0 | err += ((int64_t)e * e); |
357 | 0 | } |
358 | 0 | dat += dat_stride; |
359 | 0 | src += src_stride; |
360 | 0 | } |
361 | 0 | } |
362 | 0 | return err; |
363 | 0 | } |
364 | | #endif // CONFIG_AV1_HIGHBITDEPTH |
365 | | |
366 | | static int64_t get_pixel_proj_error(const uint8_t *src8, int width, int height, |
367 | | int src_stride, const uint8_t *dat8, |
368 | | int dat_stride, int use_highbitdepth, |
369 | | int32_t *flt0, int flt0_stride, |
370 | | int32_t *flt1, int flt1_stride, int *xqd, |
371 | 0 | const sgr_params_type *params) { |
372 | 0 | int xq[2]; |
373 | 0 | av1_decode_xq(xqd, xq, params); |
374 | |
|
375 | 0 | #if CONFIG_AV1_HIGHBITDEPTH |
376 | 0 | if (use_highbitdepth) { |
377 | 0 | return av1_highbd_pixel_proj_error(src8, width, height, src_stride, dat8, |
378 | 0 | dat_stride, flt0, flt0_stride, flt1, |
379 | 0 | flt1_stride, xq, params); |
380 | |
|
381 | 0 | } else { |
382 | 0 | return av1_lowbd_pixel_proj_error(src8, width, height, src_stride, dat8, |
383 | 0 | dat_stride, flt0, flt0_stride, flt1, |
384 | 0 | flt1_stride, xq, params); |
385 | 0 | } |
386 | | #else |
387 | | (void)use_highbitdepth; |
388 | | return av1_lowbd_pixel_proj_error(src8, width, height, src_stride, dat8, |
389 | | dat_stride, flt0, flt0_stride, flt1, |
390 | | flt1_stride, xq, params); |
391 | | #endif |
392 | 0 | } |
393 | | |
394 | | #define USE_SGRPROJ_REFINEMENT_SEARCH 1 |
395 | | static int64_t finer_search_pixel_proj_error( |
396 | | const uint8_t *src8, int width, int height, int src_stride, |
397 | | const uint8_t *dat8, int dat_stride, int use_highbitdepth, int32_t *flt0, |
398 | | int flt0_stride, int32_t *flt1, int flt1_stride, int start_step, int *xqd, |
399 | 0 | const sgr_params_type *params) { |
400 | 0 | int64_t err = get_pixel_proj_error( |
401 | 0 | src8, width, height, src_stride, dat8, dat_stride, use_highbitdepth, flt0, |
402 | 0 | flt0_stride, flt1, flt1_stride, xqd, params); |
403 | 0 | (void)start_step; |
404 | 0 | #if USE_SGRPROJ_REFINEMENT_SEARCH |
405 | 0 | int64_t err2; |
406 | 0 | int tap_min[] = { SGRPROJ_PRJ_MIN0, SGRPROJ_PRJ_MIN1 }; |
407 | 0 | int tap_max[] = { SGRPROJ_PRJ_MAX0, SGRPROJ_PRJ_MAX1 }; |
408 | 0 | for (int s = start_step; s >= 1; s >>= 1) { |
409 | 0 | for (int p = 0; p < 2; ++p) { |
410 | 0 | if ((params->r[0] == 0 && p == 0) || (params->r[1] == 0 && p == 1)) { |
411 | 0 | continue; |
412 | 0 | } |
413 | 0 | int skip = 0; |
414 | 0 | do { |
415 | 0 | if (xqd[p] - s >= tap_min[p]) { |
416 | 0 | xqd[p] -= s; |
417 | 0 | err2 = |
418 | 0 | get_pixel_proj_error(src8, width, height, src_stride, dat8, |
419 | 0 | dat_stride, use_highbitdepth, flt0, |
420 | 0 | flt0_stride, flt1, flt1_stride, xqd, params); |
421 | 0 | if (err2 > err) { |
422 | 0 | xqd[p] += s; |
423 | 0 | } else { |
424 | 0 | err = err2; |
425 | 0 | skip = 1; |
426 | | // At the highest step size continue moving in the same direction |
427 | 0 | if (s == start_step) continue; |
428 | 0 | } |
429 | 0 | } |
430 | 0 | break; |
431 | 0 | } while (1); |
432 | 0 | if (skip) break; |
433 | 0 | do { |
434 | 0 | if (xqd[p] + s <= tap_max[p]) { |
435 | 0 | xqd[p] += s; |
436 | 0 | err2 = |
437 | 0 | get_pixel_proj_error(src8, width, height, src_stride, dat8, |
438 | 0 | dat_stride, use_highbitdepth, flt0, |
439 | 0 | flt0_stride, flt1, flt1_stride, xqd, params); |
440 | 0 | if (err2 > err) { |
441 | 0 | xqd[p] -= s; |
442 | 0 | } else { |
443 | 0 | err = err2; |
444 | | // At the highest step size continue moving in the same direction |
445 | 0 | if (s == start_step) continue; |
446 | 0 | } |
447 | 0 | } |
448 | 0 | break; |
449 | 0 | } while (1); |
450 | 0 | } |
451 | 0 | } |
452 | 0 | #endif // USE_SGRPROJ_REFINEMENT_SEARCH |
453 | 0 | return err; |
454 | 0 | } |
455 | | |
456 | 0 | static int64_t signed_rounded_divide(int64_t dividend, int64_t divisor) { |
457 | 0 | if (dividend < 0) |
458 | 0 | return (dividend - divisor / 2) / divisor; |
459 | 0 | else |
460 | 0 | return (dividend + divisor / 2) / divisor; |
461 | 0 | } |
462 | | |
463 | | static AOM_INLINE void calc_proj_params_r0_r1_c( |
464 | | const uint8_t *src8, int width, int height, int src_stride, |
465 | | const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, |
466 | 0 | int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2]) { |
467 | 0 | const int size = width * height; |
468 | 0 | const uint8_t *src = src8; |
469 | 0 | const uint8_t *dat = dat8; |
470 | 0 | for (int i = 0; i < height; ++i) { |
471 | 0 | for (int j = 0; j < width; ++j) { |
472 | 0 | const int32_t u = (int32_t)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS); |
473 | 0 | const int32_t s = |
474 | 0 | (int32_t)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u; |
475 | 0 | const int32_t f1 = (int32_t)flt0[i * flt0_stride + j] - u; |
476 | 0 | const int32_t f2 = (int32_t)flt1[i * flt1_stride + j] - u; |
477 | 0 | H[0][0] += (int64_t)f1 * f1; |
478 | 0 | H[1][1] += (int64_t)f2 * f2; |
479 | 0 | H[0][1] += (int64_t)f1 * f2; |
480 | 0 | C[0] += (int64_t)f1 * s; |
481 | 0 | C[1] += (int64_t)f2 * s; |
482 | 0 | } |
483 | 0 | } |
484 | 0 | H[0][0] /= size; |
485 | 0 | H[0][1] /= size; |
486 | 0 | H[1][1] /= size; |
487 | 0 | H[1][0] = H[0][1]; |
488 | 0 | C[0] /= size; |
489 | 0 | C[1] /= size; |
490 | 0 | } |
491 | | |
492 | | static AOM_INLINE void calc_proj_params_r0_r1_high_bd_c( |
493 | | const uint8_t *src8, int width, int height, int src_stride, |
494 | | const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, |
495 | 0 | int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2]) { |
496 | 0 | const int size = width * height; |
497 | 0 | const uint16_t *src = CONVERT_TO_SHORTPTR(src8); |
498 | 0 | const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8); |
499 | 0 | for (int i = 0; i < height; ++i) { |
500 | 0 | for (int j = 0; j < width; ++j) { |
501 | 0 | const int32_t u = (int32_t)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS); |
502 | 0 | const int32_t s = |
503 | 0 | (int32_t)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u; |
504 | 0 | const int32_t f1 = (int32_t)flt0[i * flt0_stride + j] - u; |
505 | 0 | const int32_t f2 = (int32_t)flt1[i * flt1_stride + j] - u; |
506 | 0 | H[0][0] += (int64_t)f1 * f1; |
507 | 0 | H[1][1] += (int64_t)f2 * f2; |
508 | 0 | H[0][1] += (int64_t)f1 * f2; |
509 | 0 | C[0] += (int64_t)f1 * s; |
510 | 0 | C[1] += (int64_t)f2 * s; |
511 | 0 | } |
512 | 0 | } |
513 | 0 | H[0][0] /= size; |
514 | 0 | H[0][1] /= size; |
515 | 0 | H[1][1] /= size; |
516 | 0 | H[1][0] = H[0][1]; |
517 | 0 | C[0] /= size; |
518 | 0 | C[1] /= size; |
519 | 0 | } |
520 | | |
521 | | static AOM_INLINE void calc_proj_params_r0_c(const uint8_t *src8, int width, |
522 | | int height, int src_stride, |
523 | | const uint8_t *dat8, |
524 | | int dat_stride, int32_t *flt0, |
525 | | int flt0_stride, int64_t H[2][2], |
526 | 0 | int64_t C[2]) { |
527 | 0 | const int size = width * height; |
528 | 0 | const uint8_t *src = src8; |
529 | 0 | const uint8_t *dat = dat8; |
530 | 0 | for (int i = 0; i < height; ++i) { |
531 | 0 | for (int j = 0; j < width; ++j) { |
532 | 0 | const int32_t u = (int32_t)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS); |
533 | 0 | const int32_t s = |
534 | 0 | (int32_t)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u; |
535 | 0 | const int32_t f1 = (int32_t)flt0[i * flt0_stride + j] - u; |
536 | 0 | H[0][0] += (int64_t)f1 * f1; |
537 | 0 | C[0] += (int64_t)f1 * s; |
538 | 0 | } |
539 | 0 | } |
540 | 0 | H[0][0] /= size; |
541 | 0 | C[0] /= size; |
542 | 0 | } |
543 | | |
544 | | static AOM_INLINE void calc_proj_params_r0_high_bd_c( |
545 | | const uint8_t *src8, int width, int height, int src_stride, |
546 | | const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride, |
547 | 0 | int64_t H[2][2], int64_t C[2]) { |
548 | 0 | const int size = width * height; |
549 | 0 | const uint16_t *src = CONVERT_TO_SHORTPTR(src8); |
550 | 0 | const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8); |
551 | 0 | for (int i = 0; i < height; ++i) { |
552 | 0 | for (int j = 0; j < width; ++j) { |
553 | 0 | const int32_t u = (int32_t)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS); |
554 | 0 | const int32_t s = |
555 | 0 | (int32_t)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u; |
556 | 0 | const int32_t f1 = (int32_t)flt0[i * flt0_stride + j] - u; |
557 | 0 | H[0][0] += (int64_t)f1 * f1; |
558 | 0 | C[0] += (int64_t)f1 * s; |
559 | 0 | } |
560 | 0 | } |
561 | 0 | H[0][0] /= size; |
562 | 0 | C[0] /= size; |
563 | 0 | } |
564 | | |
565 | | static AOM_INLINE void calc_proj_params_r1_c(const uint8_t *src8, int width, |
566 | | int height, int src_stride, |
567 | | const uint8_t *dat8, |
568 | | int dat_stride, int32_t *flt1, |
569 | | int flt1_stride, int64_t H[2][2], |
570 | 0 | int64_t C[2]) { |
571 | 0 | const int size = width * height; |
572 | 0 | const uint8_t *src = src8; |
573 | 0 | const uint8_t *dat = dat8; |
574 | 0 | for (int i = 0; i < height; ++i) { |
575 | 0 | for (int j = 0; j < width; ++j) { |
576 | 0 | const int32_t u = (int32_t)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS); |
577 | 0 | const int32_t s = |
578 | 0 | (int32_t)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u; |
579 | 0 | const int32_t f2 = (int32_t)flt1[i * flt1_stride + j] - u; |
580 | 0 | H[1][1] += (int64_t)f2 * f2; |
581 | 0 | C[1] += (int64_t)f2 * s; |
582 | 0 | } |
583 | 0 | } |
584 | 0 | H[1][1] /= size; |
585 | 0 | C[1] /= size; |
586 | 0 | } |
587 | | |
588 | | static AOM_INLINE void calc_proj_params_r1_high_bd_c( |
589 | | const uint8_t *src8, int width, int height, int src_stride, |
590 | | const uint8_t *dat8, int dat_stride, int32_t *flt1, int flt1_stride, |
591 | 0 | int64_t H[2][2], int64_t C[2]) { |
592 | 0 | const int size = width * height; |
593 | 0 | const uint16_t *src = CONVERT_TO_SHORTPTR(src8); |
594 | 0 | const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8); |
595 | 0 | for (int i = 0; i < height; ++i) { |
596 | 0 | for (int j = 0; j < width; ++j) { |
597 | 0 | const int32_t u = (int32_t)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS); |
598 | 0 | const int32_t s = |
599 | 0 | (int32_t)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u; |
600 | 0 | const int32_t f2 = (int32_t)flt1[i * flt1_stride + j] - u; |
601 | 0 | H[1][1] += (int64_t)f2 * f2; |
602 | 0 | C[1] += (int64_t)f2 * s; |
603 | 0 | } |
604 | 0 | } |
605 | 0 | H[1][1] /= size; |
606 | 0 | C[1] /= size; |
607 | 0 | } |
608 | | |
609 | | // The function calls 3 subfunctions for the following cases : |
610 | | // 1) When params->r[0] > 0 and params->r[1] > 0. In this case all elements |
611 | | // of C and H need to be computed. |
612 | | // 2) When only params->r[0] > 0. In this case only H[0][0] and C[0] are |
613 | | // non-zero and need to be computed. |
614 | | // 3) When only params->r[1] > 0. In this case only H[1][1] and C[1] are |
615 | | // non-zero and need to be computed. |
616 | | void av1_calc_proj_params_c(const uint8_t *src8, int width, int height, |
617 | | int src_stride, const uint8_t *dat8, int dat_stride, |
618 | | int32_t *flt0, int flt0_stride, int32_t *flt1, |
619 | | int flt1_stride, int64_t H[2][2], int64_t C[2], |
620 | 0 | const sgr_params_type *params) { |
621 | 0 | if ((params->r[0] > 0) && (params->r[1] > 0)) { |
622 | 0 | calc_proj_params_r0_r1_c(src8, width, height, src_stride, dat8, dat_stride, |
623 | 0 | flt0, flt0_stride, flt1, flt1_stride, H, C); |
624 | 0 | } else if (params->r[0] > 0) { |
625 | 0 | calc_proj_params_r0_c(src8, width, height, src_stride, dat8, dat_stride, |
626 | 0 | flt0, flt0_stride, H, C); |
627 | 0 | } else if (params->r[1] > 0) { |
628 | 0 | calc_proj_params_r1_c(src8, width, height, src_stride, dat8, dat_stride, |
629 | 0 | flt1, flt1_stride, H, C); |
630 | 0 | } |
631 | 0 | } |
632 | | |
633 | | void av1_calc_proj_params_high_bd_c(const uint8_t *src8, int width, int height, |
634 | | int src_stride, const uint8_t *dat8, |
635 | | int dat_stride, int32_t *flt0, |
636 | | int flt0_stride, int32_t *flt1, |
637 | | int flt1_stride, int64_t H[2][2], |
638 | | int64_t C[2], |
639 | 0 | const sgr_params_type *params) { |
640 | 0 | if ((params->r[0] > 0) && (params->r[1] > 0)) { |
641 | 0 | calc_proj_params_r0_r1_high_bd_c(src8, width, height, src_stride, dat8, |
642 | 0 | dat_stride, flt0, flt0_stride, flt1, |
643 | 0 | flt1_stride, H, C); |
644 | 0 | } else if (params->r[0] > 0) { |
645 | 0 | calc_proj_params_r0_high_bd_c(src8, width, height, src_stride, dat8, |
646 | 0 | dat_stride, flt0, flt0_stride, H, C); |
647 | 0 | } else if (params->r[1] > 0) { |
648 | 0 | calc_proj_params_r1_high_bd_c(src8, width, height, src_stride, dat8, |
649 | 0 | dat_stride, flt1, flt1_stride, H, C); |
650 | 0 | } |
651 | 0 | } |
652 | | |
653 | | static AOM_INLINE void get_proj_subspace(const uint8_t *src8, int width, |
654 | | int height, int src_stride, |
655 | | const uint8_t *dat8, int dat_stride, |
656 | | int use_highbitdepth, int32_t *flt0, |
657 | | int flt0_stride, int32_t *flt1, |
658 | | int flt1_stride, int *xq, |
659 | 0 | const sgr_params_type *params) { |
660 | 0 | int64_t H[2][2] = { { 0, 0 }, { 0, 0 } }; |
661 | 0 | int64_t C[2] = { 0, 0 }; |
662 | | |
663 | | // Default values to be returned if the problem becomes ill-posed |
664 | 0 | xq[0] = 0; |
665 | 0 | xq[1] = 0; |
666 | |
|
667 | 0 | if (!use_highbitdepth) { |
668 | 0 | if ((width & 0x7) == 0) { |
669 | 0 | av1_calc_proj_params(src8, width, height, src_stride, dat8, dat_stride, |
670 | 0 | flt0, flt0_stride, flt1, flt1_stride, H, C, params); |
671 | 0 | } else { |
672 | 0 | av1_calc_proj_params_c(src8, width, height, src_stride, dat8, dat_stride, |
673 | 0 | flt0, flt0_stride, flt1, flt1_stride, H, C, |
674 | 0 | params); |
675 | 0 | } |
676 | 0 | } |
677 | 0 | #if CONFIG_AV1_HIGHBITDEPTH |
678 | 0 | else { // NOLINT |
679 | 0 | if ((width & 0x7) == 0) { |
680 | 0 | av1_calc_proj_params_high_bd(src8, width, height, src_stride, dat8, |
681 | 0 | dat_stride, flt0, flt0_stride, flt1, |
682 | 0 | flt1_stride, H, C, params); |
683 | 0 | } else { |
684 | 0 | av1_calc_proj_params_high_bd_c(src8, width, height, src_stride, dat8, |
685 | 0 | dat_stride, flt0, flt0_stride, flt1, |
686 | 0 | flt1_stride, H, C, params); |
687 | 0 | } |
688 | 0 | } |
689 | 0 | #endif |
690 | |
|
691 | 0 | if (params->r[0] == 0) { |
692 | | // H matrix is now only the scalar H[1][1] |
693 | | // C vector is now only the scalar C[1] |
694 | 0 | const int64_t Det = H[1][1]; |
695 | 0 | if (Det == 0) return; // ill-posed, return default values |
696 | 0 | xq[0] = 0; |
697 | 0 | xq[1] = (int)signed_rounded_divide(C[1] * (1 << SGRPROJ_PRJ_BITS), Det); |
698 | 0 | } else if (params->r[1] == 0) { |
699 | | // H matrix is now only the scalar H[0][0] |
700 | | // C vector is now only the scalar C[0] |
701 | 0 | const int64_t Det = H[0][0]; |
702 | 0 | if (Det == 0) return; // ill-posed, return default values |
703 | 0 | xq[0] = (int)signed_rounded_divide(C[0] * (1 << SGRPROJ_PRJ_BITS), Det); |
704 | 0 | xq[1] = 0; |
705 | 0 | } else { |
706 | 0 | const int64_t Det = H[0][0] * H[1][1] - H[0][1] * H[1][0]; |
707 | 0 | if (Det == 0) return; // ill-posed, return default values |
708 | | |
709 | | // If scaling up dividend would overflow, instead scale down the divisor |
710 | 0 | const int64_t div1 = H[1][1] * C[0] - H[0][1] * C[1]; |
711 | 0 | if ((div1 > 0 && INT64_MAX / (1 << SGRPROJ_PRJ_BITS) < div1) || |
712 | 0 | (div1 < 0 && INT64_MIN / (1 << SGRPROJ_PRJ_BITS) > div1)) |
713 | 0 | xq[0] = (int)signed_rounded_divide(div1, Det / (1 << SGRPROJ_PRJ_BITS)); |
714 | 0 | else |
715 | 0 | xq[0] = (int)signed_rounded_divide(div1 * (1 << SGRPROJ_PRJ_BITS), Det); |
716 | |
|
717 | 0 | const int64_t div2 = H[0][0] * C[1] - H[1][0] * C[0]; |
718 | 0 | if ((div2 > 0 && INT64_MAX / (1 << SGRPROJ_PRJ_BITS) < div2) || |
719 | 0 | (div2 < 0 && INT64_MIN / (1 << SGRPROJ_PRJ_BITS) > div2)) |
720 | 0 | xq[1] = (int)signed_rounded_divide(div2, Det / (1 << SGRPROJ_PRJ_BITS)); |
721 | 0 | else |
722 | 0 | xq[1] = (int)signed_rounded_divide(div2 * (1 << SGRPROJ_PRJ_BITS), Det); |
723 | 0 | } |
724 | 0 | } |
725 | | |
726 | | static AOM_INLINE void encode_xq(int *xq, int *xqd, |
727 | 0 | const sgr_params_type *params) { |
728 | 0 | if (params->r[0] == 0) { |
729 | 0 | xqd[0] = 0; |
730 | 0 | xqd[1] = clamp((1 << SGRPROJ_PRJ_BITS) - xq[1], SGRPROJ_PRJ_MIN1, |
731 | 0 | SGRPROJ_PRJ_MAX1); |
732 | 0 | } else if (params->r[1] == 0) { |
733 | 0 | xqd[0] = clamp(xq[0], SGRPROJ_PRJ_MIN0, SGRPROJ_PRJ_MAX0); |
734 | 0 | xqd[1] = clamp((1 << SGRPROJ_PRJ_BITS) - xqd[0], SGRPROJ_PRJ_MIN1, |
735 | 0 | SGRPROJ_PRJ_MAX1); |
736 | 0 | } else { |
737 | 0 | xqd[0] = clamp(xq[0], SGRPROJ_PRJ_MIN0, SGRPROJ_PRJ_MAX0); |
738 | 0 | xqd[1] = clamp((1 << SGRPROJ_PRJ_BITS) - xqd[0] - xq[1], SGRPROJ_PRJ_MIN1, |
739 | 0 | SGRPROJ_PRJ_MAX1); |
740 | 0 | } |
741 | 0 | } |
742 | | |
743 | | // Apply the self-guided filter across an entire restoration unit. |
744 | | static AOM_INLINE void apply_sgr(int sgr_params_idx, const uint8_t *dat8, |
745 | | int width, int height, int dat_stride, |
746 | | int use_highbd, int bit_depth, int pu_width, |
747 | | int pu_height, int32_t *flt0, int32_t *flt1, |
748 | 0 | int flt_stride) { |
749 | 0 | for (int i = 0; i < height; i += pu_height) { |
750 | 0 | const int h = AOMMIN(pu_height, height - i); |
751 | 0 | int32_t *flt0_row = flt0 + i * flt_stride; |
752 | 0 | int32_t *flt1_row = flt1 + i * flt_stride; |
753 | 0 | const uint8_t *dat8_row = dat8 + i * dat_stride; |
754 | | |
755 | | // Iterate over the stripe in blocks of width pu_width |
756 | 0 | for (int j = 0; j < width; j += pu_width) { |
757 | 0 | const int w = AOMMIN(pu_width, width - j); |
758 | 0 | const int ret = av1_selfguided_restoration( |
759 | 0 | dat8_row + j, w, h, dat_stride, flt0_row + j, flt1_row + j, |
760 | 0 | flt_stride, sgr_params_idx, bit_depth, use_highbd); |
761 | 0 | (void)ret; |
762 | 0 | assert(!ret); |
763 | 0 | } |
764 | 0 | } |
765 | 0 | } |
766 | | |
767 | | static AOM_INLINE void compute_sgrproj_err( |
768 | | const uint8_t *dat8, const int width, const int height, |
769 | | const int dat_stride, const uint8_t *src8, const int src_stride, |
770 | | const int use_highbitdepth, const int bit_depth, const int pu_width, |
771 | | const int pu_height, const int ep, int32_t *flt0, int32_t *flt1, |
772 | 0 | const int flt_stride, int *exqd, int64_t *err) { |
773 | 0 | int exq[2]; |
774 | 0 | apply_sgr(ep, dat8, width, height, dat_stride, use_highbitdepth, bit_depth, |
775 | 0 | pu_width, pu_height, flt0, flt1, flt_stride); |
776 | 0 | const sgr_params_type *const params = &av1_sgr_params[ep]; |
777 | 0 | get_proj_subspace(src8, width, height, src_stride, dat8, dat_stride, |
778 | 0 | use_highbitdepth, flt0, flt_stride, flt1, flt_stride, exq, |
779 | 0 | params); |
780 | 0 | encode_xq(exq, exqd, params); |
781 | 0 | *err = finer_search_pixel_proj_error( |
782 | 0 | src8, width, height, src_stride, dat8, dat_stride, use_highbitdepth, flt0, |
783 | 0 | flt_stride, flt1, flt_stride, 2, exqd, params); |
784 | 0 | } |
785 | | |
786 | | static AOM_INLINE void get_best_error(int64_t *besterr, const int64_t err, |
787 | | const int *exqd, int *bestxqd, |
788 | 0 | int *bestep, const int ep) { |
789 | 0 | if (*besterr == -1 || err < *besterr) { |
790 | 0 | *bestep = ep; |
791 | 0 | *besterr = err; |
792 | 0 | bestxqd[0] = exqd[0]; |
793 | 0 | bestxqd[1] = exqd[1]; |
794 | 0 | } |
795 | 0 | } |
796 | | |
797 | | static SgrprojInfo search_selfguided_restoration( |
798 | | const uint8_t *dat8, int width, int height, int dat_stride, |
799 | | const uint8_t *src8, int src_stride, int use_highbitdepth, int bit_depth, |
800 | 0 | int pu_width, int pu_height, int32_t *rstbuf, int enable_sgr_ep_pruning) { |
801 | 0 | int32_t *flt0 = rstbuf; |
802 | 0 | int32_t *flt1 = flt0 + RESTORATION_UNITPELS_MAX; |
803 | 0 | int ep, idx, bestep = 0; |
804 | 0 | int64_t besterr = -1; |
805 | 0 | int exqd[2], bestxqd[2] = { 0, 0 }; |
806 | 0 | int flt_stride = ((width + 7) & ~7) + 8; |
807 | 0 | assert(pu_width == (RESTORATION_PROC_UNIT_SIZE >> 1) || |
808 | 0 | pu_width == RESTORATION_PROC_UNIT_SIZE); |
809 | 0 | assert(pu_height == (RESTORATION_PROC_UNIT_SIZE >> 1) || |
810 | 0 | pu_height == RESTORATION_PROC_UNIT_SIZE); |
811 | 0 | if (!enable_sgr_ep_pruning) { |
812 | 0 | for (ep = 0; ep < SGRPROJ_PARAMS; ep++) { |
813 | 0 | int64_t err; |
814 | 0 | compute_sgrproj_err(dat8, width, height, dat_stride, src8, src_stride, |
815 | 0 | use_highbitdepth, bit_depth, pu_width, pu_height, ep, |
816 | 0 | flt0, flt1, flt_stride, exqd, &err); |
817 | 0 | get_best_error(&besterr, err, exqd, bestxqd, &bestep, ep); |
818 | 0 | } |
819 | 0 | } else { |
820 | | // evaluate first four seed ep in first group |
821 | 0 | for (idx = 0; idx < SGRPROJ_EP_GRP1_SEARCH_COUNT; idx++) { |
822 | 0 | ep = sgproj_ep_grp1_seed[idx]; |
823 | 0 | int64_t err; |
824 | 0 | compute_sgrproj_err(dat8, width, height, dat_stride, src8, src_stride, |
825 | 0 | use_highbitdepth, bit_depth, pu_width, pu_height, ep, |
826 | 0 | flt0, flt1, flt_stride, exqd, &err); |
827 | 0 | get_best_error(&besterr, err, exqd, bestxqd, &bestep, ep); |
828 | 0 | } |
829 | | // evaluate left and right ep of winner in seed ep |
830 | 0 | int bestep_ref = bestep; |
831 | 0 | for (ep = bestep_ref - 1; ep < bestep_ref + 2; ep += 2) { |
832 | 0 | if (ep < SGRPROJ_EP_GRP1_START_IDX || ep > SGRPROJ_EP_GRP1_END_IDX) |
833 | 0 | continue; |
834 | 0 | int64_t err; |
835 | 0 | compute_sgrproj_err(dat8, width, height, dat_stride, src8, src_stride, |
836 | 0 | use_highbitdepth, bit_depth, pu_width, pu_height, ep, |
837 | 0 | flt0, flt1, flt_stride, exqd, &err); |
838 | 0 | get_best_error(&besterr, err, exqd, bestxqd, &bestep, ep); |
839 | 0 | } |
840 | | // evaluate last two group |
841 | 0 | for (idx = 0; idx < SGRPROJ_EP_GRP2_3_SEARCH_COUNT; idx++) { |
842 | 0 | ep = sgproj_ep_grp2_3[idx][bestep]; |
843 | 0 | int64_t err; |
844 | 0 | compute_sgrproj_err(dat8, width, height, dat_stride, src8, src_stride, |
845 | 0 | use_highbitdepth, bit_depth, pu_width, pu_height, ep, |
846 | 0 | flt0, flt1, flt_stride, exqd, &err); |
847 | 0 | get_best_error(&besterr, err, exqd, bestxqd, &bestep, ep); |
848 | 0 | } |
849 | 0 | } |
850 | |
|
851 | 0 | SgrprojInfo ret; |
852 | 0 | ret.ep = bestep; |
853 | 0 | ret.xqd[0] = bestxqd[0]; |
854 | 0 | ret.xqd[1] = bestxqd[1]; |
855 | 0 | return ret; |
856 | 0 | } |
857 | | |
858 | | static int count_sgrproj_bits(SgrprojInfo *sgrproj_info, |
859 | 0 | SgrprojInfo *ref_sgrproj_info) { |
860 | 0 | int bits = SGRPROJ_PARAMS_BITS; |
861 | 0 | const sgr_params_type *params = &av1_sgr_params[sgrproj_info->ep]; |
862 | 0 | if (params->r[0] > 0) |
863 | 0 | bits += aom_count_primitive_refsubexpfin( |
864 | 0 | SGRPROJ_PRJ_MAX0 - SGRPROJ_PRJ_MIN0 + 1, SGRPROJ_PRJ_SUBEXP_K, |
865 | 0 | ref_sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0, |
866 | 0 | sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0); |
867 | 0 | if (params->r[1] > 0) |
868 | 0 | bits += aom_count_primitive_refsubexpfin( |
869 | 0 | SGRPROJ_PRJ_MAX1 - SGRPROJ_PRJ_MIN1 + 1, SGRPROJ_PRJ_SUBEXP_K, |
870 | 0 | ref_sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1, |
871 | 0 | sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1); |
872 | 0 | return bits; |
873 | 0 | } |
874 | | |
875 | | static AOM_INLINE void search_sgrproj(const RestorationTileLimits *limits, |
876 | | const AV1PixelRect *tile, |
877 | | int rest_unit_idx, void *priv, |
878 | | int32_t *tmpbuf, |
879 | 0 | RestorationLineBuffers *rlbs) { |
880 | 0 | (void)rlbs; |
881 | 0 | RestSearchCtxt *rsc = (RestSearchCtxt *)priv; |
882 | 0 | RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx]; |
883 | |
|
884 | 0 | const MACROBLOCK *const x = rsc->x; |
885 | 0 | const AV1_COMMON *const cm = rsc->cm; |
886 | 0 | const int highbd = cm->seq_params->use_highbitdepth; |
887 | 0 | const int bit_depth = cm->seq_params->bit_depth; |
888 | |
|
889 | 0 | const int64_t bits_none = x->mode_costs.sgrproj_restore_cost[0]; |
890 | | // Prune evaluation of RESTORE_SGRPROJ if 'skip_sgr_eval' is set |
891 | 0 | if (rusi->skip_sgr_eval) { |
892 | 0 | rsc->bits += bits_none; |
893 | 0 | rsc->sse += rusi->sse[RESTORE_NONE]; |
894 | 0 | rusi->best_rtype[RESTORE_SGRPROJ - 1] = RESTORE_NONE; |
895 | 0 | rusi->sse[RESTORE_SGRPROJ] = INT64_MAX; |
896 | 0 | return; |
897 | 0 | } |
898 | | |
899 | 0 | uint8_t *dgd_start = |
900 | 0 | rsc->dgd_buffer + limits->v_start * rsc->dgd_stride + limits->h_start; |
901 | 0 | const uint8_t *src_start = |
902 | 0 | rsc->src_buffer + limits->v_start * rsc->src_stride + limits->h_start; |
903 | |
|
904 | 0 | const int is_uv = rsc->plane > 0; |
905 | 0 | const int ss_x = is_uv && cm->seq_params->subsampling_x; |
906 | 0 | const int ss_y = is_uv && cm->seq_params->subsampling_y; |
907 | 0 | const int procunit_width = RESTORATION_PROC_UNIT_SIZE >> ss_x; |
908 | 0 | const int procunit_height = RESTORATION_PROC_UNIT_SIZE >> ss_y; |
909 | |
|
910 | 0 | rusi->sgrproj = search_selfguided_restoration( |
911 | 0 | dgd_start, limits->h_end - limits->h_start, |
912 | 0 | limits->v_end - limits->v_start, rsc->dgd_stride, src_start, |
913 | 0 | rsc->src_stride, highbd, bit_depth, procunit_width, procunit_height, |
914 | 0 | tmpbuf, rsc->lpf_sf->enable_sgr_ep_pruning); |
915 | |
|
916 | 0 | RestorationUnitInfo rui; |
917 | 0 | rui.restoration_type = RESTORE_SGRPROJ; |
918 | 0 | rui.sgrproj_info = rusi->sgrproj; |
919 | |
|
920 | 0 | rusi->sse[RESTORE_SGRPROJ] = try_restoration_unit(rsc, limits, tile, &rui); |
921 | |
|
922 | 0 | const int64_t bits_sgr = x->mode_costs.sgrproj_restore_cost[1] + |
923 | 0 | (count_sgrproj_bits(&rusi->sgrproj, &rsc->sgrproj) |
924 | 0 | << AV1_PROB_COST_SHIFT); |
925 | 0 | double cost_none = RDCOST_DBL_WITH_NATIVE_BD_DIST( |
926 | 0 | x->rdmult, bits_none >> 4, rusi->sse[RESTORE_NONE], bit_depth); |
927 | 0 | double cost_sgr = RDCOST_DBL_WITH_NATIVE_BD_DIST( |
928 | 0 | x->rdmult, bits_sgr >> 4, rusi->sse[RESTORE_SGRPROJ], bit_depth); |
929 | 0 | if (rusi->sgrproj.ep < 10) |
930 | 0 | cost_sgr *= |
931 | 0 | (1 + DUAL_SGR_PENALTY_MULT * rsc->lpf_sf->dual_sgr_penalty_level); |
932 | |
|
933 | 0 | RestorationType rtype = |
934 | 0 | (cost_sgr < cost_none) ? RESTORE_SGRPROJ : RESTORE_NONE; |
935 | 0 | rusi->best_rtype[RESTORE_SGRPROJ - 1] = rtype; |
936 | |
|
937 | 0 | rsc->sse += rusi->sse[rtype]; |
938 | 0 | rsc->bits += (cost_sgr < cost_none) ? bits_sgr : bits_none; |
939 | 0 | if (cost_sgr < cost_none) rsc->sgrproj = rusi->sgrproj; |
940 | 0 | } |
941 | | |
942 | | void acc_stat_one_line(const uint8_t *dgd, const uint8_t *src, int dgd_stride, |
943 | | int h_start, int h_end, uint8_t avg, |
944 | | const int wiener_halfwin, const int wiener_win2, |
945 | 0 | int32_t *M_int32, int32_t *H_int32, int count) { |
946 | 0 | int j, k, l; |
947 | 0 | int16_t Y[WIENER_WIN2]; |
948 | |
|
949 | 0 | for (j = h_start; j < h_end; j++) { |
950 | 0 | const int16_t X = (int16_t)src[j] - (int16_t)avg; |
951 | 0 | int idx = 0; |
952 | 0 | for (k = -wiener_halfwin; k <= wiener_halfwin; k++) { |
953 | 0 | for (l = -wiener_halfwin; l <= wiener_halfwin; l++) { |
954 | 0 | Y[idx] = |
955 | 0 | (int16_t)dgd[(count + l) * dgd_stride + (j + k)] - (int16_t)avg; |
956 | 0 | idx++; |
957 | 0 | } |
958 | 0 | } |
959 | 0 | assert(idx == wiener_win2); |
960 | 0 | for (k = 0; k < wiener_win2; ++k) { |
961 | 0 | M_int32[k] += (int32_t)Y[k] * X; |
962 | 0 | for (l = k; l < wiener_win2; ++l) { |
963 | | // H is a symmetric matrix, so we only need to fill out the upper |
964 | | // triangle here. We can copy it down to the lower triangle outside |
965 | | // the (i, j) loops. |
966 | 0 | H_int32[k * wiener_win2 + l] += (int32_t)Y[k] * Y[l]; |
967 | 0 | } |
968 | 0 | } |
969 | 0 | } |
970 | 0 | } |
971 | | |
972 | | void av1_compute_stats_c(int wiener_win, const uint8_t *dgd, const uint8_t *src, |
973 | | int h_start, int h_end, int v_start, int v_end, |
974 | | int dgd_stride, int src_stride, int64_t *M, int64_t *H, |
975 | 0 | int use_downsampled_wiener_stats) { |
976 | 0 | int i, k, l; |
977 | 0 | const int wiener_win2 = wiener_win * wiener_win; |
978 | 0 | const int wiener_halfwin = (wiener_win >> 1); |
979 | 0 | uint8_t avg = find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride); |
980 | 0 | int32_t M_row[WIENER_WIN2] = { 0 }; |
981 | 0 | int32_t H_row[WIENER_WIN2 * WIENER_WIN2] = { 0 }; |
982 | 0 | int downsample_factor = |
983 | 0 | use_downsampled_wiener_stats ? WIENER_STATS_DOWNSAMPLE_FACTOR : 1; |
984 | |
|
985 | 0 | memset(M, 0, sizeof(*M) * wiener_win2); |
986 | 0 | memset(H, 0, sizeof(*H) * wiener_win2 * wiener_win2); |
987 | |
|
988 | 0 | for (i = v_start; i < v_end; i = i + downsample_factor) { |
989 | 0 | if (use_downsampled_wiener_stats && |
990 | 0 | (v_end - i < WIENER_STATS_DOWNSAMPLE_FACTOR)) { |
991 | 0 | downsample_factor = v_end - i; |
992 | 0 | } |
993 | |
|
994 | 0 | memset(M_row, 0, sizeof(int32_t) * WIENER_WIN2); |
995 | 0 | memset(H_row, 0, sizeof(int32_t) * WIENER_WIN2 * WIENER_WIN2); |
996 | 0 | acc_stat_one_line(dgd, src + i * src_stride, dgd_stride, h_start, h_end, |
997 | 0 | avg, wiener_halfwin, wiener_win2, M_row, H_row, i); |
998 | |
|
999 | 0 | for (k = 0; k < wiener_win2; ++k) { |
1000 | | // Scale M matrix based on the downsampling factor |
1001 | 0 | M[k] += ((int64_t)M_row[k] * downsample_factor); |
1002 | 0 | for (l = k; l < wiener_win2; ++l) { |
1003 | | // H is a symmetric matrix, so we only need to fill out the upper |
1004 | | // triangle here. We can copy it down to the lower triangle outside |
1005 | | // the (i, j) loops. |
1006 | | // Scale H Matrix based on the downsampling factor |
1007 | 0 | H[k * wiener_win2 + l] += |
1008 | 0 | ((int64_t)H_row[k * wiener_win2 + l] * downsample_factor); |
1009 | 0 | } |
1010 | 0 | } |
1011 | 0 | } |
1012 | |
|
1013 | 0 | for (k = 0; k < wiener_win2; ++k) { |
1014 | 0 | for (l = k + 1; l < wiener_win2; ++l) { |
1015 | 0 | H[l * wiener_win2 + k] = H[k * wiener_win2 + l]; |
1016 | 0 | } |
1017 | 0 | } |
1018 | 0 | } |
1019 | | |
1020 | | #if CONFIG_AV1_HIGHBITDEPTH |
1021 | | void av1_compute_stats_highbd_c(int wiener_win, const uint8_t *dgd8, |
1022 | | const uint8_t *src8, int h_start, int h_end, |
1023 | | int v_start, int v_end, int dgd_stride, |
1024 | | int src_stride, int64_t *M, int64_t *H, |
1025 | 0 | aom_bit_depth_t bit_depth) { |
1026 | 0 | int i, j, k, l; |
1027 | 0 | int32_t Y[WIENER_WIN2]; |
1028 | 0 | const int wiener_win2 = wiener_win * wiener_win; |
1029 | 0 | const int wiener_halfwin = (wiener_win >> 1); |
1030 | 0 | const uint16_t *src = CONVERT_TO_SHORTPTR(src8); |
1031 | 0 | const uint16_t *dgd = CONVERT_TO_SHORTPTR(dgd8); |
1032 | 0 | uint16_t avg = |
1033 | 0 | find_average_highbd(dgd, h_start, h_end, v_start, v_end, dgd_stride); |
1034 | |
|
1035 | 0 | uint8_t bit_depth_divider = 1; |
1036 | 0 | if (bit_depth == AOM_BITS_12) |
1037 | 0 | bit_depth_divider = 16; |
1038 | 0 | else if (bit_depth == AOM_BITS_10) |
1039 | 0 | bit_depth_divider = 4; |
1040 | |
|
1041 | 0 | memset(M, 0, sizeof(*M) * wiener_win2); |
1042 | 0 | memset(H, 0, sizeof(*H) * wiener_win2 * wiener_win2); |
1043 | 0 | for (i = v_start; i < v_end; i++) { |
1044 | 0 | for (j = h_start; j < h_end; j++) { |
1045 | 0 | const int32_t X = (int32_t)src[i * src_stride + j] - (int32_t)avg; |
1046 | 0 | int idx = 0; |
1047 | 0 | for (k = -wiener_halfwin; k <= wiener_halfwin; k++) { |
1048 | 0 | for (l = -wiener_halfwin; l <= wiener_halfwin; l++) { |
1049 | 0 | Y[idx] = (int32_t)dgd[(i + l) * dgd_stride + (j + k)] - (int32_t)avg; |
1050 | 0 | idx++; |
1051 | 0 | } |
1052 | 0 | } |
1053 | 0 | assert(idx == wiener_win2); |
1054 | 0 | for (k = 0; k < wiener_win2; ++k) { |
1055 | 0 | M[k] += (int64_t)Y[k] * X; |
1056 | 0 | for (l = k; l < wiener_win2; ++l) { |
1057 | | // H is a symmetric matrix, so we only need to fill out the upper |
1058 | | // triangle here. We can copy it down to the lower triangle outside |
1059 | | // the (i, j) loops. |
1060 | 0 | H[k * wiener_win2 + l] += (int64_t)Y[k] * Y[l]; |
1061 | 0 | } |
1062 | 0 | } |
1063 | 0 | } |
1064 | 0 | } |
1065 | 0 | for (k = 0; k < wiener_win2; ++k) { |
1066 | 0 | M[k] /= bit_depth_divider; |
1067 | 0 | H[k * wiener_win2 + k] /= bit_depth_divider; |
1068 | 0 | for (l = k + 1; l < wiener_win2; ++l) { |
1069 | 0 | H[k * wiener_win2 + l] /= bit_depth_divider; |
1070 | 0 | H[l * wiener_win2 + k] = H[k * wiener_win2 + l]; |
1071 | 0 | } |
1072 | 0 | } |
1073 | 0 | } |
1074 | | #endif // CONFIG_AV1_HIGHBITDEPTH |
1075 | | |
1076 | 0 | static INLINE int wrap_index(int i, int wiener_win) { |
1077 | 0 | const int wiener_halfwin1 = (wiener_win >> 1) + 1; |
1078 | 0 | return (i >= wiener_halfwin1 ? wiener_win - 1 - i : i); |
1079 | 0 | } |
1080 | | |
1081 | | // Solve linear equations to find Wiener filter tap values |
1082 | | // Taps are output scaled by WIENER_FILT_STEP |
1083 | | static int linsolve_wiener(int n, int64_t *A, int stride, int64_t *b, |
1084 | 0 | int32_t *x) { |
1085 | 0 | for (int k = 0; k < n - 1; k++) { |
1086 | | // Partial pivoting: bring the row with the largest pivot to the top |
1087 | 0 | for (int i = n - 1; i > k; i--) { |
1088 | | // If row i has a better (bigger) pivot than row (i-1), swap them |
1089 | 0 | if (llabs(A[(i - 1) * stride + k]) < llabs(A[i * stride + k])) { |
1090 | 0 | for (int j = 0; j < n; j++) { |
1091 | 0 | const int64_t c = A[i * stride + j]; |
1092 | 0 | A[i * stride + j] = A[(i - 1) * stride + j]; |
1093 | 0 | A[(i - 1) * stride + j] = c; |
1094 | 0 | } |
1095 | 0 | const int64_t c = b[i]; |
1096 | 0 | b[i] = b[i - 1]; |
1097 | 0 | b[i - 1] = c; |
1098 | 0 | } |
1099 | 0 | } |
1100 | | // Forward elimination (convert A to row-echelon form) |
1101 | 0 | for (int i = k; i < n - 1; i++) { |
1102 | 0 | if (A[k * stride + k] == 0) return 0; |
1103 | 0 | const int64_t c = A[(i + 1) * stride + k]; |
1104 | 0 | const int64_t cd = A[k * stride + k]; |
1105 | 0 | for (int j = 0; j < n; j++) { |
1106 | 0 | A[(i + 1) * stride + j] -= c / 256 * A[k * stride + j] / cd * 256; |
1107 | 0 | } |
1108 | 0 | b[i + 1] -= c * b[k] / cd; |
1109 | 0 | } |
1110 | 0 | } |
1111 | | // Back-substitution |
1112 | 0 | for (int i = n - 1; i >= 0; i--) { |
1113 | 0 | if (A[i * stride + i] == 0) return 0; |
1114 | 0 | int64_t c = 0; |
1115 | 0 | for (int j = i + 1; j <= n - 1; j++) { |
1116 | 0 | c += A[i * stride + j] * x[j] / WIENER_TAP_SCALE_FACTOR; |
1117 | 0 | } |
1118 | | // Store filter taps x in scaled form. |
1119 | 0 | x[i] = (int32_t)(WIENER_TAP_SCALE_FACTOR * (b[i] - c) / A[i * stride + i]); |
1120 | 0 | } |
1121 | | |
1122 | 0 | return 1; |
1123 | 0 | } |
1124 | | |
1125 | | // Fix vector b, update vector a |
1126 | | static AOM_INLINE void update_a_sep_sym(int wiener_win, int64_t **Mc, |
1127 | 0 | int64_t **Hc, int32_t *a, int32_t *b) { |
1128 | 0 | int i, j; |
1129 | 0 | int32_t S[WIENER_WIN]; |
1130 | 0 | int64_t A[WIENER_HALFWIN1], B[WIENER_HALFWIN1 * WIENER_HALFWIN1]; |
1131 | 0 | const int wiener_win2 = wiener_win * wiener_win; |
1132 | 0 | const int wiener_halfwin1 = (wiener_win >> 1) + 1; |
1133 | 0 | memset(A, 0, sizeof(A)); |
1134 | 0 | memset(B, 0, sizeof(B)); |
1135 | 0 | for (i = 0; i < wiener_win; i++) { |
1136 | 0 | for (j = 0; j < wiener_win; ++j) { |
1137 | 0 | const int jj = wrap_index(j, wiener_win); |
1138 | 0 | A[jj] += Mc[i][j] * b[i] / WIENER_TAP_SCALE_FACTOR; |
1139 | 0 | } |
1140 | 0 | } |
1141 | 0 | for (i = 0; i < wiener_win; i++) { |
1142 | 0 | for (j = 0; j < wiener_win; j++) { |
1143 | 0 | int k, l; |
1144 | 0 | for (k = 0; k < wiener_win; ++k) { |
1145 | 0 | for (l = 0; l < wiener_win; ++l) { |
1146 | 0 | const int kk = wrap_index(k, wiener_win); |
1147 | 0 | const int ll = wrap_index(l, wiener_win); |
1148 | 0 | B[ll * wiener_halfwin1 + kk] += |
1149 | 0 | Hc[j * wiener_win + i][k * wiener_win2 + l] * b[i] / |
1150 | 0 | WIENER_TAP_SCALE_FACTOR * b[j] / WIENER_TAP_SCALE_FACTOR; |
1151 | 0 | } |
1152 | 0 | } |
1153 | 0 | } |
1154 | 0 | } |
1155 | | // Normalization enforcement in the system of equations itself |
1156 | 0 | for (i = 0; i < wiener_halfwin1 - 1; ++i) { |
1157 | 0 | A[i] -= |
1158 | 0 | A[wiener_halfwin1 - 1] * 2 + |
1159 | 0 | B[i * wiener_halfwin1 + wiener_halfwin1 - 1] - |
1160 | 0 | 2 * B[(wiener_halfwin1 - 1) * wiener_halfwin1 + (wiener_halfwin1 - 1)]; |
1161 | 0 | } |
1162 | 0 | for (i = 0; i < wiener_halfwin1 - 1; ++i) { |
1163 | 0 | for (j = 0; j < wiener_halfwin1 - 1; ++j) { |
1164 | 0 | B[i * wiener_halfwin1 + j] -= |
1165 | 0 | 2 * (B[i * wiener_halfwin1 + (wiener_halfwin1 - 1)] + |
1166 | 0 | B[(wiener_halfwin1 - 1) * wiener_halfwin1 + j] - |
1167 | 0 | 2 * B[(wiener_halfwin1 - 1) * wiener_halfwin1 + |
1168 | 0 | (wiener_halfwin1 - 1)]); |
1169 | 0 | } |
1170 | 0 | } |
1171 | 0 | if (linsolve_wiener(wiener_halfwin1 - 1, B, wiener_halfwin1, A, S)) { |
1172 | 0 | S[wiener_halfwin1 - 1] = WIENER_TAP_SCALE_FACTOR; |
1173 | 0 | for (i = wiener_halfwin1; i < wiener_win; ++i) { |
1174 | 0 | S[i] = S[wiener_win - 1 - i]; |
1175 | 0 | S[wiener_halfwin1 - 1] -= 2 * S[i]; |
1176 | 0 | } |
1177 | 0 | memcpy(a, S, wiener_win * sizeof(*a)); |
1178 | 0 | } |
1179 | 0 | } |
1180 | | |
1181 | | // Fix vector a, update vector b |
1182 | | static AOM_INLINE void update_b_sep_sym(int wiener_win, int64_t **Mc, |
1183 | 0 | int64_t **Hc, int32_t *a, int32_t *b) { |
1184 | 0 | int i, j; |
1185 | 0 | int32_t S[WIENER_WIN]; |
1186 | 0 | int64_t A[WIENER_HALFWIN1], B[WIENER_HALFWIN1 * WIENER_HALFWIN1]; |
1187 | 0 | const int wiener_win2 = wiener_win * wiener_win; |
1188 | 0 | const int wiener_halfwin1 = (wiener_win >> 1) + 1; |
1189 | 0 | memset(A, 0, sizeof(A)); |
1190 | 0 | memset(B, 0, sizeof(B)); |
1191 | 0 | for (i = 0; i < wiener_win; i++) { |
1192 | 0 | const int ii = wrap_index(i, wiener_win); |
1193 | 0 | for (j = 0; j < wiener_win; j++) { |
1194 | 0 | A[ii] += Mc[i][j] * a[j] / WIENER_TAP_SCALE_FACTOR; |
1195 | 0 | } |
1196 | 0 | } |
1197 | |
|
1198 | 0 | for (i = 0; i < wiener_win; i++) { |
1199 | 0 | for (j = 0; j < wiener_win; j++) { |
1200 | 0 | const int ii = wrap_index(i, wiener_win); |
1201 | 0 | const int jj = wrap_index(j, wiener_win); |
1202 | 0 | int k, l; |
1203 | 0 | for (k = 0; k < wiener_win; ++k) { |
1204 | 0 | for (l = 0; l < wiener_win; ++l) { |
1205 | 0 | B[jj * wiener_halfwin1 + ii] += |
1206 | 0 | Hc[i * wiener_win + j][k * wiener_win2 + l] * a[k] / |
1207 | 0 | WIENER_TAP_SCALE_FACTOR * a[l] / WIENER_TAP_SCALE_FACTOR; |
1208 | 0 | } |
1209 | 0 | } |
1210 | 0 | } |
1211 | 0 | } |
1212 | | // Normalization enforcement in the system of equations itself |
1213 | 0 | for (i = 0; i < wiener_halfwin1 - 1; ++i) { |
1214 | 0 | A[i] -= |
1215 | 0 | A[wiener_halfwin1 - 1] * 2 + |
1216 | 0 | B[i * wiener_halfwin1 + wiener_halfwin1 - 1] - |
1217 | 0 | 2 * B[(wiener_halfwin1 - 1) * wiener_halfwin1 + (wiener_halfwin1 - 1)]; |
1218 | 0 | } |
1219 | 0 | for (i = 0; i < wiener_halfwin1 - 1; ++i) { |
1220 | 0 | for (j = 0; j < wiener_halfwin1 - 1; ++j) { |
1221 | 0 | B[i * wiener_halfwin1 + j] -= |
1222 | 0 | 2 * (B[i * wiener_halfwin1 + (wiener_halfwin1 - 1)] + |
1223 | 0 | B[(wiener_halfwin1 - 1) * wiener_halfwin1 + j] - |
1224 | 0 | 2 * B[(wiener_halfwin1 - 1) * wiener_halfwin1 + |
1225 | 0 | (wiener_halfwin1 - 1)]); |
1226 | 0 | } |
1227 | 0 | } |
1228 | 0 | if (linsolve_wiener(wiener_halfwin1 - 1, B, wiener_halfwin1, A, S)) { |
1229 | 0 | S[wiener_halfwin1 - 1] = WIENER_TAP_SCALE_FACTOR; |
1230 | 0 | for (i = wiener_halfwin1; i < wiener_win; ++i) { |
1231 | 0 | S[i] = S[wiener_win - 1 - i]; |
1232 | 0 | S[wiener_halfwin1 - 1] -= 2 * S[i]; |
1233 | 0 | } |
1234 | 0 | memcpy(b, S, wiener_win * sizeof(*b)); |
1235 | 0 | } |
1236 | 0 | } |
1237 | | |
1238 | | static void wiener_decompose_sep_sym(int wiener_win, int64_t *M, int64_t *H, |
1239 | 0 | int32_t *a, int32_t *b) { |
1240 | 0 | static const int32_t init_filt[WIENER_WIN] = { |
1241 | 0 | WIENER_FILT_TAP0_MIDV, WIENER_FILT_TAP1_MIDV, WIENER_FILT_TAP2_MIDV, |
1242 | 0 | WIENER_FILT_TAP3_MIDV, WIENER_FILT_TAP2_MIDV, WIENER_FILT_TAP1_MIDV, |
1243 | 0 | WIENER_FILT_TAP0_MIDV, |
1244 | 0 | }; |
1245 | 0 | int64_t *Hc[WIENER_WIN2]; |
1246 | 0 | int64_t *Mc[WIENER_WIN]; |
1247 | 0 | int i, j, iter; |
1248 | 0 | const int plane_off = (WIENER_WIN - wiener_win) >> 1; |
1249 | 0 | const int wiener_win2 = wiener_win * wiener_win; |
1250 | 0 | for (i = 0; i < wiener_win; i++) { |
1251 | 0 | a[i] = b[i] = |
1252 | 0 | WIENER_TAP_SCALE_FACTOR / WIENER_FILT_STEP * init_filt[i + plane_off]; |
1253 | 0 | } |
1254 | 0 | for (i = 0; i < wiener_win; i++) { |
1255 | 0 | Mc[i] = M + i * wiener_win; |
1256 | 0 | for (j = 0; j < wiener_win; j++) { |
1257 | 0 | Hc[i * wiener_win + j] = |
1258 | 0 | H + i * wiener_win * wiener_win2 + j * wiener_win; |
1259 | 0 | } |
1260 | 0 | } |
1261 | |
|
1262 | 0 | iter = 1; |
1263 | 0 | while (iter < NUM_WIENER_ITERS) { |
1264 | 0 | update_a_sep_sym(wiener_win, Mc, Hc, a, b); |
1265 | 0 | update_b_sep_sym(wiener_win, Mc, Hc, a, b); |
1266 | 0 | iter++; |
1267 | 0 | } |
1268 | 0 | } |
1269 | | |
1270 | | // Computes the function x'*H*x - x'*M for the learned 2D filter x, and compares |
1271 | | // against identity filters; Final score is defined as the difference between |
1272 | | // the function values |
1273 | | static int64_t compute_score(int wiener_win, int64_t *M, int64_t *H, |
1274 | 0 | InterpKernel vfilt, InterpKernel hfilt) { |
1275 | 0 | int32_t ab[WIENER_WIN * WIENER_WIN]; |
1276 | 0 | int16_t a[WIENER_WIN], b[WIENER_WIN]; |
1277 | 0 | int64_t P = 0, Q = 0; |
1278 | 0 | int64_t iP = 0, iQ = 0; |
1279 | 0 | int64_t Score, iScore; |
1280 | 0 | int i, k, l; |
1281 | 0 | const int plane_off = (WIENER_WIN - wiener_win) >> 1; |
1282 | 0 | const int wiener_win2 = wiener_win * wiener_win; |
1283 | |
|
1284 | 0 | a[WIENER_HALFWIN] = b[WIENER_HALFWIN] = WIENER_FILT_STEP; |
1285 | 0 | for (i = 0; i < WIENER_HALFWIN; ++i) { |
1286 | 0 | a[i] = a[WIENER_WIN - i - 1] = vfilt[i]; |
1287 | 0 | b[i] = b[WIENER_WIN - i - 1] = hfilt[i]; |
1288 | 0 | a[WIENER_HALFWIN] -= 2 * a[i]; |
1289 | 0 | b[WIENER_HALFWIN] -= 2 * b[i]; |
1290 | 0 | } |
1291 | 0 | memset(ab, 0, sizeof(ab)); |
1292 | 0 | for (k = 0; k < wiener_win; ++k) { |
1293 | 0 | for (l = 0; l < wiener_win; ++l) |
1294 | 0 | ab[k * wiener_win + l] = a[l + plane_off] * b[k + plane_off]; |
1295 | 0 | } |
1296 | 0 | for (k = 0; k < wiener_win2; ++k) { |
1297 | 0 | P += ab[k] * M[k] / WIENER_FILT_STEP / WIENER_FILT_STEP; |
1298 | 0 | for (l = 0; l < wiener_win2; ++l) { |
1299 | 0 | Q += ab[k] * H[k * wiener_win2 + l] * ab[l] / WIENER_FILT_STEP / |
1300 | 0 | WIENER_FILT_STEP / WIENER_FILT_STEP / WIENER_FILT_STEP; |
1301 | 0 | } |
1302 | 0 | } |
1303 | 0 | Score = Q - 2 * P; |
1304 | |
|
1305 | 0 | iP = M[wiener_win2 >> 1]; |
1306 | 0 | iQ = H[(wiener_win2 >> 1) * wiener_win2 + (wiener_win2 >> 1)]; |
1307 | 0 | iScore = iQ - 2 * iP; |
1308 | |
|
1309 | 0 | return Score - iScore; |
1310 | 0 | } |
1311 | | |
1312 | | static AOM_INLINE void finalize_sym_filter(int wiener_win, int32_t *f, |
1313 | 0 | InterpKernel fi) { |
1314 | 0 | int i; |
1315 | 0 | const int wiener_halfwin = (wiener_win >> 1); |
1316 | |
|
1317 | 0 | for (i = 0; i < wiener_halfwin; ++i) { |
1318 | 0 | const int64_t dividend = (int64_t)f[i] * WIENER_FILT_STEP; |
1319 | 0 | const int64_t divisor = WIENER_TAP_SCALE_FACTOR; |
1320 | | // Perform this division with proper rounding rather than truncation |
1321 | 0 | if (dividend < 0) { |
1322 | 0 | fi[i] = (int16_t)((dividend - (divisor / 2)) / divisor); |
1323 | 0 | } else { |
1324 | 0 | fi[i] = (int16_t)((dividend + (divisor / 2)) / divisor); |
1325 | 0 | } |
1326 | 0 | } |
1327 | | // Specialize for 7-tap filter |
1328 | 0 | if (wiener_win == WIENER_WIN) { |
1329 | 0 | fi[0] = CLIP(fi[0], WIENER_FILT_TAP0_MINV, WIENER_FILT_TAP0_MAXV); |
1330 | 0 | fi[1] = CLIP(fi[1], WIENER_FILT_TAP1_MINV, WIENER_FILT_TAP1_MAXV); |
1331 | 0 | fi[2] = CLIP(fi[2], WIENER_FILT_TAP2_MINV, WIENER_FILT_TAP2_MAXV); |
1332 | 0 | } else { |
1333 | 0 | fi[2] = CLIP(fi[1], WIENER_FILT_TAP2_MINV, WIENER_FILT_TAP2_MAXV); |
1334 | 0 | fi[1] = CLIP(fi[0], WIENER_FILT_TAP1_MINV, WIENER_FILT_TAP1_MAXV); |
1335 | 0 | fi[0] = 0; |
1336 | 0 | } |
1337 | | // Satisfy filter constraints |
1338 | 0 | fi[WIENER_WIN - 1] = fi[0]; |
1339 | 0 | fi[WIENER_WIN - 2] = fi[1]; |
1340 | 0 | fi[WIENER_WIN - 3] = fi[2]; |
1341 | | // The central element has an implicit +WIENER_FILT_STEP |
1342 | 0 | fi[3] = -2 * (fi[0] + fi[1] + fi[2]); |
1343 | 0 | } |
1344 | | |
1345 | | static int count_wiener_bits(int wiener_win, WienerInfo *wiener_info, |
1346 | 0 | WienerInfo *ref_wiener_info) { |
1347 | 0 | int bits = 0; |
1348 | 0 | if (wiener_win == WIENER_WIN) |
1349 | 0 | bits += aom_count_primitive_refsubexpfin( |
1350 | 0 | WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1, |
1351 | 0 | WIENER_FILT_TAP0_SUBEXP_K, |
1352 | 0 | ref_wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV, |
1353 | 0 | wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV); |
1354 | 0 | bits += aom_count_primitive_refsubexpfin( |
1355 | 0 | WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1, |
1356 | 0 | WIENER_FILT_TAP1_SUBEXP_K, |
1357 | 0 | ref_wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV, |
1358 | 0 | wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV); |
1359 | 0 | bits += aom_count_primitive_refsubexpfin( |
1360 | 0 | WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1, |
1361 | 0 | WIENER_FILT_TAP2_SUBEXP_K, |
1362 | 0 | ref_wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV, |
1363 | 0 | wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV); |
1364 | 0 | if (wiener_win == WIENER_WIN) |
1365 | 0 | bits += aom_count_primitive_refsubexpfin( |
1366 | 0 | WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1, |
1367 | 0 | WIENER_FILT_TAP0_SUBEXP_K, |
1368 | 0 | ref_wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV, |
1369 | 0 | wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV); |
1370 | 0 | bits += aom_count_primitive_refsubexpfin( |
1371 | 0 | WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1, |
1372 | 0 | WIENER_FILT_TAP1_SUBEXP_K, |
1373 | 0 | ref_wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV, |
1374 | 0 | wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV); |
1375 | 0 | bits += aom_count_primitive_refsubexpfin( |
1376 | 0 | WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1, |
1377 | 0 | WIENER_FILT_TAP2_SUBEXP_K, |
1378 | 0 | ref_wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV, |
1379 | 0 | wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV); |
1380 | 0 | return bits; |
1381 | 0 | } |
1382 | | |
1383 | | #define USE_WIENER_REFINEMENT_SEARCH 1 |
1384 | | static int64_t finer_tile_search_wiener(const RestSearchCtxt *rsc, |
1385 | | const RestorationTileLimits *limits, |
1386 | | const AV1PixelRect *tile, |
1387 | | RestorationUnitInfo *rui, |
1388 | 0 | int wiener_win) { |
1389 | 0 | const int plane_off = (WIENER_WIN - wiener_win) >> 1; |
1390 | 0 | int64_t err = try_restoration_unit(rsc, limits, tile, rui); |
1391 | 0 | #if USE_WIENER_REFINEMENT_SEARCH |
1392 | 0 | int64_t err2; |
1393 | 0 | int tap_min[] = { WIENER_FILT_TAP0_MINV, WIENER_FILT_TAP1_MINV, |
1394 | 0 | WIENER_FILT_TAP2_MINV }; |
1395 | 0 | int tap_max[] = { WIENER_FILT_TAP0_MAXV, WIENER_FILT_TAP1_MAXV, |
1396 | 0 | WIENER_FILT_TAP2_MAXV }; |
1397 | |
|
1398 | 0 | WienerInfo *plane_wiener = &rui->wiener_info; |
1399 | | |
1400 | | // printf("err pre = %"PRId64"\n", err); |
1401 | 0 | const int start_step = 4; |
1402 | 0 | for (int s = start_step; s >= 1; s >>= 1) { |
1403 | 0 | for (int p = plane_off; p < WIENER_HALFWIN; ++p) { |
1404 | 0 | int skip = 0; |
1405 | 0 | do { |
1406 | 0 | if (plane_wiener->hfilter[p] - s >= tap_min[p]) { |
1407 | 0 | plane_wiener->hfilter[p] -= s; |
1408 | 0 | plane_wiener->hfilter[WIENER_WIN - p - 1] -= s; |
1409 | 0 | plane_wiener->hfilter[WIENER_HALFWIN] += 2 * s; |
1410 | 0 | err2 = try_restoration_unit(rsc, limits, tile, rui); |
1411 | 0 | if (err2 > err) { |
1412 | 0 | plane_wiener->hfilter[p] += s; |
1413 | 0 | plane_wiener->hfilter[WIENER_WIN - p - 1] += s; |
1414 | 0 | plane_wiener->hfilter[WIENER_HALFWIN] -= 2 * s; |
1415 | 0 | } else { |
1416 | 0 | err = err2; |
1417 | 0 | skip = 1; |
1418 | | // At the highest step size continue moving in the same direction |
1419 | 0 | if (s == start_step) continue; |
1420 | 0 | } |
1421 | 0 | } |
1422 | 0 | break; |
1423 | 0 | } while (1); |
1424 | 0 | if (skip) break; |
1425 | 0 | do { |
1426 | 0 | if (plane_wiener->hfilter[p] + s <= tap_max[p]) { |
1427 | 0 | plane_wiener->hfilter[p] += s; |
1428 | 0 | plane_wiener->hfilter[WIENER_WIN - p - 1] += s; |
1429 | 0 | plane_wiener->hfilter[WIENER_HALFWIN] -= 2 * s; |
1430 | 0 | err2 = try_restoration_unit(rsc, limits, tile, rui); |
1431 | 0 | if (err2 > err) { |
1432 | 0 | plane_wiener->hfilter[p] -= s; |
1433 | 0 | plane_wiener->hfilter[WIENER_WIN - p - 1] -= s; |
1434 | 0 | plane_wiener->hfilter[WIENER_HALFWIN] += 2 * s; |
1435 | 0 | } else { |
1436 | 0 | err = err2; |
1437 | | // At the highest step size continue moving in the same direction |
1438 | 0 | if (s == start_step) continue; |
1439 | 0 | } |
1440 | 0 | } |
1441 | 0 | break; |
1442 | 0 | } while (1); |
1443 | 0 | } |
1444 | 0 | for (int p = plane_off; p < WIENER_HALFWIN; ++p) { |
1445 | 0 | int skip = 0; |
1446 | 0 | do { |
1447 | 0 | if (plane_wiener->vfilter[p] - s >= tap_min[p]) { |
1448 | 0 | plane_wiener->vfilter[p] -= s; |
1449 | 0 | plane_wiener->vfilter[WIENER_WIN - p - 1] -= s; |
1450 | 0 | plane_wiener->vfilter[WIENER_HALFWIN] += 2 * s; |
1451 | 0 | err2 = try_restoration_unit(rsc, limits, tile, rui); |
1452 | 0 | if (err2 > err) { |
1453 | 0 | plane_wiener->vfilter[p] += s; |
1454 | 0 | plane_wiener->vfilter[WIENER_WIN - p - 1] += s; |
1455 | 0 | plane_wiener->vfilter[WIENER_HALFWIN] -= 2 * s; |
1456 | 0 | } else { |
1457 | 0 | err = err2; |
1458 | 0 | skip = 1; |
1459 | | // At the highest step size continue moving in the same direction |
1460 | 0 | if (s == start_step) continue; |
1461 | 0 | } |
1462 | 0 | } |
1463 | 0 | break; |
1464 | 0 | } while (1); |
1465 | 0 | if (skip) break; |
1466 | 0 | do { |
1467 | 0 | if (plane_wiener->vfilter[p] + s <= tap_max[p]) { |
1468 | 0 | plane_wiener->vfilter[p] += s; |
1469 | 0 | plane_wiener->vfilter[WIENER_WIN - p - 1] += s; |
1470 | 0 | plane_wiener->vfilter[WIENER_HALFWIN] -= 2 * s; |
1471 | 0 | err2 = try_restoration_unit(rsc, limits, tile, rui); |
1472 | 0 | if (err2 > err) { |
1473 | 0 | plane_wiener->vfilter[p] -= s; |
1474 | 0 | plane_wiener->vfilter[WIENER_WIN - p - 1] -= s; |
1475 | 0 | plane_wiener->vfilter[WIENER_HALFWIN] += 2 * s; |
1476 | 0 | } else { |
1477 | 0 | err = err2; |
1478 | | // At the highest step size continue moving in the same direction |
1479 | 0 | if (s == start_step) continue; |
1480 | 0 | } |
1481 | 0 | } |
1482 | 0 | break; |
1483 | 0 | } while (1); |
1484 | 0 | } |
1485 | 0 | } |
1486 | | // printf("err post = %"PRId64"\n", err); |
1487 | 0 | #endif // USE_WIENER_REFINEMENT_SEARCH |
1488 | 0 | return err; |
1489 | 0 | } |
1490 | | |
1491 | | static AOM_INLINE void search_wiener(const RestorationTileLimits *limits, |
1492 | | const AV1PixelRect *tile_rect, |
1493 | | int rest_unit_idx, void *priv, |
1494 | | int32_t *tmpbuf, |
1495 | 0 | RestorationLineBuffers *rlbs) { |
1496 | 0 | (void)tmpbuf; |
1497 | 0 | (void)rlbs; |
1498 | 0 | RestSearchCtxt *rsc = (RestSearchCtxt *)priv; |
1499 | 0 | RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx]; |
1500 | |
|
1501 | 0 | const MACROBLOCK *const x = rsc->x; |
1502 | 0 | const int64_t bits_none = x->mode_costs.wiener_restore_cost[0]; |
1503 | | |
1504 | | // Skip Wiener search for low variance contents |
1505 | 0 | if (rsc->lpf_sf->prune_wiener_based_on_src_var) { |
1506 | 0 | const int scale[3] = { 0, 1, 2 }; |
1507 | | // Obtain the normalized Qscale |
1508 | 0 | const int qs = av1_dc_quant_QTX(rsc->cm->quant_params.base_qindex, 0, |
1509 | 0 | rsc->cm->seq_params->bit_depth) >> |
1510 | 0 | 3; |
1511 | | // Derive threshold as sqr(normalized Qscale) * scale / 16, |
1512 | 0 | const uint64_t thresh = |
1513 | 0 | (qs * qs * scale[rsc->lpf_sf->prune_wiener_based_on_src_var]) >> 4; |
1514 | 0 | const int highbd = rsc->cm->seq_params->use_highbitdepth; |
1515 | 0 | const uint64_t src_var = |
1516 | 0 | var_restoration_unit(limits, rsc->src, rsc->plane, highbd); |
1517 | | // Do not perform Wiener search if source variance is lower than threshold |
1518 | | // or if the reconstruction error is zero |
1519 | 0 | int prune_wiener = (src_var < thresh) || (rusi->sse[RESTORE_NONE] == 0); |
1520 | 0 | if (prune_wiener) { |
1521 | 0 | rsc->bits += bits_none; |
1522 | 0 | rsc->sse += rusi->sse[RESTORE_NONE]; |
1523 | 0 | rusi->best_rtype[RESTORE_WIENER - 1] = RESTORE_NONE; |
1524 | 0 | rusi->sse[RESTORE_WIENER] = INT64_MAX; |
1525 | 0 | if (rsc->lpf_sf->prune_sgr_based_on_wiener == 2) rusi->skip_sgr_eval = 1; |
1526 | 0 | return; |
1527 | 0 | } |
1528 | 0 | } |
1529 | | |
1530 | 0 | const int wiener_win = |
1531 | 0 | (rsc->plane == AOM_PLANE_Y) ? WIENER_WIN : WIENER_WIN_CHROMA; |
1532 | |
|
1533 | 0 | int reduced_wiener_win = wiener_win; |
1534 | 0 | if (rsc->lpf_sf->reduce_wiener_window_size) { |
1535 | 0 | reduced_wiener_win = |
1536 | 0 | (rsc->plane == AOM_PLANE_Y) ? WIENER_WIN_REDUCED : WIENER_WIN_CHROMA; |
1537 | 0 | } |
1538 | |
|
1539 | 0 | int64_t M[WIENER_WIN2]; |
1540 | 0 | int64_t H[WIENER_WIN2 * WIENER_WIN2]; |
1541 | 0 | int32_t vfilter[WIENER_WIN], hfilter[WIENER_WIN]; |
1542 | |
|
1543 | 0 | #if CONFIG_AV1_HIGHBITDEPTH |
1544 | 0 | const AV1_COMMON *const cm = rsc->cm; |
1545 | 0 | if (cm->seq_params->use_highbitdepth) { |
1546 | | // TODO(any) : Add support for use_downsampled_wiener_stats SF in HBD |
1547 | | // functions |
1548 | 0 | av1_compute_stats_highbd(reduced_wiener_win, rsc->dgd_buffer, |
1549 | 0 | rsc->src_buffer, limits->h_start, limits->h_end, |
1550 | 0 | limits->v_start, limits->v_end, rsc->dgd_stride, |
1551 | 0 | rsc->src_stride, M, H, cm->seq_params->bit_depth); |
1552 | 0 | } else { |
1553 | 0 | av1_compute_stats(reduced_wiener_win, rsc->dgd_buffer, rsc->src_buffer, |
1554 | 0 | limits->h_start, limits->h_end, limits->v_start, |
1555 | 0 | limits->v_end, rsc->dgd_stride, rsc->src_stride, M, H, |
1556 | 0 | rsc->lpf_sf->use_downsampled_wiener_stats); |
1557 | 0 | } |
1558 | | #else |
1559 | | av1_compute_stats(reduced_wiener_win, rsc->dgd_buffer, rsc->src_buffer, |
1560 | | limits->h_start, limits->h_end, limits->v_start, |
1561 | | limits->v_end, rsc->dgd_stride, rsc->src_stride, M, H, |
1562 | | rsc->lpf_sf->use_downsampled_wiener_stats); |
1563 | | #endif |
1564 | |
|
1565 | 0 | wiener_decompose_sep_sym(reduced_wiener_win, M, H, vfilter, hfilter); |
1566 | |
|
1567 | 0 | RestorationUnitInfo rui; |
1568 | 0 | memset(&rui, 0, sizeof(rui)); |
1569 | 0 | rui.restoration_type = RESTORE_WIENER; |
1570 | 0 | finalize_sym_filter(reduced_wiener_win, vfilter, rui.wiener_info.vfilter); |
1571 | 0 | finalize_sym_filter(reduced_wiener_win, hfilter, rui.wiener_info.hfilter); |
1572 | | |
1573 | | // Filter score computes the value of the function x'*A*x - x'*b for the |
1574 | | // learned filter and compares it against identity filer. If there is no |
1575 | | // reduction in the function, the filter is reverted back to identity |
1576 | 0 | if (compute_score(reduced_wiener_win, M, H, rui.wiener_info.vfilter, |
1577 | 0 | rui.wiener_info.hfilter) > 0) { |
1578 | 0 | rsc->bits += bits_none; |
1579 | 0 | rsc->sse += rusi->sse[RESTORE_NONE]; |
1580 | 0 | rusi->best_rtype[RESTORE_WIENER - 1] = RESTORE_NONE; |
1581 | 0 | rusi->sse[RESTORE_WIENER] = INT64_MAX; |
1582 | 0 | if (rsc->lpf_sf->prune_sgr_based_on_wiener == 2) rusi->skip_sgr_eval = 1; |
1583 | 0 | return; |
1584 | 0 | } |
1585 | | |
1586 | 0 | rusi->sse[RESTORE_WIENER] = finer_tile_search_wiener( |
1587 | 0 | rsc, limits, tile_rect, &rui, reduced_wiener_win); |
1588 | 0 | rusi->wiener = rui.wiener_info; |
1589 | |
|
1590 | 0 | if (reduced_wiener_win != WIENER_WIN) { |
1591 | 0 | assert(rui.wiener_info.vfilter[0] == 0 && |
1592 | 0 | rui.wiener_info.vfilter[WIENER_WIN - 1] == 0); |
1593 | 0 | assert(rui.wiener_info.hfilter[0] == 0 && |
1594 | 0 | rui.wiener_info.hfilter[WIENER_WIN - 1] == 0); |
1595 | 0 | } |
1596 | |
|
1597 | 0 | const int64_t bits_wiener = |
1598 | 0 | x->mode_costs.wiener_restore_cost[1] + |
1599 | 0 | (count_wiener_bits(wiener_win, &rusi->wiener, &rsc->wiener) |
1600 | 0 | << AV1_PROB_COST_SHIFT); |
1601 | |
|
1602 | 0 | double cost_none = RDCOST_DBL_WITH_NATIVE_BD_DIST( |
1603 | 0 | x->rdmult, bits_none >> 4, rusi->sse[RESTORE_NONE], |
1604 | 0 | rsc->cm->seq_params->bit_depth); |
1605 | 0 | double cost_wiener = RDCOST_DBL_WITH_NATIVE_BD_DIST( |
1606 | 0 | x->rdmult, bits_wiener >> 4, rusi->sse[RESTORE_WIENER], |
1607 | 0 | rsc->cm->seq_params->bit_depth); |
1608 | |
|
1609 | 0 | RestorationType rtype = |
1610 | 0 | (cost_wiener < cost_none) ? RESTORE_WIENER : RESTORE_NONE; |
1611 | 0 | rusi->best_rtype[RESTORE_WIENER - 1] = rtype; |
1612 | | |
1613 | | // Set 'skip_sgr_eval' based on rdcost ratio of RESTORE_WIENER and |
1614 | | // RESTORE_NONE or based on best_rtype |
1615 | 0 | if (rsc->lpf_sf->prune_sgr_based_on_wiener == 1) { |
1616 | 0 | rusi->skip_sgr_eval = cost_wiener > (1.01 * cost_none); |
1617 | 0 | } else if (rsc->lpf_sf->prune_sgr_based_on_wiener == 2) { |
1618 | 0 | rusi->skip_sgr_eval = rusi->best_rtype[RESTORE_WIENER - 1] == RESTORE_NONE; |
1619 | 0 | } |
1620 | |
|
1621 | 0 | rsc->sse += rusi->sse[rtype]; |
1622 | 0 | rsc->bits += (cost_wiener < cost_none) ? bits_wiener : bits_none; |
1623 | 0 | if (cost_wiener < cost_none) rsc->wiener = rusi->wiener; |
1624 | 0 | } |
1625 | | |
1626 | | static AOM_INLINE void search_norestore(const RestorationTileLimits *limits, |
1627 | | const AV1PixelRect *tile_rect, |
1628 | | int rest_unit_idx, void *priv, |
1629 | | int32_t *tmpbuf, |
1630 | 0 | RestorationLineBuffers *rlbs) { |
1631 | 0 | (void)tile_rect; |
1632 | 0 | (void)tmpbuf; |
1633 | 0 | (void)rlbs; |
1634 | |
|
1635 | 0 | RestSearchCtxt *rsc = (RestSearchCtxt *)priv; |
1636 | 0 | RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx]; |
1637 | |
|
1638 | 0 | const int highbd = rsc->cm->seq_params->use_highbitdepth; |
1639 | 0 | rusi->sse[RESTORE_NONE] = sse_restoration_unit( |
1640 | 0 | limits, rsc->src, &rsc->cm->cur_frame->buf, rsc->plane, highbd); |
1641 | |
|
1642 | 0 | rsc->sse += rusi->sse[RESTORE_NONE]; |
1643 | 0 | } |
1644 | | |
1645 | | static AOM_INLINE void search_switchable(const RestorationTileLimits *limits, |
1646 | | const AV1PixelRect *tile_rect, |
1647 | | int rest_unit_idx, void *priv, |
1648 | | int32_t *tmpbuf, |
1649 | 0 | RestorationLineBuffers *rlbs) { |
1650 | 0 | (void)limits; |
1651 | 0 | (void)tile_rect; |
1652 | 0 | (void)tmpbuf; |
1653 | 0 | (void)rlbs; |
1654 | 0 | RestSearchCtxt *rsc = (RestSearchCtxt *)priv; |
1655 | 0 | RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx]; |
1656 | |
|
1657 | 0 | const MACROBLOCK *const x = rsc->x; |
1658 | |
|
1659 | 0 | const int wiener_win = |
1660 | 0 | (rsc->plane == AOM_PLANE_Y) ? WIENER_WIN : WIENER_WIN_CHROMA; |
1661 | |
|
1662 | 0 | double best_cost = 0; |
1663 | 0 | int64_t best_bits = 0; |
1664 | 0 | RestorationType best_rtype = RESTORE_NONE; |
1665 | |
|
1666 | 0 | for (RestorationType r = 0; r < RESTORE_SWITCHABLE_TYPES; ++r) { |
1667 | | // Check for the condition that wiener or sgrproj search could not |
1668 | | // find a solution or the solution was worse than RESTORE_NONE. |
1669 | | // In either case the best_rtype will be set as RESTORE_NONE. These |
1670 | | // should be skipped from the test below. |
1671 | 0 | if (r > RESTORE_NONE) { |
1672 | 0 | if (rusi->best_rtype[r - 1] == RESTORE_NONE) continue; |
1673 | 0 | } |
1674 | | |
1675 | 0 | const int64_t sse = rusi->sse[r]; |
1676 | 0 | int64_t coeff_pcost = 0; |
1677 | 0 | switch (r) { |
1678 | 0 | case RESTORE_NONE: coeff_pcost = 0; break; |
1679 | 0 | case RESTORE_WIENER: |
1680 | 0 | coeff_pcost = |
1681 | 0 | count_wiener_bits(wiener_win, &rusi->wiener, &rsc->wiener); |
1682 | 0 | break; |
1683 | 0 | case RESTORE_SGRPROJ: |
1684 | 0 | coeff_pcost = count_sgrproj_bits(&rusi->sgrproj, &rsc->sgrproj); |
1685 | 0 | break; |
1686 | 0 | default: assert(0); break; |
1687 | 0 | } |
1688 | 0 | const int64_t coeff_bits = coeff_pcost << AV1_PROB_COST_SHIFT; |
1689 | 0 | const int64_t bits = x->mode_costs.switchable_restore_cost[r] + coeff_bits; |
1690 | 0 | double cost = RDCOST_DBL_WITH_NATIVE_BD_DIST( |
1691 | 0 | x->rdmult, bits >> 4, sse, rsc->cm->seq_params->bit_depth); |
1692 | 0 | if (r == RESTORE_SGRPROJ && rusi->sgrproj.ep < 10) |
1693 | 0 | cost *= (1 + DUAL_SGR_PENALTY_MULT * rsc->lpf_sf->dual_sgr_penalty_level); |
1694 | 0 | if (r == 0 || cost < best_cost) { |
1695 | 0 | best_cost = cost; |
1696 | 0 | best_bits = bits; |
1697 | 0 | best_rtype = r; |
1698 | 0 | } |
1699 | 0 | } |
1700 | | |
1701 | 0 | rusi->best_rtype[RESTORE_SWITCHABLE - 1] = best_rtype; |
1702 | |
|
1703 | 0 | rsc->sse += rusi->sse[best_rtype]; |
1704 | 0 | rsc->bits += best_bits; |
1705 | 0 | if (best_rtype == RESTORE_WIENER) rsc->wiener = rusi->wiener; |
1706 | 0 | if (best_rtype == RESTORE_SGRPROJ) rsc->sgrproj = rusi->sgrproj; |
1707 | 0 | } |
1708 | | |
1709 | | static AOM_INLINE void copy_unit_info(RestorationType frame_rtype, |
1710 | | const RestUnitSearchInfo *rusi, |
1711 | 0 | RestorationUnitInfo *rui) { |
1712 | 0 | assert(frame_rtype > 0); |
1713 | 0 | rui->restoration_type = rusi->best_rtype[frame_rtype - 1]; |
1714 | 0 | if (rui->restoration_type == RESTORE_WIENER) |
1715 | 0 | rui->wiener_info = rusi->wiener; |
1716 | 0 | else |
1717 | 0 | rui->sgrproj_info = rusi->sgrproj; |
1718 | 0 | } |
1719 | | |
1720 | 0 | static double search_rest_type(RestSearchCtxt *rsc, RestorationType rtype) { |
1721 | 0 | static const rest_unit_visitor_t funs[RESTORE_TYPES] = { |
1722 | 0 | search_norestore, search_wiener, search_sgrproj, search_switchable |
1723 | 0 | }; |
1724 | |
|
1725 | 0 | reset_rsc(rsc); |
1726 | 0 | rsc_on_tile(rsc); |
1727 | |
|
1728 | 0 | av1_foreach_rest_unit_in_plane(rsc->cm, rsc->plane, funs[rtype], rsc, |
1729 | 0 | &rsc->tile_rect, rsc->cm->rst_tmpbuf, NULL); |
1730 | 0 | return RDCOST_DBL_WITH_NATIVE_BD_DIST( |
1731 | 0 | rsc->x->rdmult, rsc->bits >> 4, rsc->sse, rsc->cm->seq_params->bit_depth); |
1732 | 0 | } |
1733 | | |
1734 | 0 | static int rest_tiles_in_plane(const AV1_COMMON *cm, int plane) { |
1735 | 0 | const RestorationInfo *rsi = &cm->rst_info[plane]; |
1736 | 0 | return rsi->units_per_tile; |
1737 | 0 | } |
1738 | | |
1739 | 0 | void av1_pick_filter_restoration(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi) { |
1740 | 0 | AV1_COMMON *const cm = &cpi->common; |
1741 | 0 | MACROBLOCK *const x = &cpi->td.mb; |
1742 | 0 | const SequenceHeader *const seq_params = cm->seq_params; |
1743 | 0 | const int num_planes = av1_num_planes(cm); |
1744 | 0 | assert(!cm->features.all_lossless); |
1745 | |
|
1746 | 0 | av1_fill_lr_rates(&x->mode_costs, x->e_mbd.tile_ctx); |
1747 | |
|
1748 | 0 | int ntiles[2]; |
1749 | 0 | for (int is_uv = 0; is_uv < 2; ++is_uv) |
1750 | 0 | ntiles[is_uv] = rest_tiles_in_plane(cm, is_uv); |
1751 | |
|
1752 | 0 | assert(ntiles[1] <= ntiles[0]); |
1753 | 0 | RestUnitSearchInfo *rusi = |
1754 | 0 | (RestUnitSearchInfo *)aom_memalign(16, sizeof(*rusi) * ntiles[0]); |
1755 | | |
1756 | | // If the restoration unit dimensions are not multiples of |
1757 | | // rsi->restoration_unit_size then some elements of the rusi array may be |
1758 | | // left uninitialised when we reach copy_unit_info(...). This is not a |
1759 | | // problem, as these elements are ignored later, but in order to quiet |
1760 | | // Valgrind's warnings we initialise the array below. |
1761 | 0 | memset(rusi, 0, sizeof(*rusi) * ntiles[0]); |
1762 | 0 | x->rdmult = cpi->rd.RDMULT; |
1763 | | |
1764 | | // Allocate the frame buffer trial_frame_rst, which is used to temporarily |
1765 | | // store the loop restored frame. |
1766 | 0 | if (aom_realloc_frame_buffer( |
1767 | 0 | &cpi->trial_frame_rst, cm->superres_upscaled_width, |
1768 | 0 | cm->superres_upscaled_height, seq_params->subsampling_x, |
1769 | 0 | seq_params->subsampling_y, seq_params->use_highbitdepth, |
1770 | 0 | AOM_RESTORATION_FRAME_BORDER, cm->features.byte_alignment, NULL, NULL, |
1771 | 0 | NULL, 0)) |
1772 | 0 | aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR, |
1773 | 0 | "Failed to allocate trial restored frame buffer"); |
1774 | |
|
1775 | 0 | RestSearchCtxt rsc; |
1776 | 0 | const int plane_start = AOM_PLANE_Y; |
1777 | 0 | const int plane_end = num_planes > 1 ? AOM_PLANE_V : AOM_PLANE_Y; |
1778 | 0 | for (int plane = plane_start; plane <= plane_end; ++plane) { |
1779 | 0 | init_rsc(src, &cpi->common, x, &cpi->sf.lpf_sf, plane, rusi, |
1780 | 0 | &cpi->trial_frame_rst, &rsc); |
1781 | |
|
1782 | 0 | const int plane_ntiles = ntiles[plane > 0]; |
1783 | 0 | const RestorationType num_rtypes = |
1784 | 0 | (plane_ntiles > 1) ? RESTORE_TYPES : RESTORE_SWITCHABLE_TYPES; |
1785 | |
|
1786 | 0 | double best_cost = 0; |
1787 | 0 | RestorationType best_rtype = RESTORE_NONE; |
1788 | |
|
1789 | 0 | const int highbd = rsc.cm->seq_params->use_highbitdepth; |
1790 | 0 | if ((plane && !cpi->sf.lpf_sf.disable_loop_restoration_chroma) || |
1791 | 0 | (!plane && !cpi->sf.lpf_sf.disable_loop_restoration_luma)) { |
1792 | 0 | av1_extend_frame(rsc.dgd_buffer, rsc.plane_width, rsc.plane_height, |
1793 | 0 | rsc.dgd_stride, RESTORATION_BORDER, RESTORATION_BORDER, |
1794 | 0 | highbd); |
1795 | |
|
1796 | 0 | for (RestorationType r = 0; r < num_rtypes; ++r) { |
1797 | 0 | if ((force_restore_type != RESTORE_TYPES) && (r != RESTORE_NONE) && |
1798 | 0 | (r != force_restore_type)) |
1799 | 0 | continue; |
1800 | | |
1801 | 0 | double cost = search_rest_type(&rsc, r); |
1802 | |
|
1803 | 0 | if (r == 0 || cost < best_cost) { |
1804 | 0 | best_cost = cost; |
1805 | 0 | best_rtype = r; |
1806 | 0 | } |
1807 | 0 | } |
1808 | 0 | } |
1809 | |
|
1810 | 0 | cm->rst_info[plane].frame_restoration_type = best_rtype; |
1811 | 0 | if (force_restore_type != RESTORE_TYPES) |
1812 | 0 | assert(best_rtype == force_restore_type || best_rtype == RESTORE_NONE); |
1813 | |
|
1814 | 0 | if (best_rtype != RESTORE_NONE) { |
1815 | 0 | for (int u = 0; u < plane_ntiles; ++u) { |
1816 | 0 | copy_unit_info(best_rtype, &rusi[u], &cm->rst_info[plane].unit_info[u]); |
1817 | 0 | } |
1818 | 0 | } |
1819 | 0 | } |
1820 | |
|
1821 | 0 | aom_free(rusi); |
1822 | 0 | } |