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