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