LCOV - code coverage report
Current view: top level - Codec - EbRestorationPick.c (source / functions) Hit Total Coverage
Test: coverage.info Lines: 612 1095 55.9 %
Date: 2019-11-25 17:38:06 Functions: 33 46 71.7 %

          Line data    Source code
       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 <math.h>
      14             : #include "EbDefinitions.h"
      15             : #include "EbPictureBufferDesc.h"
      16             : #include "EbPsnr.h"
      17             : #include "EbPictureControlSet.h"
      18             : #include "aom_dsp_rtcd.h"
      19             : #include "EbRestoration.h"
      20             : #include "EbRestorationPick.h"
      21             : 
      22             : #include "EbRestProcess.h"
      23             : 
      24             : void av1_foreach_rest_unit_in_frame_seg(Av1Common *cm, int32_t plane,
      25             :     RestTileStartVisitor on_tile,
      26             :     RestUnitVisitor on_rest_unit,
      27             :     void *priv,
      28             :     PictureControlSet   *picture_control_set_ptr,
      29             :     uint32_t segment_index);
      30             : 
      31             : void eb_av1_selfguided_restoration_c(const uint8_t *dgd8, int32_t width, int32_t height,
      32             :     int32_t dgd_stride, int32_t *flt0, int32_t *flt1,
      33             :     int32_t flt_stride, int32_t sgr_params_idx,
      34             :     int32_t bit_depth, int32_t highbd);
      35             : void av1_foreach_rest_unit_in_frame(Av1Common *cm, int32_t plane,
      36             :     RestTileStartVisitor on_tile,
      37             :     RestUnitVisitor on_rest_unit,
      38             :     void *priv);
      39             : 
      40             : // When set to RESTORE_WIENER or RESTORE_SGRPROJ only those are allowed.
      41             : // When set to RESTORE_TYPES we allow switchable.
      42             : //static const RestorationType force_restore_type = RESTORE_TYPES;
      43             : 
      44             : // Number of Wiener iterations
      45             : #define NUM_WIENER_ITERS 5
      46             : // Working precision for Wiener filter coefficients
      47             : #define WIENER_TAP_SCALE_FACTOR ((int64_t)1 << 16)
      48             : 
      49             : typedef int64_t(*sse_part_extractor_type)(const Yv12BufferConfig *a,
      50             :     const Yv12BufferConfig *b,
      51             :     int32_t hstart, int32_t width, int32_t vstart,
      52             :     int32_t height);
      53             : 
      54             : #define NUM_EXTRACTORS (3 * (1 + 1))
      55             : 
      56             : static const sse_part_extractor_type sse_part_extractors[NUM_EXTRACTORS] = {
      57             :   eb_aom_get_y_sse_part,        eb_aom_get_u_sse_part,
      58             :   eb_aom_get_v_sse_part,        eb_aom_highbd_get_y_sse_part,
      59             :   eb_aom_highbd_get_u_sse_part, eb_aom_highbd_get_v_sse_part,
      60             : };
      61        9645 : static int64_t sse_restoration_unit(const RestorationTileLimits *limits,
      62             :     const Yv12BufferConfig *src,
      63             :     const Yv12BufferConfig *dst, int32_t plane,
      64             :     int32_t highbd) {
      65       19290 :     return sse_part_extractors[3 * highbd + plane](
      66        9645 :         src, dst, limits->h_start, limits->h_end - limits->h_start,
      67        9645 :         limits->v_start, limits->v_end - limits->v_start);
      68             : }
      69             : 
      70             : typedef struct {
      71             :     const Yv12BufferConfig *src;
      72             :     Yv12BufferConfig *dst;
      73             : 
      74             :     Av1Common *cm;
      75             :     const Macroblock *x;
      76             :     int32_t plane;
      77             :     int32_t plane_width;
      78             :     int32_t plane_height;
      79             :     RestUnitSearchInfo *rusi;
      80             :     RestUnitSearchInfo *rusi_pic;
      81             :     uint32_t  pic_num;
      82             :     Yv12BufferConfig * org_frame_to_show;
      83             :     int32_t *tmpbuf;
      84             : 
      85             :     uint8_t *dgd_buffer;
      86             :     int32_t dgd_stride;
      87             :     const uint8_t *src_buffer;
      88             :     int32_t src_stride;
      89             : 
      90             :     // sse and bits are initialised by reset_rsc in search_rest_type
      91             :     int64_t sse;
      92             :     int64_t bits;
      93             :     int32_t tile_y0, tile_stripe0;
      94             : 
      95             :     // sgrproj and wiener are initialised by rsc_on_tile when starting the first
      96             :     // tile in the frame.
      97             :     SgrprojInfo sgrproj;
      98             :     WienerInfo wiener;
      99             : } RestSearchCtxt;
     100             : 
     101        1140 : static void rsc_on_tile(int32_t tile_row, int32_t tile_col, void *priv) {
     102             :     (void)tile_col;
     103             : 
     104        1140 :     RestSearchCtxt *rsc = (RestSearchCtxt *)priv;
     105        1140 :     set_default_sgrproj(&rsc->sgrproj);
     106        1140 :     set_default_wiener(&rsc->wiener);
     107             : 
     108        1140 :     rsc->tile_stripe0 =
     109        1140 :         (tile_row == 0) ? 0 : rsc->cm->rst_end_stripe[tile_row - 1];
     110        1140 : }
     111             : 
     112         600 : static void reset_rsc(RestSearchCtxt *rsc) {
     113         600 :     rsc->sse = 0;
     114         600 :     rsc->bits = 0;
     115         600 : }
     116         180 : static void init_rsc_seg(
     117             :     Yv12BufferConfig *org_fts,
     118             :     const Yv12BufferConfig *src, Av1Common *cm,
     119             :     const Macroblock *x, int32_t plane, RestUnitSearchInfo *rusi,
     120             :     Yv12BufferConfig *dst, RestSearchCtxt *rsc) {
     121         180 :     rsc->src = src;
     122         180 :     rsc->dst = dst;
     123         180 :     rsc->cm = cm;
     124         180 :     rsc->x = x;
     125         180 :     rsc->plane = plane;
     126         180 :     rsc->rusi = rusi;
     127             : 
     128         180 :     rsc->org_frame_to_show = org_fts;
     129             : 
     130         180 :     const Yv12BufferConfig *dgd = org_fts;
     131         180 :     const int32_t is_uv = plane != AOM_PLANE_Y;
     132         180 :     rsc->plane_width = src->crop_widths[is_uv];
     133         180 :     rsc->plane_height = src->crop_heights[is_uv];
     134         180 :     rsc->src_buffer = src->buffers[plane];
     135         180 :     rsc->src_stride = src->strides[is_uv];
     136         180 :     rsc->dgd_buffer = dgd->buffers[plane];
     137         180 :     rsc->dgd_stride = dgd->strides[is_uv];
     138             :     assert(src->crop_widths[is_uv] == dgd->crop_widths[is_uv]);
     139             :     assert(src->crop_heights[is_uv] == dgd->crop_heights[is_uv]);
     140         180 : }
     141             : 
     142           0 : static void init_rsc(const Yv12BufferConfig *src, Av1Common *cm,
     143             :     const Macroblock *x, int32_t plane, RestUnitSearchInfo *rusi,
     144             :     Yv12BufferConfig *dst, RestSearchCtxt *rsc) {
     145           0 :     rsc->src = src;
     146           0 :     rsc->dst = dst;
     147           0 :     rsc->cm = cm;
     148           0 :     rsc->x = x;
     149           0 :     rsc->plane = plane;
     150           0 :     rsc->rusi = rusi;
     151             : 
     152           0 :     const Yv12BufferConfig *dgd = cm->frame_to_show;
     153           0 :     const int32_t is_uv = plane != AOM_PLANE_Y;
     154           0 :     rsc->plane_width = src->crop_widths[is_uv];
     155           0 :     rsc->plane_height = src->crop_heights[is_uv];
     156           0 :     rsc->src_buffer = src->buffers[plane];
     157           0 :     rsc->src_stride = src->strides[is_uv];
     158           0 :     rsc->dgd_buffer = dgd->buffers[plane];
     159           0 :     rsc->dgd_stride = dgd->strides[is_uv];
     160             :     assert(src->crop_widths[is_uv] == dgd->crop_widths[is_uv]);
     161             :     assert(src->crop_heights[is_uv] == dgd->crop_heights[is_uv]);
     162           0 : }
     163             : 
     164           0 : static int64_t try_restoration_unit(const RestSearchCtxt *rsc,
     165             :     const RestorationTileLimits *limits,
     166             :     const AV1PixelRect *tile_rect,
     167             :     const RestorationUnitInfo *rui) {
     168           0 :     const Av1Common *const cm = rsc->cm;
     169           0 :     const int32_t plane = rsc->plane;
     170           0 :     const int32_t is_uv = plane > 0;
     171           0 :     const RestorationInfo *rsi = &cm->rst_info[plane];
     172             :     RestorationLineBuffers rlbs;
     173           0 :     const int32_t bit_depth = cm->bit_depth;
     174           0 :     const int32_t highbd = cm->use_highbitdepth;
     175             : 
     176           0 :     const Yv12BufferConfig *fts = cm->frame_to_show;
     177             :     // TODO(yunqing): For now, only use optimized LR filter in decoder. Can be
     178             :     // also used in encoder.
     179           0 :     const int32_t optimized_lr = 0;
     180             : 
     181           0 :     eb_av1_loop_restoration_filter_unit(
     182             :         1,
     183             :         limits, rui, &rsi->boundaries, &rlbs, tile_rect, rsc->tile_stripe0,
     184           0 :         is_uv && cm->subsampling_x, is_uv && cm->subsampling_y, highbd, bit_depth,
     185           0 :         fts->buffers[plane], fts->strides[is_uv], rsc->dst->buffers[plane],
     186           0 :         rsc->dst->strides[is_uv], cm->rst_tmpbuf, optimized_lr);
     187             : 
     188           0 :     return sse_restoration_unit(limits, rsc->src, rsc->dst, plane, highbd);
     189             : }
     190        9345 : static int64_t try_restoration_unit_seg(const RestSearchCtxt *rsc,
     191             :     const RestorationTileLimits *limits,
     192             :     const AV1PixelRect *tile_rect,
     193             :     const RestorationUnitInfo *rui) {
     194        9345 :     const Av1Common *const cm = rsc->cm;
     195        9345 :     const int32_t plane = rsc->plane;
     196        9345 :     const int32_t is_uv = plane > 0;
     197        9345 :     const RestorationInfo *rsi = &cm->rst_info[plane];
     198             :     RestorationLineBuffers rlbs;
     199        9345 :     const int32_t bit_depth = cm->bit_depth;
     200        9345 :     const int32_t highbd = cm->use_highbitdepth;
     201             : 
     202        9345 :     const Yv12BufferConfig *fts = rsc->org_frame_to_show;
     203             : 
     204             :     // TODO(yunqing): For now, only use optimized LR filter in decoder. Can be
     205             :     // also used in encoder.
     206        9345 :     const int32_t optimized_lr = 0;
     207             : 
     208       12326 :     eb_av1_loop_restoration_filter_unit(
     209             :         1,
     210             :         limits, rui, &rsi->boundaries, &rlbs, tile_rect, rsc->tile_stripe0,
     211        5962 :         is_uv && cm->subsampling_x, is_uv && cm->subsampling_y, highbd, bit_depth,
     212        9345 :         fts->buffers[plane], fts->strides[is_uv], rsc->dst->buffers[plane],
     213        9345 :         rsc->dst->strides[is_uv], rsc->tmpbuf, optimized_lr);
     214             : 
     215        9345 :     return sse_restoration_unit(limits, rsc->src, rsc->dst, plane, highbd);
     216             : }
     217             : 
     218           0 : int64_t eb_av1_lowbd_pixel_proj_error_c(const uint8_t *src8, int32_t width, int32_t height,
     219             :     int32_t src_stride, const uint8_t *dat8,
     220             :     int32_t dat_stride, int32_t *flt0,
     221             :     int32_t flt0_stride, int32_t *flt1,
     222             :     int32_t flt1_stride, int32_t xq[2],
     223             :     const SgrParamsType *params) {
     224             :     int32_t i, j;
     225           0 :     const uint8_t *src = src8;
     226           0 :     const uint8_t *dat = dat8;
     227           0 :     int64_t err = 0;
     228           0 :     if (params->r[0] > 0 && params->r[1] > 0) {
     229           0 :         for (i = 0; i < height; ++i) {
     230           0 :             for (j = 0; j < width; ++j) {
     231             :                 assert(flt1[j] < (1 << 15) && flt1[j] > -(1 << 15));
     232             :                 assert(flt0[j] < (1 << 15) && flt0[j] > -(1 << 15));
     233           0 :                 const int32_t u = (int32_t)(dat[j] << SGRPROJ_RST_BITS);
     234           0 :                 int32_t v = u << SGRPROJ_PRJ_BITS;
     235           0 :                 v += xq[0] * (flt0[j] - u) + xq[1] * (flt1[j] - u);
     236           0 :                 const int32_t e =
     237           0 :                     ROUND_POWER_OF_TWO(v, SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) - src[j];
     238           0 :                 err += e * e;
     239             :             }
     240           0 :             dat += dat_stride;
     241           0 :             src += src_stride;
     242           0 :             flt0 += flt0_stride;
     243           0 :             flt1 += flt1_stride;
     244             :         }
     245             :     }
     246           0 :     else if (params->r[0] > 0) {
     247           0 :         for (i = 0; i < height; ++i) {
     248           0 :             for (j = 0; j < width; ++j) {
     249             :                 assert(flt0[j] < (1 << 15) && flt0[j] > -(1 << 15));
     250           0 :                 const int32_t u = (int32_t)(dat[j] << SGRPROJ_RST_BITS);
     251           0 :                 int32_t v = u << SGRPROJ_PRJ_BITS;
     252           0 :                 v += xq[0] * (flt0[j] - u);
     253           0 :                 const int32_t e =
     254           0 :                     ROUND_POWER_OF_TWO(v, SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) - src[j];
     255           0 :                 err += e * e;
     256             :             }
     257           0 :             dat += dat_stride;
     258           0 :             src += src_stride;
     259           0 :             flt0 += flt0_stride;
     260             :         }
     261             :     }
     262           0 :     else if (params->r[1] > 0) {
     263           0 :         for (i = 0; i < height; ++i) {
     264           0 :             for (j = 0; j < width; ++j) {
     265             :                 assert(flt1[j] < (1 << 15) && flt1[j] > -(1 << 15));
     266           0 :                 const int32_t u = (int32_t)(dat[j] << SGRPROJ_RST_BITS);
     267           0 :                 int32_t v = u << SGRPROJ_PRJ_BITS;
     268           0 :                 v += xq[1] * (flt1[j] - u);
     269           0 :                 const int32_t e =
     270           0 :                     ROUND_POWER_OF_TWO(v, SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) - src[j];
     271           0 :                 err += e * e;
     272             :             }
     273           0 :             dat += dat_stride;
     274           0 :             src += src_stride;
     275           0 :             flt1 += flt1_stride;
     276             :         }
     277             :     }
     278             :     else {
     279           0 :         for (i = 0; i < height; ++i) {
     280           0 :             for (j = 0; j < width; ++j) {
     281           0 :                 const int32_t e = (int32_t)(dat[j]) - src[j];
     282           0 :                 err += e * e;
     283             :             }
     284           0 :             dat += dat_stride;
     285           0 :             src += src_stride;
     286             :         }
     287             :     }
     288             : 
     289           0 :     return err;
     290             : }
     291             : 
     292           0 : int64_t eb_av1_highbd_pixel_proj_error_c(const uint8_t *src8, int32_t width,
     293             :     int32_t height, int32_t src_stride,
     294             :     const uint8_t *dat8, int32_t dat_stride,
     295             :     int32_t *flt0, int32_t flt0_stride,
     296             :     int32_t *flt1, int32_t flt1_stride, int32_t xq[2],
     297             :     const SgrParamsType *params) {
     298           0 :     const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
     299           0 :     const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
     300             :     int32_t i, j;
     301           0 :     int64_t err = 0;
     302           0 :     const int32_t half = 1 << (SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS - 1);
     303           0 :     if (params->r[0] > 0 && params->r[1] > 0) {
     304           0 :         int32_t xq0 = xq[0];
     305           0 :         int32_t xq1 = xq[1];
     306           0 :         for (i = 0; i < height; ++i) {
     307           0 :             for (j = 0; j < width; ++j) {
     308           0 :                 const int32_t d = dat[j];
     309           0 :                 const int32_t s = src[j];
     310           0 :                 const int32_t u = (int32_t)(d << SGRPROJ_RST_BITS);
     311           0 :                 int32_t v0 = flt0[j] - u;
     312           0 :                 int32_t v1 = flt1[j] - u;
     313           0 :                 int32_t v = half;
     314           0 :                 v += xq0 * v0;
     315           0 :                 v += xq1 * v1;
     316           0 :                 const int32_t e = (v >> (SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS)) + d - s;
     317           0 :                 err += e * e;
     318             :             }
     319           0 :             dat += dat_stride;
     320           0 :             flt0 += flt0_stride;
     321           0 :             flt1 += flt1_stride;
     322           0 :             src += src_stride;
     323             :         }
     324             :     }
     325           0 :     else if (params->r[0] > 0 || params->r[1] > 0) {
     326             :         int32_t exq;
     327             :         int32_t *flt;
     328             :         int32_t flt_stride;
     329           0 :         if (params->r[0] > 0) {
     330           0 :             exq = xq[0];
     331           0 :             flt = flt0;
     332           0 :             flt_stride = flt0_stride;
     333             :         }
     334             :         else {
     335           0 :             exq = xq[1];
     336           0 :             flt = flt1;
     337           0 :             flt_stride = flt1_stride;
     338             :         }
     339           0 :         for (i = 0; i < height; ++i) {
     340           0 :             for (j = 0; j < width; ++j) {
     341           0 :                 const int32_t d = dat[j];
     342           0 :                 const int32_t s = src[j];
     343           0 :                 const int32_t u = (int32_t)(d << SGRPROJ_RST_BITS);
     344           0 :                 int32_t v = half;
     345           0 :                 v += exq * (flt[j] - u);
     346           0 :                 const int32_t e = (v >> (SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS)) + d - s;
     347           0 :                 err += e * e;
     348             :             }
     349           0 :             dat += dat_stride;
     350           0 :             flt += flt_stride;
     351           0 :             src += src_stride;
     352             :         }
     353             :     }
     354             :     else {
     355           0 :         for (i = 0; i < height; ++i) {
     356           0 :             for (j = 0; j < width; ++j) {
     357           0 :                 const int32_t d = dat[j];
     358           0 :                 const int32_t s = src[j];
     359           0 :                 const int32_t e = d - s;
     360           0 :                 err += e * e;
     361             :             }
     362           0 :             dat += dat_stride;
     363           0 :             src += src_stride;
     364             :         }
     365             :     }
     366           0 :     return err;
     367             : }
     368             : 
     369       36589 : static int64_t get_pixel_proj_error(const uint8_t *src8, int32_t width, int32_t height,
     370             :     int32_t src_stride, const uint8_t *dat8,
     371             :     int32_t dat_stride, int32_t use_highbitdepth,
     372             :     int32_t *flt0, int32_t flt0_stride,
     373             :     int32_t *flt1, int32_t flt1_stride, int32_t *xqd,
     374             :     const SgrParamsType *params) {
     375             :     int32_t xq[2];
     376       36589 :     eb_decode_xq(xqd, xq, params);
     377       36589 :     if (!use_highbitdepth) {
     378       36589 :         return eb_av1_lowbd_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           0 :         return eb_av1_highbd_pixel_proj_error(src8, width, height, src_stride, dat8,
     384             :             dat_stride, flt0, flt0_stride, flt1,
     385             :             flt1_stride, xq, params);
     386             :     }
     387             : }
     388             : 
     389             : #define USE_SGRPROJ_REFINEMENT_SEARCH 1
     390        4800 : static int64_t finer_search_pixel_proj_error(
     391             :     const uint8_t *src8, int32_t width, int32_t height, int32_t src_stride,
     392             :     const uint8_t *dat8, int32_t dat_stride, int32_t use_highbitdepth, int32_t *flt0,
     393             :     int32_t flt0_stride, int32_t *flt1, int32_t flt1_stride, int32_t start_step, int32_t *xqd,
     394             :     const SgrParamsType *params) {
     395        4800 :     int64_t err = get_pixel_proj_error(
     396             :         src8, width, height, src_stride, dat8, dat_stride, use_highbitdepth, flt0,
     397             :         flt0_stride, flt1, flt1_stride, xqd, params);
     398             :     (void)start_step;
     399             : #if USE_SGRPROJ_REFINEMENT_SEARCH
     400             :     int64_t err2;
     401        4800 :     int32_t tap_min[] = { SGRPROJ_PRJ_MIN0, SGRPROJ_PRJ_MIN1 };
     402        4800 :     int32_t tap_max[] = { SGRPROJ_PRJ_MAX0, SGRPROJ_PRJ_MAX1 };
     403       14400 :     for (int32_t s = start_step; s >= 1; s >>= 1) {
     404       23073 :         for (int32_t p = 0; p < 2; ++p) {
     405       16930 :             if ((params->r[0] == 0 && p == 0) || (params->r[1] == 0 && p == 1))
     406        3082 :                 continue;
     407       13848 :             int32_t skip = 0;
     408             :             do {
     409       20159 :                 if (xqd[p] - s >= tap_min[p]) {
     410       17519 :                     xqd[p] -= s;
     411             :                     err2 =
     412       17519 :                         get_pixel_proj_error(src8, width, height, src_stride, dat8,
     413             :                             dat_stride, use_highbitdepth, flt0,
     414             :                             flt0_stride, flt1, flt1_stride, xqd, params);
     415       17519 :                     if (err2 > err)
     416        9840 :                         xqd[p] += s;
     417             :                     else {
     418        7679 :                         err = err2;
     419        7679 :                         skip = 1;
     420             :                         // At the highest step size continue moving in the same direction
     421        7679 :                         if (s == start_step) continue;
     422             :                     }
     423             :                 }
     424       13848 :                 break;
     425             :             } while (1);
     426       13848 :             if (skip) break;
     427             :             do {
     428       16607 :                 if (xqd[p] + s <= tap_max[p]) {
     429       14270 :                     xqd[p] += s;
     430             :                     err2 =
     431       14270 :                         get_pixel_proj_error(src8, width, height, src_stride, dat8,
     432             :                             dat_stride, use_highbitdepth, flt0,
     433             :                             flt0_stride, flt1, flt1_stride, xqd, params);
     434       14270 :                     if (err2 > err)
     435        6928 :                         xqd[p] -= s;
     436             :                     else {
     437        7342 :                         err = err2;
     438             :                         // At the highest step size continue moving in the same direction
     439        7342 :                         if (s == start_step) continue;
     440             :                     }
     441             :                 }
     442       10391 :                 break;
     443             :             } while (1);
     444             :         }
     445             :     }
     446             : #endif  // USE_SGRPROJ_REFINEMENT_SEARCH
     447        4800 :     return err;
     448             : }
     449             : 
     450             : extern void RunEmms();
     451             : 
     452           0 : void get_proj_subspace_c(const uint8_t *src8, int32_t width, int32_t height,
     453             :     int32_t src_stride, const uint8_t *dat8,
     454             :     int32_t dat_stride, int32_t use_highbitdepth,
     455             :     int32_t *flt0, int32_t flt0_stride, int32_t *flt1,
     456             :     int32_t flt1_stride, int32_t *xq,
     457             :     const SgrParamsType *params) {
     458             :     int32_t i, j;
     459           0 :     double H[2][2] = { { 0, 0 }, { 0, 0 } };
     460           0 :     double C[2] = { 0, 0 };
     461             :     double Det;
     462             :     double x[2];
     463           0 :     const int32_t size = width * height;
     464             : 
     465           0 :     aom_clear_system_state();
     466           0 :     RunEmms();
     467             : 
     468             :     // Default
     469           0 :     xq[0] = 0;
     470           0 :     xq[1] = 0;
     471           0 :     if (!use_highbitdepth) {
     472           0 :         const uint8_t *src = src8;
     473           0 :         const uint8_t *dat = dat8;
     474           0 :         for (i = 0; i < height; ++i) {
     475           0 :             for (j = 0; j < width; ++j) {
     476           0 :                 const double u = (double)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS);
     477           0 :                 const double s =
     478           0 :                     (double)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u;
     479           0 :                 const double f1 =
     480           0 :                     (params->r[0] > 0) ? (double)flt0[i * flt0_stride + j] - u : 0;
     481           0 :                 const double f2 =
     482           0 :                     (params->r[1] > 0) ? (double)flt1[i * flt1_stride + j] - u : 0;
     483           0 :                 H[0][0] += f1 * f1;
     484           0 :                 H[1][1] += f2 * f2;
     485           0 :                 H[0][1] += f1 * f2;
     486           0 :                 C[0] += f1 * s;
     487           0 :                 C[1] += f2 * s;
     488             :             }
     489             :         }
     490             :     }
     491             :     else {
     492           0 :         const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
     493           0 :         const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
     494           0 :         for (i = 0; i < height; ++i) {
     495           0 :             for (j = 0; j < width; ++j) {
     496           0 :                 const double u = (double)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS);
     497           0 :                 const double s =
     498           0 :                     (double)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u;
     499           0 :                 const double f1 =
     500           0 :                     (params->r[0] > 0) ? (double)flt0[i * flt0_stride + j] - u : 0;
     501           0 :                 const double f2 =
     502           0 :                     (params->r[1] > 0) ? (double)flt1[i * flt1_stride + j] - u : 0;
     503           0 :                 H[0][0] += f1 * f1;
     504           0 :                 H[1][1] += f2 * f2;
     505           0 :                 H[0][1] += f1 * f2;
     506           0 :                 C[0] += f1 * s;
     507           0 :                 C[1] += f2 * s;
     508             :             }
     509             :         }
     510             :     }
     511           0 :     H[0][0] /= size;
     512           0 :     H[0][1] /= size;
     513           0 :     H[1][1] /= size;
     514           0 :     H[1][0] = H[0][1];
     515           0 :     C[0] /= size;
     516           0 :     C[1] /= size;
     517           0 :     if (params->r[0] == 0) {
     518             :         // H matrix is now only the scalar H[1][1]
     519             :         // C vector is now only the scalar C[1]
     520           0 :         Det = H[1][1];
     521           0 :         if (Det < 1e-8) return;  // ill-posed, return default values
     522           0 :         x[0] = 0;
     523           0 :         x[1] = C[1] / Det;
     524             : 
     525           0 :         xq[0] = 0;
     526           0 :         xq[1] = (int32_t)rint(x[1] * (1 << SGRPROJ_PRJ_BITS));
     527             :     }
     528           0 :     else if (params->r[1] == 0) {
     529             :         // H matrix is now only the scalar H[0][0]
     530             :         // C vector is now only the scalar C[0]
     531           0 :         Det = H[0][0];
     532           0 :         if (Det < 1e-8) return;  // ill-posed, return default values
     533           0 :         x[0] = C[0] / Det;
     534           0 :         x[1] = 0;
     535             : 
     536           0 :         xq[0] = (int32_t)rint(x[0] * (1 << SGRPROJ_PRJ_BITS));
     537           0 :         xq[1] = 0;
     538             :     }
     539             :     else {
     540           0 :         Det = (H[0][0] * H[1][1] - H[0][1] * H[1][0]);
     541           0 :         if (Det < 1e-8) return;  // ill-posed, return default values
     542           0 :         x[0] = (H[1][1] * C[0] - H[0][1] * C[1]) / Det;
     543           0 :         x[1] = (H[0][0] * C[1] - H[1][0] * C[0]) / Det;
     544             : 
     545           0 :         xq[0] = (int32_t)rint(x[0] * (1 << SGRPROJ_PRJ_BITS));
     546           0 :         xq[1] = (int32_t)rint(x[1] * (1 << SGRPROJ_PRJ_BITS));
     547             :     }
     548             : }
     549             : 
     550        4800 : static INLINE void encode_xq(int32_t *xq, int32_t *xqd,
     551             :     const SgrParamsType *params)
     552             : {
     553        4800 :     if (params->r[0] == 0) {
     554        1200 :         xqd[0] = 0;
     555        1200 :         xqd[1] = clamp((1 << SGRPROJ_PRJ_BITS) - xq[1], SGRPROJ_PRJ_MIN1,
     556             :             SGRPROJ_PRJ_MAX1);
     557             :     }
     558        3600 :     else if (params->r[1] == 0) {
     559         600 :         xqd[0] = clamp(xq[0], SGRPROJ_PRJ_MIN0, SGRPROJ_PRJ_MAX0);
     560         600 :         xqd[1] = clamp((1 << SGRPROJ_PRJ_BITS) - xqd[0], SGRPROJ_PRJ_MIN1,
     561             :             SGRPROJ_PRJ_MAX1);
     562             :     }
     563             :     else {
     564        3000 :         xqd[0] = clamp(xq[0], SGRPROJ_PRJ_MIN0, SGRPROJ_PRJ_MAX0);
     565        3000 :         xqd[1] = clamp((1 << SGRPROJ_PRJ_BITS) - xqd[0] - xq[1], SGRPROJ_PRJ_MIN1,
     566             :             SGRPROJ_PRJ_MAX1);
     567             :     }
     568        4800 : }
     569             : 
     570             : // Apply the self-guided filter across an entire restoration unit.
     571        4800 : static INLINE void apply_sgr(int32_t sgr_params_idx, const uint8_t *dat8,
     572             :     int32_t width, int32_t height, int32_t dat_stride, int32_t use_highbd,
     573             :     int32_t bit_depth, int32_t pu_width, int32_t pu_height, int32_t *flt0,
     574             :     int32_t *flt1, int32_t flt_stride)
     575             : {
     576       33600 :     for (int32_t i = 0; i < height; i += pu_height)
     577             :     {
     578       28800 :         const int32_t h = AOMMIN(pu_height, height - i);
     579       28800 :         int32_t *flt0_row = flt0 + i * flt_stride;
     580       28800 :         int32_t *flt1_row = flt1 + i * flt_stride;
     581       28800 :         const uint8_t *dat8_row = dat8 + i * dat_stride;
     582             : 
     583             :         // Iterate over the stripe in blocks of width pu_width
     584      201599 :         for (int32_t j = 0; j < width; j += pu_width) {
     585      172799 :             const int32_t w = AOMMIN(pu_width, width - j);
     586             : 
     587             :             //CHKN SSE
     588      172799 :             eb_av1_selfguided_restoration(dat8_row + j, w, h, dat_stride, flt0_row + j,
     589      172799 :                 flt1_row + j, flt_stride, sgr_params_idx,
     590             :                 bit_depth, use_highbd);
     591             :         }
     592             :     }
     593        4800 : }
     594             : 
     595         300 : static SgrprojInfo search_selfguided_restoration(
     596             :     const uint8_t *dat8, int32_t width, int32_t height, int32_t dat_stride,
     597             :     const uint8_t *src8, int32_t src_stride, int32_t use_highbitdepth, int32_t bit_depth,
     598             :     int32_t pu_width, int32_t pu_height, int32_t *rstbuf
     599             :     ,
     600             :     int8_t sg_ref_frame_ep[2],
     601             :     int32_t sg_frame_ep_cnt[SGRPROJ_PARAMS],
     602             :     int8_t step
     603             : )
     604             : {
     605         300 :     int32_t *flt0 = rstbuf;
     606         300 :     int32_t *flt1 = flt0 + RESTORATION_UNITPELS_MAX;
     607         300 :     int32_t ep, bestep = 0;
     608         300 :     int64_t besterr = -1;
     609         300 :     int32_t exqd[2], bestxqd[2] = { 0, 0 };
     610         300 :     int32_t flt_stride = ((width + 7) & ~7) + 8;
     611             :     assert(pu_width == (RESTORATION_PROC_UNIT_SIZE >> 1) ||
     612             :         pu_width == RESTORATION_PROC_UNIT_SIZE);
     613             :     assert(pu_height == (RESTORATION_PROC_UNIT_SIZE >> 1) ||
     614             :         pu_height == RESTORATION_PROC_UNIT_SIZE);
     615         590 :     int8_t mid_ep = sg_ref_frame_ep[0] < 0 && sg_ref_frame_ep[1] < 0 ? 0 :
     616         290 :         sg_ref_frame_ep[1] < 0 ? sg_ref_frame_ep[0] :
     617         290 :         sg_ref_frame_ep[0] < 0 ? sg_ref_frame_ep[1] :
     618         290 :         (sg_ref_frame_ep[0] + sg_ref_frame_ep[1]) / 2;
     619             : 
     620         300 :     int8_t start_ep = sg_ref_frame_ep[0] < 0 && sg_ref_frame_ep[1] < 0 ? 0 : AOMMAX(0, mid_ep - step);
     621         300 :     int8_t end_ep = sg_ref_frame_ep[0] < 0 && sg_ref_frame_ep[1] < 0 ? SGRPROJ_PARAMS : AOMMIN(SGRPROJ_PARAMS, mid_ep + step);
     622             : 
     623        5100 :     for (ep = start_ep; ep < end_ep; ep++) {
     624             :         int32_t exq[2];
     625        4800 :         apply_sgr(ep, dat8, width, height, dat_stride, use_highbitdepth, bit_depth,
     626             :             pu_width, pu_height, flt0, flt1, flt_stride);
     627        4800 :         aom_clear_system_state();
     628        4800 :         const SgrParamsType *const params = &eb_sgr_params[ep];
     629        4800 :         get_proj_subspace(src8, width, height, src_stride, dat8, dat_stride,
     630             :             use_highbitdepth, flt0, flt_stride, flt1, flt_stride, exq,
     631             :             params);
     632        4800 :         aom_clear_system_state();
     633        4800 :         encode_xq(exq, exqd, params);
     634        4800 :         int64_t err = finer_search_pixel_proj_error(
     635             :             src8, width, height, src_stride, dat8, dat_stride, use_highbitdepth,
     636             :             flt0, flt_stride, flt1, flt_stride, 2, exqd, params);
     637        4800 :         if (besterr == -1 || err < besterr) {
     638        1273 :             bestep = ep;
     639        1273 :             besterr = err;
     640        1273 :             bestxqd[0] = exqd[0];
     641        1273 :             bestxqd[1] = exqd[1];
     642             :         }
     643             :     }
     644         300 :     sg_frame_ep_cnt[bestep]++;
     645             : 
     646             :     SgrprojInfo ret;
     647         300 :     ret.ep = bestep;
     648         300 :     ret.xqd[0] = bestxqd[0];
     649         300 :     ret.xqd[1] = bestxqd[1];
     650         300 :     return ret;
     651             : }
     652             : extern int32_t eb_aom_count_primitive_refsubexpfin(uint16_t n, uint16_t k, uint16_t ref, uint16_t v);
     653             : 
     654         414 : static int32_t count_sgrproj_bits(SgrprojInfo *sgrproj_info,
     655             :     SgrprojInfo *ref_sgrproj_info) {
     656         414 :     int32_t bits = SGRPROJ_PARAMS_BITS;
     657         414 :     const SgrParamsType *params = &eb_sgr_params[sgrproj_info->ep];
     658         414 :     if (params->r[0] > 0)
     659         411 :         bits += eb_aom_count_primitive_refsubexpfin(
     660             :             SGRPROJ_PRJ_MAX0 - SGRPROJ_PRJ_MIN0 + 1, SGRPROJ_PRJ_SUBEXP_K,
     661         411 :             (uint16_t)(ref_sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0),
     662         411 :             (uint16_t)(sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0));
     663         414 :     if (params->r[1] > 0)
     664         324 :         bits += eb_aom_count_primitive_refsubexpfin(
     665             :             SGRPROJ_PRJ_MAX1 - SGRPROJ_PRJ_MIN1 + 1, SGRPROJ_PRJ_SUBEXP_K,
     666         324 :             (uint16_t)(ref_sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1),
     667         324 :             (uint16_t)(sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1));
     668         414 :     return bits;
     669             : }
     670             : 
     671         300 : int8_t get_sg_step(int8_t  sg_filter_mode) {
     672             :     int8_t step;
     673         300 :     switch (sg_filter_mode) {
     674           0 :     case 1:
     675           0 :         step = 0;
     676           0 :         break;
     677           0 :     case 2:
     678           0 :         step = 1;
     679           0 :         break;
     680           0 :     case 3:
     681           0 :         step = 4;
     682           0 :         break;
     683         300 :     case 4:
     684         300 :         step = 16;
     685         300 :         break;
     686           0 :     default:
     687           0 :         step = 16;
     688           0 :         break;
     689             :     }
     690         300 :     return step;
     691             : }
     692             : 
     693           0 : static void search_sgrproj(const RestorationTileLimits *limits,
     694             :     const AV1PixelRect *tile, int32_t rest_unit_idx,
     695             :     void *priv) {
     696           0 :     RestSearchCtxt *rsc = (RestSearchCtxt *)priv;
     697           0 :     RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx];
     698             : 
     699           0 :     const Macroblock *const x = rsc->x;
     700           0 :     Av1Common *const cm = rsc->cm;
     701             : 
     702           0 :     const int32_t highbd = cm->use_highbitdepth;
     703           0 :     const int32_t bit_depth = cm->bit_depth;
     704             : 
     705           0 :     uint8_t *dgd_start =
     706           0 :         rsc->dgd_buffer + limits->v_start * rsc->dgd_stride + limits->h_start;
     707           0 :     const uint8_t *src_start =
     708           0 :         rsc->src_buffer + limits->v_start * rsc->src_stride + limits->h_start;
     709             : 
     710           0 :     const int32_t is_uv = rsc->plane > 0;
     711           0 :     const int32_t ss_x = is_uv && cm->subsampling_x;
     712           0 :     const int32_t ss_y = is_uv && cm->subsampling_y;
     713           0 :     const int32_t procunit_width = RESTORATION_PROC_UNIT_SIZE >> ss_x;
     714           0 :     const int32_t procunit_height = RESTORATION_PROC_UNIT_SIZE >> ss_y;
     715             : 
     716           0 :     int8_t step = get_sg_step(cm->sg_filter_mode);
     717             : 
     718           0 :     rusi->sgrproj = search_selfguided_restoration(
     719           0 :         dgd_start, limits->h_end - limits->h_start,
     720           0 :         limits->v_end - limits->v_start, rsc->dgd_stride, src_start,
     721             :         rsc->src_stride, highbd, bit_depth, procunit_width, procunit_height,
     722             :         cm->rst_tmpbuf
     723           0 :         , cm->sg_ref_frame_ep,
     724           0 :         cm->sg_frame_ep_cnt,
     725             :         step
     726             :     );
     727             : 
     728             :     RestorationUnitInfo rui;
     729           0 :     rui.restoration_type = RESTORE_SGRPROJ;
     730           0 :     rui.sgrproj_info = rusi->sgrproj;
     731             : 
     732           0 :     rusi->sse[RESTORE_SGRPROJ] = try_restoration_unit(rsc, limits, tile, &rui);
     733             : 
     734           0 :     const int64_t bits_none = x->sgrproj_restore_cost[0];
     735           0 :     const int64_t bits_sgr = x->sgrproj_restore_cost[1] +
     736           0 :         (count_sgrproj_bits(&rusi->sgrproj, &rsc->sgrproj)
     737           0 :             << AV1_PROB_COST_SHIFT);
     738             : 
     739           0 :     double cost_none =
     740           0 :         RDCOST_DBL(x->rdmult, bits_none >> 4, rusi->sse[RESTORE_NONE]);
     741           0 :     double cost_sgr =
     742           0 :         RDCOST_DBL(x->rdmult, bits_sgr >> 4, rusi->sse[RESTORE_SGRPROJ]);
     743             : 
     744           0 :     RestorationType rtype =
     745           0 :         (cost_sgr < cost_none) ? RESTORE_SGRPROJ : RESTORE_NONE;
     746           0 :     rusi->best_rtype[RESTORE_SGRPROJ - 1] = rtype;
     747             : 
     748           0 :     rsc->sse += rusi->sse[rtype];
     749           0 :     rsc->bits += (cost_sgr < cost_none) ? bits_sgr : bits_none;
     750           0 :     if (cost_sgr < cost_none) rsc->sgrproj = rusi->sgrproj;
     751           0 : }
     752             : 
     753           0 : void eb_av1_compute_stats_c(int32_t wiener_win, const uint8_t *dgd, const uint8_t *src,
     754             :     int32_t h_start, int32_t h_end, int32_t v_start, int32_t v_end,
     755             :     int32_t dgd_stride, int32_t src_stride, int64_t *M,
     756             :     int64_t *H) {
     757             :     int32_t i, j, k, l;
     758           0 :     int16_t Y[WIENER_WIN2] = { 0 };
     759           0 :     const int32_t wiener_win2 = wiener_win * wiener_win;
     760           0 :     const int32_t wiener_halfwin = (wiener_win >> 1);
     761           0 :     uint8_t avg = find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride);
     762             : 
     763           0 :     memset(M, 0, sizeof(*M) * wiener_win2);
     764           0 :     memset(H, 0, sizeof(*H) * wiener_win2 * wiener_win2);
     765           0 :     for (i = v_start; i < v_end; i++) {
     766           0 :         for (j = h_start; j < h_end; j++) {
     767           0 :             const int16_t X = (int16_t)src[i * src_stride + j] - (int16_t)avg;
     768           0 :             int32_t idx = 0;
     769           0 :             for (k = -wiener_halfwin; k <= wiener_halfwin; k++) {
     770           0 :                 for (l = -wiener_halfwin; l <= wiener_halfwin; l++) {
     771           0 :                     Y[idx] = (int16_t)dgd[(i + l) * dgd_stride + (j + k)] - (int16_t)avg;
     772           0 :                     idx++;
     773             :                 }
     774             :             }
     775             :             assert(idx == wiener_win2);
     776           0 :             for (k = 0; k < wiener_win2; ++k) {
     777           0 :                 M[k] += (int32_t)Y[k] * X;
     778           0 :                 for (l = k; l < wiener_win2; ++l) {
     779             :                     // H is a symmetric matrix, so we only need to fill out the upper
     780             :                     // triangle here. We can copy it down to the lower triangle outside
     781             :                     // the (i, j) loops.
     782           0 :                     H[k * wiener_win2 + l] += (int32_t)Y[k] * Y[l];
     783             :                 }
     784             :             }
     785             :         }
     786             :     }
     787           0 :     for (k = 0; k < wiener_win2; ++k) {
     788           0 :         for (l = k + 1; l < wiener_win2; ++l)
     789           0 :             H[l * wiener_win2 + k] = H[k * wiener_win2 + l];
     790             :     }
     791           0 : }
     792             : 
     793             : #if 0
     794             : // Note: Don't delete! This is the code base for optimizations, and 4.8x, 9x or
     795             : //       16x faster than the above original C version for win 3, 5 and 7
     796             : //       respectively, while keeping the bit-exactness.
     797             : 
     798             : // Demostration of TAP 3:
     799             : //                       (Left Edge)         (Left Edge)
     800             : //  \lj   00    10    20      01    11    21      02    12    22
     801             : // ki-------------------------------------------------------------
     802             : // 00| 00X00 00X10 00X20 | 00X01 00X11 00X21 | 00X02 00X12 00X22 |(Top Edge)
     803             : // 10|       10x10 10x20 | 10X01 10x11 10x21 | 10X02 10x12 10x22 |(Top Row)
     804             : // 20|             20x20 | 20X01 20x11 20x21 | 20X02 20x12 20x22 |
     805             : //   |                   --------------------|-------------------|
     806             : // 01|                     01x01 01x11 01x21 | 01x02 01x12 01x22 |(Top Edge)
     807             : // 11|                           11x11 11x21 | 11x02 11x12 11x22 |(Mid Rows)
     808             : // 21|                                 21x21 | 21x02 21x12 21x22 |
     809             : //   |                                       --------------------|
     810             : // 02|                                         02x02 02x12 02x22 |(Top Edge)
     811             : // 12|                                               12x12 12x22 |(Btm Row)
     812             : // 22|                                                     22x22 |
     813             : //   -------------------------------------------------------------
     814             : // kiXlj means the multiply-accumulate must be calculated from scratch, and
     815             : // starts from a[k * a_stride + i] * b[l * b_stride + j].
     816             : // kixlj means the multiply-accumulate could be derived from a neighbor.
     817             : 
     818             : static INLINE void sub_avg_block(const uint8_t *src, const int32_t src_stride,
     819             :     const uint8_t avg, const int32_t width,
     820             :     const int32_t height, int16_t *dst,
     821             :     const int32_t dst_stride) {
     822             :     for (int32_t i = 0; i < height; i++) {
     823             :         for (int32_t j = 0; j < width; j++)
     824             :             dst[i * dst_stride + j] = (int16_t)src[i * src_stride + j] - avg;
     825             :     }
     826             : }
     827             : 
     828             : void av1_compute_stats_base(int32_t wiener_win, const uint8_t *dgd,
     829             :     const uint8_t *src, int32_t h_start, int32_t h_end,
     830             :     int32_t v_start, int32_t v_end, int32_t dgd_stride,
     831             :     int32_t src_stride, int64_t *M, int64_t *H) {
     832             :     const int32_t wiener_win2 = wiener_win * wiener_win;
     833             :     const int32_t wiener_halfwin = wiener_win >> 1;
     834             :     const uint8_t avg =
     835             :         find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride);
     836             :     const int32_t width = h_end - h_start;
     837             :     const int32_t height = v_end - v_start;
     838             :     const int32_t d_stride = (width + 2 * wiener_halfwin + 15) & ~15;
     839             :     const int32_t s_stride = (width + 15) & ~15;
     840             :     // The maximum input size is width * height, which is
     841             :     // (9 / 4) * RESTORATION_UNITSIZE_MAX * RESTORATION_UNITSIZE_MAX. Enlarge to
     842             :     // 3 * RESTORATION_UNITSIZE_MAX * RESTORATION_UNITSIZE_MAX considering
     843             :     // paddings.
     844             :     int16_t d[3 * RESTORATION_UNITSIZE_MAX * RESTORATION_UNITSIZE_MAX];
     845             :     int16_t s[3 * RESTORATION_UNITSIZE_MAX * RESTORATION_UNITSIZE_MAX];
     846             :     int32_t i, j, k, l, x, y;
     847             : 
     848             :     sub_avg_block(src + v_start * src_stride + h_start,
     849             :         src_stride,
     850             :         avg,
     851             :         width,
     852             :         height,
     853             :         s,
     854             :         s_stride);
     855             :     sub_avg_block(dgd + (v_start - wiener_halfwin) * dgd_stride + h_start -
     856             :         wiener_halfwin,
     857             :         dgd_stride,
     858             :         avg,
     859             :         width + 2 * wiener_halfwin,
     860             :         height + 2 * wiener_halfwin,
     861             :         d,
     862             :         d_stride);
     863             : 
     864             :     // Step 1: Calculate the top edge of the whole matrix, i.e., the top edge of
     865             :     // each triangle and square on the top row.
     866             :     // Demostration of TAP 3:
     867             :     //                       (Left Edge)         (Left Edge)
     868             :     //  \lj   00    10    20      01    11    21      02    12    22
     869             :     // ki-------------------------------------------------------------
     870             :     // 00| 00X00 00X10 00X20 | 00X01 00X11 00X21 | 00X02 00X12 00X22 |(Top Edge)
     871             :     // 10|                                                           |(Top Row)
     872             :     // 20|                                                           |
     873             :     //   |                   --------------------|-------------------|
     874             :     // 01|                                                           |(Top Edge)
     875             :     // 11|                                                           |(Mid Rows)
     876             :     // 21|                                                           |
     877             :     //   |                                       --------------------|
     878             :     // 02|                                                           |(Top Edge)
     879             :     // 12|                                                           |(Btm Row)
     880             :     // 22|                                                           |
     881             :     //   -------------------------------------------------------------
     882             :     for (j = 0; j < wiener_win; j++) {
     883             :         for (l = 0; l < wiener_win; l++) {
     884             :             int64_t sumM = 0;
     885             :             int64_t sumH = 0;
     886             : 
     887             :             for (y = 0; y < height; y++) {
     888             :                 for (x = 0; x < width; x++) {
     889             :                     sumM += s[y * s_stride + x] * d[(l + y) * d_stride + j + x];
     890             :                     sumH += d[y * d_stride + x] * d[(l + y) * d_stride + j + x];
     891             :                 }
     892             :             }
     893             : 
     894             :             M[j * wiener_win + l] = sumM;
     895             :             H[j * wiener_win + l] = sumH;
     896             :         }
     897             :     }
     898             : 
     899             :     // Step 2: Calculate the left edge of each square on the top row.
     900             :     // Demostration of TAP 3:
     901             :     //                       (Left Edge)         (Left Edge)
     902             :     //  \lj   00    10    20      01    11    21      02    12    22
     903             :     // ki-------------------------------------------------------------
     904             :     // 00|                                                           |(Top Edge)
     905             :     // 10|                   | 10X01             | 10X02             |(Top Row)
     906             :     // 20|                   | 20X01             | 20X02             |
     907             :     //   |                   --------------------|-------------------|
     908             :     // 01|                                                           |(Top Edge)
     909             :     // 11|                                                           |(Mid Rows)
     910             :     // 21|                                                           |
     911             :     //   |                                       --------------------|
     912             :     // 02|                                                           |(Top Edge)
     913             :     // 12|                                                           |(Btm Row)
     914             :     // 22|                                                           |
     915             :     //   -------------------------------------------------------------
     916             :     for (j = 1; j < wiener_win; j++) {
     917             :         for (k = 1; k < wiener_win; k++) {
     918             :             int64_t sumH = 0;
     919             : 
     920             :             for (y = 0; y < height; y++) {
     921             :                 for (x = 0; x < width; x++)
     922             :                     sumH += d[(k + y) * d_stride + x] * d[y * d_stride + j + x];
     923             :             }
     924             : 
     925             :             H[k * wiener_win2 + j * wiener_win] = sumH;
     926             :         }
     927             :     }
     928             : 
     929             :     // Step 3: Derive the top edge of each triangle along the diagonal. No
     930             :     // triangle in top row.
     931             :     // Demostration of TAP 3:
     932             :     //                       (Left Edge)         (Left Edge)
     933             :     //  \lj   00    10    20      01    11    21      02    12    22
     934             :     // ki-------------------------------------------------------------
     935             :     // 00|                                                           |(Top Edge)
     936             :     // 10|                                                           |(Top Row)
     937             :     // 20|                                                           |
     938             :     //   |                   --------------------|-------------------|
     939             :     // 01|                     01x01 01x11 01x21 |                   |(Top Edge)
     940             :     // 11|                                                           |(Mid Rows)
     941             :     // 21|                                                           |
     942             :     //   |                                       --------------------|
     943             :     // 02|                                         02x02 02x12 02x22 |(Top Edge)
     944             :     // 12|                                                           |(Btm Row)
     945             :     // 22|                                                           |
     946             :     //   -------------------------------------------------------------
     947             :     for (i = 1; i < wiener_win; i++) {
     948             :         for (l = 0; l < wiener_win; l++) {
     949             :             int32_t delta = 0;
     950             : 
     951             :             for (y = 0; y < height; y++) {
     952             :                 delta -=
     953             :                     d[y * d_stride + i - 1] * d[(l + y) * d_stride + i - 1];
     954             :                 delta += d[y * d_stride + i - 1 + width] *
     955             :                     d[(l + y) * d_stride + i - 1 + width];
     956             :             }
     957             : 
     958             :             H[i * wiener_win * wiener_win2 + i * wiener_win + l] =
     959             :                 H[(i - 1) * wiener_win * wiener_win2 + (i - 1) * wiener_win +
     960             :                 l] +
     961             :                 delta;
     962             :         }
     963             :     }
     964             : 
     965             :     // Step 4: Derive the top and left edge of each square. No square in top and
     966             :     // bottom row. (There is only 1 square in this step for TAP 3, but there are
     967             :     // many for TAP 5 and 7.)
     968             :     // Demostration of TAP 3:
     969             :     //                       (Left Edge)         (Left Edge)
     970             :     //  \lj   00    10    20      01    11    21      02    12    22
     971             :     // ki-------------------------------------------------------------
     972             :     // 00|                                                           |(Top Edge)
     973             :     // 10|                                                           |(Top Row)
     974             :     // 20|                                                           |
     975             :     //   |                   --------------------|-------------------|
     976             :     // 01|                                       | 01x02 01x12 01x22 |(Top Edge)
     977             :     // 11|                                       | 11x02             |(Mid Rows)
     978             :     // 21|                                       | 21x02             |
     979             :     //   |                                       --------------------|
     980             :     // 02|                                                           |(Top Edge)
     981             :     // 12|                                                           |(Btm Row)
     982             :     // 22|                                                           |
     983             :     //   -------------------------------------------------------------
     984             :     for (i = 1; i < wiener_win - 1; i++) {
     985             :         for (j = i + 1; j < wiener_win; j++) {
     986             :             for (l = 0; l < wiener_win; l++) {
     987             :                 int32_t delta = 0;
     988             : 
     989             :                 for (y = 0; y < height; y++) {
     990             :                     delta -=
     991             :                         d[y * d_stride + i - 1] * d[(l + y) * d_stride + j - 1];
     992             :                     delta += d[y * d_stride + i - 1 + width] *
     993             :                         d[(l + y) * d_stride + j - 1 + width];
     994             :                 }
     995             : 
     996             :                 H[i * wiener_win * wiener_win2 + j * wiener_win + l] =
     997             :                     H[(i - 1) * wiener_win * wiener_win2 +
     998             :                     (j - 1) * wiener_win + l] +
     999             :                     delta;
    1000             :             }
    1001             : 
    1002             :             for (k = 1; k < wiener_win; k++) {
    1003             :                 int32_t delta = 0;
    1004             : 
    1005             :                 for (y = 0; y < height; y++) {
    1006             :                     delta -=
    1007             :                         d[(k + y) * d_stride + i - 1] * d[y * d_stride + j - 1];
    1008             :                     delta += d[(k + y) * d_stride + i - 1 + width] *
    1009             :                         d[y * d_stride + j - 1 + width];
    1010             :                 }
    1011             : 
    1012             :                 H[(i * wiener_win + k) * wiener_win2 + j * wiener_win] =
    1013             :                     H[((i - 1) * wiener_win + k) * wiener_win2 +
    1014             :                     (j - 1) * wiener_win] +
    1015             :                     delta;
    1016             :             }
    1017             :         }
    1018             :     }
    1019             : 
    1020             :     // Step 5: Derive other points of each square. No square in bottom row.
    1021             :     // Demostration of TAP 3:
    1022             :     //                       (Left Edge)         (Left Edge)
    1023             :     //  \lj   00    10    20      01    11    21      02    12    22
    1024             :     // ki-------------------------------------------------------------
    1025             :     // 00|                   |                   |                   |(Top Edge)
    1026             :     // 10|                   |       10x11 10x21 |       10x12 10x22 |(Top Row)
    1027             :     // 20|                   |       20x11 20x21 |       20x12 20x22 |
    1028             :     //   |                   --------------------|-------------------|
    1029             :     // 01|                                       |                   |(Top Edge)
    1030             :     // 11|                                       |       11x12 11x22 |(Mid Rows)
    1031             :     // 21|                                       |       21x12 21x22 |
    1032             :     //   |                                       --------------------|
    1033             :     // 02|                                                           |(Top Edge)
    1034             :     // 12|                                                           |(Btm Row)
    1035             :     // 22|                                                           |
    1036             :     //   -------------------------------------------------------------
    1037             :     for (i = 0; i < wiener_win - 1; i++) {
    1038             :         for (j = i + 1; j < wiener_win; j++) {
    1039             :             for (k = 1; k < wiener_win; k++) {
    1040             :                 for (l = 1; l < wiener_win; l++) {
    1041             :                     int32_t delta = 0;
    1042             : 
    1043             :                     for (x = 0; x < width; x++) {
    1044             :                         delta -= d[(k - 1) * d_stride + i + x] *
    1045             :                             d[(l - 1) * d_stride + j + x];
    1046             :                         delta += d[(k - 1 + height) * d_stride + i + x] *
    1047             :                             d[(l - 1 + height) * d_stride + j + x];
    1048             :                     }
    1049             : 
    1050             :                     H[(i * wiener_win + k) * wiener_win2 + j * wiener_win + l] =
    1051             :                         H[(i * wiener_win + k - 1) * wiener_win2 +
    1052             :                         j * wiener_win + l - 1] +
    1053             :                         delta;
    1054             :                 }
    1055             :             }
    1056             :         }
    1057             :     }
    1058             : 
    1059             :     // Step 6: Derive other points of each upper triangle along the diagonal.
    1060             :     // Demostration of TAP 3:
    1061             :     //                       (Left Edge)         (Left Edge)
    1062             :     //  \lj   00    10    20      01    11    21      02    12    22
    1063             :     // ki-------------------------------------------------------------
    1064             :     // 00|                   |                   |                   |(Top Edge)
    1065             :     // 10|       10x10 10x20 |                   |                   |(Top Row)
    1066             :     // 20|             20x20 |                   |                   |
    1067             :     //   |                   --------------------|-------------------|
    1068             :     // 01|                                       |                   |(Top Edge)
    1069             :     // 11|                           11x11 11x21 |                   |(Mid Rows)
    1070             :     // 21|                                 21x21 |                   |
    1071             :     //   |                                       --------------------|
    1072             :     // 02|                                                           |(Top Edge)
    1073             :     // 12|                                               12x12 12x22 |(Btm Row)
    1074             :     // 22|                                                     22x22 |
    1075             :     //   -------------------------------------------------------------
    1076             :     for (i = 0; i < wiener_win; i++) {
    1077             :         for (k = 1; k < wiener_win; k++) {
    1078             :             for (l = k; l < wiener_win; l++) {
    1079             :                 int32_t delta = 0;
    1080             : 
    1081             :                 for (x = 0; x < width; x++) {
    1082             :                     delta -= d[(k - 1) * d_stride + i + x] *
    1083             :                         d[(l - 1) * d_stride + i + x];
    1084             :                     delta += d[(k - 1 + height) * d_stride + i + x] *
    1085             :                         d[(l - 1 + height) * d_stride + i + x];
    1086             :                 }
    1087             : 
    1088             :                 H[(i * wiener_win + k) * wiener_win2 + i * wiener_win + l] =
    1089             :                     H[(i * wiener_win + k - 1) * wiener_win2 + i * wiener_win +
    1090             :                     l - 1] +
    1091             :                     delta;
    1092             :             }
    1093             :         }
    1094             :     }
    1095             : 
    1096             :     // H is a symmetric matrix, so we only need to fill out the upper triangle.
    1097             :     // We can copy it down to the lower triangle outside the (i, j) loops.
    1098             :     // Divided into 4x4 squares to do load-and-transpose-and-store in
    1099             :     // optimization.
    1100             :     // Demostration of TAP 3:
    1101             :     //  \lj   00    10    20    01      11    21    02    12      22
    1102             :     // ki-------------------------------------------------------------
    1103             :     // 00|                                                           |
    1104             :     //   |------                                                     |
    1105             :     // 10| 00X10                                                     |
    1106             :     // 20| 00X20 10x20                                               |
    1107             :     // 01| 00X01 10X01 20X01                                         |
    1108             :     // 11| 00X11 10x11 20x11 01x11                                   |
    1109             :     //   |--------------------------                                 |
    1110             :     // 21| 00X21 10x21 20x21 01x21 | 11x21                           |
    1111             :     // 02| 00X02 10X02 20X02 01x02 | 11x02 21x02                     |
    1112             :     // 12| 00X12 10x12 20x12 01x12 | 11x12 21x12 02x12               |
    1113             :     // 22| 00X22 10x22 20x22 01x22 | 11x22 21x22 02x22 12x22 |       |
    1114             :     //   -------------------------------------------------------------
    1115             :     for (i = 0; i < wiener_win2; i++) {
    1116             :         for (j = i + 1; j < wiener_win2; j++)
    1117             :             H[j * wiener_win2 + i] = H[i * wiener_win2 + j];
    1118             :     }
    1119             : 
    1120             :     // Finally we get the whole matrix.
    1121             :     // Demostration of TAP 3:
    1122             :     //  \lj   00    10    20      01    11    21      02    12    22
    1123             :     // ki-------------------------------------------------------------
    1124             :     // 00| 00X00 00X10 00X20 | 00X01 00X11 00X21 | 00X02 00X12 00X22 |
    1125             :     // 10| 00X10 10x10 10x20 | 10X01 10x11 10x21 | 10X02 10x12 10x22 |
    1126             :     // 20| 00X20 10x20 20x20 | 20X01 20x11 20x21 | 20X02 20x12 20x22 |
    1127             :     //   |-----------------------------------------------------------|
    1128             :     // 01| 00X01 10X01 20X01 | 01x01 01x11 01x21 | 01x02 01x12 01x22 |
    1129             :     // 11| 00X11 10x11 20x11 | 01x11 11x11 11x21 | 11x02 11x12 11x22 |
    1130             :     // 21| 00X21 10x21 20x21 | 01x21 11x21 21x21 | 21x02 21x12 21x22 |
    1131             :     //   |-----------------------------------------------------------|
    1132             :     // 02| 00X02 10X02 20X02 | 01x02 11x02 21x02 | 02x02 02x12 02x22 |
    1133             :     // 12| 00X12 10x12 20x12 | 01x12 11x12 21x12 | 02x12 12x12 12x22 |
    1134             :     // 22| 00X22 10x22 20x22 | 01x22 11x22 21x22 | 02x22 12x22 22x22 |
    1135             :     //   -------------------------------------------------------------
    1136             : }
    1137             : #endif
    1138             : 
    1139           0 : void eb_av1_compute_stats_highbd_c(int32_t wiener_win, const uint8_t *dgd8,
    1140             :     const uint8_t *src8, int32_t h_start, int32_t h_end,
    1141             :     int32_t v_start, int32_t v_end, int32_t dgd_stride,
    1142             :     int32_t src_stride, int64_t *M, int64_t *H,
    1143             :     AomBitDepth bit_depth) {
    1144             :     int32_t i, j, k, l;
    1145           0 :     int32_t Y[WIENER_WIN2] = { 0 };
    1146           0 :     const int32_t wiener_win2 = wiener_win * wiener_win;
    1147           0 :     const int32_t wiener_halfwin = (wiener_win >> 1);
    1148           0 :     const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
    1149           0 :     const uint16_t *dgd = CONVERT_TO_SHORTPTR(dgd8);
    1150             :     uint16_t avg =
    1151           0 :         find_average_highbd(dgd, h_start, h_end, v_start, v_end, dgd_stride);
    1152             : 
    1153           0 :     uint8_t bit_depth_divider = 1;
    1154           0 :     if (bit_depth == AOM_BITS_12)
    1155           0 :         bit_depth_divider = 16;
    1156           0 :     else if (bit_depth == AOM_BITS_10)
    1157           0 :         bit_depth_divider = 4;
    1158             : 
    1159           0 :     memset(M, 0, sizeof(*M) * wiener_win2);
    1160           0 :     memset(H, 0, sizeof(*H) * wiener_win2 * wiener_win2);
    1161           0 :     for (i = v_start; i < v_end; i++) {
    1162           0 :         for (j = h_start; j < h_end; j++) {
    1163           0 :             const int32_t X = (int32_t)src[i * src_stride + j] - (int32_t)avg;
    1164           0 :             int32_t idx = 0;
    1165           0 :             for (k = -wiener_halfwin; k <= wiener_halfwin; k++) {
    1166           0 :                 for (l = -wiener_halfwin; l <= wiener_halfwin; l++) {
    1167           0 :                     Y[idx] = (int32_t)dgd[(i + l) * dgd_stride + (j + k)] - (int32_t)avg;
    1168           0 :                     idx++;
    1169             :                 }
    1170             :             }
    1171             :             assert(idx == wiener_win2);
    1172           0 :             for (k = 0; k < wiener_win2; ++k) {
    1173           0 :                 M[k] += (int64_t)Y[k] * X;
    1174           0 :                 for (l = k; l < wiener_win2; ++l) {
    1175             :                     // H is a symmetric matrix, so we only need to fill out the upper
    1176             :                     // triangle here. We can copy it down to the lower triangle outside
    1177             :                     // the (i, j) loops.
    1178           0 :                     H[k * wiener_win2 + l] += (int64_t)Y[k] * Y[l];
    1179             :                 }
    1180             :             }
    1181             :         }
    1182             :     }
    1183           0 :     for (k = 0; k < wiener_win2; ++k) {
    1184           0 :         M[k] /= bit_depth_divider;
    1185           0 :         H[k * wiener_win2 + k] /= bit_depth_divider;
    1186           0 :         for (l = k + 1; l < wiener_win2; ++l) {
    1187           0 :             H[k * wiener_win2 + l] /= bit_depth_divider;
    1188           0 :             H[l * wiener_win2 + k] = H[k * wiener_win2 + l];
    1189             :         }
    1190             :     }
    1191           0 : }
    1192             : 
    1193     4206720 : static INLINE int32_t wrap_index(int32_t i, int32_t wiener_win) {
    1194     4206720 :     const int32_t wiener_halfwin1 = (wiener_win >> 1) + 1;
    1195     4206720 :     return (i >= wiener_halfwin1 ? wiener_win - 1 - i : i);
    1196             : }
    1197             : 
    1198             : // Solve linear equations to find Wiener filter tap values
    1199             : // Taps are output scaled by WIENER_FILT_STEP
    1200        2400 : static int32_t linsolve_wiener(int32_t n, int64_t *A, int32_t stride, int64_t *b,
    1201             :     int32_t *x) {
    1202        6240 :     for (int32_t k = 0; k < n - 1; k++) {
    1203             :         // Partial pivoting: bring the row with the largest pivot to the top
    1204        9120 :         for (int32_t i = n - 1; i > k; i--) {
    1205             :             // If row i has a better (bigger) pivot than row (i-1), swap them
    1206        5280 :             if (llabs(A[(i - 1) * stride + k]) < llabs(A[i * stride + k])) {
    1207           0 :                 for (int32_t j = 0; j < n; j++) {
    1208           0 :                     const int64_t c = A[i * stride + j];
    1209           0 :                     A[i * stride + j] = A[(i - 1) * stride + j];
    1210           0 :                     A[(i - 1) * stride + j] = c;
    1211             :                 }
    1212           0 :                 const int64_t c = b[i];
    1213           0 :                 b[i] = b[i - 1];
    1214           0 :                 b[i - 1] = c;
    1215             :             }
    1216             :         }
    1217             :         // Forward elimination (convert A to row-echelon form)
    1218        9120 :         for (int32_t i = k; i < n - 1; i++) {
    1219        5280 :             if (A[k * stride + k] == 0) return 0;
    1220        5280 :             const int64_t c = A[(i + 1) * stride + k];
    1221        5280 :             const int64_t cd = A[k * stride + k];
    1222       20160 :             for (int32_t j = 0; j < n; j++)
    1223       14880 :                 A[(i + 1) * stride + j] -= c / 256 * A[k * stride + j] / cd * 256;
    1224        5280 :             b[i + 1] -= c * b[k] / cd;
    1225             :         }
    1226             :     }
    1227             :     // Back-substitution
    1228        8640 :     for (int32_t i = n - 1; i >= 0; i--) {
    1229        6240 :         if (A[i * stride + i] == 0) return 0;
    1230        6240 :         int64_t c = 0;
    1231       11520 :         for (int32_t j = i + 1; j <= n - 1; j++)
    1232        5280 :             c += A[i * stride + j] * x[j] / WIENER_TAP_SCALE_FACTOR;
    1233             :         // Store filter taps x in scaled form.
    1234        6240 :         x[i] = (int32_t)(WIENER_TAP_SCALE_FACTOR * (b[i] - c) / A[i * stride + i]);
    1235             :     }
    1236             : 
    1237        2400 :     return 1;
    1238             : }
    1239             : // Fix vector b, update vector a
    1240        1200 : static void update_a_sep_sym(int32_t wiener_win, int64_t **Mc, int64_t **Hc,
    1241             :     int32_t *a, int32_t *b) {
    1242             :     int32_t i, j;
    1243             :     int32_t S[WIENER_WIN];
    1244             :     int64_t A[WIENER_HALFWIN1], B[WIENER_HALFWIN1 * WIENER_HALFWIN1];
    1245        1200 :     const int32_t wiener_win2 = wiener_win * wiener_win;
    1246        1200 :     const int32_t wiener_halfwin1 = (wiener_win >> 1) + 1;
    1247        1200 :     memset(A, 0, sizeof(A));
    1248        1200 :     memset(B, 0, sizeof(B));
    1249        8640 :     for (i = 0; i < wiener_win; i++) {
    1250       54720 :         for (j = 0; j < wiener_win; ++j) {
    1251       47280 :             const int32_t jj = wrap_index(j, wiener_win);
    1252       47280 :             A[jj] += Mc[i][j] * b[i] / WIENER_TAP_SCALE_FACTOR;
    1253             :         }
    1254             :     }
    1255        8640 :     for (i = 0; i < wiener_win; i++) {
    1256       54720 :         for (j = 0; j < wiener_win; j++) {
    1257             :             int32_t k, l;
    1258      354240 :             for (k = 0; k < wiener_win; ++k) {
    1259     2335680 :                 for (l = 0; l < wiener_win; ++l) {
    1260     2028720 :                     const int32_t kk = wrap_index(k, wiener_win);
    1261     2028720 :                     const int32_t ll = wrap_index(l, wiener_win);
    1262     2028720 :                     B[ll * wiener_halfwin1 + kk] +=
    1263     2028720 :                         Hc[j * wiener_win + i][k * wiener_win2 + l] * b[i] /
    1264     2028720 :                         WIENER_TAP_SCALE_FACTOR * b[j] / WIENER_TAP_SCALE_FACTOR;
    1265             :                 }
    1266             :             }
    1267             :         }
    1268             :     }
    1269             :     // Normalization enforcement in the system of equations itself
    1270             :     assert(wiener_halfwin1 <= WIENER_HALFWIN1);
    1271        1200 :     int64_t A_halfwin1 = A[wiener_halfwin1 - 1];
    1272        4320 :     for (i = 0; i < wiener_halfwin1 - 1; ++i) {
    1273        3120 :         A[i] -= A_halfwin1 * 2 +
    1274        3120 :             B[i * wiener_halfwin1 + wiener_halfwin1 - 1] -
    1275        3120 :             2 * B[(wiener_halfwin1 - 1) * wiener_halfwin1 + (wiener_halfwin1 - 1)];
    1276             :     }
    1277        4320 :     for (i = 0; i < wiener_halfwin1 - 1; ++i) {
    1278       11520 :         for (j = 0; j < wiener_halfwin1 - 1; ++j) {
    1279        8400 :             B[i * wiener_halfwin1 + j] -=
    1280        8400 :                 2 * (B[i * wiener_halfwin1 + (wiener_halfwin1 - 1)] +
    1281        8400 :                     B[(wiener_halfwin1 - 1) * wiener_halfwin1 + j] -
    1282        8400 :                     2 * B[(wiener_halfwin1 - 1) * wiener_halfwin1 +
    1283        8400 :                     (wiener_halfwin1 - 1)]);
    1284             :         }
    1285             :     }
    1286        1200 :     if (linsolve_wiener(wiener_halfwin1 - 1, B, wiener_halfwin1, A, S)) {
    1287        1200 :         S[wiener_halfwin1 - 1] = WIENER_TAP_SCALE_FACTOR;
    1288        4320 :         for (i = wiener_halfwin1; i < wiener_win; ++i) {
    1289        3120 :             S[i] = S[wiener_win - 1 - i];
    1290        3120 :             S[wiener_halfwin1 - 1] -= 2 * S[i];
    1291             :         }
    1292        1200 :         memcpy(a, S, wiener_win * sizeof(*a));
    1293             :     }
    1294        1200 : }
    1295             : 
    1296             : // Fix vector a, update vector b
    1297        1200 : static void update_b_sep_sym(int32_t wiener_win, int64_t **Mc, int64_t **Hc,
    1298             :     int32_t *a, int32_t *b) {
    1299             :     int32_t i, j;
    1300             :     int32_t S[WIENER_WIN];
    1301             :     int64_t A[WIENER_HALFWIN1], B[WIENER_HALFWIN1 * WIENER_HALFWIN1];
    1302        1200 :     const int32_t wiener_win2 = wiener_win * wiener_win;
    1303        1200 :     const int32_t wiener_halfwin1 = (wiener_win >> 1) + 1;
    1304        1200 :     memset(A, 0, sizeof(A));
    1305        1200 :     memset(B, 0, sizeof(B));
    1306        8640 :     for (i = 0; i < wiener_win; i++) {
    1307        7440 :         const int32_t ii = wrap_index(i, wiener_win);
    1308       54720 :         for (j = 0; j < wiener_win; j++)
    1309       47280 :             A[ii] += Mc[i][j] * a[j] / WIENER_TAP_SCALE_FACTOR;
    1310             :     }
    1311             : 
    1312        8640 :     for (i = 0; i < wiener_win; i++) {
    1313       54720 :         for (j = 0; j < wiener_win; j++) {
    1314       47280 :             const int32_t ii = wrap_index(i, wiener_win);
    1315       47280 :             const int32_t jj = wrap_index(j, wiener_win);
    1316             :             int32_t k, l;
    1317      354240 :             for (k = 0; k < wiener_win; ++k) {
    1318     2335680 :                 for (l = 0; l < wiener_win; ++l) {
    1319     2028720 :                     B[jj * wiener_halfwin1 + ii] +=
    1320     2028720 :                         Hc[i * wiener_win + j][k * wiener_win2 + l] * a[k] /
    1321     2028720 :                         WIENER_TAP_SCALE_FACTOR * a[l] / WIENER_TAP_SCALE_FACTOR;
    1322             :                 }
    1323             :             }
    1324             :         }
    1325             :     }
    1326             :     // Normalization enforcement in the system of equations itself
    1327        1200 :     int64_t A_halfwin1 = A[wiener_halfwin1 - 1];
    1328        4320 :     for (i = 0; i < wiener_halfwin1 - 1; ++i) {
    1329        3120 :         A[i] -= A_halfwin1 * 2 +
    1330        3120 :             B[i * wiener_halfwin1 + wiener_halfwin1 - 1] -
    1331        3120 :             2 * B[(wiener_halfwin1 - 1) * wiener_halfwin1 + (wiener_halfwin1 - 1)];
    1332             :     }
    1333        4320 :     for (i = 0; i < wiener_halfwin1 - 1; ++i) {
    1334       11520 :         for (j = 0; j < wiener_halfwin1 - 1; ++j) {
    1335        8400 :             B[i * wiener_halfwin1 + j] -=
    1336        8400 :                 2 * (B[i * wiener_halfwin1 + (wiener_halfwin1 - 1)] +
    1337        8400 :                     B[(wiener_halfwin1 - 1) * wiener_halfwin1 + j] -
    1338        8400 :                     2 * B[(wiener_halfwin1 - 1) * wiener_halfwin1 +
    1339        8400 :                     (wiener_halfwin1 - 1)]);
    1340             :         }
    1341             :     }
    1342        1200 :     if (linsolve_wiener(wiener_halfwin1 - 1, B, wiener_halfwin1, A, S)) {
    1343        1200 :         S[wiener_halfwin1 - 1] = WIENER_TAP_SCALE_FACTOR;
    1344        4320 :         for (i = wiener_halfwin1; i < wiener_win; ++i) {
    1345        3120 :             S[i] = S[wiener_win - 1 - i];
    1346        3120 :             S[wiener_halfwin1 - 1] -= 2 * S[i];
    1347             :         }
    1348        1200 :         memcpy(b, S, wiener_win * sizeof(*b));
    1349             :     }
    1350        1200 : }
    1351             : 
    1352         300 : static int32_t wiener_decompose_sep_sym(int32_t wiener_win, int64_t *M, int64_t *H,
    1353             :     int32_t *a, int32_t *b) {
    1354             :     static const int32_t init_filt[WIENER_WIN] = {
    1355             :         WIENER_FILT_TAP0_MIDV, WIENER_FILT_TAP1_MIDV, WIENER_FILT_TAP2_MIDV,
    1356             :         WIENER_FILT_TAP3_MIDV, WIENER_FILT_TAP2_MIDV, WIENER_FILT_TAP1_MIDV,
    1357             :         WIENER_FILT_TAP0_MIDV,
    1358             :     };
    1359             :     int64_t *Hc[WIENER_WIN2];
    1360             :     int64_t *Mc[WIENER_WIN];
    1361             :     int32_t i, j, iter;
    1362         300 :     const int32_t plane_off = (WIENER_WIN - wiener_win) >> 1;
    1363         300 :     const int32_t wiener_win2 = wiener_win * wiener_win;
    1364        2160 :     for (i = 0; i < wiener_win; i++) {
    1365        1860 :         a[i] = b[i] =
    1366        1860 :             WIENER_TAP_SCALE_FACTOR / WIENER_FILT_STEP * init_filt[i + plane_off];
    1367             :     }
    1368        2160 :     for (i = 0; i < wiener_win; i++) {
    1369        1860 :         Mc[i] = M + i * wiener_win;
    1370       13680 :         for (j = 0; j < wiener_win; j++) {
    1371       11820 :             Hc[i * wiener_win + j] =
    1372       11820 :                 H + i * wiener_win * wiener_win2 + j * wiener_win;
    1373             :         }
    1374             :     }
    1375             : 
    1376         300 :     iter = 1;
    1377        1500 :     while (iter < NUM_WIENER_ITERS) {
    1378        1200 :         update_a_sep_sym(wiener_win, Mc, Hc, a, b);
    1379        1200 :         update_b_sep_sym(wiener_win, Mc, Hc, a, b);
    1380        1200 :         iter++;
    1381             :     }
    1382         300 :     return 1;
    1383             : }
    1384         300 : static int64_t compute_score(int32_t wiener_win, int64_t *M, int64_t *H,
    1385             :     InterpKernel vfilt, InterpKernel hfilt) {
    1386             :     int32_t ab[WIENER_WIN * WIENER_WIN];
    1387             :     int16_t a[WIENER_WIN], b[WIENER_WIN];
    1388         300 :     int64_t P = 0, Q = 0;
    1389         300 :     int64_t iP = 0, iQ = 0;
    1390             :     int64_t Score, iScore;
    1391             :     int32_t i, k, l;
    1392         300 :     const int32_t plane_off = (WIENER_WIN - wiener_win) >> 1;
    1393         300 :     const int32_t wiener_win2 = wiener_win * wiener_win;
    1394             : 
    1395         300 :     aom_clear_system_state();
    1396             : 
    1397         300 :     a[WIENER_HALFWIN] = b[WIENER_HALFWIN] = WIENER_FILT_STEP;
    1398        1200 :     for (i = 0; i < WIENER_HALFWIN; ++i) {
    1399         900 :         a[i] = a[WIENER_WIN - i - 1] = vfilt[i];
    1400         900 :         b[i] = b[WIENER_WIN - i - 1] = hfilt[i];
    1401         900 :         a[WIENER_HALFWIN] -= 2 * a[i];
    1402         900 :         b[WIENER_HALFWIN] -= 2 * b[i];
    1403             :     }
    1404         300 :     memset(ab, 0, sizeof(ab));
    1405        2160 :     for (k = 0; k < wiener_win; ++k) {
    1406       13680 :         for (l = 0; l < wiener_win; ++l)
    1407       11820 :             ab[k * wiener_win + l] = a[l + plane_off] * b[k + plane_off];
    1408             :     }
    1409       12120 :     for (k = 0; k < wiener_win2; ++k) {
    1410       11820 :         P += ab[k] * M[k] / WIENER_FILT_STEP / WIENER_FILT_STEP;
    1411      519000 :         for (l = 0; l < wiener_win2; ++l) {
    1412      507180 :             Q += ab[k] * H[k * wiener_win2 + l] * ab[l] / WIENER_FILT_STEP /
    1413      507180 :                 WIENER_FILT_STEP / WIENER_FILT_STEP / WIENER_FILT_STEP;
    1414             :         }
    1415             :     }
    1416         300 :     Score = Q - 2 * P;
    1417             : 
    1418         300 :     iP = M[wiener_win2 >> 1];
    1419         300 :     iQ = H[(wiener_win2 >> 1) * wiener_win2 + (wiener_win2 >> 1)];
    1420         300 :     iScore = iQ - 2 * iP;
    1421             : 
    1422         300 :     return Score - iScore;
    1423             : }
    1424             : 
    1425         600 : static void finalize_sym_filter(int32_t wiener_win, int32_t *f, InterpKernel fi) {
    1426             :     int32_t i;
    1427         600 :     const int32_t wiener_halfwin = (wiener_win >> 1);
    1428             : 
    1429        2160 :     for (i = 0; i < wiener_halfwin; ++i) {
    1430        1560 :         const int64_t dividend = f[i] * WIENER_FILT_STEP;
    1431        1560 :         const int64_t divisor = WIENER_TAP_SCALE_FACTOR;
    1432             :         // Perform this division with proper rounding rather than truncation
    1433        1560 :         if (dividend < 0)
    1434         823 :             fi[i] = (int16_t)((dividend - (divisor / 2)) / divisor);
    1435             :         else
    1436         737 :             fi[i] = (int16_t)((dividend + (divisor / 2)) / divisor);
    1437             :     }
    1438             :     // Specialize for 7-tap filter
    1439         600 :     if (wiener_win == WIENER_WIN) {
    1440         360 :         fi[0] = CLIP(fi[0], WIENER_FILT_TAP0_MINV, WIENER_FILT_TAP0_MAXV);
    1441         360 :         fi[1] = CLIP(fi[1], WIENER_FILT_TAP1_MINV, WIENER_FILT_TAP1_MAXV);
    1442         360 :         fi[2] = CLIP(fi[2], WIENER_FILT_TAP2_MINV, WIENER_FILT_TAP2_MAXV);
    1443             :     }
    1444             :     else {
    1445         240 :         fi[2] = CLIP(fi[1], WIENER_FILT_TAP2_MINV, WIENER_FILT_TAP2_MAXV);
    1446         240 :         fi[1] = CLIP(fi[0], WIENER_FILT_TAP1_MINV, WIENER_FILT_TAP1_MAXV);
    1447         240 :         fi[0] = 0;
    1448             :     }
    1449             :     // Satisfy filter constraints
    1450         600 :     fi[WIENER_WIN - 1] = fi[0];
    1451         600 :     fi[WIENER_WIN - 2] = fi[1];
    1452         600 :     fi[WIENER_WIN - 3] = fi[2];
    1453             :     // The central element has an implicit +WIENER_FILT_STEP
    1454         600 :     fi[3] = -2 * (fi[0] + fi[1] + fi[2]);
    1455         600 : }
    1456             : 
    1457         391 : static int32_t count_wiener_bits(int32_t wiener_win, WienerInfo *wiener_info,
    1458             :     WienerInfo *ref_wiener_info) {
    1459         391 :     int32_t bits = 0;
    1460         391 :     if (wiener_win == WIENER_WIN)
    1461         271 :         bits += eb_aom_count_primitive_refsubexpfin(
    1462             :             WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
    1463             :             WIENER_FILT_TAP0_SUBEXP_K,
    1464         271 :             ref_wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV,
    1465         271 :             wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV);
    1466         782 :     bits += eb_aom_count_primitive_refsubexpfin(
    1467             :         WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
    1468             :         WIENER_FILT_TAP1_SUBEXP_K,
    1469         391 :         ref_wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV,
    1470         391 :         wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV);
    1471         782 :     bits += eb_aom_count_primitive_refsubexpfin(
    1472             :         WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
    1473             :         WIENER_FILT_TAP2_SUBEXP_K,
    1474         391 :         ref_wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV,
    1475         391 :         wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV);
    1476         391 :     if (wiener_win == WIENER_WIN)
    1477         271 :         bits += eb_aom_count_primitive_refsubexpfin(
    1478             :             WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
    1479             :             WIENER_FILT_TAP0_SUBEXP_K,
    1480         271 :             ref_wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV,
    1481         271 :             wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV);
    1482         782 :     bits += eb_aom_count_primitive_refsubexpfin(
    1483             :         WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
    1484             :         WIENER_FILT_TAP1_SUBEXP_K,
    1485         391 :         ref_wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV,
    1486         391 :         wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV);
    1487         782 :     bits += eb_aom_count_primitive_refsubexpfin(
    1488             :         WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
    1489             :         WIENER_FILT_TAP2_SUBEXP_K,
    1490         391 :         ref_wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV,
    1491         391 :         wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV);
    1492         391 :     return bits;
    1493             : }
    1494             : 
    1495             : #define USE_WIENER_REFINEMENT_SEARCH 1
    1496           0 : static int64_t finer_tile_search_wiener(const RestSearchCtxt *rsc,
    1497             :     const RestorationTileLimits *limits,
    1498             :     const AV1PixelRect *tile,
    1499             :     RestorationUnitInfo *rui,
    1500             :     int32_t wiener_win) {
    1501           0 :     const int32_t plane_off = (WIENER_WIN - wiener_win) >> 1;
    1502           0 :     int64_t err = try_restoration_unit(rsc, limits, tile, rui);
    1503             : #if USE_WIENER_REFINEMENT_SEARCH
    1504             :     int64_t err2;
    1505           0 :     int32_t tap_min[] = { WIENER_FILT_TAP0_MINV, WIENER_FILT_TAP1_MINV,
    1506             :                       WIENER_FILT_TAP2_MINV };
    1507           0 :     int32_t tap_max[] = { WIENER_FILT_TAP0_MAXV, WIENER_FILT_TAP1_MAXV,
    1508             :                       WIENER_FILT_TAP2_MAXV };
    1509             : 
    1510           0 :     WienerInfo *plane_wiener = &rui->wiener_info;
    1511             : 
    1512             :     // printf("err  pre = %"PRId64"\n", err);
    1513           0 :     const int32_t start_step = 4;
    1514           0 :     for (int32_t s = start_step; s >= 1; s >>= 1) {
    1515           0 :         for (int32_t p = plane_off; p < WIENER_HALFWIN; ++p) {
    1516           0 :             int32_t skip = 0;
    1517             :             do {
    1518           0 :                 if (plane_wiener->hfilter[p] - s >= tap_min[p]) {
    1519           0 :                     plane_wiener->hfilter[p] -= (int16_t)s;
    1520           0 :                     plane_wiener->hfilter[WIENER_WIN - p - 1] -= (int16_t)s;
    1521           0 :                     plane_wiener->hfilter[WIENER_HALFWIN] += 2 * (int16_t)s;
    1522           0 :                     err2 = try_restoration_unit(rsc, limits, tile, rui);
    1523           0 :                     if (err2 > err) {
    1524           0 :                         plane_wiener->hfilter[p] += (int16_t)s;
    1525           0 :                         plane_wiener->hfilter[WIENER_WIN - p - 1] += (int16_t)s;
    1526           0 :                         plane_wiener->hfilter[WIENER_HALFWIN] -= 2 * (int16_t)s;
    1527             :                     }
    1528             :                     else {
    1529           0 :                         err = err2;
    1530           0 :                         skip = 1;
    1531             :                         // At the highest step size continue moving in the same direction
    1532           0 :                         if (s == start_step) continue;
    1533             :                     }
    1534             :                 }
    1535           0 :                 break;
    1536             :             } while (1);
    1537           0 :             if (skip) break;
    1538             :             do {
    1539           0 :                 if (plane_wiener->hfilter[p] + s <= tap_max[p]) {
    1540           0 :                     plane_wiener->hfilter[p] += (int16_t)s;
    1541           0 :                     plane_wiener->hfilter[WIENER_WIN - p - 1] += (int16_t)s;
    1542           0 :                     plane_wiener->hfilter[WIENER_HALFWIN] -= 2 * (int16_t)s;
    1543           0 :                     err2 = try_restoration_unit(rsc, limits, tile, rui);
    1544           0 :                     if (err2 > err) {
    1545           0 :                         plane_wiener->hfilter[p] -= (int16_t)s;
    1546           0 :                         plane_wiener->hfilter[WIENER_WIN - p - 1] -= (int16_t)s;
    1547           0 :                         plane_wiener->hfilter[WIENER_HALFWIN] += 2 * (int16_t)s;
    1548             :                     }
    1549             :                     else {
    1550           0 :                         err = err2;
    1551             :                         // At the highest step size continue moving in the same direction
    1552           0 :                         if (s == start_step) continue;
    1553             :                     }
    1554             :                 }
    1555           0 :                 break;
    1556             :             } while (1);
    1557             :         }
    1558           0 :         for (int32_t p = plane_off; p < WIENER_HALFWIN; ++p) {
    1559           0 :             int32_t skip = 0;
    1560             :             do {
    1561           0 :                 if (plane_wiener->vfilter[p] - s >= tap_min[p]) {
    1562           0 :                     plane_wiener->vfilter[p] -= (int16_t)s;
    1563           0 :                     plane_wiener->vfilter[WIENER_WIN - p - 1] -= (int16_t)s;
    1564           0 :                     plane_wiener->vfilter[WIENER_HALFWIN] += 2 * (int16_t)s;
    1565           0 :                     err2 = try_restoration_unit(rsc, limits, tile, rui);
    1566           0 :                     if (err2 > err) {
    1567           0 :                         plane_wiener->vfilter[p] += (int16_t)s;
    1568           0 :                         plane_wiener->vfilter[WIENER_WIN - p - 1] += (int16_t)s;
    1569           0 :                         plane_wiener->vfilter[WIENER_HALFWIN] -= 2 * (int16_t)s;
    1570             :                     }
    1571             :                     else {
    1572           0 :                         err = err2;
    1573           0 :                         skip = 1;
    1574             :                         // At the highest step size continue moving in the same direction
    1575           0 :                         if (s == start_step) continue;
    1576             :                     }
    1577             :                 }
    1578           0 :                 break;
    1579             :             } while (1);
    1580           0 :             if (skip) break;
    1581             :             do {
    1582           0 :                 if (plane_wiener->vfilter[p] + s <= tap_max[p]) {
    1583           0 :                     plane_wiener->vfilter[p] += (int16_t)s;
    1584           0 :                     plane_wiener->vfilter[WIENER_WIN - p - 1] += (int16_t)s;
    1585           0 :                     plane_wiener->vfilter[WIENER_HALFWIN] -= 2 * (int16_t)s;
    1586           0 :                     err2 = try_restoration_unit(rsc, limits, tile, rui);
    1587           0 :                     if (err2 > err) {
    1588           0 :                         plane_wiener->vfilter[p] -= (int16_t)s;
    1589           0 :                         plane_wiener->vfilter[WIENER_WIN - p - 1] -= (int16_t)s;
    1590           0 :                         plane_wiener->vfilter[WIENER_HALFWIN] += 2 * (int16_t)s;
    1591             :                     }
    1592             :                     else {
    1593           0 :                         err = err2;
    1594             :                         // At the highest step size continue moving in the same direction
    1595           0 :                         if (s == start_step) continue;
    1596             :                     }
    1597             :                 }
    1598           0 :                 break;
    1599             :             } while (1);
    1600             :         }
    1601             :     }
    1602             :     // printf("err post = %"PRId64"\n", err);
    1603             : #endif  // USE_WIENER_REFINEMENT_SEARCH
    1604           0 :     return err;
    1605             : }
    1606         295 : static int64_t finer_tile_search_wiener_seg(const RestSearchCtxt *rsc,
    1607             :     const RestorationTileLimits *limits,
    1608             :     const AV1PixelRect *tile,
    1609             :     RestorationUnitInfo *rui,
    1610             :     int32_t wiener_win) {
    1611         295 :     const int32_t plane_off = (WIENER_WIN - wiener_win) >> 1;
    1612         295 :     int64_t err = try_restoration_unit_seg(rsc, limits, tile, rui);
    1613             : #if USE_WIENER_REFINEMENT_SEARCH
    1614             :     int64_t err2;
    1615         295 :     int32_t tap_min[] = { WIENER_FILT_TAP0_MINV, WIENER_FILT_TAP1_MINV,
    1616             :                       WIENER_FILT_TAP2_MINV };
    1617         295 :     int32_t tap_max[] = { WIENER_FILT_TAP0_MAXV, WIENER_FILT_TAP1_MAXV,
    1618             :                       WIENER_FILT_TAP2_MAXV };
    1619             : 
    1620         295 :     WienerInfo *plane_wiener = &rui->wiener_info;
    1621             : 
    1622             :     // printf("err  pre = %"PRId64"\n", err);
    1623         295 :     const int32_t start_step = 4;
    1624        1180 :     for (int32_t s = start_step; s >= 1; s >>= 1) {
    1625        2901 :         for (int32_t p = plane_off; p < WIENER_HALFWIN; ++p) {
    1626        2254 :             int32_t skip = 0;
    1627             :             do {
    1628        2324 :                 if (plane_wiener->hfilter[p] - s >= tap_min[p]) {
    1629        2311 :                     plane_wiener->hfilter[p] -= (int16_t)s;
    1630        2311 :                     plane_wiener->hfilter[WIENER_WIN - p - 1] -= (int16_t)s;
    1631        2311 :                     plane_wiener->hfilter[WIENER_HALFWIN] += 2 * (int16_t)s;
    1632        2311 :                     err2 = try_restoration_unit_seg(rsc, limits, tile, rui);
    1633        2311 :                     if (err2 > err) {
    1634        2068 :                         plane_wiener->hfilter[p] += (int16_t)s;
    1635        2068 :                         plane_wiener->hfilter[WIENER_WIN - p - 1] += (int16_t)s;
    1636        2068 :                         plane_wiener->hfilter[WIENER_HALFWIN] -= 2 * (int16_t)s;
    1637             :                     }
    1638             :                     else {
    1639         243 :                         err = err2;
    1640         243 :                         skip = 1;
    1641             :                         // At the highest step size continue moving in the same direction
    1642         243 :                         if (s == start_step) continue;
    1643             :                     }
    1644             :                 }
    1645        2254 :                 break;
    1646             :             } while (1);
    1647        2254 :             if (skip) break;
    1648             :             do {
    1649        2051 :                 if (plane_wiener->hfilter[p] + s <= tap_max[p]) {
    1650        2043 :                     plane_wiener->hfilter[p] += (int16_t)s;
    1651        2043 :                     plane_wiener->hfilter[WIENER_WIN - p - 1] += (int16_t)s;
    1652        2043 :                     plane_wiener->hfilter[WIENER_HALFWIN] -= 2 * (int16_t)s;
    1653        2043 :                     err2 = try_restoration_unit_seg(rsc, limits, tile, rui);
    1654        2043 :                     if (err2 > err) {
    1655        1873 :                         plane_wiener->hfilter[p] -= (int16_t)s;
    1656        1873 :                         plane_wiener->hfilter[WIENER_WIN - p - 1] -= (int16_t)s;
    1657        1873 :                         plane_wiener->hfilter[WIENER_HALFWIN] += 2 * (int16_t)s;
    1658             :                     }
    1659             :                     else {
    1660         170 :                         err = err2;
    1661             :                         // At the highest step size continue moving in the same direction
    1662         170 :                         if (s == start_step) continue;
    1663             :                     }
    1664             :                 }
    1665        2016 :                 break;
    1666             :             } while (1);
    1667             :         }
    1668        3060 :         for (int32_t p = plane_off; p < WIENER_HALFWIN; ++p) {
    1669        2280 :             int32_t skip = 0;
    1670             :             do {
    1671        2287 :                 if (plane_wiener->vfilter[p] - s >= tap_min[p]) {
    1672        2248 :                     plane_wiener->vfilter[p] -= (int16_t)s;
    1673        2248 :                     plane_wiener->vfilter[WIENER_WIN - p - 1] -= (int16_t)s;
    1674        2248 :                     plane_wiener->vfilter[WIENER_HALFWIN] += 2 * (int16_t)s;
    1675        2248 :                     err2 = try_restoration_unit_seg(rsc, limits, tile, rui);
    1676        2248 :                     if (err2 > err) {
    1677        2143 :                         plane_wiener->vfilter[p] += (int16_t)s;
    1678        2143 :                         plane_wiener->vfilter[WIENER_WIN - p - 1] += (int16_t)s;
    1679        2143 :                         plane_wiener->vfilter[WIENER_HALFWIN] -= 2 * (int16_t)s;
    1680             :                     }
    1681             :                     else {
    1682         105 :                         err = err2;
    1683         105 :                         skip = 1;
    1684             :                         // At the highest step size continue moving in the same direction
    1685         105 :                         if (s == start_step) continue;
    1686             :                     }
    1687             :                 }
    1688        2280 :                 break;
    1689             :             } while (1);
    1690        2280 :             if (skip) break;
    1691             :             do {
    1692        2222 :                 if (plane_wiener->vfilter[p] + s <= tap_max[p]) {
    1693        2148 :                     plane_wiener->vfilter[p] += (int16_t)s;
    1694        2148 :                     plane_wiener->vfilter[WIENER_WIN - p - 1] += (int16_t)s;
    1695        2148 :                     plane_wiener->vfilter[WIENER_HALFWIN] -= 2 * (int16_t)s;
    1696        2148 :                     err2 = try_restoration_unit_seg(rsc, limits, tile, rui);
    1697        2148 :                     if (err2 > err) {
    1698        1994 :                         plane_wiener->vfilter[p] -= (int16_t)s;
    1699        1994 :                         plane_wiener->vfilter[WIENER_WIN - p - 1] -= (int16_t)s;
    1700        1994 :                         plane_wiener->vfilter[WIENER_HALFWIN] += 2 * (int16_t)s;
    1701             :                     }
    1702             :                     else {
    1703         154 :                         err = err2;
    1704             :                         // At the highest step size continue moving in the same direction
    1705         154 :                         if (s == start_step) continue;
    1706             :                     }
    1707             :                 }
    1708        2175 :                 break;
    1709             :             } while (1);
    1710             :         }
    1711             :     }
    1712             :     // printf("err post = %"PRId64"\n", err);
    1713             : #endif  // USE_WIENER_REFINEMENT_SEARCH
    1714         295 :     return err;
    1715             : }
    1716             : 
    1717           0 : static void search_wiener(const RestorationTileLimits *limits,
    1718             :     const AV1PixelRect *tile_rect, int32_t rest_unit_idx,
    1719             :     void *priv) {
    1720           0 :     RestSearchCtxt *rsc = (RestSearchCtxt *)priv;
    1721           0 :     RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx];
    1722           0 :     const Av1Common *const cm = rsc->cm;
    1723           0 :     int32_t wn_luma = cm->wn_filter_mode == 1 ? WIENER_WIN_3TAP : cm->wn_filter_mode == 2 ? WIENER_WIN_CHROMA : WIENER_WIN;
    1724           0 :     const int32_t wiener_win = cm->wn_filter_mode == 1 ? WIENER_WIN_3TAP :
    1725           0 :         (rsc->plane == AOM_PLANE_Y) ? wn_luma : WIENER_WIN_CHROMA;
    1726             :     int64_t M[WIENER_WIN2];
    1727             :     int64_t H[WIENER_WIN2 * WIENER_WIN2];
    1728             :     int32_t vfilterd[WIENER_WIN], hfilterd[WIENER_WIN];
    1729             : 
    1730           0 :     if (cm->use_highbitdepth)
    1731             :     {
    1732           0 :         if (rsc->plane == AOM_PLANE_Y) {
    1733           0 :             eb_av1_compute_stats_highbd(wiener_win, rsc->dgd_buffer, rsc->src_buffer,
    1734             :                 limits->h_start, limits->h_end, limits->v_start,
    1735             :                 limits->v_end, rsc->dgd_stride, rsc->src_stride, M,
    1736           0 :                 H, (AomBitDepth)cm->bit_depth);
    1737             :         }
    1738             :         else {
    1739           0 :             eb_av1_compute_stats_highbd(wiener_win, rsc->dgd_buffer, rsc->src_buffer,
    1740             :                 limits->h_start, limits->h_end, limits->v_start,
    1741             :                 limits->v_end, rsc->dgd_stride, rsc->src_stride, M,
    1742           0 :                 H, (AomBitDepth)cm->bit_depth);
    1743             :         }
    1744             :     }
    1745             :     else {
    1746           0 :         eb_av1_compute_stats(wiener_win, rsc->dgd_buffer, rsc->src_buffer, limits->h_start,
    1747             :             limits->h_end, limits->v_start, limits->v_end,
    1748             :             rsc->dgd_stride, rsc->src_stride, M, H);
    1749             :     }
    1750             : 
    1751           0 :     const Macroblock *const x = rsc->x;
    1752           0 :     const int64_t bits_none = x->wiener_restore_cost[0];
    1753             : 
    1754           0 :     if (!wiener_decompose_sep_sym(wiener_win, M, H, vfilterd, hfilterd)) {
    1755           0 :         rsc->bits += bits_none;
    1756           0 :         rsc->sse += rusi->sse[RESTORE_NONE];
    1757           0 :         rusi->best_rtype[RESTORE_WIENER - 1] = RESTORE_NONE;
    1758           0 :         rusi->sse[RESTORE_WIENER] = INT64_MAX;
    1759           0 :         return;
    1760             :     }
    1761             : 
    1762             :     RestorationUnitInfo rui;
    1763           0 :     memset(&rui, 0, sizeof(rui));
    1764           0 :     rui.restoration_type = RESTORE_WIENER;
    1765           0 :     finalize_sym_filter(wiener_win, vfilterd, rui.wiener_info.vfilter);
    1766           0 :     finalize_sym_filter(wiener_win, hfilterd, rui.wiener_info.hfilter);
    1767             : 
    1768             :     // Filter score computes the value of the function x'*A*x - x'*b for the
    1769             :     // learned filter and compares it against identity filer. If there is no
    1770             :     // reduction in the function, the filter is reverted back to identity
    1771           0 :     if (compute_score(wiener_win, M, H, rui.wiener_info.vfilter,
    1772             :         rui.wiener_info.hfilter) > 0) {
    1773           0 :         rsc->bits += bits_none;
    1774           0 :         rsc->sse += rusi->sse[RESTORE_NONE];
    1775           0 :         rusi->best_rtype[RESTORE_WIENER - 1] = RESTORE_NONE;
    1776           0 :         rusi->sse[RESTORE_WIENER] = INT64_MAX;
    1777           0 :         return;
    1778             :     }
    1779             : 
    1780           0 :     aom_clear_system_state();
    1781             : 
    1782           0 :     rusi->sse[RESTORE_WIENER] =
    1783           0 :         finer_tile_search_wiener(rsc, limits, tile_rect, &rui, wiener_win);
    1784           0 :     rusi->wiener = rui.wiener_info;
    1785             : 
    1786             :     if (wiener_win != WIENER_WIN) {
    1787             :         assert(rui.wiener_info.vfilter[0] == 0 &&
    1788             :             rui.wiener_info.vfilter[WIENER_WIN - 1] == 0);
    1789             :         assert(rui.wiener_info.hfilter[0] == 0 &&
    1790             :             rui.wiener_info.hfilter[WIENER_WIN - 1] == 0);
    1791             :     }
    1792             : 
    1793           0 :     const int64_t bits_wiener =
    1794           0 :         x->wiener_restore_cost[1] +
    1795           0 :         (count_wiener_bits(wiener_win, &rusi->wiener, &rsc->wiener)
    1796           0 :             << AV1_PROB_COST_SHIFT);
    1797             : 
    1798           0 :     double cost_none =
    1799           0 :         RDCOST_DBL(x->rdmult, bits_none >> 4, rusi->sse[RESTORE_NONE]);
    1800           0 :     double cost_wiener =
    1801           0 :         RDCOST_DBL(x->rdmult, bits_wiener >> 4, rusi->sse[RESTORE_WIENER]);
    1802             : 
    1803           0 :     RestorationType rtype =
    1804           0 :         (cost_wiener < cost_none) ? RESTORE_WIENER : RESTORE_NONE;
    1805           0 :     rusi->best_rtype[RESTORE_WIENER - 1] = rtype;
    1806             : 
    1807           0 :     rsc->sse += rusi->sse[rtype];
    1808           0 :     rsc->bits += (cost_wiener < cost_none) ? bits_wiener : bits_none;
    1809           0 :     if (cost_wiener < cost_none) rsc->wiener = rusi->wiener;
    1810             : }
    1811             : 
    1812           0 : static void search_norestore(const RestorationTileLimits *limits,
    1813             :     const AV1PixelRect *tile_rect, int32_t rest_unit_idx,
    1814             :     void *priv) {
    1815             :     (void)tile_rect;
    1816             : 
    1817           0 :     RestSearchCtxt *rsc = (RestSearchCtxt *)priv;
    1818           0 :     RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx];
    1819             : 
    1820           0 :     const int32_t highbd = rsc->cm->use_highbitdepth;
    1821           0 :     rusi->sse[RESTORE_NONE] = sse_restoration_unit(
    1822           0 :         limits, rsc->src, rsc->cm->frame_to_show, rsc->plane, highbd);
    1823             : 
    1824           0 :     rsc->sse += rusi->sse[RESTORE_NONE];
    1825           0 : }
    1826             : 
    1827         180 : static void search_switchable(const RestorationTileLimits *limits,
    1828             :     const AV1PixelRect *tile_rect, int32_t rest_unit_idx,
    1829             :     void *priv) {
    1830             :     (void)limits;
    1831             :     (void)tile_rect;
    1832         180 :     RestSearchCtxt *rsc = (RestSearchCtxt *)priv;
    1833         180 :     RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx];
    1834             : 
    1835         180 :     const Macroblock *const x = rsc->x;
    1836             : 
    1837         180 :     const int32_t wiener_win =
    1838         180 :         (rsc->plane == AOM_PLANE_Y) ? WIENER_WIN : WIENER_WIN_CHROMA;
    1839             : 
    1840         180 :     double best_cost = 0;
    1841         180 :     int64_t best_bits = 0;
    1842         180 :     RestorationType best_rtype = RESTORE_NONE;
    1843             : 
    1844             :     //CHKN for (RestorationType r = 0; r < RESTORE_SWITCHABLE_TYPES; ++r) {
    1845         720 :     for (int32_t restType = 0; restType < RESTORE_SWITCHABLE_TYPES; ++restType) {
    1846         540 :         RestorationType r = (RestorationType)restType;
    1847             : 
    1848             :         // Check for the condition that wiener or sgrproj search could not
    1849             :         // find a solution or the solution was worse than RESTORE_NONE.
    1850             :         // In either case the best_rtype will be set as RESTORE_NONE. These
    1851             :         // should be skipped from the test below.
    1852         540 :         if (r > RESTORE_NONE)
    1853         360 :             if (rusi->best_rtype[r - 1] == RESTORE_NONE) continue;
    1854         390 :         const int64_t sse = rusi->sse[r];
    1855         390 :         int64_t coeff_pcost = 0;
    1856         390 :         switch (r) {
    1857         180 :         case RESTORE_NONE: coeff_pcost = 0; break;
    1858          96 :         case RESTORE_WIENER:
    1859          96 :             coeff_pcost =
    1860          96 :                 count_wiener_bits(wiener_win, &rusi->wiener, &rsc->wiener);
    1861          96 :             break;
    1862         114 :         case RESTORE_SGRPROJ:
    1863         114 :             coeff_pcost = count_sgrproj_bits(&rusi->sgrproj, &rsc->sgrproj);
    1864         114 :             break;
    1865           0 :         default: assert(0); break;
    1866             :         }
    1867         390 :         const int64_t coeff_bits = coeff_pcost << AV1_PROB_COST_SHIFT;
    1868         390 :         const int64_t bits = x->switchable_restore_cost[r] + coeff_bits;
    1869         390 :         double cost = RDCOST_DBL(x->rdmult, bits >> 4, sse);
    1870         390 :         if (r == 0 || cost < best_cost) {
    1871         335 :             best_cost = cost;
    1872         335 :             best_bits = bits;
    1873         335 :             best_rtype = r;
    1874             :         }
    1875             :     }
    1876             : 
    1877         180 :     rusi->best_rtype[RESTORE_SWITCHABLE - 1] = best_rtype;
    1878             : 
    1879         180 :     rsc->sse += rusi->sse[best_rtype];
    1880         180 :     rsc->bits += best_bits;
    1881         180 :     if (best_rtype == RESTORE_WIENER) rsc->wiener = rusi->wiener;
    1882         180 :     if (best_rtype == RESTORE_SGRPROJ) rsc->sgrproj = rusi->sgrproj;
    1883         180 : }
    1884             : 
    1885         192 : static void copy_unit_info(RestorationType frame_rtype,
    1886             :     const RestUnitSearchInfo *rusi,
    1887             :     RestorationUnitInfo *rui) {
    1888         192 :     if (frame_rtype >= 1)
    1889         192 :         rui->restoration_type = rusi->best_rtype[frame_rtype - 1];
    1890         192 :     if (rui->restoration_type == RESTORE_WIENER)
    1891          65 :         rui->wiener_info = rusi->wiener;
    1892             :     else
    1893         127 :         rui->sgrproj_info = rusi->sgrproj;
    1894         192 : }
    1895             : 
    1896           0 : static double search_rest_type(RestSearchCtxt *rsc, RestorationType rtype)
    1897             : {
    1898             :     static const RestUnitVisitor funs[RESTORE_TYPES] = {
    1899             :     search_norestore, search_wiener, search_sgrproj, search_switchable
    1900             :     };
    1901             : 
    1902           0 :     reset_rsc(rsc);
    1903             : 
    1904           0 :     av1_foreach_rest_unit_in_frame(rsc->cm, rsc->plane, rsc_on_tile, funs[rtype], rsc);
    1905             : 
    1906           0 :     return RDCOST_DBL(rsc->x->rdmult, rsc->bits >> 4, rsc->sse);
    1907             : }
    1908             : 
    1909         120 : static int32_t rest_tiles_in_plane(const Av1Common *cm, int32_t plane) {
    1910         120 :     const RestorationInfo *rsi = &cm->rst_info[plane];
    1911         120 :     return rsi->units_per_tile;
    1912             : }
    1913             : 
    1914             : void *eb_aom_memalign(size_t align, size_t size);
    1915             : void eb_aom_free(void *memblk);
    1916             : 
    1917           0 : void eb_av1_pick_filter_restoration(const Yv12BufferConfig *src, Yv12BufferConfig * trial_frame_rst /*Av1Comp *cpi*/, Macroblock *x, Av1Common *const cm) {
    1918             :     //CHKN Av1Common *const cm = &cpi->common;
    1919           0 :     const int32_t num_planes = 3;// av1_num_planes(cm);
    1920             :    // assert(!cm->all_lossless);
    1921           0 :     RestorationType force_restore_type_d = (cm->wn_filter_mode) ? RESTORE_TYPES : RESTORE_SGRPROJ;
    1922             : 
    1923             :     int32_t ntiles[2];
    1924           0 :     for (int32_t is_uv = 0; is_uv < 2; ++is_uv)
    1925           0 :         ntiles[is_uv] = rest_tiles_in_plane(cm, is_uv);
    1926             : 
    1927             :     assert(ntiles[1] <= ntiles[0]);
    1928             :     RestUnitSearchInfo *rusi =
    1929           0 :         (RestUnitSearchInfo *)eb_aom_memalign(16, sizeof(*rusi) * ntiles[0]);
    1930             : 
    1931             :     // If the restoration unit dimensions are not multiples of
    1932             :     // rsi->restoration_unit_size then some elements of the rusi array may be
    1933             :     // left uninitialised when we reach copy_unit_info(...). This is not a
    1934             :     // problem, as these elements are ignored later, but in order to quiet
    1935             :     // Valgrind's warnings we initialise the array below.
    1936           0 :     memset(rusi, 0, sizeof(*rusi) * ntiles[0]);
    1937             : 
    1938             :     RestSearchCtxt rsc;
    1939           0 :     const int32_t plane_start = AOM_PLANE_Y;
    1940           0 :     const int32_t plane_end = num_planes > 1 ? AOM_PLANE_V : AOM_PLANE_Y;
    1941           0 :     for (int32_t plane = plane_start; plane <= plane_end; ++plane) {
    1942             :         //CHKN  init_rsc(src, cpi->common, &cpi->td.mb, plane, rusi, &cpi->trial_frame_rst,&rsc);
    1943           0 :         init_rsc(src, cm, x, plane, rusi, trial_frame_rst, &rsc);
    1944             : 
    1945           0 :         const int32_t plane_ntiles = ntiles[plane > 0];
    1946           0 :         const RestorationType num_rtypes =
    1947           0 :             (plane_ntiles > 1) ? RESTORE_TYPES : RESTORE_SWITCHABLE_TYPES;
    1948             : 
    1949           0 :         double best_cost = 0;
    1950           0 :         RestorationType best_rtype = RESTORE_NONE;
    1951           0 :         const int32_t highbd = rsc.cm->use_highbitdepth;
    1952             : 
    1953           0 :         eb_extend_frame(rsc.dgd_buffer, rsc.plane_width, rsc.plane_height,
    1954             :             rsc.dgd_stride, RESTORATION_BORDER, RESTORATION_BORDER,
    1955             :             highbd);
    1956             : 
    1957             :         //CHKN  for (RestorationType r = 0; r < num_rtypes; ++r) {
    1958           0 :         for (int32_t restType = 0; restType < num_rtypes; ++restType) {
    1959           0 :             RestorationType r = (RestorationType)restType;
    1960             : 
    1961           0 :             if ((force_restore_type_d != RESTORE_TYPES) && (r != RESTORE_NONE) &&
    1962             :                 (r != force_restore_type_d))
    1963           0 :                 continue;
    1964             : 
    1965           0 :             double cost = search_rest_type(&rsc, r);
    1966             : 
    1967           0 :             if (r == 0 || cost < best_cost)
    1968             :             {
    1969           0 :                 best_cost = cost;
    1970           0 :                 best_rtype = r;
    1971             :             }
    1972             :         }
    1973           0 :         cm->rst_info[plane].frame_restoration_type = best_rtype;
    1974             :         if (force_restore_type_d != RESTORE_TYPES)
    1975             :             assert(best_rtype == force_restore_type_d || best_rtype == RESTORE_NONE);
    1976             : 
    1977           0 :         if (best_rtype != RESTORE_NONE) {
    1978           0 :             for (int32_t u = 0; u < plane_ntiles; ++u)
    1979           0 :                 copy_unit_info(best_rtype, &rusi[u], &cm->rst_info[plane].unit_info[u]);
    1980             :         }
    1981             :     }
    1982             : 
    1983           0 :     eb_aom_free(rusi);
    1984           0 : }
    1985             : 
    1986         300 : static void search_sgrproj_seg(const RestorationTileLimits *limits,
    1987             :     const AV1PixelRect *tile, int32_t rest_unit_idx,
    1988             :     void *priv) {
    1989         300 :     RestSearchCtxt *rsc = (RestSearchCtxt *)priv;
    1990         300 :     RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx];
    1991             : 
    1992         300 :     Av1Common *const cm = rsc->cm;
    1993         300 :     const int32_t highbd = cm->use_highbitdepth;
    1994         300 :     const int32_t bit_depth = cm->bit_depth;
    1995             : 
    1996         300 :     uint8_t *dgd_start =
    1997         300 :         rsc->dgd_buffer + limits->v_start * rsc->dgd_stride + limits->h_start;
    1998         300 :     const uint8_t *src_start =
    1999         300 :         rsc->src_buffer + limits->v_start * rsc->src_stride + limits->h_start;
    2000             : 
    2001         300 :     const int32_t is_uv = rsc->plane > 0;
    2002         300 :     const int32_t ss_x = is_uv && cm->subsampling_x;
    2003         300 :     const int32_t ss_y = is_uv && cm->subsampling_y;
    2004         300 :     const int32_t procunit_width = RESTORATION_PROC_UNIT_SIZE >> ss_x;
    2005         300 :     const int32_t procunit_height = RESTORATION_PROC_UNIT_SIZE >> ss_y;
    2006         300 :     int8_t step = get_sg_step(cm->sg_filter_mode);
    2007             : 
    2008         300 :     rusi->sgrproj = search_selfguided_restoration(
    2009         300 :         dgd_start, limits->h_end - limits->h_start,
    2010         300 :         limits->v_end - limits->v_start, rsc->dgd_stride, src_start,
    2011             :         rsc->src_stride, highbd, bit_depth, procunit_width, procunit_height,
    2012             :         rsc->tmpbuf
    2013         300 :         , cm->sg_ref_frame_ep,
    2014         300 :         cm->sg_frame_ep_cnt,
    2015             :         step
    2016             :     );
    2017             : 
    2018             :     RestorationUnitInfo rui;
    2019         300 :     rui.restoration_type = RESTORE_SGRPROJ;
    2020         300 :     rui.sgrproj_info = rusi->sgrproj;
    2021             : 
    2022         300 :     rusi->sse[RESTORE_SGRPROJ] = try_restoration_unit_seg(rsc, limits, tile, &rui);
    2023         300 : }
    2024             : 
    2025         300 : static void search_sgrproj_finish(const RestorationTileLimits *limits,
    2026             :     const AV1PixelRect *tile, int32_t rest_unit_idx,
    2027             :     void *priv) {
    2028             :     (void)limits;
    2029             :     (void)tile;
    2030         300 :     RestSearchCtxt *rsc = (RestSearchCtxt *)priv;
    2031         300 :     RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx];
    2032             : 
    2033         300 :     const Macroblock *const x = rsc->x;
    2034             : 
    2035         300 :     rusi->sse[RESTORE_SGRPROJ] = rsc->rusi_pic[rest_unit_idx].sse[RESTORE_SGRPROJ];
    2036         300 :     rusi->sgrproj = rsc->rusi_pic[rest_unit_idx].sgrproj;
    2037             : 
    2038         300 :     const int64_t bits_none = x->sgrproj_restore_cost[0];
    2039         600 :     const int64_t bits_sgr = x->sgrproj_restore_cost[1] +
    2040         300 :         (count_sgrproj_bits(&rusi->sgrproj, &rsc->sgrproj)
    2041         300 :             << AV1_PROB_COST_SHIFT);
    2042             : 
    2043         300 :     double cost_none =
    2044         300 :         RDCOST_DBL(x->rdmult, bits_none >> 4, rusi->sse[RESTORE_NONE]);
    2045         300 :     double cost_sgr =
    2046         300 :         RDCOST_DBL(x->rdmult, bits_sgr >> 4, rusi->sse[RESTORE_SGRPROJ]);
    2047             : 
    2048         300 :     RestorationType rtype =
    2049         300 :         (cost_sgr < cost_none) ? RESTORE_SGRPROJ : RESTORE_NONE;
    2050         300 :     rusi->best_rtype[RESTORE_SGRPROJ - 1] = rtype;
    2051             : 
    2052         300 :     rsc->sse += rusi->sse[rtype];
    2053         300 :     rsc->bits += (cost_sgr < cost_none) ? bits_sgr : bits_none;
    2054         300 :     if (cost_sgr < cost_none) rsc->sgrproj = rusi->sgrproj;
    2055         300 : }
    2056             : 
    2057         300 : static void search_wiener_seg(const RestorationTileLimits *limits,
    2058             :     const AV1PixelRect *tile_rect, int32_t rest_unit_idx,
    2059             :     void *priv) {
    2060         300 :     RestSearchCtxt *rsc = (RestSearchCtxt *)priv;
    2061         300 :     RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx];
    2062         300 :     const Av1Common *const cm = rsc->cm;
    2063         300 :     int32_t wn_luma = cm->wn_filter_mode == 1 ? WIENER_WIN_3TAP : cm->wn_filter_mode == 2 ? WIENER_WIN_CHROMA : WIENER_WIN;
    2064         600 :     const int32_t wiener_win = cm->wn_filter_mode == 1 ? WIENER_WIN_3TAP :
    2065         300 :         (rsc->plane == AOM_PLANE_Y) ? wn_luma : WIENER_WIN_CHROMA;
    2066             :     EB_ALIGN(32) int64_t M[WIENER_WIN2];
    2067             :     EB_ALIGN(32) int64_t H[WIENER_WIN2 * WIENER_WIN2];
    2068             :     int32_t vfilterd[WIENER_WIN], hfilterd[WIENER_WIN];
    2069             : 
    2070         300 :     if (cm->use_highbitdepth)
    2071             :     {
    2072           0 :         if (rsc->plane == AOM_PLANE_Y) {
    2073           0 :             eb_av1_compute_stats_highbd(wiener_win, rsc->dgd_buffer, rsc->src_buffer,
    2074             :                 limits->h_start, limits->h_end, limits->v_start,
    2075             :                 limits->v_end, rsc->dgd_stride, rsc->src_stride, M,
    2076           0 :                 H, (AomBitDepth)cm->bit_depth);
    2077             :         }
    2078             :         else {
    2079           0 :             eb_av1_compute_stats_highbd(wiener_win, rsc->dgd_buffer, rsc->src_buffer,
    2080             :                 limits->h_start, limits->h_end, limits->v_start,
    2081             :                 limits->v_end, rsc->dgd_stride, rsc->src_stride, M,
    2082           0 :                 H, (AomBitDepth)cm->bit_depth);
    2083             :         }
    2084             :     }
    2085             :     else {
    2086         300 :         eb_av1_compute_stats(wiener_win, rsc->dgd_buffer, rsc->src_buffer, limits->h_start,
    2087             :             limits->h_end, limits->v_start, limits->v_end,
    2088             :             rsc->dgd_stride, rsc->src_stride, M, H);
    2089             :     }
    2090             : 
    2091         300 :     if (!wiener_decompose_sep_sym(wiener_win, M, H, vfilterd, hfilterd)) {
    2092           0 :         printf("CHKN never get here\n");
    2093           0 :         rusi->best_rtype[RESTORE_WIENER - 1] = RESTORE_NONE;
    2094           0 :         rusi->sse[RESTORE_WIENER] = INT64_MAX;
    2095           5 :         return;
    2096             :     }
    2097             : 
    2098             :     RestorationUnitInfo rui;
    2099         300 :     memset(&rui, 0, sizeof(rui));
    2100         300 :     rui.restoration_type = RESTORE_WIENER;
    2101         300 :     finalize_sym_filter(wiener_win, vfilterd, rui.wiener_info.vfilter);
    2102         300 :     finalize_sym_filter(wiener_win, hfilterd, rui.wiener_info.hfilter);
    2103             : 
    2104             :     // Filter score computes the value of the function x'*A*x - x'*b for the
    2105             :     // learned filter and compares it against identity filer. If there is no
    2106             :     // reduction in the function, the filter is reverted back to identity
    2107         300 :     if (compute_score(wiener_win, M, H, rui.wiener_info.vfilter,
    2108             :         rui.wiener_info.hfilter) > 0) {
    2109           5 :         rusi->sse[RESTORE_WIENER] = INT64_MAX;
    2110           5 :         return;
    2111             :     }
    2112             : 
    2113         295 :     aom_clear_system_state();
    2114             : 
    2115         295 :     rusi->sse[RESTORE_WIENER] =
    2116         295 :         finer_tile_search_wiener_seg(rsc, limits, tile_rect, &rui, wiener_win);
    2117         295 :     rusi->wiener = rui.wiener_info;
    2118             : 
    2119             :     if (wiener_win != WIENER_WIN) {
    2120             :         assert(rui.wiener_info.vfilter[0] == 0 &&
    2121             :             rui.wiener_info.vfilter[WIENER_WIN - 1] == 0);
    2122             :         assert(rui.wiener_info.hfilter[0] == 0 &&
    2123             :             rui.wiener_info.hfilter[WIENER_WIN - 1] == 0);
    2124             :     }
    2125             : }
    2126         300 : static void search_wiener_finish(const RestorationTileLimits *limits,
    2127             :     const AV1PixelRect *tile_rect, int32_t rest_unit_idx,
    2128             :     void *priv) {
    2129             :     (void)limits;
    2130             :     (void)tile_rect;
    2131         300 :     RestSearchCtxt *rsc = (RestSearchCtxt *)priv;
    2132         300 :     RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx];
    2133         300 :     const Av1Common *const cm = rsc->cm;
    2134         300 :     int32_t wn_luma = cm->wn_filter_mode == 1 ? WIENER_WIN_3TAP : cm->wn_filter_mode == 2 ? WIENER_WIN_CHROMA : WIENER_WIN;
    2135         600 :     const int32_t wiener_win = cm->wn_filter_mode == 1 ? WIENER_WIN_3TAP :
    2136         300 :         (rsc->plane == AOM_PLANE_Y) ? wn_luma : WIENER_WIN_CHROMA;
    2137             : 
    2138         300 :     const Macroblock *const x = rsc->x;
    2139         300 :     const int64_t bits_none = x->wiener_restore_cost[0];
    2140             : 
    2141             :     RestorationUnitInfo rui;
    2142         300 :     memset(&rui, 0, sizeof(rui));
    2143         300 :     rui.restoration_type = RESTORE_WIENER;
    2144             : 
    2145             :     // Filter score computes the value of the function x'*A*x - x'*b for the
    2146             :     // learned filter and compares it against identity filer. If there is no
    2147             :     // reduction in the function, the filter is reverted back to identity
    2148             : 
    2149         300 :     rusi->sse[RESTORE_WIENER] = rsc->rusi_pic[rest_unit_idx].sse[RESTORE_WIENER];
    2150         300 :     if (rusi->sse[RESTORE_WIENER] == INT64_MAX)
    2151             :     {
    2152           5 :         rsc->bits += bits_none;
    2153           5 :         rsc->sse += rusi->sse[RESTORE_NONE];
    2154           5 :         rusi->best_rtype[RESTORE_WIENER - 1] = RESTORE_NONE;
    2155           5 :         rusi->sse[RESTORE_WIENER] = INT64_MAX;
    2156           5 :         return;
    2157             :     }
    2158             : 
    2159         295 :     aom_clear_system_state();
    2160             : 
    2161         295 :     rusi->wiener = rsc->rusi_pic[rest_unit_idx].wiener;
    2162             : 
    2163         295 :     const int64_t bits_wiener =
    2164         590 :         x->wiener_restore_cost[1] +
    2165         295 :         (count_wiener_bits(wiener_win, &rusi->wiener, &rsc->wiener)
    2166         295 :             << AV1_PROB_COST_SHIFT);
    2167             : 
    2168         295 :     double cost_none =
    2169         295 :         RDCOST_DBL(x->rdmult, bits_none >> 4, rusi->sse[RESTORE_NONE]);
    2170         295 :     double cost_wiener =
    2171         295 :         RDCOST_DBL(x->rdmult, bits_wiener >> 4, rusi->sse[RESTORE_WIENER]);
    2172             : 
    2173         295 :     RestorationType rtype =
    2174         295 :         (cost_wiener < cost_none) ? RESTORE_WIENER : RESTORE_NONE;
    2175         295 :     rusi->best_rtype[RESTORE_WIENER - 1] = rtype;
    2176             : 
    2177         295 :     rsc->sse += rusi->sse[rtype];
    2178         295 :     rsc->bits += (cost_wiener < cost_none) ? bits_wiener : bits_none;
    2179         295 :     if (cost_wiener < cost_none) rsc->wiener = rusi->wiener;
    2180             : }
    2181         300 : static void search_norestore_seg(const RestorationTileLimits *limits,
    2182             :     const AV1PixelRect *tile_rect, int32_t rest_unit_idx,
    2183             :     void *priv) {
    2184             :     (void)tile_rect;
    2185             : 
    2186         300 :     RestSearchCtxt *rsc = (RestSearchCtxt *)priv;
    2187         300 :     RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx];
    2188             : 
    2189         300 :     const int32_t highbd = rsc->cm->use_highbitdepth;
    2190         600 :     rusi->sse[RESTORE_NONE] = sse_restoration_unit(
    2191         300 :         limits, rsc->src, rsc->cm->frame_to_show, rsc->plane, highbd);
    2192         300 : }
    2193         300 : static void search_norestore_finish(const RestorationTileLimits *limits,
    2194             :     const AV1PixelRect *tile_rect, int32_t rest_unit_idx,
    2195             :     void *priv) {
    2196             :     (void)tile_rect;
    2197             :     (void)limits;
    2198             : 
    2199         300 :     RestSearchCtxt *rsc = (RestSearchCtxt *)priv;
    2200         300 :     RestUnitSearchInfo *rusi = &rsc->rusi[rest_unit_idx];
    2201             : 
    2202         300 :     rusi->sse[RESTORE_NONE] = rsc->rusi_pic[rest_unit_idx].sse[RESTORE_NONE];
    2203             : 
    2204         300 :     rsc->sse += rusi->sse[RESTORE_NONE];
    2205         300 : }
    2206         600 : static double search_rest_type_finish(RestSearchCtxt *rsc, RestorationType rtype)
    2207             : {
    2208             :     static const RestUnitVisitor funs[RESTORE_TYPES] = {
    2209             :     search_norestore_finish, search_wiener_finish, search_sgrproj_finish, search_switchable
    2210             :     };
    2211             : 
    2212         600 :     reset_rsc(rsc);
    2213             : 
    2214         600 :     av1_foreach_rest_unit_in_frame(rsc->cm, rsc->plane, rsc_on_tile, funs[rtype], rsc);
    2215             : 
    2216         600 :     return RDCOST_DBL(rsc->x->rdmult, rsc->bits >> 4, rsc->sse);
    2217             : }
    2218             : 
    2219          60 : void restoration_seg_search(
    2220             :     RestContext          *context_ptr,
    2221             :     Yv12BufferConfig       *org_fts,
    2222             :     const Yv12BufferConfig *src,
    2223             :     Yv12BufferConfig       *trial_frame_rst ,
    2224             :     PictureControlSet    *pcs_ptr,
    2225             :     uint32_t                segment_index )
    2226             : {
    2227          60 :     Av1Common *const cm = pcs_ptr->parent_pcs_ptr->av1_cm;
    2228          60 :     Macroblock *x = pcs_ptr->parent_pcs_ptr->av1x;
    2229          60 :     const int32_t num_planes = 3;
    2230             : 
    2231             :     // If the restoration unit dimensions are not multiples of
    2232             :     // rsi->restoration_unit_size then some elements of the rusi array may be
    2233             :     // left uninitialised when we reach copy_unit_info(...). This is not a
    2234             :     // problem, as these elements are ignored later, but in order to quiet
    2235             :     // Valgrind's warnings we initialise the array  to zero.
    2236             : 
    2237             :     RestSearchCtxt rsc; //this context is specific for this segment
    2238          60 :     RestSearchCtxt* rsc_p = &rsc;
    2239             : 
    2240          60 :     const int32_t plane_start = AOM_PLANE_Y;
    2241          60 :     const int32_t plane_end = num_planes > 1 ? AOM_PLANE_V : AOM_PLANE_Y;
    2242         240 :     for (int32_t plane = plane_start; plane <= plane_end; ++plane)
    2243             :     {
    2244         180 :         RestUnitSearchInfo *rusi = pcs_ptr->parent_pcs_ptr->rusi_picture[plane];
    2245             : 
    2246         180 :         init_rsc_seg(org_fts,src, cm, x, plane, rusi, trial_frame_rst, &rsc);
    2247             : 
    2248         180 :         rsc_p->tmpbuf = context_ptr->rst_tmpbuf;
    2249             : 
    2250         180 :         const int32_t highbd = rsc.cm->use_highbitdepth;
    2251         180 :         eb_extend_frame(rsc.dgd_buffer, rsc.plane_width, rsc.plane_height,
    2252             :             rsc.dgd_stride, RESTORATION_BORDER, RESTORATION_BORDER,
    2253             :             highbd);
    2254             : 
    2255         180 :         av1_foreach_rest_unit_in_frame_seg(rsc_p->cm, rsc_p->plane, rsc_on_tile, search_norestore_seg, rsc_p, pcs_ptr, segment_index);
    2256         180 :         if (cm->wn_filter_mode)
    2257         180 :             av1_foreach_rest_unit_in_frame_seg(rsc_p->cm, rsc_p->plane, rsc_on_tile, search_wiener_seg,  rsc_p, pcs_ptr, segment_index);
    2258         180 :         av1_foreach_rest_unit_in_frame_seg(rsc_p->cm, rsc_p->plane, rsc_on_tile, search_sgrproj_seg, rsc_p, pcs_ptr, segment_index);
    2259             :     }
    2260          60 : }
    2261          60 : void rest_finish_search(Macroblock *x, Av1Common *const cm)
    2262             : {
    2263          60 :     const int32_t num_planes = 3;
    2264          60 :     RestorationType force_restore_type_d = (cm->wn_filter_mode) ? RESTORE_TYPES : RESTORE_SGRPROJ;
    2265             :     int32_t ntiles[2];
    2266         180 :     for (int32_t is_uv = 0; is_uv < 2; ++is_uv)
    2267         120 :         ntiles[is_uv] = rest_tiles_in_plane(cm, is_uv);
    2268             : 
    2269             :     assert(ntiles[1] <= ntiles[0]);
    2270             :     RestUnitSearchInfo *rusi =
    2271          60 :         (RestUnitSearchInfo *)eb_aom_memalign(16, sizeof(*rusi) * ntiles[0]);
    2272             : 
    2273             :     // If the restoration unit dimensions are not multiples of
    2274             :     // rsi->restoration_unit_size then some elements of the rusi array may be
    2275             :     // left uninitialised when we reach copy_unit_info(...). This is not a
    2276             :     // problem, as these elements are ignored later, but in order to quiet
    2277             :     // Valgrind's warnings we initialise the array below.
    2278          60 :     memset(rusi, 0, sizeof(*rusi) * ntiles[0]);
    2279             : 
    2280             :     RestSearchCtxt rsc;
    2281          60 :     const int32_t plane_start = AOM_PLANE_Y;
    2282          60 :     const int32_t plane_end = num_planes > 1 ? AOM_PLANE_V : AOM_PLANE_Y;
    2283         240 :     for (int32_t plane = plane_start; plane <= plane_end; ++plane) {
    2284             :         //init rsc context for this plane
    2285         180 :         rsc.cm = cm;
    2286         180 :         rsc.x = x;
    2287         180 :         rsc.plane = plane;
    2288         180 :         rsc.rusi = rusi;
    2289         180 :         rsc.pic_num = (uint32_t)cm->p_pcs_ptr->picture_number;
    2290         180 :         rsc.rusi_pic = cm->p_pcs_ptr->rusi_picture[plane];
    2291             : 
    2292         180 :         const int32_t plane_ntiles = ntiles[plane > 0];
    2293         180 :         const RestorationType num_rtypes =
    2294         180 :             (plane_ntiles > 1) ? RESTORE_TYPES : RESTORE_SWITCHABLE_TYPES;
    2295             : 
    2296         180 :         double best_cost = 0;
    2297         180 :         RestorationType best_rtype = RESTORE_NONE;
    2298             : 
    2299         780 :         for (int32_t restType = 0; restType < num_rtypes; ++restType) {
    2300         600 :             RestorationType r = (RestorationType)restType;
    2301             : 
    2302         600 :             if ((force_restore_type_d != RESTORE_TYPES) && (r != RESTORE_NONE) &&
    2303             :                 (r != force_restore_type_d))
    2304           0 :                 continue;
    2305             : 
    2306         600 :             double cost = search_rest_type_finish(&rsc, r);
    2307             : 
    2308         600 :             if (r == 0 || cost < best_cost)
    2309             :             {
    2310         294 :                 best_cost = cost;
    2311         294 :                 best_rtype = r;
    2312             :             }
    2313             :         }
    2314             : 
    2315         180 :         cm->rst_info[plane].frame_restoration_type = best_rtype;
    2316             :         if (force_restore_type_d != RESTORE_TYPES)
    2317             :             assert(best_rtype == force_restore_type_d || best_rtype == RESTORE_NONE);
    2318             : 
    2319         180 :         if (best_rtype != RESTORE_NONE) {
    2320         264 :             for (int32_t u = 0; u < plane_ntiles; ++u)
    2321         192 :                 copy_unit_info(best_rtype, &rusi[u], &cm->rst_info[plane].unit_info[u]);
    2322             :         }
    2323             :     }
    2324             : 
    2325          60 :     eb_aom_free(rusi);
    2326          60 : }

Generated by: LCOV version 1.14