LCOV - code coverage report
Current view: top level - Codec - EbWarpedMotion.c (source / functions) Hit Total Coverage
Test: coverage.info Lines: 155 376 41.2 %
Date: 2019-11-25 17:38:06 Functions: 14 21 66.7 %

          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 <stdio.h>
      13             : #include <stdlib.h>
      14             : #include <math.h>
      15             : #include <assert.h>
      16             : #include "aom_dsp_rtcd.h"
      17             : #include "EbWarpedMotion.h"
      18             : 
      19             : #define WARP_ERROR_BLOCK 32
      20             : 
      21             : /* clang-format off */
      22             : const int error_measure_lut[512] = {
      23             :   // pow 0.7
      24             :   16384, 16339, 16294, 16249, 16204, 16158, 16113, 16068,
      25             :   16022, 15977, 15932, 15886, 15840, 15795, 15749, 15703,
      26             :   15657, 15612, 15566, 15520, 15474, 15427, 15381, 15335,
      27             :   15289, 15242, 15196, 15149, 15103, 15056, 15010, 14963,
      28             :   14916, 14869, 14822, 14775, 14728, 14681, 14634, 14587,
      29             :   14539, 14492, 14445, 14397, 14350, 14302, 14254, 14206,
      30             :   14159, 14111, 14063, 14015, 13967, 13918, 13870, 13822,
      31             :   13773, 13725, 13676, 13628, 13579, 13530, 13481, 13432,
      32             :   13383, 13334, 13285, 13236, 13187, 13137, 13088, 13038,
      33             :   12988, 12939, 12889, 12839, 12789, 12739, 12689, 12639,
      34             :   12588, 12538, 12487, 12437, 12386, 12335, 12285, 12234,
      35             :   12183, 12132, 12080, 12029, 11978, 11926, 11875, 11823,
      36             :   11771, 11719, 11667, 11615, 11563, 11511, 11458, 11406,
      37             :   11353, 11301, 11248, 11195, 11142, 11089, 11036, 10982,
      38             :   10929, 10875, 10822, 10768, 10714, 10660, 10606, 10552,
      39             :   10497, 10443, 10388, 10333, 10279, 10224, 10168, 10113,
      40             :   10058, 10002,  9947,  9891,  9835,  9779,  9723,  9666,
      41             :   9610, 9553, 9497, 9440, 9383, 9326, 9268, 9211,
      42             :   9153, 9095, 9037, 8979, 8921, 8862, 8804, 8745,
      43             :   8686, 8627, 8568, 8508, 8449, 8389, 8329, 8269,
      44             :   8208, 8148, 8087, 8026, 7965, 7903, 7842, 7780,
      45             :   7718, 7656, 7593, 7531, 7468, 7405, 7341, 7278,
      46             :   7214, 7150, 7086, 7021, 6956, 6891, 6826, 6760,
      47             :   6695, 6628, 6562, 6495, 6428, 6361, 6293, 6225,
      48             :   6157, 6089, 6020, 5950, 5881, 5811, 5741, 5670,
      49             :   5599, 5527, 5456, 5383, 5311, 5237, 5164, 5090,
      50             :   5015, 4941, 4865, 4789, 4713, 4636, 4558, 4480,
      51             :   4401, 4322, 4242, 4162, 4080, 3998, 3916, 3832,
      52             :   3748, 3663, 3577, 3490, 3402, 3314, 3224, 3133,
      53             :   3041, 2948, 2854, 2758, 2661, 2562, 2461, 2359,
      54             :   2255, 2148, 2040, 1929, 1815, 1698, 1577, 1452,
      55             :   1323, 1187, 1045,  894,  731,  550,  339,    0,
      56             :   339,  550,  731,  894, 1045, 1187, 1323, 1452,
      57             :   1577, 1698, 1815, 1929, 2040, 2148, 2255, 2359,
      58             :   2461, 2562, 2661, 2758, 2854, 2948, 3041, 3133,
      59             :   3224, 3314, 3402, 3490, 3577, 3663, 3748, 3832,
      60             :   3916, 3998, 4080, 4162, 4242, 4322, 4401, 4480,
      61             :   4558, 4636, 4713, 4789, 4865, 4941, 5015, 5090,
      62             :   5164, 5237, 5311, 5383, 5456, 5527, 5599, 5670,
      63             :   5741, 5811, 5881, 5950, 6020, 6089, 6157, 6225,
      64             :   6293, 6361, 6428, 6495, 6562, 6628, 6695, 6760,
      65             :   6826, 6891, 6956, 7021, 7086, 7150, 7214, 7278,
      66             :   7341, 7405, 7468, 7531, 7593, 7656, 7718, 7780,
      67             :   7842, 7903, 7965, 8026, 8087, 8148, 8208, 8269,
      68             :   8329, 8389, 8449, 8508, 8568, 8627, 8686, 8745,
      69             :   8804, 8862, 8921, 8979, 9037, 9095, 9153, 9211,
      70             :   9268, 9326, 9383, 9440, 9497, 9553, 9610, 9666,
      71             :   9723,  9779,  9835,  9891,  9947, 10002, 10058, 10113,
      72             :   10168, 10224, 10279, 10333, 10388, 10443, 10497, 10552,
      73             :   10606, 10660, 10714, 10768, 10822, 10875, 10929, 10982,
      74             :   11036, 11089, 11142, 11195, 11248, 11301, 11353, 11406,
      75             :   11458, 11511, 11563, 11615, 11667, 11719, 11771, 11823,
      76             :   11875, 11926, 11978, 12029, 12080, 12132, 12183, 12234,
      77             :   12285, 12335, 12386, 12437, 12487, 12538, 12588, 12639,
      78             :   12689, 12739, 12789, 12839, 12889, 12939, 12988, 13038,
      79             :   13088, 13137, 13187, 13236, 13285, 13334, 13383, 13432,
      80             :   13481, 13530, 13579, 13628, 13676, 13725, 13773, 13822,
      81             :   13870, 13918, 13967, 14015, 14063, 14111, 14159, 14206,
      82             :   14254, 14302, 14350, 14397, 14445, 14492, 14539, 14587,
      83             :   14634, 14681, 14728, 14775, 14822, 14869, 14916, 14963,
      84             :   15010, 15056, 15103, 15149, 15196, 15242, 15289, 15335,
      85             :   15381, 15427, 15474, 15520, 15566, 15612, 15657, 15703,
      86             :   15749, 15795, 15840, 15886, 15932, 15977, 16022, 16068,
      87             :   16113, 16158, 16204, 16249, 16294, 16339, 16384, 16384,
      88             : };
      89             : /* clang-format on */
      90             : 
      91             : // For warping, we really use a 6-tap filter, but we do blocks of 8 pixels
      92             : // at a time. The zoom/rotation/shear in the model are applied to the
      93             : // "fractional" position of each pixel, which therefore varies within
      94             : // [-1, 2) * WARPEDPIXEL_PREC_SHIFTS.
      95             : // We need an extra 2 taps to fit this in, for a total of 8 taps.
      96             : /* clang-format off */
      97             : EB_ALIGN(16) const int16_t eb_warped_filter[WARPEDPIXEL_PREC_SHIFTS * 3 + 1][8] = {
      98             : #if WARPEDPIXEL_PREC_BITS == 6
      99             :   // [-1, 0)
     100             :   { 0,   0, 127,   1,   0, 0, 0, 0 }, { 0, - 1, 127,   2,   0, 0, 0, 0 },
     101             :   { 1, - 3, 127,   4, - 1, 0, 0, 0 }, { 1, - 4, 126,   6, - 2, 1, 0, 0 },
     102             :   { 1, - 5, 126,   8, - 3, 1, 0, 0 }, { 1, - 6, 125,  11, - 4, 1, 0, 0 },
     103             :   { 1, - 7, 124,  13, - 4, 1, 0, 0 }, { 2, - 8, 123,  15, - 5, 1, 0, 0 },
     104             :   { 2, - 9, 122,  18, - 6, 1, 0, 0 }, { 2, -10, 121,  20, - 6, 1, 0, 0 },
     105             :   { 2, -11, 120,  22, - 7, 2, 0, 0 }, { 2, -12, 119,  25, - 8, 2, 0, 0 },
     106             :   { 3, -13, 117,  27, - 8, 2, 0, 0 }, { 3, -13, 116,  29, - 9, 2, 0, 0 },
     107             :   { 3, -14, 114,  32, -10, 3, 0, 0 }, { 3, -15, 113,  35, -10, 2, 0, 0 },
     108             :   { 3, -15, 111,  37, -11, 3, 0, 0 }, { 3, -16, 109,  40, -11, 3, 0, 0 },
     109             :   { 3, -16, 108,  42, -12, 3, 0, 0 }, { 4, -17, 106,  45, -13, 3, 0, 0 },
     110             :   { 4, -17, 104,  47, -13, 3, 0, 0 }, { 4, -17, 102,  50, -14, 3, 0, 0 },
     111             :   { 4, -17, 100,  52, -14, 3, 0, 0 }, { 4, -18,  98,  55, -15, 4, 0, 0 },
     112             :   { 4, -18,  96,  58, -15, 3, 0, 0 }, { 4, -18,  94,  60, -16, 4, 0, 0 },
     113             :   { 4, -18,  91,  63, -16, 4, 0, 0 }, { 4, -18,  89,  65, -16, 4, 0, 0 },
     114             :   { 4, -18,  87,  68, -17, 4, 0, 0 }, { 4, -18,  85,  70, -17, 4, 0, 0 },
     115             :   { 4, -18,  82,  73, -17, 4, 0, 0 }, { 4, -18,  80,  75, -17, 4, 0, 0 },
     116             :   { 4, -18,  78,  78, -18, 4, 0, 0 }, { 4, -17,  75,  80, -18, 4, 0, 0 },
     117             :   { 4, -17,  73,  82, -18, 4, 0, 0 }, { 4, -17,  70,  85, -18, 4, 0, 0 },
     118             :   { 4, -17,  68,  87, -18, 4, 0, 0 }, { 4, -16,  65,  89, -18, 4, 0, 0 },
     119             :   { 4, -16,  63,  91, -18, 4, 0, 0 }, { 4, -16,  60,  94, -18, 4, 0, 0 },
     120             :   { 3, -15,  58,  96, -18, 4, 0, 0 }, { 4, -15,  55,  98, -18, 4, 0, 0 },
     121             :   { 3, -14,  52, 100, -17, 4, 0, 0 }, { 3, -14,  50, 102, -17, 4, 0, 0 },
     122             :   { 3, -13,  47, 104, -17, 4, 0, 0 }, { 3, -13,  45, 106, -17, 4, 0, 0 },
     123             :   { 3, -12,  42, 108, -16, 3, 0, 0 }, { 3, -11,  40, 109, -16, 3, 0, 0 },
     124             :   { 3, -11,  37, 111, -15, 3, 0, 0 }, { 2, -10,  35, 113, -15, 3, 0, 0 },
     125             :   { 3, -10,  32, 114, -14, 3, 0, 0 }, { 2, - 9,  29, 116, -13, 3, 0, 0 },
     126             :   { 2, - 8,  27, 117, -13, 3, 0, 0 }, { 2, - 8,  25, 119, -12, 2, 0, 0 },
     127             :   { 2, - 7,  22, 120, -11, 2, 0, 0 }, { 1, - 6,  20, 121, -10, 2, 0, 0 },
     128             :   { 1, - 6,  18, 122, - 9, 2, 0, 0 }, { 1, - 5,  15, 123, - 8, 2, 0, 0 },
     129             :   { 1, - 4,  13, 124, - 7, 1, 0, 0 }, { 1, - 4,  11, 125, - 6, 1, 0, 0 },
     130             :   { 1, - 3,   8, 126, - 5, 1, 0, 0 }, { 1, - 2,   6, 126, - 4, 1, 0, 0 },
     131             :   { 0, - 1,   4, 127, - 3, 1, 0, 0 }, { 0,   0,   2, 127, - 1, 0, 0, 0 },
     132             :   // [0, 1)
     133             :   { 0,  0,   0, 127,   1,   0,  0,  0}, { 0,  0,  -1, 127,   2,   0,  0,  0},
     134             :   { 0,  1,  -3, 127,   4,  -2,  1,  0}, { 0,  1,  -5, 127,   6,  -2,  1,  0},
     135             :   { 0,  2,  -6, 126,   8,  -3,  1,  0}, {-1,  2,  -7, 126,  11,  -4,  2, -1},
     136             :   {-1,  3,  -8, 125,  13,  -5,  2, -1}, {-1,  3, -10, 124,  16,  -6,  3, -1},
     137             :   {-1,  4, -11, 123,  18,  -7,  3, -1}, {-1,  4, -12, 122,  20,  -7,  3, -1},
     138             :   {-1,  4, -13, 121,  23,  -8,  3, -1}, {-2,  5, -14, 120,  25,  -9,  4, -1},
     139             :   {-1,  5, -15, 119,  27, -10,  4, -1}, {-1,  5, -16, 118,  30, -11,  4, -1},
     140             :   {-2,  6, -17, 116,  33, -12,  5, -1}, {-2,  6, -17, 114,  35, -12,  5, -1},
     141             :   {-2,  6, -18, 113,  38, -13,  5, -1}, {-2,  7, -19, 111,  41, -14,  6, -2},
     142             :   {-2,  7, -19, 110,  43, -15,  6, -2}, {-2,  7, -20, 108,  46, -15,  6, -2},
     143             :   {-2,  7, -20, 106,  49, -16,  6, -2}, {-2,  7, -21, 104,  51, -16,  7, -2},
     144             :   {-2,  7, -21, 102,  54, -17,  7, -2}, {-2,  8, -21, 100,  56, -18,  7, -2},
     145             :   {-2,  8, -22,  98,  59, -18,  7, -2}, {-2,  8, -22,  96,  62, -19,  7, -2},
     146             :   {-2,  8, -22,  94,  64, -19,  7, -2}, {-2,  8, -22,  91,  67, -20,  8, -2},
     147             :   {-2,  8, -22,  89,  69, -20,  8, -2}, {-2,  8, -22,  87,  72, -21,  8, -2},
     148             :   {-2,  8, -21,  84,  74, -21,  8, -2}, {-2,  8, -22,  82,  77, -21,  8, -2},
     149             :   {-2,  8, -21,  79,  79, -21,  8, -2}, {-2,  8, -21,  77,  82, -22,  8, -2},
     150             :   {-2,  8, -21,  74,  84, -21,  8, -2}, {-2,  8, -21,  72,  87, -22,  8, -2},
     151             :   {-2,  8, -20,  69,  89, -22,  8, -2}, {-2,  8, -20,  67,  91, -22,  8, -2},
     152             :   {-2,  7, -19,  64,  94, -22,  8, -2}, {-2,  7, -19,  62,  96, -22,  8, -2},
     153             :   {-2,  7, -18,  59,  98, -22,  8, -2}, {-2,  7, -18,  56, 100, -21,  8, -2},
     154             :   {-2,  7, -17,  54, 102, -21,  7, -2}, {-2,  7, -16,  51, 104, -21,  7, -2},
     155             :   {-2,  6, -16,  49, 106, -20,  7, -2}, {-2,  6, -15,  46, 108, -20,  7, -2},
     156             :   {-2,  6, -15,  43, 110, -19,  7, -2}, {-2,  6, -14,  41, 111, -19,  7, -2},
     157             :   {-1,  5, -13,  38, 113, -18,  6, -2}, {-1,  5, -12,  35, 114, -17,  6, -2},
     158             :   {-1,  5, -12,  33, 116, -17,  6, -2}, {-1,  4, -11,  30, 118, -16,  5, -1},
     159             :   {-1,  4, -10,  27, 119, -15,  5, -1}, {-1,  4,  -9,  25, 120, -14,  5, -2},
     160             :   {-1,  3,  -8,  23, 121, -13,  4, -1}, {-1,  3,  -7,  20, 122, -12,  4, -1},
     161             :   {-1,  3,  -7,  18, 123, -11,  4, -1}, {-1,  3,  -6,  16, 124, -10,  3, -1},
     162             :   {-1,  2,  -5,  13, 125,  -8,  3, -1}, {-1,  2,  -4,  11, 126,  -7,  2, -1},
     163             :   { 0,  1,  -3,   8, 126,  -6,  2,  0}, { 0,  1,  -2,   6, 127,  -5,  1,  0},
     164             :   { 0,  1,  -2,   4, 127,  -3,  1,  0}, { 0,  0,   0,   2, 127,  -1,  0,  0},
     165             :   // [1, 2)
     166             :   { 0, 0, 0,   1, 127,   0,   0, 0 }, { 0, 0, 0, - 1, 127,   2,   0, 0 },
     167             :   { 0, 0, 1, - 3, 127,   4, - 1, 0 }, { 0, 0, 1, - 4, 126,   6, - 2, 1 },
     168             :   { 0, 0, 1, - 5, 126,   8, - 3, 1 }, { 0, 0, 1, - 6, 125,  11, - 4, 1 },
     169             :   { 0, 0, 1, - 7, 124,  13, - 4, 1 }, { 0, 0, 2, - 8, 123,  15, - 5, 1 },
     170             :   { 0, 0, 2, - 9, 122,  18, - 6, 1 }, { 0, 0, 2, -10, 121,  20, - 6, 1 },
     171             :   { 0, 0, 2, -11, 120,  22, - 7, 2 }, { 0, 0, 2, -12, 119,  25, - 8, 2 },
     172             :   { 0, 0, 3, -13, 117,  27, - 8, 2 }, { 0, 0, 3, -13, 116,  29, - 9, 2 },
     173             :   { 0, 0, 3, -14, 114,  32, -10, 3 }, { 0, 0, 3, -15, 113,  35, -10, 2 },
     174             :   { 0, 0, 3, -15, 111,  37, -11, 3 }, { 0, 0, 3, -16, 109,  40, -11, 3 },
     175             :   { 0, 0, 3, -16, 108,  42, -12, 3 }, { 0, 0, 4, -17, 106,  45, -13, 3 },
     176             :   { 0, 0, 4, -17, 104,  47, -13, 3 }, { 0, 0, 4, -17, 102,  50, -14, 3 },
     177             :   { 0, 0, 4, -17, 100,  52, -14, 3 }, { 0, 0, 4, -18,  98,  55, -15, 4 },
     178             :   { 0, 0, 4, -18,  96,  58, -15, 3 }, { 0, 0, 4, -18,  94,  60, -16, 4 },
     179             :   { 0, 0, 4, -18,  91,  63, -16, 4 }, { 0, 0, 4, -18,  89,  65, -16, 4 },
     180             :   { 0, 0, 4, -18,  87,  68, -17, 4 }, { 0, 0, 4, -18,  85,  70, -17, 4 },
     181             :   { 0, 0, 4, -18,  82,  73, -17, 4 }, { 0, 0, 4, -18,  80,  75, -17, 4 },
     182             :   { 0, 0, 4, -18,  78,  78, -18, 4 }, { 0, 0, 4, -17,  75,  80, -18, 4 },
     183             :   { 0, 0, 4, -17,  73,  82, -18, 4 }, { 0, 0, 4, -17,  70,  85, -18, 4 },
     184             :   { 0, 0, 4, -17,  68,  87, -18, 4 }, { 0, 0, 4, -16,  65,  89, -18, 4 },
     185             :   { 0, 0, 4, -16,  63,  91, -18, 4 }, { 0, 0, 4, -16,  60,  94, -18, 4 },
     186             :   { 0, 0, 3, -15,  58,  96, -18, 4 }, { 0, 0, 4, -15,  55,  98, -18, 4 },
     187             :   { 0, 0, 3, -14,  52, 100, -17, 4 }, { 0, 0, 3, -14,  50, 102, -17, 4 },
     188             :   { 0, 0, 3, -13,  47, 104, -17, 4 }, { 0, 0, 3, -13,  45, 106, -17, 4 },
     189             :   { 0, 0, 3, -12,  42, 108, -16, 3 }, { 0, 0, 3, -11,  40, 109, -16, 3 },
     190             :   { 0, 0, 3, -11,  37, 111, -15, 3 }, { 0, 0, 2, -10,  35, 113, -15, 3 },
     191             :   { 0, 0, 3, -10,  32, 114, -14, 3 }, { 0, 0, 2, - 9,  29, 116, -13, 3 },
     192             :   { 0, 0, 2, - 8,  27, 117, -13, 3 }, { 0, 0, 2, - 8,  25, 119, -12, 2 },
     193             :   { 0, 0, 2, - 7,  22, 120, -11, 2 }, { 0, 0, 1, - 6,  20, 121, -10, 2 },
     194             :   { 0, 0, 1, - 6,  18, 122, - 9, 2 }, { 0, 0, 1, - 5,  15, 123, - 8, 2 },
     195             :   { 0, 0, 1, - 4,  13, 124, - 7, 1 }, { 0, 0, 1, - 4,  11, 125, - 6, 1 },
     196             :   { 0, 0, 1, - 3,   8, 126, - 5, 1 }, { 0, 0, 1, - 2,   6, 126, - 4, 1 },
     197             :   { 0, 0, 0, - 1,   4, 127, - 3, 1 }, { 0, 0, 0,   0,   2, 127, - 1, 0 },
     198             :   // dummy (replicate row index 191)
     199             :   { 0, 0, 0,   0,   2, 127, - 1, 0 },
     200             : #elif WARPEDPIXEL_PREC_BITS == 5
     201             :   // [-1, 0)
     202             :   {0,   0, 127,   1,   0, 0, 0, 0}, {1,  -3, 127,   4,  -1, 0, 0, 0},
     203             :   {1,  -5, 126,   8,  -3, 1, 0, 0}, {1,  -7, 124,  13,  -4, 1, 0, 0},
     204             :   {2,  -9, 122,  18,  -6, 1, 0, 0}, {2, -11, 120,  22,  -7, 2, 0, 0},
     205             :   {3, -13, 117,  27,  -8, 2, 0, 0}, {3, -14, 114,  32, -10, 3, 0, 0},
     206             :   {3, -15, 111,  37, -11, 3, 0, 0}, {3, -16, 108,  42, -12, 3, 0, 0},
     207             :   {4, -17, 104,  47, -13, 3, 0, 0}, {4, -17, 100,  52, -14, 3, 0, 0},
     208             :   {4, -18,  96,  58, -15, 3, 0, 0}, {4, -18,  91,  63, -16, 4, 0, 0},
     209             :   {4, -18,  87,  68, -17, 4, 0, 0}, {4, -18,  82,  73, -17, 4, 0, 0},
     210             :   {4, -18,  78,  78, -18, 4, 0, 0}, {4, -17,  73,  82, -18, 4, 0, 0},
     211             :   {4, -17,  68,  87, -18, 4, 0, 0}, {4, -16,  63,  91, -18, 4, 0, 0},
     212             :   {3, -15,  58,  96, -18, 4, 0, 0}, {3, -14,  52, 100, -17, 4, 0, 0},
     213             :   {3, -13,  47, 104, -17, 4, 0, 0}, {3, -12,  42, 108, -16, 3, 0, 0},
     214             :   {3, -11,  37, 111, -15, 3, 0, 0}, {3, -10,  32, 114, -14, 3, 0, 0},
     215             :   {2,  -8,  27, 117, -13, 3, 0, 0}, {2,  -7,  22, 120, -11, 2, 0, 0},
     216             :   {1,  -6,  18, 122,  -9, 2, 0, 0}, {1,  -4,  13, 124,  -7, 1, 0, 0},
     217             :   {1,  -3,   8, 126,  -5, 1, 0, 0}, {0,  -1,   4, 127,  -3, 1, 0, 0},
     218             :   // [0, 1)
     219             :   { 0,  0,   0, 127,   1,   0,   0,  0}, { 0,  1,  -3, 127,   4,  -2,   1,  0},
     220             :   { 0,  2,  -6, 126,   8,  -3,   1,  0}, {-1,  3,  -8, 125,  13,  -5,   2, -1},
     221             :   {-1,  4, -11, 123,  18,  -7,   3, -1}, {-1,  4, -13, 121,  23,  -8,   3, -1},
     222             :   {-1,  5, -15, 119,  27, -10,   4, -1}, {-2,  6, -17, 116,  33, -12,   5, -1},
     223             :   {-2,  6, -18, 113,  38, -13,   5, -1}, {-2,  7, -19, 110,  43, -15,   6, -2},
     224             :   {-2,  7, -20, 106,  49, -16,   6, -2}, {-2,  7, -21, 102,  54, -17,   7, -2},
     225             :   {-2,  8, -22,  98,  59, -18,   7, -2}, {-2,  8, -22,  94,  64, -19,   7, -2},
     226             :   {-2,  8, -22,  89,  69, -20,   8, -2}, {-2,  8, -21,  84,  74, -21,   8, -2},
     227             :   {-2,  8, -21,  79,  79, -21,   8, -2}, {-2,  8, -21,  74,  84, -21,   8, -2},
     228             :   {-2,  8, -20,  69,  89, -22,   8, -2}, {-2,  7, -19,  64,  94, -22,   8, -2},
     229             :   {-2,  7, -18,  59,  98, -22,   8, -2}, {-2,  7, -17,  54, 102, -21,   7, -2},
     230             :   {-2,  6, -16,  49, 106, -20,   7, -2}, {-2,  6, -15,  43, 110, -19,   7, -2},
     231             :   {-1,  5, -13,  38, 113, -18,   6, -2}, {-1,  5, -12,  33, 116, -17,   6, -2},
     232             :   {-1,  4, -10,  27, 119, -15,   5, -1}, {-1,  3,  -8,  23, 121, -13,   4, -1},
     233             :   {-1,  3,  -7,  18, 123, -11,   4, -1}, {-1,  2,  -5,  13, 125,  -8,   3, -1},
     234             :   { 0,  1,  -3,   8, 126,  -6,   2,  0}, { 0,  1,  -2,   4, 127,  -3,   1,  0},
     235             :   // [1, 2)
     236             :   {0, 0, 0,   1, 127,   0,   0, 0}, {0, 0, 1,  -3, 127,   4,  -1, 0},
     237             :   {0, 0, 1,  -5, 126,   8,  -3, 1}, {0, 0, 1,  -7, 124,  13,  -4, 1},
     238             :   {0, 0, 2,  -9, 122,  18,  -6, 1}, {0, 0, 2, -11, 120,  22,  -7, 2},
     239             :   {0, 0, 3, -13, 117,  27,  -8, 2}, {0, 0, 3, -14, 114,  32, -10, 3},
     240             :   {0, 0, 3, -15, 111,  37, -11, 3}, {0, 0, 3, -16, 108,  42, -12, 3},
     241             :   {0, 0, 4, -17, 104,  47, -13, 3}, {0, 0, 4, -17, 100,  52, -14, 3},
     242             :   {0, 0, 4, -18,  96,  58, -15, 3}, {0, 0, 4, -18,  91,  63, -16, 4},
     243             :   {0, 0, 4, -18,  87,  68, -17, 4}, {0, 0, 4, -18,  82,  73, -17, 4},
     244             :   {0, 0, 4, -18,  78,  78, -18, 4}, {0, 0, 4, -17,  73,  82, -18, 4},
     245             :   {0, 0, 4, -17,  68,  87, -18, 4}, {0, 0, 4, -16,  63,  91, -18, 4},
     246             :   {0, 0, 3, -15,  58,  96, -18, 4}, {0, 0, 3, -14,  52, 100, -17, 4},
     247             :   {0, 0, 3, -13,  47, 104, -17, 4}, {0, 0, 3, -12,  42, 108, -16, 3},
     248             :   {0, 0, 3, -11,  37, 111, -15, 3}, {0, 0, 3, -10,  32, 114, -14, 3},
     249             :   {0, 0, 2,  -8,  27, 117, -13, 3}, {0, 0, 2,  -7,  22, 120, -11, 2},
     250             :   {0, 0, 1,  -6,  18, 122,  -9, 2}, {0, 0, 1,  -4,  13, 124,  -7, 1},
     251             :   {0, 0, 1,  -3,   8, 126,  -5, 1}, {0, 0, 0,  -1,   4, 127,  -3, 1},
     252             :   // dummy (replicate row index 95)
     253             :   {0, 0, 0,  -1,   4, 127,  -3, 1},
     254             : #endif  // WARPEDPIXEL_PREC_BITS == 6
     255             : };
     256             : 
     257             : /* clang-format on */
     258             : 
     259             : #define DIV_LUT_PREC_BITS 14
     260             : #define DIV_LUT_BITS 8
     261             : #define DIV_LUT_NUM (1 << DIV_LUT_BITS)
     262             : 
     263             : static const uint16_t div_lut[DIV_LUT_NUM + 1] = {
     264             :   16384, 16320, 16257, 16194, 16132, 16070, 16009, 15948, 15888, 15828, 15768,
     265             :   15709, 15650, 15592, 15534, 15477, 15420, 15364, 15308, 15252, 15197, 15142,
     266             :   15087, 15033, 14980, 14926, 14873, 14821, 14769, 14717, 14665, 14614, 14564,
     267             :   14513, 14463, 14413, 14364, 14315, 14266, 14218, 14170, 14122, 14075, 14028,
     268             :   13981, 13935, 13888, 13843, 13797, 13752, 13707, 13662, 13618, 13574, 13530,
     269             :   13487, 13443, 13400, 13358, 13315, 13273, 13231, 13190, 13148, 13107, 13066,
     270             :   13026, 12985, 12945, 12906, 12866, 12827, 12788, 12749, 12710, 12672, 12633,
     271             :   12596, 12558, 12520, 12483, 12446, 12409, 12373, 12336, 12300, 12264, 12228,
     272             :   12193, 12157, 12122, 12087, 12053, 12018, 11984, 11950, 11916, 11882, 11848,
     273             :   11815, 11782, 11749, 11716, 11683, 11651, 11619, 11586, 11555, 11523, 11491,
     274             :   11460, 11429, 11398, 11367, 11336, 11305, 11275, 11245, 11215, 11185, 11155,
     275             :   11125, 11096, 11067, 11038, 11009, 10980, 10951, 10923, 10894, 10866, 10838,
     276             :   10810, 10782, 10755, 10727, 10700, 10673, 10645, 10618, 10592, 10565, 10538,
     277             :   10512, 10486, 10460, 10434, 10408, 10382, 10356, 10331, 10305, 10280, 10255,
     278             :   10230, 10205, 10180, 10156, 10131, 10107, 10082, 10058, 10034, 10010, 9986,
     279             :   9963,  9939,  9916,  9892,  9869,  9846,  9823,  9800,  9777,  9754,  9732,
     280             :   9709,  9687,  9664,  9642,  9620,  9598,  9576,  9554,  9533,  9511,  9489,
     281             :   9468,  9447,  9425,  9404,  9383,  9362,  9341,  9321,  9300,  9279,  9259,
     282             :   9239,  9218,  9198,  9178,  9158,  9138,  9118,  9098,  9079,  9059,  9039,
     283             :   9020,  9001,  8981,  8962,  8943,  8924,  8905,  8886,  8867,  8849,  8830,
     284             :   8812,  8793,  8775,  8756,  8738,  8720,  8702,  8684,  8666,  8648,  8630,
     285             :   8613,  8595,  8577,  8560,  8542,  8525,  8508,  8490,  8473,  8456,  8439,
     286             :   8422,  8405,  8389,  8372,  8355,  8339,  8322,  8306,  8289,  8273,  8257,
     287             :   8240,  8224,  8208,  8192,
     288             : };
     289             : 
     290             : // Decomposes a divisor D such that 1/D = y/2^shift, where y is returned
     291             : // at precision of DIV_LUT_PREC_BITS along with the shift.
     292      737783 : static int16_t resolve_divisor_64(uint64_t D, int16_t *shift) {
     293             :   int64_t f;
     294     1464570 :   *shift = (int16_t)((D >> 32) ? get_msb((unsigned int)(D >> 32)) + 32
     295      726790 :                                : get_msb((unsigned int)D));
     296             :   // e is obtained from D after resetting the most significant 1 bit.
     297      737782 :   const int64_t e = D - ((uint64_t)1 << *shift);
     298             :   // Get the most significant DIV_LUT_BITS (8) bits of e into f
     299      737782 :   if (*shift > DIV_LUT_BITS)
     300      737781 :     f = ROUND_POWER_OF_TWO_64(e, *shift - DIV_LUT_BITS);
     301             :   else
     302           1 :     f = e << (DIV_LUT_BITS - *shift);
     303      737782 :   assert(f <= DIV_LUT_NUM);
     304      737782 :   *shift += DIV_LUT_PREC_BITS;
     305             :   // Use f as lookup into the precomputed table of multipliers
     306      737782 :   return div_lut[f];
     307             : }
     308             : 
     309      829496 : static int16_t resolve_divisor_32(uint32_t D, int16_t *shift) {
     310             :   int32_t f;
     311      829496 :   *shift = get_msb(D);
     312             :   // e is obtained from D after resetting the most significant 1 bit.
     313      829495 :   const int32_t e = D - ((uint32_t)1 << *shift);
     314             :   // Get the most significant DIV_LUT_BITS (8) bits of e into f
     315      829495 :   if (*shift > DIV_LUT_BITS)
     316      829496 :     f = ROUND_POWER_OF_TWO(e, *shift - DIV_LUT_BITS);
     317             :   else
     318           0 :     f = e << (DIV_LUT_BITS - *shift);
     319      829495 :   assert(f <= DIV_LUT_NUM);
     320      829495 :   *shift += DIV_LUT_PREC_BITS;
     321             :   // Use f as lookup into the precomputed table of multipliers
     322      829495 :   return div_lut[f];
     323             : }
     324             : 
     325      829495 : static int is_affine_valid(const EbWarpedMotionParams *const wm) {
     326      829495 :   const int32_t *mat = wm->wmmat;
     327      829495 :   return (mat[2] > 0);
     328             : }
     329             : 
     330      829487 : static int is_affine_shear_allowed(int16_t alpha, int16_t beta, int16_t gamma,
     331             :                                    int16_t delta) {
     332      829487 :   if ((4 * abs(alpha) + 7 * abs(beta) >= (1 << WARPEDMODEL_PREC_BITS)) ||
     333      782617 :       (4 * abs(gamma) + 4 * abs(delta) >= (1 << WARPEDMODEL_PREC_BITS)))
     334       50470 :     return 0;
     335             :   else
     336      779017 :     return 1;
     337             : }
     338             : 
     339             : // Returns 1 on success or 0 on an invalid affine set
     340      829495 : int eb_get_shear_params(EbWarpedMotionParams *wm) {
     341      829495 :   const int32_t *mat = wm->wmmat;
     342      829495 :   if (!is_affine_valid(wm)) return 0;
     343      829494 :   wm->alpha =
     344      829495 :       clamp(mat[2] - (1 << WARPEDMODEL_PREC_BITS), INT16_MIN, INT16_MAX);
     345      829494 :   wm->beta = clamp(mat[3], INT16_MIN, INT16_MAX);
     346             :   int16_t shift;
     347      829495 :   int16_t y = resolve_divisor_32(abs(mat[2]), &shift) * (mat[2] < 0 ? -1 : 1);
     348      829497 :   int64_t v = ((int64_t)mat[4] * (1 << WARPEDMODEL_PREC_BITS)) * y;
     349      829487 :   wm->gamma =
     350      829497 :       clamp((int)ROUND_POWER_OF_TWO_SIGNED_64(v, shift), INT16_MIN, INT16_MAX);
     351      829487 :   v = ((int64_t)mat[3] * mat[4]) * y;
     352      829487 :   wm->delta = clamp(mat[5] - (int)ROUND_POWER_OF_TWO_SIGNED_64(v, shift) -
     353             :                         (1 << WARPEDMODEL_PREC_BITS),
     354             :                     INT16_MIN, INT16_MAX);
     355             : 
     356      829482 :   wm->alpha = ROUND_POWER_OF_TWO_SIGNED(wm->alpha, WARP_PARAM_REDUCE_BITS) *
     357             :               (1 << WARP_PARAM_REDUCE_BITS);
     358      829482 :   wm->beta = ROUND_POWER_OF_TWO_SIGNED(wm->beta, WARP_PARAM_REDUCE_BITS) *
     359             :              (1 << WARP_PARAM_REDUCE_BITS);
     360      829482 :   wm->gamma = ROUND_POWER_OF_TWO_SIGNED(wm->gamma, WARP_PARAM_REDUCE_BITS) *
     361             :               (1 << WARP_PARAM_REDUCE_BITS);
     362      829482 :   wm->delta = ROUND_POWER_OF_TWO_SIGNED(wm->delta, WARP_PARAM_REDUCE_BITS) *
     363             :               (1 << WARP_PARAM_REDUCE_BITS);
     364             : 
     365      829482 :   if (!is_affine_shear_allowed(wm->alpha, wm->beta, wm->gamma, wm->delta))
     366       50471 :     return 0;
     367             : 
     368      779013 :   return 1;
     369             : }
     370             : 
     371           0 : static INLINE int highbd_error_measure(int err, int bd) {
     372           0 :   const int b = bd - 8;
     373           0 :   const int bmask = (1 << b) - 1;
     374           0 :   const int v = (1 << b);
     375           0 :   err = abs(err);
     376           0 :   const int e1 = err >> b;
     377           0 :   const int e2 = err & bmask;
     378           0 :   return error_measure_lut[255 + e1] * (v - e2) +
     379           0 :          error_measure_lut[256 + e1] * e2;
     380             : }
     381             : 
     382             : /* Note: For an explanation of the warp algorithm, and some notes on bit widths
     383             :     for hardware implementations, see the comments above eb_av1_warp_affine_c
     384             : */
     385           0 : void eb_av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref,
     386             :                               int width, int height, int stride, uint16_t *pred,
     387             :                               int p_col, int p_row, int p_width, int p_height,
     388             :                               int p_stride, int subsampling_x,
     389             :                               int subsampling_y, int bd,
     390             :                               ConvolveParams *conv_params, int16_t alpha,
     391             :                               int16_t beta, int16_t gamma, int16_t delta) {
     392             :   int32_t tmp[15 * 8];
     393           0 :   const int reduce_bits_horiz =
     394           0 :       conv_params->round_0 +
     395           0 :       AOMMAX(bd + FILTER_BITS - conv_params->round_0 - 14, 0);
     396           0 :   const int reduce_bits_vert = conv_params->is_compound
     397             :                                    ? conv_params->round_1
     398           0 :                                    : 2 * FILTER_BITS - reduce_bits_horiz;
     399           0 :   const int max_bits_horiz = bd + FILTER_BITS + 1 - reduce_bits_horiz;
     400           0 :   const int offset_bits_horiz = bd + FILTER_BITS - 1;
     401           0 :   const int offset_bits_vert = bd + 2 * FILTER_BITS - reduce_bits_horiz;
     402           0 :   const int round_bits =
     403           0 :       2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
     404           0 :   const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
     405             :   (void)max_bits_horiz;
     406           0 :   assert(IMPLIES(conv_params->is_compound, conv_params->dst != NULL));
     407             : 
     408           0 :   for (int i = p_row; i < p_row + p_height; i += 8) {
     409           0 :     for (int j = p_col; j < p_col + p_width; j += 8) {
     410             :       // Calculate the center of this 8x8 block,
     411             :       // project to luma coordinates (if in a subsampled chroma plane),
     412             :       // apply the affine transformation,
     413             :       // then convert back to the original coordinates (if necessary)
     414           0 :       const int32_t src_x = (j + 4) << subsampling_x;
     415           0 :       const int32_t src_y = (i + 4) << subsampling_y;
     416           0 :       const int32_t dst_x = mat[2] * src_x + mat[3] * src_y + mat[0];
     417           0 :       const int32_t dst_y = mat[4] * src_x + mat[5] * src_y + mat[1];
     418           0 :       const int32_t x4 = dst_x >> subsampling_x;
     419           0 :       const int32_t y4 = dst_y >> subsampling_y;
     420             : 
     421           0 :       const int32_t ix4 = x4 >> WARPEDMODEL_PREC_BITS;
     422           0 :       int32_t sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
     423           0 :       const int32_t iy4 = y4 >> WARPEDMODEL_PREC_BITS;
     424           0 :       int32_t sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
     425             : 
     426           0 :       sx4 += alpha * (-4) + beta * (-4);
     427           0 :       sy4 += gamma * (-4) + delta * (-4);
     428             : 
     429           0 :       sx4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
     430           0 :       sy4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
     431             : 
     432             :       // Horizontal filter
     433           0 :       for (int k = -7; k < 8; ++k) {
     434           0 :         const int iy = clamp(iy4 + k, 0, height - 1);
     435             : 
     436           0 :         int sx = sx4 + beta * (k + 4);
     437           0 :         for (int l = -4; l < 4; ++l) {
     438           0 :           int ix = ix4 + l - 3;
     439           0 :           const int offs = ROUND_POWER_OF_TWO(sx, WARPEDDIFF_PREC_BITS) +
     440             :                            WARPEDPIXEL_PREC_SHIFTS;
     441           0 :           assert(offs >= 0 && offs <= WARPEDPIXEL_PREC_SHIFTS * 3);
     442           0 :           const int16_t *coeffs = eb_warped_filter[offs];
     443             : 
     444           0 :           int32_t sum = 1 << offset_bits_horiz;
     445           0 :           for (int m = 0; m < 8; ++m) {
     446           0 :             const int sample_x = clamp(ix + m, 0, width - 1);
     447           0 :             sum += ref[iy * stride + sample_x] * coeffs[m];
     448             :           }
     449           0 :           sum = ROUND_POWER_OF_TWO(sum, reduce_bits_horiz);
     450           0 :           assert(0 <= sum && sum < (1 << max_bits_horiz));
     451           0 :           tmp[(k + 7) * 8 + (l + 4)] = sum;
     452           0 :           sx += alpha;
     453             :         }
     454             :       }
     455             : 
     456             :       // Vertical filter
     457           0 :       for (int k = -4; k < AOMMIN(4, p_row + p_height - i - 4); ++k) {
     458           0 :         int sy = sy4 + delta * (k + 4);
     459           0 :         for (int l = -4; l < AOMMIN(4, p_col + p_width - j - 4); ++l) {
     460           0 :           const int offs = ROUND_POWER_OF_TWO(sy, WARPEDDIFF_PREC_BITS) +
     461             :                            WARPEDPIXEL_PREC_SHIFTS;
     462           0 :           assert(offs >= 0 && offs <= WARPEDPIXEL_PREC_SHIFTS * 3);
     463           0 :           const int16_t *coeffs = eb_warped_filter[offs];
     464             : 
     465           0 :           int32_t sum = 1 << offset_bits_vert;
     466           0 :           for (int m = 0; m < 8; ++m)
     467           0 :             sum += tmp[(k + m + 4) * 8 + (l + 4)] * coeffs[m];
     468           0 :           if (conv_params->is_compound) {
     469           0 :             ConvBufType *p =
     470             :                 &conv_params
     471           0 :                      ->dst[(i - p_row + k + 4) * conv_params->dst_stride +
     472           0 :                            (j - p_col + l + 4)];
     473           0 :             sum = ROUND_POWER_OF_TWO(sum, reduce_bits_vert);
     474           0 :             if (conv_params->do_average) {
     475           0 :               uint16_t *dst16 =
     476           0 :                   &pred[(i - p_row + k + 4) * p_stride + (j - p_col + l + 4)];
     477           0 :               int32_t tmp32 = *p;
     478           0 :               if (conv_params->use_jnt_comp_avg) {
     479           0 :                 tmp32 = tmp32 * conv_params->fwd_offset +
     480           0 :                         sum * conv_params->bck_offset;
     481           0 :                 tmp32 = tmp32 >> DIST_PRECISION_BITS;
     482             :               } else {
     483           0 :                 tmp32 += sum;
     484           0 :                 tmp32 = tmp32 >> 1;
     485             :               }
     486           0 :               tmp32 = tmp32 - (1 << (offset_bits - conv_params->round_1)) -
     487           0 :                       (1 << (offset_bits - conv_params->round_1 - 1));
     488           0 :               *dst16 =
     489           0 :                   clip_pixel_highbd(ROUND_POWER_OF_TWO(tmp32, round_bits), bd);
     490             :             } else
     491           0 :               *p = sum;
     492             :           } else {
     493           0 :             uint16_t *p =
     494           0 :                 &pred[(i - p_row + k + 4) * p_stride + (j - p_col + l + 4)];
     495           0 :             sum = ROUND_POWER_OF_TWO(sum, reduce_bits_vert);
     496           0 :             assert(0 <= sum && sum < (1 << (bd + 2)));
     497           0 :             *p = clip_pixel_highbd(sum - (1 << (bd - 1)) - (1 << bd), bd);
     498             :           }
     499           0 :           sy += gamma;
     500             :         }
     501             :       }
     502             :     }
     503             :   }
     504           0 : }
     505             : 
     506           0 : static void highbd_warp_plane(EbWarpedMotionParams *wm, const uint8_t *const ref8,
     507             :                               int width, int height, int stride,
     508             :                               const uint8_t *const pred8, int p_col, int p_row,
     509             :                               int p_width, int p_height, int p_stride,
     510             :                               int subsampling_x, int subsampling_y, int bd,
     511             :                               ConvolveParams *conv_params) {
     512           0 :   assert(wm->wmtype <= AFFINE);
     513           0 :   if (wm->wmtype == ROTZOOM) {
     514           0 :     wm->wmmat[5] = wm->wmmat[2];
     515           0 :     wm->wmmat[4] = -wm->wmmat[3];
     516             :   }
     517           0 :   const int32_t *const mat = wm->wmmat;
     518           0 :   const int16_t alpha = wm->alpha;
     519           0 :   const int16_t beta = wm->beta;
     520           0 :   const int16_t gamma = wm->gamma;
     521           0 :   const int16_t delta = wm->delta;
     522             : 
     523           0 :   const uint16_t *const ref = (uint16_t *)ref8;
     524           0 :   uint16_t *pred = (uint16_t *)pred8;
     525           0 :   eb_av1_highbd_warp_affine_c(mat, ref, width, height, stride, pred, p_col, p_row,
     526             :                          p_width, p_height, p_stride, subsampling_x,
     527             :                          subsampling_y, bd, conv_params, alpha, beta, gamma,
     528             :                          delta);
     529           0 : }
     530             : 
     531           0 : static int64_t highbd_frame_error(const uint16_t *const ref, int stride,
     532             :                                   const uint16_t *const dst, int p_width,
     533             :                                   int p_height, int p_stride, int bd) {
     534           0 :   int64_t sum_error = 0;
     535           0 :   for (int i = 0; i < p_height; ++i) {
     536           0 :     for (int j = 0; j < p_width; ++j) {
     537           0 :       sum_error +=
     538           0 :           highbd_error_measure(dst[j + i * p_stride] - ref[j + i * stride], bd);
     539             :     }
     540             :   }
     541           0 :   return sum_error;
     542             : }
     543             : 
     544           0 : static int64_t highbd_warp_error(
     545             :     EbWarpedMotionParams *wm, const uint8_t *const ref8, int width, int height,
     546             :     int stride, const uint8_t *const dst8, int p_col, int p_row, int p_width,
     547             :     int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd,
     548             :     int64_t best_error) {
     549           0 :   int64_t gm_sumerr = 0;
     550           0 :   const int error_bsize_w = AOMMIN(p_width, WARP_ERROR_BLOCK);
     551           0 :   const int error_bsize_h = AOMMIN(p_height, WARP_ERROR_BLOCK);
     552             :   uint16_t tmp[WARP_ERROR_BLOCK * WARP_ERROR_BLOCK];
     553             : 
     554           0 :   ConvolveParams conv_params = get_conv_params(0, 0, 0, bd);
     555           0 :   conv_params.use_jnt_comp_avg = 0;
     556           0 :   for (int i = p_row; i < p_row + p_height; i += WARP_ERROR_BLOCK) {
     557           0 :     for (int j = p_col; j < p_col + p_width; j += WARP_ERROR_BLOCK) {
     558             :       // avoid warping extra 8x8 blocks in the padded region of the frame
     559             :       // when p_width and p_height are not multiples of WARP_ERROR_BLOCK
     560           0 :       const int warp_w = AOMMIN(error_bsize_w, p_col + p_width - j);
     561           0 :       const int warp_h = AOMMIN(error_bsize_h, p_row + p_height - i);
     562           0 :       highbd_warp_plane(wm, ref8, width, height, stride,
     563           0 :                         CONVERT_TO_BYTEPTR(tmp), j, i, warp_w, warp_h,
     564             :                         WARP_ERROR_BLOCK, subsampling_x, subsampling_y, bd,
     565             :                         &conv_params);
     566             : 
     567           0 :       gm_sumerr += highbd_frame_error(
     568           0 :           tmp, WARP_ERROR_BLOCK, CONVERT_TO_SHORTPTR(dst8) + j + i * p_stride,
     569             :           warp_w, warp_h, p_stride, bd);
     570           0 :       if (gm_sumerr > best_error) return gm_sumerr;
     571             :     }
     572             :   }
     573           0 :   return gm_sumerr;
     574             : }
     575             : 
     576             : /* The warp filter for ROTZOOM and AFFINE models works as follows:
     577             :    * Split the input into 8x8 blocks
     578             :    * For each block, project the point (4, 4) within the block, to get the
     579             :      overall block position. Split into integer and fractional coordinates,
     580             :      maintaining full WARPEDMODEL precision
     581             :    * Filter horizontally: Generate 15 rows of 8 pixels each. Each pixel gets a
     582             :      variable horizontal offset. This means that, while the rows of the
     583             :      intermediate buffer align with the rows of the *reference* image, the
     584             :      columns align with the columns of the *destination* image.
     585             :    * Filter vertically: Generate the output block (up to 8x8 pixels, but if the
     586             :      destination is too small we crop the output at this stage). Each pixel has
     587             :      a variable vertical offset, so that the resulting rows are aligned with
     588             :      the rows of the destination image.
     589             : 
     590             :    To accomplish these alignments, we factor the warp matrix as a
     591             :    product of two shear / asymmetric zoom matrices:
     592             :    / a b \  = /   1       0    \ * / 1+alpha  beta \
     593             :    \ c d /    \ gamma  1+delta /   \    0      1   /
     594             :    where a, b, c, d are wmmat[2], wmmat[3], wmmat[4], wmmat[5] respectively.
     595             :    The horizontal shear (with alpha and beta) is applied first,
     596             :    then the vertical shear (with gamma and delta) is applied second.
     597             : 
     598             :    The only limitation is that, to fit this in a fixed 8-tap filter size,
     599             :    the fractional pixel offsets must be at most +-1. Since the horizontal filter
     600             :    generates 15 rows of 8 columns, and the initial point we project is at (4, 4)
     601             :    within the block, the parameters must satisfy
     602             :    4 * |alpha| + 7 * |beta| <= 1   and   4 * |gamma| + 4 * |delta| <= 1
     603             :    for this filter to be applicable.
     604             : 
     605             :    Note: This function assumes that the caller has done all of the relevant
     606             :    checks, ie. that we have a ROTZOOM or AFFINE model, that wm[4] and wm[5]
     607             :    are set appropriately (if using a ROTZOOM model), and that alpha, beta,
     608             :    gamma, delta are all in range.
     609             : 
     610             :    TODO(david.barker): Maybe support scaled references?
     611             : */
     612             : /* A note on hardware implementation:
     613             :     The warp filter is intended to be implementable using the same hardware as
     614             :     the high-precision convolve filters from the loop-restoration and
     615             :     convolve-round experiments.
     616             : 
     617             :     For a single filter stage, considering all of the coefficient sets for the
     618             :     warp filter and the regular convolution filter, an input in the range
     619             :     [0, 2^k - 1] is mapped into the range [-56 * (2^k - 1), 184 * (2^k - 1)]
     620             :     before rounding.
     621             : 
     622             :     Allowing for some changes to the filter coefficient sets, call the range
     623             :     [-64 * 2^k, 192 * 2^k]. Then, if we initialize the accumulator to 64 * 2^k,
     624             :     we can replace this by the range [0, 256 * 2^k], which can be stored in an
     625             :     unsigned value with 8 + k bits.
     626             : 
     627             :     This allows the derivation of the appropriate bit widths and offsets for
     628             :     the various intermediate values: If
     629             : 
     630             :     F := FILTER_BITS = 7 (or else the above ranges need adjusting)
     631             :          So a *single* filter stage maps a k-bit input to a (k + F + 1)-bit
     632             :          intermediate value.
     633             :     H := ROUND0_BITS
     634             :     V := VERSHEAR_REDUCE_PREC_BITS
     635             :     (and note that we must have H + V = 2*F for the output to have the same
     636             :      scale as the input)
     637             : 
     638             :     then we end up with the following offsets and ranges:
     639             :     Horizontal filter: Apply an offset of 1 << (bd + F - 1), sum fits into a
     640             :                        uint{bd + F + 1}
     641             :     After rounding: The values stored in 'tmp' fit into a uint{bd + F + 1 - H}.
     642             :     Vertical filter: Apply an offset of 1 << (bd + 2*F - H), sum fits into a
     643             :                      uint{bd + 2*F + 2 - H}
     644             :     After rounding: The final value, before undoing the offset, fits into a
     645             :                     uint{bd + 2}.
     646             : 
     647             :     Then we need to undo the offsets before clamping to a pixel. Note that,
     648             :     if we do this at the end, the amount to subtract is actually independent
     649             :     of H and V:
     650             : 
     651             :     offset to subtract = (1 << ((bd + F - 1) - H + F - V)) +
     652             :                          (1 << ((bd + 2*F - H) - V))
     653             :                       == (1 << (bd - 1)) + (1 << bd)
     654             : 
     655             :     This allows us to entirely avoid clamping in both the warp filter and
     656             :     the convolve-round experiment. As of the time of writing, the Wiener filter
     657             :     from loop-restoration can encode a central coefficient up to 216, which
     658             :     leads to a maximum value of about 282 * 2^k after applying the offset.
     659             :     So in that case we still need to clamp.
     660             : */
     661           0 : void eb_av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width,
     662             :                        int height, int stride, uint8_t *pred, int p_col,
     663             :                        int p_row, int p_width, int p_height, int p_stride,
     664             :                        int subsampling_x, int subsampling_y,
     665             :                        ConvolveParams *conv_params, int16_t alpha, int16_t beta,
     666             :                        int16_t gamma, int16_t delta) {
     667             :   int32_t tmp[15 * 8];
     668           0 :   const int bd = 8;
     669           0 :   const int reduce_bits_horiz = conv_params->round_0;
     670           0 :   const int reduce_bits_vert = conv_params->is_compound
     671             :                                    ? conv_params->round_1
     672           0 :                                    : 2 * FILTER_BITS - reduce_bits_horiz;
     673           0 :   const int max_bits_horiz = bd + FILTER_BITS + 1 - reduce_bits_horiz;
     674           0 :   const int offset_bits_horiz = bd + FILTER_BITS - 1;
     675           0 :   const int offset_bits_vert = bd + 2 * FILTER_BITS - reduce_bits_horiz;
     676           0 :   const int round_bits =
     677           0 :       2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
     678           0 :   const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
     679             :   (void)max_bits_horiz;
     680           0 :   assert(IMPLIES(conv_params->is_compound, conv_params->dst != NULL));
     681           0 :   assert(IMPLIES(conv_params->do_average, conv_params->is_compound));
     682             : 
     683           0 :   for (int i = p_row; i < p_row + p_height; i += 8) {
     684           0 :     for (int j = p_col; j < p_col + p_width; j += 8) {
     685             :       // Calculate the center of this 8x8 block,
     686             :       // project to luma coordinates (if in a subsampled chroma plane),
     687             :       // apply the affine transformation,
     688             :       // then convert back to the original coordinates (if necessary)
     689           0 :       const int32_t src_x = (j + 4) << subsampling_x;
     690           0 :       const int32_t src_y = (i + 4) << subsampling_y;
     691           0 :       const int32_t dst_x = mat[2] * src_x + mat[3] * src_y + mat[0];
     692           0 :       const int32_t dst_y = mat[4] * src_x + mat[5] * src_y + mat[1];
     693           0 :       const int32_t x4 = dst_x >> subsampling_x;
     694           0 :       const int32_t y4 = dst_y >> subsampling_y;
     695             : 
     696           0 :       int32_t ix4 = x4 >> WARPEDMODEL_PREC_BITS;
     697           0 :       int32_t sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
     698           0 :       int32_t iy4 = y4 >> WARPEDMODEL_PREC_BITS;
     699           0 :       int32_t sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
     700             : 
     701           0 :       sx4 += alpha * (-4) + beta * (-4);
     702           0 :       sy4 += gamma * (-4) + delta * (-4);
     703             : 
     704           0 :       sx4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
     705           0 :       sy4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
     706             : 
     707             :       // Horizontal filter
     708           0 :       for (int k = -7; k < 8; ++k) {
     709             :         // Clamp to top/bottom edge of the frame
     710           0 :         const int iy = clamp(iy4 + k, 0, height - 1);
     711             : 
     712           0 :         int sx = sx4 + beta * (k + 4);
     713             : 
     714           0 :         for (int l = -4; l < 4; ++l) {
     715           0 :           int ix = ix4 + l - 3;
     716             :           // At this point, sx = sx4 + alpha * l + beta * k
     717           0 :           const int offs = ROUND_POWER_OF_TWO(sx, WARPEDDIFF_PREC_BITS) +
     718             :                            WARPEDPIXEL_PREC_SHIFTS;
     719           0 :           assert(offs >= 0 && offs <= WARPEDPIXEL_PREC_SHIFTS * 3);
     720           0 :           const int16_t *coeffs = eb_warped_filter[offs];
     721             : 
     722           0 :           int32_t sum = 1 << offset_bits_horiz;
     723           0 :           for (int m = 0; m < 8; ++m) {
     724             :             // Clamp to left/right edge of the frame
     725           0 :             const int sample_x = clamp(ix + m, 0, width - 1);
     726             : 
     727           0 :             sum += ref[iy * stride + sample_x] * coeffs[m];
     728             :           }
     729           0 :           sum = ROUND_POWER_OF_TWO(sum, reduce_bits_horiz);
     730           0 :           assert(0 <= sum && sum < (1 << max_bits_horiz));
     731           0 :           tmp[(k + 7) * 8 + (l + 4)] = sum;
     732           0 :           sx += alpha;
     733             :         }
     734             :       }
     735             : 
     736             :       // Vertical filter
     737           0 :       for (int k = -4; k < AOMMIN(4, p_row + p_height - i - 4); ++k) {
     738           0 :         int sy = sy4 + delta * (k + 4);
     739           0 :         for (int l = -4; l < AOMMIN(4, p_col + p_width - j - 4); ++l) {
     740             :           // At this point, sy = sy4 + gamma * l + delta * k
     741           0 :           const int offs = ROUND_POWER_OF_TWO(sy, WARPEDDIFF_PREC_BITS) +
     742             :                            WARPEDPIXEL_PREC_SHIFTS;
     743           0 :           assert(offs >= 0 && offs <= WARPEDPIXEL_PREC_SHIFTS * 3);
     744           0 :           const int16_t *coeffs = eb_warped_filter[offs];
     745             : 
     746           0 :           int32_t sum = 1 << offset_bits_vert;
     747           0 :           for (int m = 0; m < 8; ++m)
     748           0 :             sum += tmp[(k + m + 4) * 8 + (l + 4)] * coeffs[m];
     749           0 :           if (conv_params->is_compound) {
     750           0 :             ConvBufType *p =
     751             :                 &conv_params
     752           0 :                      ->dst[(i - p_row + k + 4) * conv_params->dst_stride +
     753           0 :                            (j - p_col + l + 4)];
     754           0 :             sum = ROUND_POWER_OF_TWO(sum, reduce_bits_vert);
     755           0 :             if (conv_params->do_average) {
     756           0 :               uint8_t *dst8 =
     757           0 :                   &pred[(i - p_row + k + 4) * p_stride + (j - p_col + l + 4)];
     758           0 :               int32_t tmp32 = *p;
     759           0 :               if (conv_params->use_jnt_comp_avg) {
     760           0 :                 tmp32 = tmp32 * conv_params->fwd_offset +
     761           0 :                         sum * conv_params->bck_offset;
     762           0 :                 tmp32 = tmp32 >> DIST_PRECISION_BITS;
     763             :               } else {
     764           0 :                 tmp32 += sum;
     765           0 :                 tmp32 = tmp32 >> 1;
     766             :               }
     767           0 :               tmp32 = tmp32 - (1 << (offset_bits - conv_params->round_1)) -
     768           0 :                       (1 << (offset_bits - conv_params->round_1 - 1));
     769           0 :               *dst8 = clip_pixel(ROUND_POWER_OF_TWO(tmp32, round_bits));
     770             :             } else
     771           0 :               *p = sum;
     772             :           } else {
     773           0 :             uint8_t *p =
     774           0 :                 &pred[(i - p_row + k + 4) * p_stride + (j - p_col + l + 4)];
     775           0 :             sum = ROUND_POWER_OF_TWO(sum, reduce_bits_vert);
     776           0 :             assert(0 <= sum && sum < (1 << (bd + 2)));
     777           0 :             *p = clip_pixel(sum - (1 << (bd - 1)) - (1 << bd));
     778             :           }
     779           0 :           sy += gamma;
     780             :         }
     781             :       }
     782             :     }
     783             :   }
     784           0 : }
     785             : 
     786    26608800 : static void warp_plane(EbWarpedMotionParams *wm, const uint8_t *const ref,
     787             :                        int width, int height, int stride, uint8_t *pred,
     788             :                        int p_col, int p_row, int p_width, int p_height,
     789             :                        int p_stride, int subsampling_x, int subsampling_y,
     790             :                        ConvolveParams *conv_params) {
     791    26608800 :   assert(wm->wmtype <= AFFINE);
     792    26608800 :   if (wm->wmtype == ROTZOOM) {
     793    25138600 :     wm->wmmat[5] = wm->wmmat[2];
     794    25138600 :     wm->wmmat[4] = -wm->wmmat[3];
     795             :   }
     796    26608800 :   const int32_t *const mat = wm->wmmat;
     797    26608800 :   const int16_t alpha = wm->alpha;
     798    26608800 :   const int16_t beta = wm->beta;
     799    26608800 :   const int16_t gamma = wm->gamma;
     800    26608800 :   const int16_t delta = wm->delta;
     801    26608800 :   eb_av1_warp_affine(mat, ref, width, height, stride, pred, p_col, p_row, p_width,
     802             :                   p_height, p_stride, subsampling_x, subsampling_y, conv_params,
     803             :                   alpha, beta, gamma, delta);
     804    26608400 : }
     805             : 
     806           0 : int64_t eb_av1_calc_frame_error_c(const uint8_t *const ref, int stride,
     807             :     const uint8_t *const dst, int p_width, int p_height, int p_stride) {
     808           0 :   int64_t sum_error = 0;
     809           0 :   for (int i = 0; i < p_height; ++i) {
     810           0 :     for (int j = 0; j < p_width; ++j) {
     811           0 :       sum_error +=
     812           0 :           (int64_t)error_measure(dst[j + i * p_stride] - ref[j + i * stride]);
     813             :     }
     814             :   }
     815           0 :   return sum_error;
     816             : }
     817             : 
     818       91233 : static int64_t warp_error(EbWarpedMotionParams *wm, const uint8_t *const ref,
     819             :                           int width, int height, int stride,
     820             :                           const uint8_t *const dst, int p_col, int p_row,
     821             :                           int p_width, int p_height, int p_stride,
     822             :                           int subsampling_x, int subsampling_y,
     823             :                           int64_t best_error) {
     824       91233 :   int64_t gm_sumerr = 0;
     825             :   int warp_w, warp_h;
     826       91233 :   int error_bsize_w = AOMMIN(p_width, WARP_ERROR_BLOCK);
     827       91233 :   int error_bsize_h = AOMMIN(p_height, WARP_ERROR_BLOCK);
     828             :   uint8_t tmp[WARP_ERROR_BLOCK * WARP_ERROR_BLOCK];
     829       91233 :   ConvolveParams conv_params = get_conv_params(0, 0, 0, 8);
     830       91233 :   conv_params.use_jnt_comp_avg = 0;
     831             : 
     832      997091 :   for (int i = p_row; i < p_row + p_height; i += WARP_ERROR_BLOCK) {
     833    19931400 :     for (int j = p_col; j < p_col + p_width; j += WARP_ERROR_BLOCK) {
     834             :       // avoid warping extra 8x8 blocks in the padded region of the frame
     835             :       // when p_width and p_height are not multiples of WARP_ERROR_BLOCK
     836    19025600 :       warp_w = AOMMIN(error_bsize_w, p_col + p_width - j);
     837    19025600 :       warp_h = AOMMIN(error_bsize_h, p_row + p_height - i);
     838    19025600 :       warp_plane(wm, ref, width, height, stride, tmp, j, i, warp_w, warp_h,
     839             :                  WARP_ERROR_BLOCK, subsampling_x, subsampling_y, &conv_params);
     840             : 
     841    19024300 :       gm_sumerr += eb_av1_calc_frame_error(tmp, WARP_ERROR_BLOCK, dst + j + i * p_stride,
     842             :                                warp_w, warp_h, p_stride);
     843    19025500 :       if (gm_sumerr > best_error) return gm_sumerr;
     844             :     }
     845             :   }
     846        2214 :   return gm_sumerr;
     847             : }
     848             : 
     849         113 : int64_t eb_av1_frame_error(int use_hbd, int bd, const uint8_t *ref, int stride,
     850             :                         uint8_t *dst, int p_width, int p_height, int p_stride) {
     851         113 :   if (use_hbd) {
     852           0 :     return highbd_frame_error(CONVERT_TO_SHORTPTR(ref), stride,
     853           0 :                               CONVERT_TO_SHORTPTR(dst), p_width, p_height,
     854             :                               p_stride, bd);
     855             :   }
     856         113 :   return eb_av1_calc_frame_error(ref, stride, dst, p_width, p_height, p_stride);
     857             : }
     858             : 
     859       91233 : int64_t eb_av1_warp_error(EbWarpedMotionParams *wm, int use_hbd, int bd,
     860             :                        const uint8_t *ref, int width, int height, int stride,
     861             :                        uint8_t *dst, int p_col, int p_row, int p_width,
     862             :                        int p_height, int p_stride, int subsampling_x,
     863             :                        int subsampling_y, int64_t best_error) {
     864       91233 :   if (wm->wmtype <= AFFINE)
     865       91233 :     if (!eb_get_shear_params(wm)) return 1;
     866       91233 :   if (use_hbd)
     867           0 :     return highbd_warp_error(wm, ref, width, height, stride, dst, p_col, p_row,
     868             :                              p_width, p_height, p_stride, subsampling_x,
     869             :                              subsampling_y, bd, best_error);
     870       91233 :   return warp_error(wm, ref, width, height, stride, dst, p_col, p_row, p_width,
     871             :                     p_height, p_stride, subsampling_x, subsampling_y,
     872             :                     best_error);
     873             : }
     874             : 
     875     7584170 : void eb_av1_warp_plane(EbWarpedMotionParams *wm, int use_hbd, int bd,
     876             :                     const uint8_t *ref, int width, int height, int stride,
     877             :                     uint8_t *pred, int p_col, int p_row, int p_width,
     878             :                     int p_height, int p_stride, int subsampling_x,
     879             :                     int subsampling_y, ConvolveParams *conv_params) {
     880     7584170 :   if (use_hbd)
     881           0 :     highbd_warp_plane(wm, ref, width, height, stride, pred, p_col, p_row,
     882             :                       p_width, p_height, p_stride, subsampling_x, subsampling_y,
     883             :                       bd, conv_params);
     884             :   else
     885     7584170 :     warp_plane(wm, ref, width, height, stride, pred, p_col, p_row, p_width,
     886             :                p_height, p_stride, subsampling_x, subsampling_y, conv_params);
     887     7585150 : }
     888             : 
     889             : #define LS_MV_MAX 256  // max mv in 1/8-pel
     890             : // Use LS_STEP = 8 so that 2 less bits needed for A, Bx, By.
     891             : #define LS_STEP 8
     892             : 
     893             : // Assuming LS_MV_MAX is < MAX_SB_SIZE * 8,
     894             : // the precision needed is:
     895             : //   (MAX_SB_SIZE_LOG2 + 3) [for sx * sx magnitude] +
     896             : //   (MAX_SB_SIZE_LOG2 + 4) [for sx * dx magnitude] +
     897             : //   1 [for sign] +
     898             : //   LEAST_SQUARES_SAMPLES_MAX_BITS
     899             : //        [for adding up to LEAST_SQUARES_SAMPLES_MAX samples]
     900             : // The value is 23
     901             : #define LS_MAT_RANGE_BITS \
     902             :   ((MAX_SB_SIZE_LOG2 + 4) * 2 + LEAST_SQUARES_SAMPLES_MAX_BITS)
     903             : 
     904             : // Bit-depth reduction from the full-range
     905             : #define LS_MAT_DOWN_BITS 2
     906             : 
     907             : // bits range of A, Bx and By after downshifting
     908             : #define LS_MAT_BITS (LS_MAT_RANGE_BITS - LS_MAT_DOWN_BITS)
     909             : #define LS_MAT_MIN (-(1 << (LS_MAT_BITS - 1)))
     910             : #define LS_MAT_MAX ((1 << (LS_MAT_BITS - 1)) - 1)
     911             : 
     912             : // By setting LS_STEP = 8, the least 2 bits of every elements in A, Bx, By are
     913             : // 0. So, we can reduce LS_MAT_RANGE_BITS(2) bits here.
     914             : #define LS_SQUARE(a)                                          \
     915             :   (((a) * (a)*4 + (a)*4 * LS_STEP + LS_STEP * LS_STEP * 2) >> \
     916             :    (2 + LS_MAT_DOWN_BITS))
     917             : #define LS_PRODUCT1(a, b)                                           \
     918             :   (((a) * (b)*4 + ((a) + (b)) * 2 * LS_STEP + LS_STEP * LS_STEP) >> \
     919             :    (2 + LS_MAT_DOWN_BITS))
     920             : #define LS_PRODUCT2(a, b)                                               \
     921             :   (((a) * (b)*4 + ((a) + (b)) * 2 * LS_STEP + LS_STEP * LS_STEP * 2) >> \
     922             :    (2 + LS_MAT_DOWN_BITS))
     923             : 
     924             : #define USE_LIMITED_PREC_MULT 0
     925             : 
     926             : #if USE_LIMITED_PREC_MULT
     927             : 
     928             : #define MUL_PREC_BITS 16
     929             : static uint16_t resolve_multiplier_64(uint64_t D, int16_t *shift) {
     930             :   int msb = 0;
     931             :   uint16_t mult = 0;
     932             :   *shift = 0;
     933             :   if (D != 0) {
     934             :     msb = (int16_t)((D >> 32) ? get_msb((unsigned int)(D >> 32)) + 32
     935             :                               : get_msb((unsigned int)D));
     936             :     if (msb >= MUL_PREC_BITS) {
     937             :       mult = (uint16_t)ROUND_POWER_OF_TWO_64(D, msb + 1 - MUL_PREC_BITS);
     938             :       *shift = msb + 1 - MUL_PREC_BITS;
     939             :     } else {
     940             :       mult = (uint16_t)D;
     941             :       *shift = 0;
     942             :     }
     943             :   }
     944             :   return mult;
     945             : }
     946             : 
     947             : static int32_t get_mult_shift_ndiag(int64_t Px, int16_t iDet, int shift) {
     948             :   int32_t ret;
     949             :   int16_t mshift;
     950             :   uint16_t Mul = resolve_multiplier_64(llabs(Px), &mshift);
     951             :   int32_t v = (int32_t)Mul * (int32_t)iDet * (Px < 0 ? -1 : 1);
     952             :   shift -= mshift;
     953             :   if (shift > 0) {
     954             :     return (int32_t)clamp(ROUND_POWER_OF_TWO_SIGNED(v, shift),
     955             :                           -WARPEDMODEL_NONDIAGAFFINE_CLAMP + 1,
     956             :                           WARPEDMODEL_NONDIAGAFFINE_CLAMP - 1);
     957             :   } else {
     958             :     return (int32_t)clamp(v * (1 << (-shift)),
     959             :                           -WARPEDMODEL_NONDIAGAFFINE_CLAMP + 1,
     960             :                           WARPEDMODEL_NONDIAGAFFINE_CLAMP - 1);
     961             :   }
     962             :   return ret;
     963             : }
     964             : 
     965             : static int32_t get_mult_shift_diag(int64_t Px, int16_t iDet, int shift) {
     966             :   int16_t mshift;
     967             :   uint16_t Mul = resolve_multiplier_64(llabs(Px), &mshift);
     968             :   int32_t v = (int32_t)Mul * (int32_t)iDet * (Px < 0 ? -1 : 1);
     969             :   shift -= mshift;
     970             :   if (shift > 0) {
     971             :     return (int32_t)clamp(
     972             :         ROUND_POWER_OF_TWO_SIGNED(v, shift),
     973             :         (1 << WARPEDMODEL_PREC_BITS) - WARPEDMODEL_NONDIAGAFFINE_CLAMP + 1,
     974             :         (1 << WARPEDMODEL_PREC_BITS) + WARPEDMODEL_NONDIAGAFFINE_CLAMP - 1);
     975             :   } else {
     976             :     return (int32_t)clamp(
     977             :         v * (1 << (-shift)),
     978             :         (1 << WARPEDMODEL_PREC_BITS) - WARPEDMODEL_NONDIAGAFFINE_CLAMP + 1,
     979             :         (1 << WARPEDMODEL_PREC_BITS) + WARPEDMODEL_NONDIAGAFFINE_CLAMP - 1);
     980             :   }
     981             : }
     982             : 
     983             : #else
     984             : 
     985     1475510 : static int32_t get_mult_shift_ndiag(int64_t Px, int16_t iDet, int shift) {
     986     1475510 :   int64_t v = Px * (int64_t)iDet;
     987     1475510 :   return (int32_t)clamp64(ROUND_POWER_OF_TWO_SIGNED_64(v, shift),
     988             :                           -WARPEDMODEL_NONDIAGAFFINE_CLAMP + 1,
     989             :                           WARPEDMODEL_NONDIAGAFFINE_CLAMP - 1);
     990             : }
     991             : 
     992     1475490 : static int32_t get_mult_shift_diag(int64_t Px, int16_t iDet, int shift) {
     993     1475490 :   int64_t v = Px * (int64_t)iDet;
     994     2950980 :   return (int32_t)clamp64(
     995     1475490 :       ROUND_POWER_OF_TWO_SIGNED_64(v, shift),
     996             :       (1 << WARPEDMODEL_PREC_BITS) - WARPEDMODEL_NONDIAGAFFINE_CLAMP + 1,
     997             :       (1 << WARPEDMODEL_PREC_BITS) + WARPEDMODEL_NONDIAGAFFINE_CLAMP - 1);
     998             : }
     999             : #endif  // USE_LIMITED_PREC_MULT
    1000             : 
    1001      773046 : static int find_affine_int(int np, const int *pts1, const int *pts2,
    1002             :                            BlockSize bsize, int mvy, int mvx,
    1003             :                            EbWarpedMotionParams *wm, int mi_row, int mi_col) {
    1004      773046 :   int32_t A[2][2] = { { 0, 0 }, { 0, 0 } };
    1005      773046 :   int32_t Bx[2] = { 0, 0 };
    1006      773046 :   int32_t By[2] = { 0, 0 };
    1007             :   int i;
    1008             : 
    1009      773046 :   const int bw = block_size_wide[bsize];
    1010      773046 :   const int bh = block_size_high[bsize];
    1011      773046 :   const int rsuy = (AOMMAX(bh, MI_SIZE) / 2 - 1);
    1012      773046 :   const int rsux = (AOMMAX(bw, MI_SIZE) / 2 - 1);
    1013      773046 :   const int suy = rsuy * 8;
    1014      773046 :   const int sux = rsux * 8;
    1015      773046 :   const int duy = suy + mvy;
    1016      773046 :   const int dux = sux + mvx;
    1017      773046 :   const int isuy = (mi_row * MI_SIZE + rsuy);
    1018      773046 :   const int isux = (mi_col * MI_SIZE + rsux);
    1019             : 
    1020             :   // Assume the center pixel of the block has exactly the same motion vector
    1021             :   // as transmitted for the block. First shift the origin of the source
    1022             :   // points to the block center, and the origin of the destination points to
    1023             :   // the block center added to the motion vector transmitted.
    1024             :   // Let (xi, yi) denote the source points and (xi', yi') denote destination
    1025             :   // points after origin shfifting, for i = 0, 1, 2, .... n-1.
    1026             :   // Then if  P = [x0, y0,
    1027             :   //               x1, y1
    1028             :   //               x2, y1,
    1029             :   //                ....
    1030             :   //              ]
    1031             :   //          q = [x0', x1', x2', ... ]'
    1032             :   //          r = [y0', y1', y2', ... ]'
    1033             :   // the least squares problems that need to be solved are:
    1034             :   //          [h1, h2]' = inv(P'P)P'q and
    1035             :   //          [h3, h4]' = inv(P'P)P'r
    1036             :   // where the affine transformation is given by:
    1037             :   //          x' = h1.x + h2.y
    1038             :   //          y' = h3.x + h4.y
    1039             :   //
    1040             :   // The loop below computes: A = P'P, Bx = P'q, By = P'r
    1041             :   // We need to just compute inv(A).Bx and inv(A).By for the solutions.
    1042             :   // Contribution from neighbor block
    1043     2269110 :   for (i = 0; i < np; i++) {
    1044     1496060 :     const int dx = pts2[i * 2] - dux;
    1045     1496060 :     const int dy = pts2[i * 2 + 1] - duy;
    1046     1496060 :     const int sx = pts1[i * 2] - sux;
    1047     1496060 :     const int sy = pts1[i * 2 + 1] - suy;
    1048             :     // (TODO)yunqing: This comparison wouldn't be necessary if the sample
    1049             :     // selection is done in find_samples(). Also, global offset can be removed
    1050             :     // while collecting samples.
    1051     1496060 :     if (abs(sx - dx) < LS_MV_MAX && abs(sy - dy) < LS_MV_MAX) {
    1052     1460760 :       A[0][0] += LS_SQUARE(sx);
    1053     1460760 :       A[0][1] += LS_PRODUCT1(sx, sy);
    1054     1460760 :       A[1][1] += LS_SQUARE(sy);
    1055     1460760 :       Bx[0] += LS_PRODUCT2(sx, dx);
    1056     1460760 :       Bx[1] += LS_PRODUCT1(sy, dx);
    1057     1460760 :       By[0] += LS_PRODUCT1(sx, dy);
    1058     1460760 :       By[1] += LS_PRODUCT2(sy, dy);
    1059             :     }
    1060             :   }
    1061             : 
    1062             :   // Just for debugging, and can be removed later.
    1063      773046 :   assert(A[0][0] >= LS_MAT_MIN && A[0][0] <= LS_MAT_MAX);
    1064      773065 :   assert(A[0][1] >= LS_MAT_MIN && A[0][1] <= LS_MAT_MAX);
    1065      773071 :   assert(A[1][1] >= LS_MAT_MIN && A[1][1] <= LS_MAT_MAX);
    1066      773072 :   assert(Bx[0] >= LS_MAT_MIN && Bx[0] <= LS_MAT_MAX);
    1067      773073 :   assert(Bx[1] >= LS_MAT_MIN && Bx[1] <= LS_MAT_MAX);
    1068      773076 :   assert(By[0] >= LS_MAT_MIN && By[0] <= LS_MAT_MAX);
    1069      773077 :   assert(By[1] >= LS_MAT_MIN && By[1] <= LS_MAT_MAX);
    1070             : 
    1071             :   int64_t Det;
    1072             :   int16_t iDet, shift;
    1073             : 
    1074             :   // Compute Determinant of A
    1075      773077 :   Det = (int64_t)A[0][0] * A[1][1] - (int64_t)A[0][1] * A[0][1];
    1076      773077 :   if (Det == 0) return 1;
    1077      737786 :   iDet = resolve_divisor_64(llabs(Det), &shift) * (Det < 0 ? -1 : 1);
    1078      737780 :   shift -= WARPEDMODEL_PREC_BITS;
    1079      737780 :   if (shift < 0) {
    1080           0 :     iDet <<= (-shift);
    1081           0 :     shift = 0;
    1082             :   }
    1083             : 
    1084             :   int64_t Px[2], Py[2];
    1085             : 
    1086             :   // These divided by the Det, are the least squares solutions
    1087      737780 :   Px[0] = (int64_t)A[1][1] * Bx[0] - (int64_t)A[0][1] * Bx[1];
    1088      737780 :   Px[1] = -(int64_t)A[0][1] * Bx[0] + (int64_t)A[0][0] * Bx[1];
    1089      737780 :   Py[0] = (int64_t)A[1][1] * By[0] - (int64_t)A[0][1] * By[1];
    1090      737780 :   Py[1] = -(int64_t)A[0][1] * By[0] + (int64_t)A[0][0] * By[1];
    1091             : 
    1092      737780 :   wm->wmmat[2] = get_mult_shift_diag(Px[0], iDet, shift);
    1093      737773 :   wm->wmmat[3] = get_mult_shift_ndiag(Px[1], iDet, shift);
    1094      737763 :   wm->wmmat[4] = get_mult_shift_ndiag(Py[0], iDet, shift);
    1095      737746 :   wm->wmmat[5] = get_mult_shift_diag(Py[1], iDet, shift);
    1096             : 
    1097             :   // Note: In the vx, vy expressions below, the max value of each of the
    1098             :   // 2nd and 3rd terms are (2^16 - 1) * (2^13 - 1). That leaves enough room
    1099             :   // for the first term so that the overall sum in the worst case fits
    1100             :   // within 32 bits overall.
    1101      737746 :   int32_t vx = mvx * (1 << (WARPEDMODEL_PREC_BITS - 3)) -
    1102      737746 :                (isux * (wm->wmmat[2] - (1 << WARPEDMODEL_PREC_BITS)) +
    1103      737746 :                 isuy * wm->wmmat[3]);
    1104      737746 :   int32_t vy = mvy * (1 << (WARPEDMODEL_PREC_BITS - 3)) -
    1105      737746 :                (isux * wm->wmmat[4] +
    1106      737746 :                 isuy * (wm->wmmat[5] - (1 << WARPEDMODEL_PREC_BITS)));
    1107      737747 :   wm->wmmat[0] =
    1108      737746 :       clamp(vx, -WARPEDMODEL_TRANS_CLAMP, WARPEDMODEL_TRANS_CLAMP - 1);
    1109      737737 :   wm->wmmat[1] =
    1110      737747 :       clamp(vy, -WARPEDMODEL_TRANS_CLAMP, WARPEDMODEL_TRANS_CLAMP - 1);
    1111             : 
    1112      737737 :   wm->wmmat[6] = wm->wmmat[7] = 0;
    1113      737737 :   return 0;
    1114             : }
    1115             : 
    1116      773044 : EbBool eb_find_projection(
    1117             :     int np,
    1118             :     int *pts1,
    1119             :     int *pts2,
    1120             :     BlockSize bsize,
    1121             :     int mvy,
    1122             :     int mvx,
    1123             :     EbWarpedMotionParams *wm_params,
    1124             :     int mi_row,
    1125             :     int mi_col)
    1126             : {
    1127      773044 :     if (find_affine_int(np, pts1, pts2, bsize, mvy, mvx, wm_params,
    1128             :                         mi_row, mi_col))
    1129             :     {
    1130       35291 :         return 1;
    1131             :     }
    1132             : 
    1133             :     // check compatibility with the fast warp filter
    1134      737737 :     if (!eb_get_shear_params(wm_params))
    1135       50471 :         return 1;
    1136             : 
    1137      687261 :     return 0;
    1138             : }

Generated by: LCOV version 1.14