vp9/encoder/vp9_resize.c

   1 /*
   2  *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
   3  *
   4  *  Use of this source code is governed by a BSD-style license
   5  *  that can be found in the LICENSE file in the root of the source
   6  *  tree. An additional intellectual property rights grant can be found
   7  *  in the file PATENTS.  All contributing project authors may
   8  *  be found in the AUTHORS file in the root of the source tree.
   9  */
  10
  11 #include <assert.h>
  12 #include <limits.h>
  13 #include <math.h>
  14 #include <stdio.h>
  15 #include <stdlib.h>
  16 #include <string.h>
  17
  18 #include "vpx_ports/mem.h"
  19 #include "vp9/common/vp9_common.h"
  20 #include "vp9/encoder/vp9_resize.h"
  21
  22 #define FILTER_BITS               7
  23
  24 #define INTERP_TAPS               8
  25 #define SUBPEL_BITS               5
  26 #define SUBPEL_MASK               ((1 << SUBPEL_BITS) - 1)
  27 #define INTERP_PRECISION_BITS     32
  28
  29 typedef int16_t interp_kernel[INTERP_TAPS];
  30
  31 // Filters for interpolation (0.5-band) - note this also filters integer pels.
  32 static const interp_kernel filteredinterp_filters500[(1 << SUBPEL_BITS)] = {
  33   {-3,  0, 35, 64, 35,  0, -3, 0},
  34   {-3, -1, 34, 64, 36,  1, -3, 0},
  35   {-3, -1, 32, 64, 38,  1, -3, 0},
  36   {-2, -2, 31, 63, 39,  2, -3, 0},
  37   {-2, -2, 29, 63, 41,  2, -3, 0},
  38   {-2, -2, 28, 63, 42,  3, -4, 0},
  39   {-2, -3, 27, 63, 43,  4, -4, 0},
  40   {-2, -3, 25, 62, 45,  5, -4, 0},
  41   {-2, -3, 24, 62, 46,  5, -4, 0},
  42   {-2, -3, 23, 61, 47,  6, -4, 0},
  43   {-2, -3, 21, 60, 49,  7, -4, 0},
  44   {-1, -4, 20, 60, 50,  8, -4, -1},
  45   {-1, -4, 19, 59, 51,  9, -4, -1},
  46   {-1, -4, 17, 58, 52, 10, -4, 0},
  47   {-1, -4, 16, 57, 53, 12, -4, -1},
  48   {-1, -4, 15, 56, 54, 13, -4, -1},
  49   {-1, -4, 14, 55, 55, 14, -4, -1},
  50   {-1, -4, 13, 54, 56, 15, -4, -1},
  51   {-1, -4, 12, 53, 57, 16, -4, -1},
  52   {0, -4, 10, 52, 58, 17, -4, -1},
  53   {-1, -4,  9, 51, 59, 19, -4, -1},
  54   {-1, -4,  8, 50, 60, 20, -4, -1},
  55   {0, -4,  7, 49, 60, 21, -3, -2},
  56   {0, -4,  6, 47, 61, 23, -3, -2},
  57   {0, -4,  5, 46, 62, 24, -3, -2},
  58   {0, -4,  5, 45, 62, 25, -3, -2},
  59   {0, -4,  4, 43, 63, 27, -3, -2},
  60   {0, -4,  3, 42, 63, 28, -2, -2},
  61   {0, -3,  2, 41, 63, 29, -2, -2},
  62   {0, -3,  2, 39, 63, 31, -2, -2},
  63   {0, -3,  1, 38, 64, 32, -1, -3},
  64   {0, -3,  1, 36, 64, 34, -1, -3}
  65 };
  66
  67 // Filters for interpolation (0.625-band) - note this also filters integer pels.
  68 static const interp_kernel filteredinterp_filters625[(1 << SUBPEL_BITS)] = {
  69   {-1, -8, 33, 80, 33, -8, -1, 0},
  70   {-1, -8, 30, 80, 35, -8, -1, 1},
  71   {-1, -8, 28, 80, 37, -7, -2, 1},
  72   {0, -8, 26, 79, 39, -7, -2, 1},
  73   {0, -8, 24, 79, 41, -7, -2, 1},
  74   {0, -8, 22, 78, 43, -6, -2, 1},
  75   {0, -8, 20, 78, 45, -5, -3, 1},
  76   {0, -8, 18, 77, 48, -5, -3, 1},
  77   {0, -8, 16, 76, 50, -4, -3, 1},
  78   {0, -8, 15, 75, 52, -3, -4, 1},
  79   {0, -7, 13, 74, 54, -3, -4, 1},
  80   {0, -7, 11, 73, 56, -2, -4, 1},
  81   {0, -7, 10, 71, 58, -1, -4, 1},
  82   {1, -7,  8, 70, 60,  0, -5, 1},
  83   {1, -6,  6, 68, 62,  1, -5, 1},
  84   {1, -6,  5, 67, 63,  2, -5, 1},
  85   {1, -6,  4, 65, 65,  4, -6, 1},
  86   {1, -5,  2, 63, 67,  5, -6, 1},
  87   {1, -5,  1, 62, 68,  6, -6, 1},
  88   {1, -5,  0, 60, 70,  8, -7, 1},
  89   {1, -4, -1, 58, 71, 10, -7, 0},
  90   {1, -4, -2, 56, 73, 11, -7, 0},
  91   {1, -4, -3, 54, 74, 13, -7, 0},
  92   {1, -4, -3, 52, 75, 15, -8, 0},
  93   {1, -3, -4, 50, 76, 16, -8, 0},
  94   {1, -3, -5, 48, 77, 18, -8, 0},
  95   {1, -3, -5, 45, 78, 20, -8, 0},
  96   {1, -2, -6, 43, 78, 22, -8, 0},
  97   {1, -2, -7, 41, 79, 24, -8, 0},
  98   {1, -2, -7, 39, 79, 26, -8, 0},
  99   {1, -2, -7, 37, 80, 28, -8, -1},
 100   {1, -1, -8, 35, 80, 30, -8, -1},
 101 };
 102
 103 // Filters for interpolation (0.75-band) - note this also filters integer pels.
 104 static const interp_kernel filteredinterp_filters750[(1 << SUBPEL_BITS)] = {
 105   {2, -11,  25,  96,  25, -11,   2, 0},
 106   {2, -11,  22,  96,  28, -11,   2, 0},
 107   {2, -10,  19,  95,  31, -11,   2, 0},
 108   {2, -10,  17,  95,  34, -12,   2, 0},
 109   {2,  -9,  14,  94,  37, -12,   2, 0},
 110   {2,  -8,  12,  93,  40, -12,   1, 0},
 111   {2,  -8,   9,  92,  43, -12,   1, 1},
 112   {2,  -7,   7,  91,  46, -12,   1, 0},
 113   {2,  -7,   5,  90,  49, -12,   1, 0},
 114   {2,  -6,   3,  88,  52, -12,   0, 1},
 115   {2,  -5,   1,  86,  55, -12,   0, 1},
 116   {2,  -5,  -1,  84,  58, -11,   0, 1},
 117   {2,  -4,  -2,  82,  61, -11,  -1, 1},
 118   {2,  -4,  -4,  80,  64, -10,  -1, 1},
 119   {1, -3, -5, 77, 67, -9, -1, 1},
 120   {1, -3, -6, 75, 70, -8, -2, 1},
 121   {1, -2, -7, 72, 72, -7, -2, 1},
 122   {1, -2, -8, 70, 75, -6, -3, 1},
 123   {1, -1, -9, 67, 77, -5, -3, 1},
 124   {1,  -1, -10,  64,  80,  -4,  -4, 2},
 125   {1,  -1, -11,  61,  82,  -2,  -4, 2},
 126   {1,   0, -11,  58,  84,  -1,  -5, 2},
 127   {1,   0, -12,  55,  86,   1,  -5, 2},
 128   {1,   0, -12,  52,  88,   3,  -6, 2},
 129   {0,   1, -12,  49,  90,   5,  -7, 2},
 130   {0,   1, -12,  46,  91,   7,  -7, 2},
 131   {1,   1, -12,  43,  92,   9,  -8, 2},
 132   {0,   1, -12,  40,  93,  12,  -8, 2},
 133   {0,   2, -12,  37,  94,  14,  -9, 2},
 134   {0,   2, -12,  34,  95,  17, -10, 2},
 135   {0,   2, -11,  31,  95,  19, -10, 2},
 136   {0,   2, -11,  28,  96,  22, -11, 2}
 137 };
 138
 139 // Filters for interpolation (0.875-band) - note this also filters integer pels.
 140 static const interp_kernel filteredinterp_filters875[(1 << SUBPEL_BITS)] = {
 141   {3,  -8,  13, 112,  13,  -8,   3, 0},
 142   {3,  -7,  10, 112,  17,  -9,   3, -1},
 143   {2,  -6,   7, 111,  21,  -9,   3, -1},
 144   {2,  -5,   4, 111,  24, -10,   3, -1},
 145   {2,  -4,   1, 110,  28, -11,   3, -1},
 146   {1,  -3,  -1, 108,  32, -12,   4, -1},
 147   {1,  -2,  -3, 106,  36, -13,   4, -1},
 148   {1,  -1,  -6, 105,  40, -14,   4, -1},
 149   {1,  -1,  -7, 102,  44, -14,   4, -1},
 150   {1,   0,  -9, 100,  48, -15,   4, -1},
 151   {1,   1, -11,  97,  53, -16,   4, -1},
 152   {0,   1, -12,  95,  57, -16,   4, -1},
 153   {0,   2, -13,  91,  61, -16,   4, -1},
 154   {0,   2, -14,  88,  65, -16,   4, -1},
 155   {0,   3, -15,  84,  69, -17,   4, 0},
 156   {0,   3, -16,  81,  73, -16,   3, 0},
 157   {0,   3, -16,  77,  77, -16,   3, 0},
 158   {0,   3, -16,  73,  81, -16,   3, 0},
 159   {0,   4, -17,  69,  84, -15,   3, 0},
 160   {-1,   4, -16,  65,  88, -14,   2, 0},
 161   {-1,   4, -16,  61,  91, -13,   2, 0},
 162   {-1,   4, -16,  57,  95, -12,   1, 0},
 163   {-1,   4, -16,  53,  97, -11,   1, 1},
 164   {-1,   4, -15,  48, 100,  -9,   0, 1},
 165   {-1,   4, -14,  44, 102,  -7,  -1, 1},
 166   {-1,   4, -14,  40, 105,  -6,  -1, 1},
 167   {-1,   4, -13,  36, 106,  -3,  -2, 1},
 168   {-1,   4, -12,  32, 108,  -1,  -3, 1},
 169   {-1,   3, -11,  28, 110,   1,  -4, 2},
 170   {-1,   3, -10,  24, 111,   4,  -5, 2},
 171   {-1,   3,  -9,  21, 111,   7,  -6, 2},
 172   {-1,   3,  -9,  17, 112,  10,  -7, 3}
 173 };
 174
 175 // Filters for interpolation (full-band) - no filtering for integer pixels
 176 static const interp_kernel filteredinterp_filters1000[(1 << SUBPEL_BITS)] = {
 177   {0,   0,   0, 128,   0,   0,   0, 0},
 178   {0,   1,  -3, 128,   3,  -1,   0, 0},
 179   {-1,   2,  -6, 127,   7,  -2,   1, 0},
 180   {-1,   3,  -9, 126,  12,  -4,   1, 0},
 181   {-1,   4, -12, 125,  16,  -5,   1, 0},
 182   {-1,   4, -14, 123,  20,  -6,   2, 0},
 183   {-1,   5, -15, 120,  25,  -8,   2, 0},
 184   {-1,   5, -17, 118,  30,  -9,   3, -1},
 185   {-1,   6, -18, 114,  35, -10,   3, -1},
 186   {-1,   6, -19, 111,  41, -12,   3, -1},
 187   {-1,   6, -20, 107,  46, -13,   4, -1},
 188   {-1,   6, -21, 103,  52, -14,   4, -1},
 189   {-1,   6, -21,  99,  57, -16,   5, -1},
 190   {-1,   6, -21,  94,  63, -17,   5, -1},
 191   {-1,   6, -20,  89,  68, -18,   5, -1},
 192   {-1,   6, -20,  84,  73, -19,   6, -1},
 193   {-1,   6, -20,  79,  79, -20,   6, -1},
 194   {-1,   6, -19,  73,  84, -20,   6, -1},
 195   {-1,   5, -18,  68,  89, -20,   6, -1},
 196   {-1,   5, -17,  63,  94, -21,   6, -1},
 197   {-1,   5, -16,  57,  99, -21,   6, -1},
 198   {-1,   4, -14,  52, 103, -21,   6, -1},
 199   {-1,   4, -13,  46, 107, -20,   6, -1},
 200   {-1,   3, -12,  41, 111, -19,   6, -1},
 201   {-1,   3, -10,  35, 114, -18,   6, -1},
 202   {-1,   3,  -9,  30, 118, -17,   5, -1},
 203   {0,   2,  -8,  25, 120, -15,   5, -1},
 204   {0,   2,  -6,  20, 123, -14,   4, -1},
 205   {0,   1,  -5,  16, 125, -12,   4, -1},
 206   {0,   1,  -4,  12, 126,  -9,   3, -1},
 207   {0,   1,  -2,   7, 127,  -6,   2, -1},
 208   {0,   0,  -1,   3, 128,  -3,   1, 0}
 209 };
 210
 211 // Filters for factor of 2 downsampling.
 212 static const int16_t vp9_down2_symeven_half_filter[] = {56, 12, -3, -1};
 213 static const int16_t vp9_down2_symodd_half_filter[] = {64, 35, 0, -3};
 214
 215 static const interp_kernel *choose_interp_filter(int inlength, int outlength) {
 216   int outlength16 = outlength * 16;
 217   if (outlength16 >= inlength * 16)
 218     return filteredinterp_filters1000;
 219   else if (outlength16 >= inlength * 13)
 220     return filteredinterp_filters875;
 221   else if (outlength16 >= inlength * 11)
 222     return filteredinterp_filters750;
 223   else if (outlength16 >= inlength * 9)
 224     return filteredinterp_filters625;
 225   else
 226     return filteredinterp_filters500;
 227 }
 228
 229 static void interpolate(const uint8_t *const input, int inlength,
 230                         uint8_t *output, int outlength) {
 231   const int64_t delta = (((uint64_t)inlength << 32) + outlength / 2) /
 232       outlength;
 233   const int64_t offset = inlength > outlength ?
 234       (((int64_t)(inlength - outlength) << 31) + outlength / 2) / outlength :
 235       -(((int64_t)(outlength - inlength) << 31) + outlength / 2) / outlength;
 236   uint8_t *optr = output;
 237   int x, x1, x2, sum, k, int_pel, sub_pel;
 238   int64_t y;
 239
 240   const interp_kernel *interp_filters =
 241       choose_interp_filter(inlength, outlength);
 242
 243   x = 0;
 244   y = offset;
 245   while ((y >> INTERP_PRECISION_BITS) < (INTERP_TAPS / 2 - 1)) {
 246     x++;
 247     y += delta;
 248   }
 249   x1 = x;
 250   x = outlength - 1;
 251   y = delta * x + offset;
 252   while ((y >> INTERP_PRECISION_BITS) +
 253          (int64_t)(INTERP_TAPS / 2) >= inlength) {
 254     x--;
 255     y -= delta;
 256   }
 257   x2 = x;
 258   if (x1 > x2) {
 259     for (x = 0, y = offset; x < outlength; ++x, y += delta) {
 260       const int16_t *filter;
 261       int_pel = y >> INTERP_PRECISION_BITS;
 262       sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
 263       filter = interp_filters[sub_pel];
 264       sum = 0;
 265       for (k = 0; k < INTERP_TAPS; ++k) {
 266         const int pk = int_pel - INTERP_TAPS / 2 + 1 + k;
 267         sum += filter[k] * input[(pk < 0 ? 0 :
 268                                   (pk >= inlength ? inlength - 1 : pk))];
 269       }
 270       *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
 271     }
 272   } else {
 273     // Initial part.
 274     for (x = 0, y = offset; x < x1; ++x, y += delta) {
 275       const int16_t *filter;
 276       int_pel = y >> INTERP_PRECISION_BITS;
 277       sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
 278       filter = interp_filters[sub_pel];
 279       sum = 0;
 280       for (k = 0; k < INTERP_TAPS; ++k)
 281         sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k < 0 ?
 282                                   0 :
 283                                   int_pel - INTERP_TAPS / 2 + 1 + k)];
 284       *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
 285     }
 286     // Middle part.
 287     for (; x <= x2; ++x, y += delta) {
 288       const int16_t *filter;
 289       int_pel = y >> INTERP_PRECISION_BITS;
 290       sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
 291       filter = interp_filters[sub_pel];
 292       sum = 0;
 293       for (k = 0; k < INTERP_TAPS; ++k)
 294         sum += filter[k] * input[int_pel - INTERP_TAPS / 2 + 1 + k];
 295       *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
 296     }
 297     // End part.
 298     for (; x < outlength; ++x, y += delta) {
 299       const int16_t *filter;
 300       int_pel = y >> INTERP_PRECISION_BITS;
 301       sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
 302       filter = interp_filters[sub_pel];
 303       sum = 0;
 304       for (k = 0; k < INTERP_TAPS; ++k)
 305         sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k >=
 306                                   inlength ?  inlength - 1 :
 307                                   int_pel - INTERP_TAPS / 2 + 1 + k)];
 308       *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
 309     }
 310   }
 311 }
 312
 313 static void down2_symeven(const uint8_t *const input, int length,
 314                           uint8_t *output) {
 315   // Actual filter len = 2 * filter_len_half.
 316   const int16_t *filter = vp9_down2_symeven_half_filter;
 317   const int filter_len_half = sizeof(vp9_down2_symeven_half_filter) / 2;
 318   int i, j;
 319   uint8_t *optr = output;
 320   int l1 = filter_len_half;
 321   int l2 = (length - filter_len_half);
 322   l1 += (l1 & 1);
 323   l2 += (l2 & 1);
 324   if (l1 > l2) {
 325     // Short input length.
 326     for (i = 0; i < length; i += 2) {
 327       int sum = (1 << (FILTER_BITS - 1));
 328       for (j = 0; j < filter_len_half; ++j) {
 329         sum += (input[(i - j < 0 ? 0 : i - j)] +
 330                 input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
 331             filter[j];
 332       }
 333       sum >>= FILTER_BITS;
 334       *optr++ = clip_pixel(sum);
 335     }
 336   } else {
 337     // Initial part.
 338     for (i = 0; i < l1; i += 2) {
 339       int sum = (1 << (FILTER_BITS - 1));
 340       for (j = 0; j < filter_len_half; ++j) {
 341         sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + 1 + j]) * filter[j];
 342       }
 343       sum >>= FILTER_BITS;
 344       *optr++ = clip_pixel(sum);
 345     }
 346     // Middle part.
 347     for (; i < l2; i += 2) {
 348       int sum = (1 << (FILTER_BITS - 1));
 349       for (j = 0; j < filter_len_half; ++j) {
 350         sum += (input[i - j] + input[i + 1 + j]) * filter[j];
 351       }
 352       sum >>= FILTER_BITS;
 353       *optr++ = clip_pixel(sum);
 354     }
 355     // End part.
 356     for (; i < length; i += 2) {
 357       int sum = (1 << (FILTER_BITS - 1));
 358       for (j = 0; j < filter_len_half; ++j) {
 359         sum += (input[i - j] +
 360                 input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
 361             filter[j];
 362       }
 363       sum >>= FILTER_BITS;
 364       *optr++ = clip_pixel(sum);
 365     }
 366   }
 367 }
 368
 369 static void down2_symodd(const uint8_t *const input, int length,
 370                          uint8_t *output) {
 371   // Actual filter len = 2 * filter_len_half - 1.
 372   const int16_t *filter = vp9_down2_symodd_half_filter;
 373   const int filter_len_half = sizeof(vp9_down2_symodd_half_filter) / 2;
 374   int i, j;
 375   uint8_t *optr = output;
 376   int l1 = filter_len_half - 1;
 377   int l2 = (length - filter_len_half + 1);
 378   l1 += (l1 & 1);
 379   l2 += (l2 & 1);
 380   if (l1 > l2) {
 381     // Short input length.
 382     for (i = 0; i < length; i += 2) {
 383       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
 384       for (j = 1; j < filter_len_half; ++j) {
 385         sum += (input[(i - j < 0 ? 0 : i - j)] +
 386                 input[(i + j >= length ? length - 1 : i + j)]) *
 387             filter[j];
 388       }
 389       sum >>= FILTER_BITS;
 390       *optr++ = clip_pixel(sum);
 391     }
 392   } else {
 393     // Initial part.
 394     for (i = 0; i < l1; i += 2) {
 395       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
 396       for (j = 1; j < filter_len_half; ++j) {
 397         sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + j]) * filter[j];
 398       }
 399       sum >>= FILTER_BITS;
 400       *optr++ = clip_pixel(sum);
 401     }
 402     // Middle part.
 403     for (; i < l2; i += 2) {
 404       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
 405       for (j = 1; j < filter_len_half; ++j) {
 406         sum += (input[i - j] + input[i + j]) * filter[j];
 407       }
 408       sum >>= FILTER_BITS;
 409       *optr++ = clip_pixel(sum);
 410     }
 411     // End part.
 412     for (; i < length; i += 2) {
 413       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
 414       for (j = 1; j < filter_len_half; ++j) {
 415         sum += (input[i - j] + input[(i + j >= length ? length - 1 : i + j)]) *
 416             filter[j];
 417       }
 418       sum >>= FILTER_BITS;
 419       *optr++ = clip_pixel(sum);
 420     }
 421   }
 422 }
 423
 424 static int get_down2_length(int length, int steps) {
 425   int s;
 426   for (s = 0; s < steps; ++s)
 427     length = (length + 1) >> 1;
 428   return length;
 429 }
 430
 431 static int get_down2_steps(int in_length, int out_length) {
 432   int steps = 0;
 433   int proj_in_length;
 434   while ((proj_in_length = get_down2_length(in_length, 1)) >= out_length) {
 435     ++steps;
 436     in_length = proj_in_length;
 437   }
 438   return steps;
 439 }
 440
 441 static void resize_multistep(const uint8_t *const input,
 442                              int length,
 443                              uint8_t *output,
 444                              int olength,
 445                              uint8_t *buf) {
 446   int steps;
 447   if (length == olength) {
 448     memcpy(output, input, sizeof(uint8_t) * length);
 449     return;
 450   }
 451   steps = get_down2_steps(length, olength);
 452
 453   if (steps > 0) {
 454     int s;
 455     uint8_t *out = NULL;
 456     uint8_t *tmpbuf = NULL;
 457     uint8_t *otmp, *otmp2;
 458     int filteredlength = length;
 459     if (!tmpbuf) {
 460       tmpbuf = (uint8_t *)malloc(sizeof(uint8_t) * length);
 461       otmp = tmpbuf;
 462     } else {
 463       otmp = buf;
 464     }
 465     otmp2 = otmp + get_down2_length(length, 1);
 466     for (s = 0; s < steps; ++s) {
 467       const int proj_filteredlength = get_down2_length(filteredlength, 1);
 468       const uint8_t *const in = (s == 0 ? input : out);
 469       if (s == steps - 1 && proj_filteredlength == olength)
 470         out = output;
 471       else
 472         out = (s & 1 ? otmp2 : otmp);
 473       if (filteredlength & 1)
 474         down2_symodd(in, filteredlength, out);
 475       else
 476         down2_symeven(in, filteredlength, out);
 477       filteredlength = proj_filteredlength;
 478     }
 479     if (filteredlength != olength) {
 480       interpolate(out, filteredlength, output, olength);
 481     }
 482     if (tmpbuf)
 483       free(tmpbuf);
 484   } else {
 485     interpolate(input, length, output, olength);
 486   }
 487 }
 488
 489 static void fill_col_to_arr(uint8_t *img, int stride, int len, uint8_t *arr) {
 490   int i;
 491   uint8_t *iptr = img;
 492   uint8_t *aptr = arr;
 493   for (i = 0; i < len; ++i, iptr += stride) {
 494     *aptr++ = *iptr;
 495   }
 496 }
 497
 498 static void fill_arr_to_col(uint8_t *img, int stride, int len, uint8_t *arr) {
 499   int i;
 500   uint8_t *iptr = img;
 501   uint8_t *aptr = arr;
 502   for (i = 0; i < len; ++i, iptr += stride) {
 503     *iptr = *aptr++;
 504   }
 505 }
 506
 507 void vp9_resize_plane(const uint8_t *const input,
 508                       int height,
 509                       int width,
 510                       int in_stride,
 511                       uint8_t *output,
 512                       int height2,
 513                       int width2,
 514                       int out_stride) {
 515   int i;
 516   uint8_t *intbuf = (uint8_t *)malloc(sizeof(uint8_t) * width2 * height);
 517   uint8_t *tmpbuf = (uint8_t *)malloc(sizeof(uint8_t) *
 518                                       (width < height ? height : width));
 519   uint8_t *arrbuf = (uint8_t *)malloc(sizeof(uint8_t) * (height + height2));
 520   assert(width > 0);
 521   assert(height > 0);
 522   assert(width2 > 0);
 523   assert(height2 > 0);
 524   for (i = 0; i < height; ++i)
 525     resize_multistep(input + in_stride * i, width,
 526                         intbuf + width2 * i, width2, tmpbuf);
 527   for (i = 0; i < width2; ++i) {
 528     fill_col_to_arr(intbuf + i, width2, height, arrbuf);
 529     resize_multistep(arrbuf, height, arrbuf + height, height2, tmpbuf);
 530     fill_arr_to_col(output + i, out_stride, height2, arrbuf + height);
 531   }
 532   free(intbuf);
 533   free(tmpbuf);
 534   free(arrbuf);
 535 }
 536
 537 #if CONFIG_VP9_HIGHBITDEPTH
 538 static void highbd_interpolate(const uint16_t *const input, int inlength,
 539                                uint16_t *output, int outlength, int bd) {
 540   const int64_t delta =
 541       (((uint64_t)inlength << 32) + outlength / 2) / outlength;
 542   const int64_t offset = inlength > outlength ?
 543       (((int64_t)(inlength - outlength) << 31) + outlength / 2) / outlength :
 544       -(((int64_t)(outlength - inlength) << 31) + outlength / 2) / outlength;
 545   uint16_t *optr = output;
 546   int x, x1, x2, sum, k, int_pel, sub_pel;
 547   int64_t y;
 548
 549   const interp_kernel *interp_filters =
 550       choose_interp_filter(inlength, outlength);
 551
 552   x = 0;
 553   y = offset;
 554   while ((y >> INTERP_PRECISION_BITS) < (INTERP_TAPS / 2 - 1)) {
 555     x++;
 556     y += delta;
 557   }
 558   x1 = x;
 559   x = outlength - 1;
 560   y = delta * x + offset;
 561   while ((y >> INTERP_PRECISION_BITS) +
 562          (int64_t)(INTERP_TAPS / 2) >= inlength) {
 563     x--;
 564     y -= delta;
 565   }
 566   x2 = x;
 567   if (x1 > x2) {
 568     for (x = 0, y = offset; x < outlength; ++x, y += delta) {
 569       const int16_t *filter;
 570       int_pel = y >> INTERP_PRECISION_BITS;
 571       sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
 572       filter = interp_filters[sub_pel];
 573       sum = 0;
 574       for (k = 0; k < INTERP_TAPS; ++k) {
 575         const int pk = int_pel - INTERP_TAPS / 2 + 1 + k;
 576         sum += filter[k] *
 577             input[(pk < 0 ? 0 : (pk >= inlength ? inlength - 1 : pk))];
 578       }
 579       *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
 580     }
 581   } else {
 582     // Initial part.
 583     for (x = 0, y = offset; x < x1; ++x, y += delta) {
 584       const int16_t *filter;
 585       int_pel = y >> INTERP_PRECISION_BITS;
 586       sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
 587       filter = interp_filters[sub_pel];
 588       sum = 0;
 589       for (k = 0; k < INTERP_TAPS; ++k)
 590         sum += filter[k] *
 591             input[(int_pel - INTERP_TAPS / 2 + 1 + k < 0 ?
 592                    0 : int_pel - INTERP_TAPS / 2 + 1 + k)];
 593       *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
 594     }
 595     // Middle part.
 596     for (; x <= x2; ++x, y += delta) {
 597       const int16_t *filter;
 598       int_pel = y >> INTERP_PRECISION_BITS;
 599       sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
 600       filter = interp_filters[sub_pel];
 601       sum = 0;
 602       for (k = 0; k < INTERP_TAPS; ++k)
 603         sum += filter[k] * input[int_pel - INTERP_TAPS / 2 + 1 + k];
 604       *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
 605     }
 606     // End part.
 607     for (; x < outlength; ++x, y += delta) {
 608       const int16_t *filter;
 609       int_pel = y >> INTERP_PRECISION_BITS;
 610       sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
 611       filter = interp_filters[sub_pel];
 612       sum = 0;
 613       for (k = 0; k < INTERP_TAPS; ++k)
 614         sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k >=
 615                                   inlength ?  inlength - 1 :
 616                                   int_pel - INTERP_TAPS / 2 + 1 + k)];
 617       *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
 618     }
 619   }
 620 }
 621
 622 static void highbd_down2_symeven(const uint16_t *const input, int length,
 623                                  uint16_t *output, int bd) {
 624   // Actual filter len = 2 * filter_len_half.
 625   static const int16_t *filter = vp9_down2_symeven_half_filter;
 626   const int filter_len_half = sizeof(vp9_down2_symeven_half_filter) / 2;
 627   int i, j;
 628   uint16_t *optr = output;
 629   int l1 = filter_len_half;
 630   int l2 = (length - filter_len_half);
 631   l1 += (l1 & 1);
 632   l2 += (l2 & 1);
 633   if (l1 > l2) {
 634     // Short input length.
 635     for (i = 0; i < length; i += 2) {
 636       int sum = (1 << (FILTER_BITS - 1));
 637       for (j = 0; j < filter_len_half; ++j) {
 638         sum += (input[(i - j < 0 ? 0 : i - j)] +
 639                 input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
 640             filter[j];
 641       }
 642       sum >>= FILTER_BITS;
 643       *optr++ = clip_pixel_highbd(sum, bd);
 644     }
 645   } else {
 646     // Initial part.
 647     for (i = 0; i < l1; i += 2) {
 648       int sum = (1 << (FILTER_BITS - 1));
 649       for (j = 0; j < filter_len_half; ++j) {
 650         sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + 1 + j]) * filter[j];
 651       }
 652       sum >>= FILTER_BITS;
 653       *optr++ = clip_pixel_highbd(sum, bd);
 654     }
 655     // Middle part.
 656     for (; i < l2; i += 2) {
 657       int sum = (1 << (FILTER_BITS - 1));
 658       for (j = 0; j < filter_len_half; ++j) {
 659         sum += (input[i - j] + input[i + 1 + j]) * filter[j];
 660       }
 661       sum >>= FILTER_BITS;
 662       *optr++ = clip_pixel_highbd(sum, bd);
 663     }
 664     // End part.
 665     for (; i < length; i += 2) {
 666       int sum = (1 << (FILTER_BITS - 1));
 667       for (j = 0; j < filter_len_half; ++j) {
 668         sum += (input[i - j] +
 669                 input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
 670             filter[j];
 671       }
 672       sum >>= FILTER_BITS;
 673       *optr++ = clip_pixel_highbd(sum, bd);
 674     }
 675   }
 676 }
 677
 678 static void highbd_down2_symodd(const uint16_t *const input, int length,
 679                               uint16_t *output, int bd) {
 680   // Actual filter len = 2 * filter_len_half - 1.
 681   static const int16_t *filter = vp9_down2_symodd_half_filter;
 682   const int filter_len_half = sizeof(vp9_down2_symodd_half_filter) / 2;
 683   int i, j;
 684   uint16_t *optr = output;
 685   int l1 = filter_len_half - 1;
 686   int l2 = (length - filter_len_half + 1);
 687   l1 += (l1 & 1);
 688   l2 += (l2 & 1);
 689   if (l1 > l2) {
 690     // Short input length.
 691     for (i = 0; i < length; i += 2) {
 692       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
 693       for (j = 1; j < filter_len_half; ++j) {
 694         sum += (input[(i - j < 0 ? 0 : i - j)] +
 695                 input[(i + j >= length ? length - 1 : i + j)]) *
 696             filter[j];
 697       }
 698       sum >>= FILTER_BITS;
 699       *optr++ = clip_pixel_highbd(sum, bd);
 700     }
 701   } else {
 702     // Initial part.
 703     for (i = 0; i < l1; i += 2) {
 704       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
 705       for (j = 1; j < filter_len_half; ++j) {
 706         sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + j]) * filter[j];
 707       }
 708       sum >>= FILTER_BITS;
 709       *optr++ = clip_pixel_highbd(sum, bd);
 710     }
 711     // Middle part.
 712     for (; i < l2; i += 2) {
 713       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
 714       for (j = 1; j < filter_len_half; ++j) {
 715         sum += (input[i - j] + input[i + j]) * filter[j];
 716       }
 717       sum >>= FILTER_BITS;
 718       *optr++ = clip_pixel_highbd(sum, bd);
 719     }
 720     // End part.
 721     for (; i < length; i += 2) {
 722       int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
 723       for (j = 1; j < filter_len_half; ++j) {
 724         sum += (input[i - j] + input[(i + j >= length ? length - 1 : i + j)]) *
 725             filter[j];
 726       }
 727       sum >>= FILTER_BITS;
 728       *optr++ = clip_pixel_highbd(sum, bd);
 729     }
 730   }
 731 }
 732
 733 static void highbd_resize_multistep(const uint16_t *const input,
 734                                     int length,
 735                                     uint16_t *output,
 736                                     int olength,
 737                                     uint16_t *buf,
 738                                     int bd) {
 739   int steps;
 740   if (length == olength) {
 741     memcpy(output, input, sizeof(uint16_t) * length);
 742     return;
 743   }
 744   steps = get_down2_steps(length, olength);
 745
 746   if (steps > 0) {
 747     int s;
 748     uint16_t *out = NULL;
 749     uint16_t *tmpbuf = NULL;
 750     uint16_t *otmp, *otmp2;
 751     int filteredlength = length;
 752     if (!tmpbuf) {
 753       tmpbuf = (uint16_t *)malloc(sizeof(uint16_t) * length);
 754       otmp = tmpbuf;
 755     } else {
 756       otmp = buf;
 757     }
 758     otmp2 = otmp + get_down2_length(length, 1);
 759     for (s = 0; s < steps; ++s) {
 760       const int proj_filteredlength = get_down2_length(filteredlength, 1);
 761       const uint16_t *const in = (s == 0 ? input : out);
 762       if (s == steps - 1 && proj_filteredlength == olength)
 763         out = output;
 764       else
 765         out = (s & 1 ? otmp2 : otmp);
 766       if (filteredlength & 1)
 767         highbd_down2_symodd(in, filteredlength, out, bd);
 768       else
 769         highbd_down2_symeven(in, filteredlength, out, bd);
 770       filteredlength = proj_filteredlength;
 771     }
 772     if (filteredlength != olength) {
 773       highbd_interpolate(out, filteredlength, output, olength, bd);
 774     }
 775     if (tmpbuf)
 776       free(tmpbuf);
 777   } else {
 778     highbd_interpolate(input, length, output, olength, bd);
 779   }
 780 }
 781
 782 static void highbd_fill_col_to_arr(uint16_t *img, int stride, int len,
 783                                    uint16_t *arr) {
 784   int i;
 785   uint16_t *iptr = img;
 786   uint16_t *aptr = arr;
 787   for (i = 0; i < len; ++i, iptr += stride) {
 788     *aptr++ = *iptr;
 789   }
 790 }
 791
 792 static void highbd_fill_arr_to_col(uint16_t *img, int stride, int len,
 793                                    uint16_t *arr) {
 794   int i;
 795   uint16_t *iptr = img;
 796   uint16_t *aptr = arr;
 797   for (i = 0; i < len; ++i, iptr += stride) {
 798     *iptr = *aptr++;
 799   }
 800 }
 801
 802 void vp9_highbd_resize_plane(const uint8_t *const input,
 803                              int height,
 804                              int width,
 805                              int in_stride,
 806                              uint8_t *output,
 807                              int height2,
 808                              int width2,
 809                              int out_stride,
 810                              int bd) {
 811   int i;
 812   uint16_t *intbuf = (uint16_t *)malloc(sizeof(uint16_t) * width2 * height);
 813   uint16_t *tmpbuf = (uint16_t *)malloc(sizeof(uint16_t) *
 814                                         (width < height ? height : width));
 815   uint16_t *arrbuf = (uint16_t *)malloc(sizeof(uint16_t) * (height + height2));
 816   for (i = 0; i < height; ++i) {
 817     highbd_resize_multistep(CONVERT_TO_SHORTPTR(input + in_stride * i), width,
 818                             intbuf + width2 * i, width2, tmpbuf, bd);
 819   }
 820   for (i = 0; i < width2; ++i) {
 821     highbd_fill_col_to_arr(intbuf + i, width2, height, arrbuf);
 822     highbd_resize_multistep(arrbuf, height, arrbuf + height, height2, tmpbuf,
 823                             bd);
 824     highbd_fill_arr_to_col(CONVERT_TO_SHORTPTR(output + i), out_stride, height2,
 825                            arrbuf + height);
 826   }
 827   free(intbuf);
 828   free(tmpbuf);
 829   free(arrbuf);
 830 }
 831 #endif  // CONFIG_VP9_HIGHBITDEPTH
 832
 833 void vp9_resize_frame420(const uint8_t *const y,
 834                          int y_stride,
 835                          const uint8_t *const u, const uint8_t *const v,
 836                          int uv_stride,
 837                          int height, int width,
 838                          uint8_t *oy, int oy_stride,
 839                          uint8_t *ou, uint8_t *ov, int ouv_stride,
 840                          int oheight, int owidth) {
 841   vp9_resize_plane(y, height, width, y_stride,
 842                    oy, oheight, owidth, oy_stride);
 843   vp9_resize_plane(u, height / 2, width / 2, uv_stride,
 844                    ou, oheight / 2, owidth / 2, ouv_stride);
 845   vp9_resize_plane(v, height / 2, width / 2, uv_stride,
 846                    ov, oheight / 2, owidth / 2, ouv_stride);
 847 }
 848
 849 void vp9_resize_frame422(const uint8_t *const y, int y_stride,
 850                          const uint8_t *const u, const uint8_t *const v,
 851                          int uv_stride,
 852                          int height, int width,
 853                          uint8_t *oy, int oy_stride,
 854                          uint8_t *ou, uint8_t *ov, int ouv_stride,
 855                          int oheight, int owidth) {
 856   vp9_resize_plane(y, height, width, y_stride,
 857                    oy, oheight, owidth, oy_stride);
 858   vp9_resize_plane(u, height, width / 2, uv_stride,
 859                    ou, oheight, owidth / 2, ouv_stride);
 860   vp9_resize_plane(v, height, width / 2, uv_stride,
 861                    ov, oheight, owidth / 2, ouv_stride);
 862 }
 863
 864 void vp9_resize_frame444(const uint8_t *const y, int y_stride,
 865                          const uint8_t *const u, const uint8_t *const v,
 866                          int uv_stride,
 867                          int height, int width,
 868                          uint8_t *oy, int oy_stride,
 869                          uint8_t *ou, uint8_t *ov, int ouv_stride,
 870                          int oheight, int owidth) {
 871   vp9_resize_plane(y, height, width, y_stride,
 872                    oy, oheight, owidth, oy_stride);
 873   vp9_resize_plane(u, height, width, uv_stride,
 874                    ou, oheight, owidth, ouv_stride);
 875   vp9_resize_plane(v, height, width, uv_stride,
 876                    ov, oheight, owidth, ouv_stride);
 877 }
 878
 879 #if CONFIG_VP9_HIGHBITDEPTH
 880 void vp9_highbd_resize_frame420(const uint8_t *const y,
 881                                 int y_stride,
 882                                 const uint8_t *const u, const uint8_t *const v,
 883                                 int uv_stride,
 884                                 int height, int width,
 885                                 uint8_t *oy, int oy_stride,
 886                                 uint8_t *ou, uint8_t *ov, int ouv_stride,
 887                                 int oheight, int owidth, int bd) {
 888   vp9_highbd_resize_plane(y, height, width, y_stride,
 889                           oy, oheight, owidth, oy_stride, bd);
 890   vp9_highbd_resize_plane(u, height / 2, width / 2, uv_stride,
 891                           ou, oheight / 2, owidth / 2, ouv_stride, bd);
 892   vp9_highbd_resize_plane(v, height / 2, width / 2, uv_stride,
 893                           ov, oheight / 2, owidth / 2, ouv_stride, bd);
 894 }
 895
 896 void vp9_highbd_resize_frame422(const uint8_t *const y, int y_stride,
 897                                 const uint8_t *const u, const uint8_t *const v,
 898                                 int uv_stride,
 899                                 int height, int width,
 900                                 uint8_t *oy, int oy_stride,
 901                                 uint8_t *ou, uint8_t *ov, int ouv_stride,
 902                                 int oheight, int owidth, int bd) {
 903   vp9_highbd_resize_plane(y, height, width, y_stride,
 904                           oy, oheight, owidth, oy_stride, bd);
 905   vp9_highbd_resize_plane(u, height, width / 2, uv_stride,
 906                           ou, oheight, owidth / 2, ouv_stride, bd);
 907   vp9_highbd_resize_plane(v, height, width / 2, uv_stride,
 908                           ov, oheight, owidth / 2, ouv_stride, bd);
 909 }
 910
 911 void vp9_highbd_resize_frame444(const uint8_t *const y, int y_stride,
 912                                 const uint8_t *const u, const uint8_t *const v,
 913                                 int uv_stride,
 914                                 int height, int width,
 915                                 uint8_t *oy, int oy_stride,
 916                                 uint8_t *ou, uint8_t *ov, int ouv_stride,
 917                                 int oheight, int owidth, int bd) {
 918   vp9_highbd_resize_plane(y, height, width, y_stride,
 919                           oy, oheight, owidth, oy_stride, bd);
 920   vp9_highbd_resize_plane(u, height, width, uv_stride,
 921                           ou, oheight, owidth, ouv_stride, bd);
 922   vp9_highbd_resize_plane(v, height, width, uv_stride,
 923                           ov, oheight, owidth, ouv_stride, bd);
 924 }
 925 #endif  // CONFIG_VP9_HIGHBITDEPTH