test/av1_txfm_test.cc

   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 "test/av1_txfm_test.h"
  14
  15 namespace libaom_test {
  16
  17 int get_txfm1d_size(TX_SIZE tx_size) { return tx_size_wide[tx_size]; }
  18
  19 void get_txfm1d_type(TX_TYPE txfm2d_type, TYPE_TXFM *type0, TYPE_TXFM *type1) {
  20   switch (txfm2d_type) {
  21     case DCT_DCT:
  22       *type0 = TYPE_DCT;
  23       *type1 = TYPE_DCT;
  24       break;
  25     case ADST_DCT:
  26       *type0 = TYPE_ADST;
  27       *type1 = TYPE_DCT;
  28       break;
  29     case DCT_ADST:
  30       *type0 = TYPE_DCT;
  31       *type1 = TYPE_ADST;
  32       break;
  33     case ADST_ADST:
  34       *type0 = TYPE_ADST;
  35       *type1 = TYPE_ADST;
  36       break;
  37     case FLIPADST_DCT:
  38       *type0 = TYPE_ADST;
  39       *type1 = TYPE_DCT;
  40       break;
  41     case DCT_FLIPADST:
  42       *type0 = TYPE_DCT;
  43       *type1 = TYPE_ADST;
  44       break;
  45     case FLIPADST_FLIPADST:
  46       *type0 = TYPE_ADST;
  47       *type1 = TYPE_ADST;
  48       break;
  49     case ADST_FLIPADST:
  50       *type0 = TYPE_ADST;
  51       *type1 = TYPE_ADST;
  52       break;
  53     case FLIPADST_ADST:
  54       *type0 = TYPE_ADST;
  55       *type1 = TYPE_ADST;
  56       break;
  57     case IDTX:
  58       *type0 = TYPE_IDTX;
  59       *type1 = TYPE_IDTX;
  60       break;
  61     case H_DCT:
  62       *type0 = TYPE_IDTX;
  63       *type1 = TYPE_DCT;
  64       break;
  65     case V_DCT:
  66       *type0 = TYPE_DCT;
  67       *type1 = TYPE_IDTX;
  68       break;
  69     case H_ADST:
  70       *type0 = TYPE_IDTX;
  71       *type1 = TYPE_ADST;
  72       break;
  73     case V_ADST:
  74       *type0 = TYPE_ADST;
  75       *type1 = TYPE_IDTX;
  76       break;
  77     case H_FLIPADST:
  78       *type0 = TYPE_IDTX;
  79       *type1 = TYPE_ADST;
  80       break;
  81     case V_FLIPADST:
  82       *type0 = TYPE_ADST;
  83       *type1 = TYPE_IDTX;
  84       break;
  85     default:
  86       *type0 = TYPE_DCT;
  87       *type1 = TYPE_DCT;
  88       assert(0);
  89       break;
  90   }
  91 }
  92
  93 double Sqrt2 = pow(2, 0.5);
  94 double invSqrt2 = 1 / pow(2, 0.5);
  95
  96 double dct_matrix(double n, double k, int size) {
  97   return cos(M_PI * (2 * n + 1) * k / (2 * size));
  98 }
  99
 100 void reference_dct_1d(const double *in, double *out, int size) {
 101   for (int k = 0; k < size; ++k) {
 102     out[k] = 0;
 103     for (int n = 0; n < size; ++n) {
 104       out[k] += in[n] * dct_matrix(n, k, size);
 105     }
 106     if (k == 0) out[k] = out[k] * invSqrt2;
 107   }
 108 }
 109
 110 void reference_idct_1d(const double *in, double *out, int size) {
 111   for (int k = 0; k < size; ++k) {
 112     out[k] = 0;
 113     for (int n = 0; n < size; ++n) {
 114       if (n == 0)
 115         out[k] += invSqrt2 * in[n] * dct_matrix(k, n, size);
 116       else
 117         out[k] += in[n] * dct_matrix(k, n, size);
 118     }
 119   }
 120 }
 121
 122 // TODO(any): Copied from dct.c. Should be replaced by a proper reference
 123 // function that takes 'double' input & output.
 124 static void fadst4(const tran_low_t *input, tran_low_t *output) {
 125   tran_high_t x0, x1, x2, x3;
 126   tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
 127
 128   x0 = input[0];
 129   x1 = input[1];
 130   x2 = input[2];
 131   x3 = input[3];
 132
 133   if (!(x0 | x1 | x2 | x3)) {
 134     output[0] = output[1] = output[2] = output[3] = 0;
 135     return;
 136   }
 137
 138   s0 = sinpi_1_9 * x0;
 139   s1 = sinpi_4_9 * x0;
 140   s2 = sinpi_2_9 * x1;
 141   s3 = sinpi_1_9 * x1;
 142   s4 = sinpi_3_9 * x2;
 143   s5 = sinpi_4_9 * x3;
 144   s6 = sinpi_2_9 * x3;
 145   s7 = x0 + x1 - x3;
 146
 147   x0 = s0 + s2 + s5;
 148   x1 = sinpi_3_9 * s7;
 149   x2 = s1 - s3 + s6;
 150   x3 = s4;
 151
 152   s0 = x0 + x3;
 153   s1 = x1;
 154   s2 = x2 - x3;
 155   s3 = x2 - x0 + x3;
 156
 157   // 1-D transform scaling factor is sqrt(2).
 158   output[0] = (tran_low_t)fdct_round_shift(s0);
 159   output[1] = (tran_low_t)fdct_round_shift(s1);
 160   output[2] = (tran_low_t)fdct_round_shift(s2);
 161   output[3] = (tran_low_t)fdct_round_shift(s3);
 162 }
 163
 164 void reference_adst_1d(const double *in, double *out, int size) {
 165   if (size == 4) {  // Special case.
 166     tran_low_t int_input[4];
 167     for (int i = 0; i < 4; ++i) {
 168       int_input[i] = static_cast<tran_low_t>(round(in[i]));
 169     }
 170     tran_low_t int_output[4];
 171     fadst4(int_input, int_output);
 172     for (int i = 0; i < 4; ++i) {
 173       out[i] = int_output[i];
 174     }
 175     return;
 176   }
 177
 178   for (int k = 0; k < size; ++k) {
 179     out[k] = 0;
 180     for (int n = 0; n < size; ++n) {
 181       out[k] += in[n] * sin(M_PI * (2 * n + 1) * (2 * k + 1) / (4 * size));
 182     }
 183   }
 184 }
 185
 186 void reference_idtx_1d(const double *in, double *out, int size) {
 187   double scale = 0;
 188   if (size == 4)
 189     scale = Sqrt2;
 190   else if (size == 8)
 191     scale = 2;
 192   else if (size == 16)
 193     scale = 2 * Sqrt2;
 194   else if (size == 32)
 195     scale = 4;
 196   else if (size == 64)
 197     scale = 4 * Sqrt2;
 198   for (int k = 0; k < size; ++k) {
 199     out[k] = in[k] * scale;
 200   }
 201 }
 202
 203 void reference_hybrid_1d(double *in, double *out, int size, int type) {
 204   if (type == TYPE_DCT)
 205     reference_dct_1d(in, out, size);
 206   else if (type == TYPE_ADST)
 207     reference_adst_1d(in, out, size);
 208   else
 209     reference_idtx_1d(in, out, size);
 210 }
 211
 212 double get_amplification_factor(TX_TYPE tx_type, TX_SIZE tx_size) {
 213   TXFM_2D_FLIP_CFG fwd_txfm_flip_cfg;
 214   av1_get_fwd_txfm_cfg(tx_type, tx_size, &fwd_txfm_flip_cfg);
 215   const int tx_width = tx_size_wide[fwd_txfm_flip_cfg.tx_size];
 216   const int tx_height = tx_size_high[fwd_txfm_flip_cfg.tx_size];
 217   const int8_t *shift = fwd_txfm_flip_cfg.shift;
 218   const int amplify_bit = shift[0] + shift[1] + shift[2];
 219   double amplify_factor =
 220       amplify_bit >= 0 ? (1 << amplify_bit) : (1.0 / (1 << -amplify_bit));
 221
 222   // For rectangular transforms, we need to multiply by an extra factor.
 223   const int rect_type = get_rect_tx_log_ratio(tx_width, tx_height);
 224   if (abs(rect_type) == 1) {
 225     amplify_factor *= pow(2, 0.5);
 226   }
 227   return amplify_factor;
 228 }
 229
 230 void reference_hybrid_2d(double *in, double *out, TX_TYPE tx_type,
 231                          TX_SIZE tx_size) {
 232   // Get transform type and size of each dimension.
 233   TYPE_TXFM type0;
 234   TYPE_TXFM type1;
 235   get_txfm1d_type(tx_type, &type0, &type1);
 236   const int tx_width = tx_size_wide[tx_size];
 237   const int tx_height = tx_size_high[tx_size];
 238
 239   double *const temp_in = new double[AOMMAX(tx_width, tx_height)];
 240   double *const temp_out = new double[AOMMAX(tx_width, tx_height)];
 241   double *const out_interm = new double[tx_width * tx_height];
 242   const int stride = tx_width;
 243
 244   // Transform columns.
 245   for (int c = 0; c < tx_width; ++c) {
 246     for (int r = 0; r < tx_height; ++r) {
 247       temp_in[r] = in[r * stride + c];
 248     }
 249     reference_hybrid_1d(temp_in, temp_out, tx_height, type0);
 250     for (int r = 0; r < tx_height; ++r) {
 251       out_interm[r * stride + c] = temp_out[r];
 252     }
 253   }
 254
 255   // Transform rows.
 256   for (int r = 0; r < tx_height; ++r) {
 257     reference_hybrid_1d(out_interm + r * stride, out + r * stride, tx_width,
 258                         type1);
 259   }
 260
 261   delete[] temp_in;
 262   delete[] temp_out;
 263   delete[] out_interm;
 264
 265   // These transforms use an approximate 2D DCT transform, by only keeping the
 266   // top-left quarter of the coefficients, and repacking them in the first
 267   // quarter indices.
 268   // TODO(urvang): Refactor this code.
 269   if (tx_width == 64 && tx_height == 64) {  // tx_size == TX_64X64
 270     // Zero out top-right 32x32 area.
 271     for (int row = 0; row < 32; ++row) {
 272       memset(out + row * 64 + 32, 0, 32 * sizeof(*out));
 273     }
 274     // Zero out the bottom 64x32 area.
 275     memset(out + 32 * 64, 0, 32 * 64 * sizeof(*out));
 276     // Re-pack non-zero coeffs in the first 32x32 indices.
 277     for (int row = 1; row < 32; ++row) {
 278       memcpy(out + row * 32, out + row * 64, 32 * sizeof(*out));
 279     }
 280   } else if (tx_width == 32 && tx_height == 64) {  // tx_size == TX_32X64
 281     // Zero out the bottom 32x32 area.
 282     memset(out + 32 * 32, 0, 32 * 32 * sizeof(*out));
 283     // Note: no repacking needed here.
 284   } else if (tx_width == 64 && tx_height == 32) {  // tx_size == TX_64X32
 285     // Zero out right 32x32 area.
 286     for (int row = 0; row < 32; ++row) {
 287       memset(out + row * 64 + 32, 0, 32 * sizeof(*out));
 288     }
 289     // Re-pack non-zero coeffs in the first 32x32 indices.
 290     for (int row = 1; row < 32; ++row) {
 291       memcpy(out + row * 32, out + row * 64, 32 * sizeof(*out));
 292     }
 293   } else if (tx_width == 16 && tx_height == 64) {  // tx_size == TX_16X64
 294     // Zero out the bottom 16x32 area.
 295     memset(out + 16 * 32, 0, 16 * 32 * sizeof(*out));
 296     // Note: no repacking needed here.
 297   } else if (tx_width == 64 && tx_height == 16) {  // tx_size == TX_64X16
 298     // Zero out right 32x16 area.
 299     for (int row = 0; row < 16; ++row) {
 300       memset(out + row * 64 + 32, 0, 32 * sizeof(*out));
 301     }
 302     // Re-pack non-zero coeffs in the first 32x16 indices.
 303     for (int row = 1; row < 16; ++row) {
 304       memcpy(out + row * 32, out + row * 64, 32 * sizeof(*out));
 305     }
 306   }
 307
 308   // Apply appropriate scale.
 309   const double amplify_factor = get_amplification_factor(tx_type, tx_size);
 310   for (int c = 0; c < tx_width; ++c) {
 311     for (int r = 0; r < tx_height; ++r) {
 312       out[r * stride + c] *= amplify_factor;
 313     }
 314   }
 315 }
 316
 317 template <typename Type>
 318 void fliplr(Type *dest, int width, int height, int stride) {
 319   for (int r = 0; r < height; ++r) {
 320     for (int c = 0; c < width / 2; ++c) {
 321       const Type tmp = dest[r * stride + c];
 322       dest[r * stride + c] = dest[r * stride + width - 1 - c];
 323       dest[r * stride + width - 1 - c] = tmp;
 324     }
 325   }
 326 }
 327
 328 template <typename Type>
 329 void flipud(Type *dest, int width, int height, int stride) {
 330   for (int c = 0; c < width; ++c) {
 331     for (int r = 0; r < height / 2; ++r) {
 332       const Type tmp = dest[r * stride + c];
 333       dest[r * stride + c] = dest[(height - 1 - r) * stride + c];
 334       dest[(height - 1 - r) * stride + c] = tmp;
 335     }
 336   }
 337 }
 338
 339 template <typename Type>
 340 void fliplrud(Type *dest, int width, int height, int stride) {
 341   for (int r = 0; r < height / 2; ++r) {
 342     for (int c = 0; c < width; ++c) {
 343       const Type tmp = dest[r * stride + c];
 344       dest[r * stride + c] = dest[(height - 1 - r) * stride + width - 1 - c];
 345       dest[(height - 1 - r) * stride + width - 1 - c] = tmp;
 346     }
 347   }
 348 }
 349
 350 template void fliplr<double>(double *dest, int width, int height, int stride);
 351 template void flipud<double>(double *dest, int width, int height, int stride);
 352 template void fliplrud<double>(double *dest, int width, int height, int stride);
 353
 354 int bd_arr[BD_NUM] = { 8, 10, 12 };
 355
 356 int8_t low_range_arr[BD_NUM] = { 18, 32, 32 };
 357 int8_t high_range_arr[BD_NUM] = { 32, 32, 32 };
 358
 359 void txfm_stage_range_check(const int8_t *stage_range, int stage_num,
 360                             int8_t cos_bit, int low_range, int high_range) {
 361   for (int i = 0; i < stage_num; ++i) {
 362     EXPECT_LE(stage_range[i], low_range);
 363     ASSERT_LE(stage_range[i] + cos_bit, high_range) << "stage = " << i;
 364   }
 365   for (int i = 0; i < stage_num - 1; ++i) {
 366     // make sure there is no overflow while doing half_btf()
 367     ASSERT_LE(stage_range[i + 1] + cos_bit, high_range) << "stage = " << i;
 368   }
 369 }
 370 }  // namespace libaom_test