av1_inv_txfm1d.c 78.6 KB
Newer Older
Angie Chiang's avatar
Angie Chiang committed
1
/*
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
Angie Chiang's avatar
Angie Chiang committed
3
 *
4 5 6 7 8 9
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
Angie Chiang's avatar
Angie Chiang committed
10 11
 */

12
#include <stdlib.h>
13
#include "aom_dsp/inv_txfm.h"
Yaowu Xu's avatar
Yaowu Xu committed
14
#include "av1/common/av1_inv_txfm1d.h"
Angie Chiang's avatar
Angie Chiang committed
15 16

int32_t range_check_value(int32_t value, int8_t bit) {
Angie Chiang's avatar
Angie Chiang committed
17
#if CONFIG_COEFFICIENT_RANGE_CHECKING
Angie Chiang's avatar
Angie Chiang committed
18 19 20 21 22 23 24 25 26 27 28
  const int64_t maxValue = (1LL << (bit - 1)) - 1;
  const int64_t minValue = -(1LL << (bit - 1));
  if (value < minValue || value > maxValue) {
    fprintf(stderr, "coeff out of bit range, value: %d bit %d\n", value, bit);
    assert(0);
  }
#else
  (void)bit;
#endif
  return value;
}
29

Angie Chiang's avatar
Angie Chiang committed
30
#if CONFIG_COEFFICIENT_RANGE_CHECKING
31 32
void range_check_func(int32_t stage, const int32_t *input, const int32_t *buf,
                      int32_t size, int8_t bit) {
33 34
  const int64_t maxValue = (1LL << (bit - 1)) - 1;
  const int64_t minValue = -(1LL << (bit - 1));
35

36 37
  int in_range = 1;

38 39
  for (int i = 0; i < size; ++i) {
    if (buf[i] < minValue || buf[i] > maxValue) {
40 41 42 43 44 45
      in_range = 0;
    }
  }

  if (!in_range) {
    fprintf(stderr, "Error: coeffs contain out-of-range values\n");
46
    fprintf(stderr, "size: %d\n", size);
47 48 49 50 51
    fprintf(stderr, "stage: %d\n", stage);
    fprintf(stderr, "allowed range: [%" PRId64 ";%" PRId64 "]\n", minValue,
            maxValue);

    fprintf(stderr, "coeffs: ");
52

53 54 55 56
    fprintf(stderr, "[");
    for (int j = 0; j < size; j++) {
      if (j > 0) fprintf(stderr, ", ");
      fprintf(stderr, "%d", input[j]);
57
    }
58 59 60 61 62 63 64 65 66 67
    fprintf(stderr, "]\n");

    fprintf(stderr, "   buf: ");

    fprintf(stderr, "[");
    for (int j = 0; j < size; j++) {
      if (j > 0) fprintf(stderr, ", ");
      fprintf(stderr, "%d", buf[j]);
    }
    fprintf(stderr, "]\n\n");
Angie Chiang's avatar
Angie Chiang committed
68
  }
69 70

  assert(in_range);
71 72
}

73
#define apply_range(stage, input, buf, size, bit) \
74
  range_check_func(stage, input, buf, size, bit)
Angie Chiang's avatar
Angie Chiang committed
75
#else
76 77 78 79 80 81 82 83 84
#define apply_range(stage, input, buf, size, bit) \
  clamp_buf((int32_t *)buf, size, bit)

void clamp_buf(int32_t *buf, int32_t size, int8_t bit) {
  if (bit <= 16) {
    const int32_t maxValue = (1 << 15) - 1;
    const int32_t minValue = -(1 << 15);

    for (int i = 0; i < size; ++i) buf[i] = clamp(buf[i], minValue, maxValue);
85
  }
86
}
Angie Chiang's avatar
Angie Chiang committed
87 88
#endif

Angie Chiang's avatar
Angie Chiang committed
89 90 91 92 93 94 95 96 97
int32_t clamp_value(int32_t value, int8_t bit) {
  if (bit <= 16) {
    const int32_t maxValue = (1 << 15) - 1;
    const int32_t minValue = -(1 << 15);
    return clamp(value, minValue, maxValue);
  }
  return value;
}

98
// TODO(angiebird): Make 1-d txfm functions static
99
void av1_idct4_new(const int32_t *input, int32_t *output, int8_t cos_bit,
Yaowu Xu's avatar
Yaowu Xu committed
100
                   const int8_t *stage_range) {
Angie Chiang's avatar
Angie Chiang committed
101 102 103 104 105 106 107 108
  const int32_t size = 4;
  const int32_t *cospi;

  int32_t stage = 0;
  int32_t *bf0, *bf1;
  int32_t step[4];

  // stage 0;
109
  apply_range(stage, input, input, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
110 111 112

  // stage 1;
  stage++;
113
  assert(output != input);
Angie Chiang's avatar
Angie Chiang committed
114 115 116 117 118
  bf1 = output;
  bf1[0] = input[0];
  bf1[1] = input[2];
  bf1[2] = input[1];
  bf1[3] = input[3];
119
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
120 121 122

  // stage 2
  stage++;
123
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
124 125
  bf0 = output;
  bf1 = step;
126 127 128 129
  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit);
  bf1[1] = half_btf(cospi[32], bf0[0], -cospi[32], bf0[1], cos_bit);
  bf1[2] = half_btf(cospi[48], bf0[2], -cospi[16], bf0[3], cos_bit);
  bf1[3] = half_btf(cospi[16], bf0[2], cospi[48], bf0[3], cos_bit);
130
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
131 132 133 134 135 136 137 138 139

  // stage 3
  stage++;
  bf0 = step;
  bf1 = output;
  bf1[0] = bf0[0] + bf0[3];
  bf1[1] = bf0[1] + bf0[2];
  bf1[2] = bf0[1] - bf0[2];
  bf1[3] = bf0[0] - bf0[3];
140
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
141
}
Angie Chiang's avatar
Angie Chiang committed
142

143
void av1_idct8_new(const int32_t *input, int32_t *output, int8_t cos_bit,
Yaowu Xu's avatar
Yaowu Xu committed
144
                   const int8_t *stage_range) {
Angie Chiang's avatar
Angie Chiang committed
145 146 147 148 149 150 151 152
  const int32_t size = 8;
  const int32_t *cospi;

  int32_t stage = 0;
  int32_t *bf0, *bf1;
  int32_t step[8];

  // stage 0;
153
  apply_range(stage, input, input, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
154 155 156

  // stage 1;
  stage++;
157
  assert(output != input);
Angie Chiang's avatar
Angie Chiang committed
158 159 160 161 162 163 164 165 166
  bf1 = output;
  bf1[0] = input[0];
  bf1[1] = input[4];
  bf1[2] = input[2];
  bf1[3] = input[6];
  bf1[4] = input[1];
  bf1[5] = input[5];
  bf1[6] = input[3];
  bf1[7] = input[7];
167
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
168 169 170

  // stage 2
  stage++;
171
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
172 173 174 175 176 177
  bf0 = output;
  bf1 = step;
  bf1[0] = bf0[0];
  bf1[1] = bf0[1];
  bf1[2] = bf0[2];
  bf1[3] = bf0[3];
178 179 180 181
  bf1[4] = half_btf(cospi[56], bf0[4], -cospi[8], bf0[7], cos_bit);
  bf1[5] = half_btf(cospi[24], bf0[5], -cospi[40], bf0[6], cos_bit);
  bf1[6] = half_btf(cospi[40], bf0[5], cospi[24], bf0[6], cos_bit);
  bf1[7] = half_btf(cospi[8], bf0[4], cospi[56], bf0[7], cos_bit);
182
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
183 184 185

  // stage 3
  stage++;
186
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
187 188
  bf0 = step;
  bf1 = output;
189 190 191 192
  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit);
  bf1[1] = half_btf(cospi[32], bf0[0], -cospi[32], bf0[1], cos_bit);
  bf1[2] = half_btf(cospi[48], bf0[2], -cospi[16], bf0[3], cos_bit);
  bf1[3] = half_btf(cospi[16], bf0[2], cospi[48], bf0[3], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
193 194 195 196
  bf1[4] = bf0[4] + bf0[5];
  bf1[5] = bf0[4] - bf0[5];
  bf1[6] = -bf0[6] + bf0[7];
  bf1[7] = bf0[6] + bf0[7];
197
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
198 199 200

  // stage 4
  stage++;
201
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
202 203 204 205 206 207 208
  bf0 = output;
  bf1 = step;
  bf1[0] = bf0[0] + bf0[3];
  bf1[1] = bf0[1] + bf0[2];
  bf1[2] = bf0[1] - bf0[2];
  bf1[3] = bf0[0] - bf0[3];
  bf1[4] = bf0[4];
209 210
  bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit);
  bf1[6] = half_btf(cospi[32], bf0[5], cospi[32], bf0[6], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
211
  bf1[7] = bf0[7];
212
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
213 214 215 216 217 218 219 220 221 222 223 224 225

  // stage 5
  stage++;
  bf0 = step;
  bf1 = output;
  bf1[0] = bf0[0] + bf0[7];
  bf1[1] = bf0[1] + bf0[6];
  bf1[2] = bf0[2] + bf0[5];
  bf1[3] = bf0[3] + bf0[4];
  bf1[4] = bf0[3] - bf0[4];
  bf1[5] = bf0[2] - bf0[5];
  bf1[6] = bf0[1] - bf0[6];
  bf1[7] = bf0[0] - bf0[7];
226
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
227
}
Angie Chiang's avatar
Angie Chiang committed
228

229 230
void av1_idct16_new(const int32_t *input, int32_t *output, int8_t cos_bit,
                    const int8_t *stage_range) {
Angie Chiang's avatar
Angie Chiang committed
231 232 233 234 235 236 237 238
  const int32_t size = 16;
  const int32_t *cospi;

  int32_t stage = 0;
  int32_t *bf0, *bf1;
  int32_t step[16];

  // stage 0;
239
  apply_range(stage, input, input, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
240 241 242

  // stage 1;
  stage++;
243
  assert(output != input);
Angie Chiang's avatar
Angie Chiang committed
244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260
  bf1 = output;
  bf1[0] = input[0];
  bf1[1] = input[8];
  bf1[2] = input[4];
  bf1[3] = input[12];
  bf1[4] = input[2];
  bf1[5] = input[10];
  bf1[6] = input[6];
  bf1[7] = input[14];
  bf1[8] = input[1];
  bf1[9] = input[9];
  bf1[10] = input[5];
  bf1[11] = input[13];
  bf1[12] = input[3];
  bf1[13] = input[11];
  bf1[14] = input[7];
  bf1[15] = input[15];
261
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
262 263 264

  // stage 2
  stage++;
265
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
266 267 268 269 270 271 272 273 274 275
  bf0 = output;
  bf1 = step;
  bf1[0] = bf0[0];
  bf1[1] = bf0[1];
  bf1[2] = bf0[2];
  bf1[3] = bf0[3];
  bf1[4] = bf0[4];
  bf1[5] = bf0[5];
  bf1[6] = bf0[6];
  bf1[7] = bf0[7];
276 277 278 279 280 281 282 283
  bf1[8] = half_btf(cospi[60], bf0[8], -cospi[4], bf0[15], cos_bit);
  bf1[9] = half_btf(cospi[28], bf0[9], -cospi[36], bf0[14], cos_bit);
  bf1[10] = half_btf(cospi[44], bf0[10], -cospi[20], bf0[13], cos_bit);
  bf1[11] = half_btf(cospi[12], bf0[11], -cospi[52], bf0[12], cos_bit);
  bf1[12] = half_btf(cospi[52], bf0[11], cospi[12], bf0[12], cos_bit);
  bf1[13] = half_btf(cospi[20], bf0[10], cospi[44], bf0[13], cos_bit);
  bf1[14] = half_btf(cospi[36], bf0[9], cospi[28], bf0[14], cos_bit);
  bf1[15] = half_btf(cospi[4], bf0[8], cospi[60], bf0[15], cos_bit);
284
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
285 286 287

  // stage 3
  stage++;
288
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
289 290 291 292 293 294
  bf0 = step;
  bf1 = output;
  bf1[0] = bf0[0];
  bf1[1] = bf0[1];
  bf1[2] = bf0[2];
  bf1[3] = bf0[3];
295 296 297 298
  bf1[4] = half_btf(cospi[56], bf0[4], -cospi[8], bf0[7], cos_bit);
  bf1[5] = half_btf(cospi[24], bf0[5], -cospi[40], bf0[6], cos_bit);
  bf1[6] = half_btf(cospi[40], bf0[5], cospi[24], bf0[6], cos_bit);
  bf1[7] = half_btf(cospi[8], bf0[4], cospi[56], bf0[7], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
299 300 301 302 303 304 305 306
  bf1[8] = bf0[8] + bf0[9];
  bf1[9] = bf0[8] - bf0[9];
  bf1[10] = -bf0[10] + bf0[11];
  bf1[11] = bf0[10] + bf0[11];
  bf1[12] = bf0[12] + bf0[13];
  bf1[13] = bf0[12] - bf0[13];
  bf1[14] = -bf0[14] + bf0[15];
  bf1[15] = bf0[14] + bf0[15];
307
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
308 309 310

  // stage 4
  stage++;
311
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
312 313
  bf0 = output;
  bf1 = step;
314 315 316 317
  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit);
  bf1[1] = half_btf(cospi[32], bf0[0], -cospi[32], bf0[1], cos_bit);
  bf1[2] = half_btf(cospi[48], bf0[2], -cospi[16], bf0[3], cos_bit);
  bf1[3] = half_btf(cospi[16], bf0[2], cospi[48], bf0[3], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
318 319 320 321 322
  bf1[4] = bf0[4] + bf0[5];
  bf1[5] = bf0[4] - bf0[5];
  bf1[6] = -bf0[6] + bf0[7];
  bf1[7] = bf0[6] + bf0[7];
  bf1[8] = bf0[8];
323 324
  bf1[9] = half_btf(-cospi[16], bf0[9], cospi[48], bf0[14], cos_bit);
  bf1[10] = half_btf(-cospi[48], bf0[10], -cospi[16], bf0[13], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
325 326
  bf1[11] = bf0[11];
  bf1[12] = bf0[12];
327 328
  bf1[13] = half_btf(-cospi[16], bf0[10], cospi[48], bf0[13], cos_bit);
  bf1[14] = half_btf(cospi[48], bf0[9], cospi[16], bf0[14], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
329
  bf1[15] = bf0[15];
330
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
331 332 333

  // stage 5
  stage++;
334
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
335 336 337 338 339 340 341
  bf0 = step;
  bf1 = output;
  bf1[0] = bf0[0] + bf0[3];
  bf1[1] = bf0[1] + bf0[2];
  bf1[2] = bf0[1] - bf0[2];
  bf1[3] = bf0[0] - bf0[3];
  bf1[4] = bf0[4];
342 343
  bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit);
  bf1[6] = half_btf(cospi[32], bf0[5], cospi[32], bf0[6], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
344 345 346 347 348 349 350 351 352
  bf1[7] = bf0[7];
  bf1[8] = bf0[8] + bf0[11];
  bf1[9] = bf0[9] + bf0[10];
  bf1[10] = bf0[9] - bf0[10];
  bf1[11] = bf0[8] - bf0[11];
  bf1[12] = -bf0[12] + bf0[15];
  bf1[13] = -bf0[13] + bf0[14];
  bf1[14] = bf0[13] + bf0[14];
  bf1[15] = bf0[12] + bf0[15];
353
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
354 355 356

  // stage 6
  stage++;
357
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
358 359 360 361 362 363 364 365 366 367 368 369
  bf0 = output;
  bf1 = step;
  bf1[0] = bf0[0] + bf0[7];
  bf1[1] = bf0[1] + bf0[6];
  bf1[2] = bf0[2] + bf0[5];
  bf1[3] = bf0[3] + bf0[4];
  bf1[4] = bf0[3] - bf0[4];
  bf1[5] = bf0[2] - bf0[5];
  bf1[6] = bf0[1] - bf0[6];
  bf1[7] = bf0[0] - bf0[7];
  bf1[8] = bf0[8];
  bf1[9] = bf0[9];
370 371 372 373
  bf1[10] = half_btf(-cospi[32], bf0[10], cospi[32], bf0[13], cos_bit);
  bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[12], cos_bit);
  bf1[12] = half_btf(cospi[32], bf0[11], cospi[32], bf0[12], cos_bit);
  bf1[13] = half_btf(cospi[32], bf0[10], cospi[32], bf0[13], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
374 375
  bf1[14] = bf0[14];
  bf1[15] = bf0[15];
376
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397

  // stage 7
  stage++;
  bf0 = step;
  bf1 = output;
  bf1[0] = bf0[0] + bf0[15];
  bf1[1] = bf0[1] + bf0[14];
  bf1[2] = bf0[2] + bf0[13];
  bf1[3] = bf0[3] + bf0[12];
  bf1[4] = bf0[4] + bf0[11];
  bf1[5] = bf0[5] + bf0[10];
  bf1[6] = bf0[6] + bf0[9];
  bf1[7] = bf0[7] + bf0[8];
  bf1[8] = bf0[7] - bf0[8];
  bf1[9] = bf0[6] - bf0[9];
  bf1[10] = bf0[5] - bf0[10];
  bf1[11] = bf0[4] - bf0[11];
  bf1[12] = bf0[3] - bf0[12];
  bf1[13] = bf0[2] - bf0[13];
  bf1[14] = bf0[1] - bf0[14];
  bf1[15] = bf0[0] - bf0[15];
398
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
399
}
Angie Chiang's avatar
Angie Chiang committed
400

401 402
void av1_idct32_new(const int32_t *input, int32_t *output, int8_t cos_bit,
                    const int8_t *stage_range) {
Angie Chiang's avatar
Angie Chiang committed
403 404 405 406 407 408 409 410
  const int32_t size = 32;
  const int32_t *cospi;

  int32_t stage = 0;
  int32_t *bf0, *bf1;
  int32_t step[32];

  // stage 0;
411
  apply_range(stage, input, input, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
412 413 414

  // stage 1;
  stage++;
415
  assert(output != input);
Angie Chiang's avatar
Angie Chiang committed
416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448
  bf1 = output;
  bf1[0] = input[0];
  bf1[1] = input[16];
  bf1[2] = input[8];
  bf1[3] = input[24];
  bf1[4] = input[4];
  bf1[5] = input[20];
  bf1[6] = input[12];
  bf1[7] = input[28];
  bf1[8] = input[2];
  bf1[9] = input[18];
  bf1[10] = input[10];
  bf1[11] = input[26];
  bf1[12] = input[6];
  bf1[13] = input[22];
  bf1[14] = input[14];
  bf1[15] = input[30];
  bf1[16] = input[1];
  bf1[17] = input[17];
  bf1[18] = input[9];
  bf1[19] = input[25];
  bf1[20] = input[5];
  bf1[21] = input[21];
  bf1[22] = input[13];
  bf1[23] = input[29];
  bf1[24] = input[3];
  bf1[25] = input[19];
  bf1[26] = input[11];
  bf1[27] = input[27];
  bf1[28] = input[7];
  bf1[29] = input[23];
  bf1[30] = input[15];
  bf1[31] = input[31];
449
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
450 451 452

  // stage 2
  stage++;
453
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471
  bf0 = output;
  bf1 = step;
  bf1[0] = bf0[0];
  bf1[1] = bf0[1];
  bf1[2] = bf0[2];
  bf1[3] = bf0[3];
  bf1[4] = bf0[4];
  bf1[5] = bf0[5];
  bf1[6] = bf0[6];
  bf1[7] = bf0[7];
  bf1[8] = bf0[8];
  bf1[9] = bf0[9];
  bf1[10] = bf0[10];
  bf1[11] = bf0[11];
  bf1[12] = bf0[12];
  bf1[13] = bf0[13];
  bf1[14] = bf0[14];
  bf1[15] = bf0[15];
472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487
  bf1[16] = half_btf(cospi[62], bf0[16], -cospi[2], bf0[31], cos_bit);
  bf1[17] = half_btf(cospi[30], bf0[17], -cospi[34], bf0[30], cos_bit);
  bf1[18] = half_btf(cospi[46], bf0[18], -cospi[18], bf0[29], cos_bit);
  bf1[19] = half_btf(cospi[14], bf0[19], -cospi[50], bf0[28], cos_bit);
  bf1[20] = half_btf(cospi[54], bf0[20], -cospi[10], bf0[27], cos_bit);
  bf1[21] = half_btf(cospi[22], bf0[21], -cospi[42], bf0[26], cos_bit);
  bf1[22] = half_btf(cospi[38], bf0[22], -cospi[26], bf0[25], cos_bit);
  bf1[23] = half_btf(cospi[6], bf0[23], -cospi[58], bf0[24], cos_bit);
  bf1[24] = half_btf(cospi[58], bf0[23], cospi[6], bf0[24], cos_bit);
  bf1[25] = half_btf(cospi[26], bf0[22], cospi[38], bf0[25], cos_bit);
  bf1[26] = half_btf(cospi[42], bf0[21], cospi[22], bf0[26], cos_bit);
  bf1[27] = half_btf(cospi[10], bf0[20], cospi[54], bf0[27], cos_bit);
  bf1[28] = half_btf(cospi[50], bf0[19], cospi[14], bf0[28], cos_bit);
  bf1[29] = half_btf(cospi[18], bf0[18], cospi[46], bf0[29], cos_bit);
  bf1[30] = half_btf(cospi[34], bf0[17], cospi[30], bf0[30], cos_bit);
  bf1[31] = half_btf(cospi[2], bf0[16], cospi[62], bf0[31], cos_bit);
488
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
489 490 491

  // stage 3
  stage++;
492
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
493 494 495 496 497 498 499 500 501 502
  bf0 = step;
  bf1 = output;
  bf1[0] = bf0[0];
  bf1[1] = bf0[1];
  bf1[2] = bf0[2];
  bf1[3] = bf0[3];
  bf1[4] = bf0[4];
  bf1[5] = bf0[5];
  bf1[6] = bf0[6];
  bf1[7] = bf0[7];
503 504 505 506 507 508 509 510
  bf1[8] = half_btf(cospi[60], bf0[8], -cospi[4], bf0[15], cos_bit);
  bf1[9] = half_btf(cospi[28], bf0[9], -cospi[36], bf0[14], cos_bit);
  bf1[10] = half_btf(cospi[44], bf0[10], -cospi[20], bf0[13], cos_bit);
  bf1[11] = half_btf(cospi[12], bf0[11], -cospi[52], bf0[12], cos_bit);
  bf1[12] = half_btf(cospi[52], bf0[11], cospi[12], bf0[12], cos_bit);
  bf1[13] = half_btf(cospi[20], bf0[10], cospi[44], bf0[13], cos_bit);
  bf1[14] = half_btf(cospi[36], bf0[9], cospi[28], bf0[14], cos_bit);
  bf1[15] = half_btf(cospi[4], bf0[8], cospi[60], bf0[15], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526
  bf1[16] = bf0[16] + bf0[17];
  bf1[17] = bf0[16] - bf0[17];
  bf1[18] = -bf0[18] + bf0[19];
  bf1[19] = bf0[18] + bf0[19];
  bf1[20] = bf0[20] + bf0[21];
  bf1[21] = bf0[20] - bf0[21];
  bf1[22] = -bf0[22] + bf0[23];
  bf1[23] = bf0[22] + bf0[23];
  bf1[24] = bf0[24] + bf0[25];
  bf1[25] = bf0[24] - bf0[25];
  bf1[26] = -bf0[26] + bf0[27];
  bf1[27] = bf0[26] + bf0[27];
  bf1[28] = bf0[28] + bf0[29];
  bf1[29] = bf0[28] - bf0[29];
  bf1[30] = -bf0[30] + bf0[31];
  bf1[31] = bf0[30] + bf0[31];
527
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
528 529 530

  // stage 4
  stage++;
531
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
532 533 534 535 536 537
  bf0 = output;
  bf1 = step;
  bf1[0] = bf0[0];
  bf1[1] = bf0[1];
  bf1[2] = bf0[2];
  bf1[3] = bf0[3];
538 539 540 541
  bf1[4] = half_btf(cospi[56], bf0[4], -cospi[8], bf0[7], cos_bit);
  bf1[5] = half_btf(cospi[24], bf0[5], -cospi[40], bf0[6], cos_bit);
  bf1[6] = half_btf(cospi[40], bf0[5], cospi[24], bf0[6], cos_bit);
  bf1[7] = half_btf(cospi[8], bf0[4], cospi[56], bf0[7], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
542 543 544 545 546 547 548 549 550
  bf1[8] = bf0[8] + bf0[9];
  bf1[9] = bf0[8] - bf0[9];
  bf1[10] = -bf0[10] + bf0[11];
  bf1[11] = bf0[10] + bf0[11];
  bf1[12] = bf0[12] + bf0[13];
  bf1[13] = bf0[12] - bf0[13];
  bf1[14] = -bf0[14] + bf0[15];
  bf1[15] = bf0[14] + bf0[15];
  bf1[16] = bf0[16];
551 552
  bf1[17] = half_btf(-cospi[8], bf0[17], cospi[56], bf0[30], cos_bit);
  bf1[18] = half_btf(-cospi[56], bf0[18], -cospi[8], bf0[29], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
553 554
  bf1[19] = bf0[19];
  bf1[20] = bf0[20];
555 556
  bf1[21] = half_btf(-cospi[40], bf0[21], cospi[24], bf0[26], cos_bit);
  bf1[22] = half_btf(-cospi[24], bf0[22], -cospi[40], bf0[25], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
557 558
  bf1[23] = bf0[23];
  bf1[24] = bf0[24];
559 560
  bf1[25] = half_btf(-cospi[40], bf0[22], cospi[24], bf0[25], cos_bit);
  bf1[26] = half_btf(cospi[24], bf0[21], cospi[40], bf0[26], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
561 562
  bf1[27] = bf0[27];
  bf1[28] = bf0[28];
563 564
  bf1[29] = half_btf(-cospi[8], bf0[18], cospi[56], bf0[29], cos_bit);
  bf1[30] = half_btf(cospi[56], bf0[17], cospi[8], bf0[30], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
565
  bf1[31] = bf0[31];
566
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
567 568 569

  // stage 5
  stage++;
570
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
571 572
  bf0 = step;
  bf1 = output;
573 574 575 576
  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit);
  bf1[1] = half_btf(cospi[32], bf0[0], -cospi[32], bf0[1], cos_bit);
  bf1[2] = half_btf(cospi[48], bf0[2], -cospi[16], bf0[3], cos_bit);
  bf1[3] = half_btf(cospi[16], bf0[2], cospi[48], bf0[3], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
577 578 579 580 581
  bf1[4] = bf0[4] + bf0[5];
  bf1[5] = bf0[4] - bf0[5];
  bf1[6] = -bf0[6] + bf0[7];
  bf1[7] = bf0[6] + bf0[7];
  bf1[8] = bf0[8];
582 583
  bf1[9] = half_btf(-cospi[16], bf0[9], cospi[48], bf0[14], cos_bit);
  bf1[10] = half_btf(-cospi[48], bf0[10], -cospi[16], bf0[13], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
584 585
  bf1[11] = bf0[11];
  bf1[12] = bf0[12];
586 587
  bf1[13] = half_btf(-cospi[16], bf0[10], cospi[48], bf0[13], cos_bit);
  bf1[14] = half_btf(cospi[48], bf0[9], cospi[16], bf0[14], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604
  bf1[15] = bf0[15];
  bf1[16] = bf0[16] + bf0[19];
  bf1[17] = bf0[17] + bf0[18];
  bf1[18] = bf0[17] - bf0[18];
  bf1[19] = bf0[16] - bf0[19];
  bf1[20] = -bf0[20] + bf0[23];
  bf1[21] = -bf0[21] + bf0[22];
  bf1[22] = bf0[21] + bf0[22];
  bf1[23] = bf0[20] + bf0[23];
  bf1[24] = bf0[24] + bf0[27];
  bf1[25] = bf0[25] + bf0[26];
  bf1[26] = bf0[25] - bf0[26];
  bf1[27] = bf0[24] - bf0[27];
  bf1[28] = -bf0[28] + bf0[31];
  bf1[29] = -bf0[29] + bf0[30];
  bf1[30] = bf0[29] + bf0[30];
  bf1[31] = bf0[28] + bf0[31];
605
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
606 607 608

  // stage 6
  stage++;
609
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
610 611 612 613 614 615 616
  bf0 = output;
  bf1 = step;
  bf1[0] = bf0[0] + bf0[3];
  bf1[1] = bf0[1] + bf0[2];
  bf1[2] = bf0[1] - bf0[2];
  bf1[3] = bf0[0] - bf0[3];
  bf1[4] = bf0[4];
617 618
  bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit);
  bf1[6] = half_btf(cospi[32], bf0[5], cospi[32], bf0[6], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
619 620 621 622 623 624 625 626 627 628 629
  bf1[7] = bf0[7];
  bf1[8] = bf0[8] + bf0[11];
  bf1[9] = bf0[9] + bf0[10];
  bf1[10] = bf0[9] - bf0[10];
  bf1[11] = bf0[8] - bf0[11];
  bf1[12] = -bf0[12] + bf0[15];
  bf1[13] = -bf0[13] + bf0[14];
  bf1[14] = bf0[13] + bf0[14];
  bf1[15] = bf0[12] + bf0[15];
  bf1[16] = bf0[16];
  bf1[17] = bf0[17];
630 631 632 633
  bf1[18] = half_btf(-cospi[16], bf0[18], cospi[48], bf0[29], cos_bit);
  bf1[19] = half_btf(-cospi[16], bf0[19], cospi[48], bf0[28], cos_bit);
  bf1[20] = half_btf(-cospi[48], bf0[20], -cospi[16], bf0[27], cos_bit);
  bf1[21] = half_btf(-cospi[48], bf0[21], -cospi[16], bf0[26], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
634 635 636 637
  bf1[22] = bf0[22];
  bf1[23] = bf0[23];
  bf1[24] = bf0[24];
  bf1[25] = bf0[25];
638 639 640 641
  bf1[26] = half_btf(-cospi[16], bf0[21], cospi[48], bf0[26], cos_bit);
  bf1[27] = half_btf(-cospi[16], bf0[20], cospi[48], bf0[27], cos_bit);
  bf1[28] = half_btf(cospi[48], bf0[19], cospi[16], bf0[28], cos_bit);
  bf1[29] = half_btf(cospi[48], bf0[18], cospi[16], bf0[29], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
642 643
  bf1[30] = bf0[30];
  bf1[31] = bf0[31];
644
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
645 646 647

  // stage 7
  stage++;
648
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
649 650 651 652 653 654 655 656 657 658 659 660
  bf0 = step;
  bf1 = output;
  bf1[0] = bf0[0] + bf0[7];
  bf1[1] = bf0[1] + bf0[6];
  bf1[2] = bf0[2] + bf0[5];
  bf1[3] = bf0[3] + bf0[4];
  bf1[4] = bf0[3] - bf0[4];
  bf1[5] = bf0[2] - bf0[5];
  bf1[6] = bf0[1] - bf0[6];
  bf1[7] = bf0[0] - bf0[7];
  bf1[8] = bf0[8];
  bf1[9] = bf0[9];
661 662 663 664
  bf1[10] = half_btf(-cospi[32], bf0[10], cospi[32], bf0[13], cos_bit);
  bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[12], cos_bit);
  bf1[12] = half_btf(cospi[32], bf0[11], cospi[32], bf0[12], cos_bit);
  bf1[13] = half_btf(cospi[32], bf0[10], cospi[32], bf0[13], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682
  bf1[14] = bf0[14];
  bf1[15] = bf0[15];
  bf1[16] = bf0[16] + bf0[23];
  bf1[17] = bf0[17] + bf0[22];
  bf1[18] = bf0[18] + bf0[21];
  bf1[19] = bf0[19] + bf0[20];
  bf1[20] = bf0[19] - bf0[20];
  bf1[21] = bf0[18] - bf0[21];
  bf1[22] = bf0[17] - bf0[22];
  bf1[23] = bf0[16] - bf0[23];
  bf1[24] = -bf0[24] + bf0[31];
  bf1[25] = -bf0[25] + bf0[30];
  bf1[26] = -bf0[26] + bf0[29];
  bf1[27] = -bf0[27] + bf0[28];
  bf1[28] = bf0[27] + bf0[28];
  bf1[29] = bf0[26] + bf0[29];
  bf1[30] = bf0[25] + bf0[30];
  bf1[31] = bf0[24] + bf0[31];
683
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
684 685 686

  // stage 8
  stage++;
687
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709
  bf0 = output;
  bf1 = step;
  bf1[0] = bf0[0] + bf0[15];
  bf1[1] = bf0[1] + bf0[14];
  bf1[2] = bf0[2] + bf0[13];
  bf1[3] = bf0[3] + bf0[12];
  bf1[4] = bf0[4] + bf0[11];
  bf1[5] = bf0[5] + bf0[10];
  bf1[6] = bf0[6] + bf0[9];
  bf1[7] = bf0[7] + bf0[8];
  bf1[8] = bf0[7] - bf0[8];
  bf1[9] = bf0[6] - bf0[9];
  bf1[10] = bf0[5] - bf0[10];
  bf1[11] = bf0[4] - bf0[11];
  bf1[12] = bf0[3] - bf0[12];
  bf1[13] = bf0[2] - bf0[13];
  bf1[14] = bf0[1] - bf0[14];
  bf1[15] = bf0[0] - bf0[15];
  bf1[16] = bf0[16];
  bf1[17] = bf0[17];
  bf1[18] = bf0[18];
  bf1[19] = bf0[19];
710 711 712 713 714 715 716 717
  bf1[20] = half_btf(-cospi[32], bf0[20], cospi[32], bf0[27], cos_bit);
  bf1[21] = half_btf(-cospi[32], bf0[21], cospi[32], bf0[26], cos_bit);
  bf1[22] = half_btf(-cospi[32], bf0[22], cospi[32], bf0[25], cos_bit);
  bf1[23] = half_btf(-cospi[32], bf0[23], cospi[32], bf0[24], cos_bit);
  bf1[24] = half_btf(cospi[32], bf0[23], cospi[32], bf0[24], cos_bit);
  bf1[25] = half_btf(cospi[32], bf0[22], cospi[32], bf0[25], cos_bit);
  bf1[26] = half_btf(cospi[32], bf0[21], cospi[32], bf0[26], cos_bit);
  bf1[27] = half_btf(cospi[32], bf0[20], cospi[32], bf0[27], cos_bit);
Angie Chiang's avatar
Angie Chiang committed
718 719 720 721
  bf1[28] = bf0[28];
  bf1[29] = bf0[29];
  bf1[30] = bf0[30];
  bf1[31] = bf0[31];
722
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759

  // stage 9
  stage++;
  bf0 = step;
  bf1 = output;
  bf1[0] = bf0[0] + bf0[31];
  bf1[1] = bf0[1] + bf0[30];
  bf1[2] = bf0[2] + bf0[29];
  bf1[3] = bf0[3] + bf0[28];
  bf1[4] = bf0[4] + bf0[27];
  bf1[5] = bf0[5] + bf0[26];
  bf1[6] = bf0[6] + bf0[25];
  bf1[7] = bf0[7] + bf0[24];
  bf1[8] = bf0[8] + bf0[23];
  bf1[9] = bf0[9] + bf0[22];
  bf1[10] = bf0[10] + bf0[21];
  bf1[11] = bf0[11] + bf0[20];
  bf1[12] = bf0[12] + bf0[19];
  bf1[13] = bf0[13] + bf0[18];
  bf1[14] = bf0[14] + bf0[17];
  bf1[15] = bf0[15] + bf0[16];
  bf1[16] = bf0[15] - bf0[16];
  bf1[17] = bf0[14] - bf0[17];
  bf1[18] = bf0[13] - bf0[18];
  bf1[19] = bf0[12] - bf0[19];
  bf1[20] = bf0[11] - bf0[20];
  bf1[21] = bf0[10] - bf0[21];
  bf1[22] = bf0[9] - bf0[22];
  bf1[23] = bf0[8] - bf0[23];
  bf1[24] = bf0[7] - bf0[24];
  bf1[25] = bf0[6] - bf0[25];
  bf1[26] = bf0[5] - bf0[26];
  bf1[27] = bf0[4] - bf0[27];
  bf1[28] = bf0[3] - bf0[28];
  bf1[29] = bf0[2] - bf0[29];
  bf1[30] = bf0[1] - bf0[30];
  bf1[31] = bf0[0] - bf0[31];
760
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
761
}
Angie Chiang's avatar
Angie Chiang committed
762

763 764
void av1_iadst4_new(const int32_t *input, int32_t *output, int8_t cos_bit,
                    const int8_t *stage_range) {
Yaowu Xu's avatar
Yaowu Xu committed
765 766
  const int32_t size = 4;
  int32_t stage = 0;
767
  int bit = cos_bit;
768 769
  const int32_t *sinpi = sinpi_arr(bit);
  int32_t s0, s1, s2, s3, s4, s5, s6, s7;
Yaowu Xu's avatar
Yaowu Xu committed
770
  // stage 0;
771
  apply_range(stage, input, input, size, stage_range[stage]);
772 773 774 775 776 777 778 779 780 781

  int32_t x0 = input[0];
  int32_t x1 = input[1];
  int32_t x2 = input[2];
  int32_t x3 = input[3];

  if (!(x0 | x1 | x2 | x3)) {
    output[0] = output[1] = output[2] = output[3] = 0;
    return;
  }
Angie Chiang's avatar
Angie Chiang committed
782

Angie Chiang's avatar
Angie Chiang committed
783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800
  // stage 1
  s0 = range_check_value(sinpi[1] * x0, stage_range[1]);
  s1 = range_check_value(sinpi[2] * x0, stage_range[1]);
  s2 = range_check_value(sinpi[3] * x1, stage_range[1]);
  s3 = range_check_value(sinpi[4] * x2, stage_range[1]);
  s4 = range_check_value(sinpi[1] * x2, stage_range[1]);
  s5 = range_check_value(sinpi[2] * x3, stage_range[1]);
  s6 = range_check_value(sinpi[4] * x3, stage_range[1]);
  s7 = clamp_value(x0 - x2, stage_range[1]);

  // stage 2
  s7 = clamp_value(s7 + x3, stage_range[2]);

  // stage 3
  s0 = range_check_value(s0 + s3, stage_range[3] + bit);
  s1 = range_check_value(s1 - s4, stage_range[3] + bit);
  s3 = range_check_value(s2, stage_range[3] + bit);
  s2 = range_check_value(sinpi[3] * s7, stage_range[3] + bit);
801

Angie Chiang's avatar
Angie Chiang committed
802 803 804 805 806 807 808 809 810 811 812 813
  // stage 4
  s0 = range_check_value(s0 + s5, stage_range[4] + bit);
  s1 = range_check_value(s1 - s6, stage_range[4] + bit);

  // stage 5
  x0 = range_check_value(s0 + s3, stage_range[5] + bit);
  x1 = range_check_value(s1 + s3, stage_range[5] + bit);
  x2 = range_check_value(s2, stage_range[5] + bit);
  x3 = range_check_value(s0 + s1, stage_range[5] + bit);

  // stage 6
  x3 = range_check_value(x3 - s3, stage_range[6] + bit);
814 815 816 817 818

  // 1-D transform scaling factor is sqrt(2).
  // The overall dynamic range is 14b (input) + 14b (multiplication scaling)
  // + 1b (addition) = 29b.
  // Hence the output bit depth is 15b.
Angie Chiang's avatar
Angie Chiang committed
819 820 821 822 823
  output[0] = round_shift(x0, bit);
  output[1] = round_shift(x1, bit);
  output[2] = round_shift(x2, bit);
  output[3] = round_shift(x3, bit);
  apply_range(6, input, output, size, stage_range[6]);
Angie Chiang's avatar
Angie Chiang committed
824
}
Angie Chiang's avatar
Angie Chiang committed
825

826 827
void av1_iadst8_new(const int32_t *input, int32_t *output, int8_t cos_bit,
                    const int8_t *stage_range) {
Angie Chiang's avatar
Angie Chiang committed
828 829 830 831 832 833 834 835
  const int32_t size = 8;
  const int32_t *cospi;

  int32_t stage = 0;
  int32_t *bf0, *bf1;
  int32_t step[8];

  // stage 0;
836
  apply_range(stage, input, input, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
837 838 839 840

  // stage 1;
  stage++;
  bf1 = output;
841 842 843 844 845 846 847 848
  bf1[0] = input[7];
  bf1[1] = input[0];
  bf1[2] = input[5];
  bf1[3] = input[2];
  bf1[4] = input[3];
  bf1[5] = input[4];
  bf1[6] = input[1];
  bf1[7] = input[6];
849
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
850 851 852

  // stage 2
  stage++;
853
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
854 855
  bf0 = output;
  bf1 = step;
856 857 858 859 860 861 862 863
  bf1[0] = half_btf(cospi[4], bf0[0], cospi[60], bf0[1], cos_bit);
  bf1[1] = half_btf(-cospi[4], bf0[1], cospi[60], bf0[0], cos_bit);
  bf1[2] = half_btf(cospi[20], bf0[2], cospi[44], bf0[3], cos_bit);
  bf1[3] = half_btf(-cospi[20], bf0[3], cospi[44], bf0[2], cos_bit);
  bf1[4] = half_btf(cospi[36], bf0[4], cospi[28], bf0[5], cos_bit);
  bf1[5] = half_btf(-cospi[36], bf0[5], cospi[28], bf0[4], cos_bit);
  bf1[6] = half_btf(cospi[52], bf0[6], cospi[12], bf0[7], cos_bit);
  bf1[7] = half_btf(-cospi[52], bf0[7], cospi[12], bf0[6], cos_bit);
864
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
865 866 867 868 869

  // stage 3
  stage++;
  bf0 = step;
  bf1 = output;
870 871 872 873 874 875 876 877
  bf1[0] = bf0[0] + bf0[4];
  bf1[1] = bf0[1] + bf0[5];
  bf1[2] = bf0[2] + bf0[6];
  bf1[3] = bf0[3] + bf0[7];
  bf1[4] = -bf0[4] + bf0[0];
  bf1[5] = -bf0[5] + bf0[1];
  bf1[6] = -bf0[6] + bf0[2];
  bf1[7] = -bf0[7] + bf0[3];
878
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
879 880 881

  // stage 4
  stage++;
882
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
883 884 885 886 887 888
  bf0 = output;
  bf1 = step;
  bf1[0] = bf0[0];
  bf1[1] = bf0[1];
  bf1[2] = bf0[2];
  bf1[3] = bf0[3];
889 890 891 892
  bf1[4] = half_btf(cospi[16], bf0[4], cospi[48], bf0[5], cos_bit);
  bf1[5] = half_btf(-cospi[16], bf0[5], cospi[48], bf0[4], cos_bit);
  bf1[6] = half_btf(-cospi[48], bf0[6], cospi[16], bf0[7], cos_bit);
  bf1[7] = half_btf(cospi[48], bf0[7], cospi[16], bf0[6], cos_bit);
893
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
894 895 896 897 898

  // stage 5
  stage++;
  bf0 = step;
  bf1 = output;
899 900 901 902 903 904 905 906
  bf1[0] = bf0[0] + bf0[2];
  bf1[1] = bf0[1] + bf0[3];
  bf1[2] = -bf0[2] + bf0[0];
  bf1[3] = -bf0[3] + bf0[1];
  bf1[4] = bf0[4] + bf0[6];
  bf1[5] = bf0[5] + bf0[7];
  bf1[6] = -bf0[6] + bf0[4];
  bf1[7] = -bf0[7] + bf0[5];
907
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
908 909 910

  // stage 6
  stage++;
911
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
912 913
  bf0 = output;
  bf1 = step;
914 915
  bf1[0] = bf0[0];
  bf1[1] = bf0[1];
916 917
  bf1[2] = half_btf(cospi[32], bf0[2], cospi[32], bf0[3], cos_bit);
  bf1[3] = half_btf(-cospi[32], bf0[3], cospi[32], bf0[2], cos_bit);
918 919
  bf1[4] = bf0[4];
  bf1[5] = bf0[5];
920 921
  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[7], cos_bit);
  bf1[7] = half_btf(-cospi[32], bf0[7], cospi[32], bf0[6], cos_bit);
922
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
923 924 925 926 927

  // stage 7
  stage++;
  bf0 = step;
  bf1 = output;
928 929 930 931 932 933 934 935
  bf1[0] = bf0[0];
  bf1[1] = -bf0[4];
  bf1[2] = bf0[6];
  bf1[3] = -bf0[2];
  bf1[4] = bf0[3];
  bf1[5] = -bf0[7];
  bf1[6] = bf0[5];
  bf1[7] = -bf0[1];
936
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
937
}
Angie Chiang's avatar
Angie Chiang committed
938

939 940
void av1_iadst16_new(const int32_t *input, int32_t *output, int8_t cos_bit,
                     const int8_t *stage_range) {
Angie Chiang's avatar
Angie Chiang committed
941 942 943 944 945 946 947 948
  const int32_t size = 16;
  const int32_t *cospi;

  int32_t stage = 0;
  int32_t *bf0, *bf1;
  int32_t step[16];

  // stage 0;
949
  apply_range(stage, input, input, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
950 951 952 953

  // stage 1;
  stage++;
  bf1 = output;
954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969
  bf1[0] = input[15];
  bf1[1] = input[0];
  bf1[2] = input[13];
  bf1[3] = input[2];
  bf1[4] = input[11];
  bf1[5] = input[4];
  bf1[6] = input[9];
  bf1[7] = input[6];
  bf1[8] = input[7];
  bf1[9] = input[8];
  bf1[10] = input[5];
  bf1[11] = input[10];
  bf1[12] = input[3];
  bf1[13] = input[12];
  bf1[14] = input[1];
  bf1[15] = input[14];
970
  apply_range(stage, input, bf1, size, stage_range[stage]);
Angie Chiang's avatar
Angie Chiang committed
971 972 973

  // stage 2
  stage++;
974
  cospi = cospi_arr(cos_bit);
Angie Chiang's avatar
Angie Chiang committed
975 976
  bf0 = output;
  bf1 = step;
977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992
  bf1[0] = half_btf(cospi[2], bf0[0], cospi[62], bf0[1], cos_bit);
  bf1[1] = half_btf(-cospi[2], bf0[1], cospi[62], bf0[0], cos_bit);
  bf1[2] = half_btf(cospi[10], bf0[2], cospi[54], bf0[3], cos_bit);
  bf1[3] = half_btf(-cospi[10], bf0[3], cospi[54], bf0[2], cos_bit);
  bf1[4] = half_btf(cospi[18], bf0[4], cospi[46], bf0[5], cos_bit);
  bf1[5] = half_btf(-cospi[18], bf0[5], cospi[46], bf0[4], cos_bit);
  bf1[6] = half_btf(cospi[26], bf0[6], cospi[38], bf0[7], cos_bit);
  bf1[7] = half_btf(-cospi[26], bf0[7], cospi[38], bf0[6], cos_bit);
  bf1[8] = half_btf(cospi[34], bf0[8], cospi[30], bf0[9], cos_bit);
  bf1[9] = half_btf(-cospi[34], bf0[9], cospi[30], bf0[8], cos_bit);
  bf1[10] = half_btf(cospi[42], bf0[10], cospi[22], bf0[11], cos_bit);
  bf1[11] = half_btf(-cospi[42], bf0[11], cospi[22], bf0[10], cos_bit);
  bf1[12] = half_btf(cospi[50], bf0[12], cospi[14], bf0[13], cos_bit);
  bf1[13] = half_btf(-cospi[50], bf0[13], cospi[14], bf0[12], cos_bit);
  bf1[14] = half_btf(cospi[58], bf0[14], cospi[6], bf0[15], cos_bit);
  bf1[15] = half_btf(-cospi[58], bf0[15], cospi[6], bf0[14], cos_bit);