idct.c 78.2 KB
Newer Older
Jingning Han's avatar
Jingning Han committed
1
/*
Yaowu Xu's avatar
Yaowu Xu committed
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
Jingning Han's avatar
Jingning Han committed
3
 *
Yaowu Xu's avatar
Yaowu Xu committed
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.
Jingning Han's avatar
Jingning Han committed
10 11 12 13
 */

#include <math.h>

Yaowu Xu's avatar
Yaowu Xu committed
14
#include "./aom_dsp_rtcd.h"
Geza Lore's avatar
Geza Lore committed
15 16 17
#include "./av1_rtcd.h"
#include "aom_dsp/inv_txfm.h"
#include "aom_ports/mem.h"
18
#include "av1/common/av1_inv_txfm1d_cfg.h"
19 20 21
#include "av1/common/blockd.h"
#include "av1/common/enums.h"
#include "av1/common/idct.h"
22 23 24 25
#if CONFIG_DAALA_DCT4 || CONFIG_DAALA_DCT8 || CONFIG_DAALA_DCT16 || \
    CONFIG_DAALA_DCT32 || CONFIG_DAALA_DCT64
#include "av1/common/daala_tx.h"
#endif
Jingning Han's avatar
Jingning Han committed
26

27
int av1_get_tx_scale(const TX_SIZE tx_size) {
28 29
  const int pels = tx_size_2d[tx_size];
  return (pels > 256) + (pels > 1024) + (pels > 4096);
30 31
}

32 33 34
// NOTE: The implementation of all inverses need to be aware of the fact
// that input and output could be the same buffer.

35
#if CONFIG_EXT_TX
Debargha Mukherjee's avatar
Debargha Mukherjee committed
36 37
static void iidtx4_c(const tran_low_t *input, tran_low_t *output) {
  int i;
38
  for (i = 0; i < 4; ++i) {
Debargha Mukherjee's avatar
Debargha Mukherjee committed
39
    output[i] = (tran_low_t)dct_const_round_shift(input[i] * Sqrt2);
40
  }
Debargha Mukherjee's avatar
Debargha Mukherjee committed
41 42 43 44
}

static void iidtx8_c(const tran_low_t *input, tran_low_t *output) {
  int i;
45 46 47
  for (i = 0; i < 8; ++i) {
    output[i] = input[i] * 2;
  }
Debargha Mukherjee's avatar
Debargha Mukherjee committed
48 49 50 51
}

static void iidtx16_c(const tran_low_t *input, tran_low_t *output) {
  int i;
52
  for (i = 0; i < 16; ++i) {
Debargha Mukherjee's avatar
Debargha Mukherjee committed
53
    output[i] = (tran_low_t)dct_const_round_shift(input[i] * 2 * Sqrt2);
54
  }
Debargha Mukherjee's avatar
Debargha Mukherjee committed
55 56 57 58
}

static void iidtx32_c(const tran_low_t *input, tran_low_t *output) {
  int i;
59 60 61
  for (i = 0; i < 32; ++i) {
    output[i] = input[i] * 4;
  }
Debargha Mukherjee's avatar
Debargha Mukherjee committed
62
}
63

64
#if CONFIG_TX64X64 && !CONFIG_DAALA_DCT64
65 66
static void iidtx64_c(const tran_low_t *input, tran_low_t *output) {
  int i;
67
  for (i = 0; i < 64; ++i) {
68
    output[i] = (tran_low_t)dct_const_round_shift(input[i] * 4 * Sqrt2);
69
  }
70 71
}
#endif  // CONFIG_TX64X64
72
#endif  // CONFIG_EXT_TX
Debargha Mukherjee's avatar
Debargha Mukherjee committed
73

74
// For use in lieu of ADST
75 76 77 78 79 80 81
static void ihalfright32_c(const tran_low_t *input, tran_low_t *output) {
  int i;
  tran_low_t inputhalf[16];
  // Multiply input by sqrt(2)
  for (i = 0; i < 16; ++i) {
    inputhalf[i] = (tran_low_t)dct_const_round_shift(input[i] * Sqrt2);
  }
82 83 84
  for (i = 0; i < 16; ++i) {
    output[i] = input[16 + i] * 4;
  }
Luca Barbato's avatar
Luca Barbato committed
85
  aom_idct16_c(inputhalf, output + 16);
86 87 88
  // Note overall scaling factor is 4 times orthogonal
}

89
#if CONFIG_TX64X64
90 91 92 93 94 95 96 97 98 99 100
#if CONFIG_DAALA_DCT64
static void ihalfright64_c(const tran_low_t *input, tran_low_t *output) {
  int i;
  tran_low_t inputhalf[32];
  // No scaling within; Daala transforms are all orthonormal
  for (i = 0; i < 32; ++i) {
    inputhalf[i] = input[i];
  }
  for (i = 0; i < 32; ++i) {
    output[i] = input[32 + i];
  }
101
  daala_idct32(inputhalf, output + 32);
102 103
}
#else
104 105 106 107
static void idct64_col_c(const tran_low_t *input, tran_low_t *output) {
  int32_t in[64], out[64];
  int i;
  for (i = 0; i < 64; ++i) in[i] = (int32_t)input[i];
108
  av1_idct64_new(in, out, inv_cos_bit_col_dct_64, inv_stage_range_col_dct_64);
109 110 111 112 113 114 115
  for (i = 0; i < 64; ++i) output[i] = (tran_low_t)out[i];
}

static void idct64_row_c(const tran_low_t *input, tran_low_t *output) {
  int32_t in[64], out[64];
  int i;
  for (i = 0; i < 64; ++i) in[i] = (int32_t)input[i];
116
  av1_idct64_new(in, out, inv_cos_bit_row_dct_64, inv_stage_range_row_dct_64);
117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
  for (i = 0; i < 64; ++i) output[i] = (tran_low_t)out[i];
}

// For use in lieu of ADST
static void ihalfright64_c(const tran_low_t *input, tran_low_t *output) {
  int i;
  tran_low_t inputhalf[32];
  // Multiply input by sqrt(2)
  for (i = 0; i < 32; ++i) {
    inputhalf[i] = (tran_low_t)dct_const_round_shift(input[i] * Sqrt2);
  }
  for (i = 0; i < 32; ++i) {
    output[i] = (tran_low_t)dct_const_round_shift(input[32 + i] * 4 * Sqrt2);
  }
  aom_idct32_c(inputhalf, output + 32);
  // Note overall scaling factor is 4 * sqrt(2)  times orthogonal
}
134
#endif  // CONFIG_DAALA_DCT64
135 136
#endif  // CONFIG_TX64X64

Jingning Han's avatar
Jingning Han committed
137
// Inverse identity transform and add.
138
#if CONFIG_EXT_TX
139
static void inv_idtx_add_c(const tran_low_t *input, uint8_t *dest, int stride,
Jingning Han's avatar
Jingning Han committed
140
                           int bs, int tx_type) {
141
  int r, c;
142
  const int shift = bs < 32 ? 3 : (bs < 64 ? 2 : 1);
Debargha Mukherjee's avatar
Debargha Mukherjee committed
143
  if (tx_type == IDTX) {
Jingning Han's avatar
Jingning Han committed
144 145
    for (r = 0; r < bs; ++r) {
      for (c = 0; c < bs; ++c)
Debargha Mukherjee's avatar
Debargha Mukherjee committed
146 147 148
        dest[c] = clip_pixel_add(dest[c], input[c] >> shift);
      dest += stride;
      input += bs;
Jingning Han's avatar
Jingning Han committed
149
    }
150 151
  }
}
152
#endif  // CONFIG_EXT_TX
153

clang-format's avatar
clang-format committed
154 155 156 157 158
#define FLIPUD_PTR(dest, stride, size)       \
  do {                                       \
    (dest) = (dest) + ((size)-1) * (stride); \
    (stride) = -(stride);                    \
  } while (0)
159

160
#if CONFIG_EXT_TX
clang-format's avatar
clang-format committed
161 162 163
static void maybe_flip_strides(uint8_t **dst, int *dstride, tran_low_t **src,
                               int *sstride, int tx_type, int sizey,
                               int sizex) {
164 165 166 167 168 169 170 171
  // Note that the transpose of src will be added to dst. In order to LR
  // flip the addends (in dst coordinates), we UD flip the src. To UD flip
  // the addends, we UD flip the dst.
  switch (tx_type) {
    case DCT_DCT:
    case ADST_DCT:
    case DCT_ADST:
    case ADST_ADST:
Debargha Mukherjee's avatar
Debargha Mukherjee committed
172
    case IDTX:
Jingning Han's avatar
Jingning Han committed
173 174
    case V_DCT:
    case H_DCT:
175
    case V_ADST:
clang-format's avatar
clang-format committed
176
    case H_ADST: break;
177 178
    case FLIPADST_DCT:
    case FLIPADST_ADST:
179
    case V_FLIPADST:
180
      // flip UD
181
      FLIPUD_PTR(*dst, *dstride, sizey);
182 183 184
      break;
    case DCT_FLIPADST:
    case ADST_FLIPADST:
185
    case H_FLIPADST:
186
      // flip LR
187
      FLIPUD_PTR(*src, *sstride, sizex);
188 189 190
      break;
    case FLIPADST_FLIPADST:
      // flip UD
191
      FLIPUD_PTR(*dst, *dstride, sizey);
192
      // flip LR
193
      FLIPUD_PTR(*src, *sstride, sizex);
194
      break;
clang-format's avatar
clang-format committed
195
    default: assert(0); break;
196 197
  }
}
198
#endif  // CONFIG_EXT_TX
199

200
#if CONFIG_HIGHBITDEPTH
201
#if CONFIG_EXT_TX && CONFIG_TX64X64
202
static void highbd_inv_idtx_add_c(const tran_low_t *input, uint8_t *dest8,
203
                                  int stride, int bs, int tx_type, int bd) {
204 205 206
  int r, c;
  const int shift = bs < 32 ? 3 : 2;
  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
207

Debargha Mukherjee's avatar
Debargha Mukherjee committed
208
  if (tx_type == IDTX) {
209 210
    for (r = 0; r < bs; ++r) {
      for (c = 0; c < bs; ++c)
Debargha Mukherjee's avatar
Debargha Mukherjee committed
211 212 213
        dest[c] = highbd_clip_pixel_add(dest[c], input[c] >> shift, bd);
      dest += stride;
      input += bs;
214
    }
215 216
  }
}
217
#endif  // CONFIG_EXT_TX && CONFIG_TX64X64
218
#endif  // CONFIG_HIGHBITDEPTH
219

Lester Lu's avatar
Lester Lu committed
220 221 222
#if CONFIG_LGT
void ilgt4(const tran_low_t *input, tran_low_t *output,
           const tran_high_t *lgtmtx) {
Lester Lu's avatar
Lester Lu committed
223 224
  if (!lgtmtx) assert(0);
  // evaluate s[j] = sum of all lgtmtx[j]*input[i] over i=1,...,4
Lester Lu's avatar
Lester Lu committed
225 226 227 228 229 230 231 232 233
  tran_high_t s[4] = { 0 };
  for (int i = 0; i < 4; ++i)
    for (int j = 0; j < 4; ++j) s[j] += lgtmtx[i * 4 + j] * input[i];

  for (int i = 0; i < 4; ++i) output[i] = WRAPLOW(dct_const_round_shift(s[i]));
}

void ilgt8(const tran_low_t *input, tran_low_t *output,
           const tran_high_t *lgtmtx) {
Lester Lu's avatar
Lester Lu committed
234 235
  if (!lgtmtx) assert(0);
  // evaluate s[j] = sum of all lgtmtx[j]*input[i] over i=1,...,8
Lester Lu's avatar
Lester Lu committed
236 237 238 239 240 241 242
  tran_high_t s[8] = { 0 };
  for (int i = 0; i < 8; ++i)
    for (int j = 0; j < 8; ++j) s[j] += lgtmtx[i * 8 + j] * input[i];

  for (int i = 0; i < 8; ++i) output[i] = WRAPLOW(dct_const_round_shift(s[i]));
}

Lester Lu's avatar
Lester Lu committed
243 244 245 246 247 248 249 250 251 252 253
// get_lgt4 and get_lgt8 return 1 and pick a lgt matrix if LGT is chosen to
// apply. Otherwise they return 0
int get_lgt4(const TxfmParam *txfm_param, int is_col,
             const tran_high_t **lgtmtx) {
  if (is_col && (vtx_tab[txfm_param->tx_type] == ADST_1D ||
                 vtx_tab[txfm_param->tx_type] == FLIPADST_1D)) {
    lgtmtx[0] = txfm_param->is_inter ? &lgt4_170[0][0] : &lgt4_140[0][0];
    return 1;
  } else if (!is_col && (htx_tab[txfm_param->tx_type] == ADST_1D ||
                         htx_tab[txfm_param->tx_type] == FLIPADST_1D)) {
    lgtmtx[0] = txfm_param->is_inter ? &lgt4_170[0][0] : &lgt4_140[0][0];
Lester Lu's avatar
Lester Lu committed
254 255
    return 1;
  }
Lester Lu's avatar
Lester Lu committed
256
  lgtmtx[0] = NULL;
Lester Lu's avatar
Lester Lu committed
257 258 259
  return 0;
}

Lester Lu's avatar
Lester Lu committed
260 261 262 263 264 265 266 267 268
int get_lgt8(const TxfmParam *txfm_param, int is_col,
             const tran_high_t **lgtmtx) {
  if (is_col && (vtx_tab[txfm_param->tx_type] == ADST_1D ||
                 vtx_tab[txfm_param->tx_type] == FLIPADST_1D)) {
    lgtmtx[0] = txfm_param->is_inter ? &lgt8_170[0][0] : &lgt8_150[0][0];
    return 1;
  } else if (!is_col && (htx_tab[txfm_param->tx_type] == ADST_1D ||
                         htx_tab[txfm_param->tx_type] == FLIPADST_1D)) {
    lgtmtx[0] = txfm_param->is_inter ? &lgt8_170[0][0] : &lgt8_150[0][0];
Lester Lu's avatar
Lester Lu committed
269 270
    return 1;
  }
Lester Lu's avatar
Lester Lu committed
271
  lgtmtx[0] = NULL;
Lester Lu's avatar
Lester Lu committed
272 273 274 275
  return 0;
}
#endif  // CONFIG_LGT

Yaowu Xu's avatar
Yaowu Xu committed
276
void av1_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride,
277 278
                         const TxfmParam *txfm_param) {
  int tx_type = txfm_param->tx_type;
Sarah Parker's avatar
Sarah Parker committed
279 280 281
#if CONFIG_MRC_TX
  assert(tx_type != MRC_DCT && "Invalid tx type for tx size");
#endif  // CONFIG_MRC_TX
282
#if !CONFIG_DAALA_DCT4
283 284 285 286
  if (tx_type == DCT_DCT) {
    aom_idct4x4_16_add(input, dest, stride);
    return;
  }
287
#endif
288
  static const transform_2d IHT_4[] = {
289
#if CONFIG_DAALA_DCT4
290 291 292 293
    { daala_idct4, daala_idct4 },  // DCT_DCT  = 0
    { daala_idst4, daala_idct4 },  // ADST_DCT = 1
    { daala_idct4, daala_idst4 },  // DCT_ADST = 2
    { daala_idst4, daala_idst4 },  // ADST_ADST = 3
294
#if CONFIG_EXT_TX
295 296 297 298 299
    { daala_idst4, daala_idct4 },  // FLIPADST_DCT
    { daala_idct4, daala_idst4 },  // DCT_FLIPADST
    { daala_idst4, daala_idst4 },  // FLIPADST_FLIPADST
    { daala_idst4, daala_idst4 },  // ADST_FLIPADST
    { daala_idst4, daala_idst4 },  // FLIPADST_ADST
300 301 302 303 304 305 306
    { daala_idtx4, daala_idtx4 },  // IDTX
    { daala_idct4, daala_idtx4 },  // V_DCT
    { daala_idtx4, daala_idct4 },  // H_DCT
    { daala_idst4, daala_idtx4 },  // V_ADST
    { daala_idtx4, daala_idst4 },  // H_ADST
    { daala_idst4, daala_idtx4 },  // V_FLIPADST
    { daala_idtx4, daala_idst4 },  // H_FLIPADST
307 308
#endif
#else
Luca Barbato's avatar
Luca Barbato committed
309 310 311 312
    { aom_idct4_c, aom_idct4_c },    // DCT_DCT  = 0
    { aom_iadst4_c, aom_idct4_c },   // ADST_DCT = 1
    { aom_idct4_c, aom_iadst4_c },   // DCT_ADST = 2
    { aom_iadst4_c, aom_iadst4_c },  // ADST_ADST = 3
313
#if CONFIG_EXT_TX
Luca Barbato's avatar
Luca Barbato committed
314 315 316 317 318 319 320 321 322 323 324 325
    { aom_iadst4_c, aom_idct4_c },   // FLIPADST_DCT
    { aom_idct4_c, aom_iadst4_c },   // DCT_FLIPADST
    { aom_iadst4_c, aom_iadst4_c },  // FLIPADST_FLIPADST
    { aom_iadst4_c, aom_iadst4_c },  // ADST_FLIPADST
    { aom_iadst4_c, aom_iadst4_c },  // FLIPADST_ADST
    { iidtx4_c, iidtx4_c },          // IDTX
    { aom_idct4_c, iidtx4_c },       // V_DCT
    { iidtx4_c, aom_idct4_c },       // H_DCT
    { aom_iadst4_c, iidtx4_c },      // V_ADST
    { iidtx4_c, aom_iadst4_c },      // H_ADST
    { aom_iadst4_c, iidtx4_c },      // V_FLIPADST
    { iidtx4_c, aom_iadst4_c },      // H_FLIPADST
326
#endif
327
#endif
328 329 330
  };

  int i, j;
331
  tran_low_t tmp[4][4];
332 333 334
  tran_low_t out[4][4];
  tran_low_t *outp = &out[0][0];
  int outstride = 4;
335

336 337 338 339
#if CONFIG_DCT_ONLY
  assert(tx_type == DCT_DCT);
#endif

Lester Lu's avatar
Lester Lu committed
340
#if CONFIG_LGT
Lester Lu's avatar
Lester Lu committed
341 342 343 344
  const tran_high_t *lgtmtx_col[1];
  const tran_high_t *lgtmtx_row[1];
  int use_lgt_col = get_lgt4(txfm_param, 1, lgtmtx_col);
  int use_lgt_row = get_lgt4(txfm_param, 0, lgtmtx_row);
Lester Lu's avatar
Lester Lu committed
345 346
#endif

347 348
  // inverse transform row vectors
  for (i = 0; i < 4; ++i) {
349 350
#if CONFIG_DAALA_DCT4
    tran_low_t temp_in[4];
351
    for (j = 0; j < 4; j++) temp_in[j] = input[j] * 2;
352 353
    IHT_4[tx_type].rows(temp_in, out[i]);
#else
Lester Lu's avatar
Lester Lu committed
354 355
#if CONFIG_LGT
    if (use_lgt_row)
Lester Lu's avatar
Lester Lu committed
356
      ilgt4(input, out[i], lgtmtx_row[0]);
Lester Lu's avatar
Lester Lu committed
357 358 359
    else
#endif
      IHT_4[tx_type].rows(input, out[i]);
360
#endif
clang-format's avatar
clang-format committed
361
    input += 4;
362 363 364
  }

  // transpose
365 366 367
  for (i = 0; i < 4; i++) {
    for (j = 0; j < 4; j++) {
      tmp[j][i] = out[i][j];
368
    }
369 370 371 372
  }

  // inverse transform column vectors
  for (i = 0; i < 4; ++i) {
Lester Lu's avatar
Lester Lu committed
373 374
#if CONFIG_LGT
    if (use_lgt_col)
Lester Lu's avatar
Lester Lu committed
375
      ilgt4(tmp[i], out[i], lgtmtx_col[0]);
Lester Lu's avatar
Lester Lu committed
376 377 378
    else
#endif
      IHT_4[tx_type].cols(tmp[i], out[i]);
379 380 381
  }

#if CONFIG_EXT_TX
382
  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, 4, 4);
383 384 385 386
#endif

  // Sum with the destination
  for (i = 0; i < 4; ++i) {
387
    for (j = 0; j < 4; ++j) {
388 389
      int d = i * stride + j;
      int s = j * outstride + i;
390
#if CONFIG_DAALA_DCT4
391
      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 4));
392 393 394
#else
      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 4));
#endif
395 396 397 398
    }
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
399
void av1_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int stride,
400 401
                         const TxfmParam *txfm_param) {
  int tx_type = txfm_param->tx_type;
Sarah Parker's avatar
Sarah Parker committed
402 403 404
#if CONFIG_MRC_TX
  assert(tx_type != MRC_DCT && "Invalid tx type for tx size");
#endif  // CONFIG_MRC_TX
405 406 407
#if CONFIG_DCT_ONLY
  assert(tx_type == DCT_DCT);
#endif
408
  static const transform_2d IHT_4x8[] = {
Luca Barbato's avatar
Luca Barbato committed
409 410 411 412
    { aom_idct8_c, aom_idct4_c },    // DCT_DCT
    { aom_iadst8_c, aom_idct4_c },   // ADST_DCT
    { aom_idct8_c, aom_iadst4_c },   // DCT_ADST
    { aom_iadst8_c, aom_iadst4_c },  // ADST_ADST
413
#if CONFIG_EXT_TX
Luca Barbato's avatar
Luca Barbato committed
414 415 416 417 418 419 420 421 422 423 424 425
    { aom_iadst8_c, aom_idct4_c },   // FLIPADST_DCT
    { aom_idct8_c, aom_iadst4_c },   // DCT_FLIPADST
    { aom_iadst8_c, aom_iadst4_c },  // FLIPADST_FLIPADST
    { aom_iadst8_c, aom_iadst4_c },  // ADST_FLIPADST
    { aom_iadst8_c, aom_iadst4_c },  // FLIPADST_ADST
    { iidtx8_c, iidtx4_c },          // IDTX
    { aom_idct8_c, iidtx4_c },       // V_DCT
    { iidtx8_c, aom_idct4_c },       // H_DCT
    { aom_iadst8_c, iidtx4_c },      // V_ADST
    { iidtx8_c, aom_iadst4_c },      // H_ADST
    { aom_iadst8_c, iidtx4_c },      // V_FLIPADST
    { iidtx8_c, aom_iadst4_c },      // H_FLIPADST
426
#endif
427 428
  };

429 430
  const int n = 4;
  const int n2 = 8;
431
  int i, j;
432
  tran_low_t out[4][8], tmp[4][8], outtmp[4];
433
  tran_low_t *outp = &out[0][0];
434
  int outstride = n2;
435

Lester Lu's avatar
Lester Lu committed
436
#if CONFIG_LGT
Lester Lu's avatar
Lester Lu committed
437 438 439 440
  const tran_high_t *lgtmtx_col[1];
  const tran_high_t *lgtmtx_row[1];
  int use_lgt_col = get_lgt8(txfm_param, 1, lgtmtx_col);
  int use_lgt_row = get_lgt4(txfm_param, 0, lgtmtx_row);
Lester Lu's avatar
Lester Lu committed
441 442
#endif

443
  // inverse transform row vectors and transpose
444
  for (i = 0; i < n2; ++i) {
Lester Lu's avatar
Lester Lu committed
445 446
#if CONFIG_LGT
    if (use_lgt_row)
Lester Lu's avatar
Lester Lu committed
447
      ilgt4(input, outtmp, lgtmtx_row[0]);
Lester Lu's avatar
Lester Lu committed
448 449 450
    else
#endif
      IHT_4x8[tx_type].rows(input, outtmp);
451
    for (j = 0; j < n; ++j)
452
      tmp[j][i] = (tran_low_t)dct_const_round_shift(outtmp[j] * Sqrt2);
453
    input += n;
454 455 456
  }

  // inverse transform column vectors
457
  for (i = 0; i < n; ++i) {
Lester Lu's avatar
Lester Lu committed
458 459
#if CONFIG_LGT
    if (use_lgt_col)
Lester Lu's avatar
Lester Lu committed
460
      ilgt8(tmp[i], out[i], lgtmtx_col[0]);
Lester Lu's avatar
Lester Lu committed
461 462 463
    else
#endif
      IHT_4x8[tx_type].cols(tmp[i], out[i]);
464 465
  }

466
#if CONFIG_EXT_TX
467
  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n2, n);
468
#endif
469 470

  // Sum with the destination
471 472
  for (i = 0; i < n2; ++i) {
    for (j = 0; j < n; ++j) {
473 474 475 476 477 478 479
      int d = i * stride + j;
      int s = j * outstride + i;
      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5));
    }
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
480
void av1_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int stride,
481 482
                         const TxfmParam *txfm_param) {
  int tx_type = txfm_param->tx_type;
Sarah Parker's avatar
Sarah Parker committed
483 484 485
#if CONFIG_MRC_TX
  assert(tx_type != MRC_DCT && "Invalid tx type for tx size");
#endif  // CONFIG_MRC_TX
486 487 488
#if CONFIG_DCT_ONLY
  assert(tx_type == DCT_DCT);
#endif
489
  static const transform_2d IHT_8x4[] = {
Luca Barbato's avatar
Luca Barbato committed
490 491 492 493
    { aom_idct4_c, aom_idct8_c },    // DCT_DCT
    { aom_iadst4_c, aom_idct8_c },   // ADST_DCT
    { aom_idct4_c, aom_iadst8_c },   // DCT_ADST
    { aom_iadst4_c, aom_iadst8_c },  // ADST_ADST
494
#if CONFIG_EXT_TX
Luca Barbato's avatar
Luca Barbato committed
495 496 497 498 499 500 501 502 503 504 505 506
    { aom_iadst4_c, aom_idct8_c },   // FLIPADST_DCT
    { aom_idct4_c, aom_iadst8_c },   // DCT_FLIPADST
    { aom_iadst4_c, aom_iadst8_c },  // FLIPADST_FLIPADST
    { aom_iadst4_c, aom_iadst8_c },  // ADST_FLIPADST
    { aom_iadst4_c, aom_iadst8_c },  // FLIPADST_ADST
    { iidtx4_c, iidtx8_c },          // IDTX
    { aom_idct4_c, iidtx8_c },       // V_DCT
    { iidtx4_c, aom_idct8_c },       // H_DCT
    { aom_iadst4_c, iidtx8_c },      // V_ADST
    { iidtx4_c, aom_iadst8_c },      // H_ADST
    { aom_iadst4_c, iidtx8_c },      // V_FLIPADST
    { iidtx4_c, aom_iadst8_c },      // H_FLIPADST
507
#endif
508
  };
509

510 511
  const int n = 4;
  const int n2 = 8;
512 513

  int i, j;
514
  tran_low_t out[8][4], tmp[8][4], outtmp[8];
515
  tran_low_t *outp = &out[0][0];
516
  int outstride = n;
517

Lester Lu's avatar
Lester Lu committed
518
#if CONFIG_LGT
Lester Lu's avatar
Lester Lu committed
519 520 521 522
  const tran_high_t *lgtmtx_col[1];
  const tran_high_t *lgtmtx_row[1];
  int use_lgt_col = get_lgt4(txfm_param, 1, lgtmtx_col);
  int use_lgt_row = get_lgt8(txfm_param, 0, lgtmtx_row);
Lester Lu's avatar
Lester Lu committed
523 524
#endif

525
  // inverse transform row vectors and transpose
526
  for (i = 0; i < n; ++i) {
Lester Lu's avatar
Lester Lu committed
527 528
#if CONFIG_LGT
    if (use_lgt_row)
Lester Lu's avatar
Lester Lu committed
529
      ilgt8(input, outtmp, lgtmtx_row[0]);
Lester Lu's avatar
Lester Lu committed
530 531 532
    else
#endif
      IHT_8x4[tx_type].rows(input, outtmp);
533
    for (j = 0; j < n2; ++j)
534
      tmp[j][i] = (tran_low_t)dct_const_round_shift(outtmp[j] * Sqrt2);
535
    input += n2;
536 537 538
  }

  // inverse transform column vectors
539
  for (i = 0; i < n2; ++i) {
Lester Lu's avatar
Lester Lu committed
540 541
#if CONFIG_LGT
    if (use_lgt_col)
Lester Lu's avatar
Lester Lu committed
542
      ilgt4(tmp[i], out[i], lgtmtx_col[0]);
Lester Lu's avatar
Lester Lu committed
543 544 545
    else
#endif
      IHT_8x4[tx_type].cols(tmp[i], out[i]);
546 547
  }

548
#if CONFIG_EXT_TX
549
  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n, n2);
550
#endif
551 552

  // Sum with the destination
553 554
  for (i = 0; i < n; ++i) {
    for (j = 0; j < n2; ++j) {
555 556 557 558 559 560 561
      int d = i * stride + j;
      int s = j * outstride + i;
      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5));
    }
  }
}

562
void av1_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int stride,
563 564
                          const TxfmParam *txfm_param) {
  int tx_type = txfm_param->tx_type;
Sarah Parker's avatar
Sarah Parker committed
565 566 567
#if CONFIG_MRC_TX
  assert(tx_type != MRC_DCT && "Invalid tx type for tx size");
#endif  // CONFIG_MRC_TX
568 569 570
#if CONFIG_DCT_ONLY
  assert(tx_type == DCT_DCT);
#endif
571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594
  static const transform_2d IHT_4x16[] = {
    { aom_idct16_c, aom_idct4_c },    // DCT_DCT
    { aom_iadst16_c, aom_idct4_c },   // ADST_DCT
    { aom_idct16_c, aom_iadst4_c },   // DCT_ADST
    { aom_iadst16_c, aom_iadst4_c },  // ADST_ADST
#if CONFIG_EXT_TX
    { aom_iadst16_c, aom_idct4_c },   // FLIPADST_DCT
    { aom_idct16_c, aom_iadst4_c },   // DCT_FLIPADST
    { aom_iadst16_c, aom_iadst4_c },  // FLIPADST_FLIPADST
    { aom_iadst16_c, aom_iadst4_c },  // ADST_FLIPADST
    { aom_iadst16_c, aom_iadst4_c },  // FLIPADST_ADST
    { iidtx16_c, iidtx4_c },          // IDTX
    { aom_idct16_c, iidtx4_c },       // V_DCT
    { iidtx16_c, aom_idct4_c },       // H_DCT
    { aom_iadst16_c, iidtx4_c },      // V_ADST
    { iidtx16_c, aom_iadst4_c },      // H_ADST
    { aom_iadst16_c, iidtx4_c },      // V_FLIPADST
    { iidtx16_c, aom_iadst4_c },      // H_FLIPADST
#endif
  };

  const int n = 4;
  const int n4 = 16;
  int i, j;
595
  tran_low_t out[4][16], tmp[4][16], outtmp[4];
596 597 598
  tran_low_t *outp = &out[0][0];
  int outstride = n4;

Lester Lu's avatar
Lester Lu committed
599
#if CONFIG_LGT
Lester Lu's avatar
Lester Lu committed
600 601
  const tran_high_t *lgtmtx_row[1];
  int use_lgt_row = get_lgt4(txfm_param, 0, lgtmtx_row);
Lester Lu's avatar
Lester Lu committed
602 603
#endif

604 605
  // inverse transform row vectors and transpose
  for (i = 0; i < n4; ++i) {
Lester Lu's avatar
Lester Lu committed
606 607
#if CONFIG_LGT
    if (use_lgt_row)
Lester Lu's avatar
Lester Lu committed
608
      ilgt4(input, outtmp, lgtmtx_row[0]);
Lester Lu's avatar
Lester Lu committed
609 610 611
    else
#endif
      IHT_4x16[tx_type].rows(input, outtmp);
612
    for (j = 0; j < n; ++j) tmp[j][i] = outtmp[j];
613 614 615 616
    input += n;
  }

  // inverse transform column vectors
Lester Lu's avatar
Lester Lu committed
617 618 619
  for (i = 0; i < n; ++i) {
    IHT_4x16[tx_type].cols(tmp[i], out[i]);
  }
620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635

#if CONFIG_EXT_TX
  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n4, n);
#endif

  // Sum with the destination
  for (i = 0; i < n4; ++i) {
    for (j = 0; j < n; ++j) {
      int d = i * stride + j;
      int s = j * outstride + i;
      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5));
    }
  }
}

void av1_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int stride,
636 637
                          const TxfmParam *txfm_param) {
  int tx_type = txfm_param->tx_type;
Sarah Parker's avatar
Sarah Parker committed
638 639 640
#if CONFIG_MRC_TX
  assert(tx_type != MRC_DCT && "Invalid tx type for tx size");
#endif  // CONFIG_MRC_TX
641 642 643
#if CONFIG_DCT_ONLY
  assert(tx_type == DCT_DCT);
#endif
644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663
  static const transform_2d IHT_16x4[] = {
    { aom_idct4_c, aom_idct16_c },    // DCT_DCT
    { aom_iadst4_c, aom_idct16_c },   // ADST_DCT
    { aom_idct4_c, aom_iadst16_c },   // DCT_ADST
    { aom_iadst4_c, aom_iadst16_c },  // ADST_ADST
#if CONFIG_EXT_TX
    { aom_iadst4_c, aom_idct16_c },   // FLIPADST_DCT
    { aom_idct4_c, aom_iadst16_c },   // DCT_FLIPADST
    { aom_iadst4_c, aom_iadst16_c },  // FLIPADST_FLIPADST
    { aom_iadst4_c, aom_iadst16_c },  // ADST_FLIPADST
    { aom_iadst4_c, aom_iadst16_c },  // FLIPADST_ADST
    { iidtx4_c, iidtx16_c },          // IDTX
    { aom_idct4_c, iidtx16_c },       // V_DCT
    { iidtx4_c, aom_idct16_c },       // H_DCT
    { aom_iadst4_c, iidtx16_c },      // V_ADST
    { iidtx4_c, aom_iadst16_c },      // H_ADST
    { aom_iadst4_c, iidtx16_c },      // V_FLIPADST
    { iidtx4_c, aom_iadst16_c },      // H_FLIPADST
#endif
  };
664

665 666 667 668
  const int n = 4;
  const int n4 = 16;

  int i, j;
669
  tran_low_t out[16][4], tmp[16][4], outtmp[16];
670 671 672
  tran_low_t *outp = &out[0][0];
  int outstride = n;

Lester Lu's avatar
Lester Lu committed
673
#if CONFIG_LGT
Lester Lu's avatar
Lester Lu committed
674 675
  const tran_high_t *lgtmtx_col[1];
  int use_lgt_col = get_lgt4(txfm_param, 1, lgtmtx_col);
Lester Lu's avatar
Lester Lu committed
676 677
#endif

678 679 680
  // inverse transform row vectors and transpose
  for (i = 0; i < n; ++i) {
    IHT_16x4[tx_type].rows(input, outtmp);
681
    for (j = 0; j < n4; ++j) tmp[j][i] = outtmp[j];
682 683 684 685
    input += n4;
  }

  // inverse transform column vectors
Lester Lu's avatar
Lester Lu committed
686 687 688
  for (i = 0; i < n4; ++i) {
#if CONFIG_LGT
    if (use_lgt_col)
Lester Lu's avatar
Lester Lu committed
689
      ilgt4(tmp[i], out[i], lgtmtx_col[0]);
Lester Lu's avatar
Lester Lu committed
690 691 692 693
    else
#endif
      IHT_16x4[tx_type].cols(tmp[i], out[i]);
  }
694 695 696 697 698 699 700 701 702 703 704 705 706 707 708

#if CONFIG_EXT_TX
  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n, n4);
#endif

  // Sum with the destination
  for (i = 0; i < n; ++i) {
    for (j = 0; j < n4; ++j) {
      int d = i * stride + j;
      int s = j * outstride + i;
      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5));
    }
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
709
void av1_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int stride,
710 711
                           const TxfmParam *txfm_param) {
  int tx_type = txfm_param->tx_type;
Sarah Parker's avatar
Sarah Parker committed
712 713 714
#if CONFIG_MRC_TX
  assert(tx_type != MRC_DCT && "Invalid tx type for tx size");
#endif  // CONFIG_MRC_TX
715 716 717
#if CONFIG_DCT_ONLY
  assert(tx_type == DCT_DCT);
#endif
718
  static const transform_2d IHT_8x16[] = {
Luca Barbato's avatar
Luca Barbato committed
719 720 721 722
    { aom_idct16_c, aom_idct8_c },    // DCT_DCT
    { aom_iadst16_c, aom_idct8_c },   // ADST_DCT
    { aom_idct16_c, aom_iadst8_c },   // DCT_ADST
    { aom_iadst16_c, aom_iadst8_c },  // ADST_ADST
723
#if CONFIG_EXT_TX
Luca Barbato's avatar
Luca Barbato committed
724 725 726 727 728 729 730 731 732 733 734 735
    { aom_iadst16_c, aom_idct8_c },   // FLIPADST_DCT
    { aom_idct16_c, aom_iadst8_c },   // DCT_FLIPADST
    { aom_iadst16_c, aom_iadst8_c },  // FLIPADST_FLIPADST
    { aom_iadst16_c, aom_iadst8_c },  // ADST_FLIPADST
    { aom_iadst16_c, aom_iadst8_c },  // FLIPADST_ADST
    { iidtx16_c, iidtx8_c },          // IDTX
    { aom_idct16_c, iidtx8_c },       // V_DCT
    { iidtx16_c, aom_idct8_c },       // H_DCT
    { aom_iadst16_c, iidtx8_c },      // V_ADST
    { iidtx16_c, aom_iadst8_c },      // H_ADST
    { aom_iadst16_c, iidtx8_c },      // V_FLIPADST
    { iidtx16_c, aom_iadst8_c },      // H_FLIPADST
736
#endif
737 738 739 740 741
  };

  const int n = 8;
  const int n2 = 16;
  int i, j;
742
  tran_low_t out[8][16], tmp[8][16], outtmp[8];
743 744 745
  tran_low_t *outp = &out[0][0];
  int outstride = n2;

Lester Lu's avatar
Lester Lu committed
746
#if CONFIG_LGT
Lester Lu's avatar
Lester Lu committed
747 748
  const tran_high_t *lgtmtx_row[1];
  int use_lgt_row = get_lgt8(txfm_param, 0, lgtmtx_row);
Lester Lu's avatar
Lester Lu committed
749 750
#endif

751 752
  // inverse transform row vectors and transpose
  for (i = 0; i < n2; ++i) {
Lester Lu's avatar
Lester Lu committed
753 754
#if CONFIG_LGT
    if (use_lgt_row)
Lester Lu's avatar
Lester Lu committed
755
      ilgt8(input, outtmp, lgtmtx_row[0]);
Lester Lu's avatar
Lester Lu committed
756 757 758
    else
#endif
      IHT_8x16[tx_type].rows(input, outtmp);
759
    for (j = 0; j < n; ++j)
760
      tmp[j][i] = (tran_low_t)dct_const_round_shift(outtmp[j] * Sqrt2);
clang-format's avatar
clang-format committed
761
    input += n;
762 763 764 765
  }

  // inverse transform column vectors
  for (i = 0; i < n; ++i) {
766
    IHT_8x16[tx_type].cols(tmp[i], out[i]);
767 768
  }

769
#if CONFIG_EXT_TX
770
  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n2, n);
771
#endif
772 773 774 775 776 777 778 779 780 781 782

  // Sum with the destination
  for (i = 0; i < n2; ++i) {
    for (j = 0; j < n; ++j) {
      int d = i * stride + j;
      int s = j * outstride + i;
      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6));
    }
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
783
void av1_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int stride,
784 785
                           const TxfmParam *txfm_param) {
  int tx_type = txfm_param->tx_type;
Sarah Parker's avatar
Sarah Parker committed
786 787 788
#if CONFIG_MRC_TX
  assert(tx_type != MRC_DCT && "Invalid tx type for tx size");
#endif  // CONFIG_MRC_TX
789 790 791
#if CONFIG_DCT_ONLY
  assert(tx_type == DCT_DCT);
#endif
792
  static const transform_2d IHT_16x8[] = {
Luca Barbato's avatar
Luca Barbato committed
793 794 795 796
    { aom_idct8_c, aom_idct16_c },    // DCT_DCT
    { aom_iadst8_c, aom_idct16_c },   // ADST_DCT
    { aom_idct8_c, aom_iadst16_c },   // DCT_ADST
    { aom_iadst8_c, aom_iadst16_c },  // ADST_ADST
797
#if CONFIG_EXT_TX
Luca Barbato's avatar
Luca Barbato committed
798 799 800 801 802 803 804 805 806 807 808 809
    { aom_iadst8_c, aom_idct16_c },   // FLIPADST_DCT
    { aom_idct8_c, aom_iadst16_c },   // DCT_FLIPADST
    { aom_iadst8_c, aom_iadst16_c },  // FLIPADST_FLIPADST
    { aom_iadst8_c, aom_iadst16_c },  // ADST_FLIPADST
    { aom_iadst8_c, aom_iadst16_c },  // FLIPADST_ADST
    { iidtx8_c, iidtx16_c },          // IDTX
    { aom_idct8_c, iidtx16_c },       // V_DCT
    { iidtx8_c, aom_idct16_c },       // H_DCT
    { aom_iadst8_c, iidtx16_c },      // V_ADST
    { iidtx8_c, aom_iadst16_c },      // H_ADST
    { aom_iadst8_c, iidtx16_c },      // V_FLIPADST
    { iidtx8_c, aom_iadst16_c },      // H_FLIPADST
810
#endif
811
  };
812

813 814 815 816
  const int n = 8;
  const int n2 = 16;

  int i, j;
817
  tran_low_t out[16][8], tmp[16][8], outtmp[16];
818 819 820
  tran_low_t *outp = &out[0][0];
  int outstride = n;

Lester Lu's avatar
Lester Lu committed
821
#if CONFIG_LGT
Lester Lu's avatar
Lester Lu committed
822 823
  const tran_high_t *lgtmtx_col[1];
  int use_lgt_col = get_lgt8(txfm_param, 1, lgtmtx_col);
Lester Lu's avatar
Lester Lu committed
824 825
#endif

826 827 828 829
  // inverse transform row vectors and transpose
  for (i = 0; i < n; ++i) {
    IHT_16x8[tx_type].rows(input, outtmp);
    for (j = 0; j < n2; ++j)
830
      tmp[j][i] = (tran_low_t)dct_const_round_shift(outtmp[j] * Sqrt2);
clang-format's avatar
clang-format committed
831
    input += n2;
832 833 834 835
  }

  // inverse transform column vectors
  for (i = 0; i < n2; ++i) {
Lester Lu's avatar
Lester Lu committed
836 837
#if CONFIG_LGT
    if (use_lgt_col)
Lester Lu's avatar
Lester Lu committed
838
      ilgt8(tmp[i], out[i], lgtmtx_col[0]);
Lester Lu's avatar
Lester Lu committed
839 840 841
    else
#endif
      IHT_16x8[tx_type].cols(tmp[i], out[i]);
842 843
  }

844
#if CONFIG_EXT_TX
845
  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n, n2);
846
#endif
847 848 849 850 851 852 853 854 855 856 857

  // Sum with the destination
  for (i = 0; i < n; ++i) {
    for (j = 0; j < n2; ++j) {
      int d = i * stride + j;
      int s = j * outstride + i;
      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6));
    }
  }
}

858
void av1_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int stride,
859 860
                           const TxfmParam *txfm_param) {
  int tx_type = txfm_param->tx_type;
Sarah Parker's avatar
Sarah Parker committed
861 862 863
#if CONFIG_MRC_TX
  assert(tx_type != MRC_DCT && "Invalid tx type for tx size");
#endif  // CONFIG_MRC_TX
864 865 866
#if CONFIG_DCT_ONLY
  assert(tx_type == DCT_DCT);
#endif
867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890
  static const transform_2d IHT_8x32[] = {
    { aom_idct32_c, aom_idct8_c },     // DCT_DCT
    { ihalfright32_c, aom_idct8_c },   // ADST_DCT
    { aom_idct32_c, aom_iadst8_c },    // DCT_ADST
    { ihalfright32_c, aom_iadst8_c },  // ADST_ADST
#if CONFIG_EXT_TX
    { ihalfright32_c, aom_idct8_c },   // FLIPADST_DCT
    { aom_idct32_c, aom_iadst8_c },    // DCT_FLIPADST
    { ihalfright32_c, aom_iadst8_c },  // FLIPADST_FLIPADST
    { ihalfright32_c, aom_iadst8_c },  // ADST_FLIPADST
    { ihalfright32_c, aom_iadst8_c },  // FLIPADST_ADST
    { iidtx32_c, iidtx8_c },           // IDTX
    { aom_idct32_c, iidtx8_c },        // V_DCT
    { iidtx32_c, aom_idct8_c },        // H_DCT
    { ihalfright32_c, iidtx8_c },      // V_ADST
    { iidtx32_c, aom_iadst8_c },       // H_ADST
    { ihalfright32_c, iidtx8_c },      // V_FLIPADST
    { iidtx32_c, aom_iadst8_c },       // H_FLIPADST
#endif
  };

  const int n = 8;
  const int n4 = 32;
  int i, j;
891
  tran_low_t out[8][32], tmp[8][32], outtmp[8];
892 893 894
  tran_low_t *outp = &out[0][0];
  int outstride = n4;

Lester Lu's avatar
Lester Lu committed
895
#if CONFIG_LGT
Lester Lu's avatar
Lester Lu committed
896 897
  const tran_high_t *lgtmtx_row[1];
  int use_lgt_row = get_lgt8(txfm_param, 0, lgtmtx_row);
Lester Lu's avatar
Lester Lu committed
898 899
#endif

900 901
  // inverse transform row vectors and transpose
  for (i = 0; i < n4; ++i) {
Lester Lu's avatar
Lester Lu committed
902 903
#if CONFIG_LGT
    if (use_lgt_row)
Lester Lu's avatar
Lester Lu committed
904
      ilgt8(input, outtmp, lgtmtx_row[0]);
Lester Lu's avatar
Lester Lu committed
905 906 907
    else
#endif
      IHT_8x32[tx_type].rows(input, outtmp);
908
    for (j = 0; j < n; ++j) tmp[j][i] = outtmp[j];