av1_inv_txfm2d.c 13.4 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
 */

Yaowu Xu's avatar
Yaowu Xu committed
12
#include "./av1_rtcd.h"
13
#include "aom_dsp/inv_txfm.h"
14
#include "av1/common/enums.h"
Yaowu Xu's avatar
Yaowu Xu committed
15 16
#include "av1/common/av1_txfm.h"
#include "av1/common/av1_inv_txfm1d.h"
17
#include "av1/common/av1_inv_txfm1d_cfg.h"
Angie Chiang's avatar
Angie Chiang committed
18

Yaowu Xu's avatar
Yaowu Xu committed
19
static INLINE TxfmFunc inv_txfm_type_to_func(TXFM_TYPE txfm_type) {
20
  switch (txfm_type) {
Yaowu Xu's avatar
Yaowu Xu committed
21 22 23 24
    case TXFM_TYPE_DCT4: return av1_idct4_new;
    case TXFM_TYPE_DCT8: return av1_idct8_new;
    case TXFM_TYPE_DCT16: return av1_idct16_new;
    case TXFM_TYPE_DCT32: return av1_idct32_new;
25 26 27
#if CONFIG_TX64X64
    case TXFM_TYPE_DCT64: return av1_idct64_new;
#endif  // CONFIG_TX64X64
Yaowu Xu's avatar
Yaowu Xu committed
28 29 30 31
    case TXFM_TYPE_ADST4: return av1_iadst4_new;
    case TXFM_TYPE_ADST8: return av1_iadst8_new;
    case TXFM_TYPE_ADST16: return av1_iadst16_new;
    case TXFM_TYPE_ADST32: return av1_iadst32_new;
32 33 34 35 36 37
#if CONFIG_EXT_TX
    case TXFM_TYPE_IDENTITY4: return av1_iidentity4_c;
    case TXFM_TYPE_IDENTITY8: return av1_iidentity8_c;
    case TXFM_TYPE_IDENTITY16: return av1_iidentity16_c;
    case TXFM_TYPE_IDENTITY32: return av1_iidentity32_c;
#endif  // CONFIG_EXT_TX
clang-format's avatar
clang-format committed
38
    default: assert(0); return NULL;
39 40 41
  }
}

42 43
static const TXFM_1D_CFG *inv_txfm_col_cfg_ls[TX_TYPES_1D][TX_SIZES] = {
  // DCT
44
  {
45
#if CONFIG_CHROMA_2X2
46 47
      NULL,
#endif
48 49 50
      &inv_txfm_1d_col_cfg_dct_4, &inv_txfm_1d_col_cfg_dct_8,
      &inv_txfm_1d_col_cfg_dct_16, &inv_txfm_1d_col_cfg_dct_32 },
  // ADST
51
  {
52
#if CONFIG_CHROMA_2X2
53 54
      NULL,
#endif
55 56 57 58
      &inv_txfm_1d_col_cfg_adst_4, &inv_txfm_1d_col_cfg_adst_8,
      &inv_txfm_1d_col_cfg_adst_16, &inv_txfm_1d_col_cfg_adst_32 },
#if CONFIG_EXT_TX
  // FLIPADST
59
  {
60
#if CONFIG_CHROMA_2X2
61 62
      NULL,
#endif
63 64
      &inv_txfm_1d_col_cfg_adst_4, &inv_txfm_1d_col_cfg_adst_8,
      &inv_txfm_1d_col_cfg_adst_16, &inv_txfm_1d_col_cfg_adst_32 },
65
  // IDENTITY
66
  {
67
#if CONFIG_CHROMA_2X2
68 69
      NULL,
#endif
70 71
      &inv_txfm_1d_cfg_identity_4, &inv_txfm_1d_cfg_identity_8,
      &inv_txfm_1d_cfg_identity_16, &inv_txfm_1d_cfg_identity_32 },
72 73 74 75 76
#endif  // CONFIG_EXT_TX
};

static const TXFM_1D_CFG *inv_txfm_row_cfg_ls[TX_TYPES_1D][TX_SIZES] = {
  // DCT
77
  {
78
#if CONFIG_CHROMA_2X2
79 80
      NULL,
#endif
81 82 83
      &inv_txfm_1d_row_cfg_dct_4, &inv_txfm_1d_row_cfg_dct_8,
      &inv_txfm_1d_row_cfg_dct_16, &inv_txfm_1d_row_cfg_dct_32 },
  // ADST
84
  {
85
#if CONFIG_CHROMA_2X2
86 87
      NULL,
#endif
88 89 90 91
      &inv_txfm_1d_row_cfg_adst_4, &inv_txfm_1d_row_cfg_adst_8,
      &inv_txfm_1d_row_cfg_adst_16, &inv_txfm_1d_row_cfg_adst_32 },
#if CONFIG_EXT_TX
  // FLIPADST
92
  {
93
#if CONFIG_CHROMA_2X2
94 95
      NULL,
#endif
96 97
      &inv_txfm_1d_row_cfg_adst_4, &inv_txfm_1d_row_cfg_adst_8,
      &inv_txfm_1d_row_cfg_adst_16, &inv_txfm_1d_row_cfg_adst_32 },
98
  // IDENTITY
99
  {
100
#if CONFIG_CHROMA_2X2
101 102
      NULL,
#endif
103 104
      &inv_txfm_1d_cfg_identity_4, &inv_txfm_1d_cfg_identity_8,
      &inv_txfm_1d_cfg_identity_16, &inv_txfm_1d_cfg_identity_32 },
105
#endif  // CONFIG_EXT_TX
106 107
};

Yaowu Xu's avatar
Yaowu Xu committed
108
TXFM_2D_FLIP_CFG av1_get_inv_txfm_cfg(int tx_type, int tx_size) {
109 110
  TXFM_2D_FLIP_CFG cfg;
  set_flip_cfg(tx_type, &cfg);
Sarah Parker's avatar
Sarah Parker committed
111 112 113 114
  const int tx_type_col = vtx_tab[tx_type];
  const int tx_type_row = htx_tab[tx_type];
  const int tx_size_col = txsize_vert_map[tx_size];
  const int tx_size_row = txsize_horz_map[tx_size];
115 116
  cfg.col_cfg = inv_txfm_col_cfg_ls[tx_type_col][tx_size_col];
  cfg.row_cfg = inv_txfm_row_cfg_ls[tx_type_row][tx_size_row];
117 118 119
  return cfg;
}

Yaowu Xu's avatar
Yaowu Xu committed
120
TXFM_2D_FLIP_CFG av1_get_inv_txfm_64x64_cfg(int tx_type) {
121
  TXFM_2D_FLIP_CFG cfg = { 0, 0, NULL, NULL };
122 123
  switch (tx_type) {
    case DCT_DCT:
124 125
      cfg.col_cfg = &inv_txfm_1d_col_cfg_dct_64;
      cfg.row_cfg = &inv_txfm_1d_row_cfg_dct_64;
126
      set_flip_cfg(tx_type, &cfg);
127
      break;
clang-format's avatar
clang-format committed
128
    default: assert(0);
129 130 131 132
  }
  return cfg;
}

133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157
void av1_gen_inv_stage_range(int8_t *stage_range_col, int8_t *stage_range_row,
                             const TXFM_2D_FLIP_CFG *cfg, int8_t fwd_shift,
                             int bd) {
  // Note when assigning txfm_size_col, we use the txfm_size from the
  // row configuration and vice versa. This is intentionally done to
  // accurately perform rectangular transforms. When the transform is
  // rectangular, the number of columns will be the same as the
  // txfm_size stored in the row cfg struct. It will make no difference
  // for square transforms.
  const int txfm_size_col = cfg->row_cfg->txfm_size;
  const int txfm_size_row = cfg->col_cfg->txfm_size;
  // Take the shift from the larger dimension in the rectangular case.
  const int8_t *shift = (txfm_size_col > txfm_size_row) ? cfg->row_cfg->shift
                                                        : cfg->col_cfg->shift;
  // i < MAX_TXFM_STAGE_NUM will mute above array bounds warning
  for (int i = 0; i < cfg->row_cfg->stage_num && i < MAX_TXFM_STAGE_NUM; ++i) {
    stage_range_row[i] = cfg->row_cfg->stage_range[i] + fwd_shift + bd + 1;
  }
  // i < MAX_TXFM_STAGE_NUM will mute above array bounds warning
  for (int i = 0; i < cfg->col_cfg->stage_num && i < MAX_TXFM_STAGE_NUM; ++i) {
    stage_range_col[i] =
        cfg->col_cfg->stage_range[i] + fwd_shift + shift[0] + bd + 1;
  }
}

158
static INLINE void inv_txfm2d_add_c(const int32_t *input, uint16_t *output,
159
                                    int stride, TXFM_2D_FLIP_CFG *cfg,
160 161
                                    int32_t *txfm_buf, int8_t fwd_shift,
                                    int bd) {
162 163 164 165 166 167 168 169 170
  // Note when assigning txfm_size_col, we use the txfm_size from the
  // row configuration and vice versa. This is intentionally done to
  // accurately perform rectangular transforms. When the transform is
  // rectangular, the number of columns will be the same as the
  // txfm_size stored in the row cfg struct. It will make no difference
  // for square transforms.
  const int txfm_size_col = cfg->row_cfg->txfm_size;
  const int txfm_size_row = cfg->col_cfg->txfm_size;
  // Take the shift from the larger dimension in the rectangular case.
Sarah Parker's avatar
Sarah Parker committed
171 172
  const int8_t *shift = (txfm_size_col > txfm_size_row) ? cfg->row_cfg->shift
                                                        : cfg->col_cfg->shift;
173 174 175 176 177 178
  int8_t stage_range_row[MAX_TXFM_STAGE_NUM];
  int8_t stage_range_col[MAX_TXFM_STAGE_NUM];
  assert(cfg->row_cfg->stage_num <= MAX_TXFM_STAGE_NUM);
  assert(cfg->col_cfg->stage_num <= MAX_TXFM_STAGE_NUM);
  av1_gen_inv_stage_range(stage_range_col, stage_range_row, cfg, fwd_shift, bd);

179 180 181 182
  const int8_t *cos_bit_col = cfg->col_cfg->cos_bit;
  const int8_t *cos_bit_row = cfg->row_cfg->cos_bit;
  const TxfmFunc txfm_func_col = inv_txfm_type_to_func(cfg->col_cfg->txfm_type);
  const TxfmFunc txfm_func_row = inv_txfm_type_to_func(cfg->row_cfg->txfm_type);
Angie Chiang's avatar
Angie Chiang committed
183

184
  // txfm_buf's length is  txfm_size_row * txfm_size_col + 2 * txfm_size_row
Angie Chiang's avatar
Angie Chiang committed
185 186
  // it is used for intermediate data buffering
  int32_t *temp_in = txfm_buf;
187 188
  int32_t *temp_out = temp_in + txfm_size_row;
  int32_t *buf = temp_out + txfm_size_row;
Angie Chiang's avatar
Angie Chiang committed
189
  int32_t *buf_ptr = buf;
190
  int c, r;
Angie Chiang's avatar
Angie Chiang committed
191 192

  // Rows
193
  for (r = 0; r < txfm_size_row; ++r) {
Angie Chiang's avatar
Angie Chiang committed
194
    txfm_func_row(input, buf_ptr, cos_bit_row, stage_range_row);
195 196 197 198 199 200 201 202
    round_shift_array(buf_ptr, txfm_size_col, -shift[0]);
    // Multiply everything by Sqrt2 if the transform is rectangular
    if (txfm_size_row != txfm_size_col) {
      for (c = 0; c < txfm_size_col; ++c)
        buf_ptr[c] = (int32_t)dct_const_round_shift(buf_ptr[c] * Sqrt2);
    }
    input += txfm_size_col;
    buf_ptr += txfm_size_col;
Angie Chiang's avatar
Angie Chiang committed
203 204 205
  }

  // Columns
206
  for (c = 0; c < txfm_size_col; ++c) {
207
    if (cfg->lr_flip == 0) {
208 209
      for (r = 0; r < txfm_size_row; ++r)
        temp_in[r] = buf[r * txfm_size_col + c];
210 211
    } else {
      // flip left right
212 213
      for (r = 0; r < txfm_size_row; ++r)
        temp_in[r] = buf[r * txfm_size_col + (txfm_size_col - c - 1)];
214
    }
Angie Chiang's avatar
Angie Chiang committed
215
    txfm_func_col(temp_in, temp_out, cos_bit_col, stage_range_col);
216
    round_shift_array(temp_out, txfm_size_row, -shift[1]);
217
    if (cfg->ud_flip == 0) {
218 219 220 221
      for (r = 0; r < txfm_size_row; ++r) {
        output[r * stride + c] =
            highbd_clip_pixel_add(output[r * stride + c], temp_out[r], bd);
      }
222 223
    } else {
      // flip upside down
224 225 226 227
      for (r = 0; r < txfm_size_row; ++r) {
        output[r * stride + c] = highbd_clip_pixel_add(
            output[r * stride + c], temp_out[txfm_size_row - r - 1], bd);
      }
228
    }
Angie Chiang's avatar
Angie Chiang committed
229 230 231
  }
}

232 233 234 235
static INLINE void inv_txfm2d_add_facade(const int32_t *input, uint16_t *output,
                                         int stride, int32_t *txfm_buf,
                                         int tx_type, int tx_size, int bd) {
  TXFM_2D_FLIP_CFG cfg = av1_get_inv_txfm_cfg(tx_type, tx_size);
236 237 238
  int tx_size_sqr = txsize_sqr_map[tx_size];
  inv_txfm2d_add_c(input, output, stride, &cfg, txfm_buf,
                   fwd_shift_sum[tx_size_sqr], bd);
239 240 241 242 243 244 245 246 247 248
}

void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output,
                              int stride, int tx_type, int bd) {
  int txfm_buf[4 * 8 + 8 + 8];
  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_4X8, bd);
}

void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output,
                              int stride, int tx_type, int bd) {
249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
#if CONFIG_TXMG
  int txfm_buf[8 * 4 + 8 + 8];
  int32_t rinput[8 * 4];
  uint16_t routput[8 * 4];
  int tx_size = TX_8X4;
  int rtx_size = av1_rotate_tx_size(tx_size);
  int rtx_type = av1_rotate_tx_type(tx_type);
  int w = tx_size_wide[tx_size];
  int h = tx_size_high[tx_size];
  int rw = h;
  int rh = w;
  transpose_int32(rinput, rw, input, w, w, h);
  transpose_uint16(routput, rw, output, stride, w, h);
  inv_txfm2d_add_facade(rinput, routput, rw, txfm_buf, rtx_type, rtx_size, bd);
  transpose_uint16(output, stride, routput, rw, rw, rh);
#else
265 266
  int txfm_buf[8 * 4 + 4 + 4];
  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_8X4, bd);
267
#endif
268 269 270 271 272 273 274 275 276 277
}

void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output,
                               int stride, int tx_type, int bd) {
  int txfm_buf[8 * 16 + 16 + 16];
  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_8X16, bd);
}

void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output,
                               int stride, int tx_type, int bd) {
278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293
#if CONFIG_TXMG
  int txfm_buf[16 * 8 + 16 + 16];
  int32_t rinput[16 * 8];
  uint16_t routput[16 * 8];
  int tx_size = TX_16X8;
  int rtx_size = av1_rotate_tx_size(tx_size);
  int rtx_type = av1_rotate_tx_type(tx_type);
  int w = tx_size_wide[tx_size];
  int h = tx_size_high[tx_size];
  int rw = h;
  int rh = w;
  transpose_int32(rinput, rw, input, w, w, h);
  transpose_uint16(routput, rw, output, stride, w, h);
  inv_txfm2d_add_facade(rinput, routput, rw, txfm_buf, rtx_type, rtx_size, bd);
  transpose_uint16(output, stride, routput, rw, rw, rh);
#else
294 295
  int txfm_buf[16 * 8 + 8 + 8];
  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_16X8, bd);
296
#endif
297 298 299 300 301 302 303 304 305 306
}

void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output,
                                int stride, int tx_type, int bd) {
  int txfm_buf[16 * 32 + 32 + 32];
  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_16X32, bd);
}

void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output,
                                int stride, int tx_type, int bd) {
307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322
#if CONFIG_TXMG
  int txfm_buf[32 * 16 + 32 + 32];
  int32_t rinput[32 * 16];
  uint16_t routput[32 * 16];
  int tx_size = TX_32X16;
  int rtx_size = av1_rotate_tx_size(tx_size);
  int rtx_type = av1_rotate_tx_type(tx_type);
  int w = tx_size_wide[tx_size];
  int h = tx_size_high[tx_size];
  int rw = h;
  int rh = w;
  transpose_int32(rinput, rw, input, w, w, h);
  transpose_uint16(routput, rw, output, stride, w, h);
  inv_txfm2d_add_facade(rinput, routput, rw, txfm_buf, rtx_type, rtx_size, bd);
  transpose_uint16(output, stride, routput, rw, rw, rh);
#else
323 324
  int txfm_buf[32 * 16 + 16 + 16];
  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_32X16, bd);
325
#endif
326 327 328 329 330 331
}

void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output,
                              int stride, int tx_type, int bd) {
  int txfm_buf[4 * 4 + 4 + 4];
  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_4X4, bd);
Angie Chiang's avatar
Angie Chiang committed
332 333
}

Yaowu Xu's avatar
Yaowu Xu committed
334 335
void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output,
                              int stride, int tx_type, int bd) {
Angie Chiang's avatar
Angie Chiang committed
336
  int txfm_buf[8 * 8 + 8 + 8];
337
  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_8X8, bd);
Angie Chiang's avatar
Angie Chiang committed
338 339
}

Yaowu Xu's avatar
Yaowu Xu committed
340 341
void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output,
                                int stride, int tx_type, int bd) {
Angie Chiang's avatar
Angie Chiang committed
342
  int txfm_buf[16 * 16 + 16 + 16];
343
  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_16X16, bd);
Angie Chiang's avatar
Angie Chiang committed
344 345
}

Yaowu Xu's avatar
Yaowu Xu committed
346 347
void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output,
                                int stride, int tx_type, int bd) {
Angie Chiang's avatar
Angie Chiang committed
348
  int txfm_buf[32 * 32 + 32 + 32];
349
  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_32X32, bd);
Angie Chiang's avatar
Angie Chiang committed
350
}
Angie Chiang's avatar
Angie Chiang committed
351

Yaowu Xu's avatar
Yaowu Xu committed
352 353
void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output,
                                int stride, int tx_type, int bd) {
Angie Chiang's avatar
Angie Chiang committed
354
  int txfm_buf[64 * 64 + 64 + 64];
Yaowu Xu's avatar
Yaowu Xu committed
355
  TXFM_2D_FLIP_CFG cfg = av1_get_inv_txfm_64x64_cfg(tx_type);
356 357 358 359
  inv_txfm2d_add_c(input, output, stride, &cfg, txfm_buf, -4, bd);
#if CONFIG_TX64X64
  assert(fwd_shift_sum[TX_64X64] == -4);
#endif
Angie Chiang's avatar
Angie Chiang committed
360
}