av1/common/idct.c

/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * 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.
 */

#include <math.h>

#include "config/aom_dsp_rtcd.h"
#include "config/av1_rtcd.h"

#include "aom_ports/mem.h"
#include "av1/common/av1_inv_txfm1d_cfg.h"
#include "av1/common/av1_txfm.h"
#include "av1/common/blockd.h"
#include "av1/common/enums.h"
#include "av1/common/idct.h"
#if CONFIG_DSPL_RESIDUAL
#include "av1/common/resize.h"
#include "av1/common/scan.h"
#endif  // CONFIG_DSPL_RESIDUAL
#if CONFIG_NN_RECON
#include "av1/common/cnn_tflite.h"
#include "av1/common/nn_recon.h"
#endif  // CONFIG_NN_RECON

int av1_get_tx_scale(const TX_SIZE tx_size) {
  const int pels = tx_size_2d[tx_size];
  // Largest possible pels is 4096 (64x64).
  return (pels > 256) + (pels > 1024);
}

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

// idct
void av1_highbd_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
                            int eob, int bd) {
  if (eob > 1)
    av1_highbd_iwht4x4_16_add(input, dest, stride, bd);
  else
    av1_highbd_iwht4x4_1_add(input, dest, stride, bd);
}

void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *input, uint8_t *dest,
                                   int stride, const TxfmParam *txfm_param) {
  assert(av1_ext_tx_used[txfm_param->tx_set_type][txfm_param->tx_type]);
  int eob = txfm_param->eob;
  int bd = txfm_param->bd;
  int lossless = txfm_param->lossless;
  const int32_t *src = cast_to_int32(input);
  const TX_TYPE tx_type = txfm_param->tx_type;
  if (lossless) {
    assert(tx_type == DCT_DCT);
    av1_highbd_iwht4x4_add(input, dest, stride, eob, bd);
    return;
  }

  av1_inv_txfm2d_add_4x4_c(src, CONVERT_TO_SHORTPTR(dest), stride, tx_type,
                           txfm_param->mode, bd);
}

void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest,
                                   int stride, const TxfmParam *txfm_param) {
  assert(av1_ext_tx_used[txfm_param->tx_set_type][txfm_param->tx_type]);
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_4x8_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                           txfm_param->tx_type, txfm_param->mode,
                           txfm_param->bd);
}

void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest,
                                   int stride, const TxfmParam *txfm_param) {
  assert(av1_ext_tx_used[txfm_param->tx_set_type][txfm_param->tx_type]);
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_8x4_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                           txfm_param->tx_type, txfm_param->mode,
                           txfm_param->bd);
}

void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest,
                                     int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_16x32_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                             txfm_param->tx_type, txfm_param->mode,
                             txfm_param->bd);
}

void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest,
                                     int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_32x16_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                             txfm_param->tx_type, txfm_param->mode,
                             txfm_param->bd);
}

void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_16x4_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}

void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_4x16_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}

void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_32x8_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}

void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_8x32_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}

void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest,
                                     int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_32x64_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                             txfm_param->tx_type, txfm_param->mode,
                             txfm_param->bd);
}

void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest,
                                     int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_64x32_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                             txfm_param->tx_type, txfm_param->mode,
                             txfm_param->bd);
}

void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest,
                                     int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_16x64_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                             txfm_param->tx_type, txfm_param->mode,
                             txfm_param->bd);
}

void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest,
                                     int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_64x16_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                             txfm_param->tx_type, txfm_param->mode,
                             txfm_param->bd);
}

void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *input, uint8_t *dest,
                                   int stride, const TxfmParam *txfm_param) {
  int bd = txfm_param->bd;
  const TX_TYPE tx_type = txfm_param->tx_type;
  const int32_t *src = cast_to_int32(input);

  av1_inv_txfm2d_add_8x8_c(src, CONVERT_TO_SHORTPTR(dest), stride, tx_type,
                           txfm_param->mode, bd);
}

void av1_highbd_inv_txfm_add_16x16_c(const tran_low_t *input, uint8_t *dest,
                                     int stride, const TxfmParam *txfm_param) {
  int bd = txfm_param->bd;
  const TX_TYPE tx_type = txfm_param->tx_type;
  const int32_t *src = cast_to_int32(input);

  av1_inv_txfm2d_add_16x16_c(src, CONVERT_TO_SHORTPTR(dest), stride, tx_type,
                             txfm_param->mode, bd);
}

void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_8x16_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}

void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_16x8_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}

void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest,
                                     int stride, const TxfmParam *txfm_param) {
  const int bd = txfm_param->bd;
  const TX_TYPE tx_type = txfm_param->tx_type;
  const int32_t *src = cast_to_int32(input);

  av1_inv_txfm2d_add_32x32_c(src, CONVERT_TO_SHORTPTR(dest), stride, tx_type,
                             txfm_param->mode, bd);
}

void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest,
                                     int stride, const TxfmParam *txfm_param) {
  const int bd = txfm_param->bd;
  const TX_TYPE tx_type = txfm_param->tx_type;
  const int32_t *src = cast_to_int32(input);
  assert(tx_type == DCT_DCT);
  av1_inv_txfm2d_add_64x64_c(src, CONVERT_TO_SHORTPTR(dest), stride, tx_type,
                             txfm_param->mode, bd);
}

#if CONFIG_FLEX_PARTITION
void av1_highbd_inv_txfm_add_4x32_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_4x32_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}

void av1_highbd_inv_txfm_add_32x4_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_32x4_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}

void av1_highbd_inv_txfm_add_8x64_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_8x64_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}

void av1_highbd_inv_txfm_add_64x8_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_64x8_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}

void av1_highbd_inv_txfm_add_4x64_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_4x64_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}

void av1_highbd_inv_txfm_add_64x4_c(const tran_low_t *input, uint8_t *dest,
                                    int stride, const TxfmParam *txfm_param) {
  const int32_t *src = cast_to_int32(input);
  av1_inv_txfm2d_add_64x4_c(src, CONVERT_TO_SHORTPTR(dest), stride,
                            txfm_param->tx_type, txfm_param->mode,
                            txfm_param->bd);
}
#endif  // CONFIG_FLEX_PARTITION

static void init_txfm_param(const MACROBLOCKD *xd, int plane, TX_SIZE tx_size,
                            TX_TYPE tx_type, int eob, int reduced_tx_set,
                            TxfmParam *txfm_param) {
  (void)plane;
  txfm_param->tx_type = tx_type;
  txfm_param->tx_size = tx_size;
  txfm_param->eob = eob;
  txfm_param->lossless = xd->lossless[xd->mi[0]->segment_id];
  txfm_param->bd = xd->bd;
  txfm_param->is_hbd = is_cur_buf_hbd(xd);
  txfm_param->tx_set_type = av1_get_ext_tx_set_type(
      txfm_param->tx_size, is_inter_block(xd->mi[0]), reduced_tx_set);
  txfm_param->mode = get_mode_dep_txfm_mode(xd->mi[0]);
}

void av1_highbd_inv_txfm_add_c(const tran_low_t *input, uint8_t *dest,
                               int stride, const TxfmParam *txfm_param) {
  assert(av1_ext_tx_used[txfm_param->tx_set_type][txfm_param->tx_type]);
  const TX_SIZE tx_size = txfm_param->tx_size;
  switch (tx_size) {
    case TX_32X32:
      av1_highbd_inv_txfm_add_32x32_c(input, dest, stride, txfm_param);
      break;
    case TX_16X16:
      av1_highbd_inv_txfm_add_16x16_c(input, dest, stride, txfm_param);
      break;
    case TX_8X8:
      av1_highbd_inv_txfm_add_8x8_c(input, dest, stride, txfm_param);
      break;
    case TX_4X8:
      av1_highbd_inv_txfm_add_4x8_c(input, dest, stride, txfm_param);
      break;
    case TX_8X4:
      av1_highbd_inv_txfm_add_8x4_c(input, dest, stride, txfm_param);
      break;
    case TX_8X16:
      av1_highbd_inv_txfm_add_8x16_c(input, dest, stride, txfm_param);
      break;
    case TX_16X8:
      av1_highbd_inv_txfm_add_16x8_c(input, dest, stride, txfm_param);
      break;
    case TX_16X32:
      av1_highbd_inv_txfm_add_16x32_c(input, dest, stride, txfm_param);
      break;
    case TX_32X16:
      av1_highbd_inv_txfm_add_32x16_c(input, dest, stride, txfm_param);
      break;
    case TX_64X64:
      av1_highbd_inv_txfm_add_64x64_c(input, dest, stride, txfm_param);
      break;
    case TX_32X64:
      av1_highbd_inv_txfm_add_32x64_c(input, dest, stride, txfm_param);
      break;
    case TX_64X32:
      av1_highbd_inv_txfm_add_64x32_c(input, dest, stride, txfm_param);
      break;
    case TX_16X64:
      av1_highbd_inv_txfm_add_16x64_c(input, dest, stride, txfm_param);
      break;
    case TX_64X16:
      av1_highbd_inv_txfm_add_64x16_c(input, dest, stride, txfm_param);
      break;
    case TX_4X4:
      // this is like av1_short_idct4x4 but has a special case around eob<=1
      // which is significant (not just an optimization) for the lossless
      // case.
      av1_highbd_inv_txfm_add_4x4_c(input, dest, stride, txfm_param);
      break;
    case TX_16X4:
      av1_highbd_inv_txfm_add_16x4_c(input, dest, stride, txfm_param);
      break;
    case TX_4X16:
      av1_highbd_inv_txfm_add_4x16_c(input, dest, stride, txfm_param);
      break;
    case TX_8X32:
      av1_highbd_inv_txfm_add_8x32_c(input, dest, stride, txfm_param);
      break;
    case TX_32X8:
      av1_highbd_inv_txfm_add_32x8_c(input, dest, stride, txfm_param);
      break;
#if CONFIG_FLEX_PARTITION
    case TX_4X32:
      av1_highbd_inv_txfm_add_4x32_c(input, dest, stride, txfm_param);
      break;
    case TX_32X4:
      av1_highbd_inv_txfm_add_32x4_c(input, dest, stride, txfm_param);
      break;
    case TX_8X64:
      av1_highbd_inv_txfm_add_8x64_c(input, dest, stride, txfm_param);
      break;
    case TX_64X8:
      av1_highbd_inv_txfm_add_64x8_c(input, dest, stride, txfm_param);
      break;
    case TX_4X64:
      av1_highbd_inv_txfm_add_4x64_c(input, dest, stride, txfm_param);
      break;
    case TX_64X4:
      av1_highbd_inv_txfm_add_64x4_c(input, dest, stride, txfm_param);
      break;
#endif  // CONFIG_FLEX_PARTITION
    default: assert(0 && "Invalid transform size"); break;
  }
}

void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride,
                        const TxfmParam *txfm_param) {
  const TX_SIZE tx_size = txfm_param->tx_size;
  DECLARE_ALIGNED(32, uint16_t, tmp[MAX_TX_SQUARE]);
  int tmp_stride = MAX_TX_SIZE;
  int w = tx_size_wide[tx_size];
  int h = tx_size_high[tx_size];
  for (int r = 0; r < h; ++r) {
    for (int c = 0; c < w; ++c) {
      tmp[r * tmp_stride + c] = dst[r * stride + c];
    }
  }

#if CONFIG_MODE_DEP_INTRA_TX || CONFIG_MODE_DEP_INTER_TX || CONFIG_LGT
  av1_highbd_inv_txfm_add_c(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                            txfm_param);
#elif CONFIG_DST_32X32 && CONFIG_SUPERRES_TX64 && CONFIG_DST7_16X16
  if (tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32 ||
      tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16 ||
      txsize_sqr_up_map[tx_size] == TX_64X64)
    av1_highbd_inv_txfm_add_c(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                              txfm_param);
  else
    av1_highbd_inv_txfm_add(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                            txfm_param);
#elif CONFIG_DST7_16X16 && CONFIG_SUPERRES_TX64
  if (tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16 ||
      txsize_sqr_up_map[tx_size] == TX_64X64)
    av1_highbd_inv_txfm_add_c(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                              txfm_param);
  else
    av1_highbd_inv_txfm_add(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                            txfm_param);
#elif CONFIG_DST_32X32 && CONFIG_SUPERRES_TX64
  if (tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32 ||
      txsize_sqr_up_map[tx_size] == TX_64X64)
    av1_highbd_inv_txfm_add_c(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                              txfm_param);
  else
    av1_highbd_inv_txfm_add(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                            txfm_param);
#elif CONFIG_DST_32X32 && CONFIG_DST7_16X16
  if (tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32 ||
      tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16)
    av1_highbd_inv_txfm_add_c(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                              txfm_param);
  else
    av1_highbd_inv_txfm_add(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                            txfm_param);
#elif CONFIG_DST7_16X16
  if (tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16)
    av1_highbd_inv_txfm_add_c(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                              txfm_param);
  else
    av1_highbd_inv_txfm_add(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                            txfm_param);
#elif CONFIG_DST_32X32
  if (tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32)
    av1_highbd_inv_txfm_add_c(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                              txfm_param);
  else
    av1_highbd_inv_txfm_add(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                            txfm_param);
#elif CONFIG_SUPERRES_TX64
  if (txsize_sqr_up_map[tx_size] == TX_64X64)
    av1_highbd_inv_txfm_add_c(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                              txfm_param);
  else
    av1_highbd_inv_txfm_add(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                            txfm_param);
#else
  av1_highbd_inv_txfm_add(dqcoeff, CONVERT_TO_BYTEPTR(tmp), tmp_stride,
                          txfm_param);
#endif  // CONFIG_MODE_DEP_INTRA_TX || CONFIG_MODE_DEP_INTER_TX || CONFIG_LGT

  for (int r = 0; r < h; ++r) {
    for (int c = 0; c < w; ++c) {
      dst[r * stride + c] = (uint8_t)tmp[r * tmp_stride + c];
    }
  }
}

#if CONFIG_DSPL_RESIDUAL
/*!
 * Given dequantized coefficients and transform parameters, this function
 * inverts the transform and saves result in a signed 16-bit integer buffer dst.
 *
 * This function is necessitated by the reduced output range of
 * av1_inverse_transform_block(), which produces unsigned output in [0-255],
 * [0-1023], etc depending on bitdepth. This is because after
 * av1_inverse_transform_block() inverts transform, it adds the residuals to the
 * prediction buffer to obtain the reconstruction. The reconstruction is always
 * in this bounded range.
 *
 * However the CONFIG_DSPL_RESIDUAL experiment requires access to the residuals
 * themselves, not the reconstruction. The residuals can range form
 * [-255, 255], [-1023, 1023], etc depending on bitdepth.
 *
 * To extract these residuals, this function forces a 12-bit depth transform
 * with prediction initialized to 1023. For 8 and 10 bit depths, 1023 can then
 * be subtracted from the output to compute residual.
 *
 * Currently this function only supports 8-bit depth (as that is the target
 * of CONFIG_DSPL_RESIDUAL at this moment). Extension to 10-bit depth
 * should be straightforward but has not been tested. However, extension to
 * 12-bit depth will require using inverse functions further down the call stack
 * rather than using av1_highbd_inv_txfm_add() since residuals will range in
 * [-4095, 4095].
 */
void av1_inverse_transform_block_diff(const MACROBLOCKD *xd,
                                      const tran_low_t *dqcoeff, int plane,
                                      TX_TYPE tx_type, TX_SIZE tx_size,
                                      int16_t *dst, int stride, int eob,
                                      int reduced_tx_set) {
  if (!eob) return;

  // This method has not been tested for 10-bit depth and most certainly won't
  // work with 12-bit depth
  assert(xd->bd == 8);

  // We initialize the residual buffer to 1023 and then subtract 1023 after the
  // inverse and add operation is performed by av1_highbd_inv_txfm_add() to get
  // residuals
  const uint16_t residual_init_val = 1023;
  const uint8_t txw = tx_size_wide[tx_size], txh = tx_size_high[tx_size];

  DECLARE_ALIGNED(32, uint16_t, residual16[MAX_SB_SQUARE]);
  uint8_t *const residual = CONVERT_TO_BYTEPTR(residual16);

  for (int r = 0; r < txh; ++r)
    for (int c = 0; c < txw; ++c)
      residual16[r * MAX_TX_SIZE + c] = residual_init_val;

  TxfmParam txfm_param;
  init_txfm_param(xd, plane, tx_size, tx_type, eob, reduced_tx_set,
                  &txfm_param);
  txfm_param.bd = 12;
  txfm_param.is_hbd = 1;

  av1_highbd_inv_txfm_add(dqcoeff, residual, MAX_TX_SIZE, &txfm_param);

  for (int r = 0; r < txh; ++r)
    for (int c = 0; c < txw; ++c)
      dst[r * stride + c] =
          (int16_t)residual16[r * MAX_TX_SIZE + c] - residual_init_val;
}

/*!
 * This function inverts transform blocks which have been coded using
 * downsampling in four steps: unpack coefficients, perform inverse transform,
 * upsample and add to the prediction buffer dst. For packing logic, please see
 * av1_xform_quant().
 */
void av1_inverse_dspl_transform_block(const MACROBLOCKD *xd,
                                      const tran_low_t *dqcoeff, int plane,
                                      TX_TYPE tx_type, TX_SIZE tx_size,
                                      uint8_t *dst, int stride, int eob,
                                      int reduced_tx_set) {
  const TX_SIZE new_tx_size = dspl_tx_size_map[tx_size];
  const uint8_t txw = tx_size_wide[tx_size], txh = tx_size_high[tx_size];
  const uint8_t dspl_txw = tx_size_wide[new_tx_size],
                dspl_txh = tx_size_high[new_tx_size];

  // Since we have to have enough buffer memory on the stack for MAX_TX_SIZE
  // size transforms, we will use MAX_TX_SIZE as stride for simplicity.
  const int buf_stride = MAX_TX_SIZE;
  assert(buf_stride >= dspl_txw);

  // Buffers
  DECLARE_ALIGNED(32, tran_low_t, scan_buf[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(32, tran_low_t, dqcoeff_buf[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(32, int16_t, buf_diff[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(32, int16_t, buf_up_diff[MAX_TX_SQUARE]);

  const int dspl_eob = eob;
  TxfmParam txfm_param;
  init_txfm_param(xd, plane, tx_size, tx_type, dspl_eob, reduced_tx_set,
                  &txfm_param);

  // Unpack coefficients
  const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
  const SCAN_ORDER *const dspl_scan_order = get_scan(new_tx_size, tx_type);
  const int size = av1_get_max_eob(tx_size);
  memset(scan_buf, 0, size * sizeof(tran_low_t));
  scan_array(dqcoeff, scan_buf, eob, scan_order);
  memset(dqcoeff_buf, 0, txw * txh * sizeof(tran_low_t));
  iscan_array(scan_buf, dqcoeff_buf, dspl_eob, dspl_scan_order);

  // Invert and compute difference
  av1_inverse_transform_block_diff(xd, dqcoeff_buf, plane, tx_type, new_tx_size,
                                   buf_diff, buf_stride, dspl_eob,
                                   reduced_tx_set);

  // Upsample
  av1_signed_up2(buf_diff, dspl_txh, dspl_txw, buf_stride, buf_up_diff,
                 buf_stride, 1, 1, txfm_param.bd);

  // Add to output
  for (int r = 0; r < txh; ++r) {
    for (int c = 0; c < txw; ++c) {
      const tran_low_t residue = (tran_low_t)buf_up_diff[r * buf_stride + c];
      const uint16_t out =
          highbd_clip_pixel_add(dst[r * stride + c], residue, txfm_param.bd);
      dst[r * stride + c] = (uint8_t)out;
    }
  }
}
#endif  // CONFIG_DSPL_RESIDUAL

void av1_inverse_transform_block(const MACROBLOCKD *xd,
                                 const tran_low_t *dqcoeff, int plane,
                                 TX_TYPE tx_type, TX_SIZE tx_size, uint8_t *dst,
                                 int stride, int eob, int reduced_tx_set) {
  if (!eob) return;

  assert(eob <= av1_get_max_eob(tx_size));

  TxfmParam txfm_param;
  init_txfm_param(xd, plane, tx_size, tx_type, eob, reduced_tx_set,
                  &txfm_param);
  assert(av1_ext_tx_used[txfm_param.tx_set_type][txfm_param.tx_type]);

  if (txfm_param.is_hbd) {
#if CONFIG_MODE_DEP_INTRA_TX || CONFIG_MODE_DEP_INTER_TX || CONFIG_LGT
    av1_highbd_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST7_16X16 && CONFIG_SUPERRES_TX64 && CONFIG_DST_32X32
    if (tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16 ||
        tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32 ||
        txsize_sqr_up_map[tx_size] == TX_64X64)
      av1_highbd_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_highbd_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST7_16X16 && CONFIG_SUPERRES_TX64
    if (tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16 ||
        txsize_sqr_up_map[tx_size] == TX_64X64)
      av1_highbd_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_highbd_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST7_16X16 && CONFIG_DST_32X32
    if (tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16 ||
        tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32)
      av1_highbd_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_highbd_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST_32X32 && CONFIG_SUPERRES_TX64
    if (tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32 ||
        txsize_sqr_up_map[tx_size] == TX_64X64)
      av1_highbd_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_highbd_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST7_16X16
    if (tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16)
      av1_highbd_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_highbd_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST_32X32
    if (tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32)
      av1_highbd_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_highbd_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_SUPERRES_TX64
    if (txsize_sqr_up_map[tx_size] == TX_64X64)
      av1_highbd_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_highbd_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#else
    av1_highbd_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#endif  // CONFIG_MODE_DEP_INTRA_TX || CONFIG_MODE_DEP_INTER_TX || CONFIG_LGT
  } else {
#if CONFIG_MODE_DEP_INTRA_TX || CONFIG_MODE_DEP_INTER_TX || CONFIG_LGT
    av1_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST7_16X16 && CONFIG_SUPERRES_TX64 && CONFIG_DST_32X32
    if (tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16 ||
        tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32 ||
        txsize_sqr_up_map[tx_size] == TX_64X64)
      av1_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST7_16X16 && CONFIG_SUPERRES_TX64
    if (tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16 ||
        txsize_sqr_up_map[tx_size] == TX_64X64)
      av1_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST_32X32 && CONFIG_SUPERRES_TX64
    if (tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32 ||
        txsize_sqr_up_map[tx_size] == TX_64X64)
      av1_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST7_16X16 && CONFIG_DST_32X32
    if (tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16 ||
        tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32)
      av1_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST7_16X16
    if (tx_size_wide[tx_size] == 16 || tx_size_high[tx_size] == 16)
      av1_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DST_32X32
    if (tx_size_wide[tx_size] == 32 || tx_size_high[tx_size] == 32)
      av1_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_SUPERRES_TX64
    if (txsize_sqr_up_map[tx_size] == TX_64X64)
      av1_inv_txfm_add_c(dqcoeff, dst, stride, &txfm_param);
    else
      av1_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#elif CONFIG_DSPL_RESIDUAL
    DSPL_TYPE dspl_type = xd->mi[0]->dspl_type;
    if (plane == 0 && dspl_type == DSPL_XY && xd->bd == 8) {
      av1_inverse_dspl_transform_block(xd, dqcoeff, plane, tx_type, tx_size,
                                       dst, stride, eob, reduced_tx_set);
    } else {
      av1_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
    }
#else
    av1_inv_txfm_add(dqcoeff, dst, stride, &txfm_param);
#endif  // CONFIG_MODE_DEP_INTRA_TX || CONFIG_MODE_DEP_INTER_TX || CONFIG_LGT
#if CONFIG_NN_RECON
    if (xd->mi[0]->use_nn_recon && plane == 0) {
      av1_cnn_recon(xd, dst, stride, tx_size);
    }
#endif  // CONFIG_NN_RECON
  }
}