Quelle  vp9_dct_neon.c

/*
 *  Copyright (c) 2022 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <arm_neon.h>

#include "./vpx_config.h"
#include "./vp9_rtcd.h"
#include "./vpx_dsp_rtcd.h"

#include "vpx_dsp/txfm_common.h"
#include "vpx_dsp/arm/mem_neon.h"
#include "vpx_dsp/arm/transpose_neon.h"
#include "vpx_dsp/arm/fdct_neon.h"
#include "vpx_dsp/arm/fdct4x4_neon.h"
#include "vpx_dsp/arm/fdct8x8_neon.h"
#include "vpx_dsp/arm/fdct16x16_neon.h"

static INLINE void load_buffer_4x4(const int16_t *input, int16x8_t *in,
                                   int stride) {
  // { 0, 1, 1, 1 };
  const int16x4_t nonzero_bias_a = vext_s16(vdup_n_s16(0), vdup_n_s16(1), 3);
  // { 1, 0, 0, 0 };
  const int16x4_t nonzero_bias_b = vext_s16(vdup_n_s16(1), vdup_n_s16(0), 3);
  int16x4_t mask;

  int16x4_t input_0 = vshl_n_s16(vld1_s16(input + 0 * stride), 4);
  int16x4_t input_1 = vshl_n_s16(vld1_s16(input + 1 * stride), 4);
  int16x4_t input_2 = vshl_n_s16(vld1_s16(input + 2 * stride), 4);
  int16x4_t input_3 = vshl_n_s16(vld1_s16(input + 3 * stride), 4);

  // Copy the SSE method, use a mask to avoid an 'if' branch here to increase by
  // one non-zero first elements
  mask = vreinterpret_s16_u16(vceq_s16(input_0, nonzero_bias_a));
  input_0 = vadd_s16(input_0, mask);
  input_0 = vadd_s16(input_0, nonzero_bias_b);

  in[0] = vcombine_s16(input_0, input_1);
  in[1] = vcombine_s16(input_2, input_3);
}

static INLINE void write_buffer_4x4(tran_low_t *output, int16x8_t *res) {
  const int16x8_t one_s16 = vdupq_n_s16(1);
  res[0] = vaddq_s16(res[0], one_s16);
  res[1] = vaddq_s16(res[1], one_s16);
  res[0] = vshrq_n_s16(res[0], 2);
  res[1] = vshrq_n_s16(res[1], 2);
  store_s16q_to_tran_low(output + 0 * 8, res[0]);
  store_s16q_to_tran_low(output + 1 * 8, res[1]);
}

static INLINE void fadst4x4_neon(int16x8_t *in) {
  int32x4_t u[4], t[4];
  int16x4_t s[4], out[4];

  s[0] = vget_low_s16(in[0]);   // | x_00 | x_01 | x_02 | x_03 |
  s[1] = vget_high_s16(in[0]);  // | x_10 | x_11 | x_12 | x_13 |
  s[2] = vget_low_s16(in[1]);   // | x_20 | x_21 | x_22 | x_23 |
  s[3] = vget_high_s16(in[1]);  // | x_30 | x_31 | x_32 | x_33 |

  // Must expand all elements to s32. See 'needs32' comment in fwd_txfm.c.
  // t0 = s0 * sinpi_1_9 + s1 * sinpi_2_9 + s3 * sinpi_4_9
  t[0] = vmull_n_s16(s[0], sinpi_1_9);
  t[0] = vmlal_n_s16(t[0], s[1], sinpi_2_9);
  t[0] = vmlal_n_s16(t[0], s[3], sinpi_4_9);

  // t1 = (s0 + s1) * sinpi_3_9 - s3 * sinpi_3_9
  t[1] = vmull_n_s16(s[0], sinpi_3_9);
  t[1] = vmlal_n_s16(t[1], s[1], sinpi_3_9);
  t[1] = vmlsl_n_s16(t[1], s[3], sinpi_3_9);

  // t2 = s0 * sinpi_4_9 - s1* sinpi_1_9 + s3 * sinpi_2_9
  t[2] = vmull_n_s16(s[0], sinpi_4_9);
  t[2] = vmlsl_n_s16(t[2], s[1], sinpi_1_9);
  t[2] = vmlal_n_s16(t[2], s[3], sinpi_2_9);

  // t3 = s2 * sinpi_3_9
  t[3] = vmull_n_s16(s[2], sinpi_3_9);

  /*
   * u0 = t0 + t3
   * u1 = t1
   * u2 = t2 - t3
   * u3 = t2 - t0 + t3
   */
  u[0] = vaddq_s32(t[0], t[3]);
  u[1] = t[1];
  u[2] = vsubq_s32(t[2], t[3]);
  u[3] = vaddq_s32(vsubq_s32(t[2], t[0]), t[3]);

  // fdct_round_shift
  out[0] = vrshrn_n_s32(u[0], DCT_CONST_BITS);
  out[1] = vrshrn_n_s32(u[1], DCT_CONST_BITS);
  out[2] = vrshrn_n_s32(u[2], DCT_CONST_BITS);
  out[3] = vrshrn_n_s32(u[3], DCT_CONST_BITS);

  transpose_s16_4x4d(&out[0], &out[1], &out[n style='color: green'>2], &out[3]);

  in[0] = vcombine_s16(out[0], out[1]);
  in[1] = vcombine_s16(out[2], out[3]);
}

void vp9_fht4x4_neon(const int16_t *input, tran_low_t *output, int stride,
                     int tx_type) {
  int16x8_t in[2];

  switch (tx_type) {
    case DCT_DCT: vpx_fdct4x4_neon(input, output, stride); break;
    case ADST_DCT:
      load_buffer_4x4(input, in, stride);
      fadst4x4_neon(in);
      // pass1 variant is not accurate enough
      vpx_fdct4x4_pass2_neon((int16x4_t *)in);
      write_buffer_4x4(output, in);
      break;
    case DCT_ADST:
      load_buffer_4x4(input, in, stride);
      // pass1 variant is not accurate enough
      vpx_fdct4x4_pass2_neon((int16x4_t *)in);
      fadst4x4_neon(in);
      write_buffer_4x4(output, in);
      break;
    default:
      assert(tx_type == ADST_ADST);
      load_buffer_4x4(input, in, stride);
      fadst4x4_neon(in);
      fadst4x4_neon(in);
      write_buffer_4x4(output, in);
      break;
  }
}

static INLINE void load_buffer_8x8(const int16_t *input, int16x8_t *in,
                                   int stride) {
  in[0] = vshlq_n_s16(vld1q_s16(input + 0 * stride), 2);
  in[1] = vshlq_n_s16(vld1q_s16(input + 1 * stride), 2);
  in[2] = vshlq_n_s16(vld1q_s16(input + 2 * stride), 2);
  in[3] = vshlq_n_s16(vld1q_s16(input + 3 * stride), 2);
  in[4] = vshlq_n_s16(vld1q_s16(input + 4 * stride), 2);
  in[5] = vshlq_n_s16(vld1q_s16(input + 5 * stride), 2);
  in[6] = vshlq_n_s16(vld1q_s16(input + 6 * stride), 2);
  in[7] = vshlq_n_s16(vld1q_s16(input + 7 * stride), 2);
}

/* right shift and rounding
 * first get the sign bit (bit 15).
 * If bit == 1, it's the simple case of shifting right by one bit.
 * If bit == 2, it essentially computes the expression:
 *
 * out[j * 16 + i] = (temp_out[j] + 1 + (temp_out[j] < 0)) >> 2;
 *
 * for each row.
 */
static INLINE void right_shift_8x8(int16x8_t *res, const int bit) {
  int16x8_t sign0 = vshrq_n_s16(res[0], 15);
  int16x8_t sign1 = vshrq_n_s16(res[1], 15);
  int16x8_t sign2 = vshrq_n_s16(res[2], 15);
  int16x8_t sign3 = vshrq_n_s16(res[3], 15);
  int16x8_t sign4 = vshrq_n_s16(res[4], 15);
  int16x8_t sign5 = vshrq_n_s16(res[5], 15);
  int16x8_t sign6 = vshrq_n_s16(res[6], 15);
  int16x8_t sign7 = vshrq_n_s16(res[7], 15);

  if (bit == 2) {
    const int16x8_t const_rounding = vdupq_n_s16(1);
    res[0] = vaddq_s16(res[0], const_rounding);
    res[1] = vaddq_s16(res[1], const_rounding);
    res[2] = vaddq_s16(res[2], const_rounding);
    res[3] = vaddq_s16(res[3], const_rounding);
    res[4] = vaddq_s16(res[4], const_rounding);
    res[5] = vaddq_s16(res[5], const_rounding);
    res[6] = vaddq_s16(res[6], const_rounding);
    res[7] = vaddq_s16(res[7], const_rounding);
  }

  res[0] = vsubq_s16(res[0], sign0);
  res[1] = vsubq_s16(res[1], sign1);
  res[2] = vsubq_s16(res[2], sign2);
  res[3] = vsubq_s16(res[3], sign3);
  res[4] = vsubq_s16(res[4], sign4);
  res[5] = vsubq_s16(res[5], sign5);
  res[6] = vsubq_s16(res[6], sign6);
  res[7] = vsubq_s16(res[7], sign7);

  if (bit == 1) {
    res[0] = vshrq_n_s16(res[0], 1);
    res[1] = vshrq_n_s16(res[1], 1);
    res[2] = vshrq_n_s16(res[2], 1);
    res[3] = vshrq_n_s16(res[3], 1);
    res[4] = vshrq_n_s16(res[4], 1);
    res[5] = vshrq_n_s16(res[5], 1);
    res[6] = vshrq_n_s16(res[6], 1);
    res[7] = vshrq_n_s16(res[7], 1);
  } else {
    res[0] = vshrq_n_s16(res[0], 2);
    res[1] = vshrq_n_s16(res[1], 2);
    res[2] = vshrq_n_s16(res[2], 2);
    res[3] = vshrq_n_s16(res[3], 2);
    res[4] = vshrq_n_s16(res[4], 2);
    res[5] = vshrq_n_s16(res[5], 2);
    res[6] = vshrq_n_s16(res[6], 2);
    res[7] = vshrq_n_s16(res[7], 2);
  }
}

static INLINE void write_buffer_8x8(tran_low_t *output, int16x8_t *res,
                                    int stride) {
  store_s16q_to_tran_low(output + 0 * stride, res[0]);
  store_s16q_to_tran_low(output + 1 * stride, res[1]);
  store_s16q_to_tran_low(output + 2 * stride, res[2]);
  store_s16q_to_tran_low(output + 3 * stride, res[3]);
  store_s16q_to_tran_low(output + 4 * stride, res[4]);
  store_s16q_to_tran_low(output + 5 * stride, res[5]);
  store_s16q_to_tran_low(output + 6 * stride, res[6]);
  store_s16q_to_tran_low(output + 7 * stride, res[7]);
}

static INLINE void fadst8x8_neon(int16x8_t *in) {
  int16x4_t x_lo[8], x_hi[8];
  int32x4_t s_lo[8], s_hi[8];
  int32x4_t t_lo[8], t_hi[8];

  x_lo[0] = vget_low_s16(in[7]);
  x_hi[0] = vget_high_s16(in[7]);
  x_lo[1] = vget_low_s16(in[0]);
  x_hi[1] = vget_high_s16(in[0]);
  x_lo[2] = vget_low_s16(in[5]);
  x_hi[2] = vget_high_s16(in[5]);
  x_lo[3] = vget_low_s16(in[2]);
  x_hi[3] = vget_high_s16(in[2]);
  x_lo[4] = vget_low_s16(in[3]);
  x_hi[4] = vget_high_s16(in[3]);
  x_lo[5] = vget_low_s16(in[4]);
  x_hi[5] = vget_high_s16(in[4]);
  x_lo[6] = vget_low_s16(in[1]);
  x_hi[6] = vget_high_s16(in[1]);
  x_lo[7] = vget_low_s16(in[6]);
  x_hi[7] = vget_high_s16(in[6]);

  // stage 1
  // s0 = cospi_2_64 * x0 + cospi_30_64 * x1;
  // s1 = cospi_30_64 * x0 - cospi_2_64 * x1;
  butterfly_two_coeff_s16_s32_noround(x_lo[0], x_hi[0], x_lo[1], x_hi[1],
                                      cospi_2_64, cospi_30_64, &s_lo[0],
                                      &s_hi[0], &s_lo[1], &s_hi[1]);

  // s2 = cospi_10_64 * x2 + cospi_22_64 * x3;
  // s3 = cospi_22_64 * x2 - cospi_10_64 * x3;
  butterfly_two_coeff_s16_s32_noround(x_lo[2], x_hi[2], x_lo[3], x_hi[3],
                                      cospi_10_64, cospi_22_64, &s_lo[2],
                                      &s_hi[2], &s_lo[3], &s_hi[3]);

  // s4 = cospi_18_64 * x4 + cospi_14_64 * x5;
  // s5 = cospi_14_64 * x4 - cospi_18_64 * x5;
  butterfly_two_coeff_s16_s32_noround(x_lo[4], x_hi[4], x_lo[5], x_hi[5],
                                      cospi_18_64, cospi_14_64, &s_lo[4],
                                      &s_hi[4], &s_lo[5], &s_hi[5]);

  // s6 = cospi_26_64 * x6 + cospi_6_64 * x7;
  // s7 = cospi_6_64 * x6 - cospi_26_64 * x7;
  butterfly_two_coeff_s16_s32_noround(x_lo[6], x_hi[6], x_lo[7], x_hi[7],
                                      cospi_26_64, cospi_6_64, &s_lo[6],
                                      &s_hi[6], &s_lo[7], &s_hi[7]);

  // fdct_round_shift
  t_lo[0] = vrshrq_n_s32(vaddq_s32(s_lo[0], s_lo[4]), DCT_CONST_BITS);
  t_hi[0] = vrshrq_n_s32(vaddq_s32(s_hi[0], s_hi[4]), DCT_CONST_BITS);
  t_lo[1] = vrshrq_n_s32(vaddq_s32(s_lo[1], s_lo[5]), DCT_CONST_BITS);
  t_hi[1] = vrshrq_n_s32(vaddq_s32(s_hi[1], s_hi[5]), DCT_CONST_BITS);
  t_lo[2] = vrshrq_n_s32(vaddq_s32(s_lo[2], s_lo[6]), DCT_CONST_BITS);
  t_hi[2] = vrshrq_n_s32(vaddq_s32(s_hi[2], s_hi[6]), DCT_CONST_BITS);
  t_lo[3] = vrshrq_n_s32(vaddq_s32(s_lo[3], s_lo[7]), DCT_CONST_BITS);
  t_hi[3] = vrshrq_n_s32(vaddq_s32(s_hi[3], s_hi[7]), DCT_CONST_BITS);
  t_lo[4] = vrshrq_n_s32(vsubq_s32(s_lo[0], s_lo[4]), DCT_CONST_BITS);
  t_hi[4] = vrshrq_n_s32(vsubq_s32(s_hi[0], s_hi[4]), DCT_CONST_BITS);
  t_lo[5] = vrshrq_n_s32(vsubq_s32(s_lo[1], s_lo[5]), DCT_CONST_BITS);
  t_hi[5] = vrshrq_n_s32(vsubq_s32(s_hi[1], s_hi[5]), DCT_CONST_BITS);
  t_lo[6] = vrshrq_n_s32(vsubq_s32(s_lo[2], s_lo[6]), DCT_CONST_BITS);
  t_hi[6] = vrshrq_n_s32(vsubq_s32(s_hi[2], s_hi[6]), DCT_CONST_BITS);
  t_lo[7] = vrshrq_n_s32(vsubq_s32(s_lo[3], s_lo[7]), DCT_CONST_BITS);
  t_hi[7] = vrshrq_n_s32(vsubq_s32(s_hi[3], s_hi[7]), DCT_CONST_BITS);

  // stage 2
  s_lo[0] = t_lo[0];
  s_hi[0] = t_hi[0];
  s_lo[1] = t_lo[1];
  s_hi[1] = t_hi[1];
  s_lo[2] = t_lo[2];
  s_hi[2] = t_hi[2];
  s_lo[3] = t_lo[3];
  s_hi[3] = t_hi[3];
  // s4 = cospi_8_64 * x4 + cospi_24_64 * x5;
  // s5 = cospi_24_64 * x4 - cospi_8_64 * x5;
  butterfly_two_coeff_s32_noround(t_lo[4], t_hi[4], t_lo[5], t_hi[5],
                                  cospi_8_64, cospi_24_64, &s_lo[4], &s_hi[4],
                                  &s_lo[5], &s_hi[5]);

  // s6 = -cospi_24_64 * x6 + cospi_8_64 * x7;
  // s7 = cospi_8_64 * x6 + cospi_24_64 * x7;
  butterfly_two_coeff_s32_noround(t_lo[6], t_hi[6], t_lo[7], t_hi[7],
                                  -cospi_24_64, cospi_8_64, &s_lo[6], &s_hi[6],
                                  &s_lo[7], &s_hi[7]);

  // fdct_round_shift
  // s0 + s2
  t_lo[0] = vaddq_s32(s_lo[0], s_lo[2]);
  t_hi[0] = vaddq_s32(s_hi[0], s_hi[2]);
  // s1 + s3
  t_lo[1] = vaddq_s32(s_lo[1], s_lo[3]);
  t_hi[1] = vaddq_s32(s_hi[1], s_hi[3]);
  // s0 - s2
  t_lo[2] = vsubq_s32(s_lo[0], s_lo[2]);
  t_hi[2] = vsubq_s32(s_hi[0], s_hi[2]);
  // s1 - s3
  t_lo[3] = vsubq_s32(s_lo[1], s_lo[3]);
  t_hi[3] = vsubq_s32(s_hi[1], s_hi[3]);
  // s4 + s6
  t_lo[4] = vrshrq_n_s32(vaddq_s32(s_lo[4], s_lo[6]), DCT_CONST_BITS);
  t_hi[4] = vrshrq_n_s32(vaddq_s32(s_hi[4], s_hi[6]), DCT_CONST_BITS);
  // s5 + s7
  t_lo[5] = vrshrq_n_s32(vaddq_s32(s_lo[5], s_lo[7]), DCT_CONST_BITS);
  t_hi[5] = vrshrq_n_s32(vaddq_s32(s_hi[5], s_hi[7]), DCT_CONST_BITS);
  // s4 - s6
  t_lo[6] = vrshrq_n_s32(vsubq_s32(s_lo[4], s_lo[6]), DCT_CONST_BITS);
  t_hi[6] = vrshrq_n_s32(vsubq_s32(s_hi[4], s_hi[6]), DCT_CONST_BITS);
  // s5 - s7
  t_lo[7] = vrshrq_n_s32(vsubq_s32(s_lo[5], s_lo[7]), DCT_CONST_BITS);
  t_hi[7] = vrshrq_n_s32(vsubq_s32(s_hi[5], s_hi[7]), DCT_CONST_BITS);

  // stage 3
  // cospi_16_64 * (x2 + x3)
  // cospi_16_64 * (x2 - x3)
  butterfly_one_coeff_s32_noround(t_lo[2], t_hi[2], t_lo[3], t_hi[3],
                                  cospi_16_64, &s_lo[2], &s_hi[2], &s_lo[='color: green'>3],
                                  &s_hi[3]);

  // cospi_16_64 * (x6 + x7)
  // cospi_16_64 * (x2 - x3)
  butterfly_one_coeff_s32_noround(t_lo[6], t_hi[6], t_lo[7], t_hi[7],
                                  cospi_16_64, &s_lo[6], &s_hi[6], &s_lo[='color: green'>7],
                                  &s_hi[7]);

  // final fdct_round_shift
  x_lo[2] = vrshrn_n_s32(s_lo[2], DCT_CONST_BITS);
  x_hi[2] = vrshrn_n_s32(s_hi[2], DCT_CONST_BITS);
  x_lo[3] = vrshrn_n_s32(s_lo[3], DCT_CONST_BITS);
  x_hi[3] = vrshrn_n_s32(s_hi[3], DCT_CONST_BITS);
  x_lo[6] = vrshrn_n_s32(s_lo[6], DCT_CONST_BITS);
  x_hi[6] = vrshrn_n_s32(s_hi[6], DCT_CONST_BITS);
  x_lo[7] = vrshrn_n_s32(s_lo[7], DCT_CONST_BITS);
  x_hi[7] = vrshrn_n_s32(s_hi[7], DCT_CONST_BITS);

  // x0, x1, x4, x5 narrow down to 16-bits directly
  x_lo[0] = vmovn_s32(t_lo[0]);
  x_hi[0] = vmovn_s32(t_hi[0]);
  x_lo[1] = vmovn_s32(t_lo[1]);
  x_hi[1] = vmovn_s32(t_hi[1]);
  x_lo[4] = vmovn_s32(t_lo[4]);
  x_hi[4] = vmovn_s32(t_hi[4]);
  x_lo[5] = vmovn_s32(t_lo[5]);
  x_hi[5] = vmovn_s32(t_hi[5]);

  in[0] = vcombine_s16(x_lo[0], x_hi[0]);
  in[1] = vnegq_s16(vcombine_s16(x_lo[4], x_hi[4]));
  in[2] = vcombine_s16(x_lo[6], x_hi[6]);
  in[3] = vnegq_s16(vcombine_s16(x_lo[2], x_hi[2]));
  in[4] = vcombine_s16(x_lo[3], x_hi[3]);
  in[5] = vnegq_s16(vcombine_s16(x_lo[7], x_hi[7]));
  in[6] = vcombine_s16(x_lo[5], x_hi[5]);
  in[7] = vnegq_s16(vcombine_s16(x_lo[1], x_hi[1]));

  transpose_s16_8x8(&in[0], &in[1], &in[ style='color: green'>2], &in[3], &in[4], &in[5], &in[6],
                    &in[7]);
}

void vp9_fht8x8_neon(const int16_t *input, tran_low_t *output, int stride,
                     int tx_type) {
  int16x8_t in[8];

  switch (tx_type) {
    case DCT_DCT: vpx_fdct8x8_neon(input, output, stride); break;
    case ADST_DCT:
      load_buffer_8x8(input, in, stride);
      fadst8x8_neon(in);
      // pass1 variant is not accurate enough
      vpx_fdct8x8_pass2_neon(in);
      right_shift_8x8(in, 1);
      write_buffer_8x8(output, in, 8);
      break;
    case DCT_ADST:
      load_buffer_8x8(input, in, stride);
      // pass1 variant is not accurate enough
      vpx_fdct8x8_pass2_neon(in);
      fadst8x8_neon(in);
      right_shift_8x8(in, 1);
      write_buffer_8x8(output, in, 8);
      break;
    default:
      assert(tx_type == ADST_ADST);
      load_buffer_8x8(input, in, stride);
      fadst8x8_neon(in);
      fadst8x8_neon(in);
      right_shift_8x8(in, 1);
      write_buffer_8x8(output, in, 8);
      break;
  }
}

static INLINE void load_buffer_16x16(const int16_t *input, int16x8_t *in0,
                                     int16x8_t *in1, int stride) {
  // load first 8 columns
  load_buffer_8x8(input, in0, stride);
  load_buffer_8x8(input + 8 * stride, in0 + 8, stride);

  input += 8;
  // load second 8 columns
  load_buffer_8x8(input, in1, stride);
  load_buffer_8x8(input + 8 * stride, in1 + 8, stride);
}

static INLINE void write_buffer_16x16(tran_low_t *output, int16x8_t *in0,
                                      int16x8_t *in1, int stride) {
  // write first 8 columns
  write_buffer_8x8(output, in0, stride);
  write_buffer_8x8(output + 8 * stride, in0 + 8, stride);

  // write second 8 columns
  output += 8;
  write_buffer_8x8(output, in1, stride);
  write_buffer_8x8(output + 8 * stride, in1 + 8, stride);
}

static INLINE void right_shift_16x16(int16x8_t *res0, int16x8_t *res1) {
  // perform rounding operations
  right_shift_8x8(res0, 2);
  right_shift_8x8(res0 + 8, 2);
  right_shift_8x8(res1, 2);
  right_shift_8x8(res1 + 8, 2);
}

static void fdct16_8col(int16x8_t *in) {
  // perform 16x16 1-D DCT for 8 columns
  int16x8_t i[8], s1[8], s2[8], s3[8], t[8];
  int16x4_t t_lo[8], t_hi[8];
  int32x4_t u_lo[8], u_hi[8];

  // stage 1
  i[0] = vaddq_s16(in[0], in[15]);
  i[1] = vaddq_s16(in[1], in[14]);
  i[2] = vaddq_s16(in[2], in[13]);
  i[3] = vaddq_s16(in[3], in[12]);
  i[4] = vaddq_s16(in[4], in[11]);
  i[5] = vaddq_s16(in[5], in[10]);
  i[6] = vaddq_s16(in[6], in[9]);
  i[7] = vaddq_s16(in[7], in[8]);

  // pass1 variant is not accurate enough
  vpx_fdct8x8_pass2_neon(i);
  transpose_s16_8x8(&i[0], &i[1], &i[ style='color: green'>2], &i[3], &i[4], &i[5], &i[6], &i[7]);

  // step 2
  s1[0] = vsubq_s16(in[7], in[8]);
  s1[1] = vsubq_s16(in[6], in[9]);
  s1[2] = vsubq_s16(in[5], in[10]);
  s1[3] = vsubq_s16(in[4], in[11]);
  s1[4] = vsubq_s16(in[3], in[12]);
  s1[5] = vsubq_s16(in[2], in[13]);
  s1[6] = vsubq_s16(in[1], in[14]);
  s1[7] = vsubq_s16(in[0], in[15]);

  t[2] = vsubq_s16(s1[5], s1[2]);
  t[3] = vsubq_s16(s1[4], s1[3]);
  t[4] = vaddq_s16(s1[4], s1[3]);
  t[5] = vaddq_s16(s1[5], s1[2]);

  t_lo[2] = vget_low_s16(t[2]);
  t_hi[2] = vget_high_s16(t[2]);
  t_lo[3] = vget_low_s16(t[3]);
  t_hi[3] = vget_high_s16(t[3]);
  t_lo[4] = vget_low_s16(t[4]);
  t_hi[4] = vget_high_s16(t[4]);
  t_lo[5] = vget_low_s16(t[5]);
  t_hi[5] = vget_high_s16(t[5]);

  u_lo[2] = vmull_n_s16(t_lo[2], cospi_16_64);
  u_hi[2] = vmull_n_s16(t_hi[2], cospi_16_64);
  u_lo[3] = vmull_n_s16(t_lo[3], cospi_16_64);
  u_hi[3] = vmull_n_s16(t_hi[3], cospi_16_64);
  u_lo[4] = vmull_n_s16(t_lo[4], cospi_16_64);
  u_hi[4] = vmull_n_s16(t_hi[4], cospi_16_64);
  u_lo[5] = vmull_n_s16(t_lo[5], cospi_16_64);
  u_hi[5] = vmull_n_s16(t_hi[5], cospi_16_64);

  t_lo[2] = vrshrn_n_s32(u_lo[2], DCT_CONST_BITS);
  t_hi[2] = vrshrn_n_s32(u_hi[2], DCT_CONST_BITS);
  t_lo[3] = vrshrn_n_s32(u_lo[3], DCT_CONST_BITS);
  t_hi[3] = vrshrn_n_s32(u_hi[3], DCT_CONST_BITS);
  t_lo[4] = vrshrn_n_s32(u_lo[4], DCT_CONST_BITS);
  t_hi[4] = vrshrn_n_s32(u_hi[4], DCT_CONST_BITS);
  t_lo[5] = vrshrn_n_s32(u_lo[5], DCT_CONST_BITS);
  t_hi[5] = vrshrn_n_s32(u_hi[5], DCT_CONST_BITS);

  s2[2] = vcombine_s16(t_lo[2], t_hi[2]);
  s2[3] = vcombine_s16(t_lo[3], t_hi[3]);
  s2[4] = vcombine_s16(t_lo[4], t_hi[4]);
  s2[5] = vcombine_s16(t_lo[5], t_hi[5]);

  // step 3
  s3[0] = vaddq_s16(s1[0], s2[3]);
  s3[1] = vaddq_s16(s1[1], s2[2]);
  s3[2] = vsubq_s16(s1[1], s2[2]);
  s3[3] = vsubq_s16(s1[0], s2[3]);
  s3[4] = vsubq_s16(s1[7], s2[4]);
  s3[5] = vsubq_s16(s1[6], s2[5]);
  s3[6] = vaddq_s16(s1[6], s2[5]);
  s3[7] = vaddq_s16(s1[7], s2[4]);

  // step 4
  t_lo[0] = vget_low_s16(s3[0]);
  t_hi[0] = vget_high_s16(s3[0]);
  t_lo[1] = vget_low_s16(s3[1]);
  t_hi[1] = vget_high_s16(s3[1]);
  t_lo[2] = vget_low_s16(s3[2]);
  t_hi[2] = vget_high_s16(s3[2]);
  t_lo[3] = vget_low_s16(s3[3]);
  t_hi[3] = vget_high_s16(s3[3]);
  t_lo[4] = vget_low_s16(s3[4]);
  t_hi[4] = vget_high_s16(s3[4]);
  t_lo[5] = vget_low_s16(s3[5]);
  t_hi[5] = vget_high_s16(s3[5]);
  t_lo[6] = vget_low_s16(s3[6]);
  t_hi[6] = vget_high_s16(s3[6]);
  t_lo[7] = vget_low_s16(s3[7]);
  t_hi[7] = vget_high_s16(s3[7]);

  // u[1] = -cospi_8_64 * t[1] + cospi_24_64 * t[6]
  // u[6] = cospi_24_64 * t[1] + cospi_8_64 * t[6]
  butterfly_two_coeff_s16_s32_noround(t_lo[1], t_hi[1], t_lo[6], t_hi[6],
                                      -cospi_8_64, cospi_24_64, &u_lo[1],
                                      &u_hi[1], &u_lo[6], &u_hi[6]);

  // u[5] = -cospi_24_64 * t[5] + cospi_8_64 * t[2]
  // u[2] = cospi_8_64 * t[5]   + cospi_24_64 * t[2]
  butterfly_two_coeff_s16_s32_noround(t_lo[5], t_hi[5], t_lo[2], t_hi[2],
                                      -cospi_24_64, cospi_8_64, &u_lo[5],
                                      &u_hi[5], &u_lo[2], &u_hi[2]);

  t_lo[1] = vrshrn_n_s32(u_lo[1], DCT_CONST_BITS);
  t_hi[1] = vrshrn_n_s32(u_hi[1], DCT_CONST_BITS);
  t_lo[2] = vrshrn_n_s32(u_lo[2], DCT_CONST_BITS);
  t_hi[2] = vrshrn_n_s32(u_hi[2], DCT_CONST_BITS);
  t_lo[5] = vrshrn_n_s32(u_lo[5], DCT_CONST_BITS);
  t_hi[5] = vrshrn_n_s32(u_hi[5], DCT_CONST_BITS);
  t_lo[6] = vrshrn_n_s32(u_lo[6], DCT_CONST_BITS);
  t_hi[6] = vrshrn_n_s32(u_hi[6], DCT_CONST_BITS);

  s2[1] = vcombine_s16(t_lo[1], t_hi[1]);
  s2[2] = vcombine_s16(t_lo[2], t_hi[2]);
  s2[5] = vcombine_s16(t_lo[5], t_hi[5]);
  s2[6] = vcombine_s16(t_lo[6], t_hi[6]);

  // step 5
  s1[0] = vaddq_s16(s3[0], s2[1]);
  s1[1] = vsubq_s16(s3[0], s2[1]);
  s1[2] = vaddq_s16(s3[3], s2[2]);
  s1[3] = vsubq_s16(s3[3], s2[2]);
  s1[4] = vsubq_s16(s3[4], s2[5]);
  s1[5] = vaddq_s16(s3[4], s2[5]);
  s1[6] = vsubq_s16(s3[7], s2[6]);
  s1[7] = vaddq_s16(s3[7], s2[6]);

  // step 6
  t_lo[0] = vget_low_s16(s1[0]);
  t_hi[0] = vget_high_s16(s1[0]);
  t_lo[1] = vget_low_s16(s1[1]);
  t_hi[1] = vget_high_s16(s1[1]);
  t_lo[2] = vget_low_s16(s1[2]);
  t_hi[2] = vget_high_s16(s1[2]);
  t_lo[3] = vget_low_s16(s1[3]);
  t_hi[3] = vget_high_s16(s1[3]);
  t_lo[4] = vget_low_s16(s1[4]);
  t_hi[4] = vget_high_s16(s1[4]);
  t_lo[5] = vget_low_s16(s1[5]);
  t_hi[5] = vget_high_s16(s1[5]);
  t_lo[6] = vget_low_s16(s1[6]);
  t_hi[6] = vget_high_s16(s1[6]);
  t_lo[7] = vget_low_s16(s1[7]);
  t_hi[7] = vget_high_s16(s1[7]);

  // u[0] = step1[7] * cospi_2_64 + step1[0] * cospi_30_64
  // u[7] = step1[7] * cospi_30_64 - step1[0] * cospi_2_64
  butterfly_two_coeff_s16_s32_noround(t_lo[7], t_hi[7], t_lo[0], t_hi[0],
                                      cospi_2_64, cospi_30_64, &u_lo[0],
                                      &u_hi[0], &u_lo[7], &u_hi[7]);

  // u[1] = step1[6] * cospi_18_64 + step1[1] * cospi_14_64
  // u[6] = step1[6] * cospi_14_64 - step1[1] * cospi_18_64
  butterfly_two_coeff_s16_s32_noround(t_lo[6], t_hi[6], t_lo[1], t_hi[1],
                                      cospi_18_64, cospi_14_64, &u_lo[1],
                                      &u_hi[1], &u_lo[6], &u_hi[6]);

  // u[2] = step1[5] * cospi_10_64 + step1[2] * cospi_22_64
  // u[5] = step1[5] * cospi_22_64 - step1[2] * cospi_10_64
  butterfly_two_coeff_s16_s32_noround(t_lo[5], t_hi[5], t_lo[2], t_hi[2],
                                      cospi_10_64, cospi_22_64, &u_lo[2],
                                      &u_hi[2], &u_lo[5], &u_hi[5]);

  // u[3] = step1[4] * cospi_26_64 + step1[3] * cospi_6_64
  // u[4] = step1[4] * cospi_6_64  - step1[3] * cospi_26_64
  butterfly_two_coeff_s16_s32_noround(t_lo[4], t_hi[4], t_lo[3], t_hi[3],
                                      cospi_26_64, cospi_6_64, &u_lo[3],
                                      &u_hi[3], &u_lo[4], &u_hi[4]);

  // final fdct_round_shift
  t_lo[0] = vrshrn_n_s32(u_lo[0], DCT_CONST_BITS);
  t_hi[0] = vrshrn_n_s32(u_hi[0], DCT_CONST_BITS);
  t_lo[1] = vrshrn_n_s32(u_lo[1], DCT_CONST_BITS);
  t_hi[1] = vrshrn_n_s32(u_hi[1], DCT_CONST_BITS);
  t_lo[2] = vrshrn_n_s32(u_lo[2], DCT_CONST_BITS);
  t_hi[2] = vrshrn_n_s32(u_hi[2], DCT_CONST_BITS);
  t_lo[3] = vrshrn_n_s32(u_lo[3], DCT_CONST_BITS);
  t_hi[3] = vrshrn_n_s32(u_hi[3], DCT_CONST_BITS);
  t_lo[4] = vrshrn_n_s32(u_lo[4], DCT_CONST_BITS);
  t_hi[4] = vrshrn_n_s32(u_hi[4], DCT_CONST_BITS);
  t_lo[5] = vrshrn_n_s32(u_lo[5], DCT_CONST_BITS);
  t_hi[5] = vrshrn_n_s32(u_hi[5], DCT_CONST_BITS);
  t_lo[6] = vrshrn_n_s32(u_lo[6], DCT_CONST_BITS);
  t_hi[6] = vrshrn_n_s32(u_hi[6], DCT_CONST_BITS);
  t_lo[7] = vrshrn_n_s32(u_lo[7], DCT_CONST_BITS);
  t_hi[7] = vrshrn_n_s32(u_hi[7], DCT_CONST_BITS);

  in[0] = i[0];
  in[2] = i[1];
  in[4] = i[2];
  in[6] = i[3];
  in[8] = i[4];
  in[10] = i[5];
  in[12] = i[6];
  in[14] = i[7];
  in[1] = vcombine_s16(t_lo[0], t_hi[0]);
  in[3] = vcombine_s16(t_lo[4], t_hi[4]);
  in[5] = vcombine_s16(t_lo[2], t_hi[2]);
  in[7] = vcombine_s16(t_lo[6], t_hi[6]);
  in[9] = vcombine_s16(t_lo[1], t_hi[1]);
  in[11] = vcombine_s16(t_lo[5], t_hi[5]);
  in[13] = vcombine_s16(t_lo[3], t_hi[3]);
  in[15] = vcombine_s16(t_lo[7], t_hi[7]);
}

static void fadst16_8col(int16x8_t *in) {
  // perform 16x16 1-D ADST for 8 columns
  int16x4_t x_lo[16], x_hi[16];
  int32x4_t s_lo[16], s_hi[16];
  int32x4_t t_lo[16], t_hi[16];

  x_lo[0] = vget_low_s16(in[15]);
  x_hi[0] = vget_high_s16(in[15]);
  x_lo[1] = vget_low_s16(in[0]);
  x_hi[1] = vget_high_s16(in[0]);
  x_lo[2] = vget_low_s16(in[13]);
  x_hi[2] = vget_high_s16(in[13]);
  x_lo[3] = vget_low_s16(in[2]);
  x_hi[3] = vget_high_s16(in[2]);
  x_lo[4] = vget_low_s16(in[11]);
  x_hi[4] = vget_high_s16(in[11]);
  x_lo[5] = vget_low_s16(in[4]);
  x_hi[5] = vget_high_s16(in[4]);
  x_lo[6] = vget_low_s16(in[9]);
  x_hi[6] = vget_high_s16(in[9]);
  x_lo[7] = vget_low_s16(in[6]);
  x_hi[7] = vget_high_s16(in[6]);
  x_lo[8] = vget_low_s16(in[7]);
  x_hi[8] = vget_high_s16(in[7]);
  x_lo[9] = vget_low_s16(in[8]);
  x_hi[9] = vget_high_s16(in[8]);
  x_lo[10] = vget_low_s16(in[5]);
  x_hi[10] = vget_high_s16(in[5]);
  x_lo[11] = vget_low_s16(in[10]);
  x_hi[11] = vget_high_s16(in[10]);
  x_lo[12] = vget_low_s16(in[3]);
  x_hi[12] = vget_high_s16(in[3]);
  x_lo[13] = vget_low_s16(in[12]);
  x_hi[13] = vget_high_s16(in[12]);
  x_lo[14] = vget_low_s16(in[1]);
  x_hi[14] = vget_high_s16(in[1]);
  x_lo[15] = vget_low_s16(in[14]);
  x_hi[15] = vget_high_s16(in[14]);

  // stage 1
  // s0 = cospi_1_64 * x0 + cospi_31_64 * x1;
  // s1 = cospi_31_64 * x0 - cospi_1_64 * x1;
  butterfly_two_coeff_s16_s32_noround(x_lo[0], x_hi[0], x_lo[1], x_hi[1],
                                      cospi_1_64, cospi_31_64, &s_lo[0],
                                      &s_hi[0], &s_lo[1], &s_hi[1]);
  // s2 = cospi_5_64 * x2 + cospi_27_64 * x3;
  // s3 = cospi_27_64 * x2 - cospi_5_64 * x3;
  butterfly_two_coeff_s16_s32_noround(x_lo[2], x_hi[2], x_lo[3], x_hi[3],
                                      cospi_5_64, cospi_27_64, &s_lo[2],
                                      &s_hi[2], &s_lo[3], &s_hi[3]);
  // s4 = cospi_9_64 * x4 + cospi_23_64 * x5;
  // s5 = cospi_23_64 * x4 - cospi_9_64 * x5;
  butterfly_two_coeff_s16_s32_noround(x_lo[4], x_hi[4], x_lo[5], x_hi[5],
                                      cospi_9_64, cospi_23_64, &s_lo[4],
                                      &s_hi[4], &s_lo[5], &s_hi[5]);
  // s6 = cospi_13_64 * x6 + cospi_19_64 * x7;
  // s7 = cospi_19_64 * x6 - cospi_13_64 * x7;
  butterfly_two_coeff_s16_s32_noround(x_lo[6], x_hi[6], x_lo[7], x_hi[7],
                                      cospi_13_64, cospi_19_64, &s_lo[6],
                                      &s_hi[6], &s_lo[7], &s_hi[7]);
  // s8 = cospi_17_64 * x8 + cospi_15_64 * x9;
  // s9 = cospi_15_64 * x8 - cospi_17_64 * x9;
  butterfly_two_coeff_s16_s32_noround(x_lo[8], x_hi[8], x_lo[9], x_hi[9],
                                      cospi_17_64, cospi_15_64, &s_lo[8],
                                      &s_hi[8], &s_lo[9], &s_hi[9]);
  // s10 = cospi_21_64 * x10 + cospi_11_64 * x11;
  // s11 = cospi_11_64 * x10 - cospi_21_64 * x11;
  butterfly_two_coeff_s16_s32_noround(x_lo[10], x_hi[10], x_lo[11], x_hi[11],
                                      cospi_21_64, cospi_11_64, &s_lo[10],
                                      &s_hi[10], &s_lo[11], &s_hi[11]);
  // s12 = cospi_25_64 * x12 + cospi_7_64 * x13;
  // s13 = cospi_7_64 * x12 - cospi_25_64 * x13;
  butterfly_two_coeff_s16_s32_noround(x_lo[12], x_hi[12], x_lo[13], x_hi[13],
                                      cospi_25_64, cospi_7_64, &s_lo[12],
                                      &s_hi[12], &s_lo[13], &s_hi[13]);
  // s14 = cospi_29_64 * x14 + cospi_3_64 * x15;
  // s15 = cospi_3_64 * x14 - cospi_29_64 * x15;
  butterfly_two_coeff_s16_s32_noround(x_lo[14], x_hi[14], x_lo[15], x_hi[15],
                                      cospi_29_64, cospi_3_64, &s_lo[14],
                                      &s_hi[14], &s_lo[15], &s_hi[15]);

  // fdct_round_shift
  t_lo[0] = vrshrq_n_s32(vaddq_s32(s_lo[0], s_lo[8]), DCT_CONST_BITS);
  t_hi[0] = vrshrq_n_s32(vaddq_s32(s_hi[0], s_hi[8]), DCT_CONST_BITS);
  t_lo[1] = vrshrq_n_s32(vaddq_s32(s_lo[1], s_lo[9]), DCT_CONST_BITS);
  t_hi[1] = vrshrq_n_s32(vaddq_s32(s_hi[1], s_hi[9]), DCT_CONST_BITS);
  t_lo[2] = vrshrq_n_s32(vaddq_s32(s_lo[2], s_lo[10]), DCT_CONST_BITS);
  t_hi[2] = vrshrq_n_s32(vaddq_s32(s_hi[2], s_hi[10]), DCT_CONST_BITS);
  t_lo[3] = vrshrq_n_s32(vaddq_s32(s_lo[3], s_lo[11]), DCT_CONST_BITS);
  t_hi[3] = vrshrq_n_s32(vaddq_s32(s_hi[3], s_hi[11]), DCT_CONST_BITS);
  t_lo[4] = vrshrq_n_s32(vaddq_s32(s_lo[4], s_lo[12]), DCT_CONST_BITS);
  t_hi[4] = vrshrq_n_s32(vaddq_s32(s_hi[4], s_hi[12]), DCT_CONST_BITS);
  t_lo[5] = vrshrq_n_s32(vaddq_s32(s_lo[5], s_lo[13]), DCT_CONST_BITS);
  t_hi[5] = vrshrq_n_s32(vaddq_s32(s_hi[5], s_hi[13]), DCT_CONST_BITS);
  t_lo[6] = vrshrq_n_s32(vaddq_s32(s_lo[6], s_lo[14]), DCT_CONST_BITS);
  t_hi[6] = vrshrq_n_s32(vaddq_s32(s_hi[6], s_hi[14]), DCT_CONST_BITS);
  t_lo[7] = vrshrq_n_s32(vaddq_s32(s_lo[7], s_lo[15]), DCT_CONST_BITS);
  t_hi[7] = vrshrq_n_s32(vaddq_s32(s_hi[7], s_hi[15]), DCT_CONST_BITS);
  t_lo[8] = vrshrq_n_s32(vsubq_s32(s_lo[0], s_lo[8]), DCT_CONST_BITS);
  t_hi[8] = vrshrq_n_s32(vsubq_s32(s_hi[0], s_hi[8]), DCT_CONST_BITS);
  t_lo[9] = vrshrq_n_s32(vsubq_s32(s_lo[1], s_lo[9]), DCT_CONST_BITS);
  t_hi[9] = vrshrq_n_s32(vsubq_s32(s_hi[1], s_hi[9]), DCT_CONST_BITS);
  t_lo[10] = vrshrq_n_s32(vsubq_s32(s_lo[2], s_lo[10]), DCT_CONST_BITS);
  t_hi[10] = vrshrq_n_s32(vsubq_s32(s_hi[2], s_hi[10]), DCT_CONST_BITS);
  t_lo[11] = vrshrq_n_s32(vsubq_s32(s_lo[3], s_lo[11]), DCT_CONST_BITS);
  t_hi[11] = vrshrq_n_s32(vsubq_s32(s_hi[3], s_hi[11]), DCT_CONST_BITS);
  t_lo[12] = vrshrq_n_s32(vsubq_s32(s_lo[4], s_lo[12]), DCT_CONST_BITS);
  t_hi[12] = vrshrq_n_s32(vsubq_s32(s_hi[4], s_hi[12]), DCT_CONST_BITS);
  t_lo[13] = vrshrq_n_s32(vsubq_s32(s_lo[5], s_lo[13]), DCT_CONST_BITS);
  t_hi[13] = vrshrq_n_s32(vsubq_s32(s_hi[5], s_hi[13]), DCT_CONST_BITS);
  t_lo[14] = vrshrq_n_s32(vsubq_s32(s_lo[6], s_lo[14]), DCT_CONST_BITS);
  t_hi[14] = vrshrq_n_s32(vsubq_s32(s_hi[6], s_hi[14]), DCT_CONST_BITS);
  t_lo[15] = vrshrq_n_s32(vsubq_s32(s_lo[7], s_lo[15]), DCT_CONST_BITS);
  t_hi[15] = vrshrq_n_s32(vsubq_s32(s_hi[7], s_hi[15]), DCT_CONST_BITS);

  // stage 2
  s_lo[0] = t_lo[0];
  s_hi[0] = t_hi[0];
  s_lo[1] = t_lo[1];
  s_hi[1] = t_hi[1];
  s_lo[2] = t_lo[2];
  s_hi[2] = t_hi[2];
  s_lo[3] = t_lo[3];
  s_hi[3] = t_hi[3];
  s_lo[4] = t_lo[4];
  s_hi[4] = t_hi[4];
  s_lo[5] = t_lo[5];
  s_hi[5] = t_hi[5];
  s_lo[6] = t_lo[6];
  s_hi[6] = t_hi[6];
  s_lo[7] = t_lo[7];
  s_hi[7] = t_hi[7];
  // s8 = x8 * cospi_4_64 + x9 * cospi_28_64;
  // s9 = x8 * cospi_28_64 - x9 * cospi_4_64;
  butterfly_two_coeff_s32_noround(t_lo[8], t_hi[8], t_lo[9], t_hi[9],
                                  cospi_4_64, cospi_28_64, &s_lo[8], &s_hi[8],
                                  &s_lo[9], &s_hi[9]);
  // s10 = x10 * cospi_20_64 + x11 * cospi_12_64;
  // s11 = x10 * cospi_12_64 - x11 * cospi_20_64;
  butterfly_two_coeff_s32_noround(t_lo[10], t_hi[10], t_lo[11], t_hi[11],
                                  cospi_20_64, cospi_12_64, &s_lo[10],
                                  &s_hi[10], &s_lo[11], &s_hi[11]);
  // s12 = -x12 * cospi_28_64 + x13 * cospi_4_64;
  // s13 = x12 * cospi_4_64 + x13 * cospi_28_64;
  butterfly_two_coeff_s32_noround(t_lo[13], t_hi[13], t_lo[12], t_hi[12],
                                  cospi_28_64, cospi_4_64, &s_lo[13], &s_hi[13],
                                  &s_lo[12], &s_hi[12]);
  // s14 = -x14 * cospi_12_64 + x15 * cospi_20_64;
  // s15 = x14 * cospi_20_64 + x15 * cospi_12_64;
  butterfly_two_coeff_s32_noround(t_lo[15], t_hi[15], t_lo[14], t_hi[14],
                                  cospi_12_64, cospi_20_64, &s_lo[15],
                                  &s_hi[15], &s_lo[14], &s_hi[14]);

  // s0 + s4
  t_lo[0] = vaddq_s32(s_lo[0], s_lo[4]);
  t_hi[0] = vaddq_s32(s_hi[0], s_hi[4]);
  // s1 + s5
  t_lo[1] = vaddq_s32(s_lo[1], s_lo[5]);
  t_hi[1] = vaddq_s32(s_hi[1], s_hi[5]);
  // s2 + s6
  t_lo[2] = vaddq_s32(s_lo[2], s_lo[6]);
  t_hi[2] = vaddq_s32(s_hi[2], s_hi[6]);
  // s3 + s7
  t_lo[3] = vaddq_s32(s_lo[3], s_lo[7]);
  t_hi[3] = vaddq_s32(s_hi[3], s_hi[7]);
  // s0 - s4
  t_lo[4] = vsubq_s32(s_lo[0], s_lo[4]);
  t_hi[4] = vsubq_s32(s_hi[0], s_hi[4]);
  // s1 - s7
  t_lo[5] = vsubq_s32(s_lo[1], s_lo[5]);
  t_hi[5] = vsubq_s32(s_hi[1], s_hi[5]);
  // s2 - s6
  t_lo[6] = vsubq_s32(s_lo[2], s_lo[6]);
  t_hi[6] = vsubq_s32(s_hi[2], s_hi[6]);
  // s3 - s7
  t_lo[7] = vsubq_s32(s_lo[3], s_lo[7]);
  t_hi[7] = vsubq_s32(s_hi[3], s_hi[7]);
  // s8 + s12
  t_lo[8] = vaddq_s32(s_lo[8], s_lo[12]);
  t_hi[8] = vaddq_s32(s_hi[8], s_hi[12]);
  // s9 + s13
  t_lo[9] = vaddq_s32(s_lo[9], s_lo[13]);
  t_hi[9] = vaddq_s32(s_hi[9], s_hi[13]);
  // s10 + s14
  t_lo[10] = vaddq_s32(s_lo[10], s_lo[14]);
  t_hi[10] = vaddq_s32(s_hi[10], s_hi[14]);
  // s11 + s15
  t_lo[11] = vaddq_s32(s_lo[11], s_lo[15]);
  t_hi[11] = vaddq_s32(s_hi[11], s_hi[15]);
  // s8 + s12
  t_lo[12] = vsubq_s32(s_lo[8], s_lo[12]);
  t_hi[12] = vsubq_s32(s_hi[8], s_hi[12]);
  // s9 + s13
  t_lo[13] = vsubq_s32(s_lo[9], s_lo[13]);
  t_hi[13] = vsubq_s32(s_hi[9], s_hi[13]);
  // s10 + s14
  t_lo[14] = vsubq_s32(s_lo[10], s_lo[14]);
  t_hi[14] = vsubq_s32(s_hi[10], s_hi[14]);
  // s11 + s15
  t_lo[15] = vsubq_s32(s_lo[11], s_lo[15]);
  t_hi[15] = vsubq_s32(s_hi[11], s_hi[15]);

  t_lo[8] = vrshrq_n_s32(t_lo[8], DCT_CONST_BITS);
  t_hi[8] = vrshrq_n_s32(t_hi[8], DCT_CONST_BITS);
  t_lo[9] = vrshrq_n_s32(t_lo[9], DCT_CONST_BITS);
  t_hi[9] = vrshrq_n_s32(t_hi[9], DCT_CONST_BITS);
  t_lo[10] = vrshrq_n_s32(t_lo[10], DCT_CONST_BITS);
  t_hi[10] = vrshrq_n_s32(t_hi[10], DCT_CONST_BITS);
  t_lo[11] = vrshrq_n_s32(t_lo[11], DCT_CONST_BITS);
  t_hi[11] = vrshrq_n_s32(t_hi[11], DCT_CONST_BITS);
  t_lo[12] = vrshrq_n_s32(t_lo[12], DCT_CONST_BITS);
  t_hi[12] = vrshrq_n_s32(t_hi[12], DCT_CONST_BITS);
  t_lo[13] = vrshrq_n_s32(t_lo[13], DCT_CONST_BITS);
  t_hi[13] = vrshrq_n_s32(t_hi[13], DCT_CONST_BITS);
  t_lo[14] = vrshrq_n_s32(t_lo[14], DCT_CONST_BITS);
  t_hi[14] = vrshrq_n_s32(t_hi[14], DCT_CONST_BITS);
  t_lo[15] = vrshrq_n_s32(t_lo[15], DCT_CONST_BITS);
  t_hi[15] = vrshrq_n_s32(t_hi[15], DCT_CONST_BITS);

  // stage 3
  s_lo[0] = t_lo[0];
  s_hi[0] = t_hi[0];
  s_lo[1] = t_lo[1];
  s_hi[1] = t_hi[1];
  s_lo[2] = t_lo[2];
  s_hi[2] = t_hi[2];
  s_lo[3] = t_lo[3];
  s_hi[3] = t_hi[3];
  // s4 = x4 * cospi_8_64 + x5 * cospi_24_64;
  // s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
  butterfly_two_coeff_s32_noround(t_lo[4], t_hi[4], t_lo[5], t_hi[5],
                                  cospi_8_64, cospi_24_64, &s_lo[4], &s_hi[4],
                                  &s_lo[5], &s_hi[5]);
  // s6 = -x6 * cospi_24_64 + x7 * cospi_8_64;
  // s7 = x6 * cospi_8_64 + x7 * cospi_24_64;
  butterfly_two_coeff_s32_noround(t_lo[7], t_hi[7], t_lo[6], t_hi[6],
                                  cospi_24_64, cospi_8_64, &s_lo[7], &s_hi[7],
                                  &s_lo[6], &s_hi[6]);
  s_lo[8] = t_lo[8];
  s_hi[8] = t_hi[8];
  s_lo[9] = t_lo[9];
  s_hi[9] = t_hi[9];
  s_lo[10] = t_lo[10];
  s_hi[10] = t_hi[10];
  s_lo[11] = t_lo[11];
  s_hi[11] = t_hi[11];
  // s12 = x12 * cospi_8_64 + x13 * cospi_24_64;
  // s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
  butterfly_two_coeff_s32_noround(t_lo[12], t_hi[12], t_lo[13], t_hi[13],
                                  cospi_8_64, cospi_24_64, &s_lo[12], &s_hi[12],
                                  &s_lo[13], &s_hi[13]);
  // s14 = -x14 * cospi_24_64 + x15 * cospi_8_64;
  // s15 = x14 * cospi_8_64 + x15 * cospi_24_64;
  butterfly_two_coeff_s32_noround(t_lo[15], t_hi[15], t_lo[14], t_hi[14],
                                  cospi_24_64, cospi_8_64, &s_lo[15], &s_hi[15],
                                  &s_lo[14], &s_hi[14]);

  // s0 + s4
  t_lo[0] = vaddq_s32(s_lo[0], s_lo[2]);
  t_hi[0] = vaddq_s32(s_hi[0], s_hi[2]);
  // s1 + s3
  t_lo[1] = vaddq_s32(s_lo[1], s_lo[3]);
  t_hi[1] = vaddq_s32(s_hi[1], s_hi[3]);
  // s0 - s4
  t_lo[2] = vsubq_s32(s_lo[0], s_lo[2]);
  t_hi[2] = vsubq_s32(s_hi[0], s_hi[2]);
  // s1 - s3
  t_lo[3] = vsubq_s32(s_lo[1], s_lo[3]);
  t_hi[3] = vsubq_s32(s_hi[1], s_hi[3]);
  // s4 + s6
  t_lo[4] = vaddq_s32(s_lo[4], s_lo[6]);
  t_hi[4] = vaddq_s32(s_hi[4], s_hi[6]);
  // s5 + s7
  t_lo[5] = vaddq_s32(s_lo[5], s_lo[7]);
  t_hi[5] = vaddq_s32(s_hi[5], s_hi[7]);
  // s4 - s6
  t_lo[6] = vsubq_s32(s_lo[4], s_lo[6]);
  t_hi[6] = vsubq_s32(s_hi[4], s_hi[6]);
  // s5 - s7
  t_lo[7] = vsubq_s32(s_lo[5], s_lo[7]);
  t_hi[7] = vsubq_s32(s_hi[5], s_hi[7]);
  // s8 + s10
  t_lo[8] = vaddq_s32(s_lo[8], s_lo[10]);
  t_hi[8] = vaddq_s32(s_hi[8], s_hi[10]);
  // s9 + s11
  t_lo[9] = vaddq_s32(s_lo[9], s_lo[11]);
  t_hi[9] = vaddq_s32(s_hi[9], s_hi[11]);
  // s8 - s10
  t_lo[10] = vsubq_s32(s_lo[8], s_lo[10]);
  t_hi[10] = vsubq_s32(s_hi[8], s_hi[10]);
  // s9 - s11
  t_lo[11] = vsubq_s32(s_lo[9], s_lo[11]);
  t_hi[11] = vsubq_s32(s_hi[9], s_hi[11]);
  // s12 + s14
  t_lo[12] = vaddq_s32(s_lo[12], s_lo[14]);
  t_hi[12] = vaddq_s32(s_hi[12], s_hi[14]);
  // s13 + s15
  t_lo[13] = vaddq_s32(s_lo[13], s_lo[15]);
  t_hi[13] = vaddq_s32(s_hi[13], s_hi[15]);
  // s12 - s14
  t_lo[14] = vsubq_s32(s_lo[12], s_lo[14]);
  t_hi[14] = vsubq_s32(s_hi[12], s_hi[14]);
  // s13 - s15
  t_lo[15] = vsubq_s32(s_lo[13], s_lo[15]);
  t_hi[15] = vsubq_s32(s_hi[13], s_hi[15]);

  t_lo[4] = vrshrq_n_s32(t_lo[4], DCT_CONST_BITS);
  t_hi[4] = vrshrq_n_s32(t_hi[4], DCT_CONST_BITS);
  t_lo[5] = vrshrq_n_s32(t_lo[5], DCT_CONST_BITS);
  t_hi[5] = vrshrq_n_s32(t_hi[5], DCT_CONST_BITS);
  t_lo[6] = vrshrq_n_s32(t_lo[6], DCT_CONST_BITS);
  t_hi[6] = vrshrq_n_s32(t_hi[6], DCT_CONST_BITS);
  t_lo[7] = vrshrq_n_s32(t_lo[7], DCT_CONST_BITS);
  t_hi[7] = vrshrq_n_s32(t_hi[7], DCT_CONST_BITS);
  t_lo[12] = vrshrq_n_s32(t_lo[12], DCT_CONST_BITS);
  t_hi[12] = vrshrq_n_s32(t_hi[12], DCT_CONST_BITS);
  t_lo[13] = vrshrq_n_s32(t_lo[13], DCT_CONST_BITS);
  t_hi[13] = vrshrq_n_s32(t_hi[13], DCT_CONST_BITS);
  t_lo[14] = vrshrq_n_s32(t_lo[14], DCT_CONST_BITS);
  t_hi[14] = vrshrq_n_s32(t_hi[14], DCT_CONST_BITS);
  t_lo[15] = vrshrq_n_s32(t_lo[15], DCT_CONST_BITS);
  t_hi[15] = vrshrq_n_s32(t_hi[15], DCT_CONST_BITS);

  // stage 4
  // s2 = (-cospi_16_64) * (x2 + x3);
  // s3 = cospi_16_64 * (x2 - x3);
  butterfly_one_coeff_s32_noround(t_lo[3], t_hi[3], t_lo[2], t_hi[2],
                                  -cospi_16_64, &s_lo[2], &s_hi[2], &s_lo[e='color: green'>3],
                                  &s_hi[3]);
  // s6 = cospi_16_64 * (x6 + x7);
  // s7 = cospi_16_64 * (-x6 + x7);
  butterfly_one_coeff_s32_noround(t_lo[7], t_hi[7], t_lo[6], t_hi[6],
                                  cospi_16_64, &s_lo[6], &s_hi[6], &s_lo[='color: green'>7],
                                  &s_hi[7]);
  // s10 = cospi_16_64 * (x10 + x11);
  // s11 = cospi_16_64 * (-x10 + x11);
  butterfly_one_coeff_s32_noround(t_lo[11], t_hi[11], t_lo[10], t_hi[10],
                                  cospi_16_64, &s_lo[10], &s_hi[10], &s_lo[le='color: green'>11],
                                  &s_hi[11]);
  // s14 = (-cospi_16_64) * (x14 + x15);
  // s15 = cospi_16_64 * (x14 - x15);
  butterfly_one_coeff_s32_noround(t_lo[15], t_hi[15], t_lo[14], t_hi[14],
                                  -cospi_16_64, &s_lo[14], &s_hi[14], &s_lo[yle='color: green'>15],
                                  &s_hi[15]);

  // final fdct_round_shift
  x_lo[2] = vrshrn_n_s32(s_lo[2], DCT_CONST_BITS);
  x_hi[2] = vrshrn_n_s32(s_hi[2], DCT_CONST_BITS);
  x_lo[3] = vrshrn_n_s32(s_lo[3], DCT_CONST_BITS);
  x_hi[3] = vrshrn_n_s32(s_hi[3], DCT_CONST_BITS);
  x_lo[6] = vrshrn_n_s32(s_lo[6], DCT_CONST_BITS);
  x_hi[6] = vrshrn_n_s32(s_hi[6], DCT_CONST_BITS);
  x_lo[7] = vrshrn_n_s32(s_lo[7], DCT_CONST_BITS);
  x_hi[7] = vrshrn_n_s32(s_hi[7], DCT_CONST_BITS);
  x_lo[10] = vrshrn_n_s32(s_lo[10], DCT_CONST_BITS);
  x_hi[10] = vrshrn_n_s32(s_hi[10], DCT_CONST_BITS);
  x_lo[11] = vrshrn_n_s32(s_lo[11], DCT_CONST_BITS);
  x_hi[11] = vrshrn_n_s32(s_hi[11], DCT_CONST_BITS);
  x_lo[14] = vrshrn_n_s32(s_lo[14], DCT_CONST_BITS);
  x_hi[14] = vrshrn_n_s32(s_hi[14], DCT_CONST_BITS);
  x_lo[15] = vrshrn_n_s32(s_lo[15], DCT_CONST_BITS);
  x_hi[15] = vrshrn_n_s32(s_hi[15], DCT_CONST_BITS);

  // x0, x1, x4, x5, x8, x9, x12, x13 narrow down to 16-bits directly
  x_lo[0] = vmovn_s32(t_lo[0]);
  x_hi[0] = vmovn_s32(t_hi[0]);
  x_lo[1] = vmovn_s32(t_lo[1]);
  x_hi[1] = vmovn_s32(t_hi[1]);
  x_lo[4] = vmovn_s32(t_lo[4]);
  x_hi[4] = vmovn_s32(t_hi[4]);
  x_lo[5] = vmovn_s32(t_lo[5]);
  x_hi[5] = vmovn_s32(t_hi[5]);
  x_lo[8] = vmovn_s32(t_lo[8]);
  x_hi[8] = vmovn_s32(t_hi[8]);
  x_lo[9] = vmovn_s32(t_lo[9]);
  x_hi[9] = vmovn_s32(t_hi[9]);
  x_lo[12] = vmovn_s32(t_lo[12]);
  x_hi[12] = vmovn_s32(t_hi[12]);
  x_lo[13] = vmovn_s32(t_lo[13]);
  x_hi[13] = vmovn_s32(t_hi[13]);

  in[0] = vcombine_s16(x_lo[0], x_hi[0]);
  in[1] = vnegq_s16(vcombine_s16(x_lo[8], x_hi[8]));
  in[2] = vcombine_s16(x_lo[12], x_hi[12]);
  in[3] = vnegq_s16(vcombine_s16(x_lo[4], x_hi[4]));
  in[4] = vcombine_s16(x_lo[6], x_hi[6]);
  in[5] = vcombine_s16(x_lo[14], x_hi[14]);
  in[6] = vcombine_s16(x_lo[10], x_hi[10]);
  in[7] = vcombine_s16(x_lo[2], x_hi[2]);
  in[8] = vcombine_s16(x_lo[3], x_hi[3]);
  in[9] = vcombine_s16(x_lo[11], x_hi[11]);
  in[10] = vcombine_s16(x_lo[15], x_hi[15]);
  in[11] = vcombine_s16(x_lo[7], x_hi[7]);
  in[12] = vcombine_s16(x_lo[5], x_hi[5]);
  in[13] = vnegq_s16(vcombine_s16(x_lo[13], x_hi[13]));
  in[14] = vcombine_s16(x_lo[9], x_hi[9]);
  in[15] = vnegq_s16(vcombine_s16(x_lo[1], x_hi[1]));
}

static void fdct16x16_neon(int16x8_t *in0, int16x8_t *in1) {
  // Left half.
  fdct16_8col(in0);
  // Right half.
  fdct16_8col(in1);
  transpose_s16_16x16(in0, in1);
}

static void fadst16x16_neon(int16x8_t *in0, int16x8_t *in1) {
  fadst16_8col(in0);
  fadst16_8col(in1);
  transpose_s16_16x16(in0, in1);
}

void vp9_fht16x16_neon(const int16_t *input, tran_low_t *output, int stride,
                       int tx_type) {
  int16x8_t in0[16], in1[16];

  switch (tx_type) {
    case DCT_DCT: vpx_fdct16x16_neon(input, output, stride); break;
    case ADST_DCT:
      load_buffer_16x16(input, in0, in1, stride);
      fadst16x16_neon(in0, in1);
      right_shift_16x16(in0, in1);
      fdct16x16_neon(in0, in1);
      write_buffer_16x16(output, in0, in1, 16);
      break;
    case DCT_ADST:
      load_buffer_16x16(input, in0, in1, stride);
      fdct16x16_neon(in0, in1);
      right_shift_16x16(in0, in1);
      fadst16x16_neon(in0, in1);
      write_buffer_16x16(output, in0, in1, 16);
      break;
    default:
      assert(tx_type == ADST_ADST);
      load_buffer_16x16(input, in0, in1, stride);
      fadst16x16_neon(in0, in1);
      right_shift_16x16(in0, in1);
      fadst16x16_neon(in0, in1);
      write_buffer_16x16(output, in0, in1, 16);
      break;
  }
}

#if CONFIG_VP9_HIGHBITDEPTH

static INLINE void highbd_load_buffer_4x4(const int16_t *input,
                                          int32x4_t *in /*[4]*/, int stride) {
  // { 0, 1, 1, 1 };
  const int32x4_t nonzero_bias_a = vextq_s32(vdupq_n_s32(0), vdupq_n_s32(1), 3);
  // { 1, 0, 0, 0 };
  const int32x4_t nonzero_bias_b = vextq_s32(vdupq_n_s32(1), vdupq_n_s32(0), 3);
  int32x4_t mask;

  in[0] = vshll_n_s16(vld1_s16(input + 0 * stride), 4);
  in[1] = vshll_n_s16(vld1_s16(input + 1 * stride), 4);
  in[2] = vshll_n_s16(vld1_s16(input + 2 * stride), 4);
  in[3] = vshll_n_s16(vld1_s16(input + 3 * stride), 4);

  // Copy the SSE method, use a mask to avoid an 'if' branch here to increase by
  // one non-zero first elements
  mask = vreinterpretq_s32_u32(vceqq_s32(in[0], nonzero_bias_a));
  in[0] = vaddq_s32(in[0], mask);
  in[0] = vaddq_s32(in[0], nonzero_bias_b);
}

static INLINE void highbd_write_buffer_4x4(tran_low_t *output, int32x4_t *res) {
  const int32x4_t one = vdupq_n_s32(1);
  res[0] = vshrq_n_s32(vaddq_s32(res[0], one), 2);
  res[1] = vshrq_n_s32(vaddq_s32(res[1], one), 2);
  res[2] = vshrq_n_s32(vaddq_s32(res[2], one), 2);
  res[3] = vshrq_n_s32(vaddq_s32(res[3], one), 2);
  vst1q_s32(output + 0 * 4, res[0]);
  vst1q_s32(output + 1 * 4, res[1]);
  vst1q_s32(output + 2 * 4, res[2]);
  vst1q_s32(output + 3 * 4, res[3]);
}

static INLINE void highbd_fadst4x4_neon(int32x4_t *in /*[4]*/) {
  int32x2_t s_lo[4], s_hi[4];
  int64x2_t u_lo[4], u_hi[4], t_lo[4], t_hi[4];

  s_lo[0] = vget_low_s32(in[0]);
  s_hi[0] = vget_high_s32(in[0]);
  s_lo[1] = vget_low_s32(in[1]);
  s_hi[1] = vget_high_s32(in[1]);
  s_lo[2] = vget_low_s32(in[2]);
  s_hi[2] = vget_high_s32(in[2]);
  s_lo[3] = vget_low_s32(in[3]);
  s_hi[3] = vget_high_s32(in[3]);

  // t0 = s0 * sinpi_1_9 + s1 * sinpi_2_9 + s3 * sinpi_4_9
  t_lo[0] = vmull_n_s32(s_lo[0], sinpi_1_9);
  t_lo[0] = vmlal_n_s32(t_lo[0], s_lo[1], sinpi_2_9);
  t_lo[0] = vmlal_n_s32(t_lo[0], s_lo[3], sinpi_4_9);
  t_hi[0] = vmull_n_s32(s_hi[0], sinpi_1_9);
  t_hi[0] = vmlal_n_s32(t_hi[0], s_hi[1], sinpi_2_9);
  t_hi[0] = vmlal_n_s32(t_hi[0], s_hi[3], sinpi_4_9);

  // t1 = (s0 + s1) * sinpi_3_9 - s3 * sinpi_3_9
  t_lo[1] = vmull_n_s32(s_lo[0], sinpi_3_9);
  t_lo[1] = vmlal_n_s32(t_lo[1], s_lo[1], sinpi_3_9);
  t_lo[1] = vmlsl_n_s32(t_lo[1], s_lo[3], sinpi_3_9);
  t_hi[1] = vmull_n_s32(s_hi[0], sinpi_3_9);
  t_hi[1] = vmlal_n_s32(t_hi[1], s_hi[1], sinpi_3_9);
  t_hi[1] = vmlsl_n_s32(t_hi[1], s_hi[3], sinpi_3_9);

  // t2 = s0 * sinpi_4_9 - s1* sinpi_1_9 + s3 * sinpi_2_9
  t_lo[2] = vmull_n_s32(s_lo[0], sinpi_4_9);
  t_lo[2] = vmlsl_n_s32(t_lo[2], s_lo[1], sinpi_1_9);
  t_lo[2] = vmlal_n_s32(t_lo[2], s_lo[3], sinpi_2_9);
  t_hi[2] = vmull_n_s32(s_hi[0], sinpi_4_9);
  t_hi[2] = vmlsl_n_s32(t_hi[2], s_hi[1], sinpi_1_9);
  t_hi[2] = vmlal_n_s32(t_hi[2], s_hi[3], sinpi_2_9);

  // t3 = s2 * sinpi_3_9
  t_lo[3] = vmull_n_s32(s_lo[2], sinpi_3_9);
  t_hi[3] = vmull_n_s32(s_hi[2], sinpi_3_9);

  /*
   * u0 = t0 + t3
   * u1 = t1
   * u2 = t2 - t3
   * u3 = t2 - t0 + t3
   */
  u_lo[0] = vaddq_s64(t_lo[0], t_lo[3]);
  u_hi[0] = vaddq_s64(t_hi[0], t_hi[3]);
  u_lo[1] = t_lo[1];
  u_hi[1] = t_hi[1];
  u_lo[2] = vsubq_s64(t_lo[2], t_lo[3]);
  u_hi[2] = vsubq_s64(t_hi[2], t_hi[3]);
  u_lo[3] = vaddq_s64(vsubq_s64(t_lo[2], t_lo[0]), t_lo[3]);
  u_hi[3] = vaddq_s64(vsubq_s64(t_hi[2], t_hi[0]), t_hi[3]);

  // fdct_round_shift
  in[0] = vcombine_s32(vrshrn_n_s64(u_lo[0], DCT_CONST_BITS),
                       vrshrn_n_s64(u_hi[0], DCT_CONST_BITS));
  in[1] = vcombine_s32(vrshrn_n_s64(u_lo[1], DCT_CONST_BITS),
                       vrshrn_n_s64(u_hi[1], DCT_CONST_BITS));
  in[2] = vcombine_s32(vrshrn_n_s64(u_lo[2], DCT_CONST_BITS),
                       vrshrn_n_s64(u_hi[2], DCT_CONST_BITS));
  in[3] = vcombine_s32(vrshrn_n_s64(u_lo[3], DCT_CONST_BITS),
                       vrshrn_n_s64(u_hi[3], DCT_CONST_BITS));

  transpose_s32_4x4(&in[0], &in[1], &in[ style='color: green'>2], &in[3]);
}

void vp9_highbd_fht4x4_neon(const int16_t *input, tran_low_t *output,
                            int stride, int tx_type) {
  int32x4_t in[4];
  // int i;

  switch (tx_type) {
    case DCT_DCT: vpx_highbd_fdct4x4_neon(input, output, stride); break;
    case ADST_DCT:
      highbd_load_buffer_4x4(input, in, stride);
      highbd_fadst4x4_neon(in);
      vpx_highbd_fdct4x4_pass1_neon(in);
      highbd_write_buffer_4x4(output, in);
      break;
    case DCT_ADST:
      highbd_load_buffer_4x4(input, in, stride);
      vpx_highbd_fdct4x4_pass1_neon(in);
      highbd_fadst4x4_neon(in);
      highbd_write_buffer_4x4(output, in);
      break;
    default:
      assert(tx_type == ADST_ADST);
      highbd_load_buffer_4x4(input, in, stride);
      highbd_fadst4x4_neon(in);
      highbd_fadst4x4_neon(in);
      highbd_write_buffer_4x4(output, in);
      break;
  }
}

static INLINE void highbd_load_buffer_8x8(const int16_t *input,
                                          int32x4_t *lo /*[8]*/,
                                          int32x4_t *hi /*[8]*/, int stride) {
  int16x8_t in[8];
  in[0] = vld1q_s16(input + 0 * stride);
  in[1] = vld1q_s16(input + 1 * stride);
  in[2] = vld1q_s16(input + 2 * stride);
  in[3] = vld1q_s16(input + 3 * stride);
  in[4] = vld1q_s16(input + 4 * stride);
  in[5] = vld1q_s16(input + 5 * stride);
  in[6] = vld1q_s16(input + 6 * stride);
  in[7] = vld1q_s16(input + 7 * stride);
  lo[0] = vshll_n_s16(vget_low_s16(in[0]), 2);
  hi[0] = vshll_n_s16(vget_high_s16(in[0]), 2);
  lo[1] = vshll_n_s16(vget_low_s16(in[1]), 2);
  hi[1] = vshll_n_s16(vget_high_s16(in[1]), 2);
  lo[2] = vshll_n_s16(vget_low_s16(in[2]), 2);
  hi[2] = vshll_n_s16(vget_high_s16(in[2]), 2);
  lo[3] = vshll_n_s16(vget_low_s16(in[3]), 2);
  hi[3] = vshll_n_s16(vget_high_s16(in[3]), 2);
  lo[4] = vshll_n_s16(vget_low_s16(in[4]), 2);
  hi[4] = vshll_n_s16(vget_high_s16(in[4]), 2);
  lo[5] = vshll_n_s16(vget_low_s16(in[5]), 2);
  hi[5] = vshll_n_s16(vget_high_s16(in[5]), 2);
  lo[6] = vshll_n_s16(vget_low_s16(in[6]), 2);
  hi[6] = vshll_n_s16(vget_high_s16(in[6]), 2);
  lo[7] = vshll_n_s16(vget_low_s16(in[7]), 2);
  hi[7] = vshll_n_s16(vget_high_s16(in[7]), 2);
}

/* right shift and rounding
 * first get the sign bit (bit 15).
 * If bit == 1, it's the simple case of shifting right by one bit.
 * If bit == 2, it essentially computes the expression:
 *
 * out[j * 16 + i] = (temp_out[j] + 1 + (temp_out[j] < 0)) >> 2;
 *
 * for each row.
 */
static INLINE void highbd_right_shift_8x8(int32x4_t *lo, int32x4_t *hi,
                                          const int bit) {
  int32x4_t sign_lo[8], sign_hi[8];
  sign_lo[0] = vshrq_n_s32(lo[0], 31);
  sign_hi[0] = vshrq_n_s32(hi[0], 31);
  sign_lo[1] = vshrq_n_s32(lo[1], 31);
  sign_hi[1] = vshrq_n_s32(hi[1], 31);
  sign_lo[2] = vshrq_n_s32(lo[2], 31);
  sign_hi[2] = vshrq_n_s32(hi[2], 31);
  sign_lo[3] = vshrq_n_s32(lo[3], 31);
  sign_hi[3] = vshrq_n_s32(hi[3], 31);
  sign_lo[4] = vshrq_n_s32(lo[4], 31);
  sign_hi[4] = vshrq_n_s32(hi[4], 31);
  sign_lo[5] = vshrq_n_s32(lo[5], 31);
  sign_hi[5] = vshrq_n_s32(hi[5], 31);
  sign_lo[6] = vshrq_n_s32(lo[6], 31);
  sign_hi[6] = vshrq_n_s32(hi[6], 31);
  sign_lo[7] = vshrq_n_s32(lo[7], 31);
  sign_hi[7] = vshrq_n_s32(hi[7], 31);

  if (bit == 2) {
    const int32x4_t const_rounding = vdupq_n_s32(1);
    lo[0] = vaddq_s32(lo[0], const_rounding);
    hi[0] = vaddq_s32(hi[0], const_rounding);
    lo[1] = vaddq_s32(lo[1], const_rounding);
    hi[1] = vaddq_s32(hi[1], const_rounding);
    lo[2] = vaddq_s32(lo[2], const_rounding);
    hi[2] = vaddq_s32(hi[2], const_rounding);
    lo[3] = vaddq_s32(lo[3], const_rounding);
    hi[3] = vaddq_s32(hi[3], const_rounding);
    lo[4] = vaddq_s32(lo[4], const_rounding);
    hi[4] = vaddq_s32(hi[4], const_rounding);
    lo[5] = vaddq_s32(lo[5], const_rounding);
    hi[5] = vaddq_s32(hi[5], const_rounding);
    lo[6] = vaddq_s32(lo[6], const_rounding);
    hi[6] = vaddq_s32(hi[6], const_rounding);
    lo[7] = vaddq_s32(lo[7], const_rounding);
    hi[7] = vaddq_s32(hi[7], const_rounding);
  }

  lo[0] = vsubq_s32(lo[0], sign_lo[0]);
  hi[0] = vsubq_s32(hi[0], sign_hi[0]);
  lo[1] = vsubq_s32(lo[1], sign_lo[1]);
  hi[1] = vsubq_s32(hi[1], sign_hi[1]);
  lo[2] = vsubq_s32(lo[2], sign_lo[2]);
  hi[2] = vsubq_s32(hi[2], sign_hi[2]);
  lo[3] = vsubq_s32(lo[3], sign_lo[3]);
  hi[3] = vsubq_s32(hi[3], sign_hi[3]);
  lo[4] = vsubq_s32(lo[4], sign_lo[4]);
  hi[4] = vsubq_s32(hi[4], sign_hi[4]);
  lo[5] = vsubq_s32(lo[5], sign_lo[5]);
  hi[5] = vsubq_s32(hi[5], sign_hi[5]);
  lo[6] = vsubq_s32(lo[6], sign_lo[6]);
  hi[6] = vsubq_s32(hi[6], sign_hi[6]);
  lo[7] = vsubq_s32(lo[7], sign_lo[7]);
  hi[7] = vsubq_s32(hi[7], sign_hi[7]);

  if (bit == 1) {
    lo[0] = vshrq_n_s32(lo[0], 1);
    hi[0] = vshrq_n_s32(hi[0], 1);
    lo[1] = vshrq_n_s32(lo[1], 1);
    hi[1] = vshrq_n_s32(hi[1], 1);
    lo[2] = vshrq_n_s32(lo[2], 1);
    hi[2] = vshrq_n_s32(hi[2], 1);
    lo[3] = vshrq_n_s32(lo[3], 1);
    hi[3] = vshrq_n_s32(hi[3], 1);
    lo[4] = vshrq_n_s32(lo[4], 1);
    hi[4] = vshrq_n_s32(hi[4], 1);
    lo[5] = vshrq_n_s32(lo[5], 1);
    hi[5] = vshrq_n_s32(hi[5], 1);
    lo[6] = vshrq_n_s32(lo[6], 1);
    hi[6] = vshrq_n_s32(hi[6], 1);
    lo[7] = vshrq_n_s32(lo[7], 1);
    hi[7] = vshrq_n_s32(hi[7], 1);
  } else {
    lo[0] = vshrq_n_s32(lo[0], 2);
    hi[0] = vshrq_n_s32(hi[0], 2);
    lo[1] = vshrq_n_s32(lo[1], 2);
    hi[1] = vshrq_n_s32(hi[1], 2);
    lo[2] = vshrq_n_s32(lo[2], 2);
    hi[2] = vshrq_n_s32(hi[2], 2);
    lo[3] = vshrq_n_s32(lo[3], 2);
    hi[3] = vshrq_n_s32(hi[3], 2);
    lo[4] = vshrq_n_s32(lo[4], 2);
    hi[4] = vshrq_n_s32(hi[4], 2);
    lo[5] = vshrq_n_s32(lo[5], 2);
    hi[5] = vshrq_n_s32(hi[5], 2);
    lo[6] = vshrq_n_s32(lo[6], 2);
    hi[6] = vshrq_n_s32(hi[6], 2);
    lo[7] = vshrq_n_s32(lo[7], 2);
    hi[7] = vshrq_n_s32(hi[7], 2);
  }
}

static INLINE void highbd_write_buffer_8x8(tran_low_t *output, int32x4_t *lo,
                                           int32x4_t *hi, int stride) {
  vst1q_s32(output + 0 * stride, lo[0]);
  vst1q_s32(output + 0 * stride + 4, hi[0]);
  vst1q_s32(output + 1 * stride, lo[1]);
  vst1q_s32(output + 1 * stride + 4, hi[1]);
  vst1q_s32(output + 2 * stride, lo[2]);
  vst1q_s32(output + 2 * stride + 4, hi[2]);
  vst1q_s32(output + 3 * stride, lo[3]);
  vst1q_s32(output + 3 * stride + 4, hi[3]);
  vst1q_s32(output + 4 * stride, lo[4]);
  vst1q_s32(output + 4 * stride + 4, hi[4]);
  vst1q_s32(output + 5 * stride, lo[5]);
  vst1q_s32(output + 5 * stride + 4, hi[5]);
  vst1q_s32(output + 6 * stride, lo[6]);
  vst1q_s32(output + 6 * stride + 4, hi[6]);
  vst1q_s32(output + 7 * stride, lo[7]);
  vst1q_s32(output + 7 * stride + 4, hi[7]);
}

static INLINE void highbd_fadst8x8_neon(int32x4_t *lo /*[8]*/,
                                        int32x4_t *hi /*[8]*/) {
  int32x4_t s_lo[8], s_hi[8];
  int32x4_t t_lo[8], t_hi[8];
  int32x4_t x_lo[8], x_hi[8];
  int64x2_t s64_lo[16], s64_hi[16];

  x_lo[0] = lo[7];
  x_hi[0] = hi[7];
  x_lo[1] = lo[0];
  x_hi[1] = hi[0];
  x_lo[2] = lo[5];
  x_hi[2] = hi[5];
  x_lo[3] = lo[2];
  x_hi[3] = hi[2];
  x_lo[4] = lo[3];
  x_hi[4] = hi[3];
  x_lo[5] = lo[4];
  x_hi[5] = hi[4];
  x_lo[6] = lo[1];
  x_hi[6] = hi[1];
  x_lo[7] = lo[6];
  x_hi[7] = hi[6];

  // stage 1
  // s0 = cospi_2_64 * x0 + cospi_30_64 * x1;
  // s1 = cospi_30_64 * x0 - cospi_2_64 * x1;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[0], x_hi[0], x_lo[1], x_hi[1], cospi_2_64, cospi_30_64,
      &s64_lo[2 * 0], &s64_hi[2 * 0], &s64_lo[2 * 1], &s64_hi[2 * 1]);
  // s2 = cospi_10_64 * x2 + cospi_22_64 * x3;
  // s3 = cospi_22_64 * x2 - cospi_10_64 * x3;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[2], x_hi[2], x_lo[3], x_hi[3], cospi_10_64, cospi_22_64,
      &s64_lo[2 * 2], &s64_hi[2 * 2], &s64_lo[2 * 3], &s64_hi[2 * 3]);

  // s4 = cospi_18_64 * x4 + cospi_14_64 * x5;
  // s5 = cospi_14_64 * x4 - cospi_18_64 * x5;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[4], x_hi[4], x_lo[5], x_hi[5], cospi_18_64, cospi_14_64,
      &s64_lo[2 * 4], &s64_hi[2 * 4], &s64_lo[2 * 5], &s64_hi[2 * 5]);

  // s6 = cospi_26_64 * x6 + cospi_6_64 * x7;
  // s7 = cospi_6_64 * x6 - cospi_26_64 * x7;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[6], x_hi[6], x_lo[7], x_hi[7], cospi_26_64, cospi_6_64,
      &s64_lo[2 * 6], &s64_hi[2 * 6], &s64_lo[2 * 7], &s64_hi[2 * 7]);

  // fdct_round_shift, indices are doubled
  t_lo[0] = add_s64_round_narrow(&s64_lo[2 * 0], &s64_lo[2 * 4]);
  t_hi[0] = add_s64_round_narrow(&s64_hi[2 * 0], &s64_hi[2 * 4]);
  t_lo[1] = add_s64_round_narrow(&s64_lo[2 * 1], &s64_lo[2 * 5]);
  t_hi[1] = add_s64_round_narrow(&s64_hi[2 * 1], &s64_hi[2 * 5]);
  t_lo[2] = add_s64_round_narrow(&s64_lo[2 * 2], &s64_lo[2 * 6]);
  t_hi[2] = add_s64_round_narrow(&s64_hi[2 * 2], &s64_hi[2 * 6]);
  t_lo[3] = add_s64_round_narrow(&s64_lo[2 * 3], &s64_lo[2 * 7]);
  t_hi[3] = add_s64_round_narrow(&s64_hi[2 * 3], &s64_hi[2 * 7]);
  t_lo[4] = sub_s64_round_narrow(&s64_lo[2 * 0], &s64_lo[2 * 4]);
  t_hi[4] = sub_s64_round_narrow(&s64_hi[2 * 0], &s64_hi[2 * 4]);
  t_lo[5] = sub_s64_round_narrow(&s64_lo[2 * 1], &s64_lo[2 * 5]);
  t_hi[5] = sub_s64_round_narrow(&s64_hi[2 * 1], &s64_hi[2 * 5]);
  t_lo[6] = sub_s64_round_narrow(&s64_lo[2 * 2], &s64_lo[2 * 6]);
  t_hi[6] = sub_s64_round_narrow(&s64_hi[2 * 2], &s64_hi[2 * 6]);
  t_lo[7] = sub_s64_round_narrow(&s64_lo[2 * 3], &s64_lo[2 * 7]);
  t_hi[7] = sub_s64_round_narrow(&s64_hi[2 * 3], &s64_hi[2 * 7]);

  // stage 2
  s_lo[0] = t_lo[0];
  s_hi[0] = t_hi[0];
  s_lo[1] = t_lo[1];
  s_hi[1] = t_hi[1];
  s_lo[2] = t_lo[2];
  s_hi[2] = t_hi[2];
  s_lo[3] = t_lo[3];
  s_hi[3] = t_hi[3];
  // s4 = cospi_8_64 * x4 + cospi_24_64 * x5;
  // s5 = cospi_24_64 * x4 - cospi_8_64 * x5;
  butterfly_two_coeff_s32_s64_noround(
      t_lo[4], t_hi[4], t_lo[5], t_hi[5], cospi_8_64, cospi_24_64,
      &s64_lo[2 * 4], &s64_hi[2 * 4], &s64_lo[2 * 5], &s64_hi[2 * 5]);

  // s6 = -cospi_24_64 * x6 + cospi_8_64 * x7;
  // s7 = cospi_8_64 * x6 + cospi_24_64 * x7;
  butterfly_two_coeff_s32_s64_noround(
      t_lo[6], t_hi[6], t_lo[7], t_hi[7], -cospi_24_64, cospi_8_64,
      &s64_lo[2 * 6], &s64_hi[2 * 6], &s64_lo[2 * 7], &s64_hi[2 * 7]);

  // fdct_round_shift
  // s0 + s2
  t_lo[0] = add_s32_s64_narrow(s_lo[0], s_lo[2]);
  t_hi[0] = add_s32_s64_narrow(s_hi[0], s_hi[2]);
  // s0 - s2
  t_lo[2] = sub_s32_s64_narrow(s_lo[0], s_lo[2]);
  t_hi[2] = sub_s32_s64_narrow(s_hi[0], s_hi[2]);

  // s1 + s3
  t_lo[1] = add_s32_s64_narrow(s_lo[1], s_lo[3]);
  t_hi[1] = add_s32_s64_narrow(s_hi[1], s_hi[3]);
  // s1 - s3
  t_lo[3] = sub_s32_s64_narrow(s_lo[1], s_lo[3]);
  t_hi[3] = sub_s32_s64_narrow(s_hi[1], s_hi[3]);

  // s4 + s6
  t_lo[4] = add_s64_round_narrow(&s64_lo[2 * 4], &s64_lo[2 * 6]);
  t_hi[4] = add_s64_round_narrow(&s64_hi[2 * 4], &s64_hi[2 * 6]);
  // s4 - s6
  t_lo[6] = sub_s64_round_narrow(&s64_lo[2 * 4], &s64_lo[2 * 6]);
  t_hi[6] = sub_s64_round_narrow(&s64_hi[2 * 4], &s64_hi[2 * 6]);

  // s5 + s7
  t_lo[5] = add_s64_round_narrow(&s64_lo[2 * 5], &s64_lo[2 * 7]);
  t_hi[5] = add_s64_round_narrow(&s64_hi[2 * 5], &s64_hi[2 * 7]);
  // s5 - s7
  t_lo[7] = sub_s64_round_narrow(&s64_lo[2 * 5], &s64_lo[2 * 7]);
  t_hi[7] = sub_s64_round_narrow(&s64_hi[2 * 5], &s64_hi[2 * 7]);

  // stage 3
  // s2 = cospi_16_64 * (x2 + x3)
  // s3 = cospi_16_64 * (x2 - x3)
  butterfly_one_coeff_s32_fast(t_lo[2], t_hi[2], t_lo[3], t_hi[3], cospi_16_64,
                               &s_lo[2], &s_hi[2], &s_lo[3], &s_hi[3]);

  // s6 = cospi_16_64 * (x6 + x7)
  // s7 = cospi_16_64 * (x6 - x7)
  butterfly_one_coeff_s32_fast(t_lo[6], t_hi[6], t_lo[7], t_hi[7], cospi_16_64,
                               &s_lo[6], &s_hi[6], &s_lo[7], &s_hi[7]);

  // x0, x2, x4, x6 pass through
  lo[0] = t_lo[0];
  hi[0] = t_hi[0];
  lo[2] = s_lo[6];
  hi[2] = s_hi[6];
  lo[4] = s_lo[3];
  hi[4] = s_hi[3];
  lo[6] = t_lo[5];
  hi[6] = t_hi[5];

  lo[1] = vnegq_s32(t_lo[4]);
  hi[1] = vnegq_s32(t_hi[4]);
  lo[3] = vnegq_s32(s_lo[2]);
  hi[3] = vnegq_s32(s_hi[2]);
  lo[5] = vnegq_s32(s_lo[7]);
  hi[5] = vnegq_s32(s_hi[7]);
  lo[7] = vnegq_s32(t_lo[1]);
  hi[7] = vnegq_s32(t_hi[1]);

  transpose_s32_8x8_2(lo, hi, lo, hi);
}

void vp9_highbd_fht8x8_neon(const int16_t *input, tran_low_t *output,
                            int stride, int tx_type) {
  int32x4_t lo[8], hi[8];

  switch (tx_type) {
    case DCT_DCT: vpx_highbd_fdct8x8_neon(input, output, stride); break;
    case ADST_DCT:
      highbd_load_buffer_8x8(input, lo, hi, stride);
      highbd_fadst8x8_neon(lo, hi);
      // pass1 variant is not precise enough
      vpx_highbd_fdct8x8_pass2_neon(lo, hi);
      highbd_right_shift_8x8(lo, hi, 1);
      highbd_write_buffer_8x8(output, lo, hi, 8);
      break;
    case DCT_ADST:
      highbd_load_buffer_8x8(input, lo, hi, stride);
      // pass1 variant is not precise enough
      vpx_highbd_fdct8x8_pass2_neon(lo, hi);
      highbd_fadst8x8_neon(lo, hi);
      highbd_right_shift_8x8(lo, hi, 1);
      highbd_write_buffer_8x8(output, lo, hi, 8);
      break;
    default:
      assert(tx_type == ADST_ADST);
      highbd_load_buffer_8x8(input, lo, hi, stride);
      highbd_fadst8x8_neon(lo, hi);
      highbd_fadst8x8_neon(lo, hi);
      highbd_right_shift_8x8(lo, hi, 1);
      highbd_write_buffer_8x8(output, lo, hi, 8);
      break;
  }
}

static INLINE void highbd_load_buffer_16x16(
    const int16_t *input, int32x4_t *left1 /*[16]*/, int32x4_t *right1 /*[16]*/,
    int32x4_t *left2 /*[16]*/, int32x4_t *right2 /*[16]*/, int stride) {
  // load first 8 columns
  highbd_load_buffer_8x8(input, left1, right1, stride);
  highbd_load_buffer_8x8(input + 8 * stride, left1 + 8, right1 + 8, stride);

  input += 8;
  // load second 8 columns
  highbd_load_buffer_8x8(input, left2, right2, stride);
  highbd_load_buffer_8x8(input + 8 * stride, left2 + 8, right2 + 8, stride);
}

static INLINE void highbd_write_buffer_16x16(
    tran_low_t *output, int32x4_t *left1 /*[16]*/, int32x4_t *right1 /*[16]*/,
    int32x4_t *left2 /*[16]*/, int32x4_t *right2 /*[16]*/, int stride) {
  // write first 8 columns
  highbd_write_buffer_8x8(output, left1, right1, stride);
  highbd_write_buffer_8x8(output + 8 * stride, left1 + 8, right1 + 8, stride);

  // write second 8 columns
  output += 8;
  highbd_write_buffer_8x8(output, left2, right2, stride);
  highbd_write_buffer_8x8(output + 8 * stride, left2 + 8, right2 + 8, stride);
}

static INLINE void highbd_right_shift_16x16(int32x4_t *left1 /*[16]*/,
                                            int32x4_t *right1 /*[16]*/,
                                            int32x4_t *left2 /*[16]*/,
                                            int32x4_t *right2 /*[16]*/,
                                            const int bit) {
  // perform rounding operations
  highbd_right_shift_8x8(left1, right1, bit);
  highbd_right_shift_8x8(left1 + 8, right1 + 8, bit);
  highbd_right_shift_8x8(left2, right2, bit);
  highbd_right_shift_8x8(left2 + 8, right2 + 8, bit);
}

static void highbd_fdct16_8col(int32x4_t *left, int32x4_t *right) {
  // perform 16x16 1-D DCT for 8 columns
  int32x4_t s1_lo[8], s1_hi[8], s2_lo[8], s2_hi[8], s3_lo[8], s3_hi[8];
  int32x4_t left8[8], right8[8];

  // stage 1
  left8[0] = vaddq_s32(left[0], left[15]);
  right8[0] = vaddq_s32(right[0], right[15]);
  left8[1] = vaddq_s32(left[1], left[14]);
  right8[1] = vaddq_s32(right[1], right[14]);
  left8[2] = vaddq_s32(left[2], left[13]);
  right8[2] = vaddq_s32(right[2], right[13]);
  left8[3] = vaddq_s32(left[3], left[12]);
  right8[3] = vaddq_s32(right[3], right[12]);
  left8[4] = vaddq_s32(left[4], left[11]);
  right8[4] = vaddq_s32(right[4], right[11]);
  left8[5] = vaddq_s32(left[5], left[10]);
  right8[5] = vaddq_s32(right[5], right[10]);
  left8[6] = vaddq_s32(left[6], left[9]);
  right8[6] = vaddq_s32(right[6], right[9]);
  left8[7] = vaddq_s32(left[7], left[8]);
  right8[7] = vaddq_s32(right[7], right[8]);

  // step 1
  s1_lo[0] = vsubq_s32(left[7], left[8]);
  s1_hi[0] = vsubq_s32(right[7], right[8]);
  s1_lo[1] = vsubq_s32(left[6], left[9]);
  s1_hi[1] = vsubq_s32(right[6], right[9]);
  s1_lo[2] = vsubq_s32(left[5], left[10]);
  s1_hi[2] = vsubq_s32(right[5], right[10]);
  s1_lo[3] = vsubq_s32(left[4], left[11]);
  s1_hi[3] = vsubq_s32(right[4], right[11]);
  s1_lo[4] = vsubq_s32(left[3], left[12]);
  s1_hi[4] = vsubq_s32(right[3], right[12]);
  s1_lo[5] = vsubq_s32(left[2], left[13]);
  s1_hi[5] = vsubq_s32(right[2], right[13]);
  s1_lo[6] = vsubq_s32(left[1], left[14]);
  s1_hi[6] = vsubq_s32(right[1], right[14]);
  s1_lo[7] = vsubq_s32(left[0], left[15]);
  s1_hi[7] = vsubq_s32(right[0], right[15]);

  // pass1 variant is not accurate enough
  vpx_highbd_fdct8x8_pass2_notranspose_neon(left8, right8);

  // step 2
  // step2[2] = (step1[5] - step1[2]) * cospi_16_64;
  // step2[5] = (step1[5] + step1[2]) * cospi_16_64;
  butterfly_one_coeff_s32_s64_narrow(s1_lo[5], s1_hi[5], s1_lo[2], s1_hi[2],
                                     cospi_16_64, &s2_lo[5], &s2_hi[5],
                                     &s2_lo[2], &s2_hi[2]);
  // step2[3] = (step1[4] - step1[3]) * cospi_16_64;
  // step2[4] = (step1[4] + step1[3]) * cospi_16_64;
  butterfly_one_coeff_s32_s64_narrow(s1_lo[4], s1_hi[4], s1_lo[3], s1_hi[3],
                                     cospi_16_64, &s2_lo[4], &s2_hi[4],
                                     &s2_lo[3], &s2_hi[3]);

  // step 3
  s3_lo[0] = vaddq_s32(s1_lo[0], s2_lo[3]);
  s3_hi[0] = vaddq_s32(s1_hi[0], s2_hi[3]);
  s3_lo[1] = vaddq_s32(s1_lo[1], s2_lo[2]);
  s3_hi[1] = vaddq_s32(s1_hi[1], s2_hi[2]);
  s3_lo[2] = vsubq_s32(s1_lo[1], s2_lo[2]);
  s3_hi[2] = vsubq_s32(s1_hi[1], s2_hi[2]);
  s3_lo[3] = vsubq_s32(s1_lo[0], s2_lo[3]);
  s3_hi[3] = vsubq_s32(s1_hi[0], s2_hi[3]);
  s3_lo[4] = vsubq_s32(s1_lo[7], s2_lo[4]);
  s3_hi[4] = vsubq_s32(s1_hi[7], s2_hi[4]);
  s3_lo[5] = vsubq_s32(s1_lo[6], s2_lo[5]);
  s3_hi[5] = vsubq_s32(s1_hi[6], s2_hi[5]);
  s3_lo[6] = vaddq_s32(s1_lo[6], s2_lo[5]);
  s3_hi[6] = vaddq_s32(s1_hi[6], s2_hi[5]);
  s3_lo[7] = vaddq_s32(s1_lo[7], s2_lo[4]);
  s3_hi[7] = vaddq_s32(s1_hi[7], s2_hi[4]);

  // step 4
  // s2[1] = cospi_24_64 * s3[6] - cospi_8_64 * s3[1]
  // s2[6] = cospi_8_64 * s3[6]  + cospi_24_64 * s3[1]
  butterfly_two_coeff_s32_s64_narrow(s3_lo[6], s3_hi[6], s3_lo[1], s3_hi[1],
                                     cospi_8_64, cospi_24_64, &s2_lo[6],
                                     &s2_hi[6], &s2_lo[1], &s2_hi[1]);

  // s2[5] =  cospi_8_64 * s3[2] - cospi_24_64 * s3[5]
  // s2[2] = cospi_24_64 * s3[2] + cospi_8_64 * s3[5]
  butterfly_two_coeff_s32_s64_narrow(s3_lo[2], s3_hi[2], s3_lo[5], s3_hi[5],
                                     cospi_24_64, cospi_8_64, &s2_lo[2],
                                     &s2_hi[2], &s2_lo[5], &s2_hi[5]);

  // step 5
  s1_lo[0] = vaddq_s32(s3_lo[0], s2_lo[1]);
  s1_hi[0] = vaddq_s32(s3_hi[0], s2_hi[1]);
  s1_lo[1] = vsubq_s32(s3_lo[0], s2_lo[1]);
  s1_hi[1] = vsubq_s32(s3_hi[0], s2_hi[1]);
  s1_lo[2] = vaddq_s32(s3_lo[3], s2_lo[2]);
  s1_hi[2] = vaddq_s32(s3_hi[3], s2_hi[2]);
  s1_lo[3] = vsubq_s32(s3_lo[3], s2_lo[2]);
  s1_hi[3] = vsubq_s32(s3_hi[3], s2_hi[2]);
  s1_lo[4] = vsubq_s32(s3_lo[4], s2_lo[5]);
  s1_hi[4] = vsubq_s32(s3_hi[4], s2_hi[5]);
  s1_lo[5] = vaddq_s32(s3_lo[4], s2_lo[5]);
  s1_hi[5] = vaddq_s32(s3_hi[4], s2_hi[5]);
  s1_lo[6] = vsubq_s32(s3_lo[7], s2_lo[6]);
  s1_hi[6] = vsubq_s32(s3_hi[7], s2_hi[6]);
  s1_lo[7] = vaddq_s32(s3_lo[7], s2_lo[6]);
  s1_hi[7] = vaddq_s32(s3_hi[7], s2_hi[6]);

  // step 6
  // out[1]  = step1[7] * cospi_2_64 + step1[0] * cospi_30_64
  // out[15] = step1[7] * cospi_30_64 - step1[0] * cospi_2_64
  butterfly_two_coeff_s32_s64_narrow(s1_lo[7], s1_hi[7], s1_lo[0], s1_hi[0],
                                     cospi_2_64, cospi_30_64, &left[1],
                                     &right[1], &left[15], &right[15]);

  // out[9] = step1[6] * cospi_18_64 + step1[1] * cospi_14_64
  // out[7] = step1[6] * cospi_14_64 - step1[1] * cospi_18_64
  butterfly_two_coeff_s32_s64_narrow(s1_lo[6], s1_hi[6], s1_lo[1], s1_hi[1],
                                     cospi_18_64, cospi_14_64, &left[9],
                                     &right[9], &left[7], &right[7]);

  // out[5]  = step1[5] * cospi_10_64 + step1[2] * cospi_22_64
  // out[11] = step1[5] * cospi_22_64 - step1[2] * cospi_10_64
  butterfly_two_coeff_s32_s64_narrow(s1_lo[5], s1_hi[5], s1_lo[2], s1_hi[2],
                                     cospi_10_64, cospi_22_64, &left[5],
                                     &right[5], &left[11], &right[11]);

  // out[13] = step1[4] * cospi_26_64 + step1[3] * cospi_6_64
  // out[3]  = step1[4] * cospi_6_64  - step1[3] * cospi_26_64
  butterfly_two_coeff_s32_s64_narrow(s1_lo[4], s1_hi[4], s1_lo[3], s1_hi[3],
                                     cospi_26_64, cospi_6_64, &left[13],
                                     &right[13], &left[3], &right[3]);

  left[0] = left8[0];
  right[0] = right8[0];
  left[2] = left8[1];
  right[2] = right8[1];
  left[4] = left8[2];
  right[4] = right8[2];
  left[6] = left8[3];
  right[6] = right8[3];
  left[8] = left8[4];
  right[8] = right8[4];
  left[10] = left8[5];
  right[10] = right8[5];
  left[12] = left8[6];
  right[12] = right8[6];
  left[14] = left8[7];
  right[14] = right8[7];
}

static void highbd_fadst16_8col(int32x4_t *left, int32x4_t *right) {
  // perform 16x16 1-D ADST for 8 columns
  int32x4_t x_lo[16], x_hi[16];
  int32x4_t s_lo[16], s_hi[16];
  int32x4_t t_lo[16], t_hi[16];
  int64x2_t s64_lo[32], s64_hi[32];

  x_lo[0] = left[15];
  x_hi[0] = right[15];
  x_lo[1] = left[0];
  x_hi[1] = right[0];
  x_lo[2] = left[13];
  x_hi[2] = right[13];
  x_lo[3] = left[2];
  x_hi[3] = right[2];
  x_lo[4] = left[11];
  x_hi[4] = right[11];
  x_lo[5] = left[4];
  x_hi[5] = right[4];
  x_lo[6] = left[9];
  x_hi[6] = right[9];
  x_lo[7] = left[6];
  x_hi[7] = right[6];
  x_lo[8] = left[7];
  x_hi[8] = right[7];
  x_lo[9] = left[8];
  x_hi[9] = right[8];
  x_lo[10] = left[5];
  x_hi[10] = right[5];
  x_lo[11] = left[10];
  x_hi[11] = right[10];
  x_lo[12] = left[3];
  x_hi[12] = right[3];
  x_lo[13] = left[12];
  x_hi[13] = right[12];
  x_lo[14] = left[1];
  x_hi[14] = right[1];
  x_lo[15] = left[14];
  x_hi[15] = right[14];

  // stage 1, indices are doubled
  // s0 = cospi_1_64 * x0 + cospi_31_64 * x1;
  // s1 = cospi_31_64 * x0 - cospi_1_64 * x1;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[0], x_hi[0], x_lo[1], x_hi[1], cospi_1_64, cospi_31_64,
      &s64_lo[2 * 0], &s64_hi[2 * 0], &s64_lo[2 * 1], &s64_hi[2 * 1]);
  // s2 = cospi_5_64 * x2 + cospi_27_64 * x3;
  // s3 = cospi_27_64 * x2 - cospi_5_64 * x3;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[2], x_hi[2], x_lo[3], x_hi[3], cospi_5_64, cospi_27_64,
      &s64_lo[2 * 2], &s64_hi[2 * 2], &s64_lo[2 * 3], &s64_hi[2 * 3]);
  // s4 = cospi_9_64 * x4 + cospi_23_64 * x5;
  // s5 = cospi_23_64 * x4 - cospi_9_64 * x5;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[4], x_hi[4], x_lo[5], x_hi[5], cospi_9_64, cospi_23_64,
      &s64_lo[2 * 4], &s64_hi[2 * 4], &s64_lo[2 * 5], &s64_hi[2 * 5]);
  // s6 = cospi_13_64 * x6 + cospi_19_64 * x7;
  // s7 = cospi_19_64 * x6 - cospi_13_64 * x7;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[6], x_hi[6], x_lo[7], x_hi[7], cospi_13_64, cospi_19_64,
      &s64_lo[2 * 6], &s64_hi[2 * 6], &s64_lo[2 * 7], &s64_hi[2 * 7]);
  // s8 = cospi_17_64 * x8 + cospi_15_64 * x9;
  // s9 = cospi_15_64 * x8 - cospi_17_64 * x9;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[8], x_hi[8], x_lo[9], x_hi[9], cospi_17_64, cospi_15_64,
      &s64_lo[2 * 8], &s64_hi[2 * 8], &s64_lo[2 * 9], &s64_hi[2 * 9]);
  // s10 = cospi_21_64 * x10 + cospi_11_64 * x11;
  // s11 = cospi_11_64 * x10 - cospi_21_64 * x11;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[10], x_hi[10], x_lo[11], x_hi[11], cospi_21_64, cospi_11_64,
      &s64_lo[2 * 10], &s64_hi[2 * 10], &s64_lo[2 * 11], &s64_hi[2 * 11]);
  // s12 = cospi_25_64 * x12 + cospi_7_64 * x13;
  // s13 = cospi_7_64 * x12 - cospi_25_64 * x13;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[12], x_hi[12], x_lo[13], x_hi[13], cospi_25_64, cospi_7_64,
      &s64_lo[2 * 12], &s64_hi[2 * 12], &s64_lo[2 * 13], &s64_hi[2 * 13]);
  // s14 = cospi_29_64 * x14 + cospi_3_64 * x15;
  // s15 = cospi_3_64 * x14 - cospi_29_64 * x15;
  butterfly_two_coeff_s32_s64_noround(
      x_lo[14], x_hi[14], x_lo[15], x_hi[15], cospi_29_64, cospi_3_64,
      &s64_lo[2 * 14], &s64_hi[2 * 14], &s64_lo[2 * 15], &s64_hi[2 * 15]);

  // fdct_round_shift, indices are doubled
  t_lo[0] = add_s64_round_narrow(&s64_lo[2 * 0], &s64_lo[2 * 8]);
  t_hi[0] = add_s64_round_narrow(&s64_hi[2 * 0], &s64_hi[2 * 8]);
  t_lo[1] = add_s64_round_narrow(&s64_lo[2 * 1], &s64_lo[2 * 9]);
  t_hi[1] = add_s64_round_narrow(&s64_hi[2 * 1], &s64_hi[2 * 9]);
  t_lo[2] = add_s64_round_narrow(&s64_lo[2 * 2], &s64_lo[2 * 10]);
  t_hi[2] = add_s64_round_narrow(&s64_hi[2 * 2], &s64_hi[2 * 10]);
  t_lo[3] = add_s64_round_narrow(&s64_lo[2 * 3], &s64_lo[2 * 11]);
  t_hi[3] = add_s64_round_narrow(&s64_hi[2 * 3], &s64_hi[2 * 11]);
  t_lo[4] = add_s64_round_narrow(&s64_lo[2 * 4], &s64_lo[2 * 12]);
  t_hi[4] = add_s64_round_narrow(&s64_hi[2 * 4], &s64_hi[2 * 12]);
  t_lo[5] = add_s64_round_narrow(&s64_lo[2 * 5], &s64_lo[2 * 13]);
  t_hi[5] = add_s64_round_narrow(&s64_hi[2 * 5], &s64_hi[2 * 13]);
  t_lo[6] = add_s64_round_narrow(&s64_lo[2 * 6], &s64_lo[2 * 14]);
  t_hi[6] = add_s64_round_narrow(&s64_hi[2 * 6], &s64_hi[2 * 14]);
  t_lo[7] = add_s64_round_narrow(&s64_lo[2 * 7], &s64_lo[2 * 15]);
  t_hi[7] = add_s64_round_narrow(&s64_hi[2 * 7], &s64_hi[2 * 15]);
  t_lo[8] = sub_s64_round_narrow(&s64_lo[2 * 0], &s64_lo[2 * 8]);
  t_hi[8] = sub_s64_round_narrow(&s64_hi[2 * 0], &s64_hi[2 * 8]);
  t_lo[9] = sub_s64_round_narrow(&s64_lo[2 * 1], &s64_lo[2 * 9]);
  t_hi[9] = sub_s64_round_narrow(&s64_hi[2 * 1], &s64_hi[2 * 9]);
  t_lo[10] = sub_s64_round_narrow(&s64_lo[2 * 2], &s64_lo[2 * 10]);
  t_hi[10] = sub_s64_round_narrow(&s64_hi[2 * 2], &s64_hi[2 * 10]);
  t_lo[11] = sub_s64_round_narrow(&s64_lo[2 * 3], &s64_lo[2 * 11]);
  t_hi[11] = sub_s64_round_narrow(&s64_hi[2 * 3], &s64_hi[2 * 11]);
  t_lo[12] = sub_s64_round_narrow(&s64_lo[2 * 4], &s64_lo[2 * 12]);
  t_hi[12] = sub_s64_round_narrow(&s64_hi[2 * 4], &s64_hi[2 * 12]);
  t_lo[13] = sub_s64_round_narrow(&s64_lo[2 * 5], &s64_lo[2 * 13]);
  t_hi[13] = sub_s64_round_narrow(&s64_hi[2 * 5], &s64_hi[2 * 13]);
  t_lo[14] = sub_s64_round_narrow(&s64_lo[2 * 6], &s64_lo[2 * 14]);
  t_hi[14] = sub_s64_round_narrow(&s64_hi[2 * 6], &s64_hi[2 * 14]);
  t_lo[15] = sub_s64_round_narrow(&s64_lo[2 * 7], &s64_lo[2 * 15]);
  t_hi[15] = sub_s64_round_narrow(&s64_hi[2 * 7], &s64_hi[2 * 15]);

  // stage 2
  s_lo[0] = t_lo[0];
  s_hi[0] = t_hi[0];
  s_lo[1] = t_lo[1];
  s_hi[1] = t_hi[1];
  s_lo[2] = t_lo[2];
  s_hi[2] = t_hi[2];
  s_lo[3] = t_lo[3];
  s_hi[3] = t_hi[3];
  s_lo[4] = t_lo[4];
  s_hi[4] = t_hi[4];
  s_lo[5] = t_lo[5];
  s_hi[5] = t_hi[5];
  s_lo[6] = t_lo[6];
  s_hi[6] = t_hi[6];
  s_lo[7] = t_lo[7];
  s_hi[7] = t_hi[7];
  // s8 = x8 * cospi_4_64 + x9 * cospi_28_64;
  // s9 = x8 * cospi_28_64 - x9 * cospi_4_64;
  butterfly_two_coeff_s32_s64_noround(
      t_lo[8], t_hi[8], t_lo[9], t_hi[9], cospi_4_64, cospi_28_64,
      &s64_lo[2 * 8], &s64_hi[2 * 8], &s64_lo[2 * 9], &s64_hi[2 * 9]);
  // s10 = x10 * cospi_20_64 + x11 * cospi_12_64;
  // s11 = x10 * cospi_12_64 - x11 * cospi_20_64;
  butterfly_two_coeff_s32_s64_noround(
      t_lo[10], t_hi[10], t_lo[11], t_hi[11], cospi_20_64, cospi_12_64,
      &s64_lo[2 * 10], &s64_hi[2 * 10], &s64_lo[2 * 11], &s64_hi[2 * 11]);
  // s12 = -x12 * cospi_28_64 + x13 * cospi_4_64;
  // s13 = x12 * cospi_4_64 + x13 * cospi_28_64;
  butterfly_two_coeff_s32_s64_noround(
      t_lo[13], t_hi[13], t_lo[12], t_hi[12], cospi_28_64, cospi_4_64,
      &s64_lo[2 * 13], &s64_hi[2 * 13], &s64_lo[2 * 12], &s64_hi[2 * 12]);
  // s14 = -x14 * cospi_12_64 + x15 * cospi_20_64;
  // s15 = x14 * cospi_20_64 + x15 * cospi_12_64;
  butterfly_two_coeff_s32_s64_noround(
      t_lo[15], t_hi[15], t_lo[14], t_hi[14], cospi_12_64, cospi_20_64,
      &s64_lo[2 * 15], &s64_hi[2 * 15], &s64_lo[2 * 14], &s64_hi[2 * 14]);

  // s0 + s4
  t_lo[0] = add_s32_s64_narrow(s_lo[0], s_lo[4]);
  t_hi[0] = add_s32_s64_narrow(s_hi[0], s_hi[4]);
  // s1 + s5
  t_lo[1] = add_s32_s64_narrow(s_lo[1], s_lo[5]);
  t_hi[1] = add_s32_s64_narrow(s_hi[1], s_hi[5]);
  // s2 + s6
  t_lo[2] = add_s32_s64_narrow(s_lo[2], s_lo[6]);
  t_hi[2] = add_s32_s64_narrow(s_hi[2], s_hi[6]);
  // s3 + s7
  t_lo[3] = add_s32_s64_narrow(s_lo[3], s_lo[7]);
  t_hi[3] = add_s32_s64_narrow(s_hi[3], s_hi[7]);

  // s0 - s4
  t_lo[4] = sub_s32_s64_narrow(s_lo[0], s_lo[4]);
  t_hi[4] = sub_s32_s64_narrow(s_hi[0], s_hi[4]);
  // s1 - s5
  t_lo[5] = sub_s32_s64_narrow(s_lo[1], s_lo[5]);
  t_hi[5] = sub_s32_s64_narrow(s_hi[1], s_hi[5]);
  // s2 - s6
  t_lo[6] = sub_s32_s64_narrow(s_lo[2], s_lo[6]);
  t_hi[6] = sub_s32_s64_narrow(s_hi[2], s_hi[6]);
  // s3 - s7
  t_lo[7] = sub_s32_s64_narrow(s_lo[3], s_lo[7]);
  t_hi[7] = sub_s32_s64_narrow(s_hi[3], s_hi[7]);

  // fdct_round_shift()
  // s8 + s12
  t_lo[8] = add_s64_round_narrow(&s64_lo[2 * 8], &s64_lo[2 * 12]);
  t_hi[8] = add_s64_round_narrow(&s64_hi[2 * 8], &s64_hi[2 * 12]);
  // s9 + s13
  t_lo[9] = add_s64_round_narrow(&s64_lo[2 * 9], &s64_lo[2 * 13]);
  t_hi[9] = add_s64_round_narrow(&s64_hi[2 * 9], &s64_hi[2 * 13]);
  // s10 + s14
  t_lo[10] = add_s64_round_narrow(&s64_lo[2 * 10], &s64_lo[2 * 14]);
  t_hi[10] = add_s64_round_narrow(&s64_hi[2 * 10], &s64_hi[2 * 14]);
  // s11 + s15
  t_lo[11] = add_s64_round_narrow(&s64_lo[2 * 11], &s64_lo[2 * 15]);
  t_hi[11] = add_s64_round_narrow(&s64_hi[2 * 11], &s64_hi[2 * 15]);

  // s8 - s12
  t_lo[12] = sub_s64_round_narrow(&s64_lo[2 * 8], &s64_lo[2 * 12]);
  t_hi[12] = sub_s64_round_narrow(&s64_hi[2 * 8], &s64_hi[2 * 12]);
  // s9 - s13
  t_lo[13] = sub_s64_round_narrow(&s64_lo[2 * 9], &s64_lo[2 * 13]);
  t_hi[13] = sub_s64_round_narrow(&s64_hi[2 * 9], &s64_hi[2 * 13]);
  // s10 - s14
  t_lo[14] = sub_s64_round_narrow(&s64_lo[2 * 10], &s64_lo[2 * 14]);
  t_hi[14] = sub_s64_round_narrow(&s64_hi[2 * 10], &s64_hi[2 * 14]);
  // s11 - s15
  t_lo[15] = sub_s64_round_narrow(&s64_lo[2 * 11], &s64_lo[2 * 15]);
  t_hi[15] = sub_s64_round_narrow(&s64_hi[2 * 11], &s64_hi[2 * 15]);

  // stage 3
  s_lo[0] = t_lo[0];
  s_hi[0] = t_hi[0];
  s_lo[1] = t_lo[1];
  s_hi[1] = t_hi[1];
  s_lo[2] = t_lo[2];
  s_hi[2] = t_hi[2];
  s_lo[3] = t_lo[3];
  s_hi[3] = t_hi[3];
  // s4 = x4 * cospi_8_64 + x5 * cospi_24_64;
  // s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
  butterfly_two_coeff_s32_s64_noround(
      t_lo[4], t_hi[4], t_lo[5], t_hi[5], cospi_8_64, cospi_24_64,
      &s64_lo[2 * 4], &s64_hi[2 * 4], &s64_lo[2 * 5], &s64_hi[2 * 5]);
  // s6 = -x6 * cospi_24_64 + x7 * cospi_8_64;
  // s7 = x6 * cospi_8_64 + x7 * cospi_24_64;
  butterfly_two_coeff_s32_s64_noround(
      t_lo[7], t_hi[7], t_lo[6], t_hi[6], cospi_24_64, cospi_8_64,
      &s64_lo[2 * 7], &s64_hi[2 * 7], &s64_lo[2 * 6], &s64_hi[2 * 6]);
  s_lo[8] = t_lo[8];
  s_hi[8] = t_hi[8];
  s_lo[9] = t_lo[9];
  s_hi[9] = t_hi[9];
  s_lo[10] = t_lo[10];
  s_hi[10] = t_hi[10];
  s_lo[11] = t_lo[11];
  s_hi[11] = t_hi[11];
  // s12 = x12 * cospi_8_64 + x13 * cospi_24_64;
  // s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
  butterfly_two_coeff_s32_s64_noround(
      t_lo[12], t_hi[12], t_lo[13], t_hi[13], cospi_8_64, cospi_24_64,
      &s64_lo[2 * 12], &s64_hi[2 * 12], &s64_lo[2 * 13], &s64_hi[2 * 13]);
  // s14 = -x14 * cospi_24_64 + x15 * cospi_8_64;
  // s15 = x14 * cospi_8_64 + x15 * cospi_24_64;
  butterfly_two_coeff_s32_s64_noround(
      t_lo[15], t_hi[15], t_lo[14], t_hi[14], cospi_24_64, cospi_8_64,
      &s64_lo[2 * 15], &s64_hi[2 * 15], &s64_lo[2 * 14], &s64_hi[2 * 14]);

  // s0 + s2
  t_lo[0] = add_s32_s64_narrow(s_lo[0], s_lo[2]);
  t_hi[0] = add_s32_s64_narrow(s_hi[0], s_hi[2]);
  // s1 + s3
  t_lo[1] = add_s32_s64_narrow(s_lo[1], s_lo[3]);
  t_hi[1] = add_s32_s64_narrow(s_hi[1], s_hi[3]);
  // s0 - s2
  t_lo[2] = sub_s32_s64_narrow(s_lo[0], s_lo[2]);
  t_hi[2] = sub_s32_s64_narrow(s_hi[0], s_hi[2]);
  // s1 - s3
  t_lo[3] = sub_s32_s64_narrow(s_lo[1], s_lo[3]);
  t_hi[3] = sub_s32_s64_narrow(s_hi[1], s_hi[3]);
  // fdct_round_shift()
  // s4 + s6
  t_lo[4] = add_s64_round_narrow(&s64_lo[2 * 4], &s64_lo[2 * 6]);
  t_hi[4] = add_s64_round_narrow(&s64_hi[2 * 4], &s64_hi[2 * 6]);
  // s5 + s7
  t_lo[5] = add_s64_round_narrow(&s64_lo[2 * 5], &s64_lo[2 * 7]);
  t_hi[5] = add_s64_round_narrow(&s64_hi[2 * 5], &s64_hi[2 * 7]);
  // s4 - s6
  t_lo[6] = sub_s64_round_narrow(&s64_lo[2 * 4], &s64_lo[2 * 6]);
  t_hi[6] = sub_s64_round_narrow(&s64_hi[2 * 4], &s64_hi[2 * 6]);
  // s5 - s7
  t_lo[7] = sub_s64_round_narrow(&s64_lo[2 * 5], &s64_lo[2 * 7]);
  t_hi[7] = sub_s64_round_narrow(&s64_hi[2 * 5], &s64_hi[2 * 7]);
  // s8 + s10
  t_lo[8] = add_s32_s64_narrow(s_lo[8], s_lo[10]);
  t_hi[8] = add_s32_s64_narrow(s_hi[8], s_hi[10]);
  // s9 + s11
  t_lo[9] = add_s32_s64_narrow(s_lo[9], s_lo[11]);
  t_hi[9] = add_s32_s64_narrow(s_hi[9], s_hi[11]);
  // s8 - s10
  t_lo[10] = sub_s32_s64_narrow(s_lo[8], s_lo[10]);
  t_hi[10] = sub_s32_s64_narrow(s_hi[8], s_hi[10]);
  // s9 - s11
  t_lo[11] = sub_s32_s64_narrow(s_lo[9], s_lo[11]);
  t_hi[11] = sub_s32_s64_narrow(s_hi[9], s_hi[11]);
  // fdct_round_shift()
  // s12 + s14
  t_lo[12] = add_s64_round_narrow(&s64_lo[2 * 12], &s64_lo[2 * 14]);
  t_hi[12] = add_s64_round_narrow(&s64_hi[2 * 12], &s64_hi[2 * 14]);
  // s13 + s15
  t_lo[13] = add_s64_round_narrow(&s64_lo[2 * 13], &s64_lo[2 * 15]);
  t_hi[13] = add_s64_round_narrow(&s64_hi[2 * 13], &s64_hi[2 * 15]);
  // s12 - s14
  t_lo[14] = sub_s64_round_narrow(&s64_lo[2 * 12], &s64_lo[2 * 14]);
  t_hi[14] = sub_s64_round_narrow(&s64_hi[2 * 12], &s64_hi[2 * 14]);
  // s13 - s15
  t_lo[15] = sub_s64_round_narrow(&s64_lo[2 * 13], &s64_lo[2 * 15]);
  t_hi[15] = sub_s64_round_narrow(&s64_hi[2 * 13], &s64_hi[2 * 15]);

  // stage 4, with fdct_round_shift
  // s2 = (-cospi_16_64) * (x2 + x3);
  // s3 = cospi_16_64 * (x2 - x3);
  butterfly_one_coeff_s32_s64_narrow(t_lo[3], t_hi[3], t_lo[2], t_hi[2],
                                     -cospi_16_64, &x_lo[2], &x_hi[2], &x_lo[e='color: green'>3],
                                     &x_hi[3]);
  // s6 = cospi_16_64 * (x6 + x7);
  // s7 = cospi_16_64 * (-x6 + x7);
  butterfly_one_coeff_s32_s64_narrow(t_lo[7], t_hi[7], t_lo[6], t_hi[6],
                                     cospi_16_64, &x_lo[6], &x_hi[6], &x_lo[='color: green'>7],
                                     &x_hi[7]);
  // s10 = cospi_16_64 * (x10 + x11);
  // s11 = cospi_16_64 * (-x10 + x11);
  butterfly_one_coeff_s32_s64_narrow(t_lo[11], t_hi[11], t_lo[10], t_hi[10],
                                     cospi_16_64, &x_lo[10], &x_hi[10],
                                     &x_lo[11], &x_hi[11]);
  // s14 = (-cospi_16_64) * (x14 + x15);
  // s15 = cospi_16_64 * (x14 - x15);
  butterfly_one_coeff_s32_s64_narrow(t_lo[15], t_hi[15], t_lo[14], t_hi[14],
                                     -cospi_16_64, &x_lo[14], &x_hi[14],
                                     &x_lo[15], &x_hi[15]);

  // Just copy x0, x1, x4, x5, x8, x9, x12, x13
  x_lo[0] = t_lo[0];
  x_hi[0] = t_hi[0];
  x_lo[1] = t_lo[1];
  x_hi[1] = t_hi[1];
  x_lo[4] = t_lo[4];
  x_hi[4] = t_hi[4];
  x_lo[5] = t_lo[5];
  x_hi[5] = t_hi[5];
  x_lo[8] = t_lo[8];
  x_hi[8] = t_hi[8];
  x_lo[9] = t_lo[9];
  x_hi[9] = t_hi[9];
  x_lo[12] = t_lo[12];
  x_hi[12] = t_hi[12];
  x_lo[13] = t_lo[13];
  x_hi[13] = t_hi[13];

  left[0] = x_lo[0];
  right[0] = x_hi[0];
  left[1] = vnegq_s32(x_lo[8]);
  right[1] = vnegq_s32(x_hi[8]);
  left[2] = x_lo[12];
  right[2] = x_hi[12];
  left[3] = vnegq_s32(x_lo[4]);
  right[3] = vnegq_s32(x_hi[4]);
  left[4] = x_lo[6];
  right[4] = x_hi[6];
  left[5] = x_lo[14];
  right[5] = x_hi[14];
  left[6] = x_lo[10];
  right[6] = x_hi[10];
  left[7] = x_lo[2];
  right[7] = x_hi[2];
  left[8] = x_lo[3];
  right[8] = x_hi[3];
  left[9] = x_lo[11];
  right[9] = x_hi[11];
  left[10] = x_lo[15];
  right[10] = x_hi[15];
  left[11] = x_lo[7];
  right[11] = x_hi[7];
  left[12] = x_lo[5];
  right[12] = x_hi[5];
  left[13] = vnegq_s32(x_lo[13]);
  right[13] = vnegq_s32(x_hi[13]);
  left[14] = x_lo[9];
  right[14] = x_hi[9];
  left[15] = vnegq_s32(x_lo[1]);
  right[15] = vnegq_s32(x_hi[1]);
}

static void highbd_fdct16x16_neon(int32x4_t *left1, int32x4_t *right1,
                                  int32x4_t *left2, int32x4_t *right2) {
  // Left half.
  highbd_fdct16_8col(left1, right1);
  // Right half.
  highbd_fdct16_8col(left2, right2);
  transpose_s32_16x16(left1, right1, left2, right2);
}

static void highbd_fadst16x16_neon(int32x4_t *left1, int32x4_t *right1,
                                   int32x4_t *left2, int32x4_t *right2) {
  // Left half.
  highbd_fadst16_8col(left1, right1);
  // Right half.
  highbd_fadst16_8col(left2, right2);
  transpose_s32_16x16(left1, right1, left2, right2);
}

void vp9_highbd_fht16x16_neon(const int16_t *input, tran_low_t *output,
                              int stride, int tx_type) {
  int32x4_t left1[16], right1[16], left2[16], right2[16];

  switch (tx_type) {
    case DCT_DCT: vpx_highbd_fdct16x16_neon(input, output, stride); break;
    case ADST_DCT:
      highbd_load_buffer_16x16(input, left1, right1, left2, right2, stride);
      highbd_fadst16x16_neon(left1, right1, left2, right2);
      highbd_write_buffer_16x16(output, left1, right1, left2, right2, 16);
      highbd_right_shift_16x16(left1, right1, left2, right2, 2);
      highbd_fdct16x16_neon(left1, right1, left2, right2);
      highbd_write_buffer_16x16(output, left1, right1, left2, right2, 16);
      break;
    case DCT_ADST:
      highbd_load_buffer_16x16(input, left1, right1, left2, right2, stride);
      highbd_fdct16x16_neon(left1, right1, left2, right2);
      highbd_right_shift_16x16(left1, right1, left2, right2, 2);
      highbd_fadst16x16_neon(left1, right1, left2, right2);
      highbd_write_buffer_16x16(output, left1, right1, left2, right2, 16);
      break;
    default:
      assert(tx_type == ADST_ADST);
      highbd_load_buffer_16x16(input, left1, right1, left2, right2, stride);
      highbd_fadst16x16_neon(left1, right1, left2, right2);
      highbd_right_shift_16x16(left1, right1, left2, right2, 2);
      highbd_fadst16x16_neon(left1, right1, left2, right2);
      highbd_write_buffer_16x16(output, left1, right1, left2, right2, 16);
      break;
  }
}

#endif  // CONFIG_VP9_HIGHBITDEPTH