av1/common/x86/highbd_inv_txfm_sse4.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 <assert.h>
#include <smmintrin.h> /* SSE4.1 */

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

#include "av1/common/av1_inv_txfm1d_cfg.h"
#include "av1/common/idct.h"
#include "av1/common/x86/av1_inv_txfm_ssse3.h"
#include "av1/common/x86/av1_txfm_sse2.h"
#include "av1/common/x86/av1_txfm_sse4.h"
#include "av1/common/x86/highbd_txfm_utility_sse4.h"

static INLINE __m128i highbd_clamp_epi16(__m128i u, int bd) {
  const __m128i zero = _mm_setzero_si128();
  const __m128i one = _mm_set1_epi16(1);
  const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
  __m128i clamped, mask;

  mask = _mm_cmpgt_epi16(u, max);
  clamped = _mm_andnot_si128(mask, u);
  mask = _mm_and_si128(mask, max);
  clamped = _mm_or_si128(mask, clamped);
  mask = _mm_cmpgt_epi16(clamped, zero);
  clamped = _mm_and_si128(clamped, mask);

  return clamped;
}

static INLINE void round_shift_4x4(__m128i *in, int shift) {
  if (shift != 0) {
    __m128i rnding = _mm_set1_epi32(1 << (shift - 1));
    in[0] = _mm_add_epi32(in[0], rnding);
    in[1] = _mm_add_epi32(in[1], rnding);
    in[2] = _mm_add_epi32(in[2], rnding);
    in[3] = _mm_add_epi32(in[3], rnding);

    in[0] = _mm_srai_epi32(in[0], shift);
    in[1] = _mm_srai_epi32(in[1], shift);
    in[2] = _mm_srai_epi32(in[2], shift);
    in[3] = _mm_srai_epi32(in[3], shift);
  }
}

static void round_shift_8x8(__m128i *in, int shift) {
  round_shift_4x4(&in[0], shift);
  round_shift_4x4(&in[4], shift);
  round_shift_4x4(&in[8], shift);
  round_shift_4x4(&in[12], shift);
}

static void highbd_clamp_epi32_sse4_1(__m128i *in, __m128i *out,
                                      const __m128i *clamp_lo,
                                      const __m128i *clamp_hi, int size) {
  __m128i a0, a1;
  for (int i = 0; i < size; i += 4) {
    a0 = _mm_max_epi32(in[i], *clamp_lo);
    out[i] = _mm_min_epi32(a0, *clamp_hi);

    a1 = _mm_max_epi32(in[i + 1], *clamp_lo);
    out[i + 1] = _mm_min_epi32(a1, *clamp_hi);

    a0 = _mm_max_epi32(in[i + 2], *clamp_lo);
    out[i + 2] = _mm_min_epi32(a0, *clamp_hi);

    a1 = _mm_max_epi32(in[i + 3], *clamp_lo);
    out[i + 3] = _mm_min_epi32(a1, *clamp_hi);
  }
}

static INLINE __m128i highbd_get_recon_8x8_sse4_1(const __m128i pred,
                                                  __m128i res0, __m128i res1,
                                                  const int bd) {
  __m128i x0 = _mm_cvtepi16_epi32(pred);
  __m128i x1 = _mm_cvtepi16_epi32(_mm_srli_si128(pred, 8));
  __m128i min_clip_val = _mm_setzero_si128();
  __m128i max_clip_val = _mm_set1_epi32((1 << bd) - 1);
  x0 = _mm_add_epi32(res0, x0);
  x1 = _mm_add_epi32(res1, x1);
  x0 = _mm_max_epi32(x0, min_clip_val);
  x0 = _mm_min_epi32(x0, max_clip_val);
  x1 = _mm_max_epi32(x1, min_clip_val);
  x1 = _mm_min_epi32(x1, max_clip_val);
  x0 = _mm_packus_epi32(x0, x1);
  return x0;
}

static INLINE __m128i highbd_get_recon_4xn_sse4_1(const __m128i pred,
                                                  __m128i res0, const int bd) {
  __m128i x0 = _mm_cvtepi16_epi32(pred);

  x0 = _mm_add_epi32(res0, x0);
  x0 = _mm_packus_epi32(x0, x0);
  x0 = highbd_clamp_epi16(x0, bd);
  return x0;
}

static INLINE void highbd_write_buffer_4xn_sse4_1(__m128i *in, uint16_t *output,
                                                  int stride, int flipud,
                                                  int height, const int bd) {
  int j = flipud ? (height - 1) : 0;
  const int step = flipud ? -1 : 1;
  for (int i = 0; i < height; ++i, j += step) {
    __m128i v = _mm_loadl_epi64((__m128i const *)(output + i * stride));
    __m128i u = highbd_get_recon_4xn_sse4_1(v, in[j], bd);

    _mm_storel_epi64((__m128i *)(output + i * stride), u);
  }
}

static INLINE void highbd_write_buffer_8xn_sse4_1(__m128i *in, uint16_t *output,
                                                  int stride, int flipud,
                                                  int height, const int bd) {
  int j = flipud ? (height - 1) : 0;
  const int step = flipud ? -1 : 1;
  for (int i = 0; i < height; ++i, j += step) {
    __m128i v = _mm_loadu_si128((__m128i const *)(output + i * stride));
    __m128i u = highbd_get_recon_8x8_sse4_1(v, in[j], in[j + height], bd);

    _mm_storeu_si128((__m128i *)(output + i * stride), u);
  }
}

static INLINE void load_buffer_32bit_input(const int32_t *in, int stride,
                                           __m128i *out, int out_size) {
  for (int i = 0; i < out_size; ++i) {
    out[i] = _mm_loadu_si128((const __m128i *)(in + i * stride));
  }
}

static INLINE void load_buffer_4x4(const int32_t *coeff, __m128i *in) {
  in[0] = _mm_load_si128((const __m128i *)(coeff + 0));
  in[1] = _mm_load_si128((const __m128i *)(coeff + 4));
  in[2] = _mm_load_si128((const __m128i *)(coeff + 8));
  in[3] = _mm_load_si128((const __m128i *)(coeff + 12));
}

static void addsub_sse4_1(const __m128i in0, const __m128i in1, __m128i *out0,
                          __m128i *out1, const __m128i *clamp_lo,
                          const __m128i *clamp_hi) {
  __m128i a0 = _mm_add_epi32(in0, in1);
  __m128i a1 = _mm_sub_epi32(in0, in1);

  a0 = _mm_max_epi32(a0, *clamp_lo);
  a0 = _mm_min_epi32(a0, *clamp_hi);
  a1 = _mm_max_epi32(a1, *clamp_lo);
  a1 = _mm_min_epi32(a1, *clamp_hi);

  *out0 = a0;
  *out1 = a1;
}

static void shift_and_clamp_sse4_1(__m128i *in0, __m128i *in1,
                                   const __m128i *clamp_lo,
                                   const __m128i *clamp_hi, int shift) {
  __m128i offset = _mm_set1_epi32((1 << shift) >> 1);
  __m128i in0_w_offset = _mm_add_epi32(*in0, offset);
  __m128i in1_w_offset = _mm_add_epi32(*in1, offset);

  in0_w_offset = _mm_sra_epi32(in0_w_offset, _mm_cvtsi32_si128(shift));
  in1_w_offset = _mm_sra_epi32(in1_w_offset, _mm_cvtsi32_si128(shift));

  in0_w_offset = _mm_max_epi32(in0_w_offset, *clamp_lo);
  in0_w_offset = _mm_min_epi32(in0_w_offset, *clamp_hi);
  in1_w_offset = _mm_max_epi32(in1_w_offset, *clamp_lo);
  in1_w_offset = _mm_min_epi32(in1_w_offset, *clamp_hi);

  *in0 = in0_w_offset;
  *in1 = in1_w_offset;
}

static INLINE void idct32_stage4_sse4_1(
    __m128i *bf1, const __m128i *cospim8, const __m128i *cospi56,
    const __m128i *cospi8, const __m128i *cospim56, const __m128i *cospim40,
    const __m128i *cospi24, const __m128i *cospi40, const __m128i *cospim24,
    const __m128i *rounding, int bit) {
  __m128i temp1, temp2;
  temp1 = half_btf_sse4_1(cospim8, &bf1[17], cospi56, &bf1[30], rounding, bit);
  bf1[30] = half_btf_sse4_1(cospi56, &bf1[17], cospi8, &bf1[30], rounding, bit);
  bf1[17] = temp1;

  temp2 = half_btf_sse4_1(cospim56, &bf1[18], cospim8, &bf1[29], rounding, bit);
  bf1[29] =
      half_btf_sse4_1(cospim8, &bf1[18], cospi56, &bf1[29], rounding, bit);
  bf1[18] = temp2;

  temp1 = half_btf_sse4_1(cospim40, &bf1[21], cospi24, &bf1[26], rounding, bit);
  bf1[26] =
      half_btf_sse4_1(cospi24, &bf1[21], cospi40, &bf1[26], rounding, bit);
  bf1[21] = temp1;

  temp2 =
      half_btf_sse4_1(cospim24, &bf1[22], cospim40, &bf1[25], rounding, bit);
  bf1[25] =
      half_btf_sse4_1(cospim40, &bf1[22], cospi24, &bf1[25], rounding, bit);
  bf1[22] = temp2;
}

static INLINE void idct32_stage5_sse4_1(
    __m128i *bf1, const __m128i *cospim16, const __m128i *cospi48,
    const __m128i *cospi16, const __m128i *cospim48, const __m128i *clamp_lo,
    const __m128i *clamp_hi, const __m128i *rounding, int bit) {
  __m128i temp1, temp2;
  temp1 = half_btf_sse4_1(cospim16, &bf1[9], cospi48, &bf1[14], rounding, bit);
  bf1[14] = half_btf_sse4_1(cospi48, &bf1[9], cospi16, &bf1[14], rounding, bit);
  bf1[9] = temp1;

  temp2 =
      half_btf_sse4_1(cospim48, &bf1[10], cospim16, &bf1[13], rounding, bit);
  bf1[13] =
      half_btf_sse4_1(cospim16, &bf1[10], cospi48, &bf1[13], rounding, bit);
  bf1[10] = temp2;

  addsub_sse4_1(bf1[16], bf1[19], bf1 + 16, bf1 + 19, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[17], bf1[18], bf1 + 17, bf1 + 18, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[23], bf1[20], bf1 + 23, bf1 + 20, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[22], bf1[21], bf1 + 22, bf1 + 21, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[24], bf1[27], bf1 + 24, bf1 + 27, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[25], bf1[26], bf1 + 25, bf1 + 26, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[31], bf1[28], bf1 + 31, bf1 + 28, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[30], bf1[29], bf1 + 30, bf1 + 29, clamp_lo, clamp_hi);
}

static INLINE void idct32_stage6_sse4_1(
    __m128i *bf1, const __m128i *cospim32, const __m128i *cospi32,
    const __m128i *cospim16, const __m128i *cospi48, const __m128i *cospi16,
    const __m128i *cospim48, const __m128i *clamp_lo, const __m128i *clamp_hi,
    const __m128i *rounding, int bit) {
  __m128i temp1, temp2;
  temp1 = half_btf_sse4_1(cospim32, &bf1[5], cospi32, &bf1[6], rounding, bit);
  bf1[6] = half_btf_sse4_1(cospi32, &bf1[5], cospi32, &bf1[6], rounding, bit);
  bf1[5] = temp1;

  addsub_sse4_1(bf1[8], bf1[11], bf1 + 8, bf1 + 11, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[9], bf1[10], bf1 + 9, bf1 + 10, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[15], bf1[12], bf1 + 15, bf1 + 12, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[14], bf1[13], bf1 + 14, bf1 + 13, clamp_lo, clamp_hi);

  temp1 = half_btf_sse4_1(cospim16, &bf1[18], cospi48, &bf1[29], rounding, bit);
  bf1[29] =
      half_btf_sse4_1(cospi48, &bf1[18], cospi16, &bf1[29], rounding, bit);
  bf1[18] = temp1;
  temp2 = half_btf_sse4_1(cospim16, &bf1[19], cospi48, &bf1[28], rounding, bit);
  bf1[28] =
      half_btf_sse4_1(cospi48, &bf1[19], cospi16, &bf1[28], rounding, bit);
  bf1[19] = temp2;
  temp1 =
      half_btf_sse4_1(cospim48, &bf1[20], cospim16, &bf1[27], rounding, bit);
  bf1[27] =
      half_btf_sse4_1(cospim16, &bf1[20], cospi48, &bf1[27], rounding, bit);
  bf1[20] = temp1;
  temp2 =
      half_btf_sse4_1(cospim48, &bf1[21], cospim16, &bf1[26], rounding, bit);
  bf1[26] =
      half_btf_sse4_1(cospim16, &bf1[21], cospi48, &bf1[26], rounding, bit);
  bf1[21] = temp2;
}

static INLINE void idct32_stage7_sse4_1(__m128i *bf1, const __m128i *cospim32,
                                        const __m128i *cospi32,
                                        const __m128i *clamp_lo,
                                        const __m128i *clamp_hi,
                                        const __m128i *rounding, int bit) {
  __m128i temp1, temp2;
  addsub_sse4_1(bf1[0], bf1[7], bf1 + 0, bf1 + 7, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[1], bf1[6], bf1 + 1, bf1 + 6, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[2], bf1[5], bf1 + 2, bf1 + 5, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[3], bf1[4], bf1 + 3, bf1 + 4, clamp_lo, clamp_hi);

  temp1 = half_btf_sse4_1(cospim32, &bf1[10], cospi32, &bf1[13], rounding, bit);
  bf1[13] =
      half_btf_sse4_1(cospi32, &bf1[10], cospi32, &bf1[13], rounding, bit);
  bf1[10] = temp1;
  temp2 = half_btf_sse4_1(cospim32, &bf1[11], cospi32, &bf1[12], rounding, bit);
  bf1[12] =
      half_btf_sse4_1(cospi32, &bf1[11], cospi32, &bf1[12], rounding, bit);
  bf1[11] = temp2;

  addsub_sse4_1(bf1[16], bf1[23], bf1 + 16, bf1 + 23, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[17], bf1[22], bf1 + 17, bf1 + 22, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[18], bf1[21], bf1 + 18, bf1 + 21, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[19], bf1[20], bf1 + 19, bf1 + 20, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[31], bf1[24], bf1 + 31, bf1 + 24, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[30], bf1[25], bf1 + 30, bf1 + 25, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[29], bf1[26], bf1 + 29, bf1 + 26, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[28], bf1[27], bf1 + 28, bf1 + 27, clamp_lo, clamp_hi);
}

static INLINE void idct32_stage8_sse4_1(__m128i *bf1, const __m128i *cospim32,
                                        const __m128i *cospi32,
                                        const __m128i *clamp_lo,
                                        const __m128i *clamp_hi,
                                        const __m128i *rounding, int bit) {
  __m128i temp1, temp2;
  addsub_sse4_1(bf1[0], bf1[15], bf1 + 0, bf1 + 15, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[1], bf1[14], bf1 + 1, bf1 + 14, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[2], bf1[13], bf1 + 2, bf1 + 13, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[3], bf1[12], bf1 + 3, bf1 + 12, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[4], bf1[11], bf1 + 4, bf1 + 11, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[5], bf1[10], bf1 + 5, bf1 + 10, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[6], bf1[9], bf1 + 6, bf1 + 9, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[7], bf1[8], bf1 + 7, bf1 + 8, clamp_lo, clamp_hi);

  temp1 = half_btf_sse4_1(cospim32, &bf1[20], cospi32, &bf1[27], rounding, bit);
  bf1[27] =
      half_btf_sse4_1(cospi32, &bf1[20], cospi32, &bf1[27], rounding, bit);
  bf1[20] = temp1;
  temp2 = half_btf_sse4_1(cospim32, &bf1[21], cospi32, &bf1[26], rounding, bit);
  bf1[26] =
      half_btf_sse4_1(cospi32, &bf1[21], cospi32, &bf1[26], rounding, bit);
  bf1[21] = temp2;
  temp1 = half_btf_sse4_1(cospim32, &bf1[22], cospi32, &bf1[25], rounding, bit);
  bf1[25] =
      half_btf_sse4_1(cospi32, &bf1[22], cospi32, &bf1[25], rounding, bit);
  bf1[22] = temp1;
  temp2 = half_btf_sse4_1(cospim32, &bf1[23], cospi32, &bf1[24], rounding, bit);
  bf1[24] =
      half_btf_sse4_1(cospi32, &bf1[23], cospi32, &bf1[24], rounding, bit);
  bf1[23] = temp2;
}

static INLINE void idct32_stage9_sse4_1(__m128i *bf1, __m128i *out,
                                        const int do_cols, const int bd,
                                        const int out_shift,
                                        const __m128i *clamp_lo,
                                        const __m128i *clamp_hi) {
  addsub_sse4_1(bf1[0], bf1[31], out + 0, out + 31, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[1], bf1[30], out + 1, out + 30, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[2], bf1[29], out + 2, out + 29, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[3], bf1[28], out + 3, out + 28, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[4], bf1[27], out + 4, out + 27, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[5], bf1[26], out + 5, out + 26, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[6], bf1[25], out + 6, out + 25, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[7], bf1[24], out + 7, out + 24, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[8], bf1[23], out + 8, out + 23, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[9], bf1[22], out + 9, out + 22, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[10], bf1[21], out + 10, out + 21, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[11], bf1[20], out + 11, out + 20, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[12], bf1[19], out + 12, out + 19, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[13], bf1[18], out + 13, out + 18, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[14], bf1[17], out + 14, out + 17, clamp_lo, clamp_hi);
  addsub_sse4_1(bf1[15], bf1[16], out + 15, out + 16, clamp_lo, clamp_hi);

  if (!do_cols) {
    const int log_range_out = AOMMAX(16, bd + 6);
    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);
    for (int i = 0; i < 32; i += 8) {
      round_shift_4x4(out + i, out_shift);
      round_shift_4x4(out + i + 4, out_shift);
    }
    highbd_clamp_epi32_sse4_1(out, out, &clamp_lo_out, &clamp_hi_out, 32);
  }
}

static void neg_shift_sse4_1(const __m128i in0, const __m128i in1,
                             __m128i *out0, __m128i *out1,
                             const __m128i *clamp_lo, const __m128i *clamp_hi,
                             int shift) {
  __m128i offset = _mm_set1_epi32((1 << shift) >> 1);
  __m128i a0 = _mm_add_epi32(offset, in0);
  __m128i a1 = _mm_sub_epi32(offset, in1);

  a0 = _mm_sra_epi32(a0, _mm_cvtsi32_si128(shift));
  a1 = _mm_sra_epi32(a1, _mm_cvtsi32_si128(shift));

  a0 = _mm_max_epi32(a0, *clamp_lo);
  a0 = _mm_min_epi32(a0, *clamp_hi);
  a1 = _mm_max_epi32(a1, *clamp_lo);
  a1 = _mm_min_epi32(a1, *clamp_hi);

  *out0 = a0;
  *out1 = a1;
}

static void idct4x4_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
                           int bd, int out_shift) {
  const int32_t *cospi = cospi_arr(bit);
  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
  int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
  __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
  __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
  __m128i u0, u1, u2, u3;
  __m128i v0, v1, v2, v3, x, y;

  // Stage 0
  // Stage 1
  // Stage 2
  v0 = _mm_unpacklo_epi32(in[0], in[1]);
  v1 = _mm_unpackhi_epi32(in[0], in[1]);
  v2 = _mm_unpacklo_epi32(in[2], in[3]);
  v3 = _mm_unpackhi_epi32(in[2], in[3]);

  u0 = _mm_unpacklo_epi64(v0, v2);
  u1 = _mm_unpackhi_epi64(v0, v2);
  u2 = _mm_unpacklo_epi64(v1, v3);
  u3 = _mm_unpackhi_epi64(v1, v3);

  x = _mm_mullo_epi32(u0, cospi32);
  y = _mm_mullo_epi32(u2, cospi32);
  v0 = _mm_add_epi32(x, y);
  v0 = _mm_add_epi32(v0, rnding);
  v0 = _mm_srai_epi32(v0, bit);

  v1 = _mm_sub_epi32(x, y);
  v1 = _mm_add_epi32(v1, rnding);
  v1 = _mm_srai_epi32(v1, bit);

  x = _mm_mullo_epi32(u1, cospi48);
  y = _mm_mullo_epi32(u3, cospim16);
  v2 = _mm_add_epi32(x, y);
  v2 = _mm_add_epi32(v2, rnding);
  v2 = _mm_srai_epi32(v2, bit);

  x = _mm_mullo_epi32(u1, cospi16);
  y = _mm_mullo_epi32(u3, cospi48);
  v3 = _mm_add_epi32(x, y);
  v3 = _mm_add_epi32(v3, rnding);
  v3 = _mm_srai_epi32(v3, bit);

  // Stage 3
  addsub_sse4_1(v0, v3, out + 0, out + 3, &clamp_lo, &clamp_hi);
  addsub_sse4_1(v1, v2, out + 1, out + 2, &clamp_lo, &clamp_hi);

  if (!do_cols) {
    log_range = AOMMAX(16, bd + 6);
    clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
    clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);

    shift_and_clamp_sse4_1(out + 0, out + 3, &clamp_lo, &clamp_hi, out_shift);
    shift_and_clamp_sse4_1(out + 1, out + 2, &clamp_lo, &clamp_hi, out_shift);
  }
}

static void iadst4x4_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
                            int bd, int out_shift) {
  const int32_t *sinpi = sinpi_arr(bit);
  const __m128i zero = _mm_set1_epi32(0);
  __m128i rnding = _mm_set1_epi32(1 << (bit + 4 - 1));
  rnding = _mm_unpacklo_epi32(rnding, zero);
  const __m128i mul = _mm_set1_epi32(1 << 4);
  const __m128i sinpi1 = _mm_set1_epi32((int)sinpi[1]);
  const __m128i sinpi2 = _mm_set1_epi32((int)sinpi[2]);
  const __m128i sinpi3 = _mm_set1_epi32((int)sinpi[3]);
  const __m128i sinpi4 = _mm_set1_epi32((int)sinpi[4]);
  __m128i t;
  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
  __m128i x0, x1, x2, x3;
  __m128i u0, u1, u2, u3;
  __m128i v0, v1, v2, v3;
  __m128i u0_low, u1_low, u2_low, u3_low;
  __m128i u0_high, u1_high, u2_high, u3_high;

  v0 = _mm_unpacklo_epi32(in[0], in[1]);
  v1 = _mm_unpackhi_epi32(in[0], in[1]);
  v2 = _mm_unpacklo_epi32(in[2], in[3]);
  v3 = _mm_unpackhi_epi32(in[2], in[3]);

  x0 = _mm_unpacklo_epi64(v0, v2);
  x1 = _mm_unpackhi_epi64(v0, v2);
  x2 = _mm_unpacklo_epi64(v1, v3);
  x3 = _mm_unpackhi_epi64(v1, v3);

  s0 = _mm_mullo_epi32(x0, sinpi1);
  s1 = _mm_mullo_epi32(x0, sinpi2);
  s2 = _mm_mullo_epi32(x1, sinpi3);
  s3 = _mm_mullo_epi32(x2, sinpi4);
  s4 = _mm_mullo_epi32(x2, sinpi1);
  s5 = _mm_mullo_epi32(x3, sinpi2);
  s6 = _mm_mullo_epi32(x3, sinpi4);
  t = _mm_sub_epi32(x0, x2);
  s7 = _mm_add_epi32(t, x3);

  t = _mm_add_epi32(s0, s3);
  s0 = _mm_add_epi32(t, s5);
  t = _mm_sub_epi32(s1, s4);
  s1 = _mm_sub_epi32(t, s6);
  s3 = s2;
  s2 = _mm_mullo_epi32(s7, sinpi3);

  u0 = _mm_add_epi32(s0, s3);
  u1 = _mm_add_epi32(s1, s3);
  u2 = s2;
  t = _mm_add_epi32(s0, s1);
  u3 = _mm_sub_epi32(t, s3);

  // u0
  u0_low = _mm_mul_epi32(u0, mul);
  u0_low = _mm_add_epi64(u0_low, rnding);

  u0 = _mm_srli_si128(u0, 4);
  u0_high = _mm_mul_epi32(u0, mul);
  u0_high = _mm_add_epi64(u0_high, rnding);

  u0_low = _mm_srli_si128(u0_low, 2);
  u0_high = _mm_srli_si128(u0_high, 2);

  u0 = _mm_unpacklo_epi32(u0_low, u0_high);
  u0_high = _mm_unpackhi_epi32(u0_low, u0_high);
  u0 = _mm_unpacklo_epi64(u0, u0_high);

  // u1
  u1_low = _mm_mul_epi32(u1, mul);
  u1_low = _mm_add_epi64(u1_low, rnding);

  u1 = _mm_srli_si128(u1, 4);
  u1_high = _mm_mul_epi32(u1, mul);
  u1_high = _mm_add_epi64(u1_high, rnding);

  u1_low = _mm_srli_si128(u1_low, 2);
  u1_high = _mm_srli_si128(u1_high, 2);

  u1 = _mm_unpacklo_epi32(u1_low, u1_high);
  u1_high = _mm_unpackhi_epi32(u1_low, u1_high);
  u1 = _mm_unpacklo_epi64(u1, u1_high);

  // u2
  u2_low = _mm_mul_epi32(u2, mul);
  u2_low = _mm_add_epi64(u2_low, rnding);

  u2 = _mm_srli_si128(u2, 4);
  u2_high = _mm_mul_epi32(u2, mul);
  u2_high = _mm_add_epi64(u2_high, rnding);

  u2_low = _mm_srli_si128(u2_low, 2);
  u2_high = _mm_srli_si128(u2_high, 2);

  u2 = _mm_unpacklo_epi32(u2_low, u2_high);
  u2_high = _mm_unpackhi_epi32(u2_low, u2_high);
  u2 = _mm_unpacklo_epi64(u2, u2_high);

  // u3
  u3_low = _mm_mul_epi32(u3, mul);
  u3_low = _mm_add_epi64(u3_low, rnding);

  u3 = _mm_srli_si128(u3, 4);
  u3_high = _mm_mul_epi32(u3, mul);
  u3_high = _mm_add_epi64(u3_high, rnding);

  u3_low = _mm_srli_si128(u3_low, 2);
  u3_high = _mm_srli_si128(u3_high, 2);

  u3 = _mm_unpacklo_epi32(u3_low, u3_high);
  u3_high = _mm_unpackhi_epi32(u3_low, u3_high);
  u3 = _mm_unpacklo_epi64(u3, u3_high);

  out[0] = u0;
  out[1] = u1;
  out[2] = u2;
  out[3] = u3;

  if (!do_cols) {
    const int log_range = AOMMAX(16, bd + 6);
    const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
    const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
    round_shift_4x4(out, out_shift);
    highbd_clamp_epi32_sse4_1(out, out, &clamp_lo, &clamp_hi, 4);
  }
}

static void write_buffer_4x4(__m128i *in, uint16_t *output, int stride,
                             int fliplr, int flipud, int shift, int bd) {
  const __m128i zero = _mm_setzero_si128();
  __m128i u0, u1, u2, u3;
  __m128i v0, v1, v2, v3;

  round_shift_4x4(in, shift);

  v0 = _mm_loadl_epi64((__m128i const *)(output + 0 * stride));
  v1 = _mm_loadl_epi64((__m128i const *)(output + 1 * stride));
  v2 = _mm_loadl_epi64((__m128i const *)(output + 2 * stride));
  v3 = _mm_loadl_epi64((__m128i const *)(output + 3 * stride));

  v0 = _mm_unpacklo_epi16(v0, zero);
  v1 = _mm_unpacklo_epi16(v1, zero);
  v2 = _mm_unpacklo_epi16(v2, zero);
  v3 = _mm_unpacklo_epi16(v3, zero);

  if (fliplr) {
    in[0] = _mm_shuffle_epi32(in[0], 0x1B);
    in[1] = _mm_shuffle_epi32(in[1], 0x1B);
    in[2] = _mm_shuffle_epi32(in[2], 0x1B);
    in[3] = _mm_shuffle_epi32(in[3], 0x1B);
  }

  if (flipud) {
    u0 = _mm_add_epi32(in[3], v0);
    u1 = _mm_add_epi32(in[2], v1);
    u2 = _mm_add_epi32(in[1], v2);
    u3 = _mm_add_epi32(in[0], v3);
  } else {
    u0 = _mm_add_epi32(in[0], v0);
    u1 = _mm_add_epi32(in[1], v1);
    u2 = _mm_add_epi32(in[2], v2);
    u3 = _mm_add_epi32(in[3], v3);
  }

  v0 = _mm_packus_epi32(u0, u1);
  v2 = _mm_packus_epi32(u2, u3);

  u0 = highbd_clamp_epi16(v0, bd);
  u2 = highbd_clamp_epi16(v2, bd);

  v0 = _mm_unpacklo_epi64(u0, u0);
  v1 = _mm_unpackhi_epi64(u0, u0);
  v2 = _mm_unpacklo_epi64(u2, u2);
  v3 = _mm_unpackhi_epi64(u2, u2);

  _mm_storel_epi64((__m128i *)(output + 0 * stride), v0);
  _mm_storel_epi64((__m128i *)(output + 1 * stride), v1);
  _mm_storel_epi64((__m128i *)(output + 2 * stride), v2);
  _mm_storel_epi64((__m128i *)(output + 3 * stride), v3);
}

static void iidentity4_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
                              int bd, int out_shift) {
  (void)bit;
  __m128i v[4];
  __m128i zero = _mm_set1_epi32(0);
  __m128i fact = _mm_set1_epi32(NewSqrt2);
  __m128i offset = _mm_set1_epi32(1 << (NewSqrt2Bits - 1));
  __m128i a0_low, a1_low;
  __m128i a0_high, a1_high;

  offset = _mm_unpacklo_epi32(offset, zero);

  for (int i = 0; i < 4; i++) {
    a0_low = _mm_mul_epi32(in[i], fact);
    a0_low = _mm_add_epi32(a0_low, offset);
    a0_low = _mm_srli_epi64(a0_low, NewSqrt2Bits);

    a0_high = _mm_srli_si128(in[i], 4);
    a0_high = _mm_mul_epi32(a0_high, fact);
    a0_high = _mm_add_epi32(a0_high, offset);
    a0_high = _mm_srli_epi64(a0_high, NewSqrt2Bits);

    a1_low = _mm_unpacklo_epi32(a0_low, a0_high);
    a1_high = _mm_unpackhi_epi32(a0_low, a0_high);
    out[i] = _mm_unpacklo_epi64(a1_low, a1_high);
  }

  if (!do_cols) {
    const int log_range = AOMMAX(16, bd + 6);
    const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
    const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
    round_shift_4x4(out, out_shift);
    highbd_clamp_epi32_sse4_1(out, out, &clamp_lo, &clamp_hi, 4);
  }

  // Transpose for 4x4
  v[0] = _mm_unpacklo_epi32(out[0], out[1]);
  v[1] = _mm_unpackhi_epi32(out[0], out[1]);
  v[2] = _mm_unpacklo_epi32(out[2], out[3]);
  v[3] = _mm_unpackhi_epi32(out[2], out[3]);

  out[0] = _mm_unpacklo_epi64(v[0], v[2]);
  out[1] = _mm_unpackhi_epi64(v[0], v[2]);
  out[2] = _mm_unpacklo_epi64(v[1], v[3]);
  out[3] = _mm_unpackhi_epi64(v[1], v[3]);
}
void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output,
                                   int stride, TX_TYPE tx_type, int bd) {
  __m128i in[4];
  const int8_t *shift = av1_inv_txfm_shift_ls[TX_4X4];
  const int txw_idx = get_txw_idx(TX_4X4);
  const int txh_idx = get_txh_idx(TX_4X4);

  switch (tx_type) {
    case DCT_DCT:
      load_buffer_4x4(input, in);
      idct4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      idct4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case ADST_DCT:
      load_buffer_4x4(input, in);
      idct4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case DCT_ADST:
      load_buffer_4x4(input, in);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      idct4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case ADST_ADST:
      load_buffer_4x4(input, in);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case FLIPADST_DCT:
      load_buffer_4x4(input, in);
      idct4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 0, 1, -shift[1], bd);
      break;
    case DCT_FLIPADST:
      load_buffer_4x4(input, in);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      idct4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 1, 0, -shift[1], bd);
      break;
    case FLIPADST_FLIPADST:
      load_buffer_4x4(input, in);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 1, 1, -shift[1], bd);
      break;
    case ADST_FLIPADST:
      load_buffer_4x4(input, in);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 1, 0, -shift[1], bd);
      break;
    case FLIPADST_ADST:
      load_buffer_4x4(input, in);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 0, 1, -shift[1], bd);
      break;
    case IDTX:
      load_buffer_4x4(input, in);
      iidentity4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                        0);
      iidentity4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd,
                        0);
      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case V_DCT:
      load_buffer_4x4(input, in);
      iidentity4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                        0);
      idct4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case H_DCT:
      load_buffer_4x4(input, in);
      idct4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      iidentity4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd,
                        0);
      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case V_ADST:
      load_buffer_4x4(input, in);
      iidentity4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                        0);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case H_ADST:
      load_buffer_4x4(input, in);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      iidentity4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd,
                        0);
      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case V_FLIPADST:
      load_buffer_4x4(input, in);
      iidentity4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                        0);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_4x4(in, output, stride, 0, 1, -shift[1], bd);
      break;
    case H_FLIPADST:
      load_buffer_4x4(input, in);
      iadst4x4_sse4_1(in, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd, 0);
      iidentity4_sse4_1(in, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd,
                        0);
      write_buffer_4x4(in, output, stride, 1, 0, -shift[1], bd);
      break;
    default: assert(0);
  }
}

// 8x8
static void load_buffer_8x8(const int32_t *coeff, __m128i *in) {
  in[0] = _mm_load_si128((const __m128i *)(coeff + 0));
  in[1] = _mm_load_si128((const __m128i *)(coeff + 4));
  in[2] = _mm_load_si128((const __m128i *)(coeff + 8));
  in[3] = _mm_load_si128((const __m128i *)(coeff + 12));
  in[4] = _mm_load_si128((const __m128i *)(coeff + 16));
  in[5] = _mm_load_si128((const __m128i *)(coeff + 20));
  in[6] = _mm_load_si128((const __m128i *)(coeff + 24));
  in[7] = _mm_load_si128((const __m128i *)(coeff + 28));
  in[8] = _mm_load_si128((const __m128i *)(coeff + 32));
  in[9] = _mm_load_si128((const __m128i *)(coeff + 36));
  in[10] = _mm_load_si128((const __m128i *)(coeff + 40));
  in[11] = _mm_load_si128((const __m128i *)(coeff + 44));
  in[12] = _mm_load_si128((const __m128i *)(coeff + 48));
  in[13] = _mm_load_si128((const __m128i *)(coeff + 52));
  in[14] = _mm_load_si128((const __m128i *)(coeff + 56));
  in[15] = _mm_load_si128((const __m128i *)(coeff + 60));
}

static void idct8x8_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
                           int bd, int out_shift) {
  const int32_t *cospi = cospi_arr(bit);
  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
  __m128i x, y;
  int col;

  // Note:
  //  Even column: 0, 2, ..., 14
  //  Odd column: 1, 3, ..., 15
  //  one even column plus one odd column constructs one row (8 coeffs)
  //  total we have 8 rows (8x8).
  for (col = 0; col < 2; ++col) {
    // stage 0
    // stage 1
    // stage 2
    u0 = in[0 * 2 + col];
    u1 = in[4 * 2 + col];
    u2 = in[2 * 2 + col];
    u3 = in[6 * 2 + col];

    x = _mm_mullo_epi32(in[1 * 2 + col], cospi56);
    y = _mm_mullo_epi32(in[7 * 2 + col], cospim8);
    u4 = _mm_add_epi32(x, y);
    u4 = _mm_add_epi32(u4, rnding);
    u4 = _mm_srai_epi32(u4, bit);

    x = _mm_mullo_epi32(in[1 * 2 + col], cospi8);
    y = _mm_mullo_epi32(in[7 * 2 + col], cospi56);
    u7 = _mm_add_epi32(x, y);
    u7 = _mm_add_epi32(u7, rnding);
    u7 = _mm_srai_epi32(u7, bit);

    x = _mm_mullo_epi32(in[5 * 2 + col], cospi24);
    y = _mm_mullo_epi32(in[3 * 2 + col], cospim40);
    u5 = _mm_add_epi32(x, y);
    u5 = _mm_add_epi32(u5, rnding);
    u5 = _mm_srai_epi32(u5, bit);

    x = _mm_mullo_epi32(in[5 * 2 + col], cospi40);
    y = _mm_mullo_epi32(in[3 * 2 + col], cospi24);
    u6 = _mm_add_epi32(x, y);
    u6 = _mm_add_epi32(u6, rnding);
    u6 = _mm_srai_epi32(u6, bit);

    // stage 3
    x = _mm_mullo_epi32(u0, cospi32);
    y = _mm_mullo_epi32(u1, cospi32);
    v0 = _mm_add_epi32(x, y);
    v0 = _mm_add_epi32(v0, rnding);
    v0 = _mm_srai_epi32(v0, bit);

    v1 = _mm_sub_epi32(x, y);
    v1 = _mm_add_epi32(v1, rnding);
    v1 = _mm_srai_epi32(v1, bit);

    x = _mm_mullo_epi32(u2, cospi48);
    y = _mm_mullo_epi32(u3, cospim16);
    v2 = _mm_add_epi32(x, y);
    v2 = _mm_add_epi32(v2, rnding);
    v2 = _mm_srai_epi32(v2, bit);

    x = _mm_mullo_epi32(u2, cospi16);
    y = _mm_mullo_epi32(u3, cospi48);
    v3 = _mm_add_epi32(x, y);
    v3 = _mm_add_epi32(v3, rnding);
    v3 = _mm_srai_epi32(v3, bit);

    addsub_sse4_1(u4, u5, &v4, &v5, &clamp_lo, &clamp_hi);
    addsub_sse4_1(u7, u6, &v7, &v6, &clamp_lo, &clamp_hi);

    // stage 4
    addsub_sse4_1(v0, v3, &u0, &u3, &clamp_lo, &clamp_hi);
    addsub_sse4_1(v1, v2, &u1, &u2, &clamp_lo, &clamp_hi);
    u4 = v4;
    u7 = v7;

    x = _mm_mullo_epi32(v5, cospi32);
    y = _mm_mullo_epi32(v6, cospi32);
    u6 = _mm_add_epi32(y, x);
    u6 = _mm_add_epi32(u6, rnding);
    u6 = _mm_srai_epi32(u6, bit);

    u5 = _mm_sub_epi32(y, x);
    u5 = _mm_add_epi32(u5, rnding);
    u5 = _mm_srai_epi32(u5, bit);

    // stage 5
    addsub_sse4_1(u0, u7, out + 0 * 2 + col, out + 7 * 2 + col, &clamp_lo,
                  &clamp_hi);
    addsub_sse4_1(u1, u6, out + 1 * 2 + col, out + 6 * 2 + col, &clamp_lo,
                  &clamp_hi);
    addsub_sse4_1(u2, u5, out + 2 * 2 + col, out + 5 * 2 + col, &clamp_lo,
                  &clamp_hi);
    addsub_sse4_1(u3, u4, out + 3 * 2 + col, out + 4 * 2 + col, &clamp_lo,
                  &clamp_hi);
  }

  if (!do_cols) {
    const int log_range_out = AOMMAX(16, bd + 6);
    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);
    round_shift_8x8(out, out_shift);
    highbd_clamp_epi32_sse4_1(out, out, &clamp_lo_out, &clamp_hi_out, 16);
  }
}

static void iadst8x8_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
                            int bd, int out_shift) {
  const int32_t *cospi = cospi_arr(bit);
  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
  const __m128i kZero = _mm_setzero_si128();
  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
  __m128i u[8], v[8], x;

  // Even 8 points: 0, 2, ..., 14
  // stage 0
  // stage 1
  // stage 2
  // (1)
  u[0] = _mm_mullo_epi32(in[14], cospi4);
  x = _mm_mullo_epi32(in[0], cospi60);
  u[0] = _mm_add_epi32(u[0], x);
  u[0] = _mm_add_epi32(u[0], rnding);
  u[0] = _mm_srai_epi32(u[0], bit);

  u[1] = _mm_mullo_epi32(in[14], cospi60);
  x = _mm_mullo_epi32(in[0], cospi4);
  u[1] = _mm_sub_epi32(u[1], x);
  u[1] = _mm_add_epi32(u[1], rnding);
  u[1] = _mm_srai_epi32(u[1], bit);

  // (2)
  u[2] = _mm_mullo_epi32(in[10], cospi20);
  x = _mm_mullo_epi32(in[4], cospi44);
  u[2] = _mm_add_epi32(u[2], x);
  u[2] = _mm_add_epi32(u[2], rnding);
  u[2] = _mm_srai_epi32(u[2], bit);

  u[3] = _mm_mullo_epi32(in[10], cospi44);
  x = _mm_mullo_epi32(in[4], cospi20);
  u[3] = _mm_sub_epi32(u[3], x);
  u[3] = _mm_add_epi32(u[3], rnding);
  u[3] = _mm_srai_epi32(u[3], bit);

  // (3)
  u[4] = _mm_mullo_epi32(in[6], cospi36);
  x = _mm_mullo_epi32(in[8], cospi28);
  u[4] = _mm_add_epi32(u[4], x);
  u[4] = _mm_add_epi32(u[4], rnding);
  u[4] = _mm_srai_epi32(u[4], bit);

  u[5] = _mm_mullo_epi32(in[6], cospi28);
  x = _mm_mullo_epi32(in[8], cospi36);
  u[5] = _mm_sub_epi32(u[5], x);
  u[5] = _mm_add_epi32(u[5], rnding);
  u[5] = _mm_srai_epi32(u[5], bit);

  // (4)
  u[6] = _mm_mullo_epi32(in[2], cospi52);
  x = _mm_mullo_epi32(in[12], cospi12);
  u[6] = _mm_add_epi32(u[6], x);
  u[6] = _mm_add_epi32(u[6], rnding);
  u[6] = _mm_srai_epi32(u[6], bit);

  u[7] = _mm_mullo_epi32(in[2], cospi12);
  x = _mm_mullo_epi32(in[12], cospi52);
  u[7] = _mm_sub_epi32(u[7], x);
  u[7] = _mm_add_epi32(u[7], rnding);
  u[7] = _mm_srai_epi32(u[7], bit);

  // stage 3
  addsub_sse4_1(u[0], u[4], &v[0], &v[4], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[1], u[5], &v[1], &v[5], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[2], u[6], &v[2], &v[6], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[3], u[7], &v[3], &v[7], &clamp_lo, &clamp_hi);

  // stage 4
  u[0] = v[0];
  u[1] = v[1];
  u[2] = v[2];
  u[3] = v[3];

  u[4] = _mm_mullo_epi32(v[4], cospi16);
  x = _mm_mullo_epi32(v[5], cospi48);
  u[4] = _mm_add_epi32(u[4], x);
  u[4] = _mm_add_epi32(u[4], rnding);
  u[4] = _mm_srai_epi32(u[4], bit);

  u[5] = _mm_mullo_epi32(v[4], cospi48);
  x = _mm_mullo_epi32(v[5], cospi16);
  u[5] = _mm_sub_epi32(u[5], x);
  u[5] = _mm_add_epi32(u[5], rnding);
  u[5] = _mm_srai_epi32(u[5], bit);

  u[6] = _mm_mullo_epi32(v[6], cospim48);
  x = _mm_mullo_epi32(v[7], cospi16);
  u[6] = _mm_add_epi32(u[6], x);
  u[6] = _mm_add_epi32(u[6], rnding);
  u[6] = _mm_srai_epi32(u[6], bit);

  u[7] = _mm_mullo_epi32(v[6], cospi16);
  x = _mm_mullo_epi32(v[7], cospim48);
  u[7] = _mm_sub_epi32(u[7], x);
  u[7] = _mm_add_epi32(u[7], rnding);
  u[7] = _mm_srai_epi32(u[7], bit);

  // stage 5
  addsub_sse4_1(u[0], u[2], &v[0], &v[2], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[1], u[3], &v[1], &v[3], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[4], u[6], &v[4], &v[6], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[5], u[7], &v[5], &v[7], &clamp_lo, &clamp_hi);

  // stage 6
  u[0] = v[0];
  u[1] = v[1];
  u[4] = v[4];
  u[5] = v[5];

  v[0] = _mm_mullo_epi32(v[2], cospi32);
  x = _mm_mullo_epi32(v[3], cospi32);
  u[2] = _mm_add_epi32(v[0], x);
  u[2] = _mm_add_epi32(u[2], rnding);
  u[2] = _mm_srai_epi32(u[2], bit);

  u[3] = _mm_sub_epi32(v[0], x);
  u[3] = _mm_add_epi32(u[3], rnding);
  u[3] = _mm_srai_epi32(u[3], bit);

  v[0] = _mm_mullo_epi32(v[6], cospi32);
  x = _mm_mullo_epi32(v[7], cospi32);
  u[6] = _mm_add_epi32(v[0], x);
  u[6] = _mm_add_epi32(u[6], rnding);
  u[6] = _mm_srai_epi32(u[6], bit);

  u[7] = _mm_sub_epi32(v[0], x);
  u[7] = _mm_add_epi32(u[7], rnding);
  u[7] = _mm_srai_epi32(u[7], bit);

  // stage 7
  if (do_cols) {
    out[0] = u[0];
    out[2] = _mm_sub_epi32(kZero, u[4]);
    out[4] = u[6];
    out[6] = _mm_sub_epi32(kZero, u[2]);
    out[8] = u[3];
    out[10] = _mm_sub_epi32(kZero, u[7]);
    out[12] = u[5];
    out[14] = _mm_sub_epi32(kZero, u[1]);
  } else {
    const int log_range_out = AOMMAX(16, bd + 6);
    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);

    neg_shift_sse4_1(u[0], u[4], out + 0, out + 2, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
    neg_shift_sse4_1(u[6], u[2], out + 4, out + 6, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
    neg_shift_sse4_1(u[3], u[7], out + 8, out + 10, &clamp_lo_out,
                     &clamp_hi_out, out_shift);
    neg_shift_sse4_1(u[5], u[1], out + 12, out + 14, &clamp_lo_out,
                     &clamp_hi_out, out_shift);
  }

  // Odd 8 points: 1, 3, ..., 15
  // stage 0
  // stage 1
  // stage 2
  // (1)
  u[0] = _mm_mullo_epi32(in[15], cospi4);
  x = _mm_mullo_epi32(in[1], cospi60);
  u[0] = _mm_add_epi32(u[0], x);
  u[0] = _mm_add_epi32(u[0], rnding);
  u[0] = _mm_srai_epi32(u[0], bit);

  u[1] = _mm_mullo_epi32(in[15], cospi60);
  x = _mm_mullo_epi32(in[1], cospi4);
  u[1] = _mm_sub_epi32(u[1], x);
  u[1] = _mm_add_epi32(u[1], rnding);
  u[1] = _mm_srai_epi32(u[1], bit);

  // (2)
  u[2] = _mm_mullo_epi32(in[11], cospi20);
  x = _mm_mullo_epi32(in[5], cospi44);
  u[2] = _mm_add_epi32(u[2], x);
  u[2] = _mm_add_epi32(u[2], rnding);
  u[2] = _mm_srai_epi32(u[2], bit);

  u[3] = _mm_mullo_epi32(in[11], cospi44);
  x = _mm_mullo_epi32(in[5], cospi20);
  u[3] = _mm_sub_epi32(u[3], x);
  u[3] = _mm_add_epi32(u[3], rnding);
  u[3] = _mm_srai_epi32(u[3], bit);

  // (3)
  u[4] = _mm_mullo_epi32(in[7], cospi36);
  x = _mm_mullo_epi32(in[9], cospi28);
  u[4] = _mm_add_epi32(u[4], x);
  u[4] = _mm_add_epi32(u[4], rnding);
  u[4] = _mm_srai_epi32(u[4], bit);

  u[5] = _mm_mullo_epi32(in[7], cospi28);
  x = _mm_mullo_epi32(in[9], cospi36);
  u[5] = _mm_sub_epi32(u[5], x);
  u[5] = _mm_add_epi32(u[5], rnding);
  u[5] = _mm_srai_epi32(u[5], bit);

  // (4)
  u[6] = _mm_mullo_epi32(in[3], cospi52);
  x = _mm_mullo_epi32(in[13], cospi12);
  u[6] = _mm_add_epi32(u[6], x);
  u[6] = _mm_add_epi32(u[6], rnding);
  u[6] = _mm_srai_epi32(u[6], bit);

  u[7] = _mm_mullo_epi32(in[3], cospi12);
  x = _mm_mullo_epi32(in[13], cospi52);
  u[7] = _mm_sub_epi32(u[7], x);
  u[7] = _mm_add_epi32(u[7], rnding);
  u[7] = _mm_srai_epi32(u[7], bit);

  // stage 3
  addsub_sse4_1(u[0], u[4], &v[0], &v[4], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[1], u[5], &v[1], &v[5], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[2], u[6], &v[2], &v[6], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[3], u[7], &v[3], &v[7], &clamp_lo, &clamp_hi);

  // stage 4
  u[0] = v[0];
  u[1] = v[1];
  u[2] = v[2];
  u[3] = v[3];

  u[4] = _mm_mullo_epi32(v[4], cospi16);
  x = _mm_mullo_epi32(v[5], cospi48);
  u[4] = _mm_add_epi32(u[4], x);
  u[4] = _mm_add_epi32(u[4], rnding);
  u[4] = _mm_srai_epi32(u[4], bit);

  u[5] = _mm_mullo_epi32(v[4], cospi48);
  x = _mm_mullo_epi32(v[5], cospi16);
  u[5] = _mm_sub_epi32(u[5], x);
  u[5] = _mm_add_epi32(u[5], rnding);
  u[5] = _mm_srai_epi32(u[5], bit);

  u[6] = _mm_mullo_epi32(v[6], cospim48);
  x = _mm_mullo_epi32(v[7], cospi16);
  u[6] = _mm_add_epi32(u[6], x);
  u[6] = _mm_add_epi32(u[6], rnding);
  u[6] = _mm_srai_epi32(u[6], bit);

  u[7] = _mm_mullo_epi32(v[6], cospi16);
  x = _mm_mullo_epi32(v[7], cospim48);
  u[7] = _mm_sub_epi32(u[7], x);
  u[7] = _mm_add_epi32(u[7], rnding);
  u[7] = _mm_srai_epi32(u[7], bit);

  // stage 5
  addsub_sse4_1(u[0], u[2], &v[0], &v[2], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[1], u[3], &v[1], &v[3], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[4], u[6], &v[4], &v[6], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[5], u[7], &v[5], &v[7], &clamp_lo, &clamp_hi);

  // stage 6
  u[0] = v[0];
  u[1] = v[1];
  u[4] = v[4];
  u[5] = v[5];

  v[0] = _mm_mullo_epi32(v[2], cospi32);
  x = _mm_mullo_epi32(v[3], cospi32);
  u[2] = _mm_add_epi32(v[0], x);
  u[2] = _mm_add_epi32(u[2], rnding);
  u[2] = _mm_srai_epi32(u[2], bit);

  u[3] = _mm_sub_epi32(v[0], x);
  u[3] = _mm_add_epi32(u[3], rnding);
  u[3] = _mm_srai_epi32(u[3], bit);

  v[0] = _mm_mullo_epi32(v[6], cospi32);
  x = _mm_mullo_epi32(v[7], cospi32);
  u[6] = _mm_add_epi32(v[0], x);
  u[6] = _mm_add_epi32(u[6], rnding);
  u[6] = _mm_srai_epi32(u[6], bit);

  u[7] = _mm_sub_epi32(v[0], x);
  u[7] = _mm_add_epi32(u[7], rnding);
  u[7] = _mm_srai_epi32(u[7], bit);

  // stage 7
  if (do_cols) {
    out[1] = u[0];
    out[3] = _mm_sub_epi32(kZero, u[4]);
    out[5] = u[6];
    out[7] = _mm_sub_epi32(kZero, u[2]);
    out[9] = u[3];
    out[11] = _mm_sub_epi32(kZero, u[7]);
    out[13] = u[5];
    out[15] = _mm_sub_epi32(kZero, u[1]);
  } else {
    const int log_range_out = AOMMAX(16, bd + 6);
    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);

    neg_shift_sse4_1(u[0], u[4], out + 1, out + 3, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
    neg_shift_sse4_1(u[6], u[2], out + 5, out + 7, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
    neg_shift_sse4_1(u[3], u[7], out + 9, out + 11, &clamp_lo_out,
                     &clamp_hi_out, out_shift);
    neg_shift_sse4_1(u[5], u[1], out + 13, out + 15, &clamp_lo_out,
                     &clamp_hi_out, out_shift);
  }
}

static void iidentity8_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
                              int bd, int out_shift) {
  (void)bit;
  out[0] = _mm_add_epi32(in[0], in[0]);
  out[1] = _mm_add_epi32(in[1], in[1]);
  out[2] = _mm_add_epi32(in[2], in[2]);
  out[3] = _mm_add_epi32(in[3], in[3]);
  out[4] = _mm_add_epi32(in[4], in[4]);
  out[5] = _mm_add_epi32(in[5], in[5]);
  out[6] = _mm_add_epi32(in[6], in[6]);
  out[7] = _mm_add_epi32(in[7], in[7]);

  if (!do_cols) {
    const int log_range = AOMMAX(16, bd + 6);
    const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
    const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
    round_shift_4x4(out, out_shift);
    round_shift_4x4(out + 4, out_shift);
    highbd_clamp_epi32_sse4_1(out, out, &clamp_lo, &clamp_hi, 8);
  }
}

static __m128i get_recon_8x8(const __m128i pred, __m128i res_lo, __m128i res_hi,
                             int fliplr, int bd) {
  __m128i x0, x1;
  const __m128i zero = _mm_setzero_si128();

  x0 = _mm_unpacklo_epi16(pred, zero);
  x1 = _mm_unpackhi_epi16(pred, zero);

  if (fliplr) {
    res_lo = _mm_shuffle_epi32(res_lo, 0x1B);
    res_hi = _mm_shuffle_epi32(res_hi, 0x1B);
    x0 = _mm_add_epi32(res_hi, x0);
    x1 = _mm_add_epi32(res_lo, x1);

  } else {
    x0 = _mm_add_epi32(res_lo, x0);
    x1 = _mm_add_epi32(res_hi, x1);
  }

  x0 = _mm_packus_epi32(x0, x1);
  return highbd_clamp_epi16(x0, bd);
}

static void write_buffer_8x8(__m128i *in, uint16_t *output, int stride,
                             int fliplr, int flipud, int shift, int bd) {
  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
  __m128i v0, v1, v2, v3, v4, v5, v6, v7;

  round_shift_8x8(in, shift);

  v0 = _mm_load_si128((__m128i const *)(output + 0 * stride));
  v1 = _mm_load_si128((__m128i const *)(output + 1 * stride));
  v2 = _mm_load_si128((__m128i const *)(output + 2 * stride));
  v3 = _mm_load_si128((__m128i const *)(output + 3 * stride));
  v4 = _mm_load_si128((__m128i const *)(output + 4 * stride));
  v5 = _mm_load_si128((__m128i const *)(output + 5 * stride));
  v6 = _mm_load_si128((__m128i const *)(output + 6 * stride));
  v7 = _mm_load_si128((__m128i const *)(output + 7 * stride));

  if (flipud) {
    u0 = get_recon_8x8(v0, in[14], in[15], fliplr, bd);
    u1 = get_recon_8x8(v1, in[12], in[13], fliplr, bd);
    u2 = get_recon_8x8(v2, in[10], in[11], fliplr, bd);
    u3 = get_recon_8x8(v3, in[8], in[9], fliplr, bd);
    u4 = get_recon_8x8(v4, in[6], in[7], fliplr, bd);
    u5 = get_recon_8x8(v5, in[4], in[5], fliplr, bd);
    u6 = get_recon_8x8(v6, in[2], in[3], fliplr, bd);
    u7 = get_recon_8x8(v7, in[0], in[1], fliplr, bd);
  } else {
    u0 = get_recon_8x8(v0, in[0], in[1], fliplr, bd);
    u1 = get_recon_8x8(v1, in[2], in[3], fliplr, bd);
    u2 = get_recon_8x8(v2, in[4], in[5], fliplr, bd);
    u3 = get_recon_8x8(v3, in[6], in[7], fliplr, bd);
    u4 = get_recon_8x8(v4, in[8], in[9], fliplr, bd);
    u5 = get_recon_8x8(v5, in[10], in[11], fliplr, bd);
    u6 = get_recon_8x8(v6, in[12], in[13], fliplr, bd);
    u7 = get_recon_8x8(v7, in[14], in[15], fliplr, bd);
  }

  _mm_store_si128((__m128i *)(output + 0 * stride), u0);
  _mm_store_si128((__m128i *)(output + 1 * stride), u1);
  _mm_store_si128((__m128i *)(output + 2 * stride), u2);
  _mm_store_si128((__m128i *)(output + 3 * stride), u3);
  _mm_store_si128((__m128i *)(output + 4 * stride), u4);
  _mm_store_si128((__m128i *)(output + 5 * stride), u5);
  _mm_store_si128((__m128i *)(output + 6 * stride), u6);
  _mm_store_si128((__m128i *)(output + 7 * stride), u7);
}

void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output,
                                   int stride, TX_TYPE tx_type, int bd) {
  __m128i in[16], out[16];
  const int8_t *shift = av1_inv_txfm_shift_ls[TX_8X8];
  const int txw_idx = get_txw_idx(TX_8X8);
  const int txh_idx = get_txh_idx(TX_8X8);

  switch (tx_type) {
    case DCT_DCT:
      load_buffer_8x8(input, in);
      transpose_8x8(in, out);
      idct8x8_sse4_1(out, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                     -shift[0]);
      transpose_8x8(in, out);
      idct8x8_sse4_1(out, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_8x8(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case DCT_ADST:
      load_buffer_8x8(input, in);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                      -shift[0]);
      transpose_8x8(in, out);
      idct8x8_sse4_1(out, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_8x8(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case ADST_DCT:
      load_buffer_8x8(input, in);
      transpose_8x8(in, out);
      idct8x8_sse4_1(out, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                     -shift[0]);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_8x8(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case ADST_ADST:
      load_buffer_8x8(input, in);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                      -shift[0]);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_8x8(in, output, stride, 0, 0, -shift[1], bd);
      break;
    case FLIPADST_DCT:
      load_buffer_8x8(input, in);
      transpose_8x8(in, out);
      idct8x8_sse4_1(out, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                     -shift[0]);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_8x8(in, output, stride, 0, 1, -shift[1], bd);
      break;
    case DCT_FLIPADST:
      load_buffer_8x8(input, in);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                      -shift[0]);
      transpose_8x8(in, out);
      idct8x8_sse4_1(out, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_8x8(in, output, stride, 1, 0, -shift[1], bd);
      break;
    case ADST_FLIPADST:
      load_buffer_8x8(input, in);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                      -shift[0]);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_8x8(in, output, stride, 1, 0, -shift[1], bd);
      break;
    case FLIPADST_FLIPADST:
      load_buffer_8x8(input, in);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                      -shift[0]);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_8x8(in, output, stride, 1, 1, -shift[1], bd);
      break;
    case FLIPADST_ADST:
      load_buffer_8x8(input, in);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
                      -shift[0]);
      transpose_8x8(in, out);
      iadst8x8_sse4_1(out, in, av1_inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
      write_buffer_8x8(in, output, stride, 0, 1, -shift[1], bd);
      break;
    default: assert(0);
  }
}

static void idct8x8_low1_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
                                int bd, int out_shift) {
  const int32_t *cospi = cospi_arr(bit);
  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
  __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
  __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
  __m128i x;

  // stage 0
  // stage 1
  // stage 2
  // stage 3
  x = _mm_mullo_epi32(in[0], cospi32);
  x = _mm_add_epi32(x, rnding);
  x = _mm_srai_epi32(x, bit);

  // stage 4
  // stage 5
  if (!do_cols) {
    const int log_range_out = AOMMAX(16, bd + 6);
    clamp_lo = _mm_set1_epi32(-(1 << (log_range_out - 1)));
    clamp_hi = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);

    __m128i offset = _mm_set1_epi32((1 << out_shift) >> 1);
    x = _mm_add_epi32(x, offset);
    x = _mm_sra_epi32(x, _mm_cvtsi32_si128(out_shift));
  }

  x = _mm_max_epi32(x, clamp_lo);
  x = _mm_min_epi32(x, clamp_hi);
  out[0] = x;
  out[1] = x;
  out[2] = x;
  out[3] = x;
  out[4] = x;
  out[5] = x;
  out[6] = x;
  out[7] = x;
}

static void idct8x8_new_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
                               int bd, int out_shift) {
  const int32_t *cospi = cospi_arr(bit);
  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
  __m128i x, y;

  // stage 0
  // stage 1
  // stage 2
  u0 = in[0];
  u1 = in[4];
  u2 = in[2];
  u3 = in[6];

  x = _mm_mullo_epi32(in[1], cospi56);
  y = _mm_mullo_epi32(in[7], cospim8);
  u4 = _mm_add_epi32(x, y);
  u4 = _mm_add_epi32(u4, rnding);
  u4 = _mm_srai_epi32(u4, bit);

  x = _mm_mullo_epi32(in[1], cospi8);
  y = _mm_mullo_epi32(in[7], cospi56);
  u7 = _mm_add_epi32(x, y);
  u7 = _mm_add_epi32(u7, rnding);
  u7 = _mm_srai_epi32(u7, bit);

  x = _mm_mullo_epi32(in[5], cospi24);
  y = _mm_mullo_epi32(in[3], cospim40);
  u5 = _mm_add_epi32(x, y);
  u5 = _mm_add_epi32(u5, rnding);
  u5 = _mm_srai_epi32(u5, bit);

  x = _mm_mullo_epi32(in[5], cospi40);
  y = _mm_mullo_epi32(in[3], cospi24);
  u6 = _mm_add_epi32(x, y);
  u6 = _mm_add_epi32(u6, rnding);
  u6 = _mm_srai_epi32(u6, bit);

  // stage 3
  x = _mm_mullo_epi32(u0, cospi32);
  y = _mm_mullo_epi32(u1, cospi32);
  v0 = _mm_add_epi32(x, y);
  v0 = _mm_add_epi32(v0, rnding);
  v0 = _mm_srai_epi32(v0, bit);

  v1 = _mm_sub_epi32(x, y);
  v1 = _mm_add_epi32(v1, rnding);
  v1 = _mm_srai_epi32(v1, bit);

  x = _mm_mullo_epi32(u2, cospi48);
  y = _mm_mullo_epi32(u3, cospim16);
  v2 = _mm_add_epi32(x, y);
  v2 = _mm_add_epi32(v2, rnding);
  v2 = _mm_srai_epi32(v2, bit);

  x = _mm_mullo_epi32(u2, cospi16);
  y = _mm_mullo_epi32(u3, cospi48);
  v3 = _mm_add_epi32(x, y);
  v3 = _mm_add_epi32(v3, rnding);
  v3 = _mm_srai_epi32(v3, bit);

  addsub_sse4_1(u4, u5, &v4, &v5, &clamp_lo, &clamp_hi);
  addsub_sse4_1(u7, u6, &v7, &v6, &clamp_lo, &clamp_hi);

  // stage 4
  addsub_sse4_1(v0, v3, &u0, &u3, &clamp_lo, &clamp_hi);
  addsub_sse4_1(v1, v2, &u1, &u2, &clamp_lo, &clamp_hi);
  u4 = v4;
  u7 = v7;

  x = _mm_mullo_epi32(v5, cospi32);
  y = _mm_mullo_epi32(v6, cospi32);
  u6 = _mm_add_epi32(y, x);
  u6 = _mm_add_epi32(u6, rnding);
  u6 = _mm_srai_epi32(u6, bit);

  u5 = _mm_sub_epi32(y, x);
  u5 = _mm_add_epi32(u5, rnding);
  u5 = _mm_srai_epi32(u5, bit);

  // stage 5
  addsub_sse4_1(u0, u7, out + 0, out + 7, &clamp_lo, &clamp_hi);
  addsub_sse4_1(u1, u6, out + 1, out + 6, &clamp_lo, &clamp_hi);
  addsub_sse4_1(u2, u5, out + 2, out + 5, &clamp_lo, &clamp_hi);
  addsub_sse4_1(u3, u4, out + 3, out + 4, &clamp_lo, &clamp_hi);

  if (!do_cols) {
    const int log_range_out = AOMMAX(16, bd + 6);
    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);

    round_shift_4x4(out, out_shift);
    round_shift_4x4(out + 4, out_shift);
    highbd_clamp_epi32_sse4_1(out, out, &clamp_lo_out, &clamp_hi_out, 8);
  }
}

static void iadst8x8_low1_sse4_1(__m128i *in, __m128i *out, int bit,
                                 int do_cols, int bd, int out_shift) {
  const int32_t *cospi = cospi_arr(bit);
  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
  const __m128i kZero = _mm_setzero_si128();
  __m128i u[8], x;

  // stage 0
  // stage 1
  // stage 2

  x = _mm_mullo_epi32(in[0], cospi60);
  u[0] = _mm_add_epi32(x, rnding);
  u[0] = _mm_srai_epi32(u[0], bit);

  x = _mm_mullo_epi32(in[0], cospi4);
  u[1] = _mm_sub_epi32(kZero, x);
  u[1] = _mm_add_epi32(u[1], rnding);
  u[1] = _mm_srai_epi32(u[1], bit);

  // stage 3
  // stage 4
  __m128i temp1, temp2;
  temp1 = _mm_mullo_epi32(u[0], cospi16);
  x = _mm_mullo_epi32(u[1], cospi48);
  temp1 = _mm_add_epi32(temp1, x);
  temp1 = _mm_add_epi32(temp1, rnding);
  temp1 = _mm_srai_epi32(temp1, bit);
  u[4] = temp1;

  temp2 = _mm_mullo_epi32(u[0], cospi48);
  x = _mm_mullo_epi32(u[1], cospi16);
  u[5] = _mm_sub_epi32(temp2, x);
  u[5] = _mm_add_epi32(u[5], rnding);
  u[5] = _mm_srai_epi32(u[5], bit);

  // stage 5
  // stage 6
  temp1 = _mm_mullo_epi32(u[0], cospi32);
  x = _mm_mullo_epi32(u[1], cospi32);
  u[2] = _mm_add_epi32(temp1, x);
  u[2] = _mm_add_epi32(u[2], rnding);
  u[2] = _mm_srai_epi32(u[2], bit);

  u[3] = _mm_sub_epi32(temp1, x);
  u[3] = _mm_add_epi32(u[3], rnding);
  u[3] = _mm_srai_epi32(u[3], bit);

  temp1 = _mm_mullo_epi32(u[4], cospi32);
  x = _mm_mullo_epi32(u[5], cospi32);
  u[6] = _mm_add_epi32(temp1, x);
  u[6] = _mm_add_epi32(u[6], rnding);
  u[6] = _mm_srai_epi32(u[6], bit);

  u[7] = _mm_sub_epi32(temp1, x);
  u[7] = _mm_add_epi32(u[7], rnding);
  u[7] = _mm_srai_epi32(u[7], bit);

  // stage 7
  if (do_cols) {
    out[0] = u[0];
    out[1] = _mm_sub_epi32(kZero, u[4]);
    out[2] = u[6];
    out[3] = _mm_sub_epi32(kZero, u[2]);
    out[4] = u[3];
    out[5] = _mm_sub_epi32(kZero, u[7]);
    out[6] = u[5];
    out[7] = _mm_sub_epi32(kZero, u[1]);
  } else {
    const int log_range_out = AOMMAX(16, bd + 6);
    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);

    neg_shift_sse4_1(u[0], u[4], out + 0, out + 1, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
    neg_shift_sse4_1(u[6], u[2], out + 2, out + 3, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
    neg_shift_sse4_1(u[3], u[7], out + 4, out + 5, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
    neg_shift_sse4_1(u[5], u[1], out + 6, out + 7, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
  }
}

static void iadst8x8_new_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
                                int bd, int out_shift) {
  const int32_t *cospi = cospi_arr(bit);
  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
  const __m128i kZero = _mm_setzero_si128();
  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
  __m128i u[8], v[8], x;

  // stage 0
  // stage 1
  // stage 2

  u[0] = _mm_mullo_epi32(in[7], cospi4);
  x = _mm_mullo_epi32(in[0], cospi60);
  u[0] = _mm_add_epi32(u[0], x);
  u[0] = _mm_add_epi32(u[0], rnding);
  u[0] = _mm_srai_epi32(u[0], bit);

  u[1] = _mm_mullo_epi32(in[7], cospi60);
  x = _mm_mullo_epi32(in[0], cospi4);
  u[1] = _mm_sub_epi32(u[1], x);
  u[1] = _mm_add_epi32(u[1], rnding);
  u[1] = _mm_srai_epi32(u[1], bit);

  // (2)
  u[2] = _mm_mullo_epi32(in[5], cospi20);
  x = _mm_mullo_epi32(in[2], cospi44);
  u[2] = _mm_add_epi32(u[2], x);
  u[2] = _mm_add_epi32(u[2], rnding);
  u[2] = _mm_srai_epi32(u[2], bit);

  u[3] = _mm_mullo_epi32(in[5], cospi44);
  x = _mm_mullo_epi32(in[2], cospi20);
  u[3] = _mm_sub_epi32(u[3], x);
  u[3] = _mm_add_epi32(u[3], rnding);
  u[3] = _mm_srai_epi32(u[3], bit);

  // (3)
  u[4] = _mm_mullo_epi32(in[3], cospi36);
  x = _mm_mullo_epi32(in[4], cospi28);
  u[4] = _mm_add_epi32(u[4], x);
  u[4] = _mm_add_epi32(u[4], rnding);
  u[4] = _mm_srai_epi32(u[4], bit);

  u[5] = _mm_mullo_epi32(in[3], cospi28);
  x = _mm_mullo_epi32(in[4], cospi36);
  u[5] = _mm_sub_epi32(u[5], x);
  u[5] = _mm_add_epi32(u[5], rnding);
  u[5] = _mm_srai_epi32(u[5], bit);

  // (4)
  u[6] = _mm_mullo_epi32(in[1], cospi52);
  x = _mm_mullo_epi32(in[6], cospi12);
  u[6] = _mm_add_epi32(u[6], x);
  u[6] = _mm_add_epi32(u[6], rnding);
  u[6] = _mm_srai_epi32(u[6], bit);

  u[7] = _mm_mullo_epi32(in[1], cospi12);
  x = _mm_mullo_epi32(in[6], cospi52);
  u[7] = _mm_sub_epi32(u[7], x);
  u[7] = _mm_add_epi32(u[7], rnding);
  u[7] = _mm_srai_epi32(u[7], bit);

  // stage 3
  addsub_sse4_1(u[0], u[4], &v[0], &v[4], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[1], u[5], &v[1], &v[5], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[2], u[6], &v[2], &v[6], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[3], u[7], &v[3], &v[7], &clamp_lo, &clamp_hi);

  // stage 4
  u[0] = v[0];
  u[1] = v[1];
  u[2] = v[2];
  u[3] = v[3];

  u[4] = _mm_mullo_epi32(v[4], cospi16);
  x = _mm_mullo_epi32(v[5], cospi48);
  u[4] = _mm_add_epi32(u[4], x);
  u[4] = _mm_add_epi32(u[4], rnding);
  u[4] = _mm_srai_epi32(u[4], bit);

  u[5] = _mm_mullo_epi32(v[4], cospi48);
  x = _mm_mullo_epi32(v[5], cospi16);
  u[5] = _mm_sub_epi32(u[5], x);
  u[5] = _mm_add_epi32(u[5], rnding);
  u[5] = _mm_srai_epi32(u[5], bit);

  u[6] = _mm_mullo_epi32(v[6], cospim48);
  x = _mm_mullo_epi32(v[7], cospi16);
  u[6] = _mm_add_epi32(u[6], x);
  u[6] = _mm_add_epi32(u[6], rnding);
  u[6] = _mm_srai_epi32(u[6], bit);

  u[7] = _mm_mullo_epi32(v[6], cospi16);
  x = _mm_mullo_epi32(v[7], cospim48);
  u[7] = _mm_sub_epi32(u[7], x);
  u[7] = _mm_add_epi32(u[7], rnding);
  u[7] = _mm_srai_epi32(u[7], bit);

  // stage 5
  addsub_sse4_1(u[0], u[2], &v[0], &v[2], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[1], u[3], &v[1], &v[3], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[4], u[6], &v[4], &v[6], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[5], u[7], &v[5], &v[7], &clamp_lo, &clamp_hi);

  // stage 6
  u[0] = v[0];
  u[1] = v[1];
  u[4] = v[4];
  u[5] = v[5];

  v[0] = _mm_mullo_epi32(v[2], cospi32);
  x = _mm_mullo_epi32(v[3], cospi32);
  u[2] = _mm_add_epi32(v[0], x);
  u[2] = _mm_add_epi32(u[2], rnding);
  u[2] = _mm_srai_epi32(u[2], bit);

  u[3] = _mm_sub_epi32(v[0], x);
  u[3] = _mm_add_epi32(u[3], rnding);
  u[3] = _mm_srai_epi32(u[3], bit);

  v[0] = _mm_mullo_epi32(v[6], cospi32);
  x = _mm_mullo_epi32(v[7], cospi32);
  u[6] = _mm_add_epi32(v[0], x);
  u[6] = _mm_add_epi32(u[6], rnding);
  u[6] = _mm_srai_epi32(u[6], bit);

  u[7] = _mm_sub_epi32(v[0], x);
  u[7] = _mm_add_epi32(u[7], rnding);
  u[7] = _mm_srai_epi32(u[7], bit);

  // stage 7
  if (do_cols) {
    out[0] = u[0];
    out[1] = _mm_sub_epi32(kZero, u[4]);
    out[2] = u[6];
    out[3] = _mm_sub_epi32(kZero, u[2]);
    out[4] = u[3];
    out[5] = _mm_sub_epi32(kZero, u[7]);
    out[6] = u[5];
    out[7] = _mm_sub_epi32(kZero, u[1]);
  } else {
    const int log_range_out = AOMMAX(16, bd + 6);
    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);

    neg_shift_sse4_1(u[0], u[4], out + 0, out + 1, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
    neg_shift_sse4_1(u[6], u[2], out + 2, out + 3, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
    neg_shift_sse4_1(u[3], u[7], out + 4, out + 5, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
    neg_shift_sse4_1(u[5], u[1], out + 6, out + 7, &clamp_lo_out, &clamp_hi_out,
                     out_shift);
  }
}

static void idct16x16_low1_sse4_1(__m128i *in, __m128i *out, int bit,
                                  int do_cols, int bd, int out_shift) {
  const int32_t *cospi = cospi_arr(bit);
  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
  int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
  __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
  __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
  // stage 0
  // stage 1
  // stage 2
  // stage 3
  // stage 4
  in[0] = _mm_mullo_epi32(in[0], cospi32);
  in[0] = _mm_add_epi32(in[0], rnding);
  in[0] = _mm_srai_epi32(in[0], bit);

  // stage 5
  // stage 6
  // stage 7
  if (!do_cols) {
    log_range = AOMMAX(16, bd + 6);
    clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
    clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
    if (out_shift != 0) {
      __m128i offset = _mm_set1_epi32((1 << out_shift) >> 1);
      in[0] = _mm_add_epi32(in[0], offset);
      in[0] = _mm_sra_epi32(in[0], _mm_cvtsi32_si128(out_shift));
    }
  }

  in[0] = _mm_max_epi32(in[0], clamp_lo);
  in[0] = _mm_min_epi32(in[0], clamp_hi);
  out[0] = in[0];
  out[1] = in[0];
  out[2] = in[0];
  out[3] = in[0];
  out[4] = in[0];
  out[5] = in[0];
  out[6] = in[0];
  out[7] = in[0];
  out[8] = in[0];
  out[9] = in[0];
  out[10] = in[0];
  out[11] = in[0];
  out[12] = in[0];
  out[13] = in[0];
  out[14] = in[0];
  out[15] = in[0];
}

static void idct16x16_low8_sse4_1(__m128i *in, __m128i *out, int bit,
                                  int do_cols, int bd, int out_shift) {
  const int32_t *cospi = cospi_arr(bit);
  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
  const __m128i cospim36 = _mm_set1_epi32(-cospi[36]);
  const __m128i cospim52 = _mm_set1_epi32(-cospi[52]);
  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
  __m128i u[16], x, y;
  // stage 0
  // stage 1
  u[0] = in[0];
  u[2] = in[4];
  u[4] = in[2];
  u[6] = in[6];
  u[8] = in[1];
  u[10] = in[5];
  u[12] = in[3];
  u[14] = in[7];

  // stage 2
  u[15] = half_btf_0_sse4_1(&cospi4, &u[8], &rnding, bit);
  u[8] = half_btf_0_sse4_1(&cospi60, &u[8], &rnding, bit);

  u[9] = half_btf_0_sse4_1(&cospim36, &u[14], &rnding, bit);
  u[14] = half_btf_0_sse4_1(&cospi28, &u[14], &rnding, bit);

  u[13] = half_btf_0_sse4_1(&cospi20, &u[10], &rnding, bit);
  u[10] = half_btf_0_sse4_1(&cospi44, &u[10], &rnding, bit);

  u[11] = half_btf_0_sse4_1(&cospim52, &u[12], &rnding, bit);
  u[12] = half_btf_0_sse4_1(&cospi12, &u[12], &rnding, bit);

  // stage 3
  u[7] = half_btf_0_sse4_1(&cospi8, &u[4], &rnding, bit);
  u[4] = half_btf_0_sse4_1(&cospi56, &u[4], &rnding, bit);
  u[5] = half_btf_0_sse4_1(&cospim40, &u[6], &rnding, bit);
  u[6] = half_btf_0_sse4_1(&cospi24, &u[6], &rnding, bit);

  addsub_sse4_1(u[8], u[9], &u[8], &u[9], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[11], u[10], &u[11], &u[10], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[12], u[13], &u[12], &u[13], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[15], u[14], &u[15], &u[14], &clamp_lo, &clamp_hi);

  // stage 4
  x = _mm_mullo_epi32(u[0], cospi32);
  u[0] = _mm_add_epi32(x, rnding);
  u[0] = _mm_srai_epi32(u[0], bit);
  u[1] = u[0];

  u[3] = half_btf_0_sse4_1(&cospi16, &u[2], &rnding, bit);
  u[2] = half_btf_0_sse4_1(&cospi48, &u[2], &rnding, bit);

  addsub_sse4_1(u[4], u[5], &u[4], &u[5], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[7], u[6], &u[7], &u[6], &clamp_lo, &clamp_hi);

  x = half_btf_sse4_1(&cospim16, &u[9], &cospi48, &u[14], &rnding, bit);
  u[14] = half_btf_sse4_1(&cospi48, &u[9], &cospi16, &u[14], &rnding, bit);
  u[9] = x;
  y = half_btf_sse4_1(&cospim48, &u[10], &cospim16, &u[13], &rnding, bit);
  u[13] = half_btf_sse4_1(&cospim16, &u[10], &cospi48, &u[13], &rnding, bit);
  u[10] = y;

  // stage 5
  addsub_sse4_1(u[0], u[3], &u[0], &u[3], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[1], u[2], &u[1], &u[2], &clamp_lo, &clamp_hi);

  x = _mm_mullo_epi32(u[5], cospi32);
  y = _mm_mullo_epi32(u[6], cospi32);
  u[5] = _mm_sub_epi32(y, x);
  u[5] = _mm_add_epi32(u[5], rnding);
  u[5] = _mm_srai_epi32(u[5], bit);

  u[6] = _mm_add_epi32(y, x);
  u[6] = _mm_add_epi32(u[6], rnding);
  u[6] = _mm_srai_epi32(u[6], bit);

  addsub_sse4_1(u[8], u[11], &u[8], &u[11], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[9], u[10], &u[9], &u[10], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[15], u[12], &u[15], &u[12], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[14], u[13], &u[14], &u[13], &clamp_lo, &clamp_hi);

  // stage 6
  addsub_sse4_1(u[0], u[7], &u[0], &u[7], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[1], u[6], &u[1], &u[6], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[2], u[5], &u[2], &u[5], &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[3], u[4], &u[3], &u[4], &clamp_lo, &clamp_hi);

  x = _mm_mullo_epi32(u[10], cospi32);
  y = _mm_mullo_epi32(u[13], cospi32);
  u[10] = _mm_sub_epi32(y, x);
  u[10] = _mm_add_epi32(u[10], rnding);
  u[10] = _mm_srai_epi32(u[10], bit);

  u[13] = _mm_add_epi32(x, y);
  u[13] = _mm_add_epi32(u[13], rnding);
  u[13] = _mm_srai_epi32(u[13], bit);

  x = _mm_mullo_epi32(u[11], cospi32);
  y = _mm_mullo_epi32(u[12], cospi32);
  u[11] = _mm_sub_epi32(y, x);
  u[11] = _mm_add_epi32(u[11], rnding);
  u[11] = _mm_srai_epi32(u[11], bit);

  u[12] = _mm_add_epi32(x, y);
  u[12] = _mm_add_epi32(u[12], rnding);
  u[12] = _mm_srai_epi32(u[12], bit);
  // stage 7
  addsub_sse4_1(u[0], u[15], out + 0, out + 15, &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[1], u[14], out + 1, out + 14, &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[2], u[13], out + 2, out + 13, &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[3], u[12], out + 3, out + 12, &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[4], u[11], out + 4, out + 11, &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[5], u[10], out + 5, out + 10, &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[6], u[9], out + 6, out + 9, &clamp_lo, &clamp_hi);
  addsub_sse4_1(u[7], u[8], out + 7, out + 8, &clamp_lo, &clamp_hi);

  if (!do_cols) {
    const int log_range_out = AOMMAX(16, bd + 6);
    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);
    round_shift_8x8(out, out_shift);
    highbd_clamp_epi32_sse4_1(out, out, &clamp_lo_out, &clamp_hi_out, 16);
  }
}

static void iadst16x16_low1_sse4_1(__m128i *in, __m128i *out, int bit,
                                   int do_cols, int bd, int out_shift) {
  const int32_t *cospi = cospi_arr(bit);
  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
  const __m128i zero = _mm_setzero_si128();
  __m128i v[16], x, y, temp1, temp2;
  // stage 0
  // stage 1
  // stage 2
  x = _mm_mullo_epi32(in[0], cospi62);
  v[0] = _mm_add_epi32(x, rnding);
  v[0] = _mm_srai_epi32(v[0], bit);

  x = _mm_mullo_epi32(in[0], cospi2);
  v[1] = _mm_sub_epi32(zero, x);
  v[1] = _mm_add_epi32(v[1], rnding);
  v[1] = _mm_srai_epi32(v[1], bit);

  // stage 3
  v[8] = v[0];
  v[9] = v[1];

  // stage 4
  temp1 = _mm_mullo_epi32(v[8], cospi8);
  x = _mm_mullo_epi32(v[9], cospi56);
  temp1 = _mm_add_epi32(temp1, x);
  temp1 = _mm_add_epi32(temp1, rnding);
  temp1 = _mm_srai_epi32(temp1, bit);

  temp2 = _mm_mullo_epi32(v[8], cospi56);
  x = _mm_mullo_epi32(v[9], cospi8);
  temp2 = _mm_sub_epi32(temp2, x);
  temp2 = _mm_add_epi32(temp2, rnding);
  temp2 = _mm_srai_epi32(temp2, bit);
  v[8] = temp1;
  v[9] = temp2;

  // stage 5
  v[4] = v[0];
  v[5] = v[1];
  v[12] = v[8];
  v[13] = v[9];

  // stage 6
  temp1 = _mm_mullo_epi32(v[4], cospi16);
  x = _mm_mullo_epi32(v[5], cospi48);
  temp1 = _mm_add_epi32(temp1, x);
  temp1 = _mm_add_epi32(temp1, rnding);
  temp1 = _mm_srai_epi32(temp1, bit);

  temp2 = _mm_mullo_epi32(v[4], cospi48);
  x = _mm_mullo_epi32(v[5], cospi16);
  temp2 = _mm_sub_epi32(temp2, x);
  temp2 = _mm_add_epi32(temp2, rnding);
  temp2 = _mm_srai_epi32(temp2, bit);
  v[4] = temp1;
  v[5] = temp2;

  temp1 = _mm_mullo_epi32(v[12], cospi16);
  x = _mm_mullo_epi32(v[13], cospi48);
  temp1 = _mm_add_epi32(temp1, x);
  temp1 = _mm_add_epi32(temp1, rnding);
  temp1 = _mm_srai_epi32(temp1, bit);

  temp2 = _mm_mullo_epi32(v[12], cospi48);
  x = _mm_mullo_epi32(v[13], cospi16);
  temp2 = _mm_sub_epi32(temp2, x);
  temp2 = _mm_add_epi32(temp2, rnding);
  temp2 = _mm_srai_epi32(temp2, bit);
  v[12] = temp1;
  v[13] = temp2;

  // stage 7
  v[2] = v[0];
  v[3] = v[1];
  v[6] = v[4];
  v[7] = v[5];
  v[10] = v[8];
  v[11] = v[9];
  v[14] = v[12];
  v[15] = v[13];

  // stage 8
  y = _mm_mullo_epi32(v[2], cospi32);
  x = _mm_mullo_epi32(v[3], cospi32);
  v[2] = _mm_add_epi32(y, x);
  v[2] = _mm_add_epi32(v[2], rnding);
  v[2] = _mm_srai_epi32(v[2], bit);

  v[3] = _mm_sub_epi32(y, x);
  v[3] = _mm_add_epi32(v[3], rnding);
  v[3] = _mm_srai_epi32(v[3], bit);

  y = _mm_mullo_epi32(v[6], cospi32);
  x = _mm_mullo_epi32(v[7], cospi32);
  v[6] = _mm_add_epi32(y, x);
  v[6] = _mm_add_epi32(v[6], rnding);
  v[6] = _mm_srai_epi32(v[6], bit);

  v[7] = _mm_sub_epi32(y, x);
  v[7] = _mm_add_epi32(v[7], rnding);
  v[7] = _mm_srai_epi32(v[7], bit);

  y = _mm_mullo_epi32(v[10], cospi32);
  x = _mm_mullo_epi32(v[11], cospi32);
  v[10] = _mm_add_epi32(y, x);
  v[10] = _mm_add_epi32(v[10], rnding);
  v[10] = _mm_srai_epi32(v[10], bit);

  v[11] = _mm_sub_epi32(y, x);
  v[11] = _mm_add_epi32(v[11], rnding);
  v[11] = _mm_srai_epi32(v[11], bit);

  y = _mm_mullo_epi32(v[14], cospi32);
  x = _mm_mullo_epi32(v[15], cospi32);
  v[14] = _mm_add_epi32(y, x);
  v[14] = _mm_add_epi32(v[14], rnding);
  v[14] = _mm_srai_epi32(v[14], bit);

  v[15] = _mm_sub_epi32(y, x);
  v[15] = _mm_add_epi32(v[15], rnding);
  v[15] = _mm_srai_epi32(v[15], bit);

  // stage 9
  if (do_cols) {
    out[0] = v[0];
    out[1] = _mm_sub_epi32(zero, v[8]);
    out[2] = v[12];
    out[3] = _mm_sub_epi32(zero, v[4]);
    out[4] = v[6];
    out[5] = _mm_sub_epi32(zero, v[14]);
    out[6] = v[10];
    out[7] = _mm_sub_epi32(zero, v[2]);
    out[8] = v[3];
    out[9] = _mm_sub_epi32(zero, v[11]);
    out[10] = v[15</