/*
 * Copyright (c) 2017, 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 <immintrin.h>

#include "./aom_dsp_rtcd.h"

static INLINE __m256i dc_sum_64(const uint8_t *ref) {
  const __m256i x0 = _mm256_loadu_si256((const __m256i *)ref);
  const __m256i x1 = _mm256_loadu_si256((const __m256i *)(ref + 32));
  const __m256i zero = _mm256_setzero_si256();
  __m256i y0 = _mm256_sad_epu8(x0, zero);
  __m256i y1 = _mm256_sad_epu8(x1, zero);
  y0 = _mm256_add_epi64(y0, y1);
  __m256i u0 = _mm256_permute2x128_si256(y0, y0, 1);
  y0 = _mm256_add_epi64(u0, y0);
  u0 = _mm256_unpackhi_epi64(y0, y0);
  return _mm256_add_epi16(y0, u0);
}

static INLINE __m256i dc_sum_32(const uint8_t *ref) {
  const __m256i x = _mm256_loadu_si256((const __m256i *)ref);
  const __m256i zero = _mm256_setzero_si256();
  __m256i y = _mm256_sad_epu8(x, zero);
  __m256i u = _mm256_permute2x128_si256(y, y, 1);
  y = _mm256_add_epi64(u, y);
  u = _mm256_unpackhi_epi64(y, y);
  return _mm256_add_epi16(y, u);
}

static INLINE void row_store_32xh(const __m256i *r, int height, uint8_t *dst,
                                  ptrdiff_t stride) {
  for (int i = 0; i < height; ++i) {
    _mm256_storeu_si256((__m256i *)dst, *r);
    dst += stride;
  }
}

static INLINE void row_store_32x2xh(const __m256i *r0, const __m256i *r1,
                                    int height, uint8_t *dst,
                                    ptrdiff_t stride) {
  for (int i = 0; i < height; ++i) {
    _mm256_storeu_si256((__m256i *)dst, *r0);
    _mm256_storeu_si256((__m256i *)(dst + 32), *r1);
    dst += stride;
  }
}

static INLINE void row_store_64xh(const __m256i *r, int height, uint8_t *dst,
                                  ptrdiff_t stride) {
  for (int i = 0; i < height; ++i) {
    _mm256_storeu_si256((__m256i *)dst, *r);
    _mm256_storeu_si256((__m256i *)(dst + 32), *r);
    dst += stride;
  }
}

void aom_dc_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                 const uint8_t *above, const uint8_t *left) {
  const __m256i sum_above = dc_sum_32(above);
  __m256i sum_left = dc_sum_32(left);
  sum_left = _mm256_add_epi16(sum_left, sum_above);
  const __m256i thirtytwo = _mm256_set1_epi16(32);
  sum_left = _mm256_add_epi16(sum_left, thirtytwo);
  sum_left = _mm256_srai_epi16(sum_left, 6);
  const __m256i zero = _mm256_setzero_si256();
  __m256i row = _mm256_shuffle_epi8(sum_left, zero);
  row_store_32xh(&row, 32, dst, stride);
}

void aom_dc_top_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  __m256i sum = dc_sum_32(above);
  (void)left;

  const __m256i sixteen = _mm256_set1_epi16(16);
  sum = _mm256_add_epi16(sum, sixteen);
  sum = _mm256_srai_epi16(sum, 5);
  const __m256i zero = _mm256_setzero_si256();
  __m256i row = _mm256_shuffle_epi8(sum, zero);
  row_store_32xh(&row, 32, dst, stride);
}

void aom_dc_left_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                      const uint8_t *above,
                                      const uint8_t *left) {
  __m256i sum = dc_sum_32(left);
  (void)above;

  const __m256i sixteen = _mm256_set1_epi16(16);
  sum = _mm256_add_epi16(sum, sixteen);
  sum = _mm256_srai_epi16(sum, 5);
  const __m256i zero = _mm256_setzero_si256();
  __m256i row = _mm256_shuffle_epi8(sum, zero);
  row_store_32xh(&row, 32, dst, stride);
}

void aom_dc_128_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  (void)above;
  (void)left;
  const __m256i row = _mm256_set1_epi8((uint8_t)0x80);
  row_store_32xh(&row, 32, dst, stride);
}

void aom_v_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                const uint8_t *above, const uint8_t *left) {
  const __m256i row = _mm256_loadu_si256((const __m256i *)above);
  (void)left;
  row_store_32xh(&row, 32, dst, stride);
}

// There are 32 rows togeter. This function does line:
// 0,1,2,3, and 16,17,18,19. The next call would do
// 4,5,6,7, and 20,21,22,23. So 4 times of calling
// would finish 32 rows.
static INLINE void h_predictor_32x8line(const __m256i *row, uint8_t *dst,
                                        ptrdiff_t stride) {
  __m256i t[4];
  __m256i m = _mm256_setzero_si256();
  const __m256i inc = _mm256_set1_epi8(4);
  int i;

  for (i = 0; i < 4; i++) {
    t[i] = _mm256_shuffle_epi8(*row, m);
    __m256i r0 = _mm256_permute2x128_si256(t[i], t[i], 0);
    __m256i r1 = _mm256_permute2x128_si256(t[i], t[i], 0x11);
    _mm256_storeu_si256((__m256i *)dst, r0);
    _mm256_storeu_si256((__m256i *)(dst + (stride << 4)), r1);
    dst += stride;
    m = _mm256_add_epi8(m, inc);
  }
}

void aom_h_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                const uint8_t *above, const uint8_t *left) {
  (void)above;
  const __m256i left_col = _mm256_loadu_si256((__m256i const *)left);

  __m256i u = _mm256_unpacklo_epi8(left_col, left_col);

  __m256i v = _mm256_unpacklo_epi8(u, u);
  h_predictor_32x8line(&v, dst, stride);
  dst += stride << 2;

  v = _mm256_unpackhi_epi8(u, u);
  h_predictor_32x8line(&v, dst, stride);
  dst += stride << 2;

  u = _mm256_unpackhi_epi8(left_col, left_col);

  v = _mm256_unpacklo_epi8(u, u);
  h_predictor_32x8line(&v, dst, stride);
  dst += stride << 2;

  v = _mm256_unpackhi_epi8(u, u);
  h_predictor_32x8line(&v, dst, stride);
}

// -----------------------------------------------------------------------------
// Rectangle

// TODO(luoyi) The following two functions are shared with intrapred_sse2.c.
// Use a header file, intrapred_common_x86.h
static INLINE __m128i dc_sum_16_sse2(const uint8_t *ref) {
  __m128i x = _mm_load_si128((__m128i const *)ref);
  const __m128i zero = _mm_setzero_si128();
  x = _mm_sad_epu8(x, zero);
  const __m128i high = _mm_unpackhi_epi64(x, x);
  return _mm_add_epi16(x, high);
}

static INLINE __m128i dc_sum_32_sse2(const uint8_t *ref) {
  __m128i x0 = _mm_load_si128((__m128i const *)ref);
  __m128i x1 = _mm_load_si128((__m128i const *)(ref + 16));
  const __m128i zero = _mm_setzero_si128();
  x0 = _mm_sad_epu8(x0, zero);
  x1 = _mm_sad_epu8(x1, zero);
  x0 = _mm_add_epi16(x0, x1);
  const __m128i high = _mm_unpackhi_epi64(x0, x0);
  return _mm_add_epi16(x0, high);
}

void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                 const uint8_t *above, const uint8_t *left) {
  const __m128i top_sum = dc_sum_32_sse2(above);
  __m128i left_sum = dc_sum_16_sse2(left);
  left_sum = _mm_add_epi16(top_sum, left_sum);
  uint32_t sum = _mm_cvtsi128_si32(left_sum);
  sum += 24;
  sum /= 48;
  const __m256i row = _mm256_set1_epi8((uint8_t)sum);
  row_store_32xh(&row, 16, dst, stride);
}

void aom_dc_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                 const uint8_t *above, const uint8_t *left) {
  const __m256i sum_above = dc_sum_32(above);
  __m256i sum_left = dc_sum_64(left);
  sum_left = _mm256_add_epi16(sum_left, sum_above);
  uint32_t sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left));
  sum += 48;
  sum /= 96;
  const __m256i row = _mm256_set1_epi8((uint8_t)sum);
  row_store_32xh(&row, 64, dst, stride);
}

void aom_dc_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                 const uint8_t *above, const uint8_t *left) {
  const __m256i sum_above = dc_sum_64(above);
  __m256i sum_left = dc_sum_64(left);
  sum_left = _mm256_add_epi16(sum_left, sum_above);
  uint32_t sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left));
  sum += 64;
  sum /= 128;
  const __m256i row = _mm256_set1_epi8((uint8_t)sum);
  row_store_64xh(&row, 64, dst, stride);
}

void aom_dc_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                 const uint8_t *above, const uint8_t *left) {
  const __m256i sum_above = dc_sum_64(above);
  __m256i sum_left = dc_sum_32(left);
  sum_left = _mm256_add_epi16(sum_left, sum_above);
  uint32_t sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left));
  sum += 48;
  sum /= 96;
  const __m256i row = _mm256_set1_epi8((uint8_t)sum);
  row_store_64xh(&row, 32, dst, stride);
}

void aom_dc_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                 const uint8_t *above, const uint8_t *left) {
  const __m256i sum_above = dc_sum_64(above);
  __m256i sum_left = _mm256_castsi128_si256(dc_sum_16_sse2(left));
  sum_left = _mm256_add_epi16(sum_left, sum_above);
  uint32_t sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left));
  sum += 40;
  sum /= 80;
  const __m256i row = _mm256_set1_epi8((uint8_t)sum);
  row_store_64xh(&row, 16, dst, stride);
}

void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  __m256i sum = dc_sum_32(above);
  (void)left;

  const __m256i sixteen = _mm256_set1_epi16(16);
  sum = _mm256_add_epi16(sum, sixteen);
  sum = _mm256_srai_epi16(sum, 5);
  const __m256i zero = _mm256_setzero_si256();
  __m256i row = _mm256_shuffle_epi8(sum, zero);
  row_store_32xh(&row, 16, dst, stride);
}

void aom_dc_top_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  __m256i sum = dc_sum_32(above);
  (void)left;

  const __m256i sixteen = _mm256_set1_epi16(16);
  sum = _mm256_add_epi16(sum, sixteen);
  sum = _mm256_srai_epi16(sum, 5);
  const __m256i zero = _mm256_setzero_si256();
  __m256i row = _mm256_shuffle_epi8(sum, zero);
  row_store_32xh(&row, 64, dst, stride);
}

void aom_dc_top_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  __m256i sum = dc_sum_64(above);
  (void)left;

  const __m256i thirtytwo = _mm256_set1_epi16(32);
  sum = _mm256_add_epi16(sum, thirtytwo);
  sum = _mm256_srai_epi16(sum, 6);
  const __m256i zero = _mm256_setzero_si256();
  __m256i row = _mm256_shuffle_epi8(sum, zero);
  row_store_64xh(&row, 64, dst, stride);
}

void aom_dc_top_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  __m256i sum = dc_sum_64(above);
  (void)left;

  const __m256i thirtytwo = _mm256_set1_epi16(32);
  sum = _mm256_add_epi16(sum, thirtytwo);
  sum = _mm256_srai_epi16(sum, 6);
  const __m256i zero = _mm256_setzero_si256();
  __m256i row = _mm256_shuffle_epi8(sum, zero);
  row_store_64xh(&row, 32, dst, stride);
}

void aom_dc_top_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  __m256i sum = dc_sum_64(above);
  (void)left;

  const __m256i thirtytwo = _mm256_set1_epi16(32);
  sum = _mm256_add_epi16(sum, thirtytwo);
  sum = _mm256_srai_epi16(sum, 6);
  const __m256i zero = _mm256_setzero_si256();
  __m256i row = _mm256_shuffle_epi8(sum, zero);
  row_store_64xh(&row, 16, dst, stride);
}

void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                      const uint8_t *above,
                                      const uint8_t *left) {
  __m128i sum = dc_sum_16_sse2(left);
  (void)above;

  const __m128i eight = _mm_set1_epi16(8);
  sum = _mm_add_epi16(sum, eight);
  sum = _mm_srai_epi16(sum, 4);
  const __m128i zero = _mm_setzero_si128();
  const __m128i r = _mm_shuffle_epi8(sum, zero);
  const __m256i row = _mm256_inserti128_si256(_mm256_castsi128_si256(r), r, 1);
  row_store_32xh(&row, 16, dst, stride);
}

void aom_dc_left_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                      const uint8_t *above,
                                      const uint8_t *left) {
  __m256i sum = dc_sum_64(left);
  (void)above;

  const __m256i thirtytwo = _mm256_set1_epi16(32);
  sum = _mm256_add_epi16(sum, thirtytwo);
  sum = _mm256_srai_epi16(sum, 6);
  const __m256i zero = _mm256_setzero_si256();
  __m256i row = _mm256_shuffle_epi8(sum, zero);
  row_store_32xh(&row, 64, dst, stride);
}

void aom_dc_left_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                      const uint8_t *above,
                                      const uint8_t *left) {
  __m256i sum = dc_sum_64(left);
  (void)above;

  const __m256i thirtytwo = _mm256_set1_epi16(32);
  sum = _mm256_add_epi16(sum, thirtytwo);
  sum = _mm256_srai_epi16(sum, 6);
  const __m256i zero = _mm256_setzero_si256();
  __m256i row = _mm256_shuffle_epi8(sum, zero);
  row_store_64xh(&row, 64, dst, stride);
}

void aom_dc_left_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                      const uint8_t *above,
                                      const uint8_t *left) {
  __m256i sum = dc_sum_32(left);
  (void)above;

  const __m256i sixteen = _mm256_set1_epi16(16);
  sum = _mm256_add_epi16(sum, sixteen);
  sum = _mm256_srai_epi16(sum, 5);
  const __m256i zero = _mm256_setzero_si256();
  __m256i row = _mm256_shuffle_epi8(sum, zero);
  row_store_64xh(&row, 32, dst, stride);
}

void aom_dc_left_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                      const uint8_t *above,
                                      const uint8_t *left) {
  __m128i sum = dc_sum_16_sse2(left);
  (void)above;

  const __m128i eight = _mm_set1_epi16(8);
  sum = _mm_add_epi16(sum, eight);
  sum = _mm_srai_epi16(sum, 4);
  const __m128i zero = _mm_setzero_si128();
  const __m128i r = _mm_shuffle_epi8(sum, zero);
  const __m256i row = _mm256_inserti128_si256(_mm256_castsi128_si256(r), r, 1);
  row_store_64xh(&row, 16, dst, stride);
}

void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  (void)above;
  (void)left;
  const __m256i row = _mm256_set1_epi8((uint8_t)0x80);
  row_store_32xh(&row, 16, dst, stride);
}

void aom_dc_128_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  (void)above;
  (void)left;
  const __m256i row = _mm256_set1_epi8((uint8_t)0x80);
  row_store_32xh(&row, 64, dst, stride);
}

void aom_dc_128_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  (void)above;
  (void)left;
  const __m256i row = _mm256_set1_epi8((uint8_t)0x80);
  row_store_64xh(&row, 64, dst, stride);
}

void aom_dc_128_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  (void)above;
  (void)left;
  const __m256i row = _mm256_set1_epi8((uint8_t)0x80);
  row_store_64xh(&row, 32, dst, stride);
}

void aom_dc_128_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                     const uint8_t *above,
                                     const uint8_t *left) {
  (void)above;
  (void)left;
  const __m256i row = _mm256_set1_epi8((uint8_t)0x80);
  row_store_64xh(&row, 16, dst, stride);
}

void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                const uint8_t *above, const uint8_t *left) {
  const __m256i row = _mm256_loadu_si256((const __m256i *)above);
  (void)left;
  row_store_32xh(&row, 16, dst, stride);
}

void aom_v_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                const uint8_t *above, const uint8_t *left) {
  const __m256i row = _mm256_loadu_si256((const __m256i *)above);
  (void)left;
  row_store_32xh(&row, 64, dst, stride);
}

void aom_v_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                const uint8_t *above, const uint8_t *left) {
  const __m256i row0 = _mm256_loadu_si256((const __m256i *)above);
  const __m256i row1 = _mm256_loadu_si256((const __m256i *)(above + 32));
  (void)left;
  row_store_32x2xh(&row0, &row1, 64, dst, stride);
}

void aom_v_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                const uint8_t *above, const uint8_t *left) {
  const __m256i row0 = _mm256_loadu_si256((const __m256i *)above);
  const __m256i row1 = _mm256_loadu_si256((const __m256i *)(above + 32));
  (void)left;
  row_store_32x2xh(&row0, &row1, 32, dst, stride);
}

void aom_v_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                const uint8_t *above, const uint8_t *left) {
  const __m256i row0 = _mm256_loadu_si256((const __m256i *)above);
  const __m256i row1 = _mm256_loadu_si256((const __m256i *)(above + 32));
  (void)left;
  row_store_32x2xh(&row0, &row1, 16, dst, stride);
}

// -----------------------------------------------------------------------------
// PAETH_PRED

// Return 16 16-bit pixels in one row (__m256i)
static INLINE __m256i paeth_pred(const __m256i *left, const __m256i *top,
                                 const __m256i *topleft) {
  const __m256i base =
      _mm256_sub_epi16(_mm256_add_epi16(*top, *left), *topleft);

  __m256i pl = _mm256_abs_epi16(_mm256_sub_epi16(base, *left));
  __m256i pt = _mm256_abs_epi16(_mm256_sub_epi16(base, *top));
  __m256i ptl = _mm256_abs_epi16(_mm256_sub_epi16(base, *topleft));

  __m256i mask1 = _mm256_cmpgt_epi16(pl, pt);
  mask1 = _mm256_or_si256(mask1, _mm256_cmpgt_epi16(pl, ptl));
  __m256i mask2 = _mm256_cmpgt_epi16(pt, ptl);

  pl = _mm256_andnot_si256(mask1, *left);

  ptl = _mm256_and_si256(mask2, *topleft);
  pt = _mm256_andnot_si256(mask2, *top);
  pt = _mm256_or_si256(pt, ptl);
  pt = _mm256_and_si256(mask1, pt);

  return _mm256_or_si256(pt, pl);
}

// Return 16 8-bit pixels in one row (__m128i)
static INLINE __m128i paeth_16x1_pred(const __m256i *left, const __m256i *top,
                                      const __m256i *topleft) {
  const __m256i p0 = paeth_pred(left, top, topleft);
  const __m256i p1 = _mm256_permute4x64_epi64(p0, 0xe);
  const __m256i p = _mm256_packus_epi16(p0, p1);
  return _mm256_castsi256_si128(p);
}

static INLINE __m256i get_top_vector(const uint8_t *above) {
  const __m128i x = _mm_load_si128((const __m128i *)above);
  const __m128i zero = _mm_setzero_si128();
  const __m128i t0 = _mm_unpacklo_epi8(x, zero);
  const __m128i t1 = _mm_unpackhi_epi8(x, zero);
  return _mm256_inserti128_si256(_mm256_castsi128_si256(t0), t1, 1);
}

void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t stride,
                                   const uint8_t *above, const uint8_t *left) {
  __m128i x = _mm_loadl_epi64((const __m128i *)left);
  const __m256i l = _mm256_inserti128_si256(_mm256_castsi128_si256(x), x, 1);
  const __m256i tl16 = _mm256_set1_epi16((uint16_t)above[-1]);
  __m256i rep = _mm256_set1_epi16(0x8000);
  const __m256i one = _mm256_set1_epi16(1);
  const __m256i top = get_top_vector(above);

  int i;
  for (i = 0; i < 8; ++i) {
    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
    const __m128i row = paeth_16x1_pred(&l16, &top, &tl16);

    _mm_store_si128((__m128i *)dst, row);
    dst += stride;
    rep = _mm256_add_epi16(rep, one);
  }
}

static INLINE __m256i get_left_vector(const uint8_t *left) {
  const __m128i x = _mm_load_si128((const __m128i *)left);
  return _mm256_inserti128_si256(_mm256_castsi128_si256(x), x, 1);
}

void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                    const uint8_t *above, const uint8_t *left) {
  const __m256i l = get_left_vector(left);
  const __m256i tl16 = _mm256_set1_epi16((uint16_t)above[-1]);
  __m256i rep = _mm256_set1_epi16(0x8000);
  const __m256i one = _mm256_set1_epi16(1);
  const __m256i top = get_top_vector(above);

  int i;
  for (i = 0; i < 16; ++i) {
    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
    const __m128i row = paeth_16x1_pred(&l16, &top, &tl16);

    _mm_store_si128((__m128i *)dst, row);
    dst += stride;
    rep = _mm256_add_epi16(rep, one);
  }
}

void aom_paeth_predictor_16x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                    const uint8_t *above, const uint8_t *left) {
  __m256i l = get_left_vector(left);
  const __m256i tl16 = _mm256_set1_epi16((uint16_t)above[-1]);
  __m256i rep = _mm256_set1_epi16(0x8000);
  const __m256i one = _mm256_set1_epi16(1);
  const __m256i top = get_top_vector(above);

  int i;
  for (i = 0; i < 16; ++i) {
    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
    const __m128i row = paeth_16x1_pred(&l16, &top, &tl16);

    _mm_store_si128((__m128i *)dst, row);
    dst += stride;
    rep = _mm256_add_epi16(rep, one);
  }

  l = get_left_vector(left + 16);
  rep = _mm256_set1_epi16(0x8000);
  for (i = 0; i < 16; ++i) {
    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
    const __m128i row = paeth_16x1_pred(&l16, &top, &tl16);

    _mm_store_si128((__m128i *)dst, row);
    dst += stride;
    rep = _mm256_add_epi16(rep, one);
  }
}

void aom_paeth_predictor_16x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                    const uint8_t *above, const uint8_t *left) {
  const __m256i tl16 = _mm256_set1_epi16((uint16_t)above[-1]);
  const __m256i one = _mm256_set1_epi16(1);
  const __m256i top = get_top_vector(above);

  for (int j = 0; j < 4; ++j) {
    const __m256i l = get_left_vector(left + j * 16);
    __m256i rep = _mm256_set1_epi16(0x8000);
    for (int i = 0; i < 16; ++i) {
      const __m256i l16 = _mm256_shuffle_epi8(l, rep);
      const __m128i row = paeth_16x1_pred(&l16, &top, &tl16);

      _mm_store_si128((__m128i *)dst, row);
      dst += stride;
      rep = _mm256_add_epi16(rep, one);
    }
  }
}

// Return 32 8-bit pixels in one row (__m256i)
static INLINE __m256i paeth_32x1_pred(const __m256i *left, const __m256i *top0,
                                      const __m256i *top1,
                                      const __m256i *topleft) {
  __m256i p0 = paeth_pred(left, top0, topleft);
  __m256i p1 = _mm256_permute4x64_epi64(p0, 0xe);
  const __m256i x0 = _mm256_packus_epi16(p0, p1);

  p0 = paeth_pred(left, top1, topleft);
  p1 = _mm256_permute4x64_epi64(p0, 0xe);
  const __m256i x1 = _mm256_packus_epi16(p0, p1);

  return _mm256_permute2x128_si256(x0, x1, 0x20);
}

void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                    const uint8_t *above, const uint8_t *left) {
  const __m256i l = get_left_vector(left);
  const __m256i t0 = get_top_vector(above);
  const __m256i t1 = get_top_vector(above + 16);
  const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]);
  __m256i rep = _mm256_set1_epi16(0x8000);
  const __m256i one = _mm256_set1_epi16(1);

  int i;
  for (i = 0; i < 16; ++i) {
    const __m256i l16 = _mm256_shuffle_epi8(l, rep);

    const __m256i r = paeth_32x1_pred(&l16, &t0, &t1, &tl);

    _mm256_storeu_si256((__m256i *)dst, r);

    dst += stride;
    rep = _mm256_add_epi16(rep, one);
  }
}

void aom_paeth_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                    const uint8_t *above, const uint8_t *left) {
  __m256i l = get_left_vector(left);
  const __m256i t0 = get_top_vector(above);
  const __m256i t1 = get_top_vector(above + 16);
  const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]);
  __m256i rep = _mm256_set1_epi16(0x8000);
  const __m256i one = _mm256_set1_epi16(1);

  int i;
  for (i = 0; i < 16; ++i) {
    const __m256i l16 = _mm256_shuffle_epi8(l, rep);

    const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
    const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);

    _mm_store_si128((__m128i *)dst, r0);
    _mm_store_si128((__m128i *)(dst + 16), r1);

    dst += stride;
    rep = _mm256_add_epi16(rep, one);
  }

  l = get_left_vector(left + 16);
  rep = _mm256_set1_epi16(0x8000);
  for (i = 0; i < 16; ++i) {
    const __m256i l16 = _mm256_shuffle_epi8(l, rep);

    const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
    const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);

    _mm_store_si128((__m128i *)dst, r0);
    _mm_store_si128((__m128i *)(dst + 16), r1);

    dst += stride;
    rep = _mm256_add_epi16(rep, one);
  }
}

void aom_paeth_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                    const uint8_t *above, const uint8_t *left) {
  const __m256i t0 = get_top_vector(above);
  const __m256i t1 = get_top_vector(above + 16);
  const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]);
  const __m256i one = _mm256_set1_epi16(1);

  int i, j;
  for (j = 0; j < 4; ++j) {
    const __m256i l = get_left_vector(left + j * 16);
    __m256i rep = _mm256_set1_epi16(0x8000);
    for (i = 0; i < 16; ++i) {
      const __m256i l16 = _mm256_shuffle_epi8(l, rep);

      const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
      const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);

      _mm_store_si128((__m128i *)dst, r0);
      _mm_store_si128((__m128i *)(dst + 16), r1);

      dst += stride;
      rep = _mm256_add_epi16(rep, one);
    }
  }
}

void aom_paeth_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
                                    const uint8_t *above, const uint8_t *left) {
  const __m256i t0 = get_top_vector(above);
  const __m256i t1 = get_top_vector(above + 16);
  const __m256i t2 = get_top_vector(above + 32);
  const __m256i t3 = get_top_vector(above + 48);
  const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]);
  const __m256i one = _mm256_set1_epi16(1);

  int i, j;
  for (j = 0; j < 2; ++j) {
    const __m256i l = get_left_vector(left + j * 16);
    __m256i rep = _mm256_set1_epi16(0x8000);
    for (i = 0; i < 16; ++i) {
      const __m256i l16 = _mm256_shuffle_epi8(l, rep);

      const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
      const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);
      const __m128i r2 = paeth_16x1_pred(&l16, &t2, &tl);
      const __m128i r3 = paeth_16x1_pred(&l16, &t3, &tl);

      _mm_store_si128((__m128i *)dst, r0);
      _mm_store_si128((__m128i *)(dst + 16), r1);
      _mm_store_si128((__m128i *)(dst + 32), r2);
      _mm_store_si128((__m128i *)(dst + 48), r3);

      dst += stride;
      rep = _mm256_add_epi16(rep, one);
    }
  }
}

void aom_paeth_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
                                    const uint8_t *above, const uint8_t *left) {
  const __m256i t0 = get_top_vector(above);
  const __m256i t1 = get_top_vector(above + 16);
  const __m256i t2 = get_top_vector(above + 32);
  const __m256i t3 = get_top_vector(above + 48);
  const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]);
  const __m256i one = _mm256_set1_epi16(1);

  int i, j;
  for (j = 0; j < 4; ++j) {
    const __m256i l = get_left_vector(left + j * 16);
    __m256i rep = _mm256_set1_epi16(0x8000);
    for (i = 0; i < 16; ++i) {
      const __m256i l16 = _mm256_shuffle_epi8(l, rep);

      const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
      const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);
      const __m128i r2 = paeth_16x1_pred(&l16, &t2, &tl);
      const __m128i r3 = paeth_16x1_pred(&l16, &t3, &tl);

      _mm_store_si128((__m128i *)dst, r0);
      _mm_store_si128((__m128i *)(dst + 16), r1);
      _mm_store_si128((__m128i *)(dst + 32), r2);
      _mm_store_si128((__m128i *)(dst + 48), r3);

      dst += stride;
      rep = _mm256_add_epi16(rep, one);
    }
  }
}

void aom_paeth_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
                                    const uint8_t *above, const uint8_t *left) {
  const __m256i t0 = get_top_vector(above);
  const __m256i t1 = get_top_vector(above + 16);
  const __m256i t2 = get_top_vector(above + 32);
  const __m256i t3 = get_top_vector(above + 48);
  const __m256i tl = _mm256_set1_epi16((uint16_t)above[-1]);
  const __m256i one = _mm256_set1_epi16(1);

  int i;
  const __m256i l = get_left_vector(left);
  __m256i rep = _mm256_set1_epi16(0x8000);
  for (i = 0; i < 16; ++i) {
    const __m256i l16 = _mm256_shuffle_epi8(l, rep);

    const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
    const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);
    const __m128i r2 = paeth_16x1_pred(&l16, &t2, &tl);
    const __m128i r3 = paeth_16x1_pred(&l16, &t3, &tl);

    _mm_store_si128((__m128i *)dst, r0);
    _mm_store_si128((__m128i *)(dst + 16), r1);
    _mm_store_si128((__m128i *)(dst + 32), r2);
    _mm_store_si128((__m128i *)(dst + 48), r3);

    dst += stride;
    rep = _mm256_add_epi16(rep, one);
  }
}