Quelle  highbd_intrapred_sse2.c

/*
 * 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 <emmintrin.h>

#include "config/aom_dsp_rtcd.h"

// -----------------------------------------------------------------------------
// H_PRED

void aom_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,
                                     const uint16_t *above,
                                     const uint16_t *left, int bd) {
  const __m128i left_u16 = _mm_loadl_epi64((const __m128i *)left);
  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
  (void)above;
  (void)bd;
  _mm_storel_epi64((__m128i *)dst, row0);
  dst += stride;
  _mm_storel_epi64((__m128i *)dst, row1);
  dst += stride;
  _mm_storel_epi64((__m128i *)dst, row2);
  dst += stride;
  _mm_storel_epi64((__m128i *)dst, row3);
}

void aom_highbd_h_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                     const uint16_t *above,
                                     const uint16_t *left, int bd) {
  aom_highbd_h_predictor_4x4_sse2(dst, stride, above, left, bd);
  dst += stride << 2;
  left += 4;
  aom_highbd_h_predictor_4x4_sse2(dst, stride, above, left, bd);
}

void aom_highbd_h_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
                                     const uint16_t *above,
                                     const uint16_t *left, int bd) {
  const __m128i left_u16 = _mm_load_si128((const __m128i *)left);
  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
  (void)above;
  (void)bd;
  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row0, row0));
  dst += stride;
  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row1, row1));
  dst += stride;
  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row2, row2));
  dst += stride;
  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row3, row3));
}

void aom_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                     const uint16_t *above,
                                     const uint16_t *left, int bd) {
  const __m128i left_u16 = _mm_load_si128((const __m128i *)left);
  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
  const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0);
  const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55);
  const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa);
  const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff);
  (void)above;
  (void)bd;
  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row0, row0));
  dst += stride;
  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row1, row1));
  dst += stride;
  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row2, row2));
  dst += stride;
  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row3, row3));
  dst += stride;
  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row4, row4));
  dst += stride;
  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row5, row5));
  dst += stride;
  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row6, row6));
  dst += stride;
  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row7, row7));
}

void aom_highbd_h_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                      const uint16_t *above,
                                      const uint16_t *left, int bd) {
  aom_highbd_h_predictor_8x8_sse2(dst, stride, above, left, bd);
  dst += stride << 3;
  left += 8;
  aom_highbd_h_predictor_8x8_sse2(dst, stride, above, left, bd);
}

static inline void h_store_16_unpacklo(uint16_t **dst, const ptrdiff_t stride,
                                       const __m128i *row) {
  const __m128i val = _mm_unpacklo_epi64(*row, *row);
  _mm_store_si128((__m128i *)*dst, val);
  _mm_store_si128((__m128i *)(*dst + 8), val);
  *dst += stride;
}

static inline void h_store_16_unpackhi(uint16_t **dst, const ptrdiff_t stride,
                                       const __m128i *row) {
  const __m128i val = _mm_unpackhi_epi64(*row, *row);
  _mm_store_si128((__m128i *)(*dst), val);
  _mm_store_si128((__m128i *)(*dst + 8), val);
  *dst += stride;
}

static inline void h_predictor_16x8(uint16_t *dst, ptrdiff_t stride,
                                    const uint16_t *left) {
  const __m128i left_u16 = _mm_load_si128((const __m128i *)left);
  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
  const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0);
  const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55);
  const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa);
  const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff);
  h_store_16_unpacklo(&dst, stride, &row0);
  h_store_16_unpacklo(&dst, stride, &row1);
  h_store_16_unpacklo(&dst, stride, &row2);
  h_store_16_unpacklo(&dst, stride, &row3);
  h_store_16_unpackhi(&dst, stride, &row4);
  h_store_16_unpackhi(&dst, stride, &row5);
  h_store_16_unpackhi(&dst, stride, &row6);
  h_store_16_unpackhi(&dst, stride, &row7);
}

void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                      const uint16_t *above,
                                      const uint16_t *left, int bd) {
  (void)above;
  (void)bd;
  h_predictor_16x8(dst, stride, left);
}

void aom_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                       const uint16_t *above,
                                       const uint16_t *left, int bd) {
  int i;
  (void)above;
  (void)bd;

  for (i = 0; i < 2; i++, left += 8) {
    h_predictor_16x8(dst, stride, left);
    dst += stride << 3;
  }
}

void aom_highbd_h_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
                                       const uint16_t *above,
                                       const uint16_t *left, int bd) {
  int i;
  (void)above;
  (void)bd;

  for (i = 0; i < 4; i++, left += 8) {
    h_predictor_16x8(dst, stride, left);
    dst += stride << 3;
  }
}

static inline void h_store_32_unpacklo(uint16_t **dst, const ptrdiff_t stride,
                                       const __m128i *row) {
  const __m128i val = _mm_unpacklo_epi64(*row, *row);
  _mm_store_si128((__m128i *)(*dst), val);
  _mm_store_si128((__m128i *)(*dst + 8), val);
  _mm_store_si128((__m128i *)(*dst + 16), val);
  _mm_store_si128((__m128i *)(*dst + 24), val);
  *dst += stride;
}

static inline void h_store_32_unpackhi(uint16_t **dst, const ptrdiff_t stride,
                                       const __m128i *row) {
  const __m128i val = _mm_unpackhi_epi64(*row, *row);
  _mm_store_si128((__m128i *)(*dst), val);
  _mm_store_si128((__m128i *)(*dst + 8), val);
  _mm_store_si128((__m128i *)(*dst + 16), val);
  _mm_store_si128((__m128i *)(*dst + 24), val);
  *dst += stride;
}

static inline void h_predictor_32x8(uint16_t *dst, ptrdiff_t stride,
                                    const uint16_t *left) {
  const __m128i left_u16 = _mm_load_si128((const __m128i *)left);
  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
  const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0);
  const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55);
  const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa);
  const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff);
  h_store_32_unpacklo(&dst, stride, &row0);
  h_store_32_unpacklo(&dst, stride, &row1);
  h_store_32_unpacklo(&dst, stride, &row2);
  h_store_32_unpacklo(&dst, stride, &row3);
  h_store_32_unpackhi(&dst, stride, &row4);
  h_store_32_unpackhi(&dst, stride, &row5);
  h_store_32_unpackhi(&dst, stride, &row6);
  h_store_32_unpackhi(&dst, stride, &row7);
}

void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                       const uint16_t *above,
                                       const uint16_t *left, int bd) {
  int i;
  (void)above;
  (void)bd;

  for (i = 0; i < 2; i++, left += 8) {
    h_predictor_32x8(dst, stride, left);
    dst += stride << 3;
  }
}

void aom_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,
                                       const uint16_t *above,
                                       const uint16_t *left, int bd) {
  int i;
  (void)above;
  (void)bd;

  for (i = 0; i < 4; i++, left += 8) {
    h_predictor_32x8(dst, stride, left);
    dst += stride << 3;
  }
}

// -----------------------------------------------------------------------------
// DC_TOP, DC_LEFT, DC_128

// 4x4

static inline __m128i dc_sum_4(const uint16_t *ref) {
  const __m128i _dcba = _mm_loadl_epi64((const __m128i *)ref);
  const __m128i _xxdc = _mm_shufflelo_epi16(_dcba, 0xe);
  const __m128i a = _mm_add_epi16(_dcba, _xxdc);
  return _mm_add_epi16(a, _mm_shufflelo_epi16(a, 0x1));
}

static inline void dc_store_4x4(uint16_t *dst, ptrdiff_t stride,
                                const __m128i *dc) {
  const __m128i dc_dup = _mm_shufflelo_epi16(*dc, 0x0);
  int i;
  for (i = 0; i < 4; ++i, dst += stride) {
    _mm_storel_epi64((__m128i *)dst, dc_dup);
  }
}

void aom_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,
                                           const uint16_t *above,
                                           const uint16_t *left, int bd) {
  const __m128i two = _mm_cvtsi32_si128(2);
  const __m128i sum = dc_sum_4(left);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2);
  (void)above;
  (void)bd;
  dc_store_4x4(dst, stride, &dc);
}

void aom_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,
                                          const uint16_t *above,
                                          const uint16_t *left, int bd) {
  const __m128i two = _mm_cvtsi32_si128(2);
  const __m128i sum = dc_sum_4(above);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2);
  (void)left;
  (void)bd;
  dc_store_4x4(dst, stride, &dc);
}

void aom_highbd_dc_128_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,
                                          const uint16_t *above,
                                          const uint16_t *left, int bd) {
  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
  (void)above;
  (void)left;
  dc_store_4x4(dst, stride, &dc_dup);
}

// -----------------------------------------------------------------------------
// 4x8

static inline void dc_store_4x8(uint16_t *dst, ptrdiff_t stride,
                                const __m128i *dc) {
  const __m128i dc_dup = _mm_shufflelo_epi16(*dc, 0x0);
  int i;
  for (i = 0; i < 8; ++i, dst += stride) {
    _mm_storel_epi64((__m128i *)dst, dc_dup);
  }
}

// Shared with DC 8xh
static inline __m128i dc_sum_8(const uint16_t *ref) {
  const __m128i ref_u16 = _mm_load_si128((const __m128i *)ref);
  const __m128i _dcba = _mm_add_epi16(ref_u16, _mm_srli_si128(ref_u16, 8));
  const __m128i _xxdc = _mm_shufflelo_epi16(_dcba, 0xe);
  const __m128i a = _mm_add_epi16(_dcba, _xxdc);

  return _mm_add_epi16(a, _mm_shufflelo_epi16(a, 0x1));
}

void aom_highbd_dc_left_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                           const uint16_t *above,
                                           const uint16_t *left, int bd) {
  const __m128i sum = dc_sum_8(left);
  const __m128i four = _mm_cvtsi32_si128(4);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3);
  (void)above;
  (void)bd;
  dc_store_4x8(dst, stride, &dc);
}

void aom_highbd_dc_top_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                          const uint16_t *above,
                                          const uint16_t *left, int bd) {
  const __m128i two = _mm_cvtsi32_si128(2);
  const __m128i sum = dc_sum_4(above);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2);
  (void)left;
  (void)bd;
  dc_store_4x8(dst, stride, &dc);
}

void aom_highbd_dc_128_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                          const uint16_t *above,
                                          const uint16_t *left, int bd) {
  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
  (void)above;
  (void)left;
  dc_store_4x8(dst, stride, &dc_dup);
}

// -----------------------------------------------------------------------------
// 8xh

static inline void dc_store_8xh(uint16_t *dst, ptrdiff_t stride, int height,
                                const __m128i *dc) {
  const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0);
  const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo);
  int i;
  for (i = 0; i < height; ++i, dst += stride) {
    _mm_store_si128((__m128i *)dst, dc_dup);
  }
}

// -----------------------------------------------------------------------------
// DC_TOP

static inline void dc_top_predictor_8xh(uint16_t *dst, ptrdiff_t stride,
                                        int height, const uint16_t *above) {
  const __m128i four = _mm_cvtsi32_si128(4);
  const __m128i sum = dc_sum_8(above);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3);
  dc_store_8xh(dst, stride, height, &dc);
}

void aom_highbd_dc_top_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
                                          const uint16_t *above,
                                          const uint16_t *left, int bd) {
  (void)left;
  (void)bd;
  dc_top_predictor_8xh(dst, stride, 4, above);
}

void aom_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                          const uint16_t *above,
                                          const uint16_t *left, int bd) {
  (void)left;
  (void)bd;
  dc_top_predictor_8xh(dst, stride, 8, above);
}

void aom_highbd_dc_top_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                           const uint16_t *above,
                                           const uint16_t *left, int bd) {
  (void)left;
  (void)bd;
  dc_top_predictor_8xh(dst, stride, 16, above);
}

// -----------------------------------------------------------------------------
// DC_LEFT

void aom_highbd_dc_left_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
                                           const uint16_t *above,
                                           const uint16_t *left, int bd) {
  const __m128i two = _mm_cvtsi32_si128(2);
  const __m128i sum = dc_sum_4(left);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2);
  (void)above;
  (void)bd;
  dc_store_8xh(dst, stride, 4, &dc);
}

void aom_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                           const uint16_t *above,
                                           const uint16_t *left, int bd) {
  const __m128i four = _mm_cvtsi32_si128(4);
  const __m128i sum = dc_sum_8(left);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3);
  (void)above;
  (void)bd;
  dc_store_8xh(dst, stride, 8, &dc);
}

// Shared with DC 16xh
static inline __m128i dc_sum_16(const uint16_t *ref) {
  const __m128i sum_lo = dc_sum_8(ref);
  const __m128i sum_hi = dc_sum_8(ref + 8);
  return _mm_add_epi16(sum_lo, sum_hi);
}

void aom_highbd_dc_left_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                            const uint16_t *above,
                                            const uint16_t *left, int bd) {
  const __m128i eight = _mm_cvtsi32_si128(8);
  const __m128i sum = dc_sum_16(left);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
  (void)above;
  (void)bd;
  dc_store_8xh(dst, stride, 16, &dc);
}

// -----------------------------------------------------------------------------
// DC_128

static inline void dc_128_predictor_8xh(uint16_t *dst, ptrdiff_t stride,
                                        int height, int bd) {
  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
  dc_store_8xh(dst, stride, height, &dc_dup);
}

void aom_highbd_dc_128_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
                                          const uint16_t *above,
                                          const uint16_t *left, int bd) {
  (void)above;
  (void)left;
  dc_128_predictor_8xh(dst, stride, 4, bd);
}

void aom_highbd_dc_128_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                          const uint16_t *above,
                                          const uint16_t *left, int bd) {
  (void)above;
  (void)left;
  dc_128_predictor_8xh(dst, stride, 8, bd);
}

void aom_highbd_dc_128_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                           const uint16_t *above,
                                           const uint16_t *left, int bd) {
  (void)above;
  (void)left;
  dc_128_predictor_8xh(dst, stride, 16, bd);
}

// -----------------------------------------------------------------------------
// 16xh

static inline void dc_store_16xh(uint16_t *dst, ptrdiff_t stride, int height,
                                 const __m128i *dc) {
  const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0);
  const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo);
  int i;
  for (i = 0; i < height; ++i, dst += stride) {
    _mm_store_si128((__m128i *)dst, dc_dup);
    _mm_store_si128((__m128i *)(dst + 8), dc_dup);
  }
}

// -----------------------------------------------------------------------------
// DC_LEFT

void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                            const uint16_t *above,
                                            const uint16_t *left, int bd) {
  const __m128i four = _mm_cvtsi32_si128(4);
  const __m128i sum = dc_sum_8(left);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3);
  (void)above;
  (void)bd;
  dc_store_16xh(dst, stride, 8, &dc);
}

void aom_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                             const uint16_t *above,
                                             const uint16_t *left, int bd) {
  const __m128i eight = _mm_cvtsi32_si128(8);
  const __m128i sum = dc_sum_16(left);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
  (void)above;
  (void)bd;
  dc_store_16xh(dst, stride, 16, &dc);
}

// Shared with 32xh
static inline __m128i dc_sum_32(const uint16_t *ref) {
  const __m128i zero = _mm_setzero_si128();
  const __m128i sum_a = dc_sum_16(ref);
  const __m128i sum_b = dc_sum_16(ref + 16);
  // 12 bit bd will outrange, so expand to 32 bit before adding final total
  return _mm_add_epi32(_mm_unpacklo_epi16(sum_a, zero),
                       _mm_unpacklo_epi16(sum_b, zero));
}

void aom_highbd_dc_left_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
                                             const uint16_t *above,
                                             const uint16_t *left, int bd) {
  const __m128i sixteen = _mm_cvtsi32_si128(16);
  const __m128i sum = dc_sum_32(left);
  const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5);
  (void)above;
  (void)bd;
  dc_store_16xh(dst, stride, 32, &dc);
}

// -----------------------------------------------------------------------------
// DC_TOP

void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                           const uint16_t *above,
                                           const uint16_t *left, int bd) {
  const __m128i eight = _mm_cvtsi32_si128(8);
  const __m128i sum = dc_sum_16(above);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
  (void)left;
  (void)bd;
  dc_store_16xh(dst, stride, 8, &dc);
}

void aom_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                            const uint16_t *above,
                                            const uint16_t *left, int bd) {
  const __m128i eight = _mm_cvtsi32_si128(8);
  const __m128i sum = dc_sum_16(above);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
  (void)left;
  (void)bd;
  dc_store_16xh(dst, stride, 16, &dc);
}

void aom_highbd_dc_top_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
                                            const uint16_t *above,
                                            const uint16_t *left, int bd) {
  const __m128i eight = _mm_cvtsi32_si128(8);
  const __m128i sum = dc_sum_16(above);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
  (void)left;
  (void)bd;
  dc_store_16xh(dst, stride, 32, &dc);
}

// -----------------------------------------------------------------------------
// DC_128

void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                           const uint16_t *above,
                                           const uint16_t *left, int bd) {
  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
  (void)above;
  (void)left;
  dc_store_16xh(dst, stride, 8, &dc_dup);
}

void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                            const uint16_t *above,
                                            const uint16_t *left, int bd) {
  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
  (void)above;
  (void)left;
  dc_store_16xh(dst, stride, 16, &dc_dup);
}

void aom_highbd_dc_128_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
                                            const uint16_t *above,
                                            const uint16_t *left, int bd) {
  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
  (void)above;
  (void)left;
  dc_store_16xh(dst, stride, 32, &dc_dup);
}

// -----------------------------------------------------------------------------
// 32xh

static inline void dc_store_32xh(uint16_t *dst, ptrdiff_t stride, int height,
                                 const __m128i *dc) {
  const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0);
  const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo);
  int i;
  for (i = 0; i < height; ++i, dst += stride) {
    _mm_store_si128((__m128i *)dst, dc_dup);
    _mm_store_si128((__m128i *)(dst + 8), dc_dup);
    _mm_store_si128((__m128i *)(dst + 16), dc_dup);
    _mm_store_si128((__m128i *)(dst + 24), dc_dup);
  }
}

void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                             const uint16_t *above,
                                             const uint16_t *left, int bd) {
  const __m128i eight = _mm_cvtsi32_si128(8);
  const __m128i sum = dc_sum_16(left);
  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
  (void)above;
  (void)bd;
  dc_store_32xh(dst, stride, 16, &dc);
}

void aom_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,
                                             const uint16_t *above,
                                             const uint16_t *left, int bd) {
  const __m128i sixteen = _mm_cvtsi32_si128(16);
  const __m128i sum = dc_sum_32(left);
  const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5);
  (void)above;
  (void)bd;
  dc_store_32xh(dst, stride, 32, &dc);
}

void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                            const uint16_t *above,
                                            const uint16_t *left, int bd) {
  const __m128i sixteen = _mm_cvtsi32_si128(16);
  const __m128i sum = dc_sum_32(above);
  const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5);
  (void)left;
  (void)bd;
  dc_store_32xh(dst, stride, 16, &dc);
}

void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                            const uint16_t *above,
                                            const uint16_t *left, int bd) {
  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
  (void)above;
  (void)left;
  dc_store_32xh(dst, stride, 16, &dc_dup);
}

void aom_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,
                                            const uint16_t *above,
                                            const uint16_t *left, int bd) {
  const __m128i sixteen = _mm_cvtsi32_si128(16);
  const __m128i sum = dc_sum_32(above);
  const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5);
  (void)left;
  (void)bd;
  dc_store_32xh(dst, stride, 32, &dc);
}

void aom_highbd_dc_128_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,
                                            const uint16_t *above,
                                            const uint16_t *left, int bd) {
  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
  (void)above;
  (void)left;
  dc_store_32xh(dst, stride, 32, &dc_dup);
}

// -----------------------------------------------------------------------------
// V_PRED

void aom_highbd_v_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                     const uint16_t *above,
                                     const uint16_t *left, int bd) {
  (void)left;
  (void)bd;
  const __m128i above_u16 = _mm_loadl_epi64((const __m128i *)above);
  int i;
  for (i = 0; i < 2; ++i) {
    _mm_storel_epi64((__m128i *)dst, above_u16);
    _mm_storel_epi64((__m128i *)(dst + stride), above_u16);
    _mm_storel_epi64((__m128i *)(dst + 2 * stride), above_u16);
    _mm_storel_epi64((__m128i *)(dst + 3 * stride), above_u16);
    dst += stride << 2;
  }
}

void aom_highbd_v_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
                                     const uint16_t *above,
                                     const uint16_t *left, int bd) {
  (void)left;
  (void)bd;
  const __m128i above_u16 = _mm_load_si128((const __m128i *)above);
  _mm_store_si128((__m128i *)dst, above_u16);
  _mm_store_si128((__m128i *)(dst + stride), above_u16);
  _mm_store_si128((__m128i *)(dst + 2 * stride), above_u16);
  _mm_store_si128((__m128i *)(dst + 3 * stride), above_u16);
}

void aom_highbd_v_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                      const uint16_t *above,
                                      const uint16_t *left, int bd) {
  (void)left;
  (void)bd;
  const __m128i above_u16 = _mm_load_si128((const __m128i *)above);
  int i;
  for (i = 0; i < 4; ++i) {
    _mm_store_si128((__m128i *)dst, above_u16);
    _mm_store_si128((__m128i *)(dst + stride), above_u16);
    _mm_store_si128((__m128i *)(dst + 2 * stride), above_u16);
    _mm_store_si128((__m128i *)(dst + 3 * stride), above_u16);
    dst += stride << 2;
  }
}

void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                      const uint16_t *above,
                                      const uint16_t *left, int bd) {
  (void)left;
  (void)bd;
  const __m128i above0_u16 = _mm_load_si128((const __m128i *)above);
  const __m128i above1_u16 = _mm_load_si128((const __m128i *)(above + 8));
  int i;
  for (i = 0; i < 2; ++i) {
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    dst += stride;
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    dst += stride;
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    dst += stride;
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    dst += stride;
  }
}

void aom_highbd_v_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
                                       const uint16_t *above,
                                       const uint16_t *left, int bd) {
  (void)left;
  (void)bd;
  const __m128i above0_u16 = _mm_load_si128((const __m128i *)above);
  const __m128i above1_u16 = _mm_load_si128((const __m128i *)(above + 8));
  int i;
  for (i = 0; i < 8; ++i) {
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    dst += stride;
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    dst += stride;
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    dst += stride;
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    dst += stride;
  }
}

void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                       const uint16_t *above,
                                       const uint16_t *left, int bd) {
  (void)left;
  (void)bd;
  const __m128i above0_u16 = _mm_load_si128((const __m128i *)above);
  const __m128i above1_u16 = _mm_load_si128((const __m128i *)(above + 8));
  const __m128i above2_u16 = _mm_load_si128((const __m128i *)(above + 16));
  const __m128i above3_u16 = _mm_load_si128((const __m128i *)(above + 24));
  int i;
  for (i = 0; i < 4; ++i) {
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    _mm_store_si128((__m128i *)(dst + 16), above2_u16);
    _mm_store_si128((__m128i *)(dst + 24), above3_u16);
    dst += stride;
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    _mm_store_si128((__m128i *)(dst + 16), above2_u16);
    _mm_store_si128((__m128i *)(dst + 24), above3_u16);
    dst += stride;
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    _mm_store_si128((__m128i *)(dst + 16), above2_u16);
    _mm_store_si128((__m128i *)(dst + 24), above3_u16);
    dst += stride;
    _mm_store_si128((__m128i *)dst, above0_u16);
    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
    _mm_store_si128((__m128i *)(dst + 16), above2_u16);
    _mm_store_si128((__m128i *)(dst + 24), above3_u16);
    dst += stride;
  }
}

// -----------------------------------------------------------------------------
// DC_PRED

void aom_highbd_dc_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                      const uint16_t *above,
                                      const uint16_t *left, int bd) {
  (void)bd;
  const __m128i sum_above = dc_sum_4(above);
  const __m128i sum_left = dc_sum_8(left);
  const __m128i sum = _mm_add_epi16(sum_above, sum_left);
  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
  sum32 >>= 16;
  sum32 += 6;
  sum32 /= 12;
  const __m128i row = _mm_set1_epi16((int16_t)sum32);
  int i;
  for (i = 0; i < 4; ++i) {
    _mm_storel_epi64((__m128i *)dst, row);
    dst += stride;
    _mm_storel_epi64((__m128i *)dst, row);
    dst += stride;
  }
}

void aom_highbd_dc_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
                                      const uint16_t *above,
                                      const uint16_t *left, int bd) {
  (void)bd;
  const __m128i sum_left = dc_sum_4(left);
  const __m128i sum_above = dc_sum_8(above);
  const __m128i sum = _mm_add_epi16(sum_above, sum_left);
  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
  sum32 >>= 16;
  sum32 += 6;
  sum32 /= 12;
  const __m128i row = _mm_set1_epi16((int16_t)sum32);

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

void aom_highbd_dc_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                       const uint16_t *above,
                                       const uint16_t *left, int bd) {
  (void)bd;
  __m128i sum_left = dc_sum_16(left);
  __m128i sum_above = dc_sum_8(above);
  const __m128i zero = _mm_setzero_si128();
  sum_left = _mm_unpacklo_epi16(sum_left, zero);
  sum_above = _mm_unpacklo_epi16(sum_above, zero);
  const __m128i sum = _mm_add_epi32(sum_left, sum_above);
  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
  sum32 += 12;
  sum32 /= 24;
  const __m128i row = _mm_set1_epi16((int16_t)sum32);
  int i;
  for (i = 0; i < 4; ++i) {
    _mm_store_si128((__m128i *)dst, row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    dst += stride;
  }
}

void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
                                       const uint16_t *above,
                                       const uint16_t *left, int bd) {
  (void)bd;
  __m128i sum_left = dc_sum_8(left);
  __m128i sum_above = dc_sum_16(above);
  const __m128i zero = _mm_setzero_si128();
  sum_left = _mm_unpacklo_epi16(sum_left, zero);
  sum_above = _mm_unpacklo_epi16(sum_above, zero);
  const __m128i sum = _mm_add_epi32(sum_left, sum_above);
  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
  sum32 += 12;
  sum32 /= 24;
  const __m128i row = _mm_set1_epi16((int16_t)sum32);
  int i;
  for (i = 0; i < 2; ++i) {
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    dst += stride;
  }
}

void aom_highbd_dc_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
                                        const uint16_t *above,
                                        const uint16_t *left, int bd) {
  (void)bd;
  __m128i sum_left = dc_sum_32(left);
  __m128i sum_above = dc_sum_16(above);
  const __m128i zero = _mm_setzero_si128();
  sum_above = _mm_unpacklo_epi16(sum_above, zero);
  const __m128i sum = _mm_add_epi32(sum_left, sum_above);
  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
  sum32 += 24;
  sum32 /= 48;
  const __m128i row = _mm_set1_epi16((int16_t)sum32);
  int i;
  for (i = 0; i < 8; ++i) {
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    dst += stride;
  }
}

void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
                                        const uint16_t *above,
                                        const uint16_t *left, int bd) {
  (void)bd;
  __m128i sum_left = dc_sum_16(left);
  __m128i sum_above = dc_sum_32(above);
  const __m128i zero = _mm_setzero_si128();
  sum_left = _mm_unpacklo_epi16(sum_left, zero);
  const __m128i sum = _mm_add_epi32(sum_left, sum_above);
  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
  sum32 += 24;
  sum32 /= 48;
  const __m128i row = _mm_set1_epi16((int16_t)sum32);
  int i;
  for (i = 0; i < 4; ++i) {
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    _mm_store_si128((__m128i *)(dst + 16), row);
    _mm_store_si128((__m128i *)(dst + 24), row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    _mm_store_si128((__m128i *)(dst + 16), row);
    _mm_store_si128((__m128i *)(dst + 24), row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    _mm_store_si128((__m128i *)(dst + 16), row);
    _mm_store_si128((__m128i *)(dst + 24), row);
    dst += stride;
    _mm_store_si128((__m128i *)dst, row);
    _mm_store_si128((__m128i *)(dst + 8), row);
    _mm_store_si128((__m128i *)(dst + 16), row);
    _mm_store_si128((__m128i *)(dst + 24), row);
    dst += stride;
  }
}