masked_sad_intrin_avx2.c | Explore in Territory

/*
 * Copyright (c) 2018, 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 "config/aom_config.h"
#include "config/aom_dsp_rtcd.h"

#include "aom_dsp/blend.h"
#include "aom/aom_integer.h"
#include "aom_dsp/x86/synonyms.h"
#include "aom_dsp/x86/synonyms_avx2.h"
#include "aom_dsp/x86/masked_sad_intrin_ssse3.h"

static inline unsigned int masked_sad32xh_avx2(
    const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
    const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
    int width, int height) { … }

static inline unsigned int masked_sad16xh_avx2(
    const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
    const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
    int height) { … }

static inline unsigned int aom_masked_sad_avx2(
    const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride,
    const uint8_t *second_pred, const uint8_t *msk, int msk_stride,
    int invert_mask, int m, int n) { … }

#define MASKSADMXN_AVX2(m, n) …

MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)
MASKSADMXN_AVX2(…)

#if !CONFIG_REALTIME_ONLY
MASKSADMXN_AVX2(4, 16)
MASKSADMXN_AVX2(16, 4)
MASKSADMXN_AVX2(8, 32)
MASKSADMXN_AVX2(32, 8)
MASKSADMXN_AVX2(16, 64)
MASKSADMXN_AVX2(64, 16)
#endif  // !CONFIG_REALTIME_ONLY

#if CONFIG_AV1_HIGHBITDEPTH
static inline unsigned int highbd_masked_sad8xh_avx2(
    const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
    const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
    int height) {
  const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
  const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
  const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
  int y;
  __m256i res = _mm256_setzero_si256();
  const __m256i mask_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
  const __m256i round_const =
      _mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
  const __m256i one = _mm256_set1_epi16(1);

  for (y = 0; y < height; y += 2) {
    const __m256i src = yy_loadu2_128(src_ptr + src_stride, src_ptr);
    const __m256i a = yy_loadu2_128(a_ptr + a_stride, a_ptr);
    const __m256i b = yy_loadu2_128(b_ptr + b_stride, b_ptr);
    // Zero-extend mask to 16 bits
    const __m256i m = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(
        _mm_loadl_epi64((const __m128i *)(m_ptr)),
        _mm_loadl_epi64((const __m128i *)(m_ptr + m_stride))));
    const __m256i m_inv = _mm256_sub_epi16(mask_max, m);

    const __m256i data_l = _mm256_unpacklo_epi16(a, b);
    const __m256i mask_l = _mm256_unpacklo_epi16(m, m_inv);
    __m256i pred_l = _mm256_madd_epi16(data_l, mask_l);
    pred_l = _mm256_srai_epi32(_mm256_add_epi32(pred_l, round_const),
                               AOM_BLEND_A64_ROUND_BITS);

    const __m256i data_r = _mm256_unpackhi_epi16(a, b);
    const __m256i mask_r = _mm256_unpackhi_epi16(m, m_inv);
    __m256i pred_r = _mm256_madd_epi16(data_r, mask_r);
    pred_r = _mm256_srai_epi32(_mm256_add_epi32(pred_r, round_const),
                               AOM_BLEND_A64_ROUND_BITS);

    // Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15,
    // so it is safe to do signed saturation here.
    const __m256i pred = _mm256_packs_epi32(pred_l, pred_r);
    // There is no 16-bit SAD instruction, so we have to synthesize
    // an 8-element SAD. We do this by storing 4 32-bit partial SADs,
    // and accumulating them at the end
    const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(pred, src));
    res = _mm256_add_epi32(res, _mm256_madd_epi16(diff, one));

    src_ptr += src_stride << 1;
    a_ptr += a_stride << 1;
    b_ptr += b_stride << 1;
    m_ptr += m_stride << 1;
  }
  // At this point, we have four 32-bit partial SADs stored in 'res'.
  res = _mm256_hadd_epi32(res, res);
  res = _mm256_hadd_epi32(res, res);
  int sad = _mm256_extract_epi32(res, 0) + _mm256_extract_epi32(res, 4);
  return sad;
}

static inline unsigned int highbd_masked_sad16xh_avx2(
    const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
    const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
    int width, int height) {
  const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
  const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
  const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
  int x, y;
  __m256i res = _mm256_setzero_si256();
  const __m256i mask_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
  const __m256i round_const =
      _mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
  const __m256i one = _mm256_set1_epi16(1);

  for (y = 0; y < height; y++) {
    for (x = 0; x < width; x += 16) {
      const __m256i src = _mm256_lddqu_si256((const __m256i *)&src_ptr[x]);
      const __m256i a = _mm256_lddqu_si256((const __m256i *)&a_ptr[x]);
      const __m256i b = _mm256_lddqu_si256((const __m256i *)&b_ptr[x]);
      // Zero-extend mask to 16 bits
      const __m256i m =
          _mm256_cvtepu8_epi16(_mm_lddqu_si128((const __m128i *)&m_ptr[x]));
      const __m256i m_inv = _mm256_sub_epi16(mask_max, m);

      const __m256i data_l = _mm256_unpacklo_epi16(a, b);
      const __m256i mask_l = _mm256_unpacklo_epi16(m, m_inv);
      __m256i pred_l = _mm256_madd_epi16(data_l, mask_l);
      pred_l = _mm256_srai_epi32(_mm256_add_epi32(pred_l, round_const),
                                 AOM_BLEND_A64_ROUND_BITS);

      const __m256i data_r = _mm256_unpackhi_epi16(a, b);
      const __m256i mask_r = _mm256_unpackhi_epi16(m, m_inv);
      __m256i pred_r = _mm256_madd_epi16(data_r, mask_r);
      pred_r = _mm256_srai_epi32(_mm256_add_epi32(pred_r, round_const),
                                 AOM_BLEND_A64_ROUND_BITS);

      // Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15,
      // so it is safe to do signed saturation here.
      const __m256i pred = _mm256_packs_epi32(pred_l, pred_r);
      // There is no 16-bit SAD instruction, so we have to synthesize
      // an 8-element SAD. We do this by storing 4 32-bit partial SADs,
      // and accumulating them at the end
      const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(pred, src));
      res = _mm256_add_epi32(res, _mm256_madd_epi16(diff, one));
    }

    src_ptr += src_stride;
    a_ptr += a_stride;
    b_ptr += b_stride;
    m_ptr += m_stride;
  }
  // At this point, we have four 32-bit partial SADs stored in 'res'.
  res = _mm256_hadd_epi32(res, res);
  res = _mm256_hadd_epi32(res, res);
  int sad = _mm256_extract_epi32(res, 0) + _mm256_extract_epi32(res, 4);
  return sad;
}

static inline unsigned int aom_highbd_masked_sad_avx2(
    const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride,
    const uint8_t *second_pred, const uint8_t *msk, int msk_stride,
    int invert_mask, int m, int n) {
  unsigned int sad;
  if (!invert_mask) {
    switch (m) {
      case 4:
        sad =
            aom_highbd_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride,
                                           second_pred, m, msk, msk_stride, n);
        break;
      case 8:
        sad = highbd_masked_sad8xh_avx2(src, src_stride, ref, ref_stride,
                                        second_pred, m, msk, msk_stride, n);
        break;
      default:
        sad = highbd_masked_sad16xh_avx2(src, src_stride, ref, ref_stride,
                                         second_pred, m, msk, msk_stride, m, n);
        break;
    }
  } else {
    switch (m) {
      case 4:
        sad =
            aom_highbd_masked_sad4xh_ssse3(src, src_stride, second_pred, m, ref,
                                           ref_stride, msk, msk_stride, n);
        break;
      case 8:
        sad = highbd_masked_sad8xh_avx2(src, src_stride, second_pred, m, ref,
                                        ref_stride, msk, msk_stride, n);
        break;
      default:
        sad = highbd_masked_sad16xh_avx2(src, src_stride, second_pred, m, ref,
                                         ref_stride, msk, msk_stride, m, n);
        break;
    }
  }
  return sad;
}

#define HIGHBD_MASKSADMXN_AVX2 …

HIGHBD_MASKSADMXN_AVX2(4, 4)
HIGHBD_MASKSADMXN_AVX2(4, 8)
HIGHBD_MASKSADMXN_AVX2(8, 4)
HIGHBD_MASKSADMXN_AVX2(8, 8)
HIGHBD_MASKSADMXN_AVX2(8, 16)
HIGHBD_MASKSADMXN_AVX2(16, 8)
HIGHBD_MASKSADMXN_AVX2(16, 16)
HIGHBD_MASKSADMXN_AVX2(16, 32)
HIGHBD_MASKSADMXN_AVX2(32, 16)
HIGHBD_MASKSADMXN_AVX2(32, 32)
HIGHBD_MASKSADMXN_AVX2(32, 64)
HIGHBD_MASKSADMXN_AVX2(64, 32)
HIGHBD_MASKSADMXN_AVX2(64, 64)
HIGHBD_MASKSADMXN_AVX2(64, 128)
HIGHBD_MASKSADMXN_AVX2(128, 64)
HIGHBD_MASKSADMXN_AVX2(128, 128)

#if !CONFIG_REALTIME_ONLY
HIGHBD_MASKSADMXN_AVX2(4, 16)
HIGHBD_MASKSADMXN_AVX2(16, 4)
HIGHBD_MASKSADMXN_AVX2(8, 32)
HIGHBD_MASKSADMXN_AVX2(32, 8)
HIGHBD_MASKSADMXN_AVX2(16, 64)
HIGHBD_MASKSADMXN_AVX2(64, 16)
#endif  // !CONFIG_REALTIME_ONLY
#endif  // CONFIG_AV1_HIGHBITDEPTH
chromium/third_party/libaom/source/libaom/aom_dsp/x86/masked_sad_intrin_avx2.c
