Quelle  denoising_sse2.c   Sprache: C

/*
 *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include "vp8/encoder/denoising.h"
#include "vp8/common/reconinter.h"
#include "vpx/vpx_integer.h"
#include "vpx_mem/vpx_mem.h"
#include "vp8_rtcd.h"

#include <emmintrin.h>
#include "vpx_ports/emmintrin_compat.h"

/* Compute the sum of all pixel differences of this MB. */
static INLINE unsigned int abs_sum_diff_16x1(__m128i acc_diff) {
  const __m128i k_1 = _mm_set1_epi16(1);
  const __m128i acc_diff_lo =
      _mm_srai_epi16(_mm_unpacklo_epi8(acc_diff, acc_diff), 8);
  const __m128i acc_diff_hi =
      _mm_srai_epi16(_mm_unpackhi_epi8(acc_diff, acc_diff), 8);
  const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi);
  const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1);
  const __m128i hgfe_dcba =
      _mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8));
  const __m128i hgfedcba =
      _mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4));
  unsigned int sum_diff = (unsigned int)abs(_mm_cvtsi128_si32(hgfedcba));

  return sum_diff;
}

int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
                             int mc_avg_y_stride, unsigned char *running_avg_y,
                             int avg_y_stride, unsigned char *sig,
                             int sig_stride, unsigned int motion_magnitude,
                             int increase_denoising) {
  unsigned char *running_avg_y_start = running_avg_y;
  unsigned char *sig_start = sig;
  unsigned int sum_diff_thresh;
  int r;
  int shift_inc =
      (increase_denoising && motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
          ? 1
          : 0;
  __m128i acc_diff = _mm_setzero_si128();
  const __m128i k_0 = _mm_setzero_si128();
  const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
  const __m128i k_8 = _mm_set1_epi8(8);
  const __m128i k_16 = _mm_set1_epi8(16);
  /* Modify each level's adjustment according to motion_magnitude. */
  const __m128i l3 = _mm_set1_epi8(
      (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6);
  /* Difference between level 3 and level 2 is 2. */
  const __m128i l32 = _mm_set1_epi8(2);
  /* Difference between level 2 and level 1 is 1. */
  const __m128i l21 = _mm_set1_epi8(1);

  for (r = 0; r < 16; ++r) {
    /* Calculate differences */
    const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
    const __m128i v_mc_running_avg_y =
        _mm_loadu_si128((__m128i *)(&mc_running_avg_y[0]));
    __m128i v_running_avg_y;
    const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
    const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
    /* Obtain the sign. FF if diff is negative. */
    const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
    /* Clamp absolute difference to 16 to be used to get mask. Doing this
     * allows us to use _mm_cmpgt_epi8, which operates on signed byte. */
    const __m128i clamped_absdiff =
        _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_16);
    /* Get masks for l2 l1 and l0 adjustments */
    const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
    const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
    const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
    /* Get adjustments for l2, l1, and l0 */
    __m128i adj2 = _mm_and_si128(mask2, l32);
    const __m128i adj1 = _mm_and_si128(mask1, l21);
    const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
    __m128i adj, padj, nadj;

    /* Combine the adjustments and get absolute adjustments. */
    adj2 = _mm_add_epi8(adj2, adj1);
    adj = _mm_sub_epi8(l3, adj2);
    adj = _mm_andnot_si128(mask0, adj);
    adj = _mm_or_si128(adj, adj0);

    /* Restore the sign and get positive and negative adjustments. */
    padj = _mm_andnot_si128(diff_sign, adj);
    nadj = _mm_and_si128(diff_sign, adj);

    /* Calculate filtered value. */
    v_running_avg_y = _mm_adds_epu8(v_sig, padj);
    v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
    _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);

    /* Adjustments <=7, and each element in acc_diff can fit in signed
     * char.
     */
    acc_diff = _mm_adds_epi8(acc_diff, padj);
    acc_diff = _mm_subs_epi8(acc_diff, nadj);

    /* Update pointers for next iteration. */
    sig += sig_stride;
    mc_running_avg_y += mc_avg_y_stride;
    running_avg_y += avg_y_stride;
  }

  {
    /* Compute the sum of all pixel differences of this MB. */
    unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
    sum_diff_thresh = SUM_DIFF_THRESHOLD;
    if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
    if (abs_sum_diff > sum_diff_thresh) {
      // Before returning to copy the block (i.e., apply no denoising),
      // check if we can still apply some (weaker) temporal filtering to
      // this block, that would otherwise not be denoised at all. Simplest
      // is to apply an additional adjustment to running_avg_y to bring it
      // closer to sig. The adjustment is capped by a maximum delta, and
      // chosen such that in most cases the resulting sum_diff will be
      // within the acceptable range given by sum_diff_thresh.

      // The delta is set by the excess of absolute pixel diff over the
      // threshold.
      int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
      // Only apply the adjustment for max delta up to 3.
      if (delta < 4) {
        const __m128i k_delta = _mm_set1_epi8(delta);
        sig -= sig_stride * 16;
        mc_running_avg_y -= mc_avg_y_stride * 16;
        running_avg_y -= avg_y_stride * 16;
        for (r = 0; r < 16; ++r) {
          __m128i v_running_avg_y =
              _mm_loadu_si128((__m128i *)(&running_avg_y[0]));
          // Calculate differences.
          const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
          const __m128i v_mc_running_avg_y =
              _mm_loadu_si128((__m128i *)(&mc_running_avg_y[0]));
          const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
          const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
          // Obtain the sign. FF if diff is negative.
          const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
          // Clamp absolute difference to delta to get the adjustment.
          const __m128i adj = _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
          // Restore the sign and get positive and negative adjustments.
          __m128i padj, nadj;
          padj = _mm_andnot_si128(diff_sign, adj);
          nadj = _mm_and_si128(diff_sign, adj);
          // Calculate filtered value.
          v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj);
          v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj);
          _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);

          // Accumulate the adjustments.
          acc_diff = _mm_subs_epi8(acc_diff, padj);
          acc_diff = _mm_adds_epi8(acc_diff, nadj);

          // Update pointers for next iteration.
          sig += sig_stride;
          mc_running_avg_y += mc_avg_y_stride;
          running_avg_y += avg_y_stride;
        }
        abs_sum_diff = abs_sum_diff_16x1(acc_diff);
        if (abs_sum_diff > sum_diff_thresh) {
          return COPY_BLOCK;
        }
      } else {
        return COPY_BLOCK;
      }
    }
  }

  vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride);
  return FILTER_BLOCK;
}

int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg,
                                int mc_avg_stride, unsigned char *running_avg,
                                int avg_stride, unsigned char *sig,
                                int sig_stride, unsigned int motion_magnitude,
                                int increase_denoising) {
  unsigned char *running_avg_start = running_avg;
  unsigned char *sig_start = sig;
  unsigned int sum_diff_thresh;
  int r;
  int shift_inc =
      (increase_denoising && motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV)
          ? 1
          : 0;
  __m128i acc_diff = _mm_setzero_si128();
  const __m128i k_0 = _mm_setzero_si128();
  const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
  const __m128i k_8 = _mm_set1_epi8(8);
  const __m128i k_16 = _mm_set1_epi8(16);
  /* Modify each level's adjustment according to motion_magnitude. */
  const __m128i l3 = _mm_set1_epi8(
      (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ? 7 + shift_inc : 6);
  /* Difference between level 3 and level 2 is 2. */
  const __m128i l32 = _mm_set1_epi8(2);
  /* Difference between level 2 and level 1 is 1. */
  const __m128i l21 = _mm_set1_epi8(1);

  {
    const __m128i k_1 = _mm_set1_epi16(1);
    __m128i vec_sum_block = _mm_setzero_si128();

    // Avoid denoising color signal if its close to average level.
    for (r = 0; r < 8; ++r) {
      const __m128i v_sig = _mm_loadl_epi64((__m128i *)(&sig[0]));
      const __m128i v_sig_unpack = _mm_unpacklo_epi8(v_sig, k_0);
      vec_sum_block = _mm_add_epi16(vec_sum_block, v_sig_unpack);
      sig += sig_stride;
    }
    sig -= sig_stride * 8;
    {
      const __m128i hg_fe_dc_ba = _mm_madd_epi16(vec_sum_block, k_1);
      const __m128i hgfe_dcba =
          _mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8));
      const __m128i hgfedcba =
          _mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4));
      const int sum_block = _mm_cvtsi128_si32(hgfedcba);
      if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
        return COPY_BLOCK;
      }
    }
  }

  for (r = 0; r < 4; ++r) {
    /* Calculate differences */
    const __m128i v_sig_low =
        _mm_castpd_si128(_mm_load_sd((double *)(&sig[0])));
    const __m128i v_sig = _mm_castpd_si128(_mm_loadh_pd(
        _mm_castsi128_pd(v_sig_low), (double *)(&sig[sig_stride])));
    const __m128i v_mc_running_avg_low =
        _mm_castpd_si128(_mm_load_sd((double *)(&mc_running_avg[0])));
    const __m128i v_mc_running_avg = _mm_castpd_si128(
        _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
                     (double *)(&mc_running_avg[mc_avg_stride])));
    const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
    const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
    /* Obtain the sign. FF if diff is negative. */
    const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
    /* Clamp absolute difference to 16 to be used to get mask. Doing this
     * allows us to use _mm_cmpgt_epi8, which operates on signed byte. */
    const __m128i clamped_absdiff =
        _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_16);
    /* Get masks for l2 l1 and l0 adjustments */
    const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
    const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
    const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
    /* Get adjustments for l2, l1, and l0 */
    __m128i adj2 = _mm_and_si128(mask2, l32);
    const __m128i adj1 = _mm_and_si128(mask1, l21);
    const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
    __m128i adj, padj, nadj;
    __m128i v_running_avg;

    /* Combine the adjustments and get absolute adjustments. */
    adj2 = _mm_add_epi8(adj2, adj1);
    adj = _mm_sub_epi8(l3, adj2);
    adj = _mm_andnot_si128(mask0, adj);
    adj = _mm_or_si128(adj, adj0);

    /* Restore the sign and get positive and negative adjustments. */
    padj = _mm_andnot_si128(diff_sign, adj);
    nadj = _mm_and_si128(diff_sign, adj);

    /* Calculate filtered value. */
    v_running_avg = _mm_adds_epu8(v_sig, padj);
    v_running_avg = _mm_subs_epu8(v_running_avg, nadj);

    _mm_storel_pd((double *)&running_avg[0], _mm_castsi128_pd(v_running_avg));
    _mm_storeh_pd((double *)&running_avg[avg_stride],
                  _mm_castsi128_pd(v_running_avg));

    /* Adjustments <=7, and each element in acc_diff can fit in signed
     * char.
     */
    acc_diff = _mm_adds_epi8(acc_diff, padj);
    acc_diff = _mm_subs_epi8(acc_diff, nadj);

    /* Update pointers for next iteration. */
    sig += sig_stride * 2;
    mc_running_avg += mc_avg_stride * 2;
    running_avg += avg_stride * 2;
  }

  {
    unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
    sum_diff_thresh = SUM_DIFF_THRESHOLD_UV;
    if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
    if (abs_sum_diff > sum_diff_thresh) {
      // Before returning to copy the block (i.e., apply no denoising),
      // check if we can still apply some (weaker) temporal filtering to
      // this block, that would otherwise not be denoised at all. Simplest
      // is to apply an additional adjustment to running_avg_y to bring it
      // closer to sig. The adjustment is capped by a maximum delta, and
      // chosen such that in most cases the resulting sum_diff will be
      // within the acceptable range given by sum_diff_thresh.

      // The delta is set by the excess of absolute pixel diff over the
      // threshold.
      int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
      // Only apply the adjustment for max delta up to 3.
      if (delta < 4) {
        const __m128i k_delta = _mm_set1_epi8(delta);
        sig -= sig_stride * 8;
        mc_running_avg -= mc_avg_stride * 8;
        running_avg -= avg_stride * 8;
        for (r = 0; r < 4; ++r) {
          // Calculate differences.
          const __m128i v_sig_low =
              _mm_castpd_si128(_mm_load_sd((double *)(&sig[0])));
          const __m128i v_sig = _mm_castpd_si128(_mm_loadh_pd(
              _mm_castsi128_pd(v_sig_low), (double *)(&sig[sig_stride])));
          const __m128i v_mc_running_avg_low =
              _mm_castpd_si128(_mm_load_sd((double *)(&mc_running_avg[0])));
          const __m128i v_mc_running_avg = _mm_castpd_si128(
              _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
                           (double *)(&mc_running_avg[mc_avg_stride])));
          const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
          const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
          // Obtain the sign. FF if diff is negative.
          const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
          // Clamp absolute difference to delta to get the adjustment.
          const __m128i adj = _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
          // Restore the sign and get positive and negative adjustments.
          __m128i padj, nadj;
          const __m128i v_running_avg_low =
              _mm_castpd_si128(_mm_load_sd((double *)(&running_avg[0])));
          __m128i v_running_avg = _mm_castpd_si128(
              _mm_loadh_pd(_mm_castsi128_pd(v_running_avg_low),
                           (double *)(&running_avg[avg_stride])));
          padj = _mm_andnot_si128(diff_sign, adj);
          nadj = _mm_and_si128(diff_sign, adj);
          // Calculate filtered value.
          v_running_avg = _mm_subs_epu8(v_running_avg, padj);
          v_running_avg = _mm_adds_epu8(v_running_avg, nadj);

          _mm_storel_pd((double *)&running_avg[0],
                        _mm_castsi128_pd(v_running_avg));
          _mm_storeh_pd((double *)&running_avg[avg_stride],
                        _mm_castsi128_pd(v_running_avg));

          // Accumulate the adjustments.
          acc_diff = _mm_subs_epi8(acc_diff, padj);
          acc_diff = _mm_adds_epi8(acc_diff, nadj);

          // Update pointers for next iteration.
          sig += sig_stride * 2;
          mc_running_avg += mc_avg_stride * 2;
          running_avg += avg_stride * 2;
        }
        abs_sum_diff = abs_sum_diff_16x1(acc_diff);
        if (abs_sum_diff > sum_diff_thresh) {
          return COPY_BLOCK;
        }
      } else {
        return COPY_BLOCK;
      }
    }
  }

  vp8_copy_mem8x8(running_avg_start, avg_stride, sig_start, sig_stride);
  return FILTER_BLOCK;
}