Blame src/dsp/enc_sse41.c

Packit 9c6abc
// Copyright 2015 Google Inc. All Rights Reserved.
Packit 9c6abc
//
Packit 9c6abc
// Use of this source code is governed by a BSD-style license
Packit 9c6abc
// that can be found in the COPYING file in the root of the source
Packit 9c6abc
// tree. An additional intellectual property rights grant can be found
Packit 9c6abc
// in the file PATENTS. All contributing project authors may
Packit 9c6abc
// be found in the AUTHORS file in the root of the source tree.
Packit 9c6abc
// -----------------------------------------------------------------------------
Packit 9c6abc
//
Packit 9c6abc
// SSE4 version of some encoding functions.
Packit 9c6abc
//
Packit 9c6abc
// Author: Skal (pascal.massimino@gmail.com)
Packit 9c6abc
Packit 9c6abc
#include "src/dsp/dsp.h"
Packit 9c6abc
Packit 9c6abc
#if defined(WEBP_USE_SSE41)
Packit 9c6abc
#include <smmintrin.h>
Packit 9c6abc
#include <stdlib.h>  // for abs()
Packit 9c6abc
Packit 9c6abc
#include "src/dsp/common_sse2.h"
Packit 9c6abc
#include "src/enc/vp8i_enc.h"
Packit 9c6abc
Packit 9c6abc
//------------------------------------------------------------------------------
Packit 9c6abc
// Compute susceptibility based on DCT-coeff histograms.
Packit 9c6abc
Packit 9c6abc
static void CollectHistogram_SSE41(const uint8_t* ref, const uint8_t* pred,
Packit 9c6abc
                                   int start_block, int end_block,
Packit 9c6abc
                                   VP8Histogram* const histo) {
Packit 9c6abc
  const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
Packit 9c6abc
  int j;
Packit 9c6abc
  int distribution[MAX_COEFF_THRESH + 1] = { 0 };
Packit 9c6abc
  for (j = start_block; j < end_block; ++j) {
Packit 9c6abc
    int16_t out[16];
Packit 9c6abc
    int k;
Packit 9c6abc
Packit 9c6abc
    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
Packit 9c6abc
Packit 9c6abc
    // Convert coefficients to bin (within out[]).
Packit 9c6abc
    {
Packit 9c6abc
      // Load.
Packit 9c6abc
      const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]);
Packit 9c6abc
      const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]);
Packit 9c6abc
      // v = abs(out) >> 3
Packit 9c6abc
      const __m128i abs0 = _mm_abs_epi16(out0);
Packit 9c6abc
      const __m128i abs1 = _mm_abs_epi16(out1);
Packit 9c6abc
      const __m128i v0 = _mm_srai_epi16(abs0, 3);
Packit 9c6abc
      const __m128i v1 = _mm_srai_epi16(abs1, 3);
Packit 9c6abc
      // bin = min(v, MAX_COEFF_THRESH)
Packit 9c6abc
      const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh);
Packit 9c6abc
      const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh);
Packit 9c6abc
      // Store.
Packit 9c6abc
      _mm_storeu_si128((__m128i*)&out[0], bin0);
Packit 9c6abc
      _mm_storeu_si128((__m128i*)&out[8], bin1);
Packit 9c6abc
    }
Packit 9c6abc
Packit 9c6abc
    // Convert coefficients to bin.
Packit 9c6abc
    for (k = 0; k < 16; ++k) {
Packit 9c6abc
      ++distribution[out[k]];
Packit 9c6abc
    }
Packit 9c6abc
  }
Packit 9c6abc
  VP8SetHistogramData(distribution, histo);
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
//------------------------------------------------------------------------------
Packit 9c6abc
// Texture distortion
Packit 9c6abc
//
Packit 9c6abc
// We try to match the spectral content (weighted) between source and
Packit 9c6abc
// reconstructed samples.
Packit 9c6abc
Packit 9c6abc
// Hadamard transform
Packit 9c6abc
// Returns the weighted sum of the absolute value of transformed coefficients.
Packit 9c6abc
// w[] contains a row-major 4 by 4 symmetric matrix.
Packit 9c6abc
static int TTransform_SSE41(const uint8_t* inA, const uint8_t* inB,
Packit 9c6abc
                            const uint16_t* const w) {
Packit 9c6abc
  int32_t sum[4];
Packit 9c6abc
  __m128i tmp_0, tmp_1, tmp_2, tmp_3;
Packit 9c6abc
Packit 9c6abc
  // Load and combine inputs.
Packit 9c6abc
  {
Packit 9c6abc
    const __m128i inA_0 = _mm_loadu_si128((const __m128i*)&inA[BPS * 0]);
Packit 9c6abc
    const __m128i inA_1 = _mm_loadu_si128((const __m128i*)&inA[BPS * 1]);
Packit 9c6abc
    const __m128i inA_2 = _mm_loadu_si128((const __m128i*)&inA[BPS * 2]);
Packit 9c6abc
    // In SSE4.1, with gcc 4.8 at least (maybe other versions),
Packit 9c6abc
    // _mm_loadu_si128 is faster than _mm_loadl_epi64. But for the last lump
Packit 9c6abc
    // of inA and inB, _mm_loadl_epi64 is still used not to have an out of
Packit 9c6abc
    // bound read.
Packit 9c6abc
    const __m128i inA_3 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 3]);
Packit 9c6abc
    const __m128i inB_0 = _mm_loadu_si128((const __m128i*)&inB[BPS * 0]);
Packit 9c6abc
    const __m128i inB_1 = _mm_loadu_si128((const __m128i*)&inB[BPS * 1]);
Packit 9c6abc
    const __m128i inB_2 = _mm_loadu_si128((const __m128i*)&inB[BPS * 2]);
Packit 9c6abc
    const __m128i inB_3 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 3]);
Packit 9c6abc
Packit 9c6abc
    // Combine inA and inB (we'll do two transforms in parallel).
Packit 9c6abc
    const __m128i inAB_0 = _mm_unpacklo_epi32(inA_0, inB_0);
Packit 9c6abc
    const __m128i inAB_1 = _mm_unpacklo_epi32(inA_1, inB_1);
Packit 9c6abc
    const __m128i inAB_2 = _mm_unpacklo_epi32(inA_2, inB_2);
Packit 9c6abc
    const __m128i inAB_3 = _mm_unpacklo_epi32(inA_3, inB_3);
Packit 9c6abc
    tmp_0 = _mm_cvtepu8_epi16(inAB_0);
Packit 9c6abc
    tmp_1 = _mm_cvtepu8_epi16(inAB_1);
Packit 9c6abc
    tmp_2 = _mm_cvtepu8_epi16(inAB_2);
Packit 9c6abc
    tmp_3 = _mm_cvtepu8_epi16(inAB_3);
Packit 9c6abc
    // a00 a01 a02 a03   b00 b01 b02 b03
Packit 9c6abc
    // a10 a11 a12 a13   b10 b11 b12 b13
Packit 9c6abc
    // a20 a21 a22 a23   b20 b21 b22 b23
Packit 9c6abc
    // a30 a31 a32 a33   b30 b31 b32 b33
Packit 9c6abc
  }
Packit 9c6abc
Packit 9c6abc
  // Vertical pass first to avoid a transpose (vertical and horizontal passes
Packit 9c6abc
  // are commutative because w/kWeightY is symmetric) and subsequent transpose.
Packit 9c6abc
  {
Packit 9c6abc
    // Calculate a and b (two 4x4 at once).
Packit 9c6abc
    const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
Packit 9c6abc
    const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3);
Packit 9c6abc
    const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3);
Packit 9c6abc
    const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2);
Packit 9c6abc
    const __m128i b0 = _mm_add_epi16(a0, a1);
Packit 9c6abc
    const __m128i b1 = _mm_add_epi16(a3, a2);
Packit 9c6abc
    const __m128i b2 = _mm_sub_epi16(a3, a2);
Packit 9c6abc
    const __m128i b3 = _mm_sub_epi16(a0, a1);
Packit 9c6abc
    // a00 a01 a02 a03   b00 b01 b02 b03
Packit 9c6abc
    // a10 a11 a12 a13   b10 b11 b12 b13
Packit 9c6abc
    // a20 a21 a22 a23   b20 b21 b22 b23
Packit 9c6abc
    // a30 a31 a32 a33   b30 b31 b32 b33
Packit 9c6abc
Packit 9c6abc
    // Transpose the two 4x4.
Packit 9c6abc
    VP8Transpose_2_4x4_16b(&b0, &b1, &b2, &b3, &tmp_0, &tmp_1, &tmp_2, &tmp_3);
Packit 9c6abc
  }
Packit 9c6abc
Packit 9c6abc
  // Horizontal pass and difference of weighted sums.
Packit 9c6abc
  {
Packit 9c6abc
    // Load all inputs.
Packit 9c6abc
    const __m128i w_0 = _mm_loadu_si128((const __m128i*)&w[0]);
Packit 9c6abc
    const __m128i w_8 = _mm_loadu_si128((const __m128i*)&w[8]);
Packit 9c6abc
Packit 9c6abc
    // Calculate a and b (two 4x4 at once).
Packit 9c6abc
    const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
Packit 9c6abc
    const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3);
Packit 9c6abc
    const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3);
Packit 9c6abc
    const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2);
Packit 9c6abc
    const __m128i b0 = _mm_add_epi16(a0, a1);
Packit 9c6abc
    const __m128i b1 = _mm_add_epi16(a3, a2);
Packit 9c6abc
    const __m128i b2 = _mm_sub_epi16(a3, a2);
Packit 9c6abc
    const __m128i b3 = _mm_sub_epi16(a0, a1);
Packit 9c6abc
Packit 9c6abc
    // Separate the transforms of inA and inB.
Packit 9c6abc
    __m128i A_b0 = _mm_unpacklo_epi64(b0, b1);
Packit 9c6abc
    __m128i A_b2 = _mm_unpacklo_epi64(b2, b3);
Packit 9c6abc
    __m128i B_b0 = _mm_unpackhi_epi64(b0, b1);
Packit 9c6abc
    __m128i B_b2 = _mm_unpackhi_epi64(b2, b3);
Packit 9c6abc
Packit 9c6abc
    A_b0 = _mm_abs_epi16(A_b0);
Packit 9c6abc
    A_b2 = _mm_abs_epi16(A_b2);
Packit 9c6abc
    B_b0 = _mm_abs_epi16(B_b0);
Packit 9c6abc
    B_b2 = _mm_abs_epi16(B_b2);
Packit 9c6abc
Packit 9c6abc
    // weighted sums
Packit 9c6abc
    A_b0 = _mm_madd_epi16(A_b0, w_0);
Packit 9c6abc
    A_b2 = _mm_madd_epi16(A_b2, w_8);
Packit 9c6abc
    B_b0 = _mm_madd_epi16(B_b0, w_0);
Packit 9c6abc
    B_b2 = _mm_madd_epi16(B_b2, w_8);
Packit 9c6abc
    A_b0 = _mm_add_epi32(A_b0, A_b2);
Packit 9c6abc
    B_b0 = _mm_add_epi32(B_b0, B_b2);
Packit 9c6abc
Packit 9c6abc
    // difference of weighted sums
Packit 9c6abc
    A_b2 = _mm_sub_epi32(A_b0, B_b0);
Packit 9c6abc
    _mm_storeu_si128((__m128i*)&sum[0], A_b2);
Packit 9c6abc
  }
Packit 9c6abc
  return sum[0] + sum[1] + sum[2] + sum[3];
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
static int Disto4x4_SSE41(const uint8_t* const a, const uint8_t* const b,
Packit 9c6abc
                          const uint16_t* const w) {
Packit 9c6abc
  const int diff_sum = TTransform_SSE41(a, b, w);
Packit 9c6abc
  return abs(diff_sum) >> 5;
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
static int Disto16x16_SSE41(const uint8_t* const a, const uint8_t* const b,
Packit 9c6abc
                            const uint16_t* const w) {
Packit 9c6abc
  int D = 0;
Packit 9c6abc
  int x, y;
Packit 9c6abc
  for (y = 0; y < 16 * BPS; y += 4 * BPS) {
Packit 9c6abc
    for (x = 0; x < 16; x += 4) {
Packit 9c6abc
      D += Disto4x4_SSE41(a + x + y, b + x + y, w);
Packit 9c6abc
    }
Packit 9c6abc
  }
Packit 9c6abc
  return D;
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
//------------------------------------------------------------------------------
Packit 9c6abc
// Quantization
Packit 9c6abc
//
Packit 9c6abc
Packit 9c6abc
// Generates a pshufb constant for shuffling 16b words.
Packit 9c6abc
#define PSHUFB_CST(A,B,C,D,E,F,G,H) \
Packit 9c6abc
  _mm_set_epi8(2 * (H) + 1, 2 * (H) + 0, 2 * (G) + 1, 2 * (G) + 0, \
Packit 9c6abc
               2 * (F) + 1, 2 * (F) + 0, 2 * (E) + 1, 2 * (E) + 0, \
Packit 9c6abc
               2 * (D) + 1, 2 * (D) + 0, 2 * (C) + 1, 2 * (C) + 0, \
Packit 9c6abc
               2 * (B) + 1, 2 * (B) + 0, 2 * (A) + 1, 2 * (A) + 0)
Packit 9c6abc
Packit 9c6abc
static WEBP_INLINE int DoQuantizeBlock_SSE41(int16_t in[16], int16_t out[16],
Packit 9c6abc
                                             const uint16_t* const sharpen,
Packit 9c6abc
                                             const VP8Matrix* const mtx) {
Packit 9c6abc
  const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL);
Packit 9c6abc
  const __m128i zero = _mm_setzero_si128();
Packit 9c6abc
  __m128i out0, out8;
Packit 9c6abc
  __m128i packed_out;
Packit 9c6abc
Packit 9c6abc
  // Load all inputs.
Packit 9c6abc
  __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
Packit 9c6abc
  __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
Packit 9c6abc
  const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]);
Packit 9c6abc
  const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]);
Packit 9c6abc
  const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]);
Packit 9c6abc
  const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]);
Packit 9c6abc
Packit 9c6abc
  // coeff = abs(in)
Packit 9c6abc
  __m128i coeff0 = _mm_abs_epi16(in0);
Packit 9c6abc
  __m128i coeff8 = _mm_abs_epi16(in8);
Packit 9c6abc
Packit 9c6abc
  // coeff = abs(in) + sharpen
Packit 9c6abc
  if (sharpen != NULL) {
Packit 9c6abc
    const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]);
Packit 9c6abc
    const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]);
Packit 9c6abc
    coeff0 = _mm_add_epi16(coeff0, sharpen0);
Packit 9c6abc
    coeff8 = _mm_add_epi16(coeff8, sharpen8);
Packit 9c6abc
  }
Packit 9c6abc
Packit 9c6abc
  // out = (coeff * iQ + B) >> QFIX
Packit 9c6abc
  {
Packit 9c6abc
    // doing calculations with 32b precision (QFIX=17)
Packit 9c6abc
    // out = (coeff * iQ)
Packit 9c6abc
    const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
Packit 9c6abc
    const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
Packit 9c6abc
    const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
Packit 9c6abc
    const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
Packit 9c6abc
    __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
Packit 9c6abc
    __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
Packit 9c6abc
    __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
Packit 9c6abc
    __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
Packit 9c6abc
    // out = (coeff * iQ + B)
Packit 9c6abc
    const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]);
Packit 9c6abc
    const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]);
Packit 9c6abc
    const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]);
Packit 9c6abc
    const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]);
Packit 9c6abc
    out_00 = _mm_add_epi32(out_00, bias_00);
Packit 9c6abc
    out_04 = _mm_add_epi32(out_04, bias_04);
Packit 9c6abc
    out_08 = _mm_add_epi32(out_08, bias_08);
Packit 9c6abc
    out_12 = _mm_add_epi32(out_12, bias_12);
Packit 9c6abc
    // out = QUANTDIV(coeff, iQ, B, QFIX)
Packit 9c6abc
    out_00 = _mm_srai_epi32(out_00, QFIX);
Packit 9c6abc
    out_04 = _mm_srai_epi32(out_04, QFIX);
Packit 9c6abc
    out_08 = _mm_srai_epi32(out_08, QFIX);
Packit 9c6abc
    out_12 = _mm_srai_epi32(out_12, QFIX);
Packit 9c6abc
Packit 9c6abc
    // pack result as 16b
Packit 9c6abc
    out0 = _mm_packs_epi32(out_00, out_04);
Packit 9c6abc
    out8 = _mm_packs_epi32(out_08, out_12);
Packit 9c6abc
Packit 9c6abc
    // if (coeff > 2047) coeff = 2047
Packit 9c6abc
    out0 = _mm_min_epi16(out0, max_coeff_2047);
Packit 9c6abc
    out8 = _mm_min_epi16(out8, max_coeff_2047);
Packit 9c6abc
  }
Packit 9c6abc
Packit 9c6abc
  // put sign back
Packit 9c6abc
  out0 = _mm_sign_epi16(out0, in0);
Packit 9c6abc
  out8 = _mm_sign_epi16(out8, in8);
Packit 9c6abc
Packit 9c6abc
  // in = out * Q
Packit 9c6abc
  in0 = _mm_mullo_epi16(out0, q0);
Packit 9c6abc
  in8 = _mm_mullo_epi16(out8, q8);
Packit 9c6abc
Packit 9c6abc
  _mm_storeu_si128((__m128i*)&in[0], in0);
Packit 9c6abc
  _mm_storeu_si128((__m128i*)&in[8], in8);
Packit 9c6abc
Packit 9c6abc
  // zigzag the output before storing it. The re-ordering is:
Packit 9c6abc
  //    0 1 2 3 4 5 6 7 | 8  9 10 11 12 13 14 15
Packit 9c6abc
  // -> 0 1 4[8]5 2 3 6 | 9 12 13 10 [7]11 14 15
Packit 9c6abc
  // There's only two misplaced entries ([8] and [7]) that are crossing the
Packit 9c6abc
  // reg's boundaries.
Packit 9c6abc
  // We use pshufb instead of pshuflo/pshufhi.
Packit 9c6abc
  {
Packit 9c6abc
    const __m128i kCst_lo = PSHUFB_CST(0, 1, 4, -1, 5, 2, 3, 6);
Packit 9c6abc
    const __m128i kCst_7 = PSHUFB_CST(-1, -1, -1, -1, 7, -1, -1, -1);
Packit 9c6abc
    const __m128i tmp_lo = _mm_shuffle_epi8(out0, kCst_lo);
Packit 9c6abc
    const __m128i tmp_7 = _mm_shuffle_epi8(out0, kCst_7);  // extract #7
Packit 9c6abc
    const __m128i kCst_hi = PSHUFB_CST(1, 4, 5, 2, -1, 3, 6, 7);
Packit 9c6abc
    const __m128i kCst_8 = PSHUFB_CST(-1, -1, -1, 0, -1, -1, -1, -1);
Packit 9c6abc
    const __m128i tmp_hi = _mm_shuffle_epi8(out8, kCst_hi);
Packit 9c6abc
    const __m128i tmp_8 = _mm_shuffle_epi8(out8, kCst_8);  // extract #8
Packit 9c6abc
    const __m128i out_z0 = _mm_or_si128(tmp_lo, tmp_8);
Packit 9c6abc
    const __m128i out_z8 = _mm_or_si128(tmp_hi, tmp_7);
Packit 9c6abc
    _mm_storeu_si128((__m128i*)&out[0], out_z0);
Packit 9c6abc
    _mm_storeu_si128((__m128i*)&out[8], out_z8);
Packit 9c6abc
    packed_out = _mm_packs_epi16(out_z0, out_z8);
Packit 9c6abc
  }
Packit 9c6abc
Packit 9c6abc
  // detect if all 'out' values are zeroes or not
Packit 9c6abc
  return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff);
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
#undef PSHUFB_CST
Packit 9c6abc
Packit 9c6abc
static int QuantizeBlock_SSE41(int16_t in[16], int16_t out[16],
Packit 9c6abc
                               const VP8Matrix* const mtx) {
Packit 9c6abc
  return DoQuantizeBlock_SSE41(in, out, &mtx->sharpen_[0], mtx);
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
static int QuantizeBlockWHT_SSE41(int16_t in[16], int16_t out[16],
Packit 9c6abc
                                  const VP8Matrix* const mtx) {
Packit 9c6abc
  return DoQuantizeBlock_SSE41(in, out, NULL, mtx);
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
static int Quantize2Blocks_SSE41(int16_t in[32], int16_t out[32],
Packit 9c6abc
                                 const VP8Matrix* const mtx) {
Packit 9c6abc
  int nz;
Packit 9c6abc
  const uint16_t* const sharpen = &mtx->sharpen_[0];
Packit 9c6abc
  nz  = DoQuantizeBlock_SSE41(in + 0 * 16, out + 0 * 16, sharpen, mtx) << 0;
Packit 9c6abc
  nz |= DoQuantizeBlock_SSE41(in + 1 * 16, out + 1 * 16, sharpen, mtx) << 1;
Packit 9c6abc
  return nz;
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
//------------------------------------------------------------------------------
Packit 9c6abc
// Entry point
Packit 9c6abc
Packit 9c6abc
extern void VP8EncDspInitSSE41(void);
Packit 9c6abc
WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitSSE41(void) {
Packit 9c6abc
  VP8CollectHistogram = CollectHistogram_SSE41;
Packit 9c6abc
  VP8EncQuantizeBlock = QuantizeBlock_SSE41;
Packit 9c6abc
  VP8EncQuantize2Blocks = Quantize2Blocks_SSE41;
Packit 9c6abc
  VP8EncQuantizeBlockWHT = QuantizeBlockWHT_SSE41;
Packit 9c6abc
  VP8TDisto4x4 = Disto4x4_SSE41;
Packit 9c6abc
  VP8TDisto16x16 = Disto16x16_SSE41;
Packit 9c6abc
}
Packit 9c6abc
Packit 9c6abc
#else  // !WEBP_USE_SSE41
Packit 9c6abc
Packit 9c6abc
WEBP_DSP_INIT_STUB(VP8EncDspInitSSE41)
Packit 9c6abc
Packit 9c6abc
#endif  // WEBP_USE_SSE41