quantize_sse2.c 6.29 KB
Newer Older
1 2 3 4 5 6 7 8 9 10
/*
 *  Copyright (c) 2015 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.
 */

11
#include <assert.h>
12 13 14 15 16
#include <emmintrin.h>
#include <xmmintrin.h>

#include "./vpx_dsp_rtcd.h"
#include "vpx/vpx_integer.h"
17
#include "vpx_dsp/x86/bitdepth_conversion_sse2.h"
18

19 20 21 22 23 24 25
void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
                         int skip_block, const int16_t *zbin_ptr,
                         const int16_t *round_ptr, const int16_t *quant_ptr,
                         const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
                         tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
                         uint16_t *eob_ptr, const int16_t *scan_ptr,
                         const int16_t *iscan_ptr) {
26 27 28 29 30 31 32 33 34 35 36
  const __m128i zero = _mm_setzero_si128();
  int index = 16;

  __m128i zbin, round, quant, dequant, shift;
  __m128i coeff0, coeff1, coeff0_sign, coeff1_sign;
  __m128i qcoeff0, qcoeff1;
  __m128i cmp_mask0, cmp_mask1;
  __m128i qtmp0, qtmp1;
  __m128i zero_coeff0, zero_coeff1, iscan0, iscan1;
  __m128i eob, eob0, eob1;

37
  (void)scan_ptr;
38 39
  (void)skip_block;
  assert(!skip_block);
40

41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167
  // Setup global values.
  zbin = _mm_load_si128((const __m128i *)zbin_ptr);
  round = _mm_load_si128((const __m128i *)round_ptr);
  quant = _mm_load_si128((const __m128i *)quant_ptr);
  zbin = _mm_sub_epi16(zbin, _mm_set1_epi16(1));
  dequant = _mm_load_si128((const __m128i *)dequant_ptr);
  shift = _mm_load_si128((const __m128i *)quant_shift_ptr);

  // Do DC and first 15 AC.
  coeff0 = load_tran_low(coeff_ptr);
  coeff1 = load_tran_low(coeff_ptr + 8);

  // Poor man's abs().
  coeff0_sign = _mm_srai_epi16(coeff0, 15);
  coeff1_sign = _mm_srai_epi16(coeff1, 15);
  qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
  qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
  qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
  qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);

  cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
  zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
  cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);

  qcoeff0 = _mm_adds_epi16(qcoeff0, round);
  round = _mm_unpackhi_epi64(round, round);
  qcoeff1 = _mm_adds_epi16(qcoeff1, round);

  qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
  quant = _mm_unpackhi_epi64(quant, quant);
  qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);

  qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
  qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);

  qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
  shift = _mm_unpackhi_epi64(shift, shift);
  qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);

  // Reinsert signs
  qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
  qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
  qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
  qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);

  // Mask out zbin threshold coeffs
  qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
  qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);

  store_tran_low(qcoeff0, qcoeff_ptr);
  store_tran_low(qcoeff1, qcoeff_ptr + 8);

  coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
  dequant = _mm_unpackhi_epi64(dequant, dequant);
  coeff1 = _mm_mullo_epi16(qcoeff1, dequant);

  store_tran_low(coeff0, dqcoeff_ptr);
  store_tran_low(coeff1, dqcoeff_ptr + 8);

  // Scan for eob.
  zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
  zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
  iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr));
  iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8));
  // Add one to convert from indices to counts
  iscan0 = _mm_sub_epi16(iscan0, cmp_mask0);
  iscan1 = _mm_sub_epi16(iscan1, cmp_mask1);
  eob = _mm_andnot_si128(zero_coeff0, iscan0);
  eob1 = _mm_andnot_si128(zero_coeff1, iscan1);
  eob = _mm_max_epi16(eob, eob1);

  // AC only loop.
  while (index < n_coeffs) {
    coeff0 = load_tran_low(coeff_ptr + index);
    coeff1 = load_tran_low(coeff_ptr + index + 8);

    coeff0_sign = _mm_srai_epi16(coeff0, 15);
    coeff1_sign = _mm_srai_epi16(coeff1, 15);
    qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
    qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
    qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
    qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);

    cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
    cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);

    qcoeff0 = _mm_adds_epi16(qcoeff0, round);
    qcoeff1 = _mm_adds_epi16(qcoeff1, round);

    qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
    qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);

    qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
    qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);

    qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
    qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);

    qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
    qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
    qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
    qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);

    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);

    store_tran_low(qcoeff0, qcoeff_ptr + index);
    store_tran_low(qcoeff1, qcoeff_ptr + index + 8);

    coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
    coeff1 = _mm_mullo_epi16(qcoeff1, dequant);

    store_tran_low(coeff0, dqcoeff_ptr + index);
    store_tran_low(coeff1, dqcoeff_ptr + index + 8);

    zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
    zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
    iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + index));
    iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + index + 8));
    iscan0 = _mm_sub_epi16(iscan0, cmp_mask0);
    iscan1 = _mm_sub_epi16(iscan1, cmp_mask1);
    eob0 = _mm_andnot_si128(zero_coeff0, iscan0);
    eob1 = _mm_andnot_si128(zero_coeff1, iscan1);
    eob0 = _mm_max_epi16(eob0, eob1);
    eob = _mm_max_epi16(eob, eob0);

    index += 16;
168 169
  }

170
  // Accumulate eob.
171 172 173 174 175 176 177 178 179
  {
    __m128i eob_shuffled;
    eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
    eob = _mm_max_epi16(eob, eob_shuffled);
    eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
    eob = _mm_max_epi16(eob, eob_shuffled);
    eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
    eob = _mm_max_epi16(eob, eob_shuffled);
    *eob_ptr = _mm_extract_epi16(eob, 1);
180 181
  }
}