vp10/encoder/x86/vp9_quantize_sse2.c

   1 /*
   2  *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
   3  *
   4  *  Use of this source code is governed by a BSD-style license
   5  *  that can be found in the LICENSE file in the root of the source
   6  *  tree. An additional intellectual property rights grant can be found
   7  *  in the file PATENTS.  All contributing project authors may
   8  *  be found in the AUTHORS file in the root of the source tree.
   9  */
  10
  11 #include <emmintrin.h>
  12 #include <xmmintrin.h>
  13
  14 #include "./vp10_rtcd.h"
  15 #include "vpx/vpx_integer.h"
  16
  17 void vp10_quantize_fp_sse2(const int16_t* coeff_ptr, intptr_t n_coeffs,
  18                           int skip_block, const int16_t* zbin_ptr,
  19                           const int16_t* round_ptr, const int16_t* quant_ptr,
  20                           const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr,
  21                           int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
  22                           uint16_t* eob_ptr,
  23                           const int16_t* scan_ptr,
  24                           const int16_t* iscan_ptr) {
  25   __m128i zero;
  26   __m128i thr;
  27   int16_t nzflag;
  28   (void)scan_ptr;
  29   (void)zbin_ptr;
  30   (void)quant_shift_ptr;
  31
  32   coeff_ptr += n_coeffs;
  33   iscan_ptr += n_coeffs;
  34   qcoeff_ptr += n_coeffs;
  35   dqcoeff_ptr += n_coeffs;
  36   n_coeffs = -n_coeffs;
  37   zero = _mm_setzero_si128();
  38
  39   if (!skip_block) {
  40     __m128i eob;
  41     __m128i round, quant, dequant;
  42     {
  43       __m128i coeff0, coeff1;
  44
  45       // Setup global values
  46       {
  47         round = _mm_load_si128((const __m128i*)round_ptr);
  48         quant = _mm_load_si128((const __m128i*)quant_ptr);
  49         dequant = _mm_load_si128((const __m128i*)dequant_ptr);
  50       }
  51
  52       {
  53         __m128i coeff0_sign, coeff1_sign;
  54         __m128i qcoeff0, qcoeff1;
  55         __m128i qtmp0, qtmp1;
  56         // Do DC and first 15 AC
  57         coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
  58         coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
  59
  60         // Poor man's sign extract
  61         coeff0_sign = _mm_srai_epi16(coeff0, 15);
  62         coeff1_sign = _mm_srai_epi16(coeff1, 15);
  63         qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
  64         qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
  65         qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
  66         qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
  67
  68         qcoeff0 = _mm_adds_epi16(qcoeff0, round);
  69         round = _mm_unpackhi_epi64(round, round);
  70         qcoeff1 = _mm_adds_epi16(qcoeff1, round);
  71         qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
  72         quant = _mm_unpackhi_epi64(quant, quant);
  73         qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
  74
  75         // Reinsert signs
  76         qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
  77         qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
  78         qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
  79         qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
  80
  81         _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
  82         _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
  83
  84         coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
  85         dequant = _mm_unpackhi_epi64(dequant, dequant);
  86         coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
  87
  88         _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
  89         _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
  90       }
  91
  92       {
  93         // Scan for eob
  94         __m128i zero_coeff0, zero_coeff1;
  95         __m128i nzero_coeff0, nzero_coeff1;
  96         __m128i iscan0, iscan1;
  97         __m128i eob1;
  98         zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
  99         zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
 100         nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
 101         nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
 102         iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
 103         iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
 104         // Add one to convert from indices to counts
 105         iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
 106         iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
 107         eob = _mm_and_si128(iscan0, nzero_coeff0);
 108         eob1 = _mm_and_si128(iscan1, nzero_coeff1);
 109         eob = _mm_max_epi16(eob, eob1);
 110       }
 111       n_coeffs += 8 * 2;
 112     }
 113
 114     thr = _mm_srai_epi16(dequant, 1);
 115
 116     // AC only loop
 117     while (n_coeffs < 0) {
 118       __m128i coeff0, coeff1;
 119       {
 120         __m128i coeff0_sign, coeff1_sign;
 121         __m128i qcoeff0, qcoeff1;
 122         __m128i qtmp0, qtmp1;
 123
 124         coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
 125         coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
 126
 127         // Poor man's sign extract
 128         coeff0_sign = _mm_srai_epi16(coeff0, 15);
 129         coeff1_sign = _mm_srai_epi16(coeff1, 15);
 130         qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
 131         qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
 132         qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
 133         qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
 134
 135         nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
 136             _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
 137
 138         if (nzflag) {
 139           qcoeff0 = _mm_adds_epi16(qcoeff0, round);
 140           qcoeff1 = _mm_adds_epi16(qcoeff1, round);
 141           qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
 142           qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
 143
 144           // Reinsert signs
 145           qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
 146           qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
 147           qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
 148           qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
 149
 150           _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
 151           _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
 152
 153           coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
 154           coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
 155
 156           _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
 157           _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
 158         } else {
 159           _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
 160           _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
 161
 162           _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
 163           _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
 164         }
 165       }
 166
 167       if (nzflag) {
 168         // Scan for eob
 169         __m128i zero_coeff0, zero_coeff1;
 170         __m128i nzero_coeff0, nzero_coeff1;
 171         __m128i iscan0, iscan1;
 172         __m128i eob0, eob1;
 173         zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
 174         zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
 175         nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
 176         nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
 177         iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
 178         iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
 179         // Add one to convert from indices to counts
 180         iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
 181         iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
 182         eob0 = _mm_and_si128(iscan0, nzero_coeff0);
 183         eob1 = _mm_and_si128(iscan1, nzero_coeff1);
 184         eob0 = _mm_max_epi16(eob0, eob1);
 185         eob = _mm_max_epi16(eob, eob0);
 186       }
 187       n_coeffs += 8 * 2;
 188     }
 189
 190     // Accumulate EOB
 191     {
 192       __m128i eob_shuffled;
 193       eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
 194       eob = _mm_max_epi16(eob, eob_shuffled);
 195       eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
 196       eob = _mm_max_epi16(eob, eob_shuffled);
 197       eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
 198       eob = _mm_max_epi16(eob, eob_shuffled);
 199       *eob_ptr = _mm_extract_epi16(eob, 1);
 200     }
 201   } else {
 202     do {
 203       _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
 204       _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
 205       _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
 206       _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
 207       n_coeffs += 8 * 2;
 208     } while (n_coeffs < 0);
 209     *eob_ptr = 0;
 210   }
 211 }