1 /* 2 * Copyright (c) 2010 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 * This code was originally written by: Gregory Maxwell, at the Daala 11 * project. 12 */ 13 #include <assert.h> 14 #include <stdio.h> 15 #include <stdlib.h> 16 #include <math.h> 17 18 #include "./vpx_config.h" 19 #include "./vpx_dsp_rtcd.h" 20 #include "vpx_dsp/ssim.h" 21 #include "vpx_ports/system_state.h" 22 #include "vpx_dsp/psnr.h" 23 24 #if !defined(M_PI) 25 #define M_PI (3.141592653589793238462643) 26 #endif 27 #include <string.h> 28 29 static void od_bin_fdct8x8(tran_low_t *y, int ystride, const int16_t *x, 30 int xstride) { 31 int i, j; 32 (void)xstride; 33 vpx_fdct8x8(x, y, ystride); 34 for (i = 0; i < 8; i++) 35 for (j = 0; j < 8; j++) 36 *(y + ystride * i + j) = (*(y + ystride * i + j) + 4) >> 3; 37 } 38 #if CONFIG_VP9_HIGHBITDEPTH 39 static void hbd_od_bin_fdct8x8(tran_low_t *y, int ystride, const int16_t *x, 40 int xstride) { 41 int i, j; 42 (void)xstride; 43 vpx_highbd_fdct8x8(x, y, ystride); 44 for (i = 0; i < 8; i++) 45 for (j = 0; j < 8; j++) 46 *(y + ystride * i + j) = (*(y + ystride * i + j) + 4) >> 3; 47 } 48 #endif 49 50 /* Normalized inverse quantization matrix for 8x8 DCT at the point of 51 * transparency. This is not the JPEG based matrix from the paper, 52 this one gives a slightly higher MOS agreement.*/ 53 static const double csf_y[8][8] = { 54 { 1.6193873005, 2.2901594831, 2.08509755623, 1.48366094411, 1.00227514334, 55 0.678296995242, 0.466224900598, 0.3265091542 }, 56 { 2.2901594831, 1.94321815382, 2.04793073064, 1.68731108984, 1.2305666963, 57 0.868920337363, 0.61280991668, 0.436405793551 }, 58 { 2.08509755623, 2.04793073064, 1.34329019223, 1.09205635862, 0.875748795257, 59 0.670882927016, 0.501731932449, 0.372504254596 }, 60 { 1.48366094411, 1.68731108984, 1.09205635862, 0.772819797575, 0.605636379554, 61 0.48309405692, 0.380429446972, 0.295774038565 }, 62 { 1.00227514334, 1.2305666963, 0.875748795257, 0.605636379554, 0.448996256676, 63 0.352889268808, 0.283006984131, 0.226951348204 }, 64 { 0.678296995242, 0.868920337363, 0.670882927016, 0.48309405692, 65 0.352889268808, 0.27032073436, 0.215017739696, 0.17408067321 }, 66 { 0.466224900598, 0.61280991668, 0.501731932449, 0.380429446972, 67 0.283006984131, 0.215017739696, 0.168869545842, 0.136153931001 }, 68 { 0.3265091542, 0.436405793551, 0.372504254596, 0.295774038565, 69 0.226951348204, 0.17408067321, 0.136153931001, 0.109083846276 } 70 }; 71 static const double csf_cb420[8][8] = { 72 { 1.91113096927, 2.46074210438, 1.18284184739, 1.14982565193, 1.05017074788, 73 0.898018824055, 0.74725392039, 0.615105596242 }, 74 { 2.46074210438, 1.58529308355, 1.21363250036, 1.38190029285, 1.33100189972, 75 1.17428548929, 0.996404342439, 0.830890433625 }, 76 { 1.18284184739, 1.21363250036, 0.978712413627, 1.02624506078, 1.03145147362, 77 0.960060382087, 0.849823426169, 0.731221236837 }, 78 { 1.14982565193, 1.38190029285, 1.02624506078, 0.861317501629, 0.801821139099, 79 0.751437590932, 0.685398513368, 0.608694761374 }, 80 { 1.05017074788, 1.33100189972, 1.03145147362, 0.801821139099, 0.676555426187, 81 0.605503172737, 0.55002013668, 0.495804539034 }, 82 { 0.898018824055, 1.17428548929, 0.960060382087, 0.751437590932, 83 0.605503172737, 0.514674450957, 0.454353482512, 0.407050308965 }, 84 { 0.74725392039, 0.996404342439, 0.849823426169, 0.685398513368, 85 0.55002013668, 0.454353482512, 0.389234902883, 0.342353999733 }, 86 { 0.615105596242, 0.830890433625, 0.731221236837, 0.608694761374, 87 0.495804539034, 0.407050308965, 0.342353999733, 0.295530605237 } 88 }; 89 static const double csf_cr420[8][8] = { 90 { 2.03871978502, 2.62502345193, 1.26180942886, 1.11019789803, 1.01397751469, 91 0.867069376285, 0.721500455585, 0.593906509971 }, 92 { 2.62502345193, 1.69112867013, 1.17180569821, 1.3342742857, 1.28513006198, 93 1.13381474809, 0.962064122248, 0.802254508198 }, 94 { 1.26180942886, 1.17180569821, 0.944981930573, 0.990876405848, 95 0.995903384143, 0.926972725286, 0.820534991409, 0.706020324706 }, 96 { 1.11019789803, 1.3342742857, 0.990876405848, 0.831632933426, 0.77418706195, 97 0.725539939514, 0.661776842059, 0.587716619023 }, 98 { 1.01397751469, 1.28513006198, 0.995903384143, 0.77418706195, 0.653238524286, 99 0.584635025748, 0.531064164893, 0.478717061273 }, 100 { 0.867069376285, 1.13381474809, 0.926972725286, 0.725539939514, 101 0.584635025748, 0.496936637883, 0.438694579826, 0.393021669543 }, 102 { 0.721500455585, 0.962064122248, 0.820534991409, 0.661776842059, 103 0.531064164893, 0.438694579826, 0.375820256136, 0.330555063063 }, 104 { 0.593906509971, 0.802254508198, 0.706020324706, 0.587716619023, 105 0.478717061273, 0.393021669543, 0.330555063063, 0.285345396658 } 106 }; 107 108 static double convert_score_db(double _score, double _weight, int bit_depth) { 109 int16_t pix_max = 255; 110 assert(_score * _weight >= 0.0); 111 if (bit_depth == 10) 112 pix_max = 1023; 113 else if (bit_depth == 12) 114 pix_max = 4095; 115 116 if (_weight * _score < pix_max * pix_max * 1e-10) return MAX_PSNR; 117 return 10 * (log10(pix_max * pix_max) - log10(_weight * _score)); 118 } 119 120 static double calc_psnrhvs(const unsigned char *src, int _systride, 121 const unsigned char *dst, int _dystride, double _par, 122 int _w, int _h, int _step, const double _csf[8][8], 123 uint32_t bit_depth, uint32_t _shift) { 124 double ret; 125 const uint8_t *_src8 = src; 126 const uint8_t *_dst8 = dst; 127 const uint16_t *_src16 = CONVERT_TO_SHORTPTR(src); 128 const uint16_t *_dst16 = CONVERT_TO_SHORTPTR(dst); 129 int16_t dct_s[8 * 8], dct_d[8 * 8]; 130 tran_low_t dct_s_coef[8 * 8], dct_d_coef[8 * 8]; 131 double mask[8][8]; 132 int pixels; 133 int x; 134 int y; 135 (void)_par; 136 ret = pixels = 0; 137 138 /*In the PSNR-HVS-M paper[1] the authors describe the construction of 139 their masking table as "we have used the quantization table for the 140 color component Y of JPEG [6] that has been also obtained on the 141 basis of CSF. Note that the values in quantization table JPEG have 142 been normalized and then squared." Their CSF matrix (from PSNR-HVS) 143 was also constructed from the JPEG matrices. I can not find any obvious 144 scheme of normalizing to produce their table, but if I multiply their 145 CSF by 0.38857 and square the result I get their masking table. 146 I have no idea where this constant comes from, but deviating from it 147 too greatly hurts MOS agreement. 148 149 [1] Nikolay Ponomarenko, Flavia Silvestri, Karen Egiazarian, Marco Carli, 150 Jaakko Astola, Vladimir Lukin, "On between-coefficient contrast masking 151 of DCT basis functions", CD-ROM Proceedings of the Third 152 International Workshop on Video Processing and Quality Metrics for Consumer 153 Electronics VPQM-07, Scottsdale, Arizona, USA, 25-26 January, 2007, 4 p.*/ 154 for (x = 0; x < 8; x++) 155 for (y = 0; y < 8; y++) 156 mask[x][y] = 157 (_csf[x][y] * 0.3885746225901003) * (_csf[x][y] * 0.3885746225901003); 158 for (y = 0; y < _h - 7; y += _step) { 159 for (x = 0; x < _w - 7; x += _step) { 160 int i; 161 int j; 162 double s_means[4]; 163 double d_means[4]; 164 double s_vars[4]; 165 double d_vars[4]; 166 double s_gmean = 0; 167 double d_gmean = 0; 168 double s_gvar = 0; 169 double d_gvar = 0; 170 double s_mask = 0; 171 double d_mask = 0; 172 for (i = 0; i < 4; i++) 173 s_means[i] = d_means[i] = s_vars[i] = d_vars[i] = 0; 174 for (i = 0; i < 8; i++) { 175 for (j = 0; j < 8; j++) { 176 int sub = ((i & 12) >> 2) + ((j & 12) >> 1); 177 if (bit_depth == 8 && _shift == 0) { 178 dct_s[i * 8 + j] = _src8[(y + i) * _systride + (j + x)]; 179 dct_d[i * 8 + j] = _dst8[(y + i) * _dystride + (j + x)]; 180 } else if (bit_depth == 10 || bit_depth == 12) { 181 dct_s[i * 8 + j] = _src16[(y + i) * _systride + (j + x)] >> _shift; 182 dct_d[i * 8 + j] = _dst16[(y + i) * _dystride + (j + x)] >> _shift; 183 } 184 s_gmean += dct_s[i * 8 + j]; 185 d_gmean += dct_d[i * 8 + j]; 186 s_means[sub] += dct_s[i * 8 + j]; 187 d_means[sub] += dct_d[i * 8 + j]; 188 } 189 } 190 s_gmean /= 64.f; 191 d_gmean /= 64.f; 192 for (i = 0; i < 4; i++) s_means[i] /= 16.f; 193 for (i = 0; i < 4; i++) d_means[i] /= 16.f; 194 for (i = 0; i < 8; i++) { 195 for (j = 0; j < 8; j++) { 196 int sub = ((i & 12) >> 2) + ((j & 12) >> 1); 197 s_gvar += (dct_s[i * 8 + j] - s_gmean) * (dct_s[i * 8 + j] - s_gmean); 198 d_gvar += (dct_d[i * 8 + j] - d_gmean) * (dct_d[i * 8 + j] - d_gmean); 199 s_vars[sub] += (dct_s[i * 8 + j] - s_means[sub]) * 200 (dct_s[i * 8 + j] - s_means[sub]); 201 d_vars[sub] += (dct_d[i * 8 + j] - d_means[sub]) * 202 (dct_d[i * 8 + j] - d_means[sub]); 203 } 204 } 205 s_gvar *= 1 / 63.f * 64; 206 d_gvar *= 1 / 63.f * 64; 207 for (i = 0; i < 4; i++) s_vars[i] *= 1 / 15.f * 16; 208 for (i = 0; i < 4; i++) d_vars[i] *= 1 / 15.f * 16; 209 if (s_gvar > 0) 210 s_gvar = (s_vars[0] + s_vars[1] + s_vars[2] + s_vars[3]) / s_gvar; 211 if (d_gvar > 0) 212 d_gvar = (d_vars[0] + d_vars[1] + d_vars[2] + d_vars[3]) / d_gvar; 213 #if CONFIG_VP9_HIGHBITDEPTH 214 if (bit_depth == 10 || bit_depth == 12) { 215 hbd_od_bin_fdct8x8(dct_s_coef, 8, dct_s, 8); 216 hbd_od_bin_fdct8x8(dct_d_coef, 8, dct_d, 8); 217 } 218 #endif 219 if (bit_depth == 8) { 220 od_bin_fdct8x8(dct_s_coef, 8, dct_s, 8); 221 od_bin_fdct8x8(dct_d_coef, 8, dct_d, 8); 222 } 223 for (i = 0; i < 8; i++) 224 for (j = (i == 0); j < 8; j++) 225 s_mask += dct_s_coef[i * 8 + j] * dct_s_coef[i * 8 + j] * mask[i][j]; 226 for (i = 0; i < 8; i++) 227 for (j = (i == 0); j < 8; j++) 228 d_mask += dct_d_coef[i * 8 + j] * dct_d_coef[i * 8 + j] * mask[i][j]; 229 s_mask = sqrt(s_mask * s_gvar) / 32.f; 230 d_mask = sqrt(d_mask * d_gvar) / 32.f; 231 if (d_mask > s_mask) s_mask = d_mask; 232 for (i = 0; i < 8; i++) { 233 for (j = 0; j < 8; j++) { 234 double err; 235 err = fabs((double)(dct_s_coef[i * 8 + j] - dct_d_coef[i * 8 + j])); 236 if (i != 0 || j != 0) 237 err = err < s_mask / mask[i][j] ? 0 : err - s_mask / mask[i][j]; 238 ret += (err * _csf[i][j]) * (err * _csf[i][j]); 239 pixels++; 240 } 241 } 242 } 243 } 244 if (pixels <= 0) return 0; 245 ret /= pixels; 246 return ret; 247 } 248 249 double vpx_psnrhvs(const YV12_BUFFER_CONFIG *src, 250 const YV12_BUFFER_CONFIG *dest, double *y_psnrhvs, 251 double *u_psnrhvs, double *v_psnrhvs, uint32_t bd, 252 uint32_t in_bd) { 253 double psnrhvs; 254 const double par = 1.0; 255 const int step = 7; 256 uint32_t bd_shift = 0; 257 vpx_clear_system_state(); 258 259 assert(bd == 8 || bd == 10 || bd == 12); 260 assert(bd >= in_bd); 261 262 bd_shift = bd - in_bd; 263 264 *y_psnrhvs = calc_psnrhvs(src->y_buffer, src->y_stride, dest->y_buffer, 265 dest->y_stride, par, src->y_crop_width, 266 src->y_crop_height, step, csf_y, bd, bd_shift); 267 *u_psnrhvs = calc_psnrhvs(src->u_buffer, src->uv_stride, dest->u_buffer, 268 dest->uv_stride, par, src->uv_crop_width, 269 src->uv_crop_height, step, csf_cb420, bd, bd_shift); 270 *v_psnrhvs = calc_psnrhvs(src->v_buffer, src->uv_stride, dest->v_buffer, 271 dest->uv_stride, par, src->uv_crop_width, 272 src->uv_crop_height, step, csf_cr420, bd, bd_shift); 273 psnrhvs = (*y_psnrhvs) * .8 + .1 * ((*u_psnrhvs) + (*v_psnrhvs)); 274 return convert_score_db(psnrhvs, 1.0, in_bd); 275 } 276
