1 /* 2 * Copyright 2013 The LibYuv 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 "../util/ssim.h" // NOLINT 12 13 #include <string.h> 14 15 #ifdef __cplusplus 16 extern "C" { 17 #endif 18 19 typedef unsigned int uint32; // NOLINT 20 typedef unsigned short uint16; // NOLINT 21 22 #if !defined(LIBYUV_DISABLE_X86) && !defined(__SSE2__) && \ 23 (defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP >= 2))) 24 #define __SSE2__ 25 #endif 26 #if !defined(LIBYUV_DISABLE_X86) && defined(__SSE2__) 27 #include <emmintrin.h> 28 #endif 29 30 #ifdef _OPENMP 31 #include <omp.h> 32 #endif 33 34 // SSIM 35 enum { KERNEL = 3, KERNEL_SIZE = 2 * KERNEL + 1 }; 36 37 // Symmetric Gaussian kernel: K[i] = ~11 * exp(-0.3 * i * i) 38 // The maximum value (11 x 11) must be less than 128 to avoid sign 39 // problems during the calls to _mm_mullo_epi16(). 40 static const int K[KERNEL_SIZE] = { 41 1, 3, 7, 11, 7, 3, 1 // ~11 * exp(-0.3 * i * i) 42 }; 43 static const double kiW[KERNEL + 1 + 1] = { 44 1. / 1089., // 1 / sum(i:0..6, j..6) K[i]*K[j] 45 1. / 1089., // 1 / sum(i:0..6, j..6) K[i]*K[j] 46 1. / 1056., // 1 / sum(i:0..5, j..6) K[i]*K[j] 47 1. / 957., // 1 / sum(i:0..4, j..6) K[i]*K[j] 48 1. / 726., // 1 / sum(i:0..3, j..6) K[i]*K[j] 49 }; 50 51 #if !defined(LIBYUV_DISABLE_X86) && defined(__SSE2__) 52 53 #define PWEIGHT(A, B) static_cast<uint16>(K[(A)] * K[(B)]) // weight product 54 #define MAKE_WEIGHT(L) \ 55 { { { PWEIGHT(L, 0), PWEIGHT(L, 1), PWEIGHT(L, 2), PWEIGHT(L, 3), \ 56 PWEIGHT(L, 4), PWEIGHT(L, 5), PWEIGHT(L, 6), 0 } } } 57 58 // We need this union trick to be able to initialize constant static __m128i 59 // values. We can't call _mm_set_epi16() for static compile-time initialization. 60 static const struct { 61 union { 62 uint16 i16_[8]; 63 __m128i m_; 64 } values_; 65 } W0 = MAKE_WEIGHT(0), 66 W1 = MAKE_WEIGHT(1), 67 W2 = MAKE_WEIGHT(2), 68 W3 = MAKE_WEIGHT(3); 69 // ... the rest is symmetric. 70 #undef MAKE_WEIGHT 71 #undef PWEIGHT 72 #endif 73 74 // Common final expression for SSIM, once the weighted sums are known. 75 static double FinalizeSSIM(double iw, double xm, double ym, 76 double xxm, double xym, double yym) { 77 const double iwx = xm * iw; 78 const double iwy = ym * iw; 79 double sxx = xxm * iw - iwx * iwx; 80 double syy = yym * iw - iwy * iwy; 81 // small errors are possible, due to rounding. Clamp to zero. 82 if (sxx < 0.) sxx = 0.; 83 if (syy < 0.) syy = 0.; 84 const double sxsy = sqrt(sxx * syy); 85 const double sxy = xym * iw - iwx * iwy; 86 static const double C11 = (0.01 * 0.01) * (255 * 255); 87 static const double C22 = (0.03 * 0.03) * (255 * 255); 88 static const double C33 = (0.015 * 0.015) * (255 * 255); 89 const double l = (2. * iwx * iwy + C11) / (iwx * iwx + iwy * iwy + C11); 90 const double c = (2. * sxsy + C22) / (sxx + syy + C22); 91 const double s = (sxy + C33) / (sxsy + C33); 92 return l * c * s; 93 } 94 95 // GetSSIM() does clipping. GetSSIMFullKernel() does not 96 97 // TODO(skal): use summed tables? 98 // Note: worst case of accumulation is a weight of 33 = 11 + 2 * (7 + 3 + 1) 99 // with a diff of 255, squared. The maximum error is thus 0x4388241, 100 // which fits into 32 bits integers. 101 double GetSSIM(const uint8 *org, const uint8 *rec, 102 int xo, int yo, int W, int H, int stride) { 103 uint32 ws = 0, xm = 0, ym = 0, xxm = 0, xym = 0, yym = 0; 104 org += (yo - KERNEL) * stride; 105 org += (xo - KERNEL); 106 rec += (yo - KERNEL) * stride; 107 rec += (xo - KERNEL); 108 for (int y_ = 0; y_ < KERNEL_SIZE; ++y_, org += stride, rec += stride) { 109 if (((yo - KERNEL + y_) < 0) || ((yo - KERNEL + y_) >= H)) continue; 110 const int Wy = K[y_]; 111 for (int x_ = 0; x_ < KERNEL_SIZE; ++x_) { 112 const int Wxy = Wy * K[x_]; 113 if (((xo - KERNEL + x_) >= 0) && ((xo - KERNEL + x_) < W)) { 114 const int org_x = org[x_]; 115 const int rec_x = rec[x_]; 116 ws += Wxy; 117 xm += Wxy * org_x; 118 ym += Wxy * rec_x; 119 xxm += Wxy * org_x * org_x; 120 xym += Wxy * org_x * rec_x; 121 yym += Wxy * rec_x * rec_x; 122 } 123 } 124 } 125 return FinalizeSSIM(1. / ws, xm, ym, xxm, xym, yym); 126 } 127 128 double GetSSIMFullKernel(const uint8 *org, const uint8 *rec, 129 int xo, int yo, int stride, 130 double area_weight) { 131 uint32 xm = 0, ym = 0, xxm = 0, xym = 0, yym = 0; 132 133 #if defined(LIBYUV_DISABLE_X86) || !defined(__SSE2__) 134 135 org += yo * stride + xo; 136 rec += yo * stride + xo; 137 for (int y = 1; y <= KERNEL; y++) { 138 const int dy1 = y * stride; 139 const int dy2 = y * stride; 140 const int Wy = K[KERNEL + y]; 141 142 for (int x = 1; x <= KERNEL; x++) { 143 // Compute the contributions of upper-left (ul), upper-right (ur) 144 // lower-left (ll) and lower-right (lr) points (see the diagram below). 145 // Symmetric Kernel will have same weight on those points. 146 // - - - - - - - 147 // - ul - - - ur - 148 // - - - - - - - 149 // - - - 0 - - - 150 // - - - - - - - 151 // - ll - - - lr - 152 // - - - - - - - 153 const int Wxy = Wy * K[KERNEL + x]; 154 const int ul1 = org[-dy1 - x]; 155 const int ur1 = org[-dy1 + x]; 156 const int ll1 = org[dy1 - x]; 157 const int lr1 = org[dy1 + x]; 158 159 const int ul2 = rec[-dy2 - x]; 160 const int ur2 = rec[-dy2 + x]; 161 const int ll2 = rec[dy2 - x]; 162 const int lr2 = rec[dy2 + x]; 163 164 xm += Wxy * (ul1 + ur1 + ll1 + lr1); 165 ym += Wxy * (ul2 + ur2 + ll2 + lr2); 166 xxm += Wxy * (ul1 * ul1 + ur1 * ur1 + ll1 * ll1 + lr1 * lr1); 167 xym += Wxy * (ul1 * ul2 + ur1 * ur2 + ll1 * ll2 + lr1 * lr2); 168 yym += Wxy * (ul2 * ul2 + ur2 * ur2 + ll2 * ll2 + lr2 * lr2); 169 } 170 171 // Compute the contributions of up (u), down (d), left (l) and right (r) 172 // points across the main axes (see the diagram below). 173 // Symmetric Kernel will have same weight on those points. 174 // - - - - - - - 175 // - - - u - - - 176 // - - - - - - - 177 // - l - 0 - r - 178 // - - - - - - - 179 // - - - d - - - 180 // - - - - - - - 181 const int Wxy = Wy * K[KERNEL]; 182 const int u1 = org[-dy1]; 183 const int d1 = org[dy1]; 184 const int l1 = org[-y]; 185 const int r1 = org[y]; 186 187 const int u2 = rec[-dy2]; 188 const int d2 = rec[dy2]; 189 const int l2 = rec[-y]; 190 const int r2 = rec[y]; 191 192 xm += Wxy * (u1 + d1 + l1 + r1); 193 ym += Wxy * (u2 + d2 + l2 + r2); 194 xxm += Wxy * (u1 * u1 + d1 * d1 + l1 * l1 + r1 * r1); 195 xym += Wxy * (u1 * u2 + d1 * d2 + l1 * l2 + r1 * r2); 196 yym += Wxy * (u2 * u2 + d2 * d2 + l2 * l2 + r2 * r2); 197 } 198 199 // Lastly the contribution of (x0, y0) point. 200 const int Wxy = K[KERNEL] * K[KERNEL]; 201 const int s1 = org[0]; 202 const int s2 = rec[0]; 203 204 xm += Wxy * s1; 205 ym += Wxy * s2; 206 xxm += Wxy * s1 * s1; 207 xym += Wxy * s1 * s2; 208 yym += Wxy * s2 * s2; 209 210 #else // __SSE2__ 211 212 org += (yo - KERNEL) * stride + (xo - KERNEL); 213 rec += (yo - KERNEL) * stride + (xo - KERNEL); 214 215 const __m128i zero = _mm_setzero_si128(); 216 __m128i x = zero; 217 __m128i y = zero; 218 __m128i xx = zero; 219 __m128i xy = zero; 220 __m128i yy = zero; 221 222 // Read 8 pixels at line #L, and convert to 16bit, perform weighting 223 // and acccumulate. 224 #define LOAD_LINE_PAIR(L, WEIGHT) do { \ 225 const __m128i v0 = \ 226 _mm_loadl_epi64(reinterpret_cast<const __m128i*>(org + (L) * stride)); \ 227 const __m128i v1 = \ 228 _mm_loadl_epi64(reinterpret_cast<const __m128i*>(rec + (L) * stride)); \ 229 const __m128i w0 = _mm_unpacklo_epi8(v0, zero); \ 230 const __m128i w1 = _mm_unpacklo_epi8(v1, zero); \ 231 const __m128i ww0 = _mm_mullo_epi16(w0, (WEIGHT).values_.m_); \ 232 const __m128i ww1 = _mm_mullo_epi16(w1, (WEIGHT).values_.m_); \ 233 x = _mm_add_epi32(x, _mm_unpacklo_epi16(ww0, zero)); \ 234 y = _mm_add_epi32(y, _mm_unpacklo_epi16(ww1, zero)); \ 235 x = _mm_add_epi32(x, _mm_unpackhi_epi16(ww0, zero)); \ 236 y = _mm_add_epi32(y, _mm_unpackhi_epi16(ww1, zero)); \ 237 xx = _mm_add_epi32(xx, _mm_madd_epi16(ww0, w0)); \ 238 xy = _mm_add_epi32(xy, _mm_madd_epi16(ww0, w1)); \ 239 yy = _mm_add_epi32(yy, _mm_madd_epi16(ww1, w1)); \ 240 } while (0) 241 242 #define ADD_AND_STORE_FOUR_EPI32(M, OUT) do { \ 243 uint32 tmp[4]; \ 244 _mm_storeu_si128(reinterpret_cast<__m128i*>(tmp), (M)); \ 245 (OUT) = tmp[3] + tmp[2] + tmp[1] + tmp[0]; \ 246 } while (0) 247 248 LOAD_LINE_PAIR(0, W0); 249 LOAD_LINE_PAIR(1, W1); 250 LOAD_LINE_PAIR(2, W2); 251 LOAD_LINE_PAIR(3, W3); 252 LOAD_LINE_PAIR(4, W2); 253 LOAD_LINE_PAIR(5, W1); 254 LOAD_LINE_PAIR(6, W0); 255 256 ADD_AND_STORE_FOUR_EPI32(x, xm); 257 ADD_AND_STORE_FOUR_EPI32(y, ym); 258 ADD_AND_STORE_FOUR_EPI32(xx, xxm); 259 ADD_AND_STORE_FOUR_EPI32(xy, xym); 260 ADD_AND_STORE_FOUR_EPI32(yy, yym); 261 262 #undef LOAD_LINE_PAIR 263 #undef ADD_AND_STORE_FOUR_EPI32 264 #endif 265 266 return FinalizeSSIM(area_weight, xm, ym, xxm, xym, yym); 267 } 268 269 static int start_max(int x, int y) { return (x > y) ? x : y; } 270 271 double CalcSSIM(const uint8 *org, const uint8 *rec, 272 const int image_width, const int image_height) { 273 double SSIM = 0.; 274 const int KERNEL_Y = (image_height < KERNEL) ? image_height : KERNEL; 275 const int KERNEL_X = (image_width < KERNEL) ? image_width : KERNEL; 276 const int start_x = start_max(image_width - 8 + KERNEL_X, KERNEL_X); 277 const int start_y = start_max(image_height - KERNEL_Y, KERNEL_Y); 278 const int stride = image_width; 279 280 for (int j = 0; j < KERNEL_Y; ++j) { 281 for (int i = 0; i < image_width; ++i) { 282 SSIM += GetSSIM(org, rec, i, j, image_width, image_height, stride); 283 } 284 } 285 286 #ifdef _OPENMP 287 #pragma omp parallel for reduction(+: SSIM) 288 #endif 289 for (int j = KERNEL_Y; j < image_height - KERNEL_Y; ++j) { 290 for (int i = 0; i < KERNEL_X; ++i) { 291 SSIM += GetSSIM(org, rec, i, j, image_width, image_height, stride); 292 } 293 for (int i = KERNEL_X; i < start_x; ++i) { 294 SSIM += GetSSIMFullKernel(org, rec, i, j, stride, kiW[0]); 295 } 296 if (start_x < image_width) { 297 // GetSSIMFullKernel() needs to be able to read 8 pixels (in SSE2). So we 298 // copy the 8 rightmost pixels on a cache area, and pad this area with 299 // zeros which won't contribute to the overall SSIM value (but we need 300 // to pass the correct normalizing constant!). By using this cache, we can 301 // still call GetSSIMFullKernel() instead of the slower GetSSIM(). 302 // NOTE: we could use similar method for the left-most pixels too. 303 const int kScratchWidth = 8; 304 const int kScratchStride = kScratchWidth + KERNEL + 1; 305 uint8 scratch_org[KERNEL_SIZE * kScratchStride] = { 0 }; 306 uint8 scratch_rec[KERNEL_SIZE * kScratchStride] = { 0 }; 307 308 for (int k = 0; k < KERNEL_SIZE; ++k) { 309 const int offset = 310 (j - KERNEL + k) * stride + image_width - kScratchWidth; 311 memcpy(scratch_org + k * kScratchStride, org + offset, kScratchWidth); 312 memcpy(scratch_rec + k * kScratchStride, rec + offset, kScratchWidth); 313 } 314 for (int k = 0; k <= KERNEL_X + 1; ++k) { 315 SSIM += GetSSIMFullKernel(scratch_org, scratch_rec, 316 KERNEL + k, KERNEL, kScratchStride, kiW[k]); 317 } 318 } 319 } 320 321 for (int j = start_y; j < image_height; ++j) { 322 for (int i = 0; i < image_width; ++i) { 323 SSIM += GetSSIM(org, rec, i, j, image_width, image_height, stride); 324 } 325 } 326 return SSIM; 327 } 328 329 double CalcLSSIM(double ssim) { 330 return -10.0 * log10(1.0 - ssim); 331 } 332 333 #ifdef __cplusplus 334 } // extern "C" 335 #endif 336 337