2 * Copyright (c) 2003-2013 Loren Merritt
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110 USA
20 * Computes the Structural Similarity Metric between two rawYV12 video files.
22 * Z. Wang, A. C. Bovik, H. R. Sheikh and E. P. Simoncelli,
23 * "Image quality assessment: From error visibility to structural similarity,"
24 * IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600-612, Apr. 2004.
26 * To improve speed, this implementation uses the standard approximation of
27 * overlapped 8x8 block sums, rather than the original gaussian weights.
35 #define FFSWAP(type,a,b) do{type SWAP_tmp= b; b= a; a= SWAP_tmp;}while(0)
36 #define FFMIN(a,b) ((a) > (b) ? (b) : (a))
39 #define PIXEL_MAX ((1 << BIT_DEPTH)-1)
40 typedef uint8_t pixel;
42 /****************************************************************************
43 * structural similarity metric
44 ****************************************************************************/
45 static void ssim_4x4x2_core( const pixel *pix1, intptr_t stride1,
46 const pixel *pix2, intptr_t stride2,
51 for( z = 0; z < 2; z++ )
53 uint32_t s1 = 0, s2 = 0, ss = 0, s12 = 0;
54 for( y = 0; y < 4; y++ )
55 for( x = 0; x < 4; x++ )
57 int a = pix1[x+y*stride1];
58 int b = pix2[x+y*stride2];
74 static float ssim_end1( int s1, int s2, int ss, int s12 )
76 /* Maximum value for 10-bit is: ss*64 = (2^10-1)^2*16*4*64 = 4286582784, which will overflow in some cases.
77 * s1*s1, s2*s2, and s1*s2 also obtain this value for edge cases: ((2^10-1)*16*4)^2 = 4286582784.
78 * Maximum value for 9-bit is: ss*64 = (2^9-1)^2*16*4*64 = 1069551616, which will not overflow. */
81 static const float ssim_c1 = .01*.01*PIXEL_MAX*PIXEL_MAX*64;
82 static const float ssim_c2 = .03*.03*PIXEL_MAX*PIXEL_MAX*64*63;
85 static const int ssim_c1 = (int)(.01*.01*PIXEL_MAX*PIXEL_MAX*64 + .5);
86 static const int ssim_c2 = (int)(.03*.03*PIXEL_MAX*PIXEL_MAX*64*63 + .5);
92 type vars = fss*64 - fs1*fs1 - fs2*fs2;
93 type covar = fs12*64 - fs1*fs2;
94 return (float)(2*fs1*fs2 + ssim_c1) * (float)(2*covar + ssim_c2)
95 / ((float)(fs1*fs1 + fs2*fs2 + ssim_c1) * (float)(vars + ssim_c2));
99 static float ssim_end4( int sum0[5][4], int sum1[5][4], int width )
104 for( i = 0; i < width; i++ )
105 ssim += ssim_end1( sum0[i][0] + sum0[i+1][0] + sum1[i][0] + sum1[i+1][0],
106 sum0[i][1] + sum0[i+1][1] + sum1[i][1] + sum1[i+1][1],
107 sum0[i][2] + sum0[i+1][2] + sum1[i][2] + sum1[i+1][2],
108 sum0[i][3] + sum0[i+1][3] + sum1[i][3] + sum1[i+1][3] );
113 pixel *pix1, intptr_t stride1,
114 pixel *pix2, intptr_t stride2,
115 int width, int height, void *buf, int *cnt )
120 int (*sum0)[4] = buf;
121 int (*sum1)[4] = sum0 + (width >> 2) + 3;
124 for( y = 1; y < height; y++ )
128 FFSWAP( void*, sum0, sum1 );
129 for( x = 0; x < width; x+=2 )
130 ssim_4x4x2_core( &pix1[4*(x+z*stride1)], stride1, &pix2[4*(x+z*stride2)], stride2, &sum0[x] );
132 for( x = 0; x < width-1; x += 4 )
133 ssim += ssim_end4( sum0+x, sum1+x, FFMIN(4,width-x-1) );
135 // *cnt = (height-1) * (width-1);
136 return ssim / ((height-1) * (width-1));
140 uint64_t ssd_plane( const uint8_t *pix1, const uint8_t *pix2, int size )
144 for( i=0; i<size; i++ )
146 int d = pix1[i] - pix2[i];
152 double ssd_to_psnr( uint64_t ssd, uint64_t denom )
154 return -10*log((double)ssd/(denom*255*255))/log(10);
157 int main(int argc, char* argv[])
160 uint8_t *buf[2], *plane[2][3];
162 uint64_t ssd[3] = {0,0,0};
163 double ssim[3] = {0,0,0};
164 int frame_size, w, h;
168 if( argc<4 || 2 != sscanf(argv[3], "%dx%d", &w, &h) )
170 printf("tiny_ssim <file1.yuv> <file2.yuv> <width>x<height> [<seek>]\n");
174 f[0] = fopen(argv[1], "rb");
175 f[1] = fopen(argv[2], "rb");
176 sscanf(argv[3], "%dx%d", &w, &h);
177 frame_size = w*h*3/2;
180 buf[i] = malloc(frame_size);
181 plane[i][0] = buf[i];
182 plane[i][1] = plane[i][0] + w*h;
183 plane[i][2] = plane[i][1] + w*h/4;
185 temp = malloc((2*w+12)*sizeof(*temp));
186 seek = argc<5 ? 0 : atoi(argv[4]);
187 fseek(f[seek<0], seek < 0 ? -seek : seek, SEEK_SET);
189 for( frames=0;; frames++ )
191 if( fread(buf[0], frame_size, 1, f[0]) != 1) break;
192 if( fread(buf[1], frame_size, 1, f[1]) != 1) break;
195 ssd[i] += ssd_plane ( plane[0][i], plane[1][i], w*h>>2*!!i );
196 ssim[i] += ssim_plane( plane[0][i], w>>!!i,
198 w>>!!i, h>>!!i, temp, NULL );
202 if( !frames ) return 0;
204 printf( "PSNR Y:%.3f U:%.3f V:%.3f All:%.3f\n",
205 ssd_to_psnr( ssd[0], (uint64_t)frames*w*h ),
206 ssd_to_psnr( ssd[1], (uint64_t)frames*w*h/4 ),
207 ssd_to_psnr( ssd[2], (uint64_t)frames*w*h/4 ),
208 ssd_to_psnr( ssd[0] + ssd[1] + ssd[2], (uint64_t)frames*w*h*3/2 ) );
209 printf( "SSIM Y:%.5f U:%.5f V:%.5f All:%.5f\n",
213 (ssim[0]*4 + ssim[1] + ssim[2]) / (frames*6) );