1 /*
2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 *
11 * This code was originally written by: Nathan E. Egge, at the Daala
12 * project.
13 */
14 #include <assert.h>
15 #include <math.h>
16 #include <stdlib.h>
17 #include <string.h>
18
19 #include "config/aom_config.h"
20 #include "config/aom_dsp_rtcd.h"
21
22 #include "aom_dsp/ssim.h"
23 #include "aom_ports/system_state.h"
24
25 typedef struct fs_level fs_level;
26 typedef struct fs_ctx fs_ctx;
27
28 #define SSIM_C1 (255 * 255 * 0.01 * 0.01)
29 #define SSIM_C2 (255 * 255 * 0.03 * 0.03)
30 #define SSIM_C1_10 (1023 * 1023 * 0.01 * 0.01)
31 #define SSIM_C1_12 (4095 * 4095 * 0.01 * 0.01)
32 #define SSIM_C2_10 (1023 * 1023 * 0.03 * 0.03)
33 #define SSIM_C2_12 (4095 * 4095 * 0.03 * 0.03)
34
35 #define FS_MINI(_a, _b) ((_a) < (_b) ? (_a) : (_b))
36 #define FS_MAXI(_a, _b) ((_a) > (_b) ? (_a) : (_b))
37
38 struct fs_level {
39 uint32_t *im1;
40 uint32_t *im2;
41 double *ssim;
42 int w;
43 int h;
44 };
45
46 struct fs_ctx {
47 fs_level *level;
48 int nlevels;
49 unsigned *col_buf;
50 };
51
fs_ctx_init(fs_ctx * _ctx,int _w,int _h,int _nlevels)52 static void fs_ctx_init(fs_ctx *_ctx, int _w, int _h, int _nlevels) {
53 unsigned char *data;
54 size_t data_size;
55 int lw;
56 int lh;
57 int l;
58 lw = (_w + 1) >> 1;
59 lh = (_h + 1) >> 1;
60 data_size =
61 _nlevels * sizeof(fs_level) + 2 * (lw + 8) * 8 * sizeof(*_ctx->col_buf);
62 for (l = 0; l < _nlevels; l++) {
63 size_t im_size;
64 size_t level_size;
65 im_size = lw * (size_t)lh;
66 level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
67 level_size += sizeof(*_ctx->level[l].ssim) - 1;
68 level_size /= sizeof(*_ctx->level[l].ssim);
69 level_size += im_size;
70 level_size *= sizeof(*_ctx->level[l].ssim);
71 data_size += level_size;
72 lw = (lw + 1) >> 1;
73 lh = (lh + 1) >> 1;
74 }
75 data = (unsigned char *)malloc(data_size);
76 _ctx->level = (fs_level *)data;
77 _ctx->nlevels = _nlevels;
78 data += _nlevels * sizeof(*_ctx->level);
79 lw = (_w + 1) >> 1;
80 lh = (_h + 1) >> 1;
81 for (l = 0; l < _nlevels; l++) {
82 size_t im_size;
83 size_t level_size;
84 _ctx->level[l].w = lw;
85 _ctx->level[l].h = lh;
86 im_size = lw * (size_t)lh;
87 level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
88 level_size += sizeof(*_ctx->level[l].ssim) - 1;
89 level_size /= sizeof(*_ctx->level[l].ssim);
90 level_size *= sizeof(*_ctx->level[l].ssim);
91 _ctx->level[l].im1 = (uint32_t *)data;
92 _ctx->level[l].im2 = _ctx->level[l].im1 + im_size;
93 data += level_size;
94 _ctx->level[l].ssim = (double *)data;
95 data += im_size * sizeof(*_ctx->level[l].ssim);
96 lw = (lw + 1) >> 1;
97 lh = (lh + 1) >> 1;
98 }
99 _ctx->col_buf = (unsigned *)data;
100 }
101
fs_ctx_clear(fs_ctx * _ctx)102 static void fs_ctx_clear(fs_ctx *_ctx) { free(_ctx->level); }
103
fs_downsample_level(fs_ctx * _ctx,int _l)104 static void fs_downsample_level(fs_ctx *_ctx, int _l) {
105 const uint32_t *src1;
106 const uint32_t *src2;
107 uint32_t *dst1;
108 uint32_t *dst2;
109 int w2;
110 int h2;
111 int w;
112 int h;
113 int i;
114 int j;
115 w = _ctx->level[_l].w;
116 h = _ctx->level[_l].h;
117 dst1 = _ctx->level[_l].im1;
118 dst2 = _ctx->level[_l].im2;
119 w2 = _ctx->level[_l - 1].w;
120 h2 = _ctx->level[_l - 1].h;
121 src1 = _ctx->level[_l - 1].im1;
122 src2 = _ctx->level[_l - 1].im2;
123 for (j = 0; j < h; j++) {
124 int j0offs;
125 int j1offs;
126 j0offs = 2 * j * w2;
127 j1offs = FS_MINI(2 * j + 1, h2) * w2;
128 for (i = 0; i < w; i++) {
129 int i0;
130 int i1;
131 i0 = 2 * i;
132 i1 = FS_MINI(i0 + 1, w2);
133 dst1[j * w + i] = src1[j0offs + i0] + src1[j0offs + i1] +
134 src1[j1offs + i0] + src1[j1offs + i1];
135 dst2[j * w + i] = src2[j0offs + i0] + src2[j0offs + i1] +
136 src2[j1offs + i0] + src2[j1offs + i1];
137 }
138 }
139 }
140
fs_downsample_level0(fs_ctx * _ctx,const uint8_t * _src1,int _s1ystride,const uint8_t * _src2,int _s2ystride,int _w,int _h,uint32_t shift,int buf_is_hbd)141 static void fs_downsample_level0(fs_ctx *_ctx, const uint8_t *_src1,
142 int _s1ystride, const uint8_t *_src2,
143 int _s2ystride, int _w, int _h, uint32_t shift,
144 int buf_is_hbd) {
145 uint32_t *dst1;
146 uint32_t *dst2;
147 int w;
148 int h;
149 int i;
150 int j;
151 w = _ctx->level[0].w;
152 h = _ctx->level[0].h;
153 dst1 = _ctx->level[0].im1;
154 dst2 = _ctx->level[0].im2;
155 for (j = 0; j < h; j++) {
156 int j0;
157 int j1;
158 j0 = 2 * j;
159 j1 = FS_MINI(j0 + 1, _h);
160 for (i = 0; i < w; i++) {
161 int i0;
162 int i1;
163 i0 = 2 * i;
164 i1 = FS_MINI(i0 + 1, _w);
165 if (!buf_is_hbd) {
166 dst1[j * w + i] =
167 _src1[j0 * _s1ystride + i0] + _src1[j0 * _s1ystride + i1] +
168 _src1[j1 * _s1ystride + i0] + _src1[j1 * _s1ystride + i1];
169 dst2[j * w + i] =
170 _src2[j0 * _s2ystride + i0] + _src2[j0 * _s2ystride + i1] +
171 _src2[j1 * _s2ystride + i0] + _src2[j1 * _s2ystride + i1];
172 } else {
173 uint16_t *src1s = CONVERT_TO_SHORTPTR(_src1);
174 uint16_t *src2s = CONVERT_TO_SHORTPTR(_src2);
175 dst1[j * w + i] = (src1s[j0 * _s1ystride + i0] >> shift) +
176 (src1s[j0 * _s1ystride + i1] >> shift) +
177 (src1s[j1 * _s1ystride + i0] >> shift) +
178 (src1s[j1 * _s1ystride + i1] >> shift);
179 dst2[j * w + i] = (src2s[j0 * _s2ystride + i0] >> shift) +
180 (src2s[j0 * _s2ystride + i1] >> shift) +
181 (src2s[j1 * _s2ystride + i0] >> shift) +
182 (src2s[j1 * _s2ystride + i1] >> shift);
183 }
184 }
185 }
186 }
187
fs_apply_luminance(fs_ctx * _ctx,int _l,int bit_depth)188 static void fs_apply_luminance(fs_ctx *_ctx, int _l, int bit_depth) {
189 unsigned *col_sums_x;
190 unsigned *col_sums_y;
191 uint32_t *im1;
192 uint32_t *im2;
193 double *ssim;
194 double c1;
195 int w;
196 int h;
197 int j0offs;
198 int j1offs;
199 int i;
200 int j;
201 double ssim_c1 = SSIM_C1;
202
203 if (bit_depth == 10) ssim_c1 = SSIM_C1_10;
204 if (bit_depth == 12) ssim_c1 = SSIM_C1_12;
205
206 w = _ctx->level[_l].w;
207 h = _ctx->level[_l].h;
208 col_sums_x = _ctx->col_buf;
209 col_sums_y = col_sums_x + w;
210 im1 = _ctx->level[_l].im1;
211 im2 = _ctx->level[_l].im2;
212 for (i = 0; i < w; i++) col_sums_x[i] = 5 * im1[i];
213 for (i = 0; i < w; i++) col_sums_y[i] = 5 * im2[i];
214 for (j = 1; j < 4; j++) {
215 j1offs = FS_MINI(j, h - 1) * w;
216 for (i = 0; i < w; i++) col_sums_x[i] += im1[j1offs + i];
217 for (i = 0; i < w; i++) col_sums_y[i] += im2[j1offs + i];
218 }
219 ssim = _ctx->level[_l].ssim;
220 c1 = (double)(ssim_c1 * 4096 * (1 << 4 * _l));
221 for (j = 0; j < h; j++) {
222 unsigned mux;
223 unsigned muy;
224 int i0;
225 int i1;
226 mux = 5 * col_sums_x[0];
227 muy = 5 * col_sums_y[0];
228 for (i = 1; i < 4; i++) {
229 i1 = FS_MINI(i, w - 1);
230 mux += col_sums_x[i1];
231 muy += col_sums_y[i1];
232 }
233 for (i = 0; i < w; i++) {
234 ssim[j * w + i] *= (2 * mux * (double)muy + c1) /
235 (mux * (double)mux + muy * (double)muy + c1);
236 if (i + 1 < w) {
237 i0 = FS_MAXI(0, i - 4);
238 i1 = FS_MINI(i + 4, w - 1);
239 mux += col_sums_x[i1] - col_sums_x[i0];
240 muy += col_sums_x[i1] - col_sums_x[i0];
241 }
242 }
243 if (j + 1 < h) {
244 j0offs = FS_MAXI(0, j - 4) * w;
245 for (i = 0; i < w; i++) col_sums_x[i] -= im1[j0offs + i];
246 for (i = 0; i < w; i++) col_sums_y[i] -= im2[j0offs + i];
247 j1offs = FS_MINI(j + 4, h - 1) * w;
248 for (i = 0; i < w; i++) col_sums_x[i] += im1[j1offs + i];
249 for (i = 0; i < w; i++) col_sums_y[i] += im2[j1offs + i];
250 }
251 }
252 }
253
254 #define FS_COL_SET(_col, _joffs, _ioffs) \
255 do { \
256 unsigned gx; \
257 unsigned gy; \
258 gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
259 gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
260 col_sums_gx2[(_col)] = gx * (double)gx; \
261 col_sums_gy2[(_col)] = gy * (double)gy; \
262 col_sums_gxgy[(_col)] = gx * (double)gy; \
263 } while (0)
264
265 #define FS_COL_ADD(_col, _joffs, _ioffs) \
266 do { \
267 unsigned gx; \
268 unsigned gy; \
269 gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
270 gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
271 col_sums_gx2[(_col)] += gx * (double)gx; \
272 col_sums_gy2[(_col)] += gy * (double)gy; \
273 col_sums_gxgy[(_col)] += gx * (double)gy; \
274 } while (0)
275
276 #define FS_COL_SUB(_col, _joffs, _ioffs) \
277 do { \
278 unsigned gx; \
279 unsigned gy; \
280 gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
281 gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
282 col_sums_gx2[(_col)] -= gx * (double)gx; \
283 col_sums_gy2[(_col)] -= gy * (double)gy; \
284 col_sums_gxgy[(_col)] -= gx * (double)gy; \
285 } while (0)
286
287 #define FS_COL_COPY(_col1, _col2) \
288 do { \
289 col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)]; \
290 col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)]; \
291 col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)]; \
292 } while (0)
293
294 #define FS_COL_HALVE(_col1, _col2) \
295 do { \
296 col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 0.5; \
297 col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 0.5; \
298 col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 0.5; \
299 } while (0)
300
301 #define FS_COL_DOUBLE(_col1, _col2) \
302 do { \
303 col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 2; \
304 col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 2; \
305 col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 2; \
306 } while (0)
307
fs_calc_structure(fs_ctx * _ctx,int _l,int bit_depth)308 static void fs_calc_structure(fs_ctx *_ctx, int _l, int bit_depth) {
309 uint32_t *im1;
310 uint32_t *im2;
311 unsigned *gx_buf;
312 unsigned *gy_buf;
313 double *ssim;
314 double col_sums_gx2[8];
315 double col_sums_gy2[8];
316 double col_sums_gxgy[8];
317 double c2;
318 int stride;
319 int w;
320 int h;
321 int i;
322 int j;
323 double ssim_c2 = SSIM_C2;
324 if (bit_depth == 10) ssim_c2 = SSIM_C2_10;
325 if (bit_depth == 12) ssim_c2 = SSIM_C2_12;
326
327 w = _ctx->level[_l].w;
328 h = _ctx->level[_l].h;
329 im1 = _ctx->level[_l].im1;
330 im2 = _ctx->level[_l].im2;
331 ssim = _ctx->level[_l].ssim;
332 gx_buf = _ctx->col_buf;
333 stride = w + 8;
334 gy_buf = gx_buf + 8 * stride;
335 memset(gx_buf, 0, 2 * 8 * stride * sizeof(*gx_buf));
336 c2 = ssim_c2 * (1 << 4 * _l) * 16 * 104;
337 for (j = 0; j < h + 4; j++) {
338 if (j < h - 1) {
339 for (i = 0; i < w - 1; i++) {
340 unsigned g1;
341 unsigned g2;
342 unsigned gx;
343 unsigned gy;
344 g1 = abs((int)im1[(j + 1) * w + i + 1] - (int)im1[j * w + i]);
345 g2 = abs((int)im1[(j + 1) * w + i] - (int)im1[j * w + i + 1]);
346 gx = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
347 g1 = abs((int)im2[(j + 1) * w + i + 1] - (int)im2[j * w + i]);
348 g2 = abs((int)im2[(j + 1) * w + i] - (int)im2[j * w + i + 1]);
349 gy = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
350 gx_buf[(j & 7) * stride + i + 4] = gx;
351 gy_buf[(j & 7) * stride + i + 4] = gy;
352 }
353 } else {
354 memset(gx_buf + (j & 7) * stride, 0, stride * sizeof(*gx_buf));
355 memset(gy_buf + (j & 7) * stride, 0, stride * sizeof(*gy_buf));
356 }
357 if (j >= 4) {
358 int k;
359 col_sums_gx2[3] = col_sums_gx2[2] = col_sums_gx2[1] = col_sums_gx2[0] = 0;
360 col_sums_gy2[3] = col_sums_gy2[2] = col_sums_gy2[1] = col_sums_gy2[0] = 0;
361 col_sums_gxgy[3] = col_sums_gxgy[2] = col_sums_gxgy[1] =
362 col_sums_gxgy[0] = 0;
363 for (i = 4; i < 8; i++) {
364 FS_COL_SET(i, -1, 0);
365 FS_COL_ADD(i, 0, 0);
366 for (k = 1; k < 8 - i; k++) {
367 FS_COL_DOUBLE(i, i);
368 FS_COL_ADD(i, -k - 1, 0);
369 FS_COL_ADD(i, k, 0);
370 }
371 }
372 for (i = 0; i < w; i++) {
373 double mugx2;
374 double mugy2;
375 double mugxgy;
376 mugx2 = col_sums_gx2[0];
377 for (k = 1; k < 8; k++) mugx2 += col_sums_gx2[k];
378 mugy2 = col_sums_gy2[0];
379 for (k = 1; k < 8; k++) mugy2 += col_sums_gy2[k];
380 mugxgy = col_sums_gxgy[0];
381 for (k = 1; k < 8; k++) mugxgy += col_sums_gxgy[k];
382 ssim[(j - 4) * w + i] = (2 * mugxgy + c2) / (mugx2 + mugy2 + c2);
383 if (i + 1 < w) {
384 FS_COL_SET(0, -1, 1);
385 FS_COL_ADD(0, 0, 1);
386 FS_COL_SUB(2, -3, 2);
387 FS_COL_SUB(2, 2, 2);
388 FS_COL_HALVE(1, 2);
389 FS_COL_SUB(3, -4, 3);
390 FS_COL_SUB(3, 3, 3);
391 FS_COL_HALVE(2, 3);
392 FS_COL_COPY(3, 4);
393 FS_COL_DOUBLE(4, 5);
394 FS_COL_ADD(4, -4, 5);
395 FS_COL_ADD(4, 3, 5);
396 FS_COL_DOUBLE(5, 6);
397 FS_COL_ADD(5, -3, 6);
398 FS_COL_ADD(5, 2, 6);
399 FS_COL_DOUBLE(6, 7);
400 FS_COL_ADD(6, -2, 7);
401 FS_COL_ADD(6, 1, 7);
402 FS_COL_SET(7, -1, 8);
403 FS_COL_ADD(7, 0, 8);
404 }
405 }
406 }
407 }
408 }
409
410 #define FS_NLEVELS (4)
411
412 /*These weights were derived from the default weights found in Wang's original
413 Matlab implementation: {0.0448, 0.2856, 0.2363, 0.1333}.
414 We drop the finest scale and renormalize the rest to sum to 1.*/
415
416 static const double FS_WEIGHTS[FS_NLEVELS] = {
417 0.2989654541015625, 0.3141326904296875, 0.2473602294921875, 0.1395416259765625
418 };
419
fs_average(fs_ctx * _ctx,int _l)420 static double fs_average(fs_ctx *_ctx, int _l) {
421 double *ssim;
422 double ret;
423 int w;
424 int h;
425 int i;
426 int j;
427 w = _ctx->level[_l].w;
428 h = _ctx->level[_l].h;
429 ssim = _ctx->level[_l].ssim;
430 ret = 0;
431 for (j = 0; j < h; j++)
432 for (i = 0; i < w; i++) ret += ssim[j * w + i];
433 return pow(ret / (w * h), FS_WEIGHTS[_l]);
434 }
435
convert_ssim_db(double _ssim,double _weight)436 static double convert_ssim_db(double _ssim, double _weight) {
437 assert(_weight >= _ssim);
438 if ((_weight - _ssim) < 1e-10) return MAX_SSIM_DB;
439 return 10 * (log10(_weight) - log10(_weight - _ssim));
440 }
441
calc_ssim(const uint8_t * _src,int _systride,const uint8_t * _dst,int _dystride,int _w,int _h,uint32_t _bd,uint32_t _shift,int buf_is_hbd)442 static double calc_ssim(const uint8_t *_src, int _systride, const uint8_t *_dst,
443 int _dystride, int _w, int _h, uint32_t _bd,
444 uint32_t _shift, int buf_is_hbd) {
445 fs_ctx ctx;
446 double ret;
447 int l;
448 ret = 1;
449 fs_ctx_init(&ctx, _w, _h, FS_NLEVELS);
450 fs_downsample_level0(&ctx, _src, _systride, _dst, _dystride, _w, _h, _shift,
451 buf_is_hbd);
452 for (l = 0; l < FS_NLEVELS - 1; l++) {
453 fs_calc_structure(&ctx, l, _bd);
454 ret *= fs_average(&ctx, l);
455 fs_downsample_level(&ctx, l + 1);
456 }
457 fs_calc_structure(&ctx, l, _bd);
458 fs_apply_luminance(&ctx, l, _bd);
459 ret *= fs_average(&ctx, l);
460 fs_ctx_clear(&ctx);
461 return ret;
462 }
463
aom_calc_fastssim(const YV12_BUFFER_CONFIG * source,const YV12_BUFFER_CONFIG * dest,double * ssim_y,double * ssim_u,double * ssim_v,uint32_t bd,uint32_t in_bd)464 double aom_calc_fastssim(const YV12_BUFFER_CONFIG *source,
465 const YV12_BUFFER_CONFIG *dest, double *ssim_y,
466 double *ssim_u, double *ssim_v, uint32_t bd,
467 uint32_t in_bd) {
468 double ssimv;
469 uint32_t bd_shift = 0;
470 aom_clear_system_state();
471 assert(bd >= in_bd);
472 assert(source->flags == dest->flags);
473 int buf_is_hbd = source->flags & YV12_FLAG_HIGHBITDEPTH;
474 bd_shift = bd - in_bd;
475
476 *ssim_y = calc_ssim(source->y_buffer, source->y_stride, dest->y_buffer,
477 dest->y_stride, source->y_crop_width,
478 source->y_crop_height, in_bd, bd_shift, buf_is_hbd);
479 *ssim_u = calc_ssim(source->u_buffer, source->uv_stride, dest->u_buffer,
480 dest->uv_stride, source->uv_crop_width,
481 source->uv_crop_height, in_bd, bd_shift, buf_is_hbd);
482 *ssim_v = calc_ssim(source->v_buffer, source->uv_stride, dest->v_buffer,
483 dest->uv_stride, source->uv_crop_width,
484 source->uv_crop_height, in_bd, bd_shift, buf_is_hbd);
485 ssimv = (*ssim_y) * .8 + .1 * ((*ssim_u) + (*ssim_v));
486 return convert_ssim_db(ssimv, 1.0);
487 }
488