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: Nathan E. Egge, at the Daala
11 * project.
12 */
13 #include <assert.h>
14 #include <math.h>
15 #include <stdlib.h>
16 #include <string.h>
17 #include "./vpx_config.h"
18 #include "./vpx_dsp_rtcd.h"
19 #include "vpx_dsp/ssim.h"
20 #include "vpx_ports/system_state.h"
21
22 typedef struct fs_level fs_level;
23 typedef struct fs_ctx fs_ctx;
24
25 #define SSIM_C1 (255 * 255 * 0.01 * 0.01)
26 #define SSIM_C2 (255 * 255 * 0.03 * 0.03)
27 #if CONFIG_VP9_HIGHBITDEPTH
28 #define SSIM_C1_10 (1023 * 1023 * 0.01 * 0.01)
29 #define SSIM_C1_12 (4095 * 4095 * 0.01 * 0.01)
30 #define SSIM_C2_10 (1023 * 1023 * 0.03 * 0.03)
31 #define SSIM_C2_12 (4095 * 4095 * 0.03 * 0.03)
32 #endif
33 #define FS_MINI(_a, _b) ((_a) < (_b) ? (_a) : (_b))
34 #define FS_MAXI(_a, _b) ((_a) > (_b) ? (_a) : (_b))
35
36 struct fs_level {
37 uint32_t *im1;
38 uint32_t *im2;
39 double *ssim;
40 int w;
41 int h;
42 };
43
44 struct fs_ctx {
45 fs_level *level;
46 int nlevels;
47 unsigned *col_buf;
48 };
49
fs_ctx_init(fs_ctx * _ctx,int _w,int _h,int _nlevels)50 static void fs_ctx_init(fs_ctx *_ctx, int _w, int _h, int _nlevels) {
51 unsigned char *data;
52 size_t data_size;
53 int lw;
54 int lh;
55 int l;
56 lw = (_w + 1) >> 1;
57 lh = (_h + 1) >> 1;
58 data_size =
59 _nlevels * sizeof(fs_level) + 2 * (lw + 8) * 8 * sizeof(*_ctx->col_buf);
60 for (l = 0; l < _nlevels; l++) {
61 size_t im_size;
62 size_t level_size;
63 im_size = lw * (size_t)lh;
64 level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
65 level_size += sizeof(*_ctx->level[l].ssim) - 1;
66 level_size /= sizeof(*_ctx->level[l].ssim);
67 level_size += im_size;
68 level_size *= sizeof(*_ctx->level[l].ssim);
69 data_size += level_size;
70 lw = (lw + 1) >> 1;
71 lh = (lh + 1) >> 1;
72 }
73 data = (unsigned char *)malloc(data_size);
74 _ctx->level = (fs_level *)data;
75 _ctx->nlevels = _nlevels;
76 data += _nlevels * sizeof(*_ctx->level);
77 lw = (_w + 1) >> 1;
78 lh = (_h + 1) >> 1;
79 for (l = 0; l < _nlevels; l++) {
80 size_t im_size;
81 size_t level_size;
82 _ctx->level[l].w = lw;
83 _ctx->level[l].h = lh;
84 im_size = lw * (size_t)lh;
85 level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
86 level_size += sizeof(*_ctx->level[l].ssim) - 1;
87 level_size /= sizeof(*_ctx->level[l].ssim);
88 level_size *= sizeof(*_ctx->level[l].ssim);
89 _ctx->level[l].im1 = (uint32_t *)data;
90 _ctx->level[l].im2 = _ctx->level[l].im1 + im_size;
91 data += level_size;
92 _ctx->level[l].ssim = (double *)data;
93 data += im_size * sizeof(*_ctx->level[l].ssim);
94 lw = (lw + 1) >> 1;
95 lh = (lh + 1) >> 1;
96 }
97 _ctx->col_buf = (unsigned *)data;
98 }
99
fs_ctx_clear(fs_ctx * _ctx)100 static void fs_ctx_clear(fs_ctx *_ctx) { free(_ctx->level); }
101
fs_downsample_level(fs_ctx * _ctx,int _l)102 static void fs_downsample_level(fs_ctx *_ctx, int _l) {
103 const uint32_t *src1;
104 const uint32_t *src2;
105 uint32_t *dst1;
106 uint32_t *dst2;
107 int w2;
108 int h2;
109 int w;
110 int h;
111 int i;
112 int j;
113 w = _ctx->level[_l].w;
114 h = _ctx->level[_l].h;
115 dst1 = _ctx->level[_l].im1;
116 dst2 = _ctx->level[_l].im2;
117 w2 = _ctx->level[_l - 1].w;
118 h2 = _ctx->level[_l - 1].h;
119 src1 = _ctx->level[_l - 1].im1;
120 src2 = _ctx->level[_l - 1].im2;
121 for (j = 0; j < h; j++) {
122 int j0offs;
123 int j1offs;
124 j0offs = 2 * j * w2;
125 j1offs = FS_MINI(2 * j + 1, h2) * w2;
126 for (i = 0; i < w; i++) {
127 int i0;
128 int i1;
129 i0 = 2 * i;
130 i1 = FS_MINI(i0 + 1, w2);
131 dst1[j * w + i] = src1[j0offs + i0] + src1[j0offs + i1] +
132 src1[j1offs + i0] + src1[j1offs + i1];
133 dst2[j * w + i] = src2[j0offs + i0] + src2[j0offs + i1] +
134 src2[j1offs + i0] + src2[j1offs + i1];
135 }
136 }
137 }
138
fs_downsample_level0(fs_ctx * _ctx,const uint8_t * _src1,int _s1ystride,const uint8_t * _src2,int _s2ystride,int _w,int _h,uint32_t bd,uint32_t shift)139 static void fs_downsample_level0(fs_ctx *_ctx, const uint8_t *_src1,
140 int _s1ystride, const uint8_t *_src2,
141 int _s2ystride, int _w, int _h, uint32_t bd,
142 uint32_t shift) {
143 uint32_t *dst1;
144 uint32_t *dst2;
145 int w;
146 int h;
147 int i;
148 int j;
149 w = _ctx->level[0].w;
150 h = _ctx->level[0].h;
151 dst1 = _ctx->level[0].im1;
152 dst2 = _ctx->level[0].im2;
153 for (j = 0; j < h; j++) {
154 int j0;
155 int j1;
156 j0 = 2 * j;
157 j1 = FS_MINI(j0 + 1, _h);
158 for (i = 0; i < w; i++) {
159 int i0;
160 int i1;
161 i0 = 2 * i;
162 i1 = FS_MINI(i0 + 1, _w);
163 if (bd == 8 && shift == 0) {
164 dst1[j * w + i] =
165 _src1[j0 * _s1ystride + i0] + _src1[j0 * _s1ystride + i1] +
166 _src1[j1 * _s1ystride + i0] + _src1[j1 * _s1ystride + i1];
167 dst2[j * w + i] =
168 _src2[j0 * _s2ystride + i0] + _src2[j0 * _s2ystride + i1] +
169 _src2[j1 * _s2ystride + i0] + _src2[j1 * _s2ystride + i1];
170 } else {
171 uint16_t *src1s = CONVERT_TO_SHORTPTR(_src1);
172 uint16_t *src2s = CONVERT_TO_SHORTPTR(_src2);
173 dst1[j * w + i] = (src1s[j0 * _s1ystride + i0] >> shift) +
174 (src1s[j0 * _s1ystride + i1] >> shift) +
175 (src1s[j1 * _s1ystride + i0] >> shift) +
176 (src1s[j1 * _s1ystride + i1] >> shift);
177 dst2[j * w + i] = (src2s[j0 * _s2ystride + i0] >> shift) +
178 (src2s[j0 * _s2ystride + i1] >> shift) +
179 (src2s[j1 * _s2ystride + i0] >> shift) +
180 (src2s[j1 * _s2ystride + i1] >> shift);
181 }
182 }
183 }
184 }
185
fs_apply_luminance(fs_ctx * _ctx,int _l,int bit_depth)186 static void fs_apply_luminance(fs_ctx *_ctx, int _l, int bit_depth) {
187 unsigned *col_sums_x;
188 unsigned *col_sums_y;
189 uint32_t *im1;
190 uint32_t *im2;
191 double *ssim;
192 double c1;
193 int w;
194 int h;
195 int j0offs;
196 int j1offs;
197 int i;
198 int j;
199 double ssim_c1 = SSIM_C1;
200 #if CONFIG_VP9_HIGHBITDEPTH
201 if (bit_depth == 10) ssim_c1 = SSIM_C1_10;
202 if (bit_depth == 12) ssim_c1 = SSIM_C1_12;
203 #else
204 assert(bit_depth == 8);
205 (void)bit_depth;
206 #endif
207 w = _ctx->level[_l].w;
208 h = _ctx->level[_l].h;
209 col_sums_x = _ctx->col_buf;
210 col_sums_y = col_sums_x + w;
211 im1 = _ctx->level[_l].im1;
212 im2 = _ctx->level[_l].im2;
213 for (i = 0; i < w; i++) col_sums_x[i] = 5 * im1[i];
214 for (i = 0; i < w; i++) col_sums_y[i] = 5 * im2[i];
215 for (j = 1; j < 4; j++) {
216 j1offs = FS_MINI(j, h - 1) * w;
217 for (i = 0; i < w; i++) col_sums_x[i] += im1[j1offs + i];
218 for (i = 0; i < w; i++) col_sums_y[i] += im2[j1offs + i];
219 }
220 ssim = _ctx->level[_l].ssim;
221 c1 = (double)(ssim_c1 * 4096 * (1 << 4 * _l));
222 for (j = 0; j < h; j++) {
223 unsigned mux;
224 unsigned muy;
225 int i0;
226 int i1;
227 mux = 5 * col_sums_x[0];
228 muy = 5 * col_sums_y[0];
229 for (i = 1; i < 4; i++) {
230 i1 = FS_MINI(i, w - 1);
231 mux += col_sums_x[i1];
232 muy += col_sums_y[i1];
233 }
234 for (i = 0; i < w; i++) {
235 ssim[j * w + i] *= (2 * mux * (double)muy + c1) /
236 (mux * (double)mux + muy * (double)muy + c1);
237 if (i + 1 < w) {
238 i0 = FS_MAXI(0, i - 4);
239 i1 = FS_MINI(i + 4, w - 1);
240 mux += col_sums_x[i1] - col_sums_x[i0];
241 muy += col_sums_x[i1] - col_sums_x[i0];
242 }
243 }
244 if (j + 1 < h) {
245 j0offs = FS_MAXI(0, j - 4) * w;
246 for (i = 0; i < w; i++) col_sums_x[i] -= im1[j0offs + i];
247 for (i = 0; i < w; i++) col_sums_y[i] -= im2[j0offs + i];
248 j1offs = FS_MINI(j + 4, h - 1) * w;
249 for (i = 0; i < w; i++) col_sums_x[i] += im1[j1offs + i];
250 for (i = 0; i < w; i++) col_sums_y[i] += im2[j1offs + i];
251 }
252 }
253 }
254
255 #define FS_COL_SET(_col, _joffs, _ioffs) \
256 do { \
257 unsigned gx; \
258 unsigned gy; \
259 gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
260 gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
261 col_sums_gx2[(_col)] = gx * (double)gx; \
262 col_sums_gy2[(_col)] = gy * (double)gy; \
263 col_sums_gxgy[(_col)] = gx * (double)gy; \
264 } while (0)
265
266 #define FS_COL_ADD(_col, _joffs, _ioffs) \
267 do { \
268 unsigned gx; \
269 unsigned gy; \
270 gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
271 gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
272 col_sums_gx2[(_col)] += gx * (double)gx; \
273 col_sums_gy2[(_col)] += gy * (double)gy; \
274 col_sums_gxgy[(_col)] += gx * (double)gy; \
275 } while (0)
276
277 #define FS_COL_SUB(_col, _joffs, _ioffs) \
278 do { \
279 unsigned gx; \
280 unsigned gy; \
281 gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
282 gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
283 col_sums_gx2[(_col)] -= gx * (double)gx; \
284 col_sums_gy2[(_col)] -= gy * (double)gy; \
285 col_sums_gxgy[(_col)] -= gx * (double)gy; \
286 } while (0)
287
288 #define FS_COL_COPY(_col1, _col2) \
289 do { \
290 col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)]; \
291 col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)]; \
292 col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)]; \
293 } while (0)
294
295 #define FS_COL_HALVE(_col1, _col2) \
296 do { \
297 col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 0.5; \
298 col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 0.5; \
299 col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 0.5; \
300 } while (0)
301
302 #define FS_COL_DOUBLE(_col1, _col2) \
303 do { \
304 col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 2; \
305 col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 2; \
306 col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 2; \
307 } while (0)
308
fs_calc_structure(fs_ctx * _ctx,int _l,int bit_depth)309 static void fs_calc_structure(fs_ctx *_ctx, int _l, int bit_depth) {
310 uint32_t *im1;
311 uint32_t *im2;
312 unsigned *gx_buf;
313 unsigned *gy_buf;
314 double *ssim;
315 double col_sums_gx2[8];
316 double col_sums_gy2[8];
317 double col_sums_gxgy[8];
318 double c2;
319 int stride;
320 int w;
321 int h;
322 int i;
323 int j;
324 double ssim_c2 = SSIM_C2;
325 #if CONFIG_VP9_HIGHBITDEPTH
326 if (bit_depth == 10) ssim_c2 = SSIM_C2_10;
327 if (bit_depth == 12) ssim_c2 = SSIM_C2_12;
328 #else
329 assert(bit_depth == 8);
330 (void)bit_depth;
331 #endif
332
333 w = _ctx->level[_l].w;
334 h = _ctx->level[_l].h;
335 im1 = _ctx->level[_l].im1;
336 im2 = _ctx->level[_l].im2;
337 ssim = _ctx->level[_l].ssim;
338 gx_buf = _ctx->col_buf;
339 stride = w + 8;
340 gy_buf = gx_buf + 8 * stride;
341 memset(gx_buf, 0, 2 * 8 * stride * sizeof(*gx_buf));
342 c2 = ssim_c2 * (1 << 4 * _l) * 16 * 104;
343 for (j = 0; j < h + 4; j++) {
344 if (j < h - 1) {
345 for (i = 0; i < w - 1; i++) {
346 unsigned g1;
347 unsigned g2;
348 unsigned gx;
349 unsigned gy;
350 g1 = abs((int)im1[(j + 1) * w + i + 1] - (int)im1[j * w + i]);
351 g2 = abs((int)im1[(j + 1) * w + i] - (int)im1[j * w + i + 1]);
352 gx = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
353 g1 = abs((int)im2[(j + 1) * w + i + 1] - (int)im2[j * w + i]);
354 g2 = abs((int)im2[(j + 1) * w + i] - (int)im2[j * w + i + 1]);
355 gy = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
356 gx_buf[(j & 7) * stride + i + 4] = gx;
357 gy_buf[(j & 7) * stride + i + 4] = gy;
358 }
359 } else {
360 memset(gx_buf + (j & 7) * stride, 0, stride * sizeof(*gx_buf));
361 memset(gy_buf + (j & 7) * stride, 0, stride * sizeof(*gy_buf));
362 }
363 if (j >= 4) {
364 int k;
365 col_sums_gx2[3] = col_sums_gx2[2] = col_sums_gx2[1] = col_sums_gx2[0] = 0;
366 col_sums_gy2[3] = col_sums_gy2[2] = col_sums_gy2[1] = col_sums_gy2[0] = 0;
367 col_sums_gxgy[3] = col_sums_gxgy[2] = col_sums_gxgy[1] =
368 col_sums_gxgy[0] = 0;
369 for (i = 4; i < 8; i++) {
370 FS_COL_SET(i, -1, 0);
371 FS_COL_ADD(i, 0, 0);
372 for (k = 1; k < 8 - i; k++) {
373 FS_COL_DOUBLE(i, i);
374 FS_COL_ADD(i, -k - 1, 0);
375 FS_COL_ADD(i, k, 0);
376 }
377 }
378 for (i = 0; i < w; i++) {
379 double mugx2;
380 double mugy2;
381 double mugxgy;
382 mugx2 = col_sums_gx2[0];
383 for (k = 1; k < 8; k++) mugx2 += col_sums_gx2[k];
384 mugy2 = col_sums_gy2[0];
385 for (k = 1; k < 8; k++) mugy2 += col_sums_gy2[k];
386 mugxgy = col_sums_gxgy[0];
387 for (k = 1; k < 8; k++) mugxgy += col_sums_gxgy[k];
388 ssim[(j - 4) * w + i] = (2 * mugxgy + c2) / (mugx2 + mugy2 + c2);
389 if (i + 1 < w) {
390 FS_COL_SET(0, -1, 1);
391 FS_COL_ADD(0, 0, 1);
392 FS_COL_SUB(2, -3, 2);
393 FS_COL_SUB(2, 2, 2);
394 FS_COL_HALVE(1, 2);
395 FS_COL_SUB(3, -4, 3);
396 FS_COL_SUB(3, 3, 3);
397 FS_COL_HALVE(2, 3);
398 FS_COL_COPY(3, 4);
399 FS_COL_DOUBLE(4, 5);
400 FS_COL_ADD(4, -4, 5);
401 FS_COL_ADD(4, 3, 5);
402 FS_COL_DOUBLE(5, 6);
403 FS_COL_ADD(5, -3, 6);
404 FS_COL_ADD(5, 2, 6);
405 FS_COL_DOUBLE(6, 7);
406 FS_COL_ADD(6, -2, 7);
407 FS_COL_ADD(6, 1, 7);
408 FS_COL_SET(7, -1, 8);
409 FS_COL_ADD(7, 0, 8);
410 }
411 }
412 }
413 }
414 }
415
416 #define FS_NLEVELS (4)
417
418 /*These weights were derived from the default weights found in Wang's original
419 Matlab implementation: {0.0448, 0.2856, 0.2363, 0.1333}.
420 We drop the finest scale and renormalize the rest to sum to 1.*/
421
422 static const double FS_WEIGHTS[FS_NLEVELS] = {
423 0.2989654541015625, 0.3141326904296875, 0.2473602294921875, 0.1395416259765625
424 };
425
fs_average(fs_ctx * _ctx,int _l)426 static double fs_average(fs_ctx *_ctx, int _l) {
427 double *ssim;
428 double ret;
429 int w;
430 int h;
431 int i;
432 int j;
433 w = _ctx->level[_l].w;
434 h = _ctx->level[_l].h;
435 ssim = _ctx->level[_l].ssim;
436 ret = 0;
437 for (j = 0; j < h; j++)
438 for (i = 0; i < w; i++) ret += ssim[j * w + i];
439 return pow(ret / (w * h), FS_WEIGHTS[_l]);
440 }
441
convert_ssim_db(double _ssim,double _weight)442 static double convert_ssim_db(double _ssim, double _weight) {
443 assert(_weight >= _ssim);
444 if ((_weight - _ssim) < 1e-10) return MAX_SSIM_DB;
445 return 10 * (log10(_weight) - log10(_weight - _ssim));
446 }
447
calc_ssim(const uint8_t * _src,int _systride,const uint8_t * _dst,int _dystride,int _w,int _h,uint32_t _bd,uint32_t _shift)448 static double calc_ssim(const uint8_t *_src, int _systride, const uint8_t *_dst,
449 int _dystride, int _w, int _h, uint32_t _bd,
450 uint32_t _shift) {
451 fs_ctx ctx;
452 double ret;
453 int l;
454 ret = 1;
455 fs_ctx_init(&ctx, _w, _h, FS_NLEVELS);
456 fs_downsample_level0(&ctx, _src, _systride, _dst, _dystride, _w, _h, _bd,
457 _shift);
458 for (l = 0; l < FS_NLEVELS - 1; l++) {
459 fs_calc_structure(&ctx, l, _bd);
460 ret *= fs_average(&ctx, l);
461 fs_downsample_level(&ctx, l + 1);
462 }
463 fs_calc_structure(&ctx, l, _bd);
464 fs_apply_luminance(&ctx, l, _bd);
465 ret *= fs_average(&ctx, l);
466 fs_ctx_clear(&ctx);
467 return ret;
468 }
469
vpx_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)470 double vpx_calc_fastssim(const YV12_BUFFER_CONFIG *source,
471 const YV12_BUFFER_CONFIG *dest, double *ssim_y,
472 double *ssim_u, double *ssim_v, uint32_t bd,
473 uint32_t in_bd) {
474 double ssimv;
475 uint32_t bd_shift = 0;
476 vpx_clear_system_state();
477 assert(bd >= in_bd);
478 bd_shift = bd - in_bd;
479
480 *ssim_y = calc_ssim(source->y_buffer, source->y_stride, dest->y_buffer,
481 dest->y_stride, source->y_crop_width,
482 source->y_crop_height, in_bd, bd_shift);
483 *ssim_u = calc_ssim(source->u_buffer, source->uv_stride, dest->u_buffer,
484 dest->uv_stride, source->uv_crop_width,
485 source->uv_crop_height, in_bd, bd_shift);
486 *ssim_v = calc_ssim(source->v_buffer, source->uv_stride, dest->v_buffer,
487 dest->uv_stride, source->uv_crop_width,
488 source->uv_crop_height, in_bd, bd_shift);
489
490 ssimv = (*ssim_y) * .8 + .1 * ((*ssim_u) + (*ssim_v));
491 return convert_ssim_db(ssimv, 1.0);
492 }
493