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 int 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 if (!data) return -1;
75 _ctx->level = (fs_level *)data;
76 _ctx->nlevels = _nlevels;
77 data += _nlevels * sizeof(*_ctx->level);
78 lw = (_w + 1) >> 1;
79 lh = (_h + 1) >> 1;
80 for (l = 0; l < _nlevels; l++) {
81 size_t im_size;
82 size_t level_size;
83 _ctx->level[l].w = lw;
84 _ctx->level[l].h = lh;
85 im_size = lw * (size_t)lh;
86 level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
87 level_size += sizeof(*_ctx->level[l].ssim) - 1;
88 level_size /= sizeof(*_ctx->level[l].ssim);
89 level_size *= sizeof(*_ctx->level[l].ssim);
90 _ctx->level[l].im1 = (uint32_t *)data;
91 _ctx->level[l].im2 = _ctx->level[l].im1 + im_size;
92 data += level_size;
93 _ctx->level[l].ssim = (double *)data;
94 data += im_size * sizeof(*_ctx->level[l].ssim);
95 lw = (lw + 1) >> 1;
96 lh = (lh + 1) >> 1;
97 }
98 _ctx->col_buf = (unsigned *)data;
99 return 0;
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] =
134 (uint32_t)((int64_t)src1[j0offs + i0] + src1[j0offs + i1] +
135 src1[j1offs + i0] + src1[j1offs + i1]);
136 dst2[j * w + i] =
137 (uint32_t)((int64_t)src2[j0offs + i0] + src2[j0offs + i1] +
138 src2[j1offs + i0] + src2[j1offs + i1]);
139 }
140 }
141 }
142
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)143 static void fs_downsample_level0(fs_ctx *_ctx, const uint8_t *_src1,
144 int _s1ystride, const uint8_t *_src2,
145 int _s2ystride, int _w, int _h, uint32_t bd,
146 uint32_t shift) {
147 uint32_t *dst1;
148 uint32_t *dst2;
149 int w;
150 int h;
151 int i;
152 int j;
153 w = _ctx->level[0].w;
154 h = _ctx->level[0].h;
155 dst1 = _ctx->level[0].im1;
156 dst2 = _ctx->level[0].im2;
157 for (j = 0; j < h; j++) {
158 int j0;
159 int j1;
160 j0 = 2 * j;
161 j1 = FS_MINI(j0 + 1, _h);
162 for (i = 0; i < w; i++) {
163 int i0;
164 int i1;
165 i0 = 2 * i;
166 i1 = FS_MINI(i0 + 1, _w);
167 if (bd == 8 && shift == 0) {
168 dst1[j * w + i] =
169 _src1[j0 * _s1ystride + i0] + _src1[j0 * _s1ystride + i1] +
170 _src1[j1 * _s1ystride + i0] + _src1[j1 * _s1ystride + i1];
171 dst2[j * w + i] =
172 _src2[j0 * _s2ystride + i0] + _src2[j0 * _s2ystride + i1] +
173 _src2[j1 * _s2ystride + i0] + _src2[j1 * _s2ystride + i1];
174 } else {
175 uint16_t *src1s = CONVERT_TO_SHORTPTR(_src1);
176 uint16_t *src2s = CONVERT_TO_SHORTPTR(_src2);
177 dst1[j * w + i] = (src1s[j0 * _s1ystride + i0] >> shift) +
178 (src1s[j0 * _s1ystride + i1] >> shift) +
179 (src1s[j1 * _s1ystride + i0] >> shift) +
180 (src1s[j1 * _s1ystride + i1] >> shift);
181 dst2[j * w + i] = (src2s[j0 * _s2ystride + i0] >> shift) +
182 (src2s[j0 * _s2ystride + i1] >> shift) +
183 (src2s[j1 * _s2ystride + i0] >> shift) +
184 (src2s[j1 * _s2ystride + i1] >> shift);
185 }
186 }
187 }
188 }
189
fs_apply_luminance(fs_ctx * _ctx,int _l,int bit_depth)190 static void fs_apply_luminance(fs_ctx *_ctx, int _l, int bit_depth) {
191 unsigned *col_sums_x;
192 unsigned *col_sums_y;
193 uint32_t *im1;
194 uint32_t *im2;
195 double *ssim;
196 double c1;
197 int w;
198 int h;
199 int j0offs;
200 int j1offs;
201 int i;
202 int j;
203 double ssim_c1 = SSIM_C1;
204 #if CONFIG_VP9_HIGHBITDEPTH
205 if (bit_depth == 10) ssim_c1 = SSIM_C1_10;
206 if (bit_depth == 12) ssim_c1 = SSIM_C1_12;
207 #else
208 assert(bit_depth == 8);
209 (void)bit_depth;
210 #endif
211 w = _ctx->level[_l].w;
212 h = _ctx->level[_l].h;
213 col_sums_x = _ctx->col_buf;
214 col_sums_y = col_sums_x + w;
215 im1 = _ctx->level[_l].im1;
216 im2 = _ctx->level[_l].im2;
217 for (i = 0; i < w; i++) col_sums_x[i] = 5 * im1[i];
218 for (i = 0; i < w; i++) col_sums_y[i] = 5 * im2[i];
219 for (j = 1; j < 4; j++) {
220 j1offs = FS_MINI(j, h - 1) * w;
221 for (i = 0; i < w; i++) col_sums_x[i] += im1[j1offs + i];
222 for (i = 0; i < w; i++) col_sums_y[i] += im2[j1offs + i];
223 }
224 ssim = _ctx->level[_l].ssim;
225 c1 = (double)(ssim_c1 * 4096 * (1 << 4 * _l));
226 for (j = 0; j < h; j++) {
227 int64_t mux;
228 int64_t muy;
229 int i0;
230 int i1;
231 mux = (int64_t)5 * col_sums_x[0];
232 muy = (int64_t)5 * col_sums_y[0];
233 for (i = 1; i < 4; i++) {
234 i1 = FS_MINI(i, w - 1);
235 mux += col_sums_x[i1];
236 muy += col_sums_y[i1];
237 }
238 for (i = 0; i < w; i++) {
239 ssim[j * w + i] *= (2 * mux * (double)muy + c1) /
240 (mux * (double)mux + muy * (double)muy + c1);
241 if (i + 1 < w) {
242 i0 = FS_MAXI(0, i - 4);
243 i1 = FS_MINI(i + 4, w - 1);
244 mux += (int)col_sums_x[i1] - (int)col_sums_x[i0];
245 muy += (int)col_sums_x[i1] - (int)col_sums_x[i0];
246 }
247 }
248 if (j + 1 < h) {
249 j0offs = FS_MAXI(0, j - 4) * w;
250 for (i = 0; i < w; i++) col_sums_x[i] -= im1[j0offs + i];
251 for (i = 0; i < w; i++) col_sums_y[i] -= im2[j0offs + i];
252 j1offs = FS_MINI(j + 4, h - 1) * w;
253 for (i = 0; i < w; i++)
254 col_sums_x[i] = (uint32_t)((int64_t)col_sums_x[i] + im1[j1offs + i]);
255 for (i = 0; i < w; i++)
256 col_sums_y[i] = (uint32_t)((int64_t)col_sums_y[i] + im2[j1offs + i]);
257 }
258 }
259 }
260
261 #define FS_COL_SET(_col, _joffs, _ioffs) \
262 do { \
263 unsigned gx; \
264 unsigned gy; \
265 gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
266 gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
267 col_sums_gx2[(_col)] = gx * (double)gx; \
268 col_sums_gy2[(_col)] = gy * (double)gy; \
269 col_sums_gxgy[(_col)] = gx * (double)gy; \
270 } while (0)
271
272 #define FS_COL_ADD(_col, _joffs, _ioffs) \
273 do { \
274 unsigned gx; \
275 unsigned gy; \
276 gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
277 gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
278 col_sums_gx2[(_col)] += gx * (double)gx; \
279 col_sums_gy2[(_col)] += gy * (double)gy; \
280 col_sums_gxgy[(_col)] += gx * (double)gy; \
281 } while (0)
282
283 #define FS_COL_SUB(_col, _joffs, _ioffs) \
284 do { \
285 unsigned gx; \
286 unsigned gy; \
287 gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
288 gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
289 col_sums_gx2[(_col)] -= gx * (double)gx; \
290 col_sums_gy2[(_col)] -= gy * (double)gy; \
291 col_sums_gxgy[(_col)] -= gx * (double)gy; \
292 } while (0)
293
294 #define FS_COL_COPY(_col1, _col2) \
295 do { \
296 col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)]; \
297 col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)]; \
298 col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)]; \
299 } while (0)
300
301 #define FS_COL_HALVE(_col1, _col2) \
302 do { \
303 col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 0.5; \
304 col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 0.5; \
305 col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 0.5; \
306 } while (0)
307
308 #define FS_COL_DOUBLE(_col1, _col2) \
309 do { \
310 col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 2; \
311 col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 2; \
312 col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 2; \
313 } while (0)
314
fs_calc_structure(fs_ctx * _ctx,int _l,int bit_depth)315 static void fs_calc_structure(fs_ctx *_ctx, int _l, int bit_depth) {
316 uint32_t *im1;
317 uint32_t *im2;
318 unsigned *gx_buf;
319 unsigned *gy_buf;
320 double *ssim;
321 double col_sums_gx2[8];
322 double col_sums_gy2[8];
323 double col_sums_gxgy[8];
324 double c2;
325 int stride;
326 int w;
327 int h;
328 int i;
329 int j;
330 double ssim_c2 = SSIM_C2;
331 #if CONFIG_VP9_HIGHBITDEPTH
332 if (bit_depth == 10) ssim_c2 = SSIM_C2_10;
333 if (bit_depth == 12) ssim_c2 = SSIM_C2_12;
334 #else
335 assert(bit_depth == 8);
336 (void)bit_depth;
337 #endif
338
339 w = _ctx->level[_l].w;
340 h = _ctx->level[_l].h;
341 im1 = _ctx->level[_l].im1;
342 im2 = _ctx->level[_l].im2;
343 ssim = _ctx->level[_l].ssim;
344 gx_buf = _ctx->col_buf;
345 stride = w + 8;
346 gy_buf = gx_buf + 8 * stride;
347 memset(gx_buf, 0, 2 * 8 * stride * sizeof(*gx_buf));
348 c2 = ssim_c2 * (1 << 4 * _l) * 16 * 104;
349 for (j = 0; j < h + 4; j++) {
350 if (j < h - 1) {
351 for (i = 0; i < w - 1; i++) {
352 int64_t g1;
353 int64_t g2;
354 int64_t gx;
355 int64_t gy;
356 g1 = labs((int64_t)im1[(j + 1) * w + i + 1] - (int64_t)im1[j * w + i]);
357 g2 = labs((int64_t)im1[(j + 1) * w + i] - (int64_t)im1[j * w + i + 1]);
358 gx = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
359 g1 = labs((int64_t)im2[(j + 1) * w + i + 1] - (int64_t)im2[j * w + i]);
360 g2 = labs((int64_t)im2[(j + 1) * w + i] - (int64_t)im2[j * w + i + 1]);
361 gy = ((int64_t)4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2));
362 gx_buf[(j & 7) * stride + i + 4] = (uint32_t)gx;
363 gy_buf[(j & 7) * stride + i + 4] = (uint32_t)gy;
364 }
365 } else {
366 memset(gx_buf + (j & 7) * stride, 0, stride * sizeof(*gx_buf));
367 memset(gy_buf + (j & 7) * stride, 0, stride * sizeof(*gy_buf));
368 }
369 if (j >= 4) {
370 int k;
371 col_sums_gx2[3] = col_sums_gx2[2] = col_sums_gx2[1] = col_sums_gx2[0] = 0;
372 col_sums_gy2[3] = col_sums_gy2[2] = col_sums_gy2[1] = col_sums_gy2[0] = 0;
373 col_sums_gxgy[3] = col_sums_gxgy[2] = col_sums_gxgy[1] =
374 col_sums_gxgy[0] = 0;
375 for (i = 4; i < 8; i++) {
376 FS_COL_SET(i, -1, 0);
377 FS_COL_ADD(i, 0, 0);
378 for (k = 1; k < 8 - i; k++) {
379 FS_COL_DOUBLE(i, i);
380 FS_COL_ADD(i, -k - 1, 0);
381 FS_COL_ADD(i, k, 0);
382 }
383 }
384 for (i = 0; i < w; i++) {
385 double mugx2;
386 double mugy2;
387 double mugxgy;
388 mugx2 = col_sums_gx2[0];
389 for (k = 1; k < 8; k++) mugx2 += col_sums_gx2[k];
390 mugy2 = col_sums_gy2[0];
391 for (k = 1; k < 8; k++) mugy2 += col_sums_gy2[k];
392 mugxgy = col_sums_gxgy[0];
393 for (k = 1; k < 8; k++) mugxgy += col_sums_gxgy[k];
394 ssim[(j - 4) * w + i] = (2 * mugxgy + c2) / (mugx2 + mugy2 + c2);
395 if (i + 1 < w) {
396 FS_COL_SET(0, -1, 1);
397 FS_COL_ADD(0, 0, 1);
398 FS_COL_SUB(2, -3, 2);
399 FS_COL_SUB(2, 2, 2);
400 FS_COL_HALVE(1, 2);
401 FS_COL_SUB(3, -4, 3);
402 FS_COL_SUB(3, 3, 3);
403 FS_COL_HALVE(2, 3);
404 FS_COL_COPY(3, 4);
405 FS_COL_DOUBLE(4, 5);
406 FS_COL_ADD(4, -4, 5);
407 FS_COL_ADD(4, 3, 5);
408 FS_COL_DOUBLE(5, 6);
409 FS_COL_ADD(5, -3, 6);
410 FS_COL_ADD(5, 2, 6);
411 FS_COL_DOUBLE(6, 7);
412 FS_COL_ADD(6, -2, 7);
413 FS_COL_ADD(6, 1, 7);
414 FS_COL_SET(7, -1, 8);
415 FS_COL_ADD(7, 0, 8);
416 }
417 }
418 }
419 }
420 }
421
422 #define FS_NLEVELS (4)
423
424 /*These weights were derived from the default weights found in Wang's original
425 Matlab implementation: {0.0448, 0.2856, 0.2363, 0.1333}.
426 We drop the finest scale and renormalize the rest to sum to 1.*/
427
428 static const double FS_WEIGHTS[FS_NLEVELS] = {
429 0.2989654541015625, 0.3141326904296875, 0.2473602294921875, 0.1395416259765625
430 };
431
fs_average(fs_ctx * _ctx,int _l)432 static double fs_average(fs_ctx *_ctx, int _l) {
433 double *ssim;
434 double ret;
435 int w;
436 int h;
437 int i;
438 int j;
439 w = _ctx->level[_l].w;
440 h = _ctx->level[_l].h;
441 ssim = _ctx->level[_l].ssim;
442 ret = 0;
443 for (j = 0; j < h; j++)
444 for (i = 0; i < w; i++) ret += ssim[j * w + i];
445 return pow(ret / (w * h), FS_WEIGHTS[_l]);
446 }
447
convert_ssim_db(double _ssim,double _weight)448 static double convert_ssim_db(double _ssim, double _weight) {
449 assert(_weight >= _ssim);
450 if ((_weight - _ssim) < 1e-10) return MAX_SSIM_DB;
451 return 10 * (log10(_weight) - log10(_weight - _ssim));
452 }
453
calc_ssim(const uint8_t * _src,int _systride,const uint8_t * _dst,int _dystride,int _w,int _h,uint32_t _bd,uint32_t _shift)454 static double calc_ssim(const uint8_t *_src, int _systride, const uint8_t *_dst,
455 int _dystride, int _w, int _h, uint32_t _bd,
456 uint32_t _shift) {
457 fs_ctx ctx;
458 double ret;
459 int l;
460 ret = 1;
461 if (fs_ctx_init(&ctx, _w, _h, FS_NLEVELS)) return 99.0;
462 fs_downsample_level0(&ctx, _src, _systride, _dst, _dystride, _w, _h, _bd,
463 _shift);
464 for (l = 0; l < FS_NLEVELS - 1; l++) {
465 fs_calc_structure(&ctx, l, _bd);
466 ret *= fs_average(&ctx, l);
467 fs_downsample_level(&ctx, l + 1);
468 }
469 fs_calc_structure(&ctx, l, _bd);
470 fs_apply_luminance(&ctx, l, _bd);
471 ret *= fs_average(&ctx, l);
472 fs_ctx_clear(&ctx);
473 return ret;
474 }
475
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)476 double vpx_calc_fastssim(const YV12_BUFFER_CONFIG *source,
477 const YV12_BUFFER_CONFIG *dest, double *ssim_y,
478 double *ssim_u, double *ssim_v, uint32_t bd,
479 uint32_t in_bd) {
480 double ssimv;
481 uint32_t bd_shift = 0;
482 vpx_clear_system_state();
483 assert(bd >= in_bd);
484 bd_shift = bd - in_bd;
485
486 *ssim_y = calc_ssim(source->y_buffer, source->y_stride, dest->y_buffer,
487 dest->y_stride, source->y_crop_width,
488 source->y_crop_height, in_bd, bd_shift);
489 *ssim_u = calc_ssim(source->u_buffer, source->uv_stride, dest->u_buffer,
490 dest->uv_stride, source->uv_crop_width,
491 source->uv_crop_height, in_bd, bd_shift);
492 *ssim_v = calc_ssim(source->v_buffer, source->uv_stride, dest->v_buffer,
493 dest->uv_stride, source->uv_crop_width,
494 source->uv_crop_height, in_bd, bd_shift);
495
496 ssimv = (*ssim_y) * .8 + .1 * ((*ssim_u) + (*ssim_v));
497 return convert_ssim_db(ssimv, 1.0);
498 }
499