]> git.imager.perl.org - imager.git/blame_incremental - quant.c
translate_errdiff() could leak memory on failure
[imager.git] / quant.c
... / ...
CommitLineData
1/* quant.c - provides general image quantization
2 currently only used by gif.c, but maybe we'll support producing
3 8-bit (or bigger indexed) png files at some point
4*/
5#include "imager.h"
6#include "imageri.h"
7
8static void makemap_webmap(i_quantize *);
9static void makemap_addi(i_quantize *, i_img **imgs, int count);
10static void makemap_mediancut(i_quantize *, i_img **imgs, int count);
11static void makemap_mono(i_quantize *);
12static void makemap_gray(i_quantize *, int step);
13
14static int makemap_palette(i_quantize *, i_img **imgs, int count);
15
16static
17void
18setcol(i_color *cl,unsigned char r,unsigned char g,unsigned char b,unsigned char a) {
19 cl->rgba.r=r;
20 cl->rgba.g=g;
21 cl->rgba.b=b;
22 cl->rgba.a=a;
23}
24
25
26
27/* make a colour map overwrites mc_existing/mc_count in quant Note
28 that i_makemap will be called once for each image if mc_perimage is
29 set and the format support multiple colour maps per image.
30
31 This means we don't need any special processing at this level to
32 handle multiple colour maps.
33*/
34
35/*
36=item i_quant_makemap(C<quant>, C<imgs>, C<count>)
37
38=category Image quantization
39
40Analyzes the C<count> images in C<imgs> according to the rules in
41C<quant> to build a color map (optimal or not depending on
42C<< quant->make_colors >>).
43
44=cut
45*/
46
47void
48i_quant_makemap(i_quantize *quant, i_img **imgs, int count) {
49
50 if (quant->translate == pt_giflib) {
51 /* giflib does it's own color table generation */
52 /* previously we used giflib's quantizer, but it didn't handle multiple
53 images, which made it hard to build a global color map
54 We've implemented our own median cut code so we can ignore
55 the giflib version */
56 makemap_mediancut(quant, imgs, count);
57 return;
58 }
59
60 switch (quant->make_colors & mc_mask) {
61 case mc_none:
62 /* use user's specified map */
63 break;
64 case mc_web_map:
65 makemap_webmap(quant);
66 break;
67
68 case mc_median_cut:
69 makemap_mediancut(quant, imgs, count);
70 break;
71
72 case mc_mono:
73 makemap_mono(quant);
74 break;
75
76 case mc_gray:
77 makemap_gray(quant, 1);
78 break;
79
80 case mc_gray4:
81 makemap_gray(quant, 85);
82 break;
83
84 case mc_gray16:
85 makemap_gray(quant, 17);
86 break;
87
88 case mc_addi:
89 default:
90 makemap_addi(quant, imgs, count);
91 break;
92 }
93}
94
95static void translate_closest(i_quantize *, i_img *, i_palidx *);
96static int translate_errdiff(i_quantize *, i_img *, i_palidx *);
97static void translate_addi(i_quantize *, i_img *, i_palidx *);
98
99/*
100=item i_quant_translate(C<quant>, C<img>)
101
102=category Image quantization
103
104Quantize the image given the palette in C<quant>.
105
106On success returns a pointer to a memory block of C<< img->xsize *
107img->ysize >> C<i_palidx> entries.
108
109On failure returns NULL.
110
111You should call myfree() on the returned block when you're done with
112it.
113
114This function will fail if the supplied palette contains no colors.
115
116=cut
117*/
118i_palidx *
119i_quant_translate(i_quantize *quant, i_img *img) {
120 i_palidx *result;
121 size_t bytes;
122
123 mm_log((1, "quant_translate(quant %p, img %p)\n", quant, img));
124
125 /* there must be at least one color in the paletted (though even that
126 isn't very useful */
127 if (quant->mc_count == 0) {
128 i_push_error(0, "no colors available for translation");
129 return NULL;
130 }
131
132 bytes = img->xsize * img->ysize;
133 if (bytes / img->ysize != img->xsize) {
134 i_push_error(0, "integer overflow calculating memory allocation");
135 return NULL;
136 }
137 result = mymalloc(bytes);
138
139 switch (quant->translate) {
140 case pt_closest:
141 case pt_giflib:
142 translate_closest(quant, img, result);
143 break;
144
145 case pt_errdiff:
146 if (!translate_errdiff(quant, img, result)) {
147 myfree(result);
148 return NULL;
149 }
150 break;
151
152 case pt_perturb:
153 default:
154 translate_addi(quant, img, result);
155 break;
156 }
157
158 return result;
159}
160
161static void translate_closest(i_quantize *quant, i_img *img, i_palidx *out) {
162 quant->perturb = 0;
163 translate_addi(quant, img, out);
164}
165
166#define PWR2(x) ((x)*(x))
167
168typedef int (*cmpfunc)(const void*, const void*);
169
170typedef struct {
171 unsigned char r,g,b;
172 char fixed;
173 char used;
174 int dr,dg,db;
175 int cdist;
176 int mcount;
177} cvec;
178
179typedef struct {
180 int cnt;
181 int vec[256];
182} hashbox;
183
184typedef struct {
185 int boxnum;
186 int pixcnt;
187 int cand;
188 int pdc;
189} pbox;
190
191static void prescan(i_img **im,int count, int cnum, cvec *clr, i_sample_t *line);
192static void reorder(pbox prescan[512]);
193static int pboxcmp(const pbox *a,const pbox *b);
194static void boxcenter(int box,cvec *cv);
195static float frandn(void);
196static void boxrand(int box,cvec *cv);
197static void bbox(int box,int *r0,int *r1,int *g0,int *g1,int *b0,int *b1);
198static void cr_hashindex(cvec clr[256],int cnum,hashbox hb[512]);
199static int mindist(int boxnum,cvec *cv);
200static int maxdist(int boxnum,cvec *cv);
201
202/* Some of the simpler functions are kept here to aid the compiler -
203 maybe some of them will be inlined. */
204
205static int
206pixbox(i_color *ic) { return ((ic->channel[0] & 224)<<1)+ ((ic->channel[1]&224)>>2) + ((ic->channel[2] &224) >> 5); }
207
208static int
209pixbox_ch(i_sample_t *chans) { return ((chans[0] & 224)<<1)+ ((chans[1]&224)>>2) + ((chans[2] &224) >> 5); }
210
211static unsigned char
212g_sat(int in) {
213 if (in>255) { return 255; }
214 else if (in>0) return in;
215 return 0;
216}
217
218static
219float
220frand(void) {
221 return rand()/(RAND_MAX+1.0);
222}
223
224#ifdef NOTEF
225static
226int
227eucl_d(cvec* cv,i_color *cl) { return PWR2(cv->r-cl->channel[0])+PWR2(cv->g-cl->channel[1])+PWR2(cv->b-cl->channel[2]); }
228#endif
229
230static
231int
232eucl_d_ch(cvec* cv,i_sample_t *chans) {
233 return PWR2(cv->r - chans[0]) + PWR2(cv->g - chans[1])
234 + PWR2(cv->b - chans[2]);
235}
236
237static int
238ceucl_d(i_color *c1, i_color *c2) {
239return PWR2(c1->channel[0]-c2->channel[0])
240 +PWR2(c1->channel[1]-c2->channel[1])
241 +PWR2(c1->channel[2]-c2->channel[2]);
242}
243
244static const int
245gray_samples[] = { 0, 0, 0 };
246
247/*
248
249This quantization algorithm and implementation routines are by Arnar
250M. Hrafnkelson. In case any new ideas are here they are mine since
251this was written from scratch.
252
253The algorithm uses local means in the following way:
254
255 For each point in the colormap we find which image points
256 have that point as it's closest point. We calculate the mean
257 of those points and in the next iteration it will be the new
258 entry in the colormap.
259
260In order to speed this process up (i.e. nearest neighbor problem) We
261divied the r,g,b space up in equally large 512 boxes. The boxes are
262numbered from 0 to 511. Their numbering is so that for a given vector
263it is known that it belongs to the box who is formed by concatenating the
2643 most significant bits from each component of the RGB triplet.
265
266For each box we find the list of points from the colormap who might be
267closest to any given point within the box. The exact solution
268involves finding the Voronoi map (or the dual the Delauny
269triangulation) and has many issues including numerical stability.
270
271So we use this approximation:
272
2731. Find which point has the shortest maximum distance to the box.
2742. Find all points that have a shorter minimum distance than that to the box
275
276This is a very simple task and is not computationally heavy if one
277takes into account that the minimum distances from a pixel to a box is
278always found by checking if it's inside the box or is closest to some
279side or a corner. Finding the maximum distance is also either a side
280or a corner.
281
282This approach results 2-3 times more than the actual points needed but
283is still a good gain over the complete space. Usually when one has a
284256 Colorcolor map a search over 30 is often obtained.
285
286A bit of an enhancement to this approach is to keep a seperate list
287for each side of the cube, but this will require even more memory.
288
289 Arnar M. Hrafnkelsson (addi@umich.edu);
290
291*/
292/*
293 Extracted from gifquant.c, removed dependencies on gif_lib,
294 and added support for multiple images.
295 starting from 1nov2000 by TonyC <tony@develop-help.com>.
296
297*/
298
299static void
300makemap_addi(i_quantize *quant, i_img **imgs, int count) {
301 cvec *clr;
302 int cnum, i, bst_idx=0, ld, cd, iter, currhb, img_num;
303 i_img_dim x, y;
304 i_sample_t *val;
305 float dlt, accerr;
306 hashbox *hb;
307 i_mempool mp;
308 i_img_dim maxwidth = 0;
309 i_sample_t *line;
310 const int *sample_indices;
311
312 mm_log((1, "makemap_addi(quant %p { mc_count=%d, mc_colors=%p }, imgs %p, count %d)\n",
313 quant, quant->mc_count, quant->mc_colors, imgs, count));
314
315 if (makemap_palette(quant, imgs, count))
316 return;
317
318 i_mempool_init(&mp);
319
320 clr = i_mempool_alloc(&mp, sizeof(cvec) * quant->mc_size);
321 hb = i_mempool_alloc(&mp, sizeof(hashbox) * 512);
322 for (i=0; i < quant->mc_count; ++i) {
323 clr[i].r = quant->mc_colors[i].rgb.r;
324 clr[i].g = quant->mc_colors[i].rgb.g;
325 clr[i].b = quant->mc_colors[i].rgb.b;
326 clr[i].fixed = 1;
327 clr[i].mcount = 0;
328 }
329 /* mymalloc doesn't clear memory, so I think we need this */
330 for (; i < quant->mc_size; ++i) {
331 /*clr[i].r = clr[i].g = clr[i].b = 0;*/
332 clr[i].dr = 0;
333 clr[i].dg = 0;
334 clr[i].db = 0;
335 clr[i].fixed = 0;
336 clr[i].mcount = 0;
337 }
338 cnum = quant->mc_size;
339 dlt = 1;
340
341 for (img_num = 0; img_num < count; ++img_num) {
342 if (imgs[img_num]->xsize > maxwidth)
343 maxwidth = imgs[img_num]->xsize;
344 }
345 line = i_mempool_alloc(&mp, 3 * maxwidth * sizeof(*line));
346
347 prescan(imgs, count, cnum, clr, line);
348 cr_hashindex(clr, cnum, hb);
349
350 for(iter=0;iter<3;iter++) {
351 accerr=0.0;
352
353 for (img_num = 0; img_num < count; ++img_num) {
354 i_img *im = imgs[img_num];
355 sample_indices = im->channels >= 3 ? NULL : gray_samples;
356 for(y=0;y<im->ysize;y++) {
357 i_gsamp(im, 0, im->xsize, y, line, sample_indices, 3);
358 val = line;
359 for(x=0;x<im->xsize;x++) {
360 ld=196608;
361 /*i_gpix(im,x,y,&val);*/
362 currhb=pixbox_ch(val);
363 /* printf("box = %d \n",currhb); */
364 for(i=0;i<hb[currhb].cnt;i++) {
365 /* printf("comparing: pix (%d,%d,%d) vec (%d,%d,%d)\n",val.channel[0],val.channel[1],val.channel[2],clr[hb[currhb].vec[i]].r,clr[hb[currhb].vec[i]].g,clr[hb[currhb].vec[i]].b); */
366
367 cd=eucl_d_ch(&clr[hb[currhb].vec[i]],val);
368 if (cd<ld) {
369 ld=cd; /* shortest distance yet */
370 bst_idx=hb[currhb].vec[i]; /* index of closest vector yet */
371 }
372 }
373
374 clr[bst_idx].mcount++;
375 accerr+=(ld);
376 clr[bst_idx].dr+=val[0];
377 clr[bst_idx].dg+=val[1];
378 clr[bst_idx].db+=val[2];
379
380 val += 3; /* next 3 samples (next pixel) */
381 }
382 }
383 }
384
385 for(i=0;i<cnum;i++)
386 if (clr[i].mcount) {
387 clr[i].dr/=clr[i].mcount;
388 clr[i].dg/=clr[i].mcount;
389 clr[i].db/=clr[i].mcount;
390 }
391
392 /* for(i=0;i<cnum;i++) printf("vec(%d)=(%d,%d,%d) dest=(%d,%d,%d) matchcount=%d\n",
393 i,clr[i].r,clr[i].g,clr[i].b,clr[i].dr,clr[i].dg,clr[i].db,clr[i].mcount); */
394
395 /* printf("total error: %.2f\n",sqrt(accerr)); */
396
397 for(i=0;i<cnum;i++) {
398 if (clr[i].fixed) continue; /* skip reserved colors */
399
400 if (clr[i].mcount) {
401 clr[i].used = 1;
402 clr[i].r=clr[i].r*(1-dlt)+dlt*clr[i].dr;
403 clr[i].g=clr[i].g*(1-dlt)+dlt*clr[i].dg;
404 clr[i].b=clr[i].b*(1-dlt)+dlt*clr[i].db;
405 } else {
406 /* let's try something else */
407 clr[i].used = 0;
408 clr[i].r=rand();
409 clr[i].g=rand();
410 clr[i].b=rand();
411 }
412
413 clr[i].dr=0;
414 clr[i].dg=0;
415 clr[i].db=0;
416 clr[i].mcount=0;
417 }
418 cr_hashindex(clr,cnum,hb);
419 }
420
421
422#ifdef NOTEF
423 for(i=0;i<cnum;i++) {
424 cd=eucl_d(&clr[i],&val);
425 if (cd<ld) {
426 ld=cd;
427 bst_idx=i;
428 }
429 }
430#endif
431
432 /* if defined, we only include colours with an mcount or that were
433 supplied in the fixed palette, giving us a smaller output palette */
434#define ONLY_USE_USED
435#ifdef ONLY_USE_USED
436 /* transfer the colors back */
437 quant->mc_count = 0;
438 for (i = 0; i < cnum; ++i) {
439 if (clr[i].fixed || clr[i].used) {
440 /*printf("Adding %d (%d,%d,%d)\n", i, clr[i].r, clr[i].g, clr[i].b);*/
441 quant->mc_colors[quant->mc_count].rgb.r = clr[i].r;
442 quant->mc_colors[quant->mc_count].rgb.g = clr[i].g;
443 quant->mc_colors[quant->mc_count].rgb.b = clr[i].b;
444 ++quant->mc_count;
445 }
446 }
447#else
448 /* transfer the colors back */
449 for (i = 0; i < cnum; ++i) {
450 quant->mc_colors[i].rgb.r = clr[i].r;
451 quant->mc_colors[i].rgb.g = clr[i].g;
452 quant->mc_colors[i].rgb.b = clr[i].b;
453 }
454 quant->mc_count = cnum;
455#endif
456
457#if 0
458 mm_log((1, "makemap_addi returns - quant.mc_count = %d\n", quant->mc_count));
459 for (i = 0; i < quant->mc_count; ++i)
460 mm_log((5, " map entry %d: (%d, %d, %d)\n", i, clr[i].r, clr[i].g, clr[i].b));
461#endif
462
463 i_mempool_destroy(&mp);
464
465 mm_log((1, "makemap_addi() - %d colors\n", quant->mc_count));
466}
467
468typedef struct {
469 i_sample_t rgb[3];
470 i_img_dim count;
471} quant_color_entry;
472
473#define MEDIAN_CUT_COLORS 32768
474
475#define MED_CUT_INDEX(c) ((((c).rgb.r & 0xF8) << 7) | \
476 (((c).rgb.g & 0xF8) << 2) | (((c).rgb.b & 0xF8) >> 3))
477
478#define MED_CUT_GRAY_INDEX(c) ((((c).rgb.r & 0xF8) << 7) | \
479 (((c).rgb.r & 0xF8) << 2) | (((c).rgb.r & 0xF8) >> 3))
480
481/* scale these to cover the whole range */
482#define MED_CUT_RED(index) ((((index) & 0x7C00) >> 10) * 255 / 31)
483#define MED_CUT_GREEN(index) ((((index) & 0x3E0) >> 5) * 255 / 31)
484#define MED_CUT_BLUE(index) (((index) & 0x1F) * 255 / 31)
485
486typedef struct {
487 i_sample_t min[3]; /* minimum for each channel */
488 i_sample_t max[3]; /* maximum for each channel */
489 i_sample_t width[3]; /* width for each channel */
490 int start, size; /* beginning and size of the partition */
491 i_img_dim pixels; /* number of pixels represented by this partition */
492} medcut_partition;
493
494/*
495=item calc_part(part, colors)
496
497Calculates the new color limits for the given partition.
498
499Giflib assumes that the limits for the non-split channels stay the
500same, but this strikes me as incorrect, especially if the colors tend
501to be color ramps.
502
503Of course this could be optimized by not recalculating the channel we
504just sorted on, but it's not worth the effort right now.
505
506=cut
507*/
508static void calc_part(medcut_partition *part, quant_color_entry *colors) {
509 int i, ch;
510
511 for (ch = 0; ch < 3; ++ch) {
512 part->min[ch] = 255;
513 part->max[ch] = 0;
514 }
515 for (i = part->start; i < part->start + part->size; ++i) {
516 for (ch = 0; ch < 3; ++ch) {
517 if (part->min[ch] > colors[i].rgb[ch])
518 part->min[ch] = colors[i].rgb[ch];
519 if (part->max[ch] < colors[i].rgb[ch])
520 part->max[ch] = colors[i].rgb[ch];
521 }
522 }
523 for (ch = 0; ch < 3; ++ch) {
524 part->width[ch] = part->max[ch] - part->min[ch];
525 }
526}
527
528/* simple functions to sort by each channel - we could use a global, but
529 that would be bad */
530
531static int
532color_sort_red(void const *left, void const *right) {
533 return ((quant_color_entry *)left)->rgb[0] - ((quant_color_entry *)right)->rgb[0];
534}
535
536static int
537color_sort_green(void const *left, void const *right) {
538 return ((quant_color_entry *)left)->rgb[1] - ((quant_color_entry *)right)->rgb[1];
539}
540
541static int
542color_sort_blue(void const *left, void const *right) {
543 return ((quant_color_entry *)left)->rgb[2] - ((quant_color_entry *)right)->rgb[2];
544}
545
546static int (*sorters[])(void const *, void const *) =
547{
548 color_sort_red,
549 color_sort_green,
550 color_sort_blue,
551};
552
553static void
554makemap_mediancut(i_quantize *quant, i_img **imgs, int count) {
555 quant_color_entry *colors;
556 i_mempool mp;
557 int imgn, i, ch;
558 i_img_dim x, y, max_width;
559 i_color *line;
560 int color_count;
561 i_img_dim total_pixels;
562 medcut_partition *parts;
563 int part_num;
564 int in, out;
565 /* number of channels we search for the best channel to partition
566 this isn't terribly efficient, but it should work */
567 int chan_count;
568
569 mm_log((1, "makemap_mediancut(quant %p { mc_count=%d, mc_colors=%p }, imgs %p, count %d)\n",
570 quant, quant->mc_count, quant->mc_colors, imgs, count));
571
572 if (makemap_palette(quant, imgs, count))
573 return;
574
575 i_mempool_init(&mp);
576
577 colors = i_mempool_alloc(&mp, sizeof(*colors) * MEDIAN_CUT_COLORS);
578 for (i = 0; i < MEDIAN_CUT_COLORS; ++i) {
579 colors[i].rgb[0] = MED_CUT_RED(i);
580 colors[i].rgb[1] = MED_CUT_GREEN(i);
581 colors[i].rgb[2] = MED_CUT_BLUE(i);
582 colors[i].count = 0;
583 }
584
585 max_width = -1;
586 for (imgn = 0; imgn < count; ++imgn) {
587 if (imgs[imgn]->xsize > max_width)
588 max_width = imgs[imgn]->xsize;
589 }
590 line = i_mempool_alloc(&mp, sizeof(i_color) * max_width);
591
592 /* build the stats */
593 total_pixels = 0;
594 chan_count = 1; /* assume we just have grayscale */
595 for (imgn = 0; imgn < count; ++imgn) {
596 total_pixels += imgs[imgn]->xsize * imgs[imgn]->ysize;
597 for (y = 0; y < imgs[imgn]->ysize; ++y) {
598 i_glin(imgs[imgn], 0, imgs[imgn]->xsize, y, line);
599 if (imgs[imgn]->channels > 2) {
600 chan_count = 3;
601 for (x = 0; x < imgs[imgn]->xsize; ++x) {
602 ++colors[MED_CUT_INDEX(line[x])].count;
603 }
604 }
605 else {
606 /* a gray-scale image, just use the first channel */
607 for (x = 0; x < imgs[imgn]->xsize; ++x) {
608 ++colors[MED_CUT_GRAY_INDEX(line[x])].count;
609 }
610 }
611 }
612 }
613
614 /* eliminate the empty colors */
615 out = 0;
616 for (in = 0; in < MEDIAN_CUT_COLORS; ++in) {
617 if (colors[in].count) {
618 colors[out++] = colors[in];
619 }
620 }
621 /*printf("out %d\n", out);
622
623 for (i = 0; i < out; ++i) {
624 if (colors[i].count) {
625 printf("%d: (%d,%d,%d) -> %d\n", i, colors[i].rgb[0], colors[i].rgb[1],
626 colors[i].rgb[2], colors[i].count);
627 }
628 }*/
629
630 if (out < quant->mc_size) {
631 /* just copy them into the color table */
632 for (i = 0; i < out; ++i) {
633 for (ch = 0; ch < 3; ++ch) {
634 quant->mc_colors[i].channel[ch] = colors[i].rgb[ch];
635 }
636 }
637 quant->mc_count = out;
638 }
639 else {
640 /* build the starting partition */
641 parts = i_mempool_alloc(&mp, sizeof(*parts) * quant->mc_size);
642 parts[0].start = 0;
643 parts[0].size = out;
644 parts[0].pixels = total_pixels;
645 calc_part(parts, colors);
646 color_count = 1;
647
648 while (color_count < quant->mc_size) {
649 /* initialized to avoid compiler warnings */
650 int max_index = 0, max_ch = 0; /* index/channel with biggest spread */
651 int max_size;
652 medcut_partition *workpart;
653 i_img_dim cum_total;
654 i_img_dim half;
655
656 /* find the partition with the most biggest span with more than
657 one color */
658 max_size = -1;
659 for (i = 0; i < color_count; ++i) {
660 for (ch = 0; ch < chan_count; ++ch) {
661 if (parts[i].width[ch] > max_size
662 && parts[i].size > 1) {
663 max_index = i;
664 max_ch = ch;
665 max_size = parts[i].width[ch];
666 }
667 }
668 }
669
670 /* nothing else we can split */
671 if (max_size == -1)
672 break;
673
674 workpart = parts+max_index;
675 /*printf("splitting partition %d (pixels %ld, start %d, size %d)\n", max_index, workpart->pixels, workpart->start, workpart->size);*/
676 qsort(colors + workpart->start, workpart->size, sizeof(*colors),
677 sorters[max_ch]);
678
679 /* find the median or something like it we need to make sure both
680 sides of the split have at least one color in them, so we don't
681 test at the first or last entry */
682 i = workpart->start;
683 cum_total = colors[i].count;
684 ++i;
685 half = workpart->pixels / 2;
686 while (i < workpart->start + workpart->size - 1
687 && cum_total < half) {
688 cum_total += colors[i++].count;
689 }
690 /*printf("Split at %d to make %d (half %ld, cumtotal %ld)\n", i, color_count, half, cum_total);*/
691
692 /* found the spot to split */
693 parts[color_count].start = i;
694 parts[color_count].size = workpart->start + workpart->size - i;
695 workpart->size = i - workpart->start;
696 parts[color_count].pixels = workpart->pixels - cum_total;
697 workpart->pixels = cum_total;
698
699 /* recalculate the limits */
700 calc_part(workpart, colors);
701 calc_part(parts+color_count, colors);
702 ++color_count;
703 }
704
705 /* fill in the color table - since we could still have partitions
706 that have more than one color, we need to average the colors */
707 for (part_num = 0; part_num < color_count; ++part_num) {
708 double sums[3];
709 medcut_partition *workpart;
710
711 workpart = parts+part_num;
712 for (ch = 0; ch < 3; ++ch)
713 sums[ch] = 0;
714
715 for (i = workpart->start; i < workpart->start + workpart->size; ++i) {
716 for (ch = 0; ch < 3; ++ch) {
717 sums[ch] += (int)(colors[i].rgb[ch]) * colors[i].count;
718 }
719 }
720 for (ch = 0; ch < 3; ++ch) {
721 quant->mc_colors[part_num].channel[ch] = sums[ch] / workpart->pixels;
722 }
723 }
724 quant->mc_count = color_count;
725 }
726 /*printf("out %d colors\n", quant->mc_count);*/
727 i_mempool_destroy(&mp);
728
729 mm_log((1, "makemap_mediancut() - %d colors\n", quant->mc_count));
730}
731
732static void
733makemap_mono(i_quantize *quant) {
734 quant->mc_colors[0].rgba.r = 0;
735 quant->mc_colors[0].rgba.g = 0;
736 quant->mc_colors[0].rgba.b = 0;
737 quant->mc_colors[0].rgba.a = 255;
738 quant->mc_colors[1].rgba.r = 255;
739 quant->mc_colors[1].rgba.g = 255;
740 quant->mc_colors[1].rgba.b = 255;
741 quant->mc_colors[1].rgba.a = 255;
742 quant->mc_count = 2;
743}
744
745static void
746makemap_gray(i_quantize *quant, int step) {
747 int gray = 0;
748 int i = 0;
749
750 while (gray < 256) {
751 setcol(quant->mc_colors+i, gray, gray, gray, 255);
752 ++i;
753 gray += step;
754 }
755 quant->mc_count = i;
756}
757
758static void
759makemap_webmap(i_quantize *quant) {
760 int r, g, b;
761
762 int i = 0;
763 for (r = 0; r < 256; r+=0x33)
764 for (g = 0; g < 256; g+=0x33)
765 for (b = 0; b < 256; b += 0x33)
766 setcol(quant->mc_colors+i++, r, g, b, 255);
767 quant->mc_count = i;
768}
769
770static int
771in_palette(i_color *c, i_quantize *quant, int size) {
772 int i;
773
774 for (i = 0; i < size; ++i) {
775 if (c->channel[0] == quant->mc_colors[i].channel[0]
776 && c->channel[1] == quant->mc_colors[i].channel[1]
777 && c->channel[2] == quant->mc_colors[i].channel[2]) {
778 return i;
779 }
780 }
781
782 return -1;
783}
784
785/*
786=item makemap_palette(quant, imgs, count)
787
788Tests if all the given images are paletted and have a common palette,
789if they do it builds that palette.
790
791A possible improvement might be to eliminate unused colors in the
792images palettes.
793
794=cut
795*/
796
797static int
798makemap_palette(i_quantize *quant, i_img **imgs, int count) {
799 int size = quant->mc_count;
800 int i;
801 int imgn;
802 char used[256];
803 int col_count;
804
805 mm_log((1, "makemap_palette(quant %p { mc_count=%d, mc_colors=%p }, imgs %p, count %d)\n",
806 quant, quant->mc_count, quant->mc_colors, imgs, count));
807 /* we try to build a common palette here, if we can manage that, then
808 that's the palette we use */
809 for (imgn = 0; imgn < count; ++imgn) {
810 int eliminate_unused;
811 if (imgs[imgn]->type != i_palette_type) {
812 mm_log((1, "makemap_palette() -> 0 (non-palette image)\n"));
813 return 0;
814 }
815
816 if (!i_tags_get_int(&imgs[imgn]->tags, "gif_eliminate_unused", 0,
817 &eliminate_unused)) {
818 eliminate_unused = 1;
819 }
820
821 if (eliminate_unused) {
822 i_palidx *line = mymalloc(sizeof(i_palidx) * imgs[imgn]->xsize);
823 i_img_dim x, y;
824 memset(used, 0, sizeof(used));
825
826 for (y = 0; y < imgs[imgn]->ysize; ++y) {
827 i_gpal(imgs[imgn], 0, imgs[imgn]->xsize, y, line);
828 for (x = 0; x < imgs[imgn]->xsize; ++x)
829 used[line[x]] = 1;
830 }
831
832 myfree(line);
833 }
834 else {
835 /* assume all are in use */
836 memset(used, 1, sizeof(used));
837 }
838
839 col_count = i_colorcount(imgs[imgn]);
840 for (i = 0; i < col_count; ++i) {
841 i_color c;
842
843 i_getcolors(imgs[imgn], i, &c, 1);
844 if (used[i]) {
845 if (in_palette(&c, quant, size) < 0) {
846 if (size < quant->mc_size) {
847 quant->mc_colors[size++] = c;
848 }
849 else {
850 mm_log((1, "makemap_palette() -> 0 (too many colors)\n"));
851 return 0;
852 }
853 }
854 }
855 }
856 }
857
858 mm_log((1, "makemap_palette() -> 1 (%d total colors)\n", size));
859 quant->mc_count = size;
860
861 return 1;
862}
863
864#define pboxjump 32
865
866/* Define one of the following 4 symbols to choose a colour search method
867 The idea is to try these out, including benchmarking, to see which
868 is fastest in a good spread of circumstances.
869 I'd expect IM_CFLINSEARCH to be fastest for very small palettes, and
870 IM_CFHASHBOX for large images with large palettes.
871
872 Some other possibilities include:
873 - search over entries sorted by luminance
874
875 Initially I was planning on testing using the macros and then
876 integrating the code directly into each function, but this means if
877 we find a bug at a late stage we will need to update N copies of
878 the same code. Also, keeping the code in the macros means that the
879 code in the translation functions is much more to the point,
880 there's no distracting colour search code to remove attention from
881 what makes _this_ translation function different. It may be
882 advisable to move the setup code into functions at some point, but
883 it should be possible to do this fairly transparently.
884
885 If IM_CF_COPTS is defined then CFLAGS must have an appropriate
886 definition.
887
888 Each option needs to define 4 macros:
889 CF_VARS - variables to define in the function
890 CF_SETUP - code to setup for the colour search, eg. allocating and
891 initializing lookup tables
892 CF_FIND - code that looks for the color in val and puts the best
893 matching index in bst_idx
894 CF_CLEANUP - code to clean up, eg. releasing memory
895*/
896#ifndef IM_CF_COPTS
897/*#define IM_CFLINSEARCH*/
898#define IM_CFHASHBOX
899/*#define IM_CFSORTCHAN*/
900/*#define IM_CFRAND2DIST*/
901#endif
902
903/* return true if the color map contains only grays */
904static int
905is_gray_map(const i_quantize *quant) {
906 int i;
907
908 for (i = 0; i < quant->mc_count; ++i) {
909 if (quant->mc_colors[i].rgb.r != quant->mc_colors[i].rgb.g
910 || quant->mc_colors[i].rgb.r != quant->mc_colors[i].rgb.b) {
911 mm_log((1, " not a gray map\n"));
912 return 0;
913 }
914 }
915
916 mm_log((1, " is a gray map\n"));
917 return 1;
918}
919
920#ifdef IM_CFHASHBOX
921
922/* The original version I wrote for this used the sort.
923 If this is defined then we use a sort to extract the indices for
924 the hashbox */
925#define HB_SORT
926
927/* assume i is available */
928#define CF_VARS hashbox *hb = mymalloc(sizeof(hashbox) * 512); \
929 int currhb; \
930 long ld, cd
931
932#ifdef HB_SORT
933
934static long *gdists; /* qsort is annoying */
935/* int might be smaller than long, so we need to do a real compare
936 rather than a subtraction*/
937static int distcomp(void const *a, void const *b) {
938 long ra = gdists[*(int const *)a];
939 long rb = gdists[*(int const *)b];
940 if (ra < rb)
941 return -1;
942 else if (ra > rb)
943 return 1;
944 else
945 return 0;
946}
947
948#endif
949
950/* for each hashbox build a list of colours that are in the hb or is closer
951 than other colours
952 This is pretty involved. The original gifquant generated the hashbox
953 as part of it's normal processing, but since the map generation is now
954 separated from the translation we need to do this on the spot.
955 Any optimizations, even if they don't produce perfect results would be
956 welcome.
957 */
958static void hbsetup(i_quantize *quant, hashbox *hb) {
959 long *dists, mind, maxd;
960 int cr, cb, cg, hbnum, i;
961 i_color cenc;
962#ifdef HB_SORT
963 int *indices = mymalloc(quant->mc_count * sizeof(int));
964#endif
965
966 dists = mymalloc(quant->mc_count * sizeof(long));
967 for (cr = 0; cr < 8; ++cr) {
968 for (cg = 0; cg < 8; ++cg) {
969 for (cb = 0; cb < 8; ++cb) {
970 /* centre of the hashbox */
971 cenc.channel[0] = cr*pboxjump+pboxjump/2;
972 cenc.channel[1] = cg*pboxjump+pboxjump/2;
973 cenc.channel[2] = cb*pboxjump+pboxjump/2;
974 hbnum = pixbox(&cenc);
975 hb[hbnum].cnt = 0;
976 /* order indices in the order of distance from the hashbox */
977 for (i = 0; i < quant->mc_count; ++i) {
978#ifdef HB_SORT
979 indices[i] = i;
980#endif
981 dists[i] = ceucl_d(&cenc, quant->mc_colors+i);
982 }
983#ifdef HB_SORT
984 /* it should be possible to do this without a sort
985 but so far I'm too lazy */
986 gdists = dists;
987 qsort(indices, quant->mc_count, sizeof(int), distcomp);
988 /* any colors that can match are within mind+diagonal size of
989 a hashbox */
990 mind = dists[indices[0]];
991 i = 0;
992 maxd = (sqrt(mind)+pboxjump)*(sqrt(mind)+pboxjump);
993 while (i < quant->mc_count && dists[indices[i]] < maxd) {
994 hb[hbnum].vec[hb[hbnum].cnt++] = indices[i++];
995 }
996#else
997 /* work out the minimum */
998 mind = 256*256*3;
999 for (i = 0; i < quant->mc_count; ++i) {
1000 if (dists[i] < mind) mind = dists[i];
1001 }
1002 /* transfer any colours that might be closest to a colour in
1003 this hashbox */
1004 maxd = (sqrt(mind)+pboxjump)*(sqrt(mind)+pboxjump);
1005 for (i = 0; i < quant->mc_count; ++i) {
1006 if (dists[i] < maxd)
1007 hb[hbnum].vec[hb[hbnum].cnt++] = i;
1008 }
1009#endif
1010 }
1011 }
1012 }
1013#ifdef HB_SORT
1014 myfree(indices);
1015#endif
1016 myfree(dists) ;
1017}
1018#define CF_SETUP hbsetup(quant, hb)
1019
1020#define CF_FIND \
1021 currhb = pixbox(&val); \
1022 ld = 196608; \
1023 for (i = 0; i < hb[currhb].cnt; ++i) { \
1024 cd = ceucl_d(quant->mc_colors+hb[currhb].vec[i], &val); \
1025 if (cd < ld) { ld = cd; bst_idx = hb[currhb].vec[i]; } \
1026 }
1027
1028#define CF_CLEANUP myfree(hb)
1029
1030#endif
1031
1032#ifdef IM_CFLINSEARCH
1033/* as simple as it gets */
1034#define CF_VARS long ld, cd
1035#define CF_SETUP /* none needed */
1036#define CF_FIND \
1037 ld = 196608; \
1038 for (i = 0; i < quant->mc_count; ++i) { \
1039 cd = ceucl_d(quant->mc_colors+i, &val); \
1040 if (cd < ld) { ld = cd; bst_idx = i; } \
1041 }
1042#define CF_CLEANUP
1043#endif
1044
1045#ifdef IM_CFSORTCHAN
1046static int gsortchan;
1047static i_quantize *gquant;
1048static int chansort(void const *a, void const *b) {
1049 return gquant->mc_colors[*(int const *)a].channel[gsortchan] -
1050 gquant->mc_colors[*(int const *)b].channel[gsortchan];
1051}
1052#define CF_VARS int *indices, sortchan, diff; \
1053 long ld, cd; \
1054 int vindex[256] /* where to find value i of chan */
1055
1056static void chansetup(i_img *img, i_quantize *quant, int *csortchan,
1057 int *vindex, int **cindices) {
1058 int *indices, sortchan, chan, i, chval;
1059 int chanmins[MAXCHANNELS], chanmaxs[MAXCHANNELS], maxrange;
1060
1061 /* find the channel with the maximum range */
1062 /* the maximum stddev would probably be better */
1063 for (chan = 0; chan < img->channels; ++chan) {
1064 chanmins[chan] = 256; chanmaxs[chan] = 0;
1065 for (i = 0; i < quant->mc_count; ++i) {
1066 if (quant->mc_colors[i].channel[chan] < chanmins[chan])
1067 chanmins[chan] = quant->mc_colors[i].channel[chan];
1068 if (quant->mc_colors[i].channel[chan] > chanmaxs[chan])
1069 chanmaxs[chan] = quant->mc_colors[i].channel[chan];
1070 }
1071 }
1072 maxrange = -1;
1073 for (chan = 0; chan < img->channels; ++chan) {
1074 if (chanmaxs[chan]-chanmins[chan] > maxrange) {
1075 maxrange = chanmaxs[chan]-chanmins[chan];
1076 sortchan = chan;
1077 }
1078 }
1079 indices = mymalloc(quant->mc_count * sizeof(int)) ;
1080 for (i = 0; i < quant->mc_count; ++i) {
1081 indices[i] = i;
1082 }
1083 gsortchan = sortchan;
1084 gquant = quant;
1085 qsort(indices, quant->mc_count, sizeof(int), chansort) ;
1086 /* now a lookup table to find entries faster */
1087 for (chval=0, i=0; i < quant->mc_count; ++i) {
1088 while (chval < 256 &&
1089 chval < quant->mc_colors[indices[i]].channel[sortchan]) {
1090 vindex[chval++] = i;
1091 }
1092 }
1093 while (chval < 256) {
1094 vindex[chval++] = quant->mc_count-1;
1095 }
1096 *csortchan = sortchan;
1097 *cindices = indices;
1098}
1099
1100#define CF_SETUP \
1101 chansetup(img, quant, &sortchan, vindex, &indices)
1102
1103int chanfind(i_color val, i_quantize *quant, int *indices, int *vindex,
1104 int sortchan) {
1105 int i, bst_idx, diff, maxdiff;
1106 long ld, cd;
1107
1108 i = vindex[val.channel[sortchan]];
1109 bst_idx = indices[i];
1110 ld = 196608;
1111 diff = 0;
1112 maxdiff = quant->mc_count;
1113 while (diff < maxdiff) {
1114 if (i+diff < quant->mc_count) {
1115 cd = ceucl_d(&val, quant->mc_colors+indices[i+diff]);
1116 if (cd < ld) {
1117 bst_idx = indices[i+diff];
1118 ld = cd;
1119 maxdiff = sqrt(ld);
1120 }
1121 }
1122 if (i-diff >= 0) {
1123 cd = ceucl_d(&val, quant->mc_colors+indices[i-diff]);
1124 if (cd < ld) {
1125 bst_idx = indices[i-diff];
1126 ld = cd;
1127 maxdiff = sqrt(ld);
1128 }
1129 }
1130 ++diff;
1131 }
1132
1133 return bst_idx;
1134}
1135
1136#define CF_FIND \
1137 bst_idx = chanfind(val, quant, indices, vindex, sortchan)
1138
1139
1140#define CF_CLEANUP myfree(indices)
1141
1142#endif
1143
1144#ifdef IM_CFRAND2DIST
1145
1146/* This is based on a method described by Addi in the #imager channel
1147 on the 28/2/2001. I was about 1am Sydney time at the time, so I
1148 wasn't at my most cogent. Well, that's my excuse :)
1149
1150<TonyC> what I have at the moment is: hashboxes, with optimum hash box
1151filling; simple linear search; and a lookup in the widest channel
1152(currently the channel with the maximum range)
1153<Addi> There is one more way that might be simple to implement.
1154<Addi> You want to hear?
1155<TonyC> what's that?
1156<purl> somebody said that was not true
1157<Addi> For each of the colors in the palette start by creating a
1158sorted list of the form:
1159<Addi> [distance, color]
1160<Addi> Where they are sorted by distance.
1161<TonyC> distance to where?
1162<Addi> Where the elements in the lists are the distances and colors of
1163the other colors in the palette
1164<TonyC> ok
1165<Addi> So if you are at color 0
1166<Addi> ok - now to search for the closest color when you are creating
1167the final image is done like this:
1168<Addi> a) pick a random color from the palette
1169<Addi> b) calculate the distance to it
1170<Addi> c) only check the vectors that are within double the distance
1171in the list of the color you picked from the palette.
1172<Addi> Does that seem logical?
1173<Addi> Lets imagine that we only have grayscale to make an example:
1174<Addi> Our palette has 1 4 10 20 as colors.
1175<Addi> And we want to quantize the color 11
1176<Addi> lets say we picked 10 randomly
1177<Addi> the double distance is 2
1178<Addi> since abs(10-11)*2 is 2
1179<Addi> And the list at vector 10 is this:
1180<Addi> [0, 10], [6 4], [9, 1], [10, 20]
1181<Addi> so we look at the first one (but not the second one since 6 is
1182at a greater distance than 2.
1183<Addi> Any of that make sense?
1184<TonyC> yes, though are you suggesting another random jump to one of
1185the colours with the possible choices? or an exhaustive search?
1186<Addi> TonyC: It's possible to come up with a recursive/iterative
1187enhancement but this is the 'basic' version.
1188<Addi> Which would do an iterative search.
1189<Addi> You can come up with conditions where it pays to switch to a new one.
1190<Addi> And the 'random' start can be switched over to a small tree.
1191<Addi> So you would have a little index at the start.
1192<Addi> to get you into the general direction
1193<Addi> Perhaps just an 8 split.
1194<Addi> that is - split each dimension in half.
1195<TonyC> yep
1196<TonyC> I get the idea
1197<Addi> But this would seem to be a good approach in our case since we
1198usually have few codevectors.
1199<Addi> So we only need 256*256 entries in a table.
1200<Addi> We could even only index some of them that were deemed as good
1201candidates.
1202<TonyC> I was considering adding paletted output support for PNG and
1203TIFF at some point, which support 16-bit palettes
1204<Addi> ohh.
1205<Addi> 'darn' ;)
1206
1207
1208*/
1209
1210
1211typedef struct i_dists {
1212 int index;
1213 long dist;
1214} i_dists;
1215
1216#define CF_VARS \
1217 i_dists *dists;
1218
1219static int dists_sort(void const *a, void const *b) {
1220 return ((i_dists *)a)->dist - ((i_dists *)b)->dist;
1221}
1222
1223static void rand2dist_setup(i_quantize *quant, i_dists **cdists) {
1224 i_dists *dists =
1225 mymalloc(sizeof(i_dists)*quant->mc_count*quant->mc_count);
1226 int i, j;
1227 long cd;
1228 for (i = 0; i < quant->mc_count; ++i) {
1229 i_dists *ldists = dists + quant->mc_count * i;
1230 i_color val = quant->mc_colors[i];
1231 for (j = 0; j < quant->mc_count; ++j) {
1232 ldists[j].index = j;
1233 ldists[j].dist = ceucl_d(&val, quant->mc_colors+j);
1234 }
1235 qsort(ldists, quant->mc_count, sizeof(i_dists), dists_sort);
1236 }
1237 *cdists = dists;
1238}
1239
1240#define CF_SETUP \
1241 bst_idx = rand() % quant->mc_count; \
1242 rand2dist_setup(quant, &dists)
1243
1244static int rand2dist_find(i_color val, i_quantize *quant, i_dists *dists, int index) {
1245 i_dists *cdists;
1246 long cd, ld;
1247 long maxld;
1248 int i;
1249 int bst_idx;
1250
1251 cdists = dists + index * quant->mc_count;
1252 ld = 3 * 256 * 256;
1253 maxld = 8 * ceucl_d(&val, quant->mc_colors+index);
1254 for (i = 0; i < quant->mc_count && cdists[i].dist <= maxld; ++i) {
1255 cd = ceucl_d(&val, quant->mc_colors+cdists[i].index);
1256 if (cd < ld) {
1257 bst_idx = cdists[i].index;
1258 ld = cd;
1259 }
1260 }
1261 return bst_idx;
1262}
1263
1264#define CF_FIND bst_idx = rand2dist_find(val, quant, dists, bst_idx)
1265
1266#define CF_CLEANUP myfree(dists)
1267
1268
1269#endif
1270
1271static void translate_addi(i_quantize *quant, i_img *img, i_palidx *out) {
1272 i_img_dim x, y, k;
1273 int i, bst_idx = 0;
1274 i_color val;
1275 int pixdev = quant->perturb;
1276 CF_VARS;
1277
1278 CF_SETUP;
1279
1280 if (img->channels >= 3) {
1281 if (pixdev) {
1282 k=0;
1283 for(y=0;y<img->ysize;y++) for(x=0;x<img->xsize;x++) {
1284 i_gpix(img,x,y,&val);
1285 val.channel[0]=g_sat(val.channel[0]+(int)(pixdev*frandn()));
1286 val.channel[1]=g_sat(val.channel[1]+(int)(pixdev*frandn()));
1287 val.channel[2]=g_sat(val.channel[2]+(int)(pixdev*frandn()));
1288 CF_FIND;
1289 out[k++]=bst_idx;
1290 }
1291 } else {
1292 k=0;
1293 for(y=0;y<img->ysize;y++) for(x=0;x<img->xsize;x++) {
1294 i_gpix(img,x,y,&val);
1295 CF_FIND;
1296 out[k++]=bst_idx;
1297 }
1298 }
1299 }
1300 else {
1301 if (pixdev) {
1302 k=0;
1303 for(y=0;y<img->ysize;y++) for(x=0;x<img->xsize;x++) {
1304 i_gpix(img,x,y,&val);
1305 val.channel[1] = val.channel[2] =
1306 val.channel[0]=g_sat(val.channel[0]+(int)(pixdev*frandn()));
1307 CF_FIND;
1308 out[k++]=bst_idx;
1309 }
1310 } else {
1311 k=0;
1312 for(y=0;y<img->ysize;y++) for(x=0;x<img->xsize;x++) {
1313 i_gpix(img,x,y,&val);
1314 val.channel[1] = val.channel[2] = val.channel[0];
1315 CF_FIND;
1316 out[k++]=bst_idx;
1317 }
1318 }
1319 }
1320 CF_CLEANUP;
1321}
1322
1323static int floyd_map[] =
1324{
1325 0, 0, 7,
1326 3, 5, 1
1327};
1328
1329static int jarvis_map[] =
1330{
1331 0, 0, 0, 7, 5,
1332 3, 5, 7, 5, 3,
1333 1, 3, 5, 3, 1
1334};
1335
1336static int stucki_map[] =
1337{
1338 0, 0, 0, 8, 4,
1339 2, 4, 8, 4, 2,
1340 1, 2, 4, 2, 1
1341};
1342
1343struct errdiff_map {
1344 int *map;
1345 int width, height, orig;
1346};
1347
1348static struct errdiff_map maps[] =
1349{
1350 { floyd_map, 3, 2, 1 },
1351 { jarvis_map, 5, 3, 2 },
1352 { stucki_map, 5, 3, 2 },
1353};
1354
1355typedef struct errdiff_tag {
1356 int r, g, b;
1357} errdiff_t;
1358
1359/* perform an error diffusion dither */
1360static int
1361translate_errdiff(i_quantize *quant, i_img *img, i_palidx *out) {
1362 int *map;
1363 int mapw, maph, mapo;
1364 int i;
1365 errdiff_t *err;
1366 i_img_dim errw;
1367 int difftotal;
1368 i_img_dim x, y, dx, dy;
1369 int bst_idx = 0;
1370 int is_gray = is_gray_map(quant);
1371 CF_VARS;
1372
1373 if ((quant->errdiff & ed_mask) == ed_custom) {
1374 map = quant->ed_map;
1375 mapw = quant->ed_width;
1376 maph = quant->ed_height;
1377 mapo = quant->ed_orig;
1378 }
1379 else {
1380 int index = quant->errdiff & ed_mask;
1381 if (index >= ed_custom) index = ed_floyd;
1382 map = maps[index].map;
1383 mapw = maps[index].width;
1384 maph = maps[index].height;
1385 mapo = maps[index].orig;
1386 }
1387
1388 difftotal = 0;
1389 for (i = 0; i < maph * mapw; ++i) {
1390 if (map[i] < 0) {
1391 i_push_errorf(0, "errdiff_map values must be non-negative, errdiff[%d] is negative", i);
1392 goto fail;
1393 }
1394 difftotal += map[i];
1395 }
1396
1397 if (!difftotal) {
1398 i_push_error(0, "error diffusion map must contain some non-zero values");
1399 goto fail;
1400 }
1401
1402 errw = img->xsize+mapw;
1403 err = mymalloc(sizeof(*err) * maph * errw);
1404 /*errp = err+mapo;*/
1405 memset(err, 0, sizeof(*err) * maph * errw);
1406
1407 /*printf("map:\n");
1408 for (dy = 0; dy < maph; ++dy) {
1409 for (dx = 0; dx < mapw; ++dx) {
1410 printf("%2d", map[dx+dy*mapw]);
1411 }
1412 putchar('\n');
1413 }*/
1414
1415 CF_SETUP;
1416
1417 for (y = 0; y < img->ysize; ++y) {
1418 for (x = 0; x < img->xsize; ++x) {
1419 i_color val;
1420 errdiff_t perr;
1421 i_gpix(img, x, y, &val);
1422 if (img->channels < 3) {
1423 val.channel[1] = val.channel[2] = val.channel[0];
1424 }
1425 else if (is_gray) {
1426 int gray = 0.5 + color_to_grey(&val);
1427 val.channel[0] = val.channel[1] = val.channel[2] = gray;
1428 }
1429 perr = err[x+mapo];
1430 perr.r = perr.r < 0 ? -((-perr.r)/difftotal) : perr.r/difftotal;
1431 perr.g = perr.g < 0 ? -((-perr.g)/difftotal) : perr.g/difftotal;
1432 perr.b = perr.b < 0 ? -((-perr.b)/difftotal) : perr.b/difftotal;
1433 /*printf("x %3d y %3d in(%3d, %3d, %3d) di(%4d,%4d,%4d)\n", x, y, val.channel[0], val.channel[1], val.channel[2], perr.r, perr.g, perr.b);*/
1434 val.channel[0] = g_sat(val.channel[0]-perr.r);
1435 val.channel[1] = g_sat(val.channel[1]-perr.g);
1436 val.channel[2] = g_sat(val.channel[2]-perr.b);
1437 CF_FIND;
1438 /* save error */
1439 perr.r = quant->mc_colors[bst_idx].channel[0] - val.channel[0];
1440 perr.g = quant->mc_colors[bst_idx].channel[1] - val.channel[1];
1441 perr.b = quant->mc_colors[bst_idx].channel[2] - val.channel[2];
1442 /*printf(" out(%3d, %3d, %3d) er(%4d, %4d, %4d)\n", quant->mc_colors[bst_idx].channel[0], quant->mc_colors[bst_idx].channel[1], quant->mc_colors[bst_idx].channel[2], perr.r, perr.g, perr.b);*/
1443 for (dx = 0; dx < mapw; ++dx) {
1444 for (dy = 0; dy < maph; ++dy) {
1445 err[x+dx+dy*errw].r += perr.r * map[dx+mapw*dy];
1446 err[x+dx+dy*errw].g += perr.g * map[dx+mapw*dy];
1447 err[x+dx+dy*errw].b += perr.b * map[dx+mapw*dy];
1448 }
1449 }
1450 *out++ = bst_idx;
1451 }
1452 /* shift up the error matrix */
1453 for (dy = 0; dy < maph-1; ++dy) {
1454 memcpy(err+dy*errw, err+(dy+1)*errw, sizeof(*err)*errw);
1455 }
1456 memset(err+(maph-1)*errw, 0, sizeof(*err)*errw);
1457 }
1458 CF_CLEANUP;
1459 myfree(err);
1460
1461 return 1;
1462
1463 fail:
1464 CF_CLEANUP;
1465
1466 return 0;
1467}
1468/* Prescan finds the boxes in the image that have the highest number of colors
1469 and that result is used as the initial value for the vectores */
1470
1471
1472static void prescan(i_img **imgs,int count, int cnum, cvec *clr, i_sample_t *line) {
1473 int i,k,j;
1474 i_img_dim x,y;
1475 i_sample_t *val;
1476 const int *chans;
1477
1478 pbox prebox[512];
1479 for(i=0;i<512;i++) {
1480 prebox[i].boxnum=i;
1481 prebox[i].pixcnt=0;
1482 prebox[i].cand=1;
1483 }
1484
1485 /* process each image */
1486 for (i = 0; i < count; ++i) {
1487 i_img *im = imgs[i];
1488 chans = im->channels >= 3 ? NULL : gray_samples;
1489 for(y=0;y<im->ysize;y++) {
1490 i_gsamp(im, 0, im->xsize, y, line, chans, 3);
1491 val = line;
1492 for(x=0;x<im->xsize;x++) {
1493 prebox[pixbox_ch(val)].pixcnt++;
1494 }
1495 }
1496 }
1497
1498 for(i=0;i<512;i++) prebox[i].pdc=prebox[i].pixcnt;
1499 qsort(prebox,512,sizeof(pbox),(cmpfunc)pboxcmp);
1500
1501 for(i=0;i<cnum;i++) {
1502 /* printf("Color %d\n",i);
1503 for(k=0;k<10;k++) { printf("box=%03d %04d %d %04d \n",prebox[k].boxnum,prebox[k].pixcnt,prebox[k].cand,prebox[k].pdc); }
1504 printf("\n\n"); */
1505 reorder(prebox);
1506 }
1507
1508 /* for(k=0;k<cnum;k++) { printf("box=%03d %04d %d %04d \n",prebox[k].boxnum,prebox[k].pixcnt,prebox[k].cand,prebox[k].pdc); } */
1509
1510 k=0;
1511 j=1;
1512 i=0;
1513 while(i<cnum) {
1514 /* printf("prebox[%d].cand=%d\n",k,prebox[k].cand); */
1515 if (clr[i].fixed) { i++; continue; } /* reserved go to next */
1516 if (j>=prebox[k].cand) { k++; j=1; } else {
1517 if (prebox[k].cand == 2) boxcenter(prebox[k].boxnum,&(clr[i]));
1518 else boxrand(prebox[k].boxnum,&(clr[i]));
1519 /* printf("(%d,%d) %d %d -> (%d,%d,%d)\n",k,j,prebox[k].boxnum,prebox[k].pixcnt,clr[i].r,clr[i].g,clr[i].b); */
1520 j++;
1521 i++;
1522 }
1523 }
1524}
1525
1526
1527static void reorder(pbox prescan[512]) {
1528 int nidx;
1529 pbox c;
1530
1531 nidx=0;
1532 c=prescan[0];
1533
1534 c.cand++;
1535 c.pdc=c.pixcnt/(c.cand*c.cand);
1536 /* c.pdc=c.pixcnt/c.cand; */
1537 while(nidx < 511 && c.pdc < prescan[nidx+1].pdc) {
1538 prescan[nidx]=prescan[nidx+1];
1539 nidx++;
1540 }
1541 prescan[nidx]=c;
1542}
1543
1544static int
1545pboxcmp(const pbox *a,const pbox *b) {
1546 if (a->pixcnt > b->pixcnt) return -1;
1547 if (a->pixcnt < b->pixcnt) return 1;
1548 return 0;
1549}
1550
1551static void
1552boxcenter(int box,cvec *cv) {
1553 cv->r=15+((box&448)>>1);
1554 cv->g=15+((box&56)<<2);
1555 cv->b=15+((box&7)<<5);
1556}
1557
1558static void
1559bbox(int box,int *r0,int *r1,int *g0,int *g1,int *b0,int *b1) {
1560 *r0=(box&448)>>1;
1561 *r1=(*r0)|31;
1562 *g0=(box&56)<<2;
1563 *g1=(*g0)|31;
1564 *b0=(box&7)<<5;
1565 *b1=(*b0)|31;
1566}
1567
1568static void
1569boxrand(int box,cvec *cv) {
1570 cv->r=6+(rand()%25)+((box&448)>>1);
1571 cv->g=6+(rand()%25)+((box&56)<<2);
1572 cv->b=6+(rand()%25)+((box&7)<<5);
1573}
1574
1575static float
1576frandn(void) {
1577
1578 float u1,u2,w;
1579
1580 w=1;
1581
1582 while (w >= 1 || w == 0) {
1583 u1 = 2 * frand() - 1;
1584 u2 = 2 * frand() - 1;
1585 w = u1*u1 + u2*u2;
1586 }
1587
1588 w = sqrt((-2*log(w))/w);
1589 return u1*w;
1590}
1591
1592/* Create hash index */
1593static
1594void
1595cr_hashindex(cvec clr[256],int cnum,hashbox hb[512]) {
1596
1597 int bx,mind,cd,cumcnt,i;
1598/* printf("indexing... \n");*/
1599
1600 cumcnt=0;
1601 for(bx=0; bx<512; bx++) {
1602 mind=196608;
1603 for(i=0; i<cnum; i++) {
1604 cd = maxdist(bx,&clr[i]);
1605 if (cd < mind) { mind=cd; }
1606 }
1607
1608 hb[bx].cnt=0;
1609 for(i=0;i<cnum;i++) if (mindist(bx,&clr[i])<mind) hb[bx].vec[hb[bx].cnt++]=i;
1610 /*printf("box %d -> approx -> %d\n",bx,hb[bx].cnt); */
1611 /* statbox(bx,cnum,clr); */
1612 cumcnt+=hb[bx].cnt;
1613 }
1614
1615/* printf("Average search space: %d\n",cumcnt/512); */
1616}
1617
1618static int
1619maxdist(int boxnum,cvec *cv) {
1620 int r0,r1,g0,g1,b0,b1;
1621 int r,g,b,mr,mg,mb;
1622
1623 r=cv->r;
1624 g=cv->g;
1625 b=cv->b;
1626
1627 bbox(boxnum,&r0,&r1,&g0,&g1,&b0,&b1);
1628
1629 mr=i_max(abs(b-b0),abs(b-b1));
1630 mg=i_max(abs(g-g0),abs(g-g1));
1631 mb=i_max(abs(r-r0),abs(r-r1));
1632
1633 return PWR2(mr)+PWR2(mg)+PWR2(mb);
1634}
1635
1636static int
1637mindist(int boxnum,cvec *cv) {
1638 int r0,r1,g0,g1,b0,b1;
1639 int r,g,b,mr,mg,mb;
1640
1641 r=cv->r;
1642 g=cv->g;
1643 b=cv->b;
1644
1645 bbox(boxnum,&r0,&r1,&g0,&g1,&b0,&b1);
1646
1647 /* printf("box %d, (%d,%d,%d)-(%d,%d,%d) vec (%d,%d,%d) ",boxnum,r0,g0,b0,r1,g1,b1,r,g,b); */
1648
1649 if (r0<=r && r<=r1 && g0<=g && g<=g1 && b0<=b && b<=b1) return 0;
1650
1651 mr=i_min(abs(b-b0),abs(b-b1));
1652 mg=i_min(abs(g-g0),abs(g-g1));
1653 mb=i_min(abs(r-r0),abs(r-r1));
1654
1655 mr=PWR2(mr);
1656 mg=PWR2(mg);
1657 mb=PWR2(mb);
1658
1659 if (r0<=r && r<=r1 && g0<=g && g<=g1) return mb;
1660 if (r0<=r && r<=r1 && b0<=b && b<=b1) return mg;
1661 if (b0<=b && b<=b1 && g0<=g && g<=g1) return mr;
1662
1663 if (r0<=r && r<=r1) return mg+mb;
1664 if (g0<=g && g<=g1) return mr+mb;
1665 if (b0<=b && b<=b1) return mg+mr;
1666
1667 return mr+mg+mb;
1668}
1669
1670static void transparent_threshold(i_quantize *, i_palidx *, i_img *, i_palidx);
1671static void transparent_errdiff(i_quantize *, i_palidx *, i_img *, i_palidx);
1672static void transparent_ordered(i_quantize *, i_palidx *, i_img *, i_palidx);
1673
1674/*
1675=item i_quant_transparent(C<quant>, C<data>, C<img>, C<trans_index>)
1676
1677=category Image quantization
1678
1679Dither the alpha channel on C<img> into the palette indexes in
1680C<data>. Pixels to be transparent are replaced with C<trans_pixel>.
1681
1682The method used depends on the tr_* members of C<quant>.
1683
1684=cut
1685*/
1686
1687void
1688i_quant_transparent(i_quantize *quant, i_palidx *data, i_img *img,
1689 i_palidx trans_index)
1690{
1691 switch (quant->transp) {
1692 case tr_none:
1693 break;
1694
1695 default:
1696 quant->tr_threshold = 128;
1697 /* fall through */
1698 case tr_threshold:
1699 transparent_threshold(quant, data, img, trans_index);
1700 break;
1701
1702 case tr_errdiff:
1703 transparent_errdiff(quant, data, img, trans_index);
1704 break;
1705
1706 case tr_ordered:
1707 transparent_ordered(quant, data, img, trans_index);
1708 break;
1709 }
1710}
1711
1712static void
1713transparent_threshold(i_quantize *quant, i_palidx *data, i_img *img,
1714 i_palidx trans_index)
1715{
1716 i_img_dim x, y;
1717 i_sample_t *line = mymalloc(img->xsize * sizeof(i_sample_t));
1718 int trans_chan = img->channels > 2 ? 3 : 1;
1719
1720 for (y = 0; y < img->ysize; ++y) {
1721 i_gsamp(img, 0, img->xsize, y, line, &trans_chan, 1);
1722 for (x = 0; x < img->xsize; ++x) {
1723 if (line[x] < quant->tr_threshold)
1724 data[y*img->xsize+x] = trans_index;
1725 }
1726 }
1727 myfree(line);
1728}
1729
1730static void
1731transparent_errdiff(i_quantize *quant, i_palidx *data, i_img *img,
1732 i_palidx trans_index)
1733{
1734 int *map;
1735 int index;
1736 int mapw, maph, mapo;
1737 int errw, *err, *errp;
1738 int difftotal, out, error;
1739 i_img_dim x, y, dx, dy;
1740 int i;
1741 i_sample_t *line;
1742 int trans_chan = img->channels > 2 ? 3 : 1;
1743
1744 /* no custom map for transparency (yet) */
1745 index = quant->tr_errdiff & ed_mask;
1746 if (index >= ed_custom) index = ed_floyd;
1747 map = maps[index].map;
1748 mapw = maps[index].width;
1749 maph = maps[index].height;
1750 mapo = maps[index].orig;
1751
1752 errw = img->xsize+mapw-1;
1753 err = mymalloc(sizeof(*err) * maph * errw);
1754 errp = err+mapo;
1755 memset(err, 0, sizeof(*err) * maph * errw);
1756
1757 line = mymalloc(img->xsize * sizeof(i_sample_t));
1758 difftotal = 0;
1759 for (i = 0; i < maph * mapw; ++i)
1760 difftotal += map[i];
1761 for (y = 0; y < img->ysize; ++y) {
1762 i_gsamp(img, 0, img->xsize, y, line, &trans_chan, 1);
1763 for (x = 0; x < img->xsize; ++x) {
1764 line[x] = g_sat(line[x]-errp[x]/difftotal);
1765 if (line[x] < 128) {
1766 out = 0;
1767 data[y*img->xsize+x] = trans_index;
1768 }
1769 else {
1770 out = 255;
1771 }
1772 error = out - line[x];
1773 for (dx = 0; dx < mapw; ++dx) {
1774 for (dy = 0; dy < maph; ++dy) {
1775 errp[x+dx-mapo+dy*errw] += error * map[dx+mapw*dy];
1776 }
1777 }
1778 }
1779 /* shift up the error matrix */
1780 for (dy = 0; dy < maph-1; ++dy)
1781 memcpy(err+dy*errw, err+(dy+1)*errw, sizeof(*err)*errw);
1782 memset(err+(maph-1)*errw, 0, sizeof(*err)*errw);
1783 }
1784 myfree(err);
1785 myfree(line);
1786}
1787
1788/* builtin ordered dither maps */
1789static unsigned char
1790orddith_maps[][64] =
1791{
1792 { /* random
1793 this is purely random - it's pretty awful
1794 */
1795 48, 72, 196, 252, 180, 92, 108, 52,
1796 228, 176, 64, 8, 236, 40, 20, 164,
1797 120, 128, 84, 116, 24, 28, 172, 220,
1798 68, 0, 188, 124, 184, 224, 192, 104,
1799 132, 100, 240, 200, 152, 160, 244, 44,
1800 96, 204, 144, 16, 140, 56, 232, 216,
1801 208, 4, 76, 212, 136, 248, 80, 168,
1802 156, 88, 32, 112, 148, 12, 36, 60,
1803 },
1804 {
1805 /* dot8
1806 perl spot.perl '($x-3.5)*($x-3.5)+($y-3.5)*($y-3.5)'
1807 */
1808 240, 232, 200, 136, 140, 192, 228, 248,
1809 220, 148, 100, 76, 80, 104, 152, 212,
1810 180, 116, 56, 32, 36, 60, 120, 176,
1811 156, 64, 28, 0, 8, 44, 88, 160,
1812 128, 92, 24, 12, 4, 40, 68, 132,
1813 184, 96, 48, 20, 16, 52, 108, 188,
1814 216, 144, 112, 72, 84, 124, 164, 224,
1815 244, 236, 196, 168, 172, 204, 208, 252,
1816 },
1817 { /* dot4
1818 perl spot.perl \
1819 'min(dist(1.5, 1.5),dist(5.5,1.5),dist(1.5,5.5),dist(5.5,5.5))'
1820 */
1821 196, 72, 104, 220, 200, 80, 112, 224,
1822 76, 4, 24, 136, 84, 8, 32, 144,
1823 108, 28, 52, 168, 116, 36, 56, 176,
1824 216, 140, 172, 244, 228, 148, 180, 248,
1825 204, 92, 124, 236, 192, 68, 96, 208,
1826 88, 12, 44, 156, 64, 0, 16, 128,
1827 120, 40, 60, 188, 100, 20, 48, 160,
1828 232, 152, 184, 252, 212, 132, 164, 240,
1829 },
1830 { /* hline
1831 perl spot.perl '$y-3'
1832 */
1833 160, 164, 168, 172, 176, 180, 184, 188,
1834 128, 132, 136, 140, 144, 148, 152, 156,
1835 32, 36, 40, 44, 48, 52, 56, 60,
1836 0, 4, 8, 12, 16, 20, 24, 28,
1837 64, 68, 72, 76, 80, 84, 88, 92,
1838 96, 100, 104, 108, 112, 116, 120, 124,
1839 192, 196, 200, 204, 208, 212, 216, 220,
1840 224, 228, 232, 236, 240, 244, 248, 252,
1841 },
1842 { /* vline
1843 perl spot.perl '$x-3'
1844 */
1845 180, 100, 40, 12, 44, 104, 184, 232,
1846 204, 148, 60, 16, 64, 128, 208, 224,
1847 212, 144, 76, 8, 80, 132, 216, 244,
1848 160, 112, 68, 20, 84, 108, 172, 236,
1849 176, 96, 72, 28, 88, 152, 188, 228,
1850 200, 124, 92, 0, 32, 116, 164, 240,
1851 168, 120, 36, 24, 48, 136, 192, 248,
1852 196, 140, 52, 4, 56, 156, 220, 252,
1853 },
1854 { /* slashline
1855 perl spot.perl '$y+$x-7'
1856 */
1857 248, 232, 224, 192, 140, 92, 52, 28,
1858 240, 220, 196, 144, 108, 60, 12, 64,
1859 216, 180, 148, 116, 76, 20, 80, 128,
1860 204, 152, 104, 44, 16, 72, 100, 160,
1861 164, 96, 68, 24, 56, 112, 168, 176,
1862 124, 40, 8, 36, 88, 136, 184, 212,
1863 84, 4, 32, 120, 156, 188, 228, 236,
1864 0, 48, 132, 172, 200, 208, 244, 252,
1865 },
1866 { /* backline
1867 perl spot.perl '$y-$x'
1868 */
1869 0, 32, 116, 172, 184, 216, 236, 252,
1870 56, 8, 72, 132, 136, 200, 228, 240,
1871 100, 36, 12, 40, 92, 144, 204, 220,
1872 168, 120, 60, 16, 44, 96, 156, 176,
1873 180, 164, 112, 48, 28, 52, 128, 148,
1874 208, 192, 152, 88, 84, 20, 64, 104,
1875 232, 224, 196, 140, 108, 68, 24, 76,
1876 248, 244, 212, 188, 160, 124, 80, 4,
1877 },
1878 {
1879 /* tiny
1880 good for display, bad for print
1881 hand generated
1882 */
1883 0, 128, 32, 192, 8, 136, 40, 200,
1884 224, 64, 160, 112, 232, 72, 168, 120,
1885 48, 144, 16, 208, 56, 152, 24, 216,
1886 176, 96, 240, 80, 184, 104, 248, 88,
1887 12, 140, 44, 204, 4, 132, 36, 196,
1888 236, 76, 172, 124, 228, 68, 164, 116,
1889 60, 156, 28, 220, 52, 148, 20, 212,
1890 188, 108, 252, 92, 180, 100, 244, 84,
1891 },
1892};
1893
1894static void
1895transparent_ordered(i_quantize *quant, i_palidx *data, i_img *img,
1896 i_palidx trans_index)
1897{
1898 unsigned char *spot;
1899 i_img_dim x, y;
1900 i_sample_t *line;
1901 int trans_chan = img->channels > 2 ? 3 : 1;
1902 if (quant->tr_orddith == od_custom)
1903 spot = quant->tr_custom;
1904 else
1905 spot = orddith_maps[quant->tr_orddith];
1906
1907 line = mymalloc(img->xsize * sizeof(i_sample_t));
1908 for (y = 0; y < img->ysize; ++y) {
1909 i_gsamp(img, 0, img->xsize, y, line, &trans_chan, 1);
1910 for (x = 0; x < img->xsize; ++x) {
1911 if (line[x] < spot[(x&7)+(y&7)*8])
1912 data[x+y*img->xsize] = trans_index;
1913 }
1914 }
1915 myfree(line);
1916}
1917