1 /*
2 * jquant2.c
3 *
4 * Copyright (C) 1991-1996, Thomas G. Lane.
5 * Modified 2011 by Guido Vollbeding.
6 * This file is part of the Independent JPEG Group's software.
7 * For conditions of distribution and use, see the accompanying README file.
8 *
9 * This file contains 2-pass color quantization (color mapping) routines.
10 * These routines provide selection of a custom color map for an image,
11 * followed by mapping of the image to that color map, with optional
12 * Floyd-Steinberg dithering.
13 * It is also possible to use just the second pass to map to an arbitrary
14 * externally-given color map.
15 *
16 * Note: ordered dithering is not supported, since there isn't any fast
17 * way to compute intercolor distances; it's unclear that ordered dither's
18 * fundamental assumptions even hold with an irregularly spaced color map.
19 */
20
21 #define JPEG_INTERNALS
22 #include "jinclude.h"
23 #include "jpeglib.h"
24
25 #ifdef QUANT_2PASS_SUPPORTED
26
27
28 /*
29 * This module implements the well-known Heckbert paradigm for color
30 * quantization. Most of the ideas used here can be traced back to
31 * Heckbert's seminal paper
32 * Heckbert, Paul. "Color Image Quantization for Frame Buffer Display",
33 * Proc. SIGGRAPH '82, Computer Graphics v.16 #3 (July 1982), pp 297-304.
34 *
35 * In the first pass over the image, we accumulate a histogram showing the
36 * usage count of each possible color. To keep the histogram to a reasonable
37 * size, we reduce the precision of the input; typical practice is to retain
38 * 5 or 6 bits per color, so that 8 or 4 different input values are counted
39 * in the same histogram cell.
40 *
41 * Next, the color-selection step begins with a box representing the whole
42 * color space, and repeatedly splits the "largest" remaining box until we
43 * have as many boxes as desired colors. Then the mean color in each
44 * remaining box becomes one of the possible output colors.
45 *
46 * The second pass over the image maps each input pixel to the closest output
47 * color (optionally after applying a Floyd-Steinberg dithering correction).
48 * This mapping is logically trivial, but making it go fast enough requires
49 * considerable care.
50 *
51 * Heckbert-style quantizers vary a good deal in their policies for choosing
52 * the "largest" box and deciding where to cut it. The particular policies
53 * used here have proved out well in experimental comparisons, but better ones
54 * may yet be found.
55 *
56 * In earlier versions of the IJG code, this module quantized in YCbCr color
57 * space, processing the raw upsampled data without a color conversion step.
58 * This allowed the color conversion math to be done only once per colormap
59 * entry, not once per pixel. However, that optimization precluded other
60 * useful optimizations (such as merging color conversion with upsampling)
61 * and it also interfered with desired capabilities such as quantizing to an
62 * externally-supplied colormap. We have therefore abandoned that approach.
63 * The present code works in the post-conversion color space, typically RGB.
64 *
65 * To improve the visual quality of the results, we actually work in scaled
66 * RGB space, giving G distances more weight than R, and R in turn more than
67 * B. To do everything in integer math, we must use integer scale factors.
68 * The 2/3/1 scale factors used here correspond loosely to the relative
69 * weights of the colors in the NTSC grayscale equation.
70 * If you want to use this code to quantize a non-RGB color space, you'll
71 * probably need to change these scale factors.
72 */
73
74 #define R_SCALE 2 /* scale R distances by this much */
75 #define G_SCALE 3 /* scale G distances by this much */
76 #define B_SCALE 1 /* and B by this much */
77
78 /* Relabel R/G/B as components 0/1/2, respecting the RGB ordering defined
79 * in jmorecfg.h. As the code stands, it will do the right thing for R,G,B
80 * and B,G,R orders. If you define some other weird order in jmorecfg.h,
81 * you'll get compile errors until you extend this logic. In that case
82 * you'll probably want to tweak the histogram sizes too.
83 */
84
85 #if RGB_RED == 0
86 #define C0_SCALE R_SCALE
87 #endif
88 #if RGB_BLUE == 0
89 #define C0_SCALE B_SCALE
90 #endif
91 #if RGB_GREEN == 1
92 #define C1_SCALE G_SCALE
93 #endif
94 #if RGB_RED == 2
95 #define C2_SCALE R_SCALE
96 #endif
97 #if RGB_BLUE == 2
98 #define C2_SCALE B_SCALE
99 #endif
100
101
102 /*
103 * First we have the histogram data structure and routines for creating it.
104 *
105 * The number of bits of precision can be adjusted by changing these symbols.
106 * We recommend keeping 6 bits for G and 5 each for R and B.
107 * If you have plenty of memory and cycles, 6 bits all around gives marginally
108 * better results; if you are short of memory, 5 bits all around will save
109 * some space but degrade the results.
110 * To maintain a fully accurate histogram, we'd need to allocate a "long"
111 * (preferably unsigned long) for each cell. In practice this is overkill;
112 * we can get by with 16 bits per cell. Few of the cell counts will overflow,
113 * and clamping those that do overflow to the maximum value will give close-
114 * enough results. This reduces the recommended histogram size from 256Kb
115 * to 128Kb, which is a useful savings on PC-class machines.
116 * (In the second pass the histogram space is re-used for pixel mapping data;
117 * in that capacity, each cell must be able to store zero to the number of
118 * desired colors. 16 bits/cell is plenty for that too.)
119 * Since the JPEG code is intended to run in small memory model on 80x86
120 * machines, we can't just allocate the histogram in one chunk. Instead
121 * of a true 3-D array, we use a row of pointers to 2-D arrays. Each
122 * pointer corresponds to a C0 value (typically 2^5 = 32 pointers) and
123 * each 2-D array has 2^6*2^5 = 2048 or 2^6*2^6 = 4096 entries. Note that
124 * on 80x86 machines, the pointer row is in near memory but the actual
125 * arrays are in far memory (same arrangement as we use for image arrays).
126 */
127
128 #define MAXNUMCOLORS (MAXJSAMPLE+1) /* maximum size of colormap */
129
130 /* These will do the right thing for either R,G,B or B,G,R color order,
131 * but you may not like the results for other color orders.
132 */
133 #define HIST_C0_BITS 5 /* bits of precision in R/B histogram */
134 #define HIST_C1_BITS 6 /* bits of precision in G histogram */
135 #define HIST_C2_BITS 5 /* bits of precision in B/R histogram */
136
137 /* Number of elements along histogram axes. */
138 #define HIST_C0_ELEMS (1<<HIST_C0_BITS)
139 #define HIST_C1_ELEMS (1<<HIST_C1_BITS)
140 #define HIST_C2_ELEMS (1<<HIST_C2_BITS)
141
142 /* These are the amounts to shift an input value to get a histogram index. */
143 #define C0_SHIFT (BITS_IN_JSAMPLE-HIST_C0_BITS)
144 #define C1_SHIFT (BITS_IN_JSAMPLE-HIST_C1_BITS)
145 #define C2_SHIFT (BITS_IN_JSAMPLE-HIST_C2_BITS)
146
147
148 typedef UINT16 histcell; /* histogram cell; prefer an unsigned type */
149
150 typedef histcell FAR * histptr; /* for pointers to histogram cells */
151
152 typedef histcell hist1d[HIST_C2_ELEMS]; /* typedefs for the array */
153 typedef hist1d FAR * hist2d; /* type for the 2nd-level pointers */
154 typedef hist2d * hist3d; /* type for top-level pointer */
155
156
157 /* Declarations for Floyd-Steinberg dithering.
158 *
159 * Errors are accumulated into the array fserrors[], at a resolution of
160 * 1/16th of a pixel count. The error at a given pixel is propagated
161 * to its not-yet-processed neighbors using the standard F-S fractions,
162 * ... (here) 7/16
163 * 3/16 5/16 1/16
164 * We work left-to-right on even rows, right-to-left on odd rows.
165 *
166 * We can get away with a single array (holding one row's worth of errors)
167 * by using it to store the current row's errors at pixel columns not yet
168 * processed, but the next row's errors at columns already processed. We
169 * need only a few extra variables to hold the errors immediately around the
170 * current column. (If we are lucky, those variables are in registers, but
171 * even if not, they're probably cheaper to access than array elements are.)
172 *
173 * The fserrors[] array has (#columns + 2) entries; the extra entry at
174 * each end saves us from special-casing the first and last pixels.
175 * Each entry is three values long, one value for each color component.
176 *
177 * Note: on a wide image, we might not have enough room in a PC's near data
178 * segment to hold the error array; so it is allocated with alloc_large.
179 */
180
181 #if BITS_IN_JSAMPLE == 8
182 typedef INT16 FSERROR; /* 16 bits should be enough */
183 typedef int LOCFSERROR; /* use 'int' for calculation temps */
184 #else
185 typedef INT32 FSERROR; /* may need more than 16 bits */
186 typedef INT32 LOCFSERROR; /* be sure calculation temps are big enough */
187 #endif
188
189 typedef FSERROR FAR *FSERRPTR; /* pointer to error array (in FAR storage!) */
190
191
192 /* Private subobject */
193
194 typedef struct {
195 struct jpeg_color_quantizer pub; /* public fields */
196
197 /* Space for the eventually created colormap is stashed here */
198 JSAMPARRAY sv_colormap; /* colormap allocated at init time */
199 int desired; /* desired # of colors = size of colormap */
200
201 /* Variables for accumulating image statistics */
202 hist3d histogram; /* pointer to the histogram */
203
204 boolean needs_zeroed; /* TRUE if next pass must zero histogram */
205
206 /* Variables for Floyd-Steinberg dithering */
207 FSERRPTR fserrors; /* accumulated errors */
208 boolean on_odd_row; /* flag to remember which row we are on */
209 int * error_limiter; /* table for clamping the applied error */
210 } my_cquantizer;
211
212 typedef my_cquantizer * my_cquantize_ptr;
213
214
215 /*
216 * Prescan some rows of pixels.
217 * In this module the prescan simply updates the histogram, which has been
218 * initialized to zeroes by start_pass.
219 * An output_buf parameter is required by the method signature, but no data
220 * is actually output (in fact the buffer controller is probably passing a
221 * NULL pointer).
222 */
223
224 METHODDEF(void)
225 prescan_quantize (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
226 JSAMPARRAY output_buf, int num_rows)
227 {
228 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
229 register JSAMPROW ptr;
230 register histptr histp;
231 register hist3d histogram = cquantize->histogram;
232 int row;
233 JDIMENSION col;
234 JDIMENSION width = cinfo->output_width;
235
236 for (row = 0; row < num_rows; row++) {
237 ptr = input_buf[row];
238 for (col = width; col > 0; col--) {
239 /* get pixel value and index into the histogram */
240 histp = & histogram[GETJSAMPLE(ptr[0]) >> C0_SHIFT]
241 [GETJSAMPLE(ptr[1]) >> C1_SHIFT]
242 [GETJSAMPLE(ptr[2]) >> C2_SHIFT];
243 /* increment, check for overflow and undo increment if so. */
244 if (++(*histp) <= 0)
245 (*histp)--;
246 ptr += 3;
247 }
248 }
249 }
250
251
252 /*
253 * Next we have the really interesting routines: selection of a colormap
254 * given the completed histogram.
255 * These routines work with a list of "boxes", each representing a rectangular
256 * subset of the input color space (to histogram precision).
257 */
258
259 typedef struct {
260 /* The bounds of the box (inclusive); expressed as histogram indexes */
261 int c0min, c0max;
262 int c1min, c1max;
263 int c2min, c2max;
264 /* The volume (actually 2-norm) of the box */
265 INT32 volume;
266 /* The number of nonzero histogram cells within this box */
267 long colorcount;
268 } box;
269
270 typedef box * boxptr;
271
272
273 LOCAL(boxptr)
274 find_biggest_color_pop (boxptr boxlist, int numboxes)
275 /* Find the splittable box with the largest color population */
276 /* Returns NULL if no splittable boxes remain */
277 {
278 register boxptr boxp;
279 register int i;
280 register long maxc = 0;
281 boxptr which = NULL;
282
283 for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
284 if (boxp->colorcount > maxc && boxp->volume > 0) {
285 which = boxp;
286 maxc = boxp->colorcount;
287 }
288 }
289 return which;
290 }
291
292
293 LOCAL(boxptr)
294 find_biggest_volume (boxptr boxlist, int numboxes)
295 /* Find the splittable box with the largest (scaled) volume */
296 /* Returns NULL if no splittable boxes remain */
297 {
298 register boxptr boxp;
299 register int i;
300 register INT32 maxv = 0;
301 boxptr which = NULL;
302
303 for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
304 if (boxp->volume > maxv) {
305 which = boxp;
306 maxv = boxp->volume;
307 }
308 }
309 return which;
310 }
311
312
313 LOCAL(void)
314 update_box (j_decompress_ptr cinfo, boxptr boxp)
315 /* Shrink the min/max bounds of a box to enclose only nonzero elements, */
316 /* and recompute its volume and population */
317 {
318 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
319 hist3d histogram = cquantize->histogram;
320 histptr histp;
321 int c0,c1,c2;
322 int c0min,c0max,c1min,c1max,c2min,c2max;
323 INT32 dist0,dist1,dist2;
324 long ccount;
325
326 c0min = boxp->c0min; c0max = boxp->c0max;
327 c1min = boxp->c1min; c1max = boxp->c1max;
328 c2min = boxp->c2min; c2max = boxp->c2max;
329
330 if (c0max > c0min)
331 for (c0 = c0min; c0 <= c0max; c0++)
332 for (c1 = c1min; c1 <= c1max; c1++) {
333 histp = & histogram[c0][c1][c2min];
334 for (c2 = c2min; c2 <= c2max; c2++)
335 if (*histp++ != 0) {
336 boxp->c0min = c0min = c0;
337 goto have_c0min;
338 }
339 }
340 have_c0min:
341 if (c0max > c0min)
342 for (c0 = c0max; c0 >= c0min; c0--)
343 for (c1 = c1min; c1 <= c1max; c1++) {
344 histp = & histogram[c0][c1][c2min];
345 for (c2 = c2min; c2 <= c2max; c2++)
346 if (*histp++ != 0) {
347 boxp->c0max = c0max = c0;
348 goto have_c0max;
349 }
350 }
351 have_c0max:
352 if (c1max > c1min)
353 for (c1 = c1min; c1 <= c1max; c1++)
354 for (c0 = c0min; c0 <= c0max; c0++) {
355 histp = & histogram[c0][c1][c2min];
356 for (c2 = c2min; c2 <= c2max; c2++)
357 if (*histp++ != 0) {
358 boxp->c1min = c1min = c1;
359 goto have_c1min;
360 }
361 }
362 have_c1min:
363 if (c1max > c1min)
364 for (c1 = c1max; c1 >= c1min; c1--)
365 for (c0 = c0min; c0 <= c0max; c0++) {
366 histp = & histogram[c0][c1][c2min];
367 for (c2 = c2min; c2 <= c2max; c2++)
368 if (*histp++ != 0) {
369 boxp->c1max = c1max = c1;
370 goto have_c1max;
371 }
372 }
373 have_c1max:
374 if (c2max > c2min)
375 for (c2 = c2min; c2 <= c2max; c2++)
376 for (c0 = c0min; c0 <= c0max; c0++) {
377 histp = & histogram[c0][c1min][c2];
378 for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
379 if (*histp != 0) {
380 boxp->c2min = c2min = c2;
381 goto have_c2min;
382 }
383 }
384 have_c2min:
385 if (c2max > c2min)
386 for (c2 = c2max; c2 >= c2min; c2--)
387 for (c0 = c0min; c0 <= c0max; c0++) {
388 histp = & histogram[c0][c1min][c2];
389 for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
390 if (*histp != 0) {
391 boxp->c2max = c2max = c2;
392 goto have_c2max;
393 }
394 }
395 have_c2max:
396
397 /* Update box volume.
398 * We use 2-norm rather than real volume here; this biases the method
399 * against making long narrow boxes, and it has the side benefit that
400 * a box is splittable iff norm > 0.
401 * Since the differences are expressed in histogram-cell units,
402 * we have to shift back to JSAMPLE units to get consistent distances;
403 * after which, we scale according to the selected distance scale factors.
404 */
405 dist0 = ((c0max - c0min) << C0_SHIFT) * C0_SCALE;
406 dist1 = ((c1max - c1min) << C1_SHIFT) * C1_SCALE;
407 dist2 = ((c2max - c2min) << C2_SHIFT) * C2_SCALE;
408 boxp->volume = dist0*dist0 + dist1*dist1 + dist2*dist2;
409
410 /* Now scan remaining volume of box and compute population */
411 ccount = 0;
412 for (c0 = c0min; c0 <= c0max; c0++)
413 for (c1 = c1min; c1 <= c1max; c1++) {
414 histp = & histogram[c0][c1][c2min];
415 for (c2 = c2min; c2 <= c2max; c2++, histp++)
416 if (*histp != 0) {
417 ccount++;
418 }
419 }
420 boxp->colorcount = ccount;
421 }
422
423
424 LOCAL(int)
425 median_cut (j_decompress_ptr cinfo, boxptr boxlist, int numboxes,
426 int desired_colors)
427 /* Repeatedly select and split the largest box until we have enough boxes */
428 {
429 int n,lb;
430 int c0,c1,c2,cmax;
431 register boxptr b1,b2;
432
433 while (numboxes < desired_colors) {
434 /* Select box to split.
435 * Current algorithm: by population for first half, then by volume.
436 */
437 if (numboxes*2 <= desired_colors) {
438 b1 = find_biggest_color_pop(boxlist, numboxes);
439 } else {
440 b1 = find_biggest_volume(boxlist, numboxes);
441 }
442 if (b1 == NULL) /* no splittable boxes left! */
443 break;
444 b2 = &boxlist[numboxes]; /* where new box will go */
445 /* Copy the color bounds to the new box. */
446 b2->c0max = b1->c0max; b2->c1max = b1->c1max; b2->c2max = b1->c2max;
447 b2->c0min = b1->c0min; b2->c1min = b1->c1min; b2->c2min = b1->c2min;
448 /* Choose which axis to split the box on.
449 * Current algorithm: longest scaled axis.
450 * See notes in update_box about scaling distances.
451 */
452 c0 = ((b1->c0max - b1->c0min) << C0_SHIFT) * C0_SCALE;
453 c1 = ((b1->c1max - b1->c1min) << C1_SHIFT) * C1_SCALE;
454 c2 = ((b1->c2max - b1->c2min) << C2_SHIFT) * C2_SCALE;
455 /* We want to break any ties in favor of green, then red, blue last.
456 * This code does the right thing for R,G,B or B,G,R color orders only.
457 */
458 #if RGB_RED == 0
459 cmax = c1; n = 1;
460 if (c0 > cmax) { cmax = c0; n = 0; }
461 if (c2 > cmax) { n = 2; }
462 #else
463 cmax = c1; n = 1;
464 if (c2 > cmax) { cmax = c2; n = 2; }
465 if (c0 > cmax) { n = 0; }
466 #endif
467 /* Choose split point along selected axis, and update box bounds.
468 * Current algorithm: split at halfway point.
469 * (Since the box has been shrunk to minimum volume,
470 * any split will produce two nonempty subboxes.)
471 * Note that lb value is max for lower box, so must be < old max.
472 */
473 switch (n) {
474 case 0:
475 lb = (b1->c0max + b1->c0min) / 2;
476 b1->c0max = lb;
477 b2->c0min = lb+1;
478 break;
479 case 1:
480 lb = (b1->c1max + b1->c1min) / 2;
481 b1->c1max = lb;
482 b2->c1min = lb+1;
483 break;
484 case 2:
485 lb = (b1->c2max + b1->c2min) / 2;
486 b1->c2max = lb;
487 b2->c2min = lb+1;
488 break;
489 }
490 /* Update stats for boxes */
491 update_box(cinfo, b1);
492 update_box(cinfo, b2);
493 numboxes++;
494 }
495 return numboxes;
496 }
497
498
499 LOCAL(void)
500 compute_color (j_decompress_ptr cinfo, boxptr boxp, int icolor)
501 /* Compute representative color for a box, put it in colormap[icolor] */
502 {
503 /* Current algorithm: mean weighted by pixels (not colors) */
504 /* Note it is important to get the rounding correct! */
505 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
506 hist3d histogram = cquantize->histogram;
507 histptr histp;
508 int c0,c1,c2;
509 int c0min,c0max,c1min,c1max,c2min,c2max;
510 long count;
511 long total = 0;
512 long c0total = 0;
513 long c1total = 0;
514 long c2total = 0;
515
516 c0min = boxp->c0min; c0max = boxp->c0max;
517 c1min = boxp->c1min; c1max = boxp->c1max;
518 c2min = boxp->c2min; c2max = boxp->c2max;
519
520 for (c0 = c0min; c0 <= c0max; c0++)
521 for (c1 = c1min; c1 <= c1max; c1++) {
522 histp = & histogram[c0][c1][c2min];
523 for (c2 = c2min; c2 <= c2max; c2++) {
524 if ((count = *histp++) != 0) {
525 total += count;
526 c0total += ((c0 << C0_SHIFT) + ((1<<C0_SHIFT)>>1)) * count;
527 c1total += ((c1 << C1_SHIFT) + ((1<<C1_SHIFT)>>1)) * count;
528 c2total += ((c2 << C2_SHIFT) + ((1<<C2_SHIFT)>>1)) * count;
529 }
530 }
531 }
532
533 cinfo->colormap[0][icolor] = (JSAMPLE) ((c0total + (total>>1)) / total);
534 cinfo->colormap[1][icolor] = (JSAMPLE) ((c1total + (total>>1)) / total);
535 cinfo->colormap[2][icolor] = (JSAMPLE) ((c2total + (total>>1)) / total);
536 }
537
538
539 LOCAL(void)
540 select_colors (j_decompress_ptr cinfo, int desired_colors)
541 /* Master routine for color selection */
542 {
543 boxptr boxlist;
544 int numboxes;
545 int i;
546
547 /* Allocate workspace for box list */
548 boxlist = (boxptr) (*cinfo->mem->alloc_small)
549 ((j_common_ptr) cinfo, JPOOL_IMAGE, desired_colors * SIZEOF(box));
550 /* Initialize one box containing whole space */
551 numboxes = 1;
552 boxlist[0].c0min = 0;
553 boxlist[0].c0max = MAXJSAMPLE >> C0_SHIFT;
554 boxlist[0].c1min = 0;
555 boxlist[0].c1max = MAXJSAMPLE >> C1_SHIFT;
556 boxlist[0].c2min = 0;
557 boxlist[0].c2max = MAXJSAMPLE >> C2_SHIFT;
558 /* Shrink it to actually-used volume and set its statistics */
559 update_box(cinfo, & boxlist[0]);
560 /* Perform median-cut to produce final box list */
561 numboxes = median_cut(cinfo, boxlist, numboxes, desired_colors);
562 /* Compute the representative color for each box, fill colormap */
563 for (i = 0; i < numboxes; i++)
564 compute_color(cinfo, & boxlist[i], i);
565 cinfo->actual_number_of_colors = numboxes;
566 TRACEMS1(cinfo, 1, JTRC_QUANT_SELECTED, numboxes);
567 }
568
569
570 /*
571 * These routines are concerned with the time-critical task of mapping input
572 * colors to the nearest color in the selected colormap.
573 *
574 * We re-use the histogram space as an "inverse color map", essentially a
575 * cache for the results of nearest-color searches. All colors within a
576 * histogram cell will be mapped to the same colormap entry, namely the one
577 * closest to the cell's center. This may not be quite the closest entry to
578 * the actual input color, but it's almost as good. A zero in the cache
579 * indicates we haven't found the nearest color for that cell yet; the array
580 * is cleared to zeroes before starting the mapping pass. When we find the
581 * nearest color for a cell, its colormap index plus one is recorded in the
582 * cache for future use. The pass2 scanning routines call fill_inverse_cmap
583 * when they need to use an unfilled entry in the cache.
584 *
585 * Our method of efficiently finding nearest colors is based on the "locally
586 * sorted search" idea described by Heckbert and on the incremental distance
587 * calculation described by Spencer W. Thomas in chapter III.1 of Graphics
588 * Gems II (James Arvo, ed. Academic Press, 1991). Thomas points out that
589 * the distances from a given colormap entry to each cell of the histogram can
590 * be computed quickly using an incremental method: the differences between
591 * distances to adjacent cells themselves differ by a constant. This allows a
592 * fairly fast implementation of the "brute force" approach of computing the
593 * distance from every colormap entry to every histogram cell. Unfortunately,
594 * it needs a work array to hold the best-distance-so-far for each histogram
595 * cell (because the inner loop has to be over cells, not colormap entries).
596 * The work array elements have to be INT32s, so the work array would need
597 * 256Kb at our recommended precision. This is not feasible in DOS machines.
598 *
599 * To get around these problems, we apply Thomas' method to compute the
600 * nearest colors for only the cells within a small subbox of the histogram.
601 * The work array need be only as big as the subbox, so the memory usage
602 * problem is solved. Furthermore, we need not fill subboxes that are never
603 * referenced in pass2; many images use only part of the color gamut, so a
604 * fair amount of work is saved. An additional advantage of this
605 * approach is that we can apply Heckbert's locality criterion to quickly
606 * eliminate colormap entries that are far away from the subbox; typically
607 * three-fourths of the colormap entries are rejected by Heckbert's criterion,
608 * and we need not compute their distances to individual cells in the subbox.
609 * The speed of this approach is heavily influenced by the subbox size: too
610 * small means too much overhead, too big loses because Heckbert's criterion
611 * can't eliminate as many colormap entries. Empirically the best subbox
612 * size seems to be about 1/512th of the histogram (1/8th in each direction).
613 *
614 * Thomas' article also describes a refined method which is asymptotically
615 * faster than the brute-force method, but it is also far more complex and
616 * cannot efficiently be applied to small subboxes. It is therefore not
617 * useful for programs intended to be portable to DOS machines. On machines
618 * with plenty of memory, filling the whole histogram in one shot with Thomas'
619 * refined method might be faster than the present code --- but then again,
620 * it might not be any faster, and it's certainly more complicated.
621 */
622
623
624 /* log2(histogram cells in update box) for each axis; this can be adjusted */
625 #define BOX_C0_LOG (HIST_C0_BITS-3)
626 #define BOX_C1_LOG (HIST_C1_BITS-3)
627 #define BOX_C2_LOG (HIST_C2_BITS-3)
628
629 #define BOX_C0_ELEMS (1<<BOX_C0_LOG) /* # of hist cells in update box */
630 #define BOX_C1_ELEMS (1<<BOX_C1_LOG)
631 #define BOX_C2_ELEMS (1<<BOX_C2_LOG)
632
633 #define BOX_C0_SHIFT (C0_SHIFT + BOX_C0_LOG)
634 #define BOX_C1_SHIFT (C1_SHIFT + BOX_C1_LOG)
635 #define BOX_C2_SHIFT (C2_SHIFT + BOX_C2_LOG)
636
637
638 /*
639 * The next three routines implement inverse colormap filling. They could
640 * all be folded into one big routine, but splitting them up this way saves
641 * some stack space (the mindist[] and bestdist[] arrays need not coexist)
642 * and may allow some compilers to produce better code by registerizing more
643 * inner-loop variables.
644 */
645
646 LOCAL(int)
647 find_nearby_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
648 JSAMPLE colorlist[])
649 /* Locate the colormap entries close enough to an update box to be candidates
650 * for the nearest entry to some cell(s) in the update box. The update box
651 * is specified by the center coordinates of its first cell. The number of
652 * candidate colormap entries is returned, and their colormap indexes are
653 * placed in colorlist[].
654 * This routine uses Heckbert's "locally sorted search" criterion to select
655 * the colors that need further consideration.
656 */
657 {
658 int numcolors = cinfo->actual_number_of_colors;
659 int maxc0, maxc1, maxc2;
660 int centerc0, centerc1, centerc2;
661 int i, x, ncolors;
662 INT32 minmaxdist, min_dist, max_dist, tdist;
663 INT32 mindist[MAXNUMCOLORS]; /* min distance to colormap entry i */
664
665 /* Compute true coordinates of update box's upper corner and center.
666 * Actually we compute the coordinates of the center of the upper-corner
667 * histogram cell, which are the upper bounds of the volume we care about.
668 * Note that since ">>" rounds down, the "center" values may be closer to
669 * min than to max; hence comparisons to them must be "<=", not "<".
670 */
671 maxc0 = minc0 + ((1 << BOX_C0_SHIFT) - (1 << C0_SHIFT));
672 centerc0 = (minc0 + maxc0) >> 1;
673 maxc1 = minc1 + ((1 << BOX_C1_SHIFT) - (1 << C1_SHIFT));
674 centerc1 = (minc1 + maxc1) >> 1;
675 maxc2 = minc2 + ((1 << BOX_C2_SHIFT) - (1 << C2_SHIFT));
676 centerc2 = (minc2 + maxc2) >> 1;
677
678 /* For each color in colormap, find:
679 * 1. its minimum squared-distance to any point in the update box
680 * (zero if color is within update box);
681 * 2. its maximum squared-distance to any point in the update box.
682 * Both of these can be found by considering only the corners of the box.
683 * We save the minimum distance for each color in mindist[];
684 * only the smallest maximum distance is of interest.
685 */
686 minmaxdist = 0x7FFFFFFFL;
687
688 for (i = 0; i < numcolors; i++) {
689 /* We compute the squared-c0-distance term, then add in the other two. */
690 x = GETJSAMPLE(cinfo->colormap[0][i]);
691 if (x < minc0) {
692 tdist = (x - minc0) * C0_SCALE;
693 min_dist = tdist*tdist;
694 tdist = (x - maxc0) * C0_SCALE;
695 max_dist = tdist*tdist;
696 } else if (x > maxc0) {
697 tdist = (x - maxc0) * C0_SCALE;
698 min_dist = tdist*tdist;
699 tdist = (x - minc0) * C0_SCALE;
700 max_dist = tdist*tdist;
701 } else {
702 /* within cell range so no contribution to min_dist */
703 min_dist = 0;
704 if (x <= centerc0) {
705 tdist = (x - maxc0) * C0_SCALE;
706 max_dist = tdist*tdist;
707 } else {
708 tdist = (x - minc0) * C0_SCALE;
709 max_dist = tdist*tdist;
710 }
711 }
712
713 x = GETJSAMPLE(cinfo->colormap[1][i]);
714 if (x < minc1) {
715 tdist = (x - minc1) * C1_SCALE;
716 min_dist += tdist*tdist;
717 tdist = (x - maxc1) * C1_SCALE;
718 max_dist += tdist*tdist;
719 } else if (x > maxc1) {
720 tdist = (x - maxc1) * C1_SCALE;
721 min_dist += tdist*tdist;
722 tdist = (x - minc1) * C1_SCALE;
723 max_dist += tdist*tdist;
724 } else {
725 /* within cell range so no contribution to min_dist */
726 if (x <= centerc1) {
727 tdist = (x - maxc1) * C1_SCALE;
728 max_dist += tdist*tdist;
729 } else {
730 tdist = (x - minc1) * C1_SCALE;
731 max_dist += tdist*tdist;
732 }
733 }
734
735 x = GETJSAMPLE(cinfo->colormap[2][i]);
736 if (x < minc2) {
737 tdist = (x - minc2) * C2_SCALE;
738 min_dist += tdist*tdist;
739 tdist = (x - maxc2) * C2_SCALE;
740 max_dist += tdist*tdist;
741 } else if (x > maxc2) {
742 tdist = (x - maxc2) * C2_SCALE;
743 min_dist += tdist*tdist;
744 tdist = (x - minc2) * C2_SCALE;
745 max_dist += tdist*tdist;
746 } else {
747 /* within cell range so no contribution to min_dist */
748 if (x <= centerc2) {
749 tdist = (x - maxc2) * C2_SCALE;
750 max_dist += tdist*tdist;
751 } else {
752 tdist = (x - minc2) * C2_SCALE;
753 max_dist += tdist*tdist;
754 }
755 }
756
757 mindist[i] = min_dist; /* save away the results */
758 if (max_dist < minmaxdist)
759 minmaxdist = max_dist;
760 }
761
762 /* Now we know that no cell in the update box is more than minmaxdist
763 * away from some colormap entry. Therefore, only colors that are
764 * within minmaxdist of some part of the box need be considered.
765 */
766 ncolors = 0;
767 for (i = 0; i < numcolors; i++) {
768 if (mindist[i] <= minmaxdist)
769 colorlist[ncolors++] = (JSAMPLE) i;
770 }
771 return ncolors;
772 }
773
774
775 LOCAL(void)
776 find_best_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
777 int numcolors, JSAMPLE colorlist[], JSAMPLE bestcolor[])
778 /* Find the closest colormap entry for each cell in the update box,
779 * given the list of candidate colors prepared by find_nearby_colors.
780 * Return the indexes of the closest entries in the bestcolor[] array.
781 * This routine uses Thomas' incremental distance calculation method to
782 * find the distance from a colormap entry to successive cells in the box.
783 */
784 {
785 int ic0, ic1, ic2;
786 int i, icolor;
787 register INT32 * bptr; /* pointer into bestdist[] array */
788 JSAMPLE * cptr; /* pointer into bestcolor[] array */
789 INT32 dist0, dist1; /* initial distance values */
790 register INT32 dist2; /* current distance in inner loop */
791 INT32 xx0, xx1; /* distance increments */
792 register INT32 xx2;
793 INT32 inc0, inc1, inc2; /* initial values for increments */
794 /* This array holds the distance to the nearest-so-far color for each cell */
795 INT32 bestdist[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
796
797 /* Initialize best-distance for each cell of the update box */
798 bptr = bestdist;
799 for (i = BOX_C0_ELEMS*BOX_C1_ELEMS*BOX_C2_ELEMS-1; i >= 0; i--)
800 *bptr++ = 0x7FFFFFFFL;
801
802 /* For each color selected by find_nearby_colors,
803 * compute its distance to the center of each cell in the box.
804 * If that's less than best-so-far, update best distance and color number.
805 */
806
807 /* Nominal steps between cell centers ("x" in Thomas article) */
808 #define STEP_C0 ((1 << C0_SHIFT) * C0_SCALE)
809 #define STEP_C1 ((1 << C1_SHIFT) * C1_SCALE)
810 #define STEP_C2 ((1 << C2_SHIFT) * C2_SCALE)
811
812 for (i = 0; i < numcolors; i++) {
813 icolor = GETJSAMPLE(colorlist[i]);
814 /* Compute (square of) distance from minc0/c1/c2 to this color */
815 inc0 = (minc0 - GETJSAMPLE(cinfo->colormap[0][icolor])) * C0_SCALE;
816 dist0 = inc0*inc0;
817 inc1 = (minc1 - GETJSAMPLE(cinfo->colormap[1][icolor])) * C1_SCALE;
818 dist0 += inc1*inc1;
819 inc2 = (minc2 - GETJSAMPLE(cinfo->colormap[2][icolor])) * C2_SCALE;
820 dist0 += inc2*inc2;
821 /* Form the initial difference increments */
822 inc0 = inc0 * (2 * STEP_C0) + STEP_C0 * STEP_C0;
823 inc1 = inc1 * (2 * STEP_C1) + STEP_C1 * STEP_C1;
824 inc2 = inc2 * (2 * STEP_C2) + STEP_C2 * STEP_C2;
825 /* Now loop over all cells in box, updating distance per Thomas method */
826 bptr = bestdist;
827 cptr = bestcolor;
828 xx0 = inc0;
829 for (ic0 = BOX_C0_ELEMS-1; ic0 >= 0; ic0--) {
830 dist1 = dist0;
831 xx1 = inc1;
832 for (ic1 = BOX_C1_ELEMS-1; ic1 >= 0; ic1--) {
833 dist2 = dist1;
834 xx2 = inc2;
835 for (ic2 = BOX_C2_ELEMS-1; ic2 >= 0; ic2--) {
836 if (dist2 < *bptr) {
837 *bptr = dist2;
838 *cptr = (JSAMPLE) icolor;
839 }
840 dist2 += xx2;
841 xx2 += 2 * STEP_C2 * STEP_C2;
842 bptr++;
843 cptr++;
844 }
845 dist1 += xx1;
846 xx1 += 2 * STEP_C1 * STEP_C1;
847 }
848 dist0 += xx0;
849 xx0 += 2 * STEP_C0 * STEP_C0;
850 }
851 }
852 }
853
854
855 LOCAL(void)
856 fill_inverse_cmap (j_decompress_ptr cinfo, int c0, int c1, int c2)
857 /* Fill the inverse-colormap entries in the update box that contains */
858 /* histogram cell c0/c1/c2. (Only that one cell MUST be filled, but */
859 /* we can fill as many others as we wish.) */
860 {
861 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
862 hist3d histogram = cquantize->histogram;
863 int minc0, minc1, minc2; /* lower left corner of update box */
864 int ic0, ic1, ic2;
865 register JSAMPLE * cptr; /* pointer into bestcolor[] array */
866 register histptr cachep; /* pointer into main cache array */
867 /* This array lists the candidate colormap indexes. */
868 JSAMPLE colorlist[MAXNUMCOLORS];
869 int numcolors; /* number of candidate colors */
870 /* This array holds the actually closest colormap index for each cell. */
871 JSAMPLE bestcolor[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
872
873 /* Convert cell coordinates to update box ID */
874 c0 >>= BOX_C0_LOG;
875 c1 >>= BOX_C1_LOG;
876 c2 >>= BOX_C2_LOG;
877
878 /* Compute true coordinates of update box's origin corner.
879 * Actually we compute the coordinates of the center of the corner
880 * histogram cell, which are the lower bounds of the volume we care about.
881 */
882 minc0 = (c0 << BOX_C0_SHIFT) + ((1 << C0_SHIFT) >> 1);
883 minc1 = (c1 << BOX_C1_SHIFT) + ((1 << C1_SHIFT) >> 1);
884 minc2 = (c2 << BOX_C2_SHIFT) + ((1 << C2_SHIFT) >> 1);
885
886 /* Determine which colormap entries are close enough to be candidates
887 * for the nearest entry to some cell in the update box.
888 */
889 numcolors = find_nearby_colors(cinfo, minc0, minc1, minc2, colorlist);
890
891 /* Determine the actually nearest colors. */
892 find_best_colors(cinfo, minc0, minc1, minc2, numcolors, colorlist,
893 bestcolor);
894
895 /* Save the best color numbers (plus 1) in the main cache array */
896 c0 <<= BOX_C0_LOG; /* convert ID back to base cell indexes */
897 c1 <<= BOX_C1_LOG;
898 c2 <<= BOX_C2_LOG;
899 cptr = bestcolor;
900 for (ic0 = 0; ic0 < BOX_C0_ELEMS; ic0++) {
901 for (ic1 = 0; ic1 < BOX_C1_ELEMS; ic1++) {
902 cachep = & histogram[c0+ic0][c1+ic1][c2];
903 for (ic2 = 0; ic2 < BOX_C2_ELEMS; ic2++) {
904 *cachep++ = (histcell) (GETJSAMPLE(*cptr++) + 1);
905 }
906 }
907 }
908 }
909
910
911 /*
912 * Map some rows of pixels to the output colormapped representation.
913 */
914
915 METHODDEF(void)
916 pass2_no_dither (j_decompress_ptr cinfo,
917 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
918 /* This version performs no dithering */
919 {
920 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
921 hist3d histogram = cquantize->histogram;
922 register JSAMPROW inptr, outptr;
923 register histptr cachep;
924 register int c0, c1, c2;
925 int row;
926 JDIMENSION col;
927 JDIMENSION width = cinfo->output_width;
928
929 for (row = 0; row < num_rows; row++) {
930 inptr = input_buf[row];
931 outptr = output_buf[row];
932 for (col = width; col > 0; col--) {
933 /* get pixel value and index into the cache */
934 c0 = GETJSAMPLE(*inptr++) >> C0_SHIFT;
935 c1 = GETJSAMPLE(*inptr++) >> C1_SHIFT;
936 c2 = GETJSAMPLE(*inptr++) >> C2_SHIFT;
937 cachep = & histogram[c0][c1][c2];
938 /* If we have not seen this color before, find nearest colormap entry */
939 /* and update the cache */
940 if (*cachep == 0)
941 fill_inverse_cmap(cinfo, c0,c1,c2);
942 /* Now emit the colormap index for this cell */
943 *outptr++ = (JSAMPLE) (*cachep - 1);
944 }
945 }
946 }
947
948
949 METHODDEF(void)
950 pass2_fs_dither (j_decompress_ptr cinfo,
951 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
952 /* This version performs Floyd-Steinberg dithering */
953 {
954 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
955 hist3d histogram = cquantize->histogram;
956 register LOCFSERROR cur0, cur1, cur2; /* current error or pixel value */
957 LOCFSERROR belowerr0, belowerr1, belowerr2; /* error for pixel below cur */
958 LOCFSERROR bpreverr0, bpreverr1, bpreverr2; /* error for below/prev col */
959 register FSERRPTR errorptr; /* => fserrors[] at column before current */
960 JSAMPROW inptr; /* => current input pixel */
961 JSAMPROW outptr; /* => current output pixel */
962 histptr cachep;
963 int dir; /* +1 or -1 depending on direction */
964 int dir3; /* 3*dir, for advancing inptr & errorptr */
965 int row;
966 JDIMENSION col;
967 JDIMENSION width = cinfo->output_width;
968 JSAMPLE *range_limit = cinfo->sample_range_limit;
969 int *error_limit = cquantize->error_limiter;
970 JSAMPROW colormap0 = cinfo->colormap[0];
971 JSAMPROW colormap1 = cinfo->colormap[1];
972 JSAMPROW colormap2 = cinfo->colormap[2];
973 SHIFT_TEMPS
974
975 for (row = 0; row < num_rows; row++) {
976 inptr = input_buf[row];
977 outptr = output_buf[row];
978 if (cquantize->on_odd_row) {
979 /* work right to left in this row */
980 inptr += (width-1) * 3; /* so point to rightmost pixel */
981 outptr += width-1;
982 dir = -1;
983 dir3 = -3;
984 errorptr = cquantize->fserrors + (width+1)*3; /* => entry after last column */
985 cquantize->on_odd_row = FALSE; /* flip for next time */
986 } else {
987 /* work left to right in this row */
988 dir = 1;
989 dir3 = 3;
990 errorptr = cquantize->fserrors; /* => entry before first real column */
991 cquantize->on_odd_row = TRUE; /* flip for next time */
992 }
993 /* Preset error values: no error propagated to first pixel from left */
994 cur0 = cur1 = cur2 = 0;
995 /* and no error propagated to row below yet */
996 belowerr0 = belowerr1 = belowerr2 = 0;
997 bpreverr0 = bpreverr1 = bpreverr2 = 0;
998
999 for (col = width; col > 0; col--) {
1000 /* curN holds the error propagated from the previous pixel on the
1001 * current line. Add the error propagated from the previous line
1002 * to form the complete error correction term for this pixel, and
1003 * round the error term (which is expressed * 16) to an integer.
1004 * RIGHT_SHIFT rounds towards minus infinity, so adding 8 is correct
1005 * for either sign of the error value.
1006 * Note: errorptr points to *previous* column's array entry.
1007 */
1008 cur0 = RIGHT_SHIFT(cur0 + errorptr[dir3+0] + 8, 4);
1009 cur1 = RIGHT_SHIFT(cur1 + errorptr[dir3+1] + 8, 4);
1010 cur2 = RIGHT_SHIFT(cur2 + errorptr[dir3+2] + 8, 4);
1011 /* Limit the error using transfer function set by init_error_limit.
1012 * See comments with init_error_limit for rationale.
1013 */
1014 cur0 = error_limit[cur0];
1015 cur1 = error_limit[cur1];
1016 cur2 = error_limit[cur2];
1017 /* Form pixel value + error, and range-limit to 0..MAXJSAMPLE.
1018 * The maximum error is +- MAXJSAMPLE (or less with error limiting);
1019 * this sets the required size of the range_limit array.
1020 */
1021 cur0 += GETJSAMPLE(inptr[0]);
1022 cur1 += GETJSAMPLE(inptr[1]);
1023 cur2 += GETJSAMPLE(inptr[2]);
1024 cur0 = GETJSAMPLE(range_limit[cur0]);
1025 cur1 = GETJSAMPLE(range_limit[cur1]);
1026 cur2 = GETJSAMPLE(range_limit[cur2]);
1027 /* Index into the cache with adjusted pixel value */
1028 cachep = & histogram[cur0>>C0_SHIFT][cur1>>C1_SHIFT][cur2>>C2_SHIFT];
1029 /* If we have not seen this color before, find nearest colormap */
1030 /* entry and update the cache */
1031 if (*cachep == 0)
1032 fill_inverse_cmap(cinfo, cur0>>C0_SHIFT,cur1>>C1_SHIFT,cur2>>C2_SHIFT);
1033 /* Now emit the colormap index for this cell */
1034 { register int pixcode = *cachep - 1;
1035 *outptr = (JSAMPLE) pixcode;
1036 /* Compute representation error for this pixel */
1037 cur0 -= GETJSAMPLE(colormap0[pixcode]);
1038 cur1 -= GETJSAMPLE(colormap1[pixcode]);
1039 cur2 -= GETJSAMPLE(colormap2[pixcode]);
1040 }
1041 /* Compute error fractions to be propagated to adjacent pixels.
1042 * Add these into the running sums, and simultaneously shift the
1043 * next-line error sums left by 1 column.
1044 */
1045 { register LOCFSERROR bnexterr, delta;
1046
1047 bnexterr = cur0; /* Process component 0 */
1048 delta = cur0 * 2;
1049 cur0 += delta; /* form error * 3 */
1050 errorptr[0] = (FSERROR) (bpreverr0 + cur0);
1051 cur0 += delta; /* form error * 5 */
1052 bpreverr0 = belowerr0 + cur0;
1053 belowerr0 = bnexterr;
1054 cur0 += delta; /* form error * 7 */
1055 bnexterr = cur1; /* Process component 1 */
1056 delta = cur1 * 2;
1057 cur1 += delta; /* form error * 3 */
1058 errorptr[1] = (FSERROR) (bpreverr1 + cur1);
1059 cur1 += delta; /* form error * 5 */
1060 bpreverr1 = belowerr1 + cur1;
1061 belowerr1 = bnexterr;
1062 cur1 += delta; /* form error * 7 */
1063 bnexterr = cur2; /* Process component 2 */
1064 delta = cur2 * 2;
1065 cur2 += delta; /* form error * 3 */
1066 errorptr[2] = (FSERROR) (bpreverr2 + cur2);
1067 cur2 += delta; /* form error * 5 */
1068 bpreverr2 = belowerr2 + cur2;
1069 belowerr2 = bnexterr;
1070 cur2 += delta; /* form error * 7 */
1071 }
1072 /* At this point curN contains the 7/16 error value to be propagated
1073 * to the next pixel on the current line, and all the errors for the
1074 * next line have been shifted over. We are therefore ready to move on.
1075 */
1076 inptr += dir3; /* Advance pixel pointers to next column */
1077 outptr += dir;
1078 errorptr += dir3; /* advance errorptr to current column */
1079 }
1080 /* Post-loop cleanup: we must unload the final error values into the
1081 * final fserrors[] entry. Note we need not unload belowerrN because
1082 * it is for the dummy column before or after the actual array.
1083 */
1084 errorptr[0] = (FSERROR) bpreverr0; /* unload prev errs into array */
1085 errorptr[1] = (FSERROR) bpreverr1;
1086 errorptr[2] = (FSERROR) bpreverr2;
1087 }
1088 }
1089
1090
1091 /*
1092 * Initialize the error-limiting transfer function (lookup table).
1093 * The raw F-S error computation can potentially compute error values of up to
1094 * +- MAXJSAMPLE. But we want the maximum correction applied to a pixel to be
1095 * much less, otherwise obviously wrong pixels will be created. (Typical
1096 * effects include weird fringes at color-area boundaries, isolated bright
1097 * pixels in a dark area, etc.) The standard advice for avoiding this problem
1098 * is to ensure that the "corners" of the color cube are allocated as output
1099 * colors; then repeated errors in the same direction cannot cause cascading
1100 * error buildup. However, that only prevents the error from getting
1101 * completely out of hand; Aaron Giles reports that error limiting improves
1102 * the results even with corner colors allocated.
1103 * A simple clamping of the error values to about +- MAXJSAMPLE/8 works pretty
1104 * well, but the smoother transfer function used below is even better. Thanks
1105 * to Aaron Giles for this idea.
1106 */
1107
1108 LOCAL(void)
1109 init_error_limit (j_decompress_ptr cinfo)
1110 /* Allocate and fill in the error_limiter table */
1111 {
1112 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
1113 int * table;
1114 int in, out;
1115
1116 table = (int *) (*cinfo->mem->alloc_small)
1117 ((j_common_ptr) cinfo, JPOOL_IMAGE, (MAXJSAMPLE*2+1) * SIZEOF(int));
1118 table += MAXJSAMPLE; /* so can index -MAXJSAMPLE .. +MAXJSAMPLE */
1119 cquantize->error_limiter = table;
1120
1121 #define STEPSIZE ((MAXJSAMPLE+1)/16)
1122 /* Map errors 1:1 up to +- MAXJSAMPLE/16 */
1123 out = 0;
1124 for (in = 0; in < STEPSIZE; in++, out++) {
1125 table[in] = out; table[-in] = -out;
1126 }
1127 /* Map errors 1:2 up to +- 3*MAXJSAMPLE/16 */
1128 for (; in < STEPSIZE*3; in++, out += (in&1) ? 0 : 1) {
1129 table[in] = out; table[-in] = -out;
1130 }
1131 /* Clamp the rest to final out value (which is (MAXJSAMPLE+1)/8) */
1132 for (; in <= MAXJSAMPLE; in++) {
1133 table[in] = out; table[-in] = -out;
1134 }
1135 #undef STEPSIZE
1136 }
1137
1138
1139 /*
1140 * Finish up at the end of each pass.
1141 */
1142
1143 METHODDEF(void)
1144 finish_pass1 (j_decompress_ptr cinfo)
1145 {
1146 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
1147
1148 /* Select the representative colors and fill in cinfo->colormap */
1149 cinfo->colormap = cquantize->sv_colormap;
1150 select_colors(cinfo, cquantize->desired);
1151 /* Force next pass to zero the color index table */
1152 cquantize->needs_zeroed = TRUE;
1153 }
1154
1155
1156 METHODDEF(void)
1157 finish_pass2 (j_decompress_ptr cinfo)
1158 {
1159 /* no work */
1160 }
1161
1162
1163 /*
1164 * Initialize for each processing pass.
1165 */
1166
1167 METHODDEF(void)
1168 start_pass_2_quant (j_decompress_ptr cinfo, boolean is_pre_scan)
1169 {
1170 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
1171 hist3d histogram = cquantize->histogram;
1172 int i;
1173
1174 /* Only F-S dithering or no dithering is supported. */
1175 /* If user asks for ordered dither, give him F-S. */
1176 if (cinfo->dither_mode != JDITHER_NONE)
1177 cinfo->dither_mode = JDITHER_FS;
1178
1179 if (is_pre_scan) {
1180 /* Set up method pointers */
1181 cquantize->pub.color_quantize = prescan_quantize;
1182 cquantize->pub.finish_pass = finish_pass1;
1183 cquantize->needs_zeroed = TRUE; /* Always zero histogram */
1184 } else {
1185 /* Set up method pointers */
1186 if (cinfo->dither_mode == JDITHER_FS)
1187 cquantize->pub.color_quantize = pass2_fs_dither;
1188 else
1189 cquantize->pub.color_quantize = pass2_no_dither;
1190 cquantize->pub.finish_pass = finish_pass2;
1191
1192 /* Make sure color count is acceptable */
1193 i = cinfo->actual_number_of_colors;
1194 if (i < 1)
1195 ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 1);
1196 if (i > MAXNUMCOLORS)
1197 ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
1198
1199 if (cinfo->dither_mode == JDITHER_FS) {
1200 size_t arraysize = (size_t) ((cinfo->output_width + 2) *
1201 (3 * SIZEOF(FSERROR)));
1202 /* Allocate Floyd-Steinberg workspace if we didn't already. */
1203 if (cquantize->fserrors == NULL)
1204 cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
1205 ((j_common_ptr) cinfo, JPOOL_IMAGE, arraysize);
1206 /* Initialize the propagated errors to zero. */
1207 FMEMZERO((void FAR *) cquantize->fserrors, arraysize);
1208 /* Make the error-limit table if we didn't already. */
1209 if (cquantize->error_limiter == NULL)
1210 init_error_limit(cinfo);
1211 cquantize->on_odd_row = FALSE;
1212 }
1213
1214 }
1215 /* Zero the histogram or inverse color map, if necessary */
1216 if (cquantize->needs_zeroed) {
1217 for (i = 0; i < HIST_C0_ELEMS; i++) {
1218 FMEMZERO((void FAR *) histogram[i],
1219 HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
1220 }
1221 cquantize->needs_zeroed = FALSE;
1222 }
1223 }
1224
1225
1226 /*
1227 * Switch to a new external colormap between output passes.
1228 */
1229
1230 METHODDEF(void)
1231 new_color_map_2_quant (j_decompress_ptr cinfo)
1232 {
1233 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
1234
1235 /* Reset the inverse color map */
1236 cquantize->needs_zeroed = TRUE;
1237 }
1238
1239
1240 /*
1241 * Module initialization routine for 2-pass color quantization.
1242 */
1243
1244 GLOBAL(void)
1245 jinit_2pass_quantizer (j_decompress_ptr cinfo)
1246 {
1247 my_cquantize_ptr cquantize;
1248 int i;
1249
1250 cquantize = (my_cquantize_ptr)
1251 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
1252 SIZEOF(my_cquantizer));
1253 cinfo->cquantize = (struct jpeg_color_quantizer *) cquantize;
1254 cquantize->pub.start_pass = start_pass_2_quant;
1255 cquantize->pub.new_color_map = new_color_map_2_quant;
1256 cquantize->fserrors = NULL; /* flag optional arrays not allocated */
1257 cquantize->error_limiter = NULL;
1258
1259 /* Make sure jdmaster didn't give me a case I can't handle */
1260 if (cinfo->out_color_components != 3)
1261 ERREXIT(cinfo, JERR_NOTIMPL);
1262
1263 /* Allocate the histogram/inverse colormap storage */
1264 cquantize->histogram = (hist3d) (*cinfo->mem->alloc_small)
1265 ((j_common_ptr) cinfo, JPOOL_IMAGE, HIST_C0_ELEMS * SIZEOF(hist2d));
1266 for (i = 0; i < HIST_C0_ELEMS; i++) {
1267 cquantize->histogram[i] = (hist2d) (*cinfo->mem->alloc_large)
1268 ((j_common_ptr) cinfo, JPOOL_IMAGE,
1269 HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
1270 }
1271 cquantize->needs_zeroed = TRUE; /* histogram is garbage now */
1272
1273 /* Allocate storage for the completed colormap, if required.
1274 * We do this now since it is FAR storage and may affect
1275 * the memory manager's space calculations.
1276 */
1277 if (cinfo->enable_2pass_quant) {
1278 /* Make sure color count is acceptable */
1279 int desired = cinfo->desired_number_of_colors;
1280 /* Lower bound on # of colors ... somewhat arbitrary as long as > 0 */
1281 if (desired < 8)
1282 ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 8);
1283 /* Make sure colormap indexes can be represented by JSAMPLEs */
1284 if (desired > MAXNUMCOLORS)
1285 ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
1286 cquantize->sv_colormap = (*cinfo->mem->alloc_sarray)
1287 ((j_common_ptr) cinfo,JPOOL_IMAGE, (JDIMENSION) desired, (JDIMENSION) 3);
1288 cquantize->desired = desired;
1289 } else
1290 cquantize->sv_colormap = NULL;
1291
1292 /* Only F-S dithering or no dithering is supported. */
1293 /* If user asks for ordered dither, give him F-S. */
1294 if (cinfo->dither_mode != JDITHER_NONE)
1295 cinfo->dither_mode = JDITHER_FS;
1296
1297 /* Allocate Floyd-Steinberg workspace if necessary.
1298 * This isn't really needed until pass 2, but again it is FAR storage.
1299 * Although we will cope with a later change in dither_mode,
1300 * we do not promise to honor max_memory_to_use if dither_mode changes.
1301 */
1302 if (cinfo->dither_mode == JDITHER_FS) {
1303 cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
1304 ((j_common_ptr) cinfo, JPOOL_IMAGE,
1305 (size_t) ((cinfo->output_width + 2) * (3 * SIZEOF(FSERROR))));
1306 /* Might as well create the error-limiting table too. */
1307 init_error_limit(cinfo);
1308 }
1309 }
1310
1311 #endif /* QUANT_2PASS_SUPPORTED */