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