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
147typedef UINT16 histcell;	/* histogram cell; prefer an unsigned type */
148
149typedef histcell FAR * histptr;	/* for pointers to histogram cells */
150
151typedef histcell hist1d[HIST_C2_ELEMS]; /* typedefs for the array */
152typedef hist1d FAR * hist2d;	/* type for the 2nd-level pointers */
153typedef 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
181typedef INT16 FSERROR;		/* 16 bits should be enough */
182typedef int LOCFSERROR;		/* use 'int' for calculation temps */
183#else
184typedef INT32 FSERROR;		/* may need more than 16 bits */
185typedef INT32 LOCFSERROR;	/* be sure calculation temps are big enough */
186#endif
187
188typedef FSERROR FAR *FSERRPTR;	/* pointer to error array (in FAR storage!) */
189
190
191/* Private subobject */
192
193typedef 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
211typedef 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
223METHODDEF(void)
224prescan_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
258typedef 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
269typedef box * boxptr;
270
271
272LOCAL(boxptr)
273find_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
292LOCAL(boxptr)
293find_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
312LOCAL(void)
313update_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
423LOCAL(int)
424median_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
498LOCAL(void)
499compute_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
538LOCAL(void)
539select_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
645LOCAL(int)
646find_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
774LOCAL(void)
775find_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
854LOCAL(void)
855fill_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
914METHODDEF(void)
915pass2_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
948METHODDEF(void)
949pass2_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
1107LOCAL(void)
1108init_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
1142METHODDEF(void)
1143finish_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
1155METHODDEF(void)
1156finish_pass2 (j_decompress_ptr cinfo)
1157{
1158  /* no work */
1159}
1160
1161
1162/*
1163 * Initialize for each processing pass.
1164 */
1165
1166METHODDEF(void)
1167start_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
1229METHODDEF(void)
1230new_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
1243GLOBAL(void)
1244jinit_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