source: rtems-graphics-toolkit/jpeg-8d/jquant2.c @ 86b99f7

Last change on this file since 86b99f7 was 86b99f7, checked in by Alexandru-Sever Horin <alex.sever.h@…>, on Aug 1, 2012 at 10:40:32 PM

Added jpeg-8d version. Made modifications to compile for RTEMS, without man or binaries

  • Property mode set to 100644
File size: 47.3 KB
Line 
1/*
2 * jquant2.c
3 *
4 * Copyright (C) 1991-1996, Thomas G. Lane.
5 * Modified 2011 by Guido Vollbeding.
6 * This file is part of the Independent JPEG Group's software.
7 * For conditions of distribution and use, see the accompanying README file.
8 *
9 * This file contains 2-pass color quantization (color mapping) routines.
10 * These routines provide selection of a custom color map for an image,
11 * followed by mapping of the image to that color map, with optional
12 * Floyd-Steinberg dithering.
13 * It is also possible to use just the second pass to map to an arbitrary
14 * externally-given color map.
15 *
16 * Note: ordered dithering is not supported, since there isn't any fast
17 * way to compute intercolor distances; it's unclear that ordered dither's
18 * fundamental assumptions even hold with an irregularly spaced color map.
19 */
20
21#define JPEG_INTERNALS
22#include "jinclude.h"
23#include "jpeglib.h"
24
25#ifdef QUANT_2PASS_SUPPORTED
26
27
28/*
29 * This module implements the well-known Heckbert paradigm for color
30 * quantization.  Most of the ideas used here can be traced back to
31 * Heckbert's seminal paper
32 *   Heckbert, Paul.  "Color Image Quantization for Frame Buffer Display",
33 *   Proc. SIGGRAPH '82, Computer Graphics v.16 #3 (July 1982), pp 297-304.
34 *
35 * In the first pass over the image, we accumulate a histogram showing the
36 * usage count of each possible color.  To keep the histogram to a reasonable
37 * size, we reduce the precision of the input; typical practice is to retain
38 * 5 or 6 bits per color, so that 8 or 4 different input values are counted
39 * in the same histogram cell.
40 *
41 * Next, the color-selection step begins with a box representing the whole
42 * color space, and repeatedly splits the "largest" remaining box until we
43 * have as many boxes as desired colors.  Then the mean color in each
44 * remaining box becomes one of the possible output colors.
45 *
46 * The second pass over the image maps each input pixel to the closest output
47 * color (optionally after applying a Floyd-Steinberg dithering correction).
48 * This mapping is logically trivial, but making it go fast enough requires
49 * considerable care.
50 *
51 * Heckbert-style quantizers vary a good deal in their policies for choosing
52 * the "largest" box and deciding where to cut it.  The particular policies
53 * used here have proved out well in experimental comparisons, but better ones
54 * may yet be found.
55 *
56 * In earlier versions of the IJG code, this module quantized in YCbCr color
57 * space, processing the raw upsampled data without a color conversion step.
58 * This allowed the color conversion math to be done only once per colormap
59 * entry, not once per pixel.  However, that optimization precluded other
60 * useful optimizations (such as merging color conversion with upsampling)
61 * and it also interfered with desired capabilities such as quantizing to an
62 * externally-supplied colormap.  We have therefore abandoned that approach.
63 * The present code works in the post-conversion color space, typically RGB.
64 *
65 * To improve the visual quality of the results, we actually work in scaled
66 * RGB space, giving G distances more weight than R, and R in turn more than
67 * B.  To do everything in integer math, we must use integer scale factors.
68 * The 2/3/1 scale factors used here correspond loosely to the relative
69 * weights of the colors in the NTSC grayscale equation.
70 * If you want to use this code to quantize a non-RGB color space, you'll
71 * probably need to change these scale factors.
72 */
73
74#define R_SCALE 2               /* scale R distances by this much */
75#define G_SCALE 3               /* scale G distances by this much */
76#define B_SCALE 1               /* and B by this much */
77
78/* Relabel R/G/B as components 0/1/2, respecting the RGB ordering defined
79 * in jmorecfg.h.  As the code stands, it will do the right thing for R,G,B
80 * and B,G,R orders.  If you define some other weird order in jmorecfg.h,
81 * you'll get compile errors until you extend this logic.  In that case
82 * you'll probably want to tweak the histogram sizes too.
83 */
84
85#if RGB_RED == 0
86#define C0_SCALE R_SCALE
87#endif
88#if RGB_BLUE == 0
89#define C0_SCALE B_SCALE
90#endif
91#if RGB_GREEN == 1
92#define C1_SCALE G_SCALE
93#endif
94#if RGB_RED == 2
95#define C2_SCALE R_SCALE
96#endif
97#if RGB_BLUE == 2
98#define C2_SCALE B_SCALE
99#endif
100
101
102/*
103 * First we have the histogram data structure and routines for creating it.
104 *
105 * The number of bits of precision can be adjusted by changing these symbols.
106 * We recommend keeping 6 bits for G and 5 each for R and B.
107 * If you have plenty of memory and cycles, 6 bits all around gives marginally
108 * better results; if you are short of memory, 5 bits all around will save
109 * some space but degrade the results.
110 * To maintain a fully accurate histogram, we'd need to allocate a "long"
111 * (preferably unsigned long) for each cell.  In practice this is overkill;
112 * we can get by with 16 bits per cell.  Few of the cell counts will overflow,
113 * and clamping those that do overflow to the maximum value will give close-
114 * enough results.  This reduces the recommended histogram size from 256Kb
115 * to 128Kb, which is a useful savings on PC-class machines.
116 * (In the second pass the histogram space is re-used for pixel mapping data;
117 * in that capacity, each cell must be able to store zero to the number of
118 * desired colors.  16 bits/cell is plenty for that too.)
119 * Since the JPEG code is intended to run in small memory model on 80x86
120 * machines, we can't just allocate the histogram in one chunk.  Instead
121 * of a true 3-D array, we use a row of pointers to 2-D arrays.  Each
122 * pointer corresponds to a C0 value (typically 2^5 = 32 pointers) and
123 * each 2-D array has 2^6*2^5 = 2048 or 2^6*2^6 = 4096 entries.  Note that
124 * on 80x86 machines, the pointer row is in near memory but the actual
125 * arrays are in far memory (same arrangement as we use for image arrays).
126 */
127
128#define MAXNUMCOLORS  (MAXJSAMPLE+1) /* maximum size of colormap */
129
130/* These will do the right thing for either R,G,B or B,G,R color order,
131 * but you may not like the results for other color orders.
132 */
133#define HIST_C0_BITS  5         /* bits of precision in R/B histogram */
134#define HIST_C1_BITS  6         /* bits of precision in G histogram */
135#define HIST_C2_BITS  5         /* bits of precision in B/R histogram */
136
137/* Number of elements along histogram axes. */
138#define HIST_C0_ELEMS  (1<<HIST_C0_BITS)
139#define HIST_C1_ELEMS  (1<<HIST_C1_BITS)
140#define HIST_C2_ELEMS  (1<<HIST_C2_BITS)
141
142/* These are the amounts to shift an input value to get a histogram index. */
143#define C0_SHIFT  (BITS_IN_JSAMPLE-HIST_C0_BITS)
144#define C1_SHIFT  (BITS_IN_JSAMPLE-HIST_C1_BITS)
145#define C2_SHIFT  (BITS_IN_JSAMPLE-HIST_C2_BITS)
146
147
148typedef UINT16 histcell;        /* histogram cell; prefer an unsigned type */
149
150typedef histcell FAR * histptr; /* for pointers to histogram cells */
151
152typedef histcell hist1d[HIST_C2_ELEMS]; /* typedefs for the array */
153typedef hist1d FAR * hist2d;    /* type for the 2nd-level pointers */
154typedef hist2d * hist3d;        /* type for top-level pointer */
155
156
157/* Declarations for Floyd-Steinberg dithering.
158 *
159 * Errors are accumulated into the array fserrors[], at a resolution of
160 * 1/16th of a pixel count.  The error at a given pixel is propagated
161 * to its not-yet-processed neighbors using the standard F-S fractions,
162 *              ...     (here)  7/16
163 *              3/16    5/16    1/16
164 * We work left-to-right on even rows, right-to-left on odd rows.
165 *
166 * We can get away with a single array (holding one row's worth of errors)
167 * by using it to store the current row's errors at pixel columns not yet
168 * processed, but the next row's errors at columns already processed.  We
169 * need only a few extra variables to hold the errors immediately around the
170 * current column.  (If we are lucky, those variables are in registers, but
171 * even if not, they're probably cheaper to access than array elements are.)
172 *
173 * The fserrors[] array has (#columns + 2) entries; the extra entry at
174 * each end saves us from special-casing the first and last pixels.
175 * Each entry is three values long, one value for each color component.
176 *
177 * Note: on a wide image, we might not have enough room in a PC's near data
178 * segment to hold the error array; so it is allocated with alloc_large.
179 */
180
181#if BITS_IN_JSAMPLE == 8
182typedef INT16 FSERROR;          /* 16 bits should be enough */
183typedef int LOCFSERROR;         /* use 'int' for calculation temps */
184#else
185typedef INT32 FSERROR;          /* may need more than 16 bits */
186typedef INT32 LOCFSERROR;       /* be sure calculation temps are big enough */
187#endif
188
189typedef FSERROR FAR *FSERRPTR;  /* pointer to error array (in FAR storage!) */
190
191
192/* Private subobject */
193
194typedef struct {
195  struct jpeg_color_quantizer pub; /* public fields */
196
197  /* Space for the eventually created colormap is stashed here */
198  JSAMPARRAY sv_colormap;       /* colormap allocated at init time */
199  int desired;                  /* desired # of colors = size of colormap */
200
201  /* Variables for accumulating image statistics */
202  hist3d histogram;             /* pointer to the histogram */
203
204  boolean needs_zeroed;         /* TRUE if next pass must zero histogram */
205
206  /* Variables for Floyd-Steinberg dithering */
207  FSERRPTR fserrors;            /* accumulated errors */
208  boolean on_odd_row;           /* flag to remember which row we are on */
209  int * error_limiter;          /* table for clamping the applied error */
210} my_cquantizer;
211
212typedef my_cquantizer * my_cquantize_ptr;
213
214
215/*
216 * Prescan some rows of pixels.
217 * In this module the prescan simply updates the histogram, which has been
218 * initialized to zeroes by start_pass.
219 * An output_buf parameter is required by the method signature, but no data
220 * is actually output (in fact the buffer controller is probably passing a
221 * NULL pointer).
222 */
223
224METHODDEF(void)
225prescan_quantize (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
226                  JSAMPARRAY output_buf, int num_rows)
227{
228  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
229  register JSAMPROW ptr;
230  register histptr histp;
231  register hist3d histogram = cquantize->histogram;
232  int row;
233  JDIMENSION col;
234  JDIMENSION width = cinfo->output_width;
235
236  for (row = 0; row < num_rows; row++) {
237    ptr = input_buf[row];
238    for (col = width; col > 0; col--) {
239      /* get pixel value and index into the histogram */
240      histp = & histogram[GETJSAMPLE(ptr[0]) >> C0_SHIFT]
241                         [GETJSAMPLE(ptr[1]) >> C1_SHIFT]
242                         [GETJSAMPLE(ptr[2]) >> C2_SHIFT];
243      /* increment, check for overflow and undo increment if so. */
244      if (++(*histp) <= 0)
245        (*histp)--;
246      ptr += 3;
247    }
248  }
249}
250
251
252/*
253 * Next we have the really interesting routines: selection of a colormap
254 * given the completed histogram.
255 * These routines work with a list of "boxes", each representing a rectangular
256 * subset of the input color space (to histogram precision).
257 */
258
259typedef struct {
260  /* The bounds of the box (inclusive); expressed as histogram indexes */
261  int c0min, c0max;
262  int c1min, c1max;
263  int c2min, c2max;
264  /* The volume (actually 2-norm) of the box */
265  INT32 volume;
266  /* The number of nonzero histogram cells within this box */
267  long colorcount;
268} box;
269
270typedef box * boxptr;
271
272
273LOCAL(boxptr)
274find_biggest_color_pop (boxptr boxlist, int numboxes)
275/* Find the splittable box with the largest color population */
276/* Returns NULL if no splittable boxes remain */
277{
278  register boxptr boxp;
279  register int i;
280  register long maxc = 0;
281  boxptr which = NULL;
282 
283  for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
284    if (boxp->colorcount > maxc && boxp->volume > 0) {
285      which = boxp;
286      maxc = boxp->colorcount;
287    }
288  }
289  return which;
290}
291
292
293LOCAL(boxptr)
294find_biggest_volume (boxptr boxlist, int numboxes)
295/* Find the splittable box with the largest (scaled) volume */
296/* Returns NULL if no splittable boxes remain */
297{
298  register boxptr boxp;
299  register int i;
300  register INT32 maxv = 0;
301  boxptr which = NULL;
302 
303  for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
304    if (boxp->volume > maxv) {
305      which = boxp;
306      maxv = boxp->volume;
307    }
308  }
309  return which;
310}
311
312
313LOCAL(void)
314update_box (j_decompress_ptr cinfo, boxptr boxp)
315/* Shrink the min/max bounds of a box to enclose only nonzero elements, */
316/* and recompute its volume and population */
317{
318  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
319  hist3d histogram = cquantize->histogram;
320  histptr histp;
321  int c0,c1,c2;
322  int c0min,c0max,c1min,c1max,c2min,c2max;
323  INT32 dist0,dist1,dist2;
324  long ccount;
325 
326  c0min = boxp->c0min;  c0max = boxp->c0max;
327  c1min = boxp->c1min;  c1max = boxp->c1max;
328  c2min = boxp->c2min;  c2max = boxp->c2max;
329 
330  if (c0max > c0min)
331    for (c0 = c0min; c0 <= c0max; c0++)
332      for (c1 = c1min; c1 <= c1max; c1++) {
333        histp = & histogram[c0][c1][c2min];
334        for (c2 = c2min; c2 <= c2max; c2++)
335          if (*histp++ != 0) {
336            boxp->c0min = c0min = c0;
337            goto have_c0min;
338          }
339      }
340 have_c0min:
341  if (c0max > c0min)
342    for (c0 = c0max; c0 >= c0min; c0--)
343      for (c1 = c1min; c1 <= c1max; c1++) {
344        histp = & histogram[c0][c1][c2min];
345        for (c2 = c2min; c2 <= c2max; c2++)
346          if (*histp++ != 0) {
347            boxp->c0max = c0max = c0;
348            goto have_c0max;
349          }
350      }
351 have_c0max:
352  if (c1max > c1min)
353    for (c1 = c1min; c1 <= c1max; c1++)
354      for (c0 = c0min; c0 <= c0max; c0++) {
355        histp = & histogram[c0][c1][c2min];
356        for (c2 = c2min; c2 <= c2max; c2++)
357          if (*histp++ != 0) {
358            boxp->c1min = c1min = c1;
359            goto have_c1min;
360          }
361      }
362 have_c1min:
363  if (c1max > c1min)
364    for (c1 = c1max; c1 >= c1min; c1--)
365      for (c0 = c0min; c0 <= c0max; c0++) {
366        histp = & histogram[c0][c1][c2min];
367        for (c2 = c2min; c2 <= c2max; c2++)
368          if (*histp++ != 0) {
369            boxp->c1max = c1max = c1;
370            goto have_c1max;
371          }
372      }
373 have_c1max:
374  if (c2max > c2min)
375    for (c2 = c2min; c2 <= c2max; c2++)
376      for (c0 = c0min; c0 <= c0max; c0++) {
377        histp = & histogram[c0][c1min][c2];
378        for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
379          if (*histp != 0) {
380            boxp->c2min = c2min = c2;
381            goto have_c2min;
382          }
383      }
384 have_c2min:
385  if (c2max > c2min)
386    for (c2 = c2max; c2 >= c2min; c2--)
387      for (c0 = c0min; c0 <= c0max; c0++) {
388        histp = & histogram[c0][c1min][c2];
389        for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
390          if (*histp != 0) {
391            boxp->c2max = c2max = c2;
392            goto have_c2max;
393          }
394      }
395 have_c2max:
396
397  /* Update box volume.
398   * We use 2-norm rather than real volume here; this biases the method
399   * against making long narrow boxes, and it has the side benefit that
400   * a box is splittable iff norm > 0.
401   * Since the differences are expressed in histogram-cell units,
402   * we have to shift back to JSAMPLE units to get consistent distances;
403   * after which, we scale according to the selected distance scale factors.
404   */
405  dist0 = ((c0max - c0min) << C0_SHIFT) * C0_SCALE;
406  dist1 = ((c1max - c1min) << C1_SHIFT) * C1_SCALE;
407  dist2 = ((c2max - c2min) << C2_SHIFT) * C2_SCALE;
408  boxp->volume = dist0*dist0 + dist1*dist1 + dist2*dist2;
409 
410  /* Now scan remaining volume of box and compute population */
411  ccount = 0;
412  for (c0 = c0min; c0 <= c0max; c0++)
413    for (c1 = c1min; c1 <= c1max; c1++) {
414      histp = & histogram[c0][c1][c2min];
415      for (c2 = c2min; c2 <= c2max; c2++, histp++)
416        if (*histp != 0) {
417          ccount++;
418        }
419    }
420  boxp->colorcount = ccount;
421}
422
423
424LOCAL(int)
425median_cut (j_decompress_ptr cinfo, boxptr boxlist, int numboxes,
426            int desired_colors)
427/* Repeatedly select and split the largest box until we have enough boxes */
428{
429  int n,lb;
430  int c0,c1,c2,cmax;
431  register boxptr b1,b2;
432
433  while (numboxes < desired_colors) {
434    /* Select box to split.
435     * Current algorithm: by population for first half, then by volume.
436     */
437    if (numboxes*2 <= desired_colors) {
438      b1 = find_biggest_color_pop(boxlist, numboxes);
439    } else {
440      b1 = find_biggest_volume(boxlist, numboxes);
441    }
442    if (b1 == NULL)             /* no splittable boxes left! */
443      break;
444    b2 = &boxlist[numboxes];    /* where new box will go */
445    /* Copy the color bounds to the new box. */
446    b2->c0max = b1->c0max; b2->c1max = b1->c1max; b2->c2max = b1->c2max;
447    b2->c0min = b1->c0min; b2->c1min = b1->c1min; b2->c2min = b1->c2min;
448    /* Choose which axis to split the box on.
449     * Current algorithm: longest scaled axis.
450     * See notes in update_box about scaling distances.
451     */
452    c0 = ((b1->c0max - b1->c0min) << C0_SHIFT) * C0_SCALE;
453    c1 = ((b1->c1max - b1->c1min) << C1_SHIFT) * C1_SCALE;
454    c2 = ((b1->c2max - b1->c2min) << C2_SHIFT) * C2_SCALE;
455    /* We want to break any ties in favor of green, then red, blue last.
456     * This code does the right thing for R,G,B or B,G,R color orders only.
457     */
458#if RGB_RED == 0
459    cmax = c1; n = 1;
460    if (c0 > cmax) { cmax = c0; n = 0; }
461    if (c2 > cmax) { n = 2; }
462#else
463    cmax = c1; n = 1;
464    if (c2 > cmax) { cmax = c2; n = 2; }
465    if (c0 > cmax) { n = 0; }
466#endif
467    /* Choose split point along selected axis, and update box bounds.
468     * Current algorithm: split at halfway point.
469     * (Since the box has been shrunk to minimum volume,
470     * any split will produce two nonempty subboxes.)
471     * Note that lb value is max for lower box, so must be < old max.
472     */
473    switch (n) {
474    case 0:
475      lb = (b1->c0max + b1->c0min) / 2;
476      b1->c0max = lb;
477      b2->c0min = lb+1;
478      break;
479    case 1:
480      lb = (b1->c1max + b1->c1min) / 2;
481      b1->c1max = lb;
482      b2->c1min = lb+1;
483      break;
484    case 2:
485      lb = (b1->c2max + b1->c2min) / 2;
486      b1->c2max = lb;
487      b2->c2min = lb+1;
488      break;
489    }
490    /* Update stats for boxes */
491    update_box(cinfo, b1);
492    update_box(cinfo, b2);
493    numboxes++;
494  }
495  return numboxes;
496}
497
498
499LOCAL(void)
500compute_color (j_decompress_ptr cinfo, boxptr boxp, int icolor)
501/* Compute representative color for a box, put it in colormap[icolor] */
502{
503  /* Current algorithm: mean weighted by pixels (not colors) */
504  /* Note it is important to get the rounding correct! */
505  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
506  hist3d histogram = cquantize->histogram;
507  histptr histp;
508  int c0,c1,c2;
509  int c0min,c0max,c1min,c1max,c2min,c2max;
510  long count;
511  long total = 0;
512  long c0total = 0;
513  long c1total = 0;
514  long c2total = 0;
515 
516  c0min = boxp->c0min;  c0max = boxp->c0max;
517  c1min = boxp->c1min;  c1max = boxp->c1max;
518  c2min = boxp->c2min;  c2max = boxp->c2max;
519 
520  for (c0 = c0min; c0 <= c0max; c0++)
521    for (c1 = c1min; c1 <= c1max; c1++) {
522      histp = & histogram[c0][c1][c2min];
523      for (c2 = c2min; c2 <= c2max; c2++) {
524        if ((count = *histp++) != 0) {
525          total += count;
526          c0total += ((c0 << C0_SHIFT) + ((1<<C0_SHIFT)>>1)) * count;
527          c1total += ((c1 << C1_SHIFT) + ((1<<C1_SHIFT)>>1)) * count;
528          c2total += ((c2 << C2_SHIFT) + ((1<<C2_SHIFT)>>1)) * count;
529        }
530      }
531    }
532 
533  cinfo->colormap[0][icolor] = (JSAMPLE) ((c0total + (total>>1)) / total);
534  cinfo->colormap[1][icolor] = (JSAMPLE) ((c1total + (total>>1)) / total);
535  cinfo->colormap[2][icolor] = (JSAMPLE) ((c2total + (total>>1)) / total);
536}
537
538
539LOCAL(void)
540select_colors (j_decompress_ptr cinfo, int desired_colors)
541/* Master routine for color selection */
542{
543  boxptr boxlist;
544  int numboxes;
545  int i;
546
547  /* Allocate workspace for box list */
548  boxlist = (boxptr) (*cinfo->mem->alloc_small)
549    ((j_common_ptr) cinfo, JPOOL_IMAGE, desired_colors * SIZEOF(box));
550  /* Initialize one box containing whole space */
551  numboxes = 1;
552  boxlist[0].c0min = 0;
553  boxlist[0].c0max = MAXJSAMPLE >> C0_SHIFT;
554  boxlist[0].c1min = 0;
555  boxlist[0].c1max = MAXJSAMPLE >> C1_SHIFT;
556  boxlist[0].c2min = 0;
557  boxlist[0].c2max = MAXJSAMPLE >> C2_SHIFT;
558  /* Shrink it to actually-used volume and set its statistics */
559  update_box(cinfo, & boxlist[0]);
560  /* Perform median-cut to produce final box list */
561  numboxes = median_cut(cinfo, boxlist, numboxes, desired_colors);
562  /* Compute the representative color for each box, fill colormap */
563  for (i = 0; i < numboxes; i++)
564    compute_color(cinfo, & boxlist[i], i);
565  cinfo->actual_number_of_colors = numboxes;
566  TRACEMS1(cinfo, 1, JTRC_QUANT_SELECTED, numboxes);
567}
568
569
570/*
571 * These routines are concerned with the time-critical task of mapping input
572 * colors to the nearest color in the selected colormap.
573 *
574 * We re-use the histogram space as an "inverse color map", essentially a
575 * cache for the results of nearest-color searches.  All colors within a
576 * histogram cell will be mapped to the same colormap entry, namely the one
577 * closest to the cell's center.  This may not be quite the closest entry to
578 * the actual input color, but it's almost as good.  A zero in the cache
579 * indicates we haven't found the nearest color for that cell yet; the array
580 * is cleared to zeroes before starting the mapping pass.  When we find the
581 * nearest color for a cell, its colormap index plus one is recorded in the
582 * cache for future use.  The pass2 scanning routines call fill_inverse_cmap
583 * when they need to use an unfilled entry in the cache.
584 *
585 * Our method of efficiently finding nearest colors is based on the "locally
586 * sorted search" idea described by Heckbert and on the incremental distance
587 * calculation described by Spencer W. Thomas in chapter III.1 of Graphics
588 * Gems II (James Arvo, ed.  Academic Press, 1991).  Thomas points out that
589 * the distances from a given colormap entry to each cell of the histogram can
590 * be computed quickly using an incremental method: the differences between
591 * distances to adjacent cells themselves differ by a constant.  This allows a
592 * fairly fast implementation of the "brute force" approach of computing the
593 * distance from every colormap entry to every histogram cell.  Unfortunately,
594 * it needs a work array to hold the best-distance-so-far for each histogram
595 * cell (because the inner loop has to be over cells, not colormap entries).
596 * The work array elements have to be INT32s, so the work array would need
597 * 256Kb at our recommended precision.  This is not feasible in DOS machines.
598 *
599 * To get around these problems, we apply Thomas' method to compute the
600 * nearest colors for only the cells within a small subbox of the histogram.
601 * The work array need be only as big as the subbox, so the memory usage
602 * problem is solved.  Furthermore, we need not fill subboxes that are never
603 * referenced in pass2; many images use only part of the color gamut, so a
604 * fair amount of work is saved.  An additional advantage of this
605 * approach is that we can apply Heckbert's locality criterion to quickly
606 * eliminate colormap entries that are far away from the subbox; typically
607 * three-fourths of the colormap entries are rejected by Heckbert's criterion,
608 * and we need not compute their distances to individual cells in the subbox.
609 * The speed of this approach is heavily influenced by the subbox size: too
610 * small means too much overhead, too big loses because Heckbert's criterion
611 * can't eliminate as many colormap entries.  Empirically the best subbox
612 * size seems to be about 1/512th of the histogram (1/8th in each direction).
613 *
614 * Thomas' article also describes a refined method which is asymptotically
615 * faster than the brute-force method, but it is also far more complex and
616 * cannot efficiently be applied to small subboxes.  It is therefore not
617 * useful for programs intended to be portable to DOS machines.  On machines
618 * with plenty of memory, filling the whole histogram in one shot with Thomas'
619 * refined method might be faster than the present code --- but then again,
620 * it might not be any faster, and it's certainly more complicated.
621 */
622
623
624/* log2(histogram cells in update box) for each axis; this can be adjusted */
625#define BOX_C0_LOG  (HIST_C0_BITS-3)
626#define BOX_C1_LOG  (HIST_C1_BITS-3)
627#define BOX_C2_LOG  (HIST_C2_BITS-3)
628
629#define BOX_C0_ELEMS  (1<<BOX_C0_LOG) /* # of hist cells in update box */
630#define BOX_C1_ELEMS  (1<<BOX_C1_LOG)
631#define BOX_C2_ELEMS  (1<<BOX_C2_LOG)
632
633#define BOX_C0_SHIFT  (C0_SHIFT + BOX_C0_LOG)
634#define BOX_C1_SHIFT  (C1_SHIFT + BOX_C1_LOG)
635#define BOX_C2_SHIFT  (C2_SHIFT + BOX_C2_LOG)
636
637
638/*
639 * The next three routines implement inverse colormap filling.  They could
640 * all be folded into one big routine, but splitting them up this way saves
641 * some stack space (the mindist[] and bestdist[] arrays need not coexist)
642 * and may allow some compilers to produce better code by registerizing more
643 * inner-loop variables.
644 */
645
646LOCAL(int)
647find_nearby_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
648                    JSAMPLE colorlist[])
649/* Locate the colormap entries close enough to an update box to be candidates
650 * for the nearest entry to some cell(s) in the update box.  The update box
651 * is specified by the center coordinates of its first cell.  The number of
652 * candidate colormap entries is returned, and their colormap indexes are
653 * placed in colorlist[].
654 * This routine uses Heckbert's "locally sorted search" criterion to select
655 * the colors that need further consideration.
656 */
657{
658  int numcolors = cinfo->actual_number_of_colors;
659  int maxc0, maxc1, maxc2;
660  int centerc0, centerc1, centerc2;
661  int i, x, ncolors;
662  INT32 minmaxdist, min_dist, max_dist, tdist;
663  INT32 mindist[MAXNUMCOLORS];  /* min distance to colormap entry i */
664
665  /* Compute true coordinates of update box's upper corner and center.
666   * Actually we compute the coordinates of the center of the upper-corner
667   * histogram cell, which are the upper bounds of the volume we care about.
668   * Note that since ">>" rounds down, the "center" values may be closer to
669   * min than to max; hence comparisons to them must be "<=", not "<".
670   */
671  maxc0 = minc0 + ((1 << BOX_C0_SHIFT) - (1 << C0_SHIFT));
672  centerc0 = (minc0 + maxc0) >> 1;
673  maxc1 = minc1 + ((1 << BOX_C1_SHIFT) - (1 << C1_SHIFT));
674  centerc1 = (minc1 + maxc1) >> 1;
675  maxc2 = minc2 + ((1 << BOX_C2_SHIFT) - (1 << C2_SHIFT));
676  centerc2 = (minc2 + maxc2) >> 1;
677
678  /* For each color in colormap, find:
679   *  1. its minimum squared-distance to any point in the update box
680   *     (zero if color is within update box);
681   *  2. its maximum squared-distance to any point in the update box.
682   * Both of these can be found by considering only the corners of the box.
683   * We save the minimum distance for each color in mindist[];
684   * only the smallest maximum distance is of interest.
685   */
686  minmaxdist = 0x7FFFFFFFL;
687
688  for (i = 0; i < numcolors; i++) {
689    /* We compute the squared-c0-distance term, then add in the other two. */
690    x = GETJSAMPLE(cinfo->colormap[0][i]);
691    if (x < minc0) {
692      tdist = (x - minc0) * C0_SCALE;
693      min_dist = tdist*tdist;
694      tdist = (x - maxc0) * C0_SCALE;
695      max_dist = tdist*tdist;
696    } else if (x > maxc0) {
697      tdist = (x - maxc0) * C0_SCALE;
698      min_dist = tdist*tdist;
699      tdist = (x - minc0) * C0_SCALE;
700      max_dist = tdist*tdist;
701    } else {
702      /* within cell range so no contribution to min_dist */
703      min_dist = 0;
704      if (x <= centerc0) {
705        tdist = (x - maxc0) * C0_SCALE;
706        max_dist = tdist*tdist;
707      } else {
708        tdist = (x - minc0) * C0_SCALE;
709        max_dist = tdist*tdist;
710      }
711    }
712
713    x = GETJSAMPLE(cinfo->colormap[1][i]);
714    if (x < minc1) {
715      tdist = (x - minc1) * C1_SCALE;
716      min_dist += tdist*tdist;
717      tdist = (x - maxc1) * C1_SCALE;
718      max_dist += tdist*tdist;
719    } else if (x > maxc1) {
720      tdist = (x - maxc1) * C1_SCALE;
721      min_dist += tdist*tdist;
722      tdist = (x - minc1) * C1_SCALE;
723      max_dist += tdist*tdist;
724    } else {
725      /* within cell range so no contribution to min_dist */
726      if (x <= centerc1) {
727        tdist = (x - maxc1) * C1_SCALE;
728        max_dist += tdist*tdist;
729      } else {
730        tdist = (x - minc1) * C1_SCALE;
731        max_dist += tdist*tdist;
732      }
733    }
734
735    x = GETJSAMPLE(cinfo->colormap[2][i]);
736    if (x < minc2) {
737      tdist = (x - minc2) * C2_SCALE;
738      min_dist += tdist*tdist;
739      tdist = (x - maxc2) * C2_SCALE;
740      max_dist += tdist*tdist;
741    } else if (x > maxc2) {
742      tdist = (x - maxc2) * C2_SCALE;
743      min_dist += tdist*tdist;
744      tdist = (x - minc2) * C2_SCALE;
745      max_dist += tdist*tdist;
746    } else {
747      /* within cell range so no contribution to min_dist */
748      if (x <= centerc2) {
749        tdist = (x - maxc2) * C2_SCALE;
750        max_dist += tdist*tdist;
751      } else {
752        tdist = (x - minc2) * C2_SCALE;
753        max_dist += tdist*tdist;
754      }
755    }
756
757    mindist[i] = min_dist;      /* save away the results */
758    if (max_dist < minmaxdist)
759      minmaxdist = max_dist;
760  }
761
762  /* Now we know that no cell in the update box is more than minmaxdist
763   * away from some colormap entry.  Therefore, only colors that are
764   * within minmaxdist of some part of the box need be considered.
765   */
766  ncolors = 0;
767  for (i = 0; i < numcolors; i++) {
768    if (mindist[i] <= minmaxdist)
769      colorlist[ncolors++] = (JSAMPLE) i;
770  }
771  return ncolors;
772}
773
774
775LOCAL(void)
776find_best_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
777                  int numcolors, JSAMPLE colorlist[], JSAMPLE bestcolor[])
778/* Find the closest colormap entry for each cell in the update box,
779 * given the list of candidate colors prepared by find_nearby_colors.
780 * Return the indexes of the closest entries in the bestcolor[] array.
781 * This routine uses Thomas' incremental distance calculation method to
782 * find the distance from a colormap entry to successive cells in the box.
783 */
784{
785  int ic0, ic1, ic2;
786  int i, icolor;
787  register INT32 * bptr;        /* pointer into bestdist[] array */
788  JSAMPLE * cptr;               /* pointer into bestcolor[] array */
789  INT32 dist0, dist1;           /* initial distance values */
790  register INT32 dist2;         /* current distance in inner loop */
791  INT32 xx0, xx1;               /* distance increments */
792  register INT32 xx2;
793  INT32 inc0, inc1, inc2;       /* initial values for increments */
794  /* This array holds the distance to the nearest-so-far color for each cell */
795  INT32 bestdist[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
796
797  /* Initialize best-distance for each cell of the update box */
798  bptr = bestdist;
799  for (i = BOX_C0_ELEMS*BOX_C1_ELEMS*BOX_C2_ELEMS-1; i >= 0; i--)
800    *bptr++ = 0x7FFFFFFFL;
801 
802  /* For each color selected by find_nearby_colors,
803   * compute its distance to the center of each cell in the box.
804   * If that's less than best-so-far, update best distance and color number.
805   */
806 
807  /* Nominal steps between cell centers ("x" in Thomas article) */
808#define STEP_C0  ((1 << C0_SHIFT) * C0_SCALE)
809#define STEP_C1  ((1 << C1_SHIFT) * C1_SCALE)
810#define STEP_C2  ((1 << C2_SHIFT) * C2_SCALE)
811 
812  for (i = 0; i < numcolors; i++) {
813    icolor = GETJSAMPLE(colorlist[i]);
814    /* Compute (square of) distance from minc0/c1/c2 to this color */
815    inc0 = (minc0 - GETJSAMPLE(cinfo->colormap[0][icolor])) * C0_SCALE;
816    dist0 = inc0*inc0;
817    inc1 = (minc1 - GETJSAMPLE(cinfo->colormap[1][icolor])) * C1_SCALE;
818    dist0 += inc1*inc1;
819    inc2 = (minc2 - GETJSAMPLE(cinfo->colormap[2][icolor])) * C2_SCALE;
820    dist0 += inc2*inc2;
821    /* Form the initial difference increments */
822    inc0 = inc0 * (2 * STEP_C0) + STEP_C0 * STEP_C0;
823    inc1 = inc1 * (2 * STEP_C1) + STEP_C1 * STEP_C1;
824    inc2 = inc2 * (2 * STEP_C2) + STEP_C2 * STEP_C2;
825    /* Now loop over all cells in box, updating distance per Thomas method */
826    bptr = bestdist;
827    cptr = bestcolor;
828    xx0 = inc0;
829    for (ic0 = BOX_C0_ELEMS-1; ic0 >= 0; ic0--) {
830      dist1 = dist0;
831      xx1 = inc1;
832      for (ic1 = BOX_C1_ELEMS-1; ic1 >= 0; ic1--) {
833        dist2 = dist1;
834        xx2 = inc2;
835        for (ic2 = BOX_C2_ELEMS-1; ic2 >= 0; ic2--) {
836          if (dist2 < *bptr) {
837            *bptr = dist2;
838            *cptr = (JSAMPLE) icolor;
839          }
840          dist2 += xx2;
841          xx2 += 2 * STEP_C2 * STEP_C2;
842          bptr++;
843          cptr++;
844        }
845        dist1 += xx1;
846        xx1 += 2 * STEP_C1 * STEP_C1;
847      }
848      dist0 += xx0;
849      xx0 += 2 * STEP_C0 * STEP_C0;
850    }
851  }
852}
853
854
855LOCAL(void)
856fill_inverse_cmap (j_decompress_ptr cinfo, int c0, int c1, int c2)
857/* Fill the inverse-colormap entries in the update box that contains */
858/* histogram cell c0/c1/c2.  (Only that one cell MUST be filled, but */
859/* we can fill as many others as we wish.) */
860{
861  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
862  hist3d histogram = cquantize->histogram;
863  int minc0, minc1, minc2;      /* lower left corner of update box */
864  int ic0, ic1, ic2;
865  register JSAMPLE * cptr;      /* pointer into bestcolor[] array */
866  register histptr cachep;      /* pointer into main cache array */
867  /* This array lists the candidate colormap indexes. */
868  JSAMPLE colorlist[MAXNUMCOLORS];
869  int numcolors;                /* number of candidate colors */
870  /* This array holds the actually closest colormap index for each cell. */
871  JSAMPLE bestcolor[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
872
873  /* Convert cell coordinates to update box ID */
874  c0 >>= BOX_C0_LOG;
875  c1 >>= BOX_C1_LOG;
876  c2 >>= BOX_C2_LOG;
877
878  /* Compute true coordinates of update box's origin corner.
879   * Actually we compute the coordinates of the center of the corner
880   * histogram cell, which are the lower bounds of the volume we care about.
881   */
882  minc0 = (c0 << BOX_C0_SHIFT) + ((1 << C0_SHIFT) >> 1);
883  minc1 = (c1 << BOX_C1_SHIFT) + ((1 << C1_SHIFT) >> 1);
884  minc2 = (c2 << BOX_C2_SHIFT) + ((1 << C2_SHIFT) >> 1);
885 
886  /* Determine which colormap entries are close enough to be candidates
887   * for the nearest entry to some cell in the update box.
888   */
889  numcolors = find_nearby_colors(cinfo, minc0, minc1, minc2, colorlist);
890
891  /* Determine the actually nearest colors. */
892  find_best_colors(cinfo, minc0, minc1, minc2, numcolors, colorlist,
893                   bestcolor);
894
895  /* Save the best color numbers (plus 1) in the main cache array */
896  c0 <<= BOX_C0_LOG;            /* convert ID back to base cell indexes */
897  c1 <<= BOX_C1_LOG;
898  c2 <<= BOX_C2_LOG;
899  cptr = bestcolor;
900  for (ic0 = 0; ic0 < BOX_C0_ELEMS; ic0++) {
901    for (ic1 = 0; ic1 < BOX_C1_ELEMS; ic1++) {
902      cachep = & histogram[c0+ic0][c1+ic1][c2];
903      for (ic2 = 0; ic2 < BOX_C2_ELEMS; ic2++) {
904        *cachep++ = (histcell) (GETJSAMPLE(*cptr++) + 1);
905      }
906    }
907  }
908}
909
910
911/*
912 * Map some rows of pixels to the output colormapped representation.
913 */
914
915METHODDEF(void)
916pass2_no_dither (j_decompress_ptr cinfo,
917                 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
918/* This version performs no dithering */
919{
920  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
921  hist3d histogram = cquantize->histogram;
922  register JSAMPROW inptr, outptr;
923  register histptr cachep;
924  register int c0, c1, c2;
925  int row;
926  JDIMENSION col;
927  JDIMENSION width = cinfo->output_width;
928
929  for (row = 0; row < num_rows; row++) {
930    inptr = input_buf[row];
931    outptr = output_buf[row];
932    for (col = width; col > 0; col--) {
933      /* get pixel value and index into the cache */
934      c0 = GETJSAMPLE(*inptr++) >> C0_SHIFT;
935      c1 = GETJSAMPLE(*inptr++) >> C1_SHIFT;
936      c2 = GETJSAMPLE(*inptr++) >> C2_SHIFT;
937      cachep = & histogram[c0][c1][c2];
938      /* If we have not seen this color before, find nearest colormap entry */
939      /* and update the cache */
940      if (*cachep == 0)
941        fill_inverse_cmap(cinfo, c0,c1,c2);
942      /* Now emit the colormap index for this cell */
943      *outptr++ = (JSAMPLE) (*cachep - 1);
944    }
945  }
946}
947
948
949METHODDEF(void)
950pass2_fs_dither (j_decompress_ptr cinfo,
951                 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
952/* This version performs Floyd-Steinberg dithering */
953{
954  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
955  hist3d histogram = cquantize->histogram;
956  register LOCFSERROR cur0, cur1, cur2; /* current error or pixel value */
957  LOCFSERROR belowerr0, belowerr1, belowerr2; /* error for pixel below cur */
958  LOCFSERROR bpreverr0, bpreverr1, bpreverr2; /* error for below/prev col */
959  register FSERRPTR errorptr;   /* => fserrors[] at column before current */
960  JSAMPROW inptr;               /* => current input pixel */
961  JSAMPROW outptr;              /* => current output pixel */
962  histptr cachep;
963  int dir;                      /* +1 or -1 depending on direction */
964  int dir3;                     /* 3*dir, for advancing inptr & errorptr */
965  int row;
966  JDIMENSION col;
967  JDIMENSION width = cinfo->output_width;
968  JSAMPLE *range_limit = cinfo->sample_range_limit;
969  int *error_limit = cquantize->error_limiter;
970  JSAMPROW colormap0 = cinfo->colormap[0];
971  JSAMPROW colormap1 = cinfo->colormap[1];
972  JSAMPROW colormap2 = cinfo->colormap[2];
973  SHIFT_TEMPS
974
975  for (row = 0; row < num_rows; row++) {
976    inptr = input_buf[row];
977    outptr = output_buf[row];
978    if (cquantize->on_odd_row) {
979      /* work right to left in this row */
980      inptr += (width-1) * 3;   /* so point to rightmost pixel */
981      outptr += width-1;
982      dir = -1;
983      dir3 = -3;
984      errorptr = cquantize->fserrors + (width+1)*3; /* => entry after last column */
985      cquantize->on_odd_row = FALSE; /* flip for next time */
986    } else {
987      /* work left to right in this row */
988      dir = 1;
989      dir3 = 3;
990      errorptr = cquantize->fserrors; /* => entry before first real column */
991      cquantize->on_odd_row = TRUE; /* flip for next time */
992    }
993    /* Preset error values: no error propagated to first pixel from left */
994    cur0 = cur1 = cur2 = 0;
995    /* and no error propagated to row below yet */
996    belowerr0 = belowerr1 = belowerr2 = 0;
997    bpreverr0 = bpreverr1 = bpreverr2 = 0;
998
999    for (col = width; col > 0; col--) {
1000      /* curN holds the error propagated from the previous pixel on the
1001       * current line.  Add the error propagated from the previous line
1002       * to form the complete error correction term for this pixel, and
1003       * round the error term (which is expressed * 16) to an integer.
1004       * RIGHT_SHIFT rounds towards minus infinity, so adding 8 is correct
1005       * for either sign of the error value.
1006       * Note: errorptr points to *previous* column's array entry.
1007       */
1008      cur0 = RIGHT_SHIFT(cur0 + errorptr[dir3+0] + 8, 4);
1009      cur1 = RIGHT_SHIFT(cur1 + errorptr[dir3+1] + 8, 4);
1010      cur2 = RIGHT_SHIFT(cur2 + errorptr[dir3+2] + 8, 4);
1011      /* Limit the error using transfer function set by init_error_limit.
1012       * See comments with init_error_limit for rationale.
1013       */
1014      cur0 = error_limit[cur0];
1015      cur1 = error_limit[cur1];
1016      cur2 = error_limit[cur2];
1017      /* Form pixel value + error, and range-limit to 0..MAXJSAMPLE.
1018       * The maximum error is +- MAXJSAMPLE (or less with error limiting);
1019       * this sets the required size of the range_limit array.
1020       */
1021      cur0 += GETJSAMPLE(inptr[0]);
1022      cur1 += GETJSAMPLE(inptr[1]);
1023      cur2 += GETJSAMPLE(inptr[2]);
1024      cur0 = GETJSAMPLE(range_limit[cur0]);
1025      cur1 = GETJSAMPLE(range_limit[cur1]);
1026      cur2 = GETJSAMPLE(range_limit[cur2]);
1027      /* Index into the cache with adjusted pixel value */
1028      cachep = & histogram[cur0>>C0_SHIFT][cur1>>C1_SHIFT][cur2>>C2_SHIFT];
1029      /* If we have not seen this color before, find nearest colormap */
1030      /* entry and update the cache */
1031      if (*cachep == 0)
1032        fill_inverse_cmap(cinfo, cur0>>C0_SHIFT,cur1>>C1_SHIFT,cur2>>C2_SHIFT);
1033      /* Now emit the colormap index for this cell */
1034      { register int pixcode = *cachep - 1;
1035        *outptr = (JSAMPLE) pixcode;
1036        /* Compute representation error for this pixel */
1037        cur0 -= GETJSAMPLE(colormap0[pixcode]);
1038        cur1 -= GETJSAMPLE(colormap1[pixcode]);
1039        cur2 -= GETJSAMPLE(colormap2[pixcode]);
1040      }
1041      /* Compute error fractions to be propagated to adjacent pixels.
1042       * Add these into the running sums, and simultaneously shift the
1043       * next-line error sums left by 1 column.
1044       */
1045      { register LOCFSERROR bnexterr, delta;
1046
1047        bnexterr = cur0;        /* Process component 0 */
1048        delta = cur0 * 2;
1049        cur0 += delta;          /* form error * 3 */
1050        errorptr[0] = (FSERROR) (bpreverr0 + cur0);
1051        cur0 += delta;          /* form error * 5 */
1052        bpreverr0 = belowerr0 + cur0;
1053        belowerr0 = bnexterr;
1054        cur0 += delta;          /* form error * 7 */
1055        bnexterr = cur1;        /* Process component 1 */
1056        delta = cur1 * 2;
1057        cur1 += delta;          /* form error * 3 */
1058        errorptr[1] = (FSERROR) (bpreverr1 + cur1);
1059        cur1 += delta;          /* form error * 5 */
1060        bpreverr1 = belowerr1 + cur1;
1061        belowerr1 = bnexterr;
1062        cur1 += delta;          /* form error * 7 */
1063        bnexterr = cur2;        /* Process component 2 */
1064        delta = cur2 * 2;
1065        cur2 += delta;          /* form error * 3 */
1066        errorptr[2] = (FSERROR) (bpreverr2 + cur2);
1067        cur2 += delta;          /* form error * 5 */
1068        bpreverr2 = belowerr2 + cur2;
1069        belowerr2 = bnexterr;
1070        cur2 += delta;          /* form error * 7 */
1071      }
1072      /* At this point curN contains the 7/16 error value to be propagated
1073       * to the next pixel on the current line, and all the errors for the
1074       * next line have been shifted over.  We are therefore ready to move on.
1075       */
1076      inptr += dir3;            /* Advance pixel pointers to next column */
1077      outptr += dir;
1078      errorptr += dir3;         /* advance errorptr to current column */
1079    }
1080    /* Post-loop cleanup: we must unload the final error values into the
1081     * final fserrors[] entry.  Note we need not unload belowerrN because
1082     * it is for the dummy column before or after the actual array.
1083     */
1084    errorptr[0] = (FSERROR) bpreverr0; /* unload prev errs into array */
1085    errorptr[1] = (FSERROR) bpreverr1;
1086    errorptr[2] = (FSERROR) bpreverr2;
1087  }
1088}
1089
1090
1091/*
1092 * Initialize the error-limiting transfer function (lookup table).
1093 * The raw F-S error computation can potentially compute error values of up to
1094 * +- MAXJSAMPLE.  But we want the maximum correction applied to a pixel to be
1095 * much less, otherwise obviously wrong pixels will be created.  (Typical
1096 * effects include weird fringes at color-area boundaries, isolated bright
1097 * pixels in a dark area, etc.)  The standard advice for avoiding this problem
1098 * is to ensure that the "corners" of the color cube are allocated as output
1099 * colors; then repeated errors in the same direction cannot cause cascading
1100 * error buildup.  However, that only prevents the error from getting
1101 * completely out of hand; Aaron Giles reports that error limiting improves
1102 * the results even with corner colors allocated.
1103 * A simple clamping of the error values to about +- MAXJSAMPLE/8 works pretty
1104 * well, but the smoother transfer function used below is even better.  Thanks
1105 * to Aaron Giles for this idea.
1106 */
1107
1108LOCAL(void)
1109init_error_limit (j_decompress_ptr cinfo)
1110/* Allocate and fill in the error_limiter table */
1111{
1112  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
1113  int * table;
1114  int in, out;
1115
1116  table = (int *) (*cinfo->mem->alloc_small)
1117    ((j_common_ptr) cinfo, JPOOL_IMAGE, (MAXJSAMPLE*2+1) * SIZEOF(int));
1118  table += MAXJSAMPLE;          /* so can index -MAXJSAMPLE .. +MAXJSAMPLE */
1119  cquantize->error_limiter = table;
1120
1121#define STEPSIZE ((MAXJSAMPLE+1)/16)
1122  /* Map errors 1:1 up to +- MAXJSAMPLE/16 */
1123  out = 0;
1124  for (in = 0; in < STEPSIZE; in++, out++) {
1125    table[in] = out; table[-in] = -out;
1126  }
1127  /* Map errors 1:2 up to +- 3*MAXJSAMPLE/16 */
1128  for (; in < STEPSIZE*3; in++, out += (in&1) ? 0 : 1) {
1129    table[in] = out; table[-in] = -out;
1130  }
1131  /* Clamp the rest to final out value (which is (MAXJSAMPLE+1)/8) */
1132  for (; in <= MAXJSAMPLE; in++) {
1133    table[in] = out; table[-in] = -out;
1134  }
1135#undef STEPSIZE
1136}
1137
1138
1139/*
1140 * Finish up at the end of each pass.
1141 */
1142
1143METHODDEF(void)
1144finish_pass1 (j_decompress_ptr cinfo)
1145{
1146  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
1147
1148  /* Select the representative colors and fill in cinfo->colormap */
1149  cinfo->colormap = cquantize->sv_colormap;
1150  select_colors(cinfo, cquantize->desired);
1151  /* Force next pass to zero the color index table */
1152  cquantize->needs_zeroed = TRUE;
1153}
1154
1155
1156METHODDEF(void)
1157finish_pass2 (j_decompress_ptr cinfo)
1158{
1159  /* no work */
1160}
1161
1162
1163/*
1164 * Initialize for each processing pass.
1165 */
1166
1167METHODDEF(void)
1168start_pass_2_quant (j_decompress_ptr cinfo, boolean is_pre_scan)
1169{
1170  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
1171  hist3d histogram = cquantize->histogram;
1172  int i;
1173
1174  /* Only F-S dithering or no dithering is supported. */
1175  /* If user asks for ordered dither, give him F-S. */
1176  if (cinfo->dither_mode != JDITHER_NONE)
1177    cinfo->dither_mode = JDITHER_FS;
1178
1179  if (is_pre_scan) {
1180    /* Set up method pointers */
1181    cquantize->pub.color_quantize = prescan_quantize;
1182    cquantize->pub.finish_pass = finish_pass1;
1183    cquantize->needs_zeroed = TRUE; /* Always zero histogram */
1184  } else {
1185    /* Set up method pointers */
1186    if (cinfo->dither_mode == JDITHER_FS)
1187      cquantize->pub.color_quantize = pass2_fs_dither;
1188    else
1189      cquantize->pub.color_quantize = pass2_no_dither;
1190    cquantize->pub.finish_pass = finish_pass2;
1191
1192    /* Make sure color count is acceptable */
1193    i = cinfo->actual_number_of_colors;
1194    if (i < 1)
1195      ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 1);
1196    if (i > MAXNUMCOLORS)
1197      ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
1198
1199    if (cinfo->dither_mode == JDITHER_FS) {
1200      size_t arraysize = (size_t) ((cinfo->output_width + 2) *
1201                                   (3 * SIZEOF(FSERROR)));
1202      /* Allocate Floyd-Steinberg workspace if we didn't already. */
1203      if (cquantize->fserrors == NULL)
1204        cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
1205          ((j_common_ptr) cinfo, JPOOL_IMAGE, arraysize);
1206      /* Initialize the propagated errors to zero. */
1207      FMEMZERO((void FAR *) cquantize->fserrors, arraysize);
1208      /* Make the error-limit table if we didn't already. */
1209      if (cquantize->error_limiter == NULL)
1210        init_error_limit(cinfo);
1211      cquantize->on_odd_row = FALSE;
1212    }
1213
1214  }
1215  /* Zero the histogram or inverse color map, if necessary */
1216  if (cquantize->needs_zeroed) {
1217    for (i = 0; i < HIST_C0_ELEMS; i++) {
1218      FMEMZERO((void FAR *) histogram[i],
1219               HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
1220    }
1221    cquantize->needs_zeroed = FALSE;
1222  }
1223}
1224
1225
1226/*
1227 * Switch to a new external colormap between output passes.
1228 */
1229
1230METHODDEF(void)
1231new_color_map_2_quant (j_decompress_ptr cinfo)
1232{
1233  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
1234
1235  /* Reset the inverse color map */
1236  cquantize->needs_zeroed = TRUE;
1237}
1238
1239
1240/*
1241 * Module initialization routine for 2-pass color quantization.
1242 */
1243
1244GLOBAL(void)
1245jinit_2pass_quantizer (j_decompress_ptr cinfo)
1246{
1247  my_cquantize_ptr cquantize;
1248  int i;
1249
1250  cquantize = (my_cquantize_ptr)
1251    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
1252                                SIZEOF(my_cquantizer));
1253  cinfo->cquantize = (struct jpeg_color_quantizer *) cquantize;
1254  cquantize->pub.start_pass = start_pass_2_quant;
1255  cquantize->pub.new_color_map = new_color_map_2_quant;
1256  cquantize->fserrors = NULL;   /* flag optional arrays not allocated */
1257  cquantize->error_limiter = NULL;
1258
1259  /* Make sure jdmaster didn't give me a case I can't handle */
1260  if (cinfo->out_color_components != 3)
1261    ERREXIT(cinfo, JERR_NOTIMPL);
1262
1263  /* Allocate the histogram/inverse colormap storage */
1264  cquantize->histogram = (hist3d) (*cinfo->mem->alloc_small)
1265    ((j_common_ptr) cinfo, JPOOL_IMAGE, HIST_C0_ELEMS * SIZEOF(hist2d));
1266  for (i = 0; i < HIST_C0_ELEMS; i++) {
1267    cquantize->histogram[i] = (hist2d) (*cinfo->mem->alloc_large)
1268      ((j_common_ptr) cinfo, JPOOL_IMAGE,
1269       HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
1270  }
1271  cquantize->needs_zeroed = TRUE; /* histogram is garbage now */
1272
1273  /* Allocate storage for the completed colormap, if required.
1274   * We do this now since it is FAR storage and may affect
1275   * the memory manager's space calculations.
1276   */
1277  if (cinfo->enable_2pass_quant) {
1278    /* Make sure color count is acceptable */
1279    int desired = cinfo->desired_number_of_colors;
1280    /* Lower bound on # of colors ... somewhat arbitrary as long as > 0 */
1281    if (desired < 8)
1282      ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 8);
1283    /* Make sure colormap indexes can be represented by JSAMPLEs */
1284    if (desired > MAXNUMCOLORS)
1285      ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
1286    cquantize->sv_colormap = (*cinfo->mem->alloc_sarray)
1287      ((j_common_ptr) cinfo,JPOOL_IMAGE, (JDIMENSION) desired, (JDIMENSION) 3);
1288    cquantize->desired = desired;
1289  } else
1290    cquantize->sv_colormap = NULL;
1291
1292  /* Only F-S dithering or no dithering is supported. */
1293  /* If user asks for ordered dither, give him F-S. */
1294  if (cinfo->dither_mode != JDITHER_NONE)
1295    cinfo->dither_mode = JDITHER_FS;
1296
1297  /* Allocate Floyd-Steinberg workspace if necessary.
1298   * This isn't really needed until pass 2, but again it is FAR storage.
1299   * Although we will cope with a later change in dither_mode,
1300   * we do not promise to honor max_memory_to_use if dither_mode changes.
1301   */
1302  if (cinfo->dither_mode == JDITHER_FS) {
1303    cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
1304      ((j_common_ptr) cinfo, JPOOL_IMAGE,
1305       (size_t) ((cinfo->output_width + 2) * (3 * SIZEOF(FSERROR))));
1306    /* Might as well create the error-limiting table too. */
1307    init_error_limit(cinfo);
1308  }
1309}
1310
1311#endif /* QUANT_2PASS_SUPPORTED */
Note: See TracBrowser for help on using the repository browser.