vrshoot

diff libs/libjpeg/jquant2.c @ 0:b2f14e535253

initial commit
author John Tsiombikas <nuclear@member.fsf.org>
date Sat, 01 Feb 2014 19:58:19 +0200
parents
children
line diff
     1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/libs/libjpeg/jquant2.c	Sat Feb 01 19:58:19 2014 +0200
     1.3 @@ -0,0 +1,1310 @@
     1.4 +/*
     1.5 + * jquant2.c
     1.6 + *
     1.7 + * Copyright (C) 1991-1996, Thomas G. Lane.
     1.8 + * This file is part of the Independent JPEG Group's software.
     1.9 + * For conditions of distribution and use, see the accompanying README file.
    1.10 + *
    1.11 + * This file contains 2-pass color quantization (color mapping) routines.
    1.12 + * These routines provide selection of a custom color map for an image,
    1.13 + * followed by mapping of the image to that color map, with optional
    1.14 + * Floyd-Steinberg dithering.
    1.15 + * It is also possible to use just the second pass to map to an arbitrary
    1.16 + * externally-given color map.
    1.17 + *
    1.18 + * Note: ordered dithering is not supported, since there isn't any fast
    1.19 + * way to compute intercolor distances; it's unclear that ordered dither's
    1.20 + * fundamental assumptions even hold with an irregularly spaced color map.
    1.21 + */
    1.22 +
    1.23 +#define JPEG_INTERNALS
    1.24 +#include "jinclude.h"
    1.25 +#include "jpeglib.h"
    1.26 +
    1.27 +#ifdef QUANT_2PASS_SUPPORTED
    1.28 +
    1.29 +
    1.30 +/*
    1.31 + * This module implements the well-known Heckbert paradigm for color
    1.32 + * quantization.  Most of the ideas used here can be traced back to
    1.33 + * Heckbert's seminal paper
    1.34 + *   Heckbert, Paul.  "Color Image Quantization for Frame Buffer Display",
    1.35 + *   Proc. SIGGRAPH '82, Computer Graphics v.16 #3 (July 1982), pp 297-304.
    1.36 + *
    1.37 + * In the first pass over the image, we accumulate a histogram showing the
    1.38 + * usage count of each possible color.  To keep the histogram to a reasonable
    1.39 + * size, we reduce the precision of the input; typical practice is to retain
    1.40 + * 5 or 6 bits per color, so that 8 or 4 different input values are counted
    1.41 + * in the same histogram cell.
    1.42 + *
    1.43 + * Next, the color-selection step begins with a box representing the whole
    1.44 + * color space, and repeatedly splits the "largest" remaining box until we
    1.45 + * have as many boxes as desired colors.  Then the mean color in each
    1.46 + * remaining box becomes one of the possible output colors.
    1.47 + * 
    1.48 + * The second pass over the image maps each input pixel to the closest output
    1.49 + * color (optionally after applying a Floyd-Steinberg dithering correction).
    1.50 + * This mapping is logically trivial, but making it go fast enough requires
    1.51 + * considerable care.
    1.52 + *
    1.53 + * Heckbert-style quantizers vary a good deal in their policies for choosing
    1.54 + * the "largest" box and deciding where to cut it.  The particular policies
    1.55 + * used here have proved out well in experimental comparisons, but better ones
    1.56 + * may yet be found.
    1.57 + *
    1.58 + * In earlier versions of the IJG code, this module quantized in YCbCr color
    1.59 + * space, processing the raw upsampled data without a color conversion step.
    1.60 + * This allowed the color conversion math to be done only once per colormap
    1.61 + * entry, not once per pixel.  However, that optimization precluded other
    1.62 + * useful optimizations (such as merging color conversion with upsampling)
    1.63 + * and it also interfered with desired capabilities such as quantizing to an
    1.64 + * externally-supplied colormap.  We have therefore abandoned that approach.
    1.65 + * The present code works in the post-conversion color space, typically RGB.
    1.66 + *
    1.67 + * To improve the visual quality of the results, we actually work in scaled
    1.68 + * RGB space, giving G distances more weight than R, and R in turn more than
    1.69 + * B.  To do everything in integer math, we must use integer scale factors.
    1.70 + * The 2/3/1 scale factors used here correspond loosely to the relative
    1.71 + * weights of the colors in the NTSC grayscale equation.
    1.72 + * If you want to use this code to quantize a non-RGB color space, you'll
    1.73 + * probably need to change these scale factors.
    1.74 + */
    1.75 +
    1.76 +#define R_SCALE 2		/* scale R distances by this much */
    1.77 +#define G_SCALE 3		/* scale G distances by this much */
    1.78 +#define B_SCALE 1		/* and B by this much */
    1.79 +
    1.80 +/* Relabel R/G/B as components 0/1/2, respecting the RGB ordering defined
    1.81 + * in jmorecfg.h.  As the code stands, it will do the right thing for R,G,B
    1.82 + * and B,G,R orders.  If you define some other weird order in jmorecfg.h,
    1.83 + * you'll get compile errors until you extend this logic.  In that case
    1.84 + * you'll probably want to tweak the histogram sizes too.
    1.85 + */
    1.86 +
    1.87 +#if RGB_RED == 0
    1.88 +#define C0_SCALE R_SCALE
    1.89 +#endif
    1.90 +#if RGB_BLUE == 0
    1.91 +#define C0_SCALE B_SCALE
    1.92 +#endif
    1.93 +#if RGB_GREEN == 1
    1.94 +#define C1_SCALE G_SCALE
    1.95 +#endif
    1.96 +#if RGB_RED == 2
    1.97 +#define C2_SCALE R_SCALE
    1.98 +#endif
    1.99 +#if RGB_BLUE == 2
   1.100 +#define C2_SCALE B_SCALE
   1.101 +#endif
   1.102 +
   1.103 +
   1.104 +/*
   1.105 + * First we have the histogram data structure and routines for creating it.
   1.106 + *
   1.107 + * The number of bits of precision can be adjusted by changing these symbols.
   1.108 + * We recommend keeping 6 bits for G and 5 each for R and B.
   1.109 + * If you have plenty of memory and cycles, 6 bits all around gives marginally
   1.110 + * better results; if you are short of memory, 5 bits all around will save
   1.111 + * some space but degrade the results.
   1.112 + * To maintain a fully accurate histogram, we'd need to allocate a "long"
   1.113 + * (preferably unsigned long) for each cell.  In practice this is overkill;
   1.114 + * we can get by with 16 bits per cell.  Few of the cell counts will overflow,
   1.115 + * and clamping those that do overflow to the maximum value will give close-
   1.116 + * enough results.  This reduces the recommended histogram size from 256Kb
   1.117 + * to 128Kb, which is a useful savings on PC-class machines.
   1.118 + * (In the second pass the histogram space is re-used for pixel mapping data;
   1.119 + * in that capacity, each cell must be able to store zero to the number of
   1.120 + * desired colors.  16 bits/cell is plenty for that too.)
   1.121 + * Since the JPEG code is intended to run in small memory model on 80x86
   1.122 + * machines, we can't just allocate the histogram in one chunk.  Instead
   1.123 + * of a true 3-D array, we use a row of pointers to 2-D arrays.  Each
   1.124 + * pointer corresponds to a C0 value (typically 2^5 = 32 pointers) and
   1.125 + * each 2-D array has 2^6*2^5 = 2048 or 2^6*2^6 = 4096 entries.  Note that
   1.126 + * on 80x86 machines, the pointer row is in near memory but the actual
   1.127 + * arrays are in far memory (same arrangement as we use for image arrays).
   1.128 + */
   1.129 +
   1.130 +#define MAXNUMCOLORS  (MAXJSAMPLE+1) /* maximum size of colormap */
   1.131 +
   1.132 +/* These will do the right thing for either R,G,B or B,G,R color order,
   1.133 + * but you may not like the results for other color orders.
   1.134 + */
   1.135 +#define HIST_C0_BITS  5		/* bits of precision in R/B histogram */
   1.136 +#define HIST_C1_BITS  6		/* bits of precision in G histogram */
   1.137 +#define HIST_C2_BITS  5		/* bits of precision in B/R histogram */
   1.138 +
   1.139 +/* Number of elements along histogram axes. */
   1.140 +#define HIST_C0_ELEMS  (1<<HIST_C0_BITS)
   1.141 +#define HIST_C1_ELEMS  (1<<HIST_C1_BITS)
   1.142 +#define HIST_C2_ELEMS  (1<<HIST_C2_BITS)
   1.143 +
   1.144 +/* These are the amounts to shift an input value to get a histogram index. */
   1.145 +#define C0_SHIFT  (BITS_IN_JSAMPLE-HIST_C0_BITS)
   1.146 +#define C1_SHIFT  (BITS_IN_JSAMPLE-HIST_C1_BITS)
   1.147 +#define C2_SHIFT  (BITS_IN_JSAMPLE-HIST_C2_BITS)
   1.148 +
   1.149 +
   1.150 +typedef UINT16 histcell;	/* histogram cell; prefer an unsigned type */
   1.151 +
   1.152 +typedef histcell FAR * histptr;	/* for pointers to histogram cells */
   1.153 +
   1.154 +typedef histcell hist1d[HIST_C2_ELEMS]; /* typedefs for the array */
   1.155 +typedef hist1d FAR * hist2d;	/* type for the 2nd-level pointers */
   1.156 +typedef hist2d * hist3d;	/* type for top-level pointer */
   1.157 +
   1.158 +
   1.159 +/* Declarations for Floyd-Steinberg dithering.
   1.160 + *
   1.161 + * Errors are accumulated into the array fserrors[], at a resolution of
   1.162 + * 1/16th of a pixel count.  The error at a given pixel is propagated
   1.163 + * to its not-yet-processed neighbors using the standard F-S fractions,
   1.164 + *		...	(here)	7/16
   1.165 + *		3/16	5/16	1/16
   1.166 + * We work left-to-right on even rows, right-to-left on odd rows.
   1.167 + *
   1.168 + * We can get away with a single array (holding one row's worth of errors)
   1.169 + * by using it to store the current row's errors at pixel columns not yet
   1.170 + * processed, but the next row's errors at columns already processed.  We
   1.171 + * need only a few extra variables to hold the errors immediately around the
   1.172 + * current column.  (If we are lucky, those variables are in registers, but
   1.173 + * even if not, they're probably cheaper to access than array elements are.)
   1.174 + *
   1.175 + * The fserrors[] array has (#columns + 2) entries; the extra entry at
   1.176 + * each end saves us from special-casing the first and last pixels.
   1.177 + * Each entry is three values long, one value for each color component.
   1.178 + *
   1.179 + * Note: on a wide image, we might not have enough room in a PC's near data
   1.180 + * segment to hold the error array; so it is allocated with alloc_large.
   1.181 + */
   1.182 +
   1.183 +#if BITS_IN_JSAMPLE == 8
   1.184 +typedef INT16 FSERROR;		/* 16 bits should be enough */
   1.185 +typedef int LOCFSERROR;		/* use 'int' for calculation temps */
   1.186 +#else
   1.187 +typedef INT32 FSERROR;		/* may need more than 16 bits */
   1.188 +typedef INT32 LOCFSERROR;	/* be sure calculation temps are big enough */
   1.189 +#endif
   1.190 +
   1.191 +typedef FSERROR FAR *FSERRPTR;	/* pointer to error array (in FAR storage!) */
   1.192 +
   1.193 +
   1.194 +/* Private subobject */
   1.195 +
   1.196 +typedef struct {
   1.197 +  struct jpeg_color_quantizer pub; /* public fields */
   1.198 +
   1.199 +  /* Space for the eventually created colormap is stashed here */
   1.200 +  JSAMPARRAY sv_colormap;	/* colormap allocated at init time */
   1.201 +  int desired;			/* desired # of colors = size of colormap */
   1.202 +
   1.203 +  /* Variables for accumulating image statistics */
   1.204 +  hist3d histogram;		/* pointer to the histogram */
   1.205 +
   1.206 +  boolean needs_zeroed;		/* TRUE if next pass must zero histogram */
   1.207 +
   1.208 +  /* Variables for Floyd-Steinberg dithering */
   1.209 +  FSERRPTR fserrors;		/* accumulated errors */
   1.210 +  boolean on_odd_row;		/* flag to remember which row we are on */
   1.211 +  int * error_limiter;		/* table for clamping the applied error */
   1.212 +} my_cquantizer;
   1.213 +
   1.214 +typedef my_cquantizer * my_cquantize_ptr;
   1.215 +
   1.216 +
   1.217 +/*
   1.218 + * Prescan some rows of pixels.
   1.219 + * In this module the prescan simply updates the histogram, which has been
   1.220 + * initialized to zeroes by start_pass.
   1.221 + * An output_buf parameter is required by the method signature, but no data
   1.222 + * is actually output (in fact the buffer controller is probably passing a
   1.223 + * NULL pointer).
   1.224 + */
   1.225 +
   1.226 +METHODDEF(void)
   1.227 +prescan_quantize (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
   1.228 +		  JSAMPARRAY output_buf, int num_rows)
   1.229 +{
   1.230 +  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
   1.231 +  register JSAMPROW ptr;
   1.232 +  register histptr histp;
   1.233 +  register hist3d histogram = cquantize->histogram;
   1.234 +  int row;
   1.235 +  JDIMENSION col;
   1.236 +  JDIMENSION width = cinfo->output_width;
   1.237 +
   1.238 +  for (row = 0; row < num_rows; row++) {
   1.239 +    ptr = input_buf[row];
   1.240 +    for (col = width; col > 0; col--) {
   1.241 +      /* get pixel value and index into the histogram */
   1.242 +      histp = & histogram[GETJSAMPLE(ptr[0]) >> C0_SHIFT]
   1.243 +			 [GETJSAMPLE(ptr[1]) >> C1_SHIFT]
   1.244 +			 [GETJSAMPLE(ptr[2]) >> C2_SHIFT];
   1.245 +      /* increment, check for overflow and undo increment if so. */
   1.246 +      if (++(*histp) <= 0)
   1.247 +	(*histp)--;
   1.248 +      ptr += 3;
   1.249 +    }
   1.250 +  }
   1.251 +}
   1.252 +
   1.253 +
   1.254 +/*
   1.255 + * Next we have the really interesting routines: selection of a colormap
   1.256 + * given the completed histogram.
   1.257 + * These routines work with a list of "boxes", each representing a rectangular
   1.258 + * subset of the input color space (to histogram precision).
   1.259 + */
   1.260 +
   1.261 +typedef struct {
   1.262 +  /* The bounds of the box (inclusive); expressed as histogram indexes */
   1.263 +  int c0min, c0max;
   1.264 +  int c1min, c1max;
   1.265 +  int c2min, c2max;
   1.266 +  /* The volume (actually 2-norm) of the box */
   1.267 +  INT32 volume;
   1.268 +  /* The number of nonzero histogram cells within this box */
   1.269 +  long colorcount;
   1.270 +} box;
   1.271 +
   1.272 +typedef box * boxptr;
   1.273 +
   1.274 +
   1.275 +LOCAL(boxptr)
   1.276 +find_biggest_color_pop (boxptr boxlist, int numboxes)
   1.277 +/* Find the splittable box with the largest color population */
   1.278 +/* Returns NULL if no splittable boxes remain */
   1.279 +{
   1.280 +  register boxptr boxp;
   1.281 +  register int i;
   1.282 +  register long maxc = 0;
   1.283 +  boxptr which = NULL;
   1.284 +  
   1.285 +  for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
   1.286 +    if (boxp->colorcount > maxc && boxp->volume > 0) {
   1.287 +      which = boxp;
   1.288 +      maxc = boxp->colorcount;
   1.289 +    }
   1.290 +  }
   1.291 +  return which;
   1.292 +}
   1.293 +
   1.294 +
   1.295 +LOCAL(boxptr)
   1.296 +find_biggest_volume (boxptr boxlist, int numboxes)
   1.297 +/* Find the splittable box with the largest (scaled) volume */
   1.298 +/* Returns NULL if no splittable boxes remain */
   1.299 +{
   1.300 +  register boxptr boxp;
   1.301 +  register int i;
   1.302 +  register INT32 maxv = 0;
   1.303 +  boxptr which = NULL;
   1.304 +  
   1.305 +  for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
   1.306 +    if (boxp->volume > maxv) {
   1.307 +      which = boxp;
   1.308 +      maxv = boxp->volume;
   1.309 +    }
   1.310 +  }
   1.311 +  return which;
   1.312 +}
   1.313 +
   1.314 +
   1.315 +LOCAL(void)
   1.316 +update_box (j_decompress_ptr cinfo, boxptr boxp)
   1.317 +/* Shrink the min/max bounds of a box to enclose only nonzero elements, */
   1.318 +/* and recompute its volume and population */
   1.319 +{
   1.320 +  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
   1.321 +  hist3d histogram = cquantize->histogram;
   1.322 +  histptr histp;
   1.323 +  int c0,c1,c2;
   1.324 +  int c0min,c0max,c1min,c1max,c2min,c2max;
   1.325 +  INT32 dist0,dist1,dist2;
   1.326 +  long ccount;
   1.327 +  
   1.328 +  c0min = boxp->c0min;  c0max = boxp->c0max;
   1.329 +  c1min = boxp->c1min;  c1max = boxp->c1max;
   1.330 +  c2min = boxp->c2min;  c2max = boxp->c2max;
   1.331 +  
   1.332 +  if (c0max > c0min)
   1.333 +    for (c0 = c0min; c0 <= c0max; c0++)
   1.334 +      for (c1 = c1min; c1 <= c1max; c1++) {
   1.335 +	histp = & histogram[c0][c1][c2min];
   1.336 +	for (c2 = c2min; c2 <= c2max; c2++)
   1.337 +	  if (*histp++ != 0) {
   1.338 +	    boxp->c0min = c0min = c0;
   1.339 +	    goto have_c0min;
   1.340 +	  }
   1.341 +      }
   1.342 + have_c0min:
   1.343 +  if (c0max > c0min)
   1.344 +    for (c0 = c0max; c0 >= c0min; c0--)
   1.345 +      for (c1 = c1min; c1 <= c1max; c1++) {
   1.346 +	histp = & histogram[c0][c1][c2min];
   1.347 +	for (c2 = c2min; c2 <= c2max; c2++)
   1.348 +	  if (*histp++ != 0) {
   1.349 +	    boxp->c0max = c0max = c0;
   1.350 +	    goto have_c0max;
   1.351 +	  }
   1.352 +      }
   1.353 + have_c0max:
   1.354 +  if (c1max > c1min)
   1.355 +    for (c1 = c1min; c1 <= c1max; c1++)
   1.356 +      for (c0 = c0min; c0 <= c0max; c0++) {
   1.357 +	histp = & histogram[c0][c1][c2min];
   1.358 +	for (c2 = c2min; c2 <= c2max; c2++)
   1.359 +	  if (*histp++ != 0) {
   1.360 +	    boxp->c1min = c1min = c1;
   1.361 +	    goto have_c1min;
   1.362 +	  }
   1.363 +      }
   1.364 + have_c1min:
   1.365 +  if (c1max > c1min)
   1.366 +    for (c1 = c1max; c1 >= c1min; c1--)
   1.367 +      for (c0 = c0min; c0 <= c0max; c0++) {
   1.368 +	histp = & histogram[c0][c1][c2min];
   1.369 +	for (c2 = c2min; c2 <= c2max; c2++)
   1.370 +	  if (*histp++ != 0) {
   1.371 +	    boxp->c1max = c1max = c1;
   1.372 +	    goto have_c1max;
   1.373 +	  }
   1.374 +      }
   1.375 + have_c1max:
   1.376 +  if (c2max > c2min)
   1.377 +    for (c2 = c2min; c2 <= c2max; c2++)
   1.378 +      for (c0 = c0min; c0 <= c0max; c0++) {
   1.379 +	histp = & histogram[c0][c1min][c2];
   1.380 +	for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
   1.381 +	  if (*histp != 0) {
   1.382 +	    boxp->c2min = c2min = c2;
   1.383 +	    goto have_c2min;
   1.384 +	  }
   1.385 +      }
   1.386 + have_c2min:
   1.387 +  if (c2max > c2min)
   1.388 +    for (c2 = c2max; c2 >= c2min; c2--)
   1.389 +      for (c0 = c0min; c0 <= c0max; c0++) {
   1.390 +	histp = & histogram[c0][c1min][c2];
   1.391 +	for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
   1.392 +	  if (*histp != 0) {
   1.393 +	    boxp->c2max = c2max = c2;
   1.394 +	    goto have_c2max;
   1.395 +	  }
   1.396 +      }
   1.397 + have_c2max:
   1.398 +
   1.399 +  /* Update box volume.
   1.400 +   * We use 2-norm rather than real volume here; this biases the method
   1.401 +   * against making long narrow boxes, and it has the side benefit that
   1.402 +   * a box is splittable iff norm > 0.
   1.403 +   * Since the differences are expressed in histogram-cell units,
   1.404 +   * we have to shift back to JSAMPLE units to get consistent distances;
   1.405 +   * after which, we scale according to the selected distance scale factors.
   1.406 +   */
   1.407 +  dist0 = ((c0max - c0min) << C0_SHIFT) * C0_SCALE;
   1.408 +  dist1 = ((c1max - c1min) << C1_SHIFT) * C1_SCALE;
   1.409 +  dist2 = ((c2max - c2min) << C2_SHIFT) * C2_SCALE;
   1.410 +  boxp->volume = dist0*dist0 + dist1*dist1 + dist2*dist2;
   1.411 +  
   1.412 +  /* Now scan remaining volume of box and compute population */
   1.413 +  ccount = 0;
   1.414 +  for (c0 = c0min; c0 <= c0max; c0++)
   1.415 +    for (c1 = c1min; c1 <= c1max; c1++) {
   1.416 +      histp = & histogram[c0][c1][c2min];
   1.417 +      for (c2 = c2min; c2 <= c2max; c2++, histp++)
   1.418 +	if (*histp != 0) {
   1.419 +	  ccount++;
   1.420 +	}
   1.421 +    }
   1.422 +  boxp->colorcount = ccount;
   1.423 +}
   1.424 +
   1.425 +
   1.426 +LOCAL(int)
   1.427 +median_cut (j_decompress_ptr cinfo, boxptr boxlist, int numboxes,
   1.428 +	    int desired_colors)
   1.429 +/* Repeatedly select and split the largest box until we have enough boxes */
   1.430 +{
   1.431 +  int n,lb;
   1.432 +  int c0,c1,c2,cmax;
   1.433 +  register boxptr b1,b2;
   1.434 +
   1.435 +  while (numboxes < desired_colors) {
   1.436 +    /* Select box to split.
   1.437 +     * Current algorithm: by population for first half, then by volume.
   1.438 +     */
   1.439 +    if (numboxes*2 <= desired_colors) {
   1.440 +      b1 = find_biggest_color_pop(boxlist, numboxes);
   1.441 +    } else {
   1.442 +      b1 = find_biggest_volume(boxlist, numboxes);
   1.443 +    }
   1.444 +    if (b1 == NULL)		/* no splittable boxes left! */
   1.445 +      break;
   1.446 +    b2 = &boxlist[numboxes];	/* where new box will go */
   1.447 +    /* Copy the color bounds to the new box. */
   1.448 +    b2->c0max = b1->c0max; b2->c1max = b1->c1max; b2->c2max = b1->c2max;
   1.449 +    b2->c0min = b1->c0min; b2->c1min = b1->c1min; b2->c2min = b1->c2min;
   1.450 +    /* Choose which axis to split the box on.
   1.451 +     * Current algorithm: longest scaled axis.
   1.452 +     * See notes in update_box about scaling distances.
   1.453 +     */
   1.454 +    c0 = ((b1->c0max - b1->c0min) << C0_SHIFT) * C0_SCALE;
   1.455 +    c1 = ((b1->c1max - b1->c1min) << C1_SHIFT) * C1_SCALE;
   1.456 +    c2 = ((b1->c2max - b1->c2min) << C2_SHIFT) * C2_SCALE;
   1.457 +    /* We want to break any ties in favor of green, then red, blue last.
   1.458 +     * This code does the right thing for R,G,B or B,G,R color orders only.
   1.459 +     */
   1.460 +#if RGB_RED == 0
   1.461 +    cmax = c1; n = 1;
   1.462 +    if (c0 > cmax) { cmax = c0; n = 0; }
   1.463 +    if (c2 > cmax) { n = 2; }
   1.464 +#else
   1.465 +    cmax = c1; n = 1;
   1.466 +    if (c2 > cmax) { cmax = c2; n = 2; }
   1.467 +    if (c0 > cmax) { n = 0; }
   1.468 +#endif
   1.469 +    /* Choose split point along selected axis, and update box bounds.
   1.470 +     * Current algorithm: split at halfway point.
   1.471 +     * (Since the box has been shrunk to minimum volume,
   1.472 +     * any split will produce two nonempty subboxes.)
   1.473 +     * Note that lb value is max for lower box, so must be < old max.
   1.474 +     */
   1.475 +    switch (n) {
   1.476 +    case 0:
   1.477 +      lb = (b1->c0max + b1->c0min) / 2;
   1.478 +      b1->c0max = lb;
   1.479 +      b2->c0min = lb+1;
   1.480 +      break;
   1.481 +    case 1:
   1.482 +      lb = (b1->c1max + b1->c1min) / 2;
   1.483 +      b1->c1max = lb;
   1.484 +      b2->c1min = lb+1;
   1.485 +      break;
   1.486 +    case 2:
   1.487 +      lb = (b1->c2max + b1->c2min) / 2;
   1.488 +      b1->c2max = lb;
   1.489 +      b2->c2min = lb+1;
   1.490 +      break;
   1.491 +    }
   1.492 +    /* Update stats for boxes */
   1.493 +    update_box(cinfo, b1);
   1.494 +    update_box(cinfo, b2);
   1.495 +    numboxes++;
   1.496 +  }
   1.497 +  return numboxes;
   1.498 +}
   1.499 +
   1.500 +
   1.501 +LOCAL(void)
   1.502 +compute_color (j_decompress_ptr cinfo, boxptr boxp, int icolor)
   1.503 +/* Compute representative color for a box, put it in colormap[icolor] */
   1.504 +{
   1.505 +  /* Current algorithm: mean weighted by pixels (not colors) */
   1.506 +  /* Note it is important to get the rounding correct! */
   1.507 +  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
   1.508 +  hist3d histogram = cquantize->histogram;
   1.509 +  histptr histp;
   1.510 +  int c0,c1,c2;
   1.511 +  int c0min,c0max,c1min,c1max,c2min,c2max;
   1.512 +  long count;
   1.513 +  long total = 0;
   1.514 +  long c0total = 0;
   1.515 +  long c1total = 0;
   1.516 +  long c2total = 0;
   1.517 +  
   1.518 +  c0min = boxp->c0min;  c0max = boxp->c0max;
   1.519 +  c1min = boxp->c1min;  c1max = boxp->c1max;
   1.520 +  c2min = boxp->c2min;  c2max = boxp->c2max;
   1.521 +  
   1.522 +  for (c0 = c0min; c0 <= c0max; c0++)
   1.523 +    for (c1 = c1min; c1 <= c1max; c1++) {
   1.524 +      histp = & histogram[c0][c1][c2min];
   1.525 +      for (c2 = c2min; c2 <= c2max; c2++) {
   1.526 +	if ((count = *histp++) != 0) {
   1.527 +	  total += count;
   1.528 +	  c0total += ((c0 << C0_SHIFT) + ((1<<C0_SHIFT)>>1)) * count;
   1.529 +	  c1total += ((c1 << C1_SHIFT) + ((1<<C1_SHIFT)>>1)) * count;
   1.530 +	  c2total += ((c2 << C2_SHIFT) + ((1<<C2_SHIFT)>>1)) * count;
   1.531 +	}
   1.532 +      }
   1.533 +    }
   1.534 +  
   1.535 +  cinfo->colormap[0][icolor] = (JSAMPLE) ((c0total + (total>>1)) / total);
   1.536 +  cinfo->colormap[1][icolor] = (JSAMPLE) ((c1total + (total>>1)) / total);
   1.537 +  cinfo->colormap[2][icolor] = (JSAMPLE) ((c2total + (total>>1)) / total);
   1.538 +}
   1.539 +
   1.540 +
   1.541 +LOCAL(void)
   1.542 +select_colors (j_decompress_ptr cinfo, int desired_colors)
   1.543 +/* Master routine for color selection */
   1.544 +{
   1.545 +  boxptr boxlist;
   1.546 +  int numboxes;
   1.547 +  int i;
   1.548 +
   1.549 +  /* Allocate workspace for box list */
   1.550 +  boxlist = (boxptr) (*cinfo->mem->alloc_small)
   1.551 +    ((j_common_ptr) cinfo, JPOOL_IMAGE, desired_colors * SIZEOF(box));
   1.552 +  /* Initialize one box containing whole space */
   1.553 +  numboxes = 1;
   1.554 +  boxlist[0].c0min = 0;
   1.555 +  boxlist[0].c0max = MAXJSAMPLE >> C0_SHIFT;
   1.556 +  boxlist[0].c1min = 0;
   1.557 +  boxlist[0].c1max = MAXJSAMPLE >> C1_SHIFT;
   1.558 +  boxlist[0].c2min = 0;
   1.559 +  boxlist[0].c2max = MAXJSAMPLE >> C2_SHIFT;
   1.560 +  /* Shrink it to actually-used volume and set its statistics */
   1.561 +  update_box(cinfo, & boxlist[0]);
   1.562 +  /* Perform median-cut to produce final box list */
   1.563 +  numboxes = median_cut(cinfo, boxlist, numboxes, desired_colors);
   1.564 +  /* Compute the representative color for each box, fill colormap */
   1.565 +  for (i = 0; i < numboxes; i++)
   1.566 +    compute_color(cinfo, & boxlist[i], i);
   1.567 +  cinfo->actual_number_of_colors = numboxes;
   1.568 +  TRACEMS1(cinfo, 1, JTRC_QUANT_SELECTED, numboxes);
   1.569 +}
   1.570 +
   1.571 +
   1.572 +/*
   1.573 + * These routines are concerned with the time-critical task of mapping input
   1.574 + * colors to the nearest color in the selected colormap.
   1.575 + *
   1.576 + * We re-use the histogram space as an "inverse color map", essentially a
   1.577 + * cache for the results of nearest-color searches.  All colors within a
   1.578 + * histogram cell will be mapped to the same colormap entry, namely the one
   1.579 + * closest to the cell's center.  This may not be quite the closest entry to
   1.580 + * the actual input color, but it's almost as good.  A zero in the cache
   1.581 + * indicates we haven't found the nearest color for that cell yet; the array
   1.582 + * is cleared to zeroes before starting the mapping pass.  When we find the
   1.583 + * nearest color for a cell, its colormap index plus one is recorded in the
   1.584 + * cache for future use.  The pass2 scanning routines call fill_inverse_cmap
   1.585 + * when they need to use an unfilled entry in the cache.
   1.586 + *
   1.587 + * Our method of efficiently finding nearest colors is based on the "locally
   1.588 + * sorted search" idea described by Heckbert and on the incremental distance
   1.589 + * calculation described by Spencer W. Thomas in chapter III.1 of Graphics
   1.590 + * Gems II (James Arvo, ed.  Academic Press, 1991).  Thomas points out that
   1.591 + * the distances from a given colormap entry to each cell of the histogram can
   1.592 + * be computed quickly using an incremental method: the differences between
   1.593 + * distances to adjacent cells themselves differ by a constant.  This allows a
   1.594 + * fairly fast implementation of the "brute force" approach of computing the
   1.595 + * distance from every colormap entry to every histogram cell.  Unfortunately,
   1.596 + * it needs a work array to hold the best-distance-so-far for each histogram
   1.597 + * cell (because the inner loop has to be over cells, not colormap entries).
   1.598 + * The work array elements have to be INT32s, so the work array would need
   1.599 + * 256Kb at our recommended precision.  This is not feasible in DOS machines.
   1.600 + *
   1.601 + * To get around these problems, we apply Thomas' method to compute the
   1.602 + * nearest colors for only the cells within a small subbox of the histogram.
   1.603 + * The work array need be only as big as the subbox, so the memory usage
   1.604 + * problem is solved.  Furthermore, we need not fill subboxes that are never
   1.605 + * referenced in pass2; many images use only part of the color gamut, so a
   1.606 + * fair amount of work is saved.  An additional advantage of this
   1.607 + * approach is that we can apply Heckbert's locality criterion to quickly
   1.608 + * eliminate colormap entries that are far away from the subbox; typically
   1.609 + * three-fourths of the colormap entries are rejected by Heckbert's criterion,
   1.610 + * and we need not compute their distances to individual cells in the subbox.
   1.611 + * The speed of this approach is heavily influenced by the subbox size: too
   1.612 + * small means too much overhead, too big loses because Heckbert's criterion
   1.613 + * can't eliminate as many colormap entries.  Empirically the best subbox
   1.614 + * size seems to be about 1/512th of the histogram (1/8th in each direction).
   1.615 + *
   1.616 + * Thomas' article also describes a refined method which is asymptotically
   1.617 + * faster than the brute-force method, but it is also far more complex and
   1.618 + * cannot efficiently be applied to small subboxes.  It is therefore not
   1.619 + * useful for programs intended to be portable to DOS machines.  On machines
   1.620 + * with plenty of memory, filling the whole histogram in one shot with Thomas'
   1.621 + * refined method might be faster than the present code --- but then again,
   1.622 + * it might not be any faster, and it's certainly more complicated.
   1.623 + */
   1.624 +
   1.625 +
   1.626 +/* log2(histogram cells in update box) for each axis; this can be adjusted */
   1.627 +#define BOX_C0_LOG  (HIST_C0_BITS-3)
   1.628 +#define BOX_C1_LOG  (HIST_C1_BITS-3)
   1.629 +#define BOX_C2_LOG  (HIST_C2_BITS-3)
   1.630 +
   1.631 +#define BOX_C0_ELEMS  (1<<BOX_C0_LOG) /* # of hist cells in update box */
   1.632 +#define BOX_C1_ELEMS  (1<<BOX_C1_LOG)
   1.633 +#define BOX_C2_ELEMS  (1<<BOX_C2_LOG)
   1.634 +
   1.635 +#define BOX_C0_SHIFT  (C0_SHIFT + BOX_C0_LOG)
   1.636 +#define BOX_C1_SHIFT  (C1_SHIFT + BOX_C1_LOG)
   1.637 +#define BOX_C2_SHIFT  (C2_SHIFT + BOX_C2_LOG)
   1.638 +
   1.639 +
   1.640 +/*
   1.641 + * The next three routines implement inverse colormap filling.  They could
   1.642 + * all be folded into one big routine, but splitting them up this way saves
   1.643 + * some stack space (the mindist[] and bestdist[] arrays need not coexist)
   1.644 + * and may allow some compilers to produce better code by registerizing more
   1.645 + * inner-loop variables.
   1.646 + */
   1.647 +
   1.648 +LOCAL(int)
   1.649 +find_nearby_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
   1.650 +		    JSAMPLE colorlist[])
   1.651 +/* Locate the colormap entries close enough to an update box to be candidates
   1.652 + * for the nearest entry to some cell(s) in the update box.  The update box
   1.653 + * is specified by the center coordinates of its first cell.  The number of
   1.654 + * candidate colormap entries is returned, and their colormap indexes are
   1.655 + * placed in colorlist[].
   1.656 + * This routine uses Heckbert's "locally sorted search" criterion to select
   1.657 + * the colors that need further consideration.
   1.658 + */
   1.659 +{
   1.660 +  int numcolors = cinfo->actual_number_of_colors;
   1.661 +  int maxc0, maxc1, maxc2;
   1.662 +  int centerc0, centerc1, centerc2;
   1.663 +  int i, x, ncolors;
   1.664 +  INT32 minmaxdist, min_dist, max_dist, tdist;
   1.665 +  INT32 mindist[MAXNUMCOLORS];	/* min distance to colormap entry i */
   1.666 +
   1.667 +  /* Compute true coordinates of update box's upper corner and center.
   1.668 +   * Actually we compute the coordinates of the center of the upper-corner
   1.669 +   * histogram cell, which are the upper bounds of the volume we care about.
   1.670 +   * Note that since ">>" rounds down, the "center" values may be closer to
   1.671 +   * min than to max; hence comparisons to them must be "<=", not "<".
   1.672 +   */
   1.673 +  maxc0 = minc0 + ((1 << BOX_C0_SHIFT) - (1 << C0_SHIFT));
   1.674 +  centerc0 = (minc0 + maxc0) >> 1;
   1.675 +  maxc1 = minc1 + ((1 << BOX_C1_SHIFT) - (1 << C1_SHIFT));
   1.676 +  centerc1 = (minc1 + maxc1) >> 1;
   1.677 +  maxc2 = minc2 + ((1 << BOX_C2_SHIFT) - (1 << C2_SHIFT));
   1.678 +  centerc2 = (minc2 + maxc2) >> 1;
   1.679 +
   1.680 +  /* For each color in colormap, find:
   1.681 +   *  1. its minimum squared-distance to any point in the update box
   1.682 +   *     (zero if color is within update box);
   1.683 +   *  2. its maximum squared-distance to any point in the update box.
   1.684 +   * Both of these can be found by considering only the corners of the box.
   1.685 +   * We save the minimum distance for each color in mindist[];
   1.686 +   * only the smallest maximum distance is of interest.
   1.687 +   */
   1.688 +  minmaxdist = 0x7FFFFFFFL;
   1.689 +
   1.690 +  for (i = 0; i < numcolors; i++) {
   1.691 +    /* We compute the squared-c0-distance term, then add in the other two. */
   1.692 +    x = GETJSAMPLE(cinfo->colormap[0][i]);
   1.693 +    if (x < minc0) {
   1.694 +      tdist = (x - minc0) * C0_SCALE;
   1.695 +      min_dist = tdist*tdist;
   1.696 +      tdist = (x - maxc0) * C0_SCALE;
   1.697 +      max_dist = tdist*tdist;
   1.698 +    } else if (x > maxc0) {
   1.699 +      tdist = (x - maxc0) * C0_SCALE;
   1.700 +      min_dist = tdist*tdist;
   1.701 +      tdist = (x - minc0) * C0_SCALE;
   1.702 +      max_dist = tdist*tdist;
   1.703 +    } else {
   1.704 +      /* within cell range so no contribution to min_dist */
   1.705 +      min_dist = 0;
   1.706 +      if (x <= centerc0) {
   1.707 +	tdist = (x - maxc0) * C0_SCALE;
   1.708 +	max_dist = tdist*tdist;
   1.709 +      } else {
   1.710 +	tdist = (x - minc0) * C0_SCALE;
   1.711 +	max_dist = tdist*tdist;
   1.712 +      }
   1.713 +    }
   1.714 +
   1.715 +    x = GETJSAMPLE(cinfo->colormap[1][i]);
   1.716 +    if (x < minc1) {
   1.717 +      tdist = (x - minc1) * C1_SCALE;
   1.718 +      min_dist += tdist*tdist;
   1.719 +      tdist = (x - maxc1) * C1_SCALE;
   1.720 +      max_dist += tdist*tdist;
   1.721 +    } else if (x > maxc1) {
   1.722 +      tdist = (x - maxc1) * C1_SCALE;
   1.723 +      min_dist += tdist*tdist;
   1.724 +      tdist = (x - minc1) * C1_SCALE;
   1.725 +      max_dist += tdist*tdist;
   1.726 +    } else {
   1.727 +      /* within cell range so no contribution to min_dist */
   1.728 +      if (x <= centerc1) {
   1.729 +	tdist = (x - maxc1) * C1_SCALE;
   1.730 +	max_dist += tdist*tdist;
   1.731 +      } else {
   1.732 +	tdist = (x - minc1) * C1_SCALE;
   1.733 +	max_dist += tdist*tdist;
   1.734 +      }
   1.735 +    }
   1.736 +
   1.737 +    x = GETJSAMPLE(cinfo->colormap[2][i]);
   1.738 +    if (x < minc2) {
   1.739 +      tdist = (x - minc2) * C2_SCALE;
   1.740 +      min_dist += tdist*tdist;
   1.741 +      tdist = (x - maxc2) * C2_SCALE;
   1.742 +      max_dist += tdist*tdist;
   1.743 +    } else if (x > maxc2) {
   1.744 +      tdist = (x - maxc2) * C2_SCALE;
   1.745 +      min_dist += tdist*tdist;
   1.746 +      tdist = (x - minc2) * C2_SCALE;
   1.747 +      max_dist += tdist*tdist;
   1.748 +    } else {
   1.749 +      /* within cell range so no contribution to min_dist */
   1.750 +      if (x <= centerc2) {
   1.751 +	tdist = (x - maxc2) * C2_SCALE;
   1.752 +	max_dist += tdist*tdist;
   1.753 +      } else {
   1.754 +	tdist = (x - minc2) * C2_SCALE;
   1.755 +	max_dist += tdist*tdist;
   1.756 +      }
   1.757 +    }
   1.758 +
   1.759 +    mindist[i] = min_dist;	/* save away the results */
   1.760 +    if (max_dist < minmaxdist)
   1.761 +      minmaxdist = max_dist;
   1.762 +  }
   1.763 +
   1.764 +  /* Now we know that no cell in the update box is more than minmaxdist
   1.765 +   * away from some colormap entry.  Therefore, only colors that are
   1.766 +   * within minmaxdist of some part of the box need be considered.
   1.767 +   */
   1.768 +  ncolors = 0;
   1.769 +  for (i = 0; i < numcolors; i++) {
   1.770 +    if (mindist[i] <= minmaxdist)
   1.771 +      colorlist[ncolors++] = (JSAMPLE) i;
   1.772 +  }
   1.773 +  return ncolors;
   1.774 +}
   1.775 +
   1.776 +
   1.777 +LOCAL(void)
   1.778 +find_best_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
   1.779 +		  int numcolors, JSAMPLE colorlist[], JSAMPLE bestcolor[])
   1.780 +/* Find the closest colormap entry for each cell in the update box,
   1.781 + * given the list of candidate colors prepared by find_nearby_colors.
   1.782 + * Return the indexes of the closest entries in the bestcolor[] array.
   1.783 + * This routine uses Thomas' incremental distance calculation method to
   1.784 + * find the distance from a colormap entry to successive cells in the box.
   1.785 + */
   1.786 +{
   1.787 +  int ic0, ic1, ic2;
   1.788 +  int i, icolor;
   1.789 +  register INT32 * bptr;	/* pointer into bestdist[] array */
   1.790 +  JSAMPLE * cptr;		/* pointer into bestcolor[] array */
   1.791 +  INT32 dist0, dist1;		/* initial distance values */
   1.792 +  register INT32 dist2;		/* current distance in inner loop */
   1.793 +  INT32 xx0, xx1;		/* distance increments */
   1.794 +  register INT32 xx2;
   1.795 +  INT32 inc0, inc1, inc2;	/* initial values for increments */
   1.796 +  /* This array holds the distance to the nearest-so-far color for each cell */
   1.797 +  INT32 bestdist[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
   1.798 +
   1.799 +  /* Initialize best-distance for each cell of the update box */
   1.800 +  bptr = bestdist;
   1.801 +  for (i = BOX_C0_ELEMS*BOX_C1_ELEMS*BOX_C2_ELEMS-1; i >= 0; i--)
   1.802 +    *bptr++ = 0x7FFFFFFFL;
   1.803 +  
   1.804 +  /* For each color selected by find_nearby_colors,
   1.805 +   * compute its distance to the center of each cell in the box.
   1.806 +   * If that's less than best-so-far, update best distance and color number.
   1.807 +   */
   1.808 +  
   1.809 +  /* Nominal steps between cell centers ("x" in Thomas article) */
   1.810 +#define STEP_C0  ((1 << C0_SHIFT) * C0_SCALE)
   1.811 +#define STEP_C1  ((1 << C1_SHIFT) * C1_SCALE)
   1.812 +#define STEP_C2  ((1 << C2_SHIFT) * C2_SCALE)
   1.813 +  
   1.814 +  for (i = 0; i < numcolors; i++) {
   1.815 +    icolor = GETJSAMPLE(colorlist[i]);
   1.816 +    /* Compute (square of) distance from minc0/c1/c2 to this color */
   1.817 +    inc0 = (minc0 - GETJSAMPLE(cinfo->colormap[0][icolor])) * C0_SCALE;
   1.818 +    dist0 = inc0*inc0;
   1.819 +    inc1 = (minc1 - GETJSAMPLE(cinfo->colormap[1][icolor])) * C1_SCALE;
   1.820 +    dist0 += inc1*inc1;
   1.821 +    inc2 = (minc2 - GETJSAMPLE(cinfo->colormap[2][icolor])) * C2_SCALE;
   1.822 +    dist0 += inc2*inc2;
   1.823 +    /* Form the initial difference increments */
   1.824 +    inc0 = inc0 * (2 * STEP_C0) + STEP_C0 * STEP_C0;
   1.825 +    inc1 = inc1 * (2 * STEP_C1) + STEP_C1 * STEP_C1;
   1.826 +    inc2 = inc2 * (2 * STEP_C2) + STEP_C2 * STEP_C2;
   1.827 +    /* Now loop over all cells in box, updating distance per Thomas method */
   1.828 +    bptr = bestdist;
   1.829 +    cptr = bestcolor;
   1.830 +    xx0 = inc0;
   1.831 +    for (ic0 = BOX_C0_ELEMS-1; ic0 >= 0; ic0--) {
   1.832 +      dist1 = dist0;
   1.833 +      xx1 = inc1;
   1.834 +      for (ic1 = BOX_C1_ELEMS-1; ic1 >= 0; ic1--) {
   1.835 +	dist2 = dist1;
   1.836 +	xx2 = inc2;
   1.837 +	for (ic2 = BOX_C2_ELEMS-1; ic2 >= 0; ic2--) {
   1.838 +	  if (dist2 < *bptr) {
   1.839 +	    *bptr = dist2;
   1.840 +	    *cptr = (JSAMPLE) icolor;
   1.841 +	  }
   1.842 +	  dist2 += xx2;
   1.843 +	  xx2 += 2 * STEP_C2 * STEP_C2;
   1.844 +	  bptr++;
   1.845 +	  cptr++;
   1.846 +	}
   1.847 +	dist1 += xx1;
   1.848 +	xx1 += 2 * STEP_C1 * STEP_C1;
   1.849 +      }
   1.850 +      dist0 += xx0;
   1.851 +      xx0 += 2 * STEP_C0 * STEP_C0;
   1.852 +    }
   1.853 +  }
   1.854 +}
   1.855 +
   1.856 +
   1.857 +LOCAL(void)
   1.858 +fill_inverse_cmap (j_decompress_ptr cinfo, int c0, int c1, int c2)
   1.859 +/* Fill the inverse-colormap entries in the update box that contains */
   1.860 +/* histogram cell c0/c1/c2.  (Only that one cell MUST be filled, but */
   1.861 +/* we can fill as many others as we wish.) */
   1.862 +{
   1.863 +  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
   1.864 +  hist3d histogram = cquantize->histogram;
   1.865 +  int minc0, minc1, minc2;	/* lower left corner of update box */
   1.866 +  int ic0, ic1, ic2;
   1.867 +  register JSAMPLE * cptr;	/* pointer into bestcolor[] array */
   1.868 +  register histptr cachep;	/* pointer into main cache array */
   1.869 +  /* This array lists the candidate colormap indexes. */
   1.870 +  JSAMPLE colorlist[MAXNUMCOLORS];
   1.871 +  int numcolors;		/* number of candidate colors */
   1.872 +  /* This array holds the actually closest colormap index for each cell. */
   1.873 +  JSAMPLE bestcolor[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
   1.874 +
   1.875 +  /* Convert cell coordinates to update box ID */
   1.876 +  c0 >>= BOX_C0_LOG;
   1.877 +  c1 >>= BOX_C1_LOG;
   1.878 +  c2 >>= BOX_C2_LOG;
   1.879 +
   1.880 +  /* Compute true coordinates of update box's origin corner.
   1.881 +   * Actually we compute the coordinates of the center of the corner
   1.882 +   * histogram cell, which are the lower bounds of the volume we care about.
   1.883 +   */
   1.884 +  minc0 = (c0 << BOX_C0_SHIFT) + ((1 << C0_SHIFT) >> 1);
   1.885 +  minc1 = (c1 << BOX_C1_SHIFT) + ((1 << C1_SHIFT) >> 1);
   1.886 +  minc2 = (c2 << BOX_C2_SHIFT) + ((1 << C2_SHIFT) >> 1);
   1.887 +  
   1.888 +  /* Determine which colormap entries are close enough to be candidates
   1.889 +   * for the nearest entry to some cell in the update box.
   1.890 +   */
   1.891 +  numcolors = find_nearby_colors(cinfo, minc0, minc1, minc2, colorlist);
   1.892 +
   1.893 +  /* Determine the actually nearest colors. */
   1.894 +  find_best_colors(cinfo, minc0, minc1, minc2, numcolors, colorlist,
   1.895 +		   bestcolor);
   1.896 +
   1.897 +  /* Save the best color numbers (plus 1) in the main cache array */
   1.898 +  c0 <<= BOX_C0_LOG;		/* convert ID back to base cell indexes */
   1.899 +  c1 <<= BOX_C1_LOG;
   1.900 +  c2 <<= BOX_C2_LOG;
   1.901 +  cptr = bestcolor;
   1.902 +  for (ic0 = 0; ic0 < BOX_C0_ELEMS; ic0++) {
   1.903 +    for (ic1 = 0; ic1 < BOX_C1_ELEMS; ic1++) {
   1.904 +      cachep = & histogram[c0+ic0][c1+ic1][c2];
   1.905 +      for (ic2 = 0; ic2 < BOX_C2_ELEMS; ic2++) {
   1.906 +	*cachep++ = (histcell) (GETJSAMPLE(*cptr++) + 1);
   1.907 +      }
   1.908 +    }
   1.909 +  }
   1.910 +}
   1.911 +
   1.912 +
   1.913 +/*
   1.914 + * Map some rows of pixels to the output colormapped representation.
   1.915 + */
   1.916 +
   1.917 +METHODDEF(void)
   1.918 +pass2_no_dither (j_decompress_ptr cinfo,
   1.919 +		 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
   1.920 +/* This version performs no dithering */
   1.921 +{
   1.922 +  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
   1.923 +  hist3d histogram = cquantize->histogram;
   1.924 +  register JSAMPROW inptr, outptr;
   1.925 +  register histptr cachep;
   1.926 +  register int c0, c1, c2;
   1.927 +  int row;
   1.928 +  JDIMENSION col;
   1.929 +  JDIMENSION width = cinfo->output_width;
   1.930 +
   1.931 +  for (row = 0; row < num_rows; row++) {
   1.932 +    inptr = input_buf[row];
   1.933 +    outptr = output_buf[row];
   1.934 +    for (col = width; col > 0; col--) {
   1.935 +      /* get pixel value and index into the cache */
   1.936 +      c0 = GETJSAMPLE(*inptr++) >> C0_SHIFT;
   1.937 +      c1 = GETJSAMPLE(*inptr++) >> C1_SHIFT;
   1.938 +      c2 = GETJSAMPLE(*inptr++) >> C2_SHIFT;
   1.939 +      cachep = & histogram[c0][c1][c2];
   1.940 +      /* If we have not seen this color before, find nearest colormap entry */
   1.941 +      /* and update the cache */
   1.942 +      if (*cachep == 0)
   1.943 +	fill_inverse_cmap(cinfo, c0,c1,c2);
   1.944 +      /* Now emit the colormap index for this cell */
   1.945 +      *outptr++ = (JSAMPLE) (*cachep - 1);
   1.946 +    }
   1.947 +  }
   1.948 +}
   1.949 +
   1.950 +
   1.951 +METHODDEF(void)
   1.952 +pass2_fs_dither (j_decompress_ptr cinfo,
   1.953 +		 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
   1.954 +/* This version performs Floyd-Steinberg dithering */
   1.955 +{
   1.956 +  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
   1.957 +  hist3d histogram = cquantize->histogram;
   1.958 +  register LOCFSERROR cur0, cur1, cur2;	/* current error or pixel value */
   1.959 +  LOCFSERROR belowerr0, belowerr1, belowerr2; /* error for pixel below cur */
   1.960 +  LOCFSERROR bpreverr0, bpreverr1, bpreverr2; /* error for below/prev col */
   1.961 +  register FSERRPTR errorptr;	/* => fserrors[] at column before current */
   1.962 +  JSAMPROW inptr;		/* => current input pixel */
   1.963 +  JSAMPROW outptr;		/* => current output pixel */
   1.964 +  histptr cachep;
   1.965 +  int dir;			/* +1 or -1 depending on direction */
   1.966 +  int dir3;			/* 3*dir, for advancing inptr & errorptr */
   1.967 +  int row;
   1.968 +  JDIMENSION col;
   1.969 +  JDIMENSION width = cinfo->output_width;
   1.970 +  JSAMPLE *range_limit = cinfo->sample_range_limit;
   1.971 +  int *error_limit = cquantize->error_limiter;
   1.972 +  JSAMPROW colormap0 = cinfo->colormap[0];
   1.973 +  JSAMPROW colormap1 = cinfo->colormap[1];
   1.974 +  JSAMPROW colormap2 = cinfo->colormap[2];
   1.975 +  SHIFT_TEMPS
   1.976 +
   1.977 +  for (row = 0; row < num_rows; row++) {
   1.978 +    inptr = input_buf[row];
   1.979 +    outptr = output_buf[row];
   1.980 +    if (cquantize->on_odd_row) {
   1.981 +      /* work right to left in this row */
   1.982 +      inptr += (width-1) * 3;	/* so point to rightmost pixel */
   1.983 +      outptr += width-1;
   1.984 +      dir = -1;
   1.985 +      dir3 = -3;
   1.986 +      errorptr = cquantize->fserrors + (width+1)*3; /* => entry after last column */
   1.987 +      cquantize->on_odd_row = FALSE; /* flip for next time */
   1.988 +    } else {
   1.989 +      /* work left to right in this row */
   1.990 +      dir = 1;
   1.991 +      dir3 = 3;
   1.992 +      errorptr = cquantize->fserrors; /* => entry before first real column */
   1.993 +      cquantize->on_odd_row = TRUE; /* flip for next time */
   1.994 +    }
   1.995 +    /* Preset error values: no error propagated to first pixel from left */
   1.996 +    cur0 = cur1 = cur2 = 0;
   1.997 +    /* and no error propagated to row below yet */
   1.998 +    belowerr0 = belowerr1 = belowerr2 = 0;
   1.999 +    bpreverr0 = bpreverr1 = bpreverr2 = 0;
  1.1000 +
  1.1001 +    for (col = width; col > 0; col--) {
  1.1002 +      /* curN holds the error propagated from the previous pixel on the
  1.1003 +       * current line.  Add the error propagated from the previous line
  1.1004 +       * to form the complete error correction term for this pixel, and
  1.1005 +       * round the error term (which is expressed * 16) to an integer.
  1.1006 +       * RIGHT_SHIFT rounds towards minus infinity, so adding 8 is correct
  1.1007 +       * for either sign of the error value.
  1.1008 +       * Note: errorptr points to *previous* column's array entry.
  1.1009 +       */
  1.1010 +      cur0 = RIGHT_SHIFT(cur0 + errorptr[dir3+0] + 8, 4);
  1.1011 +      cur1 = RIGHT_SHIFT(cur1 + errorptr[dir3+1] + 8, 4);
  1.1012 +      cur2 = RIGHT_SHIFT(cur2 + errorptr[dir3+2] + 8, 4);
  1.1013 +      /* Limit the error using transfer function set by init_error_limit.
  1.1014 +       * See comments with init_error_limit for rationale.
  1.1015 +       */
  1.1016 +      cur0 = error_limit[cur0];
  1.1017 +      cur1 = error_limit[cur1];
  1.1018 +      cur2 = error_limit[cur2];
  1.1019 +      /* Form pixel value + error, and range-limit to 0..MAXJSAMPLE.
  1.1020 +       * The maximum error is +- MAXJSAMPLE (or less with error limiting);
  1.1021 +       * this sets the required size of the range_limit array.
  1.1022 +       */
  1.1023 +      cur0 += GETJSAMPLE(inptr[0]);
  1.1024 +      cur1 += GETJSAMPLE(inptr[1]);
  1.1025 +      cur2 += GETJSAMPLE(inptr[2]);
  1.1026 +      cur0 = GETJSAMPLE(range_limit[cur0]);
  1.1027 +      cur1 = GETJSAMPLE(range_limit[cur1]);
  1.1028 +      cur2 = GETJSAMPLE(range_limit[cur2]);
  1.1029 +      /* Index into the cache with adjusted pixel value */
  1.1030 +      cachep = & histogram[cur0>>C0_SHIFT][cur1>>C1_SHIFT][cur2>>C2_SHIFT];
  1.1031 +      /* If we have not seen this color before, find nearest colormap */
  1.1032 +      /* entry and update the cache */
  1.1033 +      if (*cachep == 0)
  1.1034 +	fill_inverse_cmap(cinfo, cur0>>C0_SHIFT,cur1>>C1_SHIFT,cur2>>C2_SHIFT);
  1.1035 +      /* Now emit the colormap index for this cell */
  1.1036 +      { register int pixcode = *cachep - 1;
  1.1037 +	*outptr = (JSAMPLE) pixcode;
  1.1038 +	/* Compute representation error for this pixel */
  1.1039 +	cur0 -= GETJSAMPLE(colormap0[pixcode]);
  1.1040 +	cur1 -= GETJSAMPLE(colormap1[pixcode]);
  1.1041 +	cur2 -= GETJSAMPLE(colormap2[pixcode]);
  1.1042 +      }
  1.1043 +      /* Compute error fractions to be propagated to adjacent pixels.
  1.1044 +       * Add these into the running sums, and simultaneously shift the
  1.1045 +       * next-line error sums left by 1 column.
  1.1046 +       */
  1.1047 +      { register LOCFSERROR bnexterr, delta;
  1.1048 +
  1.1049 +	bnexterr = cur0;	/* Process component 0 */
  1.1050 +	delta = cur0 * 2;
  1.1051 +	cur0 += delta;		/* form error * 3 */
  1.1052 +	errorptr[0] = (FSERROR) (bpreverr0 + cur0);
  1.1053 +	cur0 += delta;		/* form error * 5 */
  1.1054 +	bpreverr0 = belowerr0 + cur0;
  1.1055 +	belowerr0 = bnexterr;
  1.1056 +	cur0 += delta;		/* form error * 7 */
  1.1057 +	bnexterr = cur1;	/* Process component 1 */
  1.1058 +	delta = cur1 * 2;
  1.1059 +	cur1 += delta;		/* form error * 3 */
  1.1060 +	errorptr[1] = (FSERROR) (bpreverr1 + cur1);
  1.1061 +	cur1 += delta;		/* form error * 5 */
  1.1062 +	bpreverr1 = belowerr1 + cur1;
  1.1063 +	belowerr1 = bnexterr;
  1.1064 +	cur1 += delta;		/* form error * 7 */
  1.1065 +	bnexterr = cur2;	/* Process component 2 */
  1.1066 +	delta = cur2 * 2;
  1.1067 +	cur2 += delta;		/* form error * 3 */
  1.1068 +	errorptr[2] = (FSERROR) (bpreverr2 + cur2);
  1.1069 +	cur2 += delta;		/* form error * 5 */
  1.1070 +	bpreverr2 = belowerr2 + cur2;
  1.1071 +	belowerr2 = bnexterr;
  1.1072 +	cur2 += delta;		/* form error * 7 */
  1.1073 +      }
  1.1074 +      /* At this point curN contains the 7/16 error value to be propagated
  1.1075 +       * to the next pixel on the current line, and all the errors for the
  1.1076 +       * next line have been shifted over.  We are therefore ready to move on.
  1.1077 +       */
  1.1078 +      inptr += dir3;		/* Advance pixel pointers to next column */
  1.1079 +      outptr += dir;
  1.1080 +      errorptr += dir3;		/* advance errorptr to current column */
  1.1081 +    }
  1.1082 +    /* Post-loop cleanup: we must unload the final error values into the
  1.1083 +     * final fserrors[] entry.  Note we need not unload belowerrN because
  1.1084 +     * it is for the dummy column before or after the actual array.
  1.1085 +     */
  1.1086 +    errorptr[0] = (FSERROR) bpreverr0; /* unload prev errs into array */
  1.1087 +    errorptr[1] = (FSERROR) bpreverr1;
  1.1088 +    errorptr[2] = (FSERROR) bpreverr2;
  1.1089 +  }
  1.1090 +}
  1.1091 +
  1.1092 +
  1.1093 +/*
  1.1094 + * Initialize the error-limiting transfer function (lookup table).
  1.1095 + * The raw F-S error computation can potentially compute error values of up to
  1.1096 + * +- MAXJSAMPLE.  But we want the maximum correction applied to a pixel to be
  1.1097 + * much less, otherwise obviously wrong pixels will be created.  (Typical
  1.1098 + * effects include weird fringes at color-area boundaries, isolated bright
  1.1099 + * pixels in a dark area, etc.)  The standard advice for avoiding this problem
  1.1100 + * is to ensure that the "corners" of the color cube are allocated as output
  1.1101 + * colors; then repeated errors in the same direction cannot cause cascading
  1.1102 + * error buildup.  However, that only prevents the error from getting
  1.1103 + * completely out of hand; Aaron Giles reports that error limiting improves
  1.1104 + * the results even with corner colors allocated.
  1.1105 + * A simple clamping of the error values to about +- MAXJSAMPLE/8 works pretty
  1.1106 + * well, but the smoother transfer function used below is even better.  Thanks
  1.1107 + * to Aaron Giles for this idea.
  1.1108 + */
  1.1109 +
  1.1110 +LOCAL(void)
  1.1111 +init_error_limit (j_decompress_ptr cinfo)
  1.1112 +/* Allocate and fill in the error_limiter table */
  1.1113 +{
  1.1114 +  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
  1.1115 +  int * table;
  1.1116 +  int in, out;
  1.1117 +
  1.1118 +  table = (int *) (*cinfo->mem->alloc_small)
  1.1119 +    ((j_common_ptr) cinfo, JPOOL_IMAGE, (MAXJSAMPLE*2+1) * SIZEOF(int));
  1.1120 +  table += MAXJSAMPLE;		/* so can index -MAXJSAMPLE .. +MAXJSAMPLE */
  1.1121 +  cquantize->error_limiter = table;
  1.1122 +
  1.1123 +#define STEPSIZE ((MAXJSAMPLE+1)/16)
  1.1124 +  /* Map errors 1:1 up to +- MAXJSAMPLE/16 */
  1.1125 +  out = 0;
  1.1126 +  for (in = 0; in < STEPSIZE; in++, out++) {
  1.1127 +    table[in] = out; table[-in] = -out;
  1.1128 +  }
  1.1129 +  /* Map errors 1:2 up to +- 3*MAXJSAMPLE/16 */
  1.1130 +  for (; in < STEPSIZE*3; in++, out += (in&1) ? 0 : 1) {
  1.1131 +    table[in] = out; table[-in] = -out;
  1.1132 +  }
  1.1133 +  /* Clamp the rest to final out value (which is (MAXJSAMPLE+1)/8) */
  1.1134 +  for (; in <= MAXJSAMPLE; in++) {
  1.1135 +    table[in] = out; table[-in] = -out;
  1.1136 +  }
  1.1137 +#undef STEPSIZE
  1.1138 +}
  1.1139 +
  1.1140 +
  1.1141 +/*
  1.1142 + * Finish up at the end of each pass.
  1.1143 + */
  1.1144 +
  1.1145 +METHODDEF(void)
  1.1146 +finish_pass1 (j_decompress_ptr cinfo)
  1.1147 +{
  1.1148 +  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
  1.1149 +
  1.1150 +  /* Select the representative colors and fill in cinfo->colormap */
  1.1151 +  cinfo->colormap = cquantize->sv_colormap;
  1.1152 +  select_colors(cinfo, cquantize->desired);
  1.1153 +  /* Force next pass to zero the color index table */
  1.1154 +  cquantize->needs_zeroed = TRUE;
  1.1155 +}
  1.1156 +
  1.1157 +
  1.1158 +METHODDEF(void)
  1.1159 +finish_pass2 (j_decompress_ptr cinfo)
  1.1160 +{
  1.1161 +  /* no work */
  1.1162 +}
  1.1163 +
  1.1164 +
  1.1165 +/*
  1.1166 + * Initialize for each processing pass.
  1.1167 + */
  1.1168 +
  1.1169 +METHODDEF(void)
  1.1170 +start_pass_2_quant (j_decompress_ptr cinfo, boolean is_pre_scan)
  1.1171 +{
  1.1172 +  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
  1.1173 +  hist3d histogram = cquantize->histogram;
  1.1174 +  int i;
  1.1175 +
  1.1176 +  /* Only F-S dithering or no dithering is supported. */
  1.1177 +  /* If user asks for ordered dither, give him F-S. */
  1.1178 +  if (cinfo->dither_mode != JDITHER_NONE)
  1.1179 +    cinfo->dither_mode = JDITHER_FS;
  1.1180 +
  1.1181 +  if (is_pre_scan) {
  1.1182 +    /* Set up method pointers */
  1.1183 +    cquantize->pub.color_quantize = prescan_quantize;
  1.1184 +    cquantize->pub.finish_pass = finish_pass1;
  1.1185 +    cquantize->needs_zeroed = TRUE; /* Always zero histogram */
  1.1186 +  } else {
  1.1187 +    /* Set up method pointers */
  1.1188 +    if (cinfo->dither_mode == JDITHER_FS)
  1.1189 +      cquantize->pub.color_quantize = pass2_fs_dither;
  1.1190 +    else
  1.1191 +      cquantize->pub.color_quantize = pass2_no_dither;
  1.1192 +    cquantize->pub.finish_pass = finish_pass2;
  1.1193 +
  1.1194 +    /* Make sure color count is acceptable */
  1.1195 +    i = cinfo->actual_number_of_colors;
  1.1196 +    if (i < 1)
  1.1197 +      ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 1);
  1.1198 +    if (i > MAXNUMCOLORS)
  1.1199 +      ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
  1.1200 +
  1.1201 +    if (cinfo->dither_mode == JDITHER_FS) {
  1.1202 +      size_t arraysize = (size_t) ((cinfo->output_width + 2) *
  1.1203 +				   (3 * SIZEOF(FSERROR)));
  1.1204 +      /* Allocate Floyd-Steinberg workspace if we didn't already. */
  1.1205 +      if (cquantize->fserrors == NULL)
  1.1206 +	cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
  1.1207 +	  ((j_common_ptr) cinfo, JPOOL_IMAGE, arraysize);
  1.1208 +      /* Initialize the propagated errors to zero. */
  1.1209 +      jzero_far((void FAR *) cquantize->fserrors, arraysize);
  1.1210 +      /* Make the error-limit table if we didn't already. */
  1.1211 +      if (cquantize->error_limiter == NULL)
  1.1212 +	init_error_limit(cinfo);
  1.1213 +      cquantize->on_odd_row = FALSE;
  1.1214 +    }
  1.1215 +
  1.1216 +  }
  1.1217 +  /* Zero the histogram or inverse color map, if necessary */
  1.1218 +  if (cquantize->needs_zeroed) {
  1.1219 +    for (i = 0; i < HIST_C0_ELEMS; i++) {
  1.1220 +      jzero_far((void FAR *) histogram[i],
  1.1221 +		HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
  1.1222 +    }
  1.1223 +    cquantize->needs_zeroed = FALSE;
  1.1224 +  }
  1.1225 +}
  1.1226 +
  1.1227 +
  1.1228 +/*
  1.1229 + * Switch to a new external colormap between output passes.
  1.1230 + */
  1.1231 +
  1.1232 +METHODDEF(void)
  1.1233 +new_color_map_2_quant (j_decompress_ptr cinfo)
  1.1234 +{
  1.1235 +  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
  1.1236 +
  1.1237 +  /* Reset the inverse color map */
  1.1238 +  cquantize->needs_zeroed = TRUE;
  1.1239 +}
  1.1240 +
  1.1241 +
  1.1242 +/*
  1.1243 + * Module initialization routine for 2-pass color quantization.
  1.1244 + */
  1.1245 +
  1.1246 +GLOBAL(void)
  1.1247 +jinit_2pass_quantizer (j_decompress_ptr cinfo)
  1.1248 +{
  1.1249 +  my_cquantize_ptr cquantize;
  1.1250 +  int i;
  1.1251 +
  1.1252 +  cquantize = (my_cquantize_ptr)
  1.1253 +    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
  1.1254 +				SIZEOF(my_cquantizer));
  1.1255 +  cinfo->cquantize = (struct jpeg_color_quantizer *) cquantize;
  1.1256 +  cquantize->pub.start_pass = start_pass_2_quant;
  1.1257 +  cquantize->pub.new_color_map = new_color_map_2_quant;
  1.1258 +  cquantize->fserrors = NULL;	/* flag optional arrays not allocated */
  1.1259 +  cquantize->error_limiter = NULL;
  1.1260 +
  1.1261 +  /* Make sure jdmaster didn't give me a case I can't handle */
  1.1262 +  if (cinfo->out_color_components != 3)
  1.1263 +    ERREXIT(cinfo, JERR_NOTIMPL);
  1.1264 +
  1.1265 +  /* Allocate the histogram/inverse colormap storage */
  1.1266 +  cquantize->histogram = (hist3d) (*cinfo->mem->alloc_small)
  1.1267 +    ((j_common_ptr) cinfo, JPOOL_IMAGE, HIST_C0_ELEMS * SIZEOF(hist2d));
  1.1268 +  for (i = 0; i < HIST_C0_ELEMS; i++) {
  1.1269 +    cquantize->histogram[i] = (hist2d) (*cinfo->mem->alloc_large)
  1.1270 +      ((j_common_ptr) cinfo, JPOOL_IMAGE,
  1.1271 +       HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
  1.1272 +  }
  1.1273 +  cquantize->needs_zeroed = TRUE; /* histogram is garbage now */
  1.1274 +
  1.1275 +  /* Allocate storage for the completed colormap, if required.
  1.1276 +   * We do this now since it is FAR storage and may affect
  1.1277 +   * the memory manager's space calculations.
  1.1278 +   */
  1.1279 +  if (cinfo->enable_2pass_quant) {
  1.1280 +    /* Make sure color count is acceptable */
  1.1281 +    int desired = cinfo->desired_number_of_colors;
  1.1282 +    /* Lower bound on # of colors ... somewhat arbitrary as long as > 0 */
  1.1283 +    if (desired < 8)
  1.1284 +      ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 8);
  1.1285 +    /* Make sure colormap indexes can be represented by JSAMPLEs */
  1.1286 +    if (desired > MAXNUMCOLORS)
  1.1287 +      ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
  1.1288 +    cquantize->sv_colormap = (*cinfo->mem->alloc_sarray)
  1.1289 +      ((j_common_ptr) cinfo,JPOOL_IMAGE, (JDIMENSION) desired, (JDIMENSION) 3);
  1.1290 +    cquantize->desired = desired;
  1.1291 +  } else
  1.1292 +    cquantize->sv_colormap = NULL;
  1.1293 +
  1.1294 +  /* Only F-S dithering or no dithering is supported. */
  1.1295 +  /* If user asks for ordered dither, give him F-S. */
  1.1296 +  if (cinfo->dither_mode != JDITHER_NONE)
  1.1297 +    cinfo->dither_mode = JDITHER_FS;
  1.1298 +
  1.1299 +  /* Allocate Floyd-Steinberg workspace if necessary.
  1.1300 +   * This isn't really needed until pass 2, but again it is FAR storage.
  1.1301 +   * Although we will cope with a later change in dither_mode,
  1.1302 +   * we do not promise to honor max_memory_to_use if dither_mode changes.
  1.1303 +   */
  1.1304 +  if (cinfo->dither_mode == JDITHER_FS) {
  1.1305 +    cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
  1.1306 +      ((j_common_ptr) cinfo, JPOOL_IMAGE,
  1.1307 +       (size_t) ((cinfo->output_width + 2) * (3 * SIZEOF(FSERROR))));
  1.1308 +    /* Might as well create the error-limiting table too. */
  1.1309 +    init_error_limit(cinfo);
  1.1310 +  }
  1.1311 +}
  1.1312 +
  1.1313 +#endif /* QUANT_2PASS_SUPPORTED */