Subversion Repositories Kolibri OS

Rev

Blame | Last modification | View Log | RSS feed

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