jquant2.c

00001 /*
00002  * jquant2.c
00003  *
00004  * Copyright (C) 1991-1996, Thomas G. Lane.
00005  * Modified 2011 by Guido Vollbeding.
00006  * This file is part of the Independent JPEG Group's software.
00007  * For conditions of distribution and use, see the accompanying README file.
00008  *
00009  * This file contains 2-pass color quantization (color mapping) routines.
00010  * These routines provide selection of a custom color map for an image,
00011  * followed by mapping of the image to that color map, with optional
00012  * Floyd-Steinberg dithering.
00013  * It is also possible to use just the second pass to map to an arbitrary
00014  * externally-given color map.
00015  *
00016  * Note: ordered dithering is not supported, since there isn't any fast
00017  * way to compute intercolor distances; it's unclear that ordered dither's
00018  * fundamental assumptions even hold with an irregularly spaced color map.
00019  */
00020 
00021 #define JPEG_INTERNALS
00022 #include "jinclude.h"
00023 #include "jpeglib.h"
00024 
00025 #ifdef QUANT_2PASS_SUPPORTED
00026 
00027 
00028 /*
00029  * This module implements the well-known Heckbert paradigm for color
00030  * quantization.  Most of the ideas used here can be traced back to
00031  * Heckbert's seminal paper
00032  *   Heckbert, Paul.  "Color Image Quantization for Frame Buffer Display",
00033  *   Proc. SIGGRAPH '82, Computer Graphics v.16 #3 (July 1982), pp 297-304.
00034  *
00035  * In the first pass over the image, we accumulate a histogram showing the
00036  * usage count of each possible color.  To keep the histogram to a reasonable
00037  * size, we reduce the precision of the input; typical practice is to retain
00038  * 5 or 6 bits per color, so that 8 or 4 different input values are counted
00039  * in the same histogram cell.
00040  *
00041  * Next, the color-selection step begins with a box representing the whole
00042  * color space, and repeatedly splits the "largest" remaining box until we
00043  * have as many boxes as desired colors.  Then the mean color in each
00044  * remaining box becomes one of the possible output colors.
00045  * 
00046  * The second pass over the image maps each input pixel to the closest output
00047  * color (optionally after applying a Floyd-Steinberg dithering correction).
00048  * This mapping is logically trivial, but making it go fast enough requires
00049  * considerable care.
00050  *
00051  * Heckbert-style quantizers vary a good deal in their policies for choosing
00052  * the "largest" box and deciding where to cut it.  The particular policies
00053  * used here have proved out well in experimental comparisons, but better ones
00054  * may yet be found.
00055  *
00056  * In earlier versions of the IJG code, this module quantized in YCbCr color
00057  * space, processing the raw upsampled data without a color conversion step.
00058  * This allowed the color conversion math to be done only once per colormap
00059  * entry, not once per pixel.  However, that optimization precluded other
00060  * useful optimizations (such as merging color conversion with upsampling)
00061  * and it also interfered with desired capabilities such as quantizing to an
00062  * externally-supplied colormap.  We have therefore abandoned that approach.
00063  * The present code works in the post-conversion color space, typically RGB.
00064  *
00065  * To improve the visual quality of the results, we actually work in scaled
00066  * RGB space, giving G distances more weight than R, and R in turn more than
00067  * B.  To do everything in integer math, we must use integer scale factors.
00068  * The 2/3/1 scale factors used here correspond loosely to the relative
00069  * weights of the colors in the NTSC grayscale equation.
00070  * If you want to use this code to quantize a non-RGB color space, you'll
00071  * probably need to change these scale factors.
00072  */
00073 
00074 #define R_SCALE 2       /* scale R distances by this much */
00075 #define G_SCALE 3       /* scale G distances by this much */
00076 #define B_SCALE 1       /* and B by this much */
00077 
00078 /* Relabel R/G/B as components 0/1/2, respecting the RGB ordering defined
00079  * in jmorecfg.h.  As the code stands, it will do the right thing for R,G,B
00080  * and B,G,R orders.  If you define some other weird order in jmorecfg.h,
00081  * you'll get compile errors until you extend this logic.  In that case
00082  * you'll probably want to tweak the histogram sizes too.
00083  */
00084 
00085 #if RGB_RED == 0
00086 #define C0_SCALE R_SCALE
00087 #endif
00088 #if RGB_BLUE == 0
00089 #define C0_SCALE B_SCALE
00090 #endif
00091 #if RGB_GREEN == 1
00092 #define C1_SCALE G_SCALE
00093 #endif
00094 #if RGB_RED == 2
00095 #define C2_SCALE R_SCALE
00096 #endif
00097 #if RGB_BLUE == 2
00098 #define C2_SCALE B_SCALE
00099 #endif
00100 
00101 
00102 /*
00103  * First we have the histogram data structure and routines for creating it.
00104  *
00105  * The number of bits of precision can be adjusted by changing these symbols.
00106  * We recommend keeping 6 bits for G and 5 each for R and B.
00107  * If you have plenty of memory and cycles, 6 bits all around gives marginally
00108  * better results; if you are short of memory, 5 bits all around will save
00109  * some space but degrade the results.
00110  * To maintain a fully accurate histogram, we'd need to allocate a "long"
00111  * (preferably unsigned long) for each cell.  In practice this is overkill;
00112  * we can get by with 16 bits per cell.  Few of the cell counts will overflow,
00113  * and clamping those that do overflow to the maximum value will give close-
00114  * enough results.  This reduces the recommended histogram size from 256Kb
00115  * to 128Kb, which is a useful savings on PC-class machines.
00116  * (In the second pass the histogram space is re-used for pixel mapping data;
00117  * in that capacity, each cell must be able to store zero to the number of
00118  * desired colors.  16 bits/cell is plenty for that too.)
00119  * Since the JPEG code is intended to run in small memory model on 80x86
00120  * machines, we can't just allocate the histogram in one chunk.  Instead
00121  * of a true 3-D array, we use a row of pointers to 2-D arrays.  Each
00122  * pointer corresponds to a C0 value (typically 2^5 = 32 pointers) and
00123  * each 2-D array has 2^6*2^5 = 2048 or 2^6*2^6 = 4096 entries.  Note that
00124  * on 80x86 machines, the pointer row is in near memory but the actual
00125  * arrays are in far memory (same arrangement as we use for image arrays).
00126  */
00127 
00128 #define MAXNUMCOLORS  (MAXJSAMPLE+1) /* maximum size of colormap */
00129 
00130 /* These will do the right thing for either R,G,B or B,G,R color order,
00131  * but you may not like the results for other color orders.
00132  */
00133 #define HIST_C0_BITS  5     /* bits of precision in R/B histogram */
00134 #define HIST_C1_BITS  6     /* bits of precision in G histogram */
00135 #define HIST_C2_BITS  5     /* bits of precision in B/R histogram */
00136 
00137 /* Number of elements along histogram axes. */
00138 #define HIST_C0_ELEMS  (1<<HIST_C0_BITS)
00139 #define HIST_C1_ELEMS  (1<<HIST_C1_BITS)
00140 #define HIST_C2_ELEMS  (1<<HIST_C2_BITS)
00141 
00142 /* These are the amounts to shift an input value to get a histogram index. */
00143 #define C0_SHIFT  (BITS_IN_JSAMPLE-HIST_C0_BITS)
00144 #define C1_SHIFT  (BITS_IN_JSAMPLE-HIST_C1_BITS)
00145 #define C2_SHIFT  (BITS_IN_JSAMPLE-HIST_C2_BITS)
00146 
00147 
00148 typedef UINT16 histcell;    /* histogram cell; prefer an unsigned type */
00149 
00150 typedef histcell FAR * histptr; /* for pointers to histogram cells */
00151 
00152 typedef histcell hist1d[HIST_C2_ELEMS]; /* typedefs for the array */
00153 typedef hist1d FAR * hist2d;    /* type for the 2nd-level pointers */
00154 typedef hist2d * hist3d;    /* type for top-level pointer */
00155 
00156 
00157 /* Declarations for Floyd-Steinberg dithering.
00158  *
00159  * Errors are accumulated into the array fserrors[], at a resolution of
00160  * 1/16th of a pixel count.  The error at a given pixel is propagated
00161  * to its not-yet-processed neighbors using the standard F-S fractions,
00162  *      ... (here)  7/16
00163  *      3/16    5/16    1/16
00164  * We work left-to-right on even rows, right-to-left on odd rows.
00165  *
00166  * We can get away with a single array (holding one row's worth of errors)
00167  * by using it to store the current row's errors at pixel columns not yet
00168  * processed, but the next row's errors at columns already processed.  We
00169  * need only a few extra variables to hold the errors immediately around the
00170  * current column.  (If we are lucky, those variables are in registers, but
00171  * even if not, they're probably cheaper to access than array elements are.)
00172  *
00173  * The fserrors[] array has (#columns + 2) entries; the extra entry at
00174  * each end saves us from special-casing the first and last pixels.
00175  * Each entry is three values long, one value for each color component.
00176  *
00177  * Note: on a wide image, we might not have enough room in a PC's near data
00178  * segment to hold the error array; so it is allocated with alloc_large.
00179  */
00180 
00181 #if BITS_IN_JSAMPLE == 8
00182 typedef INT16 FSERROR;      /* 16 bits should be enough */
00183 typedef int LOCFSERROR;     /* use 'int' for calculation temps */
00184 #else
00185 typedef INT32 FSERROR;      /* may need more than 16 bits */
00186 typedef INT32 LOCFSERROR;   /* be sure calculation temps are big enough */
00187 #endif
00188 
00189 typedef FSERROR FAR *FSERRPTR;  /* pointer to error array (in FAR storage!) */
00190 
00191 
00192 /* Private subobject */
00193 
00194 typedef struct {
00195   struct jpeg_color_quantizer pub; /* public fields */
00196 
00197   /* Space for the eventually created colormap is stashed here */
00198   JSAMPARRAY sv_colormap;   /* colormap allocated at init time */
00199   int desired;          /* desired # of colors = size of colormap */
00200 
00201   /* Variables for accumulating image statistics */
00202   hist3d histogram;     /* pointer to the histogram */
00203 
00204   boolean needs_zeroed;     /* TRUE if next pass must zero histogram */
00205 
00206   /* Variables for Floyd-Steinberg dithering */
00207   FSERRPTR fserrors;        /* accumulated errors */
00208   boolean on_odd_row;       /* flag to remember which row we are on */
00209   int * error_limiter;      /* table for clamping the applied error */
00210 } my_cquantizer;
00211 
00212 typedef my_cquantizer * my_cquantize_ptr;
00213 
00214 
00215 /*
00216  * Prescan some rows of pixels.
00217  * In this module the prescan simply updates the histogram, which has been
00218  * initialized to zeroes by start_pass.
00219  * An output_buf parameter is required by the method signature, but no data
00220  * is actually output (in fact the buffer controller is probably passing a
00221  * NULL pointer).
00222  */
00223 
00224 METHODDEF(void)
00225 prescan_quantize (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
00226           JSAMPARRAY output_buf, int num_rows)
00227 {
00228   my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
00229   register JSAMPROW ptr;
00230   register histptr histp;
00231   register hist3d histogram = cquantize->histogram;
00232   int row;
00233   JDIMENSION col;
00234   JDIMENSION width = cinfo->output_width;
00235 
00236   for (row = 0; row < num_rows; row++) {
00237     ptr = input_buf[row];
00238     for (col = width; col > 0; col--) {
00239       /* get pixel value and index into the histogram */
00240       histp = & histogram[GETJSAMPLE(ptr[0]) >> C0_SHIFT]
00241              [GETJSAMPLE(ptr[1]) >> C1_SHIFT]
00242              [GETJSAMPLE(ptr[2]) >> C2_SHIFT];
00243       /* increment, check for overflow and undo increment if so. */
00244       if (++(*histp) <= 0)
00245     (*histp)--;
00246       ptr += 3;
00247     }
00248   }
00249 }
00250 
00251 
00252 /*
00253  * Next we have the really interesting routines: selection of a colormap
00254  * given the completed histogram.
00255  * These routines work with a list of "boxes", each representing a rectangular
00256  * subset of the input color space (to histogram precision).
00257  */
00258 
00259 typedef struct {
00260   /* The bounds of the box (inclusive); expressed as histogram indexes */
00261   int c0min, c0max;
00262   int c1min, c1max;
00263   int c2min, c2max;
00264   /* The volume (actually 2-norm) of the box */
00265   INT32 volume;
00266   /* The number of nonzero histogram cells within this box */
00267   long colorcount;
00268 } box;
00269 
00270 typedef box * boxptr;
00271 
00272 
00273 LOCAL(boxptr)
00274 find_biggest_color_pop (boxptr boxlist, int numboxes)
00275 /* Find the splittable box with the largest color population */
00276 /* Returns NULL if no splittable boxes remain */
00277 {
00278   register boxptr boxp;
00279   register int i;
00280   register long maxc = 0;
00281   boxptr which = NULL;
00282   
00283   for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
00284     if (boxp->colorcount > maxc && boxp->volume > 0) {
00285       which = boxp;
00286       maxc = boxp->colorcount;
00287     }
00288   }
00289   return which;
00290 }
00291 
00292 
00293 LOCAL(boxptr)
00294 find_biggest_volume (boxptr boxlist, int numboxes)
00295 /* Find the splittable box with the largest (scaled) volume */
00296 /* Returns NULL if no splittable boxes remain */
00297 {
00298   register boxptr boxp;
00299   register int i;
00300   register INT32 maxv = 0;
00301   boxptr which = NULL;
00302   
00303   for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
00304     if (boxp->volume > maxv) {
00305       which = boxp;
00306       maxv = boxp->volume;
00307     }
00308   }
00309   return which;
00310 }
00311 
00312 
00313 LOCAL(void)
00314 update_box (j_decompress_ptr cinfo, boxptr boxp)
00315 /* Shrink the min/max bounds of a box to enclose only nonzero elements, */
00316 /* and recompute its volume and population */
00317 {
00318   my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
00319   hist3d histogram = cquantize->histogram;
00320   histptr histp;
00321   int c0,c1,c2;
00322   int c0min,c0max,c1min,c1max,c2min,c2max;
00323   INT32 dist0,dist1,dist2;
00324   long ccount;
00325   
00326   c0min = boxp->c0min;  c0max = boxp->c0max;
00327   c1min = boxp->c1min;  c1max = boxp->c1max;
00328   c2min = boxp->c2min;  c2max = boxp->c2max;
00329   
00330   if (c0max > c0min)
00331     for (c0 = c0min; c0 <= c0max; c0++)
00332       for (c1 = c1min; c1 <= c1max; c1++) {
00333     histp = & histogram[c0][c1][c2min];
00334     for (c2 = c2min; c2 <= c2max; c2++)
00335       if (*histp++ != 0) {
00336         boxp->c0min = c0min = c0;
00337         goto have_c0min;
00338       }
00339       }
00340  have_c0min:
00341   if (c0max > c0min)
00342     for (c0 = c0max; c0 >= c0min; c0--)
00343       for (c1 = c1min; c1 <= c1max; c1++) {
00344     histp = & histogram[c0][c1][c2min];
00345     for (c2 = c2min; c2 <= c2max; c2++)
00346       if (*histp++ != 0) {
00347         boxp->c0max = c0max = c0;
00348         goto have_c0max;
00349       }
00350       }
00351  have_c0max:
00352   if (c1max > c1min)
00353     for (c1 = c1min; c1 <= c1max; c1++)
00354       for (c0 = c0min; c0 <= c0max; c0++) {
00355     histp = & histogram[c0][c1][c2min];
00356     for (c2 = c2min; c2 <= c2max; c2++)
00357       if (*histp++ != 0) {
00358         boxp->c1min = c1min = c1;
00359         goto have_c1min;
00360       }
00361       }
00362  have_c1min:
00363   if (c1max > c1min)
00364     for (c1 = c1max; c1 >= c1min; c1--)
00365       for (c0 = c0min; c0 <= c0max; c0++) {
00366     histp = & histogram[c0][c1][c2min];
00367     for (c2 = c2min; c2 <= c2max; c2++)
00368       if (*histp++ != 0) {
00369         boxp->c1max = c1max = c1;
00370         goto have_c1max;
00371       }
00372       }
00373  have_c1max:
00374   if (c2max > c2min)
00375     for (c2 = c2min; c2 <= c2max; c2++)
00376       for (c0 = c0min; c0 <= c0max; c0++) {
00377     histp = & histogram[c0][c1min][c2];
00378     for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
00379       if (*histp != 0) {
00380         boxp->c2min = c2min = c2;
00381         goto have_c2min;
00382       }
00383       }
00384  have_c2min:
00385   if (c2max > c2min)
00386     for (c2 = c2max; c2 >= c2min; c2--)
00387       for (c0 = c0min; c0 <= c0max; c0++) {
00388     histp = & histogram[c0][c1min][c2];
00389     for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
00390       if (*histp != 0) {
00391         boxp->c2max = c2max = c2;
00392         goto have_c2max;
00393       }
00394       }
00395  have_c2max:
00396 
00397   /* Update box volume.
00398    * We use 2-norm rather than real volume here; this biases the method
00399    * against making long narrow boxes, and it has the side benefit that
00400    * a box is splittable iff norm > 0.
00401    * Since the differences are expressed in histogram-cell units,
00402    * we have to shift back to JSAMPLE units to get consistent distances;
00403    * after which, we scale according to the selected distance scale factors.
00404    */
00405   dist0 = ((c0max - c0min) << C0_SHIFT) * C0_SCALE;
00406   dist1 = ((c1max - c1min) << C1_SHIFT) * C1_SCALE;
00407   dist2 = ((c2max - c2min) << C2_SHIFT) * C2_SCALE;
00408   boxp->volume = dist0*dist0 + dist1*dist1 + dist2*dist2;
00409   
00410   /* Now scan remaining volume of box and compute population */
00411   ccount = 0;
00412   for (c0 = c0min; c0 <= c0max; c0++)
00413     for (c1 = c1min; c1 <= c1max; c1++) {
00414       histp = & histogram[c0][c1][c2min];
00415       for (c2 = c2min; c2 <= c2max; c2++, histp++)
00416     if (*histp != 0) {
00417       ccount++;
00418     }
00419     }
00420   boxp->colorcount = ccount;
00421 }
00422 
00423 
00424 LOCAL(int)
00425 median_cut (j_decompress_ptr cinfo, boxptr boxlist, int numboxes,
00426         int desired_colors)
00427 /* Repeatedly select and split the largest box until we have enough boxes */
00428 {
00429   int n,lb;
00430   int c0,c1,c2,cmax;
00431   register boxptr b1,b2;
00432 
00433   while (numboxes < desired_colors) {
00434     /* Select box to split.
00435      * Current algorithm: by population for first half, then by volume.
00436      */
00437     if (numboxes*2 <= desired_colors) {
00438       b1 = find_biggest_color_pop(boxlist, numboxes);
00439     } else {
00440       b1 = find_biggest_volume(boxlist, numboxes);
00441     }
00442     if (b1 == NULL)     /* no splittable boxes left! */
00443       break;
00444     b2 = &boxlist[numboxes];    /* where new box will go */
00445     /* Copy the color bounds to the new box. */
00446     b2->c0max = b1->c0max; b2->c1max = b1->c1max; b2->c2max = b1->c2max;
00447     b2->c0min = b1->c0min; b2->c1min = b1->c1min; b2->c2min = b1->c2min;
00448     /* Choose which axis to split the box on.
00449      * Current algorithm: longest scaled axis.
00450      * See notes in update_box about scaling distances.
00451      */
00452     c0 = ((b1->c0max - b1->c0min) << C0_SHIFT) * C0_SCALE;
00453     c1 = ((b1->c1max - b1->c1min) << C1_SHIFT) * C1_SCALE;
00454     c2 = ((b1->c2max - b1->c2min) << C2_SHIFT) * C2_SCALE;
00455     /* We want to break any ties in favor of green, then red, blue last.
00456      * This code does the right thing for R,G,B or B,G,R color orders only.
00457      */
00458 #if RGB_RED == 0
00459     cmax = c1; n = 1;
00460     if (c0 > cmax) { cmax = c0; n = 0; }
00461     if (c2 > cmax) { n = 2; }
00462 #else
00463     cmax = c1; n = 1;
00464     if (c2 > cmax) { cmax = c2; n = 2; }
00465     if (c0 > cmax) { n = 0; }
00466 #endif
00467     /* Choose split point along selected axis, and update box bounds.
00468      * Current algorithm: split at halfway point.
00469      * (Since the box has been shrunk to minimum volume,
00470      * any split will produce two nonempty subboxes.)
00471      * Note that lb value is max for lower box, so must be < old max.
00472      */
00473     switch (n) {
00474     case 0:
00475       lb = (b1->c0max + b1->c0min) / 2;
00476       b1->c0max = lb;
00477       b2->c0min = lb+1;
00478       break;
00479     case 1:
00480       lb = (b1->c1max + b1->c1min) / 2;
00481       b1->c1max = lb;
00482       b2->c1min = lb+1;
00483       break;
00484     case 2:
00485       lb = (b1->c2max + b1->c2min) / 2;
00486       b1->c2max = lb;
00487       b2->c2min = lb+1;
00488       break;
00489     }
00490     /* Update stats for boxes */
00491     update_box(cinfo, b1);
00492     update_box(cinfo, b2);
00493     numboxes++;
00494   }
00495   return numboxes;
00496 }
00497 
00498 
00499 LOCAL(void)
00500 compute_color (j_decompress_ptr cinfo, boxptr boxp, int icolor)
00501 /* Compute representative color for a box, put it in colormap[icolor] */
00502 {
00503   /* Current algorithm: mean weighted by pixels (not colors) */
00504   /* Note it is important to get the rounding correct! */
00505   my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
00506   hist3d histogram = cquantize->histogram;
00507   histptr histp;
00508   int c0,c1,c2;
00509   int c0min,c0max,c1min,c1max,c2min,c2max;
00510   long count;
00511   long total = 0;
00512   long c0total = 0;
00513   long c1total = 0;
00514   long c2total = 0;
00515   
00516   c0min = boxp->c0min;  c0max = boxp->c0max;
00517   c1min = boxp->c1min;  c1max = boxp->c1max;
00518   c2min = boxp->c2min;  c2max = boxp->c2max;
00519   
00520   for (c0 = c0min; c0 <= c0max; c0++)
00521     for (c1 = c1min; c1 <= c1max; c1++) {
00522       histp = & histogram[c0][c1][c2min];
00523       for (c2 = c2min; c2 <= c2max; c2++) {
00524     if ((count = *histp++) != 0) {
00525       total += count;
00526       c0total += ((c0 << C0_SHIFT) + ((1<<C0_SHIFT)>>1)) * count;
00527       c1total += ((c1 << C1_SHIFT) + ((1<<C1_SHIFT)>>1)) * count;
00528       c2total += ((c2 << C2_SHIFT) + ((1<<C2_SHIFT)>>1)) * count;
00529     }
00530       }
00531     }
00532   
00533   cinfo->colormap[0][icolor] = (JSAMPLE) ((c0total + (total>>1)) / total);
00534   cinfo->colormap[1][icolor] = (JSAMPLE) ((c1total + (total>>1)) / total);
00535   cinfo->colormap[2][icolor] = (JSAMPLE) ((c2total + (total>>1)) / total);
00536 }
00537 
00538 
00539 LOCAL(void)
00540 select_colors (j_decompress_ptr cinfo, int desired_colors)
00541 /* Master routine for color selection */
00542 {
00543   boxptr boxlist;
00544   int numboxes;
00545   int i;
00546 
00547   /* Allocate workspace for box list */
00548   boxlist = (boxptr) (*cinfo->mem->alloc_small)
00549     ((j_common_ptr) cinfo, JPOOL_IMAGE, desired_colors * SIZEOF(box));
00550   /* Initialize one box containing whole space */
00551   numboxes = 1;
00552   boxlist[0].c0min = 0;
00553   boxlist[0].c0max = MAXJSAMPLE >> C0_SHIFT;
00554   boxlist[0].c1min = 0;
00555   boxlist[0].c1max = MAXJSAMPLE >> C1_SHIFT;
00556   boxlist[0].c2min = 0;
00557   boxlist[0].c2max = MAXJSAMPLE >> C2_SHIFT;
00558   /* Shrink it to actually-used volume and set its statistics */
00559   update_box(cinfo, & boxlist[0]);
00560   /* Perform median-cut to produce final box list */
00561   numboxes = median_cut(cinfo, boxlist, numboxes, desired_colors);
00562   /* Compute the representative color for each box, fill colormap */
00563   for (i = 0; i < numboxes; i++)
00564     compute_color(cinfo, & boxlist[i], i);
00565   cinfo->actual_number_of_colors = numboxes;
00566   TRACEMS1(cinfo, 1, JTRC_QUANT_SELECTED, numboxes);
00567 }
00568 
00569 
00570 /*
00571  * These routines are concerned with the time-critical task of mapping input
00572  * colors to the nearest color in the selected colormap.
00573  *
00574  * We re-use the histogram space as an "inverse color map", essentially a
00575  * cache for the results of nearest-color searches.  All colors within a
00576  * histogram cell will be mapped to the same colormap entry, namely the one
00577  * closest to the cell's center.  This may not be quite the closest entry to
00578  * the actual input color, but it's almost as good.  A zero in the cache
00579  * indicates we haven't found the nearest color for that cell yet; the array
00580  * is cleared to zeroes before starting the mapping pass.  When we find the
00581  * nearest color for a cell, its colormap index plus one is recorded in the
00582  * cache for future use.  The pass2 scanning routines call fill_inverse_cmap
00583  * when they need to use an unfilled entry in the cache.
00584  *
00585  * Our method of efficiently finding nearest colors is based on the "locally
00586  * sorted search" idea described by Heckbert and on the incremental distance
00587  * calculation described by Spencer W. Thomas in chapter III.1 of Graphics
00588  * Gems II (James Arvo, ed.  Academic Press, 1991).  Thomas points out that
00589  * the distances from a given colormap entry to each cell of the histogram can
00590  * be computed quickly using an incremental method: the differences between
00591  * distances to adjacent cells themselves differ by a constant.  This allows a
00592  * fairly fast implementation of the "brute force" approach of computing the
00593  * distance from every colormap entry to every histogram cell.  Unfortunately,
00594  * it needs a work array to hold the best-distance-so-far for each histogram
00595  * cell (because the inner loop has to be over cells, not colormap entries).
00596  * The work array elements have to be INT32s, so the work array would need
00597  * 256Kb at our recommended precision.  This is not feasible in DOS machines.
00598  *
00599  * To get around these problems, we apply Thomas' method to compute the
00600  * nearest colors for only the cells within a small subbox of the histogram.
00601  * The work array need be only as big as the subbox, so the memory usage
00602  * problem is solved.  Furthermore, we need not fill subboxes that are never
00603  * referenced in pass2; many images use only part of the color gamut, so a
00604  * fair amount of work is saved.  An additional advantage of this
00605  * approach is that we can apply Heckbert's locality criterion to quickly
00606  * eliminate colormap entries that are far away from the subbox; typically
00607  * three-fourths of the colormap entries are rejected by Heckbert's criterion,
00608  * and we need not compute their distances to individual cells in the subbox.
00609  * The speed of this approach is heavily influenced by the subbox size: too
00610  * small means too much overhead, too big loses because Heckbert's criterion
00611  * can't eliminate as many colormap entries.  Empirically the best subbox
00612  * size seems to be about 1/512th of the histogram (1/8th in each direction).
00613  *
00614  * Thomas' article also describes a refined method which is asymptotically
00615  * faster than the brute-force method, but it is also far more complex and
00616  * cannot efficiently be applied to small subboxes.  It is therefore not
00617  * useful for programs intended to be portable to DOS machines.  On machines
00618  * with plenty of memory, filling the whole histogram in one shot with Thomas'
00619  * refined method might be faster than the present code --- but then again,
00620  * it might not be any faster, and it's certainly more complicated.
00621  */
00622 
00623 
00624 /* log2(histogram cells in update box) for each axis; this can be adjusted */
00625 #define BOX_C0_LOG  (HIST_C0_BITS-3)
00626 #define BOX_C1_LOG  (HIST_C1_BITS-3)
00627 #define BOX_C2_LOG  (HIST_C2_BITS-3)
00628 
00629 #define BOX_C0_ELEMS  (1<<BOX_C0_LOG) /* # of hist cells in update box */
00630 #define BOX_C1_ELEMS  (1<<BOX_C1_LOG)
00631 #define BOX_C2_ELEMS  (1<<BOX_C2_LOG)
00632 
00633 #define BOX_C0_SHIFT  (C0_SHIFT + BOX_C0_LOG)
00634 #define BOX_C1_SHIFT  (C1_SHIFT + BOX_C1_LOG)
00635 #define BOX_C2_SHIFT  (C2_SHIFT + BOX_C2_LOG)
00636 
00637 
00638 /*
00639  * The next three routines implement inverse colormap filling.  They could
00640  * all be folded into one big routine, but splitting them up this way saves
00641  * some stack space (the mindist[] and bestdist[] arrays need not coexist)
00642  * and may allow some compilers to produce better code by registerizing more
00643  * inner-loop variables.
00644  */
00645 
00646 LOCAL(int)
00647 find_nearby_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
00648             JSAMPLE colorlist[])
00649 /* Locate the colormap entries close enough to an update box to be candidates
00650  * for the nearest entry to some cell(s) in the update box.  The update box
00651  * is specified by the center coordinates of its first cell.  The number of
00652  * candidate colormap entries is returned, and their colormap indexes are
00653  * placed in colorlist[].
00654  * This routine uses Heckbert's "locally sorted search" criterion to select
00655  * the colors that need further consideration.
00656  */
00657 {
00658   int numcolors = cinfo->actual_number_of_colors;
00659   int maxc0, maxc1, maxc2;
00660   int centerc0, centerc1, centerc2;
00661   int i, x, ncolors;
00662   INT32 minmaxdist, min_dist, max_dist, tdist;
00663   INT32 mindist[MAXNUMCOLORS];  /* min distance to colormap entry i */
00664 
00665   /* Compute true coordinates of update box's upper corner and center.
00666    * Actually we compute the coordinates of the center of the upper-corner
00667    * histogram cell, which are the upper bounds of the volume we care about.
00668    * Note that since ">>" rounds down, the "center" values may be closer to
00669    * min than to max; hence comparisons to them must be "<=", not "<".
00670    */
00671   maxc0 = minc0 + ((1 << BOX_C0_SHIFT) - (1 << C0_SHIFT));
00672   centerc0 = (minc0 + maxc0) >> 1;
00673   maxc1 = minc1 + ((1 << BOX_C1_SHIFT) - (1 << C1_SHIFT));
00674   centerc1 = (minc1 + maxc1) >> 1;
00675   maxc2 = minc2 + ((1 << BOX_C2_SHIFT) - (1 << C2_SHIFT));
00676   centerc2 = (minc2 + maxc2) >> 1;
00677 
00678   /* For each color in colormap, find:
00679    *  1. its minimum squared-distance to any point in the update box
00680    *     (zero if color is within update box);
00681    *  2. its maximum squared-distance to any point in the update box.
00682    * Both of these can be found by considering only the corners of the box.
00683    * We save the minimum distance for each color in mindist[];
00684    * only the smallest maximum distance is of interest.
00685    */
00686   minmaxdist = 0x7FFFFFFFL;
00687 
00688   for (i = 0; i < numcolors; i++) {
00689     /* We compute the squared-c0-distance term, then add in the other two. */
00690     x = GETJSAMPLE(cinfo->colormap[0][i]);
00691     if (x < minc0) {
00692       tdist = (x - minc0) * C0_SCALE;
00693       min_dist = tdist*tdist;
00694       tdist = (x - maxc0) * C0_SCALE;
00695       max_dist = tdist*tdist;
00696     } else if (x > maxc0) {
00697       tdist = (x - maxc0) * C0_SCALE;
00698       min_dist = tdist*tdist;
00699       tdist = (x - minc0) * C0_SCALE;
00700       max_dist = tdist*tdist;
00701     } else {
00702       /* within cell range so no contribution to min_dist */
00703       min_dist = 0;
00704       if (x <= centerc0) {
00705     tdist = (x - maxc0) * C0_SCALE;
00706     max_dist = tdist*tdist;
00707       } else {
00708     tdist = (x - minc0) * C0_SCALE;
00709     max_dist = tdist*tdist;
00710       }
00711     }
00712 
00713     x = GETJSAMPLE(cinfo->colormap[1][i]);
00714     if (x < minc1) {
00715       tdist = (x - minc1) * C1_SCALE;
00716       min_dist += tdist*tdist;
00717       tdist = (x - maxc1) * C1_SCALE;
00718       max_dist += tdist*tdist;
00719     } else if (x > maxc1) {
00720       tdist = (x - maxc1) * C1_SCALE;
00721       min_dist += tdist*tdist;
00722       tdist = (x - minc1) * C1_SCALE;
00723       max_dist += tdist*tdist;
00724     } else {
00725       /* within cell range so no contribution to min_dist */
00726       if (x <= centerc1) {
00727     tdist = (x - maxc1) * C1_SCALE;
00728     max_dist += tdist*tdist;
00729       } else {
00730     tdist = (x - minc1) * C1_SCALE;
00731     max_dist += tdist*tdist;
00732       }
00733     }
00734 
00735     x = GETJSAMPLE(cinfo->colormap[2][i]);
00736     if (x < minc2) {
00737       tdist = (x - minc2) * C2_SCALE;
00738       min_dist += tdist*tdist;
00739       tdist = (x - maxc2) * C2_SCALE;
00740       max_dist += tdist*tdist;
00741     } else if (x > maxc2) {
00742       tdist = (x - maxc2) * C2_SCALE;
00743       min_dist += tdist*tdist;
00744       tdist = (x - minc2) * C2_SCALE;
00745       max_dist += tdist*tdist;
00746     } else {
00747       /* within cell range so no contribution to min_dist */
00748       if (x <= centerc2) {
00749     tdist = (x - maxc2) * C2_SCALE;
00750     max_dist += tdist*tdist;
00751       } else {
00752     tdist = (x - minc2) * C2_SCALE;
00753     max_dist += tdist*tdist;
00754       }
00755     }
00756 
00757     mindist[i] = min_dist;  /* save away the results */
00758     if (max_dist < minmaxdist)
00759       minmaxdist = max_dist;
00760   }
00761 
00762   /* Now we know that no cell in the update box is more than minmaxdist
00763    * away from some colormap entry.  Therefore, only colors that are
00764    * within minmaxdist of some part of the box need be considered.
00765    */
00766   ncolors = 0;
00767   for (i = 0; i < numcolors; i++) {
00768     if (mindist[i] <= minmaxdist)
00769       colorlist[ncolors++] = (JSAMPLE) i;
00770   }
00771   return ncolors;
00772 }
00773 
00774 
00775 LOCAL(void)
00776 find_best_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
00777           int numcolors, JSAMPLE colorlist[], JSAMPLE bestcolor[])
00778 /* Find the closest colormap entry for each cell in the update box,
00779  * given the list of candidate colors prepared by find_nearby_colors.
00780  * Return the indexes of the closest entries in the bestcolor[] array.
00781  * This routine uses Thomas' incremental distance calculation method to
00782  * find the distance from a colormap entry to successive cells in the box.
00783  */
00784 {
00785   int ic0, ic1, ic2;
00786   int i, icolor;
00787   register INT32 * bptr;    /* pointer into bestdist[] array */
00788   JSAMPLE * cptr;       /* pointer into bestcolor[] array */
00789   INT32 dist0, dist1;       /* initial distance values */
00790   register INT32 dist2;     /* current distance in inner loop */
00791   INT32 xx0, xx1;       /* distance increments */
00792   register INT32 xx2;
00793   INT32 inc0, inc1, inc2;   /* initial values for increments */
00794   /* This array holds the distance to the nearest-so-far color for each cell */
00795   INT32 bestdist[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
00796 
00797   /* Initialize best-distance for each cell of the update box */
00798   bptr = bestdist;
00799   for (i = BOX_C0_ELEMS*BOX_C1_ELEMS*BOX_C2_ELEMS-1; i >= 0; i--)
00800     *bptr++ = 0x7FFFFFFFL;
00801   
00802   /* For each color selected by find_nearby_colors,
00803    * compute its distance to the center of each cell in the box.
00804    * If that's less than best-so-far, update best distance and color number.
00805    */
00806   
00807   /* Nominal steps between cell centers ("x" in Thomas article) */
00808 #define STEP_C0  ((1 << C0_SHIFT) * C0_SCALE)
00809 #define STEP_C1  ((1 << C1_SHIFT) * C1_SCALE)
00810 #define STEP_C2  ((1 << C2_SHIFT) * C2_SCALE)
00811   
00812   for (i = 0; i < numcolors; i++) {
00813     icolor = GETJSAMPLE(colorlist[i]);
00814     /* Compute (square of) distance from minc0/c1/c2 to this color */
00815     inc0 = (minc0 - GETJSAMPLE(cinfo->colormap[0][icolor])) * C0_SCALE;
00816     dist0 = inc0*inc0;
00817     inc1 = (minc1 - GETJSAMPLE(cinfo->colormap[1][icolor])) * C1_SCALE;
00818     dist0 += inc1*inc1;
00819     inc2 = (minc2 - GETJSAMPLE(cinfo->colormap[2][icolor])) * C2_SCALE;
00820     dist0 += inc2*inc2;
00821     /* Form the initial difference increments */
00822     inc0 = inc0 * (2 * STEP_C0) + STEP_C0 * STEP_C0;
00823     inc1 = inc1 * (2 * STEP_C1) + STEP_C1 * STEP_C1;
00824     inc2 = inc2 * (2 * STEP_C2) + STEP_C2 * STEP_C2;
00825     /* Now loop over all cells in box, updating distance per Thomas method */
00826     bptr = bestdist;
00827     cptr = bestcolor;
00828     xx0 = inc0;
00829     for (ic0 = BOX_C0_ELEMS-1; ic0 >= 0; ic0--) {
00830       dist1 = dist0;
00831       xx1 = inc1;
00832       for (ic1 = BOX_C1_ELEMS-1; ic1 >= 0; ic1--) {
00833     dist2 = dist1;
00834     xx2 = inc2;
00835     for (ic2 = BOX_C2_ELEMS-1; ic2 >= 0; ic2--) {
00836       if (dist2 < *bptr) {
00837         *bptr = dist2;
00838         *cptr = (JSAMPLE) icolor;
00839       }
00840       dist2 += xx2;
00841       xx2 += 2 * STEP_C2 * STEP_C2;
00842       bptr++;
00843       cptr++;
00844     }
00845     dist1 += xx1;
00846     xx1 += 2 * STEP_C1 * STEP_C1;
00847       }
00848       dist0 += xx0;
00849       xx0 += 2 * STEP_C0 * STEP_C0;
00850     }
00851   }
00852 }
00853 
00854 
00855 LOCAL(void)
00856 fill_inverse_cmap (j_decompress_ptr cinfo, int c0, int c1, int c2)
00857 /* Fill the inverse-colormap entries in the update box that contains */
00858 /* histogram cell c0/c1/c2.  (Only that one cell MUST be filled, but */
00859 /* we can fill as many others as we wish.) */
00860 {
00861   my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
00862   hist3d histogram = cquantize->histogram;
00863   int minc0, minc1, minc2;  /* lower left corner of update box */
00864   int ic0, ic1, ic2;
00865   register JSAMPLE * cptr;  /* pointer into bestcolor[] array */
00866   register histptr cachep;  /* pointer into main cache array */
00867   /* This array lists the candidate colormap indexes. */
00868   JSAMPLE colorlist[MAXNUMCOLORS];
00869   int numcolors;        /* number of candidate colors */
00870   /* This array holds the actually closest colormap index for each cell. */
00871   JSAMPLE bestcolor[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
00872 
00873   /* Convert cell coordinates to update box ID */
00874   c0 >>= BOX_C0_LOG;
00875   c1 >>= BOX_C1_LOG;
00876   c2 >>= BOX_C2_LOG;
00877 
00878   /* Compute true coordinates of update box's origin corner.
00879    * Actually we compute the coordinates of the center of the corner
00880    * histogram cell, which are the lower bounds of the volume we care about.
00881    */
00882   minc0 = (c0 << BOX_C0_SHIFT) + ((1 << C0_SHIFT) >> 1);
00883   minc1 = (c1 << BOX_C1_SHIFT) + ((1 << C1_SHIFT) >> 1);
00884   minc2 = (c2 << BOX_C2_SHIFT) + ((1 << C2_SHIFT) >> 1);
00885   
00886   /* Determine which colormap entries are close enough to be candidates
00887    * for the nearest entry to some cell in the update box.
00888    */
00889   numcolors = find_nearby_colors(cinfo, minc0, minc1, minc2, colorlist);
00890 
00891   /* Determine the actually nearest colors. */
00892   find_best_colors(cinfo, minc0, minc1, minc2, numcolors, colorlist,
00893            bestcolor);
00894 
00895   /* Save the best color numbers (plus 1) in the main cache array */
00896   c0 <<= BOX_C0_LOG;        /* convert ID back to base cell indexes */
00897   c1 <<= BOX_C1_LOG;
00898   c2 <<= BOX_C2_LOG;
00899   cptr = bestcolor;
00900   for (ic0 = 0; ic0 < BOX_C0_ELEMS; ic0++) {
00901     for (ic1 = 0; ic1 < BOX_C1_ELEMS; ic1++) {
00902       cachep = & histogram[c0+ic0][c1+ic1][c2];
00903       for (ic2 = 0; ic2 < BOX_C2_ELEMS; ic2++) {
00904     *cachep++ = (histcell) (GETJSAMPLE(*cptr++) + 1);
00905       }
00906     }
00907   }
00908 }
00909 
00910 
00911 /*
00912  * Map some rows of pixels to the output colormapped representation.
00913  */
00914 
00915 METHODDEF(void)
00916 pass2_no_dither (j_decompress_ptr cinfo,
00917          JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
00918 /* This version performs no dithering */
00919 {
00920   my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
00921   hist3d histogram = cquantize->histogram;
00922   register JSAMPROW inptr, outptr;
00923   register histptr cachep;
00924   register int c0, c1, c2;
00925   int row;
00926   JDIMENSION col;
00927   JDIMENSION width = cinfo->output_width;
00928 
00929   for (row = 0; row < num_rows; row++) {
00930     inptr = input_buf[row];
00931     outptr = output_buf[row];
00932     for (col = width; col > 0; col--) {
00933       /* get pixel value and index into the cache */
00934       c0 = GETJSAMPLE(*inptr++) >> C0_SHIFT;
00935       c1 = GETJSAMPLE(*inptr++) >> C1_SHIFT;
00936       c2 = GETJSAMPLE(*inptr++) >> C2_SHIFT;
00937       cachep = & histogram[c0][c1][c2];
00938       /* If we have not seen this color before, find nearest colormap entry */
00939       /* and update the cache */
00940       if (*cachep == 0)
00941     fill_inverse_cmap(cinfo, c0,c1,c2);
00942       /* Now emit the colormap index for this cell */
00943       *outptr++ = (JSAMPLE) (*cachep - 1);
00944     }
00945   }
00946 }
00947 
00948 
00949 METHODDEF(void)
00950 pass2_fs_dither (j_decompress_ptr cinfo,
00951          JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
00952 /* This version performs Floyd-Steinberg dithering */
00953 {
00954   my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
00955   hist3d histogram = cquantize->histogram;
00956   register LOCFSERROR cur0, cur1, cur2; /* current error or pixel value */
00957   LOCFSERROR belowerr0, belowerr1, belowerr2; /* error for pixel below cur */
00958   LOCFSERROR bpreverr0, bpreverr1, bpreverr2; /* error for below/prev col */
00959   register FSERRPTR errorptr;   /* => fserrors[] at column before current */
00960   JSAMPROW inptr;       /* => current input pixel */
00961   JSAMPROW outptr;      /* => current output pixel */
00962   histptr cachep;
00963   int dir;          /* +1 or -1 depending on direction */
00964   int dir3;         /* 3*dir, for advancing inptr & errorptr */
00965   int row;
00966   JDIMENSION col;
00967   JDIMENSION width = cinfo->output_width;
00968   JSAMPLE *range_limit = cinfo->sample_range_limit;
00969   int *error_limit = cquantize->error_limiter;
00970   JSAMPROW colormap0 = cinfo->colormap[0];
00971   JSAMPROW colormap1 = cinfo->colormap[1];
00972   JSAMPROW colormap2 = cinfo->colormap[2];
00973   SHIFT_TEMPS
00974 
00975   for (row = 0; row < num_rows; row++) {
00976     inptr = input_buf[row];
00977     outptr = output_buf[row];
00978     if (cquantize->on_odd_row) {
00979       /* work right to left in this row */
00980       inptr += (width-1) * 3;   /* so point to rightmost pixel */
00981       outptr += width-1;
00982       dir = -1;
00983       dir3 = -3;
00984       errorptr = cquantize->fserrors + (width+1)*3; /* => entry after last column */
00985       cquantize->on_odd_row = FALSE; /* flip for next time */
00986     } else {
00987       /* work left to right in this row */
00988       dir = 1;
00989       dir3 = 3;
00990       errorptr = cquantize->fserrors; /* => entry before first real column */
00991       cquantize->on_odd_row = TRUE; /* flip for next time */
00992     }
00993     /* Preset error values: no error propagated to first pixel from left */
00994     cur0 = cur1 = cur2 = 0;
00995     /* and no error propagated to row below yet */
00996     belowerr0 = belowerr1 = belowerr2 = 0;
00997     bpreverr0 = bpreverr1 = bpreverr2 = 0;
00998 
00999     for (col = width; col > 0; col--) {
01000       /* curN holds the error propagated from the previous pixel on the
01001        * current line.  Add the error propagated from the previous line
01002        * to form the complete error correction term for this pixel, and
01003        * round the error term (which is expressed * 16) to an integer.
01004        * RIGHT_SHIFT rounds towards minus infinity, so adding 8 is correct
01005        * for either sign of the error value.
01006        * Note: errorptr points to *previous* column's array entry.
01007        */
01008       cur0 = RIGHT_SHIFT(cur0 + errorptr[dir3+0] + 8, 4);
01009       cur1 = RIGHT_SHIFT(cur1 + errorptr[dir3+1] + 8, 4);
01010       cur2 = RIGHT_SHIFT(cur2 + errorptr[dir3+2] + 8, 4);
01011       /* Limit the error using transfer function set by init_error_limit.
01012        * See comments with init_error_limit for rationale.
01013        */
01014       cur0 = error_limit[cur0];
01015       cur1 = error_limit[cur1];
01016       cur2 = error_limit[cur2];
01017       /* Form pixel value + error, and range-limit to 0..MAXJSAMPLE.
01018        * The maximum error is +- MAXJSAMPLE (or less with error limiting);
01019        * this sets the required size of the range_limit array.
01020        */
01021       cur0 += GETJSAMPLE(inptr[0]);
01022       cur1 += GETJSAMPLE(inptr[1]);
01023       cur2 += GETJSAMPLE(inptr[2]);
01024       cur0 = GETJSAMPLE(range_limit[cur0]);
01025       cur1 = GETJSAMPLE(range_limit[cur1]);
01026       cur2 = GETJSAMPLE(range_limit[cur2]);
01027       /* Index into the cache with adjusted pixel value */
01028       cachep = & histogram[cur0>>C0_SHIFT][cur1>>C1_SHIFT][cur2>>C2_SHIFT];
01029       /* If we have not seen this color before, find nearest colormap */
01030       /* entry and update the cache */
01031       if (*cachep == 0)
01032     fill_inverse_cmap(cinfo, cur0>>C0_SHIFT,cur1>>C1_SHIFT,cur2>>C2_SHIFT);
01033       /* Now emit the colormap index for this cell */
01034       { register int pixcode = *cachep - 1;
01035     *outptr = (JSAMPLE) pixcode;
01036     /* Compute representation error for this pixel */
01037     cur0 -= GETJSAMPLE(colormap0[pixcode]);
01038     cur1 -= GETJSAMPLE(colormap1[pixcode]);
01039     cur2 -= GETJSAMPLE(colormap2[pixcode]);
01040       }
01041       /* Compute error fractions to be propagated to adjacent pixels.
01042        * Add these into the running sums, and simultaneously shift the
01043        * next-line error sums left by 1 column.
01044        */
01045       { register LOCFSERROR bnexterr, delta;
01046 
01047     bnexterr = cur0;    /* Process component 0 */
01048     delta = cur0 * 2;
01049     cur0 += delta;      /* form error * 3 */
01050     errorptr[0] = (FSERROR) (bpreverr0 + cur0);
01051     cur0 += delta;      /* form error * 5 */
01052     bpreverr0 = belowerr0 + cur0;
01053     belowerr0 = bnexterr;
01054     cur0 += delta;      /* form error * 7 */
01055     bnexterr = cur1;    /* Process component 1 */
01056     delta = cur1 * 2;
01057     cur1 += delta;      /* form error * 3 */
01058     errorptr[1] = (FSERROR) (bpreverr1 + cur1);
01059     cur1 += delta;      /* form error * 5 */
01060     bpreverr1 = belowerr1 + cur1;
01061     belowerr1 = bnexterr;
01062     cur1 += delta;      /* form error * 7 */
01063     bnexterr = cur2;    /* Process component 2 */
01064     delta = cur2 * 2;
01065     cur2 += delta;      /* form error * 3 */
01066     errorptr[2] = (FSERROR) (bpreverr2 + cur2);
01067     cur2 += delta;      /* form error * 5 */
01068     bpreverr2 = belowerr2 + cur2;
01069     belowerr2 = bnexterr;
01070     cur2 += delta;      /* form error * 7 */
01071       }
01072       /* At this point curN contains the 7/16 error value to be propagated
01073        * to the next pixel on the current line, and all the errors for the
01074        * next line have been shifted over.  We are therefore ready to move on.
01075        */
01076       inptr += dir3;        /* Advance pixel pointers to next column */
01077       outptr += dir;
01078       errorptr += dir3;     /* advance errorptr to current column */
01079     }
01080     /* Post-loop cleanup: we must unload the final error values into the
01081      * final fserrors[] entry.  Note we need not unload belowerrN because
01082      * it is for the dummy column before or after the actual array.
01083      */
01084     errorptr[0] = (FSERROR) bpreverr0; /* unload prev errs into array */
01085     errorptr[1] = (FSERROR) bpreverr1;
01086     errorptr[2] = (FSERROR) bpreverr2;
01087   }
01088 }
01089 
01090 
01091 /*
01092  * Initialize the error-limiting transfer function (lookup table).
01093  * The raw F-S error computation can potentially compute error values of up to
01094  * +- MAXJSAMPLE.  But we want the maximum correction applied to a pixel to be
01095  * much less, otherwise obviously wrong pixels will be created.  (Typical
01096  * effects include weird fringes at color-area boundaries, isolated bright
01097  * pixels in a dark area, etc.)  The standard advice for avoiding this problem
01098  * is to ensure that the "corners" of the color cube are allocated as output
01099  * colors; then repeated errors in the same direction cannot cause cascading
01100  * error buildup.  However, that only prevents the error from getting
01101  * completely out of hand; Aaron Giles reports that error limiting improves
01102  * the results even with corner colors allocated.
01103  * A simple clamping of the error values to about +- MAXJSAMPLE/8 works pretty
01104  * well, but the smoother transfer function used below is even better.  Thanks
01105  * to Aaron Giles for this idea.
01106  */
01107 
01108 LOCAL(void)
01109 init_error_limit (j_decompress_ptr cinfo)
01110 /* Allocate and fill in the error_limiter table */
01111 {
01112   my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
01113   int * table;
01114   int in, out;
01115 
01116   table = (int *) (*cinfo->mem->alloc_small)
01117     ((j_common_ptr) cinfo, JPOOL_IMAGE, (MAXJSAMPLE*2+1) * SIZEOF(int));
01118   table += MAXJSAMPLE;      /* so can index -MAXJSAMPLE .. +MAXJSAMPLE */
01119   cquantize->error_limiter = table;
01120 
01121 #define STEPSIZE ((MAXJSAMPLE+1)/16)
01122   /* Map errors 1:1 up to +- MAXJSAMPLE/16 */
01123   out = 0;
01124   for (in = 0; in < STEPSIZE; in++, out++) {
01125     table[in] = out; table[-in] = -out;
01126   }
01127   /* Map errors 1:2 up to +- 3*MAXJSAMPLE/16 */
01128   for (; in < STEPSIZE*3; in++, out += (in&1) ? 0 : 1) {
01129     table[in] = out; table[-in] = -out;
01130   }
01131   /* Clamp the rest to final out value (which is (MAXJSAMPLE+1)/8) */
01132   for (; in <= MAXJSAMPLE; in++) {
01133     table[in] = out; table[-in] = -out;
01134   }
01135 #undef STEPSIZE
01136 }
01137 
01138 
01139 /*
01140  * Finish up at the end of each pass.
01141  */
01142 
01143 METHODDEF(void)
01144 finish_pass1 (j_decompress_ptr cinfo)
01145 {
01146   my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
01147 
01148   /* Select the representative colors and fill in cinfo->colormap */
01149   cinfo->colormap = cquantize->sv_colormap;
01150   select_colors(cinfo, cquantize->desired);
01151   /* Force next pass to zero the color index table */
01152   cquantize->needs_zeroed = TRUE;
01153 }
01154 
01155 
01156 METHODDEF(void)
01157 finish_pass2 (j_decompress_ptr cinfo)
01158 {
01159   /* no work */
01160 }
01161 
01162 
01163 /*
01164  * Initialize for each processing pass.
01165  */
01166 
01167 METHODDEF(void)
01168 start_pass_2_quant (j_decompress_ptr cinfo, boolean is_pre_scan)
01169 {
01170   my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
01171   hist3d histogram = cquantize->histogram;
01172   int i;
01173 
01174   /* Only F-S dithering or no dithering is supported. */
01175   /* If user asks for ordered dither, give him F-S. */
01176   if (cinfo->dither_mode != JDITHER_NONE)
01177     cinfo->dither_mode = JDITHER_FS;
01178 
01179   if (is_pre_scan) {
01180     /* Set up method pointers */
01181     cquantize->pub.color_quantize = prescan_quantize;
01182     cquantize->pub.finish_pass = finish_pass1;
01183     cquantize->needs_zeroed = TRUE; /* Always zero histogram */
01184   } else {
01185     /* Set up method pointers */
01186     if (cinfo->dither_mode == JDITHER_FS)
01187       cquantize->pub.color_quantize = pass2_fs_dither;
01188     else
01189       cquantize->pub.color_quantize = pass2_no_dither;
01190     cquantize->pub.finish_pass = finish_pass2;
01191 
01192     /* Make sure color count is acceptable */
01193     i = cinfo->actual_number_of_colors;
01194     if (i < 1)
01195       ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 1);
01196     if (i > MAXNUMCOLORS)
01197       ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
01198 
01199     if (cinfo->dither_mode == JDITHER_FS) {
01200       size_t arraysize = (size_t) ((cinfo->output_width + 2) *
01201                    (3 * SIZEOF(FSERROR)));
01202       /* Allocate Floyd-Steinberg workspace if we didn't already. */
01203       if (cquantize->fserrors == NULL)
01204     cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
01205       ((j_common_ptr) cinfo, JPOOL_IMAGE, arraysize);
01206       /* Initialize the propagated errors to zero. */
01207       FMEMZERO((void FAR *) cquantize->fserrors, arraysize);
01208       /* Make the error-limit table if we didn't already. */
01209       if (cquantize->error_limiter == NULL)
01210     init_error_limit(cinfo);
01211       cquantize->on_odd_row = FALSE;
01212     }
01213 
01214   }
01215   /* Zero the histogram or inverse color map, if necessary */
01216   if (cquantize->needs_zeroed) {
01217     for (i = 0; i < HIST_C0_ELEMS; i++) {
01218       FMEMZERO((void FAR *) histogram[i],
01219            HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
01220     }
01221     cquantize->needs_zeroed = FALSE;
01222   }
01223 }
01224 
01225 
01226 /*
01227  * Switch to a new external colormap between output passes.
01228  */
01229 
01230 METHODDEF(void)
01231 new_color_map_2_quant (j_decompress_ptr cinfo)
01232 {
01233   my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
01234 
01235   /* Reset the inverse color map */
01236   cquantize->needs_zeroed = TRUE;
01237 }
01238 
01239 
01240 /*
01241  * Module initialization routine for 2-pass color quantization.
01242  */
01243 
01244 GLOBAL(void)
01245 jinit_2pass_quantizer (j_decompress_ptr cinfo)
01246 {
01247   my_cquantize_ptr cquantize;
01248   int i;
01249 
01250   cquantize = (my_cquantize_ptr)
01251     (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
01252                 SIZEOF(my_cquantizer));
01253   cinfo->cquantize = (struct jpeg_color_quantizer *) cquantize;
01254   cquantize->pub.start_pass = start_pass_2_quant;
01255   cquantize->pub.new_color_map = new_color_map_2_quant;
01256   cquantize->fserrors = NULL;   /* flag optional arrays not allocated */
01257   cquantize->error_limiter = NULL;
01258 
01259   /* Make sure jdmaster didn't give me a case I can't handle */
01260   if (cinfo->out_color_components != 3)
01261     ERREXIT(cinfo, JERR_NOTIMPL);
01262 
01263   /* Allocate the histogram/inverse colormap storage */
01264   cquantize->histogram = (hist3d) (*cinfo->mem->alloc_small)
01265     ((j_common_ptr) cinfo, JPOOL_IMAGE, HIST_C0_ELEMS * SIZEOF(hist2d));
01266   for (i = 0; i < HIST_C0_ELEMS; i++) {
01267     cquantize->histogram[i] = (hist2d) (*cinfo->mem->alloc_large)
01268       ((j_common_ptr) cinfo, JPOOL_IMAGE,
01269        HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
01270   }
01271   cquantize->needs_zeroed = TRUE; /* histogram is garbage now */
01272 
01273   /* Allocate storage for the completed colormap, if required.
01274    * We do this now since it is FAR storage and may affect
01275    * the memory manager's space calculations.
01276    */
01277   if (cinfo->enable_2pass_quant) {
01278     /* Make sure color count is acceptable */
01279     int desired = cinfo->desired_number_of_colors;
01280     /* Lower bound on # of colors ... somewhat arbitrary as long as > 0 */
01281     if (desired < 8)
01282       ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 8);
01283     /* Make sure colormap indexes can be represented by JSAMPLEs */
01284     if (desired > MAXNUMCOLORS)
01285       ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
01286     cquantize->sv_colormap = (*cinfo->mem->alloc_sarray)
01287       ((j_common_ptr) cinfo,JPOOL_IMAGE, (JDIMENSION) desired, (JDIMENSION) 3);
01288     cquantize->desired = desired;
01289   } else
01290     cquantize->sv_colormap = NULL;
01291 
01292   /* Only F-S dithering or no dithering is supported. */
01293   /* If user asks for ordered dither, give him F-S. */
01294   if (cinfo->dither_mode != JDITHER_NONE)
01295     cinfo->dither_mode = JDITHER_FS;
01296 
01297   /* Allocate Floyd-Steinberg workspace if necessary.
01298    * This isn't really needed until pass 2, but again it is FAR storage.
01299    * Although we will cope with a later change in dither_mode,
01300    * we do not promise to honor max_memory_to_use if dither_mode changes.
01301    */
01302   if (cinfo->dither_mode == JDITHER_FS) {
01303     cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
01304       ((j_common_ptr) cinfo, JPOOL_IMAGE,
01305        (size_t) ((cinfo->output_width + 2) * (3 * SIZEOF(FSERROR))));
01306     /* Might as well create the error-limiting table too. */
01307     init_error_limit(cinfo);
01308   }
01309 }
01310 
01311 #endif /* QUANT_2PASS_SUPPORTED */