1 /* 2 * jquant2.c 3 * 4 * This file was part of the Independent JPEG Group's software: 5 * Copyright (C) 1991-1996, Thomas G. Lane. 6 * libjpeg-turbo Modifications: 7 * Copyright (C) 2009, D. R. Commander. 8 * For conditions of distribution and use, see the accompanying README file. 9 * 10 * This file contains 2-pass color quantization (color mapping) routines. 11 * These routines provide selection of a custom color map for an image, 12 * followed by mapping of the image to that color map, with optional 13 * Floyd-Steinberg dithering. 14 * It is also possible to use just the second pass to map to an arbitrary 15 * externally-given color map. 16 * 17 * Note: ordered dithering is not supported, since there isn't any fast 18 * way to compute intercolor distances; it's unclear that ordered dither's 19 * fundamental assumptions even hold with an irregularly spaced color map. 20 */ 21 22 #define JPEG_INTERNALS 23 #include "jinclude.h" 24 #include "jpeglib.h" 25 26 #ifdef QUANT_2PASS_SUPPORTED 27 28 29 /* 30 * This module implements the well-known Heckbert paradigm for color 31 * quantization. Most of the ideas used here can be traced back to 32 * Heckbert's seminal paper 33 * Heckbert, Paul. "Color Image Quantization for Frame Buffer Display", 34 * Proc. SIGGRAPH '82, Computer Graphics v.16 #3 (July 1982), pp 297-304. 35 * 36 * In the first pass over the image, we accumulate a histogram showing the 37 * usage count of each possible color. To keep the histogram to a reasonable 38 * size, we reduce the precision of the input; typical practice is to retain 39 * 5 or 6 bits per color, so that 8 or 4 different input values are counted 40 * in the same histogram cell. 41 * 42 * Next, the color-selection step begins with a box representing the whole 43 * color space, and repeatedly splits the "largest" remaining box until we 44 * have as many boxes as desired colors. Then the mean color in each 45 * remaining box becomes one of the possible output colors. 46 * 47 * The second pass over the image maps each input pixel to the closest output 48 * color (optionally after applying a Floyd-Steinberg dithering correction). 49 * This mapping is logically trivial, but making it go fast enough requires 50 * considerable care. 51 * 52 * Heckbert-style quantizers vary a good deal in their policies for choosing 53 * the "largest" box and deciding where to cut it. The particular policies 54 * used here have proved out well in experimental comparisons, but better ones 55 * may yet be found. 56 * 57 * In earlier versions of the IJG code, this module quantized in YCbCr color 58 * space, processing the raw upsampled data without a color conversion step. 59 * This allowed the color conversion math to be done only once per colormap 60 * entry, not once per pixel. However, that optimization precluded other 61 * useful optimizations (such as merging color conversion with upsampling) 62 * and it also interfered with desired capabilities such as quantizing to an 63 * externally-supplied colormap. We have therefore abandoned that approach. 64 * The present code works in the post-conversion color space, typically RGB. 65 * 66 * To improve the visual quality of the results, we actually work in scaled 67 * RGB space, giving G distances more weight than R, and R in turn more than 68 * B. To do everything in integer math, we must use integer scale factors. 69 * The 2/3/1 scale factors used here correspond loosely to the relative 70 * weights of the colors in the NTSC grayscale equation. 71 * If you want to use this code to quantize a non-RGB color space, you'll 72 * probably need to change these scale factors. 73 */ 74 75 #define R_SCALE 2 /* scale R distances by this much */ 76 #define G_SCALE 3 /* scale G distances by this much */ 77 #define B_SCALE 1 /* and B by this much */ 78 79 static const int c_scales[3]={R_SCALE, G_SCALE, B_SCALE}; 80 #define C0_SCALE c_scales[rgb_red[cinfo->out_color_space]] 81 #define C1_SCALE c_scales[rgb_green[cinfo->out_color_space]] 82 #define C2_SCALE c_scales[rgb_blue[cinfo->out_color_space]] 83 84 /* 85 * First we have the histogram data structure and routines for creating it. 86 * 87 * The number of bits of precision can be adjusted by changing these symbols. 88 * We recommend keeping 6 bits for G and 5 each for R and B. 89 * If you have plenty of memory and cycles, 6 bits all around gives marginally 90 * better results; if you are short of memory, 5 bits all around will save 91 * some space but degrade the results. 92 * To maintain a fully accurate histogram, we'd need to allocate a "long" 93 * (preferably unsigned long) for each cell. In practice this is overkill; 94 * we can get by with 16 bits per cell. Few of the cell counts will overflow, 95 * and clamping those that do overflow to the maximum value will give close- 96 * enough results. This reduces the recommended histogram size from 256Kb 97 * to 128Kb, which is a useful savings on PC-class machines. 98 * (In the second pass the histogram space is re-used for pixel mapping data; 99 * in that capacity, each cell must be able to store zero to the number of 100 * desired colors. 16 bits/cell is plenty for that too.) 101 * Since the JPEG code is intended to run in small memory model on 80x86 102 * machines, we can't just allocate the histogram in one chunk. Instead 103 * of a true 3-D array, we use a row of pointers to 2-D arrays. Each 104 * pointer corresponds to a C0 value (typically 2^5 = 32 pointers) and 105 * each 2-D array has 2^6*2^5 = 2048 or 2^6*2^6 = 4096 entries. Note that 106 * on 80x86 machines, the pointer row is in near memory but the actual 107 * arrays are in far memory (same arrangement as we use for image arrays). 108 */ 109 110 #define MAXNUMCOLORS (MAXJSAMPLE+1) /* maximum size of colormap */ 111 112 /* These will do the right thing for either R,G,B or B,G,R color order, 113 * but you may not like the results for other color orders. 114 */ 115 #define HIST_C0_BITS 5 /* bits of precision in R/B histogram */ 116 #define HIST_C1_BITS 6 /* bits of precision in G histogram */ 117 #define HIST_C2_BITS 5 /* bits of precision in B/R histogram */ 118 119 /* Number of elements along histogram axes. */ 120 #define HIST_C0_ELEMS (1<<HIST_C0_BITS) 121 #define HIST_C1_ELEMS (1<<HIST_C1_BITS) 122 #define HIST_C2_ELEMS (1<<HIST_C2_BITS) 123 124 /* These are the amounts to shift an input value to get a histogram index. */ 125 #define C0_SHIFT (BITS_IN_JSAMPLE-HIST_C0_BITS) 126 #define C1_SHIFT (BITS_IN_JSAMPLE-HIST_C1_BITS) 127 #define C2_SHIFT (BITS_IN_JSAMPLE-HIST_C2_BITS) 128 129 130 typedef UINT16 histcell; /* histogram cell; prefer an unsigned type */ 131 132 typedef histcell FAR * histptr; /* for pointers to histogram cells */ 133 134 typedef histcell hist1d[HIST_C2_ELEMS]; /* typedefs for the array */ 135 typedef hist1d FAR * hist2d; /* type for the 2nd-level pointers */ 136 typedef hist2d * hist3d; /* type for top-level pointer */ 137 138 139 /* Declarations for Floyd-Steinberg dithering. 140 * 141 * Errors are accumulated into the array fserrors[], at a resolution of 142 * 1/16th of a pixel count. The error at a given pixel is propagated 143 * to its not-yet-processed neighbors using the standard F-S fractions, 144 * ... (here) 7/16 145 * 3/16 5/16 1/16 146 * We work left-to-right on even rows, right-to-left on odd rows. 147 * 148 * We can get away with a single array (holding one row's worth of errors) 149 * by using it to store the current row's errors at pixel columns not yet 150 * processed, but the next row's errors at columns already processed. We 151 * need only a few extra variables to hold the errors immediately around the 152 * current column. (If we are lucky, those variables are in registers, but 153 * even if not, they're probably cheaper to access than array elements are.) 154 * 155 * The fserrors[] array has (#columns + 2) entries; the extra entry at 156 * each end saves us from special-casing the first and last pixels. 157 * Each entry is three values long, one value for each color component. 158 * 159 * Note: on a wide image, we might not have enough room in a PC's near data 160 * segment to hold the error array; so it is allocated with alloc_large. 161 */ 162 163 #if BITS_IN_JSAMPLE == 8 164 typedef INT16 FSERROR; /* 16 bits should be enough */ 165 typedef int LOCFSERROR; /* use 'int' for calculation temps */ 166 #else 167 typedef INT32 FSERROR; /* may need more than 16 bits */ 168 typedef INT32 LOCFSERROR; /* be sure calculation temps are big enough */ 169 #endif 170 171 typedef FSERROR FAR *FSERRPTR; /* pointer to error array (in FAR storage!) */ 172 173 174 /* Private subobject */ 175 176 typedef struct { 177 struct jpeg_color_quantizer pub; /* public fields */ 178 179 /* Space for the eventually created colormap is stashed here */ 180 JSAMPARRAY sv_colormap; /* colormap allocated at init time */ 181 int desired; /* desired # of colors = size of colormap */ 182 183 /* Variables for accumulating image statistics */ 184 hist3d histogram; /* pointer to the histogram */ 185 186 boolean needs_zeroed; /* TRUE if next pass must zero histogram */ 187 188 /* Variables for Floyd-Steinberg dithering */ 189 FSERRPTR fserrors; /* accumulated errors */ 190 boolean on_odd_row; /* flag to remember which row we are on */ 191 int * error_limiter; /* table for clamping the applied error */ 192 } my_cquantizer; 193 194 typedef my_cquantizer * my_cquantize_ptr; 195 196 197 /* 198 * Prescan some rows of pixels. 199 * In this module the prescan simply updates the histogram, which has been 200 * initialized to zeroes by start_pass. 201 * An output_buf parameter is required by the method signature, but no data 202 * is actually output (in fact the buffer controller is probably passing a 203 * NULL pointer). 204 */ 205 206 METHODDEF(void) 207 prescan_quantize (j_decompress_ptr cinfo, JSAMPARRAY input_buf, 208 JSAMPARRAY output_buf, int num_rows) 209 { 210 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize; 211 register JSAMPROW ptr; 212 register histptr histp; 213 register hist3d histogram = cquantize->histogram; 214 int row; 215 JDIMENSION col; 216 JDIMENSION width = cinfo->output_width; 217 218 for (row = 0; row < num_rows; row++) { 219 ptr = input_buf[row]; 220 for (col = width; col > 0; col--) { 221 /* get pixel value and index into the histogram */ 222 histp = & histogram[GETJSAMPLE(ptr[0]) >> C0_SHIFT] 223 [GETJSAMPLE(ptr[1]) >> C1_SHIFT] 224 [GETJSAMPLE(ptr[2]) >> C2_SHIFT]; 225 /* increment, check for overflow and undo increment if so. */ 226 if (++(*histp) <= 0) 227 (*histp)--; 228 ptr += 3; 229 } 230 } 231 } 232 233 234 /* 235 * Next we have the really interesting routines: selection of a colormap 236 * given the completed histogram. 237 * These routines work with a list of "boxes", each representing a rectangular 238 * subset of the input color space (to histogram precision). 239 */ 240 241 typedef struct { 242 /* The bounds of the box (inclusive); expressed as histogram indexes */ 243 int c0min, c0max; 244 int c1min, c1max; 245 int c2min, c2max; 246 /* The volume (actually 2-norm) of the box */ 247 INT32 volume; 248 /* The number of nonzero histogram cells within this box */ 249 long colorcount; 250 } box; 251 252 typedef box * boxptr; 253 254 255 LOCAL(boxptr) 256 find_biggest_color_pop (boxptr boxlist, int numboxes) 257 /* Find the splittable box with the largest color population */ 258 /* Returns NULL if no splittable boxes remain */ 259 { 260 register boxptr boxp; 261 register int i; 262 register long maxc = 0; 263 boxptr which = NULL; 264 265 for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) { 266 if (boxp->colorcount > maxc && boxp->volume > 0) { 267 which = boxp; 268 maxc = boxp->colorcount; 269 } 270 } 271 return which; 272 } 273 274 275 LOCAL(boxptr) 276 find_biggest_volume (boxptr boxlist, int numboxes) 277 /* Find the splittable box with the largest (scaled) volume */ 278 /* Returns NULL if no splittable boxes remain */ 279 { 280 register boxptr boxp; 281 register int i; 282 register INT32 maxv = 0; 283 boxptr which = NULL; 284 285 for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) { 286 if (boxp->volume > maxv) { 287 which = boxp; 288 maxv = boxp->volume; 289 } 290 } 291 return which; 292 } 293 294 295 LOCAL(void) 296 update_box (j_decompress_ptr cinfo, boxptr boxp) 297 /* Shrink the min/max bounds of a box to enclose only nonzero elements, */ 298 /* and recompute its volume and population */ 299 { 300 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize; 301 hist3d histogram = cquantize->histogram; 302 histptr histp; 303 int c0,c1,c2; 304 int c0min,c0max,c1min,c1max,c2min,c2max; 305 INT32 dist0,dist1,dist2; 306 long ccount; 307 308 c0min = boxp->c0min; c0max = boxp->c0max; 309 c1min = boxp->c1min; c1max = boxp->c1max; 310 c2min = boxp->c2min; c2max = boxp->c2max; 311 312 if (c0max > c0min) 313 for (c0 = c0min; c0 <= c0max; c0++) 314 for (c1 = c1min; c1 <= c1max; c1++) { 315 histp = & histogram[c0][c1][c2min]; 316 for (c2 = c2min; c2 <= c2max; c2++) 317 if (*histp++ != 0) { 318 boxp->c0min = c0min = c0; 319 goto have_c0min; 320 } 321 } 322 have_c0min: 323 if (c0max > c0min) 324 for (c0 = c0max; c0 >= c0min; c0--) 325 for (c1 = c1min; c1 <= c1max; c1++) { 326 histp = & histogram[c0][c1][c2min]; 327 for (c2 = c2min; c2 <= c2max; c2++) 328 if (*histp++ != 0) { 329 boxp->c0max = c0max = c0; 330 goto have_c0max; 331 } 332 } 333 have_c0max: 334 if (c1max > c1min) 335 for (c1 = c1min; c1 <= c1max; c1++) 336 for (c0 = c0min; c0 <= c0max; c0++) { 337 histp = & histogram[c0][c1][c2min]; 338 for (c2 = c2min; c2 <= c2max; c2++) 339 if (*histp++ != 0) { 340 boxp->c1min = c1min = c1; 341 goto have_c1min; 342 } 343 } 344 have_c1min: 345 if (c1max > c1min) 346 for (c1 = c1max; c1 >= c1min; c1--) 347 for (c0 = c0min; c0 <= c0max; c0++) { 348 histp = & histogram[c0][c1][c2min]; 349 for (c2 = c2min; c2 <= c2max; c2++) 350 if (*histp++ != 0) { 351 boxp->c1max = c1max = c1; 352 goto have_c1max; 353 } 354 } 355 have_c1max: 356 if (c2max > c2min) 357 for (c2 = c2min; c2 <= c2max; c2++) 358 for (c0 = c0min; c0 <= c0max; c0++) { 359 histp = & histogram[c0][c1min][c2]; 360 for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS) 361 if (*histp != 0) { 362 boxp->c2min = c2min = c2; 363 goto have_c2min; 364 } 365 } 366 have_c2min: 367 if (c2max > c2min) 368 for (c2 = c2max; c2 >= c2min; c2--) 369 for (c0 = c0min; c0 <= c0max; c0++) { 370 histp = & histogram[c0][c1min][c2]; 371 for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS) 372 if (*histp != 0) { 373 boxp->c2max = c2max = c2; 374 goto have_c2max; 375 } 376 } 377 have_c2max: 378 379 /* Update box volume. 380 * We use 2-norm rather than real volume here; this biases the method 381 * against making long narrow boxes, and it has the side benefit that 382 * a box is splittable iff norm > 0. 383 * Since the differences are expressed in histogram-cell units, 384 * we have to shift back to JSAMPLE units to get consistent distances; 385 * after which, we scale according to the selected distance scale factors. 386 */ 387 dist0 = ((c0max - c0min) << C0_SHIFT) * C0_SCALE; 388 dist1 = ((c1max - c1min) << C1_SHIFT) * C1_SCALE; 389 dist2 = ((c2max - c2min) << C2_SHIFT) * C2_SCALE; 390 boxp->volume = dist0*dist0 + dist1*dist1 + dist2*dist2; 391 392 /* Now scan remaining volume of box and compute population */ 393 ccount = 0; 394 for (c0 = c0min; c0 <= c0max; c0++) 395 for (c1 = c1min; c1 <= c1max; c1++) { 396 histp = & histogram[c0][c1][c2min]; 397 for (c2 = c2min; c2 <= c2max; c2++, histp++) 398 if (*histp != 0) { 399 ccount++; 400 } 401 } 402 boxp->colorcount = ccount; 403 } 404 405 406 LOCAL(int) 407 median_cut (j_decompress_ptr cinfo, boxptr boxlist, int numboxes, 408 int desired_colors) 409 /* Repeatedly select and split the largest box until we have enough boxes */ 410 { 411 int n,lb; 412 int c0,c1,c2,cmax; 413 register boxptr b1,b2; 414 415 while (numboxes < desired_colors) { 416 /* Select box to split. 417 * Current algorithm: by population for first half, then by volume. 418 */ 419 if (numboxes*2 <= desired_colors) { 420 b1 = find_biggest_color_pop(boxlist, numboxes); 421 } else { 422 b1 = find_biggest_volume(boxlist, numboxes); 423 } 424 if (b1 == NULL) /* no splittable boxes left! */ 425 break; 426 b2 = &boxlist[numboxes]; /* where new box will go */ 427 /* Copy the color bounds to the new box. */ 428 b2->c0max = b1->c0max; b2->c1max = b1->c1max; b2->c2max = b1->c2max; 429 b2->c0min = b1->c0min; b2->c1min = b1->c1min; b2->c2min = b1->c2min; 430 /* Choose which axis to split the box on. 431 * Current algorithm: longest scaled axis. 432 * See notes in update_box about scaling distances. 433 */ 434 c0 = ((b1->c0max - b1->c0min) << C0_SHIFT) * C0_SCALE; 435 c1 = ((b1->c1max - b1->c1min) << C1_SHIFT) * C1_SCALE; 436 c2 = ((b1->c2max - b1->c2min) << C2_SHIFT) * C2_SCALE; 437 /* We want to break any ties in favor of green, then red, blue last. 438 * This code does the right thing for R,G,B or B,G,R color orders only. 439 */ 440 if (rgb_red[cinfo->out_color_space] == 0) { 441 cmax = c1; n = 1; 442 if (c0 > cmax) { cmax = c0; n = 0; } 443 if (c2 > cmax) { n = 2; } 444 } 445 else { 446 cmax = c1; n = 1; 447 if (c2 > cmax) { cmax = c2; n = 2; } 448 if (c0 > cmax) { n = 0; } 449 } 450 /* Choose split point along selected axis, and update box bounds. 451 * Current algorithm: split at halfway point. 452 * (Since the box has been shrunk to minimum volume, 453 * any split will produce two nonempty subboxes.) 454 * Note that lb value is max for lower box, so must be < old max. 455 */ 456 switch (n) { 457 case 0: 458 lb = (b1->c0max + b1->c0min) / 2; 459 b1->c0max = lb; 460 b2->c0min = lb+1; 461 break; 462 case 1: 463 lb = (b1->c1max + b1->c1min) / 2; 464 b1->c1max = lb; 465 b2->c1min = lb+1; 466 break; 467 case 2: 468 lb = (b1->c2max + b1->c2min) / 2; 469 b1->c2max = lb; 470 b2->c2min = lb+1; 471 break; 472 } 473 /* Update stats for boxes */ 474 update_box(cinfo, b1); 475 update_box(cinfo, b2); 476 numboxes++; 477 } 478 return numboxes; 479 } 480 481 482 LOCAL(void) 483 compute_color (j_decompress_ptr cinfo, boxptr boxp, int icolor) 484 /* Compute representative color for a box, put it in colormap[icolor] */ 485 { 486 /* Current algorithm: mean weighted by pixels (not colors) */ 487 /* Note it is important to get the rounding correct! */ 488 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize; 489 hist3d histogram = cquantize->histogram; 490 histptr histp; 491 int c0,c1,c2; 492 int c0min,c0max,c1min,c1max,c2min,c2max; 493 long count; 494 long total = 0; 495 long c0total = 0; 496 long c1total = 0; 497 long c2total = 0; 498 499 c0min = boxp->c0min; c0max = boxp->c0max; 500 c1min = boxp->c1min; c1max = boxp->c1max; 501 c2min = boxp->c2min; c2max = boxp->c2max; 502 503 for (c0 = c0min; c0 <= c0max; c0++) 504 for (c1 = c1min; c1 <= c1max; c1++) { 505 histp = & histogram[c0][c1][c2min]; 506 for (c2 = c2min; c2 <= c2max; c2++) { 507 if ((count = *histp++) != 0) { 508 total += count; 509 c0total += ((c0 << C0_SHIFT) + ((1<<C0_SHIFT)>>1)) * count; 510 c1total += ((c1 << C1_SHIFT) + ((1<<C1_SHIFT)>>1)) * count; 511 c2total += ((c2 << C2_SHIFT) + ((1<<C2_SHIFT)>>1)) * count; 512 } 513 } 514 } 515 516 cinfo->colormap[0][icolor] = (JSAMPLE) ((c0total + (total>>1)) / total); 517 cinfo->colormap[1][icolor] = (JSAMPLE) ((c1total + (total>>1)) / total); 518 cinfo->colormap[2][icolor] = (JSAMPLE) ((c2total + (total>>1)) / total); 519 } 520 521 522 LOCAL(void) 523 select_colors (j_decompress_ptr cinfo, int desired_colors) 524 /* Master routine for color selection */ 525 { 526 boxptr boxlist; 527 int numboxes; 528 int i; 529 530 /* Allocate workspace for box list */ 531 boxlist = (boxptr) (*cinfo->mem->alloc_small) 532 ((j_common_ptr) cinfo, JPOOL_IMAGE, desired_colors * SIZEOF(box)); 533 /* Initialize one box containing whole space */ 534 numboxes = 1; 535 boxlist[0].c0min = 0; 536 boxlist[0].c0max = MAXJSAMPLE >> C0_SHIFT; 537 boxlist[0].c1min = 0; 538 boxlist[0].c1max = MAXJSAMPLE >> C1_SHIFT; 539 boxlist[0].c2min = 0; 540 boxlist[0].c2max = MAXJSAMPLE >> C2_SHIFT; 541 /* Shrink it to actually-used volume and set its statistics */ 542 update_box(cinfo, & boxlist[0]); 543 /* Perform median-cut to produce final box list */ 544 numboxes = median_cut(cinfo, boxlist, numboxes, desired_colors); 545 /* Compute the representative color for each box, fill colormap */ 546 for (i = 0; i < numboxes; i++) 547 compute_color(cinfo, & boxlist[i], i); 548 cinfo->actual_number_of_colors = numboxes; 549 TRACEMS1(cinfo, 1, JTRC_QUANT_SELECTED, numboxes); 550 } 551 552 553 /* 554 * These routines are concerned with the time-critical task of mapping input 555 * colors to the nearest color in the selected colormap. 556 * 557 * We re-use the histogram space as an "inverse color map", essentially a 558 * cache for the results of nearest-color searches. All colors within a 559 * histogram cell will be mapped to the same colormap entry, namely the one 560 * closest to the cell's center. This may not be quite the closest entry to 561 * the actual input color, but it's almost as good. A zero in the cache 562 * indicates we haven't found the nearest color for that cell yet; the array 563 * is cleared to zeroes before starting the mapping pass. When we find the 564 * nearest color for a cell, its colormap index plus one is recorded in the 565 * cache for future use. The pass2 scanning routines call fill_inverse_cmap 566 * when they need to use an unfilled entry in the cache. 567 * 568 * Our method of efficiently finding nearest colors is based on the "locally 569 * sorted search" idea described by Heckbert and on the incremental distance 570 * calculation described by Spencer W. Thomas in chapter III.1 of Graphics 571 * Gems II (James Arvo, ed. Academic Press, 1991). Thomas points out that 572 * the distances from a given colormap entry to each cell of the histogram can 573 * be computed quickly using an incremental method: the differences between 574 * distances to adjacent cells themselves differ by a constant. This allows a 575 * fairly fast implementation of the "brute force" approach of computing the 576 * distance from every colormap entry to every histogram cell. Unfortunately, 577 * it needs a work array to hold the best-distance-so-far for each histogram 578 * cell (because the inner loop has to be over cells, not colormap entries). 579 * The work array elements have to be INT32s, so the work array would need 580 * 256Kb at our recommended precision. This is not feasible in DOS machines. 581 * 582 * To get around these problems, we apply Thomas' method to compute the 583 * nearest colors for only the cells within a small subbox of the histogram. 584 * The work array need be only as big as the subbox, so the memory usage 585 * problem is solved. Furthermore, we need not fill subboxes that are never 586 * referenced in pass2; many images use only part of the color gamut, so a 587 * fair amount of work is saved. An additional advantage of this 588 * approach is that we can apply Heckbert's locality criterion to quickly 589 * eliminate colormap entries that are far away from the subbox; typically 590 * three-fourths of the colormap entries are rejected by Heckbert's criterion, 591 * and we need not compute their distances to individual cells in the subbox. 592 * The speed of this approach is heavily influenced by the subbox size: too 593 * small means too much overhead, too big loses because Heckbert's criterion 594 * can't eliminate as many colormap entries. Empirically the best subbox 595 * size seems to be about 1/512th of the histogram (1/8th in each direction). 596 * 597 * Thomas' article also describes a refined method which is asymptotically 598 * faster than the brute-force method, but it is also far more complex and 599 * cannot efficiently be applied to small subboxes. It is therefore not 600 * useful for programs intended to be portable to DOS machines. On machines 601 * with plenty of memory, filling the whole histogram in one shot with Thomas' 602 * refined method might be faster than the present code --- but then again, 603 * it might not be any faster, and it's certainly more complicated. 604 */ 605 606 607 /* log2(histogram cells in update box) for each axis; this can be adjusted */ 608 #define BOX_C0_LOG (HIST_C0_BITS-3) 609 #define BOX_C1_LOG (HIST_C1_BITS-3) 610 #define BOX_C2_LOG (HIST_C2_BITS-3) 611 612 #define BOX_C0_ELEMS (1<<BOX_C0_LOG) /* # of hist cells in update box */ 613 #define BOX_C1_ELEMS (1<<BOX_C1_LOG) 614 #define BOX_C2_ELEMS (1<<BOX_C2_LOG) 615 616 #define BOX_C0_SHIFT (C0_SHIFT + BOX_C0_LOG) 617 #define BOX_C1_SHIFT (C1_SHIFT + BOX_C1_LOG) 618 #define BOX_C2_SHIFT (C2_SHIFT + BOX_C2_LOG) 619 620 621 /* 622 * The next three routines implement inverse colormap filling. They could 623 * all be folded into one big routine, but splitting them up this way saves 624 * some stack space (the mindist[] and bestdist[] arrays need not coexist) 625 * and may allow some compilers to produce better code by registerizing more 626 * inner-loop variables. 627 */ 628 629 LOCAL(int) 630 find_nearby_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2, 631 JSAMPLE colorlist[]) 632 /* Locate the colormap entries close enough to an update box to be candidates 633 * for the nearest entry to some cell(s) in the update box. The update box 634 * is specified by the center coordinates of its first cell. The number of 635 * candidate colormap entries is returned, and their colormap indexes are 636 * placed in colorlist[]. 637 * This routine uses Heckbert's "locally sorted search" criterion to select 638 * the colors that need further consideration. 639 */ 640 { 641 int numcolors = cinfo->actual_number_of_colors; 642 int maxc0, maxc1, maxc2; 643 int centerc0, centerc1, centerc2; 644 int i, x, ncolors; 645 INT32 minmaxdist, min_dist, max_dist, tdist; 646 INT32 mindist[MAXNUMCOLORS]; /* min distance to colormap entry i */ 647 648 /* Compute true coordinates of update box's upper corner and center. 649 * Actually we compute the coordinates of the center of the upper-corner 650 * histogram cell, which are the upper bounds of the volume we care about. 651 * Note that since ">>" rounds down, the "center" values may be closer to 652 * min than to max; hence comparisons to them must be "<=", not "<". 653 */ 654 maxc0 = minc0 + ((1 << BOX_C0_SHIFT) - (1 << C0_SHIFT)); 655 centerc0 = (minc0 + maxc0) >> 1; 656 maxc1 = minc1 + ((1 << BOX_C1_SHIFT) - (1 << C1_SHIFT)); 657 centerc1 = (minc1 + maxc1) >> 1; 658 maxc2 = minc2 + ((1 << BOX_C2_SHIFT) - (1 << C2_SHIFT)); 659 centerc2 = (minc2 + maxc2) >> 1; 660 661 /* For each color in colormap, find: 662 * 1. its minimum squared-distance to any point in the update box 663 * (zero if color is within update box); 664 * 2. its maximum squared-distance to any point in the update box. 665 * Both of these can be found by considering only the corners of the box. 666 * We save the minimum distance for each color in mindist[]; 667 * only the smallest maximum distance is of interest. 668 */ 669 minmaxdist = 0x7FFFFFFFL; 670 671 for (i = 0; i < numcolors; i++) { 672 /* We compute the squared-c0-distance term, then add in the other two. */ 673 x = GETJSAMPLE(cinfo->colormap[0][i]); 674 if (x < minc0) { 675 tdist = (x - minc0) * C0_SCALE; 676 min_dist = tdist*tdist; 677 tdist = (x - maxc0) * C0_SCALE; 678 max_dist = tdist*tdist; 679 } else if (x > maxc0) { 680 tdist = (x - maxc0) * C0_SCALE; 681 min_dist = tdist*tdist; 682 tdist = (x - minc0) * C0_SCALE; 683 max_dist = tdist*tdist; 684 } else { 685 /* within cell range so no contribution to min_dist */ 686 min_dist = 0; 687 if (x <= centerc0) { 688 tdist = (x - maxc0) * C0_SCALE; 689 max_dist = tdist*tdist; 690 } else { 691 tdist = (x - minc0) * C0_SCALE; 692 max_dist = tdist*tdist; 693 } 694 } 695 696 x = GETJSAMPLE(cinfo->colormap[1][i]); 697 if (x < minc1) { 698 tdist = (x - minc1) * C1_SCALE; 699 min_dist += tdist*tdist; 700 tdist = (x - maxc1) * C1_SCALE; 701 max_dist += tdist*tdist; 702 } else if (x > maxc1) { 703 tdist = (x - maxc1) * C1_SCALE; 704 min_dist += tdist*tdist; 705 tdist = (x - minc1) * C1_SCALE; 706 max_dist += tdist*tdist; 707 } else { 708 /* within cell range so no contribution to min_dist */ 709 if (x <= centerc1) { 710 tdist = (x - maxc1) * C1_SCALE; 711 max_dist += tdist*tdist; 712 } else { 713 tdist = (x - minc1) * C1_SCALE; 714 max_dist += tdist*tdist; 715 } 716 } 717 718 x = GETJSAMPLE(cinfo->colormap[2][i]); 719 if (x < minc2) { 720 tdist = (x - minc2) * C2_SCALE; 721 min_dist += tdist*tdist; 722 tdist = (x - maxc2) * C2_SCALE; 723 max_dist += tdist*tdist; 724 } else if (x > maxc2) { 725 tdist = (x - maxc2) * C2_SCALE; 726 min_dist += tdist*tdist; 727 tdist = (x - minc2) * C2_SCALE; 728 max_dist += tdist*tdist; 729 } else { 730 /* within cell range so no contribution to min_dist */ 731 if (x <= centerc2) { 732 tdist = (x - maxc2) * C2_SCALE; 733 max_dist += tdist*tdist; 734 } else { 735 tdist = (x - minc2) * C2_SCALE; 736 max_dist += tdist*tdist; 737 } 738 } 739 740 mindist[i] = min_dist; /* save away the results */ 741 if (max_dist < minmaxdist) 742 minmaxdist = max_dist; 743 } 744 745 /* Now we know that no cell in the update box is more than minmaxdist 746 * away from some colormap entry. Therefore, only colors that are 747 * within minmaxdist of some part of the box need be considered. 748 */ 749 ncolors = 0; 750 for (i = 0; i < numcolors; i++) { 751 if (mindist[i] <= minmaxdist) 752 colorlist[ncolors++] = (JSAMPLE) i; 753 } 754 return ncolors; 755 } 756 757 758 LOCAL(void) 759 find_best_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2, 760 int numcolors, JSAMPLE colorlist[], JSAMPLE bestcolor[]) 761 /* Find the closest colormap entry for each cell in the update box, 762 * given the list of candidate colors prepared by find_nearby_colors. 763 * Return the indexes of the closest entries in the bestcolor[] array. 764 * This routine uses Thomas' incremental distance calculation method to 765 * find the distance from a colormap entry to successive cells in the box. 766 */ 767 { 768 int ic0, ic1, ic2; 769 int i, icolor; 770 register INT32 * bptr; /* pointer into bestdist[] array */ 771 JSAMPLE * cptr; /* pointer into bestcolor[] array */ 772 INT32 dist0, dist1; /* initial distance values */ 773 register INT32 dist2; /* current distance in inner loop */ 774 INT32 xx0, xx1; /* distance increments */ 775 register INT32 xx2; 776 INT32 inc0, inc1, inc2; /* initial values for increments */ 777 /* This array holds the distance to the nearest-so-far color for each cell */ 778 INT32 bestdist[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS]; 779 780 /* Initialize best-distance for each cell of the update box */ 781 bptr = bestdist; 782 for (i = BOX_C0_ELEMS*BOX_C1_ELEMS*BOX_C2_ELEMS-1; i >= 0; i--) 783 *bptr++ = 0x7FFFFFFFL; 784 785 /* For each color selected by find_nearby_colors, 786 * compute its distance to the center of each cell in the box. 787 * If that's less than best-so-far, update best distance and color number. 788 */ 789 790 /* Nominal steps between cell centers ("x" in Thomas article) */ 791 #define STEP_C0 ((1 << C0_SHIFT) * C0_SCALE) 792 #define STEP_C1 ((1 << C1_SHIFT) * C1_SCALE) 793 #define STEP_C2 ((1 << C2_SHIFT) * C2_SCALE) 794 795 for (i = 0; i < numcolors; i++) { 796 icolor = GETJSAMPLE(colorlist[i]); 797 /* Compute (square of) distance from minc0/c1/c2 to this color */ 798 inc0 = (minc0 - GETJSAMPLE(cinfo->colormap[0][icolor])) * C0_SCALE; 799 dist0 = inc0*inc0; 800 inc1 = (minc1 - GETJSAMPLE(cinfo->colormap[1][icolor])) * C1_SCALE; 801 dist0 += inc1*inc1; 802 inc2 = (minc2 - GETJSAMPLE(cinfo->colormap[2][icolor])) * C2_SCALE; 803 dist0 += inc2*inc2; 804 /* Form the initial difference increments */ 805 inc0 = inc0 * (2 * STEP_C0) + STEP_C0 * STEP_C0; 806 inc1 = inc1 * (2 * STEP_C1) + STEP_C1 * STEP_C1; 807 inc2 = inc2 * (2 * STEP_C2) + STEP_C2 * STEP_C2; 808 /* Now loop over all cells in box, updating distance per Thomas method */ 809 bptr = bestdist; 810 cptr = bestcolor; 811 xx0 = inc0; 812 for (ic0 = BOX_C0_ELEMS-1; ic0 >= 0; ic0--) { 813 dist1 = dist0; 814 xx1 = inc1; 815 for (ic1 = BOX_C1_ELEMS-1; ic1 >= 0; ic1--) { 816 dist2 = dist1; 817 xx2 = inc2; 818 for (ic2 = BOX_C2_ELEMS-1; ic2 >= 0; ic2--) { 819 if (dist2 < *bptr) { 820 *bptr = dist2; 821 *cptr = (JSAMPLE) icolor; 822 } 823 dist2 += xx2; 824 xx2 += 2 * STEP_C2 * STEP_C2; 825 bptr++; 826 cptr++; 827 } 828 dist1 += xx1; 829 xx1 += 2 * STEP_C1 * STEP_C1; 830 } 831 dist0 += xx0; 832 xx0 += 2 * STEP_C0 * STEP_C0; 833 } 834 } 835 } 836 837 838 LOCAL(void) 839 fill_inverse_cmap (j_decompress_ptr cinfo, int c0, int c1, int c2) 840 /* Fill the inverse-colormap entries in the update box that contains */ 841 /* histogram cell c0/c1/c2. (Only that one cell MUST be filled, but */ 842 /* we can fill as many others as we wish.) */ 843 { 844 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize; 845 hist3d histogram = cquantize->histogram; 846 int minc0, minc1, minc2; /* lower left corner of update box */ 847 int ic0, ic1, ic2; 848 register JSAMPLE * cptr; /* pointer into bestcolor[] array */ 849 register histptr cachep; /* pointer into main cache array */ 850 /* This array lists the candidate colormap indexes. */ 851 JSAMPLE colorlist[MAXNUMCOLORS]; 852 int numcolors; /* number of candidate colors */ 853 /* This array holds the actually closest colormap index for each cell. */ 854 JSAMPLE bestcolor[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS]; 855 856 /* Convert cell coordinates to update box ID */ 857 c0 >>= BOX_C0_LOG; 858 c1 >>= BOX_C1_LOG; 859 c2 >>= BOX_C2_LOG; 860 861 /* Compute true coordinates of update box's origin corner. 862 * Actually we compute the coordinates of the center of the corner 863 * histogram cell, which are the lower bounds of the volume we care about. 864 */ 865 minc0 = (c0 << BOX_C0_SHIFT) + ((1 << C0_SHIFT) >> 1); 866 minc1 = (c1 << BOX_C1_SHIFT) + ((1 << C1_SHIFT) >> 1); 867 minc2 = (c2 << BOX_C2_SHIFT) + ((1 << C2_SHIFT) >> 1); 868 869 /* Determine which colormap entries are close enough to be candidates 870 * for the nearest entry to some cell in the update box. 871 */ 872 numcolors = find_nearby_colors(cinfo, minc0, minc1, minc2, colorlist); 873 874 /* Determine the actually nearest colors. */ 875 find_best_colors(cinfo, minc0, minc1, minc2, numcolors, colorlist, 876 bestcolor); 877 878 /* Save the best color numbers (plus 1) in the main cache array */ 879 c0 <<= BOX_C0_LOG; /* convert ID back to base cell indexes */ 880 c1 <<= BOX_C1_LOG; 881 c2 <<= BOX_C2_LOG; 882 cptr = bestcolor; 883 for (ic0 = 0; ic0 < BOX_C0_ELEMS; ic0++) { 884 for (ic1 = 0; ic1 < BOX_C1_ELEMS; ic1++) { 885 cachep = & histogram[c0+ic0][c1+ic1][c2]; 886 for (ic2 = 0; ic2 < BOX_C2_ELEMS; ic2++) { 887 *cachep++ = (histcell) (GETJSAMPLE(*cptr++) + 1); 888 } 889 } 890 } 891 } 892 893 894 /* 895 * Map some rows of pixels to the output colormapped representation. 896 */ 897 898 METHODDEF(void) 899 pass2_no_dither (j_decompress_ptr cinfo, 900 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows) 901 /* This version performs no dithering */ 902 { 903 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize; 904 hist3d histogram = cquantize->histogram; 905 register JSAMPROW inptr, outptr; 906 register histptr cachep; 907 register int c0, c1, c2; 908 int row; 909 JDIMENSION col; 910 JDIMENSION width = cinfo->output_width; 911 912 for (row = 0; row < num_rows; row++) { 913 inptr = input_buf[row]; 914 outptr = output_buf[row]; 915 for (col = width; col > 0; col--) { 916 /* get pixel value and index into the cache */ 917 c0 = GETJSAMPLE(*inptr++) >> C0_SHIFT; 918 c1 = GETJSAMPLE(*inptr++) >> C1_SHIFT; 919 c2 = GETJSAMPLE(*inptr++) >> C2_SHIFT; 920 cachep = & histogram[c0][c1][c2]; 921 /* If we have not seen this color before, find nearest colormap entry */ 922 /* and update the cache */ 923 if (*cachep == 0) 924 fill_inverse_cmap(cinfo, c0,c1,c2); 925 /* Now emit the colormap index for this cell */ 926 *outptr++ = (JSAMPLE) (*cachep - 1); 927 } 928 } 929 } 930 931 932 METHODDEF(void) 933 pass2_fs_dither (j_decompress_ptr cinfo, 934 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows) 935 /* This version performs Floyd-Steinberg dithering */ 936 { 937 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize; 938 hist3d histogram = cquantize->histogram; 939 register LOCFSERROR cur0, cur1, cur2; /* current error or pixel value */ 940 LOCFSERROR belowerr0, belowerr1, belowerr2; /* error for pixel below cur */ 941 LOCFSERROR bpreverr0, bpreverr1, bpreverr2; /* error for below/prev col */ 942 register FSERRPTR errorptr; /* => fserrors[] at column before current */ 943 JSAMPROW inptr; /* => current input pixel */ 944 JSAMPROW outptr; /* => current output pixel */ 945 histptr cachep; 946 int dir; /* +1 or -1 depending on direction */ 947 int dir3; /* 3*dir, for advancing inptr & errorptr */ 948 int row; 949 JDIMENSION col; 950 JDIMENSION width = cinfo->output_width; 951 JSAMPLE *range_limit = cinfo->sample_range_limit; 952 int *error_limit = cquantize->error_limiter; 953 JSAMPROW colormap0 = cinfo->colormap[0]; 954 JSAMPROW colormap1 = cinfo->colormap[1]; 955 JSAMPROW colormap2 = cinfo->colormap[2]; 956 SHIFT_TEMPS 957 958 for (row = 0; row < num_rows; row++) { 959 inptr = input_buf[row]; 960 outptr = output_buf[row]; 961 if (cquantize->on_odd_row) { 962 /* work right to left in this row */ 963 inptr += (width-1) * 3; /* so point to rightmost pixel */ 964 outptr += width-1; 965 dir = -1; 966 dir3 = -3; 967 errorptr = cquantize->fserrors + (width+1)*3; /* => entry after last column */ 968 cquantize->on_odd_row = FALSE; /* flip for next time */ 969 } else { 970 /* work left to right in this row */ 971 dir = 1; 972 dir3 = 3; 973 errorptr = cquantize->fserrors; /* => entry before first real column */ 974 cquantize->on_odd_row = TRUE; /* flip for next time */ 975 } 976 /* Preset error values: no error propagated to first pixel from left */ 977 cur0 = cur1 = cur2 = 0; 978 /* and no error propagated to row below yet */ 979 belowerr0 = belowerr1 = belowerr2 = 0; 980 bpreverr0 = bpreverr1 = bpreverr2 = 0; 981 982 for (col = width; col > 0; col--) { 983 /* curN holds the error propagated from the previous pixel on the 984 * current line. Add the error propagated from the previous line 985 * to form the complete error correction term for this pixel, and 986 * round the error term (which is expressed * 16) to an integer. 987 * RIGHT_SHIFT rounds towards minus infinity, so adding 8 is correct 988 * for either sign of the error value. 989 * Note: errorptr points to *previous* column's array entry. 990 */ 991 cur0 = RIGHT_SHIFT(cur0 + errorptr[dir3+0] + 8, 4); 992 cur1 = RIGHT_SHIFT(cur1 + errorptr[dir3+1] + 8, 4); 993 cur2 = RIGHT_SHIFT(cur2 + errorptr[dir3+2] + 8, 4); 994 /* Limit the error using transfer function set by init_error_limit. 995 * See comments with init_error_limit for rationale. 996 */ 997 cur0 = error_limit[cur0]; 998 cur1 = error_limit[cur1]; 999 cur2 = error_limit[cur2]; 1000 /* Form pixel value + error, and range-limit to 0..MAXJSAMPLE. 1001 * The maximum error is +- MAXJSAMPLE (or less with error limiting); 1002 * this sets the required size of the range_limit array. 1003 */ 1004 cur0 += GETJSAMPLE(inptr[0]); 1005 cur1 += GETJSAMPLE(inptr[1]); 1006 cur2 += GETJSAMPLE(inptr[2]); 1007 cur0 = GETJSAMPLE(range_limit[cur0]); 1008 cur1 = GETJSAMPLE(range_limit[cur1]); 1009 cur2 = GETJSAMPLE(range_limit[cur2]); 1010 /* Index into the cache with adjusted pixel value */ 1011 cachep = & histogram[cur0>>C0_SHIFT][cur1>>C1_SHIFT][cur2>>C2_SHIFT]; 1012 /* If we have not seen this color before, find nearest colormap */ 1013 /* entry and update the cache */ 1014 if (*cachep == 0) 1015 fill_inverse_cmap(cinfo, cur0>>C0_SHIFT,cur1>>C1_SHIFT,cur2>>C2_SHIFT); 1016 /* Now emit the colormap index for this cell */ 1017 { register int pixcode = *cachep - 1; 1018 *outptr = (JSAMPLE) pixcode; 1019 /* Compute representation error for this pixel */ 1020 cur0 -= GETJSAMPLE(colormap0[pixcode]); 1021 cur1 -= GETJSAMPLE(colormap1[pixcode]); 1022 cur2 -= GETJSAMPLE(colormap2[pixcode]); 1023 } 1024 /* Compute error fractions to be propagated to adjacent pixels. 1025 * Add these into the running sums, and simultaneously shift the 1026 * next-line error sums left by 1 column. 1027 */ 1028 { register LOCFSERROR bnexterr, delta; 1029 1030 bnexterr = cur0; /* Process component 0 */ 1031 delta = cur0 * 2; 1032 cur0 += delta; /* form error * 3 */ 1033 errorptr[0] = (FSERROR) (bpreverr0 + cur0); 1034 cur0 += delta; /* form error * 5 */ 1035 bpreverr0 = belowerr0 + cur0; 1036 belowerr0 = bnexterr; 1037 cur0 += delta; /* form error * 7 */ 1038 bnexterr = cur1; /* Process component 1 */ 1039 delta = cur1 * 2; 1040 cur1 += delta; /* form error * 3 */ 1041 errorptr[1] = (FSERROR) (bpreverr1 + cur1); 1042 cur1 += delta; /* form error * 5 */ 1043 bpreverr1 = belowerr1 + cur1; 1044 belowerr1 = bnexterr; 1045 cur1 += delta; /* form error * 7 */ 1046 bnexterr = cur2; /* Process component 2 */ 1047 delta = cur2 * 2; 1048 cur2 += delta; /* form error * 3 */ 1049 errorptr[2] = (FSERROR) (bpreverr2 + cur2); 1050 cur2 += delta; /* form error * 5 */ 1051 bpreverr2 = belowerr2 + cur2; 1052 belowerr2 = bnexterr; 1053 cur2 += delta; /* form error * 7 */ 1054 } 1055 /* At this point curN contains the 7/16 error value to be propagated 1056 * to the next pixel on the current line, and all the errors for the 1057 * next line have been shifted over. We are therefore ready to move on. 1058 */ 1059 inptr += dir3; /* Advance pixel pointers to next column */ 1060 outptr += dir; 1061 errorptr += dir3; /* advance errorptr to current column */ 1062 } 1063 /* Post-loop cleanup: we must unload the final error values into the 1064 * final fserrors[] entry. Note we need not unload belowerrN because 1065 * it is for the dummy column before or after the actual array. 1066 */ 1067 errorptr[0] = (FSERROR) bpreverr0; /* unload prev errs into array */ 1068 errorptr[1] = (FSERROR) bpreverr1; 1069 errorptr[2] = (FSERROR) bpreverr2; 1070 } 1071 } 1072 1073 1074 /* 1075 * Initialize the error-limiting transfer function (lookup table). 1076 * The raw F-S error computation can potentially compute error values of up to 1077 * +- MAXJSAMPLE. But we want the maximum correction applied to a pixel to be 1078 * much less, otherwise obviously wrong pixels will be created. (Typical 1079 * effects include weird fringes at color-area boundaries, isolated bright 1080 * pixels in a dark area, etc.) The standard advice for avoiding this problem 1081 * is to ensure that the "corners" of the color cube are allocated as output 1082 * colors; then repeated errors in the same direction cannot cause cascading 1083 * error buildup. However, that only prevents the error from getting 1084 * completely out of hand; Aaron Giles reports that error limiting improves 1085 * the results even with corner colors allocated. 1086 * A simple clamping of the error values to about +- MAXJSAMPLE/8 works pretty 1087 * well, but the smoother transfer function used below is even better. Thanks 1088 * to Aaron Giles for this idea. 1089 */ 1090 1091 LOCAL(void) 1092 init_error_limit (j_decompress_ptr cinfo) 1093 /* Allocate and fill in the error_limiter table */ 1094 { 1095 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize; 1096 int * table; 1097 int in, out; 1098 1099 table = (int *) (*cinfo->mem->alloc_small) 1100 ((j_common_ptr) cinfo, JPOOL_IMAGE, (MAXJSAMPLE*2+1) * SIZEOF(int)); 1101 table += MAXJSAMPLE; /* so can index -MAXJSAMPLE .. +MAXJSAMPLE */ 1102 cquantize->error_limiter = table; 1103 1104 #define STEPSIZE ((MAXJSAMPLE+1)/16) 1105 /* Map errors 1:1 up to +- MAXJSAMPLE/16 */ 1106 out = 0; 1107 for (in = 0; in < STEPSIZE; in++, out++) { 1108 table[in] = out; table[-in] = -out; 1109 } 1110 /* Map errors 1:2 up to +- 3*MAXJSAMPLE/16 */ 1111 for (; in < STEPSIZE*3; in++, out += (in&1) ? 0 : 1) { 1112 table[in] = out; table[-in] = -out; 1113 } 1114 /* Clamp the rest to final out value (which is (MAXJSAMPLE+1)/8) */ 1115 for (; in <= MAXJSAMPLE; in++) { 1116 table[in] = out; table[-in] = -out; 1117 } 1118 #undef STEPSIZE 1119 } 1120 1121 1122 /* 1123 * Finish up at the end of each pass. 1124 */ 1125 1126 METHODDEF(void) 1127 finish_pass1 (j_decompress_ptr cinfo) 1128 { 1129 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize; 1130 1131 /* Select the representative colors and fill in cinfo->colormap */ 1132 cinfo->colormap = cquantize->sv_colormap; 1133 select_colors(cinfo, cquantize->desired); 1134 /* Force next pass to zero the color index table */ 1135 cquantize->needs_zeroed = TRUE; 1136 } 1137 1138 1139 METHODDEF(void) 1140 finish_pass2 (j_decompress_ptr cinfo) 1141 { 1142 /* no work */ 1143 } 1144 1145 1146 /* 1147 * Initialize for each processing pass. 1148 */ 1149 1150 METHODDEF(void) 1151 start_pass_2_quant (j_decompress_ptr cinfo, boolean is_pre_scan) 1152 { 1153 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize; 1154 hist3d histogram = cquantize->histogram; 1155 int i; 1156 1157 /* Only F-S dithering or no dithering is supported. */ 1158 /* If user asks for ordered dither, give him F-S. */ 1159 if (cinfo->dither_mode != JDITHER_NONE) 1160 cinfo->dither_mode = JDITHER_FS; 1161 1162 if (is_pre_scan) { 1163 /* Set up method pointers */ 1164 cquantize->pub.color_quantize = prescan_quantize; 1165 cquantize->pub.finish_pass = finish_pass1; 1166 cquantize->needs_zeroed = TRUE; /* Always zero histogram */ 1167 } else { 1168 /* Set up method pointers */ 1169 if (cinfo->dither_mode == JDITHER_FS) 1170 cquantize->pub.color_quantize = pass2_fs_dither; 1171 else 1172 cquantize->pub.color_quantize = pass2_no_dither; 1173 cquantize->pub.finish_pass = finish_pass2; 1174 1175 /* Make sure color count is acceptable */ 1176 i = cinfo->actual_number_of_colors; 1177 if (i < 1) 1178 ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 1); 1179 if (i > MAXNUMCOLORS) 1180 ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS); 1181 1182 if (cinfo->dither_mode == JDITHER_FS) { 1183 size_t arraysize = (size_t) ((cinfo->output_width + 2) * 1184 (3 * SIZEOF(FSERROR))); 1185 /* Allocate Floyd-Steinberg workspace if we didn't already. */ 1186 if (cquantize->fserrors == NULL) 1187 cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large) 1188 ((j_common_ptr) cinfo, JPOOL_IMAGE, arraysize); 1189 /* Initialize the propagated errors to zero. */ 1190 jzero_far((void FAR *) cquantize->fserrors, arraysize); 1191 /* Make the error-limit table if we didn't already. */ 1192 if (cquantize->error_limiter == NULL) 1193 init_error_limit(cinfo); 1194 cquantize->on_odd_row = FALSE; 1195 } 1196 1197 } 1198 /* Zero the histogram or inverse color map, if necessary */ 1199 if (cquantize->needs_zeroed) { 1200 for (i = 0; i < HIST_C0_ELEMS; i++) { 1201 jzero_far((void FAR *) histogram[i], 1202 HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell)); 1203 } 1204 cquantize->needs_zeroed = FALSE; 1205 } 1206 } 1207 1208 1209 /* 1210 * Switch to a new external colormap between output passes. 1211 */ 1212 1213 METHODDEF(void) 1214 new_color_map_2_quant (j_decompress_ptr cinfo) 1215 { 1216 my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize; 1217 1218 /* Reset the inverse color map */ 1219 cquantize->needs_zeroed = TRUE; 1220 } 1221 1222 1223 /* 1224 * Module initialization routine for 2-pass color quantization. 1225 */ 1226 1227 GLOBAL(void) 1228 jinit_2pass_quantizer (j_decompress_ptr cinfo) 1229 { 1230 my_cquantize_ptr cquantize; 1231 int i; 1232 1233 cquantize = (my_cquantize_ptr) 1234 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE, 1235 SIZEOF(my_cquantizer)); 1236 cinfo->cquantize = (struct jpeg_color_quantizer *) cquantize; 1237 cquantize->pub.start_pass = start_pass_2_quant; 1238 cquantize->pub.new_color_map = new_color_map_2_quant; 1239 cquantize->fserrors = NULL; /* flag optional arrays not allocated */ 1240 cquantize->error_limiter = NULL; 1241 1242 /* Make sure jdmaster didn't give me a case I can't handle */ 1243 if (cinfo->out_color_components != 3) 1244 ERREXIT(cinfo, JERR_NOTIMPL); 1245 1246 /* Allocate the histogram/inverse colormap storage */ 1247 cquantize->histogram = (hist3d) (*cinfo->mem->alloc_small) 1248 ((j_common_ptr) cinfo, JPOOL_IMAGE, HIST_C0_ELEMS * SIZEOF(hist2d)); 1249 for (i = 0; i < HIST_C0_ELEMS; i++) { 1250 cquantize->histogram[i] = (hist2d) (*cinfo->mem->alloc_large) 1251 ((j_common_ptr) cinfo, JPOOL_IMAGE, 1252 HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell)); 1253 } 1254 cquantize->needs_zeroed = TRUE; /* histogram is garbage now */ 1255 1256 /* Allocate storage for the completed colormap, if required. 1257 * We do this now since it is FAR storage and may affect 1258 * the memory manager's space calculations. 1259 */ 1260 if (cinfo->enable_2pass_quant) { 1261 /* Make sure color count is acceptable */ 1262 int desired = cinfo->desired_number_of_colors; 1263 /* Lower bound on # of colors ... somewhat arbitrary as long as > 0 */ 1264 if (desired < 8) 1265 ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 8); 1266 /* Make sure colormap indexes can be represented by JSAMPLEs */ 1267 if (desired > MAXNUMCOLORS) 1268 ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS); 1269 cquantize->sv_colormap = (*cinfo->mem->alloc_sarray) 1270 ((j_common_ptr) cinfo,JPOOL_IMAGE, (JDIMENSION) desired, (JDIMENSION) 3); 1271 cquantize->desired = desired; 1272 } else 1273 cquantize->sv_colormap = NULL; 1274 1275 /* Only F-S dithering or no dithering is supported. */ 1276 /* If user asks for ordered dither, give him F-S. */ 1277 if (cinfo->dither_mode != JDITHER_NONE) 1278 cinfo->dither_mode = JDITHER_FS; 1279 1280 /* Allocate Floyd-Steinberg workspace if necessary. 1281 * This isn't really needed until pass 2, but again it is FAR storage. 1282 * Although we will cope with a later change in dither_mode, 1283 * we do not promise to honor max_memory_to_use if dither_mode changes. 1284 */ 1285 if (cinfo->dither_mode == JDITHER_FS) { 1286 cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large) 1287 ((j_common_ptr) cinfo, JPOOL_IMAGE, 1288 (size_t) ((cinfo->output_width + 2) * (3 * SIZEOF(FSERROR)))); 1289 /* Might as well create the error-limiting table too. */ 1290 init_error_limit(cinfo); 1291 } 1292 } 1293 1294 #endif /* QUANT_2PASS_SUPPORTED */ 1295
