Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 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: |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 7 | * Copyright (C) 2009, 2014-2015, 2020, D. R. Commander. |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 8 | * For conditions of distribution and use, see the accompanying README.ijg |
| 9 | * file. |
| 10 | * |
| 11 | * This file contains 2-pass color quantization (color mapping) routines. |
| 12 | * These routines provide selection of a custom color map for an image, |
| 13 | * followed by mapping of the image to that color map, with optional |
| 14 | * Floyd-Steinberg dithering. |
| 15 | * It is also possible to use just the second pass to map to an arbitrary |
| 16 | * externally-given color map. |
| 17 | * |
| 18 | * Note: ordered dithering is not supported, since there isn't any fast |
| 19 | * way to compute intercolor distances; it's unclear that ordered dither's |
| 20 | * fundamental assumptions even hold with an irregularly spaced color map. |
| 21 | */ |
| 22 | |
| 23 | #define JPEG_INTERNALS |
| 24 | #include "jinclude.h" |
| 25 | #include "jpeglib.h" |
| 26 | |
| 27 | #ifdef QUANT_2PASS_SUPPORTED |
| 28 | |
| 29 | |
| 30 | /* |
| 31 | * This module implements the well-known Heckbert paradigm for color |
| 32 | * quantization. Most of the ideas used here can be traced back to |
| 33 | * Heckbert's seminal paper |
| 34 | * Heckbert, Paul. "Color Image Quantization for Frame Buffer Display", |
| 35 | * Proc. SIGGRAPH '82, Computer Graphics v.16 #3 (July 1982), pp 297-304. |
| 36 | * |
| 37 | * In the first pass over the image, we accumulate a histogram showing the |
| 38 | * usage count of each possible color. To keep the histogram to a reasonable |
| 39 | * size, we reduce the precision of the input; typical practice is to retain |
| 40 | * 5 or 6 bits per color, so that 8 or 4 different input values are counted |
| 41 | * in the same histogram cell. |
| 42 | * |
| 43 | * Next, the color-selection step begins with a box representing the whole |
| 44 | * color space, and repeatedly splits the "largest" remaining box until we |
| 45 | * have as many boxes as desired colors. Then the mean color in each |
| 46 | * remaining box becomes one of the possible output colors. |
| 47 | * |
| 48 | * The second pass over the image maps each input pixel to the closest output |
| 49 | * color (optionally after applying a Floyd-Steinberg dithering correction). |
| 50 | * This mapping is logically trivial, but making it go fast enough requires |
| 51 | * considerable care. |
| 52 | * |
| 53 | * Heckbert-style quantizers vary a good deal in their policies for choosing |
| 54 | * the "largest" box and deciding where to cut it. The particular policies |
| 55 | * used here have proved out well in experimental comparisons, but better ones |
| 56 | * may yet be found. |
| 57 | * |
| 58 | * In earlier versions of the IJG code, this module quantized in YCbCr color |
| 59 | * space, processing the raw upsampled data without a color conversion step. |
| 60 | * This allowed the color conversion math to be done only once per colormap |
| 61 | * entry, not once per pixel. However, that optimization precluded other |
| 62 | * useful optimizations (such as merging color conversion with upsampling) |
| 63 | * and it also interfered with desired capabilities such as quantizing to an |
| 64 | * externally-supplied colormap. We have therefore abandoned that approach. |
| 65 | * The present code works in the post-conversion color space, typically RGB. |
| 66 | * |
| 67 | * To improve the visual quality of the results, we actually work in scaled |
| 68 | * RGB space, giving G distances more weight than R, and R in turn more than |
| 69 | * B. To do everything in integer math, we must use integer scale factors. |
| 70 | * The 2/3/1 scale factors used here correspond loosely to the relative |
| 71 | * weights of the colors in the NTSC grayscale equation. |
| 72 | * If you want to use this code to quantize a non-RGB color space, you'll |
| 73 | * probably need to change these scale factors. |
| 74 | */ |
| 75 | |
| 76 | #define R_SCALE 2 /* scale R distances by this much */ |
| 77 | #define G_SCALE 3 /* scale G distances by this much */ |
| 78 | #define B_SCALE 1 /* and B by this much */ |
| 79 | |
| 80 | static const int c_scales[3] = { R_SCALE, G_SCALE, B_SCALE }; |
| 81 | #define C0_SCALE c_scales[rgb_red[cinfo->out_color_space]] |
| 82 | #define C1_SCALE c_scales[rgb_green[cinfo->out_color_space]] |
| 83 | #define C2_SCALE c_scales[rgb_blue[cinfo->out_color_space]] |
| 84 | |
| 85 | /* |
| 86 | * First we have the histogram data structure and routines for creating it. |
| 87 | * |
| 88 | * The number of bits of precision can be adjusted by changing these symbols. |
| 89 | * We recommend keeping 6 bits for G and 5 each for R and B. |
| 90 | * If you have plenty of memory and cycles, 6 bits all around gives marginally |
| 91 | * better results; if you are short of memory, 5 bits all around will save |
| 92 | * some space but degrade the results. |
| 93 | * To maintain a fully accurate histogram, we'd need to allocate a "long" |
| 94 | * (preferably unsigned long) for each cell. In practice this is overkill; |
| 95 | * we can get by with 16 bits per cell. Few of the cell counts will overflow, |
| 96 | * and clamping those that do overflow to the maximum value will give close- |
| 97 | * enough results. This reduces the recommended histogram size from 256Kb |
| 98 | * to 128Kb, which is a useful savings on PC-class machines. |
| 99 | * (In the second pass the histogram space is re-used for pixel mapping data; |
| 100 | * in that capacity, each cell must be able to store zero to the number of |
| 101 | * desired colors. 16 bits/cell is plenty for that too.) |
| 102 | * Since the JPEG code is intended to run in small memory model on 80x86 |
| 103 | * machines, we can't just allocate the histogram in one chunk. Instead |
| 104 | * of a true 3-D array, we use a row of pointers to 2-D arrays. Each |
| 105 | * pointer corresponds to a C0 value (typically 2^5 = 32 pointers) and |
| 106 | * each 2-D array has 2^6*2^5 = 2048 or 2^6*2^6 = 4096 entries. |
| 107 | */ |
| 108 | |
| 109 | #define MAXNUMCOLORS (MAXJSAMPLE + 1) /* maximum size of colormap */ |
| 110 | |
| 111 | /* These will do the right thing for either R,G,B or B,G,R color order, |
| 112 | * but you may not like the results for other color orders. |
| 113 | */ |
| 114 | #define HIST_C0_BITS 5 /* bits of precision in R/B histogram */ |
| 115 | #define HIST_C1_BITS 6 /* bits of precision in G histogram */ |
| 116 | #define HIST_C2_BITS 5 /* bits of precision in B/R histogram */ |
| 117 | |
| 118 | /* Number of elements along histogram axes. */ |
| 119 | #define HIST_C0_ELEMS (1 << HIST_C0_BITS) |
| 120 | #define HIST_C1_ELEMS (1 << HIST_C1_BITS) |
| 121 | #define HIST_C2_ELEMS (1 << HIST_C2_BITS) |
| 122 | |
| 123 | /* These are the amounts to shift an input value to get a histogram index. */ |
| 124 | #define C0_SHIFT (BITS_IN_JSAMPLE - HIST_C0_BITS) |
| 125 | #define C1_SHIFT (BITS_IN_JSAMPLE - HIST_C1_BITS) |
| 126 | #define C2_SHIFT (BITS_IN_JSAMPLE - HIST_C2_BITS) |
| 127 | |
| 128 | |
| 129 | typedef UINT16 histcell; /* histogram cell; prefer an unsigned type */ |
| 130 | |
| 131 | typedef histcell *histptr; /* for pointers to histogram cells */ |
| 132 | |
| 133 | typedef histcell hist1d[HIST_C2_ELEMS]; /* typedefs for the array */ |
| 134 | typedef hist1d *hist2d; /* type for the 2nd-level pointers */ |
| 135 | typedef hist2d *hist3d; /* type for top-level pointer */ |
| 136 | |
| 137 | |
| 138 | /* Declarations for Floyd-Steinberg dithering. |
| 139 | * |
| 140 | * Errors are accumulated into the array fserrors[], at a resolution of |
| 141 | * 1/16th of a pixel count. The error at a given pixel is propagated |
| 142 | * to its not-yet-processed neighbors using the standard F-S fractions, |
| 143 | * ... (here) 7/16 |
| 144 | * 3/16 5/16 1/16 |
| 145 | * We work left-to-right on even rows, right-to-left on odd rows. |
| 146 | * |
| 147 | * We can get away with a single array (holding one row's worth of errors) |
| 148 | * by using it to store the current row's errors at pixel columns not yet |
| 149 | * processed, but the next row's errors at columns already processed. We |
| 150 | * need only a few extra variables to hold the errors immediately around the |
| 151 | * current column. (If we are lucky, those variables are in registers, but |
| 152 | * even if not, they're probably cheaper to access than array elements are.) |
| 153 | * |
| 154 | * The fserrors[] array has (#columns + 2) entries; the extra entry at |
| 155 | * each end saves us from special-casing the first and last pixels. |
| 156 | * Each entry is three values long, one value for each color component. |
| 157 | */ |
| 158 | |
| 159 | #if BITS_IN_JSAMPLE == 8 |
| 160 | typedef INT16 FSERROR; /* 16 bits should be enough */ |
| 161 | typedef int LOCFSERROR; /* use 'int' for calculation temps */ |
| 162 | #else |
| 163 | typedef JLONG FSERROR; /* may need more than 16 bits */ |
| 164 | typedef JLONG LOCFSERROR; /* be sure calculation temps are big enough */ |
| 165 | #endif |
| 166 | |
| 167 | typedef FSERROR *FSERRPTR; /* pointer to error array */ |
| 168 | |
| 169 | |
| 170 | /* Private subobject */ |
| 171 | |
| 172 | typedef struct { |
| 173 | struct jpeg_color_quantizer pub; /* public fields */ |
| 174 | |
| 175 | /* Space for the eventually created colormap is stashed here */ |
| 176 | JSAMPARRAY sv_colormap; /* colormap allocated at init time */ |
| 177 | int desired; /* desired # of colors = size of colormap */ |
| 178 | |
| 179 | /* Variables for accumulating image statistics */ |
| 180 | hist3d histogram; /* pointer to the histogram */ |
| 181 | |
| 182 | boolean needs_zeroed; /* TRUE if next pass must zero histogram */ |
| 183 | |
| 184 | /* Variables for Floyd-Steinberg dithering */ |
| 185 | FSERRPTR fserrors; /* accumulated errors */ |
| 186 | boolean on_odd_row; /* flag to remember which row we are on */ |
| 187 | int *error_limiter; /* table for clamping the applied error */ |
| 188 | } my_cquantizer; |
| 189 | |
| 190 | typedef my_cquantizer *my_cquantize_ptr; |
| 191 | |
| 192 | |
| 193 | /* |
| 194 | * Prescan some rows of pixels. |
| 195 | * In this module the prescan simply updates the histogram, which has been |
| 196 | * initialized to zeroes by start_pass. |
| 197 | * An output_buf parameter is required by the method signature, but no data |
| 198 | * is actually output (in fact the buffer controller is probably passing a |
| 199 | * NULL pointer). |
| 200 | */ |
| 201 | |
| 202 | METHODDEF(void) |
| 203 | prescan_quantize(j_decompress_ptr cinfo, JSAMPARRAY input_buf, |
| 204 | JSAMPARRAY output_buf, int num_rows) |
| 205 | { |
| 206 | my_cquantize_ptr cquantize = (my_cquantize_ptr)cinfo->cquantize; |
| 207 | register JSAMPROW ptr; |
| 208 | register histptr histp; |
| 209 | register hist3d histogram = cquantize->histogram; |
| 210 | int row; |
| 211 | JDIMENSION col; |
| 212 | JDIMENSION width = cinfo->output_width; |
| 213 | |
| 214 | for (row = 0; row < num_rows; row++) { |
| 215 | ptr = input_buf[row]; |
| 216 | for (col = width; col > 0; col--) { |
| 217 | /* get pixel value and index into the histogram */ |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 218 | histp = &histogram[ptr[0] >> C0_SHIFT] |
| 219 | [ptr[1] >> C1_SHIFT] |
| 220 | [ptr[2] >> C2_SHIFT]; |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 221 | /* increment, check for overflow and undo increment if so. */ |
| 222 | if (++(*histp) <= 0) |
| 223 | (*histp)--; |
| 224 | ptr += 3; |
| 225 | } |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | |
| 230 | /* |
| 231 | * Next we have the really interesting routines: selection of a colormap |
| 232 | * given the completed histogram. |
| 233 | * These routines work with a list of "boxes", each representing a rectangular |
| 234 | * subset of the input color space (to histogram precision). |
| 235 | */ |
| 236 | |
| 237 | typedef struct { |
| 238 | /* The bounds of the box (inclusive); expressed as histogram indexes */ |
| 239 | int c0min, c0max; |
| 240 | int c1min, c1max; |
| 241 | int c2min, c2max; |
| 242 | /* The volume (actually 2-norm) of the box */ |
| 243 | JLONG volume; |
| 244 | /* The number of nonzero histogram cells within this box */ |
| 245 | long colorcount; |
| 246 | } box; |
| 247 | |
| 248 | typedef box *boxptr; |
| 249 | |
| 250 | |
| 251 | LOCAL(boxptr) |
| 252 | find_biggest_color_pop(boxptr boxlist, int numboxes) |
| 253 | /* Find the splittable box with the largest color population */ |
| 254 | /* Returns NULL if no splittable boxes remain */ |
| 255 | { |
| 256 | register boxptr boxp; |
| 257 | register int i; |
| 258 | register long maxc = 0; |
| 259 | boxptr which = NULL; |
| 260 | |
| 261 | for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) { |
| 262 | if (boxp->colorcount > maxc && boxp->volume > 0) { |
| 263 | which = boxp; |
| 264 | maxc = boxp->colorcount; |
| 265 | } |
| 266 | } |
| 267 | return which; |
| 268 | } |
| 269 | |
| 270 | |
| 271 | LOCAL(boxptr) |
| 272 | find_biggest_volume(boxptr boxlist, int numboxes) |
| 273 | /* Find the splittable box with the largest (scaled) volume */ |
| 274 | /* Returns NULL if no splittable boxes remain */ |
| 275 | { |
| 276 | register boxptr boxp; |
| 277 | register int i; |
| 278 | register JLONG maxv = 0; |
| 279 | boxptr which = NULL; |
| 280 | |
| 281 | for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) { |
| 282 | if (boxp->volume > maxv) { |
| 283 | which = boxp; |
| 284 | maxv = boxp->volume; |
| 285 | } |
| 286 | } |
| 287 | return which; |
| 288 | } |
| 289 | |
| 290 | |
| 291 | LOCAL(void) |
| 292 | update_box(j_decompress_ptr cinfo, boxptr boxp) |
| 293 | /* Shrink the min/max bounds of a box to enclose only nonzero elements, */ |
| 294 | /* and recompute its volume and population */ |
| 295 | { |
| 296 | my_cquantize_ptr cquantize = (my_cquantize_ptr)cinfo->cquantize; |
| 297 | hist3d histogram = cquantize->histogram; |
| 298 | histptr histp; |
| 299 | int c0, c1, c2; |
| 300 | int c0min, c0max, c1min, c1max, c2min, c2max; |
| 301 | JLONG dist0, dist1, dist2; |
| 302 | long ccount; |
| 303 | |
| 304 | c0min = boxp->c0min; c0max = boxp->c0max; |
| 305 | c1min = boxp->c1min; c1max = boxp->c1max; |
| 306 | c2min = boxp->c2min; c2max = boxp->c2max; |
| 307 | |
| 308 | if (c0max > c0min) |
| 309 | for (c0 = c0min; c0 <= c0max; c0++) |
| 310 | for (c1 = c1min; c1 <= c1max; c1++) { |
| 311 | histp = &histogram[c0][c1][c2min]; |
| 312 | for (c2 = c2min; c2 <= c2max; c2++) |
| 313 | if (*histp++ != 0) { |
| 314 | boxp->c0min = c0min = c0; |
| 315 | goto have_c0min; |
| 316 | } |
| 317 | } |
| 318 | have_c0min: |
| 319 | if (c0max > c0min) |
| 320 | for (c0 = c0max; c0 >= c0min; c0--) |
| 321 | for (c1 = c1min; c1 <= c1max; c1++) { |
| 322 | histp = &histogram[c0][c1][c2min]; |
| 323 | for (c2 = c2min; c2 <= c2max; c2++) |
| 324 | if (*histp++ != 0) { |
| 325 | boxp->c0max = c0max = c0; |
| 326 | goto have_c0max; |
| 327 | } |
| 328 | } |
| 329 | have_c0max: |
| 330 | if (c1max > c1min) |
| 331 | for (c1 = c1min; c1 <= c1max; c1++) |
| 332 | for (c0 = c0min; c0 <= c0max; c0++) { |
| 333 | histp = &histogram[c0][c1][c2min]; |
| 334 | for (c2 = c2min; c2 <= c2max; c2++) |
| 335 | if (*histp++ != 0) { |
| 336 | boxp->c1min = c1min = c1; |
| 337 | goto have_c1min; |
| 338 | } |
| 339 | } |
| 340 | have_c1min: |
| 341 | if (c1max > c1min) |
| 342 | for (c1 = c1max; c1 >= c1min; c1--) |
| 343 | for (c0 = c0min; c0 <= c0max; c0++) { |
| 344 | histp = &histogram[c0][c1][c2min]; |
| 345 | for (c2 = c2min; c2 <= c2max; c2++) |
| 346 | if (*histp++ != 0) { |
| 347 | boxp->c1max = c1max = c1; |
| 348 | goto have_c1max; |
| 349 | } |
| 350 | } |
| 351 | have_c1max: |
| 352 | if (c2max > c2min) |
| 353 | for (c2 = c2min; c2 <= c2max; c2++) |
| 354 | for (c0 = c0min; c0 <= c0max; c0++) { |
| 355 | histp = &histogram[c0][c1min][c2]; |
| 356 | for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS) |
| 357 | if (*histp != 0) { |
| 358 | boxp->c2min = c2min = c2; |
| 359 | goto have_c2min; |
| 360 | } |
| 361 | } |
| 362 | have_c2min: |
| 363 | if (c2max > c2min) |
| 364 | for (c2 = c2max; c2 >= c2min; c2--) |
| 365 | for (c0 = c0min; c0 <= c0max; c0++) { |
| 366 | histp = &histogram[c0][c1min][c2]; |
| 367 | for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS) |
| 368 | if (*histp != 0) { |
| 369 | boxp->c2max = c2max = c2; |
| 370 | goto have_c2max; |
| 371 | } |
| 372 | } |
| 373 | have_c2max: |
| 374 | |
| 375 | /* Update box volume. |
| 376 | * We use 2-norm rather than real volume here; this biases the method |
| 377 | * against making long narrow boxes, and it has the side benefit that |
| 378 | * a box is splittable iff norm > 0. |
| 379 | * Since the differences are expressed in histogram-cell units, |
| 380 | * we have to shift back to JSAMPLE units to get consistent distances; |
| 381 | * after which, we scale according to the selected distance scale factors. |
| 382 | */ |
| 383 | dist0 = ((c0max - c0min) << C0_SHIFT) * C0_SCALE; |
| 384 | dist1 = ((c1max - c1min) << C1_SHIFT) * C1_SCALE; |
| 385 | dist2 = ((c2max - c2min) << C2_SHIFT) * C2_SCALE; |
| 386 | boxp->volume = dist0 * dist0 + dist1 * dist1 + dist2 * dist2; |
| 387 | |
| 388 | /* Now scan remaining volume of box and compute population */ |
| 389 | ccount = 0; |
| 390 | for (c0 = c0min; c0 <= c0max; c0++) |
| 391 | for (c1 = c1min; c1 <= c1max; c1++) { |
| 392 | histp = &histogram[c0][c1][c2min]; |
| 393 | for (c2 = c2min; c2 <= c2max; c2++, histp++) |
| 394 | if (*histp != 0) { |
| 395 | ccount++; |
| 396 | } |
| 397 | } |
| 398 | boxp->colorcount = ccount; |
| 399 | } |
| 400 | |
| 401 | |
| 402 | LOCAL(int) |
| 403 | median_cut(j_decompress_ptr cinfo, boxptr boxlist, int numboxes, |
| 404 | int desired_colors) |
| 405 | /* Repeatedly select and split the largest box until we have enough boxes */ |
| 406 | { |
| 407 | int n, lb; |
| 408 | int c0, c1, c2, cmax; |
| 409 | register boxptr b1, b2; |
| 410 | |
| 411 | while (numboxes < desired_colors) { |
| 412 | /* Select box to split. |
| 413 | * Current algorithm: by population for first half, then by volume. |
| 414 | */ |
| 415 | if (numboxes * 2 <= desired_colors) { |
| 416 | b1 = find_biggest_color_pop(boxlist, numboxes); |
| 417 | } else { |
| 418 | b1 = find_biggest_volume(boxlist, numboxes); |
| 419 | } |
| 420 | if (b1 == NULL) /* no splittable boxes left! */ |
| 421 | break; |
| 422 | b2 = &boxlist[numboxes]; /* where new box will go */ |
| 423 | /* Copy the color bounds to the new box. */ |
| 424 | b2->c0max = b1->c0max; b2->c1max = b1->c1max; b2->c2max = b1->c2max; |
| 425 | b2->c0min = b1->c0min; b2->c1min = b1->c1min; b2->c2min = b1->c2min; |
| 426 | /* Choose which axis to split the box on. |
| 427 | * Current algorithm: longest scaled axis. |
| 428 | * See notes in update_box about scaling distances. |
| 429 | */ |
| 430 | c0 = ((b1->c0max - b1->c0min) << C0_SHIFT) * C0_SCALE; |
| 431 | c1 = ((b1->c1max - b1->c1min) << C1_SHIFT) * C1_SCALE; |
| 432 | c2 = ((b1->c2max - b1->c2min) << C2_SHIFT) * C2_SCALE; |
| 433 | /* We want to break any ties in favor of green, then red, blue last. |
| 434 | * This code does the right thing for R,G,B or B,G,R color orders only. |
| 435 | */ |
| 436 | if (rgb_red[cinfo->out_color_space] == 0) { |
| 437 | cmax = c1; n = 1; |
| 438 | if (c0 > cmax) { cmax = c0; n = 0; } |
| 439 | if (c2 > cmax) { n = 2; } |
| 440 | } else { |
| 441 | cmax = c1; n = 1; |
| 442 | if (c2 > cmax) { cmax = c2; n = 2; } |
| 443 | if (c0 > cmax) { n = 0; } |
| 444 | } |
| 445 | /* Choose split point along selected axis, and update box bounds. |
| 446 | * Current algorithm: split at halfway point. |
| 447 | * (Since the box has been shrunk to minimum volume, |
| 448 | * any split will produce two nonempty subboxes.) |
| 449 | * Note that lb value is max for lower box, so must be < old max. |
| 450 | */ |
| 451 | switch (n) { |
| 452 | case 0: |
| 453 | lb = (b1->c0max + b1->c0min) / 2; |
| 454 | b1->c0max = lb; |
| 455 | b2->c0min = lb + 1; |
| 456 | break; |
| 457 | case 1: |
| 458 | lb = (b1->c1max + b1->c1min) / 2; |
| 459 | b1->c1max = lb; |
| 460 | b2->c1min = lb + 1; |
| 461 | break; |
| 462 | case 2: |
| 463 | lb = (b1->c2max + b1->c2min) / 2; |
| 464 | b1->c2max = lb; |
| 465 | b2->c2min = lb + 1; |
| 466 | break; |
| 467 | } |
| 468 | /* Update stats for boxes */ |
| 469 | update_box(cinfo, b1); |
| 470 | update_box(cinfo, b2); |
| 471 | numboxes++; |
| 472 | } |
| 473 | return numboxes; |
| 474 | } |
| 475 | |
| 476 | |
| 477 | LOCAL(void) |
| 478 | compute_color(j_decompress_ptr cinfo, boxptr boxp, int icolor) |
| 479 | /* Compute representative color for a box, put it in colormap[icolor] */ |
| 480 | { |
| 481 | /* Current algorithm: mean weighted by pixels (not colors) */ |
| 482 | /* Note it is important to get the rounding correct! */ |
| 483 | my_cquantize_ptr cquantize = (my_cquantize_ptr)cinfo->cquantize; |
| 484 | hist3d histogram = cquantize->histogram; |
| 485 | histptr histp; |
| 486 | int c0, c1, c2; |
| 487 | int c0min, c0max, c1min, c1max, c2min, c2max; |
| 488 | long count; |
| 489 | long total = 0; |
| 490 | long c0total = 0; |
| 491 | long c1total = 0; |
| 492 | long c2total = 0; |
| 493 | |
| 494 | c0min = boxp->c0min; c0max = boxp->c0max; |
| 495 | c1min = boxp->c1min; c1max = boxp->c1max; |
| 496 | c2min = boxp->c2min; c2max = boxp->c2max; |
| 497 | |
| 498 | for (c0 = c0min; c0 <= c0max; c0++) |
| 499 | for (c1 = c1min; c1 <= c1max; c1++) { |
| 500 | histp = &histogram[c0][c1][c2min]; |
| 501 | for (c2 = c2min; c2 <= c2max; c2++) { |
| 502 | if ((count = *histp++) != 0) { |
| 503 | total += count; |
| 504 | c0total += ((c0 << C0_SHIFT) + ((1 << C0_SHIFT) >> 1)) * count; |
| 505 | c1total += ((c1 << C1_SHIFT) + ((1 << C1_SHIFT) >> 1)) * count; |
| 506 | c2total += ((c2 << C2_SHIFT) + ((1 << C2_SHIFT) >> 1)) * count; |
| 507 | } |
| 508 | } |
| 509 | } |
| 510 | |
| 511 | cinfo->colormap[0][icolor] = (JSAMPLE)((c0total + (total >> 1)) / total); |
| 512 | cinfo->colormap[1][icolor] = (JSAMPLE)((c1total + (total >> 1)) / total); |
| 513 | cinfo->colormap[2][icolor] = (JSAMPLE)((c2total + (total >> 1)) / total); |
| 514 | } |
| 515 | |
| 516 | |
| 517 | LOCAL(void) |
| 518 | select_colors(j_decompress_ptr cinfo, int desired_colors) |
| 519 | /* Master routine for color selection */ |
| 520 | { |
| 521 | boxptr boxlist; |
| 522 | int numboxes; |
| 523 | int i; |
| 524 | |
| 525 | /* Allocate workspace for box list */ |
| 526 | boxlist = (boxptr)(*cinfo->mem->alloc_small) |
| 527 | ((j_common_ptr)cinfo, JPOOL_IMAGE, desired_colors * sizeof(box)); |
| 528 | /* Initialize one box containing whole space */ |
| 529 | numboxes = 1; |
| 530 | boxlist[0].c0min = 0; |
| 531 | boxlist[0].c0max = MAXJSAMPLE >> C0_SHIFT; |
| 532 | boxlist[0].c1min = 0; |
| 533 | boxlist[0].c1max = MAXJSAMPLE >> C1_SHIFT; |
| 534 | boxlist[0].c2min = 0; |
| 535 | boxlist[0].c2max = MAXJSAMPLE >> C2_SHIFT; |
| 536 | /* Shrink it to actually-used volume and set its statistics */ |
| 537 | update_box(cinfo, &boxlist[0]); |
| 538 | /* Perform median-cut to produce final box list */ |
| 539 | numboxes = median_cut(cinfo, boxlist, numboxes, desired_colors); |
| 540 | /* Compute the representative color for each box, fill colormap */ |
| 541 | for (i = 0; i < numboxes; i++) |
| 542 | compute_color(cinfo, &boxlist[i], i); |
| 543 | cinfo->actual_number_of_colors = numboxes; |
| 544 | TRACEMS1(cinfo, 1, JTRC_QUANT_SELECTED, numboxes); |
| 545 | } |
| 546 | |
| 547 | |
| 548 | /* |
| 549 | * These routines are concerned with the time-critical task of mapping input |
| 550 | * colors to the nearest color in the selected colormap. |
| 551 | * |
| 552 | * We re-use the histogram space as an "inverse color map", essentially a |
| 553 | * cache for the results of nearest-color searches. All colors within a |
| 554 | * histogram cell will be mapped to the same colormap entry, namely the one |
| 555 | * closest to the cell's center. This may not be quite the closest entry to |
| 556 | * the actual input color, but it's almost as good. A zero in the cache |
| 557 | * indicates we haven't found the nearest color for that cell yet; the array |
| 558 | * is cleared to zeroes before starting the mapping pass. When we find the |
| 559 | * nearest color for a cell, its colormap index plus one is recorded in the |
| 560 | * cache for future use. The pass2 scanning routines call fill_inverse_cmap |
| 561 | * when they need to use an unfilled entry in the cache. |
| 562 | * |
| 563 | * Our method of efficiently finding nearest colors is based on the "locally |
| 564 | * sorted search" idea described by Heckbert and on the incremental distance |
| 565 | * calculation described by Spencer W. Thomas in chapter III.1 of Graphics |
| 566 | * Gems II (James Arvo, ed. Academic Press, 1991). Thomas points out that |
| 567 | * the distances from a given colormap entry to each cell of the histogram can |
| 568 | * be computed quickly using an incremental method: the differences between |
| 569 | * distances to adjacent cells themselves differ by a constant. This allows a |
| 570 | * fairly fast implementation of the "brute force" approach of computing the |
| 571 | * distance from every colormap entry to every histogram cell. Unfortunately, |
| 572 | * it needs a work array to hold the best-distance-so-far for each histogram |
| 573 | * cell (because the inner loop has to be over cells, not colormap entries). |
| 574 | * The work array elements have to be JLONGs, so the work array would need |
| 575 | * 256Kb at our recommended precision. This is not feasible in DOS machines. |
| 576 | * |
| 577 | * To get around these problems, we apply Thomas' method to compute the |
| 578 | * nearest colors for only the cells within a small subbox of the histogram. |
| 579 | * The work array need be only as big as the subbox, so the memory usage |
| 580 | * problem is solved. Furthermore, we need not fill subboxes that are never |
| 581 | * referenced in pass2; many images use only part of the color gamut, so a |
| 582 | * fair amount of work is saved. An additional advantage of this |
| 583 | * approach is that we can apply Heckbert's locality criterion to quickly |
| 584 | * eliminate colormap entries that are far away from the subbox; typically |
| 585 | * three-fourths of the colormap entries are rejected by Heckbert's criterion, |
| 586 | * and we need not compute their distances to individual cells in the subbox. |
| 587 | * The speed of this approach is heavily influenced by the subbox size: too |
| 588 | * small means too much overhead, too big loses because Heckbert's criterion |
| 589 | * can't eliminate as many colormap entries. Empirically the best subbox |
| 590 | * size seems to be about 1/512th of the histogram (1/8th in each direction). |
| 591 | * |
| 592 | * Thomas' article also describes a refined method which is asymptotically |
| 593 | * faster than the brute-force method, but it is also far more complex and |
| 594 | * cannot efficiently be applied to small subboxes. It is therefore not |
| 595 | * useful for programs intended to be portable to DOS machines. On machines |
| 596 | * with plenty of memory, filling the whole histogram in one shot with Thomas' |
| 597 | * refined method might be faster than the present code --- but then again, |
| 598 | * it might not be any faster, and it's certainly more complicated. |
| 599 | */ |
| 600 | |
| 601 | |
| 602 | /* log2(histogram cells in update box) for each axis; this can be adjusted */ |
| 603 | #define BOX_C0_LOG (HIST_C0_BITS - 3) |
| 604 | #define BOX_C1_LOG (HIST_C1_BITS - 3) |
| 605 | #define BOX_C2_LOG (HIST_C2_BITS - 3) |
| 606 | |
| 607 | #define BOX_C0_ELEMS (1 << BOX_C0_LOG) /* # of hist cells in update box */ |
| 608 | #define BOX_C1_ELEMS (1 << BOX_C1_LOG) |
| 609 | #define BOX_C2_ELEMS (1 << BOX_C2_LOG) |
| 610 | |
| 611 | #define BOX_C0_SHIFT (C0_SHIFT + BOX_C0_LOG) |
| 612 | #define BOX_C1_SHIFT (C1_SHIFT + BOX_C1_LOG) |
| 613 | #define BOX_C2_SHIFT (C2_SHIFT + BOX_C2_LOG) |
| 614 | |
| 615 | |
| 616 | /* |
| 617 | * The next three routines implement inverse colormap filling. They could |
| 618 | * all be folded into one big routine, but splitting them up this way saves |
| 619 | * some stack space (the mindist[] and bestdist[] arrays need not coexist) |
| 620 | * and may allow some compilers to produce better code by registerizing more |
| 621 | * inner-loop variables. |
| 622 | */ |
| 623 | |
| 624 | LOCAL(int) |
| 625 | find_nearby_colors(j_decompress_ptr cinfo, int minc0, int minc1, int minc2, |
| 626 | JSAMPLE colorlist[]) |
| 627 | /* Locate the colormap entries close enough to an update box to be candidates |
| 628 | * for the nearest entry to some cell(s) in the update box. The update box |
| 629 | * is specified by the center coordinates of its first cell. The number of |
| 630 | * candidate colormap entries is returned, and their colormap indexes are |
| 631 | * placed in colorlist[]. |
| 632 | * This routine uses Heckbert's "locally sorted search" criterion to select |
| 633 | * the colors that need further consideration. |
| 634 | */ |
| 635 | { |
| 636 | int numcolors = cinfo->actual_number_of_colors; |
| 637 | int maxc0, maxc1, maxc2; |
| 638 | int centerc0, centerc1, centerc2; |
| 639 | int i, x, ncolors; |
| 640 | JLONG minmaxdist, min_dist, max_dist, tdist; |
| 641 | JLONG mindist[MAXNUMCOLORS]; /* min distance to colormap entry i */ |
| 642 | |
| 643 | /* Compute true coordinates of update box's upper corner and center. |
| 644 | * Actually we compute the coordinates of the center of the upper-corner |
| 645 | * histogram cell, which are the upper bounds of the volume we care about. |
| 646 | * Note that since ">>" rounds down, the "center" values may be closer to |
| 647 | * min than to max; hence comparisons to them must be "<=", not "<". |
| 648 | */ |
| 649 | maxc0 = minc0 + ((1 << BOX_C0_SHIFT) - (1 << C0_SHIFT)); |
| 650 | centerc0 = (minc0 + maxc0) >> 1; |
| 651 | maxc1 = minc1 + ((1 << BOX_C1_SHIFT) - (1 << C1_SHIFT)); |
| 652 | centerc1 = (minc1 + maxc1) >> 1; |
| 653 | maxc2 = minc2 + ((1 << BOX_C2_SHIFT) - (1 << C2_SHIFT)); |
| 654 | centerc2 = (minc2 + maxc2) >> 1; |
| 655 | |
| 656 | /* For each color in colormap, find: |
| 657 | * 1. its minimum squared-distance to any point in the update box |
| 658 | * (zero if color is within update box); |
| 659 | * 2. its maximum squared-distance to any point in the update box. |
| 660 | * Both of these can be found by considering only the corners of the box. |
| 661 | * We save the minimum distance for each color in mindist[]; |
| 662 | * only the smallest maximum distance is of interest. |
| 663 | */ |
| 664 | minmaxdist = 0x7FFFFFFFL; |
| 665 | |
| 666 | for (i = 0; i < numcolors; i++) { |
| 667 | /* We compute the squared-c0-distance term, then add in the other two. */ |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 668 | x = cinfo->colormap[0][i]; |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 669 | if (x < minc0) { |
| 670 | tdist = (x - minc0) * C0_SCALE; |
| 671 | min_dist = tdist * tdist; |
| 672 | tdist = (x - maxc0) * C0_SCALE; |
| 673 | max_dist = tdist * tdist; |
| 674 | } else if (x > maxc0) { |
| 675 | tdist = (x - maxc0) * C0_SCALE; |
| 676 | min_dist = tdist * tdist; |
| 677 | tdist = (x - minc0) * C0_SCALE; |
| 678 | max_dist = tdist * tdist; |
| 679 | } else { |
| 680 | /* within cell range so no contribution to min_dist */ |
| 681 | min_dist = 0; |
| 682 | if (x <= centerc0) { |
| 683 | tdist = (x - maxc0) * C0_SCALE; |
| 684 | max_dist = tdist * tdist; |
| 685 | } else { |
| 686 | tdist = (x - minc0) * C0_SCALE; |
| 687 | max_dist = tdist * tdist; |
| 688 | } |
| 689 | } |
| 690 | |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 691 | x = cinfo->colormap[1][i]; |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 692 | if (x < minc1) { |
| 693 | tdist = (x - minc1) * C1_SCALE; |
| 694 | min_dist += tdist * tdist; |
| 695 | tdist = (x - maxc1) * C1_SCALE; |
| 696 | max_dist += tdist * tdist; |
| 697 | } else if (x > maxc1) { |
| 698 | tdist = (x - maxc1) * C1_SCALE; |
| 699 | min_dist += tdist * tdist; |
| 700 | tdist = (x - minc1) * C1_SCALE; |
| 701 | max_dist += tdist * tdist; |
| 702 | } else { |
| 703 | /* within cell range so no contribution to min_dist */ |
| 704 | if (x <= centerc1) { |
| 705 | tdist = (x - maxc1) * C1_SCALE; |
| 706 | max_dist += tdist * tdist; |
| 707 | } else { |
| 708 | tdist = (x - minc1) * C1_SCALE; |
| 709 | max_dist += tdist * tdist; |
| 710 | } |
| 711 | } |
| 712 | |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 713 | x = cinfo->colormap[2][i]; |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 714 | if (x < minc2) { |
| 715 | tdist = (x - minc2) * C2_SCALE; |
| 716 | min_dist += tdist * tdist; |
| 717 | tdist = (x - maxc2) * C2_SCALE; |
| 718 | max_dist += tdist * tdist; |
| 719 | } else if (x > maxc2) { |
| 720 | tdist = (x - maxc2) * C2_SCALE; |
| 721 | min_dist += tdist * tdist; |
| 722 | tdist = (x - minc2) * C2_SCALE; |
| 723 | max_dist += tdist * tdist; |
| 724 | } else { |
| 725 | /* within cell range so no contribution to min_dist */ |
| 726 | if (x <= centerc2) { |
| 727 | tdist = (x - maxc2) * C2_SCALE; |
| 728 | max_dist += tdist * tdist; |
| 729 | } else { |
| 730 | tdist = (x - minc2) * C2_SCALE; |
| 731 | max_dist += tdist * tdist; |
| 732 | } |
| 733 | } |
| 734 | |
| 735 | mindist[i] = min_dist; /* save away the results */ |
| 736 | if (max_dist < minmaxdist) |
| 737 | minmaxdist = max_dist; |
| 738 | } |
| 739 | |
| 740 | /* Now we know that no cell in the update box is more than minmaxdist |
| 741 | * away from some colormap entry. Therefore, only colors that are |
| 742 | * within minmaxdist of some part of the box need be considered. |
| 743 | */ |
| 744 | ncolors = 0; |
| 745 | for (i = 0; i < numcolors; i++) { |
| 746 | if (mindist[i] <= minmaxdist) |
| 747 | colorlist[ncolors++] = (JSAMPLE)i; |
| 748 | } |
| 749 | return ncolors; |
| 750 | } |
| 751 | |
| 752 | |
| 753 | LOCAL(void) |
| 754 | find_best_colors(j_decompress_ptr cinfo, int minc0, int minc1, int minc2, |
| 755 | int numcolors, JSAMPLE colorlist[], JSAMPLE bestcolor[]) |
| 756 | /* Find the closest colormap entry for each cell in the update box, |
| 757 | * given the list of candidate colors prepared by find_nearby_colors. |
| 758 | * Return the indexes of the closest entries in the bestcolor[] array. |
| 759 | * This routine uses Thomas' incremental distance calculation method to |
| 760 | * find the distance from a colormap entry to successive cells in the box. |
| 761 | */ |
| 762 | { |
| 763 | int ic0, ic1, ic2; |
| 764 | int i, icolor; |
| 765 | register JLONG *bptr; /* pointer into bestdist[] array */ |
| 766 | JSAMPLE *cptr; /* pointer into bestcolor[] array */ |
| 767 | JLONG dist0, dist1; /* initial distance values */ |
| 768 | register JLONG dist2; /* current distance in inner loop */ |
| 769 | JLONG xx0, xx1; /* distance increments */ |
| 770 | register JLONG xx2; |
| 771 | JLONG inc0, inc1, inc2; /* initial values for increments */ |
| 772 | /* This array holds the distance to the nearest-so-far color for each cell */ |
| 773 | JLONG bestdist[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS]; |
| 774 | |
| 775 | /* Initialize best-distance for each cell of the update box */ |
| 776 | bptr = bestdist; |
| 777 | for (i = BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS - 1; i >= 0; i--) |
| 778 | *bptr++ = 0x7FFFFFFFL; |
| 779 | |
| 780 | /* For each color selected by find_nearby_colors, |
| 781 | * compute its distance to the center of each cell in the box. |
| 782 | * If that's less than best-so-far, update best distance and color number. |
| 783 | */ |
| 784 | |
| 785 | /* Nominal steps between cell centers ("x" in Thomas article) */ |
| 786 | #define STEP_C0 ((1 << C0_SHIFT) * C0_SCALE) |
| 787 | #define STEP_C1 ((1 << C1_SHIFT) * C1_SCALE) |
| 788 | #define STEP_C2 ((1 << C2_SHIFT) * C2_SCALE) |
| 789 | |
| 790 | for (i = 0; i < numcolors; i++) { |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 791 | icolor = colorlist[i]; |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 792 | /* Compute (square of) distance from minc0/c1/c2 to this color */ |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 793 | inc0 = (minc0 - cinfo->colormap[0][icolor]) * C0_SCALE; |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 794 | dist0 = inc0 * inc0; |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 795 | inc1 = (minc1 - cinfo->colormap[1][icolor]) * C1_SCALE; |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 796 | dist0 += inc1 * inc1; |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 797 | inc2 = (minc2 - cinfo->colormap[2][icolor]) * C2_SCALE; |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 798 | dist0 += inc2 * inc2; |
| 799 | /* Form the initial difference increments */ |
| 800 | inc0 = inc0 * (2 * STEP_C0) + STEP_C0 * STEP_C0; |
| 801 | inc1 = inc1 * (2 * STEP_C1) + STEP_C1 * STEP_C1; |
| 802 | inc2 = inc2 * (2 * STEP_C2) + STEP_C2 * STEP_C2; |
| 803 | /* Now loop over all cells in box, updating distance per Thomas method */ |
| 804 | bptr = bestdist; |
| 805 | cptr = bestcolor; |
| 806 | xx0 = inc0; |
| 807 | for (ic0 = BOX_C0_ELEMS - 1; ic0 >= 0; ic0--) { |
| 808 | dist1 = dist0; |
| 809 | xx1 = inc1; |
| 810 | for (ic1 = BOX_C1_ELEMS - 1; ic1 >= 0; ic1--) { |
| 811 | dist2 = dist1; |
| 812 | xx2 = inc2; |
| 813 | for (ic2 = BOX_C2_ELEMS - 1; ic2 >= 0; ic2--) { |
| 814 | if (dist2 < *bptr) { |
| 815 | *bptr = dist2; |
| 816 | *cptr = (JSAMPLE)icolor; |
| 817 | } |
| 818 | dist2 += xx2; |
| 819 | xx2 += 2 * STEP_C2 * STEP_C2; |
| 820 | bptr++; |
| 821 | cptr++; |
| 822 | } |
| 823 | dist1 += xx1; |
| 824 | xx1 += 2 * STEP_C1 * STEP_C1; |
| 825 | } |
| 826 | dist0 += xx0; |
| 827 | xx0 += 2 * STEP_C0 * STEP_C0; |
| 828 | } |
| 829 | } |
| 830 | } |
| 831 | |
| 832 | |
| 833 | LOCAL(void) |
| 834 | fill_inverse_cmap(j_decompress_ptr cinfo, int c0, int c1, int c2) |
| 835 | /* Fill the inverse-colormap entries in the update box that contains */ |
| 836 | /* histogram cell c0/c1/c2. (Only that one cell MUST be filled, but */ |
| 837 | /* we can fill as many others as we wish.) */ |
| 838 | { |
| 839 | my_cquantize_ptr cquantize = (my_cquantize_ptr)cinfo->cquantize; |
| 840 | hist3d histogram = cquantize->histogram; |
| 841 | int minc0, minc1, minc2; /* lower left corner of update box */ |
| 842 | int ic0, ic1, ic2; |
| 843 | register JSAMPLE *cptr; /* pointer into bestcolor[] array */ |
| 844 | register histptr cachep; /* pointer into main cache array */ |
| 845 | /* This array lists the candidate colormap indexes. */ |
| 846 | JSAMPLE colorlist[MAXNUMCOLORS]; |
| 847 | int numcolors; /* number of candidate colors */ |
| 848 | /* This array holds the actually closest colormap index for each cell. */ |
| 849 | JSAMPLE bestcolor[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS]; |
| 850 | |
| 851 | /* Convert cell coordinates to update box ID */ |
| 852 | c0 >>= BOX_C0_LOG; |
| 853 | c1 >>= BOX_C1_LOG; |
| 854 | c2 >>= BOX_C2_LOG; |
| 855 | |
| 856 | /* Compute true coordinates of update box's origin corner. |
| 857 | * Actually we compute the coordinates of the center of the corner |
| 858 | * histogram cell, which are the lower bounds of the volume we care about. |
| 859 | */ |
| 860 | minc0 = (c0 << BOX_C0_SHIFT) + ((1 << C0_SHIFT) >> 1); |
| 861 | minc1 = (c1 << BOX_C1_SHIFT) + ((1 << C1_SHIFT) >> 1); |
| 862 | minc2 = (c2 << BOX_C2_SHIFT) + ((1 << C2_SHIFT) >> 1); |
| 863 | |
| 864 | /* Determine which colormap entries are close enough to be candidates |
| 865 | * for the nearest entry to some cell in the update box. |
| 866 | */ |
| 867 | numcolors = find_nearby_colors(cinfo, minc0, minc1, minc2, colorlist); |
| 868 | |
| 869 | /* Determine the actually nearest colors. */ |
| 870 | find_best_colors(cinfo, minc0, minc1, minc2, numcolors, colorlist, |
| 871 | bestcolor); |
| 872 | |
| 873 | /* Save the best color numbers (plus 1) in the main cache array */ |
| 874 | c0 <<= BOX_C0_LOG; /* convert ID back to base cell indexes */ |
| 875 | c1 <<= BOX_C1_LOG; |
| 876 | c2 <<= BOX_C2_LOG; |
| 877 | cptr = bestcolor; |
| 878 | for (ic0 = 0; ic0 < BOX_C0_ELEMS; ic0++) { |
| 879 | for (ic1 = 0; ic1 < BOX_C1_ELEMS; ic1++) { |
| 880 | cachep = &histogram[c0 + ic0][c1 + ic1][c2]; |
| 881 | for (ic2 = 0; ic2 < BOX_C2_ELEMS; ic2++) { |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 882 | *cachep++ = (histcell)((*cptr++) + 1); |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 883 | } |
| 884 | } |
| 885 | } |
| 886 | } |
| 887 | |
| 888 | |
| 889 | /* |
| 890 | * Map some rows of pixels to the output colormapped representation. |
| 891 | */ |
| 892 | |
| 893 | METHODDEF(void) |
| 894 | pass2_no_dither(j_decompress_ptr cinfo, JSAMPARRAY input_buf, |
| 895 | JSAMPARRAY output_buf, int num_rows) |
| 896 | /* This version performs no dithering */ |
| 897 | { |
| 898 | my_cquantize_ptr cquantize = (my_cquantize_ptr)cinfo->cquantize; |
| 899 | hist3d histogram = cquantize->histogram; |
| 900 | register JSAMPROW inptr, outptr; |
| 901 | register histptr cachep; |
| 902 | register int c0, c1, c2; |
| 903 | int row; |
| 904 | JDIMENSION col; |
| 905 | JDIMENSION width = cinfo->output_width; |
| 906 | |
| 907 | for (row = 0; row < num_rows; row++) { |
| 908 | inptr = input_buf[row]; |
| 909 | outptr = output_buf[row]; |
| 910 | for (col = width; col > 0; col--) { |
| 911 | /* get pixel value and index into the cache */ |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 912 | c0 = (*inptr++) >> C0_SHIFT; |
| 913 | c1 = (*inptr++) >> C1_SHIFT; |
| 914 | c2 = (*inptr++) >> C2_SHIFT; |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 915 | cachep = &histogram[c0][c1][c2]; |
| 916 | /* If we have not seen this color before, find nearest colormap entry */ |
| 917 | /* and update the cache */ |
| 918 | if (*cachep == 0) |
| 919 | fill_inverse_cmap(cinfo, c0, c1, c2); |
| 920 | /* Now emit the colormap index for this cell */ |
| 921 | *outptr++ = (JSAMPLE)(*cachep - 1); |
| 922 | } |
| 923 | } |
| 924 | } |
| 925 | |
| 926 | |
| 927 | METHODDEF(void) |
| 928 | pass2_fs_dither(j_decompress_ptr cinfo, JSAMPARRAY input_buf, |
| 929 | JSAMPARRAY output_buf, int num_rows) |
| 930 | /* This version performs Floyd-Steinberg dithering */ |
| 931 | { |
| 932 | my_cquantize_ptr cquantize = (my_cquantize_ptr)cinfo->cquantize; |
| 933 | hist3d histogram = cquantize->histogram; |
| 934 | register LOCFSERROR cur0, cur1, cur2; /* current error or pixel value */ |
| 935 | LOCFSERROR belowerr0, belowerr1, belowerr2; /* error for pixel below cur */ |
| 936 | LOCFSERROR bpreverr0, bpreverr1, bpreverr2; /* error for below/prev col */ |
| 937 | register FSERRPTR errorptr; /* => fserrors[] at column before current */ |
| 938 | JSAMPROW inptr; /* => current input pixel */ |
| 939 | JSAMPROW outptr; /* => current output pixel */ |
| 940 | histptr cachep; |
| 941 | int dir; /* +1 or -1 depending on direction */ |
| 942 | int dir3; /* 3*dir, for advancing inptr & errorptr */ |
| 943 | int row; |
| 944 | JDIMENSION col; |
| 945 | JDIMENSION width = cinfo->output_width; |
| 946 | JSAMPLE *range_limit = cinfo->sample_range_limit; |
| 947 | int *error_limit = cquantize->error_limiter; |
| 948 | JSAMPROW colormap0 = cinfo->colormap[0]; |
| 949 | JSAMPROW colormap1 = cinfo->colormap[1]; |
| 950 | JSAMPROW colormap2 = cinfo->colormap[2]; |
| 951 | SHIFT_TEMPS |
| 952 | |
| 953 | for (row = 0; row < num_rows; row++) { |
| 954 | inptr = input_buf[row]; |
| 955 | outptr = output_buf[row]; |
| 956 | if (cquantize->on_odd_row) { |
| 957 | /* work right to left in this row */ |
| 958 | inptr += (width - 1) * 3; /* so point to rightmost pixel */ |
| 959 | outptr += width - 1; |
| 960 | dir = -1; |
| 961 | dir3 = -3; |
| 962 | errorptr = cquantize->fserrors + (width + 1) * 3; /* => entry after last column */ |
| 963 | cquantize->on_odd_row = FALSE; /* flip for next time */ |
| 964 | } else { |
| 965 | /* work left to right in this row */ |
| 966 | dir = 1; |
| 967 | dir3 = 3; |
| 968 | errorptr = cquantize->fserrors; /* => entry before first real column */ |
| 969 | cquantize->on_odd_row = TRUE; /* flip for next time */ |
| 970 | } |
| 971 | /* Preset error values: no error propagated to first pixel from left */ |
| 972 | cur0 = cur1 = cur2 = 0; |
| 973 | /* and no error propagated to row below yet */ |
| 974 | belowerr0 = belowerr1 = belowerr2 = 0; |
| 975 | bpreverr0 = bpreverr1 = bpreverr2 = 0; |
| 976 | |
| 977 | for (col = width; col > 0; col--) { |
| 978 | /* curN holds the error propagated from the previous pixel on the |
| 979 | * current line. Add the error propagated from the previous line |
| 980 | * to form the complete error correction term for this pixel, and |
| 981 | * round the error term (which is expressed * 16) to an integer. |
| 982 | * RIGHT_SHIFT rounds towards minus infinity, so adding 8 is correct |
| 983 | * for either sign of the error value. |
| 984 | * Note: errorptr points to *previous* column's array entry. |
| 985 | */ |
| 986 | cur0 = RIGHT_SHIFT(cur0 + errorptr[dir3 + 0] + 8, 4); |
| 987 | cur1 = RIGHT_SHIFT(cur1 + errorptr[dir3 + 1] + 8, 4); |
| 988 | cur2 = RIGHT_SHIFT(cur2 + errorptr[dir3 + 2] + 8, 4); |
| 989 | /* Limit the error using transfer function set by init_error_limit. |
| 990 | * See comments with init_error_limit for rationale. |
| 991 | */ |
| 992 | cur0 = error_limit[cur0]; |
| 993 | cur1 = error_limit[cur1]; |
| 994 | cur2 = error_limit[cur2]; |
| 995 | /* Form pixel value + error, and range-limit to 0..MAXJSAMPLE. |
| 996 | * The maximum error is +- MAXJSAMPLE (or less with error limiting); |
| 997 | * this sets the required size of the range_limit array. |
| 998 | */ |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 999 | cur0 += inptr[0]; |
| 1000 | cur1 += inptr[1]; |
| 1001 | cur2 += inptr[2]; |
| 1002 | cur0 = range_limit[cur0]; |
| 1003 | cur1 = range_limit[cur1]; |
| 1004 | cur2 = range_limit[cur2]; |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 1005 | /* Index into the cache with adjusted pixel value */ |
| 1006 | cachep = |
| 1007 | &histogram[cur0 >> C0_SHIFT][cur1 >> C1_SHIFT][cur2 >> C2_SHIFT]; |
| 1008 | /* If we have not seen this color before, find nearest colormap */ |
| 1009 | /* entry and update the cache */ |
| 1010 | if (*cachep == 0) |
| 1011 | fill_inverse_cmap(cinfo, cur0 >> C0_SHIFT, cur1 >> C1_SHIFT, |
| 1012 | cur2 >> C2_SHIFT); |
| 1013 | /* Now emit the colormap index for this cell */ |
| 1014 | { |
| 1015 | register int pixcode = *cachep - 1; |
| 1016 | *outptr = (JSAMPLE)pixcode; |
| 1017 | /* Compute representation error for this pixel */ |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 1018 | cur0 -= colormap0[pixcode]; |
| 1019 | cur1 -= colormap1[pixcode]; |
| 1020 | cur2 -= colormap2[pixcode]; |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 1021 | } |
| 1022 | /* Compute error fractions to be propagated to adjacent pixels. |
| 1023 | * Add these into the running sums, and simultaneously shift the |
| 1024 | * next-line error sums left by 1 column. |
| 1025 | */ |
| 1026 | { |
| 1027 | register LOCFSERROR bnexterr; |
| 1028 | |
| 1029 | bnexterr = cur0; /* Process component 0 */ |
| 1030 | errorptr[0] = (FSERROR)(bpreverr0 + cur0 * 3); |
| 1031 | bpreverr0 = belowerr0 + cur0 * 5; |
| 1032 | belowerr0 = bnexterr; |
| 1033 | cur0 *= 7; |
| 1034 | bnexterr = cur1; /* Process component 1 */ |
| 1035 | errorptr[1] = (FSERROR)(bpreverr1 + cur1 * 3); |
| 1036 | bpreverr1 = belowerr1 + cur1 * 5; |
| 1037 | belowerr1 = bnexterr; |
| 1038 | cur1 *= 7; |
| 1039 | bnexterr = cur2; /* Process component 2 */ |
| 1040 | errorptr[2] = (FSERROR)(bpreverr2 + cur2 * 3); |
| 1041 | bpreverr2 = belowerr2 + cur2 * 5; |
| 1042 | belowerr2 = bnexterr; |
| 1043 | cur2 *= 7; |
| 1044 | } |
| 1045 | /* At this point curN contains the 7/16 error value to be propagated |
| 1046 | * to the next pixel on the current line, and all the errors for the |
| 1047 | * next line have been shifted over. We are therefore ready to move on. |
| 1048 | */ |
| 1049 | inptr += dir3; /* Advance pixel pointers to next column */ |
| 1050 | outptr += dir; |
| 1051 | errorptr += dir3; /* advance errorptr to current column */ |
| 1052 | } |
| 1053 | /* Post-loop cleanup: we must unload the final error values into the |
| 1054 | * final fserrors[] entry. Note we need not unload belowerrN because |
| 1055 | * it is for the dummy column before or after the actual array. |
| 1056 | */ |
| 1057 | errorptr[0] = (FSERROR)bpreverr0; /* unload prev errs into array */ |
| 1058 | errorptr[1] = (FSERROR)bpreverr1; |
| 1059 | errorptr[2] = (FSERROR)bpreverr2; |
| 1060 | } |
| 1061 | } |
| 1062 | |
| 1063 | |
| 1064 | /* |
| 1065 | * Initialize the error-limiting transfer function (lookup table). |
| 1066 | * The raw F-S error computation can potentially compute error values of up to |
| 1067 | * +- MAXJSAMPLE. But we want the maximum correction applied to a pixel to be |
| 1068 | * much less, otherwise obviously wrong pixels will be created. (Typical |
| 1069 | * effects include weird fringes at color-area boundaries, isolated bright |
| 1070 | * pixels in a dark area, etc.) The standard advice for avoiding this problem |
| 1071 | * is to ensure that the "corners" of the color cube are allocated as output |
| 1072 | * colors; then repeated errors in the same direction cannot cause cascading |
| 1073 | * error buildup. However, that only prevents the error from getting |
| 1074 | * completely out of hand; Aaron Giles reports that error limiting improves |
| 1075 | * the results even with corner colors allocated. |
| 1076 | * A simple clamping of the error values to about +- MAXJSAMPLE/8 works pretty |
| 1077 | * well, but the smoother transfer function used below is even better. Thanks |
| 1078 | * to Aaron Giles for this idea. |
| 1079 | */ |
| 1080 | |
| 1081 | LOCAL(void) |
| 1082 | init_error_limit(j_decompress_ptr cinfo) |
| 1083 | /* Allocate and fill in the error_limiter table */ |
| 1084 | { |
| 1085 | my_cquantize_ptr cquantize = (my_cquantize_ptr)cinfo->cquantize; |
| 1086 | int *table; |
| 1087 | int in, out; |
| 1088 | |
| 1089 | table = (int *)(*cinfo->mem->alloc_small) |
| 1090 | ((j_common_ptr)cinfo, JPOOL_IMAGE, (MAXJSAMPLE * 2 + 1) * sizeof(int)); |
| 1091 | table += MAXJSAMPLE; /* so can index -MAXJSAMPLE .. +MAXJSAMPLE */ |
| 1092 | cquantize->error_limiter = table; |
| 1093 | |
| 1094 | #define STEPSIZE ((MAXJSAMPLE + 1) / 16) |
| 1095 | /* Map errors 1:1 up to +- MAXJSAMPLE/16 */ |
| 1096 | out = 0; |
| 1097 | for (in = 0; in < STEPSIZE; in++, out++) { |
| 1098 | table[in] = out; table[-in] = -out; |
| 1099 | } |
| 1100 | /* Map errors 1:2 up to +- 3*MAXJSAMPLE/16 */ |
| 1101 | for (; in < STEPSIZE * 3; in++, out += (in & 1) ? 0 : 1) { |
| 1102 | table[in] = out; table[-in] = -out; |
| 1103 | } |
| 1104 | /* Clamp the rest to final out value (which is (MAXJSAMPLE+1)/8) */ |
| 1105 | for (; in <= MAXJSAMPLE; in++) { |
| 1106 | table[in] = out; table[-in] = -out; |
| 1107 | } |
| 1108 | #undef STEPSIZE |
| 1109 | } |
| 1110 | |
| 1111 | |
| 1112 | /* |
| 1113 | * Finish up at the end of each pass. |
| 1114 | */ |
| 1115 | |
| 1116 | METHODDEF(void) |
| 1117 | finish_pass1(j_decompress_ptr cinfo) |
| 1118 | { |
| 1119 | my_cquantize_ptr cquantize = (my_cquantize_ptr)cinfo->cquantize; |
| 1120 | |
| 1121 | /* Select the representative colors and fill in cinfo->colormap */ |
| 1122 | cinfo->colormap = cquantize->sv_colormap; |
| 1123 | select_colors(cinfo, cquantize->desired); |
| 1124 | /* Force next pass to zero the color index table */ |
| 1125 | cquantize->needs_zeroed = TRUE; |
| 1126 | } |
| 1127 | |
| 1128 | |
| 1129 | METHODDEF(void) |
| 1130 | finish_pass2(j_decompress_ptr cinfo) |
| 1131 | { |
| 1132 | /* no work */ |
| 1133 | } |
| 1134 | |
| 1135 | |
| 1136 | /* |
| 1137 | * Initialize for each processing pass. |
| 1138 | */ |
| 1139 | |
| 1140 | METHODDEF(void) |
| 1141 | start_pass_2_quant(j_decompress_ptr cinfo, boolean is_pre_scan) |
| 1142 | { |
| 1143 | my_cquantize_ptr cquantize = (my_cquantize_ptr)cinfo->cquantize; |
| 1144 | hist3d histogram = cquantize->histogram; |
| 1145 | int i; |
| 1146 | |
| 1147 | /* Only F-S dithering or no dithering is supported. */ |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 1148 | /* If user asks for ordered dither, give them F-S. */ |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 1149 | if (cinfo->dither_mode != JDITHER_NONE) |
| 1150 | cinfo->dither_mode = JDITHER_FS; |
| 1151 | |
| 1152 | if (is_pre_scan) { |
| 1153 | /* Set up method pointers */ |
| 1154 | cquantize->pub.color_quantize = prescan_quantize; |
| 1155 | cquantize->pub.finish_pass = finish_pass1; |
| 1156 | cquantize->needs_zeroed = TRUE; /* Always zero histogram */ |
| 1157 | } else { |
| 1158 | /* Set up method pointers */ |
| 1159 | if (cinfo->dither_mode == JDITHER_FS) |
| 1160 | cquantize->pub.color_quantize = pass2_fs_dither; |
| 1161 | else |
| 1162 | cquantize->pub.color_quantize = pass2_no_dither; |
| 1163 | cquantize->pub.finish_pass = finish_pass2; |
| 1164 | |
| 1165 | /* Make sure color count is acceptable */ |
| 1166 | i = cinfo->actual_number_of_colors; |
| 1167 | if (i < 1) |
| 1168 | ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 1); |
| 1169 | if (i > MAXNUMCOLORS) |
| 1170 | ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS); |
| 1171 | |
| 1172 | if (cinfo->dither_mode == JDITHER_FS) { |
| 1173 | size_t arraysize = |
| 1174 | (size_t)((cinfo->output_width + 2) * (3 * sizeof(FSERROR))); |
| 1175 | /* Allocate Floyd-Steinberg workspace if we didn't already. */ |
| 1176 | if (cquantize->fserrors == NULL) |
| 1177 | cquantize->fserrors = (FSERRPTR)(*cinfo->mem->alloc_large) |
| 1178 | ((j_common_ptr)cinfo, JPOOL_IMAGE, arraysize); |
| 1179 | /* Initialize the propagated errors to zero. */ |
| 1180 | jzero_far((void *)cquantize->fserrors, arraysize); |
| 1181 | /* Make the error-limit table if we didn't already. */ |
| 1182 | if (cquantize->error_limiter == NULL) |
| 1183 | init_error_limit(cinfo); |
| 1184 | cquantize->on_odd_row = FALSE; |
| 1185 | } |
| 1186 | |
| 1187 | } |
| 1188 | /* Zero the histogram or inverse color map, if necessary */ |
| 1189 | if (cquantize->needs_zeroed) { |
| 1190 | for (i = 0; i < HIST_C0_ELEMS; i++) { |
| 1191 | jzero_far((void *)histogram[i], |
| 1192 | HIST_C1_ELEMS * HIST_C2_ELEMS * sizeof(histcell)); |
| 1193 | } |
| 1194 | cquantize->needs_zeroed = FALSE; |
| 1195 | } |
| 1196 | } |
| 1197 | |
| 1198 | |
| 1199 | /* |
| 1200 | * Switch to a new external colormap between output passes. |
| 1201 | */ |
| 1202 | |
| 1203 | METHODDEF(void) |
| 1204 | new_color_map_2_quant(j_decompress_ptr cinfo) |
| 1205 | { |
| 1206 | my_cquantize_ptr cquantize = (my_cquantize_ptr)cinfo->cquantize; |
| 1207 | |
| 1208 | /* Reset the inverse color map */ |
| 1209 | cquantize->needs_zeroed = TRUE; |
| 1210 | } |
| 1211 | |
| 1212 | |
| 1213 | /* |
| 1214 | * Module initialization routine for 2-pass color quantization. |
| 1215 | */ |
| 1216 | |
| 1217 | GLOBAL(void) |
| 1218 | jinit_2pass_quantizer(j_decompress_ptr cinfo) |
| 1219 | { |
| 1220 | my_cquantize_ptr cquantize; |
| 1221 | int i; |
| 1222 | |
| 1223 | cquantize = (my_cquantize_ptr) |
| 1224 | (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE, |
| 1225 | sizeof(my_cquantizer)); |
| 1226 | cinfo->cquantize = (struct jpeg_color_quantizer *)cquantize; |
| 1227 | cquantize->pub.start_pass = start_pass_2_quant; |
| 1228 | cquantize->pub.new_color_map = new_color_map_2_quant; |
| 1229 | cquantize->fserrors = NULL; /* flag optional arrays not allocated */ |
| 1230 | cquantize->error_limiter = NULL; |
| 1231 | |
| 1232 | /* Make sure jdmaster didn't give me a case I can't handle */ |
| 1233 | if (cinfo->out_color_components != 3) |
| 1234 | ERREXIT(cinfo, JERR_NOTIMPL); |
| 1235 | |
| 1236 | /* Allocate the histogram/inverse colormap storage */ |
| 1237 | cquantize->histogram = (hist3d)(*cinfo->mem->alloc_small) |
| 1238 | ((j_common_ptr)cinfo, JPOOL_IMAGE, HIST_C0_ELEMS * sizeof(hist2d)); |
| 1239 | for (i = 0; i < HIST_C0_ELEMS; i++) { |
| 1240 | cquantize->histogram[i] = (hist2d)(*cinfo->mem->alloc_large) |
| 1241 | ((j_common_ptr)cinfo, JPOOL_IMAGE, |
| 1242 | HIST_C1_ELEMS * HIST_C2_ELEMS * sizeof(histcell)); |
| 1243 | } |
| 1244 | cquantize->needs_zeroed = TRUE; /* histogram is garbage now */ |
| 1245 | |
| 1246 | /* Allocate storage for the completed colormap, if required. |
| 1247 | * We do this now since it may affect the memory manager's space |
| 1248 | * calculations. |
| 1249 | */ |
| 1250 | if (cinfo->enable_2pass_quant) { |
| 1251 | /* Make sure color count is acceptable */ |
| 1252 | int desired = cinfo->desired_number_of_colors; |
| 1253 | /* Lower bound on # of colors ... somewhat arbitrary as long as > 0 */ |
| 1254 | if (desired < 8) |
| 1255 | ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 8); |
| 1256 | /* Make sure colormap indexes can be represented by JSAMPLEs */ |
| 1257 | if (desired > MAXNUMCOLORS) |
| 1258 | ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS); |
| 1259 | cquantize->sv_colormap = (*cinfo->mem->alloc_sarray) |
| 1260 | ((j_common_ptr)cinfo, JPOOL_IMAGE, (JDIMENSION)desired, (JDIMENSION)3); |
| 1261 | cquantize->desired = desired; |
| 1262 | } else |
| 1263 | cquantize->sv_colormap = NULL; |
| 1264 | |
| 1265 | /* Only F-S dithering or no dithering is supported. */ |
Kaido Kert | 0e7d696 | 2021-08-06 13:16:46 -0700 | [diff] [blame] | 1266 | /* If user asks for ordered dither, give them F-S. */ |
Andrew Top | 8414344 | 2018-07-12 10:44:23 -0700 | [diff] [blame] | 1267 | if (cinfo->dither_mode != JDITHER_NONE) |
| 1268 | cinfo->dither_mode = JDITHER_FS; |
| 1269 | |
| 1270 | /* Allocate Floyd-Steinberg workspace if necessary. |
| 1271 | * This isn't really needed until pass 2, but again it may affect the memory |
| 1272 | * manager's space calculations. Although we will cope with a later change |
| 1273 | * in dither_mode, we do not promise to honor max_memory_to_use if |
| 1274 | * dither_mode changes. |
| 1275 | */ |
| 1276 | if (cinfo->dither_mode == JDITHER_FS) { |
| 1277 | cquantize->fserrors = (FSERRPTR)(*cinfo->mem->alloc_large) |
| 1278 | ((j_common_ptr)cinfo, JPOOL_IMAGE, |
| 1279 | (size_t)((cinfo->output_width + 2) * (3 * sizeof(FSERROR)))); |
| 1280 | /* Might as well create the error-limiting table too. */ |
| 1281 | init_error_limit(cinfo); |
| 1282 | } |
| 1283 | } |
| 1284 | |
| 1285 | #endif /* QUANT_2PASS_SUPPORTED */ |