1 /* 2 * jcsample.c 3 * 4 * Copyright (C) 1991-1996, Thomas G. Lane. 5 * Copyright 2009 Pierre Ossman <ossman (at) cendio.se> for Cendio AB 6 * This file is part of the Independent JPEG Group's software. 7 * For conditions of distribution and use, see the accompanying README file. 8 * 9 * This file contains downsampling routines. 10 * 11 * Downsampling input data is counted in "row groups". A row group 12 * is defined to be max_v_samp_factor pixel rows of each component, 13 * from which the downsampler produces v_samp_factor sample rows. 14 * A single row group is processed in each call to the downsampler module. 15 * 16 * The downsampler is responsible for edge-expansion of its output data 17 * to fill an integral number of DCT blocks horizontally. The source buffer 18 * may be modified if it is helpful for this purpose (the source buffer is 19 * allocated wide enough to correspond to the desired output width). 20 * The caller (the prep controller) is responsible for vertical padding. 21 * 22 * The downsampler may request "context rows" by setting need_context_rows 23 * during startup. In this case, the input arrays will contain at least 24 * one row group's worth of pixels above and below the passed-in data; 25 * the caller will create dummy rows at image top and bottom by replicating 26 * the first or last real pixel row. 27 * 28 * An excellent reference for image resampling is 29 * Digital Image Warping, George Wolberg, 1990. 30 * Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7. 31 * 32 * The downsampling algorithm used here is a simple average of the source 33 * pixels covered by the output pixel. The hi-falutin sampling literature 34 * refers to this as a "box filter". In general the characteristics of a box 35 * filter are not very good, but for the specific cases we normally use (1:1 36 * and 2:1 ratios) the box is equivalent to a "triangle filter" which is not 37 * nearly so bad. If you intend to use other sampling ratios, you'd be well 38 * advised to improve this code. 39 * 40 * A simple input-smoothing capability is provided. This is mainly intended 41 * for cleaning up color-dithered GIF input files (if you find it inadequate, 42 * we suggest using an external filtering program such as pnmconvol). When 43 * enabled, each input pixel P is replaced by a weighted sum of itself and its 44 * eight neighbors. P's weight is 1-8*SF and each neighbor's weight is SF, 45 * where SF = (smoothing_factor / 1024). 46 * Currently, smoothing is only supported for 2h2v sampling factors. 47 */ 48 49 #define JPEG_INTERNALS 50 #include "jinclude.h" 51 #include "jpeglib.h" 52 #include "jsimd.h" 53 54 55 /* Pointer to routine to downsample a single component */ 56 typedef JMETHOD(void, downsample1_ptr, 57 (j_compress_ptr cinfo, jpeg_component_info * compptr, 58 JSAMPARRAY input_data, JSAMPARRAY output_data)); 59 60 /* Private subobject */ 61 62 typedef struct { 63 struct jpeg_downsampler pub; /* public fields */ 64 65 /* Downsampling method pointers, one per component */ 66 downsample1_ptr methods[MAX_COMPONENTS]; 67 } my_downsampler; 68 69 typedef my_downsampler * my_downsample_ptr; 70 71 72 /* 73 * Initialize for a downsampling pass. 74 */ 75 76 METHODDEF(void) 77 start_pass_downsample (j_compress_ptr cinfo) 78 { 79 /* no work for now */ 80 } 81 82 83 /* 84 * Expand a component horizontally from width input_cols to width output_cols, 85 * by duplicating the rightmost samples. 86 */ 87 88 LOCAL(void) 89 expand_right_edge (JSAMPARRAY image_data, int num_rows, 90 JDIMENSION input_cols, JDIMENSION output_cols) 91 { 92 register JSAMPROW ptr; 93 register JSAMPLE pixval; 94 register int count; 95 int row; 96 int numcols = (int) (output_cols - input_cols); 97 98 if (numcols > 0) { 99 for (row = 0; row < num_rows; row++) { 100 ptr = image_data[row] + input_cols; 101 pixval = ptr[-1]; /* don't need GETJSAMPLE() here */ 102 for (count = numcols; count > 0; count--) 103 *ptr++ = pixval; 104 } 105 } 106 } 107 108 109 /* 110 * Do downsampling for a whole row group (all components). 111 * 112 * In this version we simply downsample each component independently. 113 */ 114 115 METHODDEF(void) 116 sep_downsample (j_compress_ptr cinfo, 117 JSAMPIMAGE input_buf, JDIMENSION in_row_index, 118 JSAMPIMAGE output_buf, JDIMENSION out_row_group_index) 119 { 120 my_downsample_ptr downsample = (my_downsample_ptr) cinfo->downsample; 121 int ci; 122 jpeg_component_info * compptr; 123 JSAMPARRAY in_ptr, out_ptr; 124 125 for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components; 126 ci++, compptr++) { 127 in_ptr = input_buf[ci] + in_row_index; 128 out_ptr = output_buf[ci] + (out_row_group_index * compptr->v_samp_factor); 129 (*downsample->methods[ci]) (cinfo, compptr, in_ptr, out_ptr); 130 } 131 } 132 133 134 /* 135 * Downsample pixel values of a single component. 136 * One row group is processed per call. 137 * This version handles arbitrary integral sampling ratios, without smoothing. 138 * Note that this version is not actually used for customary sampling ratios. 139 */ 140 141 METHODDEF(void) 142 int_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr, 143 JSAMPARRAY input_data, JSAMPARRAY output_data) 144 { 145 int inrow, outrow, h_expand, v_expand, numpix, numpix2, h, v; 146 JDIMENSION outcol, outcol_h; /* outcol_h == outcol*h_expand */ 147 JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE; 148 JSAMPROW inptr, outptr; 149 INT32 outvalue; 150 151 h_expand = cinfo->max_h_samp_factor / compptr->h_samp_factor; 152 v_expand = cinfo->max_v_samp_factor / compptr->v_samp_factor; 153 numpix = h_expand * v_expand; 154 numpix2 = numpix/2; 155 156 /* Expand input data enough to let all the output samples be generated 157 * by the standard loop. Special-casing padded output would be more 158 * efficient. 159 */ 160 expand_right_edge(input_data, cinfo->max_v_samp_factor, 161 cinfo->image_width, output_cols * h_expand); 162 163 inrow = 0; 164 for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) { 165 outptr = output_data[outrow]; 166 for (outcol = 0, outcol_h = 0; outcol < output_cols; 167 outcol++, outcol_h += h_expand) { 168 outvalue = 0; 169 for (v = 0; v < v_expand; v++) { 170 inptr = input_data[inrow+v] + outcol_h; 171 for (h = 0; h < h_expand; h++) { 172 outvalue += (INT32) GETJSAMPLE(*inptr++); 173 } 174 } 175 *outptr++ = (JSAMPLE) ((outvalue + numpix2) / numpix); 176 } 177 inrow += v_expand; 178 } 179 } 180 181 182 /* 183 * Downsample pixel values of a single component. 184 * This version handles the special case of a full-size component, 185 * without smoothing. 186 */ 187 188 METHODDEF(void) 189 fullsize_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr, 190 JSAMPARRAY input_data, JSAMPARRAY output_data) 191 { 192 /* Copy the data */ 193 jcopy_sample_rows(input_data, 0, output_data, 0, 194 cinfo->max_v_samp_factor, cinfo->image_width); 195 /* Edge-expand */ 196 expand_right_edge(output_data, cinfo->max_v_samp_factor, 197 cinfo->image_width, compptr->width_in_blocks * DCTSIZE); 198 } 199 200 201 /* 202 * Downsample pixel values of a single component. 203 * This version handles the common case of 2:1 horizontal and 1:1 vertical, 204 * without smoothing. 205 * 206 * A note about the "bias" calculations: when rounding fractional values to 207 * integer, we do not want to always round 0.5 up to the next integer. 208 * If we did that, we'd introduce a noticeable bias towards larger values. 209 * Instead, this code is arranged so that 0.5 will be rounded up or down at 210 * alternate pixel locations (a simple ordered dither pattern). 211 */ 212 213 METHODDEF(void) 214 h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr, 215 JSAMPARRAY input_data, JSAMPARRAY output_data) 216 { 217 int outrow; 218 JDIMENSION outcol; 219 JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE; 220 register JSAMPROW inptr, outptr; 221 register int bias; 222 223 /* Expand input data enough to let all the output samples be generated 224 * by the standard loop. Special-casing padded output would be more 225 * efficient. 226 */ 227 expand_right_edge(input_data, cinfo->max_v_samp_factor, 228 cinfo->image_width, output_cols * 2); 229 230 for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) { 231 outptr = output_data[outrow]; 232 inptr = input_data[outrow]; 233 bias = 0; /* bias = 0,1,0,1,... for successive samples */ 234 for (outcol = 0; outcol < output_cols; outcol++) { 235 *outptr++ = (JSAMPLE) ((GETJSAMPLE(*inptr) + GETJSAMPLE(inptr[1]) 236 + bias) >> 1); 237 bias ^= 1; /* 0=>1, 1=>0 */ 238 inptr += 2; 239 } 240 } 241 } 242 243 244 /* 245 * Downsample pixel values of a single component. 246 * This version handles the standard case of 2:1 horizontal and 2:1 vertical, 247 * without smoothing. 248 */ 249 250 METHODDEF(void) 251 h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr, 252 JSAMPARRAY input_data, JSAMPARRAY output_data) 253 { 254 int inrow, outrow; 255 JDIMENSION outcol; 256 JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE; 257 register JSAMPROW inptr0, inptr1, outptr; 258 register int bias; 259 260 /* Expand input data enough to let all the output samples be generated 261 * by the standard loop. Special-casing padded output would be more 262 * efficient. 263 */ 264 expand_right_edge(input_data, cinfo->max_v_samp_factor, 265 cinfo->image_width, output_cols * 2); 266 267 inrow = 0; 268 for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) { 269 outptr = output_data[outrow]; 270 inptr0 = input_data[inrow]; 271 inptr1 = input_data[inrow+1]; 272 bias = 1; /* bias = 1,2,1,2,... for successive samples */ 273 for (outcol = 0; outcol < output_cols; outcol++) { 274 *outptr++ = (JSAMPLE) ((GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) + 275 GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]) 276 + bias) >> 2); 277 bias ^= 3; /* 1=>2, 2=>1 */ 278 inptr0 += 2; inptr1 += 2; 279 } 280 inrow += 2; 281 } 282 } 283 284 285 #ifdef INPUT_SMOOTHING_SUPPORTED 286 287 /* 288 * Downsample pixel values of a single component. 289 * This version handles the standard case of 2:1 horizontal and 2:1 vertical, 290 * with smoothing. One row of context is required. 291 */ 292 293 METHODDEF(void) 294 h2v2_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr, 295 JSAMPARRAY input_data, JSAMPARRAY output_data) 296 { 297 int inrow, outrow; 298 JDIMENSION colctr; 299 JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE; 300 register JSAMPROW inptr0, inptr1, above_ptr, below_ptr, outptr; 301 INT32 membersum, neighsum, memberscale, neighscale; 302 303 /* Expand input data enough to let all the output samples be generated 304 * by the standard loop. Special-casing padded output would be more 305 * efficient. 306 */ 307 expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2, 308 cinfo->image_width, output_cols * 2); 309 310 /* We don't bother to form the individual "smoothed" input pixel values; 311 * we can directly compute the output which is the average of the four 312 * smoothed values. Each of the four member pixels contributes a fraction 313 * (1-8*SF) to its own smoothed image and a fraction SF to each of the three 314 * other smoothed pixels, therefore a total fraction (1-5*SF)/4 to the final 315 * output. The four corner-adjacent neighbor pixels contribute a fraction 316 * SF to just one smoothed pixel, or SF/4 to the final output; while the 317 * eight edge-adjacent neighbors contribute SF to each of two smoothed 318 * pixels, or SF/2 overall. In order to use integer arithmetic, these 319 * factors are scaled by 2^16 = 65536. 320 * Also recall that SF = smoothing_factor / 1024. 321 */ 322 323 memberscale = 16384 - cinfo->smoothing_factor * 80; /* scaled (1-5*SF)/4 */ 324 neighscale = cinfo->smoothing_factor * 16; /* scaled SF/4 */ 325 326 inrow = 0; 327 for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) { 328 outptr = output_data[outrow]; 329 inptr0 = input_data[inrow]; 330 inptr1 = input_data[inrow+1]; 331 above_ptr = input_data[inrow-1]; 332 below_ptr = input_data[inrow+2]; 333 334 /* Special case for first column: pretend column -1 is same as column 0 */ 335 membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) + 336 GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]); 337 neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) + 338 GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) + 339 GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[2]) + 340 GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[2]); 341 neighsum += neighsum; 342 neighsum += GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[2]) + 343 GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[2]); 344 membersum = membersum * memberscale + neighsum * neighscale; 345 *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16); 346 inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2; 347 348 for (colctr = output_cols - 2; colctr > 0; colctr--) { 349 /* sum of pixels directly mapped to this output element */ 350 membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) + 351 GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]); 352 /* sum of edge-neighbor pixels */ 353 neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) + 354 GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) + 355 GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[2]) + 356 GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[2]); 357 /* The edge-neighbors count twice as much as corner-neighbors */ 358 neighsum += neighsum; 359 /* Add in the corner-neighbors */ 360 neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[2]) + 361 GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[2]); 362 /* form final output scaled up by 2^16 */ 363 membersum = membersum * memberscale + neighsum * neighscale; 364 /* round, descale and output it */ 365 *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16); 366 inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2; 367 } 368 369 /* Special case for last column */ 370 membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) + 371 GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]); 372 neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) + 373 GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) + 374 GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[1]) + 375 GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[1]); 376 neighsum += neighsum; 377 neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[1]) + 378 GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[1]); 379 membersum = membersum * memberscale + neighsum * neighscale; 380 *outptr = (JSAMPLE) ((membersum + 32768) >> 16); 381 382 inrow += 2; 383 } 384 } 385 386 387 /* 388 * Downsample pixel values of a single component. 389 * This version handles the special case of a full-size component, 390 * with smoothing. One row of context is required. 391 */ 392 393 METHODDEF(void) 394 fullsize_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr, 395 JSAMPARRAY input_data, JSAMPARRAY output_data) 396 { 397 int outrow; 398 JDIMENSION colctr; 399 JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE; 400 register JSAMPROW inptr, above_ptr, below_ptr, outptr; 401 INT32 membersum, neighsum, memberscale, neighscale; 402 int colsum, lastcolsum, nextcolsum; 403 404 /* Expand input data enough to let all the output samples be generated 405 * by the standard loop. Special-casing padded output would be more 406 * efficient. 407 */ 408 expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2, 409 cinfo->image_width, output_cols); 410 411 /* Each of the eight neighbor pixels contributes a fraction SF to the 412 * smoothed pixel, while the main pixel contributes (1-8*SF). In order 413 * to use integer arithmetic, these factors are multiplied by 2^16 = 65536. 414 * Also recall that SF = smoothing_factor / 1024. 415 */ 416 417 memberscale = 65536L - cinfo->smoothing_factor * 512L; /* scaled 1-8*SF */ 418 neighscale = cinfo->smoothing_factor * 64; /* scaled SF */ 419 420 for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) { 421 outptr = output_data[outrow]; 422 inptr = input_data[outrow]; 423 above_ptr = input_data[outrow-1]; 424 below_ptr = input_data[outrow+1]; 425 426 /* Special case for first column */ 427 colsum = GETJSAMPLE(*above_ptr++) + GETJSAMPLE(*below_ptr++) + 428 GETJSAMPLE(*inptr); 429 membersum = GETJSAMPLE(*inptr++); 430 nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) + 431 GETJSAMPLE(*inptr); 432 neighsum = colsum + (colsum - membersum) + nextcolsum; 433 membersum = membersum * memberscale + neighsum * neighscale; 434 *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16); 435 lastcolsum = colsum; colsum = nextcolsum; 436 437 for (colctr = output_cols - 2; colctr > 0; colctr--) { 438 membersum = GETJSAMPLE(*inptr++); 439 above_ptr++; below_ptr++; 440 nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) + 441 GETJSAMPLE(*inptr); 442 neighsum = lastcolsum + (colsum - membersum) + nextcolsum; 443 membersum = membersum * memberscale + neighsum * neighscale; 444 *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16); 445 lastcolsum = colsum; colsum = nextcolsum; 446 } 447 448 /* Special case for last column */ 449 membersum = GETJSAMPLE(*inptr); 450 neighsum = lastcolsum + (colsum - membersum) + colsum; 451 membersum = membersum * memberscale + neighsum * neighscale; 452 *outptr = (JSAMPLE) ((membersum + 32768) >> 16); 453 454 } 455 } 456 457 #endif /* INPUT_SMOOTHING_SUPPORTED */ 458 459 460 /* 461 * Module initialization routine for downsampling. 462 * Note that we must select a routine for each component. 463 */ 464 465 GLOBAL(void) 466 jinit_downsampler (j_compress_ptr cinfo) 467 { 468 my_downsample_ptr downsample; 469 int ci; 470 jpeg_component_info * compptr; 471 boolean smoothok = TRUE; 472 473 downsample = (my_downsample_ptr) 474 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE, 475 SIZEOF(my_downsampler)); 476 cinfo->downsample = (struct jpeg_downsampler *) downsample; 477 downsample->pub.start_pass = start_pass_downsample; 478 downsample->pub.downsample = sep_downsample; 479 downsample->pub.need_context_rows = FALSE; 480 481 if (cinfo->CCIR601_sampling) 482 ERREXIT(cinfo, JERR_CCIR601_NOTIMPL); 483 484 /* Verify we can handle the sampling factors, and set up method pointers */ 485 for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components; 486 ci++, compptr++) { 487 if (compptr->h_samp_factor == cinfo->max_h_samp_factor && 488 compptr->v_samp_factor == cinfo->max_v_samp_factor) { 489 #ifdef INPUT_SMOOTHING_SUPPORTED 490 if (cinfo->smoothing_factor) { 491 downsample->methods[ci] = fullsize_smooth_downsample; 492 downsample->pub.need_context_rows = TRUE; 493 } else 494 #endif 495 downsample->methods[ci] = fullsize_downsample; 496 } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor && 497 compptr->v_samp_factor == cinfo->max_v_samp_factor) { 498 smoothok = FALSE; 499 if (jsimd_can_h2v1_downsample()) 500 downsample->methods[ci] = jsimd_h2v1_downsample; 501 else 502 downsample->methods[ci] = h2v1_downsample; 503 } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor && 504 compptr->v_samp_factor * 2 == cinfo->max_v_samp_factor) { 505 #ifdef INPUT_SMOOTHING_SUPPORTED 506 if (cinfo->smoothing_factor) { 507 downsample->methods[ci] = h2v2_smooth_downsample; 508 downsample->pub.need_context_rows = TRUE; 509 } else 510 #endif 511 if (jsimd_can_h2v2_downsample()) 512 downsample->methods[ci] = jsimd_h2v2_downsample; 513 else 514 downsample->methods[ci] = h2v2_downsample; 515 } else if ((cinfo->max_h_samp_factor % compptr->h_samp_factor) == 0 && 516 (cinfo->max_v_samp_factor % compptr->v_samp_factor) == 0) { 517 smoothok = FALSE; 518 downsample->methods[ci] = int_downsample; 519 } else 520 ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL); 521 } 522 523 #ifdef INPUT_SMOOTHING_SUPPORTED 524 if (cinfo->smoothing_factor && !smoothok) 525 TRACEMS(cinfo, 0, JTRC_SMOOTH_NOTIMPL); 526 #endif 527 } 528
