Backed out changeset 1b14354719c0 (bug 1895254) for causing bustages on NavigationTra...
[gecko.git] / media / libjpeg / jcsample.c
blob30e6e54b40587bd657fe75b1c94764786fc5433c
1 /*
2 * jcsample.c
4 * This file was part of the Independent JPEG Group's software:
5 * Copyright (C) 1991-1996, Thomas G. Lane.
6 * Lossless JPEG Modifications:
7 * Copyright (C) 1999, Ken Murchison.
8 * libjpeg-turbo Modifications:
9 * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
10 * Copyright (C) 2014, MIPS Technologies, Inc., California.
11 * Copyright (C) 2015, 2019, 2022, D. R. Commander.
12 * For conditions of distribution and use, see the accompanying README.ijg
13 * file.
15 * This file contains downsampling routines.
17 * Downsampling input data is counted in "row groups". A row group
18 * is defined to be max_v_samp_factor pixel rows of each component,
19 * from which the downsampler produces v_samp_factor sample rows.
20 * A single row group is processed in each call to the downsampler module.
22 * The downsampler is responsible for edge-expansion of its output data
23 * to fill an integral number of DCT blocks horizontally. The source buffer
24 * may be modified if it is helpful for this purpose (the source buffer is
25 * allocated wide enough to correspond to the desired output width).
26 * The caller (the prep controller) is responsible for vertical padding.
28 * The downsampler may request "context rows" by setting need_context_rows
29 * during startup. In this case, the input arrays will contain at least
30 * one row group's worth of pixels above and below the passed-in data;
31 * the caller will create dummy rows at image top and bottom by replicating
32 * the first or last real pixel row.
34 * An excellent reference for image resampling is
35 * Digital Image Warping, George Wolberg, 1990.
36 * Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7.
38 * The downsampling algorithm used here is a simple average of the source
39 * pixels covered by the output pixel. The hi-falutin sampling literature
40 * refers to this as a "box filter". In general the characteristics of a box
41 * filter are not very good, but for the specific cases we normally use (1:1
42 * and 2:1 ratios) the box is equivalent to a "triangle filter" which is not
43 * nearly so bad. If you intend to use other sampling ratios, you'd be well
44 * advised to improve this code.
46 * A simple input-smoothing capability is provided. This is mainly intended
47 * for cleaning up color-dithered GIF input files (if you find it inadequate,
48 * we suggest using an external filtering program such as pnmconvol). When
49 * enabled, each input pixel P is replaced by a weighted sum of itself and its
50 * eight neighbors. P's weight is 1-8*SF and each neighbor's weight is SF,
51 * where SF = (smoothing_factor / 1024).
52 * Currently, smoothing is only supported for 2h2v sampling factors.
55 #define JPEG_INTERNALS
56 #include "jinclude.h"
57 #include "jpeglib.h"
58 #include "jsimd.h"
59 #include "jsamplecomp.h"
62 #if BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED)
64 /* Pointer to routine to downsample a single component */
65 typedef void (*downsample1_ptr) (j_compress_ptr cinfo,
66 jpeg_component_info *compptr,
67 _JSAMPARRAY input_data,
68 _JSAMPARRAY output_data);
70 /* Private subobject */
72 typedef struct {
73 struct jpeg_downsampler pub; /* public fields */
75 /* Downsampling method pointers, one per component */
76 downsample1_ptr methods[MAX_COMPONENTS];
77 } my_downsampler;
79 typedef my_downsampler *my_downsample_ptr;
83 * Initialize for a downsampling pass.
86 METHODDEF(void)
87 start_pass_downsample(j_compress_ptr cinfo)
89 /* no work for now */
94 * Expand a component horizontally from width input_cols to width output_cols,
95 * by duplicating the rightmost samples.
98 LOCAL(void)
99 expand_right_edge(_JSAMPARRAY image_data, int num_rows, JDIMENSION input_cols,
100 JDIMENSION output_cols)
102 register _JSAMPROW ptr;
103 register _JSAMPLE pixval;
104 register int count;
105 int row;
106 int numcols = (int)(output_cols - input_cols);
108 if (numcols > 0) {
109 for (row = 0; row < num_rows; row++) {
110 ptr = image_data[row] + input_cols;
111 pixval = ptr[-1];
112 for (count = numcols; count > 0; count--)
113 *ptr++ = pixval;
120 * Do downsampling for a whole row group (all components).
122 * In this version we simply downsample each component independently.
125 METHODDEF(void)
126 sep_downsample(j_compress_ptr cinfo, _JSAMPIMAGE input_buf,
127 JDIMENSION in_row_index, _JSAMPIMAGE output_buf,
128 JDIMENSION out_row_group_index)
130 my_downsample_ptr downsample = (my_downsample_ptr)cinfo->downsample;
131 int ci;
132 jpeg_component_info *compptr;
133 _JSAMPARRAY in_ptr, out_ptr;
135 for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
136 ci++, compptr++) {
137 in_ptr = input_buf[ci] + in_row_index;
138 out_ptr = output_buf[ci] + (out_row_group_index * compptr->v_samp_factor);
139 (*downsample->methods[ci]) (cinfo, compptr, in_ptr, out_ptr);
145 * Downsample pixel values of a single component.
146 * One row group is processed per call.
147 * This version handles arbitrary integral sampling ratios, without smoothing.
148 * Note that this version is not actually used for customary sampling ratios.
151 METHODDEF(void)
152 int_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
153 _JSAMPARRAY input_data, _JSAMPARRAY output_data)
155 int inrow, outrow, h_expand, v_expand, numpix, numpix2, h, v;
156 JDIMENSION outcol, outcol_h; /* outcol_h == outcol*h_expand */
157 int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
158 JDIMENSION output_cols = compptr->width_in_blocks * data_unit;
159 _JSAMPROW inptr, outptr;
160 JLONG outvalue;
162 h_expand = cinfo->max_h_samp_factor / compptr->h_samp_factor;
163 v_expand = cinfo->max_v_samp_factor / compptr->v_samp_factor;
164 numpix = h_expand * v_expand;
165 numpix2 = numpix / 2;
167 /* Expand input data enough to let all the output samples be generated
168 * by the standard loop. Special-casing padded output would be more
169 * efficient.
171 expand_right_edge(input_data, cinfo->max_v_samp_factor, cinfo->image_width,
172 output_cols * h_expand);
174 inrow = 0;
175 for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
176 outptr = output_data[outrow];
177 for (outcol = 0, outcol_h = 0; outcol < output_cols;
178 outcol++, outcol_h += h_expand) {
179 outvalue = 0;
180 for (v = 0; v < v_expand; v++) {
181 inptr = input_data[inrow + v] + outcol_h;
182 for (h = 0; h < h_expand; h++) {
183 outvalue += (JLONG)(*inptr++);
186 *outptr++ = (_JSAMPLE)((outvalue + numpix2) / numpix);
188 inrow += v_expand;
194 * Downsample pixel values of a single component.
195 * This version handles the special case of a full-size component,
196 * without smoothing.
199 METHODDEF(void)
200 fullsize_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
201 _JSAMPARRAY input_data, _JSAMPARRAY output_data)
203 int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
205 /* Copy the data */
206 _jcopy_sample_rows(input_data, 0, output_data, 0, cinfo->max_v_samp_factor,
207 cinfo->image_width);
208 /* Edge-expand */
209 expand_right_edge(output_data, cinfo->max_v_samp_factor, cinfo->image_width,
210 compptr->width_in_blocks * data_unit);
215 * Downsample pixel values of a single component.
216 * This version handles the common case of 2:1 horizontal and 1:1 vertical,
217 * without smoothing.
219 * A note about the "bias" calculations: when rounding fractional values to
220 * integer, we do not want to always round 0.5 up to the next integer.
221 * If we did that, we'd introduce a noticeable bias towards larger values.
222 * Instead, this code is arranged so that 0.5 will be rounded up or down at
223 * alternate pixel locations (a simple ordered dither pattern).
226 METHODDEF(void)
227 h2v1_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
228 _JSAMPARRAY input_data, _JSAMPARRAY output_data)
230 int outrow;
231 JDIMENSION outcol;
232 int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
233 JDIMENSION output_cols = compptr->width_in_blocks * data_unit;
234 register _JSAMPROW inptr, outptr;
235 register int bias;
237 /* Expand input data enough to let all the output samples be generated
238 * by the standard loop. Special-casing padded output would be more
239 * efficient.
241 expand_right_edge(input_data, cinfo->max_v_samp_factor, cinfo->image_width,
242 output_cols * 2);
244 for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
245 outptr = output_data[outrow];
246 inptr = input_data[outrow];
247 bias = 0; /* bias = 0,1,0,1,... for successive samples */
248 for (outcol = 0; outcol < output_cols; outcol++) {
249 *outptr++ = (_JSAMPLE)((inptr[0] + inptr[1] + bias) >> 1);
250 bias ^= 1; /* 0=>1, 1=>0 */
251 inptr += 2;
258 * Downsample pixel values of a single component.
259 * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
260 * without smoothing.
263 METHODDEF(void)
264 h2v2_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
265 _JSAMPARRAY input_data, _JSAMPARRAY output_data)
267 int inrow, outrow;
268 JDIMENSION outcol;
269 int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
270 JDIMENSION output_cols = compptr->width_in_blocks * data_unit;
271 register _JSAMPROW inptr0, inptr1, outptr;
272 register int bias;
274 /* Expand input data enough to let all the output samples be generated
275 * by the standard loop. Special-casing padded output would be more
276 * efficient.
278 expand_right_edge(input_data, cinfo->max_v_samp_factor, cinfo->image_width,
279 output_cols * 2);
281 inrow = 0;
282 for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
283 outptr = output_data[outrow];
284 inptr0 = input_data[inrow];
285 inptr1 = input_data[inrow + 1];
286 bias = 1; /* bias = 1,2,1,2,... for successive samples */
287 for (outcol = 0; outcol < output_cols; outcol++) {
288 *outptr++ = (_JSAMPLE)
289 ((inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1] + bias) >> 2);
290 bias ^= 3; /* 1=>2, 2=>1 */
291 inptr0 += 2; inptr1 += 2;
293 inrow += 2;
298 #ifdef INPUT_SMOOTHING_SUPPORTED
301 * Downsample pixel values of a single component.
302 * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
303 * with smoothing. One row of context is required.
306 METHODDEF(void)
307 h2v2_smooth_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
308 _JSAMPARRAY input_data, _JSAMPARRAY output_data)
310 int inrow, outrow;
311 JDIMENSION colctr;
312 int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
313 JDIMENSION output_cols = compptr->width_in_blocks * data_unit;
314 register _JSAMPROW inptr0, inptr1, above_ptr, below_ptr, outptr;
315 JLONG membersum, neighsum, memberscale, neighscale;
317 /* Expand input data enough to let all the output samples be generated
318 * by the standard loop. Special-casing padded output would be more
319 * efficient.
321 expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
322 cinfo->image_width, output_cols * 2);
324 /* We don't bother to form the individual "smoothed" input pixel values;
325 * we can directly compute the output which is the average of the four
326 * smoothed values. Each of the four member pixels contributes a fraction
327 * (1-8*SF) to its own smoothed image and a fraction SF to each of the three
328 * other smoothed pixels, therefore a total fraction (1-5*SF)/4 to the final
329 * output. The four corner-adjacent neighbor pixels contribute a fraction
330 * SF to just one smoothed pixel, or SF/4 to the final output; while the
331 * eight edge-adjacent neighbors contribute SF to each of two smoothed
332 * pixels, or SF/2 overall. In order to use integer arithmetic, these
333 * factors are scaled by 2^16 = 65536.
334 * Also recall that SF = smoothing_factor / 1024.
337 memberscale = 16384 - cinfo->smoothing_factor * 80; /* scaled (1-5*SF)/4 */
338 neighscale = cinfo->smoothing_factor * 16; /* scaled SF/4 */
340 inrow = 0;
341 for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
342 outptr = output_data[outrow];
343 inptr0 = input_data[inrow];
344 inptr1 = input_data[inrow + 1];
345 above_ptr = input_data[inrow - 1];
346 below_ptr = input_data[inrow + 2];
348 /* Special case for first column: pretend column -1 is same as column 0 */
349 membersum = inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1];
350 neighsum = above_ptr[0] + above_ptr[1] + below_ptr[0] + below_ptr[1] +
351 inptr0[0] + inptr0[2] + inptr1[0] + inptr1[2];
352 neighsum += neighsum;
353 neighsum += above_ptr[0] + above_ptr[2] + below_ptr[0] + below_ptr[2];
354 membersum = membersum * memberscale + neighsum * neighscale;
355 *outptr++ = (_JSAMPLE)((membersum + 32768) >> 16);
356 inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
358 for (colctr = output_cols - 2; colctr > 0; colctr--) {
359 /* sum of pixels directly mapped to this output element */
360 membersum = inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1];
361 /* sum of edge-neighbor pixels */
362 neighsum = above_ptr[0] + above_ptr[1] + below_ptr[0] + below_ptr[1] +
363 inptr0[-1] + inptr0[2] + inptr1[-1] + inptr1[2];
364 /* The edge-neighbors count twice as much as corner-neighbors */
365 neighsum += neighsum;
366 /* Add in the corner-neighbors */
367 neighsum += above_ptr[-1] + above_ptr[2] + below_ptr[-1] + below_ptr[2];
368 /* form final output scaled up by 2^16 */
369 membersum = membersum * memberscale + neighsum * neighscale;
370 /* round, descale and output it */
371 *outptr++ = (_JSAMPLE)((membersum + 32768) >> 16);
372 inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
375 /* Special case for last column */
376 membersum = inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1];
377 neighsum = above_ptr[0] + above_ptr[1] + below_ptr[0] + below_ptr[1] +
378 inptr0[-1] + inptr0[1] + inptr1[-1] + inptr1[1];
379 neighsum += neighsum;
380 neighsum += above_ptr[-1] + above_ptr[1] + below_ptr[-1] + below_ptr[1];
381 membersum = membersum * memberscale + neighsum * neighscale;
382 *outptr = (_JSAMPLE)((membersum + 32768) >> 16);
384 inrow += 2;
390 * Downsample pixel values of a single component.
391 * This version handles the special case of a full-size component,
392 * with smoothing. One row of context is required.
395 METHODDEF(void)
396 fullsize_smooth_downsample(j_compress_ptr cinfo, jpeg_component_info *compptr,
397 _JSAMPARRAY input_data, _JSAMPARRAY output_data)
399 int outrow;
400 JDIMENSION colctr;
401 int data_unit = cinfo->master->lossless ? 1 : DCTSIZE;
402 JDIMENSION output_cols = compptr->width_in_blocks * data_unit;
403 register _JSAMPROW inptr, above_ptr, below_ptr, outptr;
404 JLONG membersum, neighsum, memberscale, neighscale;
405 int colsum, lastcolsum, nextcolsum;
407 /* Expand input data enough to let all the output samples be generated
408 * by the standard loop. Special-casing padded output would be more
409 * efficient.
411 expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
412 cinfo->image_width, output_cols);
414 /* Each of the eight neighbor pixels contributes a fraction SF to the
415 * smoothed pixel, while the main pixel contributes (1-8*SF). In order
416 * to use integer arithmetic, these factors are multiplied by 2^16 = 65536.
417 * Also recall that SF = smoothing_factor / 1024.
420 memberscale = 65536L - cinfo->smoothing_factor * 512L; /* scaled 1-8*SF */
421 neighscale = cinfo->smoothing_factor * 64; /* scaled SF */
423 for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
424 outptr = output_data[outrow];
425 inptr = input_data[outrow];
426 above_ptr = input_data[outrow - 1];
427 below_ptr = input_data[outrow + 1];
429 /* Special case for first column */
430 colsum = (*above_ptr++) + (*below_ptr++) + inptr[0];
431 membersum = *inptr++;
432 nextcolsum = above_ptr[0] + below_ptr[0] + inptr[0];
433 neighsum = colsum + (colsum - membersum) + nextcolsum;
434 membersum = membersum * memberscale + neighsum * neighscale;
435 *outptr++ = (_JSAMPLE)((membersum + 32768) >> 16);
436 lastcolsum = colsum; colsum = nextcolsum;
438 for (colctr = output_cols - 2; colctr > 0; colctr--) {
439 membersum = *inptr++;
440 above_ptr++; below_ptr++;
441 nextcolsum = above_ptr[0] + below_ptr[0] + inptr[0];
442 neighsum = lastcolsum + (colsum - membersum) + nextcolsum;
443 membersum = membersum * memberscale + neighsum * neighscale;
444 *outptr++ = (_JSAMPLE)((membersum + 32768) >> 16);
445 lastcolsum = colsum; colsum = nextcolsum;
448 /* Special case for last column */
449 membersum = *inptr;
450 neighsum = lastcolsum + (colsum - membersum) + colsum;
451 membersum = membersum * memberscale + neighsum * neighscale;
452 *outptr = (_JSAMPLE)((membersum + 32768) >> 16);
457 #endif /* INPUT_SMOOTHING_SUPPORTED */
461 * Module initialization routine for downsampling.
462 * Note that we must select a routine for each component.
465 GLOBAL(void)
466 _jinit_downsampler(j_compress_ptr cinfo)
468 my_downsample_ptr downsample;
469 int ci;
470 jpeg_component_info *compptr;
471 boolean smoothok = TRUE;
473 if (cinfo->data_precision != BITS_IN_JSAMPLE)
474 ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
476 downsample = (my_downsample_ptr)
477 (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
478 sizeof(my_downsampler));
479 cinfo->downsample = (struct jpeg_downsampler *)downsample;
480 downsample->pub.start_pass = start_pass_downsample;
481 downsample->pub._downsample = sep_downsample;
482 downsample->pub.need_context_rows = FALSE;
484 if (cinfo->CCIR601_sampling)
485 ERREXIT(cinfo, JERR_CCIR601_NOTIMPL);
487 /* Verify we can handle the sampling factors, and set up method pointers */
488 for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
489 ci++, compptr++) {
490 if (compptr->h_samp_factor == cinfo->max_h_samp_factor &&
491 compptr->v_samp_factor == cinfo->max_v_samp_factor) {
492 #ifdef INPUT_SMOOTHING_SUPPORTED
493 if (cinfo->smoothing_factor) {
494 downsample->methods[ci] = fullsize_smooth_downsample;
495 downsample->pub.need_context_rows = TRUE;
496 } else
497 #endif
498 downsample->methods[ci] = fullsize_downsample;
499 } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
500 compptr->v_samp_factor == cinfo->max_v_samp_factor) {
501 smoothok = FALSE;
502 #ifdef WITH_SIMD
503 if (jsimd_can_h2v1_downsample())
504 downsample->methods[ci] = jsimd_h2v1_downsample;
505 else
506 #endif
507 downsample->methods[ci] = h2v1_downsample;
508 } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
509 compptr->v_samp_factor * 2 == cinfo->max_v_samp_factor) {
510 #ifdef INPUT_SMOOTHING_SUPPORTED
511 if (cinfo->smoothing_factor) {
512 #if defined(WITH_SIMD) && defined(__mips__)
513 if (jsimd_can_h2v2_smooth_downsample())
514 downsample->methods[ci] = jsimd_h2v2_smooth_downsample;
515 else
516 #endif
517 downsample->methods[ci] = h2v2_smooth_downsample;
518 downsample->pub.need_context_rows = TRUE;
519 } else
520 #endif
522 #ifdef WITH_SIMD
523 if (jsimd_can_h2v2_downsample())
524 downsample->methods[ci] = jsimd_h2v2_downsample;
525 else
526 #endif
527 downsample->methods[ci] = h2v2_downsample;
529 } else if ((cinfo->max_h_samp_factor % compptr->h_samp_factor) == 0 &&
530 (cinfo->max_v_samp_factor % compptr->v_samp_factor) == 0) {
531 smoothok = FALSE;
532 downsample->methods[ci] = int_downsample;
533 } else
534 ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL);
537 #ifdef INPUT_SMOOTHING_SUPPORTED
538 if (cinfo->smoothing_factor && !smoothok)
539 TRACEMS(cinfo, 0, JTRC_SMOOTH_NOTIMPL);
540 #endif
543 #endif /* BITS_IN_JSAMPLE != 16 || defined(C_LOSSLESS_SUPPORTED) */