1 | /* |
2 | * jdsample.c |
3 | * |
4 | * This file was part of the Independent JPEG Group's software: |
5 | * Copyright (C) 1991-1996, Thomas G. Lane. |
6 | * libjpeg-turbo Modifications: |
7 | * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB |
8 | * Copyright (C) 2010, 2015-2016, D. R. Commander. |
9 | * Copyright (C) 2014, MIPS Technologies, Inc., California. |
10 | * Copyright (C) 2015, Google, Inc. |
11 | * Copyright (C) 2019-2020, Arm Limited. |
12 | * For conditions of distribution and use, see the accompanying README.ijg |
13 | * file. |
14 | * |
15 | * This file contains upsampling routines. |
16 | * |
17 | * Upsampling input data is counted in "row groups". A row group |
18 | * is defined to be (v_samp_factor * DCT_scaled_size / min_DCT_scaled_size) |
19 | * sample rows of each component. Upsampling will normally produce |
20 | * max_v_samp_factor pixel rows from each row group (but this could vary |
21 | * if the upsampler is applying a scale factor of its own). |
22 | * |
23 | * An excellent reference for image resampling is |
24 | * Digital Image Warping, George Wolberg, 1990. |
25 | * Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7. |
26 | */ |
27 | |
28 | #include "jinclude.h" |
29 | #include "jdsample.h" |
30 | #include "jsimd.h" |
31 | #include "jpegcomp.h" |
32 | |
33 | |
34 | |
35 | /* |
36 | * Initialize for an upsampling pass. |
37 | */ |
38 | |
39 | METHODDEF(void) |
40 | start_pass_upsample(j_decompress_ptr cinfo) |
41 | { |
42 | my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample; |
43 | |
44 | /* Mark the conversion buffer empty */ |
45 | upsample->next_row_out = cinfo->max_v_samp_factor; |
46 | /* Initialize total-height counter for detecting bottom of image */ |
47 | upsample->rows_to_go = cinfo->output_height; |
48 | } |
49 | |
50 | |
51 | /* |
52 | * Control routine to do upsampling (and color conversion). |
53 | * |
54 | * In this version we upsample each component independently. |
55 | * We upsample one row group into the conversion buffer, then apply |
56 | * color conversion a row at a time. |
57 | */ |
58 | |
59 | METHODDEF(void) |
60 | sep_upsample(j_decompress_ptr cinfo, JSAMPIMAGE input_buf, |
61 | JDIMENSION *in_row_group_ctr, JDIMENSION in_row_groups_avail, |
62 | JSAMPARRAY output_buf, JDIMENSION *out_row_ctr, |
63 | JDIMENSION out_rows_avail) |
64 | { |
65 | my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample; |
66 | int ci; |
67 | jpeg_component_info *compptr; |
68 | JDIMENSION num_rows; |
69 | |
70 | /* Fill the conversion buffer, if it's empty */ |
71 | if (upsample->next_row_out >= cinfo->max_v_samp_factor) { |
72 | for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components; |
73 | ci++, compptr++) { |
74 | /* Invoke per-component upsample method. Notice we pass a POINTER |
75 | * to color_buf[ci], so that fullsize_upsample can change it. |
76 | */ |
77 | (*upsample->methods[ci]) (cinfo, compptr, |
78 | input_buf[ci] + (*in_row_group_ctr * upsample->rowgroup_height[ci]), |
79 | upsample->color_buf + ci); |
80 | } |
81 | upsample->next_row_out = 0; |
82 | } |
83 | |
84 | /* Color-convert and emit rows */ |
85 | |
86 | /* How many we have in the buffer: */ |
87 | num_rows = (JDIMENSION)(cinfo->max_v_samp_factor - upsample->next_row_out); |
88 | /* Not more than the distance to the end of the image. Need this test |
89 | * in case the image height is not a multiple of max_v_samp_factor: |
90 | */ |
91 | if (num_rows > upsample->rows_to_go) |
92 | num_rows = upsample->rows_to_go; |
93 | /* And not more than what the client can accept: */ |
94 | out_rows_avail -= *out_row_ctr; |
95 | if (num_rows > out_rows_avail) |
96 | num_rows = out_rows_avail; |
97 | |
98 | (*cinfo->cconvert->color_convert) (cinfo, upsample->color_buf, |
99 | (JDIMENSION)upsample->next_row_out, |
100 | output_buf + *out_row_ctr, (int)num_rows); |
101 | |
102 | /* Adjust counts */ |
103 | *out_row_ctr += num_rows; |
104 | upsample->rows_to_go -= num_rows; |
105 | upsample->next_row_out += num_rows; |
106 | /* When the buffer is emptied, declare this input row group consumed */ |
107 | if (upsample->next_row_out >= cinfo->max_v_samp_factor) |
108 | (*in_row_group_ctr)++; |
109 | } |
110 | |
111 | |
112 | /* |
113 | * These are the routines invoked by sep_upsample to upsample pixel values |
114 | * of a single component. One row group is processed per call. |
115 | */ |
116 | |
117 | |
118 | /* |
119 | * For full-size components, we just make color_buf[ci] point at the |
120 | * input buffer, and thus avoid copying any data. Note that this is |
121 | * safe only because sep_upsample doesn't declare the input row group |
122 | * "consumed" until we are done color converting and emitting it. |
123 | */ |
124 | |
125 | METHODDEF(void) |
126 | fullsize_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr, |
127 | JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr) |
128 | { |
129 | *output_data_ptr = input_data; |
130 | } |
131 | |
132 | |
133 | /* |
134 | * This is a no-op version used for "uninteresting" components. |
135 | * These components will not be referenced by color conversion. |
136 | */ |
137 | |
138 | METHODDEF(void) |
139 | noop_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr, |
140 | JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr) |
141 | { |
142 | *output_data_ptr = NULL; /* safety check */ |
143 | } |
144 | |
145 | |
146 | /* |
147 | * This version handles any integral sampling ratios. |
148 | * This is not used for typical JPEG files, so it need not be fast. |
149 | * Nor, for that matter, is it particularly accurate: the algorithm is |
150 | * simple replication of the input pixel onto the corresponding output |
151 | * pixels. The hi-falutin sampling literature refers to this as a |
152 | * "box filter". A box filter tends to introduce visible artifacts, |
153 | * so if you are actually going to use 3:1 or 4:1 sampling ratios |
154 | * you would be well advised to improve this code. |
155 | */ |
156 | |
157 | METHODDEF(void) |
158 | int_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr, |
159 | JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr) |
160 | { |
161 | my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample; |
162 | JSAMPARRAY output_data = *output_data_ptr; |
163 | register JSAMPROW inptr, outptr; |
164 | register JSAMPLE invalue; |
165 | register int h; |
166 | JSAMPROW outend; |
167 | int h_expand, v_expand; |
168 | int inrow, outrow; |
169 | |
170 | h_expand = upsample->h_expand[compptr->component_index]; |
171 | v_expand = upsample->v_expand[compptr->component_index]; |
172 | |
173 | inrow = outrow = 0; |
174 | while (outrow < cinfo->max_v_samp_factor) { |
175 | /* Generate one output row with proper horizontal expansion */ |
176 | inptr = input_data[inrow]; |
177 | outptr = output_data[outrow]; |
178 | outend = outptr + cinfo->output_width; |
179 | while (outptr < outend) { |
180 | invalue = *inptr++; |
181 | for (h = h_expand; h > 0; h--) { |
182 | *outptr++ = invalue; |
183 | } |
184 | } |
185 | /* Generate any additional output rows by duplicating the first one */ |
186 | if (v_expand > 1) { |
187 | jcopy_sample_rows(output_data, outrow, output_data, outrow + 1, |
188 | v_expand - 1, cinfo->output_width); |
189 | } |
190 | inrow++; |
191 | outrow += v_expand; |
192 | } |
193 | } |
194 | |
195 | |
196 | /* |
197 | * Fast processing for the common case of 2:1 horizontal and 1:1 vertical. |
198 | * It's still a box filter. |
199 | */ |
200 | |
201 | METHODDEF(void) |
202 | h2v1_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr, |
203 | JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr) |
204 | { |
205 | JSAMPARRAY output_data = *output_data_ptr; |
206 | register JSAMPROW inptr, outptr; |
207 | register JSAMPLE invalue; |
208 | JSAMPROW outend; |
209 | int inrow; |
210 | |
211 | for (inrow = 0; inrow < cinfo->max_v_samp_factor; inrow++) { |
212 | inptr = input_data[inrow]; |
213 | outptr = output_data[inrow]; |
214 | outend = outptr + cinfo->output_width; |
215 | while (outptr < outend) { |
216 | invalue = *inptr++; |
217 | *outptr++ = invalue; |
218 | *outptr++ = invalue; |
219 | } |
220 | } |
221 | } |
222 | |
223 | |
224 | /* |
225 | * Fast processing for the common case of 2:1 horizontal and 2:1 vertical. |
226 | * It's still a box filter. |
227 | */ |
228 | |
229 | METHODDEF(void) |
230 | h2v2_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr, |
231 | JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr) |
232 | { |
233 | JSAMPARRAY output_data = *output_data_ptr; |
234 | register JSAMPROW inptr, outptr; |
235 | register JSAMPLE invalue; |
236 | JSAMPROW outend; |
237 | int inrow, outrow; |
238 | |
239 | inrow = outrow = 0; |
240 | while (outrow < cinfo->max_v_samp_factor) { |
241 | inptr = input_data[inrow]; |
242 | outptr = output_data[outrow]; |
243 | outend = outptr + cinfo->output_width; |
244 | while (outptr < outend) { |
245 | invalue = *inptr++; |
246 | *outptr++ = invalue; |
247 | *outptr++ = invalue; |
248 | } |
249 | jcopy_sample_rows(output_data, outrow, output_data, outrow + 1, 1, |
250 | cinfo->output_width); |
251 | inrow++; |
252 | outrow += 2; |
253 | } |
254 | } |
255 | |
256 | |
257 | /* |
258 | * Fancy processing for the common case of 2:1 horizontal and 1:1 vertical. |
259 | * |
260 | * The upsampling algorithm is linear interpolation between pixel centers, |
261 | * also known as a "triangle filter". This is a good compromise between |
262 | * speed and visual quality. The centers of the output pixels are 1/4 and 3/4 |
263 | * of the way between input pixel centers. |
264 | * |
265 | * A note about the "bias" calculations: when rounding fractional values to |
266 | * integer, we do not want to always round 0.5 up to the next integer. |
267 | * If we did that, we'd introduce a noticeable bias towards larger values. |
268 | * Instead, this code is arranged so that 0.5 will be rounded up or down at |
269 | * alternate pixel locations (a simple ordered dither pattern). |
270 | */ |
271 | |
272 | METHODDEF(void) |
273 | h2v1_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr, |
274 | JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr) |
275 | { |
276 | JSAMPARRAY output_data = *output_data_ptr; |
277 | register JSAMPROW inptr, outptr; |
278 | register int invalue; |
279 | register JDIMENSION colctr; |
280 | int inrow; |
281 | |
282 | for (inrow = 0; inrow < cinfo->max_v_samp_factor; inrow++) { |
283 | inptr = input_data[inrow]; |
284 | outptr = output_data[inrow]; |
285 | /* Special case for first column */ |
286 | invalue = *inptr++; |
287 | *outptr++ = (JSAMPLE)invalue; |
288 | *outptr++ = (JSAMPLE)((invalue * 3 + inptr[0] + 2) >> 2); |
289 | |
290 | for (colctr = compptr->downsampled_width - 2; colctr > 0; colctr--) { |
291 | /* General case: 3/4 * nearer pixel + 1/4 * further pixel */ |
292 | invalue = (*inptr++) * 3; |
293 | *outptr++ = (JSAMPLE)((invalue + inptr[-2] + 1) >> 2); |
294 | *outptr++ = (JSAMPLE)((invalue + inptr[0] + 2) >> 2); |
295 | } |
296 | |
297 | /* Special case for last column */ |
298 | invalue = *inptr; |
299 | *outptr++ = (JSAMPLE)((invalue * 3 + inptr[-1] + 1) >> 2); |
300 | *outptr++ = (JSAMPLE)invalue; |
301 | } |
302 | } |
303 | |
304 | |
305 | /* |
306 | * Fancy processing for 1:1 horizontal and 2:1 vertical (4:4:0 subsampling). |
307 | * |
308 | * This is a less common case, but it can be encountered when losslessly |
309 | * rotating/transposing a JPEG file that uses 4:2:2 chroma subsampling. |
310 | */ |
311 | |
312 | METHODDEF(void) |
313 | h1v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr, |
314 | JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr) |
315 | { |
316 | JSAMPARRAY output_data = *output_data_ptr; |
317 | JSAMPROW inptr0, inptr1, outptr; |
318 | #if BITS_IN_JSAMPLE == 8 |
319 | int thiscolsum, bias; |
320 | #else |
321 | JLONG thiscolsum, bias; |
322 | #endif |
323 | JDIMENSION colctr; |
324 | int inrow, outrow, v; |
325 | |
326 | inrow = outrow = 0; |
327 | while (outrow < cinfo->max_v_samp_factor) { |
328 | for (v = 0; v < 2; v++) { |
329 | /* inptr0 points to nearest input row, inptr1 points to next nearest */ |
330 | inptr0 = input_data[inrow]; |
331 | if (v == 0) { /* next nearest is row above */ |
332 | inptr1 = input_data[inrow - 1]; |
333 | bias = 1; |
334 | } else { /* next nearest is row below */ |
335 | inptr1 = input_data[inrow + 1]; |
336 | bias = 2; |
337 | } |
338 | outptr = output_data[outrow++]; |
339 | |
340 | for (colctr = 0; colctr < compptr->downsampled_width; colctr++) { |
341 | thiscolsum = (*inptr0++) * 3 + (*inptr1++); |
342 | *outptr++ = (JSAMPLE)((thiscolsum + bias) >> 2); |
343 | } |
344 | } |
345 | inrow++; |
346 | } |
347 | } |
348 | |
349 | |
350 | /* |
351 | * Fancy processing for the common case of 2:1 horizontal and 2:1 vertical. |
352 | * Again a triangle filter; see comments for h2v1 case, above. |
353 | * |
354 | * It is OK for us to reference the adjacent input rows because we demanded |
355 | * context from the main buffer controller (see initialization code). |
356 | */ |
357 | |
358 | METHODDEF(void) |
359 | h2v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr, |
360 | JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr) |
361 | { |
362 | JSAMPARRAY output_data = *output_data_ptr; |
363 | register JSAMPROW inptr0, inptr1, outptr; |
364 | #if BITS_IN_JSAMPLE == 8 |
365 | register int thiscolsum, lastcolsum, nextcolsum; |
366 | #else |
367 | register JLONG thiscolsum, lastcolsum, nextcolsum; |
368 | #endif |
369 | register JDIMENSION colctr; |
370 | int inrow, outrow, v; |
371 | |
372 | inrow = outrow = 0; |
373 | while (outrow < cinfo->max_v_samp_factor) { |
374 | for (v = 0; v < 2; v++) { |
375 | /* inptr0 points to nearest input row, inptr1 points to next nearest */ |
376 | inptr0 = input_data[inrow]; |
377 | if (v == 0) /* next nearest is row above */ |
378 | inptr1 = input_data[inrow - 1]; |
379 | else /* next nearest is row below */ |
380 | inptr1 = input_data[inrow + 1]; |
381 | outptr = output_data[outrow++]; |
382 | |
383 | /* Special case for first column */ |
384 | thiscolsum = (*inptr0++) * 3 + (*inptr1++); |
385 | nextcolsum = (*inptr0++) * 3 + (*inptr1++); |
386 | *outptr++ = (JSAMPLE)((thiscolsum * 4 + 8) >> 4); |
387 | *outptr++ = (JSAMPLE)((thiscolsum * 3 + nextcolsum + 7) >> 4); |
388 | lastcolsum = thiscolsum; thiscolsum = nextcolsum; |
389 | |
390 | for (colctr = compptr->downsampled_width - 2; colctr > 0; colctr--) { |
391 | /* General case: 3/4 * nearer pixel + 1/4 * further pixel in each */ |
392 | /* dimension, thus 9/16, 3/16, 3/16, 1/16 overall */ |
393 | nextcolsum = (*inptr0++) * 3 + (*inptr1++); |
394 | *outptr++ = (JSAMPLE)((thiscolsum * 3 + lastcolsum + 8) >> 4); |
395 | *outptr++ = (JSAMPLE)((thiscolsum * 3 + nextcolsum + 7) >> 4); |
396 | lastcolsum = thiscolsum; thiscolsum = nextcolsum; |
397 | } |
398 | |
399 | /* Special case for last column */ |
400 | *outptr++ = (JSAMPLE)((thiscolsum * 3 + lastcolsum + 8) >> 4); |
401 | *outptr++ = (JSAMPLE)((thiscolsum * 4 + 7) >> 4); |
402 | } |
403 | inrow++; |
404 | } |
405 | } |
406 | |
407 | |
408 | /* |
409 | * Module initialization routine for upsampling. |
410 | */ |
411 | |
412 | GLOBAL(void) |
413 | jinit_upsampler(j_decompress_ptr cinfo) |
414 | { |
415 | my_upsample_ptr upsample; |
416 | int ci; |
417 | jpeg_component_info *compptr; |
418 | boolean need_buffer, do_fancy; |
419 | int h_in_group, v_in_group, h_out_group, v_out_group; |
420 | |
421 | if (!cinfo->master->jinit_upsampler_no_alloc) { |
422 | upsample = (my_upsample_ptr) |
423 | (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE, |
424 | sizeof(my_upsampler)); |
425 | cinfo->upsample = (struct jpeg_upsampler *)upsample; |
426 | upsample->pub.start_pass = start_pass_upsample; |
427 | upsample->pub.upsample = sep_upsample; |
428 | upsample->pub.need_context_rows = FALSE; /* until we find out differently */ |
429 | } else |
430 | upsample = (my_upsample_ptr)cinfo->upsample; |
431 | |
432 | if (cinfo->CCIR601_sampling) /* this isn't supported */ |
433 | ERREXIT(cinfo, JERR_CCIR601_NOTIMPL); |
434 | |
435 | /* jdmainct.c doesn't support context rows when min_DCT_scaled_size = 1, |
436 | * so don't ask for it. |
437 | */ |
438 | do_fancy = cinfo->do_fancy_upsampling && cinfo->_min_DCT_scaled_size > 1; |
439 | |
440 | /* Verify we can handle the sampling factors, select per-component methods, |
441 | * and create storage as needed. |
442 | */ |
443 | for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components; |
444 | ci++, compptr++) { |
445 | /* Compute size of an "input group" after IDCT scaling. This many samples |
446 | * are to be converted to max_h_samp_factor * max_v_samp_factor pixels. |
447 | */ |
448 | h_in_group = (compptr->h_samp_factor * compptr->_DCT_scaled_size) / |
449 | cinfo->_min_DCT_scaled_size; |
450 | v_in_group = (compptr->v_samp_factor * compptr->_DCT_scaled_size) / |
451 | cinfo->_min_DCT_scaled_size; |
452 | h_out_group = cinfo->max_h_samp_factor; |
453 | v_out_group = cinfo->max_v_samp_factor; |
454 | upsample->rowgroup_height[ci] = v_in_group; /* save for use later */ |
455 | need_buffer = TRUE; |
456 | if (!compptr->component_needed) { |
457 | /* Don't bother to upsample an uninteresting component. */ |
458 | upsample->methods[ci] = noop_upsample; |
459 | need_buffer = FALSE; |
460 | } else if (h_in_group == h_out_group && v_in_group == v_out_group) { |
461 | /* Fullsize components can be processed without any work. */ |
462 | upsample->methods[ci] = fullsize_upsample; |
463 | need_buffer = FALSE; |
464 | } else if (h_in_group * 2 == h_out_group && v_in_group == v_out_group) { |
465 | /* Special cases for 2h1v upsampling */ |
466 | if (do_fancy && compptr->downsampled_width > 2) { |
467 | if (jsimd_can_h2v1_fancy_upsample()) |
468 | upsample->methods[ci] = jsimd_h2v1_fancy_upsample; |
469 | else |
470 | upsample->methods[ci] = h2v1_fancy_upsample; |
471 | } else { |
472 | if (jsimd_can_h2v1_upsample()) |
473 | upsample->methods[ci] = jsimd_h2v1_upsample; |
474 | else |
475 | upsample->methods[ci] = h2v1_upsample; |
476 | } |
477 | } else if (h_in_group == h_out_group && |
478 | v_in_group * 2 == v_out_group && do_fancy) { |
479 | /* Non-fancy upsampling is handled by the generic method */ |
480 | #if defined(__arm__) || defined(__aarch64__) || \ |
481 | defined(_M_ARM) || defined(_M_ARM64) |
482 | if (jsimd_can_h1v2_fancy_upsample()) |
483 | upsample->methods[ci] = jsimd_h1v2_fancy_upsample; |
484 | else |
485 | #endif |
486 | upsample->methods[ci] = h1v2_fancy_upsample; |
487 | upsample->pub.need_context_rows = TRUE; |
488 | } else if (h_in_group * 2 == h_out_group && |
489 | v_in_group * 2 == v_out_group) { |
490 | /* Special cases for 2h2v upsampling */ |
491 | if (do_fancy && compptr->downsampled_width > 2) { |
492 | if (jsimd_can_h2v2_fancy_upsample()) |
493 | upsample->methods[ci] = jsimd_h2v2_fancy_upsample; |
494 | else |
495 | upsample->methods[ci] = h2v2_fancy_upsample; |
496 | upsample->pub.need_context_rows = TRUE; |
497 | } else { |
498 | if (jsimd_can_h2v2_upsample()) |
499 | upsample->methods[ci] = jsimd_h2v2_upsample; |
500 | else |
501 | upsample->methods[ci] = h2v2_upsample; |
502 | } |
503 | } else if ((h_out_group % h_in_group) == 0 && |
504 | (v_out_group % v_in_group) == 0) { |
505 | /* Generic integral-factors upsampling method */ |
506 | #if defined(__mips__) |
507 | if (jsimd_can_int_upsample()) |
508 | upsample->methods[ci] = jsimd_int_upsample; |
509 | else |
510 | #endif |
511 | upsample->methods[ci] = int_upsample; |
512 | upsample->h_expand[ci] = (UINT8)(h_out_group / h_in_group); |
513 | upsample->v_expand[ci] = (UINT8)(v_out_group / v_in_group); |
514 | } else |
515 | ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL); |
516 | if (need_buffer && !cinfo->master->jinit_upsampler_no_alloc) { |
517 | upsample->color_buf[ci] = (*cinfo->mem->alloc_sarray) |
518 | ((j_common_ptr)cinfo, JPOOL_IMAGE, |
519 | (JDIMENSION)jround_up((long)cinfo->output_width, |
520 | (long)cinfo->max_h_samp_factor), |
521 | (JDIMENSION)cinfo->max_v_samp_factor); |
522 | } |
523 | } |
524 | } |
525 | |