4 * This file was part of the Independent JPEG Group's software:
5 * Copyright (C) 1994-1996, Thomas G. Lane.
6 * libjpeg-turbo Modifications:
7 * Copyright (C) 2010, 2016, 2022, D. R. Commander.
8 * For conditions of distribution and use, see the accompanying README.ijg
11 * This file contains the main buffer controller for decompression.
12 * The main buffer lies between the JPEG decompressor proper and the
13 * post-processor; it holds downsampled data in the JPEG colorspace.
15 * Note that this code is bypassed in raw-data mode, since the application
16 * supplies the equivalent of the main buffer in that case.
23 #if BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED)
26 * In the current system design, the main buffer need never be a full-image
27 * buffer; any full-height buffers will be found inside the coefficient,
28 * difference, or postprocessing controllers. Nonetheless, the main controller
29 * is not trivial. Its responsibility is to provide context rows for
30 * upsampling/rescaling, and doing this in an efficient fashion is a bit
33 * Postprocessor input data is counted in "row groups". A row group
34 * is defined to be (v_samp_factor * DCT_scaled_size / min_DCT_scaled_size)
35 * sample rows of each component. (We require DCT_scaled_size values to be
36 * chosen such that these numbers are integers. In practice DCT_scaled_size
37 * values will likely be powers of two, so we actually have the stronger
38 * condition that DCT_scaled_size / min_DCT_scaled_size is an integer.)
39 * Upsampling will typically produce max_v_samp_factor pixel rows from each
40 * row group (times any additional scale factor that the upsampler is
43 * The coefficient or difference controller will deliver data to us one iMCU
44 * row at a time; each iMCU row contains v_samp_factor * DCT_scaled_size sample
45 * rows, or exactly min_DCT_scaled_size row groups. (This amount of data
46 * corresponds to one row of MCUs when the image is fully interleaved.) Note
47 * that the number of sample rows varies across components, but the number of
48 * row groups does not. Some garbage sample rows may be included in the last
49 * iMCU row at the bottom of the image.
51 * Depending on the vertical scaling algorithm used, the upsampler may need
52 * access to the sample row(s) above and below its current input row group.
53 * The upsampler is required to set need_context_rows TRUE at global selection
54 * time if so. When need_context_rows is FALSE, this controller can simply
55 * obtain one iMCU row at a time from the coefficient or difference controller
56 * and dole it out as row groups to the postprocessor.
58 * When need_context_rows is TRUE, this controller guarantees that the buffer
59 * passed to postprocessing contains at least one row group's worth of samples
60 * above and below the row group(s) being processed. Note that the context
61 * rows "above" the first passed row group appear at negative row offsets in
62 * the passed buffer. At the top and bottom of the image, the required
63 * context rows are manufactured by duplicating the first or last real sample
64 * row; this avoids having special cases in the upsampling inner loops.
66 * The amount of context is fixed at one row group just because that's a
67 * convenient number for this controller to work with. The existing
68 * upsamplers really only need one sample row of context. An upsampler
69 * supporting arbitrary output rescaling might wish for more than one row
70 * group of context when shrinking the image; tough, we don't handle that.
71 * (This is justified by the assumption that downsizing will be handled mostly
72 * by adjusting the DCT_scaled_size values, so that the actual scale factor at
73 * the upsample step needn't be much less than one.)
75 * To provide the desired context, we have to retain the last two row groups
76 * of one iMCU row while reading in the next iMCU row. (The last row group
77 * can't be processed until we have another row group for its below-context,
78 * and so we have to save the next-to-last group too for its above-context.)
79 * We could do this most simply by copying data around in our buffer, but
80 * that'd be very slow. We can avoid copying any data by creating a rather
81 * strange pointer structure. Here's how it works. We allocate a workspace
82 * consisting of M+2 row groups (where M = min_DCT_scaled_size is the number
83 * of row groups per iMCU row). We create two sets of redundant pointers to
84 * the workspace. Labeling the physical row groups 0 to M+1, the synthesized
85 * pointer lists look like this:
87 * master pointer --> 0 master pointer --> 0
96 * We read alternate iMCU rows using each master pointer; thus the last two
97 * row groups of the previous iMCU row remain un-overwritten in the workspace.
98 * The pointer lists are set up so that the required context rows appear to
99 * be adjacent to the proper places when we pass the pointer lists to the
102 * The above pictures describe the normal state of the pointer lists.
103 * At top and bottom of the image, we diddle the pointer lists to duplicate
104 * the first or last sample row as necessary (this is cheaper than copying
105 * sample rows around).
107 * This scheme breaks down if M < 2, ie, min_DCT_scaled_size is 1. In that
108 * situation each iMCU row provides only one row group so the buffering logic
109 * must be different (eg, we must read two iMCU rows before we can emit the
110 * first row group). For now, we simply do not support providing context
111 * rows when min_DCT_scaled_size is 1. That combination seems unlikely to
112 * be worth providing --- if someone wants a 1/8th-size preview, they probably
113 * want it quick and dirty, so a context-free upsampler is sufficient.
117 /* Forward declarations */
118 METHODDEF(void) process_data_simple_main(j_decompress_ptr cinfo,
119 _JSAMPARRAY output_buf,
120 JDIMENSION *out_row_ctr,
121 JDIMENSION out_rows_avail);
122 METHODDEF(void) process_data_context_main(j_decompress_ptr cinfo,
123 _JSAMPARRAY output_buf,
124 JDIMENSION *out_row_ctr,
125 JDIMENSION out_rows_avail);
126 #ifdef QUANT_2PASS_SUPPORTED
127 METHODDEF(void) process_data_crank_post(j_decompress_ptr cinfo,
128 _JSAMPARRAY output_buf,
129 JDIMENSION *out_row_ctr,
130 JDIMENSION out_rows_avail);
135 alloc_funny_pointers(j_decompress_ptr cinfo)
136 /* Allocate space for the funny pointer lists.
137 * This is done only once, not once per pass.
140 my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
142 int M = cinfo->_min_DCT_scaled_size;
143 jpeg_component_info *compptr;
146 /* Get top-level space for component array pointers.
147 * We alloc both arrays with one call to save a few cycles.
149 main_ptr->xbuffer[0] = (_JSAMPIMAGE)
150 (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
151 cinfo->num_components * 2 *
152 sizeof(_JSAMPARRAY));
153 main_ptr->xbuffer[1] = main_ptr->xbuffer[0] + cinfo->num_components;
155 for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
157 rgroup = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
158 cinfo->_min_DCT_scaled_size; /* height of a row group of component */
159 /* Get space for pointer lists --- M+4 row groups in each list.
160 * We alloc both pointer lists with one call to save a few cycles.
163 (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
164 2 * (rgroup * (M + 4)) * sizeof(_JSAMPROW));
165 xbuf += rgroup; /* want one row group at negative offsets */
166 main_ptr->xbuffer[0][ci] = xbuf;
167 xbuf += rgroup * (M + 4);
168 main_ptr->xbuffer[1][ci] = xbuf;
174 make_funny_pointers(j_decompress_ptr cinfo)
175 /* Create the funny pointer lists discussed in the comments above.
176 * The actual workspace is already allocated (in main_ptr->buffer),
177 * and the space for the pointer lists is allocated too.
178 * This routine just fills in the curiously ordered lists.
179 * This will be repeated at the beginning of each pass.
182 my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
184 int M = cinfo->_min_DCT_scaled_size;
185 jpeg_component_info *compptr;
186 _JSAMPARRAY buf, xbuf0, xbuf1;
188 for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
190 rgroup = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
191 cinfo->_min_DCT_scaled_size; /* height of a row group of component */
192 xbuf0 = main_ptr->xbuffer[0][ci];
193 xbuf1 = main_ptr->xbuffer[1][ci];
194 /* First copy the workspace pointers as-is */
195 buf = main_ptr->buffer[ci];
196 for (i = 0; i < rgroup * (M + 2); i++) {
197 xbuf0[i] = xbuf1[i] = buf[i];
199 /* In the second list, put the last four row groups in swapped order */
200 for (i = 0; i < rgroup * 2; i++) {
201 xbuf1[rgroup * (M - 2) + i] = buf[rgroup * M + i];
202 xbuf1[rgroup * M + i] = buf[rgroup * (M - 2) + i];
204 /* The wraparound pointers at top and bottom will be filled later
205 * (see set_wraparound_pointers, below). Initially we want the "above"
206 * pointers to duplicate the first actual data line. This only needs
207 * to happen in xbuffer[0].
209 for (i = 0; i < rgroup; i++) {
210 xbuf0[i - rgroup] = xbuf0[0];
217 set_bottom_pointers(j_decompress_ptr cinfo)
218 /* Change the pointer lists to duplicate the last sample row at the bottom
219 * of the image. whichptr indicates which xbuffer holds the final iMCU row.
220 * Also sets rowgroups_avail to indicate number of nondummy row groups in row.
223 my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
224 int ci, i, rgroup, iMCUheight, rows_left;
225 jpeg_component_info *compptr;
228 for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
230 /* Count sample rows in one iMCU row and in one row group */
231 iMCUheight = compptr->v_samp_factor * compptr->_DCT_scaled_size;
232 rgroup = iMCUheight / cinfo->_min_DCT_scaled_size;
233 /* Count nondummy sample rows remaining for this component */
234 rows_left = (int)(compptr->downsampled_height % (JDIMENSION)iMCUheight);
235 if (rows_left == 0) rows_left = iMCUheight;
236 /* Count nondummy row groups. Should get same answer for each component,
237 * so we need only do it once.
240 main_ptr->rowgroups_avail = (JDIMENSION)((rows_left - 1) / rgroup + 1);
242 /* Duplicate the last real sample row rgroup*2 times; this pads out the
243 * last partial rowgroup and ensures at least one full rowgroup of context.
245 xbuf = main_ptr->xbuffer[main_ptr->whichptr][ci];
246 for (i = 0; i < rgroup * 2; i++) {
247 xbuf[rows_left + i] = xbuf[rows_left - 1];
254 * Initialize for a processing pass.
258 start_pass_main(j_decompress_ptr cinfo, J_BUF_MODE pass_mode)
260 my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
264 if (cinfo->upsample->need_context_rows) {
265 main_ptr->pub._process_data = process_data_context_main;
266 make_funny_pointers(cinfo); /* Create the xbuffer[] lists */
267 main_ptr->whichptr = 0; /* Read first iMCU row into xbuffer[0] */
268 main_ptr->context_state = CTX_PREPARE_FOR_IMCU;
269 main_ptr->iMCU_row_ctr = 0;
271 /* Simple case with no context needed */
272 main_ptr->pub._process_data = process_data_simple_main;
274 main_ptr->buffer_full = FALSE; /* Mark buffer empty */
275 main_ptr->rowgroup_ctr = 0;
277 #ifdef QUANT_2PASS_SUPPORTED
278 case JBUF_CRANK_DEST:
279 /* For last pass of 2-pass quantization, just crank the postprocessor */
280 main_ptr->pub._process_data = process_data_crank_post;
284 ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
292 * This handles the simple case where no context is required.
296 process_data_simple_main(j_decompress_ptr cinfo, _JSAMPARRAY output_buf,
297 JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
299 my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
300 JDIMENSION rowgroups_avail;
302 /* Read input data if we haven't filled the main buffer yet */
303 if (!main_ptr->buffer_full) {
304 if (!(*cinfo->coef->_decompress_data) (cinfo, main_ptr->buffer))
305 return; /* suspension forced, can do nothing more */
306 main_ptr->buffer_full = TRUE; /* OK, we have an iMCU row to work with */
309 /* There are always min_DCT_scaled_size row groups in an iMCU row. */
310 rowgroups_avail = (JDIMENSION)cinfo->_min_DCT_scaled_size;
311 /* Note: at the bottom of the image, we may pass extra garbage row groups
312 * to the postprocessor. The postprocessor has to check for bottom
313 * of image anyway (at row resolution), so no point in us doing it too.
316 /* Feed the postprocessor */
317 (*cinfo->post->_post_process_data) (cinfo, main_ptr->buffer,
318 &main_ptr->rowgroup_ctr, rowgroups_avail,
319 output_buf, out_row_ctr, out_rows_avail);
321 /* Has postprocessor consumed all the data yet? If so, mark buffer empty */
322 if (main_ptr->rowgroup_ctr >= rowgroups_avail) {
323 main_ptr->buffer_full = FALSE;
324 main_ptr->rowgroup_ctr = 0;
331 * This handles the case where context rows must be provided.
335 process_data_context_main(j_decompress_ptr cinfo, _JSAMPARRAY output_buf,
336 JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
338 my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
340 /* Read input data if we haven't filled the main buffer yet */
341 if (!main_ptr->buffer_full) {
342 if (!(*cinfo->coef->_decompress_data) (cinfo,
343 main_ptr->xbuffer[main_ptr->whichptr]))
344 return; /* suspension forced, can do nothing more */
345 main_ptr->buffer_full = TRUE; /* OK, we have an iMCU row to work with */
346 main_ptr->iMCU_row_ctr++; /* count rows received */
349 /* Postprocessor typically will not swallow all the input data it is handed
350 * in one call (due to filling the output buffer first). Must be prepared
351 * to exit and restart. This switch lets us keep track of how far we got.
352 * Note that each case falls through to the next on successful completion.
354 switch (main_ptr->context_state) {
355 case CTX_POSTPONED_ROW:
356 /* Call postprocessor using previously set pointers for postponed row */
357 (*cinfo->post->_post_process_data) (cinfo,
358 main_ptr->xbuffer[main_ptr->whichptr],
359 &main_ptr->rowgroup_ctr,
360 main_ptr->rowgroups_avail, output_buf,
361 out_row_ctr, out_rows_avail);
362 if (main_ptr->rowgroup_ctr < main_ptr->rowgroups_avail)
363 return; /* Need to suspend */
364 main_ptr->context_state = CTX_PREPARE_FOR_IMCU;
365 if (*out_row_ctr >= out_rows_avail)
366 return; /* Postprocessor exactly filled output buf */
367 FALLTHROUGH /*FALLTHROUGH*/
368 case CTX_PREPARE_FOR_IMCU:
369 /* Prepare to process first M-1 row groups of this iMCU row */
370 main_ptr->rowgroup_ctr = 0;
371 main_ptr->rowgroups_avail = (JDIMENSION)(cinfo->_min_DCT_scaled_size - 1);
372 /* Check for bottom of image: if so, tweak pointers to "duplicate"
373 * the last sample row, and adjust rowgroups_avail to ignore padding rows.
375 if (main_ptr->iMCU_row_ctr == cinfo->total_iMCU_rows)
376 set_bottom_pointers(cinfo);
377 main_ptr->context_state = CTX_PROCESS_IMCU;
378 FALLTHROUGH /*FALLTHROUGH*/
379 case CTX_PROCESS_IMCU:
380 /* Call postprocessor using previously set pointers */
381 (*cinfo->post->_post_process_data) (cinfo,
382 main_ptr->xbuffer[main_ptr->whichptr],
383 &main_ptr->rowgroup_ctr,
384 main_ptr->rowgroups_avail, output_buf,
385 out_row_ctr, out_rows_avail);
386 if (main_ptr->rowgroup_ctr < main_ptr->rowgroups_avail)
387 return; /* Need to suspend */
388 /* After the first iMCU, change wraparound pointers to normal state */
389 if (main_ptr->iMCU_row_ctr == 1)
390 set_wraparound_pointers(cinfo);
391 /* Prepare to load new iMCU row using other xbuffer list */
392 main_ptr->whichptr ^= 1; /* 0=>1 or 1=>0 */
393 main_ptr->buffer_full = FALSE;
394 /* Still need to process last row group of this iMCU row, */
395 /* which is saved at index M+1 of the other xbuffer */
396 main_ptr->rowgroup_ctr = (JDIMENSION)(cinfo->_min_DCT_scaled_size + 1);
397 main_ptr->rowgroups_avail = (JDIMENSION)(cinfo->_min_DCT_scaled_size + 2);
398 main_ptr->context_state = CTX_POSTPONED_ROW;
405 * Final pass of two-pass quantization: just call the postprocessor.
406 * Source data will be the postprocessor controller's internal buffer.
409 #ifdef QUANT_2PASS_SUPPORTED
412 process_data_crank_post(j_decompress_ptr cinfo, _JSAMPARRAY output_buf,
413 JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
415 (*cinfo->post->_post_process_data) (cinfo, (_JSAMPIMAGE)NULL,
416 (JDIMENSION *)NULL, (JDIMENSION)0,
417 output_buf, out_row_ctr, out_rows_avail);
420 #endif /* QUANT_2PASS_SUPPORTED */
424 * Initialize main buffer controller.
428 _jinit_d_main_controller(j_decompress_ptr cinfo, boolean need_full_buffer)
430 my_main_ptr main_ptr;
431 int ci, rgroup, ngroups;
432 jpeg_component_info *compptr;
434 if (cinfo->data_precision != BITS_IN_JSAMPLE)
435 ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
437 main_ptr = (my_main_ptr)
438 (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
439 sizeof(my_main_controller));
440 cinfo->main = (struct jpeg_d_main_controller *)main_ptr;
441 main_ptr->pub.start_pass = start_pass_main;
443 if (need_full_buffer) /* shouldn't happen */
444 ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
446 /* Allocate the workspace.
447 * ngroups is the number of row groups we need.
449 if (cinfo->upsample->need_context_rows) {
450 if (cinfo->_min_DCT_scaled_size < 2) /* unsupported, see comments above */
451 ERREXIT(cinfo, JERR_NOTIMPL);
452 alloc_funny_pointers(cinfo); /* Alloc space for xbuffer[] lists */
453 ngroups = cinfo->_min_DCT_scaled_size + 2;
455 ngroups = cinfo->_min_DCT_scaled_size;
458 for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
460 rgroup = (compptr->v_samp_factor * compptr->_DCT_scaled_size) /
461 cinfo->_min_DCT_scaled_size; /* height of a row group of component */
462 main_ptr->buffer[ci] = (_JSAMPARRAY)(*cinfo->mem->alloc_sarray)
463 ((j_common_ptr)cinfo, JPOOL_IMAGE,
464 compptr->width_in_blocks * compptr->_DCT_scaled_size,
465 (JDIMENSION)(rgroup * ngroups));
469 #endif /* BITS_IN_JSAMPLE != 16 || defined(D_LOSSLESS_SUPPORTED) */