PageRenderTime 58ms CodeModel.GetById 26ms RepoModel.GetById 1ms app.codeStats 0ms

/src/FreeImage/Source/LibJPEG/structure.txt

https://bitbucket.org/cabalistic/ogredeps/
Plain Text | 941 lines | 768 code | 173 blank | 0 comment | 0 complexity | dd2f5307b06a411b1b0fadcd675d612d MD5 | raw file
Possible License(s): LGPL-3.0, BSD-3-Clause, CPL-1.0, Unlicense, GPL-2.0, GPL-3.0, LGPL-2.0, MPL-2.0-no-copyleft-exception, BSD-2-Clause, LGPL-2.1
  1. IJG JPEG LIBRARY: SYSTEM ARCHITECTURE
  2. Copyright (C) 1991-2012, Thomas G. Lane, Guido Vollbeding.
  3. This file is part of the Independent JPEG Group's software.
  4. For conditions of distribution and use, see the accompanying README file.
  5. This file provides an overview of the architecture of the IJG JPEG software;
  6. that is, the functions of the various modules in the system and the interfaces
  7. between modules. For more precise details about any data structure or calling
  8. convention, see the include files and comments in the source code.
  9. We assume that the reader is already somewhat familiar with the JPEG standard.
  10. The README file includes references for learning about JPEG. The file
  11. libjpeg.txt describes the library from the viewpoint of an application
  12. programmer using the library; it's best to read that file before this one.
  13. Also, the file coderules.txt describes the coding style conventions we use.
  14. In this document, JPEG-specific terminology follows the JPEG standard:
  15. A "component" means a color channel, e.g., Red or Luminance.
  16. A "sample" is a single component value (i.e., one number in the image data).
  17. A "coefficient" is a frequency coefficient (a DCT transform output number).
  18. A "block" is an array of samples or coefficients.
  19. An "MCU" (minimum coded unit) is an interleaved set of blocks of size
  20. determined by the sampling factors, or a single block in a
  21. noninterleaved scan.
  22. We do not use the terms "pixel" and "sample" interchangeably. When we say
  23. pixel, we mean an element of the full-size image, while a sample is an element
  24. of the downsampled image. Thus the number of samples may vary across
  25. components while the number of pixels does not. (This terminology is not used
  26. rigorously throughout the code, but it is used in places where confusion would
  27. otherwise result.)
  28. *** System features ***
  29. The IJG distribution contains two parts:
  30. * A subroutine library for JPEG compression and decompression.
  31. * cjpeg/djpeg, two sample applications that use the library to transform
  32. JFIF JPEG files to and from several other image formats.
  33. cjpeg/djpeg are of no great intellectual complexity: they merely add a simple
  34. command-line user interface and I/O routines for several uncompressed image
  35. formats. This document concentrates on the library itself.
  36. We desire the library to be capable of supporting all JPEG baseline, extended
  37. sequential, and progressive DCT processes. The library does not support the
  38. hierarchical or lossless processes defined in the standard.
  39. Within these limits, any set of compression parameters allowed by the JPEG
  40. spec should be readable for decompression. (We can be more restrictive about
  41. what formats we can generate.) Although the system design allows for all
  42. parameter values, some uncommon settings are not yet implemented and may
  43. never be; nonintegral sampling ratios are the prime example. Furthermore,
  44. we treat 8-bit vs. 12-bit data precision as a compile-time switch, not a
  45. run-time option, because most machines can store 8-bit pixels much more
  46. compactly than 12-bit.
  47. By itself, the library handles only interchange JPEG datastreams --- in
  48. particular the widely used JFIF file format. The library can be used by
  49. surrounding code to process interchange or abbreviated JPEG datastreams that
  50. are embedded in more complex file formats. (For example, libtiff uses this
  51. library to implement JPEG compression within the TIFF file format.)
  52. The library includes a substantial amount of code that is not covered by the
  53. JPEG standard but is necessary for typical applications of JPEG. These
  54. functions preprocess the image before JPEG compression or postprocess it after
  55. decompression. They include colorspace conversion, downsampling/upsampling,
  56. and color quantization. This code can be omitted if not needed.
  57. A wide range of quality vs. speed tradeoffs are possible in JPEG processing,
  58. and even more so in decompression postprocessing. The decompression library
  59. provides multiple implementations that cover most of the useful tradeoffs,
  60. ranging from very-high-quality down to fast-preview operation. On the
  61. compression side we have generally not provided low-quality choices, since
  62. compression is normally less time-critical. It should be understood that the
  63. low-quality modes may not meet the JPEG standard's accuracy requirements;
  64. nonetheless, they are useful for viewers.
  65. *** Portability issues ***
  66. Portability is an essential requirement for the library. The key portability
  67. issues that show up at the level of system architecture are:
  68. 1. Memory usage. We want the code to be able to run on PC-class machines
  69. with limited memory. Images should therefore be processed sequentially (in
  70. strips), to avoid holding the whole image in memory at once. Where a
  71. full-image buffer is necessary, we should be able to use either virtual memory
  72. or temporary files.
  73. 2. Near/far pointer distinction. To run efficiently on 80x86 machines, the
  74. code should distinguish "small" objects (kept in near data space) from
  75. "large" ones (kept in far data space). This is an annoying restriction, but
  76. fortunately it does not impact code quality for less brain-damaged machines,
  77. and the source code clutter turns out to be minimal with sufficient use of
  78. pointer typedefs.
  79. 3. Data precision. We assume that "char" is at least 8 bits, "short" and
  80. "int" at least 16, "long" at least 32. The code will work fine with larger
  81. data sizes, although memory may be used inefficiently in some cases. However,
  82. the JPEG compressed datastream must ultimately appear on external storage as a
  83. sequence of 8-bit bytes if it is to conform to the standard. This may pose a
  84. problem on machines where char is wider than 8 bits. The library represents
  85. compressed data as an array of values of typedef JOCTET. If no data type
  86. exactly 8 bits wide is available, custom data source and data destination
  87. modules must be written to unpack and pack the chosen JOCTET datatype into
  88. 8-bit external representation.
  89. *** System overview ***
  90. The compressor and decompressor are each divided into two main sections:
  91. the JPEG compressor or decompressor proper, and the preprocessing or
  92. postprocessing functions. The interface between these two sections is the
  93. image data that the official JPEG spec regards as its input or output: this
  94. data is in the colorspace to be used for compression, and it is downsampled
  95. to the sampling factors to be used. The preprocessing and postprocessing
  96. steps are responsible for converting a normal image representation to or from
  97. this form. (Those few applications that want to deal with YCbCr downsampled
  98. data can skip the preprocessing or postprocessing step.)
  99. Looking more closely, the compressor library contains the following main
  100. elements:
  101. Preprocessing:
  102. * Color space conversion (e.g., RGB to YCbCr).
  103. * Edge expansion and downsampling. Optionally, this step can do simple
  104. smoothing --- this is often helpful for low-quality source data.
  105. JPEG proper:
  106. * MCU assembly, DCT, quantization.
  107. * Entropy coding (sequential or progressive, Huffman or arithmetic).
  108. In addition to these modules we need overall control, marker generation,
  109. and support code (memory management & error handling). There is also a
  110. module responsible for physically writing the output data --- typically
  111. this is just an interface to fwrite(), but some applications may need to
  112. do something else with the data.
  113. The decompressor library contains the following main elements:
  114. JPEG proper:
  115. * Entropy decoding (sequential or progressive, Huffman or arithmetic).
  116. * Dequantization, inverse DCT, MCU disassembly.
  117. Postprocessing:
  118. * Upsampling. Optionally, this step may be able to do more general
  119. rescaling of the image.
  120. * Color space conversion (e.g., YCbCr to RGB). This step may also
  121. provide gamma adjustment [ currently it does not ].
  122. * Optional color quantization (e.g., reduction to 256 colors).
  123. * Optional color precision reduction (e.g., 24-bit to 15-bit color).
  124. [This feature is not currently implemented.]
  125. We also need overall control, marker parsing, and a data source module.
  126. The support code (memory management & error handling) can be shared with
  127. the compression half of the library.
  128. There may be several implementations of each of these elements, particularly
  129. in the decompressor, where a wide range of speed/quality tradeoffs is very
  130. useful. It must be understood that some of the best speedups involve
  131. merging adjacent steps in the pipeline. For example, upsampling, color space
  132. conversion, and color quantization might all be done at once when using a
  133. low-quality ordered-dither technique. The system architecture is designed to
  134. allow such merging where appropriate.
  135. Note: it is convenient to regard edge expansion (padding to block boundaries)
  136. as a preprocessing/postprocessing function, even though the JPEG spec includes
  137. it in compression/decompression. We do this because downsampling/upsampling
  138. can be simplified a little if they work on padded data: it's not necessary to
  139. have special cases at the right and bottom edges. Therefore the interface
  140. buffer is always an integral number of blocks wide and high, and we expect
  141. compression preprocessing to pad the source data properly. Padding will occur
  142. only to the next block (N-sample) boundary. In an interleaved-scan situation,
  143. additional dummy blocks may be used to fill out MCUs, but the MCU assembly and
  144. disassembly logic will create or discard these blocks internally. (This is
  145. advantageous for speed reasons, since we avoid DCTing the dummy blocks.
  146. It also permits a small reduction in file size, because the compressor can
  147. choose dummy block contents so as to minimize their size in compressed form.
  148. Finally, it makes the interface buffer specification independent of whether
  149. the file is actually interleaved or not.) Applications that wish to deal
  150. directly with the downsampled data must provide similar buffering and padding
  151. for odd-sized images.
  152. *** Poor man's object-oriented programming ***
  153. It should be clear by now that we have a lot of quasi-independent processing
  154. steps, many of which have several possible behaviors. To avoid cluttering the
  155. code with lots of switch statements, we use a simple form of object-style
  156. programming to separate out the different possibilities.
  157. For example, two different color quantization algorithms could be implemented
  158. as two separate modules that present the same external interface; at runtime,
  159. the calling code will access the proper module indirectly through an "object".
  160. We can get the limited features we need while staying within portable C.
  161. The basic tool is a function pointer. An "object" is just a struct
  162. containing one or more function pointer fields, each of which corresponds to
  163. a method name in real object-oriented languages. During initialization we
  164. fill in the function pointers with references to whichever module we have
  165. determined we need to use in this run. Then invocation of the module is done
  166. by indirecting through a function pointer; on most machines this is no more
  167. expensive than a switch statement, which would be the only other way of
  168. making the required run-time choice. The really significant benefit, of
  169. course, is keeping the source code clean and well structured.
  170. We can also arrange to have private storage that varies between different
  171. implementations of the same kind of object. We do this by making all the
  172. module-specific object structs be separately allocated entities, which will
  173. be accessed via pointers in the master compression or decompression struct.
  174. The "public" fields or methods for a given kind of object are specified by
  175. a commonly known struct. But a module's initialization code can allocate
  176. a larger struct that contains the common struct as its first member, plus
  177. additional private fields. With appropriate pointer casting, the module's
  178. internal functions can access these private fields. (For a simple example,
  179. see jdatadst.c, which implements the external interface specified by struct
  180. jpeg_destination_mgr, but adds extra fields.)
  181. (Of course this would all be a lot easier if we were using C++, but we are
  182. not yet prepared to assume that everyone has a C++ compiler.)
  183. An important benefit of this scheme is that it is easy to provide multiple
  184. versions of any method, each tuned to a particular case. While a lot of
  185. precalculation might be done to select an optimal implementation of a method,
  186. the cost per invocation is constant. For example, the upsampling step might
  187. have a "generic" method, plus one or more "hardwired" methods for the most
  188. popular sampling factors; the hardwired methods would be faster because they'd
  189. use straight-line code instead of for-loops. The cost to determine which
  190. method to use is paid only once, at startup, and the selection criteria are
  191. hidden from the callers of the method.
  192. This plan differs a little bit from usual object-oriented structures, in that
  193. only one instance of each object class will exist during execution. The
  194. reason for having the class structure is that on different runs we may create
  195. different instances (choose to execute different modules). You can think of
  196. the term "method" as denoting the common interface presented by a particular
  197. set of interchangeable functions, and "object" as denoting a group of related
  198. methods, or the total shared interface behavior of a group of modules.
  199. *** Overall control structure ***
  200. We previously mentioned the need for overall control logic in the compression
  201. and decompression libraries. In IJG implementations prior to v5, overall
  202. control was mostly provided by "pipeline control" modules, which proved to be
  203. large, unwieldy, and hard to understand. To improve the situation, the
  204. control logic has been subdivided into multiple modules. The control modules
  205. consist of:
  206. 1. Master control for module selection and initialization. This has two
  207. responsibilities:
  208. 1A. Startup initialization at the beginning of image processing.
  209. The individual processing modules to be used in this run are selected
  210. and given initialization calls.
  211. 1B. Per-pass control. This determines how many passes will be performed
  212. and calls each active processing module to configure itself
  213. appropriately at the beginning of each pass. End-of-pass processing,
  214. where necessary, is also invoked from the master control module.
  215. Method selection is partially distributed, in that a particular processing
  216. module may contain several possible implementations of a particular method,
  217. which it will select among when given its initialization call. The master
  218. control code need only be concerned with decisions that affect more than
  219. one module.
  220. 2. Data buffering control. A separate control module exists for each
  221. inter-processing-step data buffer. This module is responsible for
  222. invoking the processing steps that write or read that data buffer.
  223. Each buffer controller sees the world as follows:
  224. input data => processing step A => buffer => processing step B => output data
  225. | | |
  226. ------------------ controller ------------------
  227. The controller knows the dataflow requirements of steps A and B: how much data
  228. they want to accept in one chunk and how much they output in one chunk. Its
  229. function is to manage its buffer and call A and B at the proper times.
  230. A data buffer control module may itself be viewed as a processing step by a
  231. higher-level control module; thus the control modules form a binary tree with
  232. elementary processing steps at the leaves of the tree.
  233. The control modules are objects. A considerable amount of flexibility can
  234. be had by replacing implementations of a control module. For example:
  235. * Merging of adjacent steps in the pipeline is done by replacing a control
  236. module and its pair of processing-step modules with a single processing-
  237. step module. (Hence the possible merges are determined by the tree of
  238. control modules.)
  239. * In some processing modes, a given interstep buffer need only be a "strip"
  240. buffer large enough to accommodate the desired data chunk sizes. In other
  241. modes, a full-image buffer is needed and several passes are required.
  242. The control module determines which kind of buffer is used and manipulates
  243. virtual array buffers as needed. One or both processing steps may be
  244. unaware of the multi-pass behavior.
  245. In theory, we might be able to make all of the data buffer controllers
  246. interchangeable and provide just one set of implementations for all. In
  247. practice, each one contains considerable special-case processing for its
  248. particular job. The buffer controller concept should be regarded as an
  249. overall system structuring principle, not as a complete description of the
  250. task performed by any one controller.
  251. *** Compression object structure ***
  252. Here is a sketch of the logical structure of the JPEG compression library:
  253. |-- Colorspace conversion
  254. |-- Preprocessing controller --|
  255. | |-- Downsampling
  256. Main controller --|
  257. | |-- Forward DCT, quantize
  258. |-- Coefficient controller --|
  259. |-- Entropy encoding
  260. This sketch also describes the flow of control (subroutine calls) during
  261. typical image data processing. Each of the components shown in the diagram is
  262. an "object" which may have several different implementations available. One
  263. or more source code files contain the actual implementation(s) of each object.
  264. The objects shown above are:
  265. * Main controller: buffer controller for the subsampled-data buffer, which
  266. holds the preprocessed input data. This controller invokes preprocessing to
  267. fill the subsampled-data buffer, and JPEG compression to empty it. There is
  268. usually no need for a full-image buffer here; a strip buffer is adequate.
  269. * Preprocessing controller: buffer controller for the downsampling input data
  270. buffer, which lies between colorspace conversion and downsampling. Note
  271. that a unified conversion/downsampling module would probably replace this
  272. controller entirely.
  273. * Colorspace conversion: converts application image data into the desired
  274. JPEG color space; also changes the data from pixel-interleaved layout to
  275. separate component planes. Processes one pixel row at a time.
  276. * Downsampling: performs reduction of chroma components as required.
  277. Optionally may perform pixel-level smoothing as well. Processes a "row
  278. group" at a time, where a row group is defined as Vmax pixel rows of each
  279. component before downsampling, and Vk sample rows afterwards (remember Vk
  280. differs across components). Some downsampling or smoothing algorithms may
  281. require context rows above and below the current row group; the
  282. preprocessing controller is responsible for supplying these rows via proper
  283. buffering. The downsampler is responsible for edge expansion at the right
  284. edge (i.e., extending each sample row to a multiple of N samples); but the
  285. preprocessing controller is responsible for vertical edge expansion (i.e.,
  286. duplicating the bottom sample row as needed to make a multiple of N rows).
  287. * Coefficient controller: buffer controller for the DCT-coefficient data.
  288. This controller handles MCU assembly, including insertion of dummy DCT
  289. blocks when needed at the right or bottom edge. When performing
  290. Huffman-code optimization or emitting a multiscan JPEG file, this
  291. controller is responsible for buffering the full image. The equivalent of
  292. one fully interleaved MCU row of subsampled data is processed per call,
  293. even when the JPEG file is noninterleaved.
  294. * Forward DCT and quantization: Perform DCT, quantize, and emit coefficients.
  295. Works on one or more DCT blocks at a time. (Note: the coefficients are now
  296. emitted in normal array order, which the entropy encoder is expected to
  297. convert to zigzag order as necessary. Prior versions of the IJG code did
  298. the conversion to zigzag order within the quantization step.)
  299. * Entropy encoding: Perform Huffman or arithmetic entropy coding and emit the
  300. coded data to the data destination module. Works on one MCU per call.
  301. For progressive JPEG, the same DCT blocks are fed to the entropy coder
  302. during each pass, and the coder must emit the appropriate subset of
  303. coefficients.
  304. In addition to the above objects, the compression library includes these
  305. objects:
  306. * Master control: determines the number of passes required, controls overall
  307. and per-pass initialization of the other modules.
  308. * Marker writing: generates JPEG markers (except for RSTn, which is emitted
  309. by the entropy encoder when needed).
  310. * Data destination manager: writes the output JPEG datastream to its final
  311. destination (e.g., a file). The destination manager supplied with the
  312. library knows how to write to a stdio stream or to a memory buffer;
  313. for other behaviors, the surrounding application may provide its own
  314. destination manager.
  315. * Memory manager: allocates and releases memory, controls virtual arrays
  316. (with backing store management, where required).
  317. * Error handler: performs formatting and output of error and trace messages;
  318. determines handling of nonfatal errors. The surrounding application may
  319. override some or all of this object's methods to change error handling.
  320. * Progress monitor: supports output of "percent-done" progress reports.
  321. This object represents an optional callback to the surrounding application:
  322. if wanted, it must be supplied by the application.
  323. The error handler, destination manager, and progress monitor objects are
  324. defined as separate objects in order to simplify application-specific
  325. customization of the JPEG library. A surrounding application may override
  326. individual methods or supply its own all-new implementation of one of these
  327. objects. The object interfaces for these objects are therefore treated as
  328. part of the application interface of the library, whereas the other objects
  329. are internal to the library.
  330. The error handler and memory manager are shared by JPEG compression and
  331. decompression; the progress monitor, if used, may be shared as well.
  332. *** Decompression object structure ***
  333. Here is a sketch of the logical structure of the JPEG decompression library:
  334. |-- Entropy decoding
  335. |-- Coefficient controller --|
  336. | |-- Dequantize, Inverse DCT
  337. Main controller --|
  338. | |-- Upsampling
  339. |-- Postprocessing controller --| |-- Colorspace conversion
  340. |-- Color quantization
  341. |-- Color precision reduction
  342. As before, this diagram also represents typical control flow. The objects
  343. shown are:
  344. * Main controller: buffer controller for the subsampled-data buffer, which
  345. holds the output of JPEG decompression proper. This controller's primary
  346. task is to feed the postprocessing procedure. Some upsampling algorithms
  347. may require context rows above and below the current row group; when this
  348. is true, the main controller is responsible for managing its buffer so as
  349. to make context rows available. In the current design, the main buffer is
  350. always a strip buffer; a full-image buffer is never required.
  351. * Coefficient controller: buffer controller for the DCT-coefficient data.
  352. This controller handles MCU disassembly, including deletion of any dummy
  353. DCT blocks at the right or bottom edge. When reading a multiscan JPEG
  354. file, this controller is responsible for buffering the full image.
  355. (Buffering DCT coefficients, rather than samples, is necessary to support
  356. progressive JPEG.) The equivalent of one fully interleaved MCU row of
  357. subsampled data is processed per call, even when the source JPEG file is
  358. noninterleaved.
  359. * Entropy decoding: Read coded data from the data source module and perform
  360. Huffman or arithmetic entropy decoding. Works on one MCU per call.
  361. For progressive JPEG decoding, the coefficient controller supplies the prior
  362. coefficients of each MCU (initially all zeroes), which the entropy decoder
  363. modifies in each scan.
  364. * Dequantization and inverse DCT: like it says. Note that the coefficients
  365. buffered by the coefficient controller have NOT been dequantized; we
  366. merge dequantization and inverse DCT into a single step for speed reasons.
  367. When scaled-down output is asked for, simplified DCT algorithms may be used
  368. that need fewer coefficients and emit fewer samples per DCT block, not the
  369. full 8x8. Works on one DCT block at a time.
  370. * Postprocessing controller: buffer controller for the color quantization
  371. input buffer, when quantization is in use. (Without quantization, this
  372. controller just calls the upsampler.) For two-pass quantization, this
  373. controller is responsible for buffering the full-image data.
  374. * Upsampling: restores chroma components to full size. (May support more
  375. general output rescaling, too. Note that if undersized DCT outputs have
  376. been emitted by the DCT module, this module must adjust so that properly
  377. sized outputs are created.) Works on one row group at a time. This module
  378. also calls the color conversion module, so its top level is effectively a
  379. buffer controller for the upsampling->color conversion buffer. However, in
  380. all but the highest-quality operating modes, upsampling and color
  381. conversion are likely to be merged into a single step.
  382. * Colorspace conversion: convert from JPEG color space to output color space,
  383. and change data layout from separate component planes to pixel-interleaved.
  384. Works on one pixel row at a time.
  385. * Color quantization: reduce the data to colormapped form, using either an
  386. externally specified colormap or an internally generated one. This module
  387. is not used for full-color output. Works on one pixel row at a time; may
  388. require two passes to generate a color map. Note that the output will
  389. always be a single component representing colormap indexes. In the current
  390. design, the output values are JSAMPLEs, so an 8-bit compilation cannot
  391. quantize to more than 256 colors. This is unlikely to be a problem in
  392. practice.
  393. * Color reduction: this module handles color precision reduction, e.g.,
  394. generating 15-bit color (5 bits/primary) from JPEG's 24-bit output.
  395. Not quite clear yet how this should be handled... should we merge it with
  396. colorspace conversion???
  397. Note that some high-speed operating modes might condense the entire
  398. postprocessing sequence to a single module (upsample, color convert, and
  399. quantize in one step).
  400. In addition to the above objects, the decompression library includes these
  401. objects:
  402. * Master control: determines the number of passes required, controls overall
  403. and per-pass initialization of the other modules. This is subdivided into
  404. input and output control: jdinput.c controls only input-side processing,
  405. while jdmaster.c handles overall initialization and output-side control.
  406. * Marker reading: decodes JPEG markers (except for RSTn).
  407. * Data source manager: supplies the input JPEG datastream. The source
  408. manager supplied with the library knows how to read from a stdio stream
  409. or from a memory buffer; for other behaviors, the surrounding application
  410. may provide its own source manager.
  411. * Memory manager: same as for compression library.
  412. * Error handler: same as for compression library.
  413. * Progress monitor: same as for compression library.
  414. As with compression, the data source manager, error handler, and progress
  415. monitor are candidates for replacement by a surrounding application.
  416. *** Decompression input and output separation ***
  417. To support efficient incremental display of progressive JPEG files, the
  418. decompressor is divided into two sections that can run independently:
  419. 1. Data input includes marker parsing, entropy decoding, and input into the
  420. coefficient controller's DCT coefficient buffer. Note that this
  421. processing is relatively cheap and fast.
  422. 2. Data output reads from the DCT coefficient buffer and performs the IDCT
  423. and all postprocessing steps.
  424. For a progressive JPEG file, the data input processing is allowed to get
  425. arbitrarily far ahead of the data output processing. (This occurs only
  426. if the application calls jpeg_consume_input(); otherwise input and output
  427. run in lockstep, since the input section is called only when the output
  428. section needs more data.) In this way the application can avoid making
  429. extra display passes when data is arriving faster than the display pass
  430. can run. Furthermore, it is possible to abort an output pass without
  431. losing anything, since the coefficient buffer is read-only as far as the
  432. output section is concerned. See libjpeg.txt for more detail.
  433. A full-image coefficient array is only created if the JPEG file has multiple
  434. scans (or if the application specifies buffered-image mode anyway). When
  435. reading a single-scan file, the coefficient controller normally creates only
  436. a one-MCU buffer, so input and output processing must run in lockstep in this
  437. case. jpeg_consume_input() is effectively a no-op in this situation.
  438. The main impact of dividing the decompressor in this fashion is that we must
  439. be very careful with shared variables in the cinfo data structure. Each
  440. variable that can change during the course of decompression must be
  441. classified as belonging to data input or data output, and each section must
  442. look only at its own variables. For example, the data output section may not
  443. depend on any of the variables that describe the current scan in the JPEG
  444. file, because these may change as the data input section advances into a new
  445. scan.
  446. The progress monitor is (somewhat arbitrarily) defined to treat input of the
  447. file as one pass when buffered-image mode is not used, and to ignore data
  448. input work completely when buffered-image mode is used. Note that the
  449. library has no reliable way to predict the number of passes when dealing
  450. with a progressive JPEG file, nor can it predict the number of output passes
  451. in buffered-image mode. So the work estimate is inherently bogus anyway.
  452. No comparable division is currently made in the compression library, because
  453. there isn't any real need for it.
  454. *** Data formats ***
  455. Arrays of pixel sample values use the following data structure:
  456. typedef something JSAMPLE; a pixel component value, 0..MAXJSAMPLE
  457. typedef JSAMPLE *JSAMPROW; ptr to a row of samples
  458. typedef JSAMPROW *JSAMPARRAY; ptr to a list of rows
  459. typedef JSAMPARRAY *JSAMPIMAGE; ptr to a list of color-component arrays
  460. The basic element type JSAMPLE will typically be one of unsigned char,
  461. (signed) char, or short. Short will be used if samples wider than 8 bits are
  462. to be supported (this is a compile-time option). Otherwise, unsigned char is
  463. used if possible. If the compiler only supports signed chars, then it is
  464. necessary to mask off the value when reading. Thus, all reads of JSAMPLE
  465. values must be coded as "GETJSAMPLE(value)", where the macro will be defined
  466. as "((value) & 0xFF)" on signed-char machines and "((int) (value))" elsewhere.
  467. With these conventions, JSAMPLE values can be assumed to be >= 0. This helps
  468. simplify correct rounding during downsampling, etc. The JPEG standard's
  469. specification that sample values run from -128..127 is accommodated by
  470. subtracting 128 from the sample value in the DCT step. Similarly, during
  471. decompression the output of the IDCT step will be immediately shifted back to
  472. 0..255. (NB: different values are required when 12-bit samples are in use.
  473. The code is written in terms of MAXJSAMPLE and CENTERJSAMPLE, which will be
  474. defined as 255 and 128 respectively in an 8-bit implementation, and as 4095
  475. and 2048 in a 12-bit implementation.)
  476. We use a pointer per row, rather than a two-dimensional JSAMPLE array. This
  477. choice costs only a small amount of memory and has several benefits:
  478. * Code using the data structure doesn't need to know the allocated width of
  479. the rows. This simplifies edge expansion/compression, since we can work
  480. in an array that's wider than the logical picture width.
  481. * Indexing doesn't require multiplication; this is a performance win on many
  482. machines.
  483. * Arrays with more than 64K total elements can be supported even on machines
  484. where malloc() cannot allocate chunks larger than 64K.
  485. * The rows forming a component array may be allocated at different times
  486. without extra copying. This trick allows some speedups in smoothing steps
  487. that need access to the previous and next rows.
  488. Note that each color component is stored in a separate array; we don't use the
  489. traditional layout in which the components of a pixel are stored together.
  490. This simplifies coding of modules that work on each component independently,
  491. because they don't need to know how many components there are. Furthermore,
  492. we can read or write each component to a temporary file independently, which
  493. is helpful when dealing with noninterleaved JPEG files.
  494. In general, a specific sample value is accessed by code such as
  495. GETJSAMPLE(image[colorcomponent][row][col])
  496. where col is measured from the image left edge, but row is measured from the
  497. first sample row currently in memory. Either of the first two indexings can
  498. be precomputed by copying the relevant pointer.
  499. Since most image-processing applications prefer to work on images in which
  500. the components of a pixel are stored together, the data passed to or from the
  501. surrounding application uses the traditional convention: a single pixel is
  502. represented by N consecutive JSAMPLE values, and an image row is an array of
  503. (# of color components)*(image width) JSAMPLEs. One or more rows of data can
  504. be represented by a pointer of type JSAMPARRAY in this scheme. This scheme is
  505. converted to component-wise storage inside the JPEG library. (Applications
  506. that want to skip JPEG preprocessing or postprocessing will have to contend
  507. with component-wise storage.)
  508. Arrays of DCT-coefficient values use the following data structure:
  509. typedef short JCOEF; a 16-bit signed integer
  510. typedef JCOEF JBLOCK[DCTSIZE2]; an 8x8 block of coefficients
  511. typedef JBLOCK *JBLOCKROW; ptr to one horizontal row of 8x8 blocks
  512. typedef JBLOCKROW *JBLOCKARRAY; ptr to a list of such rows
  513. typedef JBLOCKARRAY *JBLOCKIMAGE; ptr to a list of color component arrays
  514. The underlying type is at least a 16-bit signed integer; while "short" is big
  515. enough on all machines of interest, on some machines it is preferable to use
  516. "int" for speed reasons, despite the storage cost. Coefficients are grouped
  517. into 8x8 blocks (but we always use #defines DCTSIZE and DCTSIZE2 rather than
  518. "8" and "64").
  519. The contents of a coefficient block may be in either "natural" or zigzagged
  520. order, and may be true values or divided by the quantization coefficients,
  521. depending on where the block is in the processing pipeline. In the current
  522. library, coefficient blocks are kept in natural order everywhere; the entropy
  523. codecs zigzag or dezigzag the data as it is written or read. The blocks
  524. contain quantized coefficients everywhere outside the DCT/IDCT subsystems.
  525. (This latter decision may need to be revisited to support variable
  526. quantization a la JPEG Part 3.)
  527. Notice that the allocation unit is now a row of 8x8 coefficient blocks,
  528. corresponding to N rows of samples. Otherwise the structure is much the same
  529. as for samples, and for the same reasons.
  530. On machines where malloc() can't handle a request bigger than 64Kb, this data
  531. structure limits us to rows of less than 512 JBLOCKs, or a picture width of
  532. 4000+ pixels. This seems an acceptable restriction.
  533. On 80x86 machines, the bottom-level pointer types (JSAMPROW and JBLOCKROW)
  534. must be declared as "far" pointers, but the upper levels can be "near"
  535. (implying that the pointer lists are allocated in the DS segment).
  536. We use a #define symbol FAR, which expands to the "far" keyword when
  537. compiling on 80x86 machines and to nothing elsewhere.
  538. *** Suspendable processing ***
  539. In some applications it is desirable to use the JPEG library as an
  540. incremental, memory-to-memory filter. In this situation the data source or
  541. destination may be a limited-size buffer, and we can't rely on being able to
  542. empty or refill the buffer at arbitrary times. Instead the application would
  543. like to have control return from the library at buffer overflow/underrun, and
  544. then resume compression or decompression at a later time.
  545. This scenario is supported for simple cases. (For anything more complex, we
  546. recommend that the application "bite the bullet" and develop real multitasking
  547. capability.) The libjpeg.txt file goes into more detail about the usage and
  548. limitations of this capability; here we address the implications for library
  549. structure.
  550. The essence of the problem is that the entropy codec (coder or decoder) must
  551. be prepared to stop at arbitrary times. In turn, the controllers that call
  552. the entropy codec must be able to stop before having produced or consumed all
  553. the data that they normally would handle in one call. That part is reasonably
  554. straightforward: we make the controller call interfaces include "progress
  555. counters" which indicate the number of data chunks successfully processed, and
  556. we require callers to test the counter rather than just assume all of the data
  557. was processed.
  558. Rather than trying to restart at an arbitrary point, the current Huffman
  559. codecs are designed to restart at the beginning of the current MCU after a
  560. suspension due to buffer overflow/underrun. At the start of each call, the
  561. codec's internal state is loaded from permanent storage (in the JPEG object
  562. structures) into local variables. On successful completion of the MCU, the
  563. permanent state is updated. (This copying is not very expensive, and may even
  564. lead to *improved* performance if the local variables can be registerized.)
  565. If a suspension occurs, the codec simply returns without updating the state,
  566. thus effectively reverting to the start of the MCU. Note that this implies
  567. leaving some data unprocessed in the source/destination buffer (ie, the
  568. compressed partial MCU). The data source/destination module interfaces are
  569. specified so as to make this possible. This also implies that the data buffer
  570. must be large enough to hold a worst-case compressed MCU; a couple thousand
  571. bytes should be enough.
  572. In a successive-approximation AC refinement scan, the progressive Huffman
  573. decoder has to be able to undo assignments of newly nonzero coefficients if it
  574. suspends before the MCU is complete, since decoding requires distinguishing
  575. previously-zero and previously-nonzero coefficients. This is a bit tedious
  576. but probably won't have much effect on performance. Other variants of Huffman
  577. decoding need not worry about this, since they will just store the same values
  578. again if forced to repeat the MCU.
  579. This approach would probably not work for an arithmetic codec, since its
  580. modifiable state is quite large and couldn't be copied cheaply. Instead it
  581. would have to suspend and resume exactly at the point of the buffer end.
  582. The JPEG marker reader is designed to cope with suspension at an arbitrary
  583. point. It does so by backing up to the start of the marker parameter segment,
  584. so the data buffer must be big enough to hold the largest marker of interest.
  585. Again, a couple KB should be adequate. (A special "skip" convention is used
  586. to bypass COM and APPn markers, so these can be larger than the buffer size
  587. without causing problems; otherwise a 64K buffer would be needed in the worst
  588. case.)
  589. The JPEG marker writer currently does *not* cope with suspension.
  590. We feel that this is not necessary; it is much easier simply to require
  591. the application to ensure there is enough buffer space before starting. (An
  592. empty 2K buffer is more than sufficient for the header markers; and ensuring
  593. there are a dozen or two bytes available before calling jpeg_finish_compress()
  594. will suffice for the trailer.) This would not work for writing multi-scan
  595. JPEG files, but we simply do not intend to support that capability with
  596. suspension.
  597. *** Memory manager services ***
  598. The JPEG library's memory manager controls allocation and deallocation of
  599. memory, and it manages large "virtual" data arrays on machines where the
  600. operating system does not provide virtual memory. Note that the same
  601. memory manager serves both compression and decompression operations.
  602. In all cases, allocated objects are tied to a particular compression or
  603. decompression master record, and they will be released when that master
  604. record is destroyed.
  605. The memory manager does not provide explicit deallocation of objects.
  606. Instead, objects are created in "pools" of free storage, and a whole pool
  607. can be freed at once. This approach helps prevent storage-leak bugs, and
  608. it speeds up operations whenever malloc/free are slow (as they often are).
  609. The pools can be regarded as lifetime identifiers for objects. Two
  610. pools/lifetimes are defined:
  611. * JPOOL_PERMANENT lasts until master record is destroyed
  612. * JPOOL_IMAGE lasts until done with image (JPEG datastream)
  613. Permanent lifetime is used for parameters and tables that should be carried
  614. across from one datastream to another; this includes all application-visible
  615. parameters. Image lifetime is used for everything else. (A third lifetime,
  616. JPOOL_PASS = one processing pass, was originally planned. However it was
  617. dropped as not being worthwhile. The actual usage patterns are such that the
  618. peak memory usage would be about the same anyway; and having per-pass storage
  619. substantially complicates the virtual memory allocation rules --- see below.)
  620. The memory manager deals with three kinds of object:
  621. 1. "Small" objects. Typically these require no more than 10K-20K total.
  622. 2. "Large" objects. These may require tens to hundreds of K depending on
  623. image size. Semantically they behave the same as small objects, but we
  624. distinguish them for two reasons:
  625. * On MS-DOS machines, large objects are referenced by FAR pointers,
  626. small objects by NEAR pointers.
  627. * Pool allocation heuristics may differ for large and small objects.
  628. Note that individual "large" objects cannot exceed the size allowed by
  629. type size_t, which may be 64K or less on some machines.
  630. 3. "Virtual" objects. These are large 2-D arrays of JSAMPLEs or JBLOCKs
  631. (typically large enough for the entire image being processed). The
  632. memory manager provides stripwise access to these arrays. On machines
  633. without virtual memory, the rest of the array may be swapped out to a
  634. temporary file.
  635. (Note: JSAMPARRAY and JBLOCKARRAY data structures are a combination of large
  636. objects for the data proper and small objects for the row pointers. For
  637. convenience and speed, the memory manager provides single routines to create
  638. these structures. Similarly, virtual arrays include a small control block
  639. and a JSAMPARRAY or JBLOCKARRAY working buffer, all created with one call.)
  640. In the present implementation, virtual arrays are only permitted to have image
  641. lifespan. (Permanent lifespan would not be reasonable, and pass lifespan is
  642. not very useful since a virtual array's raison d'etre is to store data for
  643. multiple passes through the image.) We also expect that only "small" objects
  644. will be given permanent lifespan, though this restriction is not required by
  645. the memory manager.
  646. In a non-virtual-memory machine, some performance benefit can be gained by
  647. making the in-memory buffers for virtual arrays be as large as possible.
  648. (For small images, the buffers might fit entirely in memory, so blind
  649. swapping would be very wasteful.) The memory manager will adjust the height
  650. of the buffers to fit within a prespecified maximum memory usage. In order
  651. to do this in a reasonably optimal fashion, the manager needs to allocate all
  652. of the virtual arrays at once. Therefore, there isn't a one-step allocation
  653. routine for virtual arrays; instead, there is a "request" routine that simply
  654. allocates the control block, and a "realize" routine (called just once) that
  655. determines space allocation and creates all of the actual buffers. The
  656. realize routine must allow for space occupied by non-virtual large objects.
  657. (We don't bother to factor in the space needed for small objects, on the
  658. grounds that it isn't worth the trouble.)
  659. To support all this, we establish the following protocol for doing business
  660. with the memory manager:
  661. 1. Modules must request virtual arrays (which may have only image lifespan)
  662. during the initial setup phase, i.e., in their jinit_xxx routines.
  663. 2. All "large" objects (including JSAMPARRAYs and JBLOCKARRAYs) must also be
  664. allocated during initial setup.
  665. 3. realize_virt_arrays will be called at the completion of initial setup.
  666. The above conventions ensure that sufficient information is available
  667. for it to choose a good size for virtual array buffers.
  668. Small objects of any lifespan may be allocated at any time. We expect that
  669. the total space used for small objects will be small enough to be negligible
  670. in the realize_virt_arrays computation.
  671. In a virtual-memory machine, we simply pretend that the available space is
  672. infinite, thus causing realize_virt_arrays to decide that it can allocate all
  673. the virtual arrays as full-size in-memory buffers. The overhead of the
  674. virtual-array access protocol is very small when no swapping occurs.
  675. A virtual array can be specified to be "pre-zeroed"; when this flag is set,
  676. never-yet-written sections of the array are set to zero before being made
  677. available to the caller. If this flag is not set, never-written sections
  678. of the array contain garbage. (This feature exists primarily because the
  679. equivalent logic would otherwise be needed in jdcoefct.c for progressive
  680. JPEG mode; we may as well make it available for possible other uses.)
  681. The first write pass on a virtual array is required to occur in top-to-bottom
  682. order; read passes, as well as any write passes after the first one, may
  683. access the array in any order. This restriction exists partly to simplify
  684. the virtual array control logic, and partly because some file systems may not
  685. support seeking beyond the current end-of-file in a temporary file. The main
  686. implication of this restriction is that rearrangement of rows (such as
  687. converting top-to-bottom data order to bottom-to-top) must be handled while
  688. reading data out of the virtual array, not while putting it in.
  689. *** Memory manager internal structure ***
  690. To isolate system dependencies as much as possible, we have broken the
  691. memory manager into two parts. There is a reasonably system-independent
  692. "front end" (jmemmgr.c) and a "back end" that contains only the code
  693. likely to change across systems. All of the memory management methods
  694. outlined above are implemented by the front end. The back end provides
  695. the following routines for use by the front end (none of these routines
  696. are known to the rest of the JPEG code):
  697. jpeg_mem_init, jpeg_mem_term system-dependent initialization/shutdown
  698. jpeg_get_small, jpeg_free_small interface to malloc and free library routines
  699. (or their equivalents)
  700. jpeg_get_large, jpeg_free_large interface to FAR malloc/free in MSDOS machines;
  701. else usually the same as
  702. jpeg_get_small/jpeg_free_small
  703. jpeg_mem_available estimate available memory
  704. jpeg_open_backing_store create a backing-store object
  705. read_backing_store, manipulate a backing-store object
  706. write_backing_store,
  707. close_backing_store
  708. On some systems there will be more than one type of backing-store object
  709. (specifically, in MS-DOS a backing store file might be an area of extended
  710. memory as well as a disk file). jpeg_open_backing_store is responsible for
  711. choosing how to implement a given object. The read/write/close routines
  712. are method pointers in the structure that describes a given object; this
  713. lets them be different for different object types.
  714. It may be necessary to ensure that backing store objects are explicitly
  715. released upon abnormal program termination. For example, MS-DOS won't free
  716. extended memory by itself. To support this, we will expect the main program
  717. or surrounding application to arrange to call self_destruct (typically via
  718. jpeg_destroy) upon abnormal termination. This may require a SIGINT signal
  719. handler or equivalent. We don't want to have the back end module install its
  720. own signal handler, because that would pre-empt the surrounding application's
  721. ability to control signal handling.
  722. The IJG distribution includes several memory manager back end implementations.
  723. Usually the same back end should be suitable for all applications on a given
  724. system, but it is possible for an application to supply its own back end at
  725. need.
  726. *** Implications of DNL marker ***
  727. Some JPEG files may use a DNL marker to postpone definition of the image
  728. height (this would be useful for a fax-like scanner's output, for instance).
  729. In these files the SOF marker claims the image height is 0, and you only
  730. find out the true image height at the end of the first scan.
  731. We could read these files as follows:
  732. 1. Upon seeing zero image height, replace it by 65535 (the maximum allowed).
  733. 2. When the DNL is found, update the image height in the global image
  734. descriptor.
  735. This implies that control modules must avoid making copies of the image
  736. height, and must re-test for termination after each MCU row. This would
  737. be easy enough to do.
  738. In cases where image-size data structures are allocated, this approach will
  739. result in very inefficient use of virtual memory or much-larger-than-necessary
  740. temporary files. This seems acceptable for something that probably won't be a
  741. mainstream usage. People might have to forgo use of memory-hogging options
  742. (such as two-pass color quantization or noninterleaved JPEG files) if they
  743. want efficient conversion of such files. (One could improve efficiency by
  744. demanding a user-supplied upper bound for the height, less than 65536; in most
  745. cases it could be much less.)
  746. The standard also permits the SOF marker to overestimate the image height,
  747. with a DNL to give the true, smaller height at the end of the first scan.
  748. This would solve the space problems if the overestimate wasn't too great.
  749. However, it implies that you don't even know whether DNL will be used.
  750. This leads to a couple of very serious objections:
  751. 1. Testing for a DNL marker must occur in the inner loop of the decompressor's
  752. Huffman decoder; this implies a speed penalty whether the feature is used
  753. or not.
  754. 2. There is no way to hide the last-minute change in image height from an
  755. application using the decoder. Thus *every* application using the IJG
  756. library would suffer a complexity penalty whether it cared about DNL or
  757. not.
  758. We currently do not support DNL because of these problems.
  759. A different approach is to insist that DNL-using files be preprocessed by a
  760. separate program that reads ahead to the DNL, then goes back and fixes the SOF
  761. marker. This is a much simpler solution and is probably far more efficient.
  762. Even if one wants piped input, buffering the first scan of the JPEG file needs
  763. a lot smaller temp file than is implied by the maximum-height method. For
  764. this approach we'd simply treat DNL as a no-op in the decompressor (at most,
  765. check that it matches the SOF image height).
  766. We will not worry about making the compressor capable of outputting DNL.
  767. Something similar to the first scheme above could be applied if anyone ever
  768. wants to make that work.