blob: 779233a8fc37ef2ff22f02d952a24d9c4b663c09 [file] [log] [blame]
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +00001IJG JPEG LIBRARY: SYSTEM ARCHITECTURE
2
Guido Vollbeding5996a252009-06-27 00:00:00 +00003Copyright (C) 1991-2009, Thomas G. Lane, Guido Vollbeding.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +00004This file is part of the Independent JPEG Group's software.
5For conditions of distribution and use, see the accompanying README file.
6
7
8This file provides an overview of the architecture of the IJG JPEG software;
9that is, the functions of the various modules in the system and the interfaces
10between modules. For more precise details about any data structure or calling
11convention, see the include files and comments in the source code.
12
13We assume that the reader is already somewhat familiar with the JPEG standard.
14The README file includes references for learning about JPEG. The file
Guido Vollbeding5996a252009-06-27 00:00:00 +000015libjpeg.txt describes the library from the viewpoint of an application
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000016programmer using the library; it's best to read that file before this one.
Guido Vollbeding5996a252009-06-27 00:00:00 +000017Also, the file coderules.txt describes the coding style conventions we use.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000018
19In this document, JPEG-specific terminology follows the JPEG standard:
20 A "component" means a color channel, e.g., Red or Luminance.
21 A "sample" is a single component value (i.e., one number in the image data).
22 A "coefficient" is a frequency coefficient (a DCT transform output number).
23 A "block" is an 8x8 group of samples or coefficients.
24 An "MCU" (minimum coded unit) is an interleaved set of blocks of size
25 determined by the sampling factors, or a single block in a
26 noninterleaved scan.
27We do not use the terms "pixel" and "sample" interchangeably. When we say
28pixel, we mean an element of the full-size image, while a sample is an element
29of the downsampled image. Thus the number of samples may vary across
30components while the number of pixels does not. (This terminology is not used
31rigorously throughout the code, but it is used in places where confusion would
32otherwise result.)
33
34
35*** System features ***
36
37The IJG distribution contains two parts:
38 * A subroutine library for JPEG compression and decompression.
Thomas G. Lanebc79e061995-08-02 00:00:00 +000039 * cjpeg/djpeg, two sample applications that use the library to transform
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000040 JFIF JPEG files to and from several other image formats.
41cjpeg/djpeg are of no great intellectual complexity: they merely add a simple
42command-line user interface and I/O routines for several uncompressed image
43formats. This document concentrates on the library itself.
44
Thomas G. Lanebc79e061995-08-02 00:00:00 +000045We desire the library to be capable of supporting all JPEG baseline, extended
46sequential, and progressive DCT processes. Hierarchical processes are not
47supported.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000048
49The library does not support the lossless (spatial) JPEG process. Lossless
50JPEG shares little or no code with lossy JPEG, and would normally be used
51without the extensive pre- and post-processing provided by this library.
52We feel that lossless JPEG is better handled by a separate library.
53
54Within these limits, any set of compression parameters allowed by the JPEG
55spec should be readable for decompression. (We can be more restrictive about
56what formats we can generate.) Although the system design allows for all
57parameter values, some uncommon settings are not yet implemented and may
58never be; nonintegral sampling ratios are the prime example. Furthermore,
59we treat 8-bit vs. 12-bit data precision as a compile-time switch, not a
60run-time option, because most machines can store 8-bit pixels much more
61compactly than 12-bit.
62
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000063By itself, the library handles only interchange JPEG datastreams --- in
64particular the widely used JFIF file format. The library can be used by
65surrounding code to process interchange or abbreviated JPEG datastreams that
Thomas G. Lanebc79e061995-08-02 00:00:00 +000066are embedded in more complex file formats. (For example, libtiff uses this
67library to implement JPEG compression within the TIFF file format.)
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000068
69The library includes a substantial amount of code that is not covered by the
70JPEG standard but is necessary for typical applications of JPEG. These
71functions preprocess the image before JPEG compression or postprocess it after
72decompression. They include colorspace conversion, downsampling/upsampling,
73and color quantization. This code can be omitted if not needed.
74
75A wide range of quality vs. speed tradeoffs are possible in JPEG processing,
76and even more so in decompression postprocessing. The decompression library
77provides multiple implementations that cover most of the useful tradeoffs,
78ranging from very-high-quality down to fast-preview operation. On the
79compression side we have generally not provided low-quality choices, since
80compression is normally less time-critical. It should be understood that the
81low-quality modes may not meet the JPEG standard's accuracy requirements;
82nonetheless, they are useful for viewers.
83
84
85*** Portability issues ***
86
87Portability is an essential requirement for the library. The key portability
88issues that show up at the level of system architecture are:
89
901. Memory usage. We want the code to be able to run on PC-class machines
91with limited memory. Images should therefore be processed sequentially (in
92strips), to avoid holding the whole image in memory at once. Where a
93full-image buffer is necessary, we should be able to use either virtual memory
94or temporary files.
95
962. Near/far pointer distinction. To run efficiently on 80x86 machines, the
97code should distinguish "small" objects (kept in near data space) from
98"large" ones (kept in far data space). This is an annoying restriction, but
99fortunately it does not impact code quality for less brain-damaged machines,
100and the source code clutter turns out to be minimal with sufficient use of
101pointer typedefs.
102
1033. Data precision. We assume that "char" is at least 8 bits, "short" and
104"int" at least 16, "long" at least 32. The code will work fine with larger
105data sizes, although memory may be used inefficiently in some cases. However,
106the JPEG compressed datastream must ultimately appear on external storage as a
107sequence of 8-bit bytes if it is to conform to the standard. This may pose a
108problem on machines where char is wider than 8 bits. The library represents
109compressed data as an array of values of typedef JOCTET. If no data type
110exactly 8 bits wide is available, custom data source and data destination
111modules must be written to unpack and pack the chosen JOCTET datatype into
1128-bit external representation.
113
114
115*** System overview ***
116
117The compressor and decompressor are each divided into two main sections:
118the JPEG compressor or decompressor proper, and the preprocessing or
119postprocessing functions. The interface between these two sections is the
120image data that the official JPEG spec regards as its input or output: this
121data is in the colorspace to be used for compression, and it is downsampled
122to the sampling factors to be used. The preprocessing and postprocessing
123steps are responsible for converting a normal image representation to or from
124this form. (Those few applications that want to deal with YCbCr downsampled
125data can skip the preprocessing or postprocessing step.)
126
127Looking more closely, the compressor library contains the following main
128elements:
129
130 Preprocessing:
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000131 * Color space conversion (e.g., RGB to YCbCr).
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000132 * Edge expansion and downsampling. Optionally, this step can do simple
133 smoothing --- this is often helpful for low-quality source data.
134 JPEG proper:
135 * MCU assembly, DCT, quantization.
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000136 * Entropy coding (sequential or progressive, Huffman or arithmetic).
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000137
138In addition to these modules we need overall control, marker generation,
139and support code (memory management & error handling). There is also a
140module responsible for physically writing the output data --- typically
141this is just an interface to fwrite(), but some applications may need to
142do something else with the data.
143
144The decompressor library contains the following main elements:
145
146 JPEG proper:
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000147 * Entropy decoding (sequential or progressive, Huffman or arithmetic).
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000148 * Dequantization, inverse DCT, MCU disassembly.
149 Postprocessing:
150 * Upsampling. Optionally, this step may be able to do more general
151 rescaling of the image.
152 * Color space conversion (e.g., YCbCr to RGB). This step may also
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000153 provide gamma adjustment [ currently it does not ].
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000154 * Optional color quantization (e.g., reduction to 256 colors).
155 * Optional color precision reduction (e.g., 24-bit to 15-bit color).
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000156 [This feature is not currently implemented.]
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000157
158We also need overall control, marker parsing, and a data source module.
159The support code (memory management & error handling) can be shared with
160the compression half of the library.
161
162There may be several implementations of each of these elements, particularly
163in the decompressor, where a wide range of speed/quality tradeoffs is very
164useful. It must be understood that some of the best speedups involve
165merging adjacent steps in the pipeline. For example, upsampling, color space
166conversion, and color quantization might all be done at once when using a
167low-quality ordered-dither technique. The system architecture is designed to
168allow such merging where appropriate.
169
170
171Note: it is convenient to regard edge expansion (padding to block boundaries)
172as a preprocessing/postprocessing function, even though the JPEG spec includes
173it in compression/decompression. We do this because downsampling/upsampling
174can be simplified a little if they work on padded data: it's not necessary to
175have special cases at the right and bottom edges. Therefore the interface
176buffer is always an integral number of blocks wide and high, and we expect
177compression preprocessing to pad the source data properly. Padding will occur
178only to the next block (8-sample) boundary. In an interleaved-scan situation,
179additional dummy blocks may be used to fill out MCUs, but the MCU assembly and
180disassembly logic will create or discard these blocks internally. (This is
181advantageous for speed reasons, since we avoid DCTing the dummy blocks.
182It also permits a small reduction in file size, because the compressor can
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000183choose dummy block contents so as to minimize their size in compressed form.
184Finally, it makes the interface buffer specification independent of whether
185the file is actually interleaved or not.) Applications that wish to deal
186directly with the downsampled data must provide similar buffering and padding
187for odd-sized images.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000188
189
190*** Poor man's object-oriented programming ***
191
192It should be clear by now that we have a lot of quasi-independent processing
193steps, many of which have several possible behaviors. To avoid cluttering the
194code with lots of switch statements, we use a simple form of object-style
195programming to separate out the different possibilities.
196
197For example, two different color quantization algorithms could be implemented
198as two separate modules that present the same external interface; at runtime,
199the calling code will access the proper module indirectly through an "object".
200
201We can get the limited features we need while staying within portable C.
202The basic tool is a function pointer. An "object" is just a struct
203containing one or more function pointer fields, each of which corresponds to
204a method name in real object-oriented languages. During initialization we
205fill in the function pointers with references to whichever module we have
206determined we need to use in this run. Then invocation of the module is done
207by indirecting through a function pointer; on most machines this is no more
208expensive than a switch statement, which would be the only other way of
209making the required run-time choice. The really significant benefit, of
210course, is keeping the source code clean and well structured.
211
212We can also arrange to have private storage that varies between different
213implementations of the same kind of object. We do this by making all the
214module-specific object structs be separately allocated entities, which will
215be accessed via pointers in the master compression or decompression struct.
216The "public" fields or methods for a given kind of object are specified by
217a commonly known struct. But a module's initialization code can allocate
218a larger struct that contains the common struct as its first member, plus
219additional private fields. With appropriate pointer casting, the module's
220internal functions can access these private fields. (For a simple example,
221see jdatadst.c, which implements the external interface specified by struct
222jpeg_destination_mgr, but adds extra fields.)
223
224(Of course this would all be a lot easier if we were using C++, but we are
225not yet prepared to assume that everyone has a C++ compiler.)
226
227An important benefit of this scheme is that it is easy to provide multiple
228versions of any method, each tuned to a particular case. While a lot of
229precalculation might be done to select an optimal implementation of a method,
230the cost per invocation is constant. For example, the upsampling step might
231have a "generic" method, plus one or more "hardwired" methods for the most
232popular sampling factors; the hardwired methods would be faster because they'd
233use straight-line code instead of for-loops. The cost to determine which
234method to use is paid only once, at startup, and the selection criteria are
235hidden from the callers of the method.
236
237This plan differs a little bit from usual object-oriented structures, in that
238only one instance of each object class will exist during execution. The
239reason for having the class structure is that on different runs we may create
240different instances (choose to execute different modules). You can think of
241the term "method" as denoting the common interface presented by a particular
242set of interchangeable functions, and "object" as denoting a group of related
243methods, or the total shared interface behavior of a group of modules.
244
245
246*** Overall control structure ***
247
248We previously mentioned the need for overall control logic in the compression
249and decompression libraries. In IJG implementations prior to v5, overall
250control was mostly provided by "pipeline control" modules, which proved to be
251large, unwieldy, and hard to understand. To improve the situation, the
252control logic has been subdivided into multiple modules. The control modules
253consist of:
254
2551. Master control for module selection and initialization. This has two
256responsibilities:
257
258 1A. Startup initialization at the beginning of image processing.
259 The individual processing modules to be used in this run are selected
260 and given initialization calls.
261
262 1B. Per-pass control. This determines how many passes will be performed
263 and calls each active processing module to configure itself
264 appropriately at the beginning of each pass. End-of-pass processing,
265 where necessary, is also invoked from the master control module.
266
267 Method selection is partially distributed, in that a particular processing
268 module may contain several possible implementations of a particular method,
269 which it will select among when given its initialization call. The master
270 control code need only be concerned with decisions that affect more than
271 one module.
272
2732. Data buffering control. A separate control module exists for each
274 inter-processing-step data buffer. This module is responsible for
275 invoking the processing steps that write or read that data buffer.
276
277Each buffer controller sees the world as follows:
278
279input data => processing step A => buffer => processing step B => output data
280 | | |
281 ------------------ controller ------------------
282
283The controller knows the dataflow requirements of steps A and B: how much data
284they want to accept in one chunk and how much they output in one chunk. Its
285function is to manage its buffer and call A and B at the proper times.
286
287A data buffer control module may itself be viewed as a processing step by a
288higher-level control module; thus the control modules form a binary tree with
289elementary processing steps at the leaves of the tree.
290
291The control modules are objects. A considerable amount of flexibility can
292be had by replacing implementations of a control module. For example:
293* Merging of adjacent steps in the pipeline is done by replacing a control
294 module and its pair of processing-step modules with a single processing-
295 step module. (Hence the possible merges are determined by the tree of
296 control modules.)
297* In some processing modes, a given interstep buffer need only be a "strip"
298 buffer large enough to accommodate the desired data chunk sizes. In other
299 modes, a full-image buffer is needed and several passes are required.
300 The control module determines which kind of buffer is used and manipulates
301 virtual array buffers as needed. One or both processing steps may be
302 unaware of the multi-pass behavior.
303
304In theory, we might be able to make all of the data buffer controllers
305interchangeable and provide just one set of implementations for all. In
306practice, each one contains considerable special-case processing for its
307particular job. The buffer controller concept should be regarded as an
308overall system structuring principle, not as a complete description of the
309task performed by any one controller.
310
311
312*** Compression object structure ***
313
314Here is a sketch of the logical structure of the JPEG compression library:
315
316 |-- Colorspace conversion
317 |-- Preprocessing controller --|
318 | |-- Downsampling
319Main controller --|
320 | |-- Forward DCT, quantize
321 |-- Coefficient controller --|
322 |-- Entropy encoding
323
324This sketch also describes the flow of control (subroutine calls) during
325typical image data processing. Each of the components shown in the diagram is
326an "object" which may have several different implementations available. One
327or more source code files contain the actual implementation(s) of each object.
328
329The objects shown above are:
330
331* Main controller: buffer controller for the subsampled-data buffer, which
332 holds the preprocessed input data. This controller invokes preprocessing to
333 fill the subsampled-data buffer, and JPEG compression to empty it. There is
334 usually no need for a full-image buffer here; a strip buffer is adequate.
335
336* Preprocessing controller: buffer controller for the downsampling input data
337 buffer, which lies between colorspace conversion and downsampling. Note
338 that a unified conversion/downsampling module would probably replace this
339 controller entirely.
340
341* Colorspace conversion: converts application image data into the desired
342 JPEG color space; also changes the data from pixel-interleaved layout to
343 separate component planes. Processes one pixel row at a time.
344
345* Downsampling: performs reduction of chroma components as required.
346 Optionally may perform pixel-level smoothing as well. Processes a "row
347 group" at a time, where a row group is defined as Vmax pixel rows of each
348 component before downsampling, and Vk sample rows afterwards (remember Vk
349 differs across components). Some downsampling or smoothing algorithms may
350 require context rows above and below the current row group; the
351 preprocessing controller is responsible for supplying these rows via proper
352 buffering. The downsampler is responsible for edge expansion at the right
353 edge (i.e., extending each sample row to a multiple of 8 samples); but the
354 preprocessing controller is responsible for vertical edge expansion (i.e.,
355 duplicating the bottom sample row as needed to make a multiple of 8 rows).
356
357* Coefficient controller: buffer controller for the DCT-coefficient data.
358 This controller handles MCU assembly, including insertion of dummy DCT
359 blocks when needed at the right or bottom edge. When performing
360 Huffman-code optimization or emitting a multiscan JPEG file, this
Thomas G. Lanea8b67c41995-03-15 00:00:00 +0000361 controller is responsible for buffering the full image. The equivalent of
362 one fully interleaved MCU row of subsampled data is processed per call,
363 even when the JPEG file is noninterleaved.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000364
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000365* Forward DCT and quantization: Perform DCT, quantize, and emit coefficients.
366 Works on one or more DCT blocks at a time. (Note: the coefficients are now
367 emitted in normal array order, which the entropy encoder is expected to
368 convert to zigzag order as necessary. Prior versions of the IJG code did
369 the conversion to zigzag order within the quantization step.)
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000370
371* Entropy encoding: Perform Huffman or arithmetic entropy coding and emit the
372 coded data to the data destination module. Works on one MCU per call.
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000373 For progressive JPEG, the same DCT blocks are fed to the entropy coder
374 during each pass, and the coder must emit the appropriate subset of
375 coefficients.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000376
377In addition to the above objects, the compression library includes these
378objects:
379
380* Master control: determines the number of passes required, controls overall
381 and per-pass initialization of the other modules.
382
383* Marker writing: generates JPEG markers (except for RSTn, which is emitted
384 by the entropy encoder when needed).
385
386* Data destination manager: writes the output JPEG datastream to its final
387 destination (e.g., a file). The destination manager supplied with the
388 library knows how to write to a stdio stream; for other behaviors, the
389 surrounding application may provide its own destination manager.
390
391* Memory manager: allocates and releases memory, controls virtual arrays
392 (with backing store management, where required).
393
394* Error handler: performs formatting and output of error and trace messages;
395 determines handling of nonfatal errors. The surrounding application may
396 override some or all of this object's methods to change error handling.
397
398* Progress monitor: supports output of "percent-done" progress reports.
399 This object represents an optional callback to the surrounding application:
400 if wanted, it must be supplied by the application.
401
402The error handler, destination manager, and progress monitor objects are
403defined as separate objects in order to simplify application-specific
404customization of the JPEG library. A surrounding application may override
405individual methods or supply its own all-new implementation of one of these
406objects. The object interfaces for these objects are therefore treated as
407part of the application interface of the library, whereas the other objects
408are internal to the library.
409
410The error handler and memory manager are shared by JPEG compression and
411decompression; the progress monitor, if used, may be shared as well.
412
413
414*** Decompression object structure ***
415
416Here is a sketch of the logical structure of the JPEG decompression library:
417
418 |-- Entropy decoding
419 |-- Coefficient controller --|
420 | |-- Dequantize, Inverse DCT
421Main controller --|
422 | |-- Upsampling
423 |-- Postprocessing controller --| |-- Colorspace conversion
424 |-- Color quantization
425 |-- Color precision reduction
426
427As before, this diagram also represents typical control flow. The objects
428shown are:
429
430* Main controller: buffer controller for the subsampled-data buffer, which
431 holds the output of JPEG decompression proper. This controller's primary
432 task is to feed the postprocessing procedure. Some upsampling algorithms
433 may require context rows above and below the current row group; when this
434 is true, the main controller is responsible for managing its buffer so as
435 to make context rows available. In the current design, the main buffer is
436 always a strip buffer; a full-image buffer is never required.
437
438* Coefficient controller: buffer controller for the DCT-coefficient data.
439 This controller handles MCU disassembly, including deletion of any dummy
440 DCT blocks at the right or bottom edge. When reading a multiscan JPEG
441 file, this controller is responsible for buffering the full image.
442 (Buffering DCT coefficients, rather than samples, is necessary to support
443 progressive JPEG.) The equivalent of one fully interleaved MCU row of
444 subsampled data is processed per call, even when the source JPEG file is
445 noninterleaved.
446
447* Entropy decoding: Read coded data from the data source module and perform
448 Huffman or arithmetic entropy decoding. Works on one MCU per call.
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000449 For progressive JPEG decoding, the coefficient controller supplies the prior
450 coefficients of each MCU (initially all zeroes), which the entropy decoder
451 modifies in each scan.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000452
453* Dequantization and inverse DCT: like it says. Note that the coefficients
454 buffered by the coefficient controller have NOT been dequantized; we
455 merge dequantization and inverse DCT into a single step for speed reasons.
456 When scaled-down output is asked for, simplified DCT algorithms may be used
457 that emit only 1x1, 2x2, or 4x4 samples per DCT block, not the full 8x8.
458 Works on one DCT block at a time.
459
460* Postprocessing controller: buffer controller for the color quantization
461 input buffer, when quantization is in use. (Without quantization, this
462 controller just calls the upsampler.) For two-pass quantization, this
463 controller is responsible for buffering the full-image data.
464
465* Upsampling: restores chroma components to full size. (May support more
466 general output rescaling, too. Note that if undersized DCT outputs have
467 been emitted by the DCT module, this module must adjust so that properly
468 sized outputs are created.) Works on one row group at a time. This module
469 also calls the color conversion module, so its top level is effectively a
470 buffer controller for the upsampling->color conversion buffer. However, in
471 all but the highest-quality operating modes, upsampling and color
472 conversion are likely to be merged into a single step.
473
474* Colorspace conversion: convert from JPEG color space to output color space,
475 and change data layout from separate component planes to pixel-interleaved.
476 Works on one pixel row at a time.
477
478* Color quantization: reduce the data to colormapped form, using either an
479 externally specified colormap or an internally generated one. This module
480 is not used for full-color output. Works on one pixel row at a time; may
481 require two passes to generate a color map. Note that the output will
482 always be a single component representing colormap indexes. In the current
483 design, the output values are JSAMPLEs, so an 8-bit compilation cannot
484 quantize to more than 256 colors. This is unlikely to be a problem in
485 practice.
486
487* Color reduction: this module handles color precision reduction, e.g.,
488 generating 15-bit color (5 bits/primary) from JPEG's 24-bit output.
489 Not quite clear yet how this should be handled... should we merge it with
490 colorspace conversion???
491
492Note that some high-speed operating modes might condense the entire
493postprocessing sequence to a single module (upsample, color convert, and
494quantize in one step).
495
496In addition to the above objects, the decompression library includes these
497objects:
498
499* Master control: determines the number of passes required, controls overall
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000500 and per-pass initialization of the other modules. This is subdivided into
501 input and output control: jdinput.c controls only input-side processing,
502 while jdmaster.c handles overall initialization and output-side control.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000503
504* Marker reading: decodes JPEG markers (except for RSTn).
505
506* Data source manager: supplies the input JPEG datastream. The source
507 manager supplied with the library knows how to read from a stdio stream;
508 for other behaviors, the surrounding application may provide its own source
509 manager.
510
511* Memory manager: same as for compression library.
512
513* Error handler: same as for compression library.
514
515* Progress monitor: same as for compression library.
516
517As with compression, the data source manager, error handler, and progress
518monitor are candidates for replacement by a surrounding application.
519
520
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000521*** Decompression input and output separation ***
522
523To support efficient incremental display of progressive JPEG files, the
524decompressor is divided into two sections that can run independently:
525
5261. Data input includes marker parsing, entropy decoding, and input into the
527 coefficient controller's DCT coefficient buffer. Note that this
528 processing is relatively cheap and fast.
529
5302. Data output reads from the DCT coefficient buffer and performs the IDCT
531 and all postprocessing steps.
532
533For a progressive JPEG file, the data input processing is allowed to get
534arbitrarily far ahead of the data output processing. (This occurs only
535if the application calls jpeg_consume_input(); otherwise input and output
536run in lockstep, since the input section is called only when the output
537section needs more data.) In this way the application can avoid making
538extra display passes when data is arriving faster than the display pass
539can run. Furthermore, it is possible to abort an output pass without
540losing anything, since the coefficient buffer is read-only as far as the
Guido Vollbeding5996a252009-06-27 00:00:00 +0000541output section is concerned. See libjpeg.txt for more detail.
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000542
543A full-image coefficient array is only created if the JPEG file has multiple
544scans (or if the application specifies buffered-image mode anyway). When
545reading a single-scan file, the coefficient controller normally creates only
546a one-MCU buffer, so input and output processing must run in lockstep in this
547case. jpeg_consume_input() is effectively a no-op in this situation.
548
549The main impact of dividing the decompressor in this fashion is that we must
550be very careful with shared variables in the cinfo data structure. Each
551variable that can change during the course of decompression must be
552classified as belonging to data input or data output, and each section must
553look only at its own variables. For example, the data output section may not
554depend on any of the variables that describe the current scan in the JPEG
555file, because these may change as the data input section advances into a new
556scan.
557
558The progress monitor is (somewhat arbitrarily) defined to treat input of the
559file as one pass when buffered-image mode is not used, and to ignore data
560input work completely when buffered-image mode is used. Note that the
561library has no reliable way to predict the number of passes when dealing
562with a progressive JPEG file, nor can it predict the number of output passes
563in buffered-image mode. So the work estimate is inherently bogus anyway.
564
565No comparable division is currently made in the compression library, because
566there isn't any real need for it.
567
568
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000569*** Data formats ***
570
571Arrays of pixel sample values use the following data structure:
572
573 typedef something JSAMPLE; a pixel component value, 0..MAXJSAMPLE
574 typedef JSAMPLE *JSAMPROW; ptr to a row of samples
575 typedef JSAMPROW *JSAMPARRAY; ptr to a list of rows
576 typedef JSAMPARRAY *JSAMPIMAGE; ptr to a list of color-component arrays
577
578The basic element type JSAMPLE will typically be one of unsigned char,
579(signed) char, or short. Short will be used if samples wider than 8 bits are
580to be supported (this is a compile-time option). Otherwise, unsigned char is
581used if possible. If the compiler only supports signed chars, then it is
582necessary to mask off the value when reading. Thus, all reads of JSAMPLE
583values must be coded as "GETJSAMPLE(value)", where the macro will be defined
584as "((value) & 0xFF)" on signed-char machines and "((int) (value))" elsewhere.
585
586With these conventions, JSAMPLE values can be assumed to be >= 0. This helps
587simplify correct rounding during downsampling, etc. The JPEG standard's
588specification that sample values run from -128..127 is accommodated by
589subtracting 128 just as the sample value is copied into the source array for
590the DCT step (this will be an array of signed ints). Similarly, during
591decompression the output of the IDCT step will be immediately shifted back to
5920..255. (NB: different values are required when 12-bit samples are in use.
593The code is written in terms of MAXJSAMPLE and CENTERJSAMPLE, which will be
594defined as 255 and 128 respectively in an 8-bit implementation, and as 4095
595and 2048 in a 12-bit implementation.)
596
597We use a pointer per row, rather than a two-dimensional JSAMPLE array. This
598choice costs only a small amount of memory and has several benefits:
599* Code using the data structure doesn't need to know the allocated width of
600 the rows. This simplifies edge expansion/compression, since we can work
601 in an array that's wider than the logical picture width.
602* Indexing doesn't require multiplication; this is a performance win on many
603 machines.
604* Arrays with more than 64K total elements can be supported even on machines
605 where malloc() cannot allocate chunks larger than 64K.
606* The rows forming a component array may be allocated at different times
607 without extra copying. This trick allows some speedups in smoothing steps
608 that need access to the previous and next rows.
609
610Note that each color component is stored in a separate array; we don't use the
611traditional layout in which the components of a pixel are stored together.
612This simplifies coding of modules that work on each component independently,
613because they don't need to know how many components there are. Furthermore,
614we can read or write each component to a temporary file independently, which
615is helpful when dealing with noninterleaved JPEG files.
616
617In general, a specific sample value is accessed by code such as
618 GETJSAMPLE(image[colorcomponent][row][col])
619where col is measured from the image left edge, but row is measured from the
620first sample row currently in memory. Either of the first two indexings can
621be precomputed by copying the relevant pointer.
622
623
624Since most image-processing applications prefer to work on images in which
625the components of a pixel are stored together, the data passed to or from the
626surrounding application uses the traditional convention: a single pixel is
627represented by N consecutive JSAMPLE values, and an image row is an array of
628(# of color components)*(image width) JSAMPLEs. One or more rows of data can
629be represented by a pointer of type JSAMPARRAY in this scheme. This scheme is
630converted to component-wise storage inside the JPEG library. (Applications
631that want to skip JPEG preprocessing or postprocessing will have to contend
632with component-wise storage.)
633
634
635Arrays of DCT-coefficient values use the following data structure:
636
637 typedef short JCOEF; a 16-bit signed integer
638 typedef JCOEF JBLOCK[DCTSIZE2]; an 8x8 block of coefficients
639 typedef JBLOCK *JBLOCKROW; ptr to one horizontal row of 8x8 blocks
640 typedef JBLOCKROW *JBLOCKARRAY; ptr to a list of such rows
641 typedef JBLOCKARRAY *JBLOCKIMAGE; ptr to a list of color component arrays
642
643The underlying type is at least a 16-bit signed integer; while "short" is big
644enough on all machines of interest, on some machines it is preferable to use
645"int" for speed reasons, despite the storage cost. Coefficients are grouped
646into 8x8 blocks (but we always use #defines DCTSIZE and DCTSIZE2 rather than
Thomas G. Lane489583f1996-02-07 00:00:00 +0000647"8" and "64").
648
649The contents of a coefficient block may be in either "natural" or zigzagged
650order, and may be true values or divided by the quantization coefficients,
651depending on where the block is in the processing pipeline. In the current
652library, coefficient blocks are kept in natural order everywhere; the entropy
653codecs zigzag or dezigzag the data as it is written or read. The blocks
654contain quantized coefficients everywhere outside the DCT/IDCT subsystems.
655(This latter decision may need to be revisited to support variable
656quantization a la JPEG Part 3.)
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000657
658Notice that the allocation unit is now a row of 8x8 blocks, corresponding to
659eight rows of samples. Otherwise the structure is much the same as for
660samples, and for the same reasons.
661
662On machines where malloc() can't handle a request bigger than 64Kb, this data
663structure limits us to rows of less than 512 JBLOCKs, or a picture width of
6644000+ pixels. This seems an acceptable restriction.
665
666
667On 80x86 machines, the bottom-level pointer types (JSAMPROW and JBLOCKROW)
668must be declared as "far" pointers, but the upper levels can be "near"
669(implying that the pointer lists are allocated in the DS segment).
670We use a #define symbol FAR, which expands to the "far" keyword when
671compiling on 80x86 machines and to nothing elsewhere.
672
673
674*** Suspendable processing ***
675
676In some applications it is desirable to use the JPEG library as an
677incremental, memory-to-memory filter. In this situation the data source or
678destination may be a limited-size buffer, and we can't rely on being able to
679empty or refill the buffer at arbitrary times. Instead the application would
680like to have control return from the library at buffer overflow/underrun, and
681then resume compression or decompression at a later time.
682
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000683This scenario is supported for simple cases. (For anything more complex, we
684recommend that the application "bite the bullet" and develop real multitasking
Guido Vollbeding5996a252009-06-27 00:00:00 +0000685capability.) The libjpeg.txt file goes into more detail about the usage and
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000686limitations of this capability; here we address the implications for library
687structure.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000688
689The essence of the problem is that the entropy codec (coder or decoder) must
690be prepared to stop at arbitrary times. In turn, the controllers that call
691the entropy codec must be able to stop before having produced or consumed all
692the data that they normally would handle in one call. That part is reasonably
693straightforward: we make the controller call interfaces include "progress
694counters" which indicate the number of data chunks successfully processed, and
695we require callers to test the counter rather than just assume all of the data
696was processed.
697
698Rather than trying to restart at an arbitrary point, the current Huffman
699codecs are designed to restart at the beginning of the current MCU after a
700suspension due to buffer overflow/underrun. At the start of each call, the
701codec's internal state is loaded from permanent storage (in the JPEG object
702structures) into local variables. On successful completion of the MCU, the
703permanent state is updated. (This copying is not very expensive, and may even
704lead to *improved* performance if the local variables can be registerized.)
705If a suspension occurs, the codec simply returns without updating the state,
706thus effectively reverting to the start of the MCU. Note that this implies
707leaving some data unprocessed in the source/destination buffer (ie, the
708compressed partial MCU). The data source/destination module interfaces are
709specified so as to make this possible. This also implies that the data buffer
710must be large enough to hold a worst-case compressed MCU; a couple thousand
711bytes should be enough.
712
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000713In a successive-approximation AC refinement scan, the progressive Huffman
714decoder has to be able to undo assignments of newly nonzero coefficients if it
715suspends before the MCU is complete, since decoding requires distinguishing
716previously-zero and previously-nonzero coefficients. This is a bit tedious
717but probably won't have much effect on performance. Other variants of Huffman
718decoding need not worry about this, since they will just store the same values
719again if forced to repeat the MCU.
720
721This approach would probably not work for an arithmetic codec, since its
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000722modifiable state is quite large and couldn't be copied cheaply. Instead it
723would have to suspend and resume exactly at the point of the buffer end.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000724
725The JPEG marker reader is designed to cope with suspension at an arbitrary
726point. It does so by backing up to the start of the marker parameter segment,
727so the data buffer must be big enough to hold the largest marker of interest.
728Again, a couple KB should be adequate. (A special "skip" convention is used
729to bypass COM and APPn markers, so these can be larger than the buffer size
730without causing problems; otherwise a 64K buffer would be needed in the worst
731case.)
732
Guido Vollbeding5996a252009-06-27 00:00:00 +0000733The JPEG marker writer currently does *not* cope with suspension.
734We feel that this is not necessary; it is much easier simply to require
735the application to ensure there is enough buffer space before starting. (An
736empty 2K buffer is more than sufficient for the header markers; and ensuring
737there are a dozen or two bytes available before calling jpeg_finish_compress()
738will suffice for the trailer.) This would not work for writing multi-scan
739JPEG files, but we simply do not intend to support that capability with
740suspension.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000741
742
743*** Memory manager services ***
744
745The JPEG library's memory manager controls allocation and deallocation of
746memory, and it manages large "virtual" data arrays on machines where the
747operating system does not provide virtual memory. Note that the same
748memory manager serves both compression and decompression operations.
749
750In all cases, allocated objects are tied to a particular compression or
751decompression master record, and they will be released when that master
752record is destroyed.
753
754The memory manager does not provide explicit deallocation of objects.
755Instead, objects are created in "pools" of free storage, and a whole pool
756can be freed at once. This approach helps prevent storage-leak bugs, and
757it speeds up operations whenever malloc/free are slow (as they often are).
758The pools can be regarded as lifetime identifiers for objects. Two
759pools/lifetimes are defined:
760 * JPOOL_PERMANENT lasts until master record is destroyed
761 * JPOOL_IMAGE lasts until done with image (JPEG datastream)
762Permanent lifetime is used for parameters and tables that should be carried
763across from one datastream to another; this includes all application-visible
764parameters. Image lifetime is used for everything else. (A third lifetime,
765JPOOL_PASS = one processing pass, was originally planned. However it was
766dropped as not being worthwhile. The actual usage patterns are such that the
767peak memory usage would be about the same anyway; and having per-pass storage
768substantially complicates the virtual memory allocation rules --- see below.)
769
770The memory manager deals with three kinds of object:
7711. "Small" objects. Typically these require no more than 10K-20K total.
7722. "Large" objects. These may require tens to hundreds of K depending on
773 image size. Semantically they behave the same as small objects, but we
774 distinguish them for two reasons:
775 * On MS-DOS machines, large objects are referenced by FAR pointers,
776 small objects by NEAR pointers.
777 * Pool allocation heuristics may differ for large and small objects.
778 Note that individual "large" objects cannot exceed the size allowed by
779 type size_t, which may be 64K or less on some machines.
7803. "Virtual" objects. These are large 2-D arrays of JSAMPLEs or JBLOCKs
781 (typically large enough for the entire image being processed). The
782 memory manager provides stripwise access to these arrays. On machines
783 without virtual memory, the rest of the array may be swapped out to a
784 temporary file.
785
786(Note: JSAMPARRAY and JBLOCKARRAY data structures are a combination of large
787objects for the data proper and small objects for the row pointers. For
788convenience and speed, the memory manager provides single routines to create
789these structures. Similarly, virtual arrays include a small control block
790and a JSAMPARRAY or JBLOCKARRAY working buffer, all created with one call.)
791
792In the present implementation, virtual arrays are only permitted to have image
793lifespan. (Permanent lifespan would not be reasonable, and pass lifespan is
794not very useful since a virtual array's raison d'etre is to store data for
795multiple passes through the image.) We also expect that only "small" objects
796will be given permanent lifespan, though this restriction is not required by
797the memory manager.
798
799In a non-virtual-memory machine, some performance benefit can be gained by
800making the in-memory buffers for virtual arrays be as large as possible.
801(For small images, the buffers might fit entirely in memory, so blind
802swapping would be very wasteful.) The memory manager will adjust the height
803of the buffers to fit within a prespecified maximum memory usage. In order
804to do this in a reasonably optimal fashion, the manager needs to allocate all
805of the virtual arrays at once. Therefore, there isn't a one-step allocation
806routine for virtual arrays; instead, there is a "request" routine that simply
807allocates the control block, and a "realize" routine (called just once) that
808determines space allocation and creates all of the actual buffers. The
809realize routine must allow for space occupied by non-virtual large objects.
810(We don't bother to factor in the space needed for small objects, on the
811grounds that it isn't worth the trouble.)
812
813To support all this, we establish the following protocol for doing business
814with the memory manager:
815 1. Modules must request virtual arrays (which may have only image lifespan)
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000816 during the initial setup phase, i.e., in their jinit_xxx routines.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000817 2. All "large" objects (including JSAMPARRAYs and JBLOCKARRAYs) must also be
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000818 allocated during initial setup.
819 3. realize_virt_arrays will be called at the completion of initial setup.
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000820 The above conventions ensure that sufficient information is available
821 for it to choose a good size for virtual array buffers.
822Small objects of any lifespan may be allocated at any time. We expect that
823the total space used for small objects will be small enough to be negligible
824in the realize_virt_arrays computation.
825
826In a virtual-memory machine, we simply pretend that the available space is
827infinite, thus causing realize_virt_arrays to decide that it can allocate all
828the virtual arrays as full-size in-memory buffers. The overhead of the
829virtual-array access protocol is very small when no swapping occurs.
830
Thomas G. Lanebc79e061995-08-02 00:00:00 +0000831A virtual array can be specified to be "pre-zeroed"; when this flag is set,
832never-yet-written sections of the array are set to zero before being made
833available to the caller. If this flag is not set, never-written sections
834of the array contain garbage. (This feature exists primarily because the
835equivalent logic would otherwise be needed in jdcoefct.c for progressive
836JPEG mode; we may as well make it available for possible other uses.)
837
838The first write pass on a virtual array is required to occur in top-to-bottom
839order; read passes, as well as any write passes after the first one, may
840access the array in any order. This restriction exists partly to simplify
841the virtual array control logic, and partly because some file systems may not
842support seeking beyond the current end-of-file in a temporary file. The main
843implication of this restriction is that rearrangement of rows (such as
844converting top-to-bottom data order to bottom-to-top) must be handled while
845reading data out of the virtual array, not while putting it in.
846
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000847
848*** Memory manager internal structure ***
849
850To isolate system dependencies as much as possible, we have broken the
851memory manager into two parts. There is a reasonably system-independent
852"front end" (jmemmgr.c) and a "back end" that contains only the code
853likely to change across systems. All of the memory management methods
854outlined above are implemented by the front end. The back end provides
855the following routines for use by the front end (none of these routines
856are known to the rest of the JPEG code):
857
858jpeg_mem_init, jpeg_mem_term system-dependent initialization/shutdown
859
860jpeg_get_small, jpeg_free_small interface to malloc and free library routines
861 (or their equivalents)
862
863jpeg_get_large, jpeg_free_large interface to FAR malloc/free in MSDOS machines;
864 else usually the same as
865 jpeg_get_small/jpeg_free_small
866
867jpeg_mem_available estimate available memory
868
869jpeg_open_backing_store create a backing-store object
870
871read_backing_store, manipulate a backing-store object
872write_backing_store,
873close_backing_store
874
875On some systems there will be more than one type of backing-store object
876(specifically, in MS-DOS a backing store file might be an area of extended
877memory as well as a disk file). jpeg_open_backing_store is responsible for
878choosing how to implement a given object. The read/write/close routines
879are method pointers in the structure that describes a given object; this
880lets them be different for different object types.
881
882It may be necessary to ensure that backing store objects are explicitly
883released upon abnormal program termination. For example, MS-DOS won't free
884extended memory by itself. To support this, we will expect the main program
885or surrounding application to arrange to call self_destruct (typically via
886jpeg_destroy) upon abnormal termination. This may require a SIGINT signal
887handler or equivalent. We don't want to have the back end module install its
888own signal handler, because that would pre-empt the surrounding application's
889ability to control signal handling.
890
891The IJG distribution includes several memory manager back end implementations.
892Usually the same back end should be suitable for all applications on a given
893system, but it is possible for an application to supply its own back end at
894need.
895
896
897*** Implications of DNL marker ***
898
899Some JPEG files may use a DNL marker to postpone definition of the image
900height (this would be useful for a fax-like scanner's output, for instance).
901In these files the SOF marker claims the image height is 0, and you only
902find out the true image height at the end of the first scan.
903
904We could read these files as follows:
9051. Upon seeing zero image height, replace it by 65535 (the maximum allowed).
9062. When the DNL is found, update the image height in the global image
907 descriptor.
908This implies that control modules must avoid making copies of the image
909height, and must re-test for termination after each MCU row. This would
910be easy enough to do.
911
912In cases where image-size data structures are allocated, this approach will
913result in very inefficient use of virtual memory or much-larger-than-necessary
914temporary files. This seems acceptable for something that probably won't be a
915mainstream usage. People might have to forgo use of memory-hogging options
916(such as two-pass color quantization or noninterleaved JPEG files) if they
917want efficient conversion of such files. (One could improve efficiency by
918demanding a user-supplied upper bound for the height, less than 65536; in most
919cases it could be much less.)
920
921The standard also permits the SOF marker to overestimate the image height,
922with a DNL to give the true, smaller height at the end of the first scan.
923This would solve the space problems if the overestimate wasn't too great.
924However, it implies that you don't even know whether DNL will be used.
925
926This leads to a couple of very serious objections:
9271. Testing for a DNL marker must occur in the inner loop of the decompressor's
928 Huffman decoder; this implies a speed penalty whether the feature is used
929 or not.
9302. There is no way to hide the last-minute change in image height from an
931 application using the decoder. Thus *every* application using the IJG
932 library would suffer a complexity penalty whether it cared about DNL or
933 not.
934We currently do not support DNL because of these problems.
935
936A different approach is to insist that DNL-using files be preprocessed by a
937separate program that reads ahead to the DNL, then goes back and fixes the SOF
938marker. This is a much simpler solution and is probably far more efficient.
939Even if one wants piped input, buffering the first scan of the JPEG file needs
940a lot smaller temp file than is implied by the maximum-height method. For
941this approach we'd simply treat DNL as a no-op in the decompressor (at most,
942check that it matches the SOF image height).
943
944We will not worry about making the compressor capable of outputting DNL.
945Something similar to the first scheme above could be applied if anyone ever
946wants to make that work.