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