Home | History | Annotate | Download | only in jpeg
      1 IJG JPEG LIBRARY:  SYSTEM ARCHITECTURE
      2 
      3 Copyright (C) 1991-1995, Thomas G. Lane.
      4 This file is part of the Independent JPEG Group's software.
      5 For conditions of distribution and use, see the accompanying README file.
      6 
      7 
      8 This file provides an overview of the architecture of the IJG JPEG software;
      9 that is, the functions of the various modules in the system and the interfaces
     10 between modules.  For more precise details about any data structure or calling
     11 convention, see the include files and comments in the source code.
     12 
     13 We assume that the reader is already somewhat familiar with the JPEG standard.
     14 The README file includes references for learning about JPEG.  The file
     15 libjpeg.doc describes the library from the viewpoint of an application
     16 programmer using the library; it's best to read that file before this one.
     17 Also, the file coderules.doc describes the coding style conventions we use.
     18 
     19 In 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.
     27 We do not use the terms "pixel" and "sample" interchangeably.  When we say
     28 pixel, we mean an element of the full-size image, while a sample is an element
     29 of the downsampled image.  Thus the number of samples may vary across
     30 components while the number of pixels does not.  (This terminology is not used
     31 rigorously throughout the code, but it is used in places where confusion would
     32 otherwise result.)
     33 
     34 
     35 *** System features ***
     36 
     37 The IJG distribution contains two parts:
     38   * A subroutine library for JPEG compression and decompression.
     39   * cjpeg/djpeg, two sample applications that use the library to transform
     40     JFIF JPEG files to and from several other image formats.
     41 cjpeg/djpeg are of no great intellectual complexity: they merely add a simple
     42 command-line user interface and I/O routines for several uncompressed image
     43 formats.  This document concentrates on the library itself.
     44 
     45 We desire the library to be capable of supporting all JPEG baseline, extended
     46 sequential, and progressive DCT processes.  Hierarchical processes are not
     47 supported.
     48 
     49 The library does not support the lossless (spatial) JPEG process.  Lossless
     50 JPEG shares little or no code with lossy JPEG, and would normally be used
     51 without the extensive pre- and post-processing provided by this library.
     52 We feel that lossless JPEG is better handled by a separate library.
     53 
     54 Within these limits, any set of compression parameters allowed by the JPEG
     55 spec should be readable for decompression.  (We can be more restrictive about
     56 what formats we can generate.)  Although the system design allows for all
     57 parameter values, some uncommon settings are not yet implemented and may
     58 never be; nonintegral sampling ratios are the prime example.  Furthermore,
     59 we treat 8-bit vs. 12-bit data precision as a compile-time switch, not a
     60 run-time option, because most machines can store 8-bit pixels much more
     61 compactly than 12-bit.
     62 
     63 For legal reasons, JPEG arithmetic coding is not currently supported, but
     64 extending the library to include it would be straightforward.
     65 
     66 By itself, the library handles only interchange JPEG datastreams --- in
     67 particular the widely used JFIF file format.  The library can be used by
     68 surrounding code to process interchange or abbreviated JPEG datastreams that
     69 are embedded in more complex file formats.  (For example, libtiff uses this
     70 library to implement JPEG compression within the TIFF file format.)
     71 
     72 The library includes a substantial amount of code that is not covered by the
     73 JPEG standard but is necessary for typical applications of JPEG.  These
     74 functions preprocess the image before JPEG compression or postprocess it after
     75 decompression.  They include colorspace conversion, downsampling/upsampling,
     76 and color quantization.  This code can be omitted if not needed.
     77 
     78 A wide range of quality vs. speed tradeoffs are possible in JPEG processing,
     79 and even more so in decompression postprocessing.  The decompression library
     80 provides multiple implementations that cover most of the useful tradeoffs,
     81 ranging from very-high-quality down to fast-preview operation.  On the
     82 compression side we have generally not provided low-quality choices, since
     83 compression is normally less time-critical.  It should be understood that the
     84 low-quality modes may not meet the JPEG standard's accuracy requirements;
     85 nonetheless, they are useful for viewers.
     86 
     87 
     88 *** Portability issues ***
     89 
     90 Portability is an essential requirement for the library.  The key portability
     91 issues that show up at the level of system architecture are:
     92 
     93 1.  Memory usage.  We want the code to be able to run on PC-class machines
     94 with limited memory.  Images should therefore be processed sequentially (in
     95 strips), to avoid holding the whole image in memory at once.  Where a
     96 full-image buffer is necessary, we should be able to use either virtual memory
     97 or temporary files.
     98 
     99 2.  Near/far pointer distinction.  To run efficiently on 80x86 machines, the
    100 code should distinguish "small" objects (kept in near data space) from
    101 "large" ones (kept in far data space).  This is an annoying restriction, but
    102 fortunately it does not impact code quality for less brain-damaged machines,
    103 and the source code clutter turns out to be minimal with sufficient use of
    104 pointer typedefs.
    105 
    106 3. Data precision.  We assume that "char" is at least 8 bits, "short" and
    107 "int" at least 16, "long" at least 32.  The code will work fine with larger
    108 data sizes, although memory may be used inefficiently in some cases.  However,
    109 the JPEG compressed datastream must ultimately appear on external storage as a
    110 sequence of 8-bit bytes if it is to conform to the standard.  This may pose a
    111 problem on machines where char is wider than 8 bits.  The library represents
    112 compressed data as an array of values of typedef JOCTET.  If no data type
    113 exactly 8 bits wide is available, custom data source and data destination
    114 modules must be written to unpack and pack the chosen JOCTET datatype into
    115 8-bit external representation.
    116 
    117 
    118 *** System overview ***
    119 
    120 The compressor and decompressor are each divided into two main sections:
    121 the JPEG compressor or decompressor proper, and the preprocessing or
    122 postprocessing functions.  The interface between these two sections is the
    123 image data that the official JPEG spec regards as its input or output: this
    124 data is in the colorspace to be used for compression, and it is downsampled
    125 to the sampling factors to be used.  The preprocessing and postprocessing
    126 steps are responsible for converting a normal image representation to or from
    127 this form.  (Those few applications that want to deal with YCbCr downsampled
    128 data can skip the preprocessing or postprocessing step.)
    129 
    130 Looking more closely, the compressor library contains the following main
    131 elements:
    132 
    133   Preprocessing:
    134     * Color space conversion (e.g., RGB to YCbCr).
    135     * Edge expansion and downsampling.  Optionally, this step can do simple
    136       smoothing --- this is often helpful for low-quality source data.
    137   JPEG proper:
    138     * MCU assembly, DCT, quantization.
    139     * Entropy coding (sequential or progressive, Huffman or arithmetic).
    140 
    141 In addition to these modules we need overall control, marker generation,
    142 and support code (memory management & error handling).  There is also a
    143 module responsible for physically writing the output data --- typically
    144 this is just an interface to fwrite(), but some applications may need to
    145 do something else with the data.
    146 
    147 The decompressor library contains the following main elements:
    148 
    149   JPEG proper:
    150     * Entropy decoding (sequential or progressive, Huffman or arithmetic).
    151     * Dequantization, inverse DCT, MCU disassembly.
    152   Postprocessing:
    153     * Upsampling.  Optionally, this step may be able to do more general
    154       rescaling of the image.
    155     * Color space conversion (e.g., YCbCr to RGB).  This step may also
    156       provide gamma adjustment [ currently it does not ].
    157     * Optional color quantization (e.g., reduction to 256 colors).
    158     * Optional color precision reduction (e.g., 24-bit to 15-bit color).
    159       [This feature is not currently implemented.]
    160 
    161 We also need overall control, marker parsing, and a data source module.
    162 The support code (memory management & error handling) can be shared with
    163 the compression half of the library.
    164 
    165 There may be several implementations of each of these elements, particularly
    166 in the decompressor, where a wide range of speed/quality tradeoffs is very
    167 useful.  It must be understood that some of the best speedups involve
    168 merging adjacent steps in the pipeline.  For example, upsampling, color space
    169 conversion, and color quantization might all be done at once when using a
    170 low-quality ordered-dither technique.  The system architecture is designed to
    171 allow such merging where appropriate.
    172 
    173 
    174 Note: it is convenient to regard edge expansion (padding to block boundaries)
    175 as a preprocessing/postprocessing function, even though the JPEG spec includes
    176 it in compression/decompression.  We do this because downsampling/upsampling
    177 can be simplified a little if they work on padded data: it's not necessary to
    178 have special cases at the right and bottom edges.  Therefore the interface
    179 buffer is always an integral number of blocks wide and high, and we expect
    180 compression preprocessing to pad the source data properly.  Padding will occur
    181 only to the next block (8-sample) boundary.  In an interleaved-scan situation,
    182 additional dummy blocks may be used to fill out MCUs, but the MCU assembly and
    183 disassembly logic will create or discard these blocks internally.  (This is
    184 advantageous for speed reasons, since we avoid DCTing the dummy blocks.
    185 It also permits a small reduction in file size, because the compressor can
    186 choose dummy block contents so as to minimize their size in compressed form.
    187 Finally, it makes the interface buffer specification independent of whether
    188 the file is actually interleaved or not.)  Applications that wish to deal
    189 directly with the downsampled data must provide similar buffering and padding
    190 for odd-sized images.
    191 
    192 
    193 *** Poor man's object-oriented programming ***
    194 
    195 It should be clear by now that we have a lot of quasi-independent processing
    196 steps, many of which have several possible behaviors.  To avoid cluttering the
    197 code with lots of switch statements, we use a simple form of object-style
    198 programming to separate out the different possibilities.
    199 
    200 For example, two different color quantization algorithms could be implemented
    201 as two separate modules that present the same external interface; at runtime,
    202 the calling code will access the proper module indirectly through an "object".
    203 
    204 We can get the limited features we need while staying within portable C.
    205 The basic tool is a function pointer.  An "object" is just a struct
    206 containing one or more function pointer fields, each of which corresponds to
    207 a method name in real object-oriented languages.  During initialization we
    208 fill in the function pointers with references to whichever module we have
    209 determined we need to use in this run.  Then invocation of the module is done
    210 by indirecting through a function pointer; on most machines this is no more
    211 expensive than a switch statement, which would be the only other way of
    212 making the required run-time choice.  The really significant benefit, of
    213 course, is keeping the source code clean and well structured.
    214 
    215 We can also arrange to have private storage that varies between different
    216 implementations of the same kind of object.  We do this by making all the
    217 module-specific object structs be separately allocated entities, which will
    218 be accessed via pointers in the master compression or decompression struct.
    219 The "public" fields or methods for a given kind of object are specified by
    220 a commonly known struct.  But a module's initialization code can allocate
    221 a larger struct that contains the common struct as its first member, plus
    222 additional private fields.  With appropriate pointer casting, the module's
    223 internal functions can access these private fields.  (For a simple example,
    224 see jdatadst.c, which implements the external interface specified by struct
    225 jpeg_destination_mgr, but adds extra fields.)
    226 
    227 (Of course this would all be a lot easier if we were using C++, but we are
    228 not yet prepared to assume that everyone has a C++ compiler.)
    229 
    230 An important benefit of this scheme is that it is easy to provide multiple
    231 versions of any method, each tuned to a particular case.  While a lot of
    232 precalculation might be done to select an optimal implementation of a method,
    233 the cost per invocation is constant.  For example, the upsampling step might
    234 have a "generic" method, plus one or more "hardwired" methods for the most
    235 popular sampling factors; the hardwired methods would be faster because they'd
    236 use straight-line code instead of for-loops.  The cost to determine which
    237 method to use is paid only once, at startup, and the selection criteria are
    238 hidden from the callers of the method.
    239 
    240 This plan differs a little bit from usual object-oriented structures, in that
    241 only one instance of each object class will exist during execution.  The
    242 reason for having the class structure is that on different runs we may create
    243 different instances (choose to execute different modules).  You can think of
    244 the term "method" as denoting the common interface presented by a particular
    245 set of interchangeable functions, and "object" as denoting a group of related
    246 methods, or the total shared interface behavior of a group of modules.
    247 
    248 
    249 *** Overall control structure ***
    250 
    251 We previously mentioned the need for overall control logic in the compression
    252 and decompression libraries.  In IJG implementations prior to v5, overall
    253 control was mostly provided by "pipeline control" modules, which proved to be
    254 large, unwieldy, and hard to understand.  To improve the situation, the
    255 control logic has been subdivided into multiple modules.  The control modules
    256 consist of:
    257 
    258 1. Master control for module selection and initialization.  This has two
    259 responsibilities:
    260 
    261    1A.  Startup initialization at the beginning of image processing.
    262         The individual processing modules to be used in this run are selected
    263         and given initialization calls.
    264 
    265    1B.  Per-pass control.  This determines how many passes will be performed
    266         and calls each active processing module to configure itself
    267         appropriately at the beginning of each pass.  End-of-pass processing,
    268 	where necessary, is also invoked from the master control module.
    269 
    270    Method selection is partially distributed, in that a particular processing
    271    module may contain several possible implementations of a particular method,
    272    which it will select among when given its initialization call.  The master
    273    control code need only be concerned with decisions that affect more than
    274    one module.
    275  
    276 2. Data buffering control.  A separate control module exists for each
    277    inter-processing-step data buffer.  This module is responsible for
    278    invoking the processing steps that write or read that data buffer.
    279 
    280 Each buffer controller sees the world as follows:
    281 
    282 input data => processing step A => buffer => processing step B => output data
    283                       |              |               |
    284               ------------------ controller ------------------
    285 
    286 The controller knows the dataflow requirements of steps A and B: how much data
    287 they want to accept in one chunk and how much they output in one chunk.  Its
    288 function is to manage its buffer and call A and B at the proper times.
    289 
    290 A data buffer control module may itself be viewed as a processing step by a
    291 higher-level control module; thus the control modules form a binary tree with
    292 elementary processing steps at the leaves of the tree.
    293 
    294 The control modules are objects.  A considerable amount of flexibility can
    295 be had by replacing implementations of a control module.  For example:
    296 * Merging of adjacent steps in the pipeline is done by replacing a control
    297   module and its pair of processing-step modules with a single processing-
    298   step module.  (Hence the possible merges are determined by the tree of
    299   control modules.)
    300 * In some processing modes, a given interstep buffer need only be a "strip"
    301   buffer large enough to accommodate the desired data chunk sizes.  In other
    302   modes, a full-image buffer is needed and several passes are required.
    303   The control module determines which kind of buffer is used and manipulates
    304   virtual array buffers as needed.  One or both processing steps may be
    305   unaware of the multi-pass behavior.
    306 
    307 In theory, we might be able to make all of the data buffer controllers
    308 interchangeable and provide just one set of implementations for all.  In
    309 practice, each one contains considerable special-case processing for its
    310 particular job.  The buffer controller concept should be regarded as an
    311 overall system structuring principle, not as a complete description of the
    312 task performed by any one controller.
    313 
    314 
    315 *** Compression object structure ***
    316 
    317 Here is a sketch of the logical structure of the JPEG compression library:
    318 
    319                                                  |-- Colorspace conversion
    320                   |-- Preprocessing controller --|
    321                   |                              |-- Downsampling
    322 Main controller --|
    323                   |                            |-- Forward DCT, quantize
    324                   |-- Coefficient controller --|
    325                                                |-- Entropy encoding
    326 
    327 This sketch also describes the flow of control (subroutine calls) during
    328 typical image data processing.  Each of the components shown in the diagram is
    329 an "object" which may have several different implementations available.  One
    330 or more source code files contain the actual implementation(s) of each object.
    331 
    332 The objects shown above are:
    333 
    334 * Main controller: buffer controller for the subsampled-data buffer, which
    335   holds the preprocessed input data.  This controller invokes preprocessing to
    336   fill the subsampled-data buffer, and JPEG compression to empty it.  There is
    337   usually no need for a full-image buffer here; a strip buffer is adequate.
    338 
    339 * Preprocessing controller: buffer controller for the downsampling input data
    340   buffer, which lies between colorspace conversion and downsampling.  Note
    341   that a unified conversion/downsampling module would probably replace this
    342   controller entirely.
    343 
    344 * Colorspace conversion: converts application image data into the desired
    345   JPEG color space; also changes the data from pixel-interleaved layout to
    346   separate component planes.  Processes one pixel row at a time.
    347 
    348 * Downsampling: performs reduction of chroma components as required.
    349   Optionally may perform pixel-level smoothing as well.  Processes a "row
    350   group" at a time, where a row group is defined as Vmax pixel rows of each
    351   component before downsampling, and Vk sample rows afterwards (remember Vk
    352   differs across components).  Some downsampling or smoothing algorithms may
    353   require context rows above and below the current row group; the
    354   preprocessing controller is responsible for supplying these rows via proper
    355   buffering.  The downsampler is responsible for edge expansion at the right
    356   edge (i.e., extending each sample row to a multiple of 8 samples); but the
    357   preprocessing controller is responsible for vertical edge expansion (i.e.,
    358   duplicating the bottom sample row as needed to make a multiple of 8 rows).
    359 
    360 * Coefficient controller: buffer controller for the DCT-coefficient data.
    361   This controller handles MCU assembly, including insertion of dummy DCT
    362   blocks when needed at the right or bottom edge.  When performing
    363   Huffman-code optimization or emitting a multiscan JPEG file, this
    364   controller is responsible for buffering the full image.  The equivalent of
    365   one fully interleaved MCU row of subsampled data is processed per call,
    366   even when the JPEG file is noninterleaved.
    367 
    368 * Forward DCT and quantization: Perform DCT, quantize, and emit coefficients.
    369   Works on one or more DCT blocks at a time.  (Note: the coefficients are now
    370   emitted in normal array order, which the entropy encoder is expected to
    371   convert to zigzag order as necessary.  Prior versions of the IJG code did
    372   the conversion to zigzag order within the quantization step.)
    373 
    374 * Entropy encoding: Perform Huffman or arithmetic entropy coding and emit the
    375   coded data to the data destination module.  Works on one MCU per call.
    376   For progressive JPEG, the same DCT blocks are fed to the entropy coder
    377   during each pass, and the coder must emit the appropriate subset of
    378   coefficients.
    379 
    380 In addition to the above objects, the compression library includes these
    381 objects:
    382 
    383 * Master control: determines the number of passes required, controls overall
    384   and per-pass initialization of the other modules.
    385 
    386 * Marker writing: generates JPEG markers (except for RSTn, which is emitted
    387   by the entropy encoder when needed).
    388 
    389 * Data destination manager: writes the output JPEG datastream to its final
    390   destination (e.g., a file).  The destination manager supplied with the
    391   library knows how to write to a stdio stream; for other behaviors, the
    392   surrounding application may provide its own destination manager.
    393 
    394 * Memory manager: allocates and releases memory, controls virtual arrays
    395   (with backing store management, where required).
    396 
    397 * Error handler: performs formatting and output of error and trace messages;
    398   determines handling of nonfatal errors.  The surrounding application may
    399   override some or all of this object's methods to change error handling.
    400 
    401 * Progress monitor: supports output of "percent-done" progress reports.
    402   This object represents an optional callback to the surrounding application:
    403   if wanted, it must be supplied by the application.
    404 
    405 The error handler, destination manager, and progress monitor objects are
    406 defined as separate objects in order to simplify application-specific
    407 customization of the JPEG library.  A surrounding application may override
    408 individual methods or supply its own all-new implementation of one of these
    409 objects.  The object interfaces for these objects are therefore treated as
    410 part of the application interface of the library, whereas the other objects
    411 are internal to the library.
    412 
    413 The error handler and memory manager are shared by JPEG compression and
    414 decompression; the progress monitor, if used, may be shared as well.
    415 
    416 
    417 *** Decompression object structure ***
    418 
    419 Here is a sketch of the logical structure of the JPEG decompression library:
    420 
    421                                                |-- Entropy decoding
    422                   |-- Coefficient controller --|
    423                   |                            |-- Dequantize, Inverse DCT
    424 Main controller --|
    425                   |                               |-- Upsampling
    426                   |-- Postprocessing controller --|   |-- Colorspace conversion
    427                                                   |-- Color quantization
    428                                                   |-- Color precision reduction
    429 
    430 As before, this diagram also represents typical control flow.  The objects
    431 shown are:
    432 
    433 * Main controller: buffer controller for the subsampled-data buffer, which
    434   holds the output of JPEG decompression proper.  This controller's primary
    435   task is to feed the postprocessing procedure.  Some upsampling algorithms
    436   may require context rows above and below the current row group; when this
    437   is true, the main controller is responsible for managing its buffer so as
    438   to make context rows available.  In the current design, the main buffer is
    439   always a strip buffer; a full-image buffer is never required.
    440 
    441 * Coefficient controller: buffer controller for the DCT-coefficient data.
    442   This controller handles MCU disassembly, including deletion of any dummy
    443   DCT blocks at the right or bottom edge.  When reading a multiscan JPEG
    444   file, this controller is responsible for buffering the full image.
    445   (Buffering DCT coefficients, rather than samples, is necessary to support
    446   progressive JPEG.)  The equivalent of one fully interleaved MCU row of
    447   subsampled data is processed per call, even when the source JPEG file is
    448   noninterleaved.
    449 
    450 * Entropy decoding: Read coded data from the data source module and perform
    451   Huffman or arithmetic entropy decoding.  Works on one MCU per call.
    452   For progressive JPEG decoding, the coefficient controller supplies the prior
    453   coefficients of each MCU (initially all zeroes), which the entropy decoder
    454   modifies in each scan.
    455 
    456 * Dequantization and inverse DCT: like it says.  Note that the coefficients
    457   buffered by the coefficient controller have NOT been dequantized; we
    458   merge dequantization and inverse DCT into a single step for speed reasons.
    459   When scaled-down output is asked for, simplified DCT algorithms may be used
    460   that emit only 1x1, 2x2, or 4x4 samples per DCT block, not the full 8x8.
    461   Works on one DCT block at a time.
    462 
    463 * Postprocessing controller: buffer controller for the color quantization
    464   input buffer, when quantization is in use.  (Without quantization, this
    465   controller just calls the upsampler.)  For two-pass quantization, this
    466   controller is responsible for buffering the full-image data.
    467 
    468 * Upsampling: restores chroma components to full size.  (May support more
    469   general output rescaling, too.  Note that if undersized DCT outputs have
    470   been emitted by the DCT module, this module must adjust so that properly
    471   sized outputs are created.)  Works on one row group at a time.  This module
    472   also calls the color conversion module, so its top level is effectively a
    473   buffer controller for the upsampling->color conversion buffer.  However, in
    474   all but the highest-quality operating modes, upsampling and color
    475   conversion are likely to be merged into a single step.
    476 
    477 * Colorspace conversion: convert from JPEG color space to output color space,
    478   and change data layout from separate component planes to pixel-interleaved.
    479   Works on one pixel row at a time.
    480 
    481 * Color quantization: reduce the data to colormapped form, using either an
    482   externally specified colormap or an internally generated one.  This module
    483   is not used for full-color output.  Works on one pixel row at a time; may
    484   require two passes to generate a color map.  Note that the output will
    485   always be a single component representing colormap indexes.  In the current
    486   design, the output values are JSAMPLEs, so an 8-bit compilation cannot
    487   quantize to more than 256 colors.  This is unlikely to be a problem in
    488   practice.
    489 
    490 * Color reduction: this module handles color precision reduction, e.g.,
    491   generating 15-bit color (5 bits/primary) from JPEG's 24-bit output.
    492   Not quite clear yet how this should be handled... should we merge it with
    493   colorspace conversion???
    494 
    495 Note that some high-speed operating modes might condense the entire
    496 postprocessing sequence to a single module (upsample, color convert, and
    497 quantize in one step).
    498 
    499 In addition to the above objects, the decompression library includes these
    500 objects:
    501 
    502 * Master control: determines the number of passes required, controls overall
    503   and per-pass initialization of the other modules.  This is subdivided into
    504   input and output control: jdinput.c controls only input-side processing,
    505   while jdmaster.c handles overall initialization and output-side control.
    506 
    507 * Marker reading: decodes JPEG markers (except for RSTn).
    508 
    509 * Data source manager: supplies the input JPEG datastream.  The source
    510   manager supplied with the library knows how to read from a stdio stream;
    511   for other behaviors, the surrounding application may provide its own source
    512   manager.
    513 
    514 * Memory manager: same as for compression library.
    515 
    516 * Error handler: same as for compression library.
    517 
    518 * Progress monitor: same as for compression library.
    519 
    520 As with compression, the data source manager, error handler, and progress
    521 monitor are candidates for replacement by a surrounding application.
    522 
    523 
    524 *** Decompression input and output separation ***
    525 
    526 To support efficient incremental display of progressive JPEG files, the
    527 decompressor is divided into two sections that can run independently:
    528 
    529 1. Data input includes marker parsing, entropy decoding, and input into the
    530    coefficient controller's DCT coefficient buffer.  Note that this
    531    processing is relatively cheap and fast.
    532 
    533 2. Data output reads from the DCT coefficient buffer and performs the IDCT
    534    and all postprocessing steps.
    535 
    536 For a progressive JPEG file, the data input processing is allowed to get
    537 arbitrarily far ahead of the data output processing.  (This occurs only
    538 if the application calls jpeg_consume_input(); otherwise input and output
    539 run in lockstep, since the input section is called only when the output
    540 section needs more data.)  In this way the application can avoid making
    541 extra display passes when data is arriving faster than the display pass
    542 can run.  Furthermore, it is possible to abort an output pass without
    543 losing anything, since the coefficient buffer is read-only as far as the
    544 output section is concerned.  See libjpeg.doc for more detail.
    545 
    546 A full-image coefficient array is only created if the JPEG file has multiple
    547 scans (or if the application specifies buffered-image mode anyway).  When
    548 reading a single-scan file, the coefficient controller normally creates only
    549 a one-MCU buffer, so input and output processing must run in lockstep in this
    550 case.  jpeg_consume_input() is effectively a no-op in this situation.
    551 
    552 The main impact of dividing the decompressor in this fashion is that we must
    553 be very careful with shared variables in the cinfo data structure.  Each
    554 variable that can change during the course of decompression must be
    555 classified as belonging to data input or data output, and each section must
    556 look only at its own variables.  For example, the data output section may not
    557 depend on any of the variables that describe the current scan in the JPEG
    558 file, because these may change as the data input section advances into a new
    559 scan.
    560 
    561 The progress monitor is (somewhat arbitrarily) defined to treat input of the
    562 file as one pass when buffered-image mode is not used, and to ignore data
    563 input work completely when buffered-image mode is used.  Note that the
    564 library has no reliable way to predict the number of passes when dealing
    565 with a progressive JPEG file, nor can it predict the number of output passes
    566 in buffered-image mode.  So the work estimate is inherently bogus anyway.
    567 
    568 No comparable division is currently made in the compression library, because
    569 there isn't any real need for it.
    570 
    571 
    572 *** Data formats ***
    573 
    574 Arrays of pixel sample values use the following data structure:
    575 
    576     typedef something JSAMPLE;		a pixel component value, 0..MAXJSAMPLE
    577     typedef JSAMPLE *JSAMPROW;		ptr to a row of samples
    578     typedef JSAMPROW *JSAMPARRAY;	ptr to a list of rows
    579     typedef JSAMPARRAY *JSAMPIMAGE;	ptr to a list of color-component arrays
    580 
    581 The basic element type JSAMPLE will typically be one of unsigned char,
    582 (signed) char, or short.  Short will be used if samples wider than 8 bits are
    583 to be supported (this is a compile-time option).  Otherwise, unsigned char is
    584 used if possible.  If the compiler only supports signed chars, then it is
    585 necessary to mask off the value when reading.  Thus, all reads of JSAMPLE
    586 values must be coded as "GETJSAMPLE(value)", where the macro will be defined
    587 as "((value) & 0xFF)" on signed-char machines and "((int) (value))" elsewhere.
    588 
    589 With these conventions, JSAMPLE values can be assumed to be >= 0.  This helps
    590 simplify correct rounding during downsampling, etc.  The JPEG standard's
    591 specification that sample values run from -128..127 is accommodated by
    592 subtracting 128 just as the sample value is copied into the source array for
    593 the DCT step (this will be an array of signed ints).  Similarly, during
    594 decompression the output of the IDCT step will be immediately shifted back to
    595 0..255.  (NB: different values are required when 12-bit samples are in use.
    596 The code is written in terms of MAXJSAMPLE and CENTERJSAMPLE, which will be
    597 defined as 255 and 128 respectively in an 8-bit implementation, and as 4095
    598 and 2048 in a 12-bit implementation.)
    599 
    600 We use a pointer per row, rather than a two-dimensional JSAMPLE array.  This
    601 choice costs only a small amount of memory and has several benefits:
    602 * Code using the data structure doesn't need to know the allocated width of
    603   the rows.  This simplifies edge expansion/compression, since we can work
    604   in an array that's wider than the logical picture width.
    605 * Indexing doesn't require multiplication; this is a performance win on many
    606   machines.
    607 * Arrays with more than 64K total elements can be supported even on machines
    608   where malloc() cannot allocate chunks larger than 64K.
    609 * The rows forming a component array may be allocated at different times
    610   without extra copying.  This trick allows some speedups in smoothing steps
    611   that need access to the previous and next rows.
    612 
    613 Note that each color component is stored in a separate array; we don't use the
    614 traditional layout in which the components of a pixel are stored together.
    615 This simplifies coding of modules that work on each component independently,
    616 because they don't need to know how many components there are.  Furthermore,
    617 we can read or write each component to a temporary file independently, which
    618 is helpful when dealing with noninterleaved JPEG files.
    619 
    620 In general, a specific sample value is accessed by code such as
    621 	GETJSAMPLE(image[colorcomponent][row][col])
    622 where col is measured from the image left edge, but row is measured from the
    623 first sample row currently in memory.  Either of the first two indexings can
    624 be precomputed by copying the relevant pointer.
    625 
    626 
    627 Since most image-processing applications prefer to work on images in which
    628 the components of a pixel are stored together, the data passed to or from the
    629 surrounding application uses the traditional convention: a single pixel is
    630 represented by N consecutive JSAMPLE values, and an image row is an array of
    631 (# of color components)*(image width) JSAMPLEs.  One or more rows of data can
    632 be represented by a pointer of type JSAMPARRAY in this scheme.  This scheme is
    633 converted to component-wise storage inside the JPEG library.  (Applications
    634 that want to skip JPEG preprocessing or postprocessing will have to contend
    635 with component-wise storage.)
    636 
    637 
    638 Arrays of DCT-coefficient values use the following data structure:
    639 
    640     typedef short JCOEF;		a 16-bit signed integer
    641     typedef JCOEF JBLOCK[DCTSIZE2];	an 8x8 block of coefficients
    642     typedef JBLOCK *JBLOCKROW;		ptr to one horizontal row of 8x8 blocks
    643     typedef JBLOCKROW *JBLOCKARRAY;	ptr to a list of such rows
    644     typedef JBLOCKARRAY *JBLOCKIMAGE;	ptr to a list of color component arrays
    645 
    646 The underlying type is at least a 16-bit signed integer; while "short" is big
    647 enough on all machines of interest, on some machines it is preferable to use
    648 "int" for speed reasons, despite the storage cost.  Coefficients are grouped
    649 into 8x8 blocks (but we always use #defines DCTSIZE and DCTSIZE2 rather than
    650 "8" and "64").
    651 
    652 The contents of a coefficient block may be in either "natural" or zigzagged
    653 order, and may be true values or divided by the quantization coefficients,
    654 depending on where the block is in the processing pipeline.  In the current
    655 library, coefficient blocks are kept in natural order everywhere; the entropy
    656 codecs zigzag or dezigzag the data as it is written or read.  The blocks
    657 contain quantized coefficients everywhere outside the DCT/IDCT subsystems.
    658 (This latter decision may need to be revisited to support variable
    659 quantization a la JPEG Part 3.)
    660 
    661 Notice that the allocation unit is now a row of 8x8 blocks, corresponding to
    662 eight rows of samples.  Otherwise the structure is much the same as for
    663 samples, and for the same reasons.
    664 
    665 On machines where malloc() can't handle a request bigger than 64Kb, this data
    666 structure limits us to rows of less than 512 JBLOCKs, or a picture width of
    667 4000+ pixels.  This seems an acceptable restriction.
    668 
    669 
    670 On 80x86 machines, the bottom-level pointer types (JSAMPROW and JBLOCKROW)
    671 must be declared as "far" pointers, but the upper levels can be "near"
    672 (implying that the pointer lists are allocated in the DS segment).
    673 We use a #define symbol FAR, which expands to the "far" keyword when
    674 compiling on 80x86 machines and to nothing elsewhere.
    675 
    676 
    677 *** Suspendable processing ***
    678 
    679 In some applications it is desirable to use the JPEG library as an
    680 incremental, memory-to-memory filter.  In this situation the data source or
    681 destination may be a limited-size buffer, and we can't rely on being able to
    682 empty or refill the buffer at arbitrary times.  Instead the application would
    683 like to have control return from the library at buffer overflow/underrun, and
    684 then resume compression or decompression at a later time.
    685 
    686 This scenario is supported for simple cases.  (For anything more complex, we
    687 recommend that the application "bite the bullet" and develop real multitasking
    688 capability.)  The libjpeg.doc file goes into more detail about the usage and
    689 limitations of this capability; here we address the implications for library
    690 structure.
    691 
    692 The essence of the problem is that the entropy codec (coder or decoder) must
    693 be prepared to stop at arbitrary times.  In turn, the controllers that call
    694 the entropy codec must be able to stop before having produced or consumed all
    695 the data that they normally would handle in one call.  That part is reasonably
    696 straightforward: we make the controller call interfaces include "progress
    697 counters" which indicate the number of data chunks successfully processed, and
    698 we require callers to test the counter rather than just assume all of the data
    699 was processed.
    700 
    701 Rather than trying to restart at an arbitrary point, the current Huffman
    702 codecs are designed to restart at the beginning of the current MCU after a
    703 suspension due to buffer overflow/underrun.  At the start of each call, the
    704 codec's internal state is loaded from permanent storage (in the JPEG object
    705 structures) into local variables.  On successful completion of the MCU, the
    706 permanent state is updated.  (This copying is not very expensive, and may even
    707 lead to *improved* performance if the local variables can be registerized.)
    708 If a suspension occurs, the codec simply returns without updating the state,
    709 thus effectively reverting to the start of the MCU.  Note that this implies
    710 leaving some data unprocessed in the source/destination buffer (ie, the
    711 compressed partial MCU).  The data source/destination module interfaces are
    712 specified so as to make this possible.  This also implies that the data buffer
    713 must be large enough to hold a worst-case compressed MCU; a couple thousand
    714 bytes should be enough.
    715 
    716 In a successive-approximation AC refinement scan, the progressive Huffman
    717 decoder has to be able to undo assignments of newly nonzero coefficients if it
    718 suspends before the MCU is complete, since decoding requires distinguishing
    719 previously-zero and previously-nonzero coefficients.  This is a bit tedious
    720 but probably won't have much effect on performance.  Other variants of Huffman
    721 decoding need not worry about this, since they will just store the same values
    722 again if forced to repeat the MCU.
    723 
    724 This approach would probably not work for an arithmetic codec, since its
    725 modifiable state is quite large and couldn't be copied cheaply.  Instead it
    726 would have to suspend and resume exactly at the point of the buffer end.
    727 
    728 The JPEG marker reader is designed to cope with suspension at an arbitrary
    729 point.  It does so by backing up to the start of the marker parameter segment,
    730 so the data buffer must be big enough to hold the largest marker of interest.
    731 Again, a couple KB should be adequate.  (A special "skip" convention is used
    732 to bypass COM and APPn markers, so these can be larger than the buffer size
    733 without causing problems; otherwise a 64K buffer would be needed in the worst
    734 case.)
    735 
    736 The JPEG marker writer currently does *not* cope with suspension.  I feel that
    737 this is not necessary; it is much easier simply to require the application to
    738 ensure there is enough buffer space before starting.  (An empty 2K buffer is
    739 more than sufficient for the header markers; and ensuring there are a dozen or
    740 two bytes available before calling jpeg_finish_compress() will suffice for the
    741 trailer.)  This would not work for writing multi-scan JPEG files, but
    742 we simply do not intend to support that capability with suspension.
    743 
    744 
    745 *** Memory manager services ***
    746 
    747 The JPEG library's memory manager controls allocation and deallocation of
    748 memory, and it manages large "virtual" data arrays on machines where the
    749 operating system does not provide virtual memory.  Note that the same
    750 memory manager serves both compression and decompression operations.
    751 
    752 In all cases, allocated objects are tied to a particular compression or
    753 decompression master record, and they will be released when that master
    754 record is destroyed.
    755 
    756 The memory manager does not provide explicit deallocation of objects.
    757 Instead, objects are created in "pools" of free storage, and a whole pool
    758 can be freed at once.  This approach helps prevent storage-leak bugs, and
    759 it speeds up operations whenever malloc/free are slow (as they often are).
    760 The pools can be regarded as lifetime identifiers for objects.  Two
    761 pools/lifetimes are defined:
    762   * JPOOL_PERMANENT	lasts until master record is destroyed
    763   * JPOOL_IMAGE		lasts until done with image (JPEG datastream)
    764 Permanent lifetime is used for parameters and tables that should be carried
    765 across from one datastream to another; this includes all application-visible
    766 parameters.  Image lifetime is used for everything else.  (A third lifetime,
    767 JPOOL_PASS = one processing pass, was originally planned.  However it was
    768 dropped as not being worthwhile.  The actual usage patterns are such that the
    769 peak memory usage would be about the same anyway; and having per-pass storage
    770 substantially complicates the virtual memory allocation rules --- see below.)
    771 
    772 The memory manager deals with three kinds of object:
    773 1. "Small" objects.  Typically these require no more than 10K-20K total.
    774 2. "Large" objects.  These may require tens to hundreds of K depending on
    775    image size.  Semantically they behave the same as small objects, but we
    776    distinguish them for two reasons:
    777      * On MS-DOS machines, large objects are referenced by FAR pointers,
    778        small objects by NEAR pointers.
    779      * Pool allocation heuristics may differ for large and small objects.
    780    Note that individual "large" objects cannot exceed the size allowed by
    781    type size_t, which may be 64K or less on some machines.
    782 3. "Virtual" objects.  These are large 2-D arrays of JSAMPLEs or JBLOCKs
    783    (typically large enough for the entire image being processed).  The
    784    memory manager provides stripwise access to these arrays.  On machines
    785    without virtual memory, the rest of the array may be swapped out to a
    786    temporary file.
    787 
    788 (Note: JSAMPARRAY and JBLOCKARRAY data structures are a combination of large
    789 objects for the data proper and small objects for the row pointers.  For
    790 convenience and speed, the memory manager provides single routines to create
    791 these structures.  Similarly, virtual arrays include a small control block
    792 and a JSAMPARRAY or JBLOCKARRAY working buffer, all created with one call.)
    793 
    794 In the present implementation, virtual arrays are only permitted to have image
    795 lifespan.  (Permanent lifespan would not be reasonable, and pass lifespan is
    796 not very useful since a virtual array's raison d'etre is to store data for
    797 multiple passes through the image.)  We also expect that only "small" objects
    798 will be given permanent lifespan, though this restriction is not required by
    799 the memory manager.
    800 
    801 In a non-virtual-memory machine, some performance benefit can be gained by
    802 making the in-memory buffers for virtual arrays be as large as possible.
    803 (For small images, the buffers might fit entirely in memory, so blind
    804 swapping would be very wasteful.)  The memory manager will adjust the height
    805 of the buffers to fit within a prespecified maximum memory usage.  In order
    806 to do this in a reasonably optimal fashion, the manager needs to allocate all
    807 of the virtual arrays at once.  Therefore, there isn't a one-step allocation
    808 routine for virtual arrays; instead, there is a "request" routine that simply
    809 allocates the control block, and a "realize" routine (called just once) that
    810 determines space allocation and creates all of the actual buffers.  The
    811 realize routine must allow for space occupied by non-virtual large objects.
    812 (We don't bother to factor in the space needed for small objects, on the
    813 grounds that it isn't worth the trouble.)
    814 
    815 To support all this, we establish the following protocol for doing business
    816 with the memory manager:
    817   1. Modules must request virtual arrays (which may have only image lifespan)
    818      during the initial setup phase, i.e., in their jinit_xxx routines.
    819   2. All "large" objects (including JSAMPARRAYs and JBLOCKARRAYs) must also be
    820      allocated during initial setup.
    821   3. realize_virt_arrays will be called at the completion of initial setup.
    822      The above conventions ensure that sufficient information is available
    823      for it to choose a good size for virtual array buffers.
    824 Small objects of any lifespan may be allocated at any time.  We expect that
    825 the total space used for small objects will be small enough to be negligible
    826 in the realize_virt_arrays computation.
    827 
    828 In a virtual-memory machine, we simply pretend that the available space is
    829 infinite, thus causing realize_virt_arrays to decide that it can allocate all
    830 the virtual arrays as full-size in-memory buffers.  The overhead of the
    831 virtual-array access protocol is very small when no swapping occurs.
    832 
    833 A virtual array can be specified to be "pre-zeroed"; when this flag is set,
    834 never-yet-written sections of the array are set to zero before being made
    835 available to the caller.  If this flag is not set, never-written sections
    836 of the array contain garbage.  (This feature exists primarily because the
    837 equivalent logic would otherwise be needed in jdcoefct.c for progressive
    838 JPEG mode; we may as well make it available for possible other uses.)
    839 
    840 The first write pass on a virtual array is required to occur in top-to-bottom
    841 order; read passes, as well as any write passes after the first one, may
    842 access the array in any order.  This restriction exists partly to simplify
    843 the virtual array control logic, and partly because some file systems may not
    844 support seeking beyond the current end-of-file in a temporary file.  The main
    845 implication of this restriction is that rearrangement of rows (such as
    846 converting top-to-bottom data order to bottom-to-top) must be handled while
    847 reading data out of the virtual array, not while putting it in.
    848 
    849 
    850 *** Memory manager internal structure ***
    851 
    852 To isolate system dependencies as much as possible, we have broken the
    853 memory manager into two parts.  There is a reasonably system-independent
    854 "front end" (jmemmgr.c) and a "back end" that contains only the code
    855 likely to change across systems.  All of the memory management methods
    856 outlined above are implemented by the front end.  The back end provides
    857 the following routines for use by the front end (none of these routines
    858 are known to the rest of the JPEG code):
    859 
    860 jpeg_mem_init, jpeg_mem_term	system-dependent initialization/shutdown
    861 
    862 jpeg_get_small, jpeg_free_small	interface to malloc and free library routines
    863 				(or their equivalents)
    864 
    865 jpeg_get_large, jpeg_free_large	interface to FAR malloc/free in MSDOS machines;
    866 				else usually the same as
    867 				jpeg_get_small/jpeg_free_small
    868 
    869 jpeg_mem_available		estimate available memory
    870 
    871 jpeg_open_backing_store		create a backing-store object
    872 
    873 read_backing_store,		manipulate a backing-store object
    874 write_backing_store,
    875 close_backing_store
    876 
    877 On some systems there will be more than one type of backing-store object
    878 (specifically, in MS-DOS a backing store file might be an area of extended
    879 memory as well as a disk file).  jpeg_open_backing_store is responsible for
    880 choosing how to implement a given object.  The read/write/close routines
    881 are method pointers in the structure that describes a given object; this
    882 lets them be different for different object types.
    883 
    884 It may be necessary to ensure that backing store objects are explicitly
    885 released upon abnormal program termination.  For example, MS-DOS won't free
    886 extended memory by itself.  To support this, we will expect the main program
    887 or surrounding application to arrange to call self_destruct (typically via
    888 jpeg_destroy) upon abnormal termination.  This may require a SIGINT signal
    889 handler or equivalent.  We don't want to have the back end module install its
    890 own signal handler, because that would pre-empt the surrounding application's
    891 ability to control signal handling.
    892 
    893 The IJG distribution includes several memory manager back end implementations.
    894 Usually the same back end should be suitable for all applications on a given
    895 system, but it is possible for an application to supply its own back end at
    896 need.
    897 
    898 
    899 *** Implications of DNL marker ***
    900 
    901 Some JPEG files may use a DNL marker to postpone definition of the image
    902 height (this would be useful for a fax-like scanner's output, for instance).
    903 In these files the SOF marker claims the image height is 0, and you only
    904 find out the true image height at the end of the first scan.
    905 
    906 We could read these files as follows:
    907 1. Upon seeing zero image height, replace it by 65535 (the maximum allowed).
    908 2. When the DNL is found, update the image height in the global image
    909    descriptor.
    910 This implies that control modules must avoid making copies of the image
    911 height, and must re-test for termination after each MCU row.  This would
    912 be easy enough to do.
    913 
    914 In cases where image-size data structures are allocated, this approach will
    915 result in very inefficient use of virtual memory or much-larger-than-necessary
    916 temporary files.  This seems acceptable for something that probably won't be a
    917 mainstream usage.  People might have to forgo use of memory-hogging options
    918 (such as two-pass color quantization or noninterleaved JPEG files) if they
    919 want efficient conversion of such files.  (One could improve efficiency by
    920 demanding a user-supplied upper bound for the height, less than 65536; in most
    921 cases it could be much less.)
    922 
    923 The standard also permits the SOF marker to overestimate the image height,
    924 with a DNL to give the true, smaller height at the end of the first scan.
    925 This would solve the space problems if the overestimate wasn't too great.
    926 However, it implies that you don't even know whether DNL will be used.
    927 
    928 This leads to a couple of very serious objections:
    929 1. Testing for a DNL marker must occur in the inner loop of the decompressor's
    930    Huffman decoder; this implies a speed penalty whether the feature is used
    931    or not.
    932 2. There is no way to hide the last-minute change in image height from an
    933    application using the decoder.  Thus *every* application using the IJG
    934    library would suffer a complexity penalty whether it cared about DNL or
    935    not.
    936 We currently do not support DNL because of these problems.
    937 
    938 A different approach is to insist that DNL-using files be preprocessed by a
    939 separate program that reads ahead to the DNL, then goes back and fixes the SOF
    940 marker.  This is a much simpler solution and is probably far more efficient.
    941 Even if one wants piped input, buffering the first scan of the JPEG file needs
    942 a lot smaller temp file than is implied by the maximum-height method.  For
    943 this approach we'd simply treat DNL as a no-op in the decompressor (at most,
    944 check that it matches the SOF image height).
    945 
    946 We will not worry about making the compressor capable of outputting DNL.
    947 Something similar to the first scheme above could be applied if anyone ever
    948 wants to make that work.
    949