structure.txt

IJG JPEG LIBRARY:  SYSTEM ARCHITECTURE

This file was part of the Independent JPEG Group's software:
Copyright (C) 1991-2012, Thomas G. Lane, Guido Vollbeding.
Lossless JPEG Modifications:
Copyright (C) 1999, Ken Murchison.
libjpeg-turbo Modifications:
Copyright (C) 2022-2023, D. R. Commander.
For conditions of distribution and use, see the accompanying README.ijg file.


This file provides an overview of the architecture of the IJG JPEG software;
that is, the functions of the various modules in the system and the interfaces
between modules.  For more precise details about any data structure or calling
convention, see the include files and comments in the source code.

We assume that the reader is already somewhat familiar with the JPEG standard.
The README.ijg file includes references for learning about JPEG.  The file
libjpeg.txt describes the library from the viewpoint of an application
programmer using the library; it's best to read that file before this one.
Also, the file coderules.txt describes the coding style conventions we use.

In this document, JPEG-specific terminology follows the JPEG standard:
  A "component" means a color channel, e.g., Red or Luminance.
  A "sample" is a single component value (i.e., one number in the image data).
  A "coefficient" is a frequency coefficient (a DCT transform output number).
  A "block" is an 8x8 group of samples or coefficients.
  A "data unit" is an abstract data type that is either a block for lossy
        (DCT-based) codecs or a sample for lossless (predictive) codecs.
  An "MCU" (minimum coded unit) is an interleaved set of data units of size
        determined by the sampling factors, or a single data unit in a
        noninterleaved scan.
We do not use the terms "pixel" and "sample" interchangeably.  When we say
pixel, we mean an element of the full-size image, while a sample is an element
of the downsampled image.  Thus the number of samples may vary across
components while the number of pixels does not.  (This terminology is not used
rigorously throughout the code, but it is used in places where confusion would
otherwise result.)


*** System features ***

The IJG distribution contains two parts:
  * A subroutine library for JPEG compression and decompression.
  * cjpeg/djpeg, two sample applications that use the library to transform
    JFIF JPEG files to and from several other image formats.
cjpeg/djpeg are of no great intellectual complexity: they merely add a simple
command-line user interface and I/O routines for several uncompressed image
formats.  This document concentrates on the library itself.

We desire the library to be capable of supporting all JPEG baseline, extended
sequential, progressive DCT, and lossless (spatial) processes.  Hierarchical
processes are not supported.

Within these limits, any set of compression parameters allowed by the JPEG
spec should be readable for decompression.  (We can be more restrictive about
what formats we can generate.)  Although the system design allows for all
parameter values, some uncommon settings are not yet implemented and may
never be; nonintegral sampling ratios are the prime example.

By itself, the library handles only interchange JPEG datastreams --- in
particular the widely used JFIF file format.  The library can be used by
surrounding code to process interchange or abbreviated JPEG datastreams that
are embedded in more complex file formats.  (For example, libtiff uses this
library to implement JPEG compression within the TIFF file format.)

The library includes a substantial amount of code that is not covered by the
JPEG standard but is necessary for typical applications of JPEG.  These
functions preprocess the image before JPEG compression or postprocess it after
decompression.  They include colorspace conversion, downsampling/upsampling,
and color quantization.  This code can be omitted if not needed.

A wide range of quality vs. speed tradeoffs are possible in JPEG processing,
and even more so in decompression postprocessing.  The decompression library
provides multiple implementations that cover most of the useful tradeoffs,
ranging from very-high-quality down to fast-preview operation.  On the
compression side we have generally not provided low-quality choices, since
compression is normally less time-critical.  It should be understood that the
low-quality modes may not meet the JPEG standard's accuracy requirements;
nonetheless, they are useful for viewers.


*** System overview ***

The compressor and decompressor are each divided into two main sections:
the JPEG compressor or decompressor proper, and the preprocessing or
postprocessing functions.  The interface between these two sections is the
image data that Rec. ITU-T T.81 | ISO/IEC 10918-1 regards as its input or
output: this data is in the colorspace to be used for compression, and it is
downsampled to the sampling factors to be used.  The preprocessing and
postprocessing steps are responsible for converting a normal image
representation to or from this form.  (Those few applications that want to deal
with YCbCr downsampled data can skip the preprocessing or postprocessing step.)

Looking more closely, the compressor library contains the following main
elements:

  Preprocessing:
    * Color space conversion (e.g., RGB to YCbCr).
    * Edge expansion and downsampling.  Optionally, this step can do simple
      smoothing --- this is often helpful for low-quality source data.
  Lossy JPEG proper:
    * MCU assembly, DCT, quantization.
    * Entropy coding (sequential or progressive, Huffman or arithmetic).
  Lossless JPEG proper:
    * Point transform.
    * Prediction, differencing.
    * Entropy coding (Huffman or arithmetic)

In addition to these modules we need overall control, marker generation,
and support code (memory management & error handling).  There is also a
module responsible for physically writing the output data --- typically
this is just an interface to fwrite(), but some applications may need to
do something else with the data.

The decompressor library contains the following main elements:

  Lossy JPEG proper:
    * Entropy decoding (sequential or progressive, Huffman or arithmetic).
    * Dequantization, inverse DCT, MCU disassembly.
  Lossless JPEG proper:
    * Entropy decoding (Huffman or arithmetic).
    * Prediction, undifferencing.
    * Point transform, sample size scaling.
  Postprocessing:
    * Upsampling.  Optionally, this step may be able to do more general
      rescaling of the image.
    * Color space conversion (e.g., YCbCr to RGB).  This step may also
      provide gamma adjustment [ currently it does not ].
    * Optional color quantization (e.g., reduction to 256 colors).
    * Optional color precision reduction (e.g., 24-bit to 15-bit color).
      [This feature is not currently implemented.]

We also need overall control, marker parsing, and a data source module.
The support code (memory management & error handling) can be shared with
the compression half of the library.

There may be several implementations of each of these elements, particularly
in the decompressor, where a wide range of speed/quality tradeoffs is very
useful.  It must be understood that some of the best speedups involve
merging adjacent steps in the pipeline.  For example, upsampling, color space
conversion, and color quantization might all be done at once when using a
low-quality ordered-dither technique.  The system architecture is designed to
allow such merging where appropriate.


Note: it is convenient to regard edge expansion (padding to block boundaries)
as a preprocessing/postprocessing function, even though
Rec. ITU-T T.81 | ISO/IEC 10918-1 includes it in compression/decompression.  We
do this because downsampling/upsampling can be simplified a little if they work
on padded data: it's not necessary to have special cases at the right and
bottom edges.  Therefore the interface buffer is always an integral number of
blocks wide and high, and we expect compression preprocessing to pad the source
data properly.  Padding will occur only to the next block (8-sample) boundary.
In an interleaved-scan situation, additional dummy blocks may be used to fill
out MCUs, but the MCU assembly and disassembly logic will create or discard
these blocks internally.  (This is advantageous for speed reasons, since we
avoid DCTing the dummy blocks.  It also permits a small reduction in file size,
because the compressor can choose dummy block contents so as to minimize their
size in compressed form.  Finally, it makes the interface buffer specification
independent of whether the file is actually interleaved or not.)  Applications
that wish to deal directly with the downsampled data must provide similar
buffering and padding for odd-sized images.


*** Poor man's object-oriented programming ***

It should be clear by now that we have a lot of quasi-independent processing
steps, many of which have several possible behaviors.  To avoid cluttering the
code with lots of switch statements, we use a simple form of object-style
programming to separate out the different possibilities.

For example, two different color quantization algorithms could be implemented
as two separate modules that present the same external interface; at runtime,
the calling code will access the proper module indirectly through an "object".

We can get the limited features we need while staying within portable C.
The basic tool is a function pointer.  An "object" is just a struct
containing one or more function pointer fields, each of which corresponds to
a method name in real object-oriented languages.  During initialization we
fill in the function pointers with references to whichever module we have
determined we need to use in this run.  Then invocation of the module is done
by indirecting through a function pointer; on most machines this is no more
expensive than a switch statement, which would be the only other way of
making the required run-time choice.  The really significant benefit, of
course, is keeping the source code clean and well structured.

We can also arrange to have private storage that varies between different
implementations of the same kind of object.  We do this by making all the
module-specific object structs be separately allocated entities, which will
be accessed via pointers in the master compression or decompression struct.
The "public" fields or methods for a given kind of object are specified by
a commonly known struct.  But a module's initialization code can allocate
a larger struct that contains the common struct as its first member, plus
additional private fields.  With appropriate pointer casting, the module's
internal functions can access these private fields.  (For a simple example,
see jdatadst.c, which implements the external interface specified by struct
jpeg_destination_mgr, but adds extra fields.)

(Of course this would all be a lot easier if we were using C++, but we are
not yet prepared to assume that everyone has a C++ compiler.)

An important benefit of this scheme is that it is easy to provide multiple
versions of any method, each tuned to a particular case.  While a lot of
precalculation might be done to select an optimal implementation of a method,
the cost per invocation is constant.  For example, the upsampling step might
have a "generic" method, plus one or more "hardwired" methods for the most
popular sampling factors; the hardwired methods would be faster because they'd
use straight-line code instead of for-loops.  The cost to determine which
method to use is paid only once, at startup, and the selection criteria are
hidden from the callers of the method.

This plan differs a little bit from usual object-oriented structures, in that
only one instance of each object class will exist during execution.  The
reason for having the class structure is that on different runs we may create
different instances (choose to execute different modules).  You can think of
the term "method" as denoting the common interface presented by a particular
set of interchangeable functions, and "object" as denoting a group of related
methods, or the total shared interface behavior of a group of modules.


*** Overall control structure ***

We previously mentioned the need for overall control logic in the compression
and decompression libraries.  In IJG implementations prior to v5, overall
control was mostly provided by "pipeline control" modules, which proved to be
large, unwieldy, and hard to understand.  To improve the situation, the
control logic has been subdivided into multiple modules.  The control modules
consist of:

1. Master control for module selection and initialization.  This has two
responsibilities:

   1A.  Startup initialization at the beginning of image processing.
        The individual processing modules to be used in this run are selected
        and given initialization calls.

   1B.  Per-pass control.  This determines how many passes will be performed
        and calls each active processing module to configure itself
        appropriately at the beginning of each pass.  End-of-pass processing,
        where necessary, is also invoked from the master control module.

   Method selection is partially distributed, in that a particular processing
   module may contain several possible implementations of a particular method,
   which it will select among when given its initialization call.  The master
   control code need only be concerned with decisions that affect more than
   one module.

2. Data buffering control.  A separate control module exists for each
   inter-processing-step data buffer.  This module is responsible for
   invoking the processing steps that write or read that data buffer.

Each buffer controller sees the world as follows:

input data => processing step A => buffer => processing step B => output data
                      |              |               |
              ------------------ controller ------------------

The controller knows the dataflow requirements of steps A and B: how much data
they want to accept in one chunk and how much they output in one chunk.  Its
function is to manage its buffer and call A and B at the proper times.

A data buffer control module may itself be viewed as a processing step by a
higher-level control module; thus the control modules form a binary tree with
elementary processing steps at the leaves of the tree.

The control modules are objects.  A considerable amount of flexibility can
be had by replacing implementations of a control module.  For example:
* Merging of adjacent steps in the pipeline is done by replacing a control
  module and its pair of processing-step modules with a single processing-
  step module.  (Hence the possible merges are determined by the tree of
  control modules.)
* In some processing modes, a given interstep buffer need only be a "strip"
  buffer large enough to accommodate the desired data chunk sizes.  In other
  modes, a full-image buffer is needed and several passes are required.
  The control module determines which kind of buffer is used and manipulates
  virtual array buffers as needed.  One or both processing steps may be
  unaware of the multi-pass behavior.

In theory, we might be able to make all of the data buffer controllers
interchangeable and provide just one set of implementations for all.  In
practice, each one contains considerable special-case processing for its
particular job.  The buffer controller concept should be regarded as an
overall system structuring principle, not as a complete description of the
task performed by any one controller.


*** Compression object structure ***

Here is a sketch of the logical structure of the JPEG compression library in
lossy mode:

                                                 |-- Colorspace conversion
                  |-- Preprocessing controller --|
                  |                              |-- Downsampling
Main controller --|
                  |                            |-- Forward DCT, quantize
                  |-- Coefficient controller --|
                                               |-- Entropy encoding

... and in lossless mode:

                                                 |-- Colorspace conversion
                  |-- Preprocessing controller --|
                  |                              |-- Downsampling
Main controller --|
                  |                           |-- Point transform
                  |                           |
                  |-- Difference controller --|-- Prediction, differencing
                                              |
                                              |-- Lossless mode entropy
                                                  encoding

This sketch also describes the flow of control (subroutine calls) during
typical image data processing.  Each of the components shown in the diagram is
an "object" which may have several different implementations available.  One
or more source code files contain the actual implementation(s) of each object.

The objects shown above are:

* Main controller: buffer controller for the subsampled-data buffer, which
  holds the preprocessed input data.  This controller invokes preprocessing to
  fill the subsampled-data buffer, and JPEG compression to empty it.  There is
  usually no need for a full-image buffer here; a strip buffer is adequate.

* Preprocessing controller: buffer controller for the downsampling input data
  buffer, which lies between colorspace conversion and downsampling.  Note
  that a unified conversion/downsampling module would probably replace this
  controller entirely.

* Colorspace conversion: converts application image data into the desired
  JPEG color space; also changes the data from pixel-interleaved layout to
  separate component planes.  Processes one pixel row at a time.

* Downsampling: performs reduction of chroma components as required.
  Optionally may perform pixel-level smoothing as well.  Processes a "row
  group" at a time, where a row group is defined as Vmax pixel rows of each
  component before downsampling, and Vk sample rows afterwards (remember Vk
  differs across components).  Some downsampling or smoothing algorithms may
  require context rows above and below the current row group; the
  preprocessing controller is responsible for supplying these rows via proper
  buffering.  The downsampler is responsible for edge expansion at the right
  edge (i.e., extending each sample row to a multiple of 8 samples); but the
  preprocessing controller is responsible for vertical edge expansion (i.e.,
  duplicating the bottom sample row as needed to make a multiple of 8 rows).

* Coefficient controller: buffer controller for the DCT-coefficient data.
  This controller handles MCU assembly, including insertion of dummy DCT
  blocks when needed at the right or bottom edge.  When performing
  Huffman-code optimization or emitting a multiscan JPEG file, this
  controller is responsible for buffering the full image.  The equivalent of
  one fully interleaved MCU row of subsampled data is processed per call,
  even when the JPEG file is noninterleaved.

* Forward DCT and quantization: Perform DCT, quantize, and emit coefficients.
  Works on one or more DCT blocks at a time.  (Note: the coefficients are now
  emitted in normal array order, which the entropy encoder is expected to
  convert to zigzag order as necessary.  Prior versions of the IJG code did
  the conversion to zigzag order within the quantization step.)

* Entropy encoding: Perform Huffman or arithmetic entropy coding and emit the
  coded data to the data destination module.  Works on one MCU per call.
  For progressive JPEG, the same DCT blocks are fed to the entropy coder
  during each pass, and the coder must emit the appropriate subset of
  coefficients.

* Difference controller: buffer controller for the spatial difference data.
  When emitting a multiscan JPEG file, this controller is responsible for
  buffering the full image.  The equivalent of one fully interleaved MCU row
  of subsampled data is processed per call, even when the JPEG file is
  noninterleaved.

* Point transform: Downscale the data by the point transform value.

* Prediction and differencing: Calculate the predictor and subtract it
  from the input.  Works on one scanline per call.  The difference
  controller supplies the prior scanline, which is used for prediction.

* Lossless mode entropy encoding: Perform Huffman or arithmetic entropy coding
  and emit the coded data to the data destination module.  This module handles
  MCU assembly.  Works on one MCU row per call.

In addition to the above objects, the compression library includes these
objects:

* Master control: determines the number of passes required, controls overall
  and per-pass initialization of the other modules.

* Marker writing: generates JPEG markers (except for RSTn, which is emitted
  by the entropy encoder when needed).

* Data destination manager: writes the output JPEG datastream to its final
  destination (e.g., a file).  The destination manager supplied with the
  library knows how to write to a stdio stream or to a memory buffer;
  for other behaviors, the surrounding application may provide its own
  destination manager.

* Memory manager: allocates and releases memory, controls virtual arrays
  (with backing store management, where required).

* Error handler: performs formatting and output of error and trace messages;
  determines handling of nonfatal errors.  The surrounding application may
  override some or all of this object's methods to change error handling.

* Progress monitor: supports output of "percent-done" progress reports.
  This object represents an optional callback to the surrounding application:
  if wanted, it must be supplied by the application.

The error handler, destination manager, and progress monitor objects are
defined as separate objects in order to simplify application-specific
customization of the JPEG library.  A surrounding application may override
individual methods or supply its own all-new implementation of one of these
objects.  The object interfaces for these objects are therefore treated as
part of the application interface of the library, whereas the other objects
are internal to the library.

The error handler and memory manager are shared by JPEG compression and
decompression; the progress monitor, if used, may be shared as well.


*** Decompression object structure ***

Here is a sketch of the logical structure of the JPEG decompression library in
lossy mode:

                                               |-- Entropy decoding
                  |-- Coefficient controller --|
                  |                            |-- Dequantize, Inverse DCT
Main controller --|
                  |                               |-- Upsampling
                  |-- Postprocessing controller --|   |-- Colorspace conversion
                                                  |-- Color quantization
                                                  |-- Color precision reduction

... and in lossless mode:

                                              |-- Lossless mode entropy
                                              |   decoding
                                              |
                  |-- Difference controller --|-- Prediction, undifferencing
                  |                           |
                  |                           |-- Point transform, sample size
                  |                               scaling
Main controller --|
                  |                               |-- Upsampling
                  |-- Postprocessing controller --|
                                                  |-- Color precision reduction

As before, this diagram also represents typical control flow.  The objects
shown are:

* Main controller: buffer controller for the subsampled-data buffer, which
  holds the output of JPEG decompression proper.  This controller's primary
  task is to feed the postprocessing procedure.  Some upsampling algorithms
  may require context rows above and below the current row group; when this
  is true, the main controller is responsible for managing its buffer so as
  to make context rows available.  In the current design, the main buffer is
  always a strip buffer; a full-image buffer is never required.

* Coefficient controller: buffer controller for the DCT-coefficient data.
  This controller handles MCU disassembly, including deletion of any dummy
  DCT blocks at the right or bottom edge.  When reading a multiscan JPEG
  file, this controller is responsible for buffering the full image.
  (Buffering DCT coefficients, rather than samples, is necessary to support
  progressive JPEG.)  The equivalent of one fully interleaved MCU row of
  subsampled data is processed per call, even when the source JPEG file is
  noninterleaved.

* Entropy decoding: Read coded data from the data source module and perform
  Huffman or arithmetic entropy decoding.  Works on one MCU per call.
  For progressive JPEG decoding, the coefficient controller supplies the prior
  coefficients of each MCU (initially all zeroes), which the entropy decoder
  modifies in each scan.

* Dequantization and inverse DCT: like it says.  Note that the coefficients
  buffered by the coefficient controller have NOT been dequantized; we
  merge dequantization and inverse DCT into a single step for speed reasons.
  When scaled-down output is asked for, simplified DCT algorithms may be used
  that emit fewer samples per DCT block, not the full 8x8.  Works on one DCT
  block at a time.

* Difference controller: buffer controller for the spatial difference data.
  When reading a multiscan JPEG file, this controller is responsible for
  buffering the full image. The equivalent of one fully interleaved MCU row
  is processed per call, even when the source JPEG file is noninterleaved.

* Lossless mode entropy decoding: Read coded data from the data source module
  and perform Huffman or arithmetic entropy decoding.  Works on one MCU row per
  call.

* Prediction and undifferencing: Calculate the predictor and add it to the
  decoded difference.  Works on one scanline per call.  The difference
  controller supplies the prior scanline, which is used for prediction.

* Point transform and sample size scaling: Upscale the data by the point
  transform value and downscale it to fit into the compiled-in sample size.

* Postprocessing controller: buffer controller for the color quantization
  input buffer, when quantization is in use.  (Without quantization, this
  controller just calls the upsampler.)  For two-pass quantization, this
  controller is responsible for buffering the full-image data.

* Upsampling: restores chroma components to full size.  (May support more
  general output rescaling, too.  Note that if undersized DCT outputs have
  been emitted by the DCT module, this module must adjust so that properly
  sized outputs are created.)  Works on one row group at a time.  This module
  also calls the color conversion module, so its top level is effectively a
  buffer controller for the upsampling->color conversion buffer.  However, in
  all but the highest-quality operating modes, upsampling and color
  conversion are likely to be merged into a single step.

* Colorspace conversion: convert from JPEG color space to output color space,
  and change data layout from separate component planes to pixel-interleaved.
  Works on one pixel row at a time.

* Color quantization: reduce the data to colormapped form, using either an
  externally specified colormap or an internally generated one.  This module
  is not used for full-color output.  Works on one pixel row at a time; may
  require two passes to generate a color map.  Note that the output will
  always be a single component representing colormap indexes.  In the current
  design, the output values are JSAMPLEs, J12SAMPLEs, or J16SAMPLEs, so the
  library cannot quantize to more than 256 colors when using 8-bit data
  precision.  This is unlikely to be a problem in practice.

* Color reduction: this module handles color precision reduction, e.g.,
  generating 15-bit color (5 bits/primary) from JPEG's 24-bit output.
  Not quite clear yet how this should be handled... should we merge it with
  colorspace conversion???

Note that some high-speed operating modes might condense the entire
postprocessing sequence to a single module (upsample, color convert, and
quantize in one step).

In addition to the above objects, the decompression library includes these
objects:

* Master control: determines the number of passes required, controls overall
  and per-pass initialization of the other modules.  This is subdivided into
  input and output control: jdinput.c controls only input-side processing,
  while jdmaster.c handles overall initialization and output-side control.

* Marker reading: decodes JPEG markers (except for RSTn).

* Data source manager: supplies the input JPEG datastream.  The source
  manager supplied with the library knows how to read from a stdio stream
  or from a memory buffer;  for other behaviors, the surrounding application
  may provide its own source manager.

* Memory manager: same as for compression library.

* Error handler: same as for compression library.

* Progress monitor: same as for compression library.

As with compression, the data source manager, error handler, and progress
monitor are candidates for replacement by a surrounding application.


*** Decompression input and output separation ***

To support efficient incremental display of progressive JPEG files, the
decompressor is divided into two sections that can run independently:

1. Data input includes marker parsing, entropy decoding, and input into the
   coefficient controller's DCT coefficient buffer.  Note that this
   processing is relatively cheap and fast.

2. Data output reads from the DCT coefficient buffer and performs the IDCT
   and all postprocessing steps.

For a progressive JPEG file, the data input processing is allowed to get
arbitrarily far ahead of the data output processing.  (This occurs only
if the application calls jpeg_consume_input(); otherwise input and output
run in lockstep, since the input section is called only when the output
section needs more data.)  In this way the application can avoid making
extra display passes when data is arriving faster than the display pass
can run.  Furthermore, it is possible to abort an output pass without
losing anything, since the coefficient buffer is read-only as far as the
output section is concerned.  See libjpeg.txt for more detail.

A full-image coefficient array is only created if the JPEG file has multiple
scans (or if the application specifies buffered-image mode anyway).  When
reading a single-scan file, the coefficient controller normally creates only
a one-MCU buffer, so input and output processing must run in lockstep in this
case.  jpeg_consume_input() is effectively a no-op in this situation.

The main impact of dividing the decompressor in this fashion is that we must
be very careful with shared variables in the cinfo data structure.  Each
variable that can change during the course of decompression must be
classified as belonging to data input or data output, and each section must
look only at its own variables.  For example, the data output section may not
depend on any of the variables that describe the current scan in the JPEG
file, because these may change as the data input section advances into a new
scan.

The progress monitor is (somewhat arbitrarily) defined to treat input of the
file as one pass when buffered-image mode is not used, and to ignore data
input work completely when buffered-image mode is used.  Note that the
library has no reliable way to predict the number of passes when dealing
with a progressive JPEG file, nor can it predict the number of output passes
in buffered-image mode.  So the work estimate is inherently bogus anyway.

No comparable division is currently made in the compression library, because
there isn't any real need for it.


*** Data formats ***

Arrays of 8-bit pixel sample values use the following data structure:

    typedef something JSAMPLE;          a pixel component value, 0..MAXJSAMPLE
    typedef JSAMPLE *JSAMPROW;          ptr to a row of samples
    typedef JSAMPROW *JSAMPARRAY;       ptr to a list of rows
    typedef JSAMPARRAY *JSAMPIMAGE;     ptr to a list of color-component arrays

Arrays of 12-bit pixel sample values use the following data structure:

    typedef something J12SAMPLE;        a pixel component value, 0..MAXJ12SAMPLE
    typedef J12SAMPLE *J12SAMPROW;      ptr to a row of samples
    typedef J12SAMPROW *J12SAMPARRAY;   ptr to a list of rows
    typedef J12SAMPARRAY *J12SAMPIMAGE; ptr to a list of color-component arrays

Arrays of 16-bit pixel sample values use the following data structure:

    typedef something J16SAMPLE;        a pixel component value, 0..MAXJ16SAMPLE
    typedef J16SAMPLE *J16SAMPROW;      ptr to a row of samples
    typedef J16SAMPROW *J16SAMPARRAY;   ptr to a list of rows
    typedef J16SAMPARRAY *J16SAMPIMAGE; ptr to a list of color-component arrays

The basic element type JSAMPLE (8-bit sample) will be unsigned char, the basic
element type J12SAMPLE (12-bit sample) will be short, and the basic element
type J16SAMPLE (16-bit sample) will be unsigned short.

With these conventions, J*SAMPLE values can be assumed to be >= 0.  This helps
simplify correct rounding during downsampling, etc.  The JPEG standard's
specification that 8-bit sample values run from -128..127 is accommodated by
subtracting 128 from the sample value in the DCT step.  Similarly, during
decompression the output of the IDCT step will be immediately shifted back to
0..255.  (NOTE: different values are required when 12-bit samples are in use.
When 8-bit samples are in use, the code uses MAXJSAMPLE and CENTERJSAMPLE,
which are defined as 255 and 128 respectively.  When 12-bit samples are in use,
the code uses MAXJ12SAMPLE and CENTERJ12SAMPLE, which are defined as 4095 and
2048 respectively.  When 16-bit samples are in use, the code uses MAXJ16SAMPLE
and CENTERJ16SAMPLE, which are defined as 65535 and 32768 respectively.)

We use a pointer per row, rather than a two-dimensional J*SAMPLE array.  This
choice costs only a small amount of memory and has several benefits:
* Code using the data structure doesn't need to know the allocated width of
  the rows.  This simplifies edge expansion/compression, since we can work
  in an array that's wider than the logical picture width.
* Indexing doesn't require multiplication; this is a performance win on many
  machines.
* Arrays with more than 64K total elements can be supported even on machines
  where malloc() cannot allocate chunks larger than 64K.
* The rows forming a component array may be allocated at different times
  without extra copying.  This trick allows some speedups in smoothing steps
  that need access to the previous and next rows.

Note that each color component is stored in a separate array; we don't use the
traditional layout in which the components of a pixel are stored together.
This simplifies coding of modules that work on each component independently,
because they don't need to know how many components there are.  Furthermore,
we can read or write each component to a temporary file independently, which
is helpful when dealing with noninterleaved JPEG files.

In general, a specific sample value is accessed by code such as
        image[colorcomponent][row][col]
where col is measured from the image left edge, but row is measured from the
first sample row currently in memory.  Either of the first two indexings can
be precomputed by copying the relevant pointer.


Since most image-processing applications prefer to work on images in which
the components of a pixel are stored together, the data passed to or from the
surrounding application uses the traditional convention: a single pixel is
represented by N consecutive J*SAMPLE values, and an image row is an array of
(# of color components)*(image width) J*SAMPLEs.  One or more rows of data can
be represented by a pointer of type J*SAMPARRAY in this scheme.  This scheme is
converted to component-wise storage inside the JPEG library.  (Applications
that want to skip JPEG preprocessing or postprocessing will have to contend
with component-wise storage.)


Arrays of DCT-coefficient values use the following data structure:

    typedef short JCOEF;                a 16-bit signed integer
    typedef JCOEF JBLOCK[DCTSIZE2];     an 8x8 block of coefficients
    typedef JBLOCK *JBLOCKROW;          ptr to one horizontal row of 8x8 blocks
    typedef JBLOCKROW *JBLOCKARRAY;     ptr to a list of such rows
    typedef JBLOCKARRAY *JBLOCKIMAGE;   ptr to a list of color component arrays

The underlying type is at least a 16-bit signed integer; while "short" is big
enough on all machines of interest, on some machines it is preferable to use
"int" for speed reasons, despite the storage cost.  Coefficients are grouped
into 8x8 blocks (but we always use #defines DCTSIZE and DCTSIZE2 rather than
"8" and "64").

The contents of a coefficient block may be in either "natural" or zigzagged
order, and may be true values or divided by the quantization coefficients,
depending on where the block is in the processing pipeline.  In the current
library, coefficient blocks are kept in natural order everywhere; the entropy
codecs zigzag or dezigzag the data as it is written or read.  The blocks
contain quantized coefficients everywhere outside the DCT/IDCT subsystems.
(This latter decision may need to be revisited to support variable
quantization a la JPEG Part 3.)

Notice that the allocation unit is now a row of 8x8 blocks, corresponding to
eight rows of samples.  Otherwise the structure is much the same as for
samples, and for the same reasons.


*** Suspendable processing ***

In some applications it is desirable to use the JPEG library as an
incremental, memory-to-memory filter.  In this situation the data source or
destination may be a limited-size buffer, and we can't rely on being able to
empty or refill the buffer at arbitrary times.  Instead the application would
like to have control return from the library at buffer overflow/underrun, and
then resume compression or decompression at a later time.

This scenario is supported for simple cases.  (For anything more complex, we
recommend that the application "bite the bullet" and develop real multitasking
capability.)  The libjpeg.txt file goes into more detail about the usage and
limitations of this capability; here we address the implications for library
structure.

The essence of the problem is that the entropy codec (coder or decoder) must
be prepared to stop at arbitrary times.  In turn, the controllers that call
the entropy codec must be able to stop before having produced or consumed all
the data that they normally would handle in one call.  That part is reasonably
straightforward: we make the controller call interfaces include "progress
counters" which indicate the number of data chunks successfully processed, and
we require callers to test the counter rather than just assume all of the data
was processed.

Rather than trying to restart at an arbitrary point, the current Huffman
codecs are designed to restart at the beginning of the current MCU after a
suspension due to buffer overflow/underrun.  At the start of each call, the
codec's internal state is loaded from permanent storage (in the JPEG object
structures) into local variables.  On successful completion of the MCU, the
permanent state is updated.  (This copying is not very expensive, and may even
lead to *improved* performance if the local variables can be registerized.)
If a suspension occurs, the codec simply returns without updating the state,
thus effectively reverting to the start of the MCU.  Note that this implies
leaving some data unprocessed in the source/destination buffer (ie, the
compressed partial MCU).  The data source/destination module interfaces are
specified so as to make this possible.  This also implies that the data buffer
must be large enough to hold a worst-case compressed MCU; a couple thousand
bytes should be enough.

In a successive-approximation AC refinement scan, the progressive Huffman
decoder has to be able to undo assignments of newly nonzero coefficients if it
suspends before the MCU is complete, since decoding requires distinguishing
previously-zero and previously-nonzero coefficients.  This is a bit tedious
but probably won't have much effect on performance.  Other variants of Huffman
decoding need not worry about this, since they will just store the same values
again if forced to repeat the MCU.

This approach would probably not work for an arithmetic codec, since its
modifiable state is quite large and couldn't be copied cheaply.  Instead it
would have to suspend and resume exactly at the point of the buffer end.

The JPEG marker reader is designed to cope with suspension at an arbitrary
point.  It does so by backing up to the start of the marker parameter segment,
so the data buffer must be big enough to hold the largest marker of interest.
Again, a couple KB should be adequate.  (A special "skip" convention is used
to bypass COM and APPn markers, so these can be larger than the buffer size
without causing problems; otherwise a 64K buffer would be needed in the worst
case.)

The JPEG marker writer currently does *not* cope with suspension.
We feel that this is not necessary; it is much easier simply to require
the application to ensure there is enough buffer space before starting.  (An
empty 2K buffer is more than sufficient for the header markers; and ensuring
there are a dozen or two bytes available before calling jpeg_finish_compress()
will suffice for the trailer.)  This would not work for writing multi-scan
JPEG files, but we simply do not intend to support that capability with
suspension.


*** Memory manager services ***

The JPEG library's memory manager controls allocation and deallocation of
memory, and it manages large "virtual" data arrays on machines where the
operating system does not provide virtual memory.  Note that the same
memory manager serves both compression and decompression operations.

In all cases, allocated objects are tied to a particular compression or
decompression master record, and they will be released when that master
record is destroyed.

The memory manager does not provide explicit deallocation of objects.
Instead, objects are created in "pools" of free storage, and a whole pool
can be freed at once.  This approach helps prevent storage-leak bugs, and
it speeds up operations whenever malloc/free are slow (as they often are).
The pools can be regarded as lifetime identifiers for objects.  Two
pools/lifetimes are defined:
  * JPOOL_PERMANENT     lasts until master record is destroyed
  * JPOOL_IMAGE         lasts until done with image (JPEG datastream)
Permanent lifetime is used for parameters and tables that should be carried
across from one datastream to another; this includes all application-visible
parameters.  Image lifetime is used for everything else.  (A third lifetime,
JPOOL_PASS = one processing pass, was originally planned.  However it was
dropped as not being worthwhile.  The actual usage patterns are such that the
peak memory usage would be about the same anyway; and having per-pass storage
substantially complicates the virtual memory allocation rules --- see below.)

The memory manager deals with three kinds of object:
1. "Small" objects.  Typically these require no more than 10K-20K total.
2. "Large" objects.  These may require tens to hundreds of K depending on
   image size.  Semantically they behave the same as small objects, but we
   distinguish them because pool allocation heuristics may differ for large and
   small objects (historically, large objects were also referenced by far
   pointers on MS-DOS machines.)  Note that individual "large" objects cannot
   exceed the size allowed by type size_t, which may be 64K or less on some
   machines.
3. "Virtual" objects.  These are large 2-D arrays of J*SAMPLEs or JBLOCKs
   (typically large enough for the entire image being processed).  The
   memory manager provides stripwise access to these arrays.  On machines
   without virtual memory, the rest of the array may be swapped out to a
   temporary file.

(Note: J*SAMPARRAY and JBLOCKARRAY data structures are a combination of large
objects for the data proper and small objects for the row pointers.  For
convenience and speed, the memory manager provides single routines to create
these structures.  Similarly, virtual arrays include a small control block
and a J*SAMPARRAY or JBLOCKARRAY working buffer, all created with one call.)

In the present implementation, virtual arrays are only permitted to have image
lifespan.  (Permanent lifespan would not be reasonable, and pass lifespan is
not very useful since a virtual array's raison d'etre is to store data for
multiple passes through the image.)  We also expect that only "small" objects
will be given permanent lifespan, though this restriction is not required by
the memory manager.

In a non-virtual-memory machine, some performance benefit can be gained by
making the in-memory buffers for virtual arrays be as large as possible.
(For small images, the buffers might fit entirely in memory, so blind
swapping would be very wasteful.)  The memory manager will adjust the height
of the buffers to fit within a prespecified maximum memory usage.  In order
to do this in a reasonably optimal fashion, the manager needs to allocate all
of the virtual arrays at once.  Therefore, there isn't a one-step allocation
routine for virtual arrays; instead, there is a "request" routine that simply
allocates the control block, and a "realize" routine (called just once) that
determines space allocation and creates all of the actual buffers.  The
realize routine must allow for space occupied by non-virtual large objects.
(We don't bother to factor in the space needed for small objects, on the
grounds that it isn't worth the trouble.)

To support all this, we establish the following protocol for doing business
with the memory manager:
  1. Modules must request virtual arrays (which may have only image lifespan)
     during the initial setup phase, i.e., in their jinit_xxx routines.
  2. All "large" objects (including J*SAMPARRAYs and JBLOCKARRAYs) must also be
     allocated during initial setup.
  3. realize_virt_arrays will be called at the completion of initial setup.
     The above conventions ensure that sufficient information is available
     for it to choose a good size for virtual array buffers.
Small objects of any lifespan may be allocated at any time.  We expect that
the total space used for small objects will be small enough to be negligible
in the realize_virt_arrays computation.

In a virtual-memory machine, we simply pretend that the available space is
infinite, thus causing realize_virt_arrays to decide that it can allocate all
the virtual arrays as full-size in-memory buffers.  The overhead of the
virtual-array access protocol is very small when no swapping occurs.

A virtual array can be specified to be "pre-zeroed"; when this flag is set,
never-yet-written sections of the array are set to zero before being made
available to the caller.  If this flag is not set, never-written sections
of the array contain garbage.  (This feature exists primarily because the
equivalent logic would otherwise be needed in jdcoefct.c for progressive
JPEG mode; we may as well make it available for possible other uses.)

The first write pass on a virtual array is required to occur in top-to-bottom
order; read passes, as well as any write passes after the first one, may
access the array in any order.  This restriction exists partly to simplify
the virtual array control logic, and partly because some file systems may not
support seeking beyond the current end-of-file in a temporary file.  The main
implication of this restriction is that rearrangement of rows (such as
converting top-to-bottom data order to bottom-to-top) must be handled while
reading data out of the virtual array, not while putting it in.


*** Memory manager internal structure ***

To isolate system dependencies as much as possible, we have broken the
memory manager into two parts.  There is a reasonably system-independent
"front end" (jmemmgr.c) and a "back end" that contains only the code
likely to change across systems.  All of the memory management methods
outlined above are implemented by the front end.  The back end provides
the following routines for use by the front end (none of these routines
are known to the rest of the JPEG code):

jpeg_mem_init, jpeg_mem_term    system-dependent initialization/shutdown

jpeg_get_small, jpeg_free_small interface to malloc and free library routines
                                (or their equivalents)

jpeg_get_large, jpeg_free_large historically was used to interface with
                                FAR malloc/free on MS-DOS machines;  now the
                                same as jpeg_get_small/jpeg_free_small

jpeg_mem_available              estimate available memory

jpeg_open_backing_store         create a backing-store object

read_backing_store,             manipulate a backing-store object
write_backing_store,
close_backing_store

On some systems there will be more than one type of backing-store object.
jpeg_open_backing_store is responsible for choosing how to implement a given
object.  The read/write/close routines are method pointers in the structure
that describes a given object; this lets them be different for different object
types.

It may be necessary to ensure that backing store objects are explicitly
released upon abnormal program termination.  To support this, we will expect
the main program or surrounding application to arrange to call self_destruct
(typically via jpeg_destroy) upon abnormal termination.  This may require a
SIGINT signal handler or equivalent.  We don't want to have the back end module
install its own signal handler, because that would pre-empt the surrounding
application's ability to control signal handling.

The IJG distribution includes several memory manager back end implementations.
Usually the same back end should be suitable for all applications on a given
system, but it is possible for an application to supply its own back end at
need.


*** Implications of DNL marker ***

Some JPEG files may use a DNL marker to postpone definition of the image
height (this would be useful for a fax-like scanner's output, for instance).
In these files the SOF marker claims the image height is 0, and you only
find out the true image height at the end of the first scan.

We could read these files as follows:
1. Upon seeing zero image height, replace it by 65535 (the maximum allowed).
2. When the DNL is found, update the image height in the global image
   descriptor.
This implies that control modules must avoid making copies of the image
height, and must re-test for termination after each MCU row.  This would
be easy enough to do.

In cases where image-size data structures are allocated, this approach will
result in very inefficient use of virtual memory or much-larger-than-necessary
temporary files.  This seems acceptable for something that probably won't be a
mainstream usage.  People might have to forgo use of memory-hogging options
(such as two-pass color quantization or noninterleaved JPEG files) if they
want efficient conversion of such files.  (One could improve efficiency by
demanding a user-supplied upper bound for the height, less than 65536; in most
cases it could be much less.)

The standard also permits the SOF marker to overestimate the image height,
with a DNL to give the true, smaller height at the end of the first scan.
This would solve the space problems if the overestimate wasn't too great.
However, it implies that you don't even know whether DNL will be used.

This leads to a couple of very serious objections:
1. Testing for a DNL marker must occur in the inner loop of the decompressor's
   Huffman decoder; this implies a speed penalty whether the feature is used
   or not.
2. There is no way to hide the last-minute change in image height from an
   application using the decoder.  Thus *every* application using the IJG
   library would suffer a complexity penalty whether it cared about DNL or
   not.
We currently do not support DNL because of these problems.

A different approach is to insist that DNL-using files be preprocessed by a
separate program that reads ahead to the DNL, then goes back and fixes the SOF
marker.  This is a much simpler solution and is probably far more efficient.
Even if one wants piped input, buffering the first scan of the JPEG file needs
a lot smaller temp file than is implied by the maximum-height method.  For
this approach we'd simply treat DNL as a no-op in the decompressor (at most,
check that it matches the SOF image height).

We will not worry about making the compressor capable of outputting DNL.
Something similar to the first scheme above could be applied if anyone ever
wants to make that work.