1
IJG JPEG LIBRARY: SYSTEM ARCHITECTURE
3
Copyright (C) 1991-2009, Thomas G. Lane, Guido Vollbeding.
4
This file is part of the Independent JPEG Group's software.
5
For conditions of distribution and use, see the accompanying README file.
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.
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.txt 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.txt describes the coding style conventions we use.
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
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
35
*** System features ***
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.
45
We desire the library to be capable of supporting all JPEG baseline, extended
46
sequential, and progressive DCT processes. Hierarchical processes are not
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.
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.
63
By itself, the library handles only interchange JPEG datastreams --- in
64
particular the widely used JFIF file format. The library can be used by
65
surrounding code to process interchange or abbreviated JPEG datastreams that
66
are embedded in more complex file formats. (For example, libtiff uses this
67
library to implement JPEG compression within the TIFF file format.)
69
The library includes a substantial amount of code that is not covered by the
70
JPEG standard but is necessary for typical applications of JPEG. These
71
functions preprocess the image before JPEG compression or postprocess it after
72
decompression. They include colorspace conversion, downsampling/upsampling,
73
and color quantization. This code can be omitted if not needed.
75
A wide range of quality vs. speed tradeoffs are possible in JPEG processing,
76
and even more so in decompression postprocessing. The decompression library
77
provides multiple implementations that cover most of the useful tradeoffs,
78
ranging from very-high-quality down to fast-preview operation. On the
79
compression side we have generally not provided low-quality choices, since
80
compression is normally less time-critical. It should be understood that the
81
low-quality modes may not meet the JPEG standard's accuracy requirements;
82
nonetheless, they are useful for viewers.
85
*** Portability issues ***
87
Portability is an essential requirement for the library. The key portability
88
issues that show up at the level of system architecture are:
90
1. Memory usage. We want the code to be able to run on PC-class machines
91
with limited memory. Images should therefore be processed sequentially (in
92
strips), to avoid holding the whole image in memory at once. Where a
93
full-image buffer is necessary, we should be able to use either virtual memory
96
2. Near/far pointer distinction. To run efficiently on 80x86 machines, the
97
code should distinguish "small" objects (kept in near data space) from
98
"large" ones (kept in far data space). This is an annoying restriction, but
99
fortunately it does not impact code quality for less brain-damaged machines,
100
and the source code clutter turns out to be minimal with sufficient use of
103
3. Data precision. We assume that "char" is at least 8 bits, "short" and
104
"int" at least 16, "long" at least 32. The code will work fine with larger
105
data sizes, although memory may be used inefficiently in some cases. However,
106
the JPEG compressed datastream must ultimately appear on external storage as a
107
sequence of 8-bit bytes if it is to conform to the standard. This may pose a
108
problem on machines where char is wider than 8 bits. The library represents
109
compressed data as an array of values of typedef JOCTET. If no data type
110
exactly 8 bits wide is available, custom data source and data destination
111
modules must be written to unpack and pack the chosen JOCTET datatype into
112
8-bit external representation.
115
*** System overview ***
117
The compressor and decompressor are each divided into two main sections:
118
the JPEG compressor or decompressor proper, and the preprocessing or
119
postprocessing functions. The interface between these two sections is the
120
image data that the official JPEG spec regards as its input or output: this
121
data is in the colorspace to be used for compression, and it is downsampled
122
to the sampling factors to be used. The preprocessing and postprocessing
123
steps are responsible for converting a normal image representation to or from
124
this form. (Those few applications that want to deal with YCbCr downsampled
125
data can skip the preprocessing or postprocessing step.)
127
Looking more closely, the compressor library contains the following main
131
* Color space conversion (e.g., RGB to YCbCr).
132
* Edge expansion and downsampling. Optionally, this step can do simple
133
smoothing --- this is often helpful for low-quality source data.
135
* MCU assembly, DCT, quantization.
136
* Entropy coding (sequential or progressive, Huffman or arithmetic).
138
In addition to these modules we need overall control, marker generation,
139
and support code (memory management & error handling). There is also a
140
module responsible for physically writing the output data --- typically
141
this is just an interface to fwrite(), but some applications may need to
142
do something else with the data.
144
The decompressor library contains the following main elements:
147
* Entropy decoding (sequential or progressive, Huffman or arithmetic).
148
* Dequantization, inverse DCT, MCU disassembly.
150
* Upsampling. Optionally, this step may be able to do more general
151
rescaling of the image.
152
* Color space conversion (e.g., YCbCr to RGB). This step may also
153
provide gamma adjustment [ currently it does not ].
154
* Optional color quantization (e.g., reduction to 256 colors).
155
* Optional color precision reduction (e.g., 24-bit to 15-bit color).
156
[This feature is not currently implemented.]
158
We also need overall control, marker parsing, and a data source module.
159
The support code (memory management & error handling) can be shared with
160
the compression half of the library.
162
There may be several implementations of each of these elements, particularly
163
in the decompressor, where a wide range of speed/quality tradeoffs is very
164
useful. It must be understood that some of the best speedups involve
165
merging adjacent steps in the pipeline. For example, upsampling, color space
166
conversion, and color quantization might all be done at once when using a
167
low-quality ordered-dither technique. The system architecture is designed to
168
allow such merging where appropriate.
171
Note: it is convenient to regard edge expansion (padding to block boundaries)
172
as a preprocessing/postprocessing function, even though the JPEG spec includes
173
it in compression/decompression. We do this because downsampling/upsampling
174
can be simplified a little if they work on padded data: it's not necessary to
175
have special cases at the right and bottom edges. Therefore the interface
176
buffer is always an integral number of blocks wide and high, and we expect
177
compression preprocessing to pad the source data properly. Padding will occur
178
only to the next block (8-sample) boundary. In an interleaved-scan situation,
179
additional dummy blocks may be used to fill out MCUs, but the MCU assembly and
180
disassembly logic will create or discard these blocks internally. (This is
181
advantageous for speed reasons, since we avoid DCTing the dummy blocks.
182
It also permits a small reduction in file size, because the compressor can
183
choose dummy block contents so as to minimize their size in compressed form.
184
Finally, it makes the interface buffer specification independent of whether
185
the file is actually interleaved or not.) Applications that wish to deal
186
directly with the downsampled data must provide similar buffering and padding
187
for odd-sized images.
190
*** Poor man's object-oriented programming ***
192
It should be clear by now that we have a lot of quasi-independent processing
193
steps, many of which have several possible behaviors. To avoid cluttering the
194
code with lots of switch statements, we use a simple form of object-style
195
programming to separate out the different possibilities.
197
For example, two different color quantization algorithms could be implemented
198
as two separate modules that present the same external interface; at runtime,
199
the calling code will access the proper module indirectly through an "object".
201
We can get the limited features we need while staying within portable C.
202
The basic tool is a function pointer. An "object" is just a struct
203
containing one or more function pointer fields, each of which corresponds to
204
a method name in real object-oriented languages. During initialization we
205
fill in the function pointers with references to whichever module we have
206
determined we need to use in this run. Then invocation of the module is done
207
by indirecting through a function pointer; on most machines this is no more
208
expensive than a switch statement, which would be the only other way of
209
making the required run-time choice. The really significant benefit, of
210
course, is keeping the source code clean and well structured.
212
We can also arrange to have private storage that varies between different
213
implementations of the same kind of object. We do this by making all the
214
module-specific object structs be separately allocated entities, which will
215
be accessed via pointers in the master compression or decompression struct.
216
The "public" fields or methods for a given kind of object are specified by
217
a commonly known struct. But a module's initialization code can allocate
218
a larger struct that contains the common struct as its first member, plus
219
additional private fields. With appropriate pointer casting, the module's
220
internal functions can access these private fields. (For a simple example,
221
see jdatadst.c, which implements the external interface specified by struct
222
jpeg_destination_mgr, but adds extra fields.)
224
(Of course this would all be a lot easier if we were using C++, but we are
225
not yet prepared to assume that everyone has a C++ compiler.)
227
An important benefit of this scheme is that it is easy to provide multiple
228
versions of any method, each tuned to a particular case. While a lot of
229
precalculation might be done to select an optimal implementation of a method,
230
the cost per invocation is constant. For example, the upsampling step might
231
have a "generic" method, plus one or more "hardwired" methods for the most
232
popular sampling factors; the hardwired methods would be faster because they'd
233
use straight-line code instead of for-loops. The cost to determine which
234
method to use is paid only once, at startup, and the selection criteria are
235
hidden from the callers of the method.
237
This plan differs a little bit from usual object-oriented structures, in that
238
only one instance of each object class will exist during execution. The
239
reason for having the class structure is that on different runs we may create
240
different instances (choose to execute different modules). You can think of
241
the term "method" as denoting the common interface presented by a particular
242
set of interchangeable functions, and "object" as denoting a group of related
243
methods, or the total shared interface behavior of a group of modules.
246
*** Overall control structure ***
248
We previously mentioned the need for overall control logic in the compression
249
and decompression libraries. In IJG implementations prior to v5, overall
250
control was mostly provided by "pipeline control" modules, which proved to be
251
large, unwieldy, and hard to understand. To improve the situation, the
252
control logic has been subdivided into multiple modules. The control modules
255
1. Master control for module selection and initialization. This has two
258
1A. Startup initialization at the beginning of image processing.
259
The individual processing modules to be used in this run are selected
260
and given initialization calls.
262
1B. Per-pass control. This determines how many passes will be performed
263
and calls each active processing module to configure itself
264
appropriately at the beginning of each pass. End-of-pass processing,
265
where necessary, is also invoked from the master control module.
267
Method selection is partially distributed, in that a particular processing
268
module may contain several possible implementations of a particular method,
269
which it will select among when given its initialization call. The master
270
control code need only be concerned with decisions that affect more than
273
2. Data buffering control. A separate control module exists for each
274
inter-processing-step data buffer. This module is responsible for
275
invoking the processing steps that write or read that data buffer.
277
Each buffer controller sees the world as follows:
279
input data => processing step A => buffer => processing step B => output data
281
------------------ controller ------------------
283
The controller knows the dataflow requirements of steps A and B: how much data
284
they want to accept in one chunk and how much they output in one chunk. Its
285
function is to manage its buffer and call A and B at the proper times.
287
A data buffer control module may itself be viewed as a processing step by a
288
higher-level control module; thus the control modules form a binary tree with
289
elementary processing steps at the leaves of the tree.
291
The control modules are objects. A considerable amount of flexibility can
292
be had by replacing implementations of a control module. For example:
293
* Merging of adjacent steps in the pipeline is done by replacing a control
294
module and its pair of processing-step modules with a single processing-
295
step module. (Hence the possible merges are determined by the tree of
297
* In some processing modes, a given interstep buffer need only be a "strip"
298
buffer large enough to accommodate the desired data chunk sizes. In other
299
modes, a full-image buffer is needed and several passes are required.
300
The control module determines which kind of buffer is used and manipulates
301
virtual array buffers as needed. One or both processing steps may be
302
unaware of the multi-pass behavior.
304
In theory, we might be able to make all of the data buffer controllers
305
interchangeable and provide just one set of implementations for all. In
306
practice, each one contains considerable special-case processing for its
307
particular job. The buffer controller concept should be regarded as an
308
overall system structuring principle, not as a complete description of the
309
task performed by any one controller.
312
*** Compression object structure ***
314
Here is a sketch of the logical structure of the JPEG compression library:
316
|-- Colorspace conversion
317
|-- Preprocessing controller --|
320
| |-- Forward DCT, quantize
321
|-- Coefficient controller --|
324
This sketch also describes the flow of control (subroutine calls) during
325
typical image data processing. Each of the components shown in the diagram is
326
an "object" which may have several different implementations available. One
327
or more source code files contain the actual implementation(s) of each object.
329
The objects shown above are:
331
* Main controller: buffer controller for the subsampled-data buffer, which
332
holds the preprocessed input data. This controller invokes preprocessing to
333
fill the subsampled-data buffer, and JPEG compression to empty it. There is
334
usually no need for a full-image buffer here; a strip buffer is adequate.
336
* Preprocessing controller: buffer controller for the downsampling input data
337
buffer, which lies between colorspace conversion and downsampling. Note
338
that a unified conversion/downsampling module would probably replace this
341
* Colorspace conversion: converts application image data into the desired
342
JPEG color space; also changes the data from pixel-interleaved layout to
343
separate component planes. Processes one pixel row at a time.
345
* Downsampling: performs reduction of chroma components as required.
346
Optionally may perform pixel-level smoothing as well. Processes a "row
347
group" at a time, where a row group is defined as Vmax pixel rows of each
348
component before downsampling, and Vk sample rows afterwards (remember Vk
349
differs across components). Some downsampling or smoothing algorithms may
350
require context rows above and below the current row group; the
351
preprocessing controller is responsible for supplying these rows via proper
352
buffering. The downsampler is responsible for edge expansion at the right
353
edge (i.e., extending each sample row to a multiple of 8 samples); but the
354
preprocessing controller is responsible for vertical edge expansion (i.e.,
355
duplicating the bottom sample row as needed to make a multiple of 8 rows).
357
* Coefficient controller: buffer controller for the DCT-coefficient data.
358
This controller handles MCU assembly, including insertion of dummy DCT
359
blocks when needed at the right or bottom edge. When performing
360
Huffman-code optimization or emitting a multiscan JPEG file, this
361
controller is responsible for buffering the full image. The equivalent of
362
one fully interleaved MCU row of subsampled data is processed per call,
363
even when the JPEG file is noninterleaved.
365
* Forward DCT and quantization: Perform DCT, quantize, and emit coefficients.
366
Works on one or more DCT blocks at a time. (Note: the coefficients are now
367
emitted in normal array order, which the entropy encoder is expected to
368
convert to zigzag order as necessary. Prior versions of the IJG code did
369
the conversion to zigzag order within the quantization step.)
371
* Entropy encoding: Perform Huffman or arithmetic entropy coding and emit the
372
coded data to the data destination module. Works on one MCU per call.
373
For progressive JPEG, the same DCT blocks are fed to the entropy coder
374
during each pass, and the coder must emit the appropriate subset of
377
In addition to the above objects, the compression library includes these
380
* Master control: determines the number of passes required, controls overall
381
and per-pass initialization of the other modules.
383
* Marker writing: generates JPEG markers (except for RSTn, which is emitted
384
by the entropy encoder when needed).
386
* Data destination manager: writes the output JPEG datastream to its final
387
destination (e.g., a file). The destination manager supplied with the
388
library knows how to write to a stdio stream; for other behaviors, the
389
surrounding application may provide its own destination manager.
391
* Memory manager: allocates and releases memory, controls virtual arrays
392
(with backing store management, where required).
394
* Error handler: performs formatting and output of error and trace messages;
395
determines handling of nonfatal errors. The surrounding application may
396
override some or all of this object's methods to change error handling.
398
* Progress monitor: supports output of "percent-done" progress reports.
399
This object represents an optional callback to the surrounding application:
400
if wanted, it must be supplied by the application.
402
The error handler, destination manager, and progress monitor objects are
403
defined as separate objects in order to simplify application-specific
404
customization of the JPEG library. A surrounding application may override
405
individual methods or supply its own all-new implementation of one of these
406
objects. The object interfaces for these objects are therefore treated as
407
part of the application interface of the library, whereas the other objects
408
are internal to the library.
410
The error handler and memory manager are shared by JPEG compression and
411
decompression; the progress monitor, if used, may be shared as well.
414
*** Decompression object structure ***
416
Here is a sketch of the logical structure of the JPEG decompression library:
419
|-- Coefficient controller --|
420
| |-- Dequantize, Inverse DCT
423
|-- Postprocessing controller --| |-- Colorspace conversion
424
|-- Color quantization
425
|-- Color precision reduction
427
As before, this diagram also represents typical control flow. The objects
430
* Main controller: buffer controller for the subsampled-data buffer, which
431
holds the output of JPEG decompression proper. This controller's primary
432
task is to feed the postprocessing procedure. Some upsampling algorithms
433
may require context rows above and below the current row group; when this
434
is true, the main controller is responsible for managing its buffer so as
435
to make context rows available. In the current design, the main buffer is
436
always a strip buffer; a full-image buffer is never required.
438
* Coefficient controller: buffer controller for the DCT-coefficient data.
439
This controller handles MCU disassembly, including deletion of any dummy
440
DCT blocks at the right or bottom edge. When reading a multiscan JPEG
441
file, this controller is responsible for buffering the full image.
442
(Buffering DCT coefficients, rather than samples, is necessary to support
443
progressive JPEG.) The equivalent of one fully interleaved MCU row of
444
subsampled data is processed per call, even when the source JPEG file is
447
* Entropy decoding: Read coded data from the data source module and perform
448
Huffman or arithmetic entropy decoding. Works on one MCU per call.
449
For progressive JPEG decoding, the coefficient controller supplies the prior
450
coefficients of each MCU (initially all zeroes), which the entropy decoder
451
modifies in each scan.
453
* Dequantization and inverse DCT: like it says. Note that the coefficients
454
buffered by the coefficient controller have NOT been dequantized; we
455
merge dequantization and inverse DCT into a single step for speed reasons.
456
When scaled-down output is asked for, simplified DCT algorithms may be used
457
that need fewer coefficients and emit fewer samples per DCT block, not the
458
full 8x8. Works on one DCT block at a time.
460
* Postprocessing controller: buffer controller for the color quantization
461
input buffer, when quantization is in use. (Without quantization, this
462
controller just calls the upsampler.) For two-pass quantization, this
463
controller is responsible for buffering the full-image data.
465
* Upsampling: restores chroma components to full size. (May support more
466
general output rescaling, too. Note that if undersized DCT outputs have
467
been emitted by the DCT module, this module must adjust so that properly
468
sized outputs are created.) Works on one row group at a time. This module
469
also calls the color conversion module, so its top level is effectively a
470
buffer controller for the upsampling->color conversion buffer. However, in
471
all but the highest-quality operating modes, upsampling and color
472
conversion are likely to be merged into a single step.
474
* Colorspace conversion: convert from JPEG color space to output color space,
475
and change data layout from separate component planes to pixel-interleaved.
476
Works on one pixel row at a time.
478
* Color quantization: reduce the data to colormapped form, using either an
479
externally specified colormap or an internally generated one. This module
480
is not used for full-color output. Works on one pixel row at a time; may
481
require two passes to generate a color map. Note that the output will
482
always be a single component representing colormap indexes. In the current
483
design, the output values are JSAMPLEs, so an 8-bit compilation cannot
484
quantize to more than 256 colors. This is unlikely to be a problem in
487
* Color reduction: this module handles color precision reduction, e.g.,
488
generating 15-bit color (5 bits/primary) from JPEG's 24-bit output.
489
Not quite clear yet how this should be handled... should we merge it with
490
colorspace conversion???
492
Note that some high-speed operating modes might condense the entire
493
postprocessing sequence to a single module (upsample, color convert, and
494
quantize in one step).
496
In addition to the above objects, the decompression library includes these
499
* Master control: determines the number of passes required, controls overall
500
and per-pass initialization of the other modules. This is subdivided into
501
input and output control: jdinput.c controls only input-side processing,
502
while jdmaster.c handles overall initialization and output-side control.
504
* Marker reading: decodes JPEG markers (except for RSTn).
506
* Data source manager: supplies the input JPEG datastream. The source
507
manager supplied with the library knows how to read from a stdio stream;
508
for other behaviors, the surrounding application may provide its own source
511
* Memory manager: same as for compression library.
513
* Error handler: same as for compression library.
515
* Progress monitor: same as for compression library.
517
As with compression, the data source manager, error handler, and progress
518
monitor are candidates for replacement by a surrounding application.
521
*** Decompression input and output separation ***
523
To support efficient incremental display of progressive JPEG files, the
524
decompressor is divided into two sections that can run independently:
526
1. Data input includes marker parsing, entropy decoding, and input into the
527
coefficient controller's DCT coefficient buffer. Note that this
528
processing is relatively cheap and fast.
530
2. Data output reads from the DCT coefficient buffer and performs the IDCT
531
and all postprocessing steps.
533
For a progressive JPEG file, the data input processing is allowed to get
534
arbitrarily far ahead of the data output processing. (This occurs only
535
if the application calls jpeg_consume_input(); otherwise input and output
536
run in lockstep, since the input section is called only when the output
537
section needs more data.) In this way the application can avoid making
538
extra display passes when data is arriving faster than the display pass
539
can run. Furthermore, it is possible to abort an output pass without
540
losing anything, since the coefficient buffer is read-only as far as the
541
output section is concerned. See libjpeg.txt for more detail.
543
A full-image coefficient array is only created if the JPEG file has multiple
544
scans (or if the application specifies buffered-image mode anyway). When
545
reading a single-scan file, the coefficient controller normally creates only
546
a one-MCU buffer, so input and output processing must run in lockstep in this
547
case. jpeg_consume_input() is effectively a no-op in this situation.
549
The main impact of dividing the decompressor in this fashion is that we must
550
be very careful with shared variables in the cinfo data structure. Each
551
variable that can change during the course of decompression must be
552
classified as belonging to data input or data output, and each section must
553
look only at its own variables. For example, the data output section may not
554
depend on any of the variables that describe the current scan in the JPEG
555
file, because these may change as the data input section advances into a new
558
The progress monitor is (somewhat arbitrarily) defined to treat input of the
559
file as one pass when buffered-image mode is not used, and to ignore data
560
input work completely when buffered-image mode is used. Note that the
561
library has no reliable way to predict the number of passes when dealing
562
with a progressive JPEG file, nor can it predict the number of output passes
563
in buffered-image mode. So the work estimate is inherently bogus anyway.
565
No comparable division is currently made in the compression library, because
566
there isn't any real need for it.
571
Arrays of pixel sample values use the following data structure:
573
typedef something JSAMPLE; a pixel component value, 0..MAXJSAMPLE
574
typedef JSAMPLE *JSAMPROW; ptr to a row of samples
575
typedef JSAMPROW *JSAMPARRAY; ptr to a list of rows
576
typedef JSAMPARRAY *JSAMPIMAGE; ptr to a list of color-component arrays
578
The basic element type JSAMPLE will typically be one of unsigned char,
579
(signed) char, or short. Short will be used if samples wider than 8 bits are
580
to be supported (this is a compile-time option). Otherwise, unsigned char is
581
used if possible. If the compiler only supports signed chars, then it is
582
necessary to mask off the value when reading. Thus, all reads of JSAMPLE
583
values must be coded as "GETJSAMPLE(value)", where the macro will be defined
584
as "((value) & 0xFF)" on signed-char machines and "((int) (value))" elsewhere.
586
With these conventions, JSAMPLE values can be assumed to be >= 0. This helps
587
simplify correct rounding during downsampling, etc. The JPEG standard's
588
specification that sample values run from -128..127 is accommodated by
589
subtracting 128 from the sample value in the DCT step. Similarly, during
590
decompression the output of the IDCT step will be immediately shifted back to
591
0..255. (NB: different values are required when 12-bit samples are in use.
592
The code is written in terms of MAXJSAMPLE and CENTERJSAMPLE, which will be
593
defined as 255 and 128 respectively in an 8-bit implementation, and as 4095
594
and 2048 in a 12-bit implementation.)
596
We use a pointer per row, rather than a two-dimensional JSAMPLE array. This
597
choice costs only a small amount of memory and has several benefits:
598
* Code using the data structure doesn't need to know the allocated width of
599
the rows. This simplifies edge expansion/compression, since we can work
600
in an array that's wider than the logical picture width.
601
* Indexing doesn't require multiplication; this is a performance win on many
603
* Arrays with more than 64K total elements can be supported even on machines
604
where malloc() cannot allocate chunks larger than 64K.
605
* The rows forming a component array may be allocated at different times
606
without extra copying. This trick allows some speedups in smoothing steps
607
that need access to the previous and next rows.
609
Note that each color component is stored in a separate array; we don't use the
610
traditional layout in which the components of a pixel are stored together.
611
This simplifies coding of modules that work on each component independently,
612
because they don't need to know how many components there are. Furthermore,
613
we can read or write each component to a temporary file independently, which
614
is helpful when dealing with noninterleaved JPEG files.
616
In general, a specific sample value is accessed by code such as
617
GETJSAMPLE(image[colorcomponent][row][col])
618
where col is measured from the image left edge, but row is measured from the
619
first sample row currently in memory. Either of the first two indexings can
620
be precomputed by copying the relevant pointer.
623
Since most image-processing applications prefer to work on images in which
624
the components of a pixel are stored together, the data passed to or from the
625
surrounding application uses the traditional convention: a single pixel is
626
represented by N consecutive JSAMPLE values, and an image row is an array of
627
(# of color components)*(image width) JSAMPLEs. One or more rows of data can
628
be represented by a pointer of type JSAMPARRAY in this scheme. This scheme is
629
converted to component-wise storage inside the JPEG library. (Applications
630
that want to skip JPEG preprocessing or postprocessing will have to contend
631
with component-wise storage.)
634
Arrays of DCT-coefficient values use the following data structure:
636
typedef short JCOEF; a 16-bit signed integer
637
typedef JCOEF JBLOCK[DCTSIZE2]; an 8x8 block of coefficients
638
typedef JBLOCK *JBLOCKROW; ptr to one horizontal row of 8x8 blocks
639
typedef JBLOCKROW *JBLOCKARRAY; ptr to a list of such rows
640
typedef JBLOCKARRAY *JBLOCKIMAGE; ptr to a list of color component arrays
642
The underlying type is at least a 16-bit signed integer; while "short" is big
643
enough on all machines of interest, on some machines it is preferable to use
644
"int" for speed reasons, despite the storage cost. Coefficients are grouped
645
into 8x8 blocks (but we always use #defines DCTSIZE and DCTSIZE2 rather than
648
The contents of a coefficient block may be in either "natural" or zigzagged
649
order, and may be true values or divided by the quantization coefficients,
650
depending on where the block is in the processing pipeline. In the current
651
library, coefficient blocks are kept in natural order everywhere; the entropy
652
codecs zigzag or dezigzag the data as it is written or read. The blocks
653
contain quantized coefficients everywhere outside the DCT/IDCT subsystems.
654
(This latter decision may need to be revisited to support variable
655
quantization a la JPEG Part 3.)
657
Notice that the allocation unit is now a row of 8x8 blocks, corresponding to
658
eight rows of samples. Otherwise the structure is much the same as for
659
samples, and for the same reasons.
661
On machines where malloc() can't handle a request bigger than 64Kb, this data
662
structure limits us to rows of less than 512 JBLOCKs, or a picture width of
663
4000+ pixels. This seems an acceptable restriction.
666
On 80x86 machines, the bottom-level pointer types (JSAMPROW and JBLOCKROW)
667
must be declared as "far" pointers, but the upper levels can be "near"
668
(implying that the pointer lists are allocated in the DS segment).
669
We use a #define symbol FAR, which expands to the "far" keyword when
670
compiling on 80x86 machines and to nothing elsewhere.
673
*** Suspendable processing ***
675
In some applications it is desirable to use the JPEG library as an
676
incremental, memory-to-memory filter. In this situation the data source or
677
destination may be a limited-size buffer, and we can't rely on being able to
678
empty or refill the buffer at arbitrary times. Instead the application would
679
like to have control return from the library at buffer overflow/underrun, and
680
then resume compression or decompression at a later time.
682
This scenario is supported for simple cases. (For anything more complex, we
683
recommend that the application "bite the bullet" and develop real multitasking
684
capability.) The libjpeg.txt file goes into more detail about the usage and
685
limitations of this capability; here we address the implications for library
688
The essence of the problem is that the entropy codec (coder or decoder) must
689
be prepared to stop at arbitrary times. In turn, the controllers that call
690
the entropy codec must be able to stop before having produced or consumed all
691
the data that they normally would handle in one call. That part is reasonably
692
straightforward: we make the controller call interfaces include "progress
693
counters" which indicate the number of data chunks successfully processed, and
694
we require callers to test the counter rather than just assume all of the data
697
Rather than trying to restart at an arbitrary point, the current Huffman
698
codecs are designed to restart at the beginning of the current MCU after a
699
suspension due to buffer overflow/underrun. At the start of each call, the
700
codec's internal state is loaded from permanent storage (in the JPEG object
701
structures) into local variables. On successful completion of the MCU, the
702
permanent state is updated. (This copying is not very expensive, and may even
703
lead to *improved* performance if the local variables can be registerized.)
704
If a suspension occurs, the codec simply returns without updating the state,
705
thus effectively reverting to the start of the MCU. Note that this implies
706
leaving some data unprocessed in the source/destination buffer (ie, the
707
compressed partial MCU). The data source/destination module interfaces are
708
specified so as to make this possible. This also implies that the data buffer
709
must be large enough to hold a worst-case compressed MCU; a couple thousand
710
bytes should be enough.
712
In a successive-approximation AC refinement scan, the progressive Huffman
713
decoder has to be able to undo assignments of newly nonzero coefficients if it
714
suspends before the MCU is complete, since decoding requires distinguishing
715
previously-zero and previously-nonzero coefficients. This is a bit tedious
716
but probably won't have much effect on performance. Other variants of Huffman
717
decoding need not worry about this, since they will just store the same values
718
again if forced to repeat the MCU.
720
This approach would probably not work for an arithmetic codec, since its
721
modifiable state is quite large and couldn't be copied cheaply. Instead it
722
would have to suspend and resume exactly at the point of the buffer end.
724
The JPEG marker reader is designed to cope with suspension at an arbitrary
725
point. It does so by backing up to the start of the marker parameter segment,
726
so the data buffer must be big enough to hold the largest marker of interest.
727
Again, a couple KB should be adequate. (A special "skip" convention is used
728
to bypass COM and APPn markers, so these can be larger than the buffer size
729
without causing problems; otherwise a 64K buffer would be needed in the worst
732
The JPEG marker writer currently does *not* cope with suspension.
733
We feel that this is not necessary; it is much easier simply to require
734
the application to ensure there is enough buffer space before starting. (An
735
empty 2K buffer is more than sufficient for the header markers; and ensuring
736
there are a dozen or two bytes available before calling jpeg_finish_compress()
737
will suffice for the trailer.) This would not work for writing multi-scan
738
JPEG files, but we simply do not intend to support that capability with
742
*** Memory manager services ***
744
The JPEG library's memory manager controls allocation and deallocation of
745
memory, and it manages large "virtual" data arrays on machines where the
746
operating system does not provide virtual memory. Note that the same
747
memory manager serves both compression and decompression operations.
749
In all cases, allocated objects are tied to a particular compression or
750
decompression master record, and they will be released when that master
753
The memory manager does not provide explicit deallocation of objects.
754
Instead, objects are created in "pools" of free storage, and a whole pool
755
can be freed at once. This approach helps prevent storage-leak bugs, and
756
it speeds up operations whenever malloc/free are slow (as they often are).
757
The pools can be regarded as lifetime identifiers for objects. Two
758
pools/lifetimes are defined:
759
* JPOOL_PERMANENT lasts until master record is destroyed
760
* JPOOL_IMAGE lasts until done with image (JPEG datastream)
761
Permanent lifetime is used for parameters and tables that should be carried
762
across from one datastream to another; this includes all application-visible
763
parameters. Image lifetime is used for everything else. (A third lifetime,
764
JPOOL_PASS = one processing pass, was originally planned. However it was
765
dropped as not being worthwhile. The actual usage patterns are such that the
766
peak memory usage would be about the same anyway; and having per-pass storage
767
substantially complicates the virtual memory allocation rules --- see below.)
769
The memory manager deals with three kinds of object:
770
1. "Small" objects. Typically these require no more than 10K-20K total.
771
2. "Large" objects. These may require tens to hundreds of K depending on
772
image size. Semantically they behave the same as small objects, but we
773
distinguish them for two reasons:
774
* On MS-DOS machines, large objects are referenced by FAR pointers,
775
small objects by NEAR pointers.
776
* Pool allocation heuristics may differ for large and small objects.
777
Note that individual "large" objects cannot exceed the size allowed by
778
type size_t, which may be 64K or less on some machines.
779
3. "Virtual" objects. These are large 2-D arrays of JSAMPLEs or JBLOCKs
780
(typically large enough for the entire image being processed). The
781
memory manager provides stripwise access to these arrays. On machines
782
without virtual memory, the rest of the array may be swapped out to a
785
(Note: JSAMPARRAY and JBLOCKARRAY data structures are a combination of large
786
objects for the data proper and small objects for the row pointers. For
787
convenience and speed, the memory manager provides single routines to create
788
these structures. Similarly, virtual arrays include a small control block
789
and a JSAMPARRAY or JBLOCKARRAY working buffer, all created with one call.)
791
In the present implementation, virtual arrays are only permitted to have image
792
lifespan. (Permanent lifespan would not be reasonable, and pass lifespan is
793
not very useful since a virtual array's raison d'etre is to store data for
794
multiple passes through the image.) We also expect that only "small" objects
795
will be given permanent lifespan, though this restriction is not required by
798
In a non-virtual-memory machine, some performance benefit can be gained by
799
making the in-memory buffers for virtual arrays be as large as possible.
800
(For small images, the buffers might fit entirely in memory, so blind
801
swapping would be very wasteful.) The memory manager will adjust the height
802
of the buffers to fit within a prespecified maximum memory usage. In order
803
to do this in a reasonably optimal fashion, the manager needs to allocate all
804
of the virtual arrays at once. Therefore, there isn't a one-step allocation
805
routine for virtual arrays; instead, there is a "request" routine that simply
806
allocates the control block, and a "realize" routine (called just once) that
807
determines space allocation and creates all of the actual buffers. The
808
realize routine must allow for space occupied by non-virtual large objects.
809
(We don't bother to factor in the space needed for small objects, on the
810
grounds that it isn't worth the trouble.)
812
To support all this, we establish the following protocol for doing business
813
with the memory manager:
814
1. Modules must request virtual arrays (which may have only image lifespan)
815
during the initial setup phase, i.e., in their jinit_xxx routines.
816
2. All "large" objects (including JSAMPARRAYs and JBLOCKARRAYs) must also be
817
allocated during initial setup.
818
3. realize_virt_arrays will be called at the completion of initial setup.
819
The above conventions ensure that sufficient information is available
820
for it to choose a good size for virtual array buffers.
821
Small objects of any lifespan may be allocated at any time. We expect that
822
the total space used for small objects will be small enough to be negligible
823
in the realize_virt_arrays computation.
825
In a virtual-memory machine, we simply pretend that the available space is
826
infinite, thus causing realize_virt_arrays to decide that it can allocate all
827
the virtual arrays as full-size in-memory buffers. The overhead of the
828
virtual-array access protocol is very small when no swapping occurs.
830
A virtual array can be specified to be "pre-zeroed"; when this flag is set,
831
never-yet-written sections of the array are set to zero before being made
832
available to the caller. If this flag is not set, never-written sections
833
of the array contain garbage. (This feature exists primarily because the
834
equivalent logic would otherwise be needed in jdcoefct.c for progressive
835
JPEG mode; we may as well make it available for possible other uses.)
837
The first write pass on a virtual array is required to occur in top-to-bottom
838
order; read passes, as well as any write passes after the first one, may
839
access the array in any order. This restriction exists partly to simplify
840
the virtual array control logic, and partly because some file systems may not
841
support seeking beyond the current end-of-file in a temporary file. The main
842
implication of this restriction is that rearrangement of rows (such as
843
converting top-to-bottom data order to bottom-to-top) must be handled while
844
reading data out of the virtual array, not while putting it in.
847
*** Memory manager internal structure ***
849
To isolate system dependencies as much as possible, we have broken the
850
memory manager into two parts. There is a reasonably system-independent
851
"front end" (jmemmgr.c) and a "back end" that contains only the code
852
likely to change across systems. All of the memory management methods
853
outlined above are implemented by the front end. The back end provides
854
the following routines for use by the front end (none of these routines
855
are known to the rest of the JPEG code):
857
jpeg_mem_init, jpeg_mem_term system-dependent initialization/shutdown
859
jpeg_get_small, jpeg_free_small interface to malloc and free library routines
860
(or their equivalents)
862
jpeg_get_large, jpeg_free_large interface to FAR malloc/free in MSDOS machines;
863
else usually the same as
864
jpeg_get_small/jpeg_free_small
866
jpeg_mem_available estimate available memory
868
jpeg_open_backing_store create a backing-store object
870
read_backing_store, manipulate a backing-store object
874
On some systems there will be more than one type of backing-store object
875
(specifically, in MS-DOS a backing store file might be an area of extended
876
memory as well as a disk file). jpeg_open_backing_store is responsible for
877
choosing how to implement a given object. The read/write/close routines
878
are method pointers in the structure that describes a given object; this
879
lets them be different for different object types.
881
It may be necessary to ensure that backing store objects are explicitly
882
released upon abnormal program termination. For example, MS-DOS won't free
883
extended memory by itself. To support this, we will expect the main program
884
or surrounding application to arrange to call self_destruct (typically via
885
jpeg_destroy) upon abnormal termination. This may require a SIGINT signal
886
handler or equivalent. We don't want to have the back end module install its
887
own signal handler, because that would pre-empt the surrounding application's
888
ability to control signal handling.
890
The IJG distribution includes several memory manager back end implementations.
891
Usually the same back end should be suitable for all applications on a given
892
system, but it is possible for an application to supply its own back end at
896
*** Implications of DNL marker ***
898
Some JPEG files may use a DNL marker to postpone definition of the image
899
height (this would be useful for a fax-like scanner's output, for instance).
900
In these files the SOF marker claims the image height is 0, and you only
901
find out the true image height at the end of the first scan.
903
We could read these files as follows:
904
1. Upon seeing zero image height, replace it by 65535 (the maximum allowed).
905
2. When the DNL is found, update the image height in the global image
907
This implies that control modules must avoid making copies of the image
908
height, and must re-test for termination after each MCU row. This would
909
be easy enough to do.
911
In cases where image-size data structures are allocated, this approach will
912
result in very inefficient use of virtual memory or much-larger-than-necessary
913
temporary files. This seems acceptable for something that probably won't be a
914
mainstream usage. People might have to forgo use of memory-hogging options
915
(such as two-pass color quantization or noninterleaved JPEG files) if they
916
want efficient conversion of such files. (One could improve efficiency by
917
demanding a user-supplied upper bound for the height, less than 65536; in most
918
cases it could be much less.)
920
The standard also permits the SOF marker to overestimate the image height,
921
with a DNL to give the true, smaller height at the end of the first scan.
922
This would solve the space problems if the overestimate wasn't too great.
923
However, it implies that you don't even know whether DNL will be used.
925
This leads to a couple of very serious objections:
926
1. Testing for a DNL marker must occur in the inner loop of the decompressor's
927
Huffman decoder; this implies a speed penalty whether the feature is used
929
2. There is no way to hide the last-minute change in image height from an
930
application using the decoder. Thus *every* application using the IJG
931
library would suffer a complexity penalty whether it cared about DNL or
933
We currently do not support DNL because of these problems.
935
A different approach is to insist that DNL-using files be preprocessed by a
936
separate program that reads ahead to the DNL, then goes back and fixes the SOF
937
marker. This is a much simpler solution and is probably far more efficient.
938
Even if one wants piped input, buffering the first scan of the JPEG file needs
939
a lot smaller temp file than is implied by the maximum-height method. For
940
this approach we'd simply treat DNL as a no-op in the decompressor (at most,
941
check that it matches the SOF image height).
943
We will not worry about making the compressor capable of outputting DNL.
944
Something similar to the first scheme above could be applied if anyone ever
945
wants to make that work.