README.md
1Background
2==========
3
4libjpeg-turbo is a JPEG image codec that uses SIMD instructions to accelerate
5baseline JPEG compression and decompression on x86, x86-64, Arm, PowerPC, and
6MIPS systems, as well as progressive JPEG compression on x86, x86-64, and Arm
7systems. On such systems, libjpeg-turbo is generally 2-6x as fast as libjpeg,
8all else being equal. On other types of systems, libjpeg-turbo can still
9outperform libjpeg by a significant amount, by virtue of its highly-optimized
10Huffman coding routines. In many cases, the performance of libjpeg-turbo
11rivals that of proprietary high-speed JPEG codecs.
12
13libjpeg-turbo implements both the traditional libjpeg API as well as the less
14powerful but more straightforward TurboJPEG API. libjpeg-turbo also features
15colorspace extensions that allow it to compress from/decompress to 32-bit and
16big-endian pixel buffers (RGBX, XBGR, etc.), as well as a full-featured Java
17interface.
18
19libjpeg-turbo was originally based on libjpeg/SIMD, an MMX-accelerated
20derivative of libjpeg v6b developed by Miyasaka Masaru. The TigerVNC and
21VirtualGL projects made numerous enhancements to the codec in 2009, and in
22early 2010, libjpeg-turbo spun off into an independent project, with the goal
23of making high-speed JPEG compression/decompression technology available to a
24broader range of users and developers.
25
26
27License
28=======
29
30libjpeg-turbo is covered by three compatible BSD-style open source licenses.
31Refer to [LICENSE.md](LICENSE.md) for a roll-up of license terms.
32
33
34Building libjpeg-turbo
35======================
36
37Refer to [BUILDING.md](BUILDING.md) for complete instructions.
38
39
40Using libjpeg-turbo
41===================
42
43libjpeg-turbo includes two APIs that can be used to compress and decompress
44JPEG images:
45
46- **TurboJPEG API**<br>
47 This API provides an easy-to-use interface for compressing and decompressing
48 JPEG images in memory. It also provides some functionality that would not be
49 straightforward to achieve using the underlying libjpeg API, such as
50 generating planar YUV images and performing multiple simultaneous lossless
51 transforms on an image. The Java interface for libjpeg-turbo is written on
52 top of the TurboJPEG API. The TurboJPEG API is recommended for first-time
53 users of libjpeg-turbo. Refer to [tjexample.c](tjexample.c) and
54 [TJExample.java](java/TJExample.java) for examples of its usage and to
55 <http://libjpeg-turbo.org/Documentation/Documentation> for API documentation.
56
57- **libjpeg API**<br>
58 This is the de facto industry-standard API for compressing and decompressing
59 JPEG images. It is more difficult to use than the TurboJPEG API but also
60 more powerful. The libjpeg API implementation in libjpeg-turbo is both
61 API/ABI-compatible and mathematically compatible with libjpeg v6b. It can
62 also optionally be configured to be API/ABI-compatible with libjpeg v7 and v8
63 (see below.) Refer to [cjpeg.c](cjpeg.c) and [djpeg.c](djpeg.c) for examples
64 of its usage and to [libjpeg.txt](libjpeg.txt) for API documentation.
65
66There is no significant performance advantage to either API when both are used
67to perform similar operations.
68
69Colorspace Extensions
70---------------------
71
72libjpeg-turbo includes extensions that allow JPEG images to be compressed
73directly from (and decompressed directly to) buffers that use BGR, BGRX,
74RGBX, XBGR, and XRGB pixel ordering. This is implemented with ten new
75colorspace constants:
76
77 JCS_EXT_RGB /* red/green/blue */
78 JCS_EXT_RGBX /* red/green/blue/x */
79 JCS_EXT_BGR /* blue/green/red */
80 JCS_EXT_BGRX /* blue/green/red/x */
81 JCS_EXT_XBGR /* x/blue/green/red */
82 JCS_EXT_XRGB /* x/red/green/blue */
83 JCS_EXT_RGBA /* red/green/blue/alpha */
84 JCS_EXT_BGRA /* blue/green/red/alpha */
85 JCS_EXT_ABGR /* alpha/blue/green/red */
86 JCS_EXT_ARGB /* alpha/red/green/blue */
87
88Setting `cinfo.in_color_space` (compression) or `cinfo.out_color_space`
89(decompression) to one of these values will cause libjpeg-turbo to read the
90red, green, and blue values from (or write them to) the appropriate position in
91the pixel when compressing from/decompressing to an RGB buffer.
92
93Your application can check for the existence of these extensions at compile
94time with:
95
96 #ifdef JCS_EXTENSIONS
97
98At run time, attempting to use these extensions with a libjpeg implementation
99that does not support them will result in a "Bogus input colorspace" error.
100Applications can trap this error in order to test whether run-time support is
101available for the colorspace extensions.
102
103When using the RGBX, BGRX, XBGR, and XRGB colorspaces during decompression, the
104X byte is undefined, and in order to ensure the best performance, libjpeg-turbo
105can set that byte to whatever value it wishes. If an application expects the X
106byte to be used as an alpha channel, then it should specify `JCS_EXT_RGBA`,
107`JCS_EXT_BGRA`, `JCS_EXT_ABGR`, or `JCS_EXT_ARGB`. When these colorspace
108constants are used, the X byte is guaranteed to be 0xFF, which is interpreted
109as opaque.
110
111Your application can check for the existence of the alpha channel colorspace
112extensions at compile time with:
113
114 #ifdef JCS_ALPHA_EXTENSIONS
115
116[jcstest.c](jcstest.c), located in the libjpeg-turbo source tree, demonstrates
117how to check for the existence of the colorspace extensions at compile time and
118run time.
119
120libjpeg v7 and v8 API/ABI Emulation
121-----------------------------------
122
123With libjpeg v7 and v8, new features were added that necessitated extending the
124compression and decompression structures. Unfortunately, due to the exposed
125nature of those structures, extending them also necessitated breaking backward
126ABI compatibility with previous libjpeg releases. Thus, programs that were
127built to use libjpeg v7 or v8 did not work with libjpeg-turbo, since it is
128based on the libjpeg v6b code base. Although libjpeg v7 and v8 are not
129as widely used as v6b, enough programs (including a few Linux distros) made
130the switch that there was a demand to emulate the libjpeg v7 and v8 ABIs
131in libjpeg-turbo. It should be noted, however, that this feature was added
132primarily so that applications that had already been compiled to use libjpeg
133v7+ could take advantage of accelerated baseline JPEG encoding/decoding
134without recompiling. libjpeg-turbo does not claim to support all of the
135libjpeg v7+ features, nor to produce identical output to libjpeg v7+ in all
136cases (see below.)
137
138By passing an argument of `-DWITH_JPEG7=1` or `-DWITH_JPEG8=1` to `cmake`, you
139can build a version of libjpeg-turbo that emulates the libjpeg v7 or v8 ABI, so
140that programs that are built against libjpeg v7 or v8 can be run with
141libjpeg-turbo. The following section describes which libjpeg v7+ features are
142supported and which aren't.
143
144### Support for libjpeg v7 and v8 Features
145
146#### Fully supported
147
148- **libjpeg API: IDCT scaling extensions in decompressor**<br>
149 libjpeg-turbo supports IDCT scaling with scaling factors of 1/8, 1/4, 3/8,
150 1/2, 5/8, 3/4, 7/8, 9/8, 5/4, 11/8, 3/2, 13/8, 7/4, 15/8, and 2/1 (only 1/4
151 and 1/2 are SIMD-accelerated.)
152
153- **libjpeg API: Arithmetic coding**
154
155- **libjpeg API: In-memory source and destination managers**<br>
156 See notes below.
157
158- **cjpeg: Separate quality settings for luminance and chrominance**<br>
159 Note that the libpjeg v7+ API was extended to accommodate this feature only
160 for convenience purposes. It has always been possible to implement this
161 feature with libjpeg v6b (see rdswitch.c for an example.)
162
163- **cjpeg: 32-bit BMP support**
164
165- **cjpeg: `-rgb` option**
166
167- **jpegtran: Lossless cropping**
168
169- **jpegtran: `-perfect` option**
170
171- **jpegtran: Forcing width/height when performing lossless crop**
172
173- **rdjpgcom: `-raw` option**
174
175- **rdjpgcom: Locale awareness**
176
177
178#### Not supported
179
180NOTE: As of this writing, extensive research has been conducted into the
181usefulness of DCT scaling as a means of data reduction and SmartScale as a
182means of quality improvement. Readers are invited to peruse the research at
183<http://www.libjpeg-turbo.org/About/SmartScale> and draw their own conclusions,
184but it is the general belief of our project that these features have not
185demonstrated sufficient usefulness to justify inclusion in libjpeg-turbo.
186
187- **libjpeg API: DCT scaling in compressor**<br>
188 `cinfo.scale_num` and `cinfo.scale_denom` are silently ignored.
189 There is no technical reason why DCT scaling could not be supported when
190 emulating the libjpeg v7+ API/ABI, but without the SmartScale extension (see
191 below), only scaling factors of 1/2, 8/15, 4/7, 8/13, 2/3, 8/11, 4/5, and
192 8/9 would be available, which is of limited usefulness.
193
194- **libjpeg API: SmartScale**<br>
195 `cinfo.block_size` is silently ignored.
196 SmartScale is an extension to the JPEG format that allows for DCT block
197 sizes other than 8x8. Providing support for this new format would be
198 feasible (particularly without full acceleration.) However, until/unless
199 the format becomes either an official industry standard or, at minimum, an
200 accepted solution in the community, we are hesitant to implement it, as
201 there is no sense of whether or how it might change in the future. It is
202 our belief that SmartScale has not demonstrated sufficient usefulness as a
203 lossless format nor as a means of quality enhancement, and thus our primary
204 interest in providing this feature would be as a means of supporting
205 additional DCT scaling factors.
206
207- **libjpeg API: Fancy downsampling in compressor**<br>
208 `cinfo.do_fancy_downsampling` is silently ignored.
209 This requires the DCT scaling feature, which is not supported.
210
211- **jpegtran: Scaling**<br>
212 This requires both the DCT scaling and SmartScale features, which are not
213 supported.
214
215- **Lossless RGB JPEG files**<br>
216 This requires the SmartScale feature, which is not supported.
217
218### What About libjpeg v9?
219
220libjpeg v9 introduced yet another field to the JPEG compression structure
221(`color_transform`), thus making the ABI backward incompatible with that of
222libjpeg v8. This new field was introduced solely for the purpose of supporting
223lossless SmartScale encoding. Furthermore, there was actually no reason to
224extend the API in this manner, as the color transform could have just as easily
225been activated by way of a new JPEG colorspace constant, thus preserving
226backward ABI compatibility.
227
228Our research (see link above) has shown that lossless SmartScale does not
229generally accomplish anything that can't already be accomplished better with
230existing, standard lossless formats. Therefore, at this time it is our belief
231that there is not sufficient technical justification for software projects to
232upgrade from libjpeg v8 to libjpeg v9, and thus there is not sufficient
233technical justification for us to emulate the libjpeg v9 ABI.
234
235In-Memory Source/Destination Managers
236-------------------------------------
237
238By default, libjpeg-turbo 1.3 and later includes the `jpeg_mem_src()` and
239`jpeg_mem_dest()` functions, even when not emulating the libjpeg v8 API/ABI.
240Previously, it was necessary to build libjpeg-turbo from source with libjpeg v8
241API/ABI emulation in order to use the in-memory source/destination managers,
242but several projects requested that those functions be included when emulating
243the libjpeg v6b API/ABI as well. This allows the use of those functions by
244programs that need them, without breaking ABI compatibility for programs that
245don't, and it allows those functions to be provided in the "official"
246libjpeg-turbo binaries.
247
248Those who are concerned about maintaining strict conformance with the libjpeg
249v6b or v7 API can pass an argument of `-DWITH_MEM_SRCDST=0` to `cmake` prior to
250building libjpeg-turbo. This will restore the pre-1.3 behavior, in which
251`jpeg_mem_src()` and `jpeg_mem_dest()` are only included when emulating the
252libjpeg v8 API/ABI.
253
254On Un*x systems, including the in-memory source/destination managers changes
255the dynamic library version from 62.2.0 to 62.3.0 if using libjpeg v6b API/ABI
256emulation and from 7.2.0 to 7.3.0 if using libjpeg v7 API/ABI emulation.
257
258Note that, on most Un*x systems, the dynamic linker will not look for a
259function in a library until that function is actually used. Thus, if a program
260is built against libjpeg-turbo 1.3+ and uses `jpeg_mem_src()` or
261`jpeg_mem_dest()`, that program will not fail if run against an older version
262of libjpeg-turbo or against libjpeg v7- until the program actually tries to
263call `jpeg_mem_src()` or `jpeg_mem_dest()`. Such is not the case on Windows.
264If a program is built against the libjpeg-turbo 1.3+ DLL and uses
265`jpeg_mem_src()` or `jpeg_mem_dest()`, then it must use the libjpeg-turbo 1.3+
266DLL at run time.
267
268Both cjpeg and djpeg have been extended to allow testing the in-memory
269source/destination manager functions. See their respective man pages for more
270details.
271
272
273Mathematical Compatibility
274==========================
275
276For the most part, libjpeg-turbo should produce identical output to libjpeg
277v6b. The one exception to this is when using the floating point DCT/IDCT, in
278which case the outputs of libjpeg v6b and libjpeg-turbo can differ for the
279following reasons:
280
281- The SSE/SSE2 floating point DCT implementation in libjpeg-turbo is ever so
282 slightly more accurate than the implementation in libjpeg v6b, but not by
283 any amount perceptible to human vision (generally in the range of 0.01 to
284 0.08 dB gain in PNSR.)
285
286- When not using the SIMD extensions, libjpeg-turbo uses the more accurate
287 (and slightly faster) floating point IDCT algorithm introduced in libjpeg
288 v8a as opposed to the algorithm used in libjpeg v6b. It should be noted,
289 however, that this algorithm basically brings the accuracy of the floating
290 point IDCT in line with the accuracy of the accurate integer IDCT. The
291 floating point DCT/IDCT algorithms are mainly a legacy feature, and they do
292 not produce significantly more accuracy than the accurate integer algorithms
293 (to put numbers on this, the typical difference in PNSR between the two
294 algorithms is less than 0.10 dB, whereas changing the quality level by 1 in
295 the upper range of the quality scale is typically more like a 1.0 dB
296 difference.)
297
298- If the floating point algorithms in libjpeg-turbo are not implemented using
299 SIMD instructions on a particular platform, then the accuracy of the
300 floating point DCT/IDCT can depend on the compiler settings.
301
302While libjpeg-turbo does emulate the libjpeg v8 API/ABI, under the hood it is
303still using the same algorithms as libjpeg v6b, so there are several specific
304cases in which libjpeg-turbo cannot be expected to produce the same output as
305libjpeg v8:
306
307- When decompressing using scaling factors of 1/2 and 1/4, because libjpeg v8
308 implements those scaling algorithms differently than libjpeg v6b does, and
309 libjpeg-turbo's SIMD extensions are based on the libjpeg v6b behavior.
310
311- When using chrominance subsampling, because libjpeg v8 implements this
312 with its DCT/IDCT scaling algorithms rather than with a separate
313 downsampling/upsampling algorithm. In our testing, the subsampled/upsampled
314 output of libjpeg v8 is less accurate than that of libjpeg v6b for this
315 reason.
316
317- When decompressing using a scaling factor > 1 and merged (AKA "non-fancy" or
318 "non-smooth") chrominance upsampling, because libjpeg v8 does not support
319 merged upsampling with scaling factors > 1.
320
321
322Performance Pitfalls
323====================
324
325Restart Markers
326---------------
327
328The optimized Huffman decoder in libjpeg-turbo does not handle restart markers
329in a way that makes the rest of the libjpeg infrastructure happy, so it is
330necessary to use the slow Huffman decoder when decompressing a JPEG image that
331has restart markers. This can cause the decompression performance to drop by
332as much as 20%, but the performance will still be much greater than that of
333libjpeg. Many consumer packages, such as Photoshop, use restart markers when
334generating JPEG images, so images generated by those programs will experience
335this issue.
336
337Fast Integer Forward DCT at High Quality Levels
338-----------------------------------------------
339
340The algorithm used by the SIMD-accelerated quantization function cannot produce
341correct results whenever the fast integer forward DCT is used along with a JPEG
342quality of 98-100. Thus, libjpeg-turbo must use the non-SIMD quantization
343function in those cases. This causes performance to drop by as much as 40%.
344It is therefore strongly advised that you use the accurate integer forward DCT
345whenever encoding images with a JPEG quality of 98 or higher.
346
347
348Memory Debugger Pitfalls
349========================
350
351Valgrind and Memory Sanitizer (MSan) can generate false positives
352(specifically, incorrect reports of uninitialized memory accesses) when used
353with libjpeg-turbo's SIMD extensions. It is generally recommended that the
354SIMD extensions be disabled, either by passing an argument of `-DWITH_SIMD=0`
355to `cmake` when configuring the build or by setting the environment variable
356`JSIMD_FORCENONE` to `1` at run time, when testing libjpeg-turbo with Valgrind,
357MSan, or other memory debuggers.
358