2.1.6 ===== ### Significant changes relative to 2.1.5.1: 1. Fixed an oversight in 1.4 beta1[8] that caused various segfaults and buffer overruns when attempting to decompress various specially-crafted malformed 12-bit-per-component JPEG images using a 12-bit-per-component build of djpeg (`-DWITH_12BIT=1`) with both color quantization and RGB565 color conversion enabled. 2. Fixed an issue whereby `jpeg_crop_scanline()` sometimes miscalculated the downsampled width for components with 4x2 or 2x4 subsampling factors if decompression scaling was enabled. This caused the components to be upsampled incompletely, which caused the color converter to read from uninitialized memory. With 12-bit data precision, this caused a buffer overrun or underrun and subsequent segfault if the sample value read from uninitialized memory was outside of the valid sample range. 3. Fixed a long-standing issue whereby the `tjTransform()` function, when used with the `TJXOP_TRANSPOSE`, `TJXOP_TRANSVERSE`, `TJXOP_ROT90`, or `TJXOP_ROT270` transform operation and without automatic JPEG destination buffer (re)allocation or lossless cropping, computed the worst-case transformed JPEG image size based on the source image dimensions rather than the transformed image dimensions. If a calling program allocated the JPEG destination buffer based on the transformed image dimensions, as the API documentation instructs, and attempted to transform a specially-crafted 4:2:2, 4:4:0, or 4:1:1 JPEG source image containing a large amount of metadata, the issue caused `tjTransform()` to overflow the JPEG destination buffer rather than fail gracefully. The issue could be worked around by setting `TJXOPT_COPYNONE`. Note that, irrespective of this issue, `tjTransform()` cannot reliably transform JPEG source images that contain a large amount of metadata unless automatic JPEG destination buffer (re)allocation is used or `TJXOPT_COPYNONE` is set. 4. Fixed an issue that caused the C Huffman encoder (which is not used by default on x86 and Arm CPUs) to read from uninitialized memory when attempting to transform a specially-crafted malformed arithmetic-coded JPEG source image into a baseline Huffman-coded JPEG destination image. 5. Fixed two minor issues in the interblock smoothing algorithm that caused mathematical (but not necessarily perceptible) edge block errors when decompressing progressive JPEG images exactly two MCU blocks in width or that use vertical chrominance subsampling. 2.1.5.1 ======= ### Significant changes relative to 2.1.5: 1. The SIMD dispatchers in libjpeg-turbo 2.1.4 and prior stored the list of supported SIMD instruction sets in a global variable, which caused an innocuous race condition whereby the variable could have been initialized multiple times if `jpeg_start_*compress()` was called simultaneously in multiple threads. libjpeg-turbo 2.1.5 included an undocumented attempt to fix this race condition by making the SIMD support variable thread-local. However, that caused another issue whereby, if `jpeg_start_*compress()` was called in one thread and `jpeg_read_*()` or `jpeg_write_*()` was called in a second thread, the SIMD support variable was never initialized in the second thread. On x86 systems, this led the second thread to incorrectly assume that AVX2 instructions were always available, and when it attempted to use those instructions on older x86 CPUs that do not support them, an illegal instruction error occurred. The SIMD dispatchers now ensure that the SIMD support variable is initialized before dispatching based on its value. 2.1.5 ===== ### Significant changes relative to 2.1.4: 1. Fixed issues in the build system whereby, when using the Ninja Multi-Config CMake generator, a static build of libjpeg-turbo (a build in which `ENABLE_SHARED` is `0`) could not be installed, a Windows installer could not be built, and the Java regression tests failed. 2. Fixed a regression introduced by 2.0 beta1[15] that caused a buffer overrun in the progressive Huffman encoder when attempting to transform a specially-crafted malformed 12-bit-per-component JPEG image into a progressive 12-bit-per-component JPEG image using a 12-bit-per-component build of libjpeg-turbo (`-DWITH_12BIT=1`.) Given that the buffer overrun was fully contained within the progressive Huffman encoder structure and did not cause a segfault or other user-visible errant behavior, given that the lossless transformer (unlike the decompressor) is not generally exposed to arbitrary data exploits, and given that 12-bit-per-component builds of libjpeg-turbo are uncommon, this issue did not likely pose a security risk. 3. Fixed an issue whereby, when using a 12-bit-per-component build of libjpeg-turbo (`-DWITH_12BIT=1`), passing samples with values greater than 4095 or less than 0 to `jpeg_write_scanlines()` caused a buffer overrun or underrun in the RGB-to-YCbCr color converter. 4. Fixed a floating point exception that occurred when attempting to use the jpegtran `-drop` and `-trim` options to losslessly transform a specially-crafted malformed JPEG image. 5. Fixed an issue in `tjBufSizeYUV2()` whereby it returned a bogus result, rather than throwing an error, if the `align` parameter was not a power of 2. Fixed a similar issue in `tjCompressFromYUV()` whereby it generated a corrupt JPEG image in certain cases, rather than throwing an error, if the `align` parameter was not a power of 2. 6. Fixed an issue whereby `tjDecompressToYUV2()`, which is a wrapper for `tjDecompressToYUVPlanes()`, used the desired YUV image dimensions rather than the actual scaled image dimensions when computing the plane pointers and strides to pass to `tjDecompressToYUVPlanes()`. This caused a buffer overrun and subsequent segfault if the desired image dimensions exceeded the scaled image dimensions. 7. Fixed an issue whereby, when decompressing a 12-bit-per-component JPEG image (`-DWITH_12BIT=1`) using an alpha-enabled output color space such as `JCS_EXT_RGBA`, the alpha channel was set to 255 rather than 4095. 8. Fixed an issue whereby the Java version of TJBench did not accept a range of quality values. 9. Fixed an issue whereby, when `-progressive` was passed to TJBench, the JPEG input image was not transformed into a progressive JPEG image prior to decompression. 2.1.4 ===== ### Significant changes relative to 2.1.3: 1. Fixed a regression introduced in 2.1.3 that caused build failures with Visual Studio 2010. 2. The `tjDecompressHeader3()` function in the TurboJPEG C API and the `TJDecompressor.setSourceImage()` method in the TurboJPEG Java API now accept "abbreviated table specification" (AKA "tables-only") datastreams, which can be used to prime the decompressor with quantization and Huffman tables that can be used when decompressing subsequent "abbreviated image" datastreams. 3. libjpeg-turbo now performs run-time detection of AltiVec instructions on OS X/PowerPC systems if AltiVec instructions are not enabled at compile time. This allows both AltiVec-equipped (PowerPC G4 and G5) and non-AltiVec-equipped (PowerPC G3) CPUs to be supported using the same build of libjpeg-turbo. 4. Fixed an error ("Bogus virtual array access") that occurred when attempting to decompress a progressive JPEG image with a height less than or equal to one iMCU (8 * the vertical sampling factor) using buffered-image mode with interblock smoothing enabled. This was a regression introduced by 2.1 beta1[6(b)]. 5. Fixed two issues that prevented partial image decompression from working properly with buffered-image mode: - Attempting to call `jpeg_crop_scanline()` after `jpeg_start_decompress()` but before `jpeg_start_output()` resulted in an error ("Improper call to JPEG library in state 207".) - Attempting to use `jpeg_skip_scanlines()` resulted in an error ("Bogus virtual array access") under certain circumstances. 2.1.3 ===== ### Significant changes relative to 2.1.2: 1. Fixed a regression introduced by 2.0 beta1[7] whereby cjpeg compressed PGM input files into full-color JPEG images unless the `-grayscale` option was used. 2. cjpeg now automatically compresses GIF and 8-bit BMP input files into grayscale JPEG images if the input files contain only shades of gray. 3. The build system now enables the intrinsics implementation of the AArch64 (Arm 64-bit) Neon SIMD extensions by default when using GCC 12 or later. 4. Fixed a segfault that occurred while decompressing a 4:2:0 JPEG image using the merged (non-fancy) upsampling algorithms (that is, with `cinfo.do_fancy_upsampling` set to `FALSE`) along with `jpeg_crop_scanline()`. Specifically, the segfault occurred if the number of bytes remaining in the output buffer was less than the number of bytes required to represent one uncropped scanline of the output image. For that reason, the issue could only be reproduced using the libjpeg API, not using djpeg. 2.1.2 ===== ### Significant changes relative to 2.1.1: 1. Fixed a regression introduced by 2.1 beta1[13] that caused the remaining GAS implementations of AArch64 (Arm 64-bit) Neon SIMD functions (which are used by default with GCC for performance reasons) to be placed in the `.rodata` section rather than in the `.text` section. This caused the GNU linker to automatically place the `.rodata` section in an executable segment, which prevented libjpeg-turbo from working properly with other linkers and also represented a potential security risk. 2. Fixed an issue whereby the `tjTransform()` function incorrectly computed the MCU block size for 4:4:4 JPEG images with non-unary sampling factors and thus unduly rejected some cropping regions, even though those regions aligned with 8x8 MCU block boundaries. 3. Fixed a regression introduced by 2.1 beta1[13] that caused the build system to enable the Arm Neon SIMD extensions when targetting Armv6 and other legacy architectures that do not support Neon instructions. 4. libjpeg-turbo now performs run-time detection of AltiVec instructions on FreeBSD/PowerPC systems if AltiVec instructions are not enabled at compile time. This allows both AltiVec-equipped and non-AltiVec-equipped CPUs to be supported using the same build of libjpeg-turbo. 5. cjpeg now accepts a `-strict` argument similar to that of djpeg and jpegtran, which causes the compressor to abort if an LZW-compressed GIF input image contains incomplete or corrupt image data. 2.1.1 ===== ### Significant changes relative to 2.1.0: 1. Fixed a regression introduced in 2.1.0 that caused build failures with non-GCC-compatible compilers for Un*x/Arm platforms. 2. Fixed a regression introduced by 2.1 beta1[13] that prevented the Arm 32-bit (AArch32) Neon SIMD extensions from building unless the C compiler flags included `-mfloat-abi=softfp` or `-mfloat-abi=hard`. 3. Fixed an issue in the AArch32 Neon SIMD Huffman encoder whereby reliance on undefined C compiler behavior led to crashes ("SIGBUS: illegal alignment") on Android systems when running AArch32/Thumb builds of libjpeg-turbo built with recent versions of Clang. 4. Added a command-line argument (`-copy icc`) to jpegtran that causes it to copy only the ICC profile markers from the source file and discard any other metadata. 5. libjpeg-turbo should now build and run on CHERI-enabled architectures, which use capability pointers that are larger than the size of `size_t`. 6. Fixed a regression (CVE-2021-37972) introduced by 2.1 beta1[5] that caused a segfault in the 64-bit SSE2 Huffman encoder when attempting to losslessly transform a specially-crafted malformed JPEG image. 2.1.0 ===== ### Significant changes relative to 2.1 beta1: 1. Fixed a regression (CVE-2021-29390) introduced by 2.1 beta1[6(b)] whereby attempting to decompress certain progressive JPEG images with one or more component planes of width 8 or less caused a buffer overrun. 2. Fixed a regression introduced by 2.1 beta1[6(b)] whereby attempting to decompress a specially-crafted malformed progressive JPEG image caused the block smoothing algorithm to read from uninitialized memory. 3. Fixed an issue in the Arm Neon SIMD Huffman encoders that caused the encoders to generate incorrect results when using the Clang compiler with Visual Studio. 4. Fixed a floating point exception (CVE-2021-20205) that occurred when attempting to compress a specially-crafted malformed GIF image with a specified image width of 0 using cjpeg. 5. Fixed a regression introduced by 2.0 beta1[15] whereby attempting to generate a progressive JPEG image on an SSE2-capable CPU using a scan script containing one or more scans with lengths divisible by 32 and non-zero successive approximation low bit positions would, under certain circumstances, result in an error ("Missing Huffman code table entry") and an invalid JPEG image. 6. Introduced a new flag (`TJFLAG_LIMITSCANS` in the TurboJPEG C API and `TJ.FLAG_LIMIT_SCANS` in the TurboJPEG Java API) and a corresponding TJBench command-line argument (`-limitscans`) that causes the TurboJPEG decompression and transform functions/operations to return/throw an error if a progressive JPEG image contains an unreasonably large number of scans. This allows applications that use the TurboJPEG API to guard against an exploit of the progressive JPEG format described in the report ["Two Issues with the JPEG Standard"](https://libjpeg-turbo.org/pmwiki/uploads/About/TwoIssueswiththeJPEGStandard.pdf). 7. The PPM reader now throws an error, rather than segfaulting (due to a buffer overrun, CVE-2021-46822) or generating incorrect pixels, if an application attempts to use the `tjLoadImage()` function to load a 16-bit binary PPM file (a binary PPM file with a maximum value greater than 255) into a grayscale image buffer or to load a 16-bit binary PGM file into an RGB image buffer. 8. Fixed an issue in the PPM reader that caused incorrect pixels to be generated when using the `tjLoadImage()` function to load a 16-bit binary PPM file into an extended RGB image buffer. 9. Fixed an issue whereby, if a JPEG buffer was automatically re-allocated by one of the TurboJPEG compression or transform functions and an error subsequently occurred during compression or transformation, the JPEG buffer pointer passed by the application was not updated when the function returned. 2.0.90 (2.1 beta1) ================== ### Significant changes relative to 2.0.6: 1. The build system, x86-64 SIMD extensions, and accelerated Huffman codec now support the x32 ABI on Linux, which allows for using x86-64 instructions with 32-bit pointers. The x32 ABI is generally enabled by adding `-mx32` to the compiler flags. Caveats: - CMake 3.9.0 or later is required in order for the build system to automatically detect an x32 build. - Java does not support the x32 ABI, and thus the TurboJPEG Java API will automatically be disabled with x32 builds. 2. Added Loongson MMI SIMD implementations of the RGB-to-grayscale, 4:2:2 fancy chroma upsampling, 4:2:2 and 4:2:0 merged chroma upsampling/color conversion, and fast integer DCT/IDCT algorithms. Relative to libjpeg-turbo 2.0.x, this speeds up: - the compression of RGB source images into grayscale JPEG images by approximately 20% - the decompression of 4:2:2 JPEG images by approximately 40-60% when using fancy upsampling - the decompression of 4:2:2 and 4:2:0 JPEG images by approximately 15-20% when using merged upsampling - the compression of RGB source images by approximately 30-45% when using the fast integer DCT - the decompression of JPEG images into RGB destination images by approximately 2x when using the fast integer IDCT The overall decompression speedup for RGB images is now approximately 2.3-3.7x (compared to 2-3.5x with libjpeg-turbo 2.0.x.) 3. 32-bit (Armv7 or Armv7s) iOS builds of libjpeg-turbo are no longer supported, and the libjpeg-turbo build system can no longer be used to package such builds. 32-bit iOS apps cannot run in iOS 11 and later, and the App Store no longer allows them. 4. 32-bit (i386) OS X/macOS builds of libjpeg-turbo are no longer supported, and the libjpeg-turbo build system can no longer be used to package such builds. 32-bit Mac applications cannot run in macOS 10.15 "Catalina" and later, and the App Store no longer allows them. 5. The SSE2 (x86 SIMD) and C Huffman encoding algorithms have been significantly optimized, resulting in a measured average overall compression speedup of 12-28% for 64-bit code and 22-52% for 32-bit code on various Intel and AMD CPUs, as well as a measured average overall compression speedup of 0-23% on platforms that do not have a SIMD-accelerated Huffman encoding implementation. 6. The block smoothing algorithm that is applied by default when decompressing progressive Huffman-encoded JPEG images has been improved in the following ways: - The algorithm is now more fault-tolerant. Previously, if a particular scan was incomplete, then the smoothing parameters for the incomplete scan would be applied to the entire output image, including the parts of the image that were generated by the prior (complete) scan. Visually, this had the effect of removing block smoothing from lower-frequency scans if they were followed by an incomplete higher-frequency scan. libjpeg-turbo now applies block smoothing parameters to each iMCU row based on which scan generated the pixels in that row, rather than always using the block smoothing parameters for the most recent scan. - When applying block smoothing to DC scans, a Gaussian-like kernel with a 5x5 window is used to reduce the "blocky" appearance. 7. Added SIMD acceleration for progressive Huffman encoding on Arm platforms. This speeds up the compression of full-color progressive JPEGs by about 30-40% on average (relative to libjpeg-turbo 2.0.x) when using modern Arm CPUs. 8. Added configure-time and run-time auto-detection of Loongson MMI SIMD instructions, so that the Loongson MMI SIMD extensions can be included in any MIPS64 libjpeg-turbo build. 9. Added fault tolerance features to djpeg and jpegtran, mainly to demonstrate methods by which applications can guard against the exploits of the JPEG format described in the report ["Two Issues with the JPEG Standard"](https://libjpeg-turbo.org/pmwiki/uploads/About/TwoIssueswiththeJPEGStandard.pdf). - Both programs now accept a `-maxscans` argument, which can be used to limit the number of allowable scans in the input file. - Both programs now accept a `-strict` argument, which can be used to treat all warnings as fatal. 10. CMake package config files are now included for both the libjpeg and TurboJPEG API libraries. This facilitates using libjpeg-turbo with CMake's `find_package()` function. For example: find_package(libjpeg-turbo CONFIG REQUIRED) add_executable(libjpeg_program libjpeg_program.c) target_link_libraries(libjpeg_program PUBLIC libjpeg-turbo::jpeg) add_executable(libjpeg_program_static libjpeg_program.c) target_link_libraries(libjpeg_program_static PUBLIC libjpeg-turbo::jpeg-static) add_executable(turbojpeg_program turbojpeg_program.c) target_link_libraries(turbojpeg_program PUBLIC libjpeg-turbo::turbojpeg) add_executable(turbojpeg_program_static turbojpeg_program.c) target_link_libraries(turbojpeg_program_static PUBLIC libjpeg-turbo::turbojpeg-static) 11. Since the Unisys LZW patent has long expired, cjpeg and djpeg can now read/write both LZW-compressed and uncompressed GIF files (feature ported from jpeg-6a and jpeg-9d.) 12. jpegtran now includes the `-wipe` and `-drop` options from jpeg-9a and jpeg-9d, as well as the ability to expand the image size using the `-crop` option. Refer to jpegtran.1 or usage.txt for more details. 13. Added a complete intrinsics implementation of the Arm Neon SIMD extensions, thus providing SIMD acceleration on Arm platforms for all of the algorithms that are SIMD-accelerated on x86 platforms. This new implementation is significantly faster in some cases than the old GAS implementation-- depending on the algorithms used, the type of CPU core, and the compiler. GCC, as of this writing, does not provide a full or optimal set of Neon intrinsics, so for performance reasons, the default when building libjpeg-turbo with GCC is to continue using the GAS implementation of the following algorithms: - 32-bit RGB-to-YCbCr color conversion - 32-bit fast and accurate inverse DCT - 64-bit RGB-to-YCbCr and YCbCr-to-RGB color conversion - 64-bit accurate forward and inverse DCT - 64-bit Huffman encoding A new CMake variable (`NEON_INTRINSICS`) can be used to override this default. Since the new intrinsics implementation includes SIMD acceleration for merged upsampling/color conversion, 1.5.1[5] is no longer necessary and has been reverted. 14. The Arm Neon SIMD extensions can now be built using Visual Studio. 15. The build system can now be used to generate a universal x86-64 + Armv8 libjpeg-turbo SDK package for both iOS and macOS. 2.0.6 ===== ### Significant changes relative to 2.0.5: 1. Fixed "using JNI after critical get" errors that occurred on Android platforms when using any of the YUV encoding/compression/decompression/decoding methods in the TurboJPEG Java API. 2. Fixed or worked around multiple issues with `jpeg_skip_scanlines()`: - Fixed segfaults (CVE-2020-35538) or "Corrupt JPEG data: premature end of data segment" errors in `jpeg_skip_scanlines()` that occurred when decompressing 4:2:2 or 4:2:0 JPEG images using merged (non-fancy) upsampling/color conversion (that is, when setting `cinfo.do_fancy_upsampling` to `FALSE`.) 2.0.0[6] was a similar fix, but it did not cover all cases. - `jpeg_skip_scanlines()` now throws an error if two-pass color quantization is enabled. Two-pass color quantization never worked properly with `jpeg_skip_scanlines()`, and the issues could not readily be fixed. - Fixed an issue whereby `jpeg_skip_scanlines()` always returned 0 when skipping past the end of an image. 3. The Arm 64-bit (Armv8) Neon SIMD extensions can now be built using MinGW toolchains targetting Arm64 (AArch64) Windows binaries. 4. Fixed unexpected visual artifacts that occurred when using `jpeg_crop_scanline()` and interblock smoothing while decompressing only the DC scan of a progressive JPEG image. 5. Fixed an issue whereby libjpeg-turbo would not build if 12-bit-per-component JPEG support (`WITH_12BIT`) was enabled along with libjpeg v7 or libjpeg v8 API/ABI emulation (`WITH_JPEG7` or `WITH_JPEG8`.) 2.0.5 ===== ### Significant changes relative to 2.0.4: 1. Worked around issues in the MIPS DSPr2 SIMD extensions that caused failures in the libjpeg-turbo regression tests. Specifically, the `jsimd_h2v1_downsample_dspr2()` and `jsimd_h2v2_downsample_dspr2()` functions in the MIPS DSPr2 SIMD extensions are now disabled until/unless they can be fixed, and other functions that are incompatible with big endian MIPS CPUs are disabled when building libjpeg-turbo for such CPUs. 2. Fixed an oversight in the `TJCompressor.compress(int)` method in the TurboJPEG Java API that caused an error ("java.lang.IllegalStateException: No source image is associated with this instance") when attempting to use that method to compress a YUV image. 3. Fixed an issue (CVE-2020-13790) in the PPM reader that caused a buffer overrun in cjpeg, TJBench, or the `tjLoadImage()` function if one of the values in a binary PPM/PGM input file exceeded the maximum value defined in the file's header and that maximum value was less than 255. libjpeg-turbo 1.5.0 already included a similar fix for binary PPM/PGM files with maximum values greater than 255. 4. The TurboJPEG API library's global error handler, which is used in functions such as `tjBufSize()` and `tjLoadImage()` that do not require a TurboJPEG instance handle, is now thread-safe on platforms that support thread-local storage. 2.0.4 ===== ### Significant changes relative to 2.0.3: 1. Fixed a regression in the Windows packaging system (introduced by 2.0 beta1[2]) whereby, if both the 64-bit libjpeg-turbo SDK for GCC and the 64-bit libjpeg-turbo SDK for Visual C++ were installed on the same system, only one of them could be uninstalled. 2. Fixed a signed integer overflow and subsequent segfault that occurred when attempting to decompress images with more than 715827882 pixels using the 64-bit C version of TJBench. 3. Fixed out-of-bounds write in `tjDecompressToYUV2()` and `tjDecompressToYUVPlanes()` (sometimes manifesting as a double free) that occurred when attempting to decompress grayscale JPEG images that were compressed with a sampling factor other than 1 (for instance, with `cjpeg -grayscale -sample 2x2`). 4. Fixed a regression introduced by 2.0.2[5] that caused the TurboJPEG API to incorrectly identify some JPEG images with unusual sampling factors as 4:4:4 JPEG images. This was known to cause a buffer overflow when attempting to decompress some such images using `tjDecompressToYUV2()` or `tjDecompressToYUVPlanes()`. 5. Fixed an issue (CVE-2020-17541), detected by ASan, whereby attempting to losslessly transform a specially-crafted malformed JPEG image containing an extremely-high-frequency coefficient block (junk image data that could never be generated by a legitimate JPEG compressor) could cause the Huffman encoder's local buffer to be overrun. (Refer to 1.4.0[9] and 1.4beta1[15].) Given that the buffer overrun was fully contained within the stack and did not cause a segfault or other user-visible errant behavior, and given that the lossless transformer (unlike the decompressor) is not generally exposed to arbitrary data exploits, this issue did not likely pose a security risk. 6. The Arm 64-bit (Armv8) Neon SIMD assembly code now stores constants in a separate read-only data section rather than in the text section, to support execute-only memory layouts. 2.0.3 ===== ### Significant changes relative to 2.0.2: 1. Fixed "using JNI after critical get" errors that occurred on Android platforms when passing invalid arguments to certain methods in the TurboJPEG Java API. 2. Fixed a regression in the SIMD feature detection code, introduced by the AVX2 SIMD extensions (2.0 beta1[1]), that was known to cause an illegal instruction exception, in rare cases, on CPUs that lack support for CPUID leaf 07H (or on which the maximum CPUID leaf has been limited by way of a BIOS setting.) 3. The 4:4:0 (h1v2) fancy (smooth) chroma upsampling algorithm in the decompressor now uses a similar bias pattern to that of the 4:2:2 (h2v1) fancy chroma upsampling algorithm, rounding up or down the upsampled result for alternate pixels rather than always rounding down. This ensures that, regardless of whether a 4:2:2 JPEG image is rotated or transposed prior to decompression (in the frequency domain) or after decompression (in the spatial domain), the final image will be similar. 4. Fixed an integer overflow and subsequent segfault that occurred when attempting to compress or decompress images with more than 1 billion pixels using the TurboJPEG API. 5. Fixed a regression introduced by 2.0 beta1[15] whereby attempting to generate a progressive JPEG image on an SSE2-capable CPU using a scan script containing one or more scans with lengths divisible by 16 would result in an error ("Missing Huffman code table entry") and an invalid JPEG image. 6. Fixed an issue whereby `tjDecodeYUV()` and `tjDecodeYUVPlanes()` would throw an error ("Invalid progressive parameters") or a warning ("Inconsistent progression sequence") if passed a TurboJPEG instance that was previously used to decompress a progressive JPEG image. 2.0.2 ===== ### Significant changes relative to 2.0.1: 1. Fixed a regression introduced by 2.0.1[5] that prevented a runtime search path (rpath) from being embedded in the libjpeg-turbo shared libraries and executables for macOS and iOS. This caused a fatal error of the form "dyld: Library not loaded" when attempting to use one of the executables, unless `DYLD_LIBRARY_PATH` was explicitly set to the location of the libjpeg-turbo shared libraries. 2. Fixed an integer overflow and subsequent segfault (CVE-2018-20330) that occurred when attempting to load a BMP file with more than 1 billion pixels using the `tjLoadImage()` function. 3. Fixed a buffer overrun (CVE-2018-19664) that occurred when attempting to decompress a specially-crafted malformed JPEG image to a 256-color BMP using djpeg. 4. Fixed a floating point exception that occurred when attempting to decompress a specially-crafted malformed JPEG image with a specified image width or height of 0 using the C version of TJBench. 5. The TurboJPEG API will now decompress 4:4:4 JPEG images with 2x1, 1x2, 3x1, or 1x3 luminance and chrominance sampling factors. This is a non-standard way of specifying 1x subsampling (normally 4:4:4 JPEGs have 1x1 luminance and chrominance sampling factors), but the JPEG format and the libjpeg API both allow it. 6. Fixed a regression introduced by 2.0 beta1[7] that caused djpeg to generate incorrect PPM images when used with the `-colors` option. 7. Fixed an issue whereby a static build of libjpeg-turbo (a build in which `ENABLE_SHARED` is `0`) could not be installed using the Visual Studio IDE. 8. Fixed a severe performance issue in the Loongson MMI SIMD extensions that occurred when compressing RGB images whose image rows were not 64-bit-aligned. 2.0.1 ===== ### Significant changes relative to 2.0.0: 1. Fixed a regression introduced with the new CMake-based Un*x build system, whereby jconfig.h could cause compiler warnings of the form `"HAVE_*_H" redefined` if it was included by downstream Autotools-based projects that used `AC_CHECK_HEADERS()` to check for the existence of locale.h, stddef.h, or stdlib.h. 2. The `jsimd_quantize_float_dspr2()` and `jsimd_convsamp_float_dspr2()` functions in the MIPS DSPr2 SIMD extensions are now disabled at compile time if the soft float ABI is enabled. Those functions use instructions that are incompatible with the soft float ABI. 3. Fixed a regression in the SIMD feature detection code, introduced by the AVX2 SIMD extensions (2.0 beta1[1]), that caused libjpeg-turbo to crash on Windows 7 if Service Pack 1 was not installed. 4. Fixed out-of-bounds read in cjpeg that occurred when attempting to compress a specially-crafted malformed color-index (8-bit-per-sample) Targa file in which some of the samples (color indices) exceeded the bounds of the Targa file's color table. 5. Fixed an issue whereby installing a fully static build of libjpeg-turbo (a build in which `CFLAGS` contains `-static` and `ENABLE_SHARED` is `0`) would fail with "No valid ELF RPATH or RUNPATH entry exists in the file." 2.0.0 ===== ### Significant changes relative to 2.0 beta1: 1. The TurboJPEG API can now decompress CMYK JPEG images that have subsampled M and Y components (not to be confused with YCCK JPEG images, in which the C/M/Y components have been transformed into luma and chroma.) Previously, an error was generated ("Could not determine subsampling type for JPEG image") when such an image was passed to `tjDecompressHeader3()`, `tjTransform()`, `tjDecompressToYUVPlanes()`, `tjDecompressToYUV2()`, or the equivalent Java methods. 2. Fixed an issue (CVE-2018-11813) whereby a specially-crafted malformed input file (specifically, a file with a valid Targa header but incomplete pixel data) would cause cjpeg to generate a JPEG file that was potentially thousands of times larger than the input file. The Targa reader in cjpeg was not properly detecting that the end of the input file had been reached prematurely, so after all valid pixels had been read from the input, the reader injected dummy pixels with values of 255 into the JPEG compressor until the number of pixels specified in the Targa header had been compressed. The Targa reader in cjpeg now behaves like the PPM reader and aborts compression if the end of the input file is reached prematurely. Because this issue only affected cjpeg and not the underlying library, and because it did not involve any out-of-bounds reads or other exploitable behaviors, it was not believed to represent a security threat. 3. Fixed an issue whereby the `tjLoadImage()` and `tjSaveImage()` functions would produce a "Bogus message code" error message if the underlying bitmap and PPM readers/writers threw an error that was specific to the readers/writers (as opposed to a general libjpeg API error.) 4. Fixed an issue (CVE-2018-1152) whereby a specially-crafted malformed BMP file, one in which the header specified an image width of 1073741824 pixels, would trigger a floating point exception (division by zero) in the `tjLoadImage()` function when attempting to load the BMP file into a 4-component image buffer. 5. Fixed an issue whereby certain combinations of calls to `jpeg_skip_scanlines()` and `jpeg_read_scanlines()` could trigger an infinite loop when decompressing progressive JPEG images that use vertical chroma subsampling (for instance, 4:2:0 or 4:4:0.) 6. Fixed a segfault in `jpeg_skip_scanlines()` that occurred when decompressing a 4:2:2 or 4:2:0 JPEG image using the merged (non-fancy) upsampling algorithms (that is, when setting `cinfo.do_fancy_upsampling` to `FALSE`.) 7. The new CMake-based build system will now disable the MIPS DSPr2 SIMD extensions if it detects that the compiler does not support DSPr2 instructions. 8. Fixed out-of-bounds read in cjpeg (CVE-2018-14498) that occurred when attempting to compress a specially-crafted malformed color-index (8-bit-per-sample) BMP file in which some of the samples (color indices) exceeded the bounds of the BMP file's color table. 9. Fixed a signed integer overflow in the progressive Huffman decoder, detected by the Clang and GCC undefined behavior sanitizers, that could be triggered by attempting to decompress a specially-crafted malformed JPEG image. This issue did not pose a security threat, but removing the warning made it easier to detect actual security issues, should they arise in the future. 1.5.90 (2.0 beta1) ================== ### Significant changes relative to 1.5.3: 1. Added AVX2 SIMD implementations of the colorspace conversion, chroma downsampling and upsampling, integer quantization and sample conversion, and accurate integer DCT/IDCT algorithms. When using the accurate integer DCT/IDCT algorithms on AVX2-equipped CPUs, the compression of RGB images is approximately 13-36% (avg. 22%) faster (relative to libjpeg-turbo 1.5.x) with 64-bit code and 11-21% (avg. 17%) faster with 32-bit code, and the decompression of RGB images is approximately 9-35% (avg. 17%) faster with 64-bit code and 7-17% (avg. 12%) faster with 32-bit code. (As tested on a 3 GHz Intel Core i7. Actual mileage may vary.) 2. Overhauled the build system to use CMake on all platforms, and removed the autotools-based build system. This decision resulted from extensive discussions within the libjpeg-turbo community. libjpeg-turbo traditionally used CMake only for Windows builds, but there was an increasing amount of demand to extend CMake support to other platforms. However, because of the unique nature of our code base (the need to support different assemblers on each platform, the need for Java support, etc.), providing dual build systems as other OSS imaging libraries do (including libpng and libtiff) would have created a maintenance burden. The use of CMake greatly simplifies some aspects of our build system, owing to CMake's built-in support for various assemblers, Java, and unit testing, as well as generally fewer quirks that have to be worked around in order to implement our packaging system. Eliminating autotools puts our project slightly at odds with the traditional practices of the OSS community, since most "system libraries" tend to be built with autotools, but it is believed that the benefits of this move outweigh the risks. In addition to providing a unified build environment, switching to CMake allows for the use of various build tools and IDEs that aren't supported under autotools, including XCode, Ninja, and Eclipse. It also eliminates the need to install autotools via MacPorts/Homebrew on OS X and allows libjpeg-turbo to be configured without the use of a terminal/command prompt. Extensive testing was conducted to ensure that all features provided by the autotools-based build system are provided by the new build system. 3. The libjpeg API in this version of libjpeg-turbo now includes two additional functions, `jpeg_read_icc_profile()` and `jpeg_write_icc_profile()`, that can be used to extract ICC profile data from a JPEG file while decompressing or to embed ICC profile data in a JPEG file while compressing or transforming. This eliminates the need for downstream projects, such as color management libraries and browsers, to include their own glueware for accomplishing this. 4. Improved error handling in the TurboJPEG API library: - Introduced a new function (`tjGetErrorStr2()`) in the TurboJPEG C API that allows compression/decompression/transform error messages to be retrieved in a thread-safe manner. Retrieving error messages from global functions, such as `tjInitCompress()` or `tjBufSize()`, is still thread-unsafe, but since those functions will only throw errors if passed an invalid argument or if a memory allocation failure occurs, thread safety is not as much of a concern. - Introduced a new function (`tjGetErrorCode()`) in the TurboJPEG C API and a new method (`TJException.getErrorCode()`) in the TurboJPEG Java API that can be used to determine the severity of the last compression/decompression/transform error. This allows applications to choose whether to ignore warnings (non-fatal errors) from the underlying libjpeg API or to treat them as fatal. - Introduced a new flag (`TJFLAG_STOPONWARNING` in the TurboJPEG C API and `TJ.FLAG_STOPONWARNING` in the TurboJPEG Java API) that causes the library to immediately halt a compression/decompression/transform operation if it encounters a warning from the underlying libjpeg API (the default behavior is to allow the operation to complete unless a fatal error is encountered.) 5. Introduced a new flag in the TurboJPEG C and Java APIs (`TJFLAG_PROGRESSIVE` and `TJ.FLAG_PROGRESSIVE`, respectively) that causes the library to use progressive entropy coding in JPEG images generated by compression and transform operations. Additionally, a new transform option (`TJXOPT_PROGRESSIVE` in the C API and `TJTransform.OPT_PROGRESSIVE` in the Java API) has been introduced, allowing progressive entropy coding to be enabled for selected transforms in a multi-transform operation. 6. Introduced a new transform option in the TurboJPEG API (`TJXOPT_COPYNONE` in the C API and `TJTransform.OPT_COPYNONE` in the Java API) that allows the copying of markers (including EXIF and ICC profile data) to be disabled for a particular transform. 7. Added two functions to the TurboJPEG C API (`tjLoadImage()` and `tjSaveImage()`) that can be used to load/save a BMP or PPM/PGM image to/from a memory buffer with a specified pixel format and layout. These functions replace the project-private (and slow) bmp API, which was previously used by TJBench, and they also provide a convenient way for first-time users of libjpeg-turbo to quickly develop a complete JPEG compression/decompression program. 8. The TurboJPEG C API now includes a new convenience array (`tjAlphaOffset[]`) that contains the alpha component index for each pixel format (or -1 if the pixel format lacks an alpha component.) The TurboJPEG Java API now includes a new method (`TJ.getAlphaOffset()`) that returns the same value. In addition, the `tjRedOffset[]`, `tjGreenOffset[]`, and `tjBlueOffset[]` arrays-- and the corresponding `TJ.getRedOffset()`, `TJ.getGreenOffset()`, and `TJ.getBlueOffset()` methods-- now return -1 for `TJPF_GRAY`/`TJ.PF_GRAY` rather than 0. This allows programs to easily determine whether a pixel format has red, green, blue, and alpha components. 9. Added a new example (tjexample.c) that demonstrates the basic usage of the TurboJPEG C API. This example mirrors the functionality of TJExample.java. Both files are now included in the libjpeg-turbo documentation. 10. Fixed two signed integer overflows in the arithmetic decoder, detected by the Clang undefined behavior sanitizer, that could be triggered by attempting to decompress a specially-crafted malformed JPEG image. These issues did not pose a security threat, but removing the warnings makes it easier to detect actual security issues, should they arise in the future. 11. Fixed a bug in the merged 4:2:0 upsampling/dithered RGB565 color conversion algorithm that caused incorrect dithering in the output image. This algorithm now produces bitwise-identical results to the unmerged algorithms. 12. The SIMD function symbols for x86[-64]/ELF, MIPS/ELF, macOS/x86[-64] (if libjpeg-turbo is built with Yasm), and iOS/Arm[64] builds are now private. This prevents those symbols from being exposed in applications or shared libraries that link statically with libjpeg-turbo. 13. Added Loongson MMI SIMD implementations of the RGB-to-YCbCr and YCbCr-to-RGB colorspace conversion, 4:2:0 chroma downsampling, 4:2:0 fancy chroma upsampling, integer quantization, and accurate integer DCT/IDCT algorithms. When using the accurate integer DCT/IDCT, this speeds up the compression of RGB images by approximately 70-100% and the decompression of RGB images by approximately 2-3.5x. 14. Fixed a build error when building with older MinGW releases (regression caused by 1.5.1[7].) 15. Added SIMD acceleration for progressive Huffman encoding on SSE2-capable x86 and x86-64 platforms. This speeds up the compression of full-color progressive JPEGs by about 85-90% on average (relative to libjpeg-turbo 1.5.x) when using modern Intel and AMD CPUs. 1.5.3 ===== ### Significant changes relative to 1.5.2: 1. Fixed a NullPointerException in the TurboJPEG Java wrapper that occurred when using the YUVImage constructor that creates an instance backed by separate image planes and allocates memory for the image planes. 2. Fixed an issue whereby the Java version of TJUnitTest would fail when testing BufferedImage encoding/decoding on big endian systems. 3. Fixed a segfault in djpeg that would occur if an output format other than PPM/PGM was selected along with the `-crop` option. The `-crop` option now works with the GIF and Targa formats as well (unfortunately, it cannot be made to work with the BMP and RLE formats due to the fact that those output engines write scanlines in bottom-up order.) djpeg will now exit gracefully if an output format other than PPM/PGM, GIF, or Targa is selected along with the `-crop` option. 4. Fixed an issue (CVE-2017-15232) whereby `jpeg_skip_scanlines()` would segfault if color quantization was enabled. 5. TJBench (both C and Java versions) will now display usage information if any command-line argument is unrecognized. This prevents the program from silently ignoring typos. 6. Fixed an access violation in tjbench.exe (Windows) that occurred when the program was used to decompress an existing JPEG image. 7. Fixed an ArrayIndexOutOfBoundsException in the TJExample Java program that occurred when attempting to decompress a JPEG image that had been compressed with 4:1:1 chrominance subsampling. 8. Fixed an issue whereby, when using `jpeg_skip_scanlines()` to skip to the end of a single-scan (non-progressive) image, subsequent calls to `jpeg_consume_input()` would return `JPEG_SUSPENDED` rather than `JPEG_REACHED_EOI`. 9. `jpeg_crop_scanline()` now works correctly when decompressing grayscale JPEG images that were compressed with a sampling factor other than 1 (for instance, with `cjpeg -grayscale -sample 2x2`). 1.5.2 ===== ### Significant changes relative to 1.5.1: 1. Fixed a regression introduced by 1.5.1[7] that prevented libjpeg-turbo from building with Android NDK platforms prior to android-21 (5.0). 2. Fixed a regression introduced by 1.5.1[1] that prevented the MIPS DSPR2 SIMD code in libjpeg-turbo from building. 3. Fixed a regression introduced by 1.5 beta1[11] that prevented the Java version of TJBench from outputting any reference images (the `-nowrite` switch was accidentally enabled by default.) 4. libjpeg-turbo should now build and run with full AltiVec SIMD acceleration on PowerPC-based AmigaOS 4 and OpenBSD systems. 5. Fixed build and runtime errors on Windows that occurred when building libjpeg-turbo with libjpeg v7 API/ABI emulation and the in-memory source/destination managers. Due to an oversight, the `jpeg_skip_scanlines()` and `jpeg_crop_scanline()` functions were not being included in jpeg7.dll when libjpeg-turbo was built with `-DWITH_JPEG7=1` and `-DWITH_MEMSRCDST=1`. 6. Fixed "Bogus virtual array access" error that occurred when using the lossless crop feature in jpegtran or the TurboJPEG API, if libjpeg-turbo was built with libjpeg v7 API/ABI emulation. This was apparently a long-standing bug that has existed since the introduction of libjpeg v7/v8 API/ABI emulation in libjpeg-turbo v1.1. 7. The lossless transform features in jpegtran and the TurboJPEG API will now always attempt to adjust the EXIF image width and height tags if the image size changed as a result of the transform. This behavior has always existed when using libjpeg v8 API/ABI emulation. It was supposed to be available with libjpeg v7 API/ABI emulation as well but did not work properly due to a bug. Furthermore, there was never any good reason not to enable it with libjpeg v6b API/ABI emulation, since the behavior is entirely internal. Note that `-copy all` must be passed to jpegtran in order to transfer the EXIF tags from the source image to the destination image. 8. Fixed several memory leaks in the TurboJPEG API library that could occur if the library was built with certain compilers and optimization levels (known to occur with GCC 4.x and clang with `-O1` and higher but not with GCC 5.x or 6.x) and one of the underlying libjpeg API functions threw an error after a TurboJPEG API function allocated a local buffer. 9. The libjpeg-turbo memory manager will now honor the `max_memory_to_use` structure member in jpeg\_memory\_mgr, which can be set to the maximum amount of memory (in bytes) that libjpeg-turbo should use during decompression or multi-pass (including progressive) compression. This limit can also be set using the `JPEGMEM` environment variable or using the `-maxmemory` switch in cjpeg/djpeg/jpegtran (refer to the respective man pages for more details.) This has been a documented feature of libjpeg since v5, but the `malloc()`/`free()` implementation of the memory manager (jmemnobs.c) never implemented the feature. Restricting libjpeg-turbo's memory usage is useful for two reasons: it allows testers to more easily work around the 2 GB limit in libFuzzer, and it allows developers of security-sensitive applications to more easily defend against one of the progressive JPEG exploits (LJT-01-004) identified in [this report](http://www.libjpeg-turbo.org/pmwiki/uploads/About/TwoIssueswiththeJPEGStandard.pdf). 10. TJBench will now run each benchmark for 1 second prior to starting the timer, in order to improve the consistency of the results. Furthermore, the `-warmup` option is now used to specify the amount of warmup time rather than the number of warmup iterations. 11. Fixed an error (`short jump is out of range`) that occurred when assembling the 32-bit x86 SIMD extensions with NASM versions prior to 2.04. This was a regression introduced by 1.5 beta1[12]. 1.5.1 ===== ### Significant changes relative to 1.5.0: 1. Previously, the undocumented `JSIMD_FORCE*` environment variables could be used to force-enable a particular SIMD instruction set if multiple instruction sets were available on a particular platform. On x86 platforms, where CPU feature detection is bulletproof and multiple SIMD instruction sets are available, it makes sense for those environment variables to allow forcing the use of an instruction set only if that instruction set is available. However, since the ARM implementations of libjpeg-turbo can only use one SIMD instruction set, and since their feature detection code is less bulletproof (parsing /proc/cpuinfo), it makes sense for the `JSIMD_FORCENEON` environment variable to bypass the feature detection code and really force the use of NEON instructions. A new environment variable (`JSIMD_FORCEDSPR2`) was introduced in the MIPS implementation for the same reasons, and the existing `JSIMD_FORCENONE` environment variable was extended to that implementation. These environment variables provide a workaround for those attempting to test ARM and MIPS builds of libjpeg-turbo in QEMU, which passes through /proc/cpuinfo from the host system. 2. libjpeg-turbo previously assumed that AltiVec instructions were always available on PowerPC platforms, which led to "illegal instruction" errors when running on PowerPC chips that lack AltiVec support (such as the older 7xx/G3 and newer e5500 series.) libjpeg-turbo now examines /proc/cpuinfo on Linux/Android systems and enables AltiVec instructions only if the CPU supports them. It also now provides two environment variables, `JSIMD_FORCEALTIVEC` and `JSIMD_FORCENONE`, to force-enable and force-disable AltiVec instructions in environments where /proc/cpuinfo is an unreliable means of CPU feature detection (such as when running in QEMU.) On OS X, libjpeg-turbo continues to assume that AltiVec support is always available, which means that libjpeg-turbo cannot be used with G3 Macs unless you set the environment variable `JSIMD_FORCENONE` to `1`. 3. Fixed an issue whereby 64-bit ARM (AArch64) builds of libjpeg-turbo would crash when built with recent releases of the Clang/LLVM compiler. This was caused by an ABI conformance issue in some of libjpeg-turbo's 64-bit NEON SIMD routines. Those routines were incorrectly using 64-bit instructions to transfer a 32-bit JDIMENSION argument, whereas the ABI allows the upper (unused) 32 bits of a 32-bit argument's register to be undefined. The new Clang/LLVM optimizer uses load combining to transfer multiple adjacent 32-bit structure members into a single 64-bit register, and this exposed the ABI conformance issue. 4. Fancy upsampling is now supported when decompressing JPEG images that use 4:4:0 (h1v2) chroma subsampling. These images are generated when losslessly rotating or transposing JPEG images that use 4:2:2 (h2v1) chroma subsampling. The h1v2 fancy upsampling algorithm is not currently SIMD-accelerated. 5. If merged upsampling isn't SIMD-accelerated but YCbCr-to-RGB conversion is, then libjpeg-turbo will now disable merged upsampling when decompressing YCbCr JPEG images into RGB or extended RGB output images. This significantly speeds up the decompression of 4:2:0 and 4:2:2 JPEGs on ARM platforms if fancy upsampling is not used (for example, if the `-nosmooth` option to djpeg is specified.) 6. The TurboJPEG API will now decompress 4:2:2 and 4:4:0 JPEG images with 2x2 luminance sampling factors and 2x1 or 1x2 chrominance sampling factors. This is a non-standard way of specifying 2x subsampling (normally 4:2:2 JPEGs have 2x1 luminance and 1x1 chrominance sampling factors, and 4:4:0 JPEGs have 1x2 luminance and 1x1 chrominance sampling factors), but the JPEG format and the libjpeg API both allow it. 7. Fixed an unsigned integer overflow in the libjpeg memory manager, detected by the Clang undefined behavior sanitizer, that could be triggered by attempting to decompress a specially-crafted malformed JPEG image. This issue affected only 32-bit code and did not pose a security threat, but removing the warning makes it easier to detect actual security issues, should they arise in the future. 8. Fixed additional negative left shifts and other issues reported by the GCC and Clang undefined behavior sanitizers when attempting to decompress specially-crafted malformed JPEG images. None of these issues posed a security threat, but removing the warnings makes it easier to detect actual security issues, should they arise in the future. 9. Fixed an out-of-bounds array reference, introduced by 1.4.90[2] (partial image decompression) and detected by the Clang undefined behavior sanitizer, that could be triggered by a specially-crafted malformed JPEG image with more than four components. Because the out-of-bounds reference was still within the same structure, it was not known to pose a security threat, but removing the warning makes it easier to detect actual security issues, should they arise in the future. 10. Fixed another ABI conformance issue in the 64-bit ARM (AArch64) NEON SIMD code. Some of the routines were incorrectly reading and storing data below the stack pointer, which caused segfaults in certain applications under specific circumstances. 1.5.0 ===== ### Significant changes relative to 1.5 beta1: 1. Fixed an issue whereby a malformed motion-JPEG frame could cause the "fast path" of libjpeg-turbo's Huffman decoder to read from uninitialized memory. 2. Added libjpeg-turbo version and build information to the global string table of the libjpeg and TurboJPEG API libraries. This is a common practice in other infrastructure libraries, such as OpenSSL and libpng, because it makes it easy to examine an application binary and determine which version of the library the application was linked against. 3. Fixed a couple of issues in the PPM reader that would cause buffer overruns in cjpeg if one of the values in a binary PPM/PGM input file exceeded the maximum value defined in the file's header and that maximum value was greater than 255. libjpeg-turbo 1.4.2 already included a similar fix for ASCII PPM/PGM files. Note that these issues were not security bugs, since they were confined to the cjpeg program and did not affect any of the libjpeg-turbo libraries. 4. Fixed an issue whereby attempting to decompress a JPEG file with a corrupt header using the `tjDecompressToYUV2()` function would cause the function to abort without returning an error and, under certain circumstances, corrupt the stack. This only occurred if `tjDecompressToYUV2()` was called prior to calling `tjDecompressHeader3()`, or if the return value from `tjDecompressHeader3()` was ignored (both cases represent incorrect usage of the TurboJPEG API.) 5. Fixed an issue in the ARM 32-bit SIMD-accelerated Huffman encoder that prevented the code from assembling properly with clang. 6. The `jpeg_stdio_src()`, `jpeg_mem_src()`, `jpeg_stdio_dest()`, and `jpeg_mem_dest()` functions in the libjpeg API will now throw an error if a source/destination manager has already been assigned to the compress or decompress object by a different function or by the calling program. This prevents these functions from attempting to reuse a source/destination manager structure that was allocated elsewhere, because there is no way to ensure that it would be big enough to accommodate the new source/destination manager. 1.4.90 (1.5 beta1) ================== ### Significant changes relative to 1.4.2: 1. Added full SIMD acceleration for PowerPC platforms using AltiVec VMX (128-bit SIMD) instructions. Although the performance of libjpeg-turbo on PowerPC was already good, due to the increased number of registers available to the compiler vs. x86, it was still possible to speed up compression by about 3-4x and decompression by about 2-2.5x (relative to libjpeg v6b) through the use of AltiVec instructions. 2. Added two new libjpeg API functions (`jpeg_skip_scanlines()` and `jpeg_crop_scanline()`) that can be used to partially decode a JPEG image. See [libjpeg.txt](libjpeg.txt) for more details. 3. The TJCompressor and TJDecompressor classes in the TurboJPEG Java API now implement the Closeable interface, so those classes can be used with a try-with-resources statement. 4. The TurboJPEG Java classes now throw unchecked idiomatic exceptions (IllegalArgumentException, IllegalStateException) for unrecoverable errors caused by incorrect API usage, and those classes throw a new checked exception type (TJException) for errors that are passed through from the C library. 5. Source buffers for the TurboJPEG C API functions, as well as the `jpeg_mem_src()` function in the libjpeg API, are now declared as const pointers. This facilitates passing read-only buffers to those functions and ensures the caller that the source buffer will not be modified. This should not create any backward API or ABI incompatibilities with prior libjpeg-turbo releases. 6. The MIPS DSPr2 SIMD code can now be compiled to support either FR=0 or FR=1 FPUs. 7. Fixed additional negative left shifts and other issues reported by the GCC and Clang undefined behavior sanitizers. Most of these issues affected only 32-bit code, and none of them was known to pose a security threat, but removing the warnings makes it easier to detect actual security issues, should they arise in the future. 8. Removed the unnecessary `.arch` directive from the ARM64 NEON SIMD code. This directive was preventing the code from assembling using the clang integrated assembler. 9. Fixed a regression caused by 1.4.1[6] that prevented 32-bit and 64-bit libjpeg-turbo RPMs from being installed simultaneously on recent Red Hat/Fedora distributions. This was due to the addition of a macro in jconfig.h that allows the Huffman codec to determine the word size at compile time. Since that macro differs between 32-bit and 64-bit builds, this caused a conflict between the i386 and x86_64 RPMs (any differing files, other than executables, are not allowed when 32-bit and 64-bit RPMs are installed simultaneously.) Since the macro is used only internally, it has been moved into jconfigint.h. 10. The x86-64 SIMD code can now be disabled at run time by setting the `JSIMD_FORCENONE` environment variable to `1` (the other SIMD implementations already had this capability.) 11. Added a new command-line argument to TJBench (`-nowrite`) that prevents the benchmark from outputting any images. This removes any potential operating system overhead that might be caused by lazy writes to disk and thus improves the consistency of the performance measurements. 12. Added SIMD acceleration for Huffman encoding on SSE2-capable x86 and x86-64 platforms. This speeds up the compression of full-color JPEGs by about 10-15% on average (relative to libjpeg-turbo 1.4.x) when using modern Intel and AMD CPUs. Additionally, this works around an issue in the clang optimizer that prevents it (as of this writing) from achieving the same performance as GCC when compiling the C version of the Huffman encoder (). For the purposes of benchmarking or regression testing, SIMD-accelerated Huffman encoding can be disabled by setting the `JSIMD_NOHUFFENC` environment variable to `1`. 13. Added ARM 64-bit (ARMv8) NEON SIMD implementations of the commonly-used compression algorithms (including the accurate integer forward DCT and h2v2 & h2v1 downsampling algorithms, which are not accelerated in the 32-bit NEON implementation.) This speeds up the compression of full-color JPEGs by about 75% on average on a Cavium ThunderX processor and by about 2-2.5x on average on Cortex-A53 and Cortex-A57 cores. 14. Added SIMD acceleration for Huffman encoding on NEON-capable ARM 32-bit and 64-bit platforms. For 32-bit code, this speeds up the compression of full-color JPEGs by about 30% on average on a typical iOS device (iPhone 4S, Cortex-A9) and by about 6-7% on average on a typical Android device (Nexus 5X, Cortex-A53 and Cortex-A57), relative to libjpeg-turbo 1.4.x. Note that the larger speedup under iOS is due to the fact that iOS builds use LLVM, which does not optimize the C Huffman encoder as well as GCC does. For 64-bit code, NEON-accelerated Huffman encoding speeds up the compression of full-color JPEGs by about 40% on average on a typical iOS device (iPhone 5S, Apple A7) and by about 7-8% on average on a typical Android device (Nexus 5X, Cortex-A53 and Cortex-A57), in addition to the speedup described in [13] above. For the purposes of benchmarking or regression testing, SIMD-accelerated Huffman encoding can be disabled by setting the `JSIMD_NOHUFFENC` environment variable to `1`. 15. pkg-config (.pc) scripts are now included for both the libjpeg and TurboJPEG API libraries on Un*x systems. Note that if a project's build system relies on these scripts, then it will not be possible to build that project with libjpeg or with a prior version of libjpeg-turbo. 16. Optimized the ARM 64-bit (ARMv8) NEON SIMD decompression routines to improve performance on CPUs with in-order pipelines. This speeds up the decompression of full-color JPEGs by nearly 2x on average on a Cavium ThunderX processor and by about 15% on average on a Cortex-A53 core. 17. Fixed an issue in the accelerated Huffman decoder that could have caused the decoder to read past the end of the input buffer when a malformed, specially-crafted JPEG image was being decompressed. In prior versions of libjpeg-turbo, the accelerated Huffman decoder was invoked (in most cases) only if there were > 128 bytes of data in the input buffer. However, it is possible to construct a JPEG image in which a single Huffman block is over 430 bytes long, so this version of libjpeg-turbo activates the accelerated Huffman decoder only if there are > 512 bytes of data in the input buffer. 18. Fixed a memory leak in tjunittest encountered when running the program with the `-yuv` option. 1.4.2 ===== ### Significant changes relative to 1.4.1: 1. Fixed an issue whereby cjpeg would segfault if a Windows bitmap with a negative width or height was used as an input image (Windows bitmaps can have a negative height if they are stored in top-down order, but such files are rare and not supported by libjpeg-turbo.) 2. Fixed an issue whereby, under certain circumstances, libjpeg-turbo would incorrectly encode certain JPEG images when quality=100 and the fast integer forward DCT were used. This was known to cause `make test` to fail when the library was built with `-march=haswell` on x86 systems. 3. Fixed an issue whereby libjpeg-turbo would crash when built with the latest & greatest development version of the Clang/LLVM compiler. This was caused by an x86-64 ABI conformance issue in some of libjpeg-turbo's 64-bit SSE2 SIMD routines. Those routines were incorrectly using a 64-bit `mov` instruction to transfer a 32-bit JDIMENSION argument, whereas the x86-64 ABI allows the upper (unused) 32 bits of a 32-bit argument's register to be undefined. The new Clang/LLVM optimizer uses load combining to transfer multiple adjacent 32-bit structure members into a single 64-bit register, and this exposed the ABI conformance issue. 4. Fixed a bug in the MIPS DSPr2 4:2:0 "plain" (non-fancy and non-merged) upsampling routine that caused a buffer overflow (and subsequent segfault) when decompressing a 4:2:0 JPEG image whose scaled output width was less than 16 pixels. The "plain" upsampling routines are normally only used when decompressing a non-YCbCr JPEG image, but they are also used when decompressing a JPEG image whose scaled output height is 1. 5. Fixed various negative left shifts and other issues reported by the GCC and Clang undefined behavior sanitizers. None of these was known to pose a security threat, but removing the warnings makes it easier to detect actual security issues, should they arise in the future. 1.4.1 ===== ### Significant changes relative to 1.4.0: 1. tjbench now properly handles CMYK/YCCK JPEG files. Passing an argument of `-cmyk` (instead of, for instance, `-rgb`) will cause tjbench to internally convert the source bitmap to CMYK prior to compression, to generate YCCK JPEG files, and to internally convert the decompressed CMYK pixels back to RGB after decompression (the latter is done automatically if a CMYK or YCCK JPEG is passed to tjbench as a source image.) The CMYK<->RGB conversion operation is not benchmarked. NOTE: The quick & dirty CMYK<->RGB conversions that tjbench uses are suitable for testing only. Proper conversion between CMYK and RGB requires a color management system. 2. `make test` now performs additional bitwise regression tests using tjbench, mainly for the purpose of testing compression from/decompression to a subregion of a larger image buffer. 3. `make test` no longer tests the regression of the floating point DCT/IDCT by default, since the results of those tests can vary if the algorithms in question are not implemented using SIMD instructions on a particular platform. See the comments in [Makefile.am](Makefile.am) for information on how to re-enable the tests and to specify an expected result for them based on the particulars of your platform. 4. The NULL color conversion routines have been significantly optimized, which speeds up the compression of RGB and CMYK JPEGs by 5-20% when using 64-bit code and 0-3% when using 32-bit code, and the decompression of those images by 10-30% when using 64-bit code and 3-12% when using 32-bit code. 5. Fixed an "illegal instruction" error that occurred when djpeg from a SIMD-enabled libjpeg-turbo MIPS build was executed with the `-nosmooth` option on a MIPS machine that lacked DSPr2 support. The MIPS SIMD routines for h2v1 and h2v2 merged upsampling were not properly checking for the existence of DSPr2. 6. Performance has been improved significantly on 64-bit non-Linux and non-Windows platforms (generally 10-20% faster compression and 5-10% faster decompression.) Due to an oversight, the 64-bit version of the accelerated Huffman codec was not being compiled in when libjpeg-turbo was built on platforms other than Windows or Linux. Oops. 7. Fixed an extremely rare bug in the Huffman encoder that caused 64-bit builds of libjpeg-turbo to incorrectly encode a few specific test images when quality=98, an optimized Huffman table, and the accurate integer forward DCT were used. 8. The Windows (CMake) build system now supports building only static or only shared libraries. This is accomplished by adding either `-DENABLE_STATIC=0` or `-DENABLE_SHARED=0` to the CMake command line. 9. TurboJPEG API functions will now return an error code if a warning is triggered in the underlying libjpeg API. For instance, if a JPEG file is corrupt, the TurboJPEG decompression functions will attempt to decompress as much of the image as possible, but those functions will now return -1 to indicate that the decompression was not entirely successful. 10. Fixed a bug in the MIPS DSPr2 4:2:2 fancy upsampling routine that caused a buffer overflow (and subsequent segfault) when decompressing a 4:2:2 JPEG image in which the right-most MCU was 5 or 6 pixels wide. 1.4.0 ===== ### Significant changes relative to 1.4 beta1: 1. Fixed a build issue on OS X PowerPC platforms (md5cmp failed to build because OS X does not provide the `le32toh()` and `htole32()` functions.) 2. The non-SIMD RGB565 color conversion code did not work correctly on big endian machines. This has been fixed. 3. Fixed an issue in `tjPlaneSizeYUV()` whereby it would erroneously return 1 instead of -1 if `componentID` was > 0 and `subsamp` was `TJSAMP_GRAY`. 3. Fixed an issue in `tjBufSizeYUV2()` whereby it would erroneously return 0 instead of -1 if `width` was < 1. 5. The Huffman encoder now uses `clz` and `bsr` instructions for bit counting on ARM64 platforms (see 1.4 beta1[5].) 6. The `close()` method in the TJCompressor and TJDecompressor Java classes is now idempotent. Previously, that method would call the native `tjDestroy()` function even if the TurboJPEG instance had already been destroyed. This caused an exception to be thrown during finalization, if the `close()` method had already been called. The exception was caught, but it was still an expensive operation. 7. The TurboJPEG API previously generated an error (`Could not determine subsampling type for JPEG image`) when attempting to decompress grayscale JPEG images that were compressed with a sampling factor other than 1 (for instance, with `cjpeg -grayscale -sample 2x2`). Subsampling technically has no meaning with grayscale JPEGs, and thus the horizontal and vertical sampling factors for such images are ignored by the decompressor. However, the TurboJPEG API was being too rigid and was expecting the sampling factors to be equal to 1 before it treated the image as a grayscale JPEG. 8. cjpeg, djpeg, and jpegtran now accept an argument of `-version`, which will print the library version and exit. 9. Referring to 1.4 beta1[15], another extremely rare circumstance was discovered under which the Huffman encoder's local buffer can be overrun when a buffered destination manager is being used and an extremely-high-frequency block (basically junk image data) is being encoded. Even though the Huffman local buffer was increased from 128 bytes to 136 bytes to address the previous issue, the new issue caused even the larger buffer to be overrun. Further analysis reveals that, in the absolute worst case (such as setting alternating AC coefficients to 32767 and -32768 in the JPEG scanning order), the Huffman encoder can produce encoded blocks that approach double the size of the unencoded blocks. Thus, the Huffman local buffer was increased to 256 bytes, which should prevent any such issue from re-occurring in the future. 10. The new `tjPlaneSizeYUV()`, `tjPlaneWidth()`, and `tjPlaneHeight()` functions were not actually usable on any platform except OS X and Windows, because those functions were not included in the libturbojpeg mapfile. This has been fixed. 11. Restored the `JPP()`, `JMETHOD()`, and `FAR` macros in the libjpeg-turbo header files. The `JPP()` and `JMETHOD()` macros were originally implemented in libjpeg as a way of supporting non-ANSI compilers that lacked support for prototype parameters. libjpeg-turbo has never supported such compilers, but some software packages still use the macros to define their own prototypes. Similarly, libjpeg-turbo has never supported MS-DOS and other platforms that have far symbols, but some software packages still use the `FAR` macro. A pretty good argument can be made that this is a bad practice on the part of the software in question, but since this affects more than one package, it's just easier to fix it here. 12. Fixed issues that were preventing the ARM 64-bit SIMD code from compiling for iOS, and included an ARMv8 architecture in all of the binaries installed by the "official" libjpeg-turbo SDK for OS X. 1.3.90 (1.4 beta1) ================== ### Significant changes relative to 1.3.1: 1. New features in the TurboJPEG API: - YUV planar images can now be generated with an arbitrary line padding (previously only 4-byte padding, which was compatible with X Video, was supported.) - The decompress-to-YUV function has been extended to support image scaling. - JPEG images can now be compressed from YUV planar source images. - YUV planar images can now be decoded into RGB or grayscale images. - 4:1:1 subsampling is now supported. This is mainly included for compatibility, since 4:1:1 is not fully accelerated in libjpeg-turbo and has no significant advantages relative to 4:2:0. - CMYK images are now supported. This feature allows CMYK source images to be compressed to YCCK JPEGs and YCCK or CMYK JPEGs to be decompressed to CMYK destination images. Conversion between CMYK/YCCK and RGB or YUV images is not supported. Such conversion requires a color management system and is thus out of scope for a codec library. - The handling of YUV images in the Java API has been significantly refactored and should now be much more intuitive. - The Java API now supports encoding a YUV image from an arbitrary position in a large image buffer. - All of the YUV functions now have a corresponding function that operates on separate image planes instead of a unified image buffer. This allows for compressing/decoding from or decompressing/encoding to a subregion of a larger YUV image. It also allows for handling YUV formats that swap the order of the U and V planes. 2. Added SIMD acceleration for DSPr2-capable MIPS platforms. This speeds up the compression of full-color JPEGs by 70-80% on such platforms and decompression by 25-35%. 3. If an application attempts to decompress a Huffman-coded JPEG image whose header does not contain Huffman tables, libjpeg-turbo will now insert the default Huffman tables. In order to save space, many motion JPEG video frames are encoded without the default Huffman tables, so these frames can now be successfully decompressed by libjpeg-turbo without additional work on the part of the application. An application can still override the Huffman tables, for instance to re-use tables from a previous frame of the same video. 4. The Mac packaging system now uses pkgbuild and productbuild rather than PackageMaker (which is obsolete and no longer supported.) This means that OS X 10.6 "Snow Leopard" or later must be used when packaging libjpeg-turbo, although the packages produced can be installed on OS X 10.5 "Leopard" or later. OS X 10.4 "Tiger" is no longer supported. 5. The Huffman encoder now uses `clz` and `bsr` instructions for bit counting on ARM platforms rather than a lookup table. This reduces the memory footprint by 64k, which may be important for some mobile applications. Out of four Android devices that were tested, two demonstrated a small overall performance loss (~3-4% on average) with ARMv6 code and a small gain (also ~3-4%) with ARMv7 code when enabling this new feature, but the other two devices demonstrated a significant overall performance gain with both ARMv6 and ARMv7 code (~10-20%) when enabling the feature. Actual mileage may vary. 6. Worked around an issue with Visual C++ 2010 and later that caused incorrect pixels to be generated when decompressing a JPEG image to a 256-color bitmap, if compiler optimization was enabled when libjpeg-turbo was built. This caused the regression tests to fail when doing a release build under Visual C++ 2010 and later. 7. Improved the accuracy and performance of the non-SIMD implementation of the floating point inverse DCT (using code borrowed from libjpeg v8a and later.) The accuracy of this implementation now matches the accuracy of the SSE/SSE2 implementation. Note, however, that the floating point DCT/IDCT algorithms are mainly a legacy feature. They generally do not produce significantly better accuracy than the accurate integer DCT/IDCT algorithms, and they are quite a bit slower. 8. Added a new output colorspace (`JCS_RGB565`) to the libjpeg API that allows for decompressing JPEG images into RGB565 (16-bit) pixels. If dithering is not used, then this code path is SIMD-accelerated on ARM platforms. 9. Numerous obsolete features, such as support for non-ANSI compilers and support for the MS-DOS memory model, were removed from the libjpeg code, greatly improving its readability and making it easier to maintain and extend. 10. Fixed a segfault that occurred when calling `output_message()` with `msg_code` set to `JMSG_COPYRIGHT`. 11. Fixed an issue whereby wrjpgcom was allowing comments longer than 65k characters to be passed on the command line, which was causing it to generate incorrect JPEG files. 12. Fixed a bug in the build system that was causing the Windows version of wrjpgcom to be built using the rdjpgcom source code. 13. Restored 12-bit-per-component JPEG support. A 12-bit version of libjpeg-turbo can now be built by passing an argument of `--with-12bit` to configure (Unix) or `-DWITH_12BIT=1` to cmake (Windows.) 12-bit JPEG support is included only for convenience. Enabling this feature disables all of the performance features in libjpeg-turbo, as well as arithmetic coding and the TurboJPEG API. The resulting library still contains the other libjpeg-turbo features (such as the colorspace extensions), but in general, it performs no faster than libjpeg v6b. 14. Added ARM 64-bit SIMD acceleration for the YCC-to-RGB color conversion and IDCT algorithms (both are used during JPEG decompression.) For unknown reasons (probably related to clang), this code cannot currently be compiled for iOS. 15. Fixed an extremely rare bug (CVE-2014-9092) that could cause the Huffman encoder's local buffer to overrun when a very high-frequency MCU is compressed using quality 100 and no subsampling, and when the JPEG output buffer is being dynamically resized by the destination manager. This issue was so rare that, even with a test program specifically designed to make the bug occur (by injecting random high-frequency YUV data into the compressor), it was reproducible only once in about every 25 million iterations. 16. Fixed an oversight in the TurboJPEG C wrapper: if any of the JPEG compression functions was called repeatedly with the same automatically-allocated destination buffer, then TurboJPEG would erroneously assume that the `jpegSize` parameter was equal to the size of the buffer, when in fact that parameter was probably equal to the size of the most recently compressed JPEG image. If the size of the previous JPEG image was not as large as the current JPEG image, then TurboJPEG would unnecessarily reallocate the destination buffer. 1.3.1 ===== ### Significant changes relative to 1.3.0: 1. On Un*x systems, `make install` now installs the libjpeg-turbo libraries into /opt/libjpeg-turbo/lib32 by default on any 32-bit system, not just x86, and into /opt/libjpeg-turbo/lib64 by default on any 64-bit system, not just x86-64. You can override this by overriding either the `prefix` or `libdir` configure variables. 2. The Windows installer now places a copy of the TurboJPEG DLLs in the same directory as the rest of the libjpeg-turbo binaries. This was mainly done to support TurboVNC 1.3, which bundles the DLLs in its Windows installation. When using a 32-bit version of CMake on 64-bit Windows, it is impossible to access the c:\WINDOWS\system32 directory, which made it impossible for the TurboVNC build scripts to bundle the 64-bit TurboJPEG DLL. 3. Fixed a bug whereby attempting to encode a progressive JPEG with arithmetic entropy coding (by passing arguments of `-progressive -arithmetic` to cjpeg or jpegtran, for instance) would result in an error, `Requested feature was omitted at compile time`. 4. Fixed a couple of issues (CVE-2013-6629 and CVE-2013-6630) whereby malformed JPEG images would cause libjpeg-turbo to use uninitialized memory during decompression. 5. Fixed an error (`Buffer passed to JPEG library is too small`) that occurred when calling the TurboJPEG YUV encoding function with a very small (< 5x5) source image, and added a unit test to check for this error. 6. The Java classes should now build properly under Visual Studio 2010 and later. 7. Fixed an issue that prevented SRPMs generated using the in-tree packaging tools from being rebuilt on certain newer Linux distributions. 8. Numerous minor fixes to eliminate compilation and build/packaging system warnings, fix cosmetic issues, improve documentation clarity, and other general source cleanup. 1.3.0 ===== ### Significant changes relative to 1.3 beta1: 1. `make test` now works properly on FreeBSD, and it no longer requires the md5sum executable to be present on other Un*x platforms. 2. Overhauled the packaging system: - To avoid conflict with vendor-supplied libjpeg-turbo packages, the official RPMs and DEBs for libjpeg-turbo have been renamed to "libjpeg-turbo-official". - The TurboJPEG libraries are now located under /opt/libjpeg-turbo in the official Linux and Mac packages, to avoid conflict with vendor-supplied packages and also to streamline the packaging system. - Release packages are now created with the directory structure defined by the configure variables `prefix`, `bindir`, `libdir`, etc. (Un\*x) or by the `CMAKE_INSTALL_PREFIX` variable (Windows.) The exception is that the docs are always located under the system default documentation directory on Un\*x and Mac systems, and on Windows, the TurboJPEG DLL is always located in the Windows system directory. - To avoid confusion, official libjpeg-turbo packages on Linux/Unix platforms (except for Mac) will always install the 32-bit libraries in /opt/libjpeg-turbo/lib32 and the 64-bit libraries in /opt/libjpeg-turbo/lib64. - Fixed an issue whereby, in some cases, the libjpeg-turbo executables on Un*x systems were not properly linking with the shared libraries installed by the same package. - Fixed an issue whereby building the "installer" target on Windows when `WITH_JAVA=1` would fail if the TurboJPEG JAR had not been previously built. - Building the "install" target on Windows now installs files into the same places that the installer does. 3. Fixed a Huffman encoder bug that prevented I/O suspension from working properly. 1.2.90 (1.3 beta1) ================== ### Significant changes relative to 1.2.1: 1. Added support for additional scaling factors (3/8, 5/8, 3/4, 7/8, 9/8, 5/4, 11/8, 3/2, 13/8, 7/4, 15/8, and 2) when decompressing. Note that the IDCT will not be SIMD-accelerated when using any of these new scaling factors. 2. The TurboJPEG dynamic library is now versioned. It was not strictly necessary to do so, because TurboJPEG uses versioned symbols, and if a function changes in an ABI-incompatible way, that function is renamed and a legacy function is provided to maintain backward compatibility. However, certain Linux distro maintainers have a policy against accepting any library that isn't versioned. 3. Extended the TurboJPEG Java API so that it can be used to compress a JPEG image from and decompress a JPEG image to an arbitrary position in a large image buffer. 4. The `tjDecompressToYUV()` function now supports the `TJFLAG_FASTDCT` flag. 5. The 32-bit supplementary package for amd64 Debian systems now provides symlinks in /usr/lib/i386-linux-gnu for the TurboJPEG libraries in /usr/lib32. This allows those libraries to be used on MultiArch-compatible systems (such as Ubuntu 11 and later) without setting the linker path. 6. The TurboJPEG Java wrapper should now find the JNI library on Mac systems without having to pass `-Djava.library.path=/usr/lib` to java. 7. TJBench has been ported to Java to provide a convenient way of validating the performance of the TurboJPEG Java API. It can be run with `java -cp turbojpeg.jar TJBench`. 8. cjpeg can now be used to generate JPEG files with the RGB colorspace (feature ported from jpeg-8d.) 9. The width and height in the `-crop` argument passed to jpegtran can now be suffixed with `f` to indicate that, when the upper left corner of the cropping region is automatically moved to the nearest iMCU boundary, the bottom right corner should be moved by the same amount. In other words, this feature causes jpegtran to strictly honor the specified width/height rather than the specified bottom right corner (feature ported from jpeg-8d.) 10. JPEG files using the RGB colorspace can now be decompressed into grayscale images (feature ported from jpeg-8d.) 11. Fixed a regression caused by 1.2.1[7] whereby the build would fail with multiple "Mismatch in operand sizes" errors when attempting to build the x86 SIMD code with NASM 0.98. 12. The in-memory source/destination managers (`jpeg_mem_src()` and `jpeg_mem_dest()`) are now included by default when building libjpeg-turbo with libjpeg v6b or v7 emulation, so that programs can take advantage of these functions without requiring the use of the backward-incompatible libjpeg v8 ABI. The "age number" of the libjpeg-turbo library on Un*x systems has been incremented by 1 to reflect this. You can disable this feature with a configure/CMake switch in order to retain strict API/ABI compatibility with the libjpeg v6b or v7 API/ABI (or with previous versions of libjpeg-turbo.) See [README.md](README.md) for more details. 13. Added ARMv7s architecture to libjpeg.a and libturbojpeg.a in the official libjpeg-turbo binary package for OS X, so that those libraries can be used to build applications that leverage the faster CPUs in the iPhone 5 and iPad 4. 1.2.1 ===== ### Significant changes relative to 1.2.0: 1. Creating or decoding a JPEG file that uses the RGB colorspace should now properly work when the input or output colorspace is one of the libjpeg-turbo colorspace extensions. 2. When libjpeg-turbo was built without SIMD support and merged (non-fancy) upsampling was used along with an alpha-enabled colorspace during decompression, the unused byte of the decompressed pixels was not being set to 0xFF. This has been fixed. TJUnitTest has also been extended to test for the correct behavior of the colorspace extensions when merged upsampling is used. 3. Fixed a bug whereby the libjpeg-turbo SSE2 SIMD code would not preserve the upper 64 bits of xmm6 and xmm7 on Win64 platforms, which violated the Win64 calling conventions. 4. Fixed a regression (CVE-2012-2806) caused by 1.2.0[6] whereby decompressing corrupt JPEG images (specifically, images in which the component count was erroneously set to a large value) would cause libjpeg-turbo to segfault. 5. Worked around a severe performance issue with "Bobcat" (AMD Embedded APU) processors. The `MASKMOVDQU` instruction, which was used by the libjpeg-turbo SSE2 SIMD code, is apparently implemented in microcode on AMD processors, and it is painfully slow on Bobcat processors in particular. Eliminating the use of this instruction improved performance by an order of magnitude on Bobcat processors and by a small amount (typically 5%) on AMD desktop processors. 6. Added SIMD acceleration for performing 4:2:2 upsampling on NEON-capable ARM platforms. This speeds up the decompression of 4:2:2 JPEGs by 20-25% on such platforms. 7. Fixed a regression caused by 1.2.0[2] whereby, on Linux/x86 platforms running the 32-bit SSE2 SIMD code in libjpeg-turbo, decompressing a 4:2:0 or 4:2:2 JPEG image into a 32-bit (RGBX, BGRX, etc.) buffer without using fancy upsampling would produce several incorrect columns of pixels at the right-hand side of the output image if each row in the output image was not evenly divisible by 16 bytes. 8. Fixed an issue whereby attempting to build the SIMD extensions with Xcode 4.3 on OS X platforms would cause NASM to return numerous errors of the form "'%define' expects a macro identifier". 9. Added flags to the TurboJPEG API that allow the caller to force the use of either the fast or the accurate DCT/IDCT algorithms in the underlying codec. 1.2.0 ===== ### Significant changes relative to 1.2 beta1: 1. Fixed build issue with Yasm on Unix systems (the libjpeg-turbo build system was not adding the current directory to the assembler include path, so Yasm was not able to find jsimdcfg.inc.) 2. Fixed out-of-bounds read in SSE2 SIMD code that occurred when decompressing a JPEG image to a bitmap buffer whose size was not a multiple of 16 bytes. This was more of an annoyance than an actual bug, since it did not cause any actual run-time problems, but the issue showed up when running libjpeg-turbo in valgrind. See for more information. 3. Added a compile-time macro (`LIBJPEG_TURBO_VERSION`) that can be used to check the version of libjpeg-turbo against which an application was compiled. 4. Added new RGBA/BGRA/ABGR/ARGB colorspace extension constants (libjpeg API) and pixel formats (TurboJPEG API), which allow applications to specify that, when decompressing to a 4-component RGB buffer, the unused byte should be set to 0xFF so that it can be interpreted as an opaque alpha channel. 5. Fixed regression issue whereby DevIL failed to build against libjpeg-turbo because libjpeg-turbo's distributed version of jconfig.h contained an `INLINE` macro, which conflicted with a similar macro in DevIL. This macro is used only internally when building libjpeg-turbo, so it was moved into config.h. 6. libjpeg-turbo will now correctly decompress erroneous CMYK/YCCK JPEGs whose K component is assigned a component ID of 1 instead of 4. Although these files are in violation of the spec, other JPEG implementations handle them correctly. 7. Added ARMv6 and ARMv7 architectures to libjpeg.a and libturbojpeg.a in the official libjpeg-turbo binary package for OS X, so that those libraries can be used to build both OS X and iOS applications. 1.1.90 (1.2 beta1) ================== ### Significant changes relative to 1.1.1: 1. Added a Java wrapper for the TurboJPEG API. See [java/README](java/README) for more details. 2. The TurboJPEG API can now be used to scale down images during decompression. 3. Added SIMD routines for RGB-to-grayscale color conversion, which significantly improves the performance of grayscale JPEG compression from an RGB source image. 4. Improved the performance of the C color conversion routines, which are used on platforms for which SIMD acceleration is not available. 5. Added a function to the TurboJPEG API that performs lossless transforms. This function is implemented using the same back end as jpegtran, but it performs transcoding entirely in memory and allows multiple transforms and/or crop operations to be batched together, so the source coefficients only need to be read once. This is useful when generating image tiles from a single source JPEG. 6. Added tests for the new TurboJPEG scaled decompression and lossless transform features to tjbench (the TurboJPEG benchmark, formerly called "jpgtest".) 7. Added support for 4:4:0 (transposed 4:2:2) subsampling in TurboJPEG, which was necessary in order for it to read 4:2:2 JPEG files that had been losslessly transposed or rotated 90 degrees. 8. All legacy VirtualGL code has been re-factored, and this has allowed libjpeg-turbo, in its entirety, to be re-licensed under a BSD-style license. 9. libjpeg-turbo can now be built with Yasm. 10. Added SIMD acceleration for ARM Linux and iOS platforms that support NEON instructions. 11. Refactored the TurboJPEG C API and documented it using Doxygen. The TurboJPEG 1.2 API uses pixel formats to define the size and component order of the uncompressed source/destination images, and it includes a more efficient version of `TJBUFSIZE()` that computes a worst-case JPEG size based on the level of chrominance subsampling. The refactored implementation of the TurboJPEG API now uses the libjpeg memory source and destination managers, which allows the TurboJPEG compressor to grow the JPEG buffer as necessary. 12. Eliminated errors in the output of jpegtran on Windows that occurred when the application was invoked using I/O redirection (`jpegtran output.jpg`.) 13. The inclusion of libjpeg v7 and v8 emulation as well as arithmetic coding support in libjpeg-turbo v1.1.0 introduced several new error constants in jerror.h, and these were mistakenly enabled for all emulation modes, causing the error enum in libjpeg-turbo to sometimes have different values than the same enum in libjpeg. This represents an ABI incompatibility, and it caused problems with rare applications that took specific action based on a particular error value. The fix was to include the new error constants conditionally based on whether libjpeg v7 or v8 emulation was enabled. 14. Fixed an issue whereby Windows applications that used libjpeg-turbo would fail to compile if the Windows system headers were included before jpeglib.h. This issue was caused by a conflict in the definition of the INT32 type. 15. Fixed 32-bit supplementary package for amd64 Debian systems, which was broken by enhancements to the packaging system in 1.1. 16. When decompressing a JPEG image using an output colorspace of `JCS_EXT_RGBX`, `JCS_EXT_BGRX`, `JCS_EXT_XBGR`, or `JCS_EXT_XRGB`, libjpeg-turbo will now set the unused byte to 0xFF, which allows applications to interpret that byte as an alpha channel (0xFF = opaque). 1.1.1 ===== ### Significant changes relative to 1.1.0: 1. Fixed a 1-pixel error in row 0, column 21 of the luminance plane generated by `tjEncodeYUV()`. 2. libjpeg-turbo's accelerated Huffman decoder previously ignored unexpected markers found in the middle of the JPEG data stream during decompression. It will now hand off decoding of a particular block to the unaccelerated Huffman decoder if an unexpected marker is found, so that the unaccelerated Huffman decoder can generate an appropriate warning. 3. Older versions of MinGW64 prefixed symbol names with underscores by default, which differed from the behavior of 64-bit Visual C++. MinGW64 1.0 has adopted the behavior of 64-bit Visual C++ as the default, so to accommodate this, the libjpeg-turbo SIMD function names are no longer prefixed with an underscore when building with MinGW64. This means that, when building libjpeg-turbo with older versions of MinGW64, you will now have to add `-fno-leading-underscore` to the `CFLAGS`. 4. Fixed a regression bug in the NSIS script that caused the Windows installer build to fail when using the Visual Studio IDE. 5. Fixed a bug in `jpeg_read_coefficients()` whereby it would not initialize `cinfo->image_width` and `cinfo->image_height` if libjpeg v7 or v8 emulation was enabled. This specifically caused the jpegoptim program to fail if it was linked against a version of libjpeg-turbo that was built with libjpeg v7 or v8 emulation. 6. Eliminated excessive I/O overhead that occurred when reading BMP files in cjpeg. 7. Eliminated errors in the output of cjpeg on Windows that occurred when the application was invoked using I/O redirection (`cjpeg output.jpg`.) 1.1.0 ===== ### Significant changes relative to 1.1 beta1: 1. The algorithm used by the SIMD quantization function cannot produce correct results when the JPEG quality is >= 98 and the fast integer forward DCT is used. Thus, the non-SIMD quantization function is now used for those cases, and libjpeg-turbo should now produce identical output to libjpeg v6b in all cases. 2. Despite the above, the fast integer forward DCT still degrades somewhat for JPEG qualities greater than 95, so the TurboJPEG wrapper will now automatically use the accurate integer forward DCT when generating JPEG images of quality 96 or greater. This reduces compression performance by as much as 15% for these high-quality images but is necessary to ensure that the images are perceptually lossless. It also ensures that the library can avoid the performance pitfall created by [1]. 3. Ported jpgtest.cxx to pure C to avoid the need for a C++ compiler. 4. Fixed visual artifacts in grayscale JPEG compression caused by a typo in the RGB-to-luminance lookup tables. 5. The Windows distribution packages now include the libjpeg run-time programs (cjpeg, etc.) 6. All packages now include jpgtest. 7. The TurboJPEG dynamic library now uses versioned symbols. 8. Added two new TurboJPEG API functions, `tjEncodeYUV()` and `tjDecompressToYUV()`, to replace the somewhat hackish `TJ_YUV` flag. 1.0.90 (1.1 beta1) ================== ### Significant changes relative to 1.0.1: 1. Added emulation of the libjpeg v7 and v8 APIs and ABIs. See [README.md](README.md) for more details. This feature was sponsored by CamTrace SAS. 2. Created a new CMake-based build system for the Visual C++ and MinGW builds. 3. Grayscale bitmaps can now be compressed from/decompressed to using the TurboJPEG API. 4. jpgtest can now be used to test decompression performance with existing JPEG images. 5. If the default install prefix (/opt/libjpeg-turbo) is used, then `make install` now creates /opt/libjpeg-turbo/lib32 and /opt/libjpeg-turbo/lib64 sym links to duplicate the behavior of the binary packages. 6. All symbols in the libjpeg-turbo dynamic library are now versioned, even when the library is built with libjpeg v6b emulation. 7. Added arithmetic encoding and decoding support (can be disabled with configure or CMake options) 8. Added a `TJ_YUV` flag to the TurboJPEG API, which causes both the compressor and decompressor to output planar YUV images. 9. Added an extended version of `tjDecompressHeader()` to the TurboJPEG API, which allows the caller to determine the type of subsampling used in a JPEG image. 10. Added further protections against invalid Huffman codes. 1.0.1 ===== ### Significant changes relative to 1.0.0: 1. The Huffman decoder will now handle erroneous Huffman codes (for instance, from a corrupt JPEG image.) Previously, these would cause libjpeg-turbo to crash under certain circumstances. 2. Fixed typo in SIMD dispatch routines that was causing 4:2:2 upsampling to be used instead of 4:2:0 when decompressing JPEG images using SSE2 code. 3. The configure script will now automatically determine whether the `INCOMPLETE_TYPES_BROKEN` macro should be defined. 1.0.0 ===== ### Significant changes relative to 0.0.93: 1. 2983700: Further FreeBSD build tweaks (no longer necessary to specify `--host` when configuring on a 64-bit system) 2. Created symlinks in the Unix/Linux packages so that the TurboJPEG include file can always be found in /opt/libjpeg-turbo/include, the 32-bit static libraries can always be found in /opt/libjpeg-turbo/lib32, and the 64-bit static libraries can always be found in /opt/libjpeg-turbo/lib64. 3. The Unix/Linux distribution packages now include the libjpeg run-time programs (cjpeg, etc.) and man pages. 4. Created a 32-bit supplementary package for amd64 Debian systems, which contains just the 32-bit libjpeg-turbo libraries. 5. Moved the libraries from */lib32 to */lib in the i386 Debian package. 6. Include distribution package for Cygwin 7. No longer necessary to specify `--without-simd` on non-x86 architectures, and unit tests now work on those architectures. 0.0.93 ====== ### Significant changes since 0.0.91: 1. 2982659: Fixed x86-64 build on FreeBSD systems 2. 2988188: Added support for Windows 64-bit systems 0.0.91 ====== ### Significant changes relative to 0.0.90: 1. Added documentation to .deb packages 2. 2968313: Fixed data corruption issues when decompressing large JPEG images and/or using buffered I/O with the libjpeg-turbo decompressor 0.0.90 ====== Initial release