DRC [Wed, 14 Jan 2015 19:39:01 +0000 (19:39 +0000)]
Add the ability to benchmark YCCK JPEG compression/decompression. This is particularly useful since that is the only way to test the performance of the "plain" upsampling routines, which are accelerated on some platforms.
DRC [Wed, 14 Jan 2015 10:45:31 +0000 (10:45 +0000)]
Fix bugs in the AltiVec fancy upsampling routines uncovered during additional testing with small image sizes. Since the input width is half the output width, the upsampler should only write a second 16-byte chuck if there are more than 8 input columns left. Additionally, if the width is < 16, then we need to insert a dummy sample (the SSE2 code does this as well, but I neglected to port that portion of the code for some reason.)
DRC [Wed, 14 Jan 2015 08:42:29 +0000 (08:42 +0000)]
In the process of developing the AltiVec extensions, it was discovered that the normal regression tests aren't sufficient to test the behavior of the library with very small image sizes and when compressing from/decompressing to a subregion of a larger image buffer. Thus, an additional regression test was added that takes advantage of the tiled compression/decompression feature in tjbench.
DRC [Wed, 14 Jan 2015 08:31:54 +0000 (08:31 +0000)]
Fix a bug in the AltiVec downsampling routines uncovered during additional testing with small image sizes. Since the output width is half the input width, the downsampler should only read a second 16-byte chunk if there are more than 8 output columns left.
DRC [Sun, 11 Jan 2015 06:34:47 +0000 (06:34 +0000)]
Use intrinsics for loading aligned data in the IDCT functions. This has no effect on performance, but it makes it more obvious what that code is doing.
DRC [Sat, 10 Jan 2015 11:32:36 +0000 (11:32 +0000)]
Overhaul the AltiVec vector loading code in the compression-side colorspace conversion routines. The existing code was sometimes overreading the source buffer (at least according to valgrind), and it was necessary to increase the complexity of the code in order to prevent this without significantly compromising performance.
DRC [Mon, 22 Dec 2014 16:04:17 +0000 (16:04 +0000)]
Use intrinsics for loading/storing data in the DCT/IDCT functions. This has no effect on the performance of the aligned loads/stores, but it makes it more obvious what that code is doing. Using intrinsics for the unaligned stores in the inverse DCT functions increases overall decompression performance by 1-2%.
DRC [Sat, 20 Dec 2014 01:16:26 +0000 (01:16 +0000)]
Use macros to allocate constants statically, rather than reading them from a table using vec_splat*(). This improves code readability and probably improves performance a bit as well.
DRC [Fri, 19 Dec 2014 15:36:39 +0000 (15:36 +0000)]
Modify the ARM64 assembly file so that it uses only syntax that the clang assembler in XCode 5.x can understand. These changes should all be cosmetic in nature-- they do not change the meaning or readability of the code nor the ability to build it for Linux. Actually, the code is now more in compliance with the ARM64 programming manual. In addition to these changes, there were a couple of instructions that clang simply doesn't support, so gas-preprocessor.pl was modified so that it now converts those into equivalent instructions that clang can handle.
DRC [Fri, 19 Dec 2014 10:44:09 +0000 (10:44 +0000)]
Add iOS architectures to the shared libraries generated by the Mac/iOS packaging system. I have no idea how useful this is for "standard" iOS application development, but it is useful in a jailbreak environment, and iOS 8 supposedly allows shared libs in "official" apps as well.
DRC [Thu, 18 Dec 2014 09:49:39 +0000 (09:49 +0000)]
Further cleanup of the AltiVec forward DCT code:
-- Use macros to represent the fast FDCT constants, to facilitate comparing the AltiVec implementation of the algorithm with the SSE2 implementation.
-- Rename slow FDCT constants for consistency.
-- Use vec_sra() in all cases in the slow FDCT code. The SSE2 implementation uses psraw, which is an arithmetic shift, so we need to do likewise with AltiVec. Using vec_sr() hasn't caused any problems yet, but it is conceivable that it might cause different behavior in certain corner cases.
DRC [Wed, 17 Dec 2014 08:04:39 +0000 (08:04 +0000)]
AltiVec SIMD implementation of slow integer forward DCT; Clean up fast integer forward DCT code so that it is easier to see how it derives from the SSE2 code and to make it play more nicely with the slow FDCT code.
DRC [Tue, 25 Nov 2014 09:48:15 +0000 (09:48 +0000)]
Restore the JPP() and JMETHOD() macros. Even though libjpeg-turbo doesn't use them anymore, other software apparently does:
https://bugzilla.redhat.com/show_bug.cgi?id=1164815
https://bugs.kde.org/show_bug.cgi?id=340944
https://bugzilla.mozilla.org/show_bug.cgi?id=1093615
DRC [Sat, 22 Nov 2014 22:09:30 +0000 (22:09 +0000)]
Fix Huffman local buffer overrun discovered by Debian developers when attempting to transform a junk image using ImageMagick:
https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=768369
DRC [Wed, 19 Nov 2014 00:55:28 +0000 (00:55 +0000)]
Sometimes the sampling factors in grayscale images can be > 1 (for instance, if compressing using 'cjpeg -sample 2x2 -grayscale'.) Technically, sampling factors have no meaning with grayscale JPEGs, and the libjpeg decompressor ignores them in that case. Thus, the TurboJPEG decompressor should ignore them as well.
The AltiVec code actually works on 32-bit PowerPC platforms as well, so change the "powerpc64" token to "powerpc". Also clean up the shift code, which wasn't building properly on OS X.
When building libjpeg-turbo on Un*x systems, INT32 is usually typedef'ed to long, not int, so we need to specify an int pointer when doing a 4-byte write to the RGB565 output buffer. On little endian systems, this doesn't matter, but when you write a 32-bit int to a 64-bit long pointer address on a big endian system, you are writing to the upper 4 bytes, not the lower 4 bytes. NOTE: this will probably break on big endian systems that use 16-bit ints (are there any of those still around?)
This patch also removes an unneeded macro from jdmerge.c.
When building libjpeg-turbo on Un*x systems, INT32 is usually typedef'ed to long, not int, so we need to specify an int pointer when doing a 4-byte write to the RGB565 output buffer. On little endian systems, this doesn't matter, but when you write a 32-bit int to a 64-bit long pointer address on a big endian system, you are writing to the upper 4 bytes, not the lower 4 bytes. NOTE: this will probably break on big endian systems that use 16-bit ints (are there any of those still around?)
This patch also removes an unneeded macro from jdmerge.c.
DRC [Sat, 30 Aug 2014 20:37:50 +0000 (20:37 +0000)]
Fix issues with RGB565 color conversion on big endian machines. The RGB565 routines are now abstracted in a separate file, with separate little-endian and big-endian versions defined at compile time through the use of macros (this is similar to how the colorspace extension routines work.) This allows big-endian machines to take advantage of the same performance optimizations as little-endian machines, and it retains the performance on little-endian machines, since the conditional branch for endianness is at a very coarse-grained level.
DRC [Fri, 29 Aug 2014 01:49:59 +0000 (01:49 +0000)]
Fix several mathematical issues discovered in the ARM64 NEON code while running the extended regression tests introduced in r1267. Specific comments can be found in the original patches:
https://sourceforge.net/p/libjpeg-turbo/patches/64/
DRC [Mon, 25 Aug 2014 15:26:09 +0000 (15:26 +0000)]
Reformat code per Siarhei's original patch (to clearly indicate that the offset instructions are completely independent) and add Siarhei as an individual author (he no longer works for Nokia.)
From aee36252be20054afce371a92406fc66ba6627b5 Mon Sep 17 00:00:00 2001
From: Siarhei Siamashka <siarhei.siamashka@gmail.com>
Date: Wed, 13 Aug 2014 03:50:22 +0300
Subject: [PATCH] ARM: Faster NEON yuv->rgb conversion for Krait and Cortex-A15
The older code was developed and tested only on ARM Cortex-A8 and ARM Cortex-A9.
Tuning it for newer ARM processors can introduce some speed-up (up to 20%).
The performance of the inner loop (conversion of 8 pixels) improves from
~27 cycles down to ~22 cycles on Qualcomm Krait 300, and from ~20 cycles
down to ~18 cycles on ARM Cortex-A15.
The performance remains exactly the same on ARM Cortex-A7 (~58 cycles),
ARM Cortex-A8 (~25 cycles) and ARM Cortex-A9 (~30 cycles) processors.
Also use larger indentation in the source code for separating two independent
instruction streams.
The performance of the inner loop (conversion of 8 pixels):
* ARM Cortex-A7: ~55 cycles
* ARM Cortex-A8: ~28 cycles
* ARM Cortex-A9: ~32 cycles
* ARM Cortex-A15: ~20 cycles
* Qualcomm Krait: ~24 cycles
Based on the Linaro rgb565 patch from
https://sourceforge.net/p/libjpeg-turbo/patches/24/
but implements better instructions scheduling.
DRC [Fri, 22 Aug 2014 19:59:51 +0000 (19:59 +0000)]
Ensure that tjFree() is used for any JPEG buffers that might have been dynamically allocated by the compress/transform functions. To keep things simple, we use tjAlloc() for the statically-allocated buffer as well, so that tjFree() can always be used to free the buffer, regardless of whether it was allocated by tjbench or by the TurboJPEG library. This fixes crashes that occurred on Windows when running tjunittest or tjbench with the -alloc flag.
DRC [Fri, 22 Aug 2014 19:27:28 +0000 (19:27 +0000)]
Ensure that tjFree() is used for any JPEG buffers that might have been dynamically allocated by the compress/transform functions. To keep things simple, we use tjAlloc() for the statically-allocated buffer as well, so that tjFree() can always be used to free the buffer, regardless of whether it was allocated by tjbench or by the TurboJPEG library. This fixes crashes that occurred on Windows when running tjunittest or tjbench with the -alloc flag.
DRC [Fri, 22 Aug 2014 18:30:44 +0000 (18:30 +0000)]
Revert r1335 and r1336. It was a valiant effort, but on Windows, xmm8-xmm15 are non-volatile, and the overhead of pushing them onto the stack at the beginning of each function and popping them at the end was causing worse performance (in the neighborhood of 3-5%) than just using the work areas and limiting the register usage to xmm0-xmm7. Best to leave the SSE2 code alone. We can optimize the register usage for AVX2, once that port takes place.
DRC [Fri, 22 Aug 2014 13:43:33 +0000 (13:43 +0000)]
Add a set of undocumented environment variables and Java system properties that allow compression features of libjpeg that are not normally exposed in the TurboJPEG API to be enabled. These features are not normally exposed because, for the most part, they aren't "turbo" features, but it is still useful to be able to benchmark them without modifying the code.
DRC [Fri, 22 Aug 2014 11:31:46 +0000 (11:31 +0000)]
.func/.endfunc are only necessary when generating STABS debug info, which basically went out of style with parachute pants and Rick Astley. At any rate, none of the platforms for which we're building the ARM code use it (DWARF is the common format these days), and the .func/.endfunc directives cause the clang integrated assembler to fail (http://llvm.org/bugs/show_bug.cgi?id=20424).
DRC [Fri, 22 Aug 2014 03:04:06 +0000 (03:04 +0000)]
Extend tjbenchtest so that it tests the dynamic JPEG buffer allocation feature in TurboJPEG. Disable the tiling feature in TJBench whenever dynamic buffer allocation is enabled (because the tiling feature requires a separate buffer for each tile, using it successfully with dynamic buffer allocation would require a separate TurboJPEG compressor instance for each tile, and it's not worth going to that trouble right now.)
DRC [Fri, 22 Aug 2014 03:00:37 +0000 (03:00 +0000)]
Run the TurboJPEG conformance tests out of a directory in /tmp (for improved performance, if the source directory is on a remote file share.) Fix an issue in TJBench.java that prevented it from working properly if the source image resided in a directory with a dot in the name.
DRC [Fri, 22 Aug 2014 02:57:34 +0000 (02:57 +0000)]
Extend tjbenchtest so that it tests the dynamic JPEG buffer allocation feature in TurboJPEG. Disable the tiling feature in TJBench whenever dynamic buffer allocation is enabled (because the tiling feature requires a separate buffer for each tile, using it successfully with dynamic buffer allocation would require a separate TurboJPEG compressor instance for each tile, and it's not worth going to that trouble right now.)
projects / libjpeg-turbo / log