Lamont Alston [Tue, 13 Oct 2009 06:32:16 +0000 (23:32 -0700)]
Make B-pyramid spec-compliant
The rules of the specification with regard to picture buffering for pyramid coding are widely ignored.
x264's b-pyramid implementation, despite being practically identical to that proposed by the original paper, was technically not compliant.
Now it is.
Two modes are now available:
1) strict b-pyramid, while worse for compression, follows the rule mandated by Blu-ray (no P-frames can reference B-frames)
2) normal b-pyramid, which is like the old mode except fully compliant.
This patch also adds MMCO support (necessary for compliant pyramid in some cases).
MB-tree still doesn't support b-pyramid (but will soon).
Fiona Glaser [Mon, 12 Oct 2009 20:14:19 +0000 (13:14 -0700)]
Reduce the aggressiveness of 2-pass VBV
Now that B-frames are properly covered, we don't have to be as aggressive.
This eliminates some issues with skyrocketing QPs in B-frames in 2-pass VBV.
Loren Merritt [Sat, 10 Oct 2009 04:43:00 +0000 (04:43 +0000)]
change all dct arrays to 1d.
the C standard doesn't allow you to iterate 1-dimensionally over 2d arrays, and nothing other than the dsp functions themselves cares about the 2dness of dct.
this fixes a miscompilation in x264_mb_optimize_chroma_dc.
Fiona Glaser [Mon, 12 Oct 2009 03:17:50 +0000 (20:17 -0700)]
Add row-based VBV for B-frames
While B-frames still aren't explicitly covered by ratecontrol, this should resolve issues of VBV underflows due to larger-than-expected B-frames.
Fiona Glaser [Sun, 11 Oct 2009 00:35:03 +0000 (17:35 -0700)]
Improve VBV, fix bug in 2-pass VBV introduced in MB-tree
Bug caused AQ'd row/frame costs to not be calculated (and thus caused underflows).
Also make VBV more aggressive with more threads in 2-pass mode.
Finally, --ratetol now affects VBV aggressiveness (higher is less aggressive).
Fiona Glaser [Thu, 8 Oct 2009 11:27:11 +0000 (04:27 -0700)]
Avoid scenecuts in flashes and similar situations
"Flashes" are defined as any scene which lasts a very short period before a previous scene returns.
A common example of this is of course a camera flash.
Accordingly, look ahead during scenecut analysis and rule out the possibility of certain frames being scenecuts.
Also handles cases of tons of short scenes in sequence and avoids making those scenecuts as well.
Can only catch flashes of 1 frame in length with b-adapt 1.
With b-adapt 2, can catch flashes of length --bframes.
Speed cost should be negligible.
Holger Lubitz [Tue, 6 Oct 2009 22:17:34 +0000 (15:17 -0700)]
SSE4 version of 4x4 idct
27->24 clocks on Nehalem.
This is really just an excuse to use "movsd" in a real function.
Add some comments to subsum-related macros in x86util.
Fiona Glaser [Mon, 5 Oct 2009 02:15:28 +0000 (19:15 -0700)]
Constrained intra prediction support
Enable with --constrained-intra. Significantly reduces compression, but required for the base layer of SVC encodes and maybe some other use-cases.
Commit sponsored by a media streaming company that wishes to remain anonymous.
Fiona Glaser [Sat, 3 Oct 2009 07:59:02 +0000 (00:59 -0700)]
Reorder reference frames optimally on second pass
About +0.1-0.2% compression at normal bitrates, up to +1% at very low bitrates.
Only works if the first pass uses the same number of refs as the second (i.e. not with fast first pass).
Thus, only worthwhile at insanely slow speeds: as such, enable slow-firstpass by default with preset placebo.
Note that this changes the stats file format!
Clip log2_max_frame_num
It's still much higher than it needs to be, but that will be fixed with the upcoming MMCO patch.
Also make sure we don't write too large a frame_num or poc in slice header.
Add support for single-frame VBV, improve compliance
This allows both constant-framesize and capped-framesize encoding.
Literal constant framesize isn't actually supported yet due to the lack of
filler support.
Example with 30fps video: --vbv-bufsize 200 --vbv-maxrate 6000 will ensure that
no frame is ever larger than 200 kilobits.
One example use-case of this is for zero-delay streaming where bandwidth costs
need to be minimized. If every frame is smaller than 200 kilobits and the
client has a 6 megabit connection, every single frame can be instantly sent
to the client and handled without any decoder-side buffer.
Fix a mistake in VBV calculation--this may have caused the VBV to be slightly
non-compliant in some situations without x264 realizing it.
Add primitive prediction handling for rows with quantizers lower than their
reference. This slightly improves VBV in CBR mode.
Various other minor improvements to VBV, mostly to make single-frame VBV work.
Commit sponsored by a media streaming company that wishes to remain anonymous.
Fix 10l in API change
frame_num was set to 1, not 0, for the first frame. This broke spec compliance.
Didn't actually seem to cause any problems though except for breaking decoding on Quicktime.
Attempt to detect miscompilation due to bug in gcc 4.2
I don't know if this bug still affects latest x264, but it can't hurt to try to detect it.
Accordingly refuse to open the encoder if detected.
Apparently VLC (on Windows) has been distributed for some time with a completely
broken x264 due to the use of a completely broken compiler (gcc 4.2). In
particular, the MV costs seem to be calculated incorrectly on win32 when linking
from an application compiled without -ffast-math to an application with
-ffast-math.
I am not entirely certain why this occurs, but the result is, unsurprisingly,
encoding quality that makes MPEG-2 look good, due to the motion search being
completely broken.
Fix bug with various bizarre commandline combinations and mbtree
Second pass would have mbtree on even though the first pass didn't (and thus encoding would immediately fail).
Add intra prediction modes to output stats
Also eliminate some NANs in stat output with intra-only encoding.
Marginal speedup: disable stat calculation if log level is below X264_LOG_INFO.
Various minor cosmetics.
Major API change: encapsulate NALs within libx264
libx264 now returns NAL units instead of raw data. x264_nal_encode is no longer a public function.
See x264.h for full documentation of changes.
New parameter: b_annexb, on by default. If disabled, startcodes are replaced by sizes as in mp4.
x264's VBV now works on a NAL level, taking into account escape codes.
VBV will also take into account the bit cost of SPS/PPS, but only if b_repeat_headers is set.
Add an overhead tracking system to VBV to better predict the constant overhead of frames (headers, NALU overhead, etc).
Make MV costs global instead of static
Fixes some extremely rare threading race conditions and makes the code cleaner.
Downside: slightly higher memory usage when calling multiple encoders from the same application.
Optimize rounding of luma and chroma DC coefficients
Reduce bitrate mostly-losslessly at low quantizers.
In some rare cases, bitrate reduction may be as high as 10%.
Luma rounding optimization (helps much less than chroma) requires trellis.
Improve x264 help
Now has three help options: --help, --longhelp, and --fullhelp.
--help only shows the most basic options; most users should not need more than these.
Add usage examples.
Fix typo in a comment.
Steven Walters [Wed, 2 Sep 2009 19:33:50 +0000 (12:33 -0700)]
Change priority handling on some OSs
Instead of setting the lookahead thread to max priority, lower all the other threads' priorities instead.
This is particularly useful when the "max priority" is "realtime", as in Windows, which can cause some problems.
Steven Walters [Wed, 2 Sep 2009 01:46:51 +0000 (18:46 -0700)]
Threaded lookahead
Move lookahead into a separate thread, set to higher priority than the other threads, for optimal performance.
Reduces the amount that lookahead bottlenecks encoding, greatly increasing performance with lookahead-intensive settings (e.g. b-adapt 2) on many-core CPUs.
Buffer size can be controlled with --sync-lookahead, which defaults to auto (threads+bframes buffer size).
Note that this buffer is separate from the rc-lookahead value.
Note also that this does not split lookahead itself into multiple threads yet; this may be added in the future.
Additionally, split frames into "fdec" and "fenc" frame types and keep the two separate.
This split greatly reduces memory usage, which helps compensate for the larger lookahead size.
Extremely special thanks to Michael Kazmier and Alex Giladi of Avail Media, the original authors of this patch.
Force a link error in case of incompatible API
This is because the number of bug reports due to miscompiled ffmpeg builds is reaching critical mass.
The name of x264_encoder_open is now #defined based on the current X264_BUILD.
Note that this changes the calling convention required for dlopen, but not for ordinary calls to x264_encoder_open.
Get rid of "CBR" descriptor from qcomp
Though technically accurate in some vague way, I have never actually seen this
option used correctly, rather it has been used by hundreds of people who can't
read the documentation and believe that qcomp=0 is what should be used for CBR
encoding.
Fiona Glaser [Mon, 31 Aug 2009 13:36:41 +0000 (06:36 -0700)]
Multi-slice encoding support
Slicing support is available through three methods (which can be mixed):
--slices sets a number of slices per frame and ensures rectangular slices (required for Blu-ray). Overridden by either of the following options:
--slice-max-mbs sets a maximum number of macroblocks per slice.
--slice-max-size sets a maximum slice size, in bytes (includes NAL overhead).
Implement macroblock re-encoding support to allow highly accurate slice size limitation. Might be useful for other things in the future, too.
Anton Mitrofanov [Fri, 28 Aug 2009 22:07:12 +0000 (15:07 -0700)]
Fix bug in intra analysis in B-frames
i8x8/i4x4 never got analysed when fast_intra was toggled and RD was off; up to a 2-3% quality improvement in non-RD mode.
With this bug dating back to r369, this is probably the second-oldest bug ever fixed in x264.
Fiona Glaser [Thu, 27 Aug 2009 22:21:22 +0000 (15:21 -0700)]
Check minimum chroma QP in addition to luma QP during CQM init
Correctly error out if the implied minimum chroma QP is too low.
Add missing emms to checkasm macroblock_tree_propagate test.
Fiona Glaser [Thu, 27 Aug 2009 21:16:45 +0000 (14:16 -0700)]
Faster mbtree propagate and x264_log2, less memory usage
Avoid an int->float conversion with a small table.
Change lowres_inter_types to a bitfield; cut its size by 75%.
Somewhat lower memory usage with lots of bframes.
Make log2/exp2 tables global to avoid duplication.
David Conrad [Tue, 25 Aug 2009 00:17:41 +0000 (17:17 -0700)]
ARM configure: enable NEON-related options by default
When compiling for ARM, x264 will compile by default for Cortex A8 unless specified otherwise.
To compile for pre-ARMv6, --disable-asm is required.
Fiona Glaser [Mon, 24 Aug 2009 10:28:11 +0000 (03:28 -0700)]
2-pass VBV fixes
Properly run slicetype frame cost with 2pass + MB-tree.
Slash the VBV rate tolerance in 2-pass mode; increasing it made sense for the highly reactive 1-pass VBV algorithm, but not for 2-pass.
2-pass's planned frame sizes are guaranteed to be reasonable, since they are based on a real first pass, while 1-pass's, based on lookahead SATD, cannot always be trusted.
David Conrad [Mon, 24 Aug 2009 08:10:30 +0000 (01:10 -0700)]
GSOC merge part 7: ARM NEON deblock assembly functions (partial)
Originally written for ffmpeg by Mans Rullgard; ported by David.
Luma and chroma inter deblocking; no intra yet.
David Conrad [Sun, 23 Aug 2009 06:40:33 +0000 (23:40 -0700)]
GSOC merge part 2: ARM stack alignment
Neither GCC nor ARMCC support 16 byte stack alignment despite the fact that NEON loads require it.
These macros only work for arrays, but fortunately that covers almost all instances of stack alignment in x264.
David Conrad [Fri, 21 Aug 2009 03:44:09 +0000 (20:44 -0700)]
Fix unaligned accesses in bitstream writer
Fixes x264 on CPUs with no unaligned access support (e.g. SPARC).
Improves performance marginally on CPUs with penalties for unaligned stores (e.g. some x86).
David Conrad [Thu, 20 Aug 2009 00:03:02 +0000 (17:03 -0700)]
GSOC merge part 1: Framework for ARM assembly optimizations
x264 will detect which ARM core it's building for and only build NEON asm if the target is ARMv6 or above, then enable NEON at runtime.
David Conrad [Wed, 19 Aug 2009 23:18:36 +0000 (16:18 -0700)]
Fix a bug in checkasm and two OSX fixes
MC chroma checkasm test could crash in some situations
Remove -lmx, as it's not needed and the iPhone doesn't have it.
Remove unused sqrtf emulation; it breaks if math.h is included.
Fiona Glaser [Wed, 19 Aug 2009 08:49:47 +0000 (01:49 -0700)]
Improve QPRD
Always check the last macroblock's QP, even if the normal search doesn't reach it.
Raise the failure threshold when moving towards the last macroblock's QP.
0.2-1% improved compression.
Fiona Glaser [Wed, 19 Aug 2009 01:37:26 +0000 (18:37 -0700)]
Add support for frame-accurate parameter changes
Parameter structs can now be passed with individual frames.
The previous method would only change the parameter of what was currently being encoded, which due to delay might be very far from an intended exact frame.
Also add support for changing aspect ratio. Only works in a stream with repeating headers and requires the caller to force an IDR to ensure instant effect.
Fiona Glaser [Tue, 18 Aug 2009 22:46:26 +0000 (15:46 -0700)]
Fix x264_encoder_reconfig with multithreading
New behavior: reconfigging the encoder will result in changes being applied
to each of the encoding threads as they finish encoding the current frame.
Fiona Glaser [Sun, 16 Aug 2009 02:02:31 +0000 (19:02 -0700)]
Lookahead VBV
Use the large-scale lookahead capability introduced in MB-tree for ratecontrol purposes.
(Does not require MB-tree, however.)
Greatly improved quality and compliance in 1-pass VBV mode, especially in CBR; +2db OPSNR or more in some cases.
Fix some other bugs in VBV, which should improve non-lookahead mode as well.
Change the tolerance algorithm in row VBV to allow for more significant mispredictions when buffer is nearly full.
Note that due to the fixing of an extremely long-standing bug (>1 year), bitrates may change by nontrivial amounts in CRF without MB-tree.
Fiona Glaser [Fri, 14 Aug 2009 00:13:33 +0000 (17:13 -0700)]
Fix a potential failure in VBV
If VBV does underflow, ratecontrol could be permanently broken for the rest of the clip.
Revert part of the previous VBV changes to fix this.
Loren Merritt [Thu, 13 Aug 2009 05:02:59 +0000 (05:02 +0000)]
fix lowres padding, which failed to extrapolate the right side for some resolutions.
fix a buffer overread in x264_mbtree_propagate_cost_sse2. no effect on actual behavior, only theoretical correctness.
fix x264_slicetype_frame_cost_recalculate on I-frames, which previously used all 0 mb costs.
shut up a valgrind warning in predict_8x8_filter_mmx.
Loren Merritt [Sat, 8 Aug 2009 14:53:27 +0000 (14:53 +0000)]
MB-tree fixes:
AQ was applied inconsistently, with some AQed costs compared to other non-AQed costs. Strangely enough, fixing this increases SSIM on some sources but decreases it on others. More investigation needed.
Account for weighted bipred.
Reduce memory, increase precision, simplify, and early terminate.
Fiona Glaser [Wed, 5 Aug 2009 00:46:33 +0000 (17:46 -0700)]
Macroblock-tree ratecontrol
On by default; can be turned off with --no-mbtree.
Uses a large lookahead to track temporal propagation of data and weight quality accordingly.
Requires a very large separate statsfile (2 bytes per macroblock) in multi-pass mode.
Doesn't work with b-pyramid yet.
Note that MB-tree inherently measures quality different from the standard qcomp method, so bitrates produced by CRF may change somewhat.
This makes the "medium" preset a bit slower. Accordingly, make "fast" slower as well, and introduce a new preset "faster" between "fast" and "veryfast".
All presets "fast" and above will have MB-tree on.
Add a new option, --rc-lookahead, to control the distance MB tree looks ahead to perform propagation analysis.
Default is 40; larger values will be slower and require more memory but give more accurate results.
This value will be used in the future to control ratecontrol lookahead (VBV).
Add a new option, --no-psy, to disable all psy optimizations that don't improve PSNR or SSIM.
This disables psy-RD/trellis, but also other more subtle internal psy optimizations that can't be controlled directly via external parameters.
Quality improvement from MB-tree is about 2-70% depending on content.
Strength of MB-tree adjustments can be tweaked using qcompress; higher values mean lower MB-tree strength.
Note that MB-tree may perform slightly suboptimally on fades; this will be fixed by weighted prediction, which is coming soon.