add an sse4 quantizer so we can use pinsrw/pextrw and keep values in xmm
registers instead of proxying through the stack. and as long as we're
bumping up, use some ssse3 instructions in the EOB detection (see ssse3
fast quantizer)
pick up about a percent on 32bit and about two on 64bit.
Change-Id: If15abba0e8b037a1d231c0edf33501545c9d9363
The dc_diff flag is used to skip loopfiltering. Instead
of setting this flag in the decoder/encoder, we now check
for this condition in the loopfilter.
Change-Id: Ie2b9cdf9e0f4e8b932bbd36e0878c05bffd28931
Code cleanup. The build inter predictor functions are
redundantly checking the mode_info_context for either
INTRA_FRAME or SPLITMV.
Change-Id: I4d58c3a5192a4c2cec5c24ab1caf608bf13aebfb
Golden and ALT reference buffers were refreshed by copying from
the new buffer. Replaced this by index manipulation.
Also moved all the reference frame updates to one function for
easier tracking.
Change-Id: Icd3e534e7e2c8c5567168d222e6a64a96aae24a1
Remove tot_key_frame_bits and prior_key_frame_size[] as they were
tracked but never used. Remove intra_frame_target, as it was only
used to initialize prior_key_frame_size.
Refactor vp8_adjust_key_frame_context() some to remove unnecessary
calculations.
Change-Id: Icbc2c83d2b90e184be03e6f9679e678f3a4bce8f
the win64 abi requires saving and restoring xmm6:xmm15. currently
SAVE_XMM and RESTORE XMM only allow for saving xmm6:xmm7. allow
specifying the highest register used and if the stack is unaligned.
Change-Id: Ica5699622ffe3346d3a486f48eef0206c51cf867
Went through the code and fixed it. Verified on Windows.
Where possible, remove dependencies on xmm[67]
Current code relies on pushing rbp to the stack to get 16 byte
alignment. This broke when rbp wasn't pushed
(vp8/encoder/x86/sad_sse3.asm). Work around this by using unaligned
memory accesses. Revisit this and the offsets in
vp8/encoder/x86/sad_sse3.asm in another change to SAVE_XMM.
Change-Id: I5f940994d3ebfd977c3d68446cef20fd78b07877
Passed SSE from sub-pixel search back to pick_inter_mode
function, which is compared with the encode_breakout to
see if we could skip evaluating the remaining modes.
Change-Id: I4a86442834f0d1b880a19e21ea52d17d505f941d
In vp8_pick_inter_mode(), for NEWMV mode, use the error result got
from motion search as distortion. This helps performance in real-
time mode.
Change-Id: I398c4e46cc5381f7d874e748cf78827ef0e0860c
The value of distortion2 returned by vp8_pick_intra4x4mby_modes
was being overwritten by the value returned by get16x16prederror
before it was tested.
Change-Id: If00e80332b272c5545c3a7e381c8041e8319b41a
Index i is used to detect early breakout from the first loop, but
its value is lost due to reuse in the second for loop. I moved
the position of the second loop and did some format cleanup.
Change-Id: I02780eae1bd89df4b6c000fb8a018b0837aac2e5
This patch cleans up the source buffer storage and copy mechanism to
allow access through a standard push/pop/peek interface. This approach
also avoids an extra copy in the case where the source is not a
multiple of 16, fixing issue #102.
Change-Id: I05808c39f5743625cb4c7af54cc841b9b10fdbd9
in encodframe.c, quant_shift is set to 0 or 1 in vp8cx_invert_quant
only use 8 bits to store this, instead of 16. will allow saving an
xmm register in an updated version of the regular quantize
Change-Id: Ie88c47fe2aff5af0283dab1147fb2791e4b12f90
This patch changes the rc_undershoot_pct and rc_overshoot_pct controls
to set the "aggressiveness" of rate adaptation, by limiting the
amount of difference between the target buffer level and the actual
buffer level which is applied to the target frame rate for this frame.
This patch was initially provided by arosenberg at logitech.com as
an attachment to issue #270. It was modified to separate these controls
from the other unrelated modifications in that patch, as well as to
use the pre-existing variables rather than introducing new ones.
Change-Id: Id542e3f5667dd92d857d5eabf29878f2fd730a62
Previous to commit de4e9e3, there was an early return in the alt-ref
case that was inadvertantly removed when the function was refactored
to return void. This patch restores the prior behavior.
Change-Id: I783ffd594a4690297e2742f99526fd7ad67698b2
The error accumulator stats values cpi->prediction_error and
cpi->intra_error were being populated with rd values not
distortion values.
These are only "currently" used in a limited way for RT compress
key frame detection.
Change-Id: I2702ba1cab6e49ab8dc096ba75b6b34ab3573021
This commit fixed an overflow in ssim calculation, added register
save and restore to make sure assembly code working for x64 platform.
It also changed the sampling points to every 4x4 instead of 8x8 and
adjusted the constants in SSIM calculation to match the scale of
previous VPXSSIM.
Change-Id: Ia4dbb8c69eac55812f4662c88ab4653b6720537b
on the same order as the sse2 fast quantize change: ~2%
except for 32bit. only a slight improvment there.
Change-Id: Iff80e5f1ce7e646eebfdc8871405458ff911986b
MV sad cost error is only used in full-pixel motion search,
which only need full-pixel resolution instead of quarter-pixel
resolution. This change reduced mvsadcost table size, and
removed unneccessary pamameter passing since this table is
constant once it is generated.
Change-Id: I9f931e55f6abc3c99011321f1dfb2f3562e6f6b0
rather than look up rc in the zig zag table, embed it in the macro. this
also allows us to shuffle some values in the macro and keep *d in rsi
gains of about the same order as the obj_int_extract implementation: ~2%
Change-Id: Ib7252dd10eee66e0af8b0e567426122781dc053d
Address calculations moved from encodemb_arm.c file to neon
optimized assembly function to save cycles in function calls.
- vp8_subtract_b_neon_func replaced with vp8_subtract_b_neon
that contains all needed address calculations
- unnecessary file encodemb_arm.c removed
- consistent with ARMv6 optimized version
Change-Id: I6cbc1a2670b56c2077f59995fcf8f70786b4990b
Detect the number of available cores and limit the thread allocation
accordingly. On decoder side limit the number of threads to the max
number of token partition.
Core detetction works on Windows and
Posix platforms, which define _SC_NPROCESSORS_ONLN or _SC_NPROC_ONLN.
Change-Id: I76cbe37c18d3b8035e508b7a1795577674efc078
Adds following ARMv6 optimized functions to encoder:
- vp8_subtract_b_armv6
- vp8_subtract_mby_armv6
- vp8_subtract_mbuv_armv6
Gives 1-5% speed-up depending on input sequence and encoding
parameters. Functions have one stall cycle inside the loop body
on Cortex pipeline.
Change-Id: I19cca5408b9861b96f378e818eefeb3855238639
Half pixel interpolations optimized in variance calculations. Separate
function calls to vp8_filter_block2d_bil_x_pass_armv6 are avoided.On
average, performance improvement is 6-7% for VGA@30fps sequences.
Change-Id: Idb5f118a9d51548e824719d2cfe5be0fa6996628
remove helper function and avoid shadowing all the arguments to the
stack on 64bit systems
when running with --good --cpu-used=0:
~2% on linux x86 and x86_64
~2% on win32 x86 msys and visual studio
more on darwin10 x86_64
significantly more on
x86_64-win64-vs9
Change-Id: Ib7be12edf511fbf2922f191afd5b33b19a0c4ae6
This declaration did not match the prototype_sad() prototype, but was
unused in this translation unit, so it is removed instead. Fixes
issue 290.
Change-Id: I168854f88a85f73ca9aaf61d1e5dc0f43fc3fdb3
Optimized fdct4x4 (8x4) for ARMv6 instruction set.
- No interlocks in Cortex-A8 pipeline
- One interlock cycle in ARM11 pipeline
- About 2.16 times faster than current C-code compiled with -O3
Change-Id: I60484ecd144365da45bb68a960d30196b59952b8
Thread synchronization was not correct when frame width was 1 MB.
Number of allocated encoding threads is limited by the sync_range.
There is no point having more because each thread lags sync_range MBs
behind the thread processing the row above.
http://code.google.com/p/webm/issues/detail?id=302
Change-Id: Icaf67a883beecc5ebf2f11e9be47b6997fdf6f26
A large number of functions were defined with external linkage, even
though they were only used from within one file. This patch changes
their linkage to static and removes the vp8_ prefix from their names,
which should make it more obvious to the reader that the function is
contained within the current translation unit. Functions that were
not referenced were removed.
These symbols were identified by:
$ nm -A libvpx.a | sort -k3 | uniq -c -f2 | grep ' [A-Z] ' \
| sort | grep '^ *1 '
Change-Id: I59609f58ab65312012c047036ae1e0634f795779
The vp8_build_intra_predictors_mby and vp8_build_intra_predictors_mby_s
functions had global function pointers rather than using the RTCD
framework. This can show up as a potential data race with tools such as
helgrind. See https://bugzilla.mozilla.org/show_bug.cgi?id=640935
for an example.
Change-Id: I29c407f828ac2bddfc039f852f138de5de888534
Clean up vp8_init_config() a bit and remove null pointer case,
as this code can't be called any more and is not an adequate
trap anyway, as a null pointer would cause exceptions before
hitting the test.
Change-Id: I937c00167cc039b3aa3f645f29c319d58ae8d3ee
Issue 291 highlighted the fact that CQ mode was not working
as expected in 1 pass mode,
This commit fixes that specific problem but in so doing I also
uncovered an overflow issue in the VBR code for 1 pass and
some data values not being correctly initialized.
For some clips (particularly short clips), the resulting
improvement is dramatic.
Change-Id: Ieefd6c6e4776eb8f1b0550dbfdfb72f86b33c960
In multithreaded mode the loopfilter is running in its own thread (filter level
calculation and frame filtering). Filtering is mostly done in parallel with the
bitstream packing. Before starting the packing the loopfilter level has
to be calculated. Also any needed reference frame copying is done in the
filter thread.
Currently the encoder will create n+1 threads, where n > 1 is the number of
threads specified by application and 1 is the extra filter thread. With n = 1
the encoder runs in single thread mode. There will never be more than n threads
running concurrently.
Change-Id: I4fb29b559a40275d6d3babb8727245c40fba931b
The firstpass motion map consists of an 8-bit flag for
each MB indicating how strongly the firstpass code
believes it should be filtered during the second pass
ARNR filtering.
For long or large format material the motion map can
become extremely large and hamper the operation of
the encoding process.
This change removes the motion map altogether, leaving
the second pass to rely on the magnitude of the motion
compensated error to determine the filter weight to
use for the MB during ARNR filtering.
Tests on the derf set indicate that the effect of this
change is neutral, with some small wins and losses. The
motion map has therefore been removed based on
a cost/benefit evaluation.
Change-Id: I53e07d236f5ce09a6f0c54e7c4ffbb490fb870f6
The previous calculation of macroblock count (w*h)/256
is not correct when the width/height are not multiples of
16. Use the precalculated macroblock count from
cpi->common instead. This manifested itself as a divide
by zero when the number of pixels was less than 256.
num_mbs updated in estimate_max_q, estimate_q,
estimate_kf_group_q, and estimate_cq
Change-Id: I92ff98587864c801b1ee5485cfead964673a9973
1. Process 16 pixels at one time instead of 8.
2. Add check for both xoffset =0 and yoffset=0, which happens
during motion search.
This change gave encoder 1%~3% performance gain.
Change-Id: Idaa39506b48f4f8b2fbbeb45aae8226fa32afb3e
GCC 4.5 and 4.6 both issue a warning about the multi-line format
string introduced in bc9c30a0, which also changed the whitespace
in the associated stt file by line-wrapping the long format string.
Instead, use multiple string constants, which the compiler will
concatenate. This maintains the original formatting, but remains
legible within the standard line length.
Change-Id: I27c9f92d46be82d408105a3a4091f145f677e00e
Disable zbin boost in SPLITMV mode as intended. Was incorrectly looking
at vp8_ref_frame_order instead of vp8_mode_order when comparing against
SPLITMV. This condition should have always been false, as SPLITMV is
not in the range of valid reference frames.
Change-Id: I0408cc7595eff68f00efef6d008e79f5b60d14bf
In some cases where clips have been encoded with
borders (eg. some wide-screen content where there is a
border top and bottom and slide shows containing portrait
format photographs (border left and right)) key frames were
not being correctly detected.
The new code looks to measure cases where a portion of
the image can be coded equally easily using intra or inter
modes and where the resulting error score is also very low.
These "neutral" areas are then discounted in the key frame
detection code.
Change-Id: I00c3e8230772b8213cdc08020e1990cf83b780d8
This code extends what was previously done for GFs, to pick
cases where insertion of a key frame after a fade (or other
transition or complex motion) followed by a still section, will
be beneficial and will reduce the number of forced key frames.
Change-Id: If8bca00457f0d5f83dc3318a587f61c17d90f135
When the modified_error_left accumulator exceeds INT_MAX, an incorrect
cast to int resulted in a negative value, causing the rate control to
allocate no bits to that keyframe group, leading to severe undershoot
and subsequent poor quality.
This error was exposed by the recent change to the rolling target and
actual spend accumulators in commit 305be4e4 which fixed them to
actually calculate the average value rather than be re-initialized
on every frame to the average per-frame bitrate. When this bug was
triggered, the target bitrate could be 0, so the rolling target
becomes small, which causes the undershoot. The code prior to 305be4e4
did not exhibit this behavior because the rolling target was always
set to a reasonable value and was independent of the actual target
bitrate. With this patch, the actual target bitrate is calculated
correctly, and the rate control tracks as expected.
This cast was likely added to silence a compiler warning on a comparison
between a double (modified_error_left) and an int (0). Instead, this
patch removes the cast and changes the comparison to be against 0.0,
which should prevent the warning from reoccuring.
This fixes issue #289. Special thanks to gnafu for his efforts in
reporting and debugging this fix.
Change-Id: Ie5cc1a7b516c578a76c3a50c892a6f04a11621fe
This fixes an overflow problem in the frame error accumulators.
The overflow condition is extreme but did trigger when Frank B.
coded some high motion interlaced HD content.
The observed effect was a catastrophic breakdown of the rate
control leading to massive undershoot and poor bit allocation.
All the error values should really be unsigned but I will look at this
separately.
Change-Id: I9745f5c5ca2783620426b66b568b2088b579151f
Currently, when the video frame width is not multiples of 16, the
source buffer has a stride of non-multiples of 16, which forces
an unaligned load in SAD function and hurts the performance. To
avoid that, this change allocates source buffers to be multiples
of 16.
Change-Id: Ib7506e3eb2cea06657d56be5a899f38dfe3eeb39
In real-time mode, vp8_sad16x16 function is called heavily in
motion search part. Improvement of this function gives 1.2%
encoding performance gain (real-time mode, tulip clip).
Change-Id: I23c401fc40c061f732a9767e8d383737a179bd58
checks added to make sure that cpi->tplist
is freed correctly in vp8_dealloc_compressor_data
and vp8_alloc_compressor_data.
Change-Id: I66149dbbd25c958800ad94f4379d723191d9680d
Eliminated unnecessary calculations. Improved performance
by 10% on keyframes and 1.6% overall for the test clip used.
Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
As mentioned in check-in "Improve motion search in real-time mode",
MV prediction calculation causes speed loss for speed 7 and above.
This change added a flag to turn off this calculation for speed>6
in real-time mode.
Change-Id: I9f4ae5a8bf449222d1784b54e7d315fc8347b2d1
Applied better MV prediction in real-time mode, which improves
the encoding quality.
Used quarter-pixel search instead of iterative sub-pixel search
for speed >=5 to improve encoding performance.
Tests on the test set showed:
1. For speed=-5, quality improvement: 1.7% on AvgPSNR and 2.1%
on SSIM, performance improvement: 3.6% (This counts in the
performance lose caused by MV prediction calculation in "Improve
MV prediction in vp8_pick_inter_mode() for speed>3").
2. For speed=-8, quality improvement: 2.1% on AvgPSNR and 2.5%
on SSIM. but, 6.9% performance decrease because of MV prediction
calculation. This should be improved later.
Change-Id: I349a96c452bd691081d8c8e3e54419e7f477bebd
