The denoiser function was modified to reduce the computational
complexity.
1. The denoiser c function modification:
The original implementation calculated pixel's filter_coefficient
based on the pixel value difference between current raw frame and last
denoised raw frame, and stored them in lookup tables. For each pixel c,
find its coefficient using
filter_coefficient[c] = LUT[abs_diff[c]];
and then apply filtering operation for the pixel.
The denoising filter costed about 12% of encoding time when it was
turned on, and half of the time was spent on finding coefficients in
lookup tables. In order to simplify the process, a short cut was taken.
The pixel adjustments vs. pixel diff value were calculated ahead of time.
adjustment = filtered_value - current_raw
= (filter_coefficient * diff + 128) >> 8
The adjustment vs. diff curve becomes flat very quick when diff increases.
This allowed us to use only several levels to get a close approximation
of the curve. Following the denoiser algorithm, the adjustments are
further modified according to how big the motion magnitude is.
2. The sse2 function was rewritten.
This change made denoiser filter function 3x faster, and improved the
encoder performance by 7% ~ 10% with the denoiser on.
Change-Id: I93a4308963b8e80c7307f96ffa8b8c667425bf50
Merges this experiment in to make it easier to run tests on
filter precision, vectorized implementation etc.
Also removes an experimental filter.
Change-Id: I1e8706bb6d4fc469815123939e9c6e0b5ae945cd
Allows building the library with the gcc -pedantic option, for improved
portabilty. In particular, this commit removes usage of C99/C++ style
single-line comments and dynamic struct initializers. This is a
continuation of the work done in commit 97b766a46, which removed most
of these warnings for decode only builds.
Change-Id: Id453d9c1d9f44cc0381b10c3869fabb0184d5966
Compares the sum of differences between the input block and the averaged
block. If they differ too much the block will not be filtered. Negligible
perfomance hit.
Change-Id: Ib1c31a265efd4d100b3abc4a1ea6675038c8ddde
Add PRIVATE macro for adding private_extern directive for yasm
to hide global symbols. This is only enabled if -DCHROMIUM is used
with YASM.
Also fixed a small problem with rtcd_defs.sh to guard TEMPORAL_DENOISING.
Change-Id: I9027fce3ebddcf20078293e4b86b396f21da7857
Removes all runtime initialization of global data in entropy.c.
Precalculated values are used for initializing all entropy related
tabels.
First patch in a series to make sure code is reentrant.
Change-Id: I9aac91a2a26f96d73c6470d772a343df63bfe633
Deprecate fast quant and strict_quant code.
Small effect on quality as fast was used in first pass but the
effect is basically neutral across the derf set.
The rationale here is to reduce the number of code paths for
now to make experimentation easier. Optimized and fast code
options can be re-introduced later along with other encode
speed options.
Change-Id: Ia30c5daf3dbc52e72c83b277a1d281e3c934cdad
This is the initial patch for supporting 1/8th pel
motion. Currently if we configure with enable-high-precision-mv,
all motion vectors would default to 1/8 pel. Encode and
decode syncs fine with the current code. In the next phase
the code will be refactored so that we can choose the 1/8
pel mode adaptively at a frame/segment/mb level.
Derf results:
http://www.corp.google.com/~debargha/vp8_results/enhinterp_hpmv.html
(about 0.83% better than 8-tap interpoaltion)
Patch 3: Rebased. Also adding 1/16th pel interpolation for U and V
Patch 4: HD results.
http://www.corp.google.com/~debargha/vp8_results/enhinterp_hd_hpmv.html
Seems impressive (unless I am doing something wrong).
Patch 5: Added mmx/sse for bilateral filtering, as well as enforced
use of c-versions of subpel filters with 8-taps and 1/16th pel;
Also redesigned the 8-tap filters to reduce the cut-off in order to
introduce a denoising effect. There is a new configure option
sixteenth-subpel-uv which will use 1/16 th pel interpolation for
uv, if the motion vectors have 1/8 pel accuracy.
With the fixes the results are promising on the derf set. The enhanced
interpolation option with 8-taps alone gives 3% improvement over thei
derf set:
http://www.corp.google.com/~debargha/vp8_results/enhinterpn.html
Results on high precision mv and on the hd set are to follow.
Patch 6: Adding a missing condition for CONFIG_SIXTEENTH_SUBPEL_UV in
vp8/common/x86/x86_systemdependent.c
Patch 7: Cleaning up various debug messages.
Patch 8: Merge conflict
Change-Id: I5b1d844457aefd7414a9e4e0e06c6ed38fd8cc04
Depending on implementation the optimized SAD functions may return early
when the calculated SAD exceeds max_sad.
Change-Id: I05ce5b2d34e6d45fb3ec2a450aa99c4f3343bf3a
Removal of the pickinter.c and .h files and calls to this
code.
Removal of some code relating to real time and one pass
settings though there is more to be done in this regard.
However, vp8_set_speed_features() now
only supports modes 0 and 1 and speeds up to 3
so rd should always be set.
Change-Id: I62c0c1b6154ab499785baef310536080e87bc4d8
In the variance calculations the difference is summed and later squared.
When the sum exceeds sqrt(2^31) the value is treated as a negative when
it is shifted which gives incorrect results.
To fix this we cast the result of the multiplication as unsigned.
The alternative fix is to shift sum down by 4 before multiplying.
However that will reduce precision.
For 16x16 blocks the maximum sum is 65280 and sqrt(2^31) is 46340 (and
change).
PPC change is untested.
Change-Id: I1bad27ea0720067def6d71a6da5f789508cec265
This is the final commit in the series converting to the new RTCD
system. It removes the encoder csystemdependent files and the remaining
global function pointers that didn't conform to the old RTCD system.
Change-Id: I9649706f1bb89f0cbf431ab0e3e7552d37be4d8e
This commit continues the process of converting to the new RTCD
system. It removes the last of the VP8_ENCODER_RTCD struct references.
Change-Id: I2a44f52d7cccf5177e1ca98a028ead570d045395
Storing vp8_bilinear_filters_mmx in an mmx file and using it in an sse2
file is bad
Moving towards allowing --disable-mmx
Change-Id: I20493b35bdedcdcfc0915e6f05fdbe6c81a4a742
Patch set 2: 64 bit build fix
Patch set 3: 64 bit crash fix
[Tero]
Patch set 4: Updated ARMv6 and NEON assembly.
Added also minor NEON optimizations to subtract
functions.
Patch set 5: x86 stride bug fix
Change-Id: I1fcca93e90c89b89ddc204e1c18f208682675c15
Removal of configure #ifdefs so that segment features
always available. Removal of code supporting old
segment feature method.
Still a good deal of tidying up to do.
Change-Id: I397855f086f8c09ab1fae0a5f65d9e06d2e3e39f
Changed 'int eob' to 'char *eob' in BLOCKD so that both encoder and
decoder will use eobs[25] array from MACROBLOCKD structure. In future,
this will enable use of the decoder side IDCT in the encoder.
Change-Id: I6e1c011628cb8864fd4a0b80f0279ce16a5ca978
This quite large check in includes the following:
Merge in some code from Ronald (mbgraph.c) that scans a Gf/arf group.
This is used as a basis for a simple segmentation for the normal frames
in a gf/arf group. This code also uses satd functions from Yaowu.
Adds functionality for coding the latest possible position of an EOB for
blocks in the segment. (Currently 0-15 only, hence just for 4x4 dct).
Where the EOB position is 0 this acts like "skip" and the normal coding
of skip at the per mb level is disabled.
Added functions (seg_common.c) for setting and reading segment feature
elements. These may want to be optimized away at some point but while the
mecahnism is in a state of flux they provide a single location for making
changes and keep things a bit cleaner.
This is still proof of concept code. Currently the tested feature set:-
Quantizer,
Loop Filter level,
Reference frame,
Prediction Mode,
EOB end stop.
TBD:-
Add functions for setting and reading the feature data with range
and validity checking.
Handling of signed and unsigned feature data. At the moment all is assumed
to be signed and a sign bit is coded but many cannot be negative.
Correct handling of EOB feature with intra coded blocks.
Testing/trapping of legal/illegal ref frame and mode combinations.
Transform size switch plus merge and test with 8c8 DCT work
Merge and test with Sumans Segmenation coding optimizations
Change-Id: Iee12e83661c7abbd1e0ce6810915eb4ec35e2d8e
Prepend . to local labels in assembly code. This
allows non unique labels within a file. Also
makes profiling information more informative
by keeping the function name with the loop name.
Change-Id: I7a983cb3a5ba2413d5dafd0a37936b268fb9e37f
Calculations were incorrectly classified as either
SSE3 or SSSE3. Only using SSE2 instructions.
Cleanup function names and make non-RTCD code work
as well.
Change-Id: I48ad0218af0cc51c5078070a08511dee43ecfe09
Calculations were incorrectly classified as either
SSE3 or SSSE3. Only using SSE2 instructions.
Cleanup function names and make non-RTCD code work
as well.
Change-Id: I29f5c2ead342b2086a468029c15e2c1d948b5d97
In sub-pixel motion search, the search range is small(+/- 3 pixels).
Preload whole search area from reference buffer into a 32-byte
aligned buffer. Then in search, load reference data from this buffer
instead. This keeps data in cache, and reduces the crossing cache-
line penalty. For tulip clip, tests on Intel Core2 Quad machine(linux)
showed encoder speed improvement:
3.4% at --rt --cpu-used =-4
2.8% at --rt --cpu-used =-3
2.3% at --rt --cpu-used =-2
2.2% at --rt --cpu-used =-1
Test on Atom notebook showed only 1.1% speed improvement(speed=-4).
Test on Xeon machine also showed less improvement, since unaligned
data access latency is greatly reduced in newer cores.
Next, I will apply similar idea to other 2 sub-pixel search functions
for encoding speed > 4.
Make this change exclusively for x86 platforms.
Change-Id: Ia7bb9f56169eac0f01009fe2b2f2ab5b61d2eb2f
The encoder defined about 4 set of similar functions to calculate sum,
variance or sse or a combination of them. This commit removed one set
of these functions, get8x8var and get16x16var, where calls to the later
function are replaced with var16x16 by using the fact on a 16x16 MB:
variance == sse - sum*sum/256
Change-Id: I803eabd1fb3ab177780a40338cbd596dffaed267
In NEWMV mode, currently, full search is used as the refining search
after n-step search. By replacing it with an iterative diamond search
of radius 1 largely reduced the computation complexity, but still
maintained the same encoding quality since the refining search is
done for every macroblock instead of only a small precentage of
macroblocks while using full search.
Tests on the test set showed a 3.4% encoding speed increase with none
psnr & ssim loss.
Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
Renamed configure option "enable-psnr" to "enable-internal-stats" to
better reflect the purpose of the option and eliminate the confusion
reported in http://code.google.com/p/webm/issues/detail?id=35
Change-Id: If72df6fdb9f1e33dab1329240ba4d8911d2f1f7a
The accumulator array is an integer array, so use paddd instead of paddw
to add values to it. Fixes overflows when using large --arnr-maxframes
(>8) values.
Change-Id: Iad83794caa02400a65f3ab5760f2517e082d66ae
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 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
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 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
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
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
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
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
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
In sub-pixel calculation, xoffset and yoffset mostly take some
specific values. Modified sub-pixel filter functions according to
these possible values to improve performance.
Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21
Remove allocation/deallocation of stats storage.
Remove full search functions in machine specific encoder inits.
Remove last pass validation in validate_config.
Change-Id: I7f29be69273981a4fef6e80ecdb6217c68cbad4e
count can be reduced to short because the max number of filtered frames
is set to 15. the max value for any frame is 32 (modifier = 16,
filter_weight = 2). 15*32 = 480 which requires 9 bits
this function goes from about 7000 us / 1000 iterations for the C code
to < 275 us / 1000 iterations for sse2 for block_size = 16 and from
about 1800 us / 1000 iters to < 100 us / 1000 iters for block_size = 8
Change-Id: I64a32607f58a2d33c39286f468b04ccd457d9e6e
Use the fast quantizer for inter mode selection and the
regular quantizer for the rest of the encode for good quality,
speed 1. Both performance and quality were improved. The
quality gains will make up for the quality loss mentioned in
I9dc089007ca08129fb6c11fe7692777ebb8647b0.
Change-Id: Ia90bc9cf326a7c65d60d31fa32f6465ab6984d21
This code is unused, as the current preproc implementation uses the
same spatial filter that postproc uses.
Change-Id: Ia06d5664917d67283f279e2480016bebed602ea7
Changed the end of block computation to use pmaxw. Removed
additional pushing and popping of registers that was not needed.
Change-Id: I08cb9b424513cd8a2c7ad8cea53b4e2adc66ef98
x86-64 passes arguments in registers. There is no need to push
them to the stack before using them.
This fixes 15acc84f10 where ebx
was not getting preserved on x86.
Change-Id: I1214b5f818a0201f75ab6ad7d5c6f448e09b16c2
(test clip: tulip)
For good quality mode with speed=1, this gave the encoder
a small (2 - 3%) performance boost.
Change-Id: I8a1d4269465944ac0819986c2f0be4b0a2ee0b35
