This change implemented same idea in change "Preload reference area
to an intermediate buffer in sub-pixel motion search." The changes
were made to vp8_find_best_sub_pixel_step() and vp8_find_best_half
_pixel_step() functions which are called when speed >= 5. Test
result (using tulip clip):
1. On Core2 Quad machine(Linux)
rt mode, speed (-5 ~ -8), encoding speed gain: 2% ~ 3%
rt mode, speed (-9 ~ -11), encoding speed gain: 1% ~ 2%
rt mode, speed (-12 ~ -14), no noticeable encoding speed gain
2. On Xeon machine(Linux)
Test on speed (-5 ~ -14) didn't show noticeable speed change.
Change-Id: I21bec2d6e7fbe541fcc0f4c0366bbdf3e2076aa2
There were some situations that the start motion vectors were
out of range. This fix adjusted range checks to make sure they
are checked and clamped.
Change-Id: Ife83b7fed0882bba6d1fa559b6e63c054fd5065d
sharpness was not recalculated in vp8cx_pick_filter_level_fast
remove last_filter_type. all values are calculated, don't need to update
the lfi data when it changes.
always use cm->sharpness_level. the extra indirection was annoying.
don't track last frame_type or sharpness_level manually. frame type
only matters for motion search and sharpness_level is taken care of in
frame_init
move function declarations to their proper header
Change-Id: I7ef037bd4bf8cf5e37d2d36bd03b5e22a2ad91db
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
This is done by expanding luma row to 32-byte alignment, since
there is currently a bunch of code that assumes that
uv_stride == y_stride/2 (see, for example, vp8/common/postproc.c,
common/reconinter.c, common/arm/neon/recon16x16mb_neon.asm,
encoder/temporal_filter.c, and possibly others; I haven't done a
full audit).
It also uses replaces the hardcoded border of 16 in a number of
encoder buffers with VP8BORDERINPIXELS (currently 32), as the
chroma rows start at an offset of border/2.
Together, these two changes have the nice advantage that simply
dumping the frame memory as a contiguous blob produces a valid,
if padded, image.
Change-Id: Iaf5ea722ae5c82d5daa50f6e2dade9de753f1003
allowing the compiler to inline this function. For real-time
encodes, this gave a boost of 1% to 2.5%, depending on the
speed setting.
Change-Id: I3929d176cca086b4261267b848419d5bcff21c02
This patch attempts to improve the handling of CBR streams with
respect to the short term buffering requirements. The "buffer level"
is changed to be an average over the rc buffer, rather than a long
running average. Overshoot is also tracked over the same interval
and the golden frame targets suppressed accordingly to correct for
overly aggressive boosting.
Testing shows that this is fairly consistently positive in one
metric or another -- some clips that show significant decreases
in quality have better buffering characteristics, others show
improvenents in both.
Change-Id: I924c89aa9bdb210271f2e03311e63de3f1f8f920
Optimized C-code of the following functions:
- vp8_tokenize_mb
- tokenize1st_order_b
- tokenize2nd_order_b
Gives ~1-5% speed-up for RT encoding on Cortex-A8/A9
depending on encoding parameters.
Change-Id: I6be86104a589a06dcbc9ed3318e8bf264ef4176c
Do mvp clamping in full-pixel precision instead of 1/8-pixel
precision to avoid error caused by right shifting operation.
Also, further fixed the motion vector limit calculation in change:
b748045470
Change-Id: Ied88a4f7ddfb0476eb9f7afc6ceeddbf209fffd7
Separate simple filter with reduced no. of parameters.
MB filter level picking based on precalculated table. Level table updated for
each frame. Inside and edge limits precalculated and updated just when
sharpness changes. HEV threshhold is constant.
ARM targets use scalars and others vectors.
Change works only with --target=generic-gnu
All other targets have to be updated!
Change-Id: I6b73aca6b525075b20129a371699b2561bd4d51c
Allow the encoder to inform the application that the encoded frame will not
be used as a reference.
Change-Id: I90e41962325ef73d44da03327deb340d6f7f4860
Motion vector limits are calculated using right shifts, which
could give wrong results for negative numbers. James Berry's
test on one clip showed encoder produced some artifacts. This
change fixed that.
Change-Id: I035fc02280b10455b7f6eb388f7c2e33b796b018
In this commit I have added an experimental function
that tests prediction quality either side of a central position
to calculate a suggested boost number for an ARF frame.
The function is passed an offset from the current position and
a number of frames to search forwards and backwards.
It returns a forward, backward and compound boost number.
The new code can be deactivated using #define NEW_BOOST 0
In its current default state the code searches forwards and backwards
from the proposed position of the next alt ref.
The the old code used a boost number calculated by scanning forward
from the previous GF up to the proposed alt ref frame position.
I have also added some code to try and prevent placement of a gf/arf
where there is a brief flash.
Change-Id: I98af789a5181148659f10dd5dd2ff2d4250cd51c
There were many instances in the code of vp8_coef_tokens and
vp8_coef_tokens-1, which was a preprocessor macro despite the naming
convention. Replace these with MAX_ENTROPY_TOKENS and ENTROPY_NODES,
respectively.
Change-Id: I72c4f6c7634c94e1fa066cd511471e5592c748da
With this commit frames can be received partition-by-partition
from the encoder and passed partition-by-partition to the
decoder.
At the encoder-side this makes it easier to split encoded
frames at partition boundaries, useful when packetizing
frames. When VPX_CODEC_USE_OUTPUT_PARTITION is enabled,
several VPX_CODEC_CX_FRAME_PKT packets will be returned
from vpx_codec_get_cx_data(), containing one partition
each. The partition_id (starting at 0) specifies the decoding
order of the partitions. All partitions but the last has
the VPX_FRAME_IS_FRAGMENT flag set.
At the decoder this opens up the possibility of decoding partition
N even though partition N-1 was lost (given that independent
partitioning has been enabled in the encoder) if more info
about the missing parts of the stream is available through
external signaling.
Each partition is passed to the decoder through the
vpx_codec_decode() function, with the data pointer pointing
to the start of the partition, and with data_sz equal to the
size of the partition. Missing partitions can be signaled to
the decoder by setting data != NULL and data_sz = 0. When
all partitions have been given to the decoder "end of data"
should be signaled by calling vpx_codec_decode() with
data = NULL and data_sz = 0.
The first partition is the first partition according to the
VP8 bitstream + the uncompressed data chunk + DCT address
offsets if multiple residual partitions are used.
Change-Id: I5bc0682b9e4112e0db77904755c694c3c7ac6e74
Adding support in the encoder for generating
independent residual partitions by forcing
equal probabilities over the prev coef entropy
contexts.
Change-Id: I402f5c353255f3ca20eae2620af739f6a498cd21
I got this idea from Pascal (Thanks). Before encoding a macroblock,
copy it to a 16x16 buffer, and then read source data from there
instead. This will help keep the source data in cache, and help
with the performance.
Change-Id: Id05f4cb601299150511d59dcba0ae62c49b5b757
This reverts commit 212f618373.
Further testing shows that the overshoot accumulation/damping is too
aggressive on some clips. Allowing the accumulated overshoot to
decay and limiting to damping to golden frames shows some promise.
But some clips show significant overshoot in the buffer window, so
I think this still needs work.
Change-Id: Ic02a9ca34f55229f9cc04786f4fab54cdc1a3ef5
RDMULT/RDDIV defines a bit worth of distortion in term of sum squared
difference. This has also been used as errorperbit in subpixel motion
search, where the distortions computed as variance of the difference.
The variance of differences is different from sum squared differences
by amount of DC squared. Typically, for inter predicted MBs, this
difference averages around 10% between the two distortion, so this patch
introduces a 110% constant in deriving errorperbit from RDMULT/RDDIV.
Test on CIF set shows small but positive gain on overall PSNR (.03%)
and SSIM (.07%), overall impact on average PSNR is 0.
Change-Id: I95425f922d037b4d96083064a10c7cdd4948ee62
The starting points are always within the limits, and bounds
checking on these points is not needed. For speed < 5, the
encoded result changes a little because different treatment
is taken while starting point equals the bounds.
Change-Id: I09a402d310f51e305a3519f1601b1d17b05c6152
Modify the second-pass code to provide a full golden-frame (GF) bit
allocation boost if the past GF group (GFG) had no alt-ref frame (ARF),
even if the current GFG does contain and ARF.
This mostly has no effect on clips, since switching ARFs on/off between
GFGs is not very common. Has a positive effect on e.g. cheer (+0.45 SSIM
at 600kbps) and football (+0.25 SSIM at 600kbps), particularly at high
bitrates. Has a negative effect (-0.04 SSIM at 300kbps) at pamphlet,
which appears only marginally related to this patch, and crew (-0.1 SSIM
at 700kbps).
Change-Id: I2e32899638b59f857e26efeac18a82e0c0b77089
The BPRED mode selection uses SSE as a distortion metric, but the early
breakout threshold being used was a variance value.
Change-Id: I42d4602fb9b548bf681a36445701fada5e73aff1
firstpass.c contains some rate adjustment code that assures that the
last few frames in a sequence abide by rate limits. If the second-to-
last group of frames contains an alt-ref frame (ARF), the last golden
frame (GF) is zero bytes, and we will thus spend a ridiculously high
number of bits on regular P-frames trying to hit the target rate. This
does slightly enhance the quality of these last few frames, but has
no perceptual value (other than hitting the target rate).
Disabling this code means we consistently (slightly) undershoot the
target rate and consequently do worse on the last few frames of a
clip, which is particularly noticeable for small clips. The quality-
per-bitrate is generally better, ~0.2% better overall on derf-set,
especially on clips such as garden, tennis, foreman at low bitrates.
Has a negative effect on hallmonitor at high bitrates.
Change-Id: I1d63452fef5fee4a0ad2fb2e9af4c9f2e0d86d23
Moved encode_intra function from firstpass.c to encodeintra.c to
prevent linking problem in real-time only build. Also changed name
of the function to vp8_encode_intra because it is not a static.
Change-Id: Ibf3c6c1de3152567347e5fbef47d1d39564620a5
Small decode performance gain (~1%) on keyframes. No
noticeable gains on encode. Also changed pick_intra4x4mby_modes()
to read the above and left block modes for keyframes only.
Change-Id: I1f4885252f5b3e9caf04d4e01e643960f910aba5
Since BPRED will be tested at most once, and SPLITMV is not enabled,
there's nothing to clobber the subblock modes, so there's no need to
save and restore them.
Change-Id: I7c3615b69190c10bd068a44df5488d6e8b85a364
In activity masking, RDO constant RDMULT is adjusted on a per MB basis
adaptive to activity with the MB. errorperbit, which is defined as
RDMULT/RDDIV, is a constant used in motion estimation. Previously, in
activity masking, errorperbit is not changed even when RDMULT is changed.
This commit changed to adjust errorperbit according to the change in
RDMULT.
Test in cif set showed a very small but consistent gain by all quality
metrics (average, overall psnr and ssim) when activity masking is on.
Change-Id: I07ded3e852919ab76757691939fe435328273823
This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7,
which made the move for the non-RD path.
Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
This commit moves the intra block mode selection from encodeframe.c
to pickinter.c (in the non-RD case). This allowed pick_intra_mbuv_mode
and pick_intra4x4mby_modes to be made static, and is a step towards
refactoring intra mode selection in the main pickinter loop. Gave a
small perf increase (~0.5%).
Change-Id: I0b67dae1f8a74902378da7bdf565e39ab832dda7
Some further re-structuring of activity masking code.
Still has various experimental switches.
Supports a metric based on intra encode.
Experimental comparison against a fixed activity target rather
than a frame average, for altering rd and zbin.
Overall the SSIM performance is similar to TT's original
code but there is a much smaller PSNR hit of circa
0.5% instead of 3.2%
Change-Id: I0fd53b2dfb60620b3f74d7415e0b81c1ac58c39a
While investigating the effect of DC values on SAD and SSE in motion
estimation, a side finding indicates the two table of constants need
be adjusted. The adjustment was done by multiplying old constants by
90% with rounding. Also absorb the 1/2 scaling constant into the two
tables. Refer to change Ifa285c3e for background of the 1/2 factor.
Cif set test showed a very small gain on all metric.
Change-Id: I04333527a823371175dd46cb04a817e5b9a8b752
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 real-time mode motion search, there is no need to calculate
variance. This change improved encoding speed by 1% ~ 2%(speed=-5).
Change-Id: I65b874901eb599ac38fe8cf9cad898c14138d431
This patch attempts to reduce the peak bitrate hit by the encoder
when using small buffer windows.
Tested on the CIF set over 200-500kbps using these settings:
--buf-sz=500 --buf-initial-sz=250 --buf-optimal-sz=250 \
--undershoot-pct=100
Two pass encodes were tested at best quality. One pass encodes were
tested only at realtime speed 4:
--rt --cpu-used=-4
The peak datarate (over the specified 500ms window) was measured
for each encode, and averaged together to get metric for
"average peak," computed as SUM(peak)/SUM(target). This patch
reduces the average peak datarate as follows:
One pass:
baseline: 1.29715
this patch: 1.23664
Two pass:
baseline: 1.32702
this patch: 1.37824
This change had a positive effect on our quality metrics as well:
One pass CBR:
Min / Mean / Max (pct)
Average PSNR -0.42 / 2.86 / 27.32
Overall PSNR -0.90 / 2.00 / 17.27
SSIM -0.05 / 3.95 / 37.46
Two pass CBR:
Min / Mean / Max (pct)
Average PSNR -4.47 / 4.35 / 35.99
Overall PSNR -3.40 / 4.18 / 36.46
SSIM -4.56 / 6.98 / 53.67
One pass VBR:
Min / Mean / Max (pct)
Average PSNR -5.21 / 0.01 / 3.30
Overall PSNR -8.10 / -0.38 / 1.21
SSIM -7.38 / -0.11 / 3.17
(note: most values here were close to the mean, there were a few
outliers on files that were very sensitive to golden frame size)
Two pass VBR:
Min / Mean / Max (pct)
Average PSNR 0.00 / 0.00 / 0.00
Overall PSNR 0.00 / 0.00 / 0.00
SSIM 0.00 / 0.00 / 0.00
Neither one pass or two pass CBR mode adheres particularly strictly
to the short term buffer constraints, and two pass is less
consistent, even in the baseline commit. This should be addressed
in a later commit. This likely will hurt the quality numbers, as it
will have to reduce the burstiness of golden frames.
Aside: My work on this commit makes it clear that we need to make
rate control modes "pluggable", where you can easily write a new
one or work on one in isolation.
Change-Id: I1ea9a48f2beedd59891f1288aabf7064956b4716
Currently, hex search couldn't guarantee the motion vector(MV)
found is within the limit of maximum MV. Therefore, very large
motion vectors resulted from big motion in the video could cause
encoding artifacts. This change adjusted hex search bounds
checking to make sure the resulted motion vector won't go out
of the range. James Berry, thank you for finding the bug.
Change-Id: If2c55edd9019e72444ad9b4b8688969eef610c55
Declared the bmi in BLOCKD as a union instead of B_MODE_INFO.
Then removed B_MODE_INFO completely.
Change-Id: Ieb7469899e265892c66f7aeac87b7f2bf38e7a67
This is basically a slightly modified version of the previous patch,
and it has a moderately positive effect (SSIM/PSNR both +0.08% avg
on derf-set). Most clips show no change, except waterfall/coastguard,
each ~ +0.8% SSIM/PSNR. You can see similar effects in other clips
by shortening their length to terminate at a very short last group
of frames.
Change-Id: I7a70de99ca1f9fe6a8b6ca7a6e30e8a4b64383e4
this commit makes the usage errorperbit and sadperbit consistent for
encoding modes and passes. Removed all different magic weight factors
associated with errorperbit. Now 1/2 is used for both sadperbit16 and
sadperbit4, the /2 operation is merged into initializations of the 2
variables.
Tests on cif set show .23%, 0.18% and 0.19% gain by avg psnr, overall
psnr and ssim respectively.
Change-Id: Ifa285c3e065ce0a5a77addfc9f95aabf54ee270d
sad_per_bit has been used for a number of motion vector search routines
with different magic weights: 1, 1/2 and 1/4. This commit remove these
magic numbers and use 1/2 for all motion search routines, also reformat
a number of source code lines to within 80 column limit.
Test on cif set shows overall effect is neutral on all metrics. <=0.01%
Change-Id: I8a382821fa4cffc9c0acf8e8431435a03df74885
vp8_fast_quantize_b_pair_neon function added to quantize
two adjacent blocks at the same time to improve performance.
- Additional 3-6% speedup compared to neon optimized fast
quantizer (Tanya VGA@30fps, 1Mbps stream, cpu-used=-5..-16)
Change-Id: I3fcbf141e5d05e9118c38ca37310458afbabaa4e
Misplaced #endif caused first_time_stamp_ever to only be initialized if
CONFIG_INTERNAL_STATS was set.
Change-Id: I2296a4ab00f7dfb767583edcc5d59b94f48c0621
Added preload instructions to armv6 encoder optimizations.
About 5% average speed-up on Tegra2 for VGA@30fps sequence.
Change-Id: I41d74737720fb71ce7a316f07555357822f3347e
in onyx_if.c update_reference_frames() make
sure that frame buffer indexes are not equal
before preforming a buffer copy. If two frames
share the same buffer the flags will already be
set correctly.
Change-Id: Ida9b5516d08e3435c90f131d2dc19d842cfb536e
Test showed using hex search in realtime mode largely speed up
encoding process, and still achieves similar quality like the
diamond search we have. Therefore, removed the diamond search
option.
Change-Id: I975767d0ec0539f9f6ed7fdfc09506e39761b66c
error_per_bit and sad_per_bit were designed as estimates of a bit worth
of sum squared error and sum absolute difference respectively. Under
this assumption, error_per_bit should be used in combination with 2nd
order errors (variance or sum squared error) while sad_per_bit should
be used in combination with 1st order SADs in motion estimation. There
were a few places where sad_per_bit has been misused with variances,
this commit changes to use error_per_bit for those places, also changes
parameter names to properly indicate which constant is being used.
On cif set, the change has a universal gain by all metrics: 0.13% by
average/overall psnr and 0.1% by ssim.
Change-Id: I4850fdcc3fd6886b30f784bd843f13dd401215fb
While profile=3, there is no sub-pixel search. Distortion and SSE
have to calculated using get_inter_mbpred_error().
Change-Id: Ifb36e17eef7750af93efa7d0e2870142ef540184
Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO.
This reduced the memory footprint by 518,400 bytes for 1080
resolutions. The decoder performance improved by ~4% for the
clip used and the encoder showed very small improvements. (0.5%)
This reduction was first mentioned to me by John K. and in a
later discussion by Yaowu.
This is WIP.
Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
fixed a bug where active_worst_quality could be set
below active_best_quality which could result in an
infinite loop.
Change-Id: I93c229c3bc5bff2a82b4c33f41f8acf4dd194039
This patch collects the twopass specific memebers of VP8_COMP into a
dedicated struct. This is a first step towards isolating the two pass
rate control and aids readability by decorating these variables with
the 'twopass.' namespace. This makes it clear to the reader in what
contexts the variable will be valid, and is a hint that a section of
code might be a good candidate to move to firstpass.c in later
refactoring. There likely will be other rate control modes that need
their own specific data as well.
This notation is probably overly verbose in firstpass.c, so an
alternative would be to access this struct through a pointer like
'rc->' instead of 'cpi->firstpass.' in that file. Feel free to make
a review comment to that effect if you prefer.
Change-Id: I0ab8254647cb4b493a77c16b5d236d0d4a94ca4d
The partition_info struct contains info just for SPLITMV,
so it should be used instead of BLOCKD. Eventually, I want
to reduce the size of B_MODE_INFO struct found in BLOCKD, so
this is the first step toward that goal.
Also, since SPLITMV is not supported in vp8_pick_inter_mode(),
the unnecessary mem copies and checks were removed. For rt
encodes, this gave a slight performance improvement.
Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
This commit restructures the mb activity masking code
to better facilitate experimentation using different metrics
etc. and also allows for adjustment of the zero bin either
for encode only or both the encode and mode selection
stages
It also uses information from the current frame rather than
the previous frame and the default strength has been
reduced.
Change-Id: Id39b19eace37574dc429f25aae810c203709629b
This patch improves the accuracy of frame rate estimation by using a
larger, 1 second window. It also more quickly adapts to step changes
in the input frame rate (ie 30fps to 15fps)
Change-Id: I39e48a8f5ac880b4c4b2ebd81049259b81a0218e
The compiler produces better assembly when using int_mv
for assignments. The compiler shifts and ors the two 16bit
values when assigning MV.
Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
Further modification and wrong implementation fix which caused
refining_search and refining_searchx4 result mismatching.
Change-Id: I80cb3a44bf5824413fd50c972e383eebb75f9b6f
This is to reflect the RD improvement in the encoder. The change has a
small positive impact on quality (0.25% by VPXSSIM and 0.05% by PSNR)
Change-Id: Ic66ffc19b10870645088c0624c85556f009fd210
