For testing implemented a fixed pattern and delta, 1 pass,
fixed Q, low delay mode.
This has not in any way been tuned or optimized.
Change-Id: Idf5ee179b277fa15d07a97f14f2ce5bbaae80a04
The one pass VBR mode selects a Q range based on a
moving average of recent Q values. This calculation
should have been excluding arf overlay frames as these
are usually coded at the highest allowed value. Their
inclusion skews the average and can cause it to drift
upwards even when the clip as a whole is undershooting.
As such it can undermine correct adaptation of the allowed
Q range especially for easy content.
Change-Id: I7d10fe4227262376aa2dc2a7aec0f1fd82bf11f9
Keep track of frame indexes for the references, and
constrain inter mode search for reference with same
temporal alignment.
Improves speed by about ~15%, no noticeable loss in
compression performance.
Change-Id: I5c407a8acca921234060c4fcef4afd7d734201c8
Lower the threshold for splitting 32x32->16x16 based on average variance,
and add lower bound condition for this split to occur. This prevents
unneccassry splitting for areas with very low variance.
Change-Id: Ibeb33b3d993632c2019f296eb87ef3b7e3568189
For non-rd variannce partition, speed >= 5:
Adjustments to reduce dragging artifcat of background area near
slow moving boundary.
-Decrease base threshold under low source noise conditions.
-Add condition to split 64x64/32x32 based on average variances
of lower level blocks.
PSNR/SSIM metrics go down ~0.7/0.9% on average on RTC set.
Visually helps to reduce dragging artifact on some rtc clips.
Change-Id: If1f0a1aef1ddacd67464520ca070e167abf82fac
This commit makes the sub8x8 block rate-distortion optimization
scheme use precise motion compensated prediction to compute the rd
cost. It fixes a potential buffer overflow issue related to sub8x8
motion search on scaled reference frame.
Change-Id: I4274992ef4f54eaacfde60db045e269c13aaa2de
This commit enables the new temporal filter system for VP9. For
speed 1, it improves the compression performance:
derf 0.54%
stdhd 1.62%
Change-Id: I041760044def943e464345223790d4efad70b91e
This change has been imported from VP9 and
alters the nature and use of exhaustive motion search.
Firstly any exhaustive search is preceded by a normal step search.
The exhaustive search is only carried out if the distortion resulting
from the step search is above a threshold value.
Secondly the simple +/- 64 exhaustive search is replaced by a
multi stage mesh based search where each stage has a range
and step/interval size. Subsequent stages use the best position from
the previous stage as the center of the search but use a reduced range
and interval size.
For example:
stage 1: Range +/- 64 interval 4
stage 2: Range +/- 32 interval 2
stage 3: Range +/- 15 interval 1
This process, especially when it follows on from a normal step
search, has shown itself to be almost as effective as a full range
exhaustive search with step 1 but greatly lowers the computational
complexity such that it can be used in some cases for speeds 0-2.
This patch also removes a double exhaustive search for sub 8x8 blocks
which also contained a bug (the two searches used different distortion
metrics).
For best quality in my test animation sequence this patch has almost
no impact on quality but improves encode speed by more than 5X.
Restricted use in good quality speeds 0-2 yields significant quality gains
on the animation test of 0.2 - 0.5 db with only a small impact on encode
speed. On most natural video clips, however, where the step search
is performing well, the quality gain and speed impact are small.
Change-Id: Iac24152ae239f42a246f39ee5f00fe62d193cb98
For non-rd variance partition: Adjust variance threhsold based
on noise level estimate. This change allows the adjustment to be
updated more frequently.
Change-Id: Ie2abf63bf3f1ee54d0bc4ff497298801fdb92b0d
For low resolutions, whem 4x4downsample is used for variance,
use the same force split (that is used for 8x8downsample) for 16x16 blocks.
No change in metrics. Small improvement visually.
Change-Id: I915b9895902d0b9a41e75d37fee1bf3714d2366d
This is so we may update level at any time (e.g., to be used
for setting thresholds in variance-based partition).
Change-Id: I32caad2271b8e03017a531f9ea456a6dbb9d49c7
Under certain denoising conditons, check for re-evaluation of
zero_last mode if best mode was golden reference.
Change-Id: Ic6cdfd175eef2f7d68606300c7173ab6654b3f6e
For non-rd variance partition: only allow minmax computation
(which currently has no arm-neon optimization) for speeds < 8.
Performance loss is small: On RTC set with speed 8, few clips lose ~2/3%,
average loss is < 1%.
Change-Id: Ia9414f4d0b77dc83c3e73ca8de5d903f64b425ce
Change initial state of noise level, and only update
denoiser with noise level when estimate is done.
Change-Id: If44090d29949d3e4927e855d88241634cdb395dc
For denoising, and for noise level above threshold, re-evaluate
ZEROMV for mode selection after denoising.
Current change only does this check if selected best mode (before denoising)
was intra.
Change-Id: I4b1435b68d26c78f7597b995ee7bff0ddd5f9511
This change makes sure last reference with zero mv
is always checked for mode selection.
No change in metrics.
Change-Id: Iaf01877bf34272b966c78bfe18daad882a0a419e
the final sum may use up to 26 bits
+ add a unit test
+ disable the sse2 as the result will rollover; this will be fixed in a
future commit
Change-Id: I2a49811dfaa06abfd9fa1e1e65ed7cd68e4c97ce
Change on affects 1 pass CBR.
On key frame, temporal layer_id is reset to 0 for 1 pass CBR,
but since "layer" is reset, the svc.layer_context[layer].is_key_frame
was not correspondingly set properly.
Change-Id: I08f6da0a55ac7429ccfbaddfb7be14479e43543b
Small changes to the best quality default speed trade off.
Some speedup settings are worth while even for best quality as they
have only a very small impact on quality but a significant impact on
encode time.
These changes give as much as a further 50-60% increase in encode
speed for my test animations clip with minimal impact on quality.
For this sequence these changes improve the best quality encode speed
to about the same level as good quality speed 0 in Q3 2015 whilst
retaining the large quality gain of over 1 db
For many natural videos though the quality difference from good 0
to best is much smaller.
Change-Id: I28b3840009d77e129817a78a7c41e29cb03e1132
This change alters the nature and use of exhaustive motion search.
Firstly any exhaustive search is preceded by a normal step search.
The exhaustive search is only carried out if the distortion resulting
from the step search is above a threshold value.
Secondly the simple +/- 64 exhaustive search is replaced by a
multi stage mesh based search where each stage has a range
and step/interval size. Subsequent stages use the best position from
the previous stage as the center of the search but use a reduced range
and interval size.
For example:
stage 1: Range +/- 64 interval 4
stage 2: Range +/- 32 interval 2
stage 3: Range +/- 15 interval 1
This process, especially when it follows on from a normal step
search, has shown itself to be almost as effective as a full range
exhaustive search with step 1 but greatly lowers the computational
complexity such that it can be used in some cases for speeds 0-2.
This patch also removes a double exhaustive search for sub 8x8 blocks
which also contained a bug (the two searches used different distortion
metrics).
For best quality in my test animation sequence this patch has almost
no impact on quality but improves encode speed by more than 5X.
Restricted use in good quality speeds 0-2 yields significant quality gains
on the animation test of 0.2 - 0.5 db with only a small impact on encode
speed. On most clips though the quality gain and speed impact are small.
Change-Id: Id22967a840e996e1db273f6ac4ff03f4f52d49aa
This reverts commit 380a5519cc.
This causes an assertion failure in debug_check_frame_counts() which
probably isn't valid with this change; leaving the investigation for
later now.
Change-Id: Ieda5ca811ed2fa50a0cc6935919a8d10dca996e0
This function now has an AVX intrinsics version which is about 80%
faster compared to the C implementation. This provides a 2-4% total
speed-up for encode, depending on encoding parameters. The function
utilizes 3 properties of the cost function lookup table, constructed
in 'cal_nmvjointsadcost' and 'cal_nmvsadcosts'.
For the joint cost:
- mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3]
For the component costs:
- For all i: mvsadcost[0][i] == mvsadcost[1][i]
(equal per component cost)
- For all i: mvsadcost[0][i] == mvsadcost[0][-i]
(Cost function is even)
These must hold, otherwise the AVX version of the function cannot be used.
Change-Id: I6c2791d43022822a9e6ab43cd124a773946d0bdc
Change is only for real-time mode, speed >= 5, and non-screen content mode.
Add bias to zero/low motion for big blocks, if noise estimation
is enabled and noise level is above threshold.
Change-Id: I3a0a4608ede6aa535bda6eca528d20f8aba738e7
For 1 pass CBR mode: increase waiting time after key frame
before we start sampling rate control behavior for determining
resize. This change need to disable one internal resize(DownUp)
temporally since it requires a longer clip to do so.
Change-Id: If21beda1be23f169ee541ab4dd642f718347887a
Use same setting for speed 5 (as it is for speed > 5).
Change is only for real-time (non-rd) mode.
Change-Id: I830250eac654328373cb318baa89d4f0e63942e1
Reduces Linux perf estimated cycle count for pack_mb_tokens on a
lossless encode on my desktop from 61858501855 to 48154040219 or from
26% of the overall profile to 21%.
Change-Id: I9ca3426d7e3272bc7f7030abda4f0d0cec87fb4a
This reverts commit f1342a7b07.
This breaks 32-bit builds:
runtime error: load of misaligned address 0xf72fdd48 for type 'const
__m128i' (vector of 2 'long long' values), which requires 16 byte
alignment
+ _mm_set1_epi64x is incompatible with some versions of visual studio
Change-Id: I6f6fc3c11403344cef78d1c432cdc9147e5c1673
Add threshold/condition on spatial_variance and brightness level.
Modification to normalization of block variance.
Change resolution limit below which we disable noise estimation.
Change-Id: If5be08a26ceda351242d8a58d2f0bc88c0a918f0
This function now has an AVX intrinsics version which is about 80%
faster compared to the C implementation. This provides a 2-4% total
speed-up for encode, depending on encoding parameters. The function
utilizes 3 properties of the cost function lookup table, constructed
in 'cal_nmvjointsadcost' and 'cal_nmvsadcosts'.
For the joint cost:
- mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3]
For the component costs:
- For all i: mvsadcost[0][i] == mvsadcost[1][i]
(equal per component cost)
- For all i: mvsadcost[0][i] == mvsadcost[0][-i]
(Cost function is even)
These must hold, otherwise the AVX version of the function cannot be used.
Change-Id: I184055b864c5a2dc37b2d8c5c9012eb801e9daf6
Change is only for real-time mode, speed > 5, and non-screen content mode.
Bias is based on block size and motion vector level (motion above some threshold).
Helps to improves stability in background from lightning changes.
PSNR/SSIM metrics on RTC set almost no change/neutral (within +/- 0.1).
Change-Id: I7eac13c1ae10be4ab1f40acc7f9f1df5653ece9d
Only use non-zero threshold(s) for breakout if
the motion level of the current tested mode is low.
Change-Id: I22aae961cc42371b49d3f648560181cc54708502
Source noise level estimate is also useful for
setting variance encoder parameters (variance thresholds,
qp-delta, mode selection, etc), so allow it to be used also
if denoising is not on.
Change-Id: I4fe23d47607b4e17a35287057f489c29114beed1
this avoids redefining vpx_codec_vp9_dx, vpx_codec_vp9_dx_algo in
vp9_encoder_parms_get_to_decoder.cc
Change-Id: I3b89e7a62497227ee32419f1a7d30e4c10a13c05
The old workaround "p = 0 ? 0 : p -1" is misleading.
?: happens before =
assigning back to p truncates to one byte.
Therefore it is equivalent to (p - 1) & 0xFF, but the check just exists
to work around a first pass bug, so let's make the work around more
clear.
https://bugs.chromium.org/p/webm/issues/detail?id=1089
Change-Id: I587c44dd61c1f3767543c0126376f881889935af
Width and height of downscaling resolution should not be lower
than min_width and min_height which can be set as needed, both
are 180 for now.
Change-Id: I34d06704ea51affbdd814246e22ee8d41d991f00
This reverts commit 7f56cb2978.
It causes uninitialized reads in the first pass setting up later cost tables.
Change-Id: I2df498df3f5c03eff359f79edf045aed0c618dc9
Adjust variance threshold, delta-qp, and intra penalty cost,
based on estimated noise level in source.
Replace denoising_on with a level value=L/M/H.
Change-Id: I0c017dae75a5d897367d2c42dec26f2f37e447c1
The old workaround "p = 0 ? 0 : p -1" is misleading.
?: happens before =
assigning back to p truncates to one byte.
Therefore it is equivalent to (p - 1) & 0xFF, but the check just exists
to work around a first pass bug, so let's make the work around more
clear.
https://code.google.com/p/webm/issues/detail?id=1089
Change-Id: Ia6dcc8922e1acbac0eeca23a4d564a355c489572
Bug relating to issue:- http://b/25090786
base_frame_target is supposed to track the idealized bit
allocation based on error score and not the actual bits
allocated to each frame.
The clamping of this value based on the VBR min and max pct values
was causing a bug where in some cases the loop that adjusts the
active max quantizer for each GF group was running out of bits at
the end of a KF group. This caused a spike in Q and some ugly artifacts.
A second change makes sure that the calculation of the active
Q range for a group DOES, however, take account of clamping.
Change-Id: I31035e97d18853530b0874b433c1da7703f607d1
Periodically estiamte noise level in source, and only denoise
if estimated noise level is above threshold.
Change-Id: I54f967b3003b0c14d0b1d3dc83cb82ce8cc2d381
A new version of vp9_highbd_error_8bit is now available which is
optimized with AVX assembly. AVX itself does not buy us too much, but
the non-destructive 3 operand format encoding of the 128bit SSEn integer
instructions helps to eliminate move instructions. The Sandy Bridge
micro-architecture cannot eliminate move instructions in the processor
front end, so AVX will help on these machines.
Further 2 optimizations are applied:
1. The common case of computing block error on 4x4 blocks is optimized
as a special case.
2. All arithmetic is speculatively done on 32 bits only. At the end of
the loop, the code detects if overflow might have happened and if so,
the whole computation is re-executed using higher precision arithmetic.
This case however is extremely rare in real use, so we can achieve a
large net gain here.
The optimizations rely on the fact that the coefficients are in the
range [-(2^15-1), 2^15-1], and that the quantized coefficients always
have the same sign as the input coefficients (in the worst case they are
0). These are the same assumptions that the old SSE2 assembly code for
the non high bitdepth configuration relied on. The unit tests have been
updated to take this constraint into consideration when generating test
input data.
Change-Id: I57d9888a74715e7145a5d9987d67891ef68f39b7
Added optimization of the 8 bit assembly quantizer routines. This makes
these functions up to 100% faster, depending on encoding parameters.
This patch maskes the encoder faster in both the high bitdepth and 8bit
configurations. In the high bitdepth configuration, it effects profile 0
only.
Based on my profiling using 1080p input the net gain is between 1-3% for
the 8 bit config, and around 2.5-4.5% for the high bitdepth config,
depending on target bitrate. The difference between the 8 bit and high
bitdepth configurations for the same encoder run is reduced by 1% in all
cases I have profiled.
Change-Id: I86714a6b7364da20cd468cd784247009663a5140
VP8E_UPD_ENTROPY, VP8E_UPD_REFERENCE and VP8E_USE_REFERENCE have been
deprecated since the initial public release
Change-Id: Ied16b441eec13434d85f1ab115d49ccaf5f2f7b0
Adjust the qp threshold and consec_zeromv threshold for
limiting cyclic refresh. Also increase the refresh period
when the limit amount is significant, and some code-cleanup.
Small gain in PSNR/SSIM metrics: ~0.25/0.3 gain on RTC set, speed 7.
Change only affects non-screen content.
Change-Id: I1ced87a89a132684c071e722616e445b2d18236a
Adjust the qp threshold based on the denoising setting; not allow
to scale directly from original resolution to one half and vise versa.
Change-Id: I032a9b22f8e1c88de6bb81cf8351367223a3e40d
For the re-encoding (at max-qp) on the detected high-content change:
update rate correction factor, reset rate over/under-shoot flags,
and update/reset the rate control for layered coding.
Change-Id: I5dc72bb235427344dc87b5235f2b0f31704a034a
Changes to the breakout behavior for partition selection.
The biggest impact is on speed 0 where encode speed in
some cases more than doubles with typically less than 1%
impact on quality.
Speed 0 encode speed impact examples
Animation test clip: +128%
Park Joy: +59%
Old town Cross: + 109%
Change-Id: I222720657e56cede1b2a5539096f788ffb2df3a1
If high bit depth configuration is enabled, but encoding in profile 0,
the code now falls back on optimized SSE2 assembler to compute the
block errors, similar to when high bit depth is not enabled.
Change-Id: I471d1494e541de61a4008f852dbc0d548856484f
The artifact occurs periodically when VP9 denoiser is on and
refresh_golden_frame happen. When refresh_golden_frame happen,
we should copy the frame buffer instead of swapping the pointers.
Change-Id: Ib3204c4b04db28ecf439c6d9e61f3d146f04196d
this reduces the number of synchronizations in decode_tiles_mt() and
improves overall performance when the number of threads is less than the
number of tiles
Change-Id: Iaee6082673dc187ffe0e3d91a701d1e470c62924
Small code cleanup. consec_zeromv refresh threshold
does not need to be computed for every super-block.
No change in behavior.
Change-Id: I8c4b1b28072f42b01d917fff6d1f62722f1e1554
Use the existing VP9_SET_SVC control to set the
first spatial layer to encode.
Since we loop over all spatial layers inside the encoder, the
setting of spatial_layer_id via VP9_SET_SVC has no relevance.
Use it instead to set the first_spatial_layer_to_encode,
which allows an application to skip encoding lower layer(s).
Change only affects the 1 pass CBR SVC.
Change-Id: I5d63ab713c3e250fdf42c637f38d5ec8f60cd1fb
The resolution check fixs the issue which resets resize_pending
unnecessarily and causes not-bitexact with previous one-step version.
Change-Id: I4e7660b3c8f34f59781e2e61ca30d61080c322de
Temporary fix to denoiser when dynamic resizing is on.
-Reallocate denoiser buffers on resized frame.
-Force golden update on resized frame.
-Don't denoise resized frame, and copy source into denoised buffers.
Change-Id: Ife7638173b76a1c49eac7da4f2a30c9c1f4e2000
For screen-content mode, with frame dropper off, put a limit
on how low encoder buffer can go.
Under hard slide changes, the buffer level can go too low and then
take long time to come back up (in particular when frame-dropping
is not used), which will affect the active_worst and target frame size.
Change-Id: Ie9fca097e05cd71141f978ec687f852daf9de332
Dynamic resizing now support two-steps scaling: first go down to
3/4 and then 1/2. This feature is under a flag which controls the
switch between two-steps scaling and one-step scaling (1/2 only).
Change-Id: I3a6c1d3d5668cf8e016a0a02aeca737565604a0f
vp9_filter_block_plane_ss11() and vp9_filter_block_plane_non420()
are only called for the uv planes.
Change-Id: Iacd3b3242c8ce581edd37c8f06d95efc8a0f88a3
The loopfilter masks are now built in the decode loop.
This is done so we can eventually reduce the number of
MODE_INFO structs required by the decoder.
The encoder builds the masks for the entire frame prior
to calling the loopfilter.
Change-Id: Ia2146b07e0acb8c50203e586dfae0c4c5b316f11
When configured with high bitdepth enabled, the 8bit transform
stopped using optimised code. This made 8bit content decode slowly.
Change-Id: I67d91f9b212921d5320f949fc0a0d3f32f90c0ea
In the decoder, map this to the output variable vpx_image_t.r_w/h.
This is intended as an improved version of VP9D_GET_DISPLAY_SIZE,
which doesn't work with parallel frame decoding. In the encoder,
map this to a codec control func (VP9E_SET_RENDER_SIZE) that takes
a w/h pair argument in a int[2] (identical to VP9D_GET_DISPLAY_SIZE).
Also add render_size to the encoder_param_get_to_decoder unit test.
See issue 1030.
Change-Id: I12124c13602d832bf4c44090db08c1009c94c7e8