The speedup in rd_pick_partition() function makes it possible
to drop use_lastframe_partitioning feature. By doing that, we
achieve good PSNR gain with small speed loss. Also, this makes
encoding loop less complicated. The code cleanup patch will
follow.
Borg tests showed:
1. At speed 2,
stdhd set: 0.201% PSNR gain, 0.133% SSIM gain;
derf set: 0.262% PSNR gain, 0.276% SSIM gain.
2. At speed 3,
stdhd set: 0.139% PSNR gain, 0.109% SSIM gain;
derf set: 0.447% PSNR gain, 0.442% SSIM gain.
The average speed loss over selected test clips is within 1%
with the worst case of 4%.
Change-Id: Icfd2ded7869372b585a6972855d933b3d0280d90
The rate costs calculated for inter modes are not precise in some
cases, which causes NEWMV is chosen instead of NEARESTMV, NEARMV,
and ZEROMV. This patch added checks for these cases, and corrected
the mode decisions.
Borg tests at speed 3 showed:
1. stdhd set: 0.102% PSNR gain and 0.088% SSIM gain.
2. derf set: 0.147% PSNR gain and 0.132% SSIM gain.
No speed change.
Change-Id: I35d17684b89ad4734fb610942d707899146426db
Removed functions:
* vp9_post_proc_down_and_across_mmx
* vp9_mbpost_proc_down_mmx
* vp9_plane_add_noise_mmx
They all have sse2 equivalent.
Change-Id: I59c1fac12b7c96ca4538d455e4400c2b7875feff
vp9_variance_sse2.c contains a mix of intrinsics and references to
assembly which uses x86inc.asm; it's conditionally included as a result.
Change-Id: I254451483a65881c0b8e18e27bf0c3ddef60c4ec
allocations within vp9_alloc_context_buffers() rely on mi_rows/mi_cols
individually, use those to determine whether to realloc rather than
stride and stride * rows. this fixes a crash with some fuzzed files for
invalid accesses into last_frame_seg_map and above_context.
Change-Id: I7b9f40dcf170d443890f3bd2acd285507943c7d4
proceeding using a corrupt (incompletely decoded) frame reference may
lead to incorrect assumptions about allocation sizes leading to a crash.
Change-Id: I76e74f2e1be127c2e2c7e1174bb3307497dfd23d
This commit turns on adaptive motion search for ARF coding, in
addition to other normal inter frame coding. It improves the
average compression efficiency:
stdhd 0.1%
derf 0.04%
For the test sequences, the speed 3 runtime is reduced:
pedestrian 1080p 2000 kbps, 149932 ms -> 144580 ms, (3.3% speed-up)
bus CIF 1000 kbps, 8050 ms -> 7895 ms, (1.9%)
highway CIF 100 bkps, 45033 ms -> 44078 ms, (2.2%)
Change-Id: I5228565b609f99e8ae04f6140a2bf2b64a831d21
This is to keep the same with VP8 denoiser.
If motion magnitude is small,
make denoiser more aggressive.
Change-Id: I942a6e2f2ed9aec6f0c4c1f9e5fa47066cadcc0c
When the first try of denoising turns out to be too much,
we will use a softer filter by adopting an adjustment to
make the result closer to original pixel (as in VP8 denoiser).
The old code made the adjustment in the wrong direction.
Change-Id: I84e28fa9e01eef47c5a37d5a2e6d3d378a06786b
This commit allows the encoder to store outcomes of single reference
frame modes and compares them to decide if the inter prediction
filter, forward transform, and quantization can be skipped.
The compression performance of speed 3 is down
derf -0.364%
stdhd -0.198%
For test sequences, the speed 3 runtime is reduced
highway CIF 100 kbps, 51976 ms -> 45033 ms, 13% speed-up
stockholm 720p 1000 kbps, 71826 ms -> 67838 ms, 5.5% speed-up
pedestrian 1080p 2000 kbps, 154924 ms -> 150702 ms, 2.6% speed-up
Change-Id: I5aa26f918d2b4b5197a2c0afa2779319f1c88e44
intra_super_block_yrd() and inter_super_block_yrd() are largely same,
this commit merges them into one to reduce code duplication.
Change-Id: I64d7042a5b099345627cf55663010c185b25ec37
From 3 to 2, which seems to be slightly positive on compression for
all test sets, also reduces encoding time by 2%-5%, varying on the
test clips.
Change-Id: If045417bd27311700c919b4a335eff0dc1130ae0
This commit removes the special case for key frame, as transform size
decision is controlled by the appropriate speed feature for all lossy
coding modes: tx_size_search_method.
Change-Id: I9677171e3f2432ec23705f7c5ea8170dd4562fae
This commit allows the encoder to skip check on compound inter
modes in the rate-distortion optimization loop, if the reference
frame bias signs are the same.
Change-Id: Ib753e6bb11cbdd338aee69dbe2b649671f75a6b0
Adds config parameter vp9_highbitdepth, to support highbitdepth profiles.
Also includes most vpx level high bit-depth functions. However
encode/decode in the highbitdepth profiles will not work until
the rest of the code is in place.
Change-Id: I34c53b253c38873611057a6cbc89a1361b8985a6
It's built based on current spatial svc code.
We only support one spatial two temporal layers at this time.
Change-Id: I1fdc8584354b910331e626bfae60473b3b701ba1
This commit skips the compound inter mode prediction check in the
rate-distortion optimization loop for ARF coding. It reduces the
runtime for certain test clips at speed 3, at no compression
performance change:
bus CIF 1000 kbps, 8260 ms -> 8090 ms, 1.8% speed-up
stockholm 720p 1000 kbps, 74453 ms -> 71826 ms, 2.9% speed-up
No visible speed-up for pedestrian area 1080p at 2000 kbps.
Change-Id: Ic68aa56837159b726563b784e2e3729e846465ad
Use unsigned int type to store the sse in the pixel domain. The
precision is sufficient to handle sse of block size up to 64x64.
The transform domain version however needs int64_t, since there is
a transfer gain applied in the forward transformation that might
cause unsigned int overflow.
Change-Id: Ifef97c38597e426262290f35341fbb093cf0a079
store the number of allocated rows in VP9LfSync, the calculated values
can not be relied on when dealing with corrupt material.
Change-Id: I13b8bcec9738c299a71df726772ab7ac05511e5b
This commit allows encoder to skip intra coding mode test, when
the known inter residual is less than the source variance. It
reduces the runtime of speed 3 for test clips:
bus cif 1000 kbps: 8587 ms -> 8260 ms, 3.8% speed-up
pedestrian 1080p 2000 kbps: 161381 ms -> 155241 ms, 3.7% speed-up.
The compression performance is down by
derf -0.36%
stdhd -0.25%
Change-Id: I75ce1e035b4da2153cb1ac14111d1a07c05a735d
This commit extends the sse and forward transform computation flag
to support the case 64x64 blocks where there are 4 32x32 2D-DCT
blocks.
Change-Id: I86a3e805dfaa0f3abd812f590520c71aa0e40473
In order to understand memory layout consider the declaration of the
following structs. The first one is a part of our API:
struct vpx_codec_ctx {
// ...
struct vpx_codec_priv *priv;
};
The second one is defined in vpx_codec_internal.h:
struct vpx_codec_priv {
// ...
};
The following struct is defined 4 times for encoder/decoder VP8/VP9:
struct vpx_codec_alg_priv {
struct vpx_codec_priv base;
// ...
};
Private data allocation for the given ctx:
struct vpx_codec_ctx *ctx = <get>
struct vpx_codec_alg_priv *alg_priv = <allocate>
ctx->priv = (struct vpx_codec_priv *)alg_priv;
The cast works because vpx_codec_alg_priv has a
vpx_codec_priv instance as a first member 'base'.
Change-Id: I10d1afc8c9a7dfda50baade8c7b0296678bdb0d0
In the partition search, the encoder checks all possible
partitionings in the superblock's partition search tree.
This patch proposed a set of criteria for partition search
early termination, which effectively decided whether or
not to terminate the search in current branch based on the
"skippable" result of the quantized transform coefficients.
The "skippable" information was gathered during the
partition mode search, and no overhead calculations were
introduced.
This patch gives significant encoding speed gains without
sacrificing the quality.
Borg test results:
1. At speed 1,
stdhd set: psnr: +0.074%, ssim: +0.093%;
derf set: psnr: -0.024%, ssim: +0.011%;
2. At speed 2,
stdhd set: psnr: +0.033%, ssim: +0.100%;
derf set: psnr: -0.062%, ssim: +0.003%;
3. At speed 3,
stdhd set: psnr: +0.060%, ssim: +0.190%;
derf set: psnr: -0.064%, ssim: -0.002%;
4. At speed 4,
stdhd set: psnr: +0.070%, ssim: +0.143%;
derf set: psnr: -0.104%, ssim: +0.039%;
The speedup ranges from several percent to 60+%.
speed1 speed2 speed3 speed4
(1080p, 100f):
old_town_cross: 48.2% 23.9% 20.8% 16.5%
park_joy: 11.4% 17.8% 29.4% 18.2%
pedestrian_area: 10.7% 4.0% 4.2% 2.4%
(720p, 200f):
mobcal: 68.1% 36.3% 34.4% 17.7%
parkrun: 15.8% 24.2% 37.1% 16.8%
shields: 45.1% 32.8% 30.1% 9.6%
(cif, 300f)
bus: 3.7% 10.4% 14.0% 7.9%
deadline: 13.6% 14.8% 12.6% 10.9%
mobile: 5.3% 11.5% 14.7% 10.7%
Change-Id: I246c38fb952ad762ce5e365711235b605f470a66
Updates the vp9_pattern_search function to return integer one-away
neighbors' sad values, for subsequent use in speeding up the
sub-pel search. Also, removes code for the do_refine option
which is not being used currently.
Updates the integer and subpel functions to pass in a 5-element
sad list for output or input.
A new pruned sub-pel search algorithm is implemented that uses
the sad returned from the integer pel search. But it is not
deployed yet.
Change-Id: Ifa9f5ad024b5b660570366d2bd900343e1891520
this change is proactive: the loop filter expects valid input and may
produce undefined results / crash in other cases.
Change-Id: I6cc1e966062a91cbc6db981c87cd03d9129fc8fe
attempting to decode a frame after the previous frame failed has the
potential of interrupting an earlier loop filter task
Change-Id: I6f2b1ddcdf5b89c3e2ee8caf5289dada2a087d66
This commit re-work the operation flow related to prediction
residual generation and the rate-distortion modeling. It saves one
call for model_rd_for_sb.
Change-Id: Icaf96c0ff09c903637ed5283448afe01d798195f
The value of switchable rate has been stored in a local variable.
This change skips the second call to vp9_get_switchable_rate() by
reusing the local variable.
Change-Id: Ib7d3fef7621cc4bde94c6d6e6b3a71f1fd4559f2
Check the mode and motion vector cost. If it is already above
the existing best rate-distortion cost, skip the rest check process
on this mode.
Change-Id: Ie065cebdfda2a3be3be18b8e8b43dc29aaa8c179
This commit makes the rate distortion modeling run in the unit of
maximum transform block size. No compression/speed change observed.
It is for the use of later fast forward transform purpose.
Change-Id: Ibaaedb69c765e8d0c5d5012f0ec07f36fd9f68fd
if the first frame was corrupt and loop filter not called, the next call
would assume the necessary allocations had been done and segfault when
accessing a NULL pointer
Change-Id: Ib6ef505e5c594e6f0fe65ab0700172bcf06b92a6
This commit addes a new strategy to reduce the search for optimal
interpolation filter type. The encoder counts and store how many each
filter type is selected and used for each of the reference frames.
A filter type that is rarely used for all three reference frames is
masked out to avoid computation.
The impact on compression is neglectible:
-0.02% on derf
+0.02% on stdhd
Encoding time is seen to reduce by 2~3%.
Change-Id: Ibafa92291b51185de40da513716222db4b230383
We can use one frame context for each layer so that we don't have
to reset the probs every frame. But we can't use prev_mi since we
may drop enhancement layers. So we have to generate a non vp9
compatible bitstream and modify it in the player.
1. We need to code all frames as invisible frame to let prev_mi
not to be used. But in the bitstream we need to code the
show_frame flag to 1 so that the publisher will know it's
supposed to be a visible frame.
2. In the player we need to change the show_frame flag to 0 for
all frames. Then add an one byte frame into the super frame
to tell the decoder which layer we want to show.
Change-Id: I75b7304cf31f0ab952f043e33c034495e88f01f3
The function was called in two places. In the first case it is replaced
with vp9_set_speed_features() call. In the second case the body of set_speed_features() is inlined.
Change-Id: If3fdf1b4168eee97677c224f69c245fe46c7f606
This patch fixes slow first pass problem. Mode could only be determined
from the deadline value during frame encode call. Unfortunately, we use
mode value before any encode calls during the first pass encoding (see
set_speed_features() logic). The mode for the first pass must be different
from BEST to make first pass fast.
Change-Id: I562a7d32004ff631695d91c09a44d8a9076fd6b5
The case where frame width increases but the overall memory
size required to hold the mi arrays does not was not
handled.
Change-Id: I72e70b912a7d1766687ad682979f1c9ee124449b
My fault, that was a float (not integer) which was converted to int64_t.
This reverts commit a885e1cbf0
Change-Id: Ic50708b959e1c3cb3e37da1429d334fafc3391d6
In the full-rd transform size search, we go through all transform
sizes to choose the one with best rd score. In this patch, an
early termination is added to stop the search once we see that the
smaller size won't give better rd score than the larger size. Also,
the search starts from largest transform size, then goes down to
smallest size.
A speed feature tx_size_search_breakout is added, which is turned off
at speed 0, and on for other speeds. The transform size search is
turned on at speed 1.
Borg test results:
1. At speed 1,
derf set: psnr gain: 0.618%, ssim gain: 0.377%;
stdhd set: psnr gain: 0.594%, ssim gain: 0.162%;
No noticeable speed change.
3. At speed 2,
derf set: psnr loss: 0.157%, ssim loss: 0.175%;
stdhd set: psnr loss: 0.090%, ssim loss: 0.101%;
speed gain: ~4%.
Change-Id: I22535cd2017b5e54f2a62bb6a38231aea4268b3f
This commit enables the encoder to record the location of the
center frame to generate alter reference frame. It then allows to
skip checking prediction modes of other reference frame types when
it comes to encode this frame.
The speed 3 runtime is reduced for the test sequences:
bus at CIF 1000 kbps, 9791 ms -> 9446 ms, i.e., 3.5% speed-up,
pedestrian at 1080p 2000 kbps, 184043 ms -> 175730 ms, i.e., 4.5%
speed-up.
No compression performance change observed.
Change-Id: Iacfde3bcc1445964e7a241f239bd6ea11cb94bd1
