The feature undergoes prior assumption that the recursive partition
size search from 4x4 to 64x64, hence utilizing information from small
blocks to determine early termination in large block rate-distortion
optimization search. The current codebase is now going from top down.
The previous function might go with not properly initialized values,
hence removed.
Tested on pedestrian_area_1080p at 4000 kbps running under speed 2.
No visible difference in runtime observed.
Change-Id: I553df415c6191413762db7ae34e8790c71d8118e
Adding RefBuffer to simplify reference buffer management. The struct has a
pointer to image data and scale factors relative to the current frame.
Change-Id: If38eb1491ff687cc11428aee339f3e052e2c5d9e
This commit takes a preliminary attempt to refine the motion search
control. It detects the SAD associated with mv predictor per reference
frame, and based on which to determine whether the encoder wants to
reduce the motion search range (if the predicted mv provides fairly
small SAD), or to skip the current reference frame (if there exists
another ref frame that gives much smaller SAD cost).
This feature is turned on in the settings of speed 1 and above.
In speed 1, compression performance changed
derf -0.018%
yt -0.043%
hd -0.045%
stdhd -0.281%
speed-up
pedestrian_area_1080p at 4000 kbps 100 frames
199651ms -> 188846ms (5.5% speed-up)
blue_sky_1080p at 6000 kbps
443531ms -> 415239ms (6.3% speed-up)
In speed 2, compression performance changed
derf -0.026%
yt -0.090%
hd -0.055%
stdhd -0.210%
speed-up
pedstrian 113949ms -> 108855ms (4.5% speed-up)
blue_sky 271057ms -> 257322ms (5% speed-up)
Change-Id: I1b74ea28278c94fea329d971d706d573983d810d
Buffer the SSE of prediction residuals in the rate-distortion
optimization loop of a given block. This information would be used
for later encoding control.
Change-Id: If4e63f3462490513c48be9407d3327c8dd438367
Moving back to scale_factors struct. We don't need anymore x_offset_q4 and
y_offset_q4 because both values are calculated locally inside vp9_scale_mv
function.
Change-Id: I78a2122ba253c428a14558bda0e78ece738d2b5b
Before mv scaling it is required to calculate x_offset_q4/y_offset_q4
by calling set_scaled_offsets(). Now offset configuration can not be
missed because it happens just before scale_mv().
Change-Id: I7dd1a85b85811a6cc67c46c9b01e6ccbbb06ce3a
This commit reworks the prediction filter rate-distortion cost update
process consistent for all block sizes.
Change-Id: I5874349ab38df380240f96c2d4ef924072bab68d
Various cleanups and refactoring.
Removes feedback of active worst qaulity and uses last_q
instead to make the interface cleaner. Active worst quality
is now decided only once for a frame being coded in the
beginning based on last_q and other stats. Also, adds other
cleaups on last_q to store also the last_q for altref frames,
and reduces the altref interval a little.
The output does change a little.
derfraw300: +0.224% (global psnr)
stdhdraw250: +0.442% (global psnr)
Change-Id: Ie634cdc032697044c472dd0fe79c109b3e7f9767
Properly handle the rd_filter_cache update, when early termination
or skip prediction filter type check is triggered.
Change-Id: Ie7b9a75fed3358f45ffd15817f2b36670c14eb2d
This commit enables an adaptive prediction filter type selection
for sub8x8 block sizes. In speed 1, it re-uses the filter type of
collocated 8x8 block if it is tested in the rate-distortion optimization
loop, for the sub8x8 blocks. Otherwise, it runs the normal test
over all the three filter types. In speed 2, it re-uses the 8x8
block's prediction filter type, if available. Otherwise, force it
to be EIGHTTAP.
Compression and speed performance wise:
speed 1
derf -0.266%
yt -0.138%
bus at 2000 kbps: 33766ms -> 30451ms (10% speed-up)
football at 600 kbps: 48173ms -> 43786ms (9% speed-up)
speed 2
derf -0.026%
yt +0.134%
bus at 2000 kbps: 18973ms -> 17698ms (6% speed-up)
football at 600 kbps: 26748ms -> 25096ms (6% speed-up)
Change-Id: I77e097533b969fd3472147225fa79fc98095d342
This commit constrains the maximal motion search range for sub8x8
blocks to be [-1023, 1023], in the unit of full pixel.
Change-Id: I955b60649364ab410f2453cafd46a496f2fcb43e
Both single frame and compound inter motion search run with luma
component only. Hence removing the block size mapping therein.
Change-Id: I217488e702432ae9fa0e95bf6f516ebb36b5c79b
In the decoder we don't need to save eobs, we can pass eob as an argument.
That's why removing eob arrays from VP9Decompressor and TileWorkerData,
and moving eob pointer from macroblockd_plane to macroblock_plane.
Change-Id: I8eb919acc837acfb3abdd8319af63d1bbca8217a
Its last remaining caller can be passed its results directly without any
additional work. Also, it's not non-4:2:0 safe.
Change-Id: Ia5089ba5f7f66c7617270483c619c9271aefd868
This commit fixes the use of uv_intra_estimate by properly restoring
the mode_info struct required by rd_pick_intra_sbuv_mode.
Change-Id: I6a156d79533c4e2e60dfd3b8c5bb0a42a8eca280
We only need qcoeff buffers in the encoder. Reducing TileWorkerData struct
and VP9Decompressor struct sizes by 24K.
Change-Id: Id148868461f7ffa3d3dd634b371503ae9c57e207
There are many places in handle_inter_mode that need to restore the
dst buffer pointers, due to buffer pointer swap and early rd search
breakout. This commit wraps these operations into an inline function
for clean-up.
Change-Id: I0462e8c41c8bc3cd8db07395489cac03d8e5be54
Commit a4a5a210 enabled lossless coding, but the commit incorrectly
disabled the usage of skip in encoder even when skip should be used.
This commit make sure that skip is enabled even in lossless mode.
Change-Id: I276954f952c6ac68f17a316ebc72f09001228a08
Overall change (using dual buffer scheme for superblocks of both inter
and intra modes) reduces speed 2 runtime:
bluesky_1080p at 6000kbps: 263553ms -> 257441ms
riverbed_1080p at 8000kbps: 233230ms -> 225308ms.
Change-Id: Idf8d70f768a4b0d97b2a8506372c57b7b4022119
This commit enables the dual buffer rate-distortion optimization
and encoding scheme. It stacks the original transform coefficients,
quantized levels, and reconstructed coefficients, in the rate-
distortion optimization search process, hence eliminates the need
to re-run residual generation, forward transform, and quantization
in the encoding stage.
Change-Id: I011bfad3a59a380a869ee552e91dae0394ec492e
Allocate memory space of dual buffer sets that store the coeff, qcoeff,
dqcoeff, and eobs. Connect the pointers of macroblock_plane and
macroblockd_plane to the actual buffer in use accordingly.
Change-Id: I2f0b5f482ca879fae39095013eaf8901db20a5a4
This to make sure that prediction residue always get coded in lossless
mode.
This commit also fixed lossless unit test
Change-Id: I537726ee55328d4e4cf0a0196393a67e12bfcde1
Removing special case handling from vp9_tree_probs_from_distribution(),
tree_merge_probs(), and vp9_tokens_from_tree_offset() functions. Replacing
inter_mode_offset() function with macro INTER_OFFSET which is used now for
vp9_inter_mode_tree definition.
Change-Id: Iff75a1499d460beb949ece543389c8754deaf178
The compound inter prediction could potentially run with initial
motion vectors of invalid value and check the mv_cost, which triggers
overheap read. This commit resolves this issue by forcing a motion
vector value check for compound inter modes of both superblock and
sub8x8 block sizes.
Change-Id: I4f4fc19ce83c8272782bc382f12c82a3f03212fc
We only update partition_probs for inter frames but they are constant
for key frames. It is not necessary to have constants inside frame
context and copy them every time. This change reduces FRAME_CONTEXT size
by at least 48 bytes.
Change-Id: If70a53be51043f37fe7d113853217937710932a7
This commit fixes the use case of plane_block_idx, which determines
the plane (Y/U/V) index based on block index. When block idx >= 4 in
sub8x8 block loop, it should be of chroma components.
Change-Id: I072705aa7b35445524ac607089ca8ce54b7ba478
This commit makes zcoeff_blk cache the case where the entire block
is quantized to be zero (without applying zero-forcing) in the rate-
distortion optimization loop, and skip the forward DCT, quantization,
inverse DCT, and reconstruction process in the encode_block stage.
It now works for all the block sizes, including sub8x8 blocks.
Change-Id: I5ae60a9c436ba3637d11666733554bec4580ef98
"<< SUBPEL_BITS" needs to be added in the calculation. Call
set_scaled_offsets() to calculate x_offset_q4 and y_offset_q4.
Change-Id: Ied130ea771510e918f51cd1dc3abe57f4c0962b5
It is enough to check just block type: intra or inter. Intra block implies
intra prediction mode, and inter block implies inter mode.
Change-Id: I3cf98731a3935f670a3cd8e2b2443483eb944be4
replaces use of cur_tile_mi_(row|col)_(start|end) by VP9_COMMON, making
it less stateful and more reusable for parallel tile decoding
Change-Id: I1df09382b4567a0e5f4434825d47c79afe2399be
Use a flag variable to determine if coded in inter mode, thus avoiding
multiple inter mode checks in super_block_yrd.
Change-Id: I0ef998b2811c38e185a2e0583f0f636cee45d2cf
The ref's scale_factors are set at frame level, and then copied for
each partition block. Since the struct members are mostly constant,
this patch separated the constant and non-constant members, and
reduced struct copying. This gave 0.5% ~ 1.4% decoder speed gain.
Change-Id: I94043bf5a6995c8042da52e5c661818dfa6f6d4c
The pointer was asigned only once with vp9_regular_quantize_b_4x4, calling
this function directly now. Also removing unused declarations:
prototype_quantize_block
prototype_quantize_block_pair
prototype_quantize_mb
vp9_regular_quantize_b_4x4_pair
vp9_regular_quantize_b_8x8
Change-Id: I14325bc2f082336820671eafbc06126651b79f73
This commit uses left_available flag to decide if the left mode_info
struct is available for left_block_mode. As discussed with James
Zern (jzern@), this prevents the codec from fetching mode_info from
blocks in the left tile, which although effectively not used might
present concerns for multi-threaded tile decoding.
This is NOT a bit-stream change.
Change-Id: I1dc8cf1bcbf056688eee27c7bc5706ac4b4e0125
Just making fdct consistent with iht/idct/fht functions which all use
stride (# of elements) as input argument.
Change-Id: I0ba3c52513a5fdd194f1e7e2901092671398985b
Renames for consistency with other constants:
NUM_FRAME_TYPES -> FRAME_TYPES
NUM_PARTITION_CONTEXTS -> PARTITION_CONTEXTS
Change-Id: I3db30acb2868eb0a424237c831087b2e264ec47f
This commit makes the buffer allocation of zcoeff_blk array in
pick_mode_context block size aware. It calculates the number of
4x4 blocks in the partition and assigns the memory space accordingly.
This process (and the uninitialization) is done once for each encoding
pass. It allows memory copy of smaller buffer when possible.
For football at 600kbps, the runtimes improve by about 1%:
speed 1, 45961ms -> 45472ms
speed 2, 23863ms -> 23598ms
Change-Id: Id2ca24906fa89f46fa5fe742ec4b8efc2a61f877
Making this change in order to move allow_high_precision_mv field
from MACROBLOCKD structure to VP9_COMMON (because it is a frame level
flag).
Change-Id: I1d006ba36d938e0caf4d40fa051e2e38df9c1108
* changes:
Use a separate MODE_INFO stream for each tile column
Get rid of "this_mi", use "mi_8x8[0]" everywhere instead
Make the static_segmentation feature work again
The only case where they were intentionally pointing to different
structures was in mbgraph, and this didn't have the expected behavior
because both of these pointers are used interchangeably through the code
Change-Id: I979251782f90885fe962305bcc845bc05907f80c
Moving code that gets band_translate array from get_scan_and_band()
function to get_band_translate() function. Renaming get_scan_and_band() to
get_scan().
Change-Id: I43047c205a1ca2a6e24be44db39dc04b7a385008
This should be similar to what x264 does with --aq-mode 1.
It works well with clips like parkjoy and touhou
(http://x264.nl/developers/Dark_Shikari/LosslessTouhou.mkv).
At low bitrates, the segmentation signaling overhead may negate the
benefits of this feature.
(PGW) Default changed to feature OFF to allow provisional merge.
Change-Id: I938abf9bb487e1d4ad3b0264ea03d9826275c70b
Updated the encoder to handle frames that are coded
intra-only. Intra-only frames must be non-showable,
that is, the "show frame" flag must be set to 0 in
the frame header.
Tested by forcing the ARF frames to be coded intra-
only.
Note: The rate control code will need to be modified
to account for intra-only frames better than they
are currently handled.
Change-Id: I6a9dd5337deddcecc599d3a44a7431909ed21079
Use the zcoeff_blk buffer of PICK_MODE_CONTEXT to store the indexes
of all-zero-coeff block of the current best mode. Remove the temporary
buffer best_zcoeff_blk defined in the rate-distortion optimization
loop. This improves the speed performance by about 0.5% in all speed
settings.
Change-Id: Ie3e15988ddfa581eafa2e19a8228d3fe4a46095c
This commit moves token_cache buffer into macroblock struct, instead
of defining as a local variable in cost_coeffs. This avoids repeatedly
re-allocating memory space in the rate-distortion optimization loop.
The runtime at speed 0 reduces:
bus 2000kbps, 161692ms to 159951ms
football 600kbps, 229505ms to 225821ms
Change-Id: If7da6b0b6d8c5138a16271a33c4548fba33d8840
Converts the constant rddiv parameter to 128 (from 100) and
implements RDCOST with bit-shift rather than multiplication.
Other parameters are also adjusted to roughly keep the same
balance between Rate and Distortion.
There is a slight speed-up of about 0.5-1% (at speed 0) as
testted on football_cif.
There is a slight change in performance due to small change
in the parameters.
derfraw300: +0.033%
stdhdraw250; +0.102%
Change-Id: I70ac69f58fa71c83108f68fe41796cd19d1fc760
The commit changes to mask available intra prediction modes for test
based on prediction block size.
With this patch, encoding time of CpuUsed 2 reduces from 10% to 20% for
HD clips with a compression drop of 0.2%
Change-Id: I65f320f1237c0f5ae3a355bf7caf447f55625455
This commit re-designs the per transformed block rate-distortion
costs tracking buffers. It removes redundant buffer usage, makes
the needed context memory allocation per VP9_COMP instance and
reuses the same buffer sets inside the rate-distortion optimization
search loop, thereby avoiding repeatedly requiring memory space.
It reduces speed 0 runtime:
bus at 2000 kbps from 166763ms to 158967ms,
football at 600 kbps from 246614ms to 234257ms.
Both about 5% speed-up. Local tests suggest about 2% to 5% speed-up
for speed 1 and 2 settings. This does not change compression
performance.
Change-Id: I363514c5276b5cf9a38c7251088ffc6ab7f9a4c3
Increases these parameters.
There is a small efficiency gain.
Change-Id: Ie5f0ddb39c907d335e0dafa5eb112365a81f4542
derfraw300: +0.091%
stdhdraw250: +0.238%
The intra mode distortion adjustment for skip_encode feature was
broken in the refactoring cc91851. This commit fixes it and tunes
the distortion models used therein.
Change-Id: I0d676e82f8e855536a90cf9b3e3fdefafcd886c6
Use b_mode_info to store the inter prediction mode of sub8x8 block,
in replacement of the use of partition_info. Remove redundant buffer
update for partition_info. For bus_cif at 2000 kbps, this seem to make
speed 0 about 1% faster.
Change-Id: Id1b3be45e75a24fb4b42335ac480c23e440978f6
We already have itxm_add member in MACROBLOCKD structure. Both
inv_txm4x4_1_add and inv_txm4x4_add are just its special cases for
different eob values. But eob logic is already implemented in
vp9_iwht4x4_add and vp9_idct4x4_add (that's why also removing
inverse_transform_b_4x4_add).
Change-Id: I80bec9b6f7d40c5e5033c613faca5c819c3e6326
This commit removes the redundant second reference frame check in
the rate-distortion optimization loop for sub8x8 blocks.
Change-Id: I13a57a6f624c4a9bcef02ff2a867fa30d8b44a93
This commit defines b_mode_info as a struct type. This will allow
us to further remove the use of PARTITION_INFO in the encoding process.
Change-Id: I975b0f7d557b5e0f66545a61b472def76b671cce
This commit separates the rate-distortion optimization loop of
superblocks from that of sub8x8 blocks. This allows better design
rate-distortion optimization search loop for each setting. It also
removes the use of SPLITMV and I4X4_PRED therein.
No performance change in speed 0 settings. For bus@CIF at 2000kbps,
the speed 1 runtime goes from 48009ms to 43894ms (about 10% faster).
The overall compression performance on derf changed by -0.021%.
Speed 2 runtime goes from 27114ms to 28700ms (6% slower), while the
overall coding efficiency goes up by 1.629% for derf, 1.236% for yt.
Change-Id: Ie6bdfa0a370148dd60bd800961077f7e97e67dd4
Substantial reworking of the speed vs quality trade offs for
speed 1 and 2.
In this patch I am attempting to freeze the "quality" meaning of
speeds 1 and 2 relative to speed 0 so that in future we can
better evaluate progress.
I am targeting :
Speed 1 quality ~-5% vs speed 0.
Speed 2 quality ~-10% vs speed 0
It is inevitable that quality will still fluctuate a little as we adjust
settings and add new features, but we will attempt to keep as
close as possible to these values. Above speed 2 things will remain
a bit more fluid for now.
In this patch speed 1 is approximately 4-5x as fast as speed 0. This
is similar to before but the quality hit is a lot less. Likewise speed 2
is approximately 2x as fast as speed 1 but is similar in quality to the
previous speed 1 configuration.
Also slight change to behavior of FLAG_EARLY_TERMINATE to insure
all reference frames get at least one rd test. Important for very low
variance regions.
WIP :- Added a new speed level with old speed 4 becoming speed 5.
Speed 3 and 4 tradeoffs still WIP
Change-Id: Ic7a38dd7b5b63ab1501f9352411972f480ac6264
Don't divide RDMULT and RDDIV by 100 when RDMULT > 1000. This was
probably done to avoid overflow when the rd cost was stored in a 32 bits
integer but this is not the case anymore. This change will make it easier
to support multiple quantizers per frame.
derf compression gain at speed 0: 0.037%
Change-Id: Ibeeb9b7cfa1a132a7af41bc90fc07a3bba0857f6
INT64_MAX may be assigned as RDCOST when RDCSOST computation is skipped
for speed, this commit to prevent INT64_MAX from being used as real
RDCOST in transform size decision.
Change-Id: I89a945134191bbdea1f1431ade70424ac079eaac
This commit enables forcing all coefficients zero per transformed
block, when its rate-distortion cost is lower than regular coeff
quantization.
The overall performance improvement (including its parent patch on
calculating rd cost per transformed block) at speed 1:
derf: 0.298%
yt: 0.452%
hd: 0.741%
stdhd: 0.006%
Change-Id: I66005fe0fd7af192c3eba32e02fd6d77952accb5
This commit makes the rate-distortion optimization loop evaluate
the rd costs of regular quantization and all zero coeffs, per
transformed block. It improves speed 1 compression performance:
derf: 0.245%
yt: 0.515%
For a large partition that consists multiple transformed blocks,
this allows more flexibility to selectively force a portion of
them coded as all zero coeffs, as well be continued in the next
patches.
Change-Id: I211518be4179747b57375696f017d1160cc91851
The sub8x8 blocks has its own motion vector reference scheme. The
mv_pred is only used blocks of sizes 8x8 and above, to find the
starting point for motion search.
This change does not change any coding behavior. It makes the
encoding process slightly faster. (0.5% speed-up for local test on
speed 1.)
Change-Id: I746ee6ef0eac19aa3621be014afa12be8d82cbb9
This commit cleans up the second reference check in the
rate-distortion optimization loop of sub8x8 blocks.
Change-Id: Ife68feaa4cddbfad2878c9b44d3012788d634f97
The commit added reset of pred_mv at the beginning of each SB64x64
partition mv search, also limited the usage of pred_mv only when
search on the largest partition is already done. This is to fix
a crash at speed 1/2 encoder where an invalid mv is used in mv
search.
Change-Id: I39010177da76d054e3c90b7899a44feb2e3a5b1b
This commit enables adaptive constraint on motion search range for
smaller partitions, given the motion vectors of collocated larger
partition as a candidate initial search point.
It makes speed 0 runtime of bus at CIF and 2000 kbps goes from
167s down to 162s (3% speed-up), at 0.01dB performance gains. In
the settings of speed 1, this makes the runtime goes from 33687 ms
to 32142 ms (4.5% speed-up), at 0.03dB performance gains.
Compression performance wise, it gains at speed 1:
derf 0.118%
yt 0.237%
hd 0.203%
stdhd 0.438%
Change-Id: Ic8b34c67810d9504a9579bef2825d3fa54b69454
Mis-merge of the following change managed to break mode order
and delete two mode options (new alt ref and near alt ref)
It also created a situation where we could test two undefined
modes off the end of the VP9_mode_order[] data structure.
"clang warnings : remove split and i4x4_pred fake modes"
"Change Id: I8ef3c*"
Initial testing on Akiyo at speed 2.
101.35 44.567 44.447 improves to
96.82 44.915 44.815
Approx 0.3-0.4db gain and 2.5% size reduction
Change-Id: Icff813e7c0778d140ad4f0eea18cf1ed203c4e34
mode_info_context was stored as a grid of MODE_INFO structs.
The grid now constists of pointers to MODE_INFO structs. The
MODE_INFO structs are now stored as a stream (decoder only),
eliminating unnecessary copies and is a little more cache
friendly.
Change-Id: I031d376284c6eb98a38ad5595b797f048a6cfc0d
Thank Paul for the suggestions. While turning on static-thresh
for static-image videos, a big jump on bitrate was seen. In this
patch, we detected static frames in the video using first-pass
stats. For different cases, disable encode breakout or reduce
encode breakout threshold to limit the skipping.
More modification need be done to break incorrect partition
picking pattern for static frames while skipping happens.
Change-Id: Ia25f47041af0f04e229c70a0185e12b0ffa6047f
A previous speed feature skipped modes not used in earlier
partitions but this not longer worked as intended following
changes to the partition coding order and in conjunction
with some other speed features (Especially speed 2 and above).
This modified mode skip feature sets a mask after the first X
modes have been tested in each partition depending on the
reference frame of the current best case.
This patch also makes some changes to the order modes are
tested to fit better with this skip functionality.
Initial testing suggests speed and rd hit count improvements
of up to 20% at speed 1. Quality results. (derf -1.9%, std hd +0.23%).
Change-Id: Idd8efa656cbc0c28f06d09690984c1f18b1115e1
Sample app: vp9_spatial_scalable_encoder
vpx_codec_control extensions:
VP9E_SET_SVC
VP9E_SET_WIDTH, VP9E_SET_HEIGHT, VP9E_SET_LAYER
VP9E_SET_MIN_Q, VP9E_SET_MAX_Q
expanded buffer size for vp9_convolve
modified setting of initial width in vp9_onyx_if.c so that layer size
can be set prior to initial encode
Default number of layers set to 3 (VPX_SS_DEFAULT_LAYERS)
Number of layers set explicitly in vpx_codec_enc_cfg.ss_number_layers
Change-Id: I2c7a6fe6d665113671337032f7ad032430ac4197
mode_info_context was stored as a grid of MODE_INFO structs.
The grid now constists of a pointer to a MODE_INFO struct and
a "in the image" flag. The MODE_INFO structs are now stored
as a stream, eliminating unnecessary copies and is a little
more cache friendly.
For the test clips used, the decoder performance improved
by ~4.3% (1080p) and ~9.7% (720p).
Patch Set 2: Re-encoded clips with latest. Now ~1.7% (1080p)
and 5.9% (720p).
Change-Id: I846f29e88610fce2523ca697a9a9ef2a182e9256
While static-thresh is on, we only need to transmit skip
flag if skip = 1. The cost of skip bit is added to the
total rate cost.
Change-Id: I64e73e482bc297eba22907026298a15fa8cc3920
Added some code to output normalized rd hit count stats.
In effect this approximates to the average number of rd
operations/tests per pixel for the sequence.
The results are not quite accurate and I have not bothered
to account for partial SB64s at frame edges and for key frames
However they do give some idea of the number of modes /
prediction methods being tested for each pixel across the
different partition sizes. This indicates how much scope their
is for further gains either by reducing the number of partitions
examined or the modes per partition through heuristics.
Patch 3 moved place where count incremented so partial rd
tests that are aborted with INT_MAX return are also counted.
Example numbers for first 50 frames of Akiyo.
Speed 0 ~84.4 rd operations / pixel
Speed 1 ~28.8
Speed 2 ~11.9
Change-Id: Ib956e787e12f7fa8b12d3a1a2f6cda19a65a6cb8
The two arrays are typically initialized to INT64_MAX, if they are not
filled with valid values before the addition, the values can overflow
and lead to wrong results.
Change-Id: I515de22cf3e8f55af4b74bdb2c8eb821a02d3059
Switching from mi_{width, height}_log2 and b_{width, height}_log2 to
num_8x8_blocks_{wide, high} and num_4x4_blocks_{wide, high}. Removing
redundant code, adding const.
Change-Id: Iaab2207590fd24d0b76999071778d1395dc5cd5d
This commit resolved a mis-alignment issue in compound inter-inter
prediction of sub8x8. This patch follows solution from dkovalev@.
Change-Id: I3cc0cf7e55b84110e0c42ef4b2e6ca7ac3f8f932
Removing references to plane_block_width and plane_block_height (we are
going to delete the latter ones).
Change-Id: I7982da4d373aebb54d2209dc8886f6192df4d287
Making code more compact, adding consts, removing redundant arguments,
adding do/while(0) for macros.
Change-Id: Ic9ec0bc58cee0910a5450b7fb8cfbf35fa9d0d16
Values now carried over frame to frame.
Change to algorithm for decreasing threshold after
a hit and to max threshold (now based on speed)
Removed some old commented out code relating to
VP8 adaptive thresholds.
The impact of these changes tested on Akiyo (50 frames)
and measured in terms of unit rd hits is as follows:
Speed 0 84.36 -> 84.67
Speed 1 29.48 -> 22.22
Speed 2 11.76 -> 8.21
Speed 3 12.32 -> 7.21
Encode speed impact is broadly in line with these.
Change-Id: I5b886efee3077a11553fa950d796fd6d00c8cb19
Most of the focus so far has been on inter frames.
At high speed settings the key frame is now taking a high %
of the cycles.
This patch puts in some masking to reduce the number
of INTRA modes searched during key frame coding (as already
happens for inter frames) at higher speed settings
TODO: Develop this further with either adaptive rd thresholds
when choosing which intra modes to consider or some other
heuristic.
Impact.
At high speed settings on some clips the key frame was starting
to dominate. In a coding of the first 50 frames of AKIYO at speed
2 limiting the key frame intra modes to DC or TM_PRED resulted in
~30% overall speedup. For Bus the number was lower at ~4-5%.
Change-Id: I7bde68aee04995f9d9beb13a1902143112e341e2
Check the minimum rate-distortion cost of regular quantization and
all zero coeffs cases in the sub8x8 inter prediction rd loop for
luma components. Use this as the cumulative rdcost sent to UV rd
estimation.
Change-Id: Ia4bc7700437d5e13d7cdad4cf9ae57ab036d3e97
Cleans up the switchable filter search logic. Also adds a
speed feature - a variance threshold - to disable filter search
if source variance is lower than this value.
Results: derfraw300
threshold = 16, psnr -0.238%, 4-5% speedup (tested on football)
threshold = 32, psnr -0.381%, 8-9% speedup (tested on football)
threshold = 64, psnr -0.611%, 12-13% speedup (tested on football)
threshold = 96, psnr -0.804%, 16-17% speedup (tested on football)
Based on these results, the threshold is chosen as 16 for speed 1,
32 for speed 2, 64 for speed 3 and 96 for speed 4.
Change-Id: Ib630d39192773b1983d3d349b97973768e170c04
This commit enables early termination in the rate-distortion
optimization search loop for chroma components. When the cumulative
rd cost is above the current best value, skip the rest per-block
transform/quantization/coeff_cost and continue to the next
prediction mode.
For bus_cif at 2000 kbps, the average run-time goes down from
168546ms -> 164678ms, (2% speed-up) at speed 0
36197ms -> 34465ms, (4% speed-up) at speed 1
Change-Id: I9d3043864126e62bd0166250d66b3170d520b3c0
Updating all foreach_transformed_block_visitor functions to work with
plane block size instead of general block. Removing a lot of duplicated
code.
Change-Id: I6a9069e27528c611f5a648e1da0c5a5fd17f1bb4
This change set is intermediate. The next one will remove all repetitive
plane_bsize calculations, because it will be passed as argument to
foreach_transformed_block_visitor.
Change-Id: Ifc12e0b330e017c6851a28746b3a5460b9bf7f0b
Return the distortion value in vp9_rd_pick_intra_mode_sb as sum of
dist_y and dist_uv. Remove the right shift operation on dist_uv,
and make it consistent with that of vp9_rd_pick_inter_mode_sb.
Change-Id: I9d564e242d9add38e32595d33b0e0dddb1d55e5b
This commit makes the rate-distortion optimization search of chroma
components consistent across all block sizes. It removes redundant
codes.
Change-Id: I7e76f54d045e8efdd41d84a164c71f55b484471b
VP9_COMMON is the right place to segmentatation struct because it has
global segmentation parameters, not something specific to macroblock
processing.
Change-Id: Ib9ada0c06c253996eb3b5f6cccf6a323fbbba708
This commit unifies the rate-distortion cost calculation process of
luma and chroma components. It allows early termination to be enabled
later in the rd search loop of chroma components, in consistent with
luma pixels.
Change-Id: I2e52a7c6496176bf2a5e3ef338d34ceb8aad9b3d
The macro block mode info context originally contained an
entry for each 16x16 macroblock. In VP9 each entry refers
to an 8x8 region not a macro block, so the naming is misleading.
This first stage clean up changes the names of 3 entries in the
structure to remove the mb_ prefix.
TODO clean up the nomenclature more widely in respect of
mbmi and bmi.
Change-Id: Ia7305c6d0cb805dfe8cdc98dad21338f502e49c6
Enable SSE2 implementation of high precision 32x32 forward DCT. The
intermediate stacks are of 32-bits. The run-time goes down from
32126 cycles to 13442 cycles.
Change-Id: Ib5ccafe3176c65bd6f2dbdef790bd47bbc880e56
'skippable' can remain unset and negatively affect later decisions
address one aspect of issue #599
Change-Id: Iffdf0ac2e49ac481c27dc27c87fa546d4167bb28
Adds a new subpel motion estimation function that uses a 2-level
tree-structured decision tree to eliminate redundant computations.
It searches fewer points than iterative search (which can search
the same point multiple times) but has the same quality roughly.
This is made the default setting at speeds 0 and 1, while at
speed 2 and above only a 1-level search is used.
Also includes various cleanups for consistency and redundancy removal.
Results:
derf: +0.012% psnr
stdhd: +0.09% psnr
Speedup of about 2-3%
Change-Id: Iedde4866f5475586dea0f0ba4cb7428fba24eee9
Removing the old one bsize_from_dim_lookup. Now we have a way to determine
block size for plane using its subsampling values (ss_size_lookup). And
then we can find the number of pixels in the block (num_pels_log2_lookup).
Change-Id: I6fc981da2ae093de81741d3d78eaefed11015db9
Removes some unused code and speed features, and organizes the
interfaces for fractional mv step functions for use in new speed
features to come.
In the process a new speed feature - number of iterations per
step during the subpel search - is exposed.
No change when this parameter is set as the original value of 3.
Results:
subpel_iters_per_step = 3: baseline
subpel_iters_per_step = 2: psnr -0.067%, 1% speedup
subpel_iters_per_step = 1: psnr -0.331%, 3-4% speedup
Change-Id: I2eba8a21f6461be8caf56af04a5337257a5693a8
Adds a few pattern searches to achieve various tradeoffs
between motion estimation complexity and performance.
The search framework is unified across these searches so that a
common pattern search function is used for all. Besides it will
be easier to experiment with various patterns or combinations
thereof at different scales in the future.
The new pattern search is multi-scale and is capable of using
different patterns at different scales.
The new hex search uses 8 points at the smallest scale
and 6 points at other scales.
Two other pattern searches - big-diamond and square are
also added. Big diamond uses 4 points at the smallest scale and
8 points in diamond shape at the larger scales.
Square is very similar conceptually to the default n-step search
but is somewhat faster since it keeps only one survivor across
all scales.
Psnr/speed-up results on derf300:
hex: -1.6% psnr%, 6-8% speed-up
big-diamond: -0.96% psnr, 4-5% speedup
square: -0.93% psnr, 4-5% speedup
Change-Id: I02a7ef5193f762601e0994e2c99399a3535a43d2
There was no benefit having this function. For example, inside
read_switchable_filter_type switchable filter context was calculated twice.
Change-Id: I79cd5bf95cbc0f6d8bf91a2e32289e01b18dcff1
Adds a speed feature to skip all intra modes other than
DC_PRED if the source variance is small. This feature is
made part of speed 1 and up.
Results on derf300: psnr -0.07%, speedup about 1-2%
Also uses the source variance to fine-tune the early
termination criteria when FLAG_EARLY_TERMINATE is on.
This feature is made part of speed 2 and up.
Results on derf300: psnr -0.52%, speedup about 5-7%
Change-Id: I59e38aa836557cfa5405ae706fc64815cbfe4232
This changeset allows to remove vp9_switchable_interp and
vp9_switchable_interp_map arrays and make code much clear. Actually we
still have to use these mapping but only inside read_interp_filter_type and
write_interp_filter_type functions.
Change-Id: I4026c6f8c4acefba6c81421b7bacbaa52cc45f50
Removing assign_and_clamp_mv function, making implementation of clamp_mv
and clamp_mv2 more clear and consistent.
Change-Id: Iecd08e1c1bf0379f8314ebe01811f8253f4ade58
The function name rd_pick_intra4x4mby_modes is confusing, so
I changed it to rd_pick_intra_sub_8x8_y_modes to better
reflect what the function does. Also added const qualifiers
to some of the input parameters and removed camel-case.
Change-Id: I23d53d4c7af5d79ed8a471acd59a09bbb47add39
This allows us to increment the position at the band-level only as
we go from one band to the next; more importantly, that allows us to
use an add instead of multiply instruction, and omit the instruction
altogether if the band doesn't change from one coef to the next, thus
being slightly faster (probably more noticeable on systems where a
multiply is expensive, like arm).
Change-Id: I4343fe35b9f9a47fa00b217bdcbf5f91ff96c381
Used 3 * standard_deviation in internal threshold calculation
instead of fit curve. This actually approached the algorithm
better.
For comparison, similar tests were done:
The overall psnr loss is less than before.
1. derf set:
when static-thresh = 1, psnr loss is 0.329%;
when static-thresh = 500, psnr loss is 0.970%;
2. stdhd set:
when static-thresh = 1, psnr loss is 0.922%;
when static-thresh = 500, psnr loss is 1.307%;
Similar speedup is achieved. For example,
clip bitrate static-thresh psnr time
akiyo(cif) 500 0 48.952 5.077s(50f)
akiyo 500 500 48.866 4.169s(50f)
parkjoy(1080p) 4000 0 30.388 78.20s(30f)
parkjoy 4000 500 30.367 70.85s(30f)
sunflower(1080p) 4000 0 44.402 74.55s(30f)
sunflower 4000 500 44.414 68.69s(30f)
Change-Id: Ic78833642ce1911dbbd1cb6c899a2d7e2dfcc1f3
This option exists in VP8, and it was rewritten in VP9 to support
skipping on different partition levels. After prediction is done,
we can check if the residuals in the partition block will be all
quantized to 0. If this is true, the skip flag is set, and only
prediction data are needed in reconstruction. Based on DCT's energy
conservation property, the skipping check can be estimated in
spatial domain.
The prediction error is calculated and compared to a threshold.
The threshold is determined by the dequant values, and also
adjusted by partition sizes. To be precise, the DC and AC parts
for Y, U, and V planes are checked to decide skipping or not.
Test showed that
1. derf set:
when static-thresh = 1, psnr loss is 0.666%;
when static-thresh = 500, psnr loss is 1.162%;
2. stdhd set:
when static-thresh = 1, psnr loss is 1.249%;
when static-thresh = 500, psnr loss is 1.668%;
For different clips, encoding speedup range is between several
percentage and 20+% when static-thresh <= 500. For example,
clip bitrate static-thresh psnr time
akiyo(cif) 500 0 48.923 5.635s(50f)
akiyo 500 500 48.863 4.402s(50f)
parkjoy(1080p) 4000 0 30.380 77.54s(30f)
parkjoy 4000 500 30.384 69.59s(30f)
sunflower(1080p) 4000 0 44.461 85.2s(30f)
sunflower 4000 500 44.418 78.1s(30f)
Higher static-thresh values give larger speedup with larger
quality loss.
Change-Id: I857031ceb466ff314ab580ac5ec5d18542203c53
Removing unused constants, macros, and function declarations. Using
ROUND_POWER_OF_TWO macro, vp9_zero, vp9_copy where possible. Moving
#include from *.h to *.c. Merging for loops for motion vectors.
Change-Id: Ic3bf841764a2bb177128bb3a6d7aa8f68229cd13
Simplified the code that extracts and uses the motion
vectors for the 4 sub-partitions in rd_pick_partition.
Change-Id: Iaf698ef7ee3aef9edd59015e1ae065dd359b17d9
Adding plane type check condition because it was always used outside of
get_tx_type_{4x4, 8x8, 16x16}.
Change-Id: I02f0bbfee8063474865bd903eb25b54d26e07230
Although local copies of the mode member variables
(mode, ref_frame) were made, they were not used in
all places. Also, made a local copy of the
second_ref_frame member.
Change-Id: I84d8c822e5cb3d8a02fc3de8a4037ca3fea8bfad
Prevents doing duplicate IDCTs; encoding of first 50 frames of bus
(speed 0) @ 1500kbps goes from 1min4.0 to 1min3.5, i.e. 0.87% faster
overall.
Change-Id: I2df39e29ed9d5ea5e7d2704a34940ba622832ddd
Encoding time of first 50 frames of bus (speed 0) @ 1500kbps goes from
1min5.4 to 1min4.0, i.e. 2.2% faster overall.
Change-Id: I8c32f2aff9a649ce7dd49d910dc5ba16b99c3bc6
All filters are interpolating now, so we don't need this array, all
values from this array are evaluated to true.
Change-Id: I9af6d8219ae0eb984063cd15e4e2296374ae4961
many structures use bw and bh and they have different meanings. This cl attempts
to start this clean up and remove unneccessary 2 step look up log and then
shift operations...
also removed partition type multiple operation code in bitstream.c.
Change-Id: I7e03e552bdfc0939738e430862e3073d30fdd5db
Moving common encoder/decoder code to update_tx_counts. Also renaming
vp9_get_pred_probs_tx_size to get_tx_probs2 and adding get_tx_probs to
call vp9_get_pred_context_tx_size inside read_selected_tx_size only once
(twice before).
Change-Id: Ia50247f3893de88ef8e9041b0d44be44a40aaa4d
Stack the rate-distortion statistics in the sub8x8 rd loop. This allows
the encoder to skip the forward transform, quantization, and coeff cost
estimation, in the sub8x8 rd optimization search, if the motion
vector(s) are of integer pixel value, and have been tested in the
previous prediction filter type rd loops of the same block.
This gives about 2% speed-up for bus_cif at 2000 kpbs, for speed 0.
Its efficacy depends how frequently the motion search will select an
integer motion vector.
Change-Id: Iee15d4283ad4adea05522c1d40b198b127e6dd97
Mode search order in rd loop changed to better reflect
observed hit counts.
Also some adjustment of the baseline mode rd thresholds
to reflect the order change and observed frequencies.
Change-Id: I47a131cc83e11551df8add6d6d8d413d78d3a63c
This commit allows the encoder to skip a few buffer update steps in
rd_pick_best_mbsegmentation, when early breakout has been triggered
in the rd_check_segment_txsize. It provides about 1% speed-up for
bus_cif at 2000 kbps, in the settings of speed 0.
Change-Id: Ica034f10a24dec572b397d8389a2b81020ebc0b9
This patch modifies the auto_mv_step_size speed feature to
use a combination of the maximum magnitude mv from the last
inter frame, and the maximum magnitude mv for the two reference
mvs with the same reference. For arf frames, the max mav step
for the resolution is used.
The bounds therefore are slightly tighter. The feature is made
a speed 1 feature.
Rebased.
Results (when this feature is turned on over speed 0):
derfraw300: -0.046% psnr, about 5+% speedup
(tested on football: goes from 4m30.760s to 4m17.410s).
Change-Id: If492797a61b0b4b3e58c0b8f86afb880165fc9f6
Renaming vp9_sb_mv_ref_tree to vp9_inter_mode_tree, and
vp9_sb_mv_ref_encoding_array to vp9_inter_mode_encodings.
Change-Id: I0e91fbf81350d3ec5a2599064c74089b5d06133a
This prevents a duplicate memcpy of a 128-byte struct every time
set_scale_factors() is called (which is a lot), thus leading to a
decrease from 3.7 MB to 1.85 MB of struct copying per 64x64 block
RD/partition loop.
Overall, this decreases encoding time of the first 50 frames of bus
@ 1500kbps (speed 0) from 1min5.9 to 1min4.9, i.e. about a 1.5%
overall speedup. We can likely get more gains by removing the copy
of the other struct (and replacing it with an indexing) as well.
Change-Id: I3dceb7e79f71e6fe911b11cc994cf89a869dde7a
This means we only do UV intra mode selection if we find any intra
mode to actually be useful at all; in addition, we only do UV intra
mode selection for the transform sizes that were selected, rather
than all sizes available in this partition.
First 50 frames of bus @ 1500kbps (speed 0) gains about 5% with this
change.
Change-Id: I7b461eb8b803247f57896c5a9505f745b55502b3
The break statement only breaks out of the nested loop, not the
top-level loop, so it doesn't always work as intended. Changing it
to a return statement does what's intended.
Change-Id: I585419823b39a04ec8826b1c8a216099b1728ba7
This could happen during golden overlay frame coding from a previous
alt-ref frame if the special overlay code was triggered.
Change-Id: I3056d0c547cd26903b260ef93c94026e96bd9868
Encoding of first 50 frames of bus (speed 0) @ 1500kbps goes from
1min6.2 to 1min5.9, i.e. 0.5% faster overall.
Change-Id: I59d8a3b2f0a75010fa041d5e2646c8caac5bd683
Encode of first 50 frames of bus @ 1500kbps (speed 0) goes from
1min7.3 to 1min6.2, i.e. 1.7% faster overall.
Change-Id: I19d2deacfbffadd61d32551cee9586757ab4a987
Encode of first 50 frames of bus @ 1500kbps (speed 0) goes from 1min12.8
to 1min7.3, i.e. 8% faster.
Change-Id: Ia22d1c7b687316c553cc60eacae988b24e175b62
About 15% faster for bus (speed 0) first 50 frames @ 1500kbps, which
goes from 1min36 to 1min24. Results become slightly better (+0.2% on
derf/yt, +0.4% on hd), probably because of a bugfix for skipmode in
super_block_yrd(). Overall speed change (on derfraw300) is roughly
-13%. This can probably be improved further by caching best_yrd
between partition searches. Also, we might be able to get more
speedups by always doing PARTITION_NONE before PARTITIONS_SPLIT, not
just at the sb8x8 level.
Change-Id: I83736949ebd5b4a3b400ee688d7661913fefc98b
Current partition checking starts from small sizes, and then goes up
to large sizes. This experiment uses the small partitions' motion
estimation result, which is already available, to speed up the
large partition's motion estimation. We can decide to skip some
patition checkings if they are unlikely choices. We could use the
motion vector(MV) result as current partition's prediction MV, limit
the search range and reference frame.
Current result at speed 1:
psnr loss: 1.19% for stdhd, 0.287% for derf.
speed gain: 14% for sunflower(hd), 11% for akiyo.
Further improvement will be done later.
Change-Id: I5abfd070e9cace2e91e2a0247d1325df313887ab
Use an estimate based on DC_PRED for intra uv cost
within the rd loop then only do a full uv mode analysis
if an intra mode is chosen.
Significant speed gains in some cases. Currently only
enabled for speed 2 pending speed/quality tests.
Change-Id: Ie851a12400d5483bce47ec0e3ccb8516041e91c0
This commit makes the encoder to perform motion search only once
per reference frame type for each 4x4/4x8/8x4 block. For bus_cif
at 2000 kbps, the runtime goes from 253812ms -> 217817ms
(14% speed-up) for speed 0.
Change-Id: I5f17599ccc8cfaf93ccb4f98fcb6008af6d79e92
This speed feature allows the encoder to largely remove the spatial
dependency between blocks inside a 64x64 superblock, thereby removing
the need to repeatedly encode superblocks per partition type in the
rate-distortion optimization loop.
A major challenge lies in the intra modes tested in the rate-distortion
optimization loop. The subsequent blocks do not have access to the
reconstructed boundary pixels without the intermediate coding steps.
This was resolved by using the original pixels for intra prediction
in the rd loop, followed by an appropriately designed distortion
modeling on the quantization parameters. Experiments also suggested
that the performance impact is more discernible at lower bit-rate/psnr
settings. Hence a quantizer dependent threshold is applied to deactivate
skip of block coding.
For bus_cif at 2000 kbps,
speed 0: runtime 269854ms -> 237774ms (12% speed-up) at 0.05dB
performance loss.
speed 1: runtime 65312ms -> 61536ms, (7% speed-up) at 0.04dB
performance loss.
This operation is currently turned on in settings of speed 1.
Change-Id: Ib689741dfff8dd38365d8c1b92860a3e176f56ec
Implements some of the helper functions more efficiently with
lookups rathers than branches. Modeling function is consolidated
to reduce some computations.
Also merged the two enums BLOCK_SIZE_TYPES and BlockSize into
one because there is no need to keep them separate (even though
the semantics are a little different).
No bitstream or output change.
About 0.5% speedup
Change-Id: I7d71a66e8031ddb340744dc493f22976052b8f9f
Adding segmentation struct to vp9_seg_common.h. Struct members are from
macroblockd and VP9Common structs. Moving segmentation related constants
and enums to vp9_seg_common.h.
Change-Id: I23fabc33f11a359249f5f80d161daf569d02ec03
This commit fixed the mis-use of the tx_type for inverse transform
in intra4x4 rate-distortion optimization loop. It improves the
overall coding performance.
Change-Id: I7fe9953175b74890357dbcee33c138573766e980
Adds a speed feature to eliminate full-rd computation if the modeled
rd or rd based on a different parameter in the same mode is already
a lot larger than the best rd yet.
Specifically, only search the sharp and smooth filters if the modeled
rd cost based on the regular filter is within a certain factor of the
best rd cost so far. Also, skip full-rd computation of non splitmv
inter modes if the modeled rd cost based on pred error is within the
same factor of the best rd cost so far.
Also adds some enhancements in the rd search for splitmv mode to
speed things up by early breakouts. Negligible impact on performance.
Resuts on derfraw300:
psnr: -0.013% with the splitmv enhancements, -0.24% with the rd
breakout feature on.
speedup: 6% with splitmv enhancements, 20% with also residual breakout
(tested on football sequence at 600 Kbps)
Change-Id: I37abc308ea9f110c1679ce649b6a7e73ab1ad5fc
Encode time of first 50 frames of bus (speed 0) @ 1500kbps goes from
2min4.9 to 2min3.1, i.e. a 1.4% speedup overall.
Change-Id: I9b25e87974430cb942caa276410bb2eda815bd83
Changes cost_mv_ref() into doing a LUT into pre-calculated cost
arrays instead. Encode time of first 50 frames of bus (speed 0)
@ 1500kbps goes from 2min11.6 to 2min10.9, i.e. 0.5% faster overall.
Change-Id: If186e92c34c201b29cbbc058785a15c9c09e433a
Encode time of first 50 frames of bus (speed 0) @ 1500kbps goes from
2min4.9 to 2min3.1, i.e. a 1.4% speedup overall.
Change-Id: Ibe8b08d159797504c5d0c5122de1b6da3b6595e0
Overall, on all test sets, this gains about +0.2% on all metrics.
City is a clip where this really hurts (-1.0% on all metrics), I'm
not quite sure why yet. Maybe interesting to look into in the future.
Change-Id: I6f0eecb20e72f0194633270d30bf00d76d9eae78
Skips mode searches for intra and compound inter modes depending
on the best mode so far and the reference frames. The various
heuristics to be used are selected by bits from a flag. The
previous direction based intra mode search pruning is also absorbed
in this framework.
Specifically the flags and their impact are:
1) FLAG_SKIP_INTRA_BESTINTER (skip intra mode search for oblique
directional modes and TM_PRED if the best so far is
an inter mode)
derfraw300: -0.15%, 10% speedup
2) FLAG_SKIP_INTRA_DIRMISMATCH (skip D27, D63, D117 and D153
mode search if the best so far is not one of the closest
hor/vert/diagonal directions.
derfraw300: -0.05%, about 9% speedup
3) FLAG_SKIP_COMP_BESTINTRA (skip compound prediction mode
search if the best so far is an intra mode)
derfraw300: -0.06%, about 7-8% speedup
4) FLAG_SKIP_COMP_REFMISMATCH (skip compound prediction search
if the best single ref inter mode does not have the same ref
as one of the two references being tested in the compound mode)
derfraw300: -0.56%, about 10% speedup
Change-Id: I1a736cd29b36325489e7af9f32698d6394b2c495
This commit allows encoder to detect the cumulative rate-distortion
cost per transformed block inside a partition. If the cumulative
rd cost is already above the best rd value, it terminates the rest
operations and continue to next prediction mode test.
It reduces the runtime of bus at target bit-rate 2000 from 308 second
to 266 second, i.e., about 13% speed-up at no performance penalty.
Change-Id: I5f15a3d8955d97031d5653006027866a00654e7a
sf->unused_mode_skip_lvl. Tests modes as normal for all
sizes at or below the given level. At larger sizes it skips
all modes that were not chosen at any smaller size.
Hence setting BLOCK_SIZE_SB64X64 is in effect off.
Setting BLOCK_SIZE_AB4X4 will only consider modes that
were chosen for one or more 4x4 blocks at larger sizes.
sf->reference_masking.
Do a test encode of the NONE partition at one size and create
a reference frame mask based on the best rd choice. In the
full search only allow this reference frame.
Currently it is testing 64x64 and repeats this in the full search.
This does not work well with Jim's Partition code just now and
is disabled by default.
Change-Id: I8f8c52d2ef4a0c08100150b0ea4155d1aaab93dd
This speed feature will skip searching the directional intra prediction
modes D63, D117, D27, D153 if the best intra mode so far is not one of
the diagonal, horizontal or vertical directions closest to the respective
directions being tested. In other words, this implements a sort of
binary search in the angular domain.
Speedup: about 9-10%
Results: -0.05% only on derfraw300.
Change-Id: I413584c41f2a3e8dabfbdeb40718c8fc4b1d63a2
(1) Refines the modeling function and uses that to add some speed
features. Specifically, intead of using a flag use_largest_txfm as
a speed feature, an enum tx_size_search_method is used, of which
two of the types are USE_FULL_RD and USE_LARGESTALL. Two other
new types are added:
USE_LARGESTINTRA (use largest only for intra)
USE_LARGESTINTRA_MODELINTER (use largest for intra, and model for
inter)
(2) Another change is that the framework for deciding transform type
is simplified to use a heuristic count based method rather than
an rd based method using txfm_cache. In practice the new method
is found to work just as well - with derf only -0.01 down.
The new method is more compatible with the new framework where
certain rd costs are based on full rd and certain others are
based on modeled rd or are not computed. In this patch the existing
rd based method is still kept for use in the USE_FULL_RD mode.
In the other modes, the count based method is used.
However the recommendation is to remove it eventually since the
benefit is limited, and will remove a lot of complications in
the code
(3) Finally a bug is fixed with the existing use_largest_txfm speed feature
that causes mismatches when the lossless mode and 4x4 WH transform is
forced.
Results on derf:
USE_FULL_RD: +0.03% (due to change in the tables), 0% encode time reduction
USE_LARGESTINTRA: -0.21%, 15% encode time reduction (this one is a
pretty good compromise)
USE_LARGESTINTRA_MODELINTER: -0.98%, 22% encode time reduction
(currently the benefit of modeling is limited for txfm size selection,
but keeping this enum as a placeholder) .
USE_LARGESTALL: -1.05%, 27% encode-time reduction (same as existing
use_largest_txfm speed feature).
Change-Id: I4d60a5f9ce78fbc90cddf2f97ed91d8bc0d4f936
Compute the rate-distortion cost per transformed block, and cumulate
the cost through all blocks inside a partition. This allows encoder
to detect if the cumulative rd cost is already above the best rd cost,
thereby enabling early termination in the rate-distortion optimization
search.
Change-Id: I0a856367a9a7b6dd0b466e7b767f54d5018d09ac
This should significantly speedup cost_coeffs(). Basically what the
patch does is to make the neighbour arrays padded by one item to
prevent an eob check in get_coef_context(), then it populates each
col/row scan and left/top edge coefficient with two times the same
neighbour - this prevents a single/double context branch in
get_coef_context(). Lastly, it populates neighbour arrays in pixel
order (rather than scan order), so we don't have to dereference the
scantable to get the correct neighbours.
Total encoding time of first 50 frames of bus (speed 0) at 1500kbps
goes from 2min10.1 to 2min5.3, i.e. a 2.6% overall speed increase.
Change-Id: I42bcd2210fd7bec03767ef0e2945a665b851df56
Total encoding time for first 50 frames of bus (speed 0) @ 1500kbps
goes 2min34.8 to 2min14.4, i.e. a 10.4% overall speedup. The code is
x86-64 only, it needs some minor modifications to be 32bit compatible,
because it uses 15 xmm registers, whereas 32bit only has 8.
Change-Id: I2df53770c2e850813ffa713e1a91b45b0082b904
Added a speed feature that focuses only on thresholds
for new motion modes.
Moved sf->comp_inter_joint_search_thresh into speed
1. This has ~+0.4% impact on quality at speed 0 as
our quality reference baseline.
Slight adjustment to baseline thresholds.
Change-Id: I7ebf104f1fe29af77ed4837b2e84be065621bbe5
Makes cost_coeffs() a lot faster:
4x4: 236 -> 181 cycles
8x8: 888 -> 588 cycles
16x16: 3550 -> 2483 cycles
32x32: 17392 -> 12010 cycles
Total encode time of first 50 frames of bus (speed 0) @ 1500kbps goes
from 2min51.6 to 2min43.9, i.e. 4.7% overall speedup.
Change-Id: I16b8d595946393c8dc661599550b3f37f5718896
Cycle timings for first 3 frames of bus (speed 0) at 1500kbps:
4x4: 298 -> 234 cycles
8x8: 1227 -> 878 cycles
16x16: 23426 -> 18134 cycles
32x32: 4906 -> 3664 cycles
Total encode time of first 50 frames of bus @ 1500kbps (speed 0) goes
from 3min0.7 to 2min51.6 seconds, i.e. 5.3% faster.
Change-Id: I68a0e1b530b0563b84a67342cca4b45146077e95
This commit replaces zrun_zbin_boost, a method of biasing non-zero
coefficients following runs of zero-coefficients to be rounded towards
zero, with an explicit skip-block choice in the RD loop.
The logic is basically that if individual coefficients should be rounded
towards zero (from a RD point of view), the trellis/optimize loop should
take care of it. If whole blocks should be zero (from a RD point of
view), a single RD check is much more efficient than a complete
serialization of the quantization loop.
Quality change: derf +0.5% psnr, +1.6% ssim; yt +0.6% psnr, +1.1% ssim.
SIMD for quantize will follow in a separate patch. Results for other
test sets pending.
Change-Id: Ife5fa641163ac5150ac428011e87188f1937c1f4
This commit enables configurable reference buffer pointer for intra
predictor. This allows later removal of spatial dependency between
blocks inside a 64x64 superblock in the rate-distortion optimization
loop.
Change-Id: I02418c2077efe19adc86e046a6b49364a980f5b1
Also tweaks to other features and experiments with
what is on and off at different speed settings.
Change-Id: I3e1d0be0d195216bf17c2ac5df67f34ce0b306b2
Each frame we reset all adaptive thresholds to MAX
rather than base. As modes are picked their thresholds
drop down.
Change-Id: Ia37f03a73003c2d9bfcda57edea07205e9a0e5e8
Renamed cpi->sf.first_step to cpi->sf.reduce_first_step_size
and changed its meaning such that it is a delta applied to
reduce the default first step size (>> x) in the motion search
rather than an absolute value.
The default first step size is already changed according to the image
dimensions (smaller for smaller images). cpi->sf.reduce_first_step_size
now applies a further correction from the default.
Change-Id: Ia94e08bc24c67b604831f980909af7e982fcd16d
Change vp9_block_error() to return a 64bit error variable, change all
callers to expect a 64bit return value (this will prevent overflows,
which we basically don't check for at all right now). Remove duplicate
block_error() function, which fixed that through truncation. Remove
old (incompatible) mmx/sse2 block_error SIMD versions and replace with
a new one that returns a 64bit value.
Encoding time of first 50 frames of bus @ 1500kbps goes from 3min29 to
3min23, i.e. a 3% overall speedup.
Change-Id: Ib71ac5508b5ee8a80f1753cd85d72df1629abe68
