Loading to single lanes in multiple registers is expensive since
it requires a read and write of each register which saturates
the register file access. Loading to single registers followed
by a separate transpose reduces this pressure.
Change-Id: I4cc35887ddbca80e5e635b50d2b1d158de9668ee
Removing assign_and_clamp_mv function, making implementation of clamp_mv
and clamp_mv2 more clear and consistent.
Change-Id: Iecd08e1c1bf0379f8314ebe01811f8253f4ade58
Adds a function to compute source variance for various
sb_types to be used for pruning mode and partition searches.
[The existing activity measure function is currently specialized
for only 16x16 MBs and needs to be updated].
Change-Id: I22a41e6f1430184201487326fdbebb9b47e6fc24
The inverse 32x32 transform detects all zero entries and skips the
computations accordingly per 8 rows in the first 1-D operation. The
function vp9_short_idct10_32x32_add performs differently and is not
used anywhere, hence removed.
Change-Id: Ic4fad422debbde7b6b6ffed47c69fbd4268a906c
If the partition is out of partition size range, we don't
need to process small partition information.
Change-Id: Ice9bfbbdebe1f2ef79271a3aee17de0ed4608376
use_min_partition_size and use_max_partition_size are not used
currently, and could be added back if needed later.
Change-Id: Ib22a9c06b064567a7c1d6d5445567ed77e0d3acc
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 commit exploits the sparsity of quantized coefficient matrix.
It detects each 32x8 array and skip the corresponding inverse
transformation if all entries are zero.
For ped1080p at 8000 kbps, this on average reduces the runtime of
32x32 inverse 2D-DCT SSE2 function from 6256 cycles -> 5200
cycles. It makes the overall encoding process about 2% faster at
speed 0. The speed-up is more pronounceable for the decoding process.
Change-Id: If20056c3566bd117642a76f8884c83e8bc8efbcf
This commit removes redundant arguments passing in the function of
rd_pick_reference_frame. This resolves the clang warnings about
potential use of uninitialized values.
Change-Id: Ic68f949a9f8fcd0a583786b0c75321104ea44739
Refactor the frame buffer referencing in choose_partition and make
it consistent with other places. This means to prevent potential
issues when we extend reference frame buffer.
Change-Id: I5ff33ed5f671e1f4cc7049622212769a9b4578d9
The tokenize_b function is only called when output flag is on. Hence
removing the conditional branch on it therein.
Change-Id: Ib709f47f23f39ca05a695faf86fa3377f11f2dd0
This commit optimizes the tokenization and detokenization operational
flow for speed-up. It makes the coding process about 0.3% faster at
speed 0.
Change-Id: I28008df7482874e4b5f237f2d418ff82a249dd56
This commit makes the encoder skip the redundant tokenization process
in the rate-distortion optimization search loop, while updating the
entropy contexts accordingly. It makes the speed 0 encoding process
about 0.5% faster at no performance change.
Change-Id: I34a4155a0b5332afeb45c93a51c7f35a294d685c
This commit provides special handle on 16x16 inverse 2D-DCT, where
only DC coefficient is quantized to be non-zero value.
Change-Id: I7bf71be7fa13384fab453dc8742b5b50e77a277c
Fixes a warning on MSVS 2012 where the alignment of vp9_default_iscan_8x8
didn't match between its declaration and definition.
Change-Id: I1466a15635f4b22594d705d570b7e399bfb6cf21
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
This commit brought back the shortcut implementation of 8x8/16x16
inverse 2D-DCT. When the eob <= 10, it skips the inverse transform
operations on row 4:7/4:15 in the first round. For bus_cif at 1000
kbps, this provides about 2% speed-up at speed 0.
Change-Id: I453e2d72956467d75be4ad8c04b4482ab889d572
This commit enables a special handle for the 8x8 inverse 2D-DCT,
where only DC coefficient is quantized to be non-zero. For bus_cif
at 2000 kbps, it provides about 1% speed-up at speed 0.
Change-Id: I2523222359eec26b144cf8fd4c63a4ad63b1b011
Speed feature experiment to set an upper and lower
partition size limit based on what has been seen
in spatial neighbors.
This seems to gives quite reasonable speed gains in local
(10-15%) and when used with speed 0 the losses are small
(0.25% derf, 0.35% stdhd). However, for now I am only
enabling it on speed 1 as there may be clashes with the existing
temporal partition selection in speed 2.
Using a tighter min / max around the range derived from the
neighbors increases speed further but at the cost of a
bigger quality loss. However, I think this spatial method could
be combined with data from either the last frame or a variance
method (or both) to refine the range of minimum and maximum
partition size. I.e. consider the min and max from spatial and
temporal neighbors and the variance recommendation.
Change-Id: I1b96bf8b84368d6aad0c7aa600fe141b4f07435f
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
Now read_inter_mode_info calls read_intra_block_part (renamed from
read_intra_block_modes) or read_inter_block_part (just added).
Change-Id: I541badea6b663e0ae692ec158665efb90ed20c03
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
This commit makes the initialization of trellis coeff optimization
a per-plane operation, thereby eliminating the redundant steps in
encode_sby and encode_sbuv. It makes the encoder at speed 0 slightly
faster.
Change-Id: Iffe9faca6a109dafc0dd69dc7273cbdec19b17cd
The feature that uses small partition results as a measure to skip
mode evaluation at larger partition requires the flags to be reset.
The reset was missing in the code path that calls rd_use_partition().
Change-Id: Ia0a3a0aee1a862b6e2333d596808db7c48033d50
Add SSE2 implementation to handle the special case of inverse 2D-DCT
where only DC coefficient is non-zero.
Change-Id: I2c6a59e21e5e77b8cf39a4af5eecf4d5ade32e2f
Adding common merge_probs and merge_probs2 functions. Changing ints to
usigned ints in some places.
Change-Id: Icf088ffdea7cf5b95284a128916409bdd53506b0
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
Counts are separate from frame context. We have several frame contexts but
need only one copy of all counts.
Change-Id: I5279b0321cb450bbea7049adaa9275306a7cef7d
The struct optimize_block_args is defined same as encode_b_args.
Remove this redundant definition, and use encode_b_args consistently.
Change-Id: I1703aeeb3bacf92e98a34f4355202712110173d9
All filters are interpolating now, so we don't need this array, all
values from this array are evaluated to true.
Change-Id: I9af6d8219ae0eb984063cd15e4e2296374ae4961
The xform_quant() module is only used by inter modes, hence removing
the redundant switches therein conditioned on tx_type.
Change-Id: Ib87ce5b2f2e4cbf3ceb133a1108afa173c933a3f
When all the transform coefficients were quantized to zero, skip
the inverse transform operation. For bus_cif at 1000 kbps, the
runtime goes from 154967ms -> 149842ms, i.e., about 3% speed-up,
at speed 0.
Change-Id: Ic0a813fff5e28972d4888ee42d8747846a6c3cc6
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
Renamed:
MAX_MB_SEGMENTS to MAX_SEGMENTS
MB_SEG_TREE_PROBS to SEG_TREE_PROBS
The minimum unit for segmentation in the segment map
is now 8x8 so it is misleading to use MB_ as macro-block
traditionally refers to a 16x16 region.
Change-Id: I0b55a6f0426bb46dd13435fcfa5bae0a30a7fa22
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
Update logic for both mode and mvref was the same, so using MODE_COUNT_SAT,
MODE_MAX_UPDATE_FACTOR, update_ct, update_ct2 for both cases. Removing
function update_tx_ct because it was identical to update_mode_ct2.
Change-Id: Iff566be27dbd6cde4c2ec04e8d988f207046b8f0
Optional change in diamond search to continue in the best move
direction until that move turns worse.
This is still WIP since the exact way the new method is to be used is
under investigation. One option is to make it an option in diamond
search and use it only when motion is large.
Overall slightly positive on derfraw300 +0.02%, stdhdraw +0.13%,
but works a lot better for high motion sequences (ex. football : +1%).
Change-Id: If88e01a6021daa0cda934680cdc70be1ee04f798
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
When CONFIG_POSTPROC is set there was a now
invalid reference to cm->filter_level.
Changed to cpi->mb.e_mbd.lf.filter_level in line with
change Iaf5fb71c33719cdfa1b991f671caf071be9ea035
Change-Id: If746e60044903f7ba8d0d346225b3d015226c7d0
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
At speed 2, due to the threshold scheme used, it is possible the rate
and distortion assigned with INT_MAX value. The patch added checking
to prevent the INT_MAX value is used in further calculation of RD
scores. The patch also changed the assertion in rd_use_partition() to
be mirror similar assertion in rd_pick_partition().
Change-Id: Idb52c543cc1e10abdf6e6a5d6e9cb535a42214dc
Adding loopfilter struct with fields from MACROBLOCKD and VP9Common.
Eventually it will be moved to vp9_loopfilter.h for better code structure.
Change-Id: Iaf5fb71c33719cdfa1b991f671caf071be9ea035
Changing fc->tx_probs back to fc->pre_tx_probs. This change actually
affects the bitstream but current test vectors work. Chrome branch is not
affected at all. Broken since:
cc662dd Adding struct tx_probs and struct tx_counts to cleanup the code.
Change-Id: I36dd4b3678e902e10aba8dd49b0012eb558c209d
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
We would skip the rectangular blocks for sub8x8 partitions because
we would conclude that PARTITION_NONE was better than PARTITION_SPLIT,
however, that conclusion was made before we actually really tested
PARTITION_SPLIT.
Change-Id: I8fa91e59894badc1d8cee3ba8a49e40ae4c4a489
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
these are only used in the encoder.
frames_since_golden / frames_till_alt_ref_frame -> VP[89]_COMP
Change-Id: Ie14a6f46987bced685ddb449b85dc261caba6dfe
This could happen during golden overlay frame coding from a previous
alt-ref frame if the special overlay code was triggered.
Change-Id: I3056d0c547cd26903b260ef93c94026e96bd9868
These arrays have constant values (no any updates). Removing two
corresponding memcpy calls. Making a little cleanup in vp9_entropymode.h
as well: removing redundant 'extern' keyword and moving all function
declarations at the end.
Change-Id: Ia16b38b46aec2e2500f5df29c40a297ae241dede
vp9_init_quantizer() is called in vp9_create_compressor(), and
should not be called in vp9_set_speed_features().
Change-Id: Ic2f1f4b0531b9d46bb841d7e1d8da9812207dad6
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
Call the individually optimized horizontal and vertical functions. This
implementation abuses the temp buffer.
This will be replaced with a custom optimized function.
Over 2x speedup.
Change-Id: I5b908d2a73d264e9810d6022bbff73207a3055dd
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
Removing tile_rows and tile_columns from VP9Common, removing redundant
constants MIN_TILE_WIDTH and MAX_TILE_WIDTH, changing signature of
vp9_get_tile_n_bits.
Change-Id: I8ff3104a38179b2c6900df965c144c1d6f602267
Cosmetic code changes, renaming 'flat' local var to 'mask', removing
unused field 'blim' from loopfilter_info_n and loop_filter_info structs.
Change-Id: I51e6ccf727fe361ad9a08e29e1201aa7abd4987f
This commit enables SSE2 implementation of 16x16 inverse ADST/DCT
hybrid transform. The runtime goes from 5742 cycles -> 1821 cycles.
This provides about 1% encoding speed-up at speed 0.
Change-Id: I1678d0988bf30b9efd524877705bbb3645edb17b
Making implementation of vp9_set_pred_flag_{seg_id, mbskip} consistent
with vp9_get_segment_id without using confusing sub(a, b) macro. Passing
mi_row and mi_col to functions explicitly instead of replying on
mb_to_right_edge and mb_to_bottom_edge.
Change-Id: I54c1087dd2ba9036f8ba7eb165b073e807d00435
In the prior code, the above context pointers used for entropy
decoding were initialized on the first frame, and not updated when
the frame size changed. The per-frame code which initializes the
contexts assumes that the contexts are contiguous, leading to an
incomplete initialization when the frame is smaller. This commit
updates the pointers so that the context is contigous whenever
the frame size changes.
Change-Id: I08b53e3a30c8289491212311682ff1b8028cff6c
This is a short term optimization till we work out a decoder
implementation requiring no frame border extension.
Change-Id: I02d15bfde4d926b50a4e58b393d8c4062d1be70f
Cycle times:
4x4: 151 to 131 cycles (15% faster)
8x8: 334 to 306 cycles (9% faster)
16x16: 1401 to 1368 cycles (2.5% faster)
32x32: 7403 to 7367 cycles (0.5% faster)
Total encode time of first 50 frames of bus @ 1500kbps (speed 0)
goes from 1min39.2 to 1min38.6, i.e. a 0.67% overall speedup.
Change-Id: I799a49460e5e3fcab01725564dd49c629bfe935f
Renaming flatmask4 to flat_mask4, flatmask5 to flat_mask5, hevmask to
hev_mask, filter to filter4, mbfilter to filter8, wide_mbfilter to
filter16.
Change-Id: Ic61c73e59c2eee505257584867aafac99833cea1
Also inline some of the block calculations to assist the compiler to
not do silly things like calculating the same offset (or converting
between raster/transform block offset or block, mi and pixel unit)
many, many, many times.
Cycle times:
4x4: 584 -> 505 cycles (16% faster)
8x8: 1651 -> 1560 cycles (6% faster)
16x16: 7897 -> 7704 cycles (2.5% faster)
32x32: 16096 -> 15852 cycles (1.5% faster)
Overall, this saves about 0.5 seconds (1min49.8 -> 1min49.3) on the
first 50 frames of bus (speed 0) @ 1500kbps, i.e. 0.5% overall.
Change-Id: If3dd62453f8e2ab9d4ee616bc4ea956fb8874b80
Removing unused DEC_DEBUG define and dec_debug variable. Changing function
signatures to eliminate code duplication, renaming function
mb_init_dequantizer to init_dequantizer. Also removing redundant curly
braces, and comments.
Change-Id: Ia56ee1b0be5f24abb0e878581845be8a4773c298
Change the mbfilter Neon code from executing both branches if all
vectors follow only one branch.
The code is about 5% faster when executing only one branch and about
1% slower when executing both branches.
-PS5: Remove local stack space from mbfilter.
Change-Id: I6a23f9b318a9f4568a2718b4c9348db988fe2182
Skip the inverse transform and reconstruction of inter-mode coded
blocks in the rate-distortion optimization loop, when skip_encode_sb
feature is turned on. This provides about 1% speed-up at speed 0,
and 1.5% speed-up at speed 1. No performance change in both settings.
Change-Id: I2932718bf4d007163702b61b16b6ff100cf9d007
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
This commit enables SSE2 implementation of 8x8 inverse ADST/DCT
transform. The runtime goes from 1216 cycles -> 266 cycles.
For bus_cif at 2000 kbps, the overall runtime reduces from
253707ms -> 248430ms, i.e., 2% speed-up at speed 0.
Change-Id: Ib0372e17e9162d7b11a10d653b1c8be547c878fb
Adding missed parenthesis around boolean expressions. Bitstream is changed.
Regenerating test vectors.
Change-Id: I4cc00b761e9473f92f180a9fc3a0c607f0aaae56
This function is actually called from set_offsets which is called right
before vp9_read_mode_info.
Change-Id: Ibb9d5ad606194bc80eab264fad85b31c9dfd8f77
Super basic conversion from the other implementations. Any changes to
one should be trivial to copy over keep in sync.
Change-Id: I1720b4128e0aba4b2779e3761f6494f8a09d3ea8
this was never fleshed out in the context of VP8, for which it was
added. for VP9 it has no meaning.
Change-Id: Iba2ecc026d9e947067b96690245d337e51e26eff
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
For some reason iOS builds take a really long time to sort this
function out.
It's not used anywhere so remove it.
Change-Id: Ia5c8513a0d9c7eb32641cca58ca1c1113e2dd9f4
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
Independent horizontal and vertical implementations.
Requires that blocks be built from 4x4 and [xy]_step_q4 == 16
6-10% improvement. CIF improved the least.
Change-Id: I137f5ceae4440adc0960bf88e4453e55a618bcda
The function encode_block is called only by inter-prediction modes,
hence removing the transform type branching there.
Change-Id: I34a3172e28ce2388835efd0f8781922211bff857
This patch is in experimental but was not merged into master.
This patch swaps ptrs instead of copying and uses the
last show_frame flag instead of setting the entire buffer
to zero.
Change-Id: Ia0950466c8ba301a2a5bf917ff3d07bc1a2c2311
With sf->auto_mv_step_size on it is questionable
whether sf->reduce_first_step_size is worthwhile.
At speed 2 it was not having a big impact.
Even at speed 2 sf->optimize_coefficients = 0 is not
having a big speed imapct so for now I have moved it
down into a higher speed setting.
Change-Id: I8a54de76d486ad37aabce76474889da2768b14c1
Enable SSE2 4x4 inverse ADST/DCT transform. The runtime goes from
292 cycles down to 89 cycles. Running bus_cif at 2000 kbps, the
overall runtime of speed 0 goes from 301s to 295s (2% speed-up).
Change-Id: I24098136e7fee7ab2fbf1c11755bdf2ca37f3628
Where possible, do the 16 pixel wide filter while doing the horizontal
filtering pass. The same approach can be taken for the mbloop_filter
when that's implemented. Doing so on the vertical pass is a little more
involved, but possible.
Change-Id: I010cb505e623464247ae8f67fa25a0cdac091320
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
This commit enables 16x16 ADST/DCT forward hybrid transform using SSE2
operations. It reduces the runtime from 5433 cycles to 1621 cycles, at
no compression performance loss.
Change-Id: I75fd7f1984e9e28846af459f810ff0d6ae125230
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
Replace case statement with lookup.
Small speed gain at low speed settings but at speed 2+ where the
number of motion searches etc. falls the impact rises to ~3-4%.
Change-Id: Idff639b7b302ee65e042b7bf836943ac0a06fad8
Change-Id: I5940719a4a161f8c26ac9a6753f1678494cec644
- The vp9 mbfilter C code will branch on flat and mask. This CL
will perform both branches and combine the data. A later CL will
perform a check to see if all patch will take one branch.
- These functions are about 1.75 times faster than the C code on
Nexus 7.
PS #3
- Changed all functions to dub limit, blimit, and thresh from
vld {dx[]}, freeing up r4-r6.
- Changed code to use vbif to reduce one instruction and free
up a d register.
Change-Id: I028dae0e434dc9891c3677bdb182e201ffb04777
This probably has a mildly negative impact on performance, but will
(in future commits - or possibly merged with this one) allow SIMD
implementations of individual intra prediction functions. We may
perhaps want to consider having separate functions per txfm-size
also (i.e. 4x4, 8x8, 16x16 and 32x32 intra prediction functions for
each intra prediction mode), but I haven't played much with that
yet.
Change-Id: Ie739985eee0a3fcbb7aed29ee6910fdb653ea269
In the rare case were 4x4 interior filtering was called for but no
8x8 or larger filtering takes place, the previous code was skipping
the filtering. This patch fixes the issue by including the interior
mask in the overall mask for the filter application loops.
Change-Id: I4a0b65056c64f97478827c2ff41e0914fc7779d0
The resulting reconstruction is never used, thus it just wastes CPU
cycles. Reduces encode time of first 50 frames of bus (speed 0) @
1500kbps from 2min2.0 to 2min1.2, i.e. a 0.65% overall speedup.
Change-Id: I74755ca3aadc21e2be220f486259060bd4088c45
Encode time for first 50 frames of bus (speed 0) @ 1500kbps goes from
2min10.9 to 2min10.5, i.e. 0.3% faster overall, basically because we
prevent the call overhead.
Change-Id: I1eab1a95dd3eae282f9b866f1f0b3dcadff073d5
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
First 50 frames of bus @ 1500kbps (speed 0) goes from 2min12.6 to
2min11.6, i.e. 0.75% overall speedup.
Change-Id: I67054f8146e82a02b6457c51a1c8627a937e5e1e
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
intermediate_height for horizontal filtering must be at least 8
pixels to be able to do vertical filtering correctly. Currently
it can be less for small block and y_step_q4 sizes.
Change-Id: I2ee28b0591b2041c2fa9844d0ae2ff8a1a59cc21
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 patch allows the frame_parallel_decoding_mode flag
to be set instead of returning a codec error.
Change-Id: I4a1631c625723ac8873290d0fd0211074a87d112
This commit adds a speed feature where only squared partition are
evaluated in partition picking. Enable this feature in cpu-used 2
reduces encoding time by ~30%.
loss of compression:
-0.9% on cif set
-1.23% on stdhd
Change-Id: Ia6fad11210f0b78365abb889f9245604513be5b9
If none of the 16 coefficients that we quantize per loop iteration
are larger than the zbin, directly skip to the next round of coeffs,
rather than doing a full quantize loop that will eventually result
in 16 zeroes. This incurs a jump cost, but saves a lot of other work.
32x32 quant goes from 1349 -> 1184 cycles. The same approach yielded
no significantly positive results for smaller transforms, so is not
used there (8x8: 103 -> 101 cycles; 16x16: 302 -> 306 cycles).
Change-Id: I8fca17dc2543fc8eed1dbcd5100145e3c3a9b647
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
Uses mapping tables instead of complicated modulo/division
operations for prob mapping for forward updates.
No bit-stream or output change.
Change-Id: Ifd9ce8ac1437835c305c94f64c18273c7a68f546
Added a speed feature in speed 1 to disable splitmv for HD (>=720)
clips. Test result on stdhd set: 0.3% psnr loss and 0.07% ssim
loss. Encoding speedup is 36%.
(For reference: The test result on derf set showed 2% psnr loss
and 1.6% ssim loss. Encoding speedup is 34%. SPLITMV should be
enabled for small resolution videos.)
Change-Id: I54f72b94f506c6d404b47c42e71acaa5374d6ee6
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
Remove the use of sf->comp_inter_joint_search_thresh
from the baseline speed 0. Approx +0.4% on derf.
Change-Id: Icc14db98909830f40e5ac66130d40e78d2e55c71
This cl converts use partition from last frame to do the following:
if part is none,horz, vert -> try split
if part != none and one of the children is not split - try none
Change-Id: I5b6c659e35f3ac9f11c051b92ba98af6d7e8aa87
Signed-off-by: Jim Bankoski <jimbankoski@google.com>
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
Encode time of bus (speed 0) 50 frames @ 1500kbps goes from 2min14.4 to
2min10.1, i.e. a 2.3% overall speed increase.
Change-Id: I3699580e74ec26c7d24e03681bc47ba25ee1ee87
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