- 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
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
This commit enables SSE2 4x4 foward hybrid transform. The runtime
goes from 249 cycles down to 74 cycles. Overall around 2% speed-up
at no compression performance change.
Change-Id: Iad4d526346e05c7be896466c05500711bb763660
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
- Added vp9_loop_filter_horizontal_edge_neon and
vp9_loop_filter_vertical_edge_neon.
- The functions are based off the vp8 loopfilter
functions.
- Matches x86 md5 checksum.
Change-Id: Id1c4dddb03584227e5ecd29f574a6ac27738fdd0
Makes first 50 frames of bus @ 1500kbps encode from 3min22.7 to 3min18.2,
i.e. 2.3% faster. In addition, use the sub_pixel_avg functions to calc
the variance of the averaging predictor. This is slightly suboptimal
because the function is subpixel-position-aware, but it will (at least
for the SSE2 version) not actually use a bilinear filter for a full-pixel
position, thus leading to approximately the same performance compared to
if we implemented an actual average-aware full-pixel variance function.
That gains another 0.3 seconds (i.e. encode time goes to 3min17.4), thus
leading to a total gain of 2.7%.
Change-Id: I3f059d2b04243921868cfed2568d4fa65d7b5acd
This commit makes use of the butterfly structure to enable the sse2
version implementation of 8x8 ADST/DCT hybrid transform coding.
The runtime of hybrid transform module goes down from 1170 cycles
to 245 cycles. Overall speed-up around 1.5%.
Change-Id: Ic808ffd21ece8a9d0410d8c0243d7b6c28ac3b3f
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
Encoding of bus @ 1500kbps (first 50 frames) goes from 3min57 to
3min35, i.e. approximately a 10.5% speedup. Note that the SIMD versions
which use a bilinear filter (x_offset & 7 || y_offset & 7) aren't
perfectly interleaved, and can probably be improved further in the
future. I've marked this with a few TODOs/FIXMEs in the code.
Change-Id: I5c9e900c0f0d32e431a50fecae213b510b2549f9
This commit makes use of dual fdct32x32 versions for rate-distortion
optimization loop and encoding process, respectively. The one for
rd loop requires only 16 bits precision for intermediate steps.
The original fdct32x32 that allows higher intermediate precision (18
bits) was retained for the encoding process only.
This allows speed-up for fdct32x32 in the rd loop. No performance
loss observed.
Change-Id: I3237770e39a8f87ed17ae5513c87228533397cc3
The encoding time for bus at CIF goes from 661s to 625s. This commit
also enabled unit test of sad8x4/4x8 in sad_test.cc.
Change-Id: If3d10ebb56bda584bdb69bcf056599d580b12cb1
The encoding time for bus at CIF goes from 661s to 625s. This commit
also enabled unit test of sad8x4/4x8 in sad_test.cc.
Change-Id: If3d10ebb56bda584bdb69bcf056599d580b12cb1
Modified to work with 8x8 blocks of memory. Will revisit
later for further optimizations. For the HD clip used, the
decoder improved by almost 20%.
Change-Id: Iaa4785be293a32a42e8db07141bd699f504b8c67
Modified to work with 8x8 blocks of memory. Will revisit
later for further optimizations. For the HD clip used, the
decoder improved my 20%.
Change-Id: Ia0057f55d66d1445882351ea6c43b595a5a980e5
This version of the loop filter supports non-4:2:0 subsampling and
a fourth plane, as well as changing the filtering order to be more
friendly to hardware implementations.
The filters are applied first to all vertical edges within the
64x64 SB, followed by the top horizontal edge and any internal
horizontal edges. Since filtering is applied on each 4x4 edge
serially, a dependency is created from filtering one block edge
to the next. It would be possible to remove this depencnecy by
building all filtering decisions from the unfiltered
reconstruction data.
Change-Id: I08f3e9683eb7bded8a76651cbc50fc0dfdd05fa7
This speed 1 - uses variance threshold stolen from static-thresh
to determine split. Any superblock with greater than the variance
set by static thresh * quantizer index squared is split. In addition
transform size is set to largest size less than or equal to partition
size, sub pixel filter is set to normal, and only 12 modes are used
at all.
Change-Id: If7a2858ee70f96d1eb989c04fd87a332b147abef
Move 4x4/4x8/8x4 partition coding out of experimental list.
This commit fixed the unit test failure issues. It also resolved
the merge conflicts between 4x4 block level partition and iterative
motion search for comp_inter_inter.
Change-Id: I898671f0631f5ddc4f5cc68d4c62ead7de9c5a58
This patch eliminates the intermediate diff buffer usage by
combining the short idct and the add residual into one function.
The encoder can use the same code as well.
Change-Id: I296604bf73579c45105de0dd1adbcc91bcc53c22
This patch eliminates the intermediate diff buffer usage by
combining the short idct and the add residual into one function.
The encoder can use the same code as well.
Change-Id: Iacfd57324fbe2b7beca5d7f3dcae25c976e67f45
These building blocks enable rate-distortion optimization search
over block sizes of 8x4 and 4x8. Need to convert them into mmx/sse
forms.
Change-Id: I570ea2d22d14ceec3fe3575128d7dfa172a577de
This patch eliminates the intermediate diff buffer usage by
combining the short idct and the add residual into one function.
The encoder can use the same code as well.
Change-Id: Iea7976b22b1927d24b8004d2a3fddae7ecca3ba1
This patch eliminates the intermediate diff buffer usage by
combining the short idct and the add residual into one function.
The encoder can use the same code as well.
Change-Id: I4ea09df0e162591e420d869b7431c2e7f89a8c1a
In current code, motion vectors got from single prediction mode are used
in compound prediction mode directly. These motion vectors may not give
accurate prediction since they are searched independently. In this patch,
we took Pascal's suggestion, and did joint motion search in compound
prediction mode to find better motion vectors in this situation.
Test results:
Overall PSNR: 0.570%(derf), 0.918%(stdhd);
SSIM: 0.572%(derf), 1.009%(stdhd);
The encoder is a little slower. This can be improved since some c
code is used in motion search.
Change-Id: Ib30c9240f6c56c9b070867b4ca89412a76d9f3c6
All members can be referenced from their per-plane counterparts, and
removes assumptions about 24 blocks per macroblock.
Change-Id: I7ff2fa72d22c29163eb558981c8193765a8113d9
This originally was "Removed update_blockd_bmi()". Now,
this patch removed bmi from blockd and uses the bmi found
in mode_info_context. Eliminates unnecessary bmi copies between
blockd and mode_info_context.
Change-Id: I287a4972974bb363f49e528daa9b2a2293f4bc76
All members can be referenced from their per-plane counterparts, and
removes assumptions about 24 blocks per macroblock.
Change-Id: I593fb0715e74cd84b48facd1c9b18c3ae1185d4b
First in a series of commits making certain MACROBLOCK members
addressable per-plane. This commit also refactors the block subtraction
functions vp9_subtract_b, vp9_subtract_sby_c, etc to be
loops-over-planes and variable subsampling aware.
Change-Id: I371d092b914ae0a495dfd852ea1a3d2467be6ec3
Mostly for cleanup purposes. Now we should be able to rework
the encoder/decoder to use a common idct/add function.
Change-Id: I1597cc59812f362ecec0a3493b6101a6cc6fa7ff
Remove the unnecessary _s_ from their names, and add a new
vp9_recon_sb() that calls the y and uv variants.
Change-Id: I7ffaa5ff5605a8472cac2a53de8cf889353039a6
Use in-place buffers (dst of MACROBLOCKD) for macroblock prediction.
This makes the macroblock buffer handling consistent with those of
superblock. Remove predictor buffer MACROBLOCKD.
Change-Id: Id1bcd898961097b1e6230c10f0130753a59fc6df
The intra predictor supports configurable block sizes. It can handle
intra prediction down to 4x4 sizes, when enabled in BLOCK_SIZE_TYPE.
Change-Id: I7399ec2512393aa98aadda9813ca0c83e19af854
This patch will use the dest buffer instead of the
predictor buffer. This will allow us in future commits
to remove the extra mem copy that occurs in the dequant
functions when eob == 0. We should also be able to remove
extra params that are passed into the dequant functions.
Change-Id: I7241bc1ab797a430418b1f3a95b5476db7455f6a
Merge various super_block_yrd and super_block_uvrd versions into one
common function that works for all sizes. Make transform size selection
size-agnostic also. This fixes a slight bug in the intra UV superblock
code where it used the wrong transform size for txsz > 8x8, and stores
the txsz selection for superblocks properly (instead of forgetting it).
Lastly, it removes the trellis search that was done for 16x16 intra
predictors, since trellis is relatively expensive and should thus only
be done after RD mode selection.
Gives basically identical results on derf (+0.009%).
Change-Id: If4485c6f0a0fe4038b3172f7a238477c35a6f8d3
Merge sb32x32 and sb64x64 functions; allow for rectangular sizes. Code
gives identical encoder results before and after. There are a few
macros for rectangular block sizes under the sbsegment experiment; this
experiment is not yet functional and should not yet be used.
Change-Id: I71f93b5d2a1596e99a6f01f29c3f0a456694d728
Start grouping data per-plane, as part of refactoring to support
additional planes, and chroma planes with other-than 4:2:0
subsampling.
Change-Id: Idb76a0e23ab239180c818025bae1f36f1608bb23
Wrote sse2 version of vp9_short_idct_32x32 function. Compared
to c version, the sse2 version is 5X faster.
Change-Id: I071ab7378358346ab4d9c6e2980f713c3c209864
Adds an experiment to use a weighted prediction of two INTER
predictors, where the weight is one of (1/4, 3/4), (3/8, 5/8),
(1/2, 1/2), (5/8, 3/8) or (3/4, 1/4), and is chosen implicitly
based on consistency of the predictors to the already
reconstructed pixels to the top and left of the current macroblock
or superblock.
Currently the weighting is not applied to SPLITMV modes, which
default to the usual (1/2, 1/2) weighting. However the code is in
place controlled by a macro. The same weighting is used for Y and
UV components, where the weight is derived from analyzing the Y
component only.
Results (over compound inter-intra experiment)
derf: +0.18%
yt: +0.34%
hd: +0.49%
stdhd: +0.23%
The experiment suggests bigger benefit for explicitly signaled weights.
Change-Id: I5438539ff4485c5752874cd1eb078ff14bf5235a
Wrote sse2 version of vp9_short_idct10_16x16 function. Compared
to c version, the sse2 version is 2.3X faster.
Change-Id: I314c4f09369648721798321eeed6f58e38857f26
Wrote sse2 version of vp9_short_idct16x16 function. Compared to c
version, the sse2 version is over 2.5X faster.
Change-Id: I38536e2b846427a2cc5c5423aaf305fd0e605d61
Wrote sse2 functions of vp9_short_idct8x8 and vp9_short_idct10_8x8.
Compared to c version, the sse2 version is 2X faster. The decoder
test didn't show noticeable gain since 8x8 idct doesn't take much
of decoding time (less than 1% in my test).
Change-Id: I56313e18cd481700b3b52c4eda5ca204ca6365f3
Added SSE2 idct4_1d which is called by vp9_short_iht4x4. Also,
modified the parameter type passed to vp9_short_iht functions to
make it work with rtcd prototype.
Change-Id: I81ba7cb4db6738f1923383b52a06deb760923ffe
Split macroblock and superblock tokenization and detokenization
functions and coefficient-related data structs so that the bitstream
layout and related code of superblock coefficients looks less like it's
a hack to fit macroblocks in superblocks.
In addition, unify chroma transform size selection from luma transform
size (i.e. always use the same size, as long as it fits the predictor);
in practice, this means 32x32 and 64x64 superblocks using the 16x16 luma
transform will now use the 16x16 (instead of the 8x8) chroma transform,
and 64x64 superblocks using the 32x32 luma transform will now use the
32x32 (instead of the 16x16) chroma transform.
Lastly, add a trellis optimize function for 32x32 transform blocks.
HD gains about 0.3%, STDHD about 0.15% and derf about 0.1%. There's
a few negative points here and there that I might want to analyze
a little closer.
Change-Id: Ibad7c3ddfe1acfc52771dfc27c03e9783e054430
Simplified idct32x32 calculation when there are only 10 or less
non-zero coefficients in 32x32 block. This helps the decoder
performance.
Change-Id: If7f8893d27b64a9892b4b2621a37fdf4ac0c2a6d
sse4_1 code used uint16_t for returning sad, but that
won't work for 32x32 or 64x64. This code fixes the
assembly for those and also reenables sse4_1 on linux
Change-Id: I5ce7288d581db870a148e5f7c5092826f59edd81
This function was part of an optimization used in VP8 that required
caching two macroblocks. This is unused in VP9, and might not
survive refactoring to support superblocks, so removing it for now.
Change-Id: I744e585206ccc1ef9a402665c33863fc9fb46f0d
Wrote SSE2 version of vp9_dc_only_idct_add_c function. In order to
improve performance, clipped the absolute diff values to [0, 255].
This allowed us to keep the additions/subtractions in 8 bits.
Test showed an over 2% decoder performance increase.
Change-Id: Ie1a236d23d207e4ffcd1fc9f3d77462a9c7fe09d
Rebased.
Remove the old matrix multiplication transform computation. The 16x16
ADST/DCT can be switched on/off and evaluated by setting ACTIVE_HT16
300/0 in vp9/common/vp9_blockd.h.
Change-Id: Icab2dbd18538987e1dc4e88c45abfc4cfc6e133f
This patch includes 4x4, 8x8, and 16x16 forward butterfly ADST/DCT
hybrid transform. The kernel of 4x4 ADST is sin((2k+1)*(n+1)/(2N+1)).
The kernel of 8x8/16x16 ADST is of the form sin((2k+1)*(2n+1)/4N).
Change-Id: I8f1ab3843ce32eb287ab766f92e0611e1c5cb4c1
rebased.
This patch includes 16x16 butterfly inverse ADST/DCT hybrid
transform. It uses the variant ADST of kernel
sin((2k+1)*(2n+1)/4N),
which allows a butterfly implementation.
The coding gains as compared to DCT 16x16 are about 0.1% for
both derf and std-hd. It is noteworthy that for std-hd sets
many sequences gains about 0.5%, some 0.2%. There are also few
points that provides -1% to -3% performance. Hence the average
goes to about 0.1%.
Change-Id: Ie80ac84cf403390f6e5d282caa58723739e5ec17
Initial ssse3 convolve avg functions and is one step closer
to using x86inc.asm. The decoder performance improved by 8% for
the test clip used. This should be revisited later to see if
averaging outside the loop is better than having many similar
filter functions.
Change-Id: Ice3fafb423b02710b0448ffca18b296bcac649e9
fixed format issues.
Implement the inverse 4x4 ADST using 9 multiplications. For this
particular dimension, the original ADST transform can be
factorized into simpler operations, hence is retained.
Change-Id: Ie5d9749942468df299ab74e90d92cd899569e960
Also port the 4x4, 16x16, 8x16 and 16x8 versions to x86inc.asm; this
makes them all slightly faster, particularly on x86-64. Remove SSE3
sad16x16 version, since the SSE2 version is now faster.
About 1.5% overall encoding speedup.
Change-Id: Id4011a78cce7839f554b301d0800d5ca021af797
This commit adds the 8 tap SSSE3 subpixel filters back into the code
underneath the convolve API. The C code is still called for 4x4
blocks, as well as compound prediction modes. This restores the
encode performance to be within about 8% of the baseline.
Change-Id: Ife0d81477075ae33c05b53c65003951efdc8b09c
Refactor the 8x8 inverse hybrid transform. It is now consistent
with the new inverse DCT. Overall performance loss (due to the
use of this variant ADST, and the rounding errors in the butterfly
implementation) for std-hd is -0.02.
Fixed BUILD warning.
Devise a variant of the original ADST, which allows butterfly
computation structure. This new transform has kernel of the
form: sin((2k+1)*(2n+1) / (4N)). One of its butterfly structures
using floating-point multiplications was reported in Z. Wang,
"Fast algorithms for the discrete W transform and for the discrete
Fourier transform", IEEE Trans. on ASSP, 1984.
This patch includes the butterfly implementation of the inverse
ADST/DCT hybrid transform of dimension 8x8.
Change-Id: I3533cb715f749343a80b9087ce34b3e776d1581d
This patch adds column-based tiling. The idea is to make each tile
independently decodable (after reading the common frame header) and
also independendly encodable (minus within-frame cost adjustments in
the RD loop) to speed-up hardware & software en/decoders if they used
multi-threading. Column-based tiling has the added advantage (over
other tiling methods) that it minimizes realtime use-case latency,
since all threads can start encoding data as soon as the first SB-row
worth of data is available to the encoder.
There is some test code that does random tile ordering in the decoder,
to confirm that each tile is indeed independently decodable from other
tiles in the same frame. At tile edges, all contexts assume default
values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode),
and motion vector search and ordering do not cross tiles in the same
frame.
t log
Tile independence is not maintained between frames ATM, i.e. tile 0 of
frame 1 is free to use motion vectors that point into any tile of frame
0. We support 1 (i.e. no tiling), 2 or 4 column-tiles.
The loopfilter crosses tile boundaries. I discussed this briefly with Aki
and he says that's OK. An in-loop loopfilter would need to do some sync
between tile threads, but that shouldn't be a big issue.
Resuls: with tiling disabled, we go up slightly because of improved edge
use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf,
~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5%
on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is
concentrated in the low-bitrate end of clips, and most of it is because
of the loss of edges at tile boundaries and the resulting loss of intra
predictors.
TODO:
- more tiles (perhaps allow row-based tiling also, and max. 8 tiles)?
- maybe optionally (for EC purposes), motion vectors themselves
should not cross tile edges, or we should emulate such borders as
if they were off-frame, to limit error propagation to within one
tile only. This doesn't have to be the default behaviour but could
be an optional bitstream flag.
Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f
Update the code to call the new convolution functions to do subpixel
prediction rather than the existing functions. Remove the old C and
assembly code, since it is unused. This causes a 50% performance
reduction on the decoder, but that will be resolved when the asm for
the new functions is available.
There is no consensus for whether 6-tap or 2-tap predictors will be
supported in the final codec, so these filters are implemented in
terms of the 8-tap code, so that quality testing of these modes
can continue. Implementing the lower complexity algorithms is a
simple exercise, should it be necessary.
This code produces slightly better results in the EIGHTTAP_SMOOTH
case, since the filter is now applied in only one direction when
the subpel motion is only in one direction. Like the previous code,
the filtering is skipped entirely on full-pel MVs. This combination
seems to give the best quality gains, but this may be indicative of a
bug in the encoder's filter selection, since the encoder could
achieve the result of skipping the filtering on full-pel by selecting
one of the other filters. This should be revisited.
Quality gains on derf positive on almost all clips. The only clip
that seemed to be hurt at all datarates was football
(-0.115% PSNR average, -0.587% min). Overall averages 0.375% PSNR,
0.347% SSIM.
Change-Id: I7d469716091b1d89b4b08adde5863999319d69ff
This commit introduces a new convolution function which will be used to
replace the existing subpixel interpolation functions. It is much the
same as the existing functions, but allows for changing the filter
kernel on a per-pixel basis, and doesn't bake in knowledge of the
filter to be applied or the size of the resulting block into the
function name.
Replacing the existing subpel filters will come in a later commit.
Change-Id: Ic9a5615f2f456cb77f96741856fc650d6d78bb91
and called this function in vp9_dequant_idct_add_32x32_c when
eob == 1. For the test clip used, the decoder performance improved
by 21+%. Based on Yaowu's 16 point idct work.
Change-Id: Ib579a90fed531d45777980e04bf0c9b23c093c43
fixed a function prototypes to prevent compiler warnings;
removed a function not in use;
un-capitialize "Refstride" to ref_stride
Change-Id: Ib4472b6084f357d96328c6a06e795b6813a9edba
This commit changes the inverse 16 point dct to use the same algorithm
as the one for 32 point idct. In fact, now 16 point dct uses the exact
version of the souce code for even portion of the 32 point idct.
Tests showed current implementation has significant better accuracy
than the previous version. With this implementation and the minor bug
fix on forward 16 point dct, encoding tests showed about 0.2% better
compression of CIF set, test results on std-hd setting pending.
Change-Id: I68224b60c816ba03434e9f08bee147c7e344fb63
This experiment gives little gains and adds relatively much code
complexity (and it hinders other experiments), so let's get rid of
it.
Change-Id: Id25e79a137a1b8a01138aa27a1fa0ba4a2df274a
Updated the rtcd_defs and used the sse2 uv version
of the loopfilter. The performance improved by ~8%
for the test clip used.
Change-Id: I5a0bca3b6674198d40ca4a77b8cc722ddde79c36
