Adds an early exit based on ptest. Slightly slower than ssse3 in the
full case because of the extra check, but potentially faster if lots of
rows can be skipped.
Very close in speed to the assembly.
Can run in 32 bit, unlike the assembly. Allows reworking the function
prototype to use structs.
Change-Id: If80e2b9ba059370a4cad3c973196e82a97b4330e
About 4x faster when values are below the dequant threshold and 10x
faster if everything needs to be calculated.
Both numbers would improve if the division for dqcoeff could be
simplified.
BUG=webm:1426
Change-Id: I8da67c1f3fcb4abed8751990c1afe00bc841f4b2
This condition is handled before this code is reached. The ssse3 version
of the function has always crashed when attempting to handle the
skip_block condition.
Add assert() and comments regarding the usage of skip_block.
Removing the parameter is a fairly involved process so leave it be for
the moment.
Change-Id: Ib299f6fc6589d7ee102262cc74a7aeb60110bc5a
Having a very small value for "lag_in_frames" can result in
corrupt arf groups including displayed frames that update
the arf buffer and fake overlay frames that are not in fact
overlays of real arfs but are nevertheless starved of bits.
Leaving lag_in_frames at the default of 25 for these 5 frame two
pass VBR tests should now give rise to a valid ARF coding pattern
as follows:- K(ey), A(rf), N(ormal), N, N, O(verlay).
This change is part of a response to BUG=webm:1454 where broken
arf groups interacted badly with a change that corrects for large rate
misses. However, it may still in some cases increase encode time by
virtue of the fact that the unit test now codes a correct coding pattern
with "hidden" ARF frames.
Change-Id: Ifd0246a4c1d0be247247c754024d7a4ed5f66a6b
Use input with a narrow range because the filter only applies when the
frames are similar.
Run CompareReferenceRandom more times. Especially before narrowing the
input range, the filter frequently did not apply.
Change-Id: Ie249bedf6d0d33dfa5884611cb1835788e418b38
this test fails with the configuration similar to the assembly prior to:
d52cb5972 quantize: copy ssse3 optimizations to intrinsics
BUG=webm:1458
Change-Id: Idc5c0b84c0598259fc49609a9f0756de531d3baf
With skip block the neon is about twice as fast as C.
The neon has no shortcut for coeff < zbin so it always takes the
same amount of time. Even if the C can take the shortcut, it is over
twice as fast in neon. If it can't, that gap increases to over 10x.
BUG=webm:1426
Change-Id: I400722146c1b5a5f6289f67d85fd642463d2bfc6
Fairly minor differences from sse2. pabsw and psignw are the big gains.
Also re-uses some values in eob calculation to avoid an extra pcmp.
Fixes test failures in HBD and OS X builds.
Allows using it in 32bit builds, where it is about 40% faster than sse2.
Substantially faster than the assembly for skip_block. 10-20% faster the
rest of the time.
Change-Id: If783bb3567e561e47667e10133b9c84414a334e2
Originally, for the purpose of keeping a fast first pass, the first-pass
stats between row_mt_mode = 0 and row_mt_mode = 1 are not bit exact, but
that difference is very small that doesn't cause a mismatch between the
final bitstreams. However, if the encoder changes, this minor difference
may cause a mismatch. Thus, this patch always forces the first pass to
be bit exact.
BUG=webm:1453
Change-Id: I2b67cf529dee81f660f9d9e7fe9a60ea3c7b12b8
With skip block or coeff < zbin it is about twice as fast as C.
If most coeff values are > zbin it is about 10-15x as fast as C.
BUG=webm:1426
Change-Id: I5d3c007b014a372d5ef0882b39bb48983b4131c7
This reverts commit 03f5e300d6.
This causes test failures under OSX:
SSSE3/VP9QuantizeTest.EOBCheck/0
SSSE3/VP9QuantizeTest.OperationCheck/0
Change-Id: I122732717ead1f7af5b04c529a6948e382e5e59b
Avoid unsigned overflow warning:
unsigned integer overflow: 19974 - 32703 cannot be represented in type
'unsigned int'
Change-Id: Ifebee014342e4c6f3b53306c0cad6ae0b465ac12
Officially the quant structures are 8 elements, with one dc element and
7 repeated ac elements. The low bit depth optimizations take advantage
of this to fill the xmm registers. The high bit depth version manually
duplicates the values.
If all the optimizations were unified, the structure sizes could be
greatly reduced.
Change-Id: Ibd7a0337a7832ce2a1a05ee433c310077e1059ae
Use only valid values for quantize inputs. These were determined by
looping over vp9_init_quantizer and looking for max and min values.
This allows extending the test to the low bit depth functions which were
not designed to handle all possible inputs but only valid inputs.
Change-Id: I94e1d8863a49ac227845b65c6b50130e10e6319e
Although the low bitdepth functions are identical (excepting the need
for larger intermediate values) they do not pass these tests. This
improves the error output to aid debugging.
Simplify buffer usage with Buffer and removing unnecessarily aligned
variables.
eob is a single element and never written using aligned instructions.
BUG=webm:1426
Change-Id: Ic95789a135cf1e8a3846d85270f2b818f6ec7e35
* changes:
sad neon: avg for 64x[32,64]
sad neon: macroize 64xN definitions
sad neon: avg for 32x[16,32,64]
sad neon: macroize 32xN definitions
sad neon: avg for 16x[8,16,32]
sad neon: macroize 16xN definitions
the file was empty after the struct removal. the only remaining use was
within vp9_dx_iface, but the wrapper became unnecessary after the
removal of frame_parallel_decode.
BUG=webm:1395
Change-Id: I515ab585d701e77d388d12b2802d844c424f9bcd
testing::Range does not include the end parameter in the set of values.
also adjust the start to 2 as the single threaded case is already
covered in another instantiation
Change-Id: Iae3bf3ed4363dd434eccfa5ad4e3c5e553fbee60
Add a method to acm_random.h to generate ranges of values
Add a way to call that method to buffer.h
Adjust dct_[partial_]test.cc to use it.
Change-Id: I8c23ae9d27612c28f050b0e44c41cb4ad2494086
VPX_CODEC_USE_FRAME_THREADING was made a no-op in:
01d23109a vp9: make VPX_CODEC_USE_FRAME_THREADING a no-op
and the tests in this file have been disabled since:
6ab0870d4 disable VP9MultiThreadedFrameParallel tests
BUG=webm:1395
Change-Id: I2c7a250acb65cf9522cf8a7bb724bb92070e41c6
this was made a no-op in:
01d23109a vp9: make VPX_CODEC_USE_FRAME_THREADING a no-op
and the test hitting this branch has been disabled since:
6ab0870d4 disable VP9MultiThreadedFrameParallel tests
rename the test to VP9MultiThreaded to exercise the tile-based threading
BUG=webm:1395
Change-Id: I35564a75eb5a7d7f7ccb923133b1b07295201f4c
Always return an int32_t. Since it needs to be moved to a register for
shifting, this doesn't really penalize the smaller transforms.
The values could potentially be summed and shifted in place.
BUG=webm:1424
Change-Id: Id5beb35d79c7574ebd99285fc4182788cf2bb972
For the 8x8_1, the highbd output fit nicely in the existing function. 12
bit input will overflow this implementation of 16x16_1.
BUG=webm:1424
Change-Id: I2945fe5478b18f996f1a5de80110fa30f3f4e7ec
The function was originally written with HBD in mind. Enable it and
configure the tests.
BUG=webm:1424
Change-Id: I78a2eba8d4d9d59db98a344ba0840d4a60ebe9a1
* changes:
sad neon: rewrite 64x64 and add 64x32
sad neon: rewrite 32x32, add 32x16 and 32x64
sad neon: rewrite 16x8, 16x16, add 16x32
sad neon: rewrite 8x8 and 8x16
sad neon: rewrite 4x4 and add 4x8
Test the _1 variant of the fdct, which simply sums the block and applies
a modifying shift based on the block size.
BUG=webm:1424
Change-Id: Ic80d6008abba0c596b575fa0484d5b5855321468
this normalizes these tests with the regular variance ones both in
implementation and test list output
Change-Id: I387aea81456f94b8223b8fb2a28cab94bc1aa9d5
this is consistent with other threaded tests and ensures gtest_filters
meant to operate on these pick them up
Change-Id: I99ce53720553a22c4b9905a2882273c2be2c031b
vpx_idct32x32_1024_add_ssse3() is actually a sse2 function and faster
than vpx_idct32x32_1024_add_sse2(). Replace the slow one. All are
code relocations, no new code.
Change-Id: I5dac0e98cc411a4ce05660406921118986638d19
'in' is used for the reference fdct. 'coeff' is input to the idct being
tested and 'dst[16]' is output
Fixes a segfault on unaligned memory access on x86.
Change-Id: I3691b1380ed49986897dd89a63ce63a80a0e0962
this was deleted in:
98967645a Remove vpx_idct8x8_64_add_ssse3()
but this was merged in:
9e03eedf6 Merge changes Ib26dd515,Ie60dabc3
after:
a92991133 Merge "dct tests: run all possible sizes in one test"
which added a new reference
Change-Id: I8da4a6c80d27b237a378ff15eead1daab89e7e25
Modify fdct4x4_test.cc to support all size combinations. This does not
add any new tests and in fact fails a few. There were minimal changes
made to the tests so it's not entirely surprising that some of the
larger 12 bit transforms are failing since it was initially only used
for 4x4.
In follow up patches the tests in fdct8x8_test.cc, dct16x16_test.cc and
dct32x32_test.cc will be evaluated and moved to dct_test.cc.
BUG=webm:1424
Change-Id: I72a23430f457d7fae8c91e706adc0e77c25abc8f
It's almost identical with vpx_idct8x8_64_add_sse2(), except little
difference in instructions order.
Change-Id: Ie60dabc35eaa6ebae7c755e6cff00a710aad284f
the check for error correction being disabled was overriding the data
length checks. this avoids returning incorrect information (width /
height) for the decoded frame which could result in inconsistent sizes
returned in to an application causing it to read beyond the bounds of
the frame allocation.
BUG=webm:1443
BUG=b/62458770
Change-Id: I063459674e01b57c0990cb29372e0eb9a1fbf342
Roughly 2x speedup. Since the only change for HBD is to store(), the
improvement appears to hold there as well.
BUG=webm:1424
Change-Id: I15b813d50deb2e47b49a6b0705945de748e83c19
Split vp8/vp9 implementations on yv12_copy_frame_c.
Remove high-bitdepth codes from vp8_yv12_extend_frame_borders_c.
Clean up vp8 codes usage in vp9.
BUG=webm:1435
Change-Id: Ic68e79e9d71e1b20ddfc451fb8dcf2447861236d
Continue processing sets of 16 values. Plenty of improvement for 4x8
(doubles the speed) but only about 30% for 4x4.
BUG=webm:1422
Change-Id: Ib8dd96f75d474f0348800271d11e58356b620905
Add PartialIDctTest::PrintDiff() to help debugging.
In RunQuantCheck, try all combinations of +/-mask_ input for 4x4 idct.
Update PartialIDctTest::InitInput().
Change-Id: I13fd163954a4c1a3a6cfeb5e4a4d3d0e7ff901f4
For SVC 1 pass non-rd pickmode, the interpolation filter for the
upsampling of the golden (spatial) reference was not being explicitly
set and instead was takin gwhatever value was set in the previous
mode/block (which would be either EIGHTTAP or EIGHTAP_SMOOTH).
Fix it to the default EIGHTTAP for now, to be updated/selected
adaptively in a later change.
Minor adjustmemt to rate targeting thresholds in datarate unittests.
Change-Id: I52085048674072c6cfb7163e11e9a2658d773826
Makes more sense to call the corresponding partial idct C function
instead of the full idct C function as the reference.
Change-Id: Ibb7681dd063edd6307ba582c10c26c4c6a4b78c6
An intended behavior change disabling exhaustive searches in speed
feature causes VP9/DatarateTestVP9LargeDenoiser.4threads test failure.
Change the threshold to make it pass.
BUG=webm:1429
Change-Id: Ibcbe2314c6b2525799894f5d7204fc8eb4ec2a1e
Add support for everything except block sizes of 4.
Performance is better but numbers will improve again when the variance
optimizations land.
BUG=webm:1423
Change-Id: I92eb4312b20be423fa2fe6fdb18167a604ff4d80
Some of the mixed sizes were missing. They can be implemented trivially
using the existing helper function.
When comparing the previous 16x8 and 8x16 implementations, the helper
function is about 10% faster than the 16x8 version. The 8x16 is very
close, but the existing version appears to be faster.
BUG=webm:1422
Change-Id: Ib0e856083c1893e1bd399373c5fbcd6271a7f004
Approximates division using multiply and shift.
Speeds up both sizes (8x8 and 16x16) by 30 times.
Fix the call sites to use the RTCD function.
Delete sse2 and mips implementation. They were based on a previous
implementation of the filter. It was changed in Dec 2015:
ece4fd5d22
BUG=webm:1378
Change-Id: I0818e767a802966520b5c6e7999584ad13159276
This patch followed allow_exhaustive_searches feature modification and
continued to modify the encoder to achieve the determinism in the row
based multi-threaded encoding. While row-mt = 1 and using multiple
threads, the adaptive feature in encoder was disabled, which gave
BDRate gain(at speed 1, -0.6% ~ -0.7%; at speed 2, -0.46% ~ -0.59%),
but some encoder speed losses(7% ~ 10% at speed 1 and 3% ~ 6% at
speed 2). These speed losses were acceptable considering the speed
gains obtained from row-mt.
Change-Id: I60d87a25346ebc487a864b57d559f560b7e398bb
Re-enable the SVC tests, wrap the non-zero expectation
in GetMismatchFrames around #if CONFIG_VP9_DECODER.
Change-Id: I0e8a2d78b868c32f18fe597540f397d3a1b303b5
