| Age | Commit message (Collapse) | Author |
|---|
|
Signed-off-by: James Almer <jamrial@gmail.com>
|
|
Will be used for regression tests.
|
|
Since n equals either 25 or 9, the dav1d_sgr_one_by_x table can be
replaced with a ternary operation.
|
|
Test case: clusterfuzz-testcase-minimized-dav1d_fuzzer-5733716937146368,
credits to oss-fuzz.
|
|
This rewrites msac.c to the point of there being no libaom project
code left, hence changing the license of the file to the dav1d
project's license.
The rewrite much simplifies and optimizes entropy decoding.
Some encoder specific code also remained, such as tell_offs, to
tell the fractional amount of bits left, which the decoder does
not need.
ctx_refill is much simpler and has a tighter loop with less
instructions, which on some CPUs can actually be ran in one cycle.
The old mechanism which checked to see if the buffer reached the
end to disable calling ctx_refill is gone, as all it saved was
a mostly well predicted branch.
The optimizations regarading this function enabled us to use
an ec_win of 64 bits whilst improving performance. This was not
possible with the old needlessly robust system.
Some msac-specific API changes were made - msac_decode_bool now
takes a scaled value directly rather than doing scaling itself.
This saves a shift in most use cases as the function is mainly
used to read equiprobable bools rather than ones with specific
probabilities.
There's still room for optimizations, mainly in that update_cdf
could be SIMD'd. This commit prepares for some of them by
moving the init function at the bottom of the file.
Overall decoder speedup seems to be around 3%-5%, specific on
bitrate and encoder as expected.
|
|
Fix #102
|
|
Avoid out of memory (2Gb) or time out (25s) issues in oss-fuzz.
|
|
|
|
This simplifies running specifically this test by "meson test checkasm".
|
|
|
|
This would result in weird output, as checkasm outputs several lines.
|
|
|
|
Fixes a regression since d8996b181d2013abd01a2e9f5dceae5e09b3afda, where
the relevant fields started being written to and read from the wrong struct.
|
|
This is identical to what libaom does, and fixes #119. After
gather_left/top_partition_prob(), probabilities *can* overlap,
and in extreme cases the output CDF[0] can therefore be 0, which
will cause an assert failure when used with msac_decode_bool().
|
|
Fix #120.
|
|
Fix #115. Fix 'assert(seg_id < 8)' in
clusterfuzz-testcase-minimized-dav1d_fuzzer-5740590025670656 due to
decoding error in the primary reference picture. Credits to oss-fuzz.
|
|
cdef_filter_4x4_8bpc_c: 2273.6
cdef_filter_4x4_8bpc_avx2: 113.6
Decoding time reduces to 15.51s for first 1000 frames of chimera 1080p,
from 23.1 before cdef_filter SIMD or 17.86 with only 8x8 cdef_filter
SIMD.
|
|
cdef_filter_8x8_8bpc_c: 7913.0
cdef_filter_8x8_8bpc_avx2: 309.9
First 1000 frames of Chimera 1080p:
before: 0m23.100s
after: 0m17.863s
|
|
|
|
|
|
Fix a theoretical bug found by the clang static analyzer:
In theory num_muxers could be smaller than res, never evaluating the
loop (as the condition would be false) but not satisfying the
i == num_muxers case, so proceeding and dereferencing the never
initialized impl pointer.
Fixing this is simple: Changed the num_muxers and index variables
to unsigned so num_muxers can never be smaller than i.
|
|
Fixes #123.
|
|
oss-fuzz's FuzzingEngine requires libc++.
|
|
This prevents warnings on some old compilers that don't like structs that
start with an array being initialized with { 0 }.
|
|
Fixes #112.
|
|
Fixes #111.
|
|
Prevent decoding a frame with inconsistent sequence and frame headers.
Fix #124, #125. Fix negative size param in pixel_copy due to inconsistent
sb128 state between frame header (parsed with sb128 == 0) and sequence
header and frame decoding with sb128 == 1. Fix
clusterfuzz-testcase-minimized-dav1d_fuzzer-5707479116152832. Credits to
oss-fuzz.
|
|
Fixes #116.
|
|
cdef_dir_8bpc_c: 629.3
cdef_dir_8bpc_avx2: 82.4
First 1000 frames of Chimera 1080p:
before: 0m23.084s
after: 0m21.860s
|
|
|
|
|
|
Fix 122.
|
|
|
|
This is similar to --noblit of libaom
|
|
oss-fuzz builds the fuzzing binary out of meson. Local (quoted) includes
should be preferred for headers located in the same directory.
|
|
Instead of bit reverting the register subtract 24 from clz. This avoids
two empty filler values in the jump table with the same instruction
count.
|
|
Missed in 46e2a2d0cc. Arm asm will be hard to template so move them to
the plain source list.
Fix #96.
|
|
Fix #96
|
|
Add duplicate prototypes to ref_mvs.c for now, since ref_mvs.h can't
be included.
Signed-off-by: James Almer <jamrial@gmail.com>
|
|
For arm/arm64, there's no need to align any buffer to 32 bytes
as the assembly doesn't need it and doesn't benefit from it.
This would be much more elegant if defined like this:
#define MAX_ALIGN 16
#define ALIGN(align) __attribute__((aligned(MIN(align, MAX_ALIGN))))
This works for GCC and Clang, but the MSVC alignment __declspec
needs a literal alignment value, it can't handle an expression.
|
|
Fixes an undefined left shift in
clusterfuzz-testcase-minimized-dav1d_fuzzer-5717082881130496. Credits to
oss-fuzz. Fixes #110
|
|
Due to the stripe offset of 8 rows the last stripe can have a height of
128 + 8 rows if the resolution is a multiple of 128. Fixes a
stack-buffer-overflow in
clusterfuzz-testcase-minimized-dav1d_fuzzer-5731418676658176. Credits to
oss-fuzz.
|
|
Fix issue #70
|
|
These went out of sync in 227b7eae7a.
|
|
checkasm --bench numbers from a Snapdragon 835:
nop: 23.0
avg_w4_8bpc_c: 385.0
avg_w4_8bpc_neon: 34.0
avg_w8_8bpc_c: 590.5
avg_w8_8bpc_neon: 65.5
avg_w16_8bpc_c: 1304.4
avg_w16_8bpc_neon: 161.3
avg_w32_8bpc_c: 4098.4
avg_w32_8bpc_neon: 589.2
avg_w64_8bpc_c: 8405.0
avg_w64_8bpc_neon: 1367.1
avg_w128_8bpc_c: 19667.9
avg_w128_8bpc_neon: 3409.0
w_avg_w4_8bpc_c: 453.8
w_avg_w4_8bpc_neon: 50.0
w_avg_w8_8bpc_c: 749.0
w_avg_w8_8bpc_neon: 105.7
w_avg_w16_8bpc_c: 1851.2
w_avg_w16_8bpc_neon: 283.7
w_avg_w32_8bpc_c: 5991.5
w_avg_w32_8bpc_neon: 1080.9
w_avg_w64_8bpc_c: 12763.5
w_avg_w64_8bpc_neon: 2544.4
w_avg_w128_8bpc_c: 30311.3
w_avg_w128_8bpc_neon: 6350.5
mask_w4_8bpc_c: 492.9
mask_w4_8bpc_neon: 57.7
mask_w8_8bpc_c: 1108.5
mask_w8_8bpc_neon: 123.0
mask_w16_8bpc_c: 2880.3
mask_w16_8bpc_neon: 349.2
mask_w32_8bpc_c: 8996.4
mask_w32_8bpc_neon: 1368.1
mask_w64_8bpc_c: 19570.3
mask_w64_8bpc_neon: 3263.5
mask_w128_8bpc_c: 46757.4
mask_w128_8bpc_neon: 8743.1
|
|
Fixes #100.
|
|
|
|
Fixes #104.
|
|
|
|
Fixes #93.
|
