While this function on its own passes all of fate-hevc, there's
indications that the function might need to handle widths that
aren't a multiple of 8 (noted in commit
f63f9be37c, which later was
reverted).
Signed-off-by: Martin Storsjö <martin@martin.st>
Passing an uninitialized variable as argument to a function is
undefined behaviour (UB). The compiler can assume that UB does not
happen.
Hence, the compiler can assume that the variables are never
uninitialized when passed as argument, which means that the codepaths
that initializes them must be taken.
In ff_seek_frame_binary, this means that the compiler can assume
that the codepaths that initialize pos_min and pos_max are taken,
which means that the conditions "if (sti->index_entries)" and
"if (index >= 0)" can be optimized out.
Current Clang git versions (upcoming Clang 14) enabled an optimization
that does this, which broke the current version of this function
(which intentionally left the variables uninitialized, but silencing
warnings about being uninitialized). See [1] for discussion on
the matter.
[1] https://reviews.llvm.org/D105169#3069555
Signed-off-by: Martin Storsjö <martin@martin.st>
Don't use the loaded registers directly, avoiding stalls on in
order cores. Use vrhadd.u8 with q registers where easily possible.
Signed-off-by: Martin Storsjö <martin@martin.st>
The refactoring in 844890b1bc caused
fate-source to point out that this could be av_clip_uintp2 (or
rather av_clip_uint8).
Signed-off-by: Martin Storsjö <martin@martin.st>
If no --cpu= option was passed to configure, we detect what the
compiler defaults to. This detected value was then fed back to the
rest of the configure logic, as if it was an explicit choice.
This breaks on Ubuntu 21.10 with GCC 11.1.
Since GCC 8, it's possible to add configure extra features via the
-march option, like e.g. -march=armv7-a+neon. If the -mfpu= option
is configured to default to 'auto', the fpu setting gets taken
from the -march option.
GCC 11.1 in Ubuntu seems to be configured to use -mfpu=auto. This
has the effect of breaking any compilation command that specifies
-march=armv7-a, because the driver implicitly also adds -mfloat-abi=hard,
and that combination results in this error:
cc1: error: ‘-mfloat-abi=hard’: selected processor lacks an FPU
One can compile successfully by passing e.g. -march=armv7-a+fp.
Therefore, restructure configure. If no specific preference was set
(and the 'cpu' configure variable was set as the output of
probe_arm_arch), the value we tried to set via -march= was the same
value that we just tried to detect as the compiler default.
So instead, just try to detect what the compiler defaults to, with
to allow setting other configure settings (such as 'fast_unaligned'),
but don't try to spell out the compiler's default via the -march flag.
Signed-off-by: Martin Storsjö <martin@martin.st>
This unbreaks the fate-checkasm-hevc_pel test on arm targets.
The assembly assumed that the width passed to the DSP functions is
a multiple of 8, while the checkasm test used other widths too.
This wasn't noticed before, because the hevc_pel checkasm tests
(that were added in 9c513edb79 in
January) weren't run as part of fate until in
b492cacffd in August.
As this hasn't been an issue in practice with actual full decoding
tests, it seems like the actual decoder doesn't call these functions
with such widths. Therefore, we could alternatively fix the test
to only test things that the real decoder does, and this modification
could be reverted.
Signed-off-by: Martin Storsjö <martin@martin.st>
If autoflushing on a new packet (e.g. due to the frag_every_frame
flag being set), there's no samples to be written in the new fragment,
so we can't overwrite the track duration in order to make it line
up with the next packet to be written.
Signed-off-by: Martin Storsjö <martin@martin.st>
Instead use --preprocessor-arg; in binutils 2.36, the --preprocessor
flag was changed so that it no longer accepts a string containing
multiple arguments, but the whole --preprocessor argument is
treated as the path to the preprocessor executable (where the path
can contain spaces).
It's currently unclear whether this behaviour will stay or if it
is going to be reverted in the future, see discussion at [1]. Just
to be safe, avoid using the --preprocessor argument. Don't redeclare
the full preprocessing command, but just add the $(CC_DEPFLAGS) options.
Based on a patch by Kyle Schwartz.
[1] https://sourceware.org/bugzilla/show_bug.cgi?id=27594
Signed-off-by: Martin Storsjö <martin@martin.st>
This is marginally slower, but correct for all input values.
The previous implementation failed with certain input seeds, e.g.
"checkasm --test=hevc_idct 98".
Signed-off-by: Martin Storsjö <martin@martin.st>
Move the loop counter decrement further from the branch instruction,
this hides the latency of the decrement.
In loops that first load, then store (the horizontal prediction cases),
do the decrement after the load (where the next instruction would
stall a bit anyway, waiting for the result of the load).
In loops that store twice using the same destination register,
also do the decrement between the two stores (as the second store
would need to wait for the updated destination register from the
first instruction).
In loops that store twice to two different destination registers,
do the decrement before both stores, to do it as soon before the
branch as possible.
This gives minor (1-2 cycle) speedups in most cases (modulo measurement
noise), but the horizontal prediction functions get a rather notable
speedup on the Cortex A53.
Before: Cortex A53 A72 A73
pred8x8_dc_8_neon: 60.7 46.2 39.2
pred8x8_dc_128_8_neon: 30.7 18.0 14.0
pred8x8_horizontal_8_neon: 42.2 29.2 18.5
pred8x8_left_dc_8_neon: 52.7 36.2 32.2
pred8x8_mad_cow_dc_0l0_8_neon: 48.2 27.7 25.7
pred8x8_mad_cow_dc_0lt_8_neon: 52.5 33.2 34.7
pred8x8_mad_cow_dc_l0t_8_neon: 52.5 31.7 33.2
pred8x8_mad_cow_dc_l00_8_neon: 43.2 27.0 25.5
pred8x8_plane_8_neon: 112.2 86.2 88.2
pred8x8_top_dc_8_neon: 40.7 23.0 21.2
pred8x8_vertical_8_neon: 27.2 15.5 14.0
pred16x16_dc_8_neon: 91.0 73.2 70.5
pred16x16_dc_128_8_neon: 43.0 34.7 30.7
pred16x16_horizontal_8_neon: 86.0 49.7 44.7
pred16x16_left_dc_8_neon: 87.0 67.2 67.5
pred16x16_plane_8_neon: 236.0 175.7 173.0
pred16x16_top_dc_8_neon: 53.2 39.0 41.7
pred16x16_vertical_8_neon: 41.7 29.7 31.0
After:
pred8x8_dc_8_neon: 59.0 46.7 42.5
pred8x8_dc_128_8_neon: 28.2 18.0 14.0
pred8x8_horizontal_8_neon: 34.2 29.2 18.5
pred8x8_left_dc_8_neon: 51.0 38.2 32.7
pred8x8_mad_cow_dc_0l0_8_neon: 46.7 28.2 26.2
pred8x8_mad_cow_dc_0lt_8_neon: 55.2 33.7 37.5
pred8x8_mad_cow_dc_l0t_8_neon: 51.2 31.7 37.2
pred8x8_mad_cow_dc_l00_8_neon: 41.7 27.5 26.0
pred8x8_plane_8_neon: 111.5 86.5 89.5
pred8x8_top_dc_8_neon: 39.0 23.2 21.0
pred8x8_vertical_8_neon: 27.2 16.0 14.0
pred16x16_dc_8_neon: 85.0 70.2 70.5
pred16x16_dc_128_8_neon: 42.0 30.0 30.7
pred16x16_horizontal_8_neon: 66.5 49.5 42.5
pred16x16_left_dc_8_neon: 81.0 66.5 67.5
pred16x16_plane_8_neon: 235.0 175.7 173.0
pred16x16_top_dc_8_neon: 52.0 39.0 41.7
pred16x16_vertical_8_neon: 40.2 33.2 31.0
Despite this, a number of these functions still are slower than
what e.g. GCC 7 generates - this shows the relative speedup of the
neon codepaths over the compiler generated ones:
Cortex A53 A72 A73
pred8x8_dc_8_neon: 0.86 0.65 1.04
pred8x8_dc_128_8_neon: 0.59 0.44 0.62
pred8x8_horizontal_8_neon: 1.51 0.58 1.30
pred8x8_left_dc_8_neon: 0.72 0.56 0.89
pred8x8_mad_cow_dc_0l0_8_neon: 0.93 0.93 1.37
pred8x8_mad_cow_dc_0lt_8_neon: 1.37 1.41 1.68
pred8x8_mad_cow_dc_l0t_8_neon: 1.21 1.17 1.32
pred8x8_mad_cow_dc_l00_8_neon: 1.24 1.19 1.60
pred8x8_plane_8_neon: 3.36 3.58 3.76
pred8x8_top_dc_8_neon: 0.97 0.99 1.43
pred8x8_vertical_8_neon: 0.86 0.78 1.18
pred16x16_dc_8_neon: 1.20 1.06 1.49
pred16x16_dc_128_8_neon: 0.83 0.95 0.99
pred16x16_horizontal_8_neon: 1.78 0.96 1.59
pred16x16_left_dc_8_neon: 1.06 0.96 1.32
pred16x16_plane_8_neon: 5.78 6.49 7.19
pred16x16_top_dc_8_neon: 1.48 1.53 1.94
pred16x16_vertical_8_neon: 1.39 1.34 1.98
In particular, on Cortex A72, many of these functions are slower
than the compiler generated code, while they're more beneficial on
e.g. the Cortex A73.
Signed-off-by: Martin Storsjö <martin@martin.st>
This fixes building with MSVC after
a2a38b1606.
Remove the stray semicolon, and add casts for the input argument
(which is an intptr_t*) to the right type (PVOID volatile *).
Signed-off-by: Martin Storsjö <martin@martin.st>
Print every error in the stack, if more than one, and don't print
bogus errors if there's none logged within OpenSSL.
Retain the underlying IO error code, print an error message out of
it, and pass the error code on to the caller.
Signed-off-by: Martin Storsjö <martin@martin.st>
If b-frames were enabled implicitly (if max_b_frames wasn't set by
the caller at all, since a0949d0bcb),
we wouldn't offset dts at all, producing invalid pts/dts combinations
(causing loud warnings by ffmpeg, or muxer errors if passed without
an extra cleanup pass).
Instead use frameIntervalP for offsetting, which should always be
accurate.
Signed-off-by: Timo Rothenpieler <timo@rothenpieler.org>
This is much less precise than the cycle counter register, but
the cycle counter register is not available on apple platforms
(and on linux, it requires a kernel module for allowing user mode
access).
Signed-off-by: Martin Storsjö <martin@martin.st>
If the edit lists remove parts of the output timeline, or add a
delay to it, this should be included in the mvhd/tkhd/mdhd durations,
which should correspond to the edit lists.
For tracks starting with pts < 0, the edit list trims out the segment
before pts=0. For tracks starting with pts > 0, a delay element is
added in the edit list, delaying the start of the track data.
In both cases, the practical effect is that the post-edit output
is as if the track had started with pts = 0. Thus calculate the range
from pts=0 to end_pts, for the purposes of mvhd/tkhd/mdhd, unless
edit lists explicitly are disabled.
mov_write_edts_tag needs to operate on the actual pts duration of
the track samples, not the duration that already takes the edit
list effect into account.
Signed-off-by: Martin Storsjö <martin@martin.st>
When the check was added (in 3668701f96, in 2015), some IO
functions returned 0 on EOF (in particular, the TCP protocol
did, but the TLS protocol returned AVERROR_EOF). Since
0e1f771d22 in 2017, the TCP protocol also returns AVERROR_EOF
instead of 0, making the check for premature end never have the
intended effect.
Signed-off-by: Martin Storsjö <martin@martin.st>
The previous threshold, 4 KB, maybe was reasonable when it was set
(in 2010), but in today's settings and with typical network speeds
and data sizes, it's pretty small. 32 KB probably is a more reasonable
default now, regardless of input.
This changes the test references for two seek tests.
When using the normal seek function, which boils down to the lseek(2)
function, a seek to an out of bounds position doesn't return an error,
but that condition is only reported when doing the subsequent read
(which returns EOF). When doing more seeks by fast forwarding, the
fact that the seeked to destination is out of bounds is noticed and
reported sooner in these cases.
Signed-off-by: Martin Storsjö <martin@martin.st>
In listen mode with UDP transport, once the sender has sent
the TEARDOWN and closed the connection, poll will indicate that
one can read from the connection (indicating that the socket has
reached EOF and should be closed by the receiver as well). In this
case, parse_rtsp_message won't try to parse the command (because
it's no longer in state STREAMING), but previously just returned
zero.
Prior to f6161fccf8, this caused
udp_read_packet to return zero, which is treated as EOF by
read_packet. But after that commit, udp_read_packet would continue
if parse_rtsp_message didn't return an explicit error code.
To keep the original behaviour from before that commit, more
explicitly return an error in parse_rtsp_message when in the wrong
state.
Fixes: #8840
Signed-off-by: Martin Storsjö <martin@martin.st>
Previously, the hls-fmp4 and hls-fmp4_ac3 tests used the same file
names for init and segment files, which occasionally could cause
corruption and failed tests, if the input files for both tests are
generated in parallel, as they could overwrite each other.
This happened to work some of the time, as the fmp4_ac3 test actually
only checked the init segment file (which the fmp4 test case never
wrote, due to using the incorrect hls_segment_type option) and the
fmp4 test case always regenerated the input files due to mismatched
target and file names.
Signed-off-by: Martin Storsjö <martin@martin.st>
Previously, with the file name not matching the target, the files
were regenerated every time fate is rerun - contrary to the other
test targets in the same file. (While regenerating it every time
might be desireable, as that's what the test is about, the file
at least has a dependency on the ffmpeg executable, making them
regenerated every time the executable is updated - and this change
at least makes it consistent with the rest.)
Signed-off-by: Martin Storsjö <martin@martin.st>
Including codecapi.h and uuids.h in UWP mode doesn't define all defines
properly, ending up with constructs that MSVC silently tolerates, but
that clang errors out on, like this:
DEFINE_GUIDEX(CODECAPI_AVEncCommonFormatConstraint);
Just avoid including codecapi.h completely and hardcode the last few
enum values we use from there. We already use local versions of most
enums from there, due to older mingw-w64 headers being incomplete.
Signed-off-by: Martin Storsjö <martin@martin.st>
This might have been used originally for the decoder parts of
the MediaFoundation wrapper, which aren't merged yet.
Signed-off-by: Martin Storsjö <martin@martin.st>
Just including icodecapi.h + codecapi.h apparently misses some bits,
that MSVC doesn't complain about, but Clang does. Thus add a missing
header to the custom hacky include path (required for keeping
ICodecAPI visible when building in UWP mode, as it is intended to).
The issue in Windows SDK headers about ICodecAPI availability in
UWP/app mode has been reported upstream at
https://developercommunity.visualstudio.com/content/problem/1037125/icodecapi-not-visible-when-compiling-in-uwp-mode.html,
and later also filed in Feedback Hub (which can't be linked to
externally).
Signed-off-by: Martin Storsjö <martin@martin.st>
The framerate field is the one users are supposed to set, but not
all users might be setting it, so it might be good to fall back
time_base in that case.
Signed-off-by: Martin Storsjö <martin@martin.st>
Move the copying of the frame to vos_data further up in the function,
so that when writing the actual frame data for the first frame, it's
clear that the stream really is in annex b format, for the cases where
we create extradata from the first frame.
Alternatively - we could invert the checks for bitstream format. If
extradata is missing, we can't pretend that the bitstream is in
mp4 form, because we can't even know the NAL unit length prefix size
in that case.
Also avoid creating extradata for AVC intra. If the track tag is
an AVC intra tag, don't copy the frame into vos_data - this matches
other existing cases of how vos_data and TAG_IS_AVCI interact in
other places.
Signed-off-by: Martin Storsjö <martin@martin.st>
Sticking a full frame in the extradata works, as the code for writing
the avcC/hvcC extracts the relevant parameter set NAL units - provided
that they actually exist in the frame.
Some encoders don't provide split out extradata directly on init (or
at all). In particular, the MediaFoundation encoder wrapper doesn't
always (depending on the actual encoder device) - this is the case for
Qualcomm's HEVC encoder on SD835, and also on some QSV H264 encoders).
This only works for cases where the moov hasn't already been written
(e.g. when not writing fragmented mp4 with empty_moov, unless using
the delay_moov option).
Signed-off-by: Martin Storsjö <martin@martin.st>
Also fill x8-x17 with garbage before calling the function.
Figure out the number of stack parameters and make sure that the
value on the stack after those is untouched.
Signed-off-by: Martin Storsjö <martin@martin.st>
Figure out the number of stack parameters and make sure that the
value on the stack after those is untouched.
Signed-off-by: Martin Storsjö <martin@martin.st>
We should just use a normal bl here, and the linker will add the 'x'
bit if necessary.
This fixes calling the checkasm_fail_func on windows, where the
code is built in thumb mode (and the linker doesn't clear the 'x'
bit in the blx instruction).
Signed-off-by: Martin Storsjö <martin@martin.st>
On windows and darwin (and modern android), the x18 register is reserved
and shouldn't be modified by user code, while it is freely available on
linux. Strictly avoid it, to keep the assembly code portable.
This would have helped catch the issue fixed in 872790b1f9
immediately.
Signed-off-by: Martin Storsjö <martin@martin.st>
Use this in vf_spp.c, where the get_pixels operation is done on
unaligned source addresses.
Hook up the x86 (mmx and sse) versions of get_pixels to this
function pointer, as those implementations seem to support unaligned
use.
This fixes fate-filter-spp on armv7.
Signed-off-by: Martin Storsjö <martin@martin.st>
Apparently the changes from 3c9185bf3a
aren't enough; even with that in place, I got errors like this
when trying to build for iOS:
src/libavdevice/avfoundation.m:135:5: error:
'AVCaptureDeviceTransportControlsPlaybackMode' is unavailable: not
available on iOS
AVCaptureDeviceTransportControlsPlaybackMode observed_mode;
^
This fixes tests on 32 bit x86 mingw with clang, which uses x87
fpu by default.
In this setup, while the get_expected function is declared to
return float, the compiler is (especially given the optimization
flags set) free to keep the intermediate values (in this case,
the return value from the inlined function) in higher precision.
This results in the situation where 7.28 (which actually, as
a float, ends up as 7.2800002098), multiplied by 100, is
728.000000 when really forced into a 32 bit float, but 728.000021
when kept with higher intermediate precision.
For the multiplication case, a more suitable epsilon would e.g.
be 2*FLT_EPSILON*fabs(expected_output), but just increase the
current hardcoded threshold for now.
Signed-off-by: Martin Storsjö <martin@martin.st>
The x18 is a reserved platform register on Darwin and Windows.
x8/w8 seems to be unused in this function though (and same about
x10 and x14), so there's really no reason to use x18 here - just change
the uses of x18/w18 into x8/w8 instead without any further rewrites.
Signed-off-by: Martin Storsjö <martin@martin.st>
Previously, it was always left in the automatic mode, if the option
was set to the only special (negative) value. Now there's two separate
special values for this option, -1 for automatic (metadata based)
and -2 for explicitly disabled.
Signed-off-by: Martin Storsjö <martin@martin.st>
It was disabled by default in 2dbd35b00c
as it added delay, but now we compensate for the delay properly
by offsetting timestamps.
Signed-off-by: Martin Storsjö <martin@martin.st>
This fixes the tests filter-refcmp-ssim-yuv and filter-refcmp-ssim-rgb
on i386 after breaking in fcc0424c93.
Signed-off-by: Martin Storsjö <martin@martin.st>
When clang works in MSVC mode, it does have the _byteswap_ulong
builtin, but one has to include stdlib.h before using it.
Signed-off-by: Martin Storsjö <martin@martin.st>
SetConsoleTextAttribute used to be unavailable for Windows Store apps,
but is available to them now. But GetStdHandle still is unavailable,
thus make sure to check for both functions before using code that
requires both.
Signed-off-by: Martin Storsjö <martin@martin.st>
Keep all the existing data fields as they are (there's lots and
lots of nontrivial calculation and heuristics based on them in
their current form), but derive the duration as the difference
between the pts of the first packet to the maximum pts+duration
(not necessarily the last packet); use this duration in any box
where the actual presentation duration is supposed to be.
Fixes: 8420
Signed-off-by: Martin Storsjö <martin@martin.st>
After 06ec9c4746 we check for these
functions in configure (which will succeed in cygwin), but cmdutils.c
only includes windows.h if _WIN32 is defined (which it isn't in cygwin).
Retain the old intent from before 06ec9c4746,
that these functions only would be used when _WIN32 is defined, while
only using them if configure has agreed that they do exist.
Signed-off-by: Martin Storsjö <martin@martin.st>
The stereo_interpolate functions add h_step to the values h
BUF_SIZE times. Within the stereo_interpolate C functions, the
values h (h0-h3, h00-h13) are declared as local float variables,
but the compiler is free to keep them in a register with extra
precision.
If the accumulation is rounded to 32 bit float precision after
each step, the less significant bits of h_step end up ignored
and the sum can deviate, affecting the end result more than
the currently set EPS.
By clearing the log2(BUF_SIZE) lower bits of h_step, we make sure
that the accumulation shouldn't differ significantly, regardless
of any extra precision in the accmulating register/variable.
This fixes the aacpsdsp checkasm test when built with clang for
mingw/x86_32.
Signed-off-by: Martin Storsjö <martin@martin.st>
In these cases, we must pass the full path of the file to ffprobe
(as the current working dir on the remote system, e.g. when invoked
with "ssh remote ffprobe ..." isn't the wanted one).
The input filename passed to ffprobe is also included in the output,
which is part of the reference test data. Add a new option to
ffprobe to allow overriding what path is printed, to keep the
original relative path in the tests.
An alternative approach could be an option to allow requesting omitting
the file name from the dumped data, and updating the test references
accordingly.
Signed-off-by: Martin Storsjö <martin@martin.st>
When testing on a memory limited system, these tests consume a
significant amount of memory and can often fail if testing by running
multiple processes in parallel.
Signed-off-by: Martin Storsjö <martin@martin.st>
These functions aren't available when building for the restricted
UWP/WinRT/WinStore API subsets.
Normally when building in this mode, one is probably only building
the libraries, but being able to build ffmpeg.exe still is useful
(and a ffmpeg.exe targeting these API subsets still can be run
e.g. in wine, for testing).
Signed-off-by: Martin Storsjö <martin@martin.st>
As the values generated by av_bmg_get can be arbitrarily large
(only the stddev is specified), we can't use a fixed tolerance.
Calculate a dynamic tolerance (like in float_dsp from 38f966b222),
based on the individual steps of the calculation.
This fixes running this test with certain seeds, when built with
clang for mingw/x86_32.
Signed-off-by: Martin Storsjö <martin@martin.st>
This particular function is only required to return nonzero on
errors, but use the common AVERROR() pattern for consistency.
Signed-off-by: Martin Storsjö <martin@martin.st>
These dependencies are evaluted by make and must be expressed with
the paths as in the local filesystem.
Signed-off-by: Martin Storsjö <martin@martin.st>
The tremolo filter uses floating point internally, and uses
multiplication factors derived from sin(fmod()), neither of
which is bitexact for use with framecrc.
This fixes running this test when built with for mingw/x86_32
with clang.
In this case, a 1 ulp difference in the output from fmod() would
end up in an output from the filter that differs by 1 ulp, but
which makes the lrint() in swresample/audioconvert.c round in a
different direction.
Signed-off-by: Martin Storsjö <martin@martin.st>
As the values generated by av_bmg_get can be arbitrarily large
(only the stddev is specified), we can't use a fixed tolerance.
This matches what was done for test_vector_dmul_scalar in
38f966b222.
This fixes the float_dsp checkasm test for some seeds, when built
with clang for mingw/x86_32.
Signed-off-by: Martin Storsjö <martin@martin.st>
Use a macro to redirect calling code from the official name to the
ff_ prefixed one.
Detecting these functions in configure can be tricky (on mingw, they
are conditionally available depending on posix feature defines).
If configure didn't detect them, but they still are visible at
compile time (due to an unrelated header defining the posix feature
defines), providing the local fallback versions with a prefixed
name is safer.
Signed-off-by: Martin Storsjö <martin@martin.st>
These functions are available in time.h (conditional on posix thread
safe functions) on mingw.
Previously, these functions weren't detected by configure, and
libavutil/time_internal.h provided replacements, even if time.h
actually contained definitions of them.
These mingw inline functions are currently defined as
"extern __inline __attribute__((__gnu_inline__))". In this case,
redefining a new static inline version of the same function with the
same name is tolerated.
Signed-off-by: Martin Storsjö <martin@martin.st>
Mingw headers have got header inline implementations of localtime_r
and gmtime_r, but only visible if certain posix thread safe functions
have been requested.
This is a preparatory step for improving the detection of those
functions.
Signed-off-by: Martin Storsjö <martin@martin.st>
Even if NEON would be disabled, the init functions should be built
as they are called as long as ARCH_AARCH64 is set.
These functions are part of a generic DSP subsytem, not tied directly
to one decoder. (They should be built if the vp7 decoder is enabled,
even if the vp8 decoder is disabled.)
Signed-off-by: Martin Storsjö <martin@martin.st>
(cherry picked from commit b4b27dce95)
This also partially fixes assembling with MS armasm64 (via
gas-preprocessor).
The movrel macro invocations need to pass the offset via a separate
parameter. Mach-o and COFF relocations don't allow a negative
offset to a symbol, which is handled properly if the offset is passed
via the parameter. If no offset parameter is given, the macro
evaluates to something like "adrp x17, subpel_filters-16+(0)", which
older clang versions also fail to parse (the older clang versions
only support one single offset term, although it can be a parenthesis.
Signed-off-by: Martin Storsjö <martin@martin.st>
(cherry picked from commit 26d7af4c38)
The previous version was a pretty exact translation of the arm
version. This version does do some unnecessary arithemetic (it does
more operations on vectors that are only half filled; it does 4
uaddw and 4 sqxtun instead of 2 of each), but it reduces the overhead
of packing data together (which could be done for free in the arm
version).
This gives a decent speedup on Cortex A53, a minor speedup on
A72 and a very minor slowdown on Cortex A73.
Before: Cortex A53 A72 A73
vp8_idct_add_neon: 79.7 67.5 65.0
After:
vp8_idct_add_neon: 67.7 64.8 66.7
Signed-off-by: Martin Storsjö <martin@martin.st>
The original arm version didn't do saturation here. This probably
doesn't make any difference for performance, but reduces the
differences.
Signed-off-by: Martin Storsjö <martin@martin.st>
This makes it similar to put_epel16_v6, and gives a large speedup
on Cortex A53, a minor speedup on A72 and a very minor slowdown on
A73.
Before: Cortex A53 A72 A73
vp8_put_epel16_h6v6_neon: 2211.4 1586.5 1431.7
After:
vp8_put_epel16_h6v6_neon: 1736.9 1522.0 1448.1
Signed-off-by: Martin Storsjö <martin@martin.st>
This makes it similar to put_epel16_v6, and gives a 10-25%
speedup of this function.
Before: Cortex A7 A8 A9 A53 A72
vp8_put_epel16_h6v6_neon: 3058.0 2218.5 2459.8 2183.0 1572.2
After:
vp8_put_epel16_h6v6_neon: 2670.8 1934.2 2244.4 1729.4 1503.9
Signed-off-by: Martin Storsjö <martin@martin.st>
Martin Storsjö