This is by no means perfect, since at least ddagrab will return scRGB
data with values outside of 0.0f to 1.0f for HDR values.
Its primary purpose is to be able to work with the format at all.
_Float16 support was available on arm/aarch64 for a while, and with gcc
12 was enabled on x86 as long as SSE2 is supported.
If the target arch supports f16c, gcc emits fairly efficient assembly,
taking advantage of it. This is the case on x86-64-v3 or higher.
Same goes on arm, which has native float16 support.
On x86, without f16c, it emulates it in software using sse2 instructions.
This has shown to perform rather poorly:
_Float16 full SSE2 emulation:
frame=50074 fps=848 q=-0.0 size=N/A time=00:33:22.96 bitrate=N/A speed=33.9x
_Float16 f16c accelerated (Zen2, --cpu=znver2):
frame=50636 fps=1965 q=-0.0 Lsize=N/A time=00:33:45.40 bitrate=N/A speed=78.6x
classic half2float full software implementation:
frame=49926 fps=1605 q=-0.0 Lsize=N/A time=00:33:17.00 bitrate=N/A speed=64.2x
Hence an additional check was introduced, that only enables use of
_Float16 on x86 if f16c is being utilized.
On aarch64, a similar uplift in performance is seen:
RPi4 half2float full software implementation:
frame= 6088 fps=126 q=-0.0 Lsize=N/A time=00:04:03.48 bitrate=N/A speed=5.06x
RPi4 _Float16:
frame= 6103 fps=158 q=-0.0 Lsize=N/A time=00:04:04.08 bitrate=N/A speed=6.32x
Since arm/aarch64 always natively support 16 bit floats, it can always
be considered fast there.
I'm not aware of any additional platforms that currently support
_Float16. And if there are, they should be considered non-fast until
proven fast.
IEEE-754 differentiates two different kind of NaNs.
Quiet and Signaling ones. They are differentiated by the MSB of the
mantissa.
For whatever reason, actual hardware conversion of half to single always
sets the signaling bit to 1 if the mantissa is != 0, and to 0 if it's 0.
So our code has to follow suite or fate-testing hardware float16 will be
impossible.
This avoids triggering overflows in the filters, and avoids stray
test failures in the approximate functions on x86; due to rounding
differences, one implementation might overflow while another one
doesn't.
Signed-off-by: Martin Storsjö <martin@martin.st>
Some muxers, such as GPAC, create files with only one sidx, but two streams
muxed into the same fragments pointed to by this sidx.
Prevously, in such a case, when we seeked in such files, we fell back
to, for example, using the sidx associated with the video stream, to
seek the audio stream, leaving the seekhead in the wrong place.
We can still do this, but we need to take care to compare timestamps
in the same time base.
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
They are already synced generically in update_context_from_thread()
in pthread_frame.c.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It is unused since 3575a495f6
(and the error message is dangerous: av_get_pix_fmt_name(format)
returns NULL iff av_pix_fmt_desc_get(format) returns NULL
and using a NULL string for %s would be UB).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Improves the grepability of the code.
(Furthermore, I hope that no compiler will really call memset
for 28 bytes.)
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It is done later in ff_mpv_frame_start() (and nobody uses
current_picture_ptr between setting it in ff_mpv_frame_start()).
(The reason the vsynth*-h263-obmc ref files change is because
the call to ff_find_unused_picture() now happens after the older
pictures have been unreferenced in ff_mpv_frame_start(),
so that their slots in the picture array can be immediately
reused; the obmc code is somehow buggy and changes its output
depending on the earlier contents of the motion_val buffer.)
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Provide optimized implementation of pix_abs8 function for arm64.
Performance comparison tests are shown below.
- pix_abs_1_0_c: 101.2
- pix_abs_1_0_neon: 22.5
- sad_1_c: 101.2
- sad_1_neon: 22.5
Benchmarks and tests are run with checkasm tool on AWS Graviton 3.
Signed-off-by: Martin Storsjö <martin@martin.st>
Provide optimized implementation of sse8 function for arm64.
Performance comparison tests are shown below.
- sse_1_c: 130.7
- sse_1_neon: 29.7
Benchmarks and tests run with checkasm tool on AWS Graviton 3.
Signed-off-by: Hubert Mazur <hum@semihalf.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Provide optimized implementation of pix_abs16_y2 function for arm64.
Performance comparison tests are shown below.
pix_abs_0_2_c: 317.2
pix_abs_0_2_neon: 37.5
Benchmarks and tests run with checkasm tool on AWS Graviton 3.
Signed-off-by: Hubert Mazur <hum@semihalf.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Provide neon implementation for sse4 function.
Performance comparison tests are shown below.
- sse_2_c: 80.7
- sse_2_neon: 31.0
Benchmarks and tests are run with checkasm tool on AWS Graviton 3.
Signed-off-by: Hubert Mazur <hum@semihalf.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Provide neon implementation for sse16 function.
Performance comparison tests are shown below.
- sse_0_c: 268.2
- sse_0_neon: 43.5
Benchmarks and tests run with checkasm tool on AWS Graviton 3.
Signed-off-by: Hubert Mazur <hum@semihalf.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
c11fb46731 led to a regression whereby the return code for missing
input or input probe is overridden by writer close return code and
hence not conveyed in the exit code.
Since d69d12a5b9 these av_assert2()
(or more exactly, the ones in hadamard8_diff8x8_c() and
hadamard8_intra8x8_c()) are hit. So just remove all of these asserts.
(If the test were improved to know which functions expect h == 8
and which support any value, the asserts could be readded
at the appropriate places.)
Reviewed-by: Martin Storsjö <martin@martin.st>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This directory dependency is normally added implicitly by rules
in ffbuild/common.mak; for tools it's created by a rule for TOOLOBJS.
TOOLOBJS is populated implicitly from TOOLS, and decode_simple.o
doesn't end up there because it's an odd occurrance of a lone
object file in the tools subdirectory, not belonging to any other
tool.
Signed-off-by: Martin Storsjö <martin@martin.st>
Previously, the checkasm test always passed h=8, so no other cases
were tested.
Out of the me_cmp functions, in practice, some functions are hardcoded
to always assume a 8x8 block (ignoring the h parameter), while others
do use the parameter. For those with hardcoded height, both the
reference C function and the assembly implementations ignore the
parameter similarly.
The documentation for the functions indicate that heights between
w/2 and 2*w, within the range of 4 to 16, should be supported. This
patch just tests random heights in that range, without knowing what
width the current function actually uses.
Signed-off-by: Martin Storsjö <martin@martin.st>
The height is hardcoded in some of the me_cmp functions, but not
in all of them. But in the case of all other functions, it's hardcoded
in the same place in SIMD functions as in the C reference functions,
while this one function differs from the behaviour of the C code.
(Before 542765ce3e, there were a
couple other sad8_*_mmx functions with similar hardcoded height.)
Signed-off-by: Martin Storsjö <martin@martin.st>
This commit adds new code paths for vscale when filterSize is 2, 4, or
8. By using specialized code with unrolling to match the filterSize we
can improve performance.
On AWS c7g (Graviton 3, Neoverse V1) instances:
before after
yuv2yuvX_2_0_512_accurate_neon: 558.8 268.9
yuv2yuvX_4_0_512_accurate_neon: 637.5 434.9
yuv2yuvX_8_0_512_accurate_neon: 1144.8 806.2
yuv2yuvX_16_0_512_accurate_neon: 2080.5 1853.7
Signed-off-by: Jonathan Swinney <jswinney@amazon.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Use scalar times vector multiply accumlate instructions instead of
vector times vector to remove the need for replicating load instructions
which are slightly slower.
On AWS c7g (Graviton 3, Neoverse V1) instances:
yuv2yuvX_8_0_512_accurate_neon: 1144.8 987.4
yuv2yuvX_16_0_512_accurate_neon: 2080.5 1869.4
Signed-off-by: Jonathan Swinney <jswinney@amazon.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Change the reference to exactly match the C reference in swscale,
instead of exactly matching the x86 SIMD implementations (which
differs slightly). Test with and without SWS_ACCURATE_RND - if this
flag isn't set, the output must match the C reference exactly,
otherwise it is allowed to be off by 2.
Mark a couple x86 functions as unavailable when SWS_ACCURATE_RND
is set - apparently this discrepancy hasn't been noticed in other
exact tests before.
Add a test for yuv2plane1.
Signed-off-by: Jonathan Swinney <jswinney@amazon.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Use it instead of AVStream.codecpar in the main thread. While
AVStream.codecpar is documented to only be updated when the stream is
added or avformat_find_stream_info(), it is actually updated during
demuxing. Accessing it from a different thread then constitutes a race.
Ideally, some mechanism should eventually be provided for signalling
parameter updates to the user. Then the demuxing thread could pick up
the changes and propagate them to the decoder.
This specialization handles the case where filtersize is 4 mod 8, e.g.
12, 20, etc. Aarch64 was previously using the c function for this case.
This implementation speeds up that case significantly.
hscale_8_to_15__fs_12_dstW_512_c: 6234.1
hscale_8_to_15__fs_12_dstW_512_neon: 1505.6
Signed-off-by: Jonathan Swinney <jswinney@amazon.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
frag_stream_info->index_entry isn't the first sample/trun index.
cenc.frag_index_entry_base failed to catch the case since
current_index > 0.
Fix ticket #9807.
Signed-off-by: Zhao Zhili <zhilizhao@tencent.com>
frag_index.current is used by cenc_filter, and is updated inside
mov_read_moof. It can out of sync regarding to mov_read_packet.
Partly fix ticket #9807.
Signed-off-by: Zhao Zhili <zhilizhao@tencent.com>
Convert the input from a scatter to a gather instead,
which is faster and better for SIMD.
Also, add a pre-shuffled exptab version to avoid
gathering there at all. This doubles the exptab size,
but the speedup makes it worth it. In SIMD, the
exptab will likely be purged to a higher cache
anyway because of the FFT in the middle, and
the amount of loads stays identical.
For a 960-point inverse MDCT, the speedup is 10%.
This makes it possible to write sane and fast SIMD
versions of inverse MDCTs.
A gateway can see everything, and we should not be shipping a hardcoded
default from a third party company; it's a security risk.
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
Timo Rothenpieler
Martin Storsjö
Derek Buitenhuis
Andreas Rheinhardt
Chema Gonzalez
Alan Kelly
J. Dekker
Hubert Mazur
Gyan Doshi
Swinney, Jonathan
Anton Khirnov
Zhao Zhili
Lynne