Intel's Instruction Set Reference (as of September 2015) clearly states
that cvtpi2ps switches to MMX state. Actual CPUs do not switch if the
source is a memory location. The Instruction Set Reference from 1999
(Order Number 243191) describes this behaviour but all later versions
I've seen have make no distinction whether MMX registers or memory is
used as source.
The documentation for the matching SSE2 instruction to convert to double
(cvtpi2pd) was fixed (see the valgrind bug
https://bugs.kde.org/show_bug.cgi?id=210264).
It will take time to get a clarification and fixes in place. In the
meantime it makes sense to change ff_int32_to_float_fmul_scalar_sse to
be correct according to the documentation. The vast majority of users
will have SSE2 so a change to the SSE version has little effect.
Fixes fate-checkasm on x86 valgrind targets.
Valgrind 'bug' reported as https://bugs.kde.org/show_bug.cgi?id=357059
Check the full FPU tag word instead of only the lower half and simplify
the comparison.
Use upper-case function base name as macro name to instantiate both
checked_call variants.
Not every asm routine is expected clear the MMX state after returning.
It is however a requisite for testing floating point code in checkasm.
Annotate functions requiring cleanup with declare_func_emms() and issue
emms after the call. The remaining functions are checked for having a
cleared MMX state after return.
Quite a bit faster than int32_to_float_fmul_array8_c calling
ff_int32_to_float_fmul_scalar_neon through FmtConvertContext.
Number of cycles per int32_to_float_fmul_array8 call while decoding
padded.dts on exynos5422:
before after change
cortex-a7: 1270 951 -25%
cortex-a15: 434 285 -34%
checkasm --bench cycle counts: cortex-a15 cortex-a7
int32_to_float_fmul_array8_c: 1730.4 4384.5
int32_to_float_fmul_array8_neon_c: 571.5 1694.3
int32_to_float_fmul_array8_neon: 374.0 1448.8
Interesting are the differences between
int32_to_float_fmul_array8_neon_c and int32_to_float_fmul_array8_neon.
The former is current behaviour of calling
ff_int32_to_float_fmul_scalar_neon repeatedly from the c function,
The raw numbers differ since checkasm uses different lengths than the
dca decoder.
~25% faster dts decoding overall. The checkasm CPU cycles numbers are
not that useful since synth_filter_float() calls FFTContext.imdct_half().
cortex-a57 cortex-a53
synth_filter_float_c: 1866.2 3490.9
synth_filter_float_neon: 915.0 1531.5
With fftc.imdct_half forced to imdct_half_neon:
cortex-a57 cortex-a53
synth_filter_float_c: 1718.4 3025.3
synth_filter_float_neon: 926.2 1530.1
The vector mode was deprecated in ARMv7-A/VFPv3 and various cpu
implementations do not support it in hardware. Vector mode code will
depending the OS either be emulated in software or result in an illegal
instruction on cpus which does not support it. This was not really
problem in practice since NEON implementations of the same functions are
preferred. It will however become a problem for checkasm which tests
every cpu flag separately.
Since this is a cpu feature newer cpu do not support anymore the
behaviour of this flag differs from the other flags. It can be only
activated by runtime cpu feature selection.
The ISB (instruction synchronization barrier) might be too heavy for
START/STOPTIMER use but should be more accurate in checkasm where the
timing overhead is subtracted.
Include macros.h explicitly in common.h so that external code using
FFALIGN does not break. It was already implicitly included through
version.h. Include macros.h in lls.h and internal.h for FFALIGN.
lls.h was including common.h only for FFALIGN and internal.h was
missing the include for FFALIGN. `make checkheaders` did not catch it
because it's an internal header.
display_matrix_size is only initialized when av_stream_get_side_data()
returns a side data pointer. The code is safe since the only effect this
has is setting the display_matrix pointer to NULL which it was already
anyway.
The existing meridian audio test does not test
ff_mlp_rematrix_channel_arm. This sample (first 640k of
https://samples.libav.org/A-codecs/TrueHD/TrueHD.raw) uses
ff_mlp_rematrix_channel_arm. Since this sample has 5.1 channels it also
allows testing the integrated downmixing.
MpegEncContext based decoders are only fully initialized after the first
ff_thread_get_buffer() call. The RV30/40 decoders may fail before a frame
buffer was requested. ff_mpeg_update_thread_context() fails on half
initialized MpegEncContexts. Since this can only happen before a the
first frame was decoded there is no need to call
ff_mpeg_update_thread_context().
Based on patches by John Stebbins and tested by John Stebbins.
CC: libav-stable@libav.org
This makes the default of '1' more explicit than defaulting to '1' in
fate-run.sh and regression-funcs.sh if THREADS is not set.
Fixes the reported thread count in fate-cpu if THREADS is not set.
llvm's integrated assembler does not accept spaces as macro argument
delimiter when targeting darwin. Using a explicit delimiter is a good
idea in principle since it makes case like 'macro 4 -2' vs 'macro 4 - 2'
clear.
libavutil/cpu-test prints raw and effective cpu flags to STDERR. Detected
cpu flags can be useful for debugging fate errors.
No comparison of the result against a expected result since that would
require fate config specific references.
Adding 'branch=release/10' to the fate config file will check the
release/10 branch instead of master. If no branch is specified it will
use 'master' so that existing config are still valid.
The server side changes are already deployed, see
https://fate.libav.org/v10/ for an example. The server supports only the
release/* branches.
The server enforces that a single slot tests always the same branch.
Please append "-v$RELEASE" to the slot of release branch configs or make
the slot otherwise unique.
A different fate samples dir is needed for each release branch. make
fate-rsync has the correct URL in each branch.
Since error resilience uses AVFrame pointers instead of references it
has to copy NULL pointers too. After a codec flush the last/next frame
pointers in MpegEncContext are NULL and the old pointers remaining in
ERContext are invalid. Fixes a crash in vlc for android thumbnailer.
Reported and debugged by Adrien Maglo <magsoft@videolan.org>.
Clang's integrated assembler (after 3.4) does not yet support the
'.altmacro' directive which is only used in arm asm. Support is planned:
http://llvm.org/bugs/show_bug.cgi?id=18918
Move the GNU as check before the arch specific asm checks since the .dn
check requires gas compatible assembler.
Disable the VC-1 motion compensation NEON asm which is the only part
using that directive. The integrated assembler in the upcoming clang 3.5
does not support .dn/.qn without plans to change that. Too much effort
to implement it while it is rarely used.
http://llvm.org/bugs/show_bug.cgi?id=18199.
Both gnu as and clang treat lines starting with '#' as comments if they
aren't consumed by the C-style preprocessor.
Using '//' does not work with clang since comments are removed before
macro expansion.
avcodec might depend on avresample and with --as-needed required symbols
might be get removed if avresample is linked before avcodec.
Fixes link failures of avprobe and avplay on aarch64 with
--enable-neon-clobber-test.
avctx->coded_{height,width} will always equal h->{height,width} since
init_dimensions() does that explicitly, Size changes are detected by
changes in mb_{height,width} earlier and propagated through the
needs_reinit variable.
Ported from arm NEON and added vector_dmul_scalar.
Functions between 1.5 and 5 times faster than the C implementations
using Apple's clang-503.0.19 on A7.
Based on the aarch64 asm. CPU cycle counts on cortex-a9 compared to
gcc 4.8.2:
before: 475 decicycles in get_cabac_noinline, 67106035 runs, 2829 skips
after: 393 decicycles in get_cabac_noinline, 67106474 runs, 2390 skips
Overall speedup is above 2%. Code generated by clang 3.4 is slower on
the same hardware and the relative change is a little larger.
Based on the x86 branchless get_cabac asm. get_cabac_noinline() gets
approximately 20% faster (no cycle counts available) compared to clang
from Xcode 5.1 beta5. More than 6% faster overall. A part of the overall
speedup might be explained by additional inlining of get_cabac().
The overread avoidance fix in cbddee1cca
broke the computation for the last row since it prevented the safe
reading from the height+1-th row.
CC: libav-stable@libav.org
Based on a patch from Christophe Gisquet.
Unrolling of the m == 0 case avoids a possible use of the uninitilized
value sum when s->predictor_history is not set. I failed to find a
sample for it. It also reduced the cycle count from 220 to 150 on
sandy bridge, x86_64 linux, gcc 4.8.2 compared to his patch.
Moving cpunop from the HAVE_LIST to the ARCH_EXT_LIST_X86 has the side
effect of enabling it. The semantics of the check have to be changed
from enable if successful to disable if unsuccessful. This was missing
in 2b0bb69997 causing build errors with
nasm.
Janne Grunau
