Fixes mismatch in first keyframe in sample
ffvp9_fails_where_libvpx.succeeds.webm from ticket 3849. There's still
a second mismatch a few frames into the sample.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Such files can be created using the --bff x264 option.
Sample-Id: h264_direct_temporal_mvs_bff.mkv
Signed-off-by: Luca Barbato <lu_zero@gentoo.org>
Signed-off-by: Vittorio Giovara <vittorio.giovara@gmail.com>
Linear interpolation goes from 63 (llvm) or 58 (gcc) to 48 (yasm)
cycles/sample on 64bit, or from 66 (llvm/gcc) to 52 (yasm) cycles/
sample on 32bit. Bon-linear goes from 43 (llvm) or 38 (gcc) to
32 (yasm) cycles/sample on 64bit, or from 46 (llvm) or 44 (gcc) to
38 (yasm) cycles/sample on 32bit (all testing on OSX 10.9.2, llvm
5.1 and gcc 4.8/9).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
DSP bits of swri_resample go into their own mini-DSP functions; DSP
init goes from a per-call branch in multiple_resample to a proper
DSP init routine; x86 bits go into x86/; swri_resample() moves out of
resample_template.c into resample.c because it's independent of DSP
code or sample type; multiple_resample() is simplified.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
I don't see dst_incr/dst_incr_frac ever being changed from their
initial value (which is the inverse of this operation), so it seems
to me that this is a no-op.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
I think there's an off-by-one in terms of the switchpoint where we
switch from dst_incr to ideal_dst_incr, I don't think that's a massive
issue, but just be aware of that. It's probably trivial to prevent but
I don't care.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
I could not reproduce any off by 1 error, results are bit exact (michael)
This removes a branch at a performance-sensitive point (in the middle
of the loop). In fate-swr-resample-s32p-8000-2626, this makes the code
about 10% faster. It also simplifies the loops, allowing us to rewrite
it in yasm at some later point.
The compensation_distance != 0 code and index < 0 code are still kind
of hairy. For compensation_distance != 0, this should likely be handled
in the caller, so that it calls swri_resample twice (once until the
dst_incr switch-point, and once with the remainder of the samples). For
index < 0, the code should probably be rewritten to break out of the
loop once sample_index >= 0, and then resume (e.g. as a tail-call) to
the common or linear resampling loops.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
The advantage of this is that the is32x32 division branch in
decode_coeffs_b is removed from the inner loop to outside the block
coef decoding loop in decode_coeffs. Also, it allows us to merge the
txsz branches from the block coef decoding loop, the context merge
and the context split.
The old one didn't use segmentation. One uses segmentation in all frame
types (--aq-mode=1), and the other uses all segmentation features, but
only in inter frames (mbgraph).
Signed-off-by: Anton Khirnov <anton@khirnov.net>
This disables backward probability updates, which makes the codec more
friendly for frame-level multi-threading.
Signed-off-by: Anton Khirnov <anton@khirnov.net>
The directional intra predictors either don't care about order (dc, h,
dc_left, tm), or they prefer inverted order (vr, dr, hd). This allows
more efficient SIMD implementations.
The spec doesn't describe how it should be decoded so this is probably
the safest thing to do. Fixes valgrind errors on fuzzed11.ivf and fixes
valgrind errors on fuzzed10.ivf differently.
Should fix crashes or corrupt output on pre-SSE2 CPUs when they were
using SSE2-code (e.g. AMD Athlon XP 2400+ or Intel Pentium III) in
hfix or hvar single-edge (left/right) extension functions.
Signed-off-by: Janne Grunau <janne-libav@jannau.net>
Fixes invalid reads and crashes in vp90-2-05-resize.webm and fuzzed6.ivf.
The output is still not identical to what libvpx does (because we don't
actually scale in MC).
Reviewed-by: ubitux
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Prevents some invalid memory accesses after resolution change in
vp90-2-05-resize.webm, and libvpx does this too.
Reviewed-by: ubitux
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Before this patch, we explicitly modify rsp, which isn't necessarily
universally acceptable, since the space under the stack pointer might
be modified in things like signal handlers. Therefore, use an explicit
register to hold the stack pointer relative to the bottom of the stack
(i.e. rsp). This will also clear out valgrind errors about the use of
uninitialized data that started occurring after the idct16x16/ssse3
optimizations were first merged.
Runtime of the full 32x32 idct goes from 2446 to 2441 cycles (intra) or
from 1425 to 1306 cycles (inter). Overall runtime is not significantly
affected.
Runtime of all IDCTs together goes from 3327 to 2473 cycles (intra, i.e.
~35% faster) or from 2312 to 1448 cycles (inter, i.e. ~60% faster). Total
decode time of ped1080p.webm goes from 8.086sec to 7.974sec (1.4% faster).
Sub-IDCTs will follow later. ped1080.webm goes from 9.295s to 8.191s
(13.5% faster). The IDCT itself goes from 4372 (intra) or 4337 (inter)
to 403 (intra) or 329 (inter) cycles for the DC-only form, 23755 (intra)
or 23723 (inter) to 3497 (intra) or 3607 (inter) cycles for the no-DC
form, which averages from 23393 (intra) or 16612 (inter) to 3449 (intra)
or 2392 (inter) for all 32x32s together, i.e. about ~7x faster (all
tests done on ped1080p.webm).
Sub8x8 speed (w/o dc-only case) goes from ~750 cycles (inter) or ~735
cycles (intra) to ~415 cycles (inter) or ~430 cycles (intra). Average
overall 16x16 idct speed goes from ~635 cycles (inter) or ~720 cycles
(intra) to ~415 cycles (inter) or ~545 (intra) - all measurements done
using ped1080p.webm.
The reference map is never used in such cases, but we accidently copied
it anyway. This could cause crashes if this map has not yet been
allocated. Fixes trac ticket 3188.
Reviewed-by: Clément Bœsch <u@pkh.me>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Currently only dc-only and full 16x16. Other subforms will follow in the
near future. Total decoding time of ped1080p.webm goes from 9.7 to 9.3
seconds. DC-only goes from 957 -> 131 cycles, and the full IDCT goes
from ~4050 to ~745 cycles.
It was previously int and would return error if decode_coeffs_b()
returns an error; however, that can never happen, so refactor all
that code to make all dependent functions return void also (all the
way up to decode_coeffs_sb()).
The old one didn't use segmentation. One uses segmentation in all frame
types (--aq-mode=1), and the other uses all segmentation features, but
only in inter frames (mbgraph).
For a random 1080p sample, decoding time went from 9.7sec (1 threads)
to 6.0sec (2 threads) and 5.2sec (4 threads) in 2-pass decoding mode.
I don't have any samples that use the parallelmode feature, but the
gains should be higher.
We need more information from last/cur_frame than from reference
buffers, so we can use a simplified structure for reference buffers,
and then store mvs and segmentation map information in last/cur.
Don't use word-size multiplications if size == 2, and if we're using
SIMD instructions (size >= 8), complete leftover 4byte sets using movd,
not mov. Both of these changes lead to minor speedups.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Should fix crashes or corrupt output on pre-SSE2 CPUs when they were
using SSE2-code (e.g. AMD Athlon XP 2400+ or Intel Pentium III) in
hfix or hvar single-edge (left/right) extension functions.
Tested-by: Ingo Brückl <ib@wupperonline.de>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This allows supporting files for which the image stride is smaller than
the max. block size + number of subpel mc taps, e.g. a 64x64 VP9 file
or a 16x16 VP8 file with -fflags +emu_edge.
