This improves commit 59c7022740.
In ff_thread_report_progress(), the fast code path can load
progress[field] with the relaxed memory order, and the slow code path
can store progress[field] with the release memory order. These changes
are mainly intended to avoid confusion when one inspects the source code.
They are unlikely to have measurable performance improvement.
ff_thread_report_progress() and ff_thread_await_progress() form a pair.
ff_thread_await_progress() reads progress[field] with the acquire memory
order (in the fast code path). Therefore, one expects to see
ff_thread_report_progress() write progress[field] with the matching
release memory order.
In the fast code path in ff_thread_report_progress(), the atomic load of
progress[field] doesn't need the acquire memory order because the
calling thread is trying to make the data it just decoded visible to the
other threads, rather than trying to read the data decoded by other
threads.
In ff_thread_get_buffer(), initialize progress[0] and progress[1] using
atomic_init().
Signed-off-by: Wan-Teh Chang <wtc@google.com>
Signed-off-by: Anton Khirnov <anton@khirnov.net>
This could happen when there was a frame number gap and frame threading was used.
Debugging-by: Ronald S. Bultje <rsbultje@gmail.com>
Debugging-by: Justin Ruggles <justin.ruggles@gmail.com>
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
CC:libav-stable@libav.org
Signed-off-by: Anton Khirnov <anton@khirnov.net>
It is more natural for this codec and allows to avoid awkward constructs
like "consuming 0 bytes from input". Also, keep a reference to the input
packet to avoid unnecessary copying.
Currently, the new decoding API is pretty much just a wrapper around the
old deprecated one. This is problematic, since it interferes with making
full use of the flexibility added by the new API. The old API should
also be removed at some future point.
Reorganize the code so that the new send_packet/receive_frame functions
call the actual decoding directly and change the old deprecated
avcodec_decode_* functions into wrappers around the new API.
The new internal API for decoders is now changing as well. Before this
commit, it mirrors the public API, so the decoders need to implement
send_packet() and receive_frame() callbacks. This turns out to require
awkward constructs in both the decoders and the generic code. After this
commit, the decoders only implement the receive_frame() callback and
call a new internal function, ff_decode_get_packet() to obtain input
data, in the same manner to how the bitstream filters now work.
avcodec will now always make a reference to the input packet, which means
that non-refcounted input packets will be copied. Keeping the previous
behaviour, where this copy could sometimes be avoided, would make the
code significantly more complex and fragile for only dubious gains,
since packets are typically small and everyone who cares about
performance should use refcounted packets anyway.
The current code stores a pointer to the packet passed to the decoder,
which is then used during get_buffer() for timestamps and side data
passthrough. However, since this is a pointer to user data which we do
not own, storing it is potentially dangerous. It is also ill defined for
the new decoding API with split input/output.
Fix this problem by making an explicit internally owned copy of the
packet properties.
It is useful for testing/debugging and will also be used as the default
filter in the following commit adding pre-decode filtering to avoid
having a separate non-filtered codepath.
Also preserve the return value from ff_get_buffer().
Signed-off-by: Andreas Cadhalpun <Andreas.Cadhalpun@googlemail.com>
Signed-off-by: Vittorio Giovara <vittorio.giovara@gmail.com>
This is how we initialize refcount in libavutil/buffer.c.
Signed-off-by: Wan-Teh Chang <wtc@google.com>
Signed-off-by: Vittorio Giovara <vittorio.giovara@gmail.com>
This work is sponsored by, and copyright, Google.
Previously all subpartitions except the eob=1 (DC) case ran with
the same runtime:
vp9_inv_dct_dct_16x16_sub16_add_neon: 1373.2
vp9_inv_dct_dct_32x32_sub32_add_neon: 8089.0
By skipping individual 8x16 or 8x32 pixel slices in the first pass,
we reduce the runtime of these functions like this:
vp9_inv_dct_dct_16x16_sub1_add_neon: 235.3
vp9_inv_dct_dct_16x16_sub2_add_neon: 1036.7
vp9_inv_dct_dct_16x16_sub4_add_neon: 1036.7
vp9_inv_dct_dct_16x16_sub8_add_neon: 1036.7
vp9_inv_dct_dct_16x16_sub12_add_neon: 1372.1
vp9_inv_dct_dct_16x16_sub16_add_neon: 1372.1
vp9_inv_dct_dct_32x32_sub1_add_neon: 555.1
vp9_inv_dct_dct_32x32_sub2_add_neon: 5190.2
vp9_inv_dct_dct_32x32_sub4_add_neon: 5180.0
vp9_inv_dct_dct_32x32_sub8_add_neon: 5183.1
vp9_inv_dct_dct_32x32_sub12_add_neon: 6161.5
vp9_inv_dct_dct_32x32_sub16_add_neon: 6155.5
vp9_inv_dct_dct_32x32_sub20_add_neon: 7136.3
vp9_inv_dct_dct_32x32_sub24_add_neon: 7128.4
vp9_inv_dct_dct_32x32_sub28_add_neon: 8098.9
vp9_inv_dct_dct_32x32_sub32_add_neon: 8098.8
I.e. in general a very minor overhead for the full subpartition case due
to the additional cmps, but a significant speedup for the cases when we
only need to process a small part of the actual input data.
Signed-off-by: Martin Storsjö <martin@martin.st>
This work is sponsored by, and copyright, Google.
Previously all subpartitions except the eob=1 (DC) case ran with
the same runtime:
Cortex A7 A8 A9 A53
vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0
vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3
By skipping individual 4x16 or 4x32 pixel slices in the first pass,
we reduce the runtime of these functions like this:
vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8
vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7
vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5
vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7
vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1
vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1
vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9
vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5
vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8
vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5
vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2
vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6
vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6
vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0
vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7
vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1
I.e. in general a very minor overhead for the full subpartition case due
to the additional loads and cmps, but a significant speedup for the cases
when we only need to process a small part of the actual input data.
In common VP9 content in a few inspected clips, 70-90% of the non-dc-only
16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left
8x8 or 16x16 subpartitions respectively.
Signed-off-by: Martin Storsjö <martin@martin.st>
This avoids reloading them if they haven't been clobbered, if the
first pass also was idct.
This is similar to what was done in the aarch64 version.
Signed-off-by: Martin Storsjö <martin@martin.st>
Diego Biurrun
Wan-Teh Chang
Derek Buitenhuis
Anton Khirnov
Alexandra Hajkova
Mark Thompson
Andreas Cadhalpun
Timothy Gu
Vittorio Giovara
James Almer
Martin Storsjö
Clément Bœsch