It used to be allocated separately, so that the pointer to it
is copied to all HEVCContexts, so that all slice-threads
use the same. This is completely unnecessary now that there
is only one HEVCContext any more. There is just one minor
complication left: The slice-threads only get a pointer to
const HEVCContext, but they need to modify the common CABAC
state. Fix this by adding a pointer to the common CABAC state
to HEVCLocalContext and document why it exists.
Reviewed-by: Anton Khirnov <anton@khirnov.net>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
While just at it, also use av_calloc() instead of zeroing
the array ourselves in a loop.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Actually, ff_slice_thread_allocz_entries() always already
allocates zeroed entries, so ff_reset_entries() was already
unnecessary. Make this more clear by renaming it to
ff_slice_thread_allocz_entries().
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The HEVC decoder has both HEVCContext and HEVCLocalContext
structures. The latter is supposed to be the structure
containing the per-slicethread state.
Yet up until now that is not how it is handled in practice:
Each HEVCLocalContext has a unique HEVCContext allocated for it
and each of these coincides except in exactly one field: The
corresponding HEVCLocalContext. This makes it possible to pass
the HEVCContext everywhere where logically a HEVCLocalContext
should be used. And up until recently, this is how it has been done.
Yet the preceding patches changed this, making it possible
to avoid allocating redundant HEVCContexts.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Right now the code passes a list of ints whose entry #i
is just i as opaque parameter to hls_decode_entry_wpp
via execute2; said list is even constantly allocated and freed.
This commit stops doing so and instead passes the list of
HEVCLocalContext* instead, so that the main HEVCContext
can be avoided in accessing the HEVCLocalContext.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The HEVC decoder has both HEVCContext and HEVCLocalContext
structures. The latter is supposed to be the structure
containing the per-slicethread state.
Yet that is not how it is handled in practice: Each HEVCLocalContext
has a unique HEVCContext allocated for it and each of these
coincides except in exactly one field: The corresponding
HEVCLocalContext. This makes it possible to pass the HEVCContext
everywhere where logically a HEVCLocalContext should be used.
This commit stops doing this for lavc/hevcdec.c itself.
It also constifies what can be constified in order to make
the nonconst stuff stand out more.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The HEVC decoder has both HEVCContext and HEVCLocalContext
structures. The latter is supposed to be the structure
containing the per-slicethread state.
Yet that is not how it is handled in practice: Each HEVCLocalContext
has a unique HEVCContext allocated for it and each of these
coincides except in exactly one field: The corresponding
HEVCLocalContext. This makes it possible to pass the HEVCContext
everywhere where logically a HEVCLocalContext should be used.
This commit stops doing this for lavc/hevcpred as well as
the corresponding mips code; the latter is untested.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The HEVC decoder has both HEVCContext and HEVCLocalContext
structures. The latter is supposed to be the structure
containing the per-slicethread state.
Yet that is not how it is handled in practice: Each HEVCLocalContext
has a unique HEVCContext allocated for it and each of these
coincides except in exactly one field: The corresponding
HEVCLocalContext. This makes it possible to pass the HEVCContext
everywhere where logically a HEVCLocalContext should be used.
This commit stops doing this for lavc/hevc_cabac.c; it also constifies
everything that is possible in order to ensure that no slice thread
accidentally modifies the main HEVCContext state.
Reviewed-by: Anton Khirnov <anton@khirnov.net>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The HEVC decoder has both HEVCContext and HEVCLocalContext
structures. The latter is supposed to be the structure
containing the per-slicethread state.
Yet that is not how it is handled in practice: Each HEVCLocalContext
has a unique HEVCContext allocated for it and each of these
coincides with the main HEVCContext except in exactly one field:
The corresponding HEVCLocalContext.
This makes it possible to pass the HEVCContext everywhere where
logically a HEVCLocalContext should be used.
This led to confusion in the first version of what eventually became
commit c8bc0f66a875bc3708d8dc11b757f2198606ffd7:
Before said commit, the initialization of the Rice parameter derivation
state was incorrect; the fix for single-threaded as well as
frame-threaded decoding was to add backup stats to HEVCContext
that are used when the cabac state is updated*, see
https://ffmpeg.org/pipermail/ffmpeg-devel/2020-August/268861.html
Yet due to what has been said above, this does not work for
slice-threading, because the each HEVCLocalContext has its own
HEVCContext, so the Rice parameter state would not be transferred
between threads.
This is fixed in c8bc0f66a8
by a hack: It rederives what the previous thread was and accesses
the corresponding HEVCContext.
Fix this by treating the Rice parameter state the same way
the ordinary CABAC parameters are shared between threads:
Make them part of the same struct that is shared between
slice threads. This does not cause races, because
the parts of the code that access these Rice parameters
are a subset of the parts of code that access the CABAC parameters.
*: And if the persistent_rice_adaptation_enabled_flag is set.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The HEVC decoder has both HEVCContext and HEVCLocalContext
structures. The latter is supposed to be the structure
containing the per-slicethread state.
Yet that is not how it is handled in practice: Each HEVCLocalContext
has a unique HEVCContext allocated for it and each of these
coincides with the main HEVCContext except in exactly one field:
The corresponding HEVCLocalContext.
This makes it possible to pass the HEVCContext everywhere where
logically a HEVCLocalContext should be used.
This commit stops doing this for lavc/hevc_filter.c; it also constifies
everything that is possible in order to ensure that no slice thread
accidentally modifies the main HEVCContext state.
There are places where this was not possible, namely with the SAOParams
in sao_filter_CTB() or with sao_pixels_buffer_h in copy_CTB_to_hv().
Both of these instances lead to data races, see
https://fate.ffmpeg.org/report.cgi?time=20220629145651&slot=x86_64-archlinux-gcc-tsan-slices
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The HEVC decoder has both HEVCContext and HEVCLocalContext
structures. The latter is supposed to be the structure
containing the per-slicethread state.
Yet that is not how it is handled in practice: Each HEVCLocalContext
has a unique HEVCContext allocated for it and each of these
coincides except in exactly one field: The corresponding
HEVCLocalContext. This makes it possible to pass the HEVCContext
everywhere where logically a HEVCLocalContext should be used.
This commit stops doing this for lavc/hevc_mvs.c; it also constifies
everything that is possible in order to ensure that no slice thread
accidentally modifies the main HEVCContext state.
Reviewed-by: Anton Khirnov <anton@khirnov.net>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It is safe for a slice thread to read the main context
and therefore it is safe to add a pointer to const HEVCContext
(namely the parent context) to each HEVCLocalContext.
It is also safe (and actually redundant) to add a pointer
to a logcontext to HEVCLocalContext.
Doing so allows to pass the HEVCLocalContext as context in
the parts of the code that is run slice-threaded when slice-threading
is in use (currently these parts of the code use ordinary
HEVCContext*). This way one is not tempted to modify
the main context from the slice contexts.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The slicethread contexts need to be initialized for
every frame, not only the first one, so one can
remove the initialization when allocating these contexts,
because the ordinary per-frame initialization will
initialize them again just a few lines below.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It is overridden by ff_add_bytes_l2_sse2() on any non-ancient CPU.
Reviewed-by: Henrik Gramner <henrik@gramner.com>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Otherwise, there is no guarantee that the various av_log-messages
are not interrupted by another log statement. The latter may originate
from anywhere else, even the HEVC decoder itself, as happens when
one uses frame-threading to decode the BUMPING_A_ericsson_1.bit
sample from the FATE-suite.
Furthermore, the earlier approach suffered from the fact that
various parts of the logmsg were output with different loglevels
and that checking stopped after having encountered the first
plane with MD5 mismatch, although it is probably interesting to
know whether other planes are incorrect, too.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Dump all input/output names to OVModel struct. In case other funcs use
them for reporting errors or locating issues.
Signed-off-by: Ting Fu <ting.fu@intel.com>
encode_block() in svq1enc.c looks like the following:
static int encode_block(int block[7][256], int level)
{
int best_score = 0;
for (unsigned x = 0; x < level; x++) {
int v = block[1][x];
block[level][x] = 0;
best_score += v * v;
}
if (level > 0 && best_score > 64) {
int score = 0;
score += encode_block(block, level - 1);
score += encode_block(block, level - 1);
if (score < best_score) {
best_score = score;
}
}
return best_score;
}
When called from outside of encode_block(), it is always called with
level == 5.
This triggers a bug [1] in GCC: On -O3, it creates eight clones of
encode_block with different values of level inlined into it. The clones
with negative values are of course useless*, but they also lead to
-Warray-bounds warnings, because they access block[-1].
This has been mitigated in GCC 12: It no longer creates clones
for parameters that it knows are impossible. Somehow switching levels
to unsigned makes GCC know this. Therefore this commit does this.
(For GCC 11, this changes the warning to "array subscript 4294967295 is
above array bounds" from "array subscript -1 is below array bounds".)
[1]: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102513
*: These clones can actually be discarded when compiling with
-ffunction-sections.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
filter_mb_mbaff_edgev() and filter_mb_mbaff_edgecv()
have a function parameter whose expected size depends upon
another parameter: It is 2 * bsi + 1 (with bsi always being 1 or 2).
This array is declared as const int16_t[7], yet some of the callers
with bsi == 1 call it with only an const int16_t[4] available.
This leads to -Wstringop-overread warnings from GCC 12.1.
This commit fixes these by replacing [7] with [/* 2 * bsi + 1 */],
so that the expected range and its dependence on bsi is immediately
visible.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
check_block_inter() currently does this when calling check_block().
This leads to a -Wstringop-overflow= warning when compiling with
GCC 12.1.
Given that the main part of the body of check_block() consists
of an "if (intra) { ... } else { ... }" which is true iff
check_block() is not called from check_block_inter(),
it makes sense to fix this by just inlining check_block()
check_block_inter() and turning check_block() into a new
check_block_intra() (with the inter parts removed, of course).
This should also not make much of a difference for the generated code
given that both check_block() as well as check_block_inter()
are already marked as av_always_inline, so this commit follows
this route to fix the issue.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
multiswap_step() and multiswap_inv_step() both only require
six keys; in all current callers, these keys are part of
an array of twelve keys, yet in some of these callers the keys
given to these functions point to the second half of these
twelve keys, so that only six keys are available to these functions.
This led to -Wstringop-overread warnings when compiling with GCC 12.1.
Fix these by adapting the declaration of these functions.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Using tail calls with functions returning void is forbidden
(C99/C11 6.8.6.4: "A return statement with an expression shall not appear
in a function whose return type is void.") GCC emits a warning
because of this when using -pedantic: "ISO C forbids ‘return’ with
expression, in function returning void"
Reviewed-by: Hendrik Leppkes <h.leppkes@gmail.com>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It retrieves the muxer's internal timestamp with under-defined
semantics. Continuing to use this value would also require
synchronization once the muxer is moved to a separate thread.
Replace the value with last_mux_dts.
This field means different things when the video is encoded (number of
frames emitted to the encoding sync queue/encoder by the video sync
code) or copied (number of packets sent to the muxer sync queue).
Print the value of packets_written instead, which means the same thing
in both cases. It is also more accurate, since packets may be dropped by
the sync queue or bitstream filters.
Same issues apply to it as to -shortest.
Changes the results of the following tests:
- matroska-flac-extradata-update
The test reencodes two input FLAC streams into three output FLAC
streams. The last output stream is limited to 8 frames. The current
code results in the first two output streams having 12 frames, after
this commit all three streams have 8 frames and are the same length.
This new result is better, since it is predictable.
- mkv-1242
The test streamcopies one video and one audio stream, video is limited
to 11 frames. The new result shortens the audio stream so that it is
not longer than the video.
The -shortest option (which finishes the output file at the time the
shortest stream ends) is currently implemented by faking the -t option
when an output stream ends. This approach is fragile, since it depends
on the frames/packets being processed in a specific order. E.g. there
are currently some situations in which the output file length will
depend unpredictably on unrelated factors like encoder delay. More
importantly, the present work aiming at splitting various ffmpeg
components into different threads will make this approach completely
unworkable, since the frames/packets will arrive in effectively random
order.
This commit introduces a "sync queue", which is essentially a collection
of FIFOs, one per stream. Frames/packets are submitted to these FIFOs
and are then released for further processing (encoding or muxing) when
it is ensured that the frame in question will not cause its stream to
get ahead of the other streams (the logic is similar to libavformat's
interleaving queue).
These sync queues are then used for encoding and/or muxing when the
-shortest option is specified.
A new option – -shortest_buf_duration – controls the maximum number of
queued packets, to avoid runaway memory usage.
This commit changes the results of the following tests:
- copy-shortest[12]: the last audio frame is now gone. This is
correct, since it actually outlasts the last video frame.
- shortest-sub: the video packets following the last subtitle packet are
now gone. This is also correct.
James Almer
Andreas Rheinhardt
Xu Guangxin
Marvin Scholz
Zane van Iperen
Anton Khirnov
Ting Fu