Seems like the last refactor to this code broke playing flipped images,
at least with --opengl-pbo --gpu-api=opengl.
Flipping is rather a shitmess. The main problem is that OpenGL does not
support flipped uploading. The original vo_gl implementation considered
it important to handle the flipped case efficiently, so instead of
uploading the image line by line backwards, it uploaded it flipped, and
then flipped it in the renderer (basically the upload path ignored the
flipping). The ra code and backends probably have an insane and
inconsistent mix of semantics, so fix this by never passing it flipped
images in the first place.
In the future, the backends should probably support flipped images
directly.
Fixes#5097.
If subtitles are part of the stream, determining the seekable range
becomes harder. Subtitles are sparse, and can have packets in irregular
intervals, or even completely lack packets. The usual logic of computing
the seek range by the min/max packet timestamps fails.
Solve this by making the only assumption we can make: subtitle packets
are implicitly demuxed along with other packets. We also assume perfect
interleaving for this, but you really can't do anything with sparse
packets that makes sense without this assumption.
One special case is if we prune sparse packets within the current
seekable range. Obviously this should limit the seekable range to after
the pruned packet.
Instead of weirdly deciding this on every packet read and having the
code far away from where it's actually needed, just run it where it's
actually needed.
A typical idiom for calling functions that remove something from the
array being iterated, but it's not needed here. I have no idea why this
was ever done.
Setting ds->refreshing for unselected streams could lead to a
nonsensical queue overflow warning, because read_packet() took it as a
reason to continue reading.
Also add some more information to the queue overflow warning (even if
that one doesn't have anything to do with this bug - it was for
unselected streams only).
This fixes an endless loop with threading disabled, such as for example
when playing a file with an external subtitle file, and seeking to the
beginning. Something will set in->seeking, but the seek is never
executed, which made demux_read_packet() loop endlessly. (External
subtitles will use non-threaded mode for whatever reasons.)
Fix this by by making the unthreaded code to execute the worker thread
body, which reduces the difference in logic.
Until now, the demuxer cache was limited to a single range. Extend this
to multiple range. Should be useful for slow network streams.
This commit changes a lot in the internal demuxer cache logic, so
there's a lot of room for bugs and regressions. The logic without
demuxer cache is mostly untouched, but also involved with the code
changes. Or in other words, this commit probably fucks up shit.
There are two things which makes multiple cached ranges rather hard:
1. the need to resume the demuxer at the end of a cached range when
seeking to it
2. joining two adjacent ranges when the lowe range "grows" into it (and
resuming the demuxer at the end of the new joined range)
"Resuming" the demuxer means that we perform a low level seek to the end
of a cached range, and properly append new packets to it, without adding
packets multiple times or creating holes due to missing packets.
Since audio and video never line up exactly, there is no clean "cut"
possible, at which you could resume the demuxer cleanly (for 1.) or
which you could use to detect that two ranges are perfectly adjacent
(for 2.). The way how the demuxer interleaves multiple streams is also
unpredictable. Typically you will have to expect that it randomly allows
one of the streams to be ahead by a bit, and so on.
To deal with this, we have heuristics in place to detect when one packet
equals or is "behind" a packet that was demuxed earlier. We reuse the
refresh seek logic (used to "reread" packets into the demuxer cache when
enabling a track), which checks for certain packet invariants.
Currently, it observes whether either the raw packet position, or the
packet DTS is strictly monotonically increasing. If none of them are
true, we discard old ranges when creating a new one.
This heavily depends on the file format and the demuxer behavior. For
example, not all file formats have DTS, and the packet position can be
unset due to libavformat not always setting it (e.g. when parsers are
used).
At the same time, we must deal with all the complicated state used to
track prefetching and seek ranges. In some complicated corner cases, we
just give up and discard other seek ranges, even if the previously
mentioned packet invariants are fulfilled.
To handle joining, we're being particularly dumb, and require a small
overlap to be confident that two ranges join perfectly. (This could be
done incrementally with as little overlap as 1 packet, but corner cases
would eat us: each stream needs to be joined separately, and the cache
pruning logic could remove overlapping packets for other streams again.)
Another restriction is that switching the cached range will always
trigger an asynchronous low level seek to resume demuxing at the new
range. Some users might find this annoying.
Dealing with interleaved subtitles is not fully handled yet. It will
clamp the seekable range to where subtitle packets are.
Build ffmpeg-mpv, shaderc and crossc from source, since they are not
packaged in MSYS2. Also, add some more explicit --enable flags to the
mpv build to make sure things like D3D11, D3D11VA hwaccels and Vulkan
are auto-detected.
Like the manual says, this is technically undefined behaviour. See:
https://msdn.microsoft.com/en-us/library/windows/desktop/ff476085.aspx
In particular, MSDN says texture arrays created with the BIND_DECODER
flag cannot be used with CreateShaderResourceView, which means they
can't be sampled through SRVs like normal Direct3D textures. However,
some programs (Google Chrome included) do this anyway for performance
and power-usage reasons, and it appears to work with most drivers.
Older AMD drivers had a "bug" with zero-copy decoding, but this appears
to have been fixed. See #3255, #3464 and http://crbug.com/623029.
The shader cache in ra_d3d11 caches the result of shaderc, crossc and
the D3DCompiler DLL, so it should be invalidated when any of those
components are updated. This should make the cache more reliable, which
makes it safer to enable gpu-shader-cache-dir. Shader compilation is
slow with D3D11, so gpu-shader-cache-dir is highly necessary
Some shaders take a _long_ time to compile with the Direct3D compiler.
The ANGLE backend had this problem too, to a certain extent. Logging
should help identify which shaders cause long stalls and could also help
with benchmarking ways of reducing compile times.
ra_d3d11 uses the SPIR-V compiler to translate GLSL to SPIR-V, which is
then translated to HLSL. This means it always exposes the same GLSL
version that the SPIR-V compiler supports (4.50 for shaderc/glslang.)
Despite claiming to support GLSL 4.50, some features that are tied to
the GLSL version in OpenGL are not supported by ra_d3d11 when targeting
legacy Direct3D feature levels.
This includes two features that mpv relies on:
- Reading from gl_FragCoord in the fragment shader (requires FL 10_0)
- textureGather from any texture component (requires FL 11_0)
These features have been exposed as new RA caps.
This is a new RA/vo_gpu backend that uses Direct3D 11. The GLSL
generated by vo_gpu is cross-compiled to HLSL with SPIRV-Cross.
What works:
- All of mpv's internal shaders should work, including compute shaders.
- Some external shaders have been tested and work, including RAVU and
adaptive-sharpen.
- Non-dumb mode works, even on very old hardware. Most features work at
feature level 9_3 and all features work at feature level 10_0. Some
features also work at feature level 9_1 and 9_2, but without high-bit-
depth FBOs, it's not very useful. (Hardware this old is probably not
fast enough for advanced features anyway.)
Note: This is more compatible than ANGLE, which requires 9_3 to work
at all (GLES 2.0,) and 10_1 for non-dumb-mode (GLES 3.0.)
- Hardware decoding with D3D11VA, including decoding of 10-bit formats
without truncation to 8-bit.
What doesn't work / can be improved:
- PBO upload and direct rendering does not work yet. Direct rendering
requires persistent-mapped PBOs because the decoder needs to be able
to read data from images that have already been decoded and uploaded.
Unfortunately, it seems like persistent-mapped PBOs are fundamentally
incompatible with D3D11, which requires all resources to use driver-
managed memory and requires memory to be unmapped (and hence pointers
to be invalidated) when a resource is used in a draw or copy
operation.
However it might be possible to use D3D11's limited multithreading
capabilities to emulate some features of PBOs, like asynchronous
texture uploading.
- The blit() and clear() operations don't have equivalents in the D3D11
API that handle all cases, so in most cases, they have to be emulated
with a shader. This is currently done inside ra_d3d11, but ideally it
would be done in generic code, so it can take advantage of mpv's
shader generation utilities.
- SPIRV-Cross is used through a NIH C-compatible wrapper library, since
it does not expose a C interface itself.
The library is available here: https://github.com/rossy/crossc
- The D3D11 context could be made to support more modern DXGI features
in future. For example, it should be possible to add support for
high-bit-depth and HDR output with DXGI 1.5/1.6.
Backported from @haasn's change to libplacebo, except in the current RA,
there's nothing to indicate an ra_format can be bound as a storage
image, so there's no way to force all of these formats to have a
glsl_format. Instead, the layout qualifier will be removed if
glsl_format is NULL.
This is needed for the upcoming ra_d3d11 backend. In Direct3D 11, while
loading float values from unorm images often works as expected, it's
technically undefined behaviour, and in Windows 10, it will cause the
debug layer to spam the log with error messages. Also, apparently in
GLSL, the format name must match the image's format exactly (but in
Direct3D, it just has to have the same component type.)
Backported from @haasn's change to libplacebo. More flexible than the
previous "shared || non-shared" distinction. The extra flexibility is
needed for Direct3D 11, but it also doesn't hurt code-wise.
For some reason vo_lavc's draw_image can buffer the frame and encode it
only later. Also, there is logic for rendering the OSD (i.e. subtitles)
only when needed.
In theory this can lead to subtitles being pruned before it tries to
render them (as the subtitle logic doesn't know that the VO still needs
them later), although this probably never happens in reality.
The worse issue, that actually happened, is that if the last frame gets
buffered, it attempts to render subtitles in the uninit callback. At
this point, the subtitle decoder is already torn down and all subtitles
removed, thus it will draw nothing. This didn't always happen. I'm not
sure why - potentially in the working cases, the frame wasn't buffered.
Since this logic doesn't have much worth, except a minor performance
advantage if frames with subtitles are dropped, just remove it.
Hopefully fixes#4689.
libavcodec can't deal with them, because its API doesn't distinguish
between 0 sized packets and sending EOF. As such, keeping them doesn't
do any good, ever. This actually fixes some obscure mkv sample (see
previous commit).
Fixes some obscure sample that uses fixed size laces with 0-sized lace
size. Some broken shit. (Maybe the decoder wouldn't care about these
packets, but the demuxer attempted to resync after these packet reading
errors, even though they were perfectly recoverable. But I don't care
enough about this.)
Sample link: https://samples.ffmpeg.org/Matroska/switzler084d_dl.mkv
This directly reads individual mkv sub-packets (block laces) into a
dedicated AVBufferRefs, which can be directly used for creating packets
without a additional copy of the packet data. This also means we switch
parsing of block header fields and lacing metadata to read directly from
the stream, instead of a memory buffer.
This could have been much easier if libavcodec didn't require padding
the packet data with zero bytes. We could just have each packet
reference a slice of the block data. But as it is, the only way to get
padding without a copy is to read the laces into individually allocated
(and padded) memory block, which required a larger rewrite.
This probably makes recovering from broken mkv files slightly worse if
the transport is unseekable. We just read, and then check if we've
overread. But I think that shouldn't be a real concern.
No actual measureable performance change. Potential for some
regressions, as this is quite intrusive, and touches weird obscure shit
like mkv lacing. Still keeping it because I like how it removes some
redundant EBML parsing functions.
This adds a bunch of stuff (mostly unused or redundant) as preparation
for supporting multiple seek ranges. Actual support is probably still
far away.
One change that messes deeper with the actual code is that we account
for total buffered bytes instead of just the backwards bytes now. This
way, prune_old_packets() doesn't have to iterate over all seek ranges to
determine whether something needs pruning.
The main purpose of this commit is avoiding any hidden O(n^2) algorithms
in the code for pruning the demuxer cache, and for determining the
seekable boundaries of the cache. The old code could loop over the whole
packet queue on every packet pruned in certain corner cases.
There are two ways how to reach the goal:
1) commit a cardinal sin
2) do everything incrementally
The cardinal sin is adding an extra field to demux_packet, which caches
the determined seekable range for a keyframe range. demux_packet is a
rather general data structure and thus shouldn't have any fields that
are not inherent to its use, and are only needed as an implementation
detail of code using it. But what are you gonna do, sue me?
In the future, demux.c might have its own packet struct though. Then the
other existing cardinal sin (the "next" field, from MPlayer times) could
be removed as well.
This commit also changes slightly how the seek end is determined. There
is a note on the manpage in case anyone finds the new behavior
confusing. It's somewhat cleaner and might be needed for supporting
multiple ranges (although that's unclear).
The demuxer cache seeking mechanism looks at keyframe ranges to
determine the earlierst PTS of a packet. Instead of looping over all
packets twice (once to find the next keyframe, a second time to find the
seek PTS), do it in one go.
For that basically turn recompute_keyframe_target_pts() into an
iteration functionn. Functionality should be unchanged with this commit.
The base_ts field is used to guess the decoder position, and when set to
NOPTS, it just read ahead arbitrarily. Also demux_add_packet() sets
base_ts to the new timestamp when appending a packet, which would also
make it readahead by a too large amount.
Fix this by setting base_ts after a seek. This assumes that normally, a
cached seek target will always have the timestamp set. This is actually
not quite clear (as it calls recompute_keyframe_target_pts(), which
looks at multiple packets), but maybe it works well enough.
Don't do any of the extra work related to pruning the backbuffer if
demuxer cache seeking is disabled.
As a small extra, always prune packets with no timestamps immediately,
or queue heads that are not keyframes. (Both of them would be pruned
from the backbuffer by the normal logic anyway.)
If fulfilling --demuxer-readahead-secs requires more memory than allowed
by --demuxer-max-bytes, the queue obviously overflows. But the warning
is normally only for the case when trying to find the next video or
audio packet fails, and decoding can't continue.
Use a separate variable for determining whether to prefetch, and if the
queue has overflown, skip the message. In fact, skip the EOF setting and
waking up the decoder thread as well, because the EOF flag should not be
(have been) set in this situation, and waking up the reader thread helps
only if the EOF state changed.
Repeating frames (for display-sync) is not supposed to render the entire
frame again. When using hardware decoding, it unfortunately did: the
renderer uses the frame ID to check whether the frame data changed, and
unmapping the hwdec frame clears it.
Essentially reverts commit 761eeacf54. Back then I probably
thought it would be a good idea to release the hwdec image quickly in
order to return it to the decoder, but they're referenced anyway.
This should increase the performance and reduce GPU work.
In a shit show of subtle corner case interactions, making the demuxer
cache buffer the entire file can display a small buffered time if
subtitles are enabled. The reason is that some subtitle decoders may
read in advance infinitely, i.e. they read the entire subtitle stream.
Then, since the other streams (audio/video) have logically reached EOF,
and the subtitle stream is set to ds->active==true. This will have to be
fixed properly later to account buffering for subtitle-only files
(another corner case) correctly, but for now this is less annoying.
We don't hope to auto-detect them at load time, as that would be too
much of a pain - even FFmpeg requires fetching and parsing of video
packets, and exposes the information only via deprecated API.
But there still needs to be a way to select them by default. This is
also needed to get the first CC packet at all (without seeking back).
This commit also attempts to clean up locking a bit, which is a PITA,
but it's better be careful & clean.
Normally such code is didsabled by have_mglsl==false in
check_gl_features(), but apparently not this one.
Just fix it. Seems also more readable.
Fixes#5069.
Apparently this is required, but it doesn't check for it. To be fair,
this was tested by creating a compatibility context and pretending it's
GL 2.1. GL_ARB_shader_storage_buffer_object actually requires GL 4.0 or
up, but GL_ARB_uniform_buffer_object requires only GL 2.0.
According to
https://github.com/libarchive/libarchive/pull/773#issuecomment-334892291
we're not allowed to "continue reading" (post above) or performing "more
operations" (comments in archive.h header), whatever that means. Assume
closing and freeing the archive is still ok.
Since the codec already includes logic for closing and reopening the
archive for seeking in unseekable archives, this probably isn't too bad.
Untested due to lack of crashing sample (I lost my original test case,
and as recently user-provided one didn't crash).