Also refactors the usage of tex_upload to make ra_tex_upload_pbo a
RA-internal thing again.
ra_buf_pool has the main advantage of being dynamically sized depending
on buf_poll, so for OpenGL we'll end up only using one buffer (when not
persistently mapping) - while for vulkan we'll use as many as necessary,
which depends on the swapchain depth anyway.
- tex_uploads args are moved to a struct
- the ability to directly upload texture data without going through a
buffer is made explicit
- the concept of buffer updates and buffer polling is made more explicit
and generalized to buf_update as well (not just mapped buffers)
- the ability to call tex_upload/buf_update on a tex/buf is made
explicit during tex/buf creation
- uploading from buffers now uses an explicit offset instead of
implicitly comparing *src against buf->data, because not all buffers
may actually be persistently mapped
- the initial_data = immutable requirement is dropped. (May be re-added
later for D3D11 if that ever becomes a thing)
This change helps the vulkan abstraction immensely and also helps move
common code (like the PBO pooling) out of ra_gl and into the
opengl/utils.c
This also technically has the side-benefit / side-constraint of using
PBOs for OSD texture uploads as well, which actually seems to help
performance on machines where --opengl-pbo is faster than the naive code
path. Because of this, I decided to hook up the OSD code to the
opengl-pbo option as well.
One drawback of this refactor is that the GL_STREAM_COPY hack for
texture uploads "got lost", but I think I'm happy with that going away
anyway since DR almost fully deprecates it, and it's not the "right
thing" anyway - but instead an nvidia-only hack to make this stuff work
somewhat better on NUMA systems with discrete GPUs.
Another change is that due to the way fencing works with ra_buf (we get
one fence per ra_buf per upload) we have to use multiple ra_bufs instead
of offsets into a shared buffer. But for OpenGL this is probably better
anyway. It's possible that in future, we could support having
independent “buffer slices” (each with their own fence/sync object), but
this would be an optimization more than anything. I also think that we
could address the underlying problem (memory closeness) differently by
making the ra_vk memory allocator smart enough to chunk together
allocations under the hood.
The vp_w/vp_h variables and parameters were not really used anymore
(they were redundant with ra_tex w/h) - but vp_h was still used to
identify whether rendering should be done mirrored.
Simplify this by adding a fbodst struct (some bad naming), which
contains the render target texture, and some parameters how it should be
rendered to (for now only flipping). It would not be appropriate to make
this a member of ra_tex, so it's a separate struct.
Introduces a weird regression for the first frame rendered after
interpolation is toggled at runtime, but seems to work otherwise. This
is possibly due to the change that blit() now mirrors, instead of just
copying. (This is also why ra_fns.blit is changed.)
Fixes#4719.
Since the GL *gl is no longer needed for the timers, we can get rid of
the sc->gl dependency. This requires moving a utility function (which is
not GL-specific anyway) out of gl_utils.h and into utils.h
In order to prevent code duplication and keep the ra abstraction as
small as possible, `ra` only implements the actual timer queries,
it does not do pooling/averaging of the results. This is instead moved
to a ra-neutral struct timer_pool in utils.c.
Actually GL-specific parts go into gl_utils.c/h, the shader cache
(gl_sc*) into shader_cache.c/h.
No semantic changes of any kind, except that the VAO helper is made
public again as part of gl_utils.c (all while the goal for gl_utils.c
itself is to be included by GL-specific code).
Further work removing GL dependencies from the actual video renderer,
and moving them into ra backends.
Use of glInvalidateFramebuffer() falls away. I'd like to keep this, but
it's better to readd it once shader runs are in ra.
Also add some more helpers.
Fix the broken math.h include statement.
utils.c uses ra_gl.h internals, which it shouldn't, and which will be
removed again as soon as this code gets converted to ra fully.
The textures not having an FBO actually caused regressions when trying
to render the subtitles on top of this texture (--blend-subtitles),
which still relied on an FBO.
So just kill off the logic entirely. Why worry about a single FBO wasted
when we're allocating like 10 anyway.
Fixes#4657.
This is done via compute shaders. As a consequence, the tone mapping
algorithms had to be rewritten to compute their known constants in GLSL
(ahead of time), instead of doing it once. Didn't affect performance.
Using shmem/SSBO atomics in this way is extremely fast on nvidia, but it
might be slow on other platforms. Needs testing.
Unfortunately, setting up the SSBO still requires OpenGL calls, which
means I can't have it in video_shaders.c, where it belongs. But I'll
defer worrying about that until the backend refactor, since then I'll be
breaking up the video/video_shaders structure anyway.
These can either be invoked as dispatch_compute to do a single
computation, or finish_pass_fbo (after setting compute_size_minimum) to
render to a new texture using a compute shader. To make this stuff all
work transparently, we try really, really hard to make compute shaders
as identical to fragment shaders as possible in their behavior.
This removes VAO handling from video.c. Instead the shader cache will
create the VAO as needed. The consequence is that this creates a VAO
per shader, which might be a bit wasteful, but doesn't matter anyway.
Reduce this to 1 draw call per OSD pass. This removes the need for some
annoying special handling regarding 3D video support (we supported
duplicating the OSD/subtitles for side-by-side 3D output etc.).
Remove the unneeded texture sampler uniform thing.
Instead of allocating three PBOs and cycling through them, we allocate
one PBO that's three times as large, and cycle through the subregion
offsets.
This results in arguably simpler code and faster initialization
performance. Especially for 4K textures, initializing PBOs can take
quite some time (e.g. 180ms -> 110ms). For 1080p, it's more like 66ms ->
52ms for me.
The alignment to 4096 is completely unnecessary by spec, but we do it
anyway just for peace of mind.
This replaces `vo-performance` by `vo-passes`, bringing with it a number
of changes and improvements:
1. mpv users can now introspect the vo_opengl passes, which is something
that has been requested multiple times.
2. performance data is now measured per-pass, which helps both
development and debugging.
3. since adding more passes is cheap, we can now report information for
more passes (e.g. the blit pass, and the osd pass). Note: we also
switch to nanosecond scale, to be able to measure these passes
better.
4. `--user-shaders` authors can now describe their own passes, helping
users both identify which user shaders are active at any given time
as well as helping shader authors identify performance issues.
5. the timing data per pass is now exported as a full list of samples,
so projects like Argon-/mpv-stats can immediately read out all of the
samples and render a graph without having to manually poll this
option constantly.
Due to gl_timer's design being complicated (directly reading performance
data would block, so we delay the actual read-back until the next _start
command), it's vital not to conflate different passes that might be
doing different things from one frame to another. To accomplish this,
the actual timers are stored as part of the gl_shader_cache's sc_entry,
which makes them unique for that exact shader.
Starting and stopping the time measurement is easy to unify with the
gl_sc architecture, because the existing API already relies on a
"generate, render, reset" flow, so we can just put timer_start and
timer_stop in sc_generate and sc_reset, respectively.
The ugliest thing about this code is that due to the need to keep pass
information relatively stable in between frames, we need to distinguish
between "new" and "redrawn" frames, which bloats the code somewhat and
also feels hacky and vo_opengl-specific. (But then again, this entire
thing is vo_opengl-specific)
Mostly because of ANGLE (sadly).
The implementation became unpleasantly big, but at least it's relatively
self-contained.
I'm not sure to what degree shaders from different drivers are
compatible as in whether a driver would randomly misbehave if it's fed
a binary created by another driver. The useless binayFormat parameter
won't help it, as they can probably easily clash. As usual, OpenGL is
pretty shit here.
In some cases, such as when using the libmpv opengl-cb API, or with
certain vo_opengl backends, the main framebuffer is never accessed.
Instead, rendering is done to a FBO that acts as back buffer. This meant
an incorrect/broken bit depth could be used for dithering.
Change it to read the framebuffer depth lazily on the first render call.
Also move the main FBO field out of the GL struct to MPGLContext,
because the renderer's init function does not need to access it anymore.
Apparently we don't always set the viewport to window dimensions
anymore, e.g. if nothing is actually rendered. This means the viewport
can contain old values.
The window screenshot code uses the viewport values to guess the default
framebuffer dimensions. With --force-window --idle --no-osc (which draws
nothing and issues a glClear() command only), taking a screenshot would
yield an image with the wrong size and possibly garbage in it. Fix this
by explicitly passing the currently known window dimensions. Abusing the
values stored in the viewport was questionable anyway.
A minor cleanup that makes the code simpler, and guarantees that we
cleanup the GL state properly at any point.
We do this by reusing the uniform caching, and assigning each sampler
uniform its own texture unit by incrementing a counter. This has various
subtle consequences for the GL driver, which hopefully don't matter. For
example, it will bind fewer textures at a time, but also rebind them
more often.
For some reason we keep TEXUNIT_VIDEO_NUM, because it limits the number
of hook passes that can be bound at the same time.
OSD rendering is an exception: we do many passes with the same shader,
and rebinding the texture each pass. For now, this is handled in an
unclean way, and we make the shader cache reserve texture unit 0 for the
OSD texture. At a later point, we should allocate that one dynamically
too, and just pass the texture unit to the OSD rendering code. Right now
I feel like vo_rpi.c (may it rot in hell) is in the way.
The caller now has to call gl_sc_reset(), and _after_ rendering. This
way we can unset OpenGL state that was setup for rendering. This affects
the shader program, for example. The next commit uses this to
automatically manage texture units via the shader cache.
vo_rpi.c changes untested.
This is how PBOs are normally supposed to be used.
Unfortunately I can't see an any absolute improvement on nVidia binary
drivers and playing 4K material. Compared to the "old" PBO path with 1
buffer, the measured GL time decreases significantly, though.
This introduces a gl_pbo_upload_tex() function, which works almost like
our gl_upload_tex() glTexSubImage2D() wrapper, except it takes a struct
which caches the PBO handles. It also takes the full texture size (to
make allocating an ideal buffer size easier), and a parameter to disable
PBOs (so that the caller doesn't have to duplicate the gl_upload_tex()
call if PBOs are disabled or unavailable).
This also removes warnings and fallbacks on PBO failure. We just
silently try using PBOs on every frame, and if that fails at some point,
revert to normal texture uploads. Probably doesn't matter.
To avoid blocking the CPU, we use 8 time objects and rotate through
them, only blocking until the last possible moment (before we need
access to them on the next iteration through the ring buffer). I tested
it out on my machine and 4 query objects were enough to guarantee
block-free querying, but the extra margin shouldn't hurt.
Frame render times are just output at the end of each frame, via MP_DBG.
This might be improved in the future. (In particular, I want to expose
these numbers as properties so that users get some more visible feedback
about render times)
Currently, we measure pass_render_frame and pass_draw_to_screen
separately because the former might be called multiple times due to
interpolation. Doing it this way gives more faithful numbers. Same goes
for frame upload times.
Rename it to get out of OpenGL's namespace. The gl_ prefix is used by
other mpv functions, but no OpenGL ones.
The "slice" parameter was never actually used, and all callers passed 0
for it.
The main change is actually that e first copy to a "staging" memory
frame, and then upload this at once. The old non-PBO code called
glTexsubImage2D for each OSD sub-bitmap.
The new non-PBO code path is a bit faster now if there are many small
sub-bitmaps (on Linux/nVidia). It's also a bit simpler, so this is a
win.
(Although I don't particularly appreciate the mixed normal/PBO texture
code.)
Not sure how much can be gained with this, as we can't use it properly
yet. For now, this is used only before rendering, which probably does
overwhelmingly nothing.
In the future, this should be used after temporary passes, which could
possibly reduce memory usage and even memory bandwidth usage, depending
on the drivers.
This allows users to add their own near-arbitrary hooks to the vo_opengl
processing pipeline, greatly enhancing the flexibility of user shaders.
This enables, among other things, user shaders such as CrossBilateral,
SuperRes, LumaSharpen and many more.
To make parsing the user shaders easier, shaders are now loaded as
bstrs, and the hooks are set up during video reconfig instead of on
every single frame.
The hook mechanism allows arbitrary processing stages to get dispatched
whenever certain named textures have been "finalized" by the code.
This is mostly meant to serve as a change that opens up the internal
processing in pass_read_video to user scripts, but as a side benefit all
of the code dealing with offsets and plane alignment and other such
confusing things has been rewritten.
This hook mechanism is powerful enough to cover the needs of both
debanding and prescaling (and more), so as a result they can be removed
from pass_read_video entirely and implemented through hooks.
Some avenues for optimization:
- The prescale hook is currently somewhat distributed code-wise. It might be
cleaner to split it into superxbr and NNEDI3 hooks which can each be
self-contained.
- It might be possible to move a large part of the hook code out to an
external file (including the hook definitions for debanding and
prescaling), which would be very much desired.
- Currently, some stages (chroma merging, integer conversion) will
*always* run even if unnecessary. I'm planning another series of
refactors (deferred img_tex) to allow dropping unnecessary shader
stages like these, but that's probably some ways away. In the meantime
it would be doable to re-add some of the logic to skip these stages if
we know we don't need them.
- More hook locations could be added (?)
This merges all knowledge about texture format into a central table.
Most of the work done here is actually identifying which formats exactly
are supported by OpenGL(ES) under which circumstances, and keeping this
information in the format table in a somewhat declarative way. (Although
only to the extend needed by mpv.) In particular, ES and float formats
are a horrible mess.
Again this is a big refactor that might cause regression on "obscure"
configurations.
gl_transform_vec() assumed column-major, while everything else seemed to
assumed row-major memory organization for gl_transform.m. Also,
gl_transform_trans() seems to contain additional confusion.
This didn't matter until now, as everything has been orthogonal, this
the swapped matrix entries were always 0.
This is a pretty major rewrite of the internal texture binding
mechanic, which makes it more flexible.
In general, the difference between the old and current approaches is
that now, all texture description is held in a struct img_tex and only
explicitly bound with pass_bind. (Once bound, a texture unit is assumed
to be set in stone and no longer tied to the img_tex)
This approach makes the code inside pass_read_video significantly more
flexible and cuts down on the number of weird special cases and
spaghetti logic.
It also has some improvements, e.g. cutting down greatly on the number
of unnecessary conversion passes inside pass_read_video (which was
previously mostly done to cope with the fact that the alternative would
have resulted in a combinatorial explosion of code complexity).
Some other notable changes (and potential improvements):
- texture expansion is now *always* handled in pass_read_video, and the
colormatrix never does this anymore. (Which means the code could
probably be removed from the colormatrix generation logic, modulo some
other VOs)
- struct fbo_tex now stores both its "physical" and "logical"
(configured) size, which cuts down on the amount of width/height
baggage on some function calls
- vo_opengl can now technically support textures with different bit
depths (e.g. 10 bit luma, 8 bit chroma) - but the APIs it queries
inside img_format.c doesn't export this (nor does ffmpeg support it,
really) so the status quo of using the same tex_mul for all planes is
kept.
- dumb_mode is now only needed because of the indirect_fbo being in the
main rendering pipeline. If we reintroduce p->use_indirect and thread
a transform through the entire program this could be skipped where
unnecessary, allowing for the removal of dumb_mode. But I'm not sure
how to do this in a clean way. (Which is part of why it got introduced
to begin with)
- It would be trivial to resurrect source-shader now (it would just be
one extra 'if' inside pass_read_video).
GLES does not support high bit depth fixed point textures for unknown
reasons, so direct 10 bit input is not possible. But we can still use
integer textures, which are supported by GLES 3.0. These store integer
data just like the standard fixed point textures, except they are not
normalized on sampling. They also don't support bilinear filtering, and
require a special sampler ("usampler2D").
While these texture formats enable us to shuffle the data to the GPU,
they're rather impractical with the requirements mentioned above and our
current architecture. One problem is that most code assumes it can
always use bilinear scaling (even if bilinear is never used when using
appropriate scale/cscale options). Another is that we don't have any
concept of running a function on a texture in an uniform way.
So for now, run a simple conversion step through a FBO. The FBO will use
the rgba16f format normally, which gives enough bits for 10 bit, and
will at least gracefully degrade with higher depth input.
This is bound to be much slower than a more "direct" method, but at
least it works and is simple to implement.
The odd change of function call order in init_video() is to properly
disable "dumb mode" (no FBO use) if these texture formats are in use.
Do this to make the license situation less confusing.
This change should be of no consequence, since LGPL is compatible with
GPL anyway, and making it LGPL-only does not restrict the use with GPL
code.
Additionally, the wording implies that this is allowed, and that we can
just remove the GPL part.
Implement NNEDI3, a neural network based deinterlacer.
The shader is reimplemented in GLSL and supports both 8x4 and 8x6
sampling window now. This allows the shader to be licensed
under LGPL2.1 so that it can be used in mpv.
The current implementation supports uploading the NN weights (up to
51kb with placebo setting) in two different way, via uniform buffer
object or hard coding into shader source. UBO requires OpenGL 3.1,
which only guarantee 16kb per block. But I find that 64kb seems to be
a default setting for recent card/driver (which nnedi3 is targeting),
so I think we're fine here (with default nnedi3 setting the size of
weights is 9kb). Hard-coding into shader requires OpenGL 3.3, for the
"intBitsToFloat()" built-in function. This is necessary to precisely
represent these weights in GLSL. I tried several human readable
floating point number format (with really high precision as for
single precision float), but for some reason they are not working
nicely, bad pixels (with NaN value) could be produced with some
weights set.
We could also add support to upload these weights with texture, just
for compatibility reason (etc. upscaling a still image with a low end
graphics card). But as I tested, it's rather slow even with 1D
texture (we probably had to use 2D texture due to dimension size
limitation). Since there is always better choice to do NNEDI3
upscaling for still image (vapoursynth plugin), it's not implemented
in this commit. If this turns out to be a popular demand from the
user, it should be easy to add it later.
For those who wants to optimize the performance a bit further, the
bottleneck seems to be:
1. overhead to upload and access these weights, (in particular,
the shader code will be regenerated for each frame, it's on CPU
though).
2. "dot()" performance in the main loop.
3. "exp()" performance in the main loop, there are various fast
implementation with some bit tricks (probably with the help of the
intBitsToFloat function).
The code is tested with nvidia card and driver (355.11), on Linux.
Closes#2230
Add the Super-xBR filter for image doubling, and the prescaling framework
to support it.
The shader code was ported from MPDN extensions project, with
modification to process luma only.
This commit is largely inspired by code from #2266, with
`gl_transform_trans()` authored by @haasn taken directly.
Niklas Haas
wm4
Bin Jin
Niklas Haas
Bin Jin