The dirty bit was a dirty hack to let session recovery force reconfigures
on startup.
Now that we have a surface flush we can achieve the same thing by just
discarding all buffers immediately.
For SW engine we need to verify that OSMesa is present. The patch
fb048e7312 broke the logic.
Tested by temporarily removing OSMesa from my system.
Fixes T6617 (again)
osmesa needs llvm. llvm apparently just by dlopening or linking to the
lib (libLLVM...) gets you 3.5mb of dirty pages just in this lib. that's
a whole lib entirely dirty pages. odd and horrible. in fact once i
stoppd dlopening OSMesa all the time on engine init (and only when gl
is needed)... the amount of dirty pages went from 17208 to 8860.
that's a whopping drop of 8mb! 8mb saved! in fact just dlopening
osmesa and doing the other gl init stuff led to more anonymuse
mappings with dirty pages. 2 of them (2072k and 2076k) which baffled
me as that didn't seem like heap or efl's own data. these disappeared
along with libLLVM-5.0.so (3520k + 60k dirty pages). we stopped
linking/loading libedit (12k dirty), libglapi (20k dirty),
libLLVM-5.0 (3580k dirty), libncursesw (72k dirty),
libOSMesa.so (260k dirty), libtinfo (20k dirty). ... or at least
stopped until absolutely needed. total 17208k of dirty pages went down
to 8860.
my test case was just launching terminology (and doing nothing with it).
@fix memory bloating
Summary:
The default backend overrides this operation depending on the fill color
but the cairo backend dosen't hence cairo will always use bled mode while drwaing the vector.
Reviewers: jpeg
Subscribers: vtorri, cedric, jpeg
Differential Revision: https://phab.enlightenment.org/D5724
This is very useful to specify precisely which kind of RGBA -> Alpha
conversion you want. If all you wanted was the alpha layer to use as a
mask, set this flag to true.
@feature
Summary:
The Encoding key is no longer required, all desktop files are assumed to
be UTF-8 encoded. See details at:
https://standards.freedesktop.org/desktop-entry-spec/1.1/apc.html
Fix various typos and misspellings
lintian, Debian's package checker, uses strings to check for common typos
in compiled binaries. This change fixes the ones it identified in 1.20.6.
Reviewers: cedric
Reviewed By: cedric
Subscribers: cedric, jpeg
Differential Revision: https://phab.enlightenment.org/D5584
Signed-off-by: Cedric BAIL <cedric@osg.samsung.com>
This small patch just checks that we have a valid surface and bo that
we can pass to gbm_surface_release_buffer. If they are not valid, this
causes a hard crash.
NB: This does not actually Fix the ticket issue....
ref T6483
Signed-off-by: Chris Michael <cp.michael@samsung.com>
If we want to share a gl context (we do) between multiple instances of
gl_drm, we need to make sure they all use the same gbm_device.
This resolves a blocker for multi-output on the gl_drm backend.
Multi-head is hitting corner cases where there are lots of locked buffers
and it looks like right now 5 is the magic number that makes the problem
go away.
Make it possible to set 5 or more (via env var) for testing, make a macro
for MAX_BUFFERS instead of just a number.
We no longer allocate 3 buffers at startup, we now allocate only as needed.
Trimming the queue will come later, as there are some situations where we
might need 3 buffers and later drop down to 2 (when on a hardware plane)
Most clients will only ever need 2 buffers, so this is a reasonable RAM
savings.
It does us no good to be able to allocate dmabuf capable memory if the
compositor can't handle it. This should fix failures on systems where
allocation is possible but the compositor doesn't advertise dmabuf.
This moves all the platform specific buffer allocation into ecore_wl2
instead of the engine.
Note that this makes an internal struct available in the header. This
will be removed shortly.
Currently the buffer code looks up the alpha stats from the surface code.
This won't be possible when we move the buffer code into a library, so
prepare for it now.
The new library function provides the same functionality and will allow
us to stop tracking things in the engine that the library already knows
about windows (compositor_version)
The dmabuf code has been creating ARGB buffers all the time. The old
wl_shm code correctly respected the alpha field, but that was lost in
the new version.
I'm not sure we ever create non-alpha wayland buffers, since CSD has
to have alpha to do shadows, but if there's any way to do it it'll
work now.
This is what the old shm code has been doing, and it's probably better
than what the dmabuf code was doing.
We currently allocate 3 buffers. The usual case has us swapping between
two of those buffers and saving that third buffer for emergencies - if
we ever need that third buffer it'll require a full redraw.
If we return the oldest available buffer the usual case requires a little
more damage but we should never hit the full redraw case, which can cause
a frame drop on slow hardware.
Now that we're dependent on create_immed there's no possibility of falling
back to non dmabuf allocation.
The only failing case we really need to handle is failing the first
allocation, which is currently broken and I'll be adding an advance test
for it shortly.
wl_shm and dmabuf only really need to differ in how they allocate a buffer,
but right now we've got them in separate files. This dramatically
reduces the complexity of the wl_shm code and shares much more
implementation with the dmabuf code.
This throws away at least one "optimization" wl_shm used - over-allocating
buffers so that window resizing doesn't always require a new buffer
allocation. If people feel that window resizing has become too slow now
this can be added to the dmabuf code to the benefit of both allocators.
Disabling dmabuf by env var still uses the old wl_shm implementation for
now, but soon that code will be removed entirely and the env var disable
will use this path.
Summary:
Refactoring.
It is good to store values from that struct in a parsing/loading context
static variable is a big NO NO:
1. Ugly code design,
2. Might not work when trying to load more than one SVG file.
@fix
Reviewers: jpeg, smohanty
Subscribers: jenkins, cedric
Differential Revision: https://phab.enlightenment.org/D5399
Accodring to https://www.w3.org/TR/SVG/types.html#Length
length ::= number ("em" | "ex" | "px" | "in" | "cm" | "mm" | "pt" | "pc" | "%")
This is still work in progress since some of lengths should be treated
differently, for example gradient lengths
Just as a starter to make a working background that, later on, will go
through Svg_Node's and build a certain source code to be saved in SVG
picture as a file
Coverity reports that _evas_dmabuf_buffer_init function here can
potentially free the surface that was passed into it. If that happens,
we should not be calling the _fallback function with surface as the
parameter as that will directly dereference the freed pointer.
Fixes Coverity CID1381707
@fix
Signed-off-by: Chris Michael <cp.michael@samsung.com>
In most engines which inherit from software_generic, they do not
implement the outbuf_free_region_for_update function. Most engines
have it as an unused function. If we simply add a check here, then we
can reduce the need for having useless function in multiple engines.
@fix
Signed-off-by: Chris Michael <cp.michael@samsung.com>
If the image has no data, it may get an allocated surface of 1x1 but it
is not sane to return the pointer to that data, as the user would expect
a normally sized image (in my case, 1920x1080).
I do not fully understand what is going on with this image. But at least
this transforms a crash into a simple ERR in ~/.xessions-errors
Two similar crashes happened:
- SIGSEGV by writing data outside of the image data
- abort() in free() because the malloc metadata has been overridden
when writing outside of the image data (newly allocated 1x1).
Fixes T5957
@fix
this is normal - brute force trying loaders until one succeeds is
normal is etn doesnt help identify it or it fails the first
guess-by-extension. printing errors is not good as this is an ok and
EXPECTED error. slience!
@fix
We've really always needed to do immediate creates. On a surface resize
there's no place to wait for the round trip for the new buffers to exist.
We've gotten away with this until now by good luck because we dispatch
wayland events during asynchronous render. However, with async render off
or if schedule happens in an unfortunate order, we can end up with
tearing.
Outbuf shouldn't have to track its hidden status, that should be ecore_evas
problem. Until now we were doing this because our kludgey wayland
ticking made things difficult, but I think it's safe to remove now.
Coverity reports passing a null pointer 'im->gc' to
evas_gl_common_context_flush which directly dereferences it, so lets
be sure that 'im->gc' is valid before passing it to context_flush
Fixes CID1374273
@fix
Signed-off-by: Chris Michael <cp.michael@samsung.com>
Coverity reports that there may be a null pointer dereference here so
check that 'error' exists before trying to set it.
Fixes CID1374272
@fix
Signed-off-by: Chris Michael <cp.michael@samsung.com>
Summary:
When evas selects a strike of embedded bitmap font,
calculate ratio and use it for scaling embedded bitmap.
@feature
Reviewers: jpeg, tasn, woohyun, raster, herdsman
Reviewed By: raster
Subscribers: charlesmilette, Francesco149, cedric
Differential Revision: https://phab.enlightenment.org/D2713
Small patch to fix Coverity reported issues of uninitialized variables
Fixes CID1381306, CID1381305, CID1381304, CID1381303
Signed-off-by: Chris Michael <cp.michael@samsung.com>
This patch adds support for software rotation in the evas drm engine.
This is a fallback codepath in case hardware rotation is not supported
for a given rotation amount. This patch also fixes a leak of and
pending updates during output buffer free.
ref T5999
Signed-off-by: Chris Michael <cp.michael@samsung.com>
This patch provides an override in the evas drm engine for the output
resize function. We override this function so that we can reconfigure
the output buffer.
ref T5999
Signed-off-by: Chris Michael <cp.michael@samsung.com>
enlightenment internal windows insta segv e on rpi. after much hunting
it seems a fallback is happening and bunk ptrs are being used. this at
least will make the problems more reliable with null ptrs.
Turns out the "device_open" function pretty much just tests calloc
functionality, and doesn't open any device. So let's allocate a
tiny bo and discard it to make sure we're actually on exynos.
Summary:
Function argument was renamed, but in function body still uses old
variable name.
Test Plan: Build on Windows host
Reviewers: cedric, vtorri
Reviewed By: vtorri
Subscribers: jpeg
Tags: #windows, #efl
Differential Revision: https://phab.enlightenment.org/D5152
We had a hack in place to flush the display from an idle enterer instead
of after a surface commit. This led to a problem where the idle
enterer dispatch order was:
renderer for main canvas
wayland dispatch idle enterer
renderer for mouse cursor canvas
The surface commit for the mouse cursor was never dispatched, so the mouse
cursor animation would only update at the rate other events occurred.
By flushing at the appropriate times instead we ensure a proper update.
ref T5850
gl_generic_context_find() returns the gl shared context struct but
this is not just a read-only operation. It in turn calls window_use
which may call make_current. This can invalidate the work of evas gl
when the API tried to switch to a specific context.
This fixes evas gl with multiple outputs.
Signed-off-by: Cedric BAIL <cedric@osg.samsung.com>
This backend has received no patch and maintenance from anyone who could
actually test it over the last few years. After talking with KaKaRoTo it
is best to remove it. If anyone want to take over its maintenance, you
are welcome to revert this patch.
Since the EO APIs are defined as weak symbols, invalid definitions of
EAPI lead to runtime crashes on non-public APIs. This is a fix following
a series of changes wrt. EAPI definitions.
Intended to simplify the upcoming commit that merges device find and
device open into a single function that returns a device.
The fd is something callers shouldn't really need to get their hands on,
right now there are still a few places where it's needed, but those will
be gone soon too.
one if condition is always true by virtual of previous if statements
and drop-through so can remove. not actually any bug but analysers
don't like it
found by PVS studio
only the else changes finfo so reset inside there. not really any bug
at all byt style-wise a bit better and analysers don't like it
found by PVS studio
so a type we handle earlir inan if we re-handle as invalid later. this
wouldnt lead to a crash or bugs as the if's would ned to be evaluated
in order normally, but it's good to get it right.
found by PVS studio
Just because the #define is present doesn't mean the extension is, so we're
BAILing on egl completely on some systems for no good reason at all. (saw
this on an SGX stack)
This is still wrong. I don't want to try too hard until after the upcoming
release, though.
We should actually be testing for the presence of client extensions before
attempting to do any of this. It's entirely possible that a gl stack will
return bogus functions for these from eglGetProcAddress
Ooops! When mapping for writing we can't use the same code
path as when tofree is true. This restores the path for image
writing. This basically fixes rage with GL engine.
See also 45c8e5e983
A recent commit broke texture_from_pixmap for NVIDIA EGL
(again), because eglCreateImage is a symbol in libEGL.so
but isn't in fact implemented by the driver.
That's because eglCreateImage() is exposed by libglvnd but
the underlying EGL implementation is NVIDIA and its version
is only 1.4, not 1.5 (where the API was introduced as core).
Instead of reverting the patch, it's better to cover our
bases properly and use dlsym() only if the version is right.
Note that GetProcAddress() may return garbage function
pointers for ALL functions as dynamic virtual functions may
be created on the fly by libglvnd. So it is absolutely
necessary to check the extension string as well.
See 0255f14dc2
eglCreateImage is objectively better than eglCreateImageKHR - it allows
attributes large enough to hold pointer values. We should use it when
available and only use the older extension version as fallback.
Also, eglCreateImage is core EGL functionality so don't depend on extensions
to be present to use it. Theoretically we should be testing for
EGL version >= 1.5 but it's probably safe not to.
This was only necessary due to bugs in the wayland_egl and gl_drm engine
that have been corrected.
Wayland has no bizarre requirements making this necessary.
Commit 2e6587a14b adds a gl extension string to glsym_evas_gl_symbols()
to prevent using functions that are provided by extensions that aren't
available in the current gl context.
However, we can't query the extension strings until after we create an egl
context.
Split the regular symbol lookup stuff from the gl symbol lookup stuff so
we can query at the appropriate time.
These aren't a thing. the GL_OES_EGL_image_base extension doesn't exist
and the GL_OES_EGL_image extension doesn't extend egl, because it's a gl
extension.
So let's stop being the only google match for these nonsense strings.
this fixes an issue that has cropped up in the past few months - only
nvidia drivers with egl/gles in x11... and compositing won't work
(native surface) and the introduction of libglvnd
it's a combination of libglvnd lying that it has symbols it can't
later find, new features to get core functions via procaddress that we
hadn't migrated to use AND use preferring core functions that libglvnd
will expose, so switching to KHR extensions by preference. we also
need to symmetrically use destroy image khr too...
oddly enough using procaddress purely for create/destroy image makes
wayland fail ... sofor now i'm taking advantage of the fact that
wayland has no extensions string passed in at the moment and still
doing dlsym... this is odd though.
@fix
Summary:
Summary : String buffer returned by eina_strbuf_new() is not freed in some cases
@Fix
Signed-off-by: Uma Devika <u.bodapati@samsung.com>
Reviewers: cedric, tasn, jpeg, raster, singh.amitesh
Subscribers: tanwar.umesh07, yashu21985, cedric, jpeg
Differential Revision: https://phab.enlightenment.org/D5000
This has been a stub for a long time, but now we have a native_set, so
we should have a native_get too.
This is required for hw plane usage on the software engine.
Functions to assign a plane for a native surface, and release a plane
that's been assigned to a native surface.
These are empty for now as they'll need to be overridden in any backend
that can handle planes.
so modern systems seem to have abandoned rgb.txt files. this leads to
us breaking the loading of xpm files tha use color names ... i added
the rgbt.txt from vim but that didn't seem to help... odd... so to just
stop adding path after path to load... ship our own. we could ship the
file... but then we'd still have to load and parse it... every time we
look up a color. so i munged the data with awk and now we compile it
in. it should consume the same space the rgb.txt does in the shared lib
binary. if not read it shouldn't be paged in. it should end up cheaper
than our floaing of the file and mmaping it when xpm module is
loaded/initted... so either way more efficient, uses a little less ram
(12306 bytes vs 17780 for the rgb.txt) ... and bonus - dont have
an extenral out-of-code data blob to find, manage install etc...
this means we should load xpms with colornames correctly again on
systems without an rgb.txt provided by x11 ... which seems to be the
common case now. :(
@fix
2 more. /etc/X11/rgb.txt /usr/share/vim/vim80/rgb.txt ...what i do notice
is that this file seems to have vanished from modern systems... so we'll have
lots of un-fun loading old xpm's with colornames if we cant figure out what
color names map to what colors...
If/When we exit _evas_outbuf_egl_setup function, we should be freeing
the allocated cfgs variable else we leak it.
@fix
Signed-off-by: Chris Michael <cp.michael@samsung.com>
Set this env var to "300 es" to test GLSL 300 ES as shader
version. This is for brokenshakles.
Example:
export EVAS_GL_GET_PROGRAM_BINARY=0
export EVAS_GL_SHADER_GLSL_VERSION="300 es"
export ELM_ACCEL=gl
elementary_test
Removes the previous "busy" flag, as now we might have an fb attached to
multiple outputs at once, and need to be careful to destroy them only
after they've been removed from all outputs.
Removed the old "busy_set" API which nothing used, and renames fb_destroy
to fb_discard to make it more clear that it's not immediately destroyed.
It's all beta api, so I can do this.
Omg... Thanks Daekwang Ryu for pointing me to my error. I remember
struggling a lot with this OpenGL API and libGLdispatch (glvnd) when
in fact this was all just a typo in the code.
GLES 3.1 and the upcoming 3.2 support need a proper test case...
See c68a409874
@fix
Somehow this long standing bug wasn't obvious until wayland 1.13.0 made
some additions to an opaque structure.
This changed the frequency that new buffers came to us with the exact
same pointer value of a buffer that had just been freed.
This shortcut in eng_image_native_set has always been wrong - we need to
proceed to the end to make sure we pick up new dmabuf attributes.
Summary:
Evas can't open tiff file because of no implement in client read api.
I wrote codes simply for open.
Test Plan: self
Reviewers: jpeg, cedric, jypark
Subscribers: stefan_schmidt
Differential Revision: https://phab.enlightenment.org/D4857
This is the first step toward handling multi output. This patch
remove engine.data.output from Evas structure and use an Eina_List
for it instead. It also start moving code around to fetch an output
or an engine context (which are the same at the moment, but will be
split in a later patch).
This might not be used as over two consecutive runs all the
same buffers should be used. But it could happen if some
parameters in the filter change (eg. blur radius).
Fixes major (GPU) memory leaks. Reuse mode is still leaking.
An odd-sized image scaled down by 2 was losing 1 pixel during the
downscale, and it was not restored after scaling up. The same
happened with downscaling by 4 except the effect was even more
visible.
This meant that a moving snapshot with a large blur would trigger
some really ugly sampling issues if the content below was precise
(such a text).
This dramatically improves the performance and now seems
to give acceptable results. Eventually we need a quality flag
in order to enable this or not. Alternatively, "gaussian" blur
mode would skip this optimization, while "default" would trigger
it.
This can help with performance when a large region of the
filtered image (eg. snapshot) is fully hidden by an opaque
object. For instance the window border is hidden by the
opaque window content.
This make save() work on snapshot objects, provided the call
is done from inside render_post.
Also, this saves the filtered output of an image, rather than
its source pixels. Any call to save() on a filtered image must
be done from post-render as well.
Fixes T2102
@feature
If we delete the image that was the target surface for gl
rendering, a crash would occur on the next render cycle.
Unlikely but not impossible to trigger from app side.
@fix
This was a poor attempt at improving the performance but
obviously the root cause isn't fixed (too many texel fetches).
Uniform should (theoretically) work better than an attribute
the for loop. Just a guess here.
This also makes GL blur use a float value as radius, allowing
future extension to non-integer blur radii, as well as using
linear scaling as a fast blur approximation.
This optimizes the GL blur algorithm by reducing the number of
texel fetches (roughly half the number of before this patch). This
works by exploiting GL's interpolation capabilities.
By simply splitting X and Y blurs in two passes we can improve
the performance of the blur filter a lot.
There is still much to be done to make it really fast and nice
looking:
- implement true gaussian blur (not sine-based approximation,
right now the actual blurs look different in SW and GL)
- exploit linear interpolation for R tap instead of R*2+1 taps
(a tap being a texel fetch)
- downscale & upscale large images with large blur radii
Wait a second though, this implementation is not only incomplete
(no support for box vs. gaussian blur), it's also insanely bad in
terms of performance. Small radii may work fine, but at least blurs
render properly in GL with this patch (no more glReadPixels!).
The shader needs a lot of love, including in particular:
- support for 1D box blur single pass
- support for 1D gaussian (or sine) blur
- use linear interpolation and N-tap filters
- separation of 2D blur in two passes (high-level logic)
- potentially separation of large 1D blurs in 2 or more passes
knowing that 2sigma == sigma + sigma when it comes to the gaussian
bell curve.
This one was a bit more... "fun". I had to add a new vertex
attribute and obviously using a VertexAttribPointer led to
incomprehensible crashes. But a simple glVertexAttrib2fv makes
it work like a charm!
A rare option is not handled yet.
This reuses the existing mask infrastructure, but adds a color
flag to use the whole RGBA range, rather than just the Alpha
channel.
Filters are still very slow (glReadPixels and non-optimized use of
GL buffers...), but this is progress :)
This corrects two things:
- the blur filter high-level logic, that lead to reusing some
temporary buffers which contained garbage;
- the versatile gl buffer implementation so that it now properly
switches between the RGBA_Image and the FBO content (yes, this
is insanely slow and inefficient... but it works and that was
the only point).
Alright, so this is a massive patch that is the result of
trying to get rid of unused or poorly implemented classes in
ector. Originally ector was meant to support VG but extend to
things like filters as well. At the moment, ector's design
makes it quite hard to plug in the filters.
For now I think it's easier to implement the GL support for
the filters directly in the engine, where I hope to interfere
as little as possible.
This massive patch keeps only the required minimum to support
a versatile gl buffer that can be mapped, drawn or rendered to (FBO).
It's extremely inefficient as it relies on glReadPixels and lots
of texture uploads, as well as conversions between ARGB and Alpha.
Another type of GL buffer is a wrap around an existing GL image,
but that one is read-only (map or draw: no write map, no FBO).
No, all the filters run fine, and the high-level implementation
(evas_filters.c) does not need to know whether the underlying engine
is SW or GL. One problem though appears with the blending or blurring
of some Alpha buffers, the colors are wrong.
This patch removes more lines than it adds so it must be good ;)