addition to evas gl backend.
Working on different mobile devices, we've noticed that the cost of context
switch (MakeCurrent) in GL can be very expensive depending on the driver
implementation. To minimize the poorly written driver's context switch
overhead, I've implemented a state tracking layer on top of the driver
implemented GL.
Essentially, this layer wraps all the GL/Glue(GLX/EGL) APIs and manages its own
state changes. Internally, only one real GL context is created and logical
contexts are created every time a user requests context creation. The logical
contexts keep track of its states and sets only the necessary states
(the ones that are different than the current ones)
when there is a MakeCurrent request. The real MakeCurrent gets called when
there is a Surface/Window change request.
The GL library is dlopened and all the APIs are dlsym'ed and wrapped
accordingly. All the GL functions are in evas_gl_core.{h,c}.
Here's a very simply flow of the code.
- all the APIs are exported as function pointers (*glsym_glBegin),
(*glsm_eglCreatContext), and etc.
- all the native GL/Glue(GLX/EGL) APIs are dlsym'ed as _sym_glBegin,
_sym_eglCreateContext, and etc.
- all the fastpath APIs are implmemnted as fpgl_glBegin,
fpgl_eglCreateContext, and etc.
- if faspath is seletected, the exported APIs are set accordingly
ie. glsym_glBegin = fpgl_glBegin;
- default mode is the regular gl symbols are directly set.
ie. glsym_glBegin = _sym_glBegin;
I have an Environment variable where you can set it to three different Modes
EVAS_GL_FASTPATH = 0 // Default mode. Regular GL symbols are directly loaded
EVAS_GL_FASTPATH = 1 // Fastpath mode. Takes the path described above.
EVAS_GL_FASTPATH = 2 // Wrapped mode. All the regular GL functions are
wrapped once. This can be used for various
purposes
Since all the GL symbols are now loaded in this library, I took out all
the gl symbol loading parts in the evas gl backend as you'll see in the patch.
The changes to the engine and the backend itself is pretty minor.
There are still some known issues to hammer out but I thought we're at a good
place for an initial version so that my source doesn't diverge too much.
Known Issues and To Do's
* Current GL Fastpath version doesn't support multiple threads. Instead of
having one global real context, I would need to do it for each thread. I'll
get on this soon.
* Issues running Evas GL on certain conditions. When running the elementary
test (with gl engine), if you run ELMGLview test that runs in ON_DEMAND mode,
everything works fine. BUT, when you run the ELMGLView test in ALWAYS
mode, the subsequent elm tests shows blank screen. When you destroy the
GLView window, everything else comes on fine.
* Resource protection code. This actually applies to Evas GL code in general
as well. Since all the resources are shared among all the contexts that get
created, I would like to eventually have a resource protecting mechanism that
prevents access to resources outside of its context unless specifically
specified.
I'm attaching three files
- evas_gl_core.h, evas_gl_core.c, fastpath.patch
To get the code running...
- copy evas_gl_core.{c,h} to src/modules/engine/gl_common/
- apply the fastpath.patch
- compile/install evas
- to run with fastpath GL (ie. % EVAS_GL_FASTPATH=1 ./evasgl_sample1)
SVN revision: 65891
Subject: Fix Neon build with Thumb-2
In assembly part of function evas_common_convert_rgba_to_32bpp_rgb_8888_rot_90:
Don't use 3-operand add instructions (e.g. add r2, r5, #2) as this is
not supported in unified syntax.
SVN revision: 65768
tweet: hey, SeoZ here, I really like Microsoft. I mean REALLY like them. I think they're a fantastic company and I hope everyone uses winmo!
SVN revision: 65469
quickly when uf your-encoded anything due to using IFAST decoder (and
encoder). this does drop speed for decode and encode (except encoding
< 60% quality where it now uses IFAST), but we don't see progressively
worse artifacts.
SVN revision: 65293
in most cases, it is performed twice inside and outside of the function in inefficient way.
and calling of events_frozen in the passes_through() is not understable in the view of functional consistency.
Check is separately would be better.
SVN revision: 65269
Subject: [E-devel] [Patch][Evas] Fix wrong location of
_evas_touch_point_remove()
I have a small patch to fix the wrong location of _evas_touch_point_remove().
_evas_touch_point_remove() should be called in the evas_event_feed_mouse_up(),
but it is called in the evas_event_feed_mouse_cancel() in the current code.
Would you apply attached patch?
SVN revision: 65005
There was a possible segfault because we don't check if the current item
is a text item or a format item. I just removed the loop which triggered it
because it's not needed anyway, and now it works. Removing the loop also
let me remove some code that was only needed in the case of a loop.
SVN revision: 64816
Fix multiple storage bug.
* __forceinline is the equivalent of __always_inline__ on Windows. It has
'extern' as storage, so static must not be used with it
* use __always_inline__ and not always_inline as attribute value instead.
No need to add storage class with __always_inline__ too.
* static inline is fine
SVN revision: 64767
Cedric BAIL
Carsten Haitzler
Mike Blumenkrantz
Jaehwan Kim
Tom Hacohen
Christopher Michael
Rafal Krypa
ChunEon Park
Vincent Torri
Mike McCormack
Jiyoun Park
EunMi Lee
Youness Alaoui