I saw some GL error messages (with the GLERR() macro activated),
that were caused by GL_INVALID_VALUE.
For some (many) textures, tex->y was 0 but since we now use a 2D
atlas (rectangle allocator), the first row of pixels should be
repeated. This caused uploads to Y = tex->y - 1 = -1, which is
invalid.
Now, the evas loader is supposed to advertise the actual border
size in case of compressed texture formats.
The only case where the border was non zero was ETC formats,
from the TGV loader, so I think we don't need to keep the
previous behaviour (auto-calculate borders for ETC).
Summary:
This patch introduce various new logic for packing/unpacking of Eina Rectangle in a pool.
It is then used by Evas GL backend texture allocation to improve how efficiently we pack
image in texture atlas. This lead to improved memory usage and reduced power consumption
with usually a more stable higher FPS (as it use less texture to do the same task, their
is less texture switch, so saving memory and speed at the same time).
This patch was developped on Cedric's suggestions to optimize the packing logic using Skyline
algorithm. This patch is based on master and is a new submission for earlier phab link
https://phab.enlightenment.org/D774.
Signed-off-by: Sanjay Nirankari <sanjay.n1@samsung.com>
Signed-off-by: Rajeev Ranjan <rajeev.r@samsung.com>
Signed-off-by: Sreedeep Moulik <sreedeep.m@samsung.com>
Reviewers: cedric, raster
CC: wonsik, jpeg, sreedeep.m, sanjay, govi
Differential Revision: https://phab.enlightenment.org/D1063
Signed-off-by: Cedric BAIL <c.bail@partner.samsung.com>
ETC2 textures were allocated using the same empty slots pool
as RGBA textures.
ETC1 texture atlasses should still be tested.
An extension string is required to avoid using the bad
fallback in place right now. But there is no such thing yet.
We prefer ETC2 textures when ETC2 support has been detected.
According to the spec, glCompressedTexSubImage2D should work
for ETC2.
Try even with ETC1. This may fail at runtime. The fallback path
is very dubious right now but without a proper test case I'm
not sure which approach to take.
We can also imagine cases where the GPU supports TexSubImage for
ETC1 but ETC2 is not supported at all. This will need testing, as
this case is not handled.
@feature
ETC1/2 textures are (for now) allocated one by one for each image,
because we use only glCompressedTexImage (not SubImage). So,
the texture height must fit that of its content.
This commit bypasses the usual pool allocation logic where textures
are rounded up to the next multiple of 16 pixels in height, in
case of ETC1/2.
Since the introduction of color spaces other than RGBA8888 in
the GL engines, there was an issue with border images scaled in
GL. The left and right edges were simply not properly copied.
This would then show artifacts when scaling very thin images
(typically 2px wide).
if alpha4 is possible (desktopgl) then use it for fonts as this should
cut memory in half for them and possibly speed things up due to less
memory bandwidth needed
NOTE: when using Evas_Object image preload infrastructure the GL texture
upload was uploaded from the main loop during the rendering stage. This
could lead to some frame drop during fast animation due to the time needed
to upload that texture.
This patch fix this problem by uploading a small texture quickly (16x16)
and waiting for going back to the main loop to be able to use the same GL
context from another thread to do the texture upload asynchronously without
blocking the main loop.
Basically, EGL does not define a GL_DRAW_FRAMEBUFFER_BINDING so we
test for that and if not found then define it to be
GL_FRAMEBUFFER_BINDING
Signed-off-by: Chris Michael <devilhorns@comcast.net>
(Bad antognolli, No cookie !!)
NB: Fixes build break !!! If we are going to use defines from glext.h,
let's be sure to have the necessary header included Yes ?? ;)
Signed-off-by: Chris Michael <devilhorns@comcast.net>
If we don't do this, the framebuffer object used on shader_array_flush()
will be wrong.
NOTE: shader_array_flush() should take care of setting the fbo
correctly.
downscaling. This makes quality much better and "at best"
equates to a 16 point sample (2x2 linear interpolation samples,
where a linear interpolation sample equates to a 2x2 sample).
This will have perfomance impact, but the quality is worth it and
makes it closer to software downscaling in quality. It supports
2x2, 2x1 and 1x2 oversampling. YUV not done, nor image mask
(font shaders not needed).