* The existence of TW_ROTATION that implements this feature at the
level of calls to libpixelflinger API closely mirrors the existence of
ro.sf.hwrotation for surfaceflinger in LineageOS.
* A brute-force approach was previously attempted via the
BOARD_HAS_FLIPPED_SCREEN makefile flag. That code iterated over the
active display surface in a double-buffered setup, and performed a
"smart" memcpy from the UI drawing surface (gr_draw) onto the display
surface. The problem was that, without heavy loop optimizations, that
code could have never scaled for 90 and 270 degree rotation.
I tried and you could literally see the for loop with the naked eye
while the display surface was updating.
* That code is now gone, but support for BOARD_HAS_FLIPPED_SCREEN := true
is still there (now means TW_ROTATION := 180).
* This patch relies on the assumption that it is impossibly difficult
and non-portable to rotate whole framebuffer display surfaces, in a
way that is not dependent upon the graphics backend (adf, fbdev, drm,
overlay etc). Therefore, it identifies the rendering primitives that
the TWRP graphics stack exposes to the GUI application above, and
implements hwrotation inside each of those calls instead:
- gr_line(), gr_fill() - 2D geometric shapes (lines, rectangles)
- gr_blit() - graphical image resources
- gr_ttf_textExWH() - font rendering
- gr_fb_width(), gr_fb_height() - framebuffer resolution
* The gist is to keep the backend and framebuffer (dimensions, row size
etc) unchanged (because making changes there is asking for trouble),
but present an altogether different reality to the calling API,
according to the compile-time constant TW_ROTATION.
* All (x, y) API coordinates and shapes are transformed before being
actually rendered as (x_disp, y_disp) display coordinates.
* With TW_ROTATION := 90 or 270 you can turn a landscape device into
a portrait one, because the GUI is fooled by the reversed dimensions
reported by gr_fb_width() and gr_fb_height() and renders the UI as
for a different device.
* For blit and text rendering operations, figuring out the transformed
coordinates in display space is not enough, as the surfaces that are
to be rendered have to be rotated themselves. This is handled by
allocating an intermediary rotated surface on each rendering
operation (not ideal), so the code with the intermediary surface
is compiled out for the TW_ROTATION := 0 case.
* This is still not as bad as rotating the whole framebuffer though, and
on a msm8976 device the performance hit is not even noticeable (for
software rendering).
* Currently there is no attempt to make a connection between the
TW_ROTATION and the { RECOVERY_TOUCHSCREEN_SWAP_XY,
RECOVERY_TOUCHSCREEN_FLIP_X, RECOVERY_TOUCHSCREEN_FLIP_Y } settings.
Change-Id: Ic8966ad5360c8a499649fdb16e242286640fd992
Signed-off-by: Vladimir Oltean <olteanv@gmail.com>
Changes to fb_var_screeninfo can cause changes to
fb_fix_screeninfo. So, write back the force changes before fi is
read.
Change-Id: I721a960fa401ac5203ffc90bd3bfa2d7d0cfb292
On some devices TWRP tries to access the framebuffer before all
device nodes have been created. Retry opening instead of crashing later.
Change-Id: I189a8fe80a8906b46fb6cece53c0bf83c00c0e09
Fix: use row_bytes instead of xres (should help on Shield tablet)
Speed: Moving the calculations out of the inner loop
Change-Id: Ie43ae5e94ae88822360900c7b4d852b7aab4379b
While not all of these old build flags are active or implemented
yet (and they may not ever be implemeted until we find a device
that needs them), we need to make sure that the old build flags
are only applied when actually needed. Since there are a lot of
device trees that probably won't get updated due to lack of device
maintainers, we will rename the build flags and fix up devices on
a case by case basis.
Also added some pixel format related build flags to the make file
from AOSP so that AOSP devices should work in TWRP without
additional build flags.
Change-Id: I11ab475297d02b6aeffe89404fe50b4799a36be3
Note: events.cpp is still old code renamed to cpp to make it
easier to call functions like gr_fb_width().
I had to modify AOSP fbdev code to provide a separate memory
surface for drawing to as drawing directly to the framebuffer
resulted in rendering taking about 5 times longer.
I also modified AOSP adf code to provide a separate memory surface
for drawing for the same performance reasons. The Nexus 9 supports
adf graphics.
Overlay graphics work on at least one device. Overlay provides a
separate memory buffer already so performance is good.
I do not have a drm device yet that I know of. I made some attempt
to update the drm code to determine the correct pixel format based
on the drm graphics format, but what is available in pixel flinger
and what is available in drm do not line up all that well. Reports
are that the Pixel C is using drm graphics, but performance is
slow, likely due to the use of a mmap instead of a memory buffyer.
Change-Id: Ibd45bccca6ac2cb826037aa9b2aa5065cf683eed
