GitLab

The mali r3p2 blob does not seem to be affected by the window
resize bug anymore, so there is no need to apply the r3p0
workaround (and it does not even work correctly).

When using the hardware overlay, now we need to allocate secure id 1
and 2 for back/front ump buffers. Otherwise the blob is trying to
request a new dri2 buffer for each frame, and this kills performance
(thanks to redundant ump buffers mapping/unmapping). This is now
fixed.

But when not using the hardware overlay, the r3p2 blob is still
requesting dri2 buffer per each frame, and this needs an additional
tweak.

This avoids a kernel oops due to the badly implemented and badly
checked ump interface.
Signed-off-by: Luc Verhaegen <libv@skynet.be>

The binary driver is unaffected by it, only when mesa-dri is fully
installed does it do something.
Signed-off-by: Luc Verhaegen <libv@skynet.be>

Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Marvell PJ4 core used in CuBox very poorly handles VFP uncached
reads from the framebuffer. Using WMMX or ARM LDM reads is much
faster, with LDM instructions having a minor advantage. This
improves framebuffer read performance from ~50MB/s to ~100MB/s.

WMMX runtime detection and PJ4 core identification is also added
as part of this fix.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Benchmarking with x11perf, modified to support wider range of sizes
for the scroll operation. Tests have been run at the stock 700MHz CPU
clock frequency and with 1280x720 32bpp desktop.

$ DISPLAY=:0 ./x11perf -scroll5 -scroll10 -scroll15 -scroll20 \
                       -scroll30 -scroll50 -scroll100

== CPU ==

1000000 trep @   0.0289 msec ( 34600.0/sec): Scroll 5x5 pixels
1000000 trep @   0.0387 msec ( 25800.0/sec): Scroll 10x10 pixels
1000000 trep @   0.0459 msec ( 21800.0/sec): Scroll 15x15 pixels
 450000 trep @   0.0576 msec ( 17300.0/sec): Scroll 20x20 pixels
 350000 trep @   0.0817 msec ( 12200.0/sec): Scroll 30x30 pixels
 200000 trep @   0.1564 msec (  6390.0/sec): Scroll 50x50 pixels
 100000 trep @   0.4446 msec (  2250.0/sec): Scroll 100x100 pixels

== fb_copyarea (DMA) acceleration ==

1000000 trep @   0.0307 msec ( 32500.0/sec): Scroll 5x5 pixels
1000000 trep @   0.0353 msec ( 28300.0/sec): Scroll 10x10 pixels
1000000 trep @   0.0397 msec ( 25200.0/sec): Scroll 15x15 pixels
1000000 trep @   0.0464 msec ( 21600.0/sec): Scroll 20x20 pixels
 400000 trep @   0.0645 msec ( 15500.0/sec): Scroll 30x30 pixels
 250000 trep @   0.1177 msec (  8500.0/sec): Scroll 50x50 pixels
 100000 trep @   0.2783 msec (  3590.0/sec): Scroll 100x100 pixels

This shows that the ioctls overhead and the DMA setup cost is not so
significant for the Raspberry Pi. DMA already becomes a bit faster
than CPU at 10x10 size of the blit operation.

Even though there is no significant difference between CPU and DMA
for extremely small sizes of operations (the other overhead is clearly
dominating), setting a threshold is not going to harm:

== mixed CPU / fb_copyarea (DMA) with 90 pixels threshold ==

1000000 trep @   0.0291 msec ( 34300.0/sec): Scroll 5x5 pixels
1000000 trep @   0.0345 msec ( 29000.0/sec): Scroll 10x10 pixels
1000000 trep @   0.0395 msec ( 25300.0/sec): Scroll 15x15 pixels
1000000 trep @   0.0466 msec ( 21400.0/sec): Scroll 20x20 pixels
 400000 trep @   0.0650 msec ( 15400.0/sec): Scroll 30x30 pixels
 250000 trep @   0.1181 msec (  8470.0/sec): Scroll 50x50 pixels
 100000 trep @   0.2784 msec (  3590.0/sec): Scroll 100x100 pixels

If some other ARM devices also implement Raspberry Pi compatible
accelerated fb_copyarea ioctl, then the threshold selection may
be reconsidered.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Now acceleration is only used in the case if the AccelMethod option
is not set (so that it is assumed to be a default choice) or when
it is explicitly set to "COPYAREA". Any other value (for example
"CPU") disables acceleration.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

This provides basic 2D acceleration support for Raspberry Pi
to speed up moving windows and scrolling.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Because a wide range of embedded ARM devices are actually supported
(Allwinner A1X/A20, Raspberry Pi, ODROID-X, Rockchip, ...) and
are getting some sort of performance improvement and/or hardware
acceleration, the DDX driver needs a vendor neutral name.

Resolves https://github.com/ssvb/xf86-video-fbturbo/issues/10

Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

In the case if the framebuffer reservation size is too small for
efficient use of the hardware overlays and zero-copy buffers flipping,
log a hint about fixing this problem in /var/log/Xorg.0.log
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Even though we are primarily using the UMP buffer obtained by the
GET_UMP_SECURE_ID_SUNXI_FB ioctl, another UMP buffer obtained by
the GET_UMP_SECURE_ID_BUF1 ioctl should also span over the whole
framebuffer. Otherwise we may have troubles with the window resize
bug recovery and buffer flipping.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

The Allwinner A10/A13 display controller hardware is expected to
support negative coordinates of the top left corners of the layers.
But there is some bug either in the kernel driver or in the hardware,
which messes up the picture on screen when the Y coordinate is negative
for YUV layer. Negative X coordinates are not affected. RGB formats
are not affected too (no matter whether the RGB layer is scaled or not).

We fix this by just recalculating which part of the buffer in memory
corresponds to Y=0 on screen and adjust the input buffer settings.

Fixes https://github.com/ssvb/xf86-video-sunxifb/issues/16

Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Now zero copy and tear free buffer swapping is also supported
for 16bpp desktop.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Now the scaler is enabled for the sunxi disp layer only when we want
to use it for YUV format with XV. Whenever the layer is configured
for RGB format or deactivated, the scaler gets disabled.

This should make the driver more friendly to the other potential
scaled layer users. The total number of available scalers is only
2 for Allwinner A10 and only 1 for Allwinner A13.

The potential drawback is that now we may get an error when trying
to enable the scaler (if somebody else has used up all the available
scalers) instead of always having it reserved and ready for use.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Recent changes broke the configuration when "DRI2HWOverlay" option
is set to "false". This patch adds the missing UMP secure ids
initialization and resolves the problem.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Do this to keep the variables naming style consistent across the
source file (earlier these variables had different names like
'self', 'drvpriv', 'private').
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

In double buffer mode, explicitly mark the buffers as designated
for odd or even frame position when putting them into queue. And
when swapping the buffers, use these flags to re-synchronize if
it is necessary. This prevents problems after window resize (when
gles-rgb-cycle-demo could expose a mismatch between the color
name in the window title and the actual window color).
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

If DEBUG_WITH_RGB_PATTERN is defined, then we check that the
frames colors are changed as "R -> G -> B -> R -> G -> ..."
pattern and print debugging messages when this is not the
case. Such color change pattern can be generated by the
"test/gles-rgb-cycle-demo.c" program.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Do this mostly for security reasons. We don't want any application
to see whatever was last rendered by the previous GLES application
by just peeking into a freshly allocated DRI2 buffer.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

We manage only a single hardware overlay. That's a precious shared
resource, which we want to use for zero-copy fullscreen compositing
in gnome-shell. The strange 1x1 window does not really need it.
Fixes https://github.com/ssvb/xf86-video-sunxifb/issues/2

Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

When enabled, it tries to avoid tearing in OpenGL ES applications.
Works on sunxi hardware in the case if the hardware overlay (sunxi
disp layer) is used for a DRI2 window. The name of this option and
the description in the man page has been borrowed from intel and
radeon drivers.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

That's the right thing to do and fixes issues such as
    https://github.com/ssvb/xf86-video-sunxifb/issues/6



As a result, now the framebuffer size may need to be larger in
order to accomodate two DRI2 buffers in the offscreen part of
the framebuffer. The users of sunxi hardware are advised to
increase the value of fb0_framebuffer_num variable in fex file
to 3 for 32bpp mode and to 5 for 16bpp mode.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Should fix https://github.com/ssvb/xf86-video-sunxifb/issues/14


and prevent FTBFS on some systems.
Reported-by: Fred Chien <cfsghost@gmail.com>
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

When moving further to our own DRI2 buffers bookkeeping, we can't
really trust the information from DRI2BufferRec anymore. So just add
a copy of all the missing bits of information to UMPBufferInfoRec
and use it instead.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

The recent commit 9e0a8731

 (its part
that suppressed buffers reuse in the Xorg DRI2 framework) introduced
a regression. Half of the frames stoppped reaching the screen on
the CPU copy fallback path because the Mali blob now ended up
rendering them to the "wrong" buffer.

It just confirms that we need to completely move from the standard
DRI2 framework in the Xorg server to our own buffers bookkeeping
logic. This patch fixes the regression by introducing a single UMP
buffer per window, which is shared between back and front DRI2
buffers. We can do this because double buffering does not make much
sense on the fallback path at the moment (we can't set scanout from
this buffer and anyway have to copy this data elsewhere immediately
after we get it from Mali).
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Bail out earlier for the uninteresting types of DRI2 buffer
requests (by just returning a dummy null UMP buffer). Makes
the code a bit more simple on the common path.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Using the secure id 1 (framebuffer) to trick the Mali blob into
requesting DRI2 buffers again was not a very good idea. The problem
is that the blob still writes something there and corrupts the
framebuffer. So instead we try to assign secure id 2 to a dummy
4KiB UMP buffer allocated in memory and use it for the same purpose.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

The Mali blob is doing something like this:

 1. Request BackLeft DRI2 buffer (buffer A) and render to it
 2. Swap buffers
 3. Request BackLeft DRI2 buffer (buffer B)
 4. Check window size, and if it has changed - go back to step 1.
 5. Render to the current back buffer (either buffer A or B)
 6. Swap buffers
 7. Go back to step 4

A very serious show stopper problem is that the Mali blob ignores
DRI2-InvalidateBuffers events and just uses GetGeometry polling
to check whether the window size has changed. Unfortunately this
is racy and we may end up with a size mismatch between buffer A
and buffer B. This is particularly easy to trigger when the window
size changes exactly between steps 1 and 3.

See test/gles-yellow-blue-flip.c program which demonstrates this.
Qt5 applications also trigger this bug.

We workaround the issue by explicitly tracking the requests for
BackLeft buffers and checking whether the sizes of these buffers
match at step 1 and step 3. However the real challenge here is
notifying the client application that these buffers are no good,
so that it can request them again. As DRI2-InvalidateBuffers
events are ignored, we are in a pretty difficult situation.
Fortunately I remembered a weird behaviour observed earlier:

    https://groups.google.com/forum/#!msg/linux-sunxi/qnxpVaqp1Ys/aVTq09DVih0J



Actually if we return UMP secure ID value 1 for the second DRI2
buffer request, the blob responds to this by spitting out the
following error message:

    [EGL-X11] [2274] DETECTED ONLY ONE FRAMEBUFFER - FORCING A RESIZE
    [EGL-X11] [2274] DRI2 UMP ID 0x3 retrieved
    [EGL-X11] [2274] DRI2 WINDOW UMP SECURE ID CHANGED (0x3 -> 0x3)

And then it proceeds by re-trying to request a pair of DRI2 buffers.
But that's exactly the behaviour we want! As a down side, some ugly
flashing can be seen on screen at the time when this workaround kicks
in, but then everything normalizes. And unfortunately, the race
condition is still not totally eliminated because the blob is
apparently getting DRI2 buffer sizes from the separate GetGeometry
requests instead of using the information provided by DRI2GetBuffers.
But now the problem is at least very hard to trigger.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Since version 1.0, gstreamer (when using xvimagesink) often
allocates a larger XV image for the video with padding on all
four sides and then calls XvPutImage() to render a part of this
image. With the current XV implementation this results in
artifacts on the borders of the image, with a green bar at the
bottom.

I am observing this when playing a 1280x720 video on a 1920x1080
screen at 32bpp, the size of the video window doesn't matter.

This problem seems to be an exaggeration of the one described in
https://bugzilla.gnome.org/show_bug.cgi?id=685305

.

The solution appears to be to use the source area dimensions as
requested in the XvPutImage() call, as opposed to the dimensions
of the originally allocated image, and to honour the offsets
(src_x, src_y) when setting the source region on the display
controller. With this relatively simple change, the problem seems
to go away, and gstreamer 1.0 (which is faster than gstreamer 0.10
due to a zero-copy strategy) provides an acceptable solution for
video playback.
Signed-off-by: Harm Hanemaaijer <fgenfb@yahoo.com>

In the case if an attempt to reserve a scalable sunxi-disp layer
failed, don't initialize XV at all. Otherwise any attempt to use
XV overlay is not going to work correctly and just results in
the following dmesg spam:

[  728.280000] [DISP] not supported yuv channel format:18 in img_sw_para_to_reg

This may happen on Allwinner A13 if scaler mode is enabled in
.fex file (A13 only has one DEFE scaler). Allwinner A10 also
can have similar troubles in dual-head configuration if scaler
mode is enabled for one or both screens (A10 has two DEFE scalers).
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Add the "XVHWOverlay" boolean xorg.conf option to make it possible
to disable the XV acceleration feature using display layers on
sunxi hardware.
Signed-off-by: Harm Hanemaaijer <fgenfb@yahoo.com>

Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Proper layer sharing between XV and DRI2 still needs to be implemented.
Additionally we still need NEON and/or G2D "textured overlay" as a
fallback solution for the composited desktop (NEON optimized XV is going
to be useful for a wide range of ARM devices). A bit of performance
tuning is also necessary.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

They are needed for a basic XV extension implementation.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Benchmark tests reveal that xorg's fb layer PutImage implementation
does not follow on optimal code path for requests without special
raster operations, which is due to the use of a slower general blit
function instead of the pixman library. This affects Xlib PutImage
requests and some ShmPutImage requests. In the case of ShmPutImage,
xorg directs ShmPutImage requests to PutImage only if the width of
the part of the image to be copied is equal to the full width of
the image, resulting in relatively poor performance. If the width
of the part of the image that is copied is smaller than the full
image, then xorg uses CopyArea which results in the use of the
already optimal pixman blit functions. The sub-optimal path is
commonly triggered by applications such as window managers and web
browsers.

To fix this unnecessary performance flaw, PutImage is replaced with
a version that uses pixman for the common case of GXcopy and all
plane masks sets. This change is device-independent and only uses
pixman CPU blit functions that is already present in the xorg server.

Using the low-level benchmark program benchx
(https://github.com/hglm/benchx.git

), the following speed-ups were
measured (1920x1080x32bpp) on an Allwinner A10 device:

ShmPutImageFullWidth (5 x 5): Speed up 9%
ShmPutImageFullWidth (7 x 7): Slow down 5%
ShmPutImageFullWidth (22 x 22): Speed up 8%
ShmPutImageFullWidth (49 x 49): Speed up 19%
ShmPutImageFullWidth (73 x 73): Speed up 55%
ShmPutImageFullWidth (109 x 109): Speed up 50%
ShmPutImageFullWidth (163 x 163): Speed up 37%
ShmPutImageFullWidth (244 x 244): Speed up 111%
ShmPutImageFullWidth (366 x 366): Speed up 77%
ShmPutImageFullWidth (549 x 549): Speed up 92%
AlignedShmPutImageFullWidth (5 x 5): Slow down 14%
AlignedShmPutImageFullWidth (7 x 7): Slow down 6%
AlignedShmPutImageFullWidth (15 x 15): Speed up 10%
AlignedShmPutImageFullWidth (22 x 22): Speed up 9%
AlignedShmPutImageFullWidth (33 x 33): Speed up 21%
AlignedShmPutImageFullWidth (49 x 49): Speed up 28%
AlignedShmPutImageFullWidth (73 x 73): Speed up 30%
AlignedShmPutImageFullWidth (109 x 109): Speed up 47%
AlignedShmPutImageFullWidth (163 x 163): Speed up 38%
AlignedShmPutImageFullWidth (244 x 244): Speed up 63%
AlignedShmPutImageFullWidth (366 x 366): Speed up 84%
AlignedShmPutImageFullWidth (549 x 549): Speed up 89%

At 16bpp the speed-up is even greater:

ShmPutImageFullWidth (5 x 5): Slow down 8%
ShmPutImageFullWidth (7 x 7): Slow down 8%
ShmPutImageFullWidth (10 x 10): Slow down 6%
ShmPutImageFullWidth (22 x 22): Speed up 9%
ShmPutImageFullWidth (33 x 33): Speed up 20%
ShmPutImageFullWidth (49 x 49): Speed up 27%
ShmPutImageFullWidth (73 x 73): Speed up 69%
ShmPutImageFullWidth (109 x 109): Speed up 74%
ShmPutImageFullWidth (163 x 163): Speed up 100%
ShmPutImageFullWidth (244 x 244): Speed up 111%
ShmPutImageFullWidth (366 x 366): Speed up 133%
ShmPutImageFullWidth (549 x 549): Speed up 123%
AlignedShmPutImageFullWidth (5 x 5): Speed up 6%
AlignedShmPutImageFullWidth (7 x 7): Slow down 9%
AlignedShmPutImageFullWidth (10 x 10): Slow down 10%
AlignedShmPutImageFullWidth (33 x 33): Speed up 17%
AlignedShmPutImageFullWidth (49 x 49): Speed up 34%
AlignedShmPutImageFullWidth (73 x 73): Speed up 49%
AlignedShmPutImageFullWidth (109 x 109): Speed up 53%
AlignedShmPutImageFullWidth (163 x 163): Speed up 69%
AlignedShmPutImageFullWidth (244 x 244): Speed up 82%
AlignedShmPutImageFullWidth (366 x 366): Speed up 116%
AlignedShmPutImageFullWidth (549 x 549): Speed up 110%
Signed-off-by: Harm Hanemaaijer <fgenfb@yahoo.com>

This should be useful for Raspberry Pi. When reading uncached source buffers,
the VFP optimized overlapped two-pass blit is roughly 2-3 times slower than
memcpy in cached memory. Which makes it reasonably competitive compared to
ShadowFB (considering that ShadowFB allocates an extra buffer, does extra
memory copies which take time and thrash L2 cache, etc.). It even provides
a slight performance advantage in a more or less realistic use case
(scrolling in xterm), which needs reads from the framebuffer:

==== Before (xf86-video-fbdev with ShadowFB) ====

$ time DISPLAY=:0 xterm +j -maximized -e cat longtext.txt

real    1m50.245s
user    0m1.750s
sys     0m0.800s

==== After (xf86-video-sunxifb without ShadowFB) ====

$ time DISPLAY=:0 xterm +j -maximized -e cat longtext.txt

real    1m27.709s
user    0m1.690s
sys     0m0.920s

We get decent results even when reading from the framebuffer. However
in many typical workloads (excluding scrolling and dragging windows)
the framebuffer is primarily used as write-only. In write-only use
cases ShadowFB is just pure overhead. So getting rid of it is a
very good idea as this improves overall graphics performance.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

A small typo in a function argument and C compiler happily accepting
void pointers instead of something else is a dangerous combo. Need to
be more careful.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

This patch implements a heuristics, which enables backing store for some
windows. When backing store is enabled for a window, the window gets a
backing pixmap (via automatic redirection provided by composite extension).
It acts a bit similar to ShadowFB, but for individual windows.

The advantage of backing store is that we can avoid "expose event -> redraw"
animated trail in the exposed area when dragging another window on top of it.
Dragging windows becomes much smoother and faster.

But the disadvantage of backing store is the same as for ShadowFB. That's a
loss of precious RAM, extra buffer copy when somebody tries to update window
content, potentially skip of some frames on fast animation (they just do
not reach screen). Also hardware accelerated scrolling does not currently
work for the windows with backing store enabled.

We try to make the best use of backing store by enabling backing store for
all the windows that are direct children of root, except the one which has
keyboard focus (either directly or via one of its children). In practice this
heuristics seems to provide nearly perfect results:
 1) dragging windows is fast and smooth.
 2) the top level window with the keyboard focus (typically the application
    that a user is working with) is G2D accelerated and does not suffer from
    any intermediate buffer copy overhead.
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>

With the fallback to CPU backend for unsupported blits and also
threshold for avoiding small blits, now G2D should always provide
best overall performance.

The users of recent versions of xf86-video-sunxifb are supposed
to also have a reasonably recent version of linux-sunxi kernel.
Which includes the following fix:
  https://github.com/linux-sunxi/linux-sunxi/commit/3d49345343a1535b



The users of old kernels are going to see screen corruption on
dragging windows and scrolling. They just should upgrade :)
Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>