These hints allow an acceleration architecture to optimize allocation of certain
types of pixmaps, such as pixmaps that will serve as backing pixmaps for
redirected windows.
* hw/kdrive/src/Makefile.am: use fb/fbcmap_mi.c
and not fb/fbcmap.c. This allows kdrive to take advantage of
extensions redefining the entry points of micmap.c stuff.
For instance it allows Xephyr to have a working GL, which is not
possible otherwise, because GL redefines mInitVisualsProc
to initialise its visuals.
This was an attempt to avoid scratch gc creation and validation for paintwin
because that was expensive. This is not the case in current servers, and the
danger of failure to implement it correctly (as seen in all previous
implementations) is high enough to justify removing it. No performance
difference detected with x11perf -create -move -resize -circulate on Xvfb.
Leave the screen hooks for PaintWindow* in for now to avoid ABI change.
over to new system.
Need to update documentation and address some remaining vestiges of
old system such as CursorRec structure, fb "offman" structure, and
FontRec privates.
Composite's automatic redirection is a more general mechanism than the
ad-hoc BS machinery, so it's much prettier to implement the one in terms
of the other. Composite now wraps ChangeWindowAttributes and activates
automatic redirection for windows with backing store requested. The old
backing store infrastructure is completely gutted: ABI-visible structures
retain the function pointers, but they never get called, and all the
open-coded conditionals throughout the DIX layer to implement BS are gone.
Note that this is still not a strictly complete implementation of backing
store, since Composite will throw the bits away on unmap and therefore
WhenMapped and Always hints are equivalent.
I exported the evdev driver to Xephyr server. I'm running it using something
like:
$ ./hw/kdrive/ephyr/Xephyr :1 -mouse evdev,,device=/dev/input/event4 -keybd \
evdev,,device=/dev/input/event1,xkbmodel=abnt2,xkblayout=br
It also closes /#5668.
and the Xephyr virtual mouse keeps alive. With this patch the semantic changes
turning '-pointer' && 'Xephyr virtual mouse' always false.
Now we can open a device pointer and pass its options in Xephyr's command line
without having other pointer unused.
The former <X11/extensions/XKBsrv.h> has been pulled into the server now as
include/xkbsrv.h, and the world updated to look for it in the new place,
since it made no sense to define server API in an extension header. Any
further work along this line will need to do similar things with XKBgeom.h
and friends.
Move the bell into an OS function, and use that if it's declared; else,
fall back to using the driver's function.
Remove the Linux keyboard bell function; just move it into the OS layer.
Use named initialisers when converting the old structures, and eliminate
unused functions.
Add KdOsAddInputDrivers, which adds all relevant input drivers.
Could possibly be refactored to KdAddInputDrivers, which called through
OsFuncs to a new function, if it existed.
Move the keymap copying to event processing time (in
ProcessInputEvents), instead of being at event enqueuing time.
Break SetCore{Pointer,Keyboard} out into separate functions.
Change mieqEnqueue to take a device pointer, that asks for the
_original_ device associated with this event.
Don't allocate events on every GKE/GKVE/GPE call, just have the DDX manage
it instead. Introduce GetMaximumEventsNum(), which is the maximum number
of events these functions will ever produce.
Do a linear n -> n initialisation on the map up until KD_MAX_BUTTON in
KdNewPointer, moving it out of both KdParsePointer, and KdPointerProc.
Also remove dead pointer acceleration code.
Convert KDrive to GPE/GKE interface.
Add first-class drivers and enumerate every device separately through
Xi, instead of lamely attempting to aggregate them.
Add XKB support to the Linux keyboard driver.
Add 'thumb button' support to the tslib driver.
Rejig InitInput, so each DDX has to add a list of drivers it supports.
Support NewInputDeviceRequest, et al.
Add #include <sys/file.h> on Solaris for FNONBLOCK/FASYNC definitions
hw/kdriver/linux/Makefile.am Move agp.c & agp.h to KDRIVE_HW_SOURCES since
they're not needed for Xephyr-only builds
Add -lrt to XEPHYR_LIBS if needed to get nanosleep().
Changed the semantics of the Convolution filter a bit. It now doesn't try
to normalize the filter values but leaves this to the client. This
gives more reasonable behaviour in the limit where the filter
parameters sum up to 0.
in the kaa structure: markSync and waitMarker. The first, if set,
returns a hardware-dependent marker number which can then be waited for
with waitMarker. If markSync is absent (which is the case on all
drivers currently), waitMarker must wait for idle on any given marker
number. The intention is to allow for more parallelism when we get
downloading from framebuffer, or more fine-grained idling.
- Replace the KdMarkSync/KdCheckSync functions with kaaMarkSync and
kaaWaitSync. These will need to be refined when KAA starts being smart
about using them. Merge kpict.c into kasync.c since kasyn.c has all the
rest of these fallback funcs.
- Restructure all drivers to initialize a KaaInfo structure by hand rather
than statically in dubious order.
- Whack the i810 driver into shape in hopes that it'll work after this
change (it certainly wouldn't have before this). Doesn't support my
i845 though.
- Make a new KXV helper to avoid duplicated code to fill the region with
the necessary color key. Use it in i810 and mach64 (tested).
might be able to do Precise in hardware, so leave it up to the driver.
- Add a helper function for computing a set of offsets for smooth trapezoid
rasterizing using many sharp trapezoids.
RasterizeTrapezoid screen function. These hooks will be called for
imprecise, non-sharp trapezoids with A8 destinations.
Note that the current main consumer of trapezoids, cairo, is requesting
precise, sharp trapezoids by not changing the default Picture
attributes, but gets non-sharp effects in software because fb bases its
choice of sharp/non-sharp on the mask format being A8 vs A1, and cairo
asks for A8. Follow fb's (poor?) example by ignoring the sharp setting
and basing the choice off of the mask being A8.
heuristic, delay the decision until the first
kaaPixmapUse{Screen|Memory}, and put it in framebuffer if UseScreen was
called. Provides a significant improvement in cairo speeds (100%
speedup in cairogears here) and is likely to improve text performance
as well.
rate selection necessary for Mac fbdev
Don't know about fb changes to pixmaps, so can't track dirt.
Add Mac specific 1280x854 mode. Warn when requested mode isn't found.
Add ability to soft-boot video cards.
Add region expand request. FIXME: need test cases
header include. I don't know how this happened, but I'm going to blame
a "few" fscks having happened between testing the code and committing
(due to other driver changes that made me not blame dirty for the
failures I saw) for disappearance of kasync.c changes. Fixes a lot of
corruption.
offscreen pixmap area, and set it when any rendering occurs. When
moving a pixmap out of offscreen, don't read data back if it wasn't
dirtied (compared to the system memory copy).
out when allocation can't find a free area of the requested size. When
offscreen pixmaps get used, the offscreen area's score is increased by
a constant value. Every certain number of increases, all offscreen area
scores get decreased by a fraction. When choosing a set of areas to
remove for a new allocation, the set of areas with the smallest total
score is chosen for removal. While this is not the smartest system, it
prevents things like always removing the first offscreen area in memory
(likely the most recent) to be kicked out when doing replacing.
solid-fill-based composite acceleration.
- Use a real pixmap when doing an UploadToScratch (For pDrawable->type ==
DRAWABLE_WINDOW, you need to get the backing pixmap).
- Pass back the x/y offsets from kaaGetOffscreenPixmap unconditionally,
because they'll be used in the scratch case.
- Turn on the Render acceleration for Rage 128 and Radeon 100-series at
last!
Fix a few allocation bugs with alternate visual ids Allow for non-8/8/8
alternate visuals
Turn off any existing shadow before enabling it again (avoids
re-registering existing damage)
Add some validation code to catch re-registered damages
addition to the ARGB32 one. This allows 'glitz' to run on top of any X
server using mesa.
Switch to using 32bpp for depth 24 pixmaps (even when the frame buffer is
not depth 24).
KdCheckSync -- the boolean used in the latter won't be set yet.
Oops. == instead of =.
Must sync hardware before rasterizing trapezoids in case the mask is in
off-screen memory and has just been erased. Yes, it is silly to place
masks in off-screen memory. That's a separate issue.
Off-screen reallocation could have used a stale pointer.
Separate framebuffer mapping computation from actual frame buffer mapping.
Now map the frame buffer from vesaEnable so that VT switch shares the
same mapping code. This makes sure any shadow framebuffer is allocated
again.
pixmap into temporary offscreen storage. Subsequent UploadToScratch may
clobber the data of previous ones. This allows hardware acceleration of
composite operations on glyphs.
- Add a new UploadToScreen kaa hook for doing the actual moving of data to
framebuffer. This would allow us to do things like hostdata blits or
memcpy to agp and then blit.
- Add an UploadToScreen on ATI which is just memcpy, but which will be
replaced with a hostdata blit soon.
- Add UploadToScratch on ATI and reserve 64k of scratch space. This
provided a 3x speedup of rgb24text on my Radeon.
functions. Along with making things more readable, it fixes a problem
where the coordinates would get messed up if acceleration failed due to
things like pixmaps being in the wrong locations.
pixmaps are offscreen and don't have alpha maps. It is the last case
checked before going to software fallback
- Use the new Composite hook in the ati driver to implement acceleration of
most Composites that get done in an xcompmgr environment on r100 series
cards. It is only available when using the DRM. There are still some
corruption issues, but the DRI is still non-default and I need to get
this into version control.
- Add libdrm and libdri. Portions of the DRI extension are stubbed out.
- Use the DRM in the ATI driver when available. This provides a minor
performance improvement in x11perf, and opens the possibility of using
the 3d hardware for acceleration in the future.
- Implement solid fill acceleration for Composite in KAA.
- Implement Blend hook for Composite and use it on r128.
- Fix a bug of mine that resulted in overuse of offscreen memory.
- Fix many miscellaneous bugs in ATI driver and add PCI IDs.
available, with fbdev being used by default.
- Use depth 16 by default when vesa backend is used.
- Add MMIO defines for PowerPC (should be in a common location).
Many thanks for Michel Daenzer for much of this code.
help with ati, and future DRM drivers.
- Add new "ati" kdrive driver. It has ancestry in the r128 driver from
andersca, but took a detour through being the WIP SiS 300 driver on the
way. It supports Radeons (tested on QD VIVO and 7500) and Rage 128.
Current limitations include that it requires depth 16 and that the
other Rage 128 PCI IDs aren't included yet.
change, just cleaning up a bit)
Reset the screen->memory_base value when frame buffer is remapped. This
makes sure new off-screen allocations point at the newly mapped region
rather than the old (now unmapped) region.
Sufficient, but not always necessary.
Add borderClip to damage on creation so that clients needn't guess.
Fix API to FbDots functions to make PolyPoint work with screen_x/screen_y
offsets
Add debugging code to make sure no pictures are left pointing at freed
pixmaps. "Can't" happen, but it did once.
Change KdOffscreenArea structure to eliminate separate private structure,
eliminate the ScreenPtr, change from doubly linked to singly linked
list.
Don't damage BackgroundNone windows on PW_BACKGROUND. Re-clip damage to
borderClip in DamageSubtract.
drawable pixmap rather than directly into the frame buffer. Rewrite
logic in kdoffscreen to make space for new allocations, now deals
correctly with locked areas.
unwrapping pixmap operations. This makes sure the accelerated code gets
invoked for pixmaps
Add pixmap migration support to kaa. Can't quite automatically migrate
pixmaps off-screen, but soon we will. Can kick objects out of video
memory. Move per-screen linked list pointers to pixmap private. Wrap
Composite in preparation for migrating pixmaps. Have kasync ignore
drawable type so that pixmaps trigger sync Add KdOffscreenFini to
cleanup on server reset. Switch off screen area to have only a 'save'
function; moving objects to off screen memory is done by saving then
freeing the area.