Support still isn't completely correct, but it's better. piglit
point-sprite now passes. However, glean's pointSprite test fails. In
that test the texture on the sprite is somehow inverted as though
GL_POINT_SPRITE_COORD_ORIGIN were set to GL_LOWER_LEFT. i915 hardware
shouldn't be able to do that!
I believe there are also problems when not all texture units have
GL_COORD_REPLACE set. The hardware enable seems to be all or nothing.
Fixes bug #25313.
Basically don't round up shared textures. This fixes compiz,
but I'm afraid that rounding up texture sizes here in the driver
is doomed, as it will inevitably break texture wrap modes.
This helps improve the surface cache behaviour in the face of the
large number of single-use render targets generated by EXA and the xorg
state tracker. Without this we can reference hundreds of individual
render targets from a command buffer, which leaves little scope for
sharing or reuse of those targets.
Flushing early means we can start reusing textures much sooner.
This shouldn't have much effect on normal 3d rendering as it's pretty
rare to have a command buffer with >8 different render targets in that
world.
This enum is only supported for OpenGL 2.0. If a driver supports
OpenGL 1.4 and GL_ARB_point_sprite, using this enum should generate an
error. This is important because, for example, i915 and i830 can
support GL_ARB_point_sprite, but they cannot support
GL_POINT_SPRITE_COORD_ORIGIN.
This commit just removes the check for NV_point_sprite, which is
completely wrong, and add some comments describing what the code
should do. I don't see an easy way to check for version >= 2.0 from
inside Mesa. Perhaps we should add an extension
GL_MESA_point_sprite_20 (like Intel's old GL_EXT_packed_pixels_12) to
indicate that this added bit of functionality is available.
Also note that glean's pointSprite test only checks for
GL_ARB_point_sprite before trying to use
GL_POINT_SPRITE_COORD_ORIGIN. Naturally, that fails on
non-2.0 implementations (i.e., Mac OS X on GMA 950).
Thanks to Intel code which I've just stolen pretty much as usual.
This fixes fdo bug 22851 which is a dri1 regression since rewrite.
Tested by: fpiobaf (Fabio) on #radeon
Signed-off-by: Dave Airlie <airlied@redhat.com>
Similar to the classic swapbuffer throttling in GL drivers, put an
upper bound on the number of outstanding chunks of rendering the state
tracker can generate -- where calling the block handler denotes a
chunk.
Currently that number is set at around 4 "chunks", but could be
tweaked up or down. If a better measure for the amount of outstanding
rendering is found, that would be fine too.
As it stands, this improves interactivity by preventing the X server
from queueing up arbitary amounts of rendering.
These currently don't reference any constants. Can add this back
if newer shaders need them, but in the meantime don't create
a new constant buffer every time we do a blit.
The two users of composite (Composite and Solid) now call a new
xorg_composite_done() from their Done functions, while CopyDone
is directly implemented on top of xorg_renderer.c.
