| Age | Commit message (Collapse) | Author |
|---|
|
|
|
|
|
|
|
|
|
|
|
Currently we used a single buffer for each fragment programs, leading to
rendering synchronization. This patch uses a doubly linked list of BOs,
which is dynamically resized if all the BOs are busy.
Note that inline image transfers could be an alternative option: this
will be explored later.
This removes one of the big performance limitations of the current
driver.
We also stop using pipe_resource internally in favor of using nouveau_bo
directly.
|
|
|
|
Thanks to Patrice Mandin for debugging this.
|
|
[airlied -
Convert sprite coord index to a per-coord enable bit
set the rasteriser block up correctly for point sprites.
The inputs to the RS hw block change for sprite coords, so fix them up
properly - this fixes piglit point-sprite test.
]
Signed-off-by: Dave Airlie <airlied@redhat.com>
|
|
glsl generates these.
|
|
|
|
Trivially adds SEQ, SGT, SLE, SNE, SFL, STR and SSG which were missing.
|
|
This patch allocates a bigger chunk of memory to store queries in,
increasing the (hidden) outstanding query limit.
|
|
Currently on nv30/nv40 an assert will be triggered once 32 queries are
outstanding.
This violates the OpenGL/Gallium interface, which requires support for
an unlimited number of fences.
This patch fixes the problem by putting queries in a linked list and
waiting on the oldest one if allocation fails.
nVidia seems to use a similar strategy, but with 1024 instead of 32 fences.
The next patch will improve this.
|
|
Fixes SCons build.
|
|
Was broken during unification
|
|
This patch adds support for two-sided vertex color to nv30/nv40.
When set, the COLOR0/1 fs inputs on back faces will be wired to vs outputs BCOLOR0/1.
This makes OpenGL two sided lighting work, which can be tested with progs/demos/projtex.
|
|
|
|
They only apparently work on nv40 grclass cards, and this was the
previous behavior of the driver.
This really needs to be investigated more.
|
|
This is implemented in nvfx_state_fb and fragtex but was missing
in nvfx_screen.
This allows to avoid glCopyTexSubImage CPU fallbacks and makes Doom 3
much faster as a result.
|
|
|
|
Should improve performance and fix serious regressions on AGP cards.
|
|
|
|
No longer used, finally!
|
|
This makes the code faster due to the lack of indirect calls and also
makes it much easier to understand what is actually going on.
|
|
|
|
We could use an sb here instead, but there is no obvious performance
advantage, and perhaps there could be a disadvantage.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Rather than emitting relocations on flush notifications, emit them
in nvfx_state_start.
|
|
Just a dumb buffer, allowed by the RING_3D/fixed subchannel binding and
no support for relocations.
This is *much* faster than state objects.
|
|
Autobinding creates additional pushbuffer usage which may not be
accounted in callers, and is also slow.
The next relocations patch depends on this for correctness.
Assert instead if the objects are not bound, which should happen at
screen creation time.
|
|
RING_3D creates a method start for subchannel 7.
Bind the 3D engine to a fixed subchannel to make it work
This is much faster than the old BEGIN_RING, since we don't need
to waste cycles trying to "autobind" stuff, when a fast static binding
is perfectly good.
Subchannel 7 is chosen because the kernel takes up the lowest ones.
|
|
Currently we miscalculate the space needed to push vertices, causing
flushes where they should not happen.
Use a much more conservative estimate to fix it.
It will be done better in the future (e.g. using the nv50 primitive
splitter).
|
|
|
|
|
|
|
|
|
|
|
