The early Z stuff is supposed to be unsafe without some more work in the
enable/disable path (in particular, how do we want to get it disabled on
the way out to the X Server?), but at the moment is 6% in OA.
This is about a 30% performance win in OA with high settings on my GM45,
and experiments with 915GM indicate that it'll be around a 20% win there.
Currently, 915-class hardware is seriously hurt by the fact that we use
fence regs to control the tiling even for 3D instructions that could live
without them, so we spend a bunch of time waiting on previous rendering in
order to pull fences off. Thus, the texture_tiling driconf option defaults
off there for now.
One warning message:
drm_i915_getparam: -22
was still being sent to fprintf(). This causes all Piglit tests to fail,
even with MESA_DEBUG=0.
Using _mesa_warning() to emit the message allows the general Mesa controls
for messages like this to be applied.
This speeds up OA on my GM45 by 21% (more than the original CPU cost of
the upload path). We might still be able to squeeze a few more percent out
by avoiding repeatedly mapping/unmapping buffers as we upload elements into
them.
This requires upgrading the interface so that the argument to
glXBindTexImageEXT isn't just dropped on the floor. Note that this only
fixes the accelerated path on Intel, as Mesa's texture format support is
missing x8r8g8b8 support (right now, GL_RGB textures get uploaded as a8r8gb8,
but in this case we're not doing the upload so we can't really work around it
that way).
Fixes bugs with compositors trying to use shaders that use alpha channels, on
windows without a valid alpha channel. Bug #19910 and likely others as well.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
While it's a nice idea to be able to allow clients to choose a smaller
(or bigger for 16bpp screens!) depth size, right now DRI2 hands back a buffer
with a size that matches the drawable, rather than being based off of the
visual. This led to problems in readback as parts of the driver disagreed
on what format the depth buffer was really in.
Fixes the remainder of bug #19447.
I was lured into a false sense of security by the fact that the spans code was
already there, and a bunch of tests didn't catch the problem. oglconform's
mask.c did, though.
Bug #19970.
(cherry picked from commit f7d80aa006)
This also involved adding the new MSAA fields to driCreateConfigs().
Also, re-add prog_instructions->Sampler field for i965 driver. Will
have to revisit that.
We can support any combination of (a8r8g8b8, x8r8g8b8, r5g6b5) x (z0,z24,z24s8)
on either class of chipsets. The only restriction is no mixing bpp when also
mixing tiling. This shouldn't be occurring currently.
The upstream linux kernel headers and libdrm kernel headers disagree on the
tag name for the sarea struct: _drm_i915_sarea vs drm_i915_sarea. They
both typedef it to drm_i915_sarea_t though, so just use that.
Intel docs state that only 830/845 have VBOs, 855/865 don't. So
lets just not use them on i8xx at all.
This restores the old pre-vbo code and uses it on all 8xx hw.
Mesa requires that we be able to share objects between contexts, which means
that the objects need to be created by the same bufmgr, and the bufmgr
internally requires pthread protection for thread safety.
Rely on the bufmgr having appropriate locking.
Conflicts:
src/mesa/drivers/dri/intel/intel_span.c
src/mesa/main/fbobject.c
This converts the i915 driver to use the GEM interfaces for object management.
Effectively default to vblank_mode=3 on Intel to avoid tearing by default.
Users wanting to go "as fast as possible" (despite not being able to see frames
faster than their refresh rate allows) can still set the vblank_mode manually.
The boolean that the server gives us for whether the region is tiled was
getting used as the enum for what tiling mode. Instead, guess the correct
tiling in screen setup.
Also, fix the Y-tiling pitch setup. The pitch to the next tile in Y is
32 scanlines, not 8.
It turns out that it's not just deviceID dependent, and there's some additional
undefined factor that determines the bit 6 swizzling. It's now controllable
with swizzle_mode=[012] until we get a response on how to automatically detect.
The objects are swappable, so we're less concerned by excessive object
allocation now, and it's about a 20% performance improvement. If we get
concerns about the memory consumption from others, we can look into a
compromise position later.
