Flush in middle of rendering in KMS is not allowed because
buffers are discarded in flush.
Fixes crash when emiting split indices with RADEON_DEBUG=all.
Flush after transition would emit wrong state that could cause
wrong state emited for pending rendering operation.
Fixes wan once from extrement tuxracer that is using per vertex
materials.
vbo_split_copy was passing one past the max_index to draw function
which caused _tnl_draw_prims function to read uninitialized values
from copied array.
Bug was spoted in valgrind report of progs/tests/cva_huge.
Set the _NEW_BUFFERS flag and remove the code which updated the
parent framebuffer size. Normal Mesa state validation will do that.
Fixes issues with Warsow on r300g and possibly other bugs.
(cherry picked from commit c472ce5302)
The assertion is checking that the low-order bits of offset are not
set. It does this by anding the inverted offset mask with the
offset. This is clearly intended to be a bit-wise "invert".
Fixes bug #25984.
(cherry picked from commit 062a208814)
The assertion is checking that the low-order bits of offset are not
set. It does this by anding the inverted offset mask with the
offset. This is clearly intended to be a bit-wise "invert".
Fixes bug #25984.
(cherry picked from commit fda5078324)
Due to a quantization error, different cliprects of scaled video windows may
not have identical x / y scale.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Quake3 truncates the extension string, and GL_EXT_compiled_vertex_array
wasn't being detected, making it very slow.
This is a quick fix. The IMHO best way to address this in a more general
fashion is to sort by year.
Allows compiz to work in depth 16.
The DRI2 getBuffersWithFormat hook is only required for 16/32 bit depth
buffers, for colour buffers the only requirement is that the format matches
the drawable depth, which we can't check here.
There was a DrawBuffer/ReadBuffer typo and we were neglecting to invert
the texture coords when copying from a window to an FBO.
Plus, add some surface dump/debug code (disabled).
The xorg state tracker gets two new options to let the user choose
whether to enable / disable dirty throttling and swapbuffer throttling.
The default value of these options are enabled, unless the winsys
supplies a customizer with other values. The customizer record has been
extended to allow this, and also to set winsys-based throttling on a per-
context basis.
The vmware part of this patch disables the dirty throttling if the kernel
supports command submission throttling, and also in that case sets kernel
based throttling for everything but swapbuffers. The vmware winsys does not
set throttling per context, even if it theoretically could, but instead
sets throttling per screen. This should perhaps be changed, should the
xorg state tracker start to use multiple rendering contexts. Kernel throttling
is off by default for all new screens/contexts, so the dri state tracker
is not affected.
This significantly improves interactivity of the vmware xorg driver.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
The winsys may need to extract the svga_winsys_context from a
pipe_context. Add a function to enable that functionality.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Before, we only cleared the flags for the active tiles (the ones inside
the framebuffer bound). The problem is if we later bound a different,
larger surface to the tile cache we'd have some stale clear-flags still
set (and mistakenly clear some tiles in the new surface).
Fixes fd.o bug 26932.
(cherry picked from commit e1762fb870)
There's no reason to release the renderbuffer from the framebuffer object
or release the gallium surface in this function (they're reference counted).
In fact, we don't want to do this because we may later use the texture as a
pixel source (ex: glBlitFramebuffer) and need the surface.
Fixes fd.o bug 26923 and is part of the fix for bug 26932.
(cherry picked from commit 80dc54e308)
The array stack space wasn't allocated to the proper size. Fixes out of
bounds memory writes when the client/array stack depth exceeds one.
See fd.o bug 26768.
The rb->InternalFormat field will be set by the caller if the allocation
succeeds. Until then, this field's value can't be used. Fixes a failed
assertion with FlightGear.
(cherry picked from commit fe25476c04)
FRAG & TILE buffer are unused but still they need
to be associated with a valid relocation so that
userspace can't try to abuse them to overwritte
GART and then try to write anywhere in system
memory.
The kernel lets you clear depth without getting a depth offset
from userspace, mesa used to emit state before clear, but that got
lost in the refactoring, which made the kernel bug show up. Fix
mesa driver to emit the state properly now.
cherry-pick + squash master commits.
Signed-off-by: Dave Airlie <airlied@redhat.com>
This reverts commit 9d17ad2891.
Fun stuff so fixing exposes another bug which I'm having trouble
tracking down. So for now I'm just going to revert this untill
I can fix the real bug. Sorry about this.
This fixes a regression with Lightsmark, where more compact TGSI from Mesa
was causing a zero mask MOV to be emitted for shadow map compare, causing
problems in some backends.
Add a few more assertions to catch cases like this.
glxinfo needed fixing after commit 4bccd69.
Move fragment program only parameters into their own list
so that they are not queried for a vertex program.
Signed-off-by: Xavier Chantry <chantry.xavier@gmail.com>
Signed-off-by: Brian Paul <brianp@vmware.com>
This serves several purposes:
- disable caching in situations were is it useless or undesired
- share caches among all trees
- simplify purging the cache (when it's a single location)
- move the cache out of the tree, since that slows downs IDEs considerably
To retain previous behavior just define do
export SCONS_CACHE_DIR=$PWD/build/cache
before invoking scons.
This avoids exposing the ms driver structure to the winsys,
and nicely encapsulates driver customizable stuff.
In the future more things might be customizable by the winsys, like
throttling, 3D readback etc.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
The optimized Z-test functions assumed that the array of incoming quads
are adjacent, but that's not always true. The fragment shader can cull
intermediate quads, for example.
Now these Z-test functions can cope with non-adjacent quads. A little bit
of performance is probably lost, but it's probably not worth worring about.
This fixes broken glBitmap() Z testing, among other things.
Fixes crash in Homeworld2 which tries to create a 14MB buffer, because we
now avoid creating GMR buffers larger than 8MB to ensure progress given
we have only a 16MB pool.
Make sure we always test for XEXT version.
Make sure that the user has a recent version of libkms and libdrm installed.
Not that the st/xorg code requires so new versions to work but these are
needed to make a proper vmwgfx_drv.so driver which is the only real user.
Since the execbuffer change actually changed size off the ioctl
struct and not just a reuse of padded bits, we can't support
old kernels as easily as the scanout change was.
It seems there are still some places where draw can happen with mapped
buffers... Remove the assertions since there is not much more than can
be done at this moment.
This partially reverts commit af2023e31c.
Remove const qualifier from _mesa_HashLookup() table parameter to
avoid LOCK/UNLOCK warnings in the function body.
Signed-off-by: Brian Paul <brianp@vmware.com>
Purpose is two fold:
- when doing user-space memory management mapping a buffer stored in the
graphics aperture effectively pins it there, increasing the likelyhood
of failure of validating other buffers when flushing
- certain hardware/platform combinations do not allow a buffer to be
simultaneously mapped and validated
This fixes assertion failures in HL Uplink with the svga driver, where
vbo was holding a map to a buffer which was also referred in the command
stream.
Note: this a non-invasive fix and shouldn't be just cherry-picked into
master as-is -- a cleaner fix for this problem should be searched.
My first reading of MS docs was wrong. It says:
All rendering contexts of a shared display list must use an identical
pixel format. Otherwise the results depend on the implementation of
OpenGL used.
That is, it is OK to share contexts with different pixel formats.
Adobe Premiere Pro tries to do that: share lists between a rgbx8 and a
rgba8 pixel format.
This is a quick-fix for the time being...
The per-face mipmap LOD computation was invalid at cube edges. In
mip_filter_nearest/linear() we were trying to compute LOD using
texcoords that were sometimes indexes into different cube faces.
The subtraction used to compute the partial derivatives basically
gave random values, so the LOD was unpredictable. This fix simply
uses the same cube face for all four pixels in the quad. The per-
face texcoords all reference the same cube face so the partial
deriviates are computed properly.
A more elaborate fix would involve computing the LOD at the same
time as we choose the cube faces. But for now, this solution works
well and allows the piglit/cubemap test to pass.
we actually need to specify the formats for different attachements, otherwise
if the color buffer is 24bpp and the app asks for 16bpp depth buffer than
we end up fetching the depth from the drawable which is 24bpp and end up
creating the wrong depth buffer. use the new getBuffersWithFormat extension
to pass the depth correctly.
When our DLL is unloaded, even if we leave the data structures in memory
for sake of future calls, the MS CRT will destroy the heap. Instead we
make all calls no-ops by setting stw_dev to NULL.
Simply skip cleanup when contexts are still active.
This addresses two issues:
- in some situations the ICD DLL may be unloaded before the DLL that is
using GL contexts is, so we may receive GL calls after stw_cleanup.
- when aborting (exception, or control-c) the contexts may have been left
in an inconsistent state and attempting to destroy can cause
unpredictable results.
It's possible for min_index and max_index to be ~0 if the min/max
values were not actually set or computed in the state tracker.
Skip some assertions in that case.
This only effects the debug build error checking.
Windows doesn't really expect things to fail at this point -- it
will try many times until it finally gives up, worse, something bad
happens.
The WGL state tracker will gracefully decline to do work even when it
is loaded.
This object can be shared with another context, so we cannot just
delete it when the owning context is being destroyed.
Ensuring that buffer objects are properly refcounted guarantees
NullBufferObj is destroyed when all references to it are removed.
Note that when detaching the PBO from the region and making a new BO
for the region, we don't make it tiled even if the region originally
was.
Fixes piglit pbo-teximage-tiling.
(cherry picked from commit f20e83210e)
we piggy back flags in the top four bits of the element indices, so if
the maximum index stored in any of the elements is greater than
sizeof(ushort) - sizeof(flags we piggy back) then we overflow. fix it by
simply falling back to the slow path if we notice the overflow.
The piglit vp-address-01 test uses negative address-relative
offsets. In this test we're indexing into PROGRAM_STATE_VAR
which, in turn, contains references to ENV vars.
We previously fixed this issue for PROGRAM_CONSTANT.
piglit/vp-address-01 (the version from Feb 5) passes now.
Without this patch, the two symbols get an underscore prepended
and an "@4" appended when compiling with VC8.
Signed-off-by: Brian Paul <brianp@vmware.com>
PXOR user in code were causing the lowest SP float register to have NaN
values which made all math operations in that slot fail. Correct istruction
to clear float registers is XORPS which handles single precission floats
correctly.
Fixes progs/tests/fog in swrast SSE mode.
Now the correct commit instead of 9d9c1f17dc which is not even close
of correct fix for the bug.
Store XPD results in temporaries before writing to the dest register.
This issue has already been fixed in Mesa/master for XPD (and all
other instructions?).
Fixes progs/demos/bump.c when using TGSI interpreter (GALLIUM_NOSSE=1).
A new glean/glsl1 test has been added to test this case.
Texture derivatives are potentially undefined inside dynamic branches,
so hardwire lod zero in this case. Treating all if/endif and loop
constructs as dynamic branches.
The fogcoord calue was not pushed to GPU because of implicit float to int conversion.
Fix is to use float pointer to buffer object so no conversion is done in assigment.
This fixes a bug reported by Christoph Bumiller on mesa3d-dev.
When a texture is first created as RGBA, then re-defined with
glTexImage(internalFormat=GL_DEPTH_COMPONENT) we failed to deallocate
the original texture. When this texture was bound as a FBO surface,
the depth/Z surface format was RGBA instead of Z. Depending on the
driver this led to a failed assertion or FBO validation failure.
This patch does three things:
1. Remove ancient code that mysteriously tested if we were replacing
the smallest mipmap level and tested if the texture was not a cube map
texture. I can't see any reason for those tests.
2. Move the width=height=depth=0 test to after the code which frees
texture data. Calling glTexImage with width=height=depth=0 and data=NULL
is a way to free a single mipmap level.
3. Update the code comments.
There are no apparent conform, glean or piglit regressions from this change.
Before, _mesa_find_free_register() would scan the given shader to
find a free/unused register of the given type. But subsequent calls
would return the same register again. This caused a failure in the
_mesa_remove_output_reads() function which sometimes needs several
free temps.
Now use a new function which build a vector of 'used' flags and another
function which searches that vector for an unused register starting at
a position that's incremented for each call.
Fixes fd.o bug 26317. Note that a regression test for this has been
added to the glean/glsl1 test.
Lets see if this is not too pedantic. Obj pointers are never exposed to
GL apps so it should be possible to get this right.
Furthermore apps with GL widgets and test suits create and destroy many
contexts and objects, so bad reference counting is not really an option.
Always use _mesa_reference_buffer_object, and never call
ctx->Driver.DeleteBuffer() directly to prevent dangling pointers to the
null buffer object.
This fixes crash/assertions in sharedtex_mt and Autodesk Mudbox.
Commit 2708ddfb06 caused a few regressions.
We need to check/validate state after calling bind_arrays() because
it might set the _NEW_ARRAYS flag if the varying VP inputs change.
The symptom of this problem was some attribute arrays being ignored
(or interpreted as constant-valued) in glDrawRangeElements or
glMultiDrawElements.
A follow-on patch will add some additional asserts to try to catch
this kind of thing in the future.
...rather than checking/validating before all the calls to
_mesa_valid_to_render() and valid_to_render().
The next patch will actually fix some bugs...
Not since 6094e79f4e.
Drivers now need to flush draw module explicitely (which explains why
all those previous commits adding draw_flushes calls were necessary).
This is a good thing, but it's tricky to get this right in face of user buffers
(it's not even clear who has the responsibility to flush when a user buffer
is seen -- statetracker or pipe driver), so just force flush (temporarily)
since it's not a bottleneck now.
That is:
- check for no op
- update/flush draw module
- update bound state and mark it as dirty
In particular flushing the draw module is important since it may contain
unflushed primitives which would otherwise be draw with wrong state.
The driver (or preferably the hardware) should handle out of bounds
indices.
If there are problems then it's better to detect those in the debug
builds.
Do this by extending the nearest range to cover the new range.
This fixes an access fault in Call of Duty which was doing many disjoint
glBufferSubData calls.
So if we don't actually emit an atom to the hw because we don't
need it in the current state (e.g. lighting related atoms when
lighting is off) then don't mark it as clean, because when
lighting gets switched on we won't emit it at all.
This fixes funky gears colors.
Signed-off-by: Dave Airlie <airlied@redhat.com>
If GL_EXT_framebuffer_blit was not supported _mesa_DeleteFramebuffersEXT
would raise an error when deleting the currently bound framebuffer. This
because it tried to bind the default DRAW- and READ_FRAMEBUFFER separately.
This patch binds the default FRAMEBUFFER instead in that case.
Encountered in the fbo/fbo-copyteximage piglit test on R600.
Patch cleaned up a bit by Brian Paul.
Eric added some new code to check if offset < obj_size before computing
_MaxElement but my original code was still present afterward and it
clobbered the _MaxElement value.
Not sure if this came from a bad merge or what.
Based on a patch by Xavier Chantry <chantry.xavier@gmail.com>:
If x+width or y+height is negative, then maxx or maxy will get a bogus value
when converting that to unsigned. Fix this by setting 0 as minimal value.
This was also triggered by teeworlds, but only with some combination of
resolution and map section. For example upper part of dm2 at 1280x1024.
commit 53174afeeb introduced a portability change that converted GLint x,y
to GLuint. That breaks when x and y are negative, which seems to be allowed,
and which at least one game uses : teeworlds.
Rather than simply reverting the change, it seems possible to convert the
16bit unsigned to GLint so that comparisons are made between signed integers
instead. This hopefully does not break anything while keeping MSVC happy.
Signed-off-by: Xavier Chantry <chantry.xavier@gmail.com>
Signed-off-by: Brian Paul <brianp@vmware.com>
This is faster and ensures that NaN floats get stored properly.
Before, NaN values (which might be used with UP2H, UP2US, UP4B and
UP4UB) weren't getting stored properly with gcc -O3.
This is the second part of the fix for the piglit fp-unpack-01 failure
(bug 25973).
The UP2H, UP2US, UP4B and UP4UB instructions interpret the float
registers as integers. With gcc -O3 some bits were getting mixed up
somewhere. This is part of the fix for the piglit fp-unpack-01 test
failure (bug 25973).
The previous memory leak fix didn't always work properly.
Now check the xm_buffer::smh field (now documented!) to
see if the buffer points to shared memory.
Depending on first/last provoking vertex either the first or last
element/index may have the extra DRAW_PIPE_x flags. Mask off those
bits for all vertex elements to be safe (esp. for the quad case).
This may be overly cautious, but it's a cheap operation.
Fixes a segfault caused by large/bogus vertex indexes otherwise.
This is a minimum change to keep things building and running minimally -- it
might cause applications to fail to allocate buffers due to out of
GMR memory.
A proper update will be commited later.
It's necessary to download buffers from the host always, except if the
buffer is undefined, because:
- just PIPE_BUFFER_USAGE_CPU_WRITE doesn't guarantee all data is written
-- old contents may still pierce through
- PIPE_BUFFER_USAGE_DISCARD refers to a range, not the whole buffer, so
unless we track which parts have been modified and not we still need
to download the data.
__glXSetCurrentContextNull currently does not set the GL context to null
in the direct rendering case.
This can result in a segfault trying to flush an invalid old context
in glXMakeCurrent.
This fixes a crash starting the Unigine demos (they still don't work due
to missing extensions though).
Signed-off-by: Brian Paul <brianp@vmware.com>
glXDestroyContext does not destroy the context if it's still
connected to some window. Unbind context from window to test it.
Signed-off-by: Brian Paul <brianp@vmware.com>
Thanks to X server commit 2e2c5b216cc1c7a9bc26bd2c68226aaed5fc52ca
it is possible to be ABI compatible in version 2 with version 1 of
the DRI2 interface.
We use the presence of the getBuffersWithFormat extension function to
detect whether the server automatically adds a front.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
We were calling this from the CI span function, but not the RGBA
span function.
I don't know of a test program for the GL_EXT_depth_bounds_test
extension...
The integer Z clamping range depends on the number of bits
in the Z buffer because that's the scale factor used when we
transform NDC coords by the viewport/depth range.
Fixes fd.o bug #25972 but only for Z buffers up to a depth
of 30 bits. Beyond that we get into messy integer overflow
issues and things fall apart.
This should be the correct, but sub-optimal way to handle things on
pre-1.7 servers. For servers that automatically adds a fake front
we should never request the fake front and handle the reply differently.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
This reverts commit 1336989ec6.
The commit breaks frontbuffer rendering and the possibility to do
on-demand frontbuffer requests on all Xservers prior to 1.7.
The commit should be conditioned on a dri2 version check.
I've submitted a patch to Xserver to bump dri2 minor, and it's acked by Ian.
To be able to check the version from within the state tracker, we will
probably need another dri2 extension function that returns the server
version. The protocol has support for it.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Rearrange things so that the writes to the dest registers happen
after we've fetched/used all src registers.
The problematic instruction was: XPD TEMP[2].xyz, TEMP[0], TEMP[2];
Note that the dst reg is also a src reg.
This fixes bad shading with progs/glsl/bump.c since Eric's changes to the
Mesa program optimizer in commit d6690ce15f.
The optimizer rearranges some registers so we occasionally wind up with
something like the above.
Check whether the index is within bounds before accessing the array.
Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
Signed-off-by: Brian Paul <brianp@vmware.com>
NOUVEAU_BO_RD is defined (1 << 2), and `|' has higher precedence than `?'
so the second argument of nouveau_bo_map was always 0.
Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
Signed-off-by: Brian Paul <brianp@vmware.com>
I had this patch on my hard drive for long time. It doesn't fully
address SVGA multi-threading issues, but causes no regressions, so decided
to commit while it still applies cleanly.
Attention: merging this into master will cause issues due to recent
changes in reference counting to fix strict aliasing rules violation.
A src register's index can be negative if we're doing indirect
addressing into the constant buffer. Ex: MOV OUT[1], CONST[ADDR[0].x-3]
This fixes the piglit vp-arl-neg-array.vpfp test.
Before this change we were going out of bounds of the t->constants[]
array and getting garbage that later triggered an assertion.
References to program local and enviroment parameters are put into the
unified program parameters list as PROGRAM_STATE_VAR entries which point
into the local or environment arrays. So the param_binding_type field
should be PROGRAM_STATE_VAR.
This fixes the piglit vpfp-generic vp-arl-env-array.vpfp and
vp-arl-local-array.vpfp test failures.
We need to disable constant consolidation when building an array of
constants which might be indexed indirectly.
Fixes regression in piglit vpfp-generic vp-arl-constant-array.vpfp
test caused by earlier constant consolidation patch.
Only search for an existing, matching constant if swizzleOut is non-null.
We need to be able to disable constant consolidation when building an
array of constants which will be indexed indirectly.
Even shader ID doesn't change. Otherwise the token change is not
effective.
This reverts regression from commit
cdb445f3a9. Regression could be seen in
Quake3's loading screen -- the upper right corner of the screen would
be black.
If we ever had a non-tile-aligned tiled renderbuffer, the math was all
off. Use the existing x,y coordinates instead of trying to
reconstruct them from an incorrectly-calculated offset value.
GLSL 1.10 disallows initializers for uniforms but GLSL 1.20 and later
allows them. This patch uses the #version directive to allow/disallow
uniform initializers.
This addresses bug 25807, but piglit also needs to be fixed to specify
the GLSL version in the shader.
This function will search the constant parameters in an effort to re-use
constant slots. For example, {1,2,3,4} and {4,1,1,2} can be stored in
one constant slot and accessed with different swizzles. The swizzle info
must be propogated though the parsing code in a few places.
Fixes Piglit "vpfp-generic tests/shaders/generic/big-param.vpfp" failure.
Modify the non-debug (ie disabled) version of assert to expose the
value in the expression to the compiler (avoiding the unused variable
messages) while still expanding to a noop.
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
