These mistakenly computed 't' instead of t * t * (3.0 - 2.0 * t).
Also, properly vectorize the smoothstep(float, float, vec) variants.
NOTE: This is a candidate for the 7.9 and 7.10 branches.
No need to register function prototypes in the module now that we call
the C function pointer directly -- less LLVM objects lying around.
Limited testing with lp_test_format.
Even though a bound texture stays bound when calling set_fragment_sampler_views,
it must be assigned a new cache region depending on the occupancy of other
texture units.
This fixes:
https://bugs.freedesktop.org/show_bug.cgi?id=28800
Thanks to Álmos <aaalmosss@gmail.com> for finding the bug in the code.
NOTE: This is a candidate for both the 7.9 and 7.10 branches.
We need to keep using the pipe_get_tile_swizzle() even though there's
no swizzling because we need to explicitly pass in the surface format.
Fixes http://bugs.freedesktop.org/show_bug.cgi?id=32459
The only mismatch between the two is that we have to clear the
destination's alpha to 1.0. Fixes WOW performance on my Ironlake,
from a few frames a second to almost playable.
Before, we were going off of a couple of known (hopeful) matches
between internalFormats and the cpp of the read buffer. Instead, we
can now just look at the gl_format of the two to see if they match.
We should avoid bad blits that might have been possible before, but
also allow different internalFormats to work without having to
enumerate each one.
The blit that follows appears in the command stream so it's serialized
with previous rendering. Any queued vertices in the tnl layer were
already flushed up in mesa/main/.
Create a constant int pointer to the C function, then cast it to the
function's type. This avoids using trampoline code which seem to be
inadvertantly freed by LLVM in some situations (which leads to segfaults).
The root issue and work-around were found by José.
NOTE: This is a candidate for the 7.10 branch
Michel Hermier reported libdrm segfault (and kernel crash) on nv40 using
gallium :
http://www.mail-archive.com/nouveau@lists.freedesktop.org/msg06563.html
It turns out these were caused by some missing WAIT_RING (or wrong
computation of the WAIT_RING sizes). Unlike all other libdrm_nouveau users,
nvfx gallium tried to use a mininum calls of WAIT_RING, one WAIT_RING could
apply to many methods for different code paths and spread across several
functions. This made it too tricky to find out what the missing or wrong
WAIT_RING was.
By restoring BEGIN_RING, we force one WAIT_RING per method, and it's much
easier to check if the free size required in the pushbuffer is correct. As
curro said, "let's keep it simple for the maintainers until the big
bottlenecks are gone"
Benchmarked on nv35 with openarena, nexuiz and ut2004 and no performance
regression.
The core of this patch was made with Coccinelle, with minor manual fixes
made on top.
Tested-by: Michel Hermier <hermier@frugalware.org>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
This is the hack for input interactivity of frontbuffer rendering
(like we do for backbuffer at intelDRI2Flush()) by waiting for the n-2
frame to complete before starting a new one. However, for an
application doing multiple contexts or regular rebinding of a single
context, this would end up lockstepping the CPU to the GPU because
every unbind was considered the end of a frame.
Improves WOW performance on my Ironlake by 48.8% (+/- 2.3%, n=5)
Given a dispatch slot, entry_get_public returns the address of the
corresponding public entry point. There may be more than one of them.
But since they are all equivalent, it is fine to return any one of them.
With entry_get_public, the address of any public entry point can be
calculated at runtime when an assembly dispatcher is used. There is no
need to have a mapping table in such case. This omits the unnecessary
relocations from the binary.
Hidden entries are just like normal entries except that they are not
exported. Since it is not always possible to hide them, and two hidden
aliases can share the same entry, the name of hidden aliases are mangled
to '_dispatch_stub_<slot>'.
Split out function name generation from _c_decl to _c_function, and use
it everywhere. Add an optional 'export' argument to _cdecl. It is
prepended to the returned string.
Really no idea why I didn't see this before, but these values were opposite
the register spec.
this seems to fix rv530 HiZ on my laptop, will reenable in next commit.
Signed-off-by: Dave Airlie <airlied@redhat.com>
For GL fragColor semantics we need to tell the pipe drivers that the fragment
shader color result is to be replicated to all bound color buffers, this
adds the basic TGSI + documentation.
v2: fix missing comma pointed out by Tilman on mesa-dev.
Signed-off-by: Dave Airlie <airlied@redhat.com>
If 'start' is odd, render the first triangle with indices embedded
in the command stream, which adds 3 to 'start' and makes it even.
Then continue with the fast path.
This is a bandaid on the problem that if some formats were not renderable
(like luminance_alpha), st/mesa fell back to some RGBA format, so basically
some non-renderable formats were actually not used at all. This is only
a problem with hardware drivers, softpipe can render to anything.
Instead, require only RGB8/RGBA8 to be renderable.
Radeon GPUs can do this. R600 can even do render-to-texture.
Packing and extracting aren't implemented, but we shouldn't hit them (I think).
Tested with swrast, softpipe, and r300g.
Just like everywhere else, I never trust my constant uploads to
correctly put constants in the right places, even though that's so
rarely where the issue is.
It's mostly like gen4 message descriptor setup, except that the sizes
of type/control changed to be like gen5. Fixes 21 piglit cases on
gm45, including the regressions in bug #32311 from increased VS
constant buffer usage.
Fixes this GCC warning.
lp_bld_const.h: In function 'lp_build_const_int_pointer':
lp_bld_const.h:137: warning: cast from pointer to integer of different size
Determine header present for fb write by msg length is not right
for SIMD16 dispatch, and if there're more output attributes, header
present is not easy to tell from msg length. This explicitly adds
new param for fb write to say header present or not.
Fixes many cases' hang and failure in GL conformance test.
Set window_bit only when the visual id is greater than zero. Correct
visual types. Skip slow configs as they are not relevant. Finally, do
not return duplicated configs.
All single-buffered configs were ignored before to make sure
EGL_RENDER_BUFFER is settable for window surfaces. It is better to
allow single-buffered configs and set EGL_WINDOW_BIT only for
double-buffered ones. This way there can be single-buffered pixmaps.
The SNB alt-mode math does the denorm and inf reduction even for a
"raw MOV" like we do for g0 message header setup, where we are moving
values that aren't actually floats. Just use UD type, where raw MOVs
really are raw MOVs.
Fixes glxgears since c52adfc2e1, but no
piglit tests had regressed(!)
Can't get away from referencing upload buffer as after flush a vertex buffer
using the upload buffer might still be active. Likely need to simplify the
pipe_refence a bit so we don't waste so much cpu time in it.
candidates for 7.10 branch
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Instead of when we read texture tiles. Now swizzling happens after
the shadow depth compare step. This fixes the piglit glsl-fs-shadow2d*
tests (except for proj+bias because of a GLSL bug).
We need to swizzle after the shadow comparison so that the GL_DEPTH_MODE
functionality is handled properly.
This fixes all the piglit glsl-fs-shadow2d*.shader_test cases, except
for glsl-fs-shadow2dproj-bias.shader_test which fails because of a
bug in the GLSL compiler (fd.o 32395).
Note the support for non float vertex draw likely regressed need to
find what we want to do there.
candidates for 7.10 branches
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
This is still awful, but my ability to care about reworking the old
backend so we can just get a temporary value into a POW is awfully low
since the new backend does this all sensibly.
Fixes:
fp1-LIT test 1
fp1-LIT test 3 (case x < 0)
fp1-POW test (exponentiation)
fp-lit-mask
commit 4f106f44a32eaddb6cf3fea6ba5ee9787bff609a
Author: Brian Paul <brianp@vmware.com>
Date: Mon Dec 13 14:06:08 2010 -0700
st/mesa: reorganize vertex program translation code
Now it looks like the fragment and geometry program code.
Also remove the serial number fields from programs. It was used to
determine when new translations were needed. Now the variant key is
used for that. And the st_program_string_notify() callback removes all
variants when the program's code is changed.
commit e12d6791c5e4bff60bb2e6c04414b1b4d1325f3e
Author: Brian Paul <brianp@vmware.com>
Date: Mon Dec 13 13:38:12 2010 -0700
st/mesa: implement geometry shader varients
Only needed in order to support per-context gallium shaders.
commit c5751c673644808ab069259a852f24c4c0e92b9d
Author: Brian Paul <brianp@vmware.com>
Date: Sun Dec 12 15:28:57 2010 -0700
st/mesa: restore glDraw/CopyPixels using new fragment program variants
Clean up the logic for fragment programs for glDraw/CopyPixels. We now
generate fragment program variants for glDraw/CopyPixels as needed which
do texture sampling, pixel scale/bias, pixelmap lookups, etc.
commit 7b0bb99bab6547f503a0176b5c0aef1482b02c97
Author: Brian Paul <brianp@vmware.com>
Date: Fri Dec 10 17:03:23 2010 -0700
st/mesa: checkpoint: implement fragment program variants
The fragment programs variants are per-context, as the vertex programs.
NOTE: glDrawPixels is totally broken at this point.
commit 2cc926183f957f8abac18d71276dd5bbd1f27be2
Author: Brian Paul <brianp@vmware.com>
Date: Fri Dec 10 14:59:32 2010 -0700
st/mesa: make vertex shader variants per-context
Gallium shaders are per-context but OpenGL shaders aren't. So we need
to make a different variant for each context.
During context tear-down we need to walk over all shaders/programs and
free all variants for the context being destroyed.
The new GLSL compiler doesn't support geom shaders yet so disable the
GL_ARB_geometry_shader4 extension. Undo this when geom shaders work again.
NOTE: This is a candidate for the 7.10 branch.
This is the same as what the array dereference handler does.
Fixes piglit test glsl-link-struct-array (bugzilla #31648).
NOTE: This is a candidate for the 7.9 and 7.10 branches.
The hardware supports zero stride just fine. This is a port
of 2af8a19831 from r300g.
NOTE: This is a candidate for both the 7.9 and 7.10 branches.
Signed-off-by: Alex Deucher <alexdeucher@gmail.com>
The RS690 memory controller prefers things to be on a different
boundary than the discrete GPUs, we had an attempt to fix this,
but it still failed, this consolidates the stride calculation
into one place and removes the really special case check.
This fixes gnome-shell and 16 piglit tests on my rs690 system.
NOTE: This is a candidate for both the 7.9 and 7.10 branches.
Signed-off-by: Dave Airlie <airlied@redhat.com>
When the driver is the last reference to libEGL.so, unloading it will
cause libEGL.so to be unmapped and give problems. Disable the unloading
for now. Still have to figure out the right timing to unload drivers.
The hardware apparently does support a zero stride, so let's use it.
This fixes missing objects in ETQW, but might also fix a ton of other
similar-looking bugs.
NOTE: This is a candidate for both the 7.9 and 7.10 branches.
If a source operand has a non-native swizzle (e.g. the KIL instruction
cannot have a swizzle other than .xyzw), the lowering pass uses one or more
MOV instructions to move the operand to an intermediate temporary with
native swizzles.
This commit fixes that the presubtract information was lost during
the lowering.
NOTE: This is a candidate for both the 7.9 and 7.10 branches.
This fixes broken rendering of trees in ETQW. The trees still disappear
for an unknown reason when they are close.
Broken since:
2ff9d4474b
r300/compiler: make lowering passes possibly use up to two less temps
NOTE: This is a candidate for the 7.10 branch.
do_assignment may apply implicit conversions to coerce the base type
of initializer to the base type of the variable being declared. Fixes
piglit test glsl-implicit-conversion-02 (bugzilla #32287). This
probably also fixes bugzilla #32273.
NOTE: This is a candidate for the 7.9 branch and the 7.10 branch.
There are exceptions to the table for depth texturing or rendering to
not-quite-supported formats thanks to the non-orthogonal component
selection for surface formats, but it's still a lot simpler.
Fixes piglit valgrind glsl-array-bounds-04 failure (FDO bug 29946).
NOTE:
This is a candidate for the 7.10 branch.
This is a candidate for the 7.9 branch.
The current state is allowed to be undefined past DrawElements et al.
Consequently omit that copying at least in the display list code.
This pays us some percents performance.
Signed-off-by: Brian Paul <brianp@vmware.com>
assert(current_save_state < MAX_META_OPS_DEPTH) did not compile.
Rename current_save_state to SaveStackDepth to be more consistent with
the style of the other fields.
VS places color attributes together so that SF unit can fetch the right
attribute according to object orientation. This fixes light issue in
mesa demo geartrain, projtex.
_mesa_meta_CopyPixels results in nested meta operations on Sandybridge.
Previoulsy the second meta operation overrides all states saved by the
first meta function.
When the application is not linked to any libGL*.so, loading st_GL.so
would give
/usr/local/lib/egl/st_GL.so: undefined symbol: _glapi_tls_Context
In that case, load libGL.so and try again. This works because
util_dl_open loads with RTLD_GLOBAL.
Fix "clear" OpenGL ES 1.1 demo.
Fixes piglit glx-shader-sharing crash.
When shaders are shared by multiple contexts, the shader's draw context
pointer may point to a previously destroyed context. Dereferencing the
context pointer will lead to a crash.
In this case, simply removing the flushing code avoids the crash (the
exec and sse shader paths don't flush here either).
There's a deeper issue here, however, that needs examination. Shaders
should not keep pointers to contexts since contexts might get destroyed
at any time.
NOTE: This is a candidate for the 7.10 branch (after this has been
tested for a while).
Currently we only unroll loops with conditional breaks at the end, which is
the form that lower_jumps generates.
However, if breaks are not lowered, they tend to appear at the beginning, so
add support for a conditional break anywhere.
Signed-off-by: Luca Barbieri <luca@luca-barbieri.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Found this bug by code inspection. Based off the comments just before
this code, the intent is to find whether the break exists in the "then"
branch or the "else" branch. However, the code actually looked at the
last instruction in the "then" branch twice.
If you used a constant array index to access the matrix, we'd flag a
bunch of wrong inputs/outputs as being used because the index was
multiplied by matrix columns and the actual used index was left out.
Fixes glsl-mat-attribute.
Fixes this GCC warning.
brw_fs.cpp: In function 'brw_reg brw_reg_from_fs_reg(fs_reg*)':
brw_fs.cpp:3255: warning: 'brw_reg' may be used uninitialized in this function
Allow important performance increase by doing hw specific implementation
of the upload manager helper. Drop the range flushing that is not hit with
this code (and wasn't with previous neither). Performance improvement are
mostly visible on slow CPU.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
r600g is up to a point where all small CPU cycle matter and pb* turn
high on profile. It's mostly because pb try to be generic and thus
trigger unecessary check for r600g driver. To avoid having too much
abstraction & too much depth in the call embedded everythings into
r600_bo. Make code simpler & faster. The performance win highly depend
on the CPU & application considered being more important on slower CPU
and marginal/unoticeable on faster one.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
This eases the gen6 implementation, which can only handle up to 32
registers of constants, while likely not penalizing real apps using
reladdr since all of those I've seen also end up hitting the pull
constant buffer. On gen6, the constant map means that simple NV VPs
fit under the 32-reg limit and now succeed. Fixes around 10 testcases.
Raise error if a sampler array is indexed with a non-constant expression.
From section 4.1.7 of the GLSL 1.30 spec:
"Samplers aggregated into arrays within a shader (using square
brackets [ ]) can only be indexed with integral constant
expressions [...]."
Attribute 0 has no special meaning in GLES2. Add VertexAttrib4f_nopos
for that purpose and make _es_VertexAttrib* call the new function.
Rename _vbo_* to _es_* to avoid confusion. These functions are only
used by GLES, and now some of them (_es_VertexAttrib*) even behave
differently than vbo_VertexAttrib*.
CMP may now use two less temps, other non-native instructions may end up
using one less temp, except for SIN/COS/SCS, which I am leaving unchanged
for now.
This may reduce register pressure inside loops, because the register
allocator doesn't do a very good job there.
That's what I get for not running piglit before pushing.
Don't try to patch types of unsized arrays when linking fails.
Don't try to patch types of unsized arrays that are shared between
shader stages.
Fixes piglit test case glsl-vec-array (bugzilla #31908).
NOTE: This bug does not affect 7.9, but I think this patch is a
candiate for the 7.9 branch anyway.
Types of declared variables and their initializer must match excatly
except for unsized arrays. Previously the type inherritance for
unsized arrays happened implicitly in the emitted assignment.
However, this assignment is never emitted for uniforms. Now that type
is explicitly copied unconditionally.
Fixes piglit test array-compare-04.vert (bugzilla #32035) and
glsl-array-uniform-length (bugzilla #31985).
NOTE: This is a candidate for the 7.9 branch.
With the change of extended math from having the arguments moved into
mrfs and handed off through message passing to being directly hooked
up to the EU, it looks like the piece for doing source modifiers
(negate and abs) was left out.
Fixes:
fog-modes
glean/fp1-ARB_fog_exp test
glean/fp1-ARB_fog_exp2 test
glean/fp1-Computed fog exp test
glean/fp1-Computed fog exp2 test
ext_fog_coord-modes
Previously the code only handled scalars and vectors. This new code
is modeled somewhat after similar code in ir_to_mesa.
Reviewed-by: Eric Anholt <eric@anholt.net>
R6XX GPU doesn't like to have two partial flush writting
back to memory in row without a prior flush of the pipeline.
Add PS_PARTIAL_FLUSH to flush all work between the CP and
the ES, GS, VS, PS shaders.
Thanks a lot to Alban Browaeys (prahal on irc) for investigating
this issue.
Signed-off-by: Alban Browaeys <prahal@yahoo.com>
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
gen6 builtin RSQ apparently clamps negative values to 0 instead of
returning the RSQ of the absolute value like ARB_fragment_program
desires and pre-gen6 apparently does.
Fixes:
glean/fp1-RSQ test 2 (reciprocal square root of negative value)
glean/vp1-RSQ test 2 (reciprocal square root of negative value)
Fixes this GCC warning.
r200_maos_arrays.c: In function 'r200EmitArrays':
r200_maos_arrays.c:113: warning: 'emitsize' may be used uninitialized in this function
It's not always possible to preprocess the content of 3D_LOAD_VBPNTR
in a command buffer, because the offset to all vertex buffers (which
the packet depends on) is derived from the "start" parameter of draw_arrays
and the "indexBias" parameter of draw_elements, but we can at least lazily
make a command buffer for the case when offset == 0, which should occur
most of the time.
The original lazy built-in importing patch did not add the newly created
function to the symbol table, nor actually emit it into the IR stream.
Adding it to the symbol table is non-trivial since importing occurs when
generating some ir_call in a nested scope. A new add_global_function
method, backed by new symbol_table code in the previous patch, handles
this.
Fixes bug #32030.
Avoid rebuilding constant shader state at each draw call,
factor out spi update that might change at each draw call.
Best would be to update spi only when revealent states
change (likely only flat shading & sprite point).
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Vertex elements change are less frequent than draw call, those to
avoid rebuilding fetch shader to often build the fetch shader along
vertex elements. This also allow to move vertex buffer setup out
of draw path and make update to it less frequent.
Shader update can still be improved to only update SPI regs (based
on some rasterizer state like flat shading or point sprite ...).
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
More than 1023 temporaries were being used for a Cinebench shader before
doing temporary optimization, causing the index value to wrap around to
-1024.
In st_finalize_texture() we were looking at the st_texture_object::
lastLevel field instead of the pipe_resource::last_level field to
determine which resource to store the mipmap in.
Then, in st_generate_mipmap() we need to call st_finalize_texture() to
make sure the destination resource is properly allocated.
These changes fix the broken piglit fbo-generatemipmap-formats test.
When X or Y or Z is close to W the outcome of the floating point clip
test comparision may be different between the C and x86 asm paths.
That's OK; don't report an error.
See fd.o bug 32093
It was only used for gen6 fragment programs (not GLSL shaders) at this
point, and it was clearly unsuited to the task -- missing opcodes,
corrupted texturing, and assertion failures hit various applications
of all sorts. It was easier to patch up the non-glsl for remaining
gen6 changes than to make brw_wm_glsl.c complete.
Bug #30530
Since the 8-wide first-quarter and 16-wide first-half have the same
bit encoding, we now need to track "do you want instruction
compression" in the compile state.
The FS backend is fine with register level granularity. But for the
brw_wm_emit.c backend, it expects pairs of regs to be used for the
constants, because the whole world is pairs of regs. If an odd number
got used, we went looking for interpolation in the wrong place.
In the SF and brw_fs.cpp fixes to set up interpolation sanely on gen6,
the setup for 16-wide interpolation was left behind. This brings
relative sanity to that path too.
Payload reg setup on gen6 depends more on the dispatch width as well
as the uses_depth, computes_depth, and other flags. That's something
we want to decide at compile time, not at cache lookup. As a bonus,
the fragment shader program cache lookup should be cheaper now that
there's less to compute for the hash key.
The preprocessor magic in mapi was nothing but obfuscation. Rewrite
mapi_abi.py to generate real C code.
This commit removes the hack added in
43121f2086.
Need to check the required primitive type for GS on Sandybridge,
and when GS is disabled, the new state has to be issued too, instead
of only updating URB state with no GS entry, that caused hang on Sandybridge.
This fixes hang issue during conformance suite testing.
This fixes endless vertex shader recompilations in find_translated_vp
if the shader contains an edge flag output.
NOTE: This is a candidate for the 7.9 branch.
Signed-off-by Brian Paul <brianp@vmware.com>
Finished up by Marek Olšák.
We can set the constant space to use a different area per-call to the shader,
we can avoid flushing the PVS as often as we do by spreading out the constants
across the whole constant space.
Signed-off-by: Marek Olšák <maraeo@gmail.com>
It appears to be a constant buffer index (in case there are more constant
buffers explicitly used by a shader), i.e. something that Gallium currently
does not use. We treated it incorrectly as the offset to a constant buffer.
Sometimes I'm on the train and want to just read what's generated
under INTEL_DEBUG=vs,wm for some code on another generation. Or, for
the next gen enablement we'll want to dump aub files before we have
the actual hardware. This will let us do that.
rgb_src_factor and rgb_dst_factor should be PIPE_BLENDFACTOR_ONE for
VG_BLEND_SRC_IN and VG_BLEND_DST_IN respectively. VG_BLEND_SRC_OVER can
be supported only when the fb has no alpha channel. VG_BLEND_DST_OVER
and VG_BLEND_ADDITIVE have to be supported with a shader.
Note that Porter-Duff blending rules assume premultiplied alpha.
Convert color values to and back from premultiplied form for blending.
Finally the rendering result of the blend demo looks much closer to that
of the reference implementation.
Drawing an image in VG_DRAW_IMAGE_STENCIL mode produces per-channel
alpha for use in blending. Add a new shader stage to produce and save
it in TEMP[1].
For other modes that do not need per-channel alpha, the stage does
MOV TEMP[1], TEMP[0].wwww