This was a hack to avoid choosing to schedule all texturing before
consumption of any texture results due to the way dependency chains worked
out in the presence of MRFs. On gen7, we don't have MRFs, so the problem
doesn't apply, and this was just badly constraining our scheduling.
total instructions in shared programs: 1615306 -> 1612534 (-0.17%)
instructions in affected programs: 9958 -> 7186 (-27.84%)
GAINED: 259
LOST: 9
Reviewed-by: Matt Turner <mattst88@gmail.com>
The LIFO plan was simple: Take the most recently made available
instructions, and pick those first.
But because of the order we were pushing things onto our list of
available-to-schedule instructions, it meant that when a set of
instructions was made available at the same time (for example, everything
at the start of the program that didn't depend on other instructions) we'd
schedule them in reverse order.
If you had 10 texture calls in a row in your program, each with
independent argument setup, we'd set up the last texture call's args and
execute it first, even though we wouldn't be able to consume its results
until we'd finished the other 9 texture calls (assuming consumption of
texture results happens near each texture call, and combines it with
another texture result, which is normal for a convolution shader).
To fix this, walk the list for doing LIFO in the order that instructions
were originally generated in the program, but choose to push
newly-made-available instructions to the other end of the list instead.
total instructions in shared programs: 1587242 -> 1586290 (-0.06%)
instructions in affected programs: 7801 -> 6849 (-12.20%)
GAINED: 76
LOST: 67
Thanks to Chia-I Wu for pointing out the bug in my first version of the
patch that made it a huge loss.
Reviewed-by: Matt Turner <mattst88@gmail.com>
When PIPE_CAP_MIXED_FRAMEBUFFER_SIZES is not provided, parts of
ARB_framebuffer_object can't be supported, such as on NV30.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
This CAP will determine whether ARB_framebuffer_object can be enabled.
The nv30 driver does not allow mixing swizzled and linear zsbuf/cbuf
textures.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
In commit 1b4a737 (glsl: Support redeclaration of VS and GS
gl_PerVertex output), I added code to ensure that when an unnamed
gl_PerVertex interface block is redeclared, any ir_variables that
weren't included in the redeclaration are removed from the IR (and the
symbol table). This ensures that only those variables that were
explicitly redeclared may be used.
However, when I wrote this code, I neglected to match the variable
mode when finding variables to remove. This meant that redeclaring a
built-in output block might cause the built-in input gl_in to be
accidentally removed.
Fixes piglit test gs-redeclares-pervertex-out-only.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The GLSL 4.10 rules for redeclaration of built-in interface blocks
(which we've chosen to regard as clarifications of GLSL 1.50) only
require gl_PerVertex blocks to match in shaders that actually use
those blocks. The easiest way to implement this is to detect
situations where a compiled shader doesn't refer to any elements of
gl_PerVertex, and remove all the associated ir_variables from the
shader at the end of ast-to-ir conversion.
Fixes piglit tests
linker/interstage-{pervertex,pervertex-in,pervertex-out}-redeclaration-unneeded.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Normally when a built-in array (such as gl_ClipDistance) is
redeclared, we call get_variable_being_redeclared() to do the
redeclaration, and it in turn calls check_builtin_array_max_size() to
make sure that the redeclared array size isn't too large.
However when a built-in array is redeclared as part of redeclaring
gl_in, we don't call get_variable_being_redeclared() (since the
individual built-ins aren't each represented by their own ir_variable
anymore). So we need to add an explicit call to
check_builtin_array_max_size() to make sure the new array size isn't
too large.
Note: at the moment this is redundant with a test that's done at link
time, so there's no change to piglit results. But the patch that
follows will prevent link errors from being reported if gl_PerVertex
isn't used, so in order to prevent that patch from causing
regressions, we need to add the compile check now. Besides, it's
nicer to report this error at compile time anyhow.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The queries GEOMETRY_VERTICES_OUT, GEOMETRY_INPUT_TYPE, and
GEOMETRY_OUTPUT_TYPE (defined by GL 3.2) differ from the corresponding
queries in ARB_geometry_shader4 in the following ways:
- They use different enum values
- They can only be queried; they cannot be set.
- Attempting to query them yields INVALID_OPERATION if the program is
not linked, or lacks a geometry shader.
This patch switches us over from the ARB_geometry_shader4 behaviour to
the GL 3.2 behaviour.
Fixes piglit test query-gs-prim-types.
v2: Improve comment above has_core_gs.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
When program_resource_visitor visits variables that were created by
lower_named_interface_blocks, it needs to do extra work to un-do the
effects of lower_named_interface_blocks and construct the proper API
names.
Fixes piglit test
spec/glsl-1.50/execution/interface-blocks-api-access-members.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
These variables will need to be treated specially by
program_resource_visitor, so that they can be addressed through the
API using their interface block name (and array index, for interface
block arrays).
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Later patches will use this information to do proper error checking of
interpolation qualifiers that appear inside of interface blocks.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
In future patches, we will need this in order to interpret
interpolation qualifiers that appear inside interface blocks.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Although in principle there is no hardware limitation that prevents
gl_MaxGeometryInputComponents from being set to 128 on Gen7, we have
the following limitations in the vec4 compiler back end:
- Registers assigned to geometry shader inputs can't be spilled or
later re-used for any other purpose.
- The last 16 registers are set aside for the "MRF hack", meaning they
can only be used to send messages, and not for general purpose
computation.
- Up to 32 registers may be reserved for push constants, even if there
is sufficient register pressure to make this impractical.
A shader using 128 geometry input components, and having an input type
of triangles_adjacency, would use up:
- 1 register for r0 (which holds URB handles and various pieces of
control information).
- 1 register for gl_PrimitiveID.
- 102 registers for geometry shader inputs (17 registers per input
vertex, assuming DUAL_INSTANCED dispatch mode and allowing for one
register of overhead for gl_Position and gl_PointSize, which are
present in the URB map even if they are not used).
- Up to 32 registers for push constants.
- 16 registers for the "MRF hack".
That's a total of 152 registers, which is well over the 128 registers
the hardware supports.
Fortunately, the GLSL 1.50 spec allows us to reduce
gl_MaxGeometryInputComponents to 64. Doing that frees up 48
registers, brining the total down to 104 registers, leaving 24
registers available to do computation.
Fixes piglit test
spec/glsl-1.50/execution/geometry/max-input-components.
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This is similar to what we do for 16-wide vs 8-wide fragment shaders.
First we try compiling the geometry shader in DUAL_OBJECT mode. If we
can't do that without spilling, we fall back on DUAL_INSTANCED mode,
which should require less spilling (since it uses an interleaved
layout of payload registers).
In an ideal world we'd fall back to SINGLE mode, which would allow us
to interleave general-purpose registers too (resulting in even less
likelihood of spilling). But at the moment, the vec4 generator and
visitor classes don't have the infrastructure to interleave general
purpose registers, so DUAL_INSTANCED is the best we can do.
As a side benefit this paves the way for implementing instanced
geometry shaders (which are incompatible with DUAL_OBJECT mode).
Since most geometry shaders used in piglit testing are small,
DUAL_INSTANCED mode won't get exercised very much in a normal piglit
run. To force DUAL_INSTANCED mode to be used for all geometry
shaders, set INTEL_DEBUG=nodualobj.
Reviewed-by: Eric Anholt <eric@anholt.net>
Geometry shaders that run in "DUAL_INSTANCED" mode store their inputs
in vec4's. This means that when compiling gl_PointSize input
swizzling (a MOV instruction which uses a geometry shader input as
both source and destination), we need to do two things:
- Set force_writemask_all to ensure that the MOV happens regardless of
which channels are enabled.
- Set the source register region to <4;4,1> (instead of <0;4,1> to
satisfy register region restrictions.
v2: move the source register region fixup to the top of
vec4_generator::generate_vec4_instruction(), so that it applies to all
instructions rather than just MOV.
Reviewed-by: Eric Anholt <eric@anholt.net>
In future patches, this will allow us to first try compiling a
geometry shader in DUAL_OBJECT mode (which is more efficient but uses
more registers) and then if spilling is required, fall back on
DUAL_INSTANCED mode.
Reviewed-by: Eric Anholt <eric@anholt.net>
Otherwise the scheduler would be invoked with prog_data->total_grf ==
0, causing havoc.
In a future patch, this will allow us to try compiling a geometry
shader in DUAL_OBJECT mode with spilling disabled, and then fall back
to DUAL_INSTANCED mode if that failed.
Reviewed-by: Eric Anholt <eric@anholt.net>
When geometry shaders are operated in "single" or "dual instanced"
mode, a single set of geometry shader inputs is interleaved into the
thread payload (with each payload register containing a pair of
inputs) in order to save register space.
This patch modifies vec4_visitor::lower_attributes_to_hw_regs so that
it can handle the interleaved format.
Reviewed-by: Eric Anholt <eric@anholt.net>
All geometry shaders begin this instruction:
mov(1) g0.2<1>:ud 0x0:ud { align1 }
which sets up GRF0 properly for scratch reads and writes. Since this
instruction has a SIMD size of 1, it will only have an effect if the
first channel is enabled. In practice, the hardware seems to always
dispatch geometry shaders with the first channel enabled, but I can't
find anything in the docs to guarantee that.
So to be on the safe side, set force_writemask_all on the instruction,
which guarantees that it will have the desired effect regardless of
which channels are enabled.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
When lower_named_interface_blocks lowers a built-in interface block
member to an ir_variable, it needs to set explicit_location in the
ir_variable. Otherwise the linker gets confused and treats the
variable as a generic varying.
Fixes the following piglit tests, which were regressed by commit
63974c0 (glsl: Simplify the interface to
link_invalidate_variable_locations):
- clip-distance-bulk-copy
- clip-distance-in-bulk-read
- clip-distance-in-explicitly-sized
- clip-distance-in-param
- clip-distance-in-values
- core-inputs
- gs-redeclares-both-pervertex-blocks
- gs-redeclares-pervertex-in-only
- redeclare-pervertex-subset-vs-to-gs
- unsized-in-named-interface-block-gs
- unsized-in-named-interface-block-multiple
- unsized-in-unnamed-interface-block-gs
- unsized-in-unnamed-interface-block-multiple
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=70820
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This should never have been in the program key in the first place,
since it's determined by the shader source, not by GL state. Change
the code to just refer to gl_program::UsesClipDistanceOut directly.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This will make it easier for back-ends to share code between geometry
shader and vertex shader compilation. Also, it is renamed to
"UsesClipDistanceOut" to clarify that (a) in geometry shaders, it
refers to the gl_ClipDistance output rather than the gl_ClipDistance
input, and (b) it is irrelevant in fragment shaders.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Since gl_ClipDistance is lowered from an array of floats to an array
of vec4's during compilation, transform feedback has special logic to
keep track of the pre-lowered array size so that attempting to perform
transform feedback on gl_ClipDistance produces a result with the
correct size.
Previously, this special logic always consulted the vertex shader's
size for gl_ClipDistance. This patch fixes it so that it uses the
geometry shader's size for gl_ClipDistance when a geometry shader is
in use.
Fixes piglit test spec/glsl-1.50/transform-feedback-type-and-size.
v2: Change the type of LastClipDistanceArraySize to "unsigned", and
clarify the comment above it.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
We've always overriden
ctx->Const.{Vertex,Fragment}Program.MaxTextureImageUnits to reflect
the number of texture image units supported by the hardware (rather
than using the default values assigned by Mesa core) so it seems
sensible to do that for GeometryProgram.MaxTextureImageUnits too. We
set it to 0 if geometry shaders aren't supported.
Once that is done, we can just unconditionally add
GeometryProgram.MaxTextureImageUnits to MaxCombinedTextureImageUnits.
Fixes piglit test "spec/glsl-1.50/built-in
constants/gl_MaxCombinedTextureImageUnits".
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The encoding of constant, relative, and relative-const src registers is
a bit more complex than originally thought, which gives an extra bit to
encode const reg # at expense of taking a bit from relative offset.
In most cases a3xx seems to actually use a scheme whereby it can encode
an extra bit for const register. You have three possible encodings in
thirteen bits:
register: (11 bits for N.c)
00........... rN.c
relative: (10 bits for N)
010.......... r<a0.x + N>
011.......... c<a0.x + N>
const: (12 bits for N.c)
1............ cN.c
Which means we can deal w/ more consts than previously thought.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
d3d10 requires that cube corners are filtered with accurate weights (that
is, the weight of the non-existing corner texel should be evenly distributed
to the other 3 texels). OpenGL does not require this (but recommends it).
This requires us to use different filtering code, since we need per-texel
weights which our 2d lerp doesn't (and can't) do. And of course the (now
per element) weights need to be adjusted too for it to work.
Invoke the new filtering code whenever there's an edge to keep things simpler,
as it will work for edges too not just corners but of course it's only needed
with corners.
More ugly code for not much gain but at least a hacked up cubemap demo
shows very nice corners now... Not sure yet if and how this should be
configurable...
v2: incorporate feedback from Jose, only use special corner filtering code
when there's a corner not when there's only an edge (as corner filtering code
is slower, though a perf difference was only measureable when always
forcing edge code). Plus some minor style fixes.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
No driver uses it any more, and it's been replaced by megadrivers.
v2: Remove always-on conditional for NEED_LIBPROGRAM (review by Emil)
Reviewed-by: Matt Turner <mattst88@gmail.com> (v1)
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
v2: drop dridir now that it's unused.
v3: Fix linking after rebase when building just swrast from classic but a
drm-using gallium driver.
v4: Consistently put spaces around += in the updated Makefile.am block.
v5: Set a global driverAPI variable so loaders don't have to update to
createNewScreen2() (though they may want to for thread safety).
Reviewed-by: Matt Turner <mattst88@gmail.com> (v3)
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
This required some reordering of headers to ensure that the symbol name
redefines happened before any prototypes.
v2: drop dridir now that it's unused.
v3: Consistently put spaces around += in the updated Makefile.am blocks.
v4: Set a global driverAPI variable so loaders don't have to update to
createNewScreen2() (though they may want to for thread safety).
Reviewed-by: Matt Turner <mattst88@gmail.com> (v2)
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
i915 has symbols for formerly-shared code that conflict with i965, so we
define them away using gen-symbol-redefs.py. Options considered:
- This option. Downsides: The symbols in profiling and debugging don't
match the source. The symbol list may change in the future and we won't
notice without manually running the tool again.
- Use objcopy --localize-hidden to automatically demote our symbols to
locals. This didn't work on i965 due to c++ weak symbols (which can't
be localized), but could work on i915. We could do it on i915 only, but
it does produce libtool warnings at link time due to libtool not knowing
if the resulting .o file is safe to link (stupid libtool). Plus you end
up with different symbols of the same name, which is confusing for
debugging too. On the other hand, no future symbol conflicts long term.
- Write our own libelf tool that handles c++ weak symbols like we want and
apply it to all drivers. All the downsides of above, but applies
uniformly across drivers.
- Edit the files to just rename all the i915 or i965 symbols that
conflict. There are on the order of 100 that have a prefix we used to
share, so it would take a bit of typing. Fewest downsides, but still
can have conflicts long term.
Ultimately, this is the least invasive change at the moment, and we can
see if the "more symbol conflicts appear later" thing is a real concern or
not.
Note that the ability to compile a version of i915 without INTEL_DEBUG env
support is dropped. It's too useful.
v2: drop dridir now that it's unused.
v3: Consistently put spaces around += in the updated Makefile.am block.
v4: Set a global driverAPI variable so loaders don't have to update to
createNewScreen2() (though they may want to for thread safety).
Reviewed-by: Matt Turner <mattst88@gmail.com> (v2)
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
v2: drop dridir now that it's unused.
v3: Consistently put spaces around += in the updated Makefile.am block.
v4: Set a global driverAPI variable so loaders don't have to update to
createNewScreen2() (though they may want to for thread safety).
v5: Fix missed public symbol in nouveau. (caught by Emil)
Reviewed-by: Matt Turner <mattst88@gmail.com> (v2)
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
Previously, we've split things such that mesa core is in libdricore,
exposing the whole Mesa core interface in the global namespace, and the
i965_dri.so code all links against that. Along with polluting application
namespace terribly, it requires extra PLT indirections and prevents LTO.
Instead, we can build all of the driver contents into the same .so with
just a few symbols exposed to be referenced from the actual driver .so
file, allowing LTO and reducing our exposed symbol count massively.
FPS improvement on GLB2.7 with INTEL_NO_HW=1: 2.61061% +/- 1.16957% (n=50)
(without LTO, just the PLT reductions from this commit)
Note that the X Server requires commit
7ecfab47eb221dbb996ea6c033348b8eceaeb893 to successfully load this driver!
v2: Set a global driverAPI variable so loaders don't have to update to
createNewScreen2() (though they may want to for thread safety).
v3: Drop AM_CPPFLAGS addition (Emil pointed out I'd missed some cflags
that would be necessary, though only if we actually relied on them).
v4: Fix install with DESTDIR set.
Reviewed-by: Matt Turner <mattst88@gmail.com> (v1)
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com> (v2)
