We can bundle two load/store together. This eliminates the need for
explicit load/store pairing in a prepass, as well.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Conditions for branches don't have a swizzle explicitly in the emitted
binary, but they do implicitly get swizzled in whatever instruction
wrote r31, so we need to handle that.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Conditional instructions (csel and conditional branches) require their
condition to be written to a special condition pipeline register (r31.w
for scalar, r31.xyzw for vector). However, pipeline registers are live
only for the duration of a single bundle. As such, the logic to schedule
conditionals correct is surprisingly complex. Essentially, we see if we
could stuff the conditional within the same bundle as the csel/branch
without breaking anything; if we can, we do that. If we can't, we add a
dummy move to make room.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
A bit of a kludge but allows setting an implicit dependency of synthetic
conditional moves on the actual condition, fixing code generated like:
vmul.feq r0, ..
sadd.imov r31, .., r0
vadd.fcsel [...]
The imov runs simultaneous with feq so it gets garbage results, but it's
too late to add an actual dependency practically speaking, since the new
synthetic imov doesn't have a node associated.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
In the future, we will want to keep track of which components of
constants of various sizes correspond to which parts of the bundle
constants, like in the old scheduler. For now, let's just stub it out
for a simple rule of one instruction with embedded constants per bundle.
We can eventually do better, of course.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
We don't actually do any scheduling here yet, but add per-tag helpers to
consume an instruction, print it, pop it off the worklist.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
It's not always obvious what the optimal bundle type should be. Let's
break out the logic to decide.
Currently set for purely in-order operation.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
After we've chosen an instruction, popped it off, and processed it, it's
time to update the worklist, removing that instruction from the
dependency graph to allow its dependents to be put onto the worklist.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
In the future, this routine will implement the core scheduling logic to
decide which instruction out of the worklist will be scheduled next, in
a way that minimizes cycle count and register pressure.
In the present, we are more interested in replicating in-order
scheduling with the much-more-powerful out-of-order model. So rather
than discriminating by a register pressure estimate, we simply choose
the latest possible instruction in the worklist.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
We would like to flatten a linked list of midgard_instructions into an
array of midgard_instruction pointers on the heap.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
It's not based on the writemask and it can't be inferred; it's just
intrinsic to the op itself.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
This worked better than my original v3d-local pass for just subs, and is a
huge win over not producing subs.
total instructions in shared programs: 6408469 -> 6167932 (-3.75%)
total threads in shared programs: 153784 -> 154104 (0.21%)
total uniforms in shared programs: 2157078 -> 1905823 (-11.65%)
total max-temps in shared programs: 904546 -> 895796 (-0.97%)
total spills in shared programs: 4959 -> 4993 (0.69%)
total fills in shared programs: 6558 -> 6670 (1.71%)
total sfu-stalls in shared programs: 25845 -> 25175 (-2.59%)
total inst-and-stalls in shared programs: 6434314 -> 6193107 (-3.75%)
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
There are some optimizations which are only implemented for additions
and some optimizations which assume that subtractions have been lowered.
By lowering all subtractions first and later recombine for backends
which prefer this option, we don't have to implement them twice.
This patch also moves lower_negate to nir_opt_algebraic_late() to enable
these optimizations for backends which make use of it.
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Upcoming changes to sub optimization will make this pass required. Over
the course of that series, we see uniforms +.46%, instructions -.24%
(seems like a fine tradeoff -- uniforms are 1/2 the size of instructions
as far as cache occupancy)
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Without this, it was theoretically possible for the jobs to run before
the docker image was ready.
v2:
* Use - list syntax instead of [] (Eric Engestrom)
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
This allows most build jobs to run before the stretch or arm64 docker
images are ready.
v2:
* Use - list syntax instead of [] (Eric Engestrom)
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
This allows the *-old-llvm jobs to run before the buster docker images
are ready.
v2:
* Use - list syntax instead of [] (Eric Engestrom)
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This patch is based on 28e3f85e09/mingw-w64-mesa/link-ole32.patch but with tweaks to avoid MSVC build break when applied.
v2: Create Mingw platform alias pointing to windows host platform define to avoid spurious crosscompilation;
v3: Fix obviously wrong compiler flags for swr driver;
v4: Update original patch URL because it has been relocated;
v5: Don't bother patching autools stuff as it's not used by MSYS2 Mingw-w64 build and it's days are numbered anyway;
v6: After Mingw posix flag fix in 295851eb things are far simpler as we don't need more linking of uuid, ole32, version and shell32 than what is already in place.
As X86AsmPrinter component is gone, LLVMX86AsmPrinter got replaced
with LLVMRemarks, LLVMBitstreamReader and LLVMDebugInfoDWARF.
Tests done with llvm-config on both LLVM 8 and 9 indicate that
mcjit, bitwriter and x86asmprinter fully fit inside engine component.
On other platforms and with meson build mcdisassembler was used to replace
X86AsmPrinter but mcdisassembler also fully fits inside engine component
for LLVM>=8 according to same tests.
v2: Avoid duplicating code related to Mingw pthreads.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Cc: 19.1 19.2 <mesa-stable@lists.freedesktop.org>
On 19.1 this patch does not apply cleanly without 88eb2a1f
Looks like blob uses following values for uniforms buffer:
0 for 8 bytes
1 for 16 bytes
2 for 24 bytes
2 for 32 bytes
3 for 40 bytes
3 for 48 bytes
3 for 56 bytes
3 for 64 bytes
4 for 72 bytes
It all looks like log2(size / 8) rounded up, so let's do the same.
Fixes: 931fc2a7b3f9("lima: do not set the PP uniforms address lowest bits")
Reviewed-by: Icenowy Zheng <icenowy@aosc.io>
Signed-off-by: Vasily Khoruzhick <anarsoul@gmail.com>
SCons 3.1 has moved to python 3, requiring this fix
to continue supporting scons builds.
Closes: #944
Cc: mesa-stable@lists.freedesktop.org
Acked-by: Eric Engestrom <eric@engestrom.ch>
Tested-by: Eric Engestrom <eric@engestrom.ch>
Android building rules are added in src/amd/Android.compiler.mk
libmesa_aco static library is built conditionally to radeonsi
as done for vulkan.radv module
This will prevent Android build errors for non x86 systems
filter-out compiler/aco_instruction_selection_setup.cpp source,
as already included by compiler/aco_instruction_selection.cpp
and would cause several multiple definition linker errors
NOTE: libLLVM requires AMDGPU Disassembler to build radv with aco
Fixes: 93c8ebf ("aco: Initial commit of independent AMD compiler")
Fixes: a70a998 ("radv/aco: Setup alternate path in RADV to support the experimental ACO compiler")
Signed-off-by: Mauro Rossi <issor.oruam@gmail.com>
Fixes a few building errors similar to the following:
In file included from external/mesa/src/amd/compiler/aco_instruction_selection.cpp:26:
In file included from external/libcxx/include/algorithm:639:
external/libcxx/include/utility:321:9:
error: implicit instantiation of undefined template 'std::__1::array<aco::Temp, 4>'
_T2 second;
^
Fixes: 93c8ebf ("aco: Initial commit of independent AMD compiler")
Signed-off-by: Mauro Rossi <issor.oruam@gmail.com>
Fixes the following piglit test: fragdepth_gles2 (for ETNA_MESA_DEBUG=nir)
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Fixes the following piglit test: fragdepth_gles2
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
It just makes thing more complicated for no reason.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Note it can still be improved a bit:
* Use alu swizzle to determine if src is scalar
* Take into account new immediates in the multiple uniform src lowering
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
This can improve performance by allowing the LAST_VARYING_2X bit to be
set when possible (and possibility more benefits on HALTI5 where the
number of components is set for each varying).
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
LOAD starts reading into the first enabled destination component, and
doesn't skip disabled components, so we need to allocate a destination with
contiguous components.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Only invert front facing when glFrontFace is GL_CW.
Fixes following deqp test:
dEQP-GLES2.functional.shaders.builtin_variable.frontfacing
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
The PP uniforms address register in render state is not a direct pointer
to the uniforms storage -- instead, it points to an one-item array, and
the array item is the real pointer to the uniforms storage.
This register reuses some of its LSBs as a size field. Currently the
size is set according to the length of the real uniforms storage.
However, as the register itself contains only a pointer to the one-item
array, the size field should be set to the length of the one-item array
and subtract it by 1, which means a fixed value of 0. That means we can
just omit it now.
Test shows this should be the correct approach to set this register.
Signed-off-by: Icenowy Zheng <icenowy@aosc.io>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
glsl 4.4 spec section '5.9 expressions':
"The operator is multiply (*), where both operands are matrices or one operand is a vector and the
other a matrix. A right vector operand is treated as a column vector and a left vector operand as a
row vector. In all these cases, it is required that the number of columns of the left operand is equal
to the number of rows of the right operand. Then, the multiply (*) operation does a linear
algebraic multiply, yielding an object that has the same number of rows as the left operand and the
same number of columns as the right operand. Section 5.10 “Vector and Matrix Operations”
explains in more detail how vectors and matrices are operated on."
This fix disallows a multiplication of incompatible matrices like:
mat4x3(..) * mat4x3(..)
mat4x2(..) * mat4x2(..)
mat3x2(..) * mat3x2(..)
....
CC: <mesa-stable@lists.freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111664
Signed-off-by: Andrii Simiklit <andrii.simiklit@globallogic.com>
According to the 1.1.123 spec:
"The implementation will attempt to create all pipelines, and only
return VK_NULL_HANDLE values for those that actually failed."
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>