pan/midgard: Use new scheduler

We still emit in-order but we switch to using the bundles created from
the new scheduler, which will allow greater flexibility and room for
out-of-order optimization.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Add distance metric to choose_instruction

We require chosen instructions to be "close", to avoid ballooning
register pressure. This is a kludge that will go away once we have
proper liveness tracking in the scheduler, but for now it prevents a lot
of needless spilling.

v2: Lower threshold to 6 (from 8). Schedule is hurt, but a few shaders
that spilled excessively are fixed.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Derp

pan/midgard: Add mir_choose_alu helper

Based on a given unit.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Implement load/store pairing

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>

pan/midgard: Extend csel_swizzle to branches

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>

pan/midgard: Add helpers for scheduling conditionals

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>

pan/midgard: Implement predicate->unit

This allows ALUs to select for each unit of the bundle separately.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Add predicate->exclude

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>

pan/midgard: Add constant intersection filters

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>

pan/midgard: Remove csel constant unit force

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Add mir_schedule_texture/ldst/alu helpers

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>

pan/midgard: Add mir_choose_bundle helper

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>

pan/midgard: Add mir_update_worklist helper

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>

pan/midgard: Add mir_choose_instruction stub

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>

pan/midgard: Initialize worklist

This flows naturally from the dependency graph

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Calculate dependency graph

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Add flatten_mir helper

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>

pan/midgard: Squeeze indices before scheduling

This allows node_count to be correct while scheduling.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Fix component count handling for ldst

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>

pan/midgard: Add missing parans in SWIZZLE definition

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

nouveau: set lower_sub = true

Subtractions are already implemented as additions anyway.

Reviewed-by: Connor Abbott <cwabbott0@gmail.com>

v3d: Enable the late algebraic optimizations to get real subs.

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>

aco: call nir_opt_algebraic_late() exhaustively

57559 shaders in 28980 tests
Totals:
SGPRS: 2963407 -> 2959935 (-0.12 %)
VGPRS: 2014812 -> 2016328 (0.08 %)
Spilled SGPRs: 1077 -> 1077 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 10348 -> 10348 (0.00 %) dwords per thread
Code Size: 114545436 -> 114498084 (-0.04 %) bytes
LDS: 933 -> 933 (0.00 %) blocks
Max Waves: 375997 -> 375866 (-0.03 %)

Reviewed-by: Connor Abbott <cwabbott0@gmail.com>

radv/aco: Don't lower subtractions

40228 shaders in 20236 tests
Totals:
SGPRS: 2045512 -> 2046496 (0.05 %)
VGPRS: 1430856 -> 1430464 (-0.03 %)
Spilled SGPRs: 1077 -> 1077 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 10348 -> 10348 (0.00 %) dwords per thread
Code Size: 77202840 -> 77151832 (-0.07 %) bytes
LDS: 863 -> 863 (0.00 %) blocks
Max Waves: 260729 -> 260754 (0.01 %)

Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>

nir: Remove unnecessary subtraction optimizations

These optimizations are already covered after lowering.

Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>

nir: recombine nir_op_*sub when lower_sub = false

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>

freedreno: Enable the nir_opt_algebraic_late() pass.

Reviewed-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>

vc4: Enable the nir_opt_algebraic_late() pass.

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>

gitlab-ci: Add test-container:arm64 to needs: for arm64 test jobs

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>

gitlab-ci: Add needs: for x86 buster docker image

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>

gitlab-ci: Declare needs: for stretch docker image

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>

scons: Fix MSYS2 Mingw-w64 build.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This patch is based on https://github.com/msys2/MINGW-packages/blob/28e3f85e09b6947ea80036c49f6c38f1394f93ca/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.

scons/windows: Support build with LLVM 9.

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

lima: set uniforms_address lower bits properly

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: add py3 support

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: aco: add support for libmesa_aco

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>

android: compiler/nir: build nir_divergence_analysis.c

Prerequisite to avoid following radv linking error happening with aco

FAILED: out/target/product/x86_64/obj_x86/SHARED_LIBRARIES/vulkan.radv_intermediates/LINKED/vulkan.radv.so
...
external/mesa/src/amd/compiler/aco_instruction_selection_setup.cpp:178:
error: undefined reference to 'nir_divergence_analysis'
clang.real: error: linker command failed with exit code 1 (use -v to see invocation)

Fixes: df86c5f ("nir: add divergence analysis pass.")
Signed-off-by: Mauro Rossi <issor.oruam@gmail.com>

android: aco: fix undefined template 'std::__1::array' build errors

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>

etnaviv: nir: fix gl_FragDepth

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>

etnaviv: disable earlyZ when shader writes fragment depth

Fixes the following piglit test: fragdepth_gles2

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>

etnaviv: nir: make lower_alu easier to follow

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>

etnaviv: remove extra allocation for shader code

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>

etnaviv: nir: remove "options" struct

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>

etnaviv: nir: use store_deref instead of store_output

Allows some simplification.

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>

etnaviv: nir: add native integers (HALTI2+)

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>

qetnaviv: nir: use new immediates when possible

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>

etnaviv: nir: set num_components for inputs/outputs

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>

etnaviv: nir: allocate contiguous components for LOAD destination

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>

etnaviv: nir: fix gl_FrontFacing

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>

lima: do not set the PP uniforms address lowest bits

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: disallow incompatible matrices multiplication

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>

turnip: Fix failure behavior of vkCreateGraphicsPipelines.

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>

turnip: Silence compiler warning about uninit pipeline.

The code was fine as far as I see, but the warning was irritating.

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>

turnip: Add a .editorconfig and .dir-locals.el

I was inheriting the one from src/freedreno with funny tabs, while
this driver is written with normal Mesa 3-space indents.
Unfortunately I have to add both files, because I use emacs and emacs
prefers .dir-locals to .editorconfig :(

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>

shader_enums: Move MAX_DRAW_BUFFERS to this file.

We include shader_enums.h from freedreno's compiler for both GL and
Vulkan, and the main/config.h include resulted in polluting the
namespace with things like MAX_VIEWPORTS that other Vulkan drivers use
as their driver-specific maximums.

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>

intel/fs: Fix fs_inst::flags_read for ANY/ALL predicates

Without this, we were DCEing flag writes because we didn't think their
results were used because we didn't understand that an ANY32 predicate
actually read all the flags.

Fixes: df1aec763eb "i965/fs: Define methods to calculate the flag..."
Reviewed-by: Matt Turner <mattst88@gmail.com>

etnaviv: support ARB_framebuffer_object

Passes most of piglit's tests regarding arb_framebuffer_object
and unlocks some more piglit tests.

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Jonathan Marek <jonathan@marek.ca>

etnaviv: etna_resource_copy_region(..): drop assert

We are using util_resource_copy_region(..) as fallback which supports
different formats for src and dst. Improves the experience when running
deqp or piglit with a debug build.

Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Jonathan Marek <jonathan@marek.ca>

meson: Link xvmc with libxv

Prior to xvmc 1.0.12 libxvmc incorrectly required libxv, but that was
fixed. This results in compilation failures for the gallium xvmc tracker
and tools. This patch fixes that by explicitly linking to libxv.

Fixes: 22a817af8a89eb3c762fc3e07b443a3ce37d7416
       ("meson: build gallium xvmc state tracker")
Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/1844
Reviewed-by: Adam Jackson <ajax@redhat.com>

meson: Try finding libxvmcw via pkg-config before using find_library

This fixes cross compiling issues, because pkg-config is less likely to
get the wrong libs.

v2: - Fix typo in comment

Fixes: 22a817af8a89eb3c762fc3e07b443a3ce37d7416
       ("meson: build gallium xvmc state tracker")
Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/939
Reviewed-by: Adam Jackson <ajax@redhat.com>

drisw: Fix shared memory leak on drawable resize

XDestroyImage will mark the segment as to-be-destroyed, but it will
persist until we detach it, and we weren't doing so.

Cc: mesa-stable@lists.freedesktop.org
Gitlab: https://gitlab.freedesktop.org/mesa/mesa/issues/121
Reviewed-by: Adam Jackson <ajax@redhat.com>

drisw: Fix and simplify drawable setup

We don't want to require a visual for the drawable, because there exist
fbconfigs that don't correspond to any visual (say a 565 pixmap|pbuffer
config on a depth-24 display). Fortunately, we don't need one either.
Passing the visual to XCreateImage serves only to fill in the XImage's
{red,green,blue}_mask fields, which libX11 itself never uses, they exist
only for the client's convenience, and we don't care. And we already
have the drawable depth in glx_config::rgbBits. So replace the
XVisualInfo field in the drawable private with a pointer to the
glx_config.

Having done that driswCreateGCs becomes trivial, so inline it into its
caller.

Gitlab: https://gitlab.freedesktop.org/mesa/mesa/issues/1194
Reviewed-by: Eric Anholt <eric@anholt.net>

drisw: Simplify GC setup

There's no reason to have two GCs here. The only difference between
them is that swapgc would generate graphics exposures, except we only
ever use this GC for PutImage, and PutImage doesn't generate graphics
exposures. We also don't need to explicitly ChangeGC to GXCopy, because
that's the default.

Reviewed-by: Eric Anholt <eric@anholt.net>

turnip: Add todo for d24_s8 copies

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

turnip: Disallow NPoT formats.

Copying is a mess for these formats for now.

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

turnip: Always use UINT formats for copies.

Looks like r16_unorm might have precision issues.

dEQP-VK.api.copy_and_blit.core.image_to_image.all_formats.color.r16_unorm.r16_unorm.general_general

fails, but the dumped images in the xml are the same so
I'd guess the low bits are the issue.

r8_unorm and r16_uint work.

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

turnip: Add image->image blitting.

3D blits & format reinterpretation are still TBD.

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

aco: don't remove the loop exec mask in transition_to_Exact()

No pipeline-db changes.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>

aco: set loop_info::has_discard for demotes

We need the loop header phis for the outer exec masks. Needed for
dEQP-VK.glsl.demote.dynamic_loop_texture

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>

iris: Only resolve for image levels/layers which are actually in use.

There's no need to resolve everything.

lima/ppir: add NIR pass to split varying loads

NIR may emit a single instrinsic to load several packed varyings,
but that's suboptimal for Utgard PP for several reasons:
- varyings that are used as sampler inputs can be passed using
  pipeline register with increased precision
- we have small number of regs, so using a vec4 regs for storing
  two vec2 varyings increases reg pressure.

Add NIR pass to split a single load into several loads and utilize
it in lima.

Reviewed-by: Qiang Yu <yuq825@gmail.com>
Signed-off-by: Vasily Khoruzhick <anarsoul@gmail.com>

radv: Fix L2 cache rinse programming.

According to radeonsi, GLM doesn't support WB alone, so
we have to set INV too when WB is set.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

turnip: emit texture and uniform state

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

turnip: add some shader information in pipeline state

This information is needed by texture/uniform descriptors.

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

turnip: use nir_opt_copy_prop_vars

Avoids getting a "load_output" in a case like this:

   gl_Position = ubuf.MVP * ubuf.position[gl_VertexIndex];
   frag_pos = gl_Position.xyz;

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

turnip: lower samplers and uniform buffer indices

Lower these to something compatible with ir3, and save the descriptor set
and binding information.

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

turnip: basic descriptor sets (uniform buffer and samplers)

Mostly copy-paste from radv, with a few modifications.

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

turnip: enable linear filtering

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

turnip: align layer_size

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

turnip: use linear tiling for scanout image

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

turnip: implement image view descriptor

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

turnip: implement sampler state

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

turnip: fix vertex_id

ir3 uses non-zero based vertex id for a6xx

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

turnip: emit shader immediates

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Acked-by: Eric Anholt <eric@anholt.net>

util/rb_tree: Stop relying on &iter->field != NULL

The old version of the iterators relies on a &iter->field != NULL check
which works fine on older GCC but newer GCC versions and clang have
optimizations that break if you do pointer math on a null pointer.  The
correct solution to this is to do the null comparisons before we do any
sort of &iter->field or use rb_node_data to do the reverse operation.

Acked-by: Michel Dänzer <mdaenzer@redhat.com>
Tested-by: Michel Dänzer <mdaenzer@redhat.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

util/rb_tree: Also test _safe iterators

Acked-by: Michel Dänzer <mdaenzer@redhat.com>
Tested-by: Michel Dänzer <mdaenzer@redhat.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

freedreno/a3xx: Mostly fix min-vs-mag filtering decisions on non-mipmap tex.

This is based on the fix I used for the same problem on V3D.  In this
case, it fixes all but the the
dEQP-GLES2.functional.texture.filtering.2d.*_npot cases of
dEQP-GLES2.functional.texture.filtering.2d.*'s failures.

Acked-by: Rob Clark <robdclark@chromium.org>

intel/compiler: avoid truncating int64_t to int

Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Maya Rashish <maya@netbsd.org>

lima: support rectangle texture

As Vasily discovered, the bit 7 of the word 1 of the texture descriptor
is set when reloading the framebuffer, to use framebuffer-based offset
rather than normalized one. This bit also works for regular textures to
enable accessing with non-normalized offset.

Add support for rectangle texture by setting this bit for
PIPE_TEXTURE_RECT.

Suggested-by: Vasily Khoruzhick <anarsoul@gmail.com>
Signed-off-by: Icenowy Zheng <icenowy@aosc.io>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>

loader: Avoid use-after-free / use of uninitialized local variables

Per the valgrind output below, we were returning the pointer to freed
memory if none of the later conditional pointer assignments were
executed. This caused dEQP CI jobs to crash on certain runners,
presumably due to a double-free down the line.

Also, we were skipping to the out: label before the vendor_id & chip_id
variables used by it were initialized, resulting in broken
LIBGL_DEBUG=verbose output such as

libGL: pci id for fd 4: 51108f00:51108f00, driver radeonsi

Fixes: 5a545e355b23 "loader: always map the "amdgpu" kernel driver name to radeonsi (v2)"
==403== Invalid read of size 1
==403==    at 0x4AFD576: surfaceless_probe_device (platform_surfaceless.c:316)
==403==    by 0x4AFD915: dri2_initialize_surfaceless (platform_surfaceless.c:391)
==403==    by 0x4AF5EEA: dri2_initialize (egl_dri2.c:984)
==403==    by 0x4AF5EEA: dri2_initialize (egl_dri2.c:958)
==403==    by 0x4AF1EEC: _eglMatchAndInitialize (egldriver.c:75)
==403==    by 0x4AF1F3B: _eglMatchDriver (egldriver.c:96)
==403==    by 0x4AE9367: eglInitialize (eglapi.c:617)
==403==    by 0x1D99C9: tcu::surfaceless::EglRenderContext::EglRenderContext(glu::RenderConfig const&, tcu::CommandLine const&) [clone .constprop.57] (in /deqp/modules/gles2/deqp-gles2)
==403==    by 0x1DABB0: tcu::surfaceless::ContextFactory::createContext(glu::RenderConfig const&, tcu::CommandLine const&, glu::RenderContext const*) const (in /deqp/modules/gles2/deqp-gles2)
==403==    by 0x53EBD1: glu::createRenderContext(tcu::Platform&, tcu::CommandLine const&, glu::RenderConfig const&, glu::RenderContext const*) (in /deqp/modules/gles2/deqp-gles2)
==403==    by 0x53EFE9: glu::createDefaultRenderContext(tcu::Platform&, tcu::CommandLine const&, glu::ApiType) (in /deqp/modules/gles2/deqp-gles2)
==403==    by 0x1DE07A: deqp::gles2::Context::Context(tcu::TestContext&) (in /deqp/modules/gles2/deqp-gles2)
==403==    by 0x1DB5EF: deqp::gles2::TestPackage::init() (in /deqp/modules/gles2/deqp-gles2)
==403==  Address 0x56bd340 is 0 bytes inside a block of size 4 free'd
==403==    at 0x48369AB: free (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
==403==    by 0x4B01767: loader_get_driver_for_fd (loader.c:464)
==403==    by 0x4AFD553: surfaceless_probe_device (platform_surfaceless.c:308)
==403==    by 0x4AFD915: dri2_initialize_surfaceless (platform_surfaceless.c:391)
==403==    by 0x4AF5EEA: dri2_initialize (egl_dri2.c:984)
==403==    by 0x4AF5EEA: dri2_initialize (egl_dri2.c:958)
==403==    by 0x4AF1EEC: _eglMatchAndInitialize (egldriver.c:75)
==403==    by 0x4AF1F3B: _eglMatchDriver (egldriver.c:96)
==403==    by 0x4AE9367: eglInitialize (eglapi.c:617)
==403==    by 0x1D99C9: tcu::surfaceless::EglRenderContext::EglRenderContext(glu::RenderConfig const&, tcu::CommandLine const&) [clone .constprop.57] (in /deqp/modules/gles2/deqp-gles2)
==403==    by 0x1DABB0: tcu::surfaceless::ContextFactory::createContext(glu::RenderConfig const&, tcu::CommandLine const&, glu::RenderContext const*) const (in /deqp/modules/gles2/deqp-gles2)
==403==    by 0x53EBD1: glu::createRenderContext(tcu::Platform&, tcu::CommandLine const&, glu::RenderConfig const&, glu::RenderContext const*) (in /deqp/modules/gles2/deqp-gles2)
==403==    by 0x53EFE9: glu::createDefaultRenderContext(tcu::Platform&, tcu::CommandLine const&, glu::ApiType) (in /deqp/modules/gles2/deqp-gles2)
==403==  Block was alloc'd at
==403==    at 0x483577F: malloc (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
==403==    by 0x4EE5E09: strndup (strndup.c:43)
==403==    by 0x4B010B1: loader_get_kernel_driver_name (loader.c:101)
==403==    by 0x4B016AF: loader_get_driver_for_fd (loader.c:462)
==403==    by 0x4AFD553: surfaceless_probe_device (platform_surfaceless.c:308)
==403==    by 0x4AFD915: dri2_initialize_surfaceless (platform_surfaceless.c:391)
==403==    by 0x4AF5EEA: dri2_initialize (egl_dri2.c:984)
==403==    by 0x4AF5EEA: dri2_initialize (egl_dri2.c:958)
==403==    by 0x4AF1EEC: _eglMatchAndInitialize (egldriver.c:75)
==403==    by 0x4AF1F3B: _eglMatchDriver (egldriver.c:96)
==403==    by 0x4AE9367: eglInitialize (eglapi.c:617)
==403==    by 0x1D99C9: tcu::surfaceless::EglRenderContext::EglRenderContext(glu::RenderConfig const&, tcu::CommandLine const&) [clone .constprop.57] (in /deqp/modules/gles2/deqp-gles2)
==403==    by 0x1DABB0: tcu::surfaceless::ContextFactory::createContext(glu::RenderConfig const&, tcu::CommandLine const&, glu::RenderContext const*) const (in /deqp/modules/gles2/deqp-gles2)

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>

Revert "glx: Lift sending the MakeCurrent request to top-level code"

Apparently this provokes crashes elsewhere in code unrelated to
MakeCurrent. I hate GLX so very very much.

This reverts commit 999c2aed8826f403b071f52b040ce25b56d35f9d.

Gitlab: https://gitlab.freedesktop.org/mesa/mesa/issues/1207

Revert "glx: Implement GLX_EXT_no_config_context"

This reverts commit 0d635ccc912d7122f35f81eec27d8b2c0a2a7a28.

Gitlab: https://gitlab.freedesktop.org/mesa/mesa/issues/1207

radv: Add debug option to dump meta shaders.

This new option can help debug shader compiler problems when
there are issues with the meta shaders.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

amd/common: Introduce ac_get_fs_input_vgpr_cnt.

Add a function called ac_get_fs_input_vgpr_cnt which will return
the number of input VGPRs used by an AMD shader. Previously,
radv and radeonsi had the same code duplicated, but this commit also
allows them to share this code.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radv: Set shared VGPR count in radv_postprocess_config.

This commit allows RADV to set the shared VGPR count according to
the shader config.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

amd/common: Add num_shared_vgprs to ac_shader_config for GFX10.

In GFX10 wave64 mode, shared VGPRs allow the two wave halves to
share some data with each other.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

amd/common: Extract some helper functions to ac_shader_util.

This commit moves ac_get_tbuffer_format, ac_get_sampler_dim and
ac_get_image_dim into ac_shader_util, thus enabling them to be used
by compilers other than LLVM.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

amd/common: Move ac_export_mrt_z to ac_llvm_build.

The aim of this commit is to keep ac_shader_util LLVM-free,
since we would like to use it in ACO later.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

aco: CSE readlane/readfirstlane/permute/reduce with the same exec mask

v2: rename pass_temp to pass_flags
v2: also CSE reductions
v3: add ds_swizzle_b32 support
v3: check gds/offset0/offset1 fields

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>

aco: don't CSE v_readlane_b32/v_readfirstlane_b32

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>