lima/ppir: add lod-bias support

Signed-off-by: Arno Messiaen <arnomessiaen@gmail.com>
Reviewed-by: Erico Nunes <nunes.erico@gmail.com>

Revert "i965/fs: Merge CMP and SEL into CSEL on Gen8+"

This reverts commit 52c7df1643ec9af119fd66f916f7fbdbcc798d2d.  The pass,
while clearly useful for some shaders, has at least three bugs that I
was able to find fairly quickly:

 1. It doesn't work for type-converting MOVs because f > 0 is not the
    same as f2i(f) > 0

 2. CSEL is a 3src instruction and only supports one source type; it
    doesn't take this into account and tries to create instructions
    which do a F compare and a D select.  This is especially nasty to
    debug because you don't see that in the dumped assembly because we
    don't properly assert that types are the same in codegen.

 3. While you can handle 2, in theory, by reinterpreting types, you
    can't do that in the presence of source modifiers.  This pass
    doesn't even attempt to detect that.

Those are just the ones I found with the one almost trival shader I was
debugging.  There very likely may be more and.  Best thing to do for now
is just shut it off until someone has the time to figure out how to do
this properly and write tests to ensure it's correct.

Fixes: 3cb085e6d61a "i965/fs: Merge CMP and SEL into CSEL on Gen8+"
Reviewed-by: Brian Paul <brianp@vmware.com>

radv: Enable Subgroup Arithmetic and Clustered for SI

This patch also allows to enable VK_AMD_shader_ballot on SI.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

amd/llvm: Add Subgroup Scan functions for SI

The idea of this implementation is taken from the ROCm Device Libs:
https://github.com/RadeonOpenCompute/ROCm-Device-Libs/blob/master/ockl/src/wfredscan.cl

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

lima/streamparser: Add findings introduced with gl_PointSize

Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Signed-off-by: Andreas Baierl <ichgeh@imkreisrum.de>

lima/streamparser: Fix typo in vs semaphore parser

Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Signed-off-by: Andreas Baierl <ichgeh@imkreisrum.de>

meson: Fix linkage of libgallium_nine with libgalliumvl

Do not link libgallium_nine with libgalliumvl_stub if it's already
linked with libgalliumvl. Linking with stub leads to "duplicate
symbol" errors.

Fixes: 6b4c7047d57178d3362a710ad503057c6a582ca3
       ("meson: build gallium nine state_tracker")
Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/2040
Signed-off-by: Yevhenii Kolesnikov <yevhenii.kolesnikov@globallogic.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>

docs/release-calendar: Update for extended 19.3 rc period

docs: update calendar, add news item and link release notes for 19.2.5

docs/relnotes/19.2.5: Add SHA256 sum

docs: Add relnotes for 19.2.5

nir/large_constants: use nir_index_vars and nir_variable::index

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>

nir: add nir_variable::index and nir_index_vars

This will be useful as a deterministic identifier/index for the variable.

v2: fix comment style

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com> (v1)

nir: make nir_variable::{num_members,num_state_slots} a uint16_t

Doesn't shrink it (at least, on x86-64) and leaves space for more members.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>

docs: add missing new features for RADV

[skip ci]

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

freedreno/ir3: enable half precision for pre-fs texture fetch

Reviewed-by: Rob Clark <robdclark@gmail.com>

freedreno/ir3: fixup when changing to mad.f16

Reviewed-by: Rob Clark <robdclark@gmail.com>

freedreno/ir3: fix printing output registers of FS.

Fixes: cea39af2fbf1 ("freedreno/ir3: Generalize ir3_shader_disasm()")
Reviewed-by: Rob Clark <robdclark@gmail.com>

freedreno/ir3: Enabling lowering 16-bit flrp

Reviewed-by: Rob Clark <robdclark@gmail.com>

freedreno: support 16b for the sampler opcode

Reviewed-by: Rob Clark <robdclark@gmail.com>

freedreno/ir3: Implement f2b16 and i2b16

Reviewed-by: Rob Clark <robdclark@gmail.com>

freedreno/ir3: Add implementation of nir_op_b16csel

Reviewed-by: Rob Clark <robdclark@gmail.com>

freedreno/ir3: Support 16-bit comparison instructions

v2. [Hyunjun Ko (zzoon@igalia.com)]
Avoid using too much open code like "instr->regs[n]->flags |= FOO"

v3. [Hyunjun Ko (zzoon@igalia.com)]
Remove redundant code for both 16b and 32b operations.

Reviewed-by: Rob Clark <robdclark@gmail.com>

freedreno/ir3: cleanup by removing repeated code

Prep-work for the corresponding patch.

Reviewed-by: Rob Clark <robdclark@gmail.com>

nir/lower_alu_to_scalar: Support lowering 8- and 16-bit reduce ops

Reviewed-by: Rob Clark <robdclark@gmail.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

nir: Add a 8-bit bool type

Adds nir_type_bool8 as well as 8-bit versions of all the bool
opcodes.

Reviewed-by: Rob Clark <robdclark@gmail.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

nir: Add a 16-bit bool type

Adds nir_type_bool16 as well as 16-bit versions of all the bool
opcodes.

Reviewed-by: Rob Clark <robdclark@gmail.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

nir/opcodes: Add a helper function to generate reduce opcodes

Adds binop_reduce_all_sizes which generates both 1-bit and 32-bit
versions of the reduce operation. This reduces the code duplication a
bit and will make it easier to later add 16-bit versions as well.

Reviewed-by: Rob Clark <robdclark@gmail.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

nir/opcodes: Add a helper function to generate the comparison binops

Adds binop_compare_all_sizes which generates both 1-bit and 32-bit
versions of the comparison operation. This reduces the code
duplication a bit and will make it easier to later add 16-bit versions
as well.

Reviewed-by: Rob Clark <robdclark@gmail.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

radv: enable VK_KHR_shader_subgroup_extended_types on GFX6-GFX7

Most of DEQP-VK.subgroups are skipped because 16-bit float aren't
supported but others pass.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

v3d: adds an extra MOV for any sig.ld*

Specifically when we are in non-uniform control flow, as we would need
to set the condition for the last instruction. If (for example) a
image atomic load stores directly their value on a NIR register,
last_inst would be a nop, and would fail when set the condition.

Fixes piglit test:
spec/glsl-es-3.10/execution/cs-ssbo-atomic-if-else-2.shader_test

Fixes: 6281f26f064ada ("v3d: Add support for shader_image_load_store.")
v2: (Changes suggested by Eric Anholt)
   * Cover all sig.ld* signals, not just ldunif and ldtmu, as all of
     them have the same restriction.
   * Update comment explaining why we add a MOV in that case
   * Tweak commit message.

v3:
   * Drop extra set of parens (Eric)
   * Add missing ld signal to is_ld_signal to fix shader-db regression.

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

v3d: Fix predication with atomic image operations

Fixes dEQP test:
dEQP-GLES31.functional.synchronization.inter_call.with_memory_barrier.image_atomic_multiple_interleaved_write_read

Fixes piglit test:
spec/glsl-es-3.10/execution/cs-image-atomic-if-else.shader_test

Fixes: 6281f26f064ada ("v3d: Add support for shader_image_load_store.")
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

panfrost: Don't print the midgard_blend_rt structs on SFBD

Signed-off-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

gitlab-ci: Fix dir name for VK-GL-CTS sources

Signed-off-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Rework buffers in SFBD

Support cases such as depth-only renders and only set stencil buffers
when needed, to match the blob's behaviour.

Signed-off-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

panfrost: Just print tiler fields as-is for Tx20

The tiler unit in these GPUs is quite different and we haven't reverse
engineered enough of it yet to validate and pretty print it.

Signed-off-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

pan/midgard: Introduce quirks checks

Rather than open-coding checks on gpu_id in the compiler, let's track
quirks applying to whatever we're compiling for, to allow us to manage
the complexity of many heterogenous GPUs in the compiler.

It was discovered that a workaround used on T720 is also required on
T820 (and presumably T830), so let's fix this. This will also decrease
friction as we continue improving T720 support.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>

gitlab-ci: update for arb_shading_language_include

gitlab-ci: bump piglit checkout commit

mesa: enable ARB_shading_language_include

Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/999
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: implement glCompileShaderIncludeARB()

Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: add shader include lookup support for relative paths

Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: add support cursor support for relative path shader includes

This will allow us to continue searching the current path for
relative shader includes.

From the ARB_shading_language_include spec:

   "If it is quoted with double quotes in a previously included
   string, then the first search point will be the tree location
   where the previously included string had been found."

Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

glsl: delay compilation skip if shader contains an include

If the shader contains an include when need to first run the
preprocessor before deciding if we can skip compilation based
on the shader cache.

Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

glsl: add can_skip_compile() helper

We will reuse this in the following commit.

Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

glsl: error if #include used while extension is disabled

In other words make sure the shader does this:

Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

glsl: add preprocessor #include support

Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

glsl: pass gl_context to glcpp_parser_create()

This is a small tidy up and will be useful in the following commit.

Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

glsl: add ARB_shading_language_include support to #line

From the ARB_shading_language_include spec:

   "#line must have, after macro substitution, one of the following
    forms:

       #line <line>
       #line <line> <source-string-number>
       #line <line> "<path>"

    where <line> and <source-string-number> are constant integer
    expressions and <path> is a valid string for a path supplied in the
    #include directive. After processing this directive (including its
    new-line), the implementation will behave as if it is compiling at
    line number <line> and source string number <source-string-number>
    or <path> path. Subsequent source strings will be numbered
    sequentially, until another #line directive overrides that
    numbering."

Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: implement glDeleteNamedStringARB()

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: split _mesa_lookup_shader_include() in two

The new local function lookup_shader_include() will be used by
glDeleteNamedStringARB() in the following patch.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: implement glGetNamedStringivARB()

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: implement glIsNamedStringARB()

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: make error checking optional in _mesa_lookup_shader_include()

This will be usefull when implementing glIsNamedStringARB() which
doesn't do error checking, it just returns false for invalid
lookups instead.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: implement glGetNamedStringARB()

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: add glNamedStringARB() support

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: add copy_string() helper

This will be used by the various ARB_shading_language_include
functions in the following patches.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: add _mesa_lookup_shader_include() helper

This will be used both by the glsl compiler and the GL API.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: add helper to validate tokenise shader include path

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: add ARB_shading_language_include infrastructure to gl_shared_state

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

glsl: add infrastructure for ARB_shading_language_include

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

mesa: add ARB_shading_language_include stubs

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Witold Baryluk <witold.baryluk@gmail.com>

radv: Do not change scratch settings while shaders are active.

When the scratch ringbuffer settings are changed, the shader unit has
to be idle or we will have shaders using old and new settings.

That combination is not supported on the HW (likely the offset is
ringbuffer idx * WAVESIZE * 1024).

CC: <mesa-stable@lists.freedesktop.org>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

turnip: Drop the copy of the formats table.

Now that we can (mostly) generate a pipe format for a VkFormat, use that
to answer queries about formats.  This will let us refactor the freedreno
format table surface layout code to be shared between gallium and vulkan.

This causes us to expose fewer formats for now (on a 1/100 CTS run I'm
doing, skips go from 3671 to 3835 out of 5145 tests).  Fails stay about
the same (478 -> 434, but the run is pretty flaky and we're doing fewer
tests now).

v2: Rebase on master, throw a finishme on missing vk-to-pipe formats that
    tu used to support.

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com> (v1)
Reviewed-by: Jonathan Marek <jonathan@marek.ca>

util: Add a mapping from VkFormat to PIPE_FORMAT.

I'm planning on using this from radv and tu for queries about formats.

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Jonathan Marek <jonathan@marek.ca>

winsys/amdgpu: detect noop dependencies on the same ring correctly

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

ac: fill num_rings for remaining IPs

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

ac: add radeon_info::num_rings and move ring_type to amd_family.h

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

nir: don't use GLenum16 in nir.h

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

nir: move data.descriptor_set above data.index for better packing

4 bytes down

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

glsl_to_nir: rename image_access to mem_access

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

nir/print: only print image.format for image variables

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

nir: move data.image.access to data.access

The size of the data structure doesn't change.

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

st/mesa: call nir_serialize only once per shader

It was called twice.

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

st/mesa: keep serialized NIR instead of nir_shader in st_program

This decreases memory usage, because serialized NIR is more compact.

If shader_has_one_variant is true and the shader is uncached, the first
variant is created from nir_shader, otherwise the first variant and
all other variants are created from serialized NIR.

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

st/mesa: call nir_sweep in st_finalize_nir

This is invoked sooner before (pre-)compiling the first variant and is
also applied to fixed-func and ARB programs.

st/mesa: subclass st_vertex_program for VP-specific members

Inheritance:
    gl_program -> st_program -> st_vertex_program

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

st/mesa: more cleanups after unification of st_vertex/common_program

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

st/mesa: rename occurences of stcp to stp to correspond to st_program

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

st/mesa: cleanups after unification of st_vertex/common program

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

st/mesa: rename st_common_program to st_program

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

st/mesa: trivially merge st_vertex_program into st_common_program

a later commit will add back st_vertex_program as a subclass of
st_common_program

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

st/mesa: consolidate and simplify code flagging program::affected_states

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

st/mesa: initialize affected_states and uniform storage earlier in deserialize

This matches the uncached codepath.

affected_states was used before initialization, which was technically
a bug, but probably not reproducible due to _NEW_PROGRAM rebinding
everything.

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

st/mesa: start deduplicating some program code

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

st/mesa: decrease the size of st_fp_variant_key from 48 to 40 bytes

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

st/mesa: rename delete_basic_variant -> delete_common_variant

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

anv: add missing "fall-through" annotation

CoverityID: 1455884
Fixes: c1c346f1667375e9330a ("anv: implement VK_KHR_separate_depth_stencil_layouts")
Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>

egl: use EGL_CAST() macro in eglmesaext.h

Allows eglmesaext.h to be used in C++ code.

This aligns this file with the rest of EGL.

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-By: Tapani Pälli <tapani.palli@intel.com>

vulkan: delete typo'd header

Two files exist in that directory:
- vulkan_xlib_randr.h
- vulkan_xlib_xrandr.h

Both were imported in 205c271562db8cb2effc ("vulkan: Update the XML and
headers to 1.1.70") with identical contents (ie. the
VK_EXT_acquire_xlib_display extension), but the former was never
included anywhere and can't be found upstream [1], while the latter is
included in vulkan.h and found upstream.

[1] https://github.com/KhronosGroup/Vulkan-Headers/tree/master/include/vulkan

Fixes: 205c271562db8cb2effc ("vulkan: Update the XML and headers to 1.1.70")
Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>

CL: sync C++ headers with Khronos

https://github.com/KhronosGroup/OpenCL-CLHPP at commit
cf9fc1035e8298c7ce65ee33066a660fd9892ebb

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

CL: sync C headers with Khronos

https://github.com/KhronosGroup/OpenCL-Headers at commit
0d5f18c6e7196863bc1557a693f1509adfcee056

Signed-off-by: Eric Engestrom <eric.engestrom@intel.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>

intel: Add workaround for stencil state.

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>

turnip: fix sRGB GMEM clear

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>

turnip: implement CmdClearColorImage/CmdClearDepthStencilImage

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>

radv/aco: enable VK_KHR_shader_subgroup_extended_types

We could enable it on GFX10 if LLVM wasn't used as a fallback for
unsupported stages. Note that the CTS only tests it if
VK_KHR_shader_float16_int8 is enabled, even though it's not a
requirement.

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

aco: implement 64-bit integer reductions

The multiplication reduction is larger than it could be, but it should be
easier to implement this way.

No failures with dEQP-VK.subgroups.*int64* except those caused by LLVM
being used for other stages.

v2: don't call setFixed() for v_add carry-out, since setHint sets physReg
v3: add and use emit_vadd32() helper
v4: use num_opcodes instead of last_opcode

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev> (v3)

aco: refactor reduction lowering helpers

Should make 64-bit integer reductions easier to implement.

v4: use num_opcodes instead of last_opcode

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev> (v3)

radv: advertise VK_KHR_shader_subgroup_extended_types on GFX8-GFX9

This extension allows to use subgroup operations with 8 and 16-bits

Untested on GFX6-GFX7, and most of subgroup operations are broken
on GFX10, so don't enable it for now. Not enabled on ACO because
it's still doesn't support 8-bits/16-bits.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

ac: add 16-bit float support to ac_build_alu_op()

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>