intel/fs: Add surface OWORD BLOCK opcodes

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7448>

intel/fs: Implement nir_intrinsic_{load,store}_global_block_intel

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7448>

intel/fs: Add A64 OWORD BLOCK opcodes

Based on a patch for OWORD BLOCK READ from Jason Ekstrand.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7448>

spirv: Implement SpvCapabilitySubgroupBufferBlockIOINTEL

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7448>

nir: Add nir_intrinsic_{load,store}_deref_block_intel

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7448>

spirv: Implement SpvCapabilitySubgroupShuffleINTEL from SPV_INTEL_subgroups

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7448>

iris: Flush dmabufs during context flushes

Currently, every modifier that uses CCS also lacks support for
fast-clears. On gen9+, dmabufs may gain fast-cleared blocks through
clear calls. On gen12, fast-clearing can occur during any rendering
operation. Mark when dmabufs gain fast-cleared blocks and flush them
during a context flush operation.

Cc: mesa-stable
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/3425
Tested-by: Simon Ser <contact@emersion.fr>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7384>

iris: Fix fast-clears of swizzled alpha formats

Resources with alpha formats that are mapped to R are fast-cleared with
the wrong clear color.

When such resources with are cleared via iris_clear_texture,
isl_color_value_unpack places channel data in the R channel.
convert_fast_clear_color then overwrites the channel with 0.

To avoid zeroing the clear color, move convert_fast_clear_color to the
other callers of clear_color: iris_clear and iris_clear_render_target.

Enables iris to pass the "A" case of the fcc-clear-tex piglit test.

v2. Rename convert_fast_clear_color to convert_clear_color. (Ken)

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/3670
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7345>

iris: Fix SINT assert in convert_fast_clear_color

Don't assert that the size of every channel is greater than zero. This
assert doesn't work for integer formats with less than 4 channels.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7345>

iris: Fix fast-clears of swizzled LA formats

Resources with luminance alpha formats that are mapped to RG are
fast-cleared with the wrong clear color.

When such resources with are cleared via iris_clear_texture,
isl_color_value_unpack places channel data in the R and G channels.
convert_fast_clear_color then overwrites the G channel with R.

Delete the clear color override that's specific to luminance alpha
formats.

Enables iris to pass the "LA" case of the fcc-clear-tex piglit test.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7345>

iris: fix source/destination layers for 3D blits

See commit ea326912575fad09af59486ad62d126c4ea0ede7 for the
corresponding fix in anv.

Fixes Piglit's fbo-generatemipmap-3d.

Cc: mesa-stable
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7321>

st/nir: Drop setting interp mode on system values in builtins.

It's initialized to INTERP_MODE_NONE on creation, which makes more sense
for sysvals than FLAT, and is also the interp mode that GLSL IR sets up
for sysvals.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7320>

st/nir: Fix the st->pbo.use_gs case.

This case hadn't been ported to NIR before, and I missed that when
removing the TGSI path and replacing it with NIR -> NTT for TGSI drivers.
This caused breakage in nv50 on piglit's pbo-teximage.

In the process, the !use_gs gets its layer output fixed to be an int
instead of a vec4, which I suspect would fix validation in that path.

Fixes: 57effa342b75 ("st/mesa: Drop the TGSI paths for PBOs and use nir-to-tgsi if needed.")
Closes: #3680
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7320>

pan/bi: Correctly calculate render target index

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

pan/bi: Lower depth/stencil stores

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

pan/bi: Emit +ZS_EMIT as needed

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

pan/bi: Stub handling for nir_intrinsic_store_combined_output_pan

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

pan/bi: Factor out bi_emit_blend

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

pan/bi: Factor out bi_emit_atest

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

pan/bi: Infer z/stencil flags from sources passed

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

pan/bi: Add +ZS_EMIT instruction to IR

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

panfrost: Deduplicate shader properties

Between Midgard and Bifrost.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

panfrost: Pass through src_type

Needed since Bifrost blends are typed well.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

pan/mdg: Move writeout lowering to common panfrost

These will be used in the Bifrost compiler, albeit for a slightly
different purpose.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

pan/mdg: Deduplicate nir_find_variable_with_driver_location

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

nir: Add SRC_TYPE to store_combined_output_pan

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7446>

v3dv: add a v3dv_bo_init helper

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7440>

aco/ra: Fix counting of subdword variables in get_reg_create_vector

The loop variable "k" shadowed another variable in the outer scope, so
this loop had no actual effect.

Fixes: 52cc1f8237d ("aco: improve p_create_vector RA for sub-dword operands")
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7427>

aco: implement 8/16-bit instructions which can be trivially widened

When nir_lower_bit_size becomes more capable, we might want to revert some
of this.

fossil-db (parallel-rdp, Navi):
Totals from 217 (31.77% of 683) affected shaders:
SGPRs: 11320 -> 10200 (-9.89%)
VGPRs: 7156 -> 7364 (+2.91%)
CodeSize: 1453948 -> 1430136 (-1.64%); split: -1.66%, +0.02%
Instrs: 258530 -> 254840 (-1.43%); split: -1.44%, +0.01%
Cycles: 37334360 -> 37247936 (-0.23%); split: -0.26%, +0.03%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4791>

aco: implement some 16-bit arithmetic instead of lowering

fossil-db (parallel-rdp, Navi):
Totals from 210 (30.75% of 683) affected shaders:
SGPRs: 9704 -> 10248 (+5.61%)
VGPRs: 5884 -> 5368 (-8.77%)
CodeSize: 1155564 -> 1098752 (-4.92%)
Instrs: 199927 -> 189940 (-5.00%)
Cycles: 20438392 -> 19860124 (-2.83%)

v2: use divergence analysis to determine which instructions to lower.

Co-Authored-by: Daniel Schürmann <daniel@schuermann.dev>
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4791>

radv: rework nir_lower_bit_size callback and run DA on GFX8+

No fossil-db changes.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4791>

radv: do nir_lower_bit_size after algebraic optimizations

There are too many algebraic optimizations to be certain that one of them
couldn't create instructions which need lowering. It also creates better
code for some reason.

fossil-db (parallel-rdp, Navi):
Totals from 217 (31.77% of 683) affected shaders:
VGPRs: 7716 -> 7672 (-0.57%)
CodeSize: 1516152 -> 1510688 (-0.36%); split: -0.38%, +0.02%
MaxWaves: 3964 -> 3982 (+0.45%)
Instrs: 269445 -> 268508 (-0.35%); split: -0.36%, +0.02%
Cycles: 37963416 -> 37912592 (-0.13%); split: -0.15%, +0.01%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4791>

radv: move a few passes to after load/store vectorization

load/store vectorization can create 8/16-bit alu to do packing/unpacking,
which would make shader_info::bit_sizes_used out of date.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4791>

nir/lower_bit_size: optimize upcast of b2i8/b2i16

This also seems to be done by nir_opt_algebraic, but RADV will be moving
nir_lower_bit_size() to after that (so it doesn't create unsupported
8/16-bit instructions) and it doesn't seem worth creating a new pass just
for this simple optimization.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4791>

nir: add shader_info::bit_sizes_used

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4791>

va: support VA_RT_FORMAT_PROTECTED

Reviewed-by: Boyuan Zhang <boyuan.zhang@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7006>

va/picture: make sure destination buffer is protected if needed

Reviewed-by: Boyuan Zhang <boyuan.zhang@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7006>

frontends/va: Added protected playback support for VP9

Add VP9 header handling in slice data buffer.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Boyuan Zhang <boyuan.zhang@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7006>

radeon/vcn: program drm message buffer

Add a function to handle drm message buffer using input decryption parameters.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7006>

radeon/vcn: delay dec->ctx and dec->dpb allocation

This will allow to allocate them as encrypted if needed.

Reviewed-by: Boyuan Zhang <boyuan.zhang@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7006>

radeon: add si_vid_create_tmz_buffer helper

Same code as si_vid_create_buffer except that the buffer is using TMZ.

Reviewed-by: Boyuan Zhang <boyuan.zhang@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7006>

radeon/vcn: add defines for drm message buffer

Add defines and structure for drm message buffer.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7006>

radeon: add decryption params definition header

Add a header file for decryption parameters.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7006>

frontends/va: handle protected slice data buffer

Add a function to handle VaProtectedSliceDataBuffer, which is used for
sending decryption parameters. Also, for protected playback, there is
no need to check start code since data is encrypted.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7006>

vl: add flag and definition for protected playback

Add a flag to indicate if playback is protected/encrypted.
Add a pointer to decryption key for later decryption use.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7006>

nir/find_array_copies: Don't assume all children exist

Fixes: 9f3c595dfc4cd "nir/find_array_copies: Handle cast derefs"
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7437>

radesonsi: Remove unnecessary shader->selector NULL check.

shader->selector->info.stage == MESA_SHADER_COMPUTE at this case statement.

Fix defect reported by Coverity Scan.

Dereference before null check (REVERSE_INULL)
check_after_deref: Null-checking shader->selector suggests that it may
be null, but it has already been dereferenced on all paths leading to
the check.

Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7266>

lavapipe: request correct sample mask behaviour

Fixes:
dEQP-VK.rasterization.frag_side_effect*

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7435>

llvmpipe: respect the sample mask in non-multisample flag

This partly revert
50987644 llvmpipe: don't use sample mask with 0 samples

since Vulkan wants this behaviour.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7435>

gallium: add a non-multisample sample mask out behaviour flag.

Vulkan/DX want to use output sample mask even when not multisampling
GL wants it ignored.

Add a rasterizer flag to lavapipe can get correct behaviour.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7435>

nir/opt_intrinsic: Optimize bcsel(b, shuffle(x, i), shuffle(x, j))

The shuffles provided by the SPV_INTEL_subgroups extension generate

    bcsel(b, shuffle(x, i), shuffle(y, j))

In the case where x and y are the same, we can turn this into a shuffle
with the bcsel on the index which lets us drop a whole shuffle.

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7366>

nir/opt_intrinsics: Refactor a bit

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7366>

nir/constant_folding: Fold subgroup shuffle intrinsics

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7366>

nir: Move constant folding of vote to opt_constant_folding

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7366>

nir/constant_folding: Use the standard variable naming convention

Typically, if we have one alu instruction, we call it "alu" and if we
have one intrinsic we call it "intrin".

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7366>

nir/constant_folding: Use a switch in try_fold_intrinsic

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7366>

nir/opt_intrinsics: Report progress for the gl_SampleMask optimization

Fixes: d3ce8a7f6b93 "nir: optimize gl_SampleMaskIn to gl_HelperInvocation..."
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7366>

nir: use nir_alu_src_is_trivial_ssa() in nir_ssa_for_alu_src()

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7426>

nir: skip bcsel with non-trivial swizzle in opt_simplify_bcsel_of_phi()

Fixes validation error in a Dota 2 shader.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Fixes: b031c643491 ("nir: Convert a bcsel with only phi node sources to a phi node")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7426>

nir: add nir_alu_src_is_trivial_ssa()

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7426>

nir/lower_io: Add a new 62bit_generic address format

Unlike most address formats, this address format is capable of handling
all of the fancy generic pointers stuff like is_global and friends.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/lower_io: Support generic pointer access

If the pointer is generic and we haven't yet figured out what kind of
pointer it is yet, we emit an if-ladder based on a mode check.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/lower_io: Add support for lowering deref_mode_is

The guts are still missing so it will blow up if it sees any
deref_mode_is intrinsic that it can't constant-fold from the mode.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/lower_io: Add support for 32/64bit_global for shared

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/lower_io: Add a mode parameter to addr_format_is_*

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/lower_io: Add a mode parameter to build_addr_iadd

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/opt_deref: Add an optimization for deref_mode_is

If opt_restrict_deref_modes makes progress, we may be able to figure out
the mode well enough to turn a deref_mode_is intrinsic into a constant.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/opt_deref: Add a deref mode specialization optimization

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

spirv: Add generic pointer support

Most of this is fairly straightforward; we just set all the modes on any
derefs which are generic.  The one tricky bit is OpGenericCastToPtrExplicit.
Instead of adding NIR intrinsics to do the cast, we add NIR intrinsics
to do a storage class check and then bcsel based on that.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir: Add support for generic pointers

The way they're handled is that deref->modes is treated as a bitfield of
possible modes.  Variables are required to have a specific mode and
derefs with deref_type_var are as well.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir: Make nir_deref_instr::mode a bitfield

We rename it to "modes" to make it clear that it may contain more than
one mode and adjust all the uses of nir_deref_instr::modes to attempt to
handle multiple modes.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/split_*_vars: Prepare for generic pointers

All three passes check the variables for complex uses and don't split
them if they have any complex uses.  Most of these checks are just early
returns to avoid chasing the deref to the variable and a hash table
lookup if we can quickly determine it has the wrong mode.  In a couple
of cases, we need to re-arrange or add other checks to ensure that it's
safe for generic pointers.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/find_array_copies: Prepare for generic pointers

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir: Use nir_deref_mode_may_be in deref optimizations

All the checks being replaced are fore potential aliasing so we want to
flush stores whenever the mode might be something that aliases.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/vec3_to_vec4: Use nir_deref_must_be

We use the same nir_deref_mode_is_in_set helper that we use in
nir_lower_vars_to_explicit_types for the same reason.  If there are any
generic pointers in play, we have to lower all generic pointer modes at
the same time or else we risk types getting out-of-sync.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/vars_to_ssa: Use nir_deref_must_be

We can only lower a deref to SSA in this pass if it's guaranteed to be
nir_var_function_temp.  We already flag any variables with complex uses
(i.e. casts) as not being lowerable and refuse to lower any derefs to
them so we don't have to worry about false negatives.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir: Only force loop unrolling if we know it's a in/out/temp

If we don't know the actual mode then we can't get to the variable so
it's going to be a scratch or other indirect load anyway and we aren't
saving ourselves anything by unrolling the loop.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/phis_to_scalar,gcm: Use nir_deref_mode_may_be

In both cases, we're trying to determine if a load is scalarizable.  We
don't want to scalarize if it's a function_temp or shader_temp because
it might turn into something we can't scalarize.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/lower_io: Use nir_deref_mode_* helpers

For non-explicit nir_lower_io, we use nir_deref_mode_is because there's
no way it works for generic pointers.  For nir_lower_vars_to_explicit_types,
and nir_lower_explicit_io, we use nir_deref_mode_is_in_set to ensure we
never get type confusion.  For generic pointers, this means that they
must be called with the full set of generic pointer modes.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/lower_array_deref_of_vec: Use nir_deref_mode_must_be

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir: Add and use some deref mode helpers

NIR derefs currently have exactly one variable mode.  This is about to
change so we can handle OpenCL generic pointers.  In order to transition
safely, we need to audit every deref->mode check.  This commit adds a
set of helpers that provide more nuanced mode checks and converts most
of NIR to use them.

For simple cases, we add nir_deref_mode_is and nir_deref_mode_is_one_of
helpers.  These can be used in passes which don't have to bother with
generic pointers and just want to know what mode a thing is.  If the
pass ever encounters generic pointers in a way that this check would be
unsafe, it will assert-fail to alert developers that they need to think
harder about things and fix the pass.

For more complex passes which require a more nuanced understanding of
modes, we add nir_deref_mode_may_be and nir_deref_mode_must_be helpers
which accurately describe the compiler's best knowledge about the given
deref.  Unfortunately, we may not be able to exactly identify the mode
in a generic pointers scenario so we have to be very careful when we use
these.  Conversion of these passes is left to later commits.

For the case of mass lowering of a particular mode (nir_lower_explicit_io
is one good example), we add nir_deref_mode_is_in_set.  This is also
pretty assert-happy like nir_deref_mode_is but is for a set containment
comparison on deref modes where you expect the deref to either be all-in
or all-out.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/opt_find_array_copies: Allow copies from mem_constant

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir: Disallow writes to system values and mem_constant

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir: Use var->data.mode instead of deref->mode in a few cases

We already have the variable so we know the mode exactly.  Just use that
instead of the deref mode.  If these paths ever have to handle variable
pointers (not likely since they're OpenGL-specific), we can fix them to
handle crazy deref modes then.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir: Handle incomplete derefs in split_struct_vars

In split_var_list_structs where we initalize the splitting, we already
use get_complex_used_vars to avoid splitting any variables that have a
complex use.  However, we weren't actually handling the complex uses
properly in the case where we can't actually find the variable.

Fixes: f1cb3348f1 "nir/split_vars: Properly bail in the presence of ..."
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/phis_to_scalar: Use a deny-list for load_deref modes

I can't think of any reason why shared and output aren't in this list.
The real thing we're trying to do is avoid premature scalarization
because of a shader or function temporary variable because we might
lower it to something we don't want scalarized later.  Also fix the
version we copy+pasted into GCM.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

nir/builder: Add a nir_ieq_imm helper

This shows up surprisingly often.

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6332>

lavapipe: don't advertise linear filtering on integer textures.

The backend doesn't support this.

Reviewed-by: Adam Jackson <ajax@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7416>

lavapipe: use clear_buffer callback

llvmpipe needs the clear buffer callback for CL, make lavapipe
use it as well.

Reviewed-by: Adam Jackson <ajax@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7416>

llvmpipe: add clear_buffer callback. (v2)

This fixes CL CTS thread dimensions test

v2: optimise for 1 and 4. (ajax)

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Reviewed-by: Adam Jackson <ajax@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7416>

lavapipe: stop crashes with 3D z blits

This code just didn't handle 3D Z blits properly, rewrite
to handle change in direction here.

Reviewed-by: Adam Jackson <ajax@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7416>

lavapipe: fix 3d compressed texture copies.

The img stride was being calculated incorrectly.

Fixes crashes in:
dEQP-VK.api.copy_and_blit.core.image_to_image.all_formats.color.3d.bc1_rgb_srgb_block.bc1_rgb_srgb_block.general_general

Reviewed-by: Adam Jackson <ajax@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7416>

gallivm/nir: fix vulkan vertex inputs

the tgsi file max is used to determine the number of input slots,
the old code was pretty bogus for the lavapipe cases (it worked
by accident).

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7416>

gallivm/nir: handle dvec3/4 inputs properly.

This code works but isn't entirely correct, for a dvec3 it would
fetch loc 0,1 2,3 4,5 but really each loc only has 4 entries,
instead catch this and read loc 0,1 2,3 loc+1 0,1

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7416>

lavapipe: fix dEQP-VK.info.device_properties

Fix bounds and widelines aren't supported for now.

Reviewed-by: Adam Jackson <ajax@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7416>

lavapipe: constify state pointers into command buffers.

for render pass information pointers into the command buffer are stored,
but command buffers are immutable content so make sure to use const ptrs
to avoid problems like was seen with clears.

Reviewed-by: Adam Jackson <ajax@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7416>

lavapipe: don't write to pending clear aspects in cmd buffer

When the cmd buffer is recorded, we record the attachments state
into it, however we use the pending clear aspects to keep track
of clears, but it should be kept in the state no overwrritten
in the cmd buffer.

Allocate some memory to store this hanging off the state.

This fixes gears and radialblur demos.

Reviewed-by: Adam Jackson <ajax@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7416>

gallivm: fix f16 quantize.

Add the correct flush to 0 behaviour.

Fixes:
dEQP-VK.spirv_assembly.instruction.compute.opquantize.flush_to_zero

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7416>

r600: amend space check for chips older than EVERGREEN

evergreen_emit_atomic_buffer_setup_count is only called if chip >= EVERGREEN
otherwise atomic_used_mask is left uninitialized when unconditionally used by
r600_need_cs_space so it might want more space than needed

fix this by always initializing atomic_used_mask

Fixes: 32529e60849 ("r600/eg: rework atomic counter emission with flushes")
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7417>

zink: break up dynamic access lowering

this is gross and slow, but to handle dominance properly, now we just load all
the ubos and bcsel away

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7199>