gallium: decompose quad strips into quads if supported

This changes gallium to decompose quad strips into quads instead of triangles
when the driver advertises support for them.

This should result in a more correct result when those are drawn
with the line raster primitve (avoids showing the diagonal line).

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21987>

mesa: correctly allocate space for converted primtives

Previosuly it was assumed that primitives where always converted to
triangles if the driver did not support all primitives, however that's
not true for a driver that supports quads but not quad strips.

Fixes piglit spec@!opengl 1.1@dlist-fdo3129-01 on Panfrost

Fixes: dcbf2423d28 ("vbo/dlist: add vertices to incomplete primitives")
Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21987>

radv: pass a radv_shader to radv_emit_dispatch_packets()

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22102>

radv: move cs_regalloc_hang_bug to radv_shader_info

This is more like a shader property.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22102>

amd: remove duplicate from .clang-format

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22077>

aco: remove duplicates from .clang-format

The latest clang-format doesn't support this anymore.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22077>

anv: add utrace support for queue debug utils

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22006>

anv: rename anv_utrace_flush_copy in anv_utrace_submit

We want to use this for submission of traces outside command buffers,
so it won't just execute copies of timestamp buffers.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22006>

anv: fixup locking for utrace submission increments

This is supposed to happen under the device lock.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: e760c5b37b ("anv: add perfetto source")
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22006>

anv: fix incorrect utrace bo release

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: c67c9688c3 ("anv/utrace: use a bo pool for utrace buffers")
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22006>

intel/ds: rename frame timeline row to queue

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22006>

vulkan/runtime: also copy strings on queue debug utils

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22006>

util/u_trace: move needs_cs_param option to tracepoints

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22006>

d3d12: AV1 Dec - Set anchor_frame_idx only when large_scale_tile equals 1

Signed-off-by: Teng, Jin Chung <jin.chung.teng@intel.com>
Reviewed-by: Sil Vilerino <sivileri@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22048>

frontend/va: Add large_scale_tile from VADecPictureParameterBufferAV1

Signed-off-by: Teng, Jin Chung <jin.chung.teng@intel.com>
Reviewed-by: Sil Vilerino <sivileri@microsoft.com>
Reviewed-by: Boyuan Zhang <boyuan.zhang@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22048>

zink: flag rp layout change if zsbuf usedness changes on dsa/fs state bind

this should (correctly) trigger the expected zsbuf elimination and
avoid hitting asserts

fixes #8679

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22086>

panfrost: Remove Midgard RSD fields from Bifrost

These were removed and replaced by new Bifrost RSD fields, don't print the wrong
values. Harmless but noises up the decoding.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panvk: Lower blending late

Since 50b82ca8181 ("nir/lower_blend,agx,panfrost: Use lowered I/O"),
nir_lower_blend needs to be called after lowering I/O rather than before.
Furthermore, after lowering blend, we need (in general) to lower the resulting
load_output intrinsics. Now that we have a proper preprocess_nir hook, there is
a natural place in panvk_vX_shader to do this.

Fixes dEQP-VK.pipeline.blend.*

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/bi: Call pan_nir_lower_zs_store late

This will give the driver [notably, PanVK] a chance to lower dual source
blending without having the dual stores turned into store_combined_output_pan.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/bi: Export bifrost_nir_lower_load_output

If new load_output are created after preprocessing NIR (namely, from blend
lowering in panvk), this lowering needs to be called to lower load_output to the
vendor intrinsic with conversion descriptor.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Move panfrost_sysvals to GL driver

This shouldn't be used by anything else at this point.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panvk: Remove unused function

Erroneously referencing sysvals.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Move sysvals to GL driver struct

Only the GL driver produces/consumes these, they shouldn't be in the common
shader_info.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Lower sysvals in GL

Drop the backend compiler sysval handling in favour of the pass in the GL
driver, bringing us into compliance with Ekstrand's rule.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Add NIR-based sysval lowering pass

Add a NIR pass to lower all the sysvals seen in the GL driver to load_ubo
intrinsics. These load_ubo intrinsics will be pushed to uniforms by the backend
compiler as usual. This will let us remove all sysval handling from the backend
compilers.

This is a direct NIR port of the existing pan_sysvals.c infrastructure and the
consumers in the Midgard/Bifrost compilers. It aims to be bug-for-bug compatible
to ease bisection.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panvk: Inline blend constants as syvals

Blend constants are sysvals, it's just that they can sometimes be inlined
depending on the pipeline state. The old "inline blend constant" pass is a
special case of the new "lower all sysvals" pass in panvk.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panvk: Don't use vec4 for vertex_instance_offsets

Not needed with the new lowering.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panvk: Lower sysvals in NIR

Per Ekstrand's Rule. This avoids the "fixed sysval" hack that Faith introduced
to get this behaviour with the GL sysval handling.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Remove stale TODO

While the text here is still nominally accurate, we should be seeing so few
shader variants at this point that the locking contention isn't a big deal.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Preprocess shaders at CSO create time

Now the only passes that depend on the shader key can run late, so we can
preprocess ahead-of-time once and throw away the original shader. This reduces
the cost of shader variants, as well as deduplicates some lowering for
transform feedback shaders.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Effectively lower gl_FragColor late

nir_lower_fragcolor takes the number of colour buffers as input, but it's an
early pass, so we don't want to use the key for it. Instead, we can overestimate
and then optimize out late with an easy pass.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Lower texcoords late

Use the _late version of this lowering instead of the early one.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Lower clip_fs late

This pass works both early and late, so this is an easy one to sink down.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/lower_framebuffer: Lower MSAA blend shaders

Do it explicitly in NIR rather than implicitly in the Midgard compiler. This
avoids a nasty sideband input for the render target formats and sample count,
for blend shaders on midgard only.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Preprocess shaders in the driver

This is a flag-day change to how we compile. We split preprocessing NIR into a
separate step from compiling, giving the driver a chance to apply its own
lowerings on the preprocessed NIR before the final optimization loop. During
that time, the different producers of NIR (panfrost, panvk, blend shaders, blit
shaders...) will be able to (differently) lower system values.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/blit: Lower load_sampler_lod_parameters_pan

This will be needed to decouple the lowering in the Midgard compiler from the
specific sampler descriptors used in the blit code.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/lower_framebuffer: Use nir_shader_instructions_pass

Removes a lot of indentation, and improves metadata handling.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/lower_framebuffer: Only call for FS

It doesn't make sense for shader stages other than fragment (and blend which is
fragment-like), assert this.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/mdg: Split out early preprocessing from late

To prepare for the new compile flow, where this will be called by the driver
instead of internally in the compiler.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Italo Nicola <italonicola@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/bi: Split out early preprocessing from late

To prepare for the new compile flow, where this will be called by the driver
instead of internally in the compiler.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/mdg: Only lower once

Nothing in the optimization loop should remat the lowered instructions, so
there's no need to do it inside the loop.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Italo Nicola <italonicola@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/bi: Only lower once

Nothing in the optimization loop should remat the lowered instructions, so
there's no need to do it inside the loop.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Remove unused inputs.nr_cbufs

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Remove inputs->blend.rt

This sideband input is now unused, as the information is available locally
within the NIR as it should be.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/mdg: Use I/O semantics for MRT blend stores

This avoids the silly reliance on the sideband.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Italo Nicola <italonicola@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/bi: Remove bi_load_sysval

It is unused and should stay unused, as any use is a violation of Ekstrand's
rule.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/bi: Lower gl_VertexID in NIR

This gets rid of the hidden gl_BaseVertex system value which violates Ekstrand's
rule.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/bi: Allow specializing bifrost_nir_options by arch

We need different settings for Bifrost and Valhall. Keeping everything static
simplifies lifetimes.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/bi: Lower load_output to make sysval explicit

See previous commits for justification. Later, we'll split up NIR processing in
a few steps to give the caller a chance to lower the sysval, at which point the
goofy inputs here will go away.

v2: Only lower in fragment shaders. Likely harmless to run elsewhere but still
wrong because the location enum is defined per-stage.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com> [v1]
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/bi: Lower sample mask writes in NIR

This uses the new NIR sysvals to avoid materializing magic sysvals in the
driver, getting us closer to the Ekstrand Rule.

v2: Only lower for fragment shaders. Lowering in vertex shaders should be a
no-op, except that FRAG_RESULT_SAMPLE_MASK shadows a VARYING_SLOT for fog
coords, causing v1 of this patch to regress fog. Caught by the G52 piglit job in
CI. Thank you, Marge.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com> [v1]
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

pan/bi: Don't duplicate texture op cases

These two switches are redundant.

Furthermore, bi_tex_op could previously assume its input was a supported texop,
so it returned undefined values for unsupported texops. Now, without the guard
in front of it, bi_tex_op should check for supported texops, so we need to drop
the unsupported texops from the switch.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

panfrost: Use 0/~0 boolean for MSAA sysval

For consistency with NIR.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

nir: Add Mali load_output taking converison

Mali's LD_TILE instruction (mapping to NIR's load_output) requires a "conversion
descriptor" specifying how to convert from the register foramt to the tilebuffer
format. To implement framebuffer fetch on OpenGL without shader variants, we
generate these descriptors in the driver and pass them in a uniform. However, to
comply with the Ekstrand Rule, we can't have magically materialized system
values -- they should come only from the NIR where the driver can lower as it
pleases (e.g. PanVK can lower to a constant because it knows the framebuffer
format at pipeline create time). Add intrinsics to model this.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

nir: Add Panfrost intrinsics to lower sample mask

We want to lower this in NIR instead of the backend IR to give the driver a
chance to lower the "is multisampled?" system value, which makes more sense to
do in NIR. This gets rid of one of the magic compiler materialized sysvals.

Plus, this will let us constant fold away the lowering in Vulkan when we know
that the pipeline is single-sampled / multi-sampled.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20906>

intel/dev: use GFX_VERx10 to detect genX compilation

Depending on the ordering of includes, GFX_VER may not defined for
intel_device_info.h.  The failure mode of this case is silent:
BITSET_TEST will be called when it could be compiled out.

GFX_VERx10 should be used in place of GFX_VER.  GFX_VERx10 is defined
by a compiler flag, and is always present for genX compilation units.

Fixes: 3c9a8f7a6d2 ("intel/dev: generate helpers to identify platform workarounds")
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21908>

intel/dev: fix macro naming convention in gen_wa_helpers.py

intel_device_info.h tests macros in the form `INTEL_WA_{id}_GFX_VER`.
gen_wa_helpers.py produced macros in the form `INTEL_GFX_VER_WA_{id}`

Change the generated code to follow intel_device_info.h

Fixes: 3c9a8f7a6d2 ("intel/dev: generate helpers to identify platform workarounds")
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21908>

intel/dev: fix macro string concatenation for INTEL_WA_{id}_GFX_VER

`INTEL_WA_##id_GFX_VER` evaluates to `INTEL_WA_id_GFX_VER`
instead of numbered identifiers like `INTEL_WA_220579888_GFX_VER`.

Fixes: 3c9a8f7a6d2 ("intel/dev: generate helpers to identify platform workarounds")
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21908>

dzn: Use mesa_loge for DXIL validation errors

This allows them to be printed to OutputDebugString

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Enable variable size bindings

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Support descriptor indexing via bindless

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Add a debug option for enabling bindless mode

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Apply bindless lowering when compiling pipelines

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Only bind descriptor sets up to the used amount of the current layout

Prevents setting a root SRV into a slot that's not declared as an SRV
in the root signature if a set is bound into a higher slot from a previous
draw/dispatch op.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Ensure root signatures are re-bound after a meta op

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: When binding a bindless root signature, bind descriptor heaps first

The D3D spec says it must be so, and not doing this causes problems
on some hardware.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Add a binding classification in the pipeline layout remapping

This is needed so that we can handle two special cases:
* Dynamic buffer data is allocated out of a command-buffer-owned buffer,
  rather than a descriptor-set-owned buffer, so the remapping puts them
  in their own register space.
* Static samplers should be left alone and not converted to bindless.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Bind buffers for bindless descriptor sets

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Use separate dirty bits for descriptor sets/dynamic buffers

We'll be able to take advantage of this granularity in bindless mode

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Don't dirty bindings if root signature doesn't change

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Allocate descriptor sets in buffers for bindless mode

Modify the root signature stored in the pipeline layout too.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Delete unused function

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Remove defragmenting of descriptor pools

Rather than trying to perfectly defrag, let's just allow re-use.
When a set is allocated for the first time, it locks in its range of
the heap that it'll use. If the last set in the heap is used, then
those descriptors go back to being free, but if a set in the middle
of the heap is freed, those descriptors remain assigned to that set.
A later allocation attempt can reclaim them, as long as the new set
fits.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: When bindless, only allocate one descriptor per layout entry

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Add initial bindless infrastructure

When operating in "bindless" mode, the device will own 2 descriptor
heaps, one for views, and one for samplers. Every time a view is
created (image view, buffer view), a slot is allocated for it out
of the device view heap for each usage type (sampled vs storage).

Then, in a future change, descriptor sets will just contain view/
sampler indices instead of actual descriptors. Instead of copying
these to a cmdbuf-owned descriptor heap, we can directly bind the
descriptor set as a buffer. We'll also modify shaders to perform
an indirection and index into the device heap.

Buffers also get views set up on creation. In a perfect world, we
could just put addresses/sizes in the descriptor set, but DXIL
doesn't support loading from addresses, we need descriptors. When
robust buffer access is disabled *or* descriptor set buffer views
reference the remainder of the buffer, we can just re-use a view
from the buffer and use an offset.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Skip setting up UAVs for depth resources

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Set up SRV descs for 3D textures correctly

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Don't use plane slice 1 for depth+stencil SRVs

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Consistently order depth formats before stencil

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Put UAVs first for storage images/buffers in descriptor tables

When running in a bindless mode, we won't ever be using SRVs for these.
Change terminology for determining descriptor offsets from "writable"
to "alt" to match naming already used elsewhere.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Add some docs around descriptor sets and remove redundant/unused data

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Fix a leak in descriptor set layout creation

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Remove descriptor heap type from descriptor heap wrapper

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

dzn: Remove device pointers from descriptor heaps

A future change is going to add descriptor heaps *to* the dzn_device,
and having 3x ID3D12Device pointers in a single object just seems
wrong. All of the callers already had a device, so just pass it
along where needed.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

spirv2dxil: Support descriptor indexing capabilities

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

spirv2dxil: Only lower readonly images to SRVs when the option is set

This handles the case where readonly is explicitly marked in the shader,
rather than just inferred based on opt_access.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

spirv2dxil: Add a pass to lower deref tex/image and vulkan ubo/ssbo to bindless

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

microsoft/compiler: Update header docs for binding modes supported by compiler

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

microsoft/compiler: Use store_dest instead of store_dest_value more

The 16bit SSBO load support missed setting the 16bit shader flag.
Make it harder to miss that by only using store_dest_value for things
that don't have a type.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

microsoft/compiler: Support descriptor heap indexing for UBO/SSBO

Treat load_vulkan_descriptor on an input that didn't come from
vulkan_resource_index as a descriptor heap index instead of a
binding index.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

microsoft/compiler: Handle "bindless" image/tex sources as heap indices

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

microsoft/compiler: Split handle annotation into two parts

The first part gets the resource props struct, and the second
actually emits the annotate instruction. Later changes will
get the resource props struct from different sources.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

microsoft/compiler: Add helpers for getting res_props structs

Currently we can get one by looking up already-emitted resource
metadata, but in the future we'll want to be able to get this
info from a call site alone. Depending on the type of call site,
we'll have different sets of info, so add helpers for the
various different kinds of call sites we can support.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

microsoft/compiler: Refactor type -> resource kind helper

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

microsoft/compiler: Only set typed UAV load feature bit for multi-comp loads

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

microsoft/compiler: Fix setting bit 31 in feature flags

Fixes: a84208ee ("microsoft/compiler: Fill out and sort the shader/module flags")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21913>

intel/fs: use generated helpers for Wa_14013363432 / Wa_14012688258

Wa_14013363432 is a clone of Wa_14012688258.  It does not apply to all
gfx 12.5 platforms.

Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21745>

freedreno/decode: fix possible overflow

```
../src/freedreno/decode/rddecompiler.c:242:65: error: 'sscanf' may overflow; destination buffer in argument 3 has size 32, but the corresponding specifier may require size 33 [-Werror,-Wfortify-source]
         if (sscanf(info->name, "%32[A-Z0-6_][%32[x0-9]].%32s", reg_name,
                                                                ^
../src/freedreno/decode/rddecompiler.c:243:21: error: 'sscanf' may overflow; destination buffer in argument 4 has size 32, but the corresponding specifier may require size 33 [-Werror,-Wfortify-source]
                    reg_idx, field_name) != 3) {
                    ^
../src/freedreno/decode/rddecompiler.c:243:30: error: 'sscanf' may overflow; destination buffer in argument 5 has size 32, but the corresponding specifier may require size 33 [-Werror,-Wfortify-source]
                    reg_idx, field_name) != 3) {
                             ^
```

Reviewed-by: Rob Clark <robclark@freedesktop.org>
Signed-off-by: David Heidelberg <david.heidelberg@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22015>

ci/freedreno: do not build tools executables without explicitly enabling them

Reviewed-by: Rob Clark <robclark@freedesktop.org>
Signed-off-by: David Heidelberg <david.heidelberg@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22015>

d3d12: Remove now-unused UAV format from shader info

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22093>

d3d12: Fix buffer SRV/UAV creation

Get the format right for UAVs (using the image view instead of
relying on the shader having a format specified) and limit the size
to the API-reported size.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22093>