ci/test-vk: uprev vkd3d

This commit uprevs vkd3d, which flags
test_primitive_restart_list_topology_stream_output as being a flake on
RADV / NAVI21.

Suggested-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Martin Roukala (né Peres) <martin.roukala@mupuf.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20278>

ac/nir: add ac_nir_lower_legacy_gs

For legacy (non-NGG) GS to lower outputs to memory stores and add
shader query when required.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20158>

ac/nir: move ngg_gs_shader_query to a common function

To be shared by NGG GS and legacy GS. Legacy GS need this when
GFX10 which mix use NGG and legacy GS. For example when streamout
is enabled, it uses legacy GS, otherwise uses NGG GS. So legacy
GS also need to update query emulation which is a sum of NGG and
legacy GS results.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20158>

nir,ac/llvm,radv: add stream id index to nir_load_ring_gsvs_amd

For used by legacy GS to store output to different ring according
to stream id.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20158>

ac/llvm,radv,radeonsi: pass instruction to intrinsic_load abi

For simple intrinsic which also need other fields to translate
to LLVM like stream_id.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20158>

nir: add nir_load_ring_gs2vs_offset_amd

Used by legacy GS output lowering.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20158>

ac/nir,radeonsi: replace llvm ngg gs invocation query with nir

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20156>

nir,radv,radeonsi: add nir_atomic_add_gs_invocation_count_amd

For shader query emulation.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20156>

freedreno/registers: Add some docs for PM4 events.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19733>

doc/freedreno: Add a bunch of docs of the hardware and drivers.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19733>

freedreno: Clarify that CACHE_FLUSH is pre-5xx.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19733>

ci/docs: Always expose the built docs.

When testing my docs changes outside of an MR, there was no way to look at
the result.

Acked-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19733>

intel: Fix crashes for importing drm buffer

image_aspect_to_binding() converts aspect to index by subrracting
VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT, however these enum values
are bitfields, not consecutive numbers, so comparing and subtracting
them won't work.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20269>

hasvk: pipelineStageCreationFeedbackCount is allowed to be 0

Fixes: 6601e5d6fc6 ("anv: implement VK_EXT_pipeline_creation_feedback")

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20216>

vdpau: allow building vdpau st with virgl only

Fixes: 6b5aecb1955 ("virgl: add support for hardware video acceleration")

Signed-off-by: Jan Palus <jpalus@fastmail.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20108>

glsl: fix function inlining for images

Here we skip replacing parameters with their actual values for
images as glsl_to_nir() expects them to be copied to temps first.
Tree grafting has a similiar rule to avoid this happening also.

Fixes: 8d10a6835f32 ("glsl: dont create temps for builtin function inputs")

Tested-by: Martin Roukala <martin.roukala@mupuf.org>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20274>

radv: Improved export conflict bug workaround.

Based on f129db911bd ("radeonsi/gfx11: use a better workaround for the export conflict bug")
which claims better performance.

I couldn't be bothered to do the refactor to check the sample count with
dynamic sample counts, so this is just conservative there.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20273>

mesa: move the _mesa_set_varying_vp_inputs call to where the state changes

This removes the _mesa_update_vao_state() call from all glDraw* functions
where it was executed unconditionally.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: move _DrawVAOEnabledAttribs determination into st_update_array

This can just be computed where it's used.

Now the non-glDraw paths like glRasterPos, glBegin/End, and the slow
display list path have to save and restore _VPModeInputFilter, which
is the only field that's different from the glDraw* path.

_VPModeInputFilter is a bitmask of VP inputs that might have to be bound.
The only difference with glBegin/End and the slow display list path is
that they also add VERT_BIT_MAT_ALL to the bitmask, whereas the glDraw* path
doesn't have that.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: inline _mesa_draw_array_bits & _mesa_draw_current_bits

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: don't AND with VERT_BIT_ALL because it's ~0u

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: simplify VBO state flagging in _mesa_update_vao_state

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: remove VAO::NewVertexBuffers/Elements, set the dirty flags directly

These intermediate dirty flags are unnecessary now.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: move calling _mesa_update_vao_derived_arrays into st_update_array

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: remove the NewVAO flag, set directly what we need to set

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

st/mesa: move st_atom.c contents into st_context.c

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

st/mesa: inline st_validate_state and remove redundant checking in callers

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

st/mesa: remove the switch from st_validate_state by passing state mask directly

Instead of passing the enum that represents which states should be updated
to st_validate_state, pass the state bitmask to st_validate_state directly.
This removes the switch statement over the enums.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

st/mesa: remove st_context::dirty, use gl_context::NewDirtyState instead

We stored the same state in 2 different variables, and then we OR'd them.

This changes the st_validate_state logic slightly, but should be identical.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: remove per_vertex_edgeflags parameter from DrawGalliumVertexState

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: completely rewrite the edge flag state update logic

This removes the edge flag logic from st_validate_state, which was always
executed and added overhead there. Now it's done only when needed.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: don't call _mesa_set_varying_vp_inputs in set_vertex_processing_mode

It's redundant because it's always called in _mesa_update_draw_vao.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: don't call _mesa_update_vao_derived_arrays for dynamic VAOs

This just moves the exit into the caller. There is nothing else in
_mesa_update_vao_derived_arrays that dynamic VAOs care about.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: merge 2 switch statements into one in _mesa_get_derived_vao_masks

_mesa_vao_enable_to_vp_inputs has a switch statement and we call it twice.
Let's inline it.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: remove derived _EffEnabled* fields and compute them at their only use

We were passing the fields to their only use and used them only once,
so let's just compute them there.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: merge _mesa_draw_(user_array,nonzero_divisor)_bits into one helper

preparation for the next change

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: remove unused inline functions from arrayobj.h

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: keep the current VAO also set as _DrawVAO

This removes setting _DrawVAO from all glDraw* functions that draw
primitives because now _DrawVAO is always equal to the current VAO.

glEnd, glRasterPos, and the slow display list path now have to save
_DrawVAO before changing it, and restore it after they are done.

This also removes _EmptyVAO, which becomes unused because we no longer
bind "no VAO" into _DrawVAO.

Rules:
- only glBindVertexArray and non-glDraw* paths should call
  _mesa_set_draw_vao, other places shouldn't
- only draws should call _mesa_update_vao_state, other places shouldn't

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

mesa: split _mesa_set_draw_vao into set and update functions

This is the beginning of splitting _mesa_set_draw_vao and moving pieces
of it out of draw functions.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19953>

radv: Handle NULL miss shaders

Fixes reflections in DOOM Eternal.

Fixes: 85580faa ("radv: Add ray traversal loop.")
Closes: #6210
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20270>

radv: Change shaderGroupBaseAlignment to RADV_RT_HANDLE_SIZE

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20270>

radv: Force non uniform texture sampling with NV Remix

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20243>

radv: Add a driconf option to force non uniform texture sampling

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20243>

spirv: Add a debug option to force non uniform texture sampling

Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20243>

radv/rt: Hash the pipeline key

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20243>

radv/rt: Propagate radv_pipeline_key

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20243>

radv/rt: Work around incorrect NOT_DEFERRED handling in Portal RTX

Allows Portal RTX to reach the main menu.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20243>

tu: Implement VK_EXT_descriptor_buffer

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19849>

vk/descriptor_set_layout: Add optional destructor

Drivers implementing descriptor buffers will want to allocate and free
descriptors with the layout for embedded samplers, so we need a hook to
allow them to free any GPU buffers.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19849>

tu: Fall back to ldg for variable-size inline uniform blocks

With descriptor buffers, we have no way to know how big the descriptor
set actually is, so we have no idea how many constants we can safely
push. If we use a UBO then it will still get pushed, because we normally
assume that we can freely access UBOs without any fear of faults due to
the range checking. This does the easiest thing of using raw pointer
loads, although performance will fall off a cliff, because we don't have
many better options.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19849>

ir3/a6xx: Fix immediate offset stg/ldg path

The immediate offset is in units of bytes, whereas the register offset
is in dwords. We need to compensate for that.

Also, fix an off-by-one when checking the range - the offset field is 13
bits, but the sign bit means we can only represent up to 1 << 12 in
bytes or 1 << 10 in dwords.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19849>

microsoft/compiler: Fix codegen when a loop ends in a jump

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/7792
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20255>

microsoft/compiler: Handle holes in driver_location when adding sysvals

All of the full runtime+compiler stacks reassign these driver_location
values to compact them and sort between shader stages, but for the
spirv2dxil tool, we leave the original shader's "location" intact. That
means that there can be holes in the driver_location space, and simply
counting how many inputs there are can lead to collisions. So instead
place all sysvals after the last-used driver_location.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/7811
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20253>

ci: Bump virglrenderer version

Update virglrenderer to the latest version on time.

Signed-off-by: Corentin Noël <corentin.noel@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20277>

ir3/freedreno: Find regs for FS inputs when printing info

FS inputs are not directly loaded into regs, but require additional
instruction to do so. So in order to print in which reg the input
is loaded we have to scan the shader for the instruction
which loads the input.

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20247>

radv: fix assertion on gpu hang detection

fixes assert in RADV_DECL_PIPELINE_DOWNCAST when bound pipline is a compute
pipeline

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20276>

intel/compiler: Remove leftover declarations of old NIR passes

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19805>

intel/nir/rt: fixup primitive id

There is a delta index value in the hit structure, we forgot to add it
to the base value.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 046571479028 ("intel/nir/rt: add more helpers for ray queries")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/7565
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19346>

radv: fix hashing descriptor set layout

Shouldn't have pointers.

Fixes: 19f8d338761 ("radv: Use vk_descriptor_set_layout")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20250>

nir/lower_shader_calls: Remove phis after dead control flow

This potentially gets rid of some more phis without sources.

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19960>

nir: Do not consider phis with incompatible dests equal

CSE tries to collapse equal instructions, and collapsing two phis with incompatible dests is illegal.

Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Fixes: 6bdce55c ("nir: Add a basic CSE pass")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19960>

docs/release-calendar: add 22.3.x dates

Signed-off-by: Eric Engestrom <eric@engestrom.ch>
Reviewed-by: David Heidelberg <david.heidelberg@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20131>

zink: Don't set dynamic color attachment state for 0 attachments.

Fixes some validation failures like:

VUID-vkCmdSetColorBlendEquationEXT-attachmentCount-arraylength(ERROR / SPEC): msgNum: -175001922 - Validation Error: [ VUID-vkCmdSetColorBlendEquationEXT-attachmentCount-arraylength ] Object 0: handle = 0xaaaae7632fa0, type = VK_OBJECT_TYPE_DEVICE; | MessageID = 0xf591aebe | vkCmdSetColorBlendEquationEXT: parameter attachmentCount must be greater than 0. The Vulkan spec states: attachmentCount must be greater than 0 (https://www.khronos.org/registry/vulkan/specs/1.3-extensions/html/vkspec.html#VUID-vkCmdSetColorBlendEquationEXT-attachmentCount-arraylength)

However, we still have some around dynamic color attachment state:

    Objects: 1
        [0] 0xaaaafcab4150, type: 6, name: NULL
VUID_Undefined(ERROR / SPEC): msgNum: 2044605652 - Validation Error: [ VUID_Undefined ] Object 0: handle = 0xaaaafcab4150, type = VK_OBJECT_TYPE_COMMAND_BUFFER; | MessageID = 0x79de34d4 | VkCommandBuffer 0xaaaafcab4150[]: Dynamic color blend enable state not set for this command buffer.
    Objects: 1
        [0] 0xaaaafcab4150, type: 6, name: NULL
VUID_Undefined(ERROR / SPEC): msgNum: 2044605652 - Validation Error: [ VUID_Undefined ] Object 0: handle = 0xaaaafcab4150, type = VK_OBJECT_TYPE_COMMAND_BUFFER; | MessageID = 0x79de34d4 | VkCommandBuffer 0xaaaafcab4150[]: Dynamic color blend equation state not set for this command buffer.
    Objects: 1
        [0] 0xaaaafcab4150, type: 6, name: NULL
VUID_Undefined(ERROR / SPEC): msgNum: 2044605652 - Validation Error: [ VUID_Undefined ] Object 0: handle = 0xaaaafcab4150, type = VK_OBJECT_TYPE_COMMAND_BUFFER; | MessageID = 0x79de34d4 | VkCommandBuffer 0xaaaafcab4150[]: Dynamic color write mask state not set for this command buffer.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20186>

radv: Use correct watermark for early loop exit.

The previous check assumed the stack starts at offset=0, which isn't
necessarily true for ray queries.

Note that this didn't cause correctness issues, just made an optimization
not apply. Found when I accidentally made this load-bearing in a
refactor.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20265>

radv: Rename stack_base to stack_low_watermark.

Better covers the purpose.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20265>

Revert "r600/sfn: Make use of variable length DOT"

This reverts commit fcafe1ffc85573b6a0df2d64a438f1bac2bae3a2.

Variable length DOT products are not supported for pre EG
cards, and the read port evaluation is not correctly
checked, so that scheduling might fail. Revert for now to
fix the issues below and get gack with a better implementation
later.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/7876
Related: https://gitlab.freedesktop.org/mesa/mesa/-/issues/7878

Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20268>

cso: remove cso_draw_vbo from all draws, call the driver or u_vbuf directly

Instead of calling like this:
    st_draw_gallium -> cso_draw_vbo -> driver_draw_vbo

Do it like this:
    st_draw_gallium -> driver_draw_vbo
OR
    st_draw_gallium -> u_vbuf_draw_vbo

It's accomplished by adding a draw_vbo function pointer into cso_context.
The pointer is equal to pipe_context::draw_vbo when needed, so there is
no call overhead from this if cso's draw_vbo callback is indeed equal to
driver_draw_vbo. We just call cso_context_base::draw_vbo to jump into
the driver directly, or u_vbuf if needed.

The cso function with the indirect function call is inlined, so draws
don't actually visit any cso_context function.

Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20025>

cso: add a base class cso_context_base holding pipe_context*

We'll add more stuff there. The first change is that we need pipe_context*
there.

Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20025>

cso: unify cso_draw_vbo and cso_multi_draw

This is going to be inlined.

Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20025>

gallium/u_vbuf: change u_vbuf_draw_vbo to accept pipe_context as first param

This makes the parameters equal to pipe_context::draw_vbo.

Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20025>

gallium: add the u_vbuf pointer into pipe_context

This will allow removing the draw_vbo wrapping in cso_context.

Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20025>

asahi: Implement occlusion queries

While the hardware supports both counter and boolean occlusion queries, the
programming model is quite different from OpenGL. In AGX (and in Metal),
there is a single "visibility result buffer" associated with the render pass.
Each draw that uses occlusion queries writes into this render pass global
visibility result buffer at a particular index. By contrast, the OpenGL
occlusion query model supposes that each query has independent state that can be
mixed and matched within a render pass. We can't simply allocate backing memory
for a query and write to it from a job. We can't allocate visibility result
buffers for each batch up front and statically assign OpenGL queries to indices,
because the OpenGL query can span multiple batches. Finally we can't use a
global visibility result buffer without introducing additional synchronization,
given that we now support multiple batches in-flight at once.

In this patch, I've elected to use a simple solution: allocate visibility result
buffers and indices on the fly as needed, and accumulate the results on the CPU
at the end of the render pass. When we have proper synchronization we'll want to
revisit this, but as everything is stalling at submit time now, I'm not inclined
to "optimize" something I can't test.

Passes dEQP-GLES3.functional.occlusion_query.* and the relevant piglit tests.
The piglits are considerably more thorough, checking lots of "we hate tilers"
conditions that dEQP skips over.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Move query functions to agx_query.c

New file. They're just stubs now but will get nontrivial in a moment.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Don't upload samplers for clears/stores

Unlikely to help but makes the traces neater.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Avoid reloads with staging blits

Noticed by inspection. Not likely to matter unless these staging blits are in a
hot path, but it's an easy win.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Don't reload uninitialized surfaces

Pointless. This should save some bandwidth in some cases (possibly mipmap
generation?)

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Fix Z32S8 harder

Fixes dEQP-GLES3.functional.texture.format.sized.2d.depth32f_stencil8_pot after
stencil texturing broke it.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Don't shadow idle resources

Pointless allocation+memcpy.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Model alignment of occlusion query indices

8-byte offsets.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Identify some more fields used with layered

These values depend on the framebuffer width/height and maybe other
stuff. Maybe strides.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Identify XML for anisotropic filtering

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Check-box implement rasterizer discard

Passes dEQP-GLES3.functional.rasterizer_discard.*

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Wire in 1D (array) textures

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Dirty track VS/FS key updates

drawoverhead 1 score doubled to 7668.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Separate VS/FS shader keys

First remove agx_shader_key from asahi_shader_key. It's trivial. agx_shader_key
is going to go away soon now that we lower everything in NIR.

Then everything else is mutually exclusive between stages. That means much
less to hash. drawoverhead test 1 from 2331 to 3443.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Implement invalidate_resource

From Panfrost. This lets us avoid storing depth/stencil attachments at the end
of the frame in GLES.

On my 4K monitor, glmark2 -btexture at fullscreen goes from 705fps to 1150fps. I
assume gains on real workloads will be smaller.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Track batch masks on ZS/blend CSO

Adapted from panfrost, with the work happening at CSO create time instead of
draw time allowing us to do more sophisticated analysis. We'll use these for
accurate masks in a moment.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Dirty track blend state

We'll want this to reduce variant lookups eventually.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

asahi: Enable dirty tracking

Whoops. drawoverhead test 1 score from 496 -> 2377.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

ail: Save level_offsets_compressed_B

So we can bind specific mip levels for rendering into compressed Z/S.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

panfrost: implement clear_depth_stencil

Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: David Heidelberg <david.heidelberg@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20238>

nine: enable on panfrost

Also, enable required kmsro dependencies.

Tested-by: Aleksey Komarov <q4arus@ya.ru>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Signed-off-by: David Heidelberg <david.heidelberg@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20238>

panfrost: implement clear_render_target

Signed-off-by: David Heidelberg <david.heidelberg@collabora.com>
Co-authored-by: Aleksey Komarov <q4arus@ya.ru>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Tested-by: Aleksey Komarov <q4arus@ya.ru>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20238>

panfrost: Handle resources without depth in batch_to_fb_info

Prevent preloading data from resources which doesn't exist.

Signed-off-by: David Heidelberg <david.heidelberg@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Tested-by: Aleksey Komarov <q4arus@ya.ru>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20238>

panfrost: Implement GL_EXT_clip_control

Signed-off-by: David Heidelberg <david.heidelberg@collabora.com>
Co-authored-by: Aleksey Komarov <q4arus@ya.ru>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Tested-by: Aleksey Komarov <q4arus@ya.ru>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20238>

intel: add devinfo->has_64bit_float_via_math_pipe

Unusual hardware features that require special hanlding usually get a
devinfo field, so do this for MTL's unordered DF types. This will
guarantee that any platform based on MTL (thus inheriting from
MTL_FEATURES) will automatically be handled in these special cases.

v2: s/has_unordered_64bit_float/has_64bit_float_via_math_pipe/ (Curro).

Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20072>

intel/compiler: fix intel_swsb_decode for newer platforms

In the previous patch we adjusted the scoreboard pass to take into
consideration a new case of unordered operations for TGL. Fix the
decoding as well.

v2: use intel_device_info_is_mtl()  (Curro, Jordan)
v3: the part where we export num_sources_from_inst() is now a separate patch
    (Curro).
v4: Work around false positive maybe-unitialized warning since Marge
    uses -Werror=maybe-uninitialized (Marge).

Reviewed-by: Francisco Jerez <currojerez@riseup.net> (v3)
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20072>

intel/compiler: export brw_num_sources_from_inst

We want to call this from brw_disasm.c, so move it out to brw_eu.c
since it's about to become more of a shared utility function than
something specific to the EU validator.

Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20072>

intel/compiler: avoid 64bit SEL_EXEC on MTL

On MTL, instructions with DF type are unordered, executed in the math
pipe. This means that they require different SWSB dependency handling,
and also that in some cases such as MOVs it's generally faster to
simply use 2 smaller ordered moves than a single unordered MOV.

One problem we have with the current code is that generate_code() is
not setting the proper SWSB dependencies for the generated DF MOVs,
causing some tests to fail.

One solution would be to fix generate_code() by making it set the
appropriate dependencies. This was the first patch I wrote. Another
solution to this problem, pointed to us by Curro, is to change
required_exec_type() so we use UD instructions instead of DF, just
like we do with platforms that don't have 64 bit instructions, which
means there won't be anything to fix in generate_code(). The second
solution is what this patch implements.

This fixes at least:
 - dEQP-VK.subgroups.arithmetic.framebuffer.subgroupmin_double_vertex

Thanks to Francisco Jerez for all the major help provided with this
problem.

Credits-to: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20072>

intel/compiler: avoid (RegDist, SBID) on DF instructions on MTL

When we use this form there's no way to specify which pipe RegDist
refers to, so there are a few rules to figure this out, which is what
inferred_sync_pipe() implements. But for MTL there's no long pipe and
the documentation does not explicitly explain what should be the
inferred type for its long (DF) instructions - which are out-of-order,
by the way.  One way to interpret this is that such case should be
avoided.  So add the extra check to entirely avoid this case.

Notice that this is not actually fixing any bug, since returning
TGL_PIPE_LONG (what we do today) will actually make these DF
instructions incompatible with every in-order instruction, so we'll
never opt to use the (RegDist, SBID) form anyway. But still, it's
better to have this case explicitly documented instead of having it
covered by a semi coincidence.

v2: use intel_device_info_is_mtl()  (Curro, Jordan)

Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20072>