nvk: Fix geometry shader active stream mask

This sets StreamOutMask in the shader header correctly based on:
https://download.nvidia.com/open-gpu-doc/Shader-Program-Header/1/Shader-Program-Header.html

After this fix transform feedback CTS tests with geometry shader streams
are passing.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Disable PRIMITIVE_RESTART_VERTEX_ARRAY by default

This made various strip geometry tests fail.  Tests with primitive
restart do not seem affected after this.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Increase the push runout to 512 dwords

This is needed in order to do all of viewport state in a single push.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Add A8B8G8R8_*_PACK32 format support

Here it is; the long-awaited change (as requested by 2 people on Discord 🥺)

This change originally added one Vulkan format which was needed for STK's
Vulkan renderer, but @gfxstrand suggested to add all of the A8B8G8R8_*_PACK32
formats in a GitLab thread)

This obviously fixes SuperTuxKart crashing with the Vulkan renderer 🐸

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Advertise VK_KHR_multiview

Closes: https://gitlab.freedesktop.org/nouveau/mesa/-/issues/52

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Add input attachments support for multiview

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Add query support for multiview

When multiview is enabled, queries must use N consecutive query indices in the
query pool, where N is the number of bits set in the view mask in the subpass
the query is used in.

In our case, only the first query is used, so we emit zeros for the remaining
queries.

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Add draw support for multiview

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Break the inner MME draw loop into a helper

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Spill DRAW_IDX pre-Turing

Spill draw_idx to shadow scratch once we've written it to the root
descriptor.  This saves us a bit of register pressure around a pretty
tight seciton.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Rework spill helpers and DRAW_COUNT spilling

The new helpers, now moved to nvk_mme.h, take an nvk_mme_scratch value
to ensure no overlap.  They also ensure that we use exactly the same
register for the unspilled value as the spilled value.  This makes
spilling safe for loop-carried values and the like.  The old
nvk_mme_fill() helper just loads a spilled value to a new, possibly
different register.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Add a helper for re-allocating registers

This helper unlike alloc, tries to use the same value as the register
had before.  This is intended for manual spill/unspill code which wants
to keep register numbers the same while carefully spilling around high-
pressure regions.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Use the new load/store_scratch helpers for DRAW_PAD_DW

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Add begin to mme_scratch

Introduce helpers to make it easier to work with the nvk mme shadow
scratch. Store begin in the shadow scratch in vkCmdDraw* related code to
save registers, which is necessary for Fermi.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Lower nir_intrinsic_load_view_index

Lower view_index intrinsics to a load_ubo using NIR.

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Add view_index to root descriptor table

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Fix while loops pre-Turing

Pre-Turing, we don't have a full condition built into the loop so, for
mme_while(b, ine, x, y) we have to generate the following:

[top]   BR -> [cond]
        /* loop body */
[cond]  XOR tmp, x, y
        BRZ !tmp -> [top]

However, due to an accounting error, we were generating

[top]   BR -> [cond]
        /* loop body */
        XOR tmp, x, y
[cond]  BRZ !tmp -> [top]

which meant that the XOR (or ADD if one is an immediate) was getting
skipped, leading to the loop either never terminating or always
terminating.  The way to fix this accounting error is to close the while
first, then compute the condition value, then do the jump.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Use ADD for ine with an immediate

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Add a macro to set MMIO registers via falcons

Signed-off-by: Karol Herbst <kherbst@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: add a macro exit helper

At the moment it's turing+ only.

Signed-off-by: Karol Herbst <kherbst@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: add test for BEQ with magic exit offset

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Fix some cast defines

This removes compiler warnings on 32-bit builds 🐸

Also other drivers do define casts this way, so this should actually
be an uncontroversial change 😁

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Add instace WSI entrypoints

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Advertise ICD/loader interface version 4

We can't advertise v5 yet because we don't support Vulkan 1.1 yet.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Use the same lock for the submit and the memory objects list

We need to lock the memory objects list around the entire push builder
process.  Otherwise, a memory object could get destroyed between adding
it to the BO list and doing the submit, leading to the kernel rejecting
our pushbuf.

When the time comes that we switch to the new uAPI both uses of
nvk_device::mutex will go away.  We'll no longer be passing lists of BOs
to the kernel and we'll move to syncobj and be able to drop all the
nvk_bo_sync nonsense.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: advertise DemoteToHelperInvocation

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: fix r32_sint format support

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: add min_lod to spirv caps.

We expose the vulkan feature

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: add compute support for ampere

ampere has support for QMDv3, it's not mandatory but lets use it for now.

the more important change is the PCAS2_B call.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Clamp viewport clip to max range

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk/barrier: handle host bit.

Fixes dEQP-VK.pipeline.monolithic.timestamp.transfer_tests.host_stage* crashes

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Support vertex shader transform feedback on Fermi

Byte counts should be fetched with a different way in case of Fermi.
We cannot read the buffer directly in the mme.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Support transform feedback queries

We need 2 reports for these. One for successfully written primitives
and one for needed primitives. Also we need to use report semaphore's
sub_report member to specify the query's corresponding vertex stream.
Finally availability index is 2 in the query results, as we have 2
results for succeeded and needed primitives.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Support transform feedback geometry streams

We just need to set the correct limits and features.  The meaningful
change here is the selected stream for rasterization which is configured
through the graphics pipeline.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Support transform feedback indirect draws

This draw needs to read the byte count indirectly so it is implemented
as an mme function.  Some refactoring was done in the draw functions, so
that nvk_mme_build_draw could be used for this case.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: De-duplicate MME code for setting draw params

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Ensure that zero-initizlied mme_value is ZERO

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Support for vertex shader transform feedback

For starters, we support only vertex shader output transform feedback.
Optional support for queries, streams and indirect draw are left out for
now.

Pipeline and shader state is based on nvc0 code. Most of the shader
state is going to change with the new compiler.

Required support for pause/resume is implemented with an mme function
that loads the offset indirectly from the counter buffer on resume.
For pause, we store the offset indirectly with a SET_REPORT_SEMAPHORE.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk/nil: don't ask for compressed image kind

These need special kernel handling around the mmu, avoid asking
for them, nvidia don't seem to use them with gsp either yet.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/winsys: allocate unique object handles across channels.

GSP requires unique object handles across channels, otherwise
it reports an error.

This fixes vulkaninfo on gsp enabled fw

Reviewed-by: Karol Herbst <kherbst@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Add a bfe helper

This extracts a fixed number of bits from a possibly variable position
in the source value.  Very helpful for testing for a bit.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Advertise shaderImageGatherExtended

Closes: https://gitlab.freedesktop.org/nouveau/mesa/-/issues/62
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Rework features enables

Use the new feature enables framework.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Rework extension enables

Take an nv_device_info instead of a full nvk_physical_device.  This
prevents errors where nvk_get_device_extensions() may access an
uninitialized field of nvk_physical_device.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: advertise VK_KHR_maintenance4

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: implement vkGetDeviceBufferMemoryRequirementsKHR()

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: add stub for vkGetDeviceImageSparseMemoryRequirementsKHR()

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: implement vkGetDeviceImageMemoryRequirementsKHR()

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Add a helper to dump instructions

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Fix nested while instructions on Turing+

Closes: https://gitlab.freedesktop.org/nouveau/mesa/-/issues/49
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Bump init context batch size

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: set device info before use in nvk_get_device_extensions

nvk_get_device_extensions uses device->info to enable extensions
only for supported device architectures.

Only info is needed, but it seemed cleaner to keep all the three
changed lines together.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Implement VK_KHR_draw_indirect_count on Turing+

v2: handle maxDrawCount parameter
v3: assert on pre-turing. Free regsisters after use.
v4: less register presure. Update to pass new tests

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: add warning on non-nouveau drm driver

Check that the drm kernel driver in use is nouveau, and
add a short description of the issue.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Return VK_ERROR_INCOMPATIBLE_DRIVER if the PCI vendor isn't NVIDIA

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Support VK_KHR_shader_draw_parameters

We need to supply the base vertex, the base instance and draw index to
the shader for this extension. Similar to the base instance we supported
before, we load base vertex and draw index as root constants at draw time.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Simplify mme build function argument

We do not need the full nvk_device struct for the mme builder functions,
just the nv_device_info is enough. We keep a pointer in the
mme_builder so we can use this instead. Also, when we run mme builder
tests we do not need to initialize a device struct.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Keep device info in mme_builder

We keep a pointer to nv_device_info instead of just copying the GPU's
3D engine class.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Advertise VK_EXT_robustness2

Closes: https://gitlab.freedesktop.org/nouveau/mesa/-/issues/39

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Advertise VK_EXT_image_robustness

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Set image index to zero for NULL nvk_buffer_view

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Advertise VK_EXT_mutable_descriptor_type

Closes: https://gitlab.freedesktop.org/nouveau/mesa/-/issues/41

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Enable VK_KHR_maintenance3

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Implement nvk_GetDescriptorSetLayoutSupport

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Don't over-size push descriptor sets

We were over-allocating them by 4x.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Replace asserts with conditional that sets type_list = NULL

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Update mutable descriptor struct type

Signed-off-by: Rebecca Mckeever <rebecca.mckeever@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: fix print inst for case MME_FERMI_OP_MERGE

By applying the same coding style of the other case statements,
this patch fixes the following building errors:

../src/nouveau/mme/mme_fermi.c:209:10: error: expected expression
         uint32_t src_bit  = inst->bitfield.src_bit;
         ^
../src/nouveau/mme/mme_fermi.c:216:39: error: use of undeclared identifier 'src_bit'; did you mean 'dst_bit'?
         fprintf(fp, " (%u, %u, %u)", src_bit, size, dst_bit);
                                      ^~~~~~~
                                      dst_bit
../src/nouveau/mme/mme_fermi.c:211:19: note: 'dst_bit' declared here
         uint32_t dst_bit  = inst->bitfield.dst_bit;
                  ^
2 errors generated.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: use fermi enum in fermi builder

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Add a build test for MMEs

This simple test compiles all the MMEs for Turing and Volta to ensure
that they build and don't run out of registers.  We could, in theory, do
actual unit testing here with gtest but just building is enough to
ensure that no one breaks driver start-up on hardware they don't own.
They can still break the driver functionally, of course, but at least
it'll initialize.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

vulkan/meta: Use the new NIR texture helpers

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Implement Draw*Indirect on pre-Turing

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/push: Add an NV_PUSH_MAX_COUNT #define

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Reduce register pressure in nvk_mme_build_draw*()

Re-order things a bit so we deal with vertex offset and first instance
separately and free those regs when we're done.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Don't allow WaW dependencies in the same Turing instruction

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Make some MME builder names more consistent

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Add a helper for pushing indirect data

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Rework nvk_cmd_push a bit

Instead of taking an nvk_cmd_bo take a nouveau_ws_bo and keep the map
pointer separate.  We'll need to do this when we start pushing stuff
that isn't nvk_cmd_bo.  Also, rework the bounds to be in bytes.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk/drm: Take a byte offset/range in push_add_push

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: nouveau/mme: Add builder tests for SLL and SRL

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Use a uint32_t for size in mme_fermi_bfe()

It will never be a non-fixed value.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Add builder tests for SUB

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Add some generic builder tests

These don't need hardware and instead run entirely in the simulator.
The goal is to have something that we can run in CI and which ensures
consistency between the Turing and Fermi MMEs.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Unify some of the test framework

Most of the actual nouveau device enumeration and test running framework
was copy+paste between the two.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Take a const nv_device_info in mme_builder_init

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Don't swap x and y in mme_fermi_merge_to()

This was probably an artifact of the fermi/turing merge.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/mme: Fix is_int18 for negative numbers

util_mask_sign_extend returns a uint64_t which we need to cast to
uint32_t in order to get the right comparison.  Otherwise, we risk
having ~0 in the top 32 bits and the comparison failing.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nil: Only choose a PTE kind for tiled images

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Report image alignments from NIL

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nil: Fix image align and size constraints

In order to use compressed images we have to align to at least (1<<16)
though the kernel also likes (1<<21).

This fixes the alignments/sizes so they can work with vm allocs

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: set the device fd

This is needed later for syncobj support

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk/winsys: store device ptr into bo instead of ptr

This will be needed later.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: init dev->physical_device earlier.

When device is passed to bo destroy this is needed earlier.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Cosmetic clean-ups to Create/DestroyDevice

Use dev and pdev instead of device and physical_device.  Also, clean up
whitespace a bit while we're here.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Handle zero-size index and vertex buffers pre-Turing

According to the nouveau GL driver, this should be the addres of the
last byte, not the byte after the end.  The calculation will overflow if
range is 0 so we need to handle that special.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Use a GS for layerered rendering pre-MaxwellB

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

vulkan/meta: Insert a geometry shader when needed

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Compile all NIR before running codegen

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Compile geometry shaders

This enables compiling geometry shaders.  Based primarily on gallium
nvc0/nvc0_program.c.

We need to enable/disable user clip planes based on the last geometry
stage.  Some asserts in codegen need to be changed too because the
compacted clip distance inputs are arrayed in the geometry shader. So we
have an array of clip distances for each input vertex.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nvk: Properly align root descriptor tables for pre-Pascal

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>