nvk: Don't free the NIR in nvk_compile_nir

It's a layering violation and the NIR is currently leaked if we the
compile fails.  It's better to make the caller clean up.

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

nvk: Switch to vk_pipeline_shader_stage_to_nir

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

nvk: Split pipeline binding into helpers

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

nvk: Add graphics state to command buffers

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

nouveau: Generate 3D headers

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

nouveau: Null terminate the debug flag list

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

nouveau/push: Make P_IMMD more versatile

It's now able to use either the immediate form or method form depending
on whether or not the value fits into 13 bits.  We wrap the check with
__builtin_constant_p() so that this choice is compile-time determined
instead of runtime.  I've verified with SET_CT_SELECT() that GCC is able
to determine which bits are used even if not all of them are constants.

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

nouveau/parser: Handle arrays properly in P_IMMD()

We make a new macro VA_NVCLS_FOO which is either an array macro or a
regular macro and, either way, redirects to V_NVCLS_FOO.  This lets us
unconditionally use VA_NVCLS_FOO in P_IMMD() and, if the caller hands in
a FOO(i), it will get unwrapped and mangled to produce to V_NVCLS_FOO.

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

nouveau/parser: Better dump float data

For now, we only do something with floats when we dump.  You still need
to use fui() when setting them.

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

nouveau/parser: Properly dump most arrays used by 3D

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

nouveau/parser: Fix an integer overflow and a typo

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

nvk: Init WSI after setting up supported_sync_types

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

nouveau/ws: allocate copy subchan as well

Starting with Ampere we have to do this. It is a safe thing for previous
gens to do.
Very old kernels might reject this, but too lazy to check.

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

nouveau/ws: allocate 3D subchan

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

nouveau/headers: initial 3D headers import

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

nouveau/headers: add path for 3D headers

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

nvk: set deviceName

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

nouveau/push: Add a P_INLINE_FLOAT helper

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

nvk: Make image/buffer address helpers const

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

nvk: Include nvk_private.h in everything

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

nouveau/ws: remove the drm.h header

It can cause compilation issues on Arch, there is a workaround by
including it as <drm.h> but this solution is simpler.

Signed-off-by: Yusuf Khan <yusisamerican@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nouveau/ws: fix building error in nouveau_ws_push_dump()

PRIxPTR macro solves the problem

Fixes the following build error:

../src/nouveau/winsys/nouveau_push.c:119:43: error: format specifies type 'unsigned long' but the argument has type 'int' [-Werror,-Wformat]
      printf("[0x%08lx] HDR %x subch %i", cur - push->orig_map, hdr, subchan);
                 ~~~~~                    ^~~~~~~~~~~~~~~~~~~~
                 %08x
1 error generated.

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

nouveau/ws: use new NVIF interface to query oclasses

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

nouveau/ws: get rid of libdrm

and just completely rework device and context creation

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

nouveau/ws: depend on generated class header files

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

nvk: Padd shader BOs by 4K to avoid I-cache overflow

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

nvk: Bind immutable samplers on descriptor set creation

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

nvk: Rework OOM handling for descriptor pools

We now properly return VK_ERROR_OUT_OF_POOL_MEMORY if we run out instead
of silently assigning a NULL bo and map.

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

nvk: Advertise minUniformBufferOffsetAlignment

Fixes a bunch of dEQP-VK.image.store.with_format.*_constant tests

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

nvk: No-op sparse image format properties

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

nvk: Unconditionally zero image format properties

It seems like this shouldn't be neded but some of the tests complain if
we don't because they call vkGetPhysicalDeviceImageFormatProperties()
and vkGetPhysicalDeviceImageFormatProperties2() and compare if they
return identical results even if they return error.  Zero is the
reasonable thing to do if you don't support a format anyway because then
it's a maximum extent, sample count, etc. of zero.

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

nvk: Initial vkCmdClearImage support

Doing something for some formats is better than silently doing nothing.

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

nvk: Assert that we don't double-free descriptors

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

nvk: Delete the storage TIC in nvk_image_view_destroy

Also, assert that descriptors are non-zero.  Zero is reserved for the
null image descriptor.

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

nvk: Support inline uniform blocks

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

nvk: Use a switch for descriptor types in load_descriptor

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

nvk: Fix descriptor offset alignment

This fixes two bugs.  First, we were aligning when setting the bindign
offset but not actually applying that alignment to buffer_size so when
we incremented by stride * descriptorCount, it incremented starting with
the unaligned value.  Second, in some cases we return stride = align = 0
and the ALIGN_POT macro will return 0 when something's aligned to 0.

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

nvk: Properly indent a comment

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

nvk: Implement push constants

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

nvk: Load the requested descriptor size

We were hard-coding to vec4s for some reason.  This results in vec4
texture handles which, while properly ignroed by codegen, are just
weird and not what we want.

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

nvk: Handle cube storage images properly

TIC entries for cubes divide the array size by 6 so we don't want that
when using it as a storage image.  This means nvk_image_view needs
separate TIC entries for storage vs. sampled now.

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

nvk: Use a descriptor type instead of a hand-rolled thing

This was an attempt to prematurely optimize some things.  Sure, it's a
hot-path, but there's other control-flow in there and we can optimize it
later if needed.

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

nvk: Advertise non-zero descriptor set limits

Right now, everything's pretty much unlimited.  We may want to limit
UBOs in the future to just what we can bind but everything else will
probably remain pretty high.

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

nvk: Advertise descriptor array indexing

This isn't as dynamic as it looks.  It still requires the index to be
uniform, it just lets it be not a constant.  Given how our descriptor
sets work, this is all handled trivially already.

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

nvk: Better advertise image format features

We now advertise the top-level features and we don't falsely claim
write-without-format support per-format.  In theory, NV hardware should
be able to do shaderStorageImageReadWithoutFormat but codegen doesn't
know how yet.

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

nvk: Add support for dynamic buffers

These are recorded in the CPU portion of the descriptor set and combined
with the dynamic offset at vkCmdBindDescriptorSets time and placed into
the root descriptor table.

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

nvk: Use the correct root descriptor table size for CmdDispatch

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

nvk: Advertise KHR_dedicated_allocation

We're not requiring or even requesting anything to be dedicated yet but
we may as well put in the plumbing.  This gets more tests running.

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

nvk: Implement buffer views

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

nvk: Move is_storage_image_format to nvk_format.c

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

nil: Add a helper for filling out buffer TIC entries

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

nvk: Implement samplers

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

nvk: Re-format nvk_sampler.c

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

nvk: Teture pool sizes are maximums not sizes

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

nil: Fix image array layer alignments

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

nil: Refactor TIC image extent setup

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

nil: Don't minify image dimensions when setting up TIC

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

nil: Drop miptail support for now

We don't have any idea how they work or even which TIC bits have to do
with them.  Better off just leave them alone for now.

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

nouveau/classes: add 906f header support.

This needed some tweaks to the generator script

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

nvk: Drop vk_sync BO refs after push_submit

Otherwise, all syncs are cumulative and, if the same command buffer is
submitted multiple times, it will sync on the union of all syncs it's
ever seen.

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

nouveau/ws: add a push reset just for references.

This is to be used with the empty submission pushbuf.

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

nvk: fix empty cmd submission.

The sync obj stuff does a second queue submit with just signals
and no command buffers. Make sure things get sent to the kernel
so sync works.

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

nvk: Support compressed images in copy commands

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

nvk: Add a more competent GetPhysicalDeviceImageFormatProperties

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

nvk: Rework format features queries

Split the buffer and image portions into vk_buffer_view.c and vk_image.c
respectively.  This way they're a lot closer to each other.  Put the
main combined helper in nvk_format.c since it's a format thing.  Also,
restructure things a bit based on ANV to make the code look nicer.
Also, implement VK_KHR_format_feature_flags2 while we're here.

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

nvk: Reset and properly clean up command buffer upload areas

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

nvk: Implement VkGetImageSubresourceLyout

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

nvk: Re-indent image entrypoints

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

nvk: Convert to using NIL for image layout

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

nvk: Manually offset for array layers in copy/blit

The more I start at the hardware the more convinced I become that arrays
and 3D are not the same thing.  I'm sure they're similar enough in
certain circumstances but they don't appear to be in the way they're
treated in the 2D hardware, especially when miplevels are involved.
Instead of trying to use the 2D engine's depth/slice logic, offset
manually.  We have to emit a copy command for each slice anyway so
making the GPU do those calculations doesn't really gain us anything.

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

Revert "nvk: Stop returning VK_ERROR_FORMAT_NOT_SUPPORTED for non-blitable"

This reverts commit 74ad75f2624c50f732db35946a127eb11edee3e5.

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

nvk: Move Fill/UpdateBuffer to nvk_cmd_copy

Even if they technically use blit-like HW commands, they're much more a
copy operation than a blit since there's no scaling going on.

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

nvk: Align arguments consistently in copy/blit code

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

nvk: Use helpers for push_ref

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

nvk: Add an nvk_get_format helper

This makes the array size internal to nvk_format.c so we don't have to
keep it in sync when we add extra elements.  Also, there's no reason why
we need to be hand-rolling that loop all the time.

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

nil: Add a nil_view and code to fill out TIC entries

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

nil: Add the TIC format table from nouveau

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

nil: Create images

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

nouveau: Add stubs for an image layout library called NIL

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

nouveau/ws: initial debugging options for command submissions

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

nouveau/headers: Generate parser functions

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

nvk/blit: assert that formats are supported

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

nouveau/ws: assert on broken channel

no point in continuing

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

nouveau/ws: make sure we don't submit nonsense

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

nvk: Rework descriptor writes

In particular, this makes us only touch the descriptor UBO memory as a
memcpy destination unless we're doing a descriptor copy.  This should be
much better for mapped GPU memory than the read-modify-write required by
our previous implementation of image descriptor writes with their 12:20
configuration.

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

nvk: Reserve a null image descriptor

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

nvk: GART os host-cache-coherent

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

nvk: Make texture descriptors a bit more acceptable to codegen

Codegen assumes that it gets a combined descriptor on both the texture
and sampler handle sources and that it's 64-bit.  It's not too hard to
just give it that in NIR.  We may make codegen a bit more flexible in
future but, for now, this gets things working.  While we're here, do a
quick optimization for combined texture+sampler descriptors because it's
easy.

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

nvk: Add a stub implementation of buffer views

We don't actually populate the TIC entries yet but we can at least
pretend and stick them in descriptor sets.

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

nvk: Re-format nvk_buffer.c a bit

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

nvk: Re-format nvk_image_view.h a bit

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

nvk: Unify descriptor loading in lower_descriptors

The code we were using for images/textures was broken since it didn't
track with the change to always use root descriptors.

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

nvk: Set b->cursor when lowering image intrinsics

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

nvk: Fill out TIC table entries for image views

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

nvk: Add an image descriptor table to the device

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

nvk/image_view: Reformat and fix Create/DestroyImageView

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

nvk: Copy in the nouveau TIC format table

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

nvk: Add a descriptor table data structure

This will be used for image and sampler descriptors

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

nvk: Mark nvk_push_descriptor_set_ref() inline

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

nvk: Allow R32_UINT

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