nvk: fix zero page refing

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

nouveau/parser: Put the dump helpers in C files

This should make the build faster since we don't have to compile those
giant functions with giant switch statements constantly.

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

nouveau/parser: Use more idiomatic python

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

nouveau/parser: Convert to mako

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

nvk: alloc a zero page and use it for vertex runouts

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

nvk: Initialize fixed draw/default state once

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

nvk: Start every command buffer with a nop

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

nvk: Split nvk_queue into its own file

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

nvk: fix mem leaks

device would leak in various situations.
Separate out lables were needed in the earlier
versions of this function but now we can just
use a single label

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

nvk: delete commented code

nvk_CmdClearColorImage first lived in nvk_cmd_copy,
then moved to nvk_cmd_meta, and finally in nvk_cmd_clear

I seems that this was left behind in one of the moves

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

nvk: fix implicit-fallthrough warnings with clang

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

nvk: use common physical device enumeration

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

nvk: fix OOB read inside nvk_get_va_format

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

nvk: Update to the new command buffer infrastructure

We leave nvk_command_pool because we'll want to use it for recycling
individual pushbufs.  Drop reset_on_submit becuse there was never really
that much point in that micro-optimization and it isn't thread-safe
anyway if we recycle things in the pool.

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

nouveau/ws: push chaining

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

nouveau/ws: rework refing push buffer bos

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

nouveau/ws: bound check nouveau_ws_push_append

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

nvk: Add some asserts for nv50 compiler image restrictions

These just ensure that we assert instead of doing something that the
nv50 compiler is likely to screw up.

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

nvk: Zero unused descriptors

When a descriptor gets freed because a buffer/image veiw or sampler is
deleted, zero out the descriptor memory.  This prevents the descriptor
from being used later and faulting on possibly non-existant memory.

While we're here, clean up the function a bit.

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

nvk: Implement descriptor table growing

Now that command buffers are no longer referencing descriptor table BOs
directly and we're not handing out maps, we can support growing.

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

nvk: Rework nvk_descriptor_table_add/remove

The new api returns a VkResult for more accurate error checking.  More
importantly, it also does copies the descriptor data into the table
instead of returning a map.  This means the users of the descriptor
table will never see a map and we can make sure that the copy happens
under the lock in case we have to grow.  Since each descriptor is only
32 bytes, this isn't going to be a perf problem.

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

nvk/sampler: Free descriptor table entries

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

nvk: Invalidate texture header and sampler caches each submit

It's possible descriptors have been added or removed.  We could track
it, but there's really no point.  Per-submit shouldn't be terrible.

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

nvk: Rework TLS/SLM and image/sampler table handling

When Dave originally added SLM, he did so with a preamble pushbuf which
sets up the SLM area prior to executing any actual command buffers.
This has the advantage that the SLM BO is never directly referenced in
any command buffer and we can swap it out even after command buffers
have been built.  When I added image/sampler tables, I sort-of went
along with it but then did something different for 3D.

Thanks to buffer reference counting, we can employ a fairly simple
locking scheme.  Each of the globals: images, samplers, and TSL has a
lock which gets taken whenever the table is modified.  When we go to do
a submit, we take each lock and take a queue-local reference to the BO
and grab a copy of its size parameter under the lock.  If anything has
changed, we update the queue-local preamble pushbuf.  Because the queue
also holds a reference, we can safely proceed to submit command buffers
which reference those global BOs without the individual global locks
taken.  If SLM or one of the descriptor tables resizes while we're
mid-submit, our reference to the BO will remain valid until either the
destruction of the queue or the next submit.

The one small bit of threaded cleverness (one always has to be careful
with that) is that the SLM size can be checked without taking the lock
because the SLM size only ever increases.

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

nvk: Drop most buffer tracking

If it's backed by an nvk_device_memory, we don't need to track it since
we always add all of those to every submit so that bindless works.

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

nvk/dispatch: Rename push to p

I took to using this convention in nvk_cmd_draw.c because it's shorter.
Let's use it other places too.

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

nvk/blit: Rename push to p

I took to using this convention in nvk_cmd_draw.c because it's shorter.
Let's use it other places too.

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

nvk/copy: Rename push to p

I took to using this convention in nvk_cmd_draw.c because it's shorter.
Let's use it other places too.

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

nvk: Inline nouveau_copy_linear

It has exactly one caller so the helper's not really doing us much.

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

nvk: clean up bo mappings

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

nouveau/ws: add a bo unmap helper function

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

nouveau/ws: add API to query if the context was killed

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

nvk: Fix dynamic buffer descriptor copies

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

nvk: Set up clip and cull distances

Closes: https://gitlab.freedesktop.org/nouveau/mesa/-/issues/12
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24326>

nil: Fix nonnormalized coordinates

Also, replace our TODO with a real comment explaining how nonnormalized
coordinate select works between image and texture header.

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

nil: Fix max mip level

The MAX_MIP_LEVEL field needs the actual number from the image because
it's used to compute the stride between layers.  The restrictions on mip
level come entirely from the RES_VIEW fields.

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

nil/format: Depth/stencil formats appear as red

Previously, it was treated as luminance and the depth or stencil
component was broadcast to all channels except alpha.  In Vulkan, depth
and stencil textures are treated as red.  If we want it treated
differntly by some other API, that driver can do that with a swizzle.

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

nvk: Fix border color alpha

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

nvk: VkSamplerCreateInfo::mipLodBias is signed

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

nvk: Assert on CmdExecuteCommands

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

nvk: Fix max anisotropy

Now that we're using the bitpack helpers, we don't need the extra shift

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

nvk: Don't advertise tiling on non-power-of-two formats

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

nvk: Use a UINT format to clear non-renderable images

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

nvk: Implement vkCmdClear*Image directly

Instead of invoking the meta code, implement it ourselves so we get the
HW clear path.  We could probably put these implementations somewhere
common since they only set up a few things and invoke Begin/EndRendering
but there's not that much point for now.

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

nvk: Add image_view_init/finish functions

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

nvk: Use hardware clears for attachment clears

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

nouveau/parser: Add array and float tags for clear values

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

vulkan/meta: fix gnu-empty-initializer build error

Fixes the following building error:

../src/vulkan/runtime/vk_meta_clear.c:213:42: error: use of GNU empty initializer extension [-Werror,-Wgnu-empty-initializer]
   struct vk_meta_clear_push_data push = { };
                                         ^
1 error generated.

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

nvk: Save/restore push constants around meta ops

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

nvk: Wire up early z and post depth coverage

Was crawling through header bits and found that we weren't plumbing
these through.

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

nvk: add basic primitive restart

fixes crucible
func.draw-index16-restart
func.draw-index32-restart

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

nvk: support GetImageSparseMemoryRequirements2

KHR_get_memory_requirements2 was enabled but GetImageMemoryRequirem
ents2 was not. This stubs out GetImageMemoryRequirements2 to fix it.

Signed-off-by: Yusuf Khan <yusisamerican@gmail.com>
v2: remove the v1 version of the command as its provided by the
common runtime

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

nvk: implement GetDeviceMemoryCommitment

vulkan spec says:

If the memory object is allocated from a heap with the
VK_MEMORY_PROPERTY_LAZILY_ALLOCATED_BIT bit set, that object’s
backing memory may be provided by the implementation lazily. The
actual committed size of the memory may initially be as small as
zero (or as large as the requested size), and monotonically
increases as additional memory is needed.

As far as I can tell we ignore the bit meaning that we allocate the
requested size so return that as the size.

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

nvk: Allocate descriptors for input attachments

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

nvk: Assert samples == 1

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

nvk: Re-flow a couple function prototypes

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

nvk: Align descriptor buffers to NVK_MIN_UBO_ALIGNMENT

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

nvk: Add nvk_queue_init/finish() helpers

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

nil: Add a PTE kind for Z32_FLOAT

This values was found by trial-and-error poking at the hardware until it
stopped complaining.  It's not clear if we need separate values for
compressed vs. not.  This appears to work on Turing regardless of what
we set for SET_Z_COMPRESSION.

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

nvk: Return OUT_OF_DEVICE_MEMORY if bo_new fails

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

nvk: Set GEOMETRY_SHADER_SELECTS_LAYER properly

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

nvk: Call nir_lower_input_attachments

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

nvk: Don't destroy NULL descriptor pool BOs

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

nvk: Advertise depth/stencil support

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

nvk: Flip the front-face setting

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

nvk: Enable two-sided stencil

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

nvk: Fix viewport Z scale

Because Z needs to end up in [0, 1] and not [-1, 1], we don't want to
divide by 2.

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

nvk: Fix depth/stencil render pass clears

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

nvk: Do internal dedicated allocations for ZS images

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

nvk: Add internal helpers for device memory allocation

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

nouveau: Add a function to allocate a tiled buffer

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

nil: Compute PTE kinds and tile modes for images

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

nvk: Disable sparse Z on Maxwell+

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

nvk: Depth buffers don't allow Z-tiling

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

vulkan/meta: Make stencil reference dynamic for clears

We call CmdSetStencilReference() but don't set it to dynamic in the
pipeline.

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

nvk: Clear dynamic state dirty after flushing it all

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

nvk: Advertise more format features

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

nil: Add more format support helpers

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

nil: Re-introduce the format capabilities

The bits are already in the table we copied from the nouveau GL drivers,
we just need to set up some enums and plumb it through.  With that,
we're now properly advertising which formats are renderable.

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

nil: Rename rendering to color_target

Also, add a helper for depth/stencil formats and use
nil_format_supports_color_targets() helper to compute Vulkan format
features rather than checking for a non-zero color format.

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

nil: Rename rt to czt in the format info struct

It's used for both depth/stencil and color so rt isn't quite the name we
want.  Also, add a comment.

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

nil: Use nvidia headers for ZS format enums

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

nil: Use the nvidia headers for render target format enums

Two notable changes here:

 1. BITMAP no longer exists
 2. B10G10R10A2_UINT was wrongly mapped to RGB10_A2_UINT before

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

nil: Use nvidia headers for texture format enums

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

nil: Properly depend on nouveau winsys and nvidia-headers

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

nouveau: Add class information to nouveau_ws_device

Annoyingly, with the current uAPI, this requires creating a dummy
context and throwing it away.  Fortunately, we should only do this once
on driver load so it's really not that big a problem.

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

nvk: Depth bounds need fui()

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

nouveau/parser: Dump more fields as float

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

nvk: Clean up try_create_physical_device error handling

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

nvk: Advertise vertex buffer format featues

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

nvk: Move the vertex format table into nvk_format.h

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

nvk: Re-initialize dynamic_graphics_state.vi when recycling

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

nvk/meta: Save and restore VI state

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

nvk: Drop sample locations structs

IDK if NVIDIA hardware even supports setting sample locations.  I've not
been able to find the packets.  In any case, if and when we implement
the extension, we can wire it up then.

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

nvk: Implement copies for D24_UNORM_S8_UINT images

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

nvk/copy: Mape bpp part of nouveau_copy_buffer

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

nvk: handle NULL to destroy descriptor pool

Fixes dEQP-VK.api.null_handle.destroy_descriptor_pool

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

nvk: lower io to temps to avoid output reads in vertex shaders

fixes a bunch of these
nv50_ir_lowering_nvc0.cpp:2878: void nv50_ir::NVC0LoweringPass::handleLDST(nv50_ir::Instruction*): Assertion `prog->getType() == Program::TYPE_TESSELLATION_CONTROL' failed.

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

nvk: assign vertex locations according to input attrib index

This copies what lavapipe does. is there a better plan?

Fixes
dEQP-VK.draw.renderpass.simple_draw.simple_draw_triangle_list

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

vulkan: Allow scissors or viewports to be set without counts

This is unlikely but can happen if you have the following sequence:

 1. vkCmdSetScissors()
 2. No pipeline bind
 3. vkCmdClearImage() which causes a meta save
 4. Meta restore with vk_cmd_set_dynamic_graphics_state()

In that case, we don't have scissor counts but need to restore the
scissors set with `vkCmdSetScissors()` before the meta save.  We can
safely copy all of them, it's just more memory traffic than maybe we'd
like.  Fortunately, this can only happen at the start of a command
buffer and only with a fairly silly sequence of commands.

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