v3dv: implement device detection on actual hardware

When running on the real hardware we have two devices: the v3d render
node and the vc4 display node. We need the latter to allocate
winsys BOs for v3d to render into. Since exporting these BOs is
a privileged operation, we need to obtain the fd for this device
through the display server. For now we only support doing this through
the XCB DRI3 platform.

Also, do not duplicate or re-open the DRM devices when creating logical
devices. The simulator checks that the file descriptor is exactly
the same we used to initialize it when we created the physical device
and aborts if it sees a different fd number, even if it points to the
same device.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/pipeline: adding some nir-based linking

Among other things, gets a constants output from a vs, used as input
to a fs, to get lowered and moved as a load const on the fs.

Heavily based on st_glsl_to_nir, already used by the v3d
driver. Slightly adapted to our needs, but there are still room for
customization.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3d/cmd_buffer: emit flat_shade/noperspective/centroid flags

We just emit the default ZERO_ALL values, to cover at least the cases
where that is enough. Need to emit non-default values if needed.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: stubs for Create/DestroyPipelineCache

Good enough to avoid vkrunner crashing.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: hook up WSI support

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement DRM modifier setup for WSI

This is only really relevant when running on real hardware, since
when we run on the simulator we don't care about the format of the
swapchain images.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkGetImageSubresourceLayout

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: fix copy image to buffer

If either of bufferRowLength or bufferImageHeight are zero, then that
aspect alone of the image is tighly packed. We were assuming that if
either was zero both aspects were tightly packed, which is wrong.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement VK_KHR_external_memory{_fd,_dma_buf}

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement VK_KHR_external_memory_capabilities

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/format: add R32G32B32A32_SFLOAT format

This is clearly supported, and we already confirmed that work with
some of our tests.

Needed to avoid an assertion of such tests when validation layers are
enabled.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement VK_KHR_get_physical_device_properties2

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: include Vulkan version 1.1 as unsupported.

This is so we can generate entry points for extensions that have been
promoted to core in 1.1.

Entry points for promoted extensions are aliased without the KHR suffix
in the Vulkan API XML, and the entry point generation scripts are designed
to point the dispatch tables to entry points generated from the non-aliased
function names, however, these are not included in the header file without
this change.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkGetPhysicalDeviceSparseImageFormatProperties

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: provide default values for input attributes

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: vertex input support

This includes:

   * Implementation for CmdBindVertexBuffers

   * Gather vertex input info during CreateGraphicsPipelines
     (pipeline_init) and SHADER_STATE_ATTRIBUTE_RECORD prepacking

   * Final emission of such packet during CmdDraw
     (cmd_buffer_emit_graphics_pipeline)

Default attributes values will be handled on a following patch.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/pipeline: lower fs/vs inputs/outputs

For now mostly call to nir_assign_io_var_locations

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: support queue submissions with multiple command buffers

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement fences

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement semaphore waits and signals on queue submissions

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vk{Create,Destroy}Semaphore

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkDeviceWaitIdle

This is really a hack, but it is better than not having anything. In the
future we should have a syncobject that we could wait on.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: prepack VCM_CACHE_SIZE

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: partial prepack of the gl_shader_state_record

We can't prepack all the record, as addresses need the job, and
uniforms depend on dynamic value.

Also due cl_emit_with_prepacked and v3dv_pack asserting correct
values, we need to define two values twice, that lead to move
vpm_config to the pipeline. In any case, the latter will be useful
when we start to prepack more stuff.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/cmd_buffer: emit CFG_BITS

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: handle VK_ATTACHMENT_UNUSED properly in more places

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/cmd_buffer: rename render pass RCL emission helpers to be more explicit

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: create a helper to start a new frame

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: rewrite the attachment needs clearing condition

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: fix tile buffer loading

So we do what we actually state in the comment. Particularly, the load
operation only affects the first subpass that uses the attachment,
after that we always want to load, but we were only doing that for
attachments marked as CLEAR.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: merge subpasses into the same job when possible

Generally, we can do this when they render to the same collection of
attachments and we only need to emit a single RCL for them.

To implement this, we need to track the first subpass that is included
in the job and rewrite our loads and stores in the RCL to refer to that
subpass instead of the current subpass (which would be the last included
in the RCL).

When we merge jobs we also reuse the tile state/alloc BOs and we only
emit the binning setup once.

The environment variable V3DV_NO_MERGE_JOBS can be set to disable
job merging and have each subpass be in a separate job. This can be
useful for debugging issues spawning from incorrect subpass merges.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: only clear attachments on the first subpass that uses them

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: rewrite attachment state tracking

Keep track of attachment state for each attachment in the renderpass
instead of tracking it only for the attachments in the current subpass.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: add a note on interactions between clearing and scissor

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: fix clipping against render area

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: do not automatically emit a binner flush when finishing jobs

Binner flushes should be emitted naturally at the end of each draw,
if we are finishing a job and it doesn't have the binner flush, it
probably means that we have bogus emission code somewhere.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkCmdPipelineBarrier

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: use the correct miplevel slice for the tile load operation

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkCmdNextSubpass

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3d/compiler: update uses_vid/uses_iid check

In order to take into account the vulkan specific system values
SYSTEM_VALUE_INSTANCE_INDEX and SYSTEM_VALUE_VERTEX_ID_ZERO_BASE.

Reviewed-by: Iago Toral <itoral@igalia.com>
Reviewed-by: Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/cmd_buffer: start jobs with CmdBeginRenderPass

We want to create jobs when we start a new render pass so that
commands in the first subpass have a job.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: add the concept of a job

As we make progress towards more complex submissions we will need to split
our command buffers into smaller executable units (jobs) that we can
submit indepdently to the kernel. This will be required to implement
pipeline barriers, split subpasses that have depedencies on previous
subpasses, split render passes that use more than 4 render targets, etc.

For now we keep things simple and we only keep one job as current
recording target in the command buffer, and we generate a new one
with every subpass or with any commands we see outside of a render pass
(only vkCopyImageToBuffer for now). In the future we probably want to
optimize this by merging subpasses into the same job when possible,
etc.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkCmdCopyImageToBuffer

For now we only support the TLB path, which limits us to copying
regions that start at offset (0,0). In the future, we will need to add
a fallback path that uses blitting to copy regions with an offset.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: rename and make compute_tile_size_for_framebuffer() public

We will use this when we implement copying images to buffers using the
TLB, where we'll need to setup a framebuffer and tiling configuration
for the TLB store to the destination buffer.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: split framebuffer internal bpp calculations from tiling calculations

We want to reuse the latter aspect in a context were we already have
the internal bpp available.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: add v3dv_write_uniforms

Initial port of the equivalent v3d_write_uniforms, to be used by the
cmd_buffer when emitting the drawing packets.

Initially doesn't include all the quniform types, only those needed by
the initial basic vulkan tests.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/cmd_buffer: cache viewport translate/scale

Before that commit we were calling get_viewport_xform to get those
values twice (to emit scissor and viewport), and we found that we
would need that info even more times. So let's just compute that info
when setting the viewport, and reuse the values.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: clif format dumping support

Add support for V3D_DEBUG=clif. Useful to compare clif_dumps from
vulkan small-programs and the equivalent opengl ones.

As we are here we expand clif_dump_packet wrapper to use
v3d42_clif_dump_packet if needed, as the vulkan driver would use that
packet version.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/cmd_buffer: add shader source bos to cmd_buffer

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/cmd_buffer: start to emit draw packets

Starting using only the packet VERTEX_ARRAY_PRIMS

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/cmd_buffer: emit shader_state packets

Values still doesn't take into account having vertex elements data,
but keeps some of that half-done code in comments. It would be better
to do that when we get an example using it.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: flush at the end of each subpass

We can't start a new BCL if we haven't flushed the current one, so
make sure we do that at the end of each subpass for now.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: emit tile loads

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/cmd_buffer: emit Viewport packets

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/cmd_buffer: emit Scissor packets

As this is done as part of CmdDraw, we also add the initial stub of
such function here

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/cmd_buffer: init command buffer dynamic state during pipeline bind

So copying it from the pipeline to the cmd_buffer if defined.

Starting with Viewport/Scissor dynamic state.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/pipeline: start to track dynamic state

Starting with Viewport/Scissor data from VkGraphicsPipelineCreateInfo.

Note that initially this can be somewhat counter-intuitive. What we
are really doing it is filling up the structs with the dynamic stuff
from the pipeline, when such is not defined as dynamic. This is what
anv/radv does, and basically means that we treat both in the same way,
so easier after that.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: CmdSetViewport and CmdSetScissor implementation

This commit also introduces adding dynamic state definitions, dirty
flags, and setting them on such methods, although this commit still
doesn't use all that info yet.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: initial stub for CmdBindPipeline

Right now it only sets the pipeline to the command buffer state.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: initial CreateGraphicsPipeline/DestroyPipeline implementation

The basic to get the spirv built to nir, including calling some common
nir passes. Pending deep review if all those are needed or if we miss
some, but for that it would be better to be able to run existing
tests.

Enough to get assembly generated for simple tests.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv/format: add v3dv_get_format_swizzle

Heavily based on the equivalent v3d v3d_get_formar_swizzle.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3d/compiler: num_tex_used on v3d_key

We would need on OpenGL to update values for all the textures used. On
OpenGL that value can be always took from the context or the nir
shader, but there are cases on Vulkan that it is not the case, or
would force up to recompute it.

Acked-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: setup color clear values at subpass setup time

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: be more conservative resetting command buffer state

We don't want to reset everything, but setting some pointer fields
to NULL should help debug issues better.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkQueueSubmit

For now this only supports a single command buffer submit without any kind
of sempaphores or fences involved.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: clear set of BOs in the command buffer on reset

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: keep track of the numbre of BOs in a command buffer

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: create the command buffer BO set before we init CLs

The init of the CLs doesn't create any BOs for now, but this looks
like a sensible decision to future-proof things.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkEndCommandBuffer

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: emit the render command list

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: emit scissor to render area and precompute hw color clear values

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: precompute more tiling info at framebuffer creation time

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: plug leak when destroying device

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: make v3dv_layer_offset public

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkDestroyImage and vkDestroyImageView

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: revert the decision that the command buffer takes ownership of BOs

The CLs in the command buffer will reference BOs allocated by the application
such as images and buffers, that will be destroyed by the application, so
destroying them with the command buffer won't be correct.

For now, let's just assume that the comman buffer only owns the BOs
it explicitly allocates and that other abstractions own their own
BOs and are responsible from freeing them.

In the future, we might consider a ref/unref system similar to v3d's, but
for now this should work.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: add the tile state and alloc BOs to the command buffer BO list

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: store base mip level in the image view

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: add a few more API stubs

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: make the command buffer own the command list BOs

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkCmdBeginRenderPass

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: compute tile size for framebuffer

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: make v3dv_bo_alloc allocate memory for the bo struct

This is just going to be more convenient going forward.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkGetPhysicalDeviceImageFormatProperties

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: start handling command buffer status

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkBeginCommandBuffer

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: add a concept of a command list

Just barebones for now, will expand as necessary as we start emitting
actual commands into command lists.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: create a v3dv_bo struct and reference it from v3dv_device_memory

So we have a lower level representation of a buffer object that we can
manipulate that is not tied to a Vulkan representation of memory. This
will be useful as we start allocating driver internal buffers, such as
command lists.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vk{Allocate,Free}CommandBuffers

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkCreateCommandPool

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vk{Create,Destroy}Framebuffer

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vk{Create,Destroy}RenderPass

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkCreateRenderPass

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkBindBufferMemory

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkGetBufferMemoryRequirements

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vk{Create,Destroy}Buffer

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkCreateImageView

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkBindImageMemory

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkUnmapMemory

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: implement vkMapMemory

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>