This made for an unfortunately large amount of work since we were using it
fairly heavily internally. However, gl_shader_stage does basically the
same things, so it's not too bad.
The stuff to take descriptor sets and turn them into binding tables and
sampler tables is still in anv_cmd_buffer.c. We may want to consider
putting it in anv_descriptor_set.c eventually.
Previously, we were including gen7_pack.h, gen75_pack.h, and gen8_pack.h
in anv_private.h. As we add more gens, this is going to become untenable.
This commit moves things around so that we only use the pack headers when
and if we need them.
What we had before was kind of a hack where we made certain untrue
assumptions about the incoming data. This new support, while it still
doesn't support indirects properly (that will come), at least pulls the
offsets and strides from SPIR-V like it's supposed to.
Now that we've done the refactoring upstream, it's much easier to to get
hooked up. We haven't tested things well enough to know that we're setting
up the GPU state correctly for them yet but at least we can compile them now.
Now that we have a decent interface in upstream mesa, we can get rid of all
our hacks. As of this commit, we no longer use any fake GL state objects
and all of shader compilation is moved into anv_pipeline.c. This should
make way for actually implementing a shader cache one of these days.
As a nice side-benifit, this commit also gains us an extra 300 passing CTS
tests because we're actually filling out the texture swizzle information
for vertex shaders.
Prefix all anv_cmd_state dirty bit tokens with ANV_CMD_DIRTY. For
example:
old -> new
ANV_DYNAMIC_VIEWPORT_DIRTY -> ANV_CMD_DIRTY_DYNAMIC_VIEWPORT
ANV_CMD_BUFFER_PIPELINE_DIRTY -> ANV_CMD_DIRTY_PIPELINE
Change type of anv_cmd_state::dirty and ::compute_dirty from uint32_t to
the self-documenting type anv_cmd_dirty_mask_t.
The Vulkan spec allows VkGraphicsPipelineCreateInfo::pDepthStencilState
to be NULL when the pipeline's subpass contains no depthstencil
attachment (see spec quote below). anv_pipeline_init_dynamic_state()
required it unconditionally.
This path fixes anv_pipeline_init_dynamic_state() to access
pDepthStencilState only when there is a depthstencil attachment.
From the Vulkan spec (20 Oct 2015, git-aa308cb)
pDepthStencilState [...] may only be NULL if renderPass and subpass
specify a subpass that has no depth/stencil attachment.
The Vulkan spec (20 Oct 2015, git-aa308cb) states that some fields of
VkGraphicsPipelineCreateInfo are required under certain conditions.
Add a new function, anv_pipeline_validate_create_info() that asserts the
requirements hold.
The assertions helped me discover bugs in Crucible and anv_meta.c.
Aparently, we had the dynamic state array in the pipeline backwards.
Instead of enabling the bits in the pipeline, it disables them and marks
them as "dynamic".
This patch reworks a bunch of stuff in the way we do descriptor set
layouts. Our previous approach had a couple of problems. First, it was
based on a misunderstanding of arrays in descriptor sets. Second, it
didn't properly handle descriptor sets where some bindings were missing
stages. The new apporach should be correct and also makes some operations,
particularly those on the hot-path, a bit easier.
We use the descriptor set layout for four things:
1) To determine the map from bindings to the actual flattened descriptor
set in vkUpdateDescriptorSets().
2) To determine the descriptor <-> binding table entry mapping to use in
anv_cmd_buffer_flush_descriptor_sets().
3) To determine the mappings of dynamic indices.
4) To determine the (set, binding, array index) -> binding table entry
mapping inside of shaders.
The new approach is directly taylored towards these operations.
gen8_graphics_pipeline_create had a bunch of stuff in it that's already set
up by anv_pipeline_init. The duplication was causing double-initialization
of a state stream and made valgrind very angry.
With all the previous commits in place, we can now drop in support for
multiple platforms. First up is gen7 (Ivybridge).
Signed-off-by: Kristian Høgsberg Kristensen <kristian.h.kristensen@intel.com>
Previously anv_batch.relocs was an actual relocation list. However, this
is limiting if the implementation of the batch wants to change the
relocation list as the batch progresses.
