Dynamic color/blend state can be NULL in case we're not rendering to
color targets (only output to depth and/or stencil). Initialize
cmd_buffer->cb_state to NULL so we can reliably detect whether it's been
set or not.
This supports the three Vulkan border color types for float color
formats. The support for integer formats is a little trickier, as we
don't know the format of the texture at this time.
According to the bspec, you're supposed to emit a PIPE_CONTROL with a CS
stall and a render target flush prior to chainging STATE_BASE_ADDRESS. A
little experimentation, however, shows that this is not enough. It also
appears as if you have to flush the texture cache after chainging base
address or things won't propagate properly.
This commit allows for us to create a whole new surface state buffer when
the old one runs out of room. We simply re-emit the state base address for
the new state, re-emit binding tables, and keep going.
Before, we were emitting surface states up-front when binding tables were
updated. Now, we wait to emit the surface states until we emit the binding
table. This makes meta simpler and should make it easier to deal with
swapping out the surface state buffer.
We need to make sure we use the right index into dynamic offset
array. Dynamic descriptors can be present or not in different stages and
to get the right offset, we need to compute the index at
vkCreateDescriptorSetLayout time.
We do this by creating a surface state on the fly that incorporates the
dynamic offset. This patch also refactor the descriptor set layout
constructor a bit to be less clever with switch statement fall
through. Instead of duplicating the subtle code to update the sampler
and surface slot map, we just use two switch statements.
We rely on these being initialized to NULL so meta can reliably detect
whether or not they've been set. ds_state is also allowed to not be
present so we need a well-defined value for that.
This mega-commit primarily does two things. First, is to turn anv_batch
into a better abstraction of a batch. Instead of actually having a BO, it
now has a few pointers to some piece of memory that are used to add data to
the "batch". If it gets to the end, there is a function pointer that it
can call to attempt to grow the batch.
The second change is to start using chained batch buffers. When the end of
the current batch BO is reached, it automatically creates a new one and
ineserts an MI_BATCH_BUFFER_START command to chain to it. In this way, our
batch buffers are effectively infinite in length.
If a meta operation is called before the pipeline is set, this can cause
uses of undefined values. They *should* be harmless, but we might as well
shut up valgrind on this one too.
Previously, we just blasted out whatever VB's we had marked as "dirty"
regardless of which ones were used by the pipeline. Given that the stride
of the VB is embedded in the pipeline this can cause problems. One problem
is if the pipeline doesn't use the given VB binding we emit a bogus stride.
Another problem is that we weren't properly resetting the dirty bits when
the pipeline changed.
Previously, we waited until later and did a pass through the used surfaces
and did the relocations then. This lead to doing double-relocations which
was causing us to get bogus surface offsets.
Since the binding table pointer is only 16 bits, we can only have 64kb
of binding table state allocated at any given time. With a block size of
1kb, that amounts to just 64 command buffers, which is not enough.
The binding table pointers packet only allows for a 16-bit binding table
address so all binding tables have to be in the first 64 KB of the surface
state BO. We solve this by adding a slave block pool that pulls off the
first 64 KB worth of blocks and reserves them for binding tables.
