Rather than keeping separate {vs,hs,ds,gs}_start fields, we now store an
array indexed by the shader stage (MESA_SHADER_*). The 3DSTATE_URB_*
commands are also sequentially numbered. This makes it easy to just
emit them in a loop.
This simplifies the code a little, and also will make it easier to add
more credible HS and DS code later.
Unfortunately, glslang gives us cull/clip distance and GS streams even if
the shader doesn't use it whenever a shader is declared as version 450.
This is a glslang bug, but we can easily enough ignore it for now.
The descriptor sizes array gives the total number of each type of
descriptor that will ever be allocated from the pool, not the total amount
that may be in any particular set. In our case, this simply means that we
have to sum a bunch of things up and there we go.
We can't use a global descriptor pool like we were because it's not
thread-safe. For now, we'll allocate them on-the-fly and that should work
fine. At some point in the future, we could do something where we
stack-allocate them or allocate them out of one of the state streams.
Descriptor pools are an optimization that lets applications allocate
descriptor sets through an externally synchronized object (that is,
unlocked). In our case it's also plugging a memory leak, since we
didn't track all allocated sets and failed to free them in
vkResetDescriptorPool() and vkDestroyDescriptorPool().
I've had people ask about the design of the pack functions, for example,
why aren't we using bitfields. I wrote up a bit of background on why and
how we ended up with the current design and we might as well keep that
with the code.
As with anv_CmdCopyBufferToImage, compressed textures require special
handling during copies.
Reviewed-by: Kristian Høgsberg Kristensen <kristian.h.kristensen@intel.com>
This gives us the chance to pack the binding table down to just what the
shaders actually need. Some applications use very large descriptor sets
and only ever use a handful of entries. Compacted binding tables should be
much more efficient in this case. It comes at the down-side of having to
re-emit binding tables every time we switch pipelines, but that's
considered an acceptable cost.
We ignore unused dimensions in the isl surface; do the same for the
resulting anv_image.
Reviewed-by: Kristian Høgsberg Kristensen <kristian.h.kristensen@intel.com>
"Result is the same as computing the sum of the absolute values of
OpDPdx and OpDPdy on P."
We were doing sum of absolute values of OpDPdx of P and OpDPdx of NULL.
