Otherwise we run into problems by putting this optimization loop
before I/O lowering, where there might still be 8-bit values that
haven't been lowered to 16 or 32. Once that's done, any remaining
movs or vec ops will have higher bit sizes already.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23173>
Add support for 16-bit UBO loads, delete handling of byte-addressed
UBO loads (which I think was never used anyway) and add handling
for the component const index to optimize out unneeded extractResults.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23173>
Derefs have index-based access semantics, which means we don't need
custom intrinsics to encode an index instead of a byte offset.
Remove the "masked" store intrinsics and just emit the pair of atomics
directly. This massively reduces duplication between scratch, shared,
and constant, while also moving more things into nir so more optimizations
can be done.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23173>
We still use shader_temp as a temporary variable mode to differentiate
which variables have simple deref patterns vs ones that need to be
lowered to ssbo, but then we put it back to mem_constant when we're
done to restore sanity.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23173>
There's a few changes in here that are very inter-related.
First, we stop lowering load_deref on shader_temp to load_ptr_dxil,
and just leave it as load_deref. In order for that to work, we need
the derefs to be in a shape that's acceptable to DXIL, so the only
current producer of shader_temp loads (the CLC frontend) needs to
run some lowering passes on them first.
The DXIL backend is augmented to just write out deref indices while
walking a deref chain, which will get combined in the load op into
a GEP instruction. For non-mesh/raytracing shaders, these are required
to be single-level scalar arrays, but the complexity here is preparation
for when we don't need to do that anymore.
Additionally, the const lookups are changed from using a hash table
to just putting an index on the variable.
All of this together is enough to enable the authored-forever-ago test
which uses indirect array access into a const packed struct. The
load_ptr_dxil handling didn't deal with packed structs / unaligned
accesses, but now that we're in a logical address space with derefs
instead of physical, there's no alignment to deal with anymore and
the fact that it's packed goes out the window.
This removes one custom DXIL intrinsic.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23173>
Instead of treating all 16-bit values as "native 16-bit types,"
differentiate between concrete casts and mediump casts, where the
former requires native 16-bit types, and the latter only requires
DXIL min-precision. Additionally, UBO/SSBO loads/stores require
native 16-bit types.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23344>
Fixes some tests when bindless is disabled, where the image format is
R32, we do atomics on it, but we didn't set the "typed UAV load with
additional formats" feature bit because when we loaded from it, we
only loaded one component. Since the image format on the DXIL side
was declared as U32x4, the DXIL validator said that we should have.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23266>
Using nir_gather_ssa_types works much better. There's 2 differences
compared to what I was doing before:
1. Multiple passes to allow data to propagate forward and backward
through the whole shader.
2. Allowing a value to have indeterminate types due to having both
int and float usages.
So this deletes some code and gets better results. Wish I'd known
this existed last week.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23062>
For ALU ops where input types are known, we can store that info on
the input sources. This can be used to produce the correct overloads
of load instructions that don't immediately need to be followed by
bitcasts, or similarly to produce a constant value which can be directly
consumed by the relevant instruction without needing a bitcast.
Similarly for values that will be stored in an output, we know type
information. And using that info, we can use more-correct information
for phis instead of forcing all phi sources to be bitcast to int just
to be bitcast back to float on the other side for an alu or an output
store.
One missing piece is SSBO stores, where we can support int or float.
If the input is coming from a phi, we don't influence the phi's type,
so it'll be int, even though the incoming sources might've been float.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22972>
For each phi src, ensure that it's only used as a phi src.
This lets us give each phi their own unique types without worrying
about them stomping on each other. Also scalarize phis.
For each constant, ensure that it's only used once. The DXIL backend
will already dedupe these consts within the module, but this lets a
single load_const have multiple types depending on how it's used.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22972>
The base nir options were assuming all bit sizes were supported at
shader model 6.2. Multiple callers were then changing properties
based on actual support.
Standardize behavior by providing the majority of things that can
impact nir options when getting them. Some callers (e.g. meta blit
shaders or libclc) don't bother, because they are known to have
contents that are unaffected by these options. Other callers might
munge more properties afterwards, but this minimizes that.
Note that lower_helper_invocation was incorrectly being turned off
for SM6.6+ by some callers, despite load_helper_invocation being
unimplemented by the backend.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22952>
Unlike DXBC, DXIL's shift instructions don't have the implicit behavior
that they only take the 5 bits. This is observable if you try to have
DXC do a shift of a dynamic value, e.g. a constant buffer value, where
the compiler inserts the appropriate 'and' op. We need to do the same.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22225>
Mike Blumenkrantz