New in MSL 2.3 is a template that can be used in the place of a scalar
type in a stage-in struct. This template has methods which interpolate
the varying at the given points. Curiously, you can't set interpolation
attributes on such a varying; perspective-correctness is encoded in the
type, while interpolation must be done using one of the methods. This
makes using this somewhat awkward from SPIRV-Cross, requiring us to jump
through a bunch of hoops to make this all work.
Using varyings from functions in particular is a pain point, requiring
us to pass the stage-in struct itself around. An alternative is to pass
references to the interpolants; except this will fall over badly with
composite types, which naturally must be flattened. As with
tessellation, dynamic indexing isn't supported with pull-model
interpolation. This is because of the need to reference the original
struct member in order to call one of the pull-model interpolation
methods on it. Also, this is done at the variable level; this means that
if one varying in a struct is used with the pull-model functions, then
the entire struct is emitted as pull-model interpolants.
For some reason, this was not documented in the MSL spec, though there
is a property on `MTLDevice`, `supportsPullModelInterpolation`,
indicating support for this, which *is* documented. This does not appear
to be implemented yet for AMD: it returns `NO` from
`supportsPullModelInterpolation`, and pipelines with shaders using the
templates fail to compile. It *is* implemeted for Intel. It's probably
also implemented for Apple GPUs: on Apple Silicon, OpenGL calls down to
Metal, and it wouldn't be possible to use the interpolation functions
without this implemented in Metal.
Based on my testing, where SPIR-V and GLSL have the offset relative to
the pixel center, in Metal it appears to be relative to the pixel's
upper-left corner, as in HLSL. Therefore, I've added an offset 0.4375,
i.e. one half minus one sixteenth, to all arguments to
`interpolate_at_offset()`.
This also fixes a long-standing bug: if a pull-model interpolation
function is used on a varying, make sure that varying is declared. We
were already doing this only for the AMD pull-model function,
`interpolateAtVertexAMD()`; for reasons which are completely beyond me,
we weren't doing this for the base interpolation functions. I also note
that there are no tests for the interpolation functions for GLSL or
HLSL.
In some cases, we need to get a literal value from a spec constant op.
Mostly relevant when emitting buffers, so implement a 32-bit integer
scalar subset of the evaluator. Can be extended as needed to support
evaluating any specialization constant operation.
When inside a loop, treat any read of outer expressions to happen
multiple times, forcing a temporary of said outer expressions.
This avoids the problem where we can end up relying on loop-invariant code motion to happen in the
compiler when converting optimized shaders.
When loading and storing array types which belong to buffer objects, we
need to treat these values as not being value types. Also, need to
handle array load/store from/to more address space combinations.
We had output dependent on complex_continue being set, but setting that
flag was dependent on unordered_set declaration order. Make it invariant
to ordering and change the implementation so it knows about the new
temporary hoisting for access chains.
Some fallout where internal functions are using stronger types.
Overkill to move everything over to strong types right now, but perhaps
move over to it slowly over time.
This was straightforward to implement in GLSL. The
`ShadingRateInterlockOrderedEXT` and `ShadingRateInterlockUnorderedEXT`
modes aren't implemented yet, because we don't support
`SPV_NV_shading_rate` or `SPV_EXT_fragment_invocation_density` yet.
HLSL and MSL were more interesting. They don't support this directly,
but they do support marking resources as "rasterizer ordered," which
does roughly the same thing. So this implementation scans all accesses
inside the critical section and marks all storage resources found
therein as rasterizer ordered. They also don't support the fine-grained
controls on pixel- vs. sample-level interlock and disabling ordering
guarantees that GLSL and SPIR-V do, but that's OK. "Unordered" here
merely means the order is undefined; that it just so happens to be the
same as rasterizer order is immaterial. As for pixel- vs. sample-level
interlock, Vulkan explicitly states:
> With sample shading enabled, [the `PixelInterlockOrderedEXT` and
> `PixelInterlockUnorderedEXT`] execution modes are treated like
> `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT`
> respectively.
and:
> If [the `SampleInterlockOrderedEXT` or `SampleInterlockUnorderedEXT`]
> execution modes are used in single-sample mode they are treated like
> `PixelInterlockOrderedEXT` or `PixelInterlockUnorderedEXT`
> respectively.
So this will DTRT for MoltenVK and gfx-rs, at least.
MSL additionally supports multiple raster order groups; resources that
are not accessed together can be placed in different ROGs to allow them
to be synchronized separately. A more sophisticated analysis might be
able to place resources optimally, but that's outside the scope of this
change. For now, we assign all resources to group 0, which should do for
our purposes.
`glslang` doesn't support the `RasterizerOrdered` UAVs this
implementation produces for HLSL, so the test case needs `fxc.exe`.
It also insists on GLSL 4.50 for `GL_ARB_fragment_shader_interlock`,
even though the spec says it needs either 4.20 or
`GL_ARB_shader_image_load_store`; and it doesn't support the
`GL_NV_fragment_shader_interlock` extension at all. So I haven't been
able to test those code paths.
Fixes#1002.
This change introduces functions and in one case, a class, to support
the `VK_KHR_sampler_ycbcr_conversion` extension. Except in the case of
GBGR8 and BGRG8 formats, for which Metal natively supports implicit
chroma reconstruction, we're on our own here. We have to do everything
ourselves. Much of the complexity comes from the need to support
multiple planes, which must now be passed to functions that use the
corresponding combined image-samplers. The rest is from the actual
Y'CbCr conversion itself, which requires additional post-processing of
the sample retrieved from the image.
Passing sampled images to a function was a particular problem. To
support this, I've added a new class which is emitted to MSL shaders
that pass sampled images with Y'CbCr conversions attached around. It
can handle sampled images with or without Y'CbCr conversion. This is an
awful abomination that should not exist, but I'm worried that there's
some shader out there which does this. This support requires Metal 2.0
to work properly, because it uses default-constructed texture objects,
which were only added in MSL 2. I'm not even going to get into arrays of
combined image-samplers--that's a whole other can of worms. They are
deliberately unsupported in this change.
I've taken the liberty of refactoring the support for texture swizzling
while I'm at it. It's now treated as a post-processing step similar to
Y'CbCr conversion. I'd like to think this is cleaner than having
everything in `to_function_name()`/`to_function_args()`. It still looks
really hairy, though. I did, however, get rid of the explicit type
arguments to `spvGatherSwizzle()`/`spvGatherCompareSwizzle()`.
Update the C API. In addition to supporting this new functionality, add
some compiler options that I added in previous changes, but for which I
neglected to update the C API.
When merging combined image samplers, we only looked at sampler, but DXC
emits RelaxedPrecision only for texture. Does not hurt to check for more
things.
This is quite complex since we cannot flush Phi inside the case labels,
we have to do it outside by emitting a lot of manual branches ourselves.
This should be extremely rare, but we need to handle this case.
We used to use the Binding decoration for this, but this method is
hopelessly broken. If no explicit MSL resource remapping exists, we
remap automatically in a manner which should always "just work".
This is rather shaky, but we don't have many choices here except add a
lot of awkward and unintuitive options. Try to deduce this from OpSource
and fallback to heuristic.
There is a risk that we try to preserve a loop variable through multiple
iterations, even though the dominating block is inside a loop.
Fix this by analyzing if a block starts off by writing to a variable. In
that case, there cannot be any preservation going on. If we don't, pretend the
loop header is reading the variable, which moves the variable to an
appropriate scope.
MSL generally emits the aliases, which means we cannot always place the
master type first, unlike GLSL and HLSL. The logic fix is just to
reorder after we have tagged types with packing information, rather than
doing it in the parser fixup.
Buffer objects can contain arbitrary pointers to blocks.
We can also implement ConvertPtrToU and ConvertUToPtr.
The latter can cast a uint64_t to any type as it pleases,
so we will need to generate fake buffer reference blocks to be able to
cast the type.
If we generate an access chain in a loop body, and it is consumed in the
loop continue block, we have a problem because we cannot emit a
temporary here holding the access chain reference. Force a complex loop
body to workaround this exceptionally rare case.
- Replace ostringstream with custom implementation.
~30% performance uplift on vector-shuffle-oom test.
Allocations are measurably reduced in Valgrind.
- Replace std::vector with SmallVector.
Classic malloc optimization, small vectors are backed by inline data.
~ 7-8% gain on vector-shuffle-oom on GCC 8 on Linux.
- Use an object pool for IVariant type.
We generally allocate a lot of SPIR* objects. We can amortize these
allocations neatly by pooling them.
- ~15% overall uplift on ./test_shaders.py --iterations 10000 shaders/.
We had a bug where error conditions in DoWhileLoop emit path would not
detect that statements were being emitted due to the masking behavior
which happens when force_recompile is true. Fix this.
Also, refactor force_recompile into member functions so we can properly
break on any situation where this is set, without having to rely on
watchpoints in debuggers.
This is a pragmatic trick to avoid symbol collision where a project
links against SPIRV-Cross statically, while linking to other projects
which also use SPIRV-Cross statically. We can end up with very awkward
symbol collisions which can resolve themselves silently because
SPIRV-Cross is pulled in as necessary. To fix this, we must use
different symbols and embed two copies of SPIRV-Cross in this scenario,
now with different namespaces, which in turn leads to different symbols.
We can trivially deal with cases where the loop tests are simply
inverted. We can also deal with cases where the condition block branches
to the merge block via other noop blocks.
This makes SPIR-V codegen easier when targeting SPIRV-Cross.
This adds a new C API for SPIRV-Cross which is intended to be stable,
both API and ABI wise.
The C++ API has been refactored a bit to make the C wrapper easier and
cleaner to write. Especially the vertex attribute / resource interfaces
for MSL has been rewritten to avoid taking mutable pointers into the
interface. This would be very annoying to wrap and it didn't fit well
with the rest of the C++ API to begin with. While doing this, I went
ahead and removed all the old deprecated interfaces.
The CMake build system has also seen an overhaul.
It is now possible to build static/shared/CLI separately with -D
options.
The shared library only exposes the C API, as it is the only ABI-stable
API. pkg-configs as well as CMake modules are exported and installed for
the shared library configuration.
We were using std::locale::global() to force a C locale which is not
safe when SPIRV-Cross is used in a multi-threaded environment.
To fix this, we could tap into various per-platform specific locale
handling to get safe thread-local locales, but since locales only affect
the decimal point in floats, we simply query the locale instead and do
the necessary radix replacement ourselves, without touching the locale.
This should be much safer and cleaner than the alternative.
In the bizarre case where the ID of a loaded opaque type aliased with a
literal which was used as part of another texturing instruction, we
could end up with a case where domination analysis assumed the loaded
opaque type needed to be moved to a different scope.
Fix the issue by never doing dominance analysis for opaque temporaries,
and be more robust when analyzing texturing instructions.
Also make sure reflection output is deterministic.
This patch slightly alterered output for some unknown reason, but it came from an
unordered_map, so it's fine.
These are mapped to Metal's post-tessellation vertex functions. The
semantic difference is much less here, so this change should be simpler
than the previous one. There are still some hairy parts, though.
In MSL, the array of control point data is represented by a special
type, `patch_control_point<T>`, where `T` is a valid stage-input type.
This object must be embedded inside the patch-level stage input. For
this reason, I've added a new type to the type system to represent this.
On Mac, the number of input control points to the function must be
specified in the `patch()` attribute. This is optional on iOS.
SPIRV-Cross takes this from the `OutputVertices` execution mode; the
intent is that if it's not set in the shader itself, MoltenVK will set
it from the tessellation control shader. If you're translating these
offline, you'll have to update the control point count manually, since
this number must match the number that is passed to the
`drawPatches:...` family of methods.
Fixes#120.
These are transpiled to kernel functions that write the output of the
shader to three buffers: one for per-vertex varyings, one for per-patch
varyings, and one for the tessellation levels. This structure is
mandated by the way Metal works, where the tessellation factors are
supplied to the draw method in their own buffer, while the per-patch and
per-vertex varyings are supplied as though they were vertex attributes;
since they have different step rates, they must be in separate buffers.
The kernel is expected to be run in a workgroup whose size is the
greater of the number of input or output control points. It uses Metal's
support for vertex-style stage input to a compute shader to get the
input values; therefore, at least one instance must run per input point.
Meanwhile, Vulkan mandates that it run at least once per output point.
Overrunning the output array is a concern, but any values written should
either be discarded or overwritten by subsequent patches. I'm probably
going to put some slop space in the buffer when I integrate this into
MoltenVK to be on the safe side.
A flat array was consuming way too much memory and was far too slow to
initialize properly with a very large ID bound (8 million IDs, showed up as #1 hotspot in perf).
Meta struct does not have to be in-order as we never iterate over it in
a meaningful way, so using a hashmap here is reasonable. Very few IDs
should need decorations or meta-data, so this should also be a quite
decent memory save.
For the pathological case, a 6x uplift was observed.
This is a fairly fundamental change on how IDs are handled.
It serves many purposes:
- Improve performance. We only need to iterate over IDs which are
relevant at any one time.
- Makes sure we iterate through IDs in SPIR-V module declaration order
rather than ID space. IDs don't have to be monotonically increasing,
which was an assumption SPIRV-Cross used to have. It has apparently
never been a problem until now.
- Support LUTs of structs. We do this by interleaving declaration of
constants and struct types in SPIR-V module order.
To support this, the ParsedIR interface needed to change slightly.
Before setting any ID with variant_set<T> we let ParsedIR know
that an ID with a specific type has been added. The surface for change
should be minimal.
ParsedIR will maintain a per-type list of IDs which the cross-compiler
will need to consider for later.
Instead of looping over ir.ids[] (which can be extremely large), we loop
over types now, using:
ir.for_each_typed_id<SPIRVariable>([&](uint32_t id, SPIRVariable &var) {
handle_variable(var);
});
Now we make sure that we're never looking at irrelevant types.
This allows shaders to declare and use pointer-type variables. Pointers
may be loaded and stored, be the result of an `OpSelect`, be passed to
and returned from functions, and even be passed as inputs to the `OpPhi`
instruction. All types of pointers may be used as variable pointers.
Variable pointers to storage buffers and workgroup memory may even be
loaded from and stored to, as though they were ordinary variables. In
addition, this enables using an interior pointer to an array as though
it were an array pointer itself using the `OpPtrAccessChain`
instruction.
This is a rather large and involved change, mostly because this is
somewhat complicated with a lot of moving parts. It's a wonder
SPIRV-Cross's output is largely unchanged. Indeed, many of these changes
are to accomplish exactly that! Perhaps the largest source of changes
was the violation of the assumption that, when emitting types, the
pointer type didn't matter.
One of the test cases added by the change doesn't optimize very well;
the output of `spirv-opt` here is invalid SPIR-V. I need to file a bug
with SPIRV-Tools about this.
I wanted to test that variable pointers to images worked too, but I
couldn't figure out how to propagate the access qualifier properly--in
MSL, it's part of the type, so getting this right is important. I've
punted on that for now.
A block name cannot alias with any name in its own scope,
and it cannot alias with any other "global" name.
To solve this, we need to complicate the name cache updates a little bit
where we have a "primary" namespace and "secondary" namespace.
This is required to avoid relying on complex sub-expression elimination
in compilers, and generates cleaner code.
The problem case is if a complex expression is used in an access chain,
like:
Composite comp = buffer[texture(...)];
vec4 a = comp.a + comp.b + comp.c;
Before, we did not have common subexpression tracking for
OpLoad/OpAccessChain, so we easily ended up with code like:
vec4 a = buffer[texture(...)].a + buffer[texture(...)].b + buffer[texture(...)].c;
A good compiler will optimize this, but we should not rely on it, and
forcing texture(...) to a temporary also looks better.
The solution is to add a vector "implied_expression_reads", which works
similarly to expression_dependencies. We also need an extra mechanism in
to_expression which lets us skip expression read checking and do it
later. E.g. for expr -> access chain -> load, we should only trigger
a read of expr when using the loaded expression.
This is a large refactor which splits out the SPIR-V parser from
Compiler and moves it into its more appropriately named Parser module.
The Parser is responsible for building a ParsedIR structure which is
then consumed by one or more compilers.
Compiler can take a ParsedIR by value or move reference. This should
allow for optimal case for both multiple compilations and single
compilation scenarios.
Implement this by flattening outputs and unflattening inputs explicitly.
This allows us to pass down a single struct instead of dealing with the
insanity that would be passing down each flattened member separately.
Remove stage_uniforms_var_id.
Seems to be dead code. Naked uniforms do not exist in SPIR-V for Vulkan,
which this seems to have been intended for. It was also unused elsewhere.
Two varyings (vertex outputs/fragment inputs) might have the same
location but be in different components--e.g. the compiler may have
packed what were two different varyings into a single varying vector.
Giving both varyings the same `[[user]]` attribute won't work--it may
yield unexpected results, or flat out fail to link. We could eventually
pack such varyings into a single vector, but that would require us to
handle the case where the varyings are different types--e.g. a `float`
and a `uint` packed into the same vector. For now, it seems most
prudent to give them unique `[[user]]` locations and let Apple's
compiler work out the best way to pack them.
