We only considered invalid names, and overwrote the alias for the
function. The correct fix is to replace illegal names early, do the
reserved fixup, then copy back alias to entry point name.
In Metal, the `[[position]]` input to a fragment shader remains at
fragment center, even at sample rate, like OpenGL and Direct3D. In
Vulkan, however, when the fragment shader runs at sample rate, the
`FragCoord` builtin moves to the sample position in the framebuffer,
instead of the fragment center. To account for this difference, adjust
the `FragCoord`, if present, by the sample position. The -0.5 offset is
because the fragment center is at (0.5, 0.5).
Also, add an option to force sample-rate shading in a fragment shader.
Since Metal has no explicit control for this, this is done by adding a
dummy `[[sample_id]]` which is otherwise unused, if none is already
present. This is intended to be used from e.g. MoltenVK when a
pipeline's `minSampleShading` value is nonzero.
Instead of checking if any `Input` variables have `Sample`
interpolation, I've elected to check that the `SampleRateShading`
capability is present. Since `SampleId`, `SamplePosition`, and the
`Sample` interpolation decoration require this cap, this should be
equivalent for any valid SPIR-V module. If this isn't acceptable, let me
know.
Add support for declaring a fixed subgroup size. Metal, like Vulkan with
`VK_EXT_subgroup_size_control`, allows the thread execution width to
vary depending on factors such as register usage. Unfortunately, this
breaks several tests that depend on the subgroup size being what the
device says it is. So we'll fix the subgroup size at the size the device
declares. The extra invocations in the subgroup will appear to be
inactive. Because of this, the ballot mask builtins are now ANDed with
the active subgroup mask.
Add support for emulating a subgroup of size 1. This is intended to be
used by Vulkan Portability implementations (e.g. MoltenVK) when the
hardware/software combo provides insufficient support for subgroups.
Luckily for us, Vulkan 1.1 only requires that the subgroup size be at
least 1.
Add support for quadgroup and SIMD-group functions which were added to
iOS in Metal 2.2 and 2.3. This will allow clients to take advantage of
expanded quadgroup and SIMD-group support in recent Metal versions and
on recent Apple GPUs (families 6 and 7).
Gut emulation of subgroup builtins in fragment shaders. It turns out
codegen for the SIMD-group functions in fragment wasn't implemented for
AMD on Mojave; it's a safe bet that it wasn't implemented for the other
drivers either. Subgroup support in fragment shaders now requires Metal
2.2.
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.
(GL_EXT_nonuniform_qualifier/SPV_EXT_descriptor_indexing).
MSLResourceBinding includes array size through API, and substitutes
in that size if the image or sampler array is not explicitly sized.
OpCopyObject supports SPIRCombinedImageSampler type in MSL.
Metal doesn't support broadcasting or shuffling boolean values, but we
can work around that by casting it to `ushort`, then casting it back to
`bool`. I used `ushort` instead of `uint` because 16-bit values give
better throughput on Apple GPUs.
Only the least *n* bits are significant, where *n* is the subgroup size.
The Vulkan CTS actually checks this.
The `FindLSB` tests weren't actually failing, but I masked that anyway,
in case there's some corner case the CTS is missing.
`half` cannot be bitcasted to `float`, because the two types are not the
same size. Use an expanding cast instead.
We were already doing this for stores to the tessellation levels; why I
didn't also do this for loads is beyond me.
Fix reversed coordinates: `y` should be used to calculate the row
address. Align row address to the row stride.
I've made the row alignment a function constant; this makes it possible
to override it at pipeline compile time.
Honestly, I don't know how this worked at all for Epic. It definitely
didn't work in the CTS prior to this.
These need to use arrayed texture types, or Metal will complain when
binding the resource. The target layer is addressed relative to the
Layer output by the vertex pipeline, or to the ViewIndex if in a
multiview pipeline. Unlike with the s/t coordinates, Vulkan does not
forbid non-zero layer coordinates here, though this cannot be expressed
in Vulkan GLSL.
Supporting 3D textures will require additional work. Part of the problem
is that Metal does not allow texture views to subset a 3D texture, so we
need some way to pass the base depth to the shader.
Some older iOS devices don't support layered rendering. In that case,
don't set `[[render_target_array_index]]`, because the compiler will
reject the shader in that case. The client will then have to unroll the
render pass manually.
In Metal render pipelines don't have an option to set a sampleMask
parameter, the only way to get that functionality is to set the
sample_mask output of the fragment shader to this value directly.
We also need to take care to combine the fixed sample mask with the
one that the shader might possibly output.
This should hopefully reduce underutilization of the GPU, especially on
GPUs where the thread execution width is greater than the number of
control points.
This also simplifies initialization by reading the buffer directly
instead of using Metal's vertex-attribute-in-compute support. It turns
out the only way in which shader stages are allowed to differ in their
interfaces is in the number of components per vector; the base type must
be the same. Since we are using the raw buffer instead of attributes, we
can now also emit arrays and matrices directly into the buffer, instead
of flattening them and then unpacking them. Structs are still flattened,
however; this is due to the need to handle vectors with fewer components
than were output, and I think handling this while also directly emitting
structs could get ugly.
Another advantage of this scheme is that the extra invocations needed to
read the attributes when there were more input than output points are
now no more. The number of threads per workgroup is now lcm(SIMD-size,
output control points). This should ensure we always process a whole
number of patches per workgroup.
To avoid complexity handling indices in the tessellation control shader,
I've also changed the way vertex shaders for tessellation are handled.
They are now compute kernels using Metal's support for vertex-style
stage input. This lets us always emit vertices into the buffer in order
of vertex shader execution. Now we no longer have to deal with indexing
in the tessellation control shader. This also fixes a long-standing
issue where if an index were greater than the number of vertices to
draw, the vertex shader would wind up writing outside the buffer, and
the vertex would be lost.
This is a breaking change, and I know SPIRV-Cross has other clients, so
I've hidden this behind an option for now. In the future, I want to
remove this option and make it the default.
On MSL, the compiler refuses to allow access chains into a normal vector type.
What happens in practice instead is a read-modify-write where a vector type is
loaded, modified and written back.
The workaround is to convert a vector into a pointer-to-scalar before
the access chain continues to add the scalar index.
Metal is picky about interface matching. If the types don't match
exactly, down to the number of vector components, Metal fails pipline
compilation. To support pipelines where the number of components
consumed by the fragment shader is less than that produced by the vertex
shader, we have to fix up the fragment shader to accept all the
components produced.
Like with `point_size` when not rendering points, Metal complains when
writing to a variable using the `[[depth]]` qualifier when no depth
buffer be attached. In that case, we must avoid emitting `FragDepth`,
just like with `PointSize`.
I assume it will also complain if there be no stencil attachment and the
shader write to `[[stencil]]`, or it write to `[[color(n)]]` but there
be no color attachment at n.
Limit inline blocks to one per descriptor set.
This should avoid the need for complicated code to calculate the
argument buffer ID stride of an inline uniform block. If there's demand
for more inline blocks, we can revisit this.
Here, the inline uniform block is explicit: we instantiate the buffer
block itself in the argument buffer, instead of a pointer to the buffer.
I just hope this will work with the `MTLArgumentDescriptor` API...
Note that Metal recursively assigns individual members of embedded
structs IDs. This means for automatic assignment that we have to
calculate the binding stride for a given buffer block. For MoltenVK,
we'll simply increment the ID by the size of the inline uniform block.
Then the later IDs will never conflict with the inline uniform block. We
can get away with this because Metal doesn't require that IDs be
contiguous, only monotonically increasing.
There was a hack to workaround a bug in DXC where control point -> patch
constant phase was passed in Function storage, but we have to use
Workgroup here. We will not support these kinds of hacks for invalid
SPIR-V, so hack the reference files to use the "proper" fix and remove
the hack for time being.
To support loading array of array properly in tessellation, we need a
rewrite of how tessellation access chains are handled.
The major change is to remove the implicit unflatten step inside
access_chain which does not take into account the case where you load
directly from a control point array variable.
We defer unflatten step until OpLoad time instead.
This fixes cases where we load array of {array,matrix,struct}.
Removes the hacky path for MSL access chain index workaround.
Add CompilerMSL::Options::texture_1D_as_2D.
Metal imposes significant restrictions on 1D textures, including not being
renderable, clearable, or permitting mipmaps. This option allows SPIR-V 1D
textures to be treated as 2D textures to permit this additional behaviour.
App must of course supply the textures to Metal as 2D textures.
There is an implicit tristate with {-1, 0, +1} values, but it was not
obvious how this was supposed to work before studying the implementation,
so refactor into a tristate enum class.
If there are enough members in an IAB, we cannot use the constant
address space as MSL compiler complains about there being too many
members. Support emitting the device address space instead.
First, when generating from HLSL before invoking the code that comes from the HLSL patch-function a control-flow and full memory-barrier are required to ensure that all the temporary values in thread-local storage for the patch are available.
Second, the inputs to control and evaluation shaders must be properly forwarded from the global variables in SPIRV to the member variables in the relevant input structure.
Finally when arrays of interpolators are used for input or output we need to add an extra level of array indirection because Metal works at a different granularity than SPIRV.
Five parts.
1. Fix tessellation patch function processing.
2. Fix loads from tessellation control inputs not being forwarded to the gl_in structure array.
3. Fix loads from tessellation evaluation inputs not being forwarded to the stage_in structure array.
4. Workaround SPIRV losing an array indirection in tessellation shaders - not the best solution but enough to keep things progressing.
5. Apparently gl_TessLevelInner/Outer is special and needs to not be placed into the input array.
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.
Vulkan has two types of buffer descriptors,
`VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC` and
`VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC`, which allow the client to
offset the buffers by an amount given when the descriptor set is bound
to a pipeline. Metal provides no direct support for this when the buffer
in question is in an argument buffer, so once again we're on our own.
These offsets cannot be stored or associated in any way with the
argument buffer itself, because they are set at bind time. Different
pipelines may have different offsets set. Therefore, we must use a
separate buffer, not in any argument buffer, to hold these offsets. Then
the shader must manually offset the buffer pointer.
This change fully supports arrays, including arrays of arrays, even
though Vulkan forbids them. It does not, however, support runtime
arrays. Perhaps later.
Writable textures cannot use argument buffers on iOS. They must be
passed as arguments directly to the shader function. Since we won't know
if a given storage image will have the `NonWritable` decoration at the
time we encode the argument buffer, we must therefore pass all storage
images as discrete arguments. Previously, we were throwing an error if
we encountered an argument buffer with a writable texture in it on iOS.
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.
These methods have largely the same logic, with minor differences. That
I felt compelled to duplicate the logic into another method was one of
the things that bothered me about the variable pointers change. This
cleans that part of the code up; now we don't have two places to change.
This command allows the caller to set the base value of
`BuiltInWorkgroupId`, and thus of `BuiltInGlobalInvocationId`. Metal
provides no direct support for this... but it does provide a builtin,
`[[grid_origin]]`, normally used to pass the base values for the stage
input region, which we will now abuse to pass the dispatch base and
avoid burning a buffer binding.
`[[grid_origin]]`, as part of Metal's support for compute stage input,
requires MSL 1.2. For 1.0 and 1.1, we're forced to provide a buffer.
(Curiously, this builtin was undocumented until the MSL 2.2 release. Go
figure.)
Make sure to test everything with scalar as well to catch any weird edge
cases.
Not all opcodes are covered here, just the arithmetic ones. FP64 packing
is also ignored.
The only piece added by this extension is the `DeviceIndex` builtin,
which tells the shader which device in a grouped logical device it is
running on.
Metal's pipeline state objects are owned by the `MTLDevice` that created
them. Since Metal doesn't support logical grouping of devices the way
Vulkan does, we'll thus have to create a pipeline state for each device
in a grouped logical device. The upcoming peer group support in Metal 3
will not change this. For this reason, for Metal, the device index is
supplied as a constant at pipeline compile time.
There's an interaction between `VK_KHR_device_group` and
`VK_KHR_multiview` in the
`VK_PIPELINE_CREATE_VIEW_INDEX_FROM_DEVICE_INDEX_BIT`, which defines the
view index to be the same as the device index. The new
`view_index_from_device_index` MSL option supports this functionality.
This maps them to their MSL equivalents. I've mapped `Coherent` to
`volatile` since MSL doesn't have anything weaker than `volatile` but
stronger than nothing.
As part of this, I had to remove the implicit `volatile` added for
atomic operation casts. If the buffer is already `coherent` or
`volatile`, then we would add a second `volatile`, which would be
redundant. I think this is OK even when the buffer *doesn't* have
`coherent`: `T *` is implicitly convertible to `volatile T *`, but not
vice-versa. It seems to compile OK at any rate. (Note that the
non-`volatile` overloads of the atomic functions documented in the spec
aren't present in the MSL 2.2 stdlib headers.)
`restrict` is tricky, because in MSL, as in C++, it needs to go *after*
the asterisk or ampersand for the pointer type it's modifying.
Another issue is that, in the `Simple`, `GLSL450`, and `Vulkan` memory
models, `Restrict` is the default (i.e. does not need to be specified);
but MSL likely follows the `OpenCL` model where `Aliased` is the
default. We probably need to implicitly set either `Restrict` or
`Aliased` depending on the module's declared memory model.
The old method of using a different unpacked matrix type doesn't work
for scalar alignment. It certainly wouldn't have any effect for a square
matrix, since the number of columns and rows are the same. So now we'll
store them as arrays of packed vectors.
Relaxed block layout relaxed the restrictions on vector alignment,
allowing them to be aligned on scalar boundaries. Scalar block layout
relaxes this further, allowing *any* member to be aligned on a scalar
boundary. The requirement that a vector not improperly straddle a
16-byte boundary is also relaxed.
I've also added a test showing that `std430` layout works with UBOs.
I'm troubled by the dual meaning of the `Packed` extended decoration. In
some instances (struct, `float[]`, and `vec2[]` members), it actually
means the exact opposite, that the member needs extra padding. This is
especially problematic for `vec2[]`, because now we need to distinguish
the two cases by checking the array stride. I wonder if this should
actually be split into two decorations.
This is needed to support `VK_KHR_multiview`, which is in turn needed
for Vulkan 1.1 support. Unfortunately, Metal provides no native support
for this, and Apple is once again less than forthcoming, so we have to
implement it all ourselves.
Tessellation and geometry shaders are deliberately unsupported for now.
The problem is that the current implementation encodes the `ViewIndex`
as part of the `InstanceIndex`, which in the SPIR-V environment at least
only exists in the vertex shader. So we need to work out a way to pass
the view index along to the later stages.
This implementation runs vertex shaders for all views up to the highest
bit set in the view mask, even those whose bits are clear. The fragments
for the inactive views are then discarded. Avoiding this is difficult:
calculating the view indices becomes far more complicated if we can only
run for those views which are set in the mask.
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".
Older API was oriented around IDs which are not available unless you're
doing full reflection, which is awkward for certain use cases which know
their set/bindings up front.
Optimize resource bindings to be hashmap rather than doing linear seeks
all the time.
In multiple-entry-point modules, we declared builtin inputs which were
not supposed to be used for that entry point.
Fix this, by being more strict when checking which builtins to emit.
This gets rather complicated because MSL does not support OpArrayLength
natively. We need to pass down a buffer which contains buffer sizes, and
we compute the array length on-demand.
Support both discrete descriptors as well as argument buffers.
Change aux buffer to swizzle buffer.
There is no good reason to expand the aux buffer, so name it
appropriately.
Make the code cleaner by emitting a straight pointer to uint rather than
a dummy struct which only contains a single unsized array member anyways.
This will also end up being very similar to how we implement swizzle
buffers for argument buffers.
Do not use implied binding if it overflows int32_t.
Some support for subgroups is present starting in Metal 2.0 on both iOS
and macOS. macOS gains more complete support in 10.14 (Metal 2.1).
Some restrictions are present. On iOS and on macOS 10.13, the
implementation of `OpGroupNonUniformElect` is incorrect: if thread 0 has
already terminated or is not executing a conditional branch, the first
thread that *is* will falsely believe itself not to be. Unfortunately,
this operation is part of the "basic" feature set; without it, subgroups
cannot be supported at all.
The `SubgroupSize` and `SubgroupLocalInvocationId` builtins are only
available in compute shaders (and, by extension, tessellation control
shaders), despite SPIR-V making them available in all stages. This
limits the usefulness of some of the subgroup operations in fragment
shaders.
Although Metal on macOS supports some clustered, inclusive, and
exclusive operations, it does not support them all. In particular,
inclusive and exclusive min, max, and, or, and xor; as well as cluster
sizes other than 4 are not supported. If this becomes a problem, they
could be emulated, but at a significant performance cost due to the need
for non-uniform operations.
Atomics are not supported on images or texture_buffers in MSL.
Properly throw an error if OpImageTexelPointer is used (since it can
only be used for atomic operations anyways).
- 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 cannot deduce if OpLoad needs ArrayCopy templates early since it's
heavily context dependent, and we might only know on 3rd iteration of
the compile loop.
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.
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.
The tessellation levels in Metal are stored as a densely-packed array of
half-precision floating point values. But, stage-in attributes in Metal
have to have offsets and strides aligned to a multiple of four, so we
can't add them individually. Luckily for us, the arrays have lengths
less than 4. So, let's use vectors for them!
Triangles get a single attribute with a `float4`, where the outer levels
are in `.xyz` and the inner levels are in `.w`. The arrays are unpacked
as though we had added the elements individually. Quads get two: a
`float4` with the outer levels and a `float2` with the inner levels.
Further, since vectors can be indexed as arrays, there's no need to
unpack them in this case.
This also saves on precious vertex attributes. Before, we were using up
to 6 of them. Now we need two at most.
Builtin attributes in SPIR-V aren't linked by location, but by their
built-in-ness. This poses a problem for MSL, since builtin inputs in
the vertex pipeline are just regular attributes. We must then assign
them locations so that they can be matched up to the attributes in the
stage input descriptor--and also to avoid duplicate attribute numbers in
tessellation evaluation shaders, where there are two different
stage-in structs, so the member index therein is no longer unique!
In SPIR-V, there are always two inner levels and four outer levels, even
if the input patch isn't a quad patch. But in MSL, due to requirements
imposed by Metal, only one inner level and three outer levels exist when
the input patch is a triangle patch. We must explicitly ignore any write
to the nonexistent second inner and fourth outer levels in this case.
This is intended to be used to support `VK_KHR_maintenance2`'s
tessellation domain origin feature. If `tess_domain_origin_lower_left`
is `true`, the `v` coordinate will be inverted with respect to the
domain. Additionally, in `Triangles` mode, the `v` and `w` coordinates
will be swapped. This is because the winding order is interpreted
differently in lower-left mode.
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.