WGSL-Specific Functionalities¶

This chapter provides information for WGSL (WebGPU Shading Language)-specific functionalities and behaviors.

System-Value semantics¶

The system-value semantics are translated to the following WGSL code.

SV semantic name	WGSL code
SV_Barycentrics	Not supported
SV_ClipDistance	Not supported
SV_CullDistance	Not supported
SV_Coverage	`@builtin(sample_mask)`
SV_CullPrimitive	Not supported
SV_Depth	`@builtin(frag_depth)`
SV_DepthGreaterEqual	Not supported
SV_DepthLessEqual	Not supported
SV_DispatchThreadID	`@builtin(global_invocation_id)`
SV_DomainLocation	Not supported
SV_GSInstanceID	Not supported
SV_GroupID	`@builtin(workgroup_id)`
SV_GroupIndex	`@builtin(local_invocation_index)`
SV_GroupThreadID	`@builtin(local_invocation_id)`
SV_InnerCoverage	Not supported
SV_InsideTessFactor	Not supported
SV_InstanceID	`@builtin(instance_index)`
SV_IntersectionAttributes	Not supported
SV_IsFrontFace	`@builtin(front_facing)`
SV_OutputControlPointID	Not supported
SV_PointSize	Not supported
SV_Position	`@builtin(position)`
SV_PrimitiveID	Not supported
SV_RenderTargetArrayIndex	Not supported
SV_SampleIndex	`@builtin(sample_index)`
SV_ShadingRate	Not supported
SV_StartVertexLocation	Not supported
SV_StartInstanceLocation	Not supported
SV_StencilRef	Not supported
SV_Target	Not supported
SV_TessFactor	Not supported
SV_VertexID	`@builtin(vertex_index)`
SV_ViewID	Not supported
SV_ViewportArrayIndex	Not supported
SV_VulkanInstanceID	`@builtin(instance_index)`
SV_VulkanVertexID	`@builtin(vertex_index)`

Supported HLSL features when targeting WGSL¶

The following table lists Slang’s support for various HLSL feature sets, when targeting WGSL.

Feature set	Supported
ray tracing	No
inline ray tracing	No
mesh shader	No
tessellation shader	No
geometry shader	No
wave intrinsics	No
barriers	Yes
atomics	Yes

Supported atomic types¶

The following table shows what is supported when targeting WGSL:

	32-bit integer	64-bit integer	32-bit float	64-bit float	16-bit float
Supported?	Yes	No	No	No	No

ConstantBuffer, (RW/RasterizerOrdered)StructuredBuffer, (RW/RasterizerOrdered)ByteAddressBuffer¶

ConstantBuffer translates to the uniform address space with read access mode in WGSL. ByteAddressBuffer and RWByteAddressBuffer translate to array<u32> in the storage address space, with the read and read_write access modes in WGSL, respectively. StructuredBuffer and RWStructuredBuffer with struct type T translate to array<T> in the storage address space, with with the read and read_write access modes in WGSL, respectively.

Specialization Constants¶

Specialization constants are not supported when targeting WGSL, at the moment. They should map to ‘override declarations’ in WGSL, however this is not yet implemented.

Interlocked operations¶

The InterlockedAdd, InterlockedAnd, etc… functions are not supported when targeting WGSL. Instead, operations on Atomic<T> types should be used.

Entry Point Parameter Handling¶

Slang performs several transformations on entry point parameters when targeting WGSL:

Struct parameters and returned structs are flattened to eliminate nested structures.
System value semantics are translated to WGSL built-ins. (See the @builtin attribute, and the table above.)
Parameters without semantics are given automatic location indices. (See the @location attribute.)

Parameter blocks¶

Each ParameterBlock is assigned its own bind group in WGSL.

Write-only Textures¶

Many image formats supported by WebGPU can only be accessed in compute shader as a write-only image. Use WTexture2D type (similar to RWTexture2D) to write to an image when possible. The write-only texture types are also supported when targeting HLSL/GLSL/SPIR-V/Metal and CUDA.

Pointers¶

out and inout parameters in Slang are translated to pointer-typed parameters in WGSL. At callsites, a pointer value is formed and passed as argument using the & operator in WGSL.

Since WGSL cannot form pointers to fields of structs (or fields of fields of structs, etc…), the described transformation cannot be done in a direct way when a function argument expression is an “access chain” like myStruct.myField or myStruct.myStructField.someField. In those cases, the argument is copied to a local variable, the address of the local variable is passed to the function, and then the local variable is written back to the struct field after the function call.

Address Space Assignment¶

WGSL requires explicit address space qualifiers. Slang automatically assigns appropriate address spaces:

Variable Type	WGSL Address Space
Local Variables	`function`
Global Variables	`private`
Uniform Buffers	`uniform`
RW/Structured Buffers	`storage`
Group Shared	`workgroup`
Parameter Blocks	`uniform`

Matrix type translation¶

A m-row-by-n-column matrix in Slang, represented as floatmxn or matrix<T, m, n>, is translated to mat[n]x[m] in WGSL, i.e. a matrix with n columns and m rows. The rationale for this inversion of terminology is the same as the rationale for SPIR-V. Since the WGSL matrix multiplication convention is the normal one, where inner products of rows of the matrix on the left are taken with columns of the matrix on the right, the order of matrix products is also reversed in WGSL. This is relying on the fact that the transpose of a matrix product equals the product of the transposed matrix operands in reverse order.

Explicit Parameter Binding¶

The [vk::binding(index,set)] attribute is respected when emitting WGSL code, and will translate to @binding(index) @group(set) in WGSL.

If the [vk::binding()] attribute is not specified but a :register() semantic is present, Slang will derive the binding from the register semantic the same way as the SPIR-V and GLSL backends.

The [vk::location(N)] attributes on stage input/output parameters are respected.

Specialization Constants¶

Specialization constants declared with the [SpecializationConstant] or [vk::constant_id] attribute will be translated into a global override declaration when generating WGSL source. For example:

[vk::constant_id(7)]
const int a = 2;

Translates to:

@id(7) override a : i32 = 2;