Skip to content

Latest commit

 

History

History
4213 lines (3412 loc) · 215 KB

SPIR-V.rst

File metadata and controls

4213 lines (3412 loc) · 215 KB
Raw

HLSL to SPIR-V Feature Mapping Manual

Introduction

This document describes the mappings from HLSL features to SPIR-V for Vulkan adopted by the SPIR-V codegen. For how to build, use, or contribute to the SPIR-V codegen and its internals, please see the wiki page.

SPIR-V is a binary intermediate language for representing graphical-shader stages and compute kernels for multiple Khronos APIs, such as Vulkan, OpenGL, and OpenCL. At the moment we only intend to support the Vulkan flavor of SPIR-V.

DirectXShaderCompiler is the reference compiler for HLSL. Adding SPIR-V codegen in DirectXShaderCompiler will enable the usage of HLSL as a frontend language for Vulkan shader programming. Sharing the same code base also means we can track the evolution of HLSL more closely and always deliver the best of HLSL to developers. Moreover, developers will also have a unified compiler toolchain for targeting both DirectX and Vulkan. We believe this effort will benefit the general graphics ecosystem.

Note that this document is expected to be an ongoing effort and grow as we implement more and more HLSL features.

Overview

Although they share the same basic concepts, DirectX and Vulkan are still different graphics APIs with semantic gaps. HLSL is the native shading language for DirectX, so certain HLSL features do not have corresponding mappings in Vulkan, and certain Vulkan specific information does not have native ways to express in HLSL source code. This section describes the general translation paradigms and how we close some of the major semantic gaps.

Note that the term "semantic" is overloaded. In HLSL, it can mean the string attached to shader input or output. For such cases, we refer it as "HLSL semantic" or "semantic string". For other cases, we just use the normal "semantic" term.

Shader entry function

HLSL entry functions can read data from the previous shader stage and write data to the next shader stage via function parameters and return value. On the contrary, Vulkan requires all SPIR-V entry functions taking no parameters and returning void. All data passing between stages should use global variables in the Input and Output storage class.

To handle this difference, we emit a wrapper function as the SPIR-V entry function around the HLSL source code entry function. The wrapper function is responsible to read data from SPIR-V Input global variables and prepare them to the types required in the source code entry function signature, call the source code entry function, and then decompose the contents in return value (and out/inout parameters) to the types required by the SPIR-V Output global variables, and then write out. For details about the wrapper function, please refer to the entry function wrapper section.

Shader stage IO interface matching

HLSL leverages semantic strings to link variables and pass data between shader stages. Great flexibility is allowed as for how to use the semantic strings. They can appear on function parameters, function returns, and struct members. In Vulkan, linking variables and passing data between shader stages is done via numeric Location decorations on SPIR-V global variables in the Input and Output storage class.

To help handling such differences, we provide Vulkan specific attributes to let the developer to express precisely their intents. The compiler will also try its best to deduce the mapping from semantic strings to SPIR-V Location numbers when such explicit Vulkan specific attributes are absent. Please see the HLSL semantic and Vulkan Location section for more details about the mapping and Location assignment.

What makes the story complicated is Vulkan's strict requirements on interface matching. Basically, a variable in the previous stage is considered a match to a variable in the next stage if and only if they are decorated with the same Location number and with the exact same type, except for the outermost arrayness in hull/domain/geometry shader, which can be ignored regarding interface matching. This is causing problems together with the flexibility of HLSL semantic strings.

Some HLSL system-value (SV) semantic strings will be mapped into SPIR-V variables with builtin decorations, some are not. HLSL non-SV semantic strings should all be mapped to SPIR-V variables without builtin decorations (but with Location decorations).

With these complications, if we are grouping multiple semantic strings in a struct in the HLSL source code, that struct should be flattened and each of its members should be mapped separately. For example, for the following:

struct T {
  float2 clip0 : SV_ClipDistance0;
  float3 cull0 : SV_CullDistance0;
  float4 foo   : FOO;
};

struct S {
  float4 pos   : SV_Position;
  float2 clip1 : SV_ClipDistance1;
  float3 cull1 : SV_CullDistance1;
  float4 bar   : BAR;
  T      t;
};

If we have an S input parameter in pixel shader, we should flatten it recursively to generate five SPIR-V Input variables. Three of them are decorated by the Position, ClipDistance, CullDistance builtin, and two of them are decorated by the Location decoration. (Note that clip0 and clip1 are concatenated, also cull0 and cull1. The ClipDistance and CullDistance builtins are special and explained in the ClipDistance & CullDistance section.)

Flattening is infective because of Vulkan interface matching rules. If we flatten a struct in the output of a previous stage, which may create multiple variables decorated with different Location numbers, we also need to flatten it in the input of the next stage. otherwise we may have Location mismatch even if we share the same definition of the struct. Because hull/domain/geometry shader is optional, we can have different chains of shader stages, which means we need to flatten all shader stage interfaces. For hull/domain/geometry shader, their inputs/outputs have an additional arrayness. So if we are seeing an array of structs in these shaders, we need to flatten them into arrays of its fields.

Vulkan specific features

We try to implement Vulkan specific features using the most intuitive and non-intrusive ways in HLSL, which means we will prefer native language constructs when possible. If that is inadequate, we then consider attaching Vulkan specific attributes to them, or introducing new syntax.

Descriptors

The compiler provides multiple mechanisms to specify which Vulkan descriptor a particular resource binds to.

In the source code, you can use the [[vk::binding(X[, Y])]] and [[vk::counter_binding(X)]] attribute. The native :register() attribute is also respected.

On the command-line, you can use the -fvk-{b|s|t|u}-shift or -fvk-bind-register option.

If you can modify the source code, the [[vk::binding(X[, Y])]] and [[vk::counter_binding(X)]] attribute gives you find-grained control over descriptor assignment.

If you cannot modify the source code, you can use command-line options to change how :register() attribute is handled by the compiler. -fvk-bind-register lets you to specify the descriptor for the source at a certain register. -fvk-{b|s|t|u}-shift lets you to apply shifts to all register numbers of a certain register type. They cannot be used together, though.

When the [[vk::combinedImageSampler]] attribute is applied, only the -fvk-t-shift value will be used to apply shifts to combined texture and sampler resource bindings and any -fvk-s-shift value will be ignored.

Without attribute and command-line option, :register(xX, spaceY) will be mapped to binding X in descriptor set Y. Note that register type x is ignored, so this may cause overlap.

The more specific a mechanism is, the higher precedence it has, and command-line option has higher precedence over source code attribute.

For more details, see HLSL register and Vulkan binding, Vulkan specific attributes, and Vulkan-specific options.

Subpass inputs

Within a Vulkan rendering pass, a subpass can write results to an output target that can then be read by the next subpass as an input subpass. The "Subpass Input" feature regards the ability to read an output target.

Subpasses are read through two new builtin resource types, available only in pixel shader:

class SubpassInput<T> {
  T SubpassLoad();
};

class SubpassInputMS<T> {
  T SubpassLoad(int sampleIndex);
};

In the above, T is a scalar or vector type. If omitted, it will defaults to float4.

Subpass inputs are implicitly addressed by the pixel's (x, y, layer) coordinate. These objects support reading the subpass input through the methods as shown in the above.

A subpass input is selected by using a new attribute vk::input_attachment_index. For example:

[[vk::input_attachment_index(i)]] SubpassInput input;

An vk::input_attachment_index of i selects the ith entry in the input pass list. (See Vulkan API spec for more information.)

Push constants

Vulkan push constant blocks are represented using normal global variables of struct types in HLSL. The variables (not the underlying struct types) should be annotated with the [[vk::push_constant]] attribute.

Please note as per the requirements of Vulkan, "there must be no more than one push constant block statically used per shader entry point."

Specialization constants

To use Vulkan specialization constants, annotate global constants with the [[vk::constant_id(X)]] attribute. For example,

[[vk::constant_id(1)]] const bool  specConstBool  = true;
[[vk::constant_id(2)]] const int   specConstInt   = 42;
[[vk::constant_id(3)]] const float specConstFloat = 1.5;

Shader Record Buffer

SPV_NV_ray_tracing exposes user managed buffer in shader binding table by using storage class ShaderRecordBufferNV. ConstantBuffer or cbuffer blocks can now be mapped to this storage class under HLSL by using [[vk::shader_record_nv]] annotation. It is applicable only on ConstantBuffer and cbuffer declarations.

Please note as per the requirements of VK_NV_ray_tracing, "there must be no more than one shader_record_nv block statically used per shader entry point otherwise results are undefined."

The official Khronos ray tracing extension also comes with a SPIR-V storage class that has the same functionality. The [[vk::shader_record_ext]] annotation can be used when targeting the SPV_KHR_ray_tracing extension.

Builtin variables

Some of the Vulkan builtin variables have no equivalents in native HLSL language. To support them, [[vk::builtin("<builtin>")]] is introduced. Right now the following <builtin> are supported:

  • PointSize: The GLSL equivalent is gl_PointSize.
  • HelperInvocation: For Vulkan 1.3 or above, we use its GLSL equivalent gl_HelperInvocation and decorate it with HelperInvocation builtin since Vulkan 1.3 or above supports Volatile decoration for builtin variables. For Vulkan 1.2 or earlier, we do not create a builtin variable for HelperInvocation. Instead, we create a variable with Private storage class and set its value as the result of OpIsHelperInvocationEXT instruction.
  • BaseVertex: The GLSL equivalent is gl_BaseVertexARB. Need SPV_KHR_shader_draw_parameters extension.
  • BaseInstance: The GLSL equivalent is gl_BaseInstanceARB. Need SPV_KHR_shader_draw_parameters extension.
  • DrawIndex: The GLSL equivalent is gl_DrawIDARB. Need SPV_KHR_shader_draw_parameters extension.
  • DeviceIndex: The GLSL equivalent is gl_DeviceIndex. Need SPV_KHR_device_group extension.
  • ViewportMaskNV: The GLSL equivalent is gl_ViewportMask.

Please see Vulkan spec. 14.6. Built-In Variables for detailed explanation of these builtins.

Supported extensions

  • SPV_KHR_16bit_storage
  • SPV_KHR_device_group
  • SPV_KHR_fragment_shading_rate
  • SPV_KHR_multivew
  • SPV_KHR_post_depth_coverage
  • SPV_KHR_shader_draw_parameters
  • SPV_EXT_descriptor_indexing
  • SPV_EXT_fragment_fully_covered
  • SPV_EXT_mesh_shader
  • SPV_EXT_shader_stencil_support
  • SPV_AMD_shader_early_and_late_fragment_tests
  • SPV_GOOGLE_hlsl_functionality1
  • SPV_GOOGLE_user_type
  • SPV_NV_mesh_shader
  • SPV_KHR_fragment_shading_barycentric

Vulkan specific attributes

C++ attribute specifier sequence is a non-intrusive way of providing Vulkan specific information in HLSL.

The namespace vk will be used for all Vulkan attributes:

  • location(X): For specifying the location (X) numbers for stage input/output variables. Allowed on function parameters, function returns, and struct fields.
  • binding(X[, Y]): For specifying the descriptor set (Y) and binding (X) numbers for resource variables. The descriptor set (Y) is optional; if missing, it will be set to 0. Allowed on global variables.
  • counter_binding(X): For specifying the binding number (X) for the associated counter for RW/Append/Consume structured buffer. The descriptor set number for the associated counter is always the same as the main resource.
  • push_constant: For marking a variable as the push constant block. Allowed on global variables of struct type. At most one variable can be marked as push_constant in a shader.
  • offset(X): For manually layout struct members. Annotating a struct member with this attribute will force the compiler to put the member at offset X w.r.t. the beginning of the struct. Only allowed on struct members.
  • constant_id(X): For marking a global constant as a specialization constant. Allowed on global variables of boolean/integer/float types.
  • input_attachment_index(X): To associate the Xth entry in the input pass list to the annotated object. Only allowed on objects whose type are SubpassInput or SubpassInputMS.
  • builtin("X"): For specifying an entity should be translated into a certain Vulkan builtin variable. Allowed on function parameters, function returns, and struct fields.
  • index(X): For specifying the index at a specific pixel shader output location. Used for dual-source blending.
  • post_depth_coverage: The input variable decorated with SampleMask will reflect the result of the EarlyFragmentTests. Only valid on pixel shader entry points.
  • combinedImageSampler: For specifying a Texture (e.g., Texture2D, Texture1DArray, TextureCube) and SamplerState to use the combined image sampler (or sampled image) type with the same descriptor set and binding numbers (see wiki page for more detail).
  • early_and_late_tests: Marks an entry point as enabling early and late depth tests. If depth is written via SV_Depth, depth_unchanged must also be specified (SV_DepthLess and SV_DepthGreater can be written freely). If a stencil reference value is written via SV_StencilRef, one of stencil_ref_unchanged_front, stencil_ref_greater_equal_front, or stencil_ref_less_equal_front and one of stencil_ref_unchanged_back, stencil_ref_greater_equal_back, or stencil_ref_less_equal_back must be specified.
  • depth_unchanged: Specifies that any depth written to SV_Depth will not invalidate the result of early depth tests. Sets the DepthUnchanged execution mode in SPIR-V. Only valid on pixel shader entry points.
  • stencil_ref_unchanged_front: Specifies that any stencil ref written to SV_StencilRef will not invalidate the result of early stencil tests when the fragment is front facing. Sets the StencilRefUnchangedFrontAMD execution mode in SPIR-V. Only valid on pixel shader entry points.
  • stencil_ref_greater_equal_front: Specifies that any stencil ref written to SV_StencilRef will be greater than or equal to the stencil reference value set by the API when the fragment is front facing. Sets the StencilRefGreaterFrontAMD execution mode in SPIR-V. Only valid on pixel shader entry points.
  • stencil_ref_less_equal_front: Specifies that any stencil ref written to SV_StencilRef will be less than or equal to the stencil reference value set by the API when the fragment is front facing. Sets the StencilRefLessFrontAMD execution mode in SPIR-V. Only valid on pixel shader entry points.
  • stencil_ref_unchanged_back: Specifies that any stencil ref written to SV_StencilRef will not invalidate the result of early stencil tests when the fragment is back facing. Sets the StencilRefUnchangedBackAMD execution mode in SPIR-V. Only valid on pixel shader entry points.
  • stencil_ref_greater_equal_back: Specifies that any stencil ref written to SV_StencilRef will be greater than or equal to the stencil reference value set by the API when the fragment is back facing. Sets the StencilRefGreaterBackAMD execution mode in SPIR-V. Only valid on pixel shader entry points.
  • stencil_ref_less_equal_back: Specifies that any stencil ref written to SV_StencilRef will be less than or equal to the stencil reference value set by the API when the fragment is back facing. Sets the StencilRefLessBackAMD execution mode in SPIR-V. Only valid on pixel shader entry points.

Only vk:: attributes in the above list are supported. Other attributes will result in warnings and be ignored by the compiler. All C++11 attributes will only trigger warnings and be ignored if not compiling towards SPIR-V.

For example, to specify the layout of resource variables and the location of interface variables:

struct S { ... };

[[vk::binding(X, Y), vk::counter_binding(Z)]]
RWStructuredBuffer<S> mySBuffer;

[[vk::location(M)]] float4
main([[vk::location(N)]] float4 input: A) : B
{ ... }

Macro for SPIR-V

If SPIR-V CodeGen is enabled and -spirv flag is used as one of the command line options (meaning that "generates SPIR-V code"), it defines an implicit macro __spirv__. For example, this macro definition can be used for SPIR-V specific part of the HLSL code:

#ifdef __spirv__
[[vk::binding(X, Y), vk::counter_binding(Z)]]
#endif
RWStructuredBuffer<S> mySBuffer;

SPIR-V version and extension

SPIR-V CodeGen provides two command-line options for fine-grained SPIR-V target environment (hence SPIR-V version) and SPIR-V extension control:

  • -fspv-target-env=: for specifying SPIR-V target environment
  • -fspv-extension=: for specifying allowed SPIR-V extensions

-fspv-target-env= accepts a Vulkan target environment (see -help for supported values). If such an option is not given, the CodeGen defaults to vulkan1.0. When targeting vulkan1.0, trying to use features that are only available in Vulkan 1.1 (SPIR-V 1.3), like Shader Model 6.0 wave intrinsics, will trigger a compiler error.

If -fspv-extension= is not specified, the CodeGen will select suitable SPIR-V extensions to translate the source code. Otherwise, only extensions supplied via -fspv-extension= will be used. If that does not suffice, errors will be emitted explaining what additional extensions are required to translate what specific feature in the source code. If you want to allow all KHR extensions, you can use -fspv-extension=KHR.

Legalization, optimization, validation

After initial translation of the HLSL source code, SPIR-V CodeGen will further conduct legalization (if needed), optimization (if requested), and validation (if not turned off). All these three stages are outsourced to SPIRV-Tools. Here are the options controlling these stages:

  • -fcgl: turn off legalization and optimization
  • -Od: turn off optimization
  • -Vd: turn off validation

Legalization

HLSL is a fairly permissive language considering the flexibility it provides for manipulating resource objects. The developer can create local copies, pass them around as function parameters and return values, as long as after certain transformations (function inlining, constant evaluation and propagating, dead code elimination, etc.), the compiler can remove all temporary copies and pinpoint all uses to unique global resource objects.

Resulting from the above property of HLSL, if we translate into SPIR-V for Vulkan literally from the input HLSL source code, we will sometimes generate illegal SPIR-V. Certain transformations are needed to legalize the literally translated SPIR-V. Performing such transformations at the frontend AST level is cumbersome or impossible (e.g., function inlining). They are better to be conducted at SPIR-V level. Therefore, legalization is delegated to SPIRV-Tools.

Specifically, we need to legalize the following HLSL source code patterns:

  • Using resource types in struct types
  • Creating aliases of global resource objects
  • Control flows invovling the above cases

Legalization transformations will not run unless the above patterns are encountered in the source code.

For more details, please see the SPIR-V cookbook, which contains examples of what HLSL code patterns will be accepted and generate valid SPIR-V for Vulkan.

Optimization

Optimization is also delegated to SPIRV-Tools. Right now there are no difference between optimization levels greater than zero; they will all invoke the same optimization recipe. That is, the recipe behind spirv-opt -O. If you want to run a custom optimization recipe, you can do so using the command line option -Oconfig= and specifying a comma-separated list of your desired passes. The passes are invoked in the specified order.

For example, you can specify -Oconfig=--loop-unroll,--scalar-replacement=300,--eliminate-dead-code-aggressive to firstly invoke loop unrolling, then invoke scalar replacement of aggregates, lastly invoke aggressive dead code elimination. All valid options to spirv-opt are accepted as components to the comma-separated list.

Here are the typical passes in alphabetical order:

  • --ccp
  • --cfg-cleanup
  • --convert-local-access-chains
  • --copy-propagate-arrays
  • --eliminate-dead-branches
  • --eliminate-dead-code-aggressive
  • --eliminate-dead-functions
  • --eliminate-local-multi-store
  • --eliminate-local-single-block
  • --eliminate-local-single-store
  • --flatten-decorations
  • --if-conversion
  • --inline-entry-points-exhaustive
  • --local-redundancy-elimination
  • --loop-fission
  • --loop-fusion
  • --loop-unroll
  • --loop-unroll-partial=[<n>]
  • --loop-peeling (requires --loop-peeling-threshold)
  • --merge-blocks
  • --merge-return
  • --loop-unswitch
  • --private-to-local
  • --reduce-load-size
  • --redundancy-elimination
  • --remove-duplicates
  • --replace-invalid-opcode
  • --ssa-rewrite
  • --scalar-replacement[=<n>]
  • --simplify-instructions
  • --vector-dce

Besides, there are two special batch options; each stands for a recommended recipe by itself:

  • -O: A bunch of passes in an appropriate order that attempt to improve performance of generated code. Same as spirv-opt -O. Also same as SPIR-V CodeGen's default recipe.
  • -Os: A bunch of passes in an appropriate order that attempt to reduce the size of the generated code. Same as spirv-opt -Os.

So if you want to run loop unrolling additionally after the default optimization recipe, you can specify -Oconfig=-O,--loop-unroll.

For the whole list of accepted passes and details about each one, please see spirv-opt's help manual (spirv-opt --help), or the SPIRV-Tools optimizer header file.

Validation

Validation is turned on by default as the last stage of SPIR-V CodeGen. Failing validation, which indicates there is a CodeGen bug, will trigger a fatal error. Please file an issue if you see that.

Debugging

By default, the compiler will only emit names for types and variables as debug information, to aid reading of the generated SPIR-V. The -Zi option will let the compiler emit the following additional debug information:

  • Full path of the main source file using OpSource
  • Preprocessed source code using OpSource and OpSourceContinued
  • Line information for certain instructions using OpLine (WIP)
  • DXC Git commit hash using OpModuleProcessed (requires Vulkan 1.1)
  • DXC command-line options used to compile the shader using OpModuleProcessed (requires Vulkan 1.1)

We chose to embed preprocessed source code instead of original source code to avoid pulling in lots of contents unrelated to the current entry point, and boilerplate contents generated by engines. We may add a mode for selecting between preprocessed single source code and original separated source code in the future.

One thing to note is that to keep the line numbers in consistent with the embedded source, the compiler is invoked twice; the first time is for preprocessing the source code, and the second time is for feeding the preprocessed source code as input for a whole compilation. So using -Zi means performance penality.

If you want to have fine-grained control over the categories of emitted debug information, you can use -fspv-debug=. It accepts:

  • file: for emitting full path of the main source file
  • source: for emitting preprocessed source code (turns on file implicitly)
  • line: for emitting line information (turns on source implicitly)
  • tool: for emitting DXC Git commit hash and command-line options

These -fspv-debug= options overrule -Zi. And you can provide multiple instances of -fspv-debug=. For example, you can use -fspv-debug=file -fspv-debug=tool to turn on emitting file path and DXC information; source code and line information will not be emitted.

If you want to generate NonSemantic.Shader.DebugInfo.100 extended instructions, you can use -fspv-debug=vulkan-with-source. These instructions support source-level shader debugging with tools such as RenderDoc, even if the SPIR-V is optimized. This option overrules the other -fspv-debug options above.

Reflection

Making reflection easier is one of the goals of SPIR-V CodeGen. This section provides guidelines about how to reflect on certain facts.

Note that we generate OpName/OpMemberName instructions for various types/variables both explicitly defined in the source code and interally created by the compiler. These names are primarily for debugging purposes in the compiler. They have "no semantic impact and can safely be removed" according to the SPIR-V spec. And they are subject to changes without notice. So we do not suggest to use them for reflection.

Source code shader profile

The source code shader profile version can be re-discovered by the "Version" operand in OpSource instruction. For *s_<major>_<minor>, the "Verison" operand in OpSource will be set as <major> * 100 + <minor> * 10. For example, vs_5_1 will have 510, ps_6_2 will have 620.

HLSL Semantic

HLSL semantic strings are by default not emitted into the SPIR-V binary module. If you need them, by specifying -fspv-reflect, the compiler will use the Op*DecorateStringGOOGLE instruction in SPV_GOOGLE_hlsl_funtionality1 extension to emit them.

HLSL User Types

HLSL type information is by default not emitted into the SPIR-V binary module. If you need them, by specifying -fspv-reflect, the compiler will emit OpDecorateString* instructions with a UserTypeGOOGLE decoration and the SPV_GOOGLE_user_type extension. A string name for the unambiguous type of the decorated object will be included in the user's source using the lowercase type name followed by template params. For example, Texture2DMSArray<float4, 64> arr would be decorated with OpDecorateString %arr UserTypeGOOGLE "texture2dmsarray:<float4,64>".

Counter buffers for RW/Append/Consume StructuredBuffer

The association between a counter buffer and its main RW/Append/Consume StructuredBuffer is conveyed by OpDecorateId <structured-buffer-id> HLSLCounterBufferGOOGLE <counter-buffer-id> instruction from the SPV_GOOGLE_hlsl_funtionality1 extension. This information is by default missing; you need to specify -fspv-reflect to direct the compiler to emit them.

Read-only vs. read-write resource types

There are no clear and consistent decorations in the SPIR-V to show whether a resource type is translated from a read-only (RO) or read-write (RW) HLSL resource type. Instead, you need to use different checks for reflecting different resource types:

  • HLSL samplers: RO.
  • HLSL Buffer/RWBuffer/Texture*/RWTexture*: Check the "Sampled" operand in the OpTypeImage instruction they translated into. "2" means RW, "1" means RO.
  • HLSL constant/texture/structured/byte buffers: Check both Block/BufferBlock and NonWritable decoration. If decorated with Block (cbuffer & ConstantBuffer), then RO; if decorated with BufferBlock and NonWritable (tbuffer, TextureBuffer, StructuredBuffer), then RO; Otherwise, RW.

HLSL Types

This section lists how various HLSL types are mapped.

Normal scalar types

Normal scalar types in HLSL are relatively easy to handle and can be mapped directly to SPIR-V type instructions:

HLSL Command Line Option SPIR-V Capability Extension
bool   OpTypeBool    
int/int32_t   OpTypeInt 32 1    
int16_t -enable-16bit-types OpTypeInt 16 1 Int16  
uint/dword/uin32_t   OpTypeInt 32 0    
uint16_t -enable-16bit-types OpTypeInt 16 0 Int16  
half   OpTypeFloat 32    
half/float16_t -enable-16bit-types OpTypeFloat 16   SPV_AMD_gpu_shader_half_float
float/float32_t   OpTypeFloat 32    
snorm float   OpTypeFloat 32    
unorm float   OpTypeFloat 32    
double/float64_t   OpTypeFloat 64 Float64  

Please note that half is translated into 32-bit floating point numbers if without -enable-16bit-types because MSDN says that "this data type is provided only for language compatibility. Direct3D 10 shader targets map all half data types to float data types."

Minimal precision scalar types

HLSL also supports various minimal precision scalar types, which graphics drivers can implement by using any precision greater than or equal to their specified bit precision. There are no direct mappings in SPIR-V for these types. We translate them into the corresponding 16-bit or 32-bit scalar types with the RelaxedPrecision decoration. We use the 16-bit variants if '-enable-16bit-types' command line option is present. For more information on these types, please refer to: https://github.com/Microsoft/DirectXShaderCompiler/wiki/16-Bit-Scalar-Types

HLSL Command Line Option SPIR-V Decoration Capability Extension
min16float   OpTypeFloat 32 RelaxedPrecision    
min10float   OpTypeFloat 32 RelaxedPrecision    
min16int   OpTypeInt 32 1 RelaxedPrecision    
min12int   OpTypeInt 32 1 RelaxedPrecision    
min16uint   OpTypeInt 32 0 RelaxedPrecision    
min16float -enable-16bit-types OpTypeFloat 16     SPV_AMD_gpu_shader_half_float
min10float -enable-16bit-types OpTypeFloat 16     SPV_AMD_gpu_shader_half_float
min16int -enable-16bit-types OpTypeInt 16 1   Int16  
min12int -enable-16bit-types OpTypeInt 16 1   Int16  
min16uint -enable-16bit-types OpTypeInt 16 0   Int16  

Vectors and matrices

Vectors and matrices are translated into:

HLSL SPIR-V
|type|N (N > 1) OpTypeVector |type| N
|type|1 The scalar type for |type|
|type|MxN (M > 1, N > 1) %v = OpTypeVector |type| N OpTypeMatrix %v M
|type|Mx1 (M > 1) OpTypeVector |type| M
|type|1xN (N > 1) OpTypeVector |type| N
|type|1x1 The scalar type for |type|

The above table is for float matrices.

A MxN HLSL float matrix is translated into a SPIR-V matrix with M vectors, each with N elements. Conceptually HLSL matrices are row-major while SPIR-V matrices are column-major, thus all HLSL matrices are represented by their transposes. Doing so may require special handling of certain matrix operations:

  • Indexing: no special handling required. matrix[m][n] will still access the correct element since m/n means the m-th/n-th row/column in HLSL but m-th/n-th vector/element in SPIR-V.
  • Per-element operation: no special handling required.
  • Matrix multiplication: need to swap the operands. mat1 x mat2 should be translated as transpose(mat2) x transpose(mat1). Then the result is transpose(mat1 x mat2).
  • Storage layout: row_major/column_major will be translated into SPIR-V ColMajor/RowMajor decoration. This is because HLSL matrix row/column becomes SPIR-V matrix column/row. If elements in a row/column are packed together, they should be loaded into a column/row correspondingly.

See Appendix A. Matrix Representation for further explanation regarding these design choices.

Since the Shader capability in SPIR-V does not allow to parameterize matrix types with non-floating-point types, a non-floating-point MxN matrix is translated into an array with M elements, with each element being a vector with N elements.

Structs

Structs in HLSL are defined in the a format similar to C structs. They are translated into SPIR-V OpTypeStruct. Depending on the storage classes of the instances, a single struct definition may generate multiple OpTypeStruct instructions in SPIR-V. For example, for the following HLSL source code:

struct S { ... }

ConstantBuffer<S>   myCBuffer;
StructuredBuffer<S> mySBuffer;

float4 main() : A {
  S myLocalVar;
  ...
}

There will be three different OpTypeStruct generated, one for each variable defined in the above source code. This is because the OpTypeStruct for both myCBuffer and mySBuffer will have layout decorations (Offset, MatrixStride, ArrayStride, RowMajor, ColMajor). However, their layout rules are different (by default); myCBuffer will use vector-relaxed OpenGL std140 while mySBuffer will use vector-relaxed OpenGL std430. myLocalVar will have its OpTypeStruct without layout decorations. Read more about storage classes in the Constant/Texture/Structured/Byte Buffers section.

Structs used as stage inputs/outputs will have semantics attached to their members. These semantics are handled in the entry function wrapper.

Structs used as pixel shader inputs can have optional interpolation modifiers for their members, which will be translated according to the following table:

HLSL Interpolation Modifier SPIR-V Decoration SPIR-V Capability
linear <none>  
centroid Centroid  
nointerpolation Flat  
noperspective NoPerspective  
sample Sample SampleRateShading

Arrays

Sized (either explicitly or implicitly) arrays are translated into SPIR-V OpTypeArray. Unsized arrays are translated into OpTypeRuntimeArray.

Arrays, if used for external resources (residing in SPIR-V Uniform or UniformConstant storage class), will need layout decorations like SPIR-V ArrayStride decoration. For arrays of opaque types, e.g., HLSL textures or samplers, we don't decorate with ArrayStride decorations since there is no meaningful strides. Similarly for arrays of structured/byte buffers.

User-defined types

User-defined types are type aliases introduced by typedef. No new types are introduced and we can rely on Clang to resolve to the original types.

Samplers

All sampler types will be translated into SPIR-V OpTypeSampler.

SPIR-V OpTypeSampler is an opaque type that cannot be parameterized; therefore state assignments on sampler types is not supported (yet).

Textures

Texture types are translated into SPIR-V OpTypeImage, with parameters:

HLSL Vulkan SPIR-V
Texture Type Descriptor Type RO/RW Storage Class Dim Depth Arrayed MS Sampled Image Format Capability
Texture1D Sampled Image RO UniformConstant 1D 2 0 0 1 Unknown  
Texture2D Sampled Image RO UniformConstant 2D 2 0 0 1 Unknown  
Texture3D Sampled Image RO UniformConstant 3D 2 0 0 1 Unknown  
TextureCube Sampled Image RO UniformConstant Cube 2 0 0 1 Unknown  
Texture1DArray Sampled Image RO UniformConstant 1D 2 1 0 1 Unknown  
Texture2DArray Sampled Image RO UniformConstant 2D 2 1 0 1 Unknown  
Texture2DMS Sampled Image RO UniformConstant 2D 2 0 1 1 Unknown  
Texture2DMSArray Sampled Image RO UniformConstant 2D 2 1 1 1 Unknown  
TextureCubeArray Sampled Image RO UniformConstant 3D 2 1 0 1 Unknown  
Buffer<T> Uniform Texel Buffer RO UniformConstant Buffer 2 0 0 1 Depends on T SampledBuffer
RWBuffer<T> Storage Texel Buffer RW UniformConstant Buffer 2 0 0 2 Depends on T SampledBuffer
RWTexture1D<T> Storage Image RW UniformConstant 1D 2 0 0 2 Depends on T  
RWTexture2D<T> Storage Image RW UniformConstant 2D 2 0 0 2 Depends on T  
RWTexture3D<T> Storage Image RW UniformConstant 3D 2 0 0 2 Depends on T  
RWTexture1DArray<T> Storage Image RW UniformConstant 1D 2 1 0 2 Depends on T  
RWTexture2DArray<T> Storage Image RW UniformConstant 2D 2 1 0 2 Depends on T  

The meanings of the headers in the above table is explained in OpTypeImage of the SPIR-V spec.

Vulkan specific Image Formats

Since HLSL lacks the syntax for fully specifying image formats for textures in SPIR-V, we introduce [[vk::image_format("FORMAT")]] attribute for texture types. For example,

[[vk::image_format("rgba8")]]
RWBuffer<float4> Buf;

[[vk::image_format("rg16f")]]
RWTexture2D<float2> Tex;

RWTexture2D<float2> Tex2; // Works like before

rgba8 means Rgba8 SPIR-V Image Format. The following table lists the mapping between FORMAT of [[vk::image_format("FORMAT")]] and its corresponding SPIR-V Image Format.

FORMAT SPIR-V Image Format
unknown Unknown
rgba32f Rgba32f
rgba16f Rgba16f
r32f R32f
rgba8 Rgba8
rgba8snorm Rgba8Snorm
rg32f Rg32f
rg16f Rg16f
r11g11b10f R11fG11fB10f
r16f R16f
rgba16 Rgba16
rgb10a2 Rgb10A2
rg16 Rg16
rg8 Rg8
r16 R16
r8 R8
rgba16snorm Rgba16Snorm
rg16snorm Rg16Snorm
rg8snorm Rg8Snorm
r16snorm R16Snorm
r8snorm R8Snorm
rgba32i Rgba32i
rgba16i Rgba16i
rgba8i Rgba8i
r32i R32i
rg32i Rg32i
rg16i Rg16i
rg8i Rg8i
r16i R16i
r8i R8i
rgba32ui Rgba32ui
rgba16ui Rgba16ui
rgba8ui Rgba8ui
r32ui R32ui
rgb10a2ui Rgb10a2ui
rg32ui Rg32ui
rg16ui Rg16ui
rg8ui Rg8ui
r16ui R16ui
r8ui R8ui
r64ui R64ui
r64i R64i

Constant/Texture/Structured/Byte Buffers

There are serveral buffer types in HLSL:

  • cbuffer and ConstantBuffer
  • tbuffer and TextureBuffer
  • StructuredBuffer and RWStructuredBuffer
  • AppendStructuredBuffer and ConsumeStructuredBuffer
  • ByteAddressBuffer and RWByteAddressBuffer

Note that Buffer and RWBuffer are considered as texture object in HLSL. They are listed in the above section.

Please see the following sections for the details of each type. As a summary:

HLSL Type Vulkan Buffer Type Default Memory Layout Rule SPIR-V Storage Class SPIR-V Decoration
cbuffer Uniform Buffer Vector-relaxed OpenGL std140 Uniform Block
ConstantBuffer Uniform Buffer Vector-relaxed OpenGL std140 Uniform Block
tbuffer Storage Buffer Vector-relaxed OpenGL std430 Uniform BufferBlock
TextureBuffer Storage Buffer Vector-relaxed OpenGL std430 Uniform BufferBlock
StructuredBuffer Storage Buffer Vector-relaxed OpenGL std430 Uniform BufferBlock
RWStructuredBuffer Storage Buffer Vector-relaxed OpenGL std430 Uniform BufferBlock
AppendStructuredBuffer Storage Buffer Vector-relaxed OpenGL std430 Uniform BufferBlock
ConsumeStructuredBuffer Storage Buffer Vector-relaxed OpenGL std430 Uniform BufferBlock
ByteAddressBuffer Storage Buffer Vector-relaxed OpenGL std430 Uniform BufferBlock
RWByteAddressBuffer Storage Buffer Vector-relaxed OpenGL std430 Uniform BufferBlock

To know more about the Vulkan buffer types, please refer to the Vulkan spec 13.1 Descriptor Types.

Memory layout rules

SPIR-V CodeGen supports four sets of memory layout rules for buffer resources right now:

  1. Vector-relaxed OpenGL std140 for uniform buffers and vector-relaxed OpenGL std430 for storage buffers: these rules satisfy Vulkan "Standard Uniform Buffer Layout" and "Standard Storage Buffer Layout", respectively. They are the default.
  2. DirectX memory layout rules for uniform buffers and storage buffers: they allow packing data on the application side that can be shared with DirectX. They can be enabled by -fvk-use-dx-layout.
  3. Strict OpenGL std140 for uniform buffers and strict OpenGL std430 for storage buffers: they allow packing data on the application side that can be shared with OpenGL. They can be enabled by -fvk-use-gl-layout.
  4. Scalar layout rules introduced via VK_EXT_scalar_block_layout, which basically aligns all aggregrate types according to their elements' natural alignment. They can be enabled by -fvk-use-scalar-layout.

To use scalar layout, the application side need to request VK_EXT_scalar_block_layout. This is also true for using DirectX memory layout since there is no dedicated DirectX layout extension for Vulkan (at least for now). So we must request something more permissive.

In the above, "vector-relaxed OpenGL std140/std430" rules mean OpenGL std140/std430 rules with the following modification for vector type alignment:

  1. The alignment of a vector type is set to be the alignment of its element type
  2. If the above causes an improper straddle, the alignment will be set to 16 bytes.

As an exmaple, for the following HLSL definition:

struct S {
    float3 f;
};

struct T {
              float    a_float;
              float3   b_float3;
              S        c_S_float3;
              float2x3 d_float2x3;
    row_major float2x3 e_float2x3;
              int      f_int_3[3];
              float2   g_float2_2[2];
};

We will have the following offsets for each member:

HLSL Uniform Buffer Storage Buffer
Member 1 (VK) 2 (DX) 3 (GL) 4 (Scalar) 1 (VK) 2 (DX) 3 (GL) 4 (Scalar)
a_float 0 0 0 0 0 0 0 0
b_float3 4 4 16 4 4 4 16 4
c_S_float3 16 16 32 16 16 16 32 16
d_float2x3 32 32 48 28 32 28 48 28
e_float2x3 80 80 96 52 64 52 80 52
f_int_3 112 112 128 76 96 76 112 76
g_float2_2 160 160 176 88 112 88 128 88

If the above layout rules do not satisfy your needs and you want to manually control the layout of struct members, you can use either

  • The native HLSL :packoffset() attribute: only available for cbuffers; or
  • The Vulkan-specific [[vk::offset()]] attribute: applies to all resources.

[[vk::offset]] overrules :packoffset. Attaching [[vk::offset]] to a struct memeber affects all variables of the struct type in question. So sharing the same struct definition having [[vk::offset]] annotations means also sharing the layout.

For global variables (which are collected into the $Globals cbuffer), you can use the native HLSL :register(c#) attribute. Note that [[vk::offset]] and :packoffset cannot be applied to these variables.

If register(cX) is used on any global variable, the offset for that variable is set to X * 16, and the offset for all other global variables without the register(c#) annotation will be set to the next available address after the highest explicit address. For example:

float x : register(c10);   // Offset = 160 (10 * 16)
float y;                   // Offset = 164 (160 + 4)
float z: register(c1);     // Offset = 16  (1  * 16)

These attributes give great flexibility but also responsibility to the developer; the compiler will just take in what is specified in the source code and emit it to SPIR-V with no error checking.

cbuffer and ConstantBuffer

These two buffer types are treated as uniform buffers using Vulkan's terminology. They are translated into an OpTypeStruct with the necessary layout decorations (Offset, ArrayStride, MatrixStride, RowMajor, ColMajor) and the Block decoration. The layout rule used is vector-relaxed OpenGL std140 (by default). A variable declared as one of these types will be placed in the Uniform storage class.

For example, for the following HLSL source code:

struct T {
  float  a;
  float3 b;
};

ConstantBuffer<T> myCBuffer;

will be translated into

; Layout decoration
OpMemberDecorate %type_ConstantBuffer_T 0 Offset 0
OpMemberDecorate %type_ConstantBuffer_T 0 Offset 4
; Block decoration
OpDecorate %type_ConstantBuffer_T Block

; Types
%type_ConstantBuffer_T = OpTypeStruct %float %v3float
%_ptr_Uniform_type_ConstantBuffer_T = OpTypePointer Uniform %type_ConstantBuffer_T

; Variable
%myCbuffer = OpVariable %_ptr_Uniform_type_ConstantBuffer_T Uniform

tbuffer and TextureBuffer

These two buffer types are treated as storage buffers using Vulkan's terminology. They are translated into an OpTypeStruct with the necessary layout decorations (Offset, ArrayStride, MatrixStride, RowMajor, ColMajor) and the BufferBlock decoration. All the struct members are also decorated with NonWritable decoration. The layout rule used is vector-relaxed OpenGL std430 (by default). A variable declared as one of these types will be placed in the Uniform storage class.

StructuredBuffer and RWStructuredBuffer

StructuredBuffer<T>/RWStructuredBuffer<T> is treated as storage buffer using Vulkan's terminology. It is translated into an OpTypeStruct containing an OpTypeRuntimeArray of type T, with necessary layout decorations (Offset, ArrayStride, MatrixStride, RowMajor, ColMajor) and the BufferBlock decoration. The default layout rule used is vector-relaxed OpenGL std430. A variable declared as one of these types will be placed in the Uniform storage class.

For RWStructuredBuffer<T>, each variable will have an associated counter variable generated. The counter variable will be of OpTypeStruct type, which only contains a 32-bit integer. The counter variable takes its own binding number. .IncrementCounter()/.DecrementCounter() will modify this counter variable.

For example, for the following HLSL source code:

struct T {
  float  a;
  float3 b;
};

StructuredBuffer<T> mySBuffer;

will be translated into

; Layout decoration
OpMemberDecorate %T 0 Offset 0
OpMemberDecorate %T 1 Offset 4
OpDecorate %_runtimearr_T ArrayStride 16
OpMemberDecorate %type_StructuredBuffer_T 0 Offset 0
OpMemberDecorate %type_StructuredBuffer_T 0 NoWritable
; BufferBlock decoration
OpDecorate %type_StructuredBuffer_T BufferBlock

; Types
%T = OpTypeStruct %float %v3float
%_runtimearr_T = OpTypeRuntimeArray %T
%type_StructuredBuffer_T = OpTypeStruct %_runtimearr_T
%_ptr_Uniform_type_StructuredBuffer_T = OpTypePointer Uniform %type_StructuredBuffer_T

; Variable
%myCbuffer = OpVariable %_ptr_Uniform_type_ConstantBuffer_T Uniform

AppendStructuredBuffer and ConsumeStructuredBuffer

AppendStructuredBuffer<T>/ConsumeStructuredBuffer<T> is treated as storage buffer using Vulkan's terminology. It is translated into an OpTypeStruct containing an OpTypeRuntimeArray of type T, with necessary layout decorations (Offset, ArrayStride, MatrixStride, RowMajor, ColMajor) and the BufferBlock decoration. The default layout rule used is vector-relaxed OpenGL std430.

A variable declared as one of these types will be placed in the Uniform storage class. Besides, each variable will have an associated counter variable generated. The counter variable will be of OpTypeStruct type, which only contains a 32-bit integer. The integer is the total number of elements in the buffer. The counter variable takes its own binding number. .Append()/.Consume() will use the counter variable as the index and adjust it accordingly.

For example, for the following HLSL source code:

struct T {
  float  a;
  float3 b;
};

AppendStructuredBuffer<T> mySBuffer;

will be translated into

; Layout decorations
OpMemberDecorate %T 0 Offset 0
OpMemberDecorate %T 1 Offset 4
OpDecorate %_runtimearr_T ArrayStride 16
OpMemberDecorate %type_AppendStructuredBuffer_T 0 Offset 0
OpDecorate %type_AppendStructuredBuffer_T BufferBlock
OpMemberDecorate %type_ACSBuffer_counter 0 Offset 0
OpDecorate %type_ACSBuffer_counter BufferBlock

; Binding numbers
OpDecorate %myASbuffer DescriptorSet 0
OpDecorate %myASbuffer Binding 0
OpDecorate %counter_var_myASbuffer DescriptorSet 0
OpDecorate %counter_var_myASbuffer Binding 1

; Types
%T = OpTypeStruct %float %v3float
%_runtimearr_T = OpTypeRuntimeArray %T
%type_AppendStructuredBuffer_T = OpTypeStruct %_runtimearr_T
%_ptr_Uniform_type_AppendStructuredBuffer_T = OpTypePointer Uniform %type_AppendStructuredBuffer_T
%type_ACSBuffer_counter = OpTypeStruct %int
%_ptr_Uniform_type_ACSBuffer_counter = OpTypePointer Uniform %type_ACSBuffer_counter

; Variables
%myASbuffer = OpVariable %_ptr_Uniform_type_AppendStructuredBuffer_T Uniform
%counter_var_myASbuffer = OpVariable %_ptr_Uniform_type_ACSBuffer_counter Uniform

ByteAddressBuffer and RWByteAddressBuffer

ByteAddressBuffer/RWByteAddressBuffer is treated as storage buffer using Vulkan's terminology. It is translated into an OpTypeStruct containing an OpTypeRuntimeArray of 32-bit unsigned integers, with BufferBlock decoration.

A variable declared as one of these types will be placed in the Uniform storage class.

For example, for the following HLSL source code:

ByteAddressBuffer   myBuffer1;
RWByteAddressBuffer myBuffer2;

will be translated into

; Layout decorations

OpDecorate %_runtimearr_uint ArrayStride 4

OpDecorate %type_ByteAddressBuffer BufferBlock
OpMemberDecorate %type_ByteAddressBuffer 0 Offset 0
OpMemberDecorate %type_ByteAddressBuffer 0 NonWritable

OpDecorate %type_RWByteAddressBuffer BufferBlock
OpMemberDecorate %type_RWByteAddressBuffer 0 Offset 0

; Types

%_runtimearr_uint = OpTypeRuntimeArray %uint

%type_ByteAddressBuffer = OpTypeStruct %_runtimearr_uint
%_ptr_Uniform_type_ByteAddressBuffer = OpTypePointer Uniform %type_ByteAddressBuffer

%type_RWByteAddressBuffer = OpTypeStruct %_runtimearr_uint
%_ptr_Uniform_type_RWByteAddressBuffer = OpTypePointer Uniform %type_RWByteAddressBuffer

; Variables

%myBuffer1 = OpVariable %_ptr_Uniform_type_ByteAddressBuffer Uniform
%myBuffer2 = OpVariable %_ptr_Uniform_type_RWByteAddressBuffer Uniform

Rasterizer Ordered Views

The following types are rasterizer ordered views:

  • RasterizerOrderedBuffer
  • RasterizerOrderedByteAddressBuffer
  • RasterizerOrderedStructuredBuffer
  • RasterizerOrderedTexture1D
  • RasterizerOrderedTexture1DArray
  • RasterizerOrderedTexture2D
  • RasterizerOrderedTexture2DArray
  • RasterizerOrderedTexture3D

These are translated to the same types as their equivalent RW* types - for example, a RasterizerOrderedBuffer is translated to the same SPIR-V type as an RWBuffer. The sole difference lies in how loads and stores to these values are treated.

The access order guarantee made by ROVs is implemented in SPIR-V using the SPV_EXT_fragment_shader_interlock. When you load or store a value from or to a rasterizer ordered view, using either the Load*() or Store*() methods or the indexing operator, OpBeginInvocationInterlockEXT will be inserted before the first access and OpEndInvocationInterlockEXT will be inserted after the last access.

An execution mode will be added to the entry point, depending on the sample frequency, which will be deduced based on the semantics inputted by the entry point. PixelInterlockOrderedEXT will be selected by default, SampleInterlockOrderedEXT will be selected if the SV_SampleIndex semantic is input, and ShadingRateInterlockOrderedEXT will be selected if the SV_ShadingRate semantic is input.

HLSL Variables and Resources

This section lists how various HLSL variables and resources are mapped.

According to Shader Constants,

There are two default constant buffers available, $Global and $Param. Variables that are placed in the global scope are added implicitly to the $Global cbuffer, using the same packing method that is used for cbuffers. Uniform parameters in the parameter list of a function appear in the $Param constant buffer when a shader is compiled outside of the effects framework.

So all global externally-visible non-resource-type stand-alone variables will be collected into a cbuffer named as $Globals, no matter whether they are statically referenced by the entry point or not. The $Globals cbuffer follows the layout rules like normal cbuffer.

Storage class

Normal local variables (without any modifier) will be placed in the Function SPIR-V storage class. Normal global variables (without any modifer) will be placed in the Uniform or UniformConstant storage class.

  • static
    • Global variables with static modifier will be placed in the Private SPIR-V storage class. Initalizers of such global variables will be translated into SPIR-V OpVariable initializers if possible; otherwise, they will be initialized at the very beginning of the entry function wrapper using SPIR-V OpStore.
    • Local variables with static modifier will also be placed in the Private SPIR-V storage class. initializers of such local variables will also be translated into SPIR-V OpVariable initializers if possible; otherwise, they will be initialized at the very beginning of the enclosing function. To make sure that such a local variable is only initialized once, a second boolean variable of the Private SPIR-V storage class will be generated to mark its initialization status.
  • groupshared
    • Global variables with groupshared modifier will be placed in the Workgroup storage class.
    • Note that this modifier overrules static; if both groupshared and static are applied to a variable, static will be ignored.
  • uinform
    • This does not affect codegen. Variables will be treated like normal global variables.
  • extern
    • This does not affect codegen. Variables will be treated like normal global variables.
  • shared
    • This is a hint to the compiler. It will be ingored.
  • volatile
    • This is a hint to the compiler. It will be ingored.

HLSL semantic and Vulkan Location

Direct3D uses HLSL "semantics" to compose and match the interfaces between subsequent stages. These semantic strings can appear after struct members, function parameters and return values. E.g.,

struct VSInput {
  float4 pos  : POSITION;
  float3 norm : NORMAL;
};

float4 VSMain(in  VSInput input,
              in  float4  tex   : TEXCOORD,
              out float4  pos   : SV_Position) : TEXCOORD {
  pos = input.pos;
  return tex;
}

In contrary, Vulkan stage input and output interface matching is via explicit Location numbers. Details can be found here.

To translate HLSL to SPIR-V for Vulkan, semantic strings need to be mapped to Vulkan Location numbers properly. This can be done either explicitly via information provided by the developer or implicitly by the compiler.

Explicit Location number assignment

[[vk::location(X)]] can be attached to the entities where semantic are allowed to attach (struct fields, function parameters, and function returns). For the above exmaple we can have:

struct VSInput {
  [[vk::location(0)]] float4 pos  : POSITION;
  [[vk::location(1)]] float3 norm : NORMAL;
};

[[vk::location(1)]]
float4 VSMain(in  VSInput input,
              [[vk::location(2)]]
              in  float4  tex     : TEXCOORD,
              out float4  pos     : SV_Position) : TEXCOORD {
  pos = input.pos;
  return tex;
}

In the above, input POSITION, NORMAL, and TEXCOORD will be mapped to Location 0, 1, and 2, respectively, and output TEXCOORD will be mapped to Location 1.

[TODO] Another explicit way: using command-line options

Please note that the compiler prohibits mixing the explicit and implicit approach for the same SigPoint to avoid complexity and fallibility. However, for a certain shader stage, one SigPoint using the explicit approach while the other adopting the implicit approach is permitted.

Implicit Location number assignment

Without hints from the developer, the compiler will try its best to map semantics to Location numbers. However, there is no single rule for this mapping; semantic strings should be handled case by case.

Firstly, under certain SigPoints, some system-value (SV) semantic strings will be translated into SPIR-V BuiltIn decorations:

HLSL Semantic SigPoint SPIR-V BuiltIn SPIR-V Execution Mode SPIR-V Capability
SV_Position VSOut Position N/A Shader
HSCPIn Position N/A Shader
HSCPOut Position N/A Shader
DSCPIn Position N/A Shader
DSOut Position N/A Shader
GSVIn Position N/A Shader
GSOut Position N/A Shader
PSIn FragCoord N/A Shader
MSOut Position N/A Shader
SV_ClipDistance VSOut ClipDistance N/A ClipDistance
HSCPIn ClipDistance N/A ClipDistance
HSCPOut ClipDistance N/A ClipDistance
DSCPIn ClipDistance N/A ClipDistance
DSOut ClipDistance N/A ClipDistance
GSVIn ClipDistance N/A ClipDistance
GSOut ClipDistance N/A ClipDistance
PSIn ClipDistance N/A ClipDistance
MSOut ClipDistance N/A ClipDistance
SV_CullDistance VSOut CullDistance N/A CullDistance
HSCPIn CullDistance N/A CullDistance
HSCPOut CullDistance N/A CullDistance
DSCPIn CullDistance N/A CullDistance
DSOut CullDistance N/A CullDistance
GSVIn CullDistance N/A CullDistance
GSOut CullDistance N/A CullDistance
PSIn CullDistance N/A CullDistance
MSOut CullDistance N/A CullDistance
SV_VertexID VSIn VertexIndex N/A Shader
SV_InstanceID VSIn InstanceIndex or InstanceIndex - BaseInstance with -fvk-support-nonzero-base-instance N/A Shader
SV_Depth PSOut FragDepth N/A Shader
SV_DepthGreaterEqual PSOut FragDepth DepthGreater Shader
SV_DepthLessEqual PSOut FragDepth DepthLess Shader
SV_IsFrontFace PSIn FrontFacing N/A Shader
SV_DispatchThreadID CSIn GlobalInvocationId N/A Shader
MSIn GlobalInvocationId N/A Shader
ASIn GlobalInvocationId N/A Shader
SV_GroupID CSIn WorkgroupId N/A Shader
MSIn WorkgroupId N/A Shader
ASIn WorkgroupId N/A Shader
SV_GroupThreadID CSIn LocalInvocationId N/A Shader
MSIn LocalInvocationId N/A Shader
ASIn LocalInvocationId N/A Shader
SV_GroupIndex CSIn LocalInvocationIndex N/A Shader
MSIn LocalInvocationIndex N/A Shader
ASIn LocalInvocationIndex N/A Shader
SV_OutputControlPointID HSIn InvocationId N/A Tessellation
SV_GSInstanceID GSIn InvocationId N/A Geometry
SV_DomainLocation DSIn TessCoord N/A Tessellation
SV_PrimitiveID HSIn PrimitiveId N/A Tessellation
PCIn PrimitiveId N/A Tessellation
DsIn PrimitiveId N/A Tessellation
GSIn PrimitiveId N/A Geometry
GSOut PrimitiveId N/A Geometry
PSIn PrimitiveId N/A Geometry
MSOut PrimitiveId N/A

MeshShadingNV

MeshShadingEXT
SV_TessFactor PCOut TessLevelOuter N/A Tessellation
DSIn TessLevelOuter N/A Tessellation
SV_InsideTessFactor PCOut TessLevelInner N/A Tessellation
DSIn TessLevelInner N/A Tessellation
SV_SampleIndex PSIn SampleId N/A SampleRateShading
SV_StencilRef PSOut FragStencilRefEXT N/A StencilExportEXT
SV_Barycentrics PSIn BaryCoord*KHR N/A FragmentBarycentricKHR
SV_RenderTargetArrayIndex GSOut Layer N/A Geometry
PSIn Layer N/A Geometry
MSOut Layer N/A

MeshShadingNV

MeshShadingEXT
SV_ViewportArrayIndex GSOut ViewportIndex N/A MultiViewport
PSIn ViewportIndex N/A MultiViewport
MSOut ViewportIndex N/A

MeshShadingNV

MeshShadingEXT
SV_Coverage PSIn SampleMask N/A Shader
PSOut SampleMask N/A Shader
SV_InnerCoverage PSIn FullyCoveredEXT N/A FragmentFullyCoveredEXT
SV_ViewID VSIn ViewIndex N/A MultiView
HSIn ViewIndex N/A MultiView
DSIn ViewIndex N/A MultiView
GSIn ViewIndex N/A MultiView
PSIn ViewIndex N/A MultiView
MSIn ViewIndex N/A MultiView
SV_ShadingRate VSOut PrimitiveShadingRateKHR N/A FragmentShadingRate
GSOut PrimitiveShadingRateKHR N/A FragmentShadingRate
PSIn ShadingRateKHR N/A FragmentShadingRate
MSOut PrimitiveShadingRateKHR N/A FragmentShadingRate
SV_CullPrimitive MSOut CullPrimitiveEXT N/A ``MeshShadingEXT ``

For entities (function parameters, function return values, struct fields) with the above SV semantic strings attached, SPIR-V variables of the Input/Output storage class will be created. They will have the corresponding SPIR-V Builtin decorations according to the above table.

SV semantic strings not translated into SPIR-V BuiltIn decorations will be handled similarly as non-SV (arbitrary) semantic strings: a SPIR-V variable of the Input/Output storage class will be created for each entity with such semantic string. Then sort all semantic strings according to declaration (the default, or if -fvk-stage-io-order=decl is given) or alphabetical (if -fvk-stage-io-order=alpha is given) order, and assign Location numbers sequentially to the corresponding SPIR-V variables. Note that this means flattening all structs if structs are used as function parameters or returns.

There is an exception to the above rule for SV_Target[N]. It will always be mapped to Location number N.

ClipDistance & CullDistance

Variables decorated with SV_ClipDistanceX can be float or vector of float type. To map them into one float array in the struct, we firstly sort them asecendingly according to X, and then concatenate them tightly. For example,

struct T {
  float clip0: SV_ClipDistance0,
};

struct S {
  float3 clip5: SV_ClipDistance5;
  ...
};

void main(T t, S s, float2 clip2 : SV_ClipDistance2) { ... }

Then we have an float array of size (1 + 2 + 3 =) 6 for ClipDistance, with clip0 at offset 0, clip2 at offset 1, clip5 at offset 3.

Decorating a variable or struct member with the ClipDistance builtin but not requiring the ClipDistance capability is legal as long as we don't read or write the variable or struct member. But as per the way we handle shader entry function, this is not satisfied because we need to read their contents to prepare for the source code entry function call or write back them after the call. So annotating a variable or struct member with SV_ClipDistanceX means requiring the ClipDistance capability in the generated SPIR-V.

Variables decorated with SV_CullDistanceX are mapped similarly as above.

Signature packing

In usual, Vulkan drivers have a limitation of the number of available locations. It varies depending on the device. To avoid the driver crash caused by the limitation, we added an experimental signature packing support using Component decoration (see the Vulkan spec "15.1.5. Component Assignment"). -pack-optimized is the command line option to enable it.

In a high level, for a stage variable that needs M components in N locations e.g., stage variable float3 foo[2] needs 3 components in 2 locations, we find a minimum K where each of N continuous locations in [K, K + N) has M continuous unused Component slots. We create a Location decoration instruction for the stage variable with K and a Component decoration instruction with the first unused component number of the M continuous unused Component slots.

HLSL register and Vulkan binding

In shaders for DirectX, resources are accessed via registers; while in shaders for Vulkan, it is done via descriptor set and binding numbers. The developer can explicitly annotate variables in HLSL to specify descriptor set and binding numbers, or leave it to the compiler to derive implicitly from registers.

Explicit binding number assignment

[[vk::binding(X[, Y])]] can be attached to global variables to specify the descriptor set as Y and binding number as X. The descriptor set number is optional; if missing, it will be zero (If -auto-binding-space N command line option is used, then descriptor set #N will be used instead of descriptor set #0). RW/append/consume structured buffers have associated counters, which will occupy their own Vulkan descriptors. [vk::counter_binding(Z)] can be attached to a RW/append/consume structured buffers to specify the binding number for the associated counter to Z. Note that the set number of the counter is always the same as the main buffer.

Implicit binding number assignment

Without explicit annotations, the compiler will try to deduce descriptor sets and binding numbers in the following way:

If there is :register(xX, spaceY) specified for the given global variable, the corresponding resource will be assigned to descriptor set Y and binding number X, regardless of the register type x. Note that this will cause binding number collision if, say, two resources are of different register type but the same register number. To solve this problem, four command-line options, -fvk-b-shift N M, -fvk-s-shift N M, -fvk-t-shift N M, and -fvk-u-shift N M, are provided to shift by N all binding numbers inferred for register type b, s, t, and u in space M, respectively.

If there is no register specification, the corresponding resource will be assigned to the next available binding number, starting from 0, in descriptor set #0 (If -auto-binding-space N command line option is used, then descriptor set #N will be used instead of descriptor set #0).

If there is no register specification AND -fvk-auto-shift-bindings is specified, then the register type will be automatically identified based on the resource type (according to the following table), and the appropriate shift will automatically be applied according to -fvk-*shift N M.

t - for shader resource views (SRV)
    TEXTURE1D
    TEXTURE1DARRAY
    TEXTURE2D
    TEXTURE2DARRAY
    TEXTURE3D
    TEXTURECUBE
    TEXTURECUBEARRAY
    TEXTURE2DMS
    TEXTURE2DMSARRAY
    STRUCTUREDBUFFER
    BYTEADDRESSBUFFER
    BUFFER
    TBUFFER

s - for samplers
    SAMPLER
    SAMPLER1D
    SAMPLER2D
    SAMPLER3D
    SAMPLERCUBE
    SAMPLERSTATE
    SAMPLERCOMPARISONSTATE

u - for unordered access views (UAV)
    RWBYTEADDRESSBUFFER
    RWSTRUCTUREDBUFFER
    APPENDSTRUCTUREDBUFFER
    CONSUMESTRUCTUREDBUFFER
    RWBUFFER
    RWTEXTURE1D
    RWTEXTURE1DARRAY
    RWTEXTURE2D
    RWTEXTURE2DARRAY
    RWTEXTURE3D

b - for constant buffer views (CBV)
    CBUFFER
    CONSTANTBUFFER

Binding number assignment for resources in cbuffer

Basically, we use the same binding assignment rule described above for a cbuffer, but when a cbuffer contains one or more resources, it is inevitable to use multiple binding numbers for a single cbuffer. For this type of cbuffers, we first assign the next available binding number to the resources. Based the order of the appearance in the cbuffer, a resource that appears early uses a smaller (earlier available) binding number than a resource that appears later. After assigning binding numbers to all resource members, if the cbuffer contains one or more members with non-resource types, it creates a struct for the remaining members and assign the next available binding number to the variable with the struct type.

For example, the binding numbers for the following resources and cbuffers

cbuffer buf0 : register(b0) {
  float4 non_resource0;
};
cbuffer buf1 : register(b4) {
  float4 non_resource1;
};
cbuffer buf2 {
  float4 non_resource2;
  Texture2D resource0;
  SamplerState resource1;
};
cbuffer buf3 : register(b2) {
  SamplerState resource2;
}

will be

  • buf0: 0 because of register(b0)
  • buf1: 4 because of register(b4)
  • resource2: 2 because of register(b2). Note that buf3 is empty without resource2. We do not assign a binding number to an empty struct.
  • resource0: 1 because it is the next available binding number.
  • resource1: 3 because it is the next available binding number.
  • buf2 including only non_resource2: 5 because it is the next available binding number.

Summary

In summary, the compiler essentially assigns binding numbers in three passes.

  • Firstly it handles all declarations with explicit [[vk::binding(X[, Y])]] annotation.
  • Then the compiler processes all remaining declarations with :register(xX, spaceY) annotation, by applying the shift passed in using command-line option -fvk-{b|s|t|u}-shift N M, if provided.
    • If :register assignment is missing and -fvk-auto-shift-bindings is specified, the register type will be automatically detected based on the resource type, and the -fvk-{b|s|t|u}-shift N M will be applied.
  • Finally, the compiler assigns next available binding numbers to the rest in the declaration order.

As an example, for the following code:

struct S { ... };

ConstantBuffer<S> cbuffer1 : register(b0);
Texture2D<float4> texture1 : register(t0);
Texture2D<float4> texture2 : register(t1, space1);
SamplerState      sampler1;
[[vk::binding(3)]]
RWBuffer<float4> rwbuffer1 : register(u5, space2);

If we compile with -fvk-t-shift 10 0 -fvk-t-shift 20 1:

  • rwbuffer1 will take binding #3 in set #0, since explicit binding assignment has precedence over the rest.
  • cbuffer1 will take binding #0 in set #0, since that's what deduced from the register assignment, and there is no shift requested from command line.
  • texture1 will take binding #10 in set #0, and texture2 will take binding #21 in set #1, since we requested an 10 shift on t-type registers.
  • sampler1 will take binding 1 in set #0, since that's the next available binding number in set #0.

HLSL global variables and Vulkan binding

As mentioned above, all global externally-visible non-resource-type stand-alone variables will be collected into a cbuffer named $Globals. By default, the $Globals cbuffer is placed in descriptor set #0, and the binding number would be the next available binding number in that set. Meaning, the binding number depends on where the very first global variable is in the code.

Example 1:

float4 someColors;
  // $Globals cbuffer placed at DescriptorSet #0, Binding #0
Texture2D<float4> texture1;
  // texture1         placed at DescriptorSet #0, Binding #1

Example 2:

Texture2D<float4> texture1;
  // texture1         placed at DescriptorSet #0, Binding #0
float4 someColors;
  // $Globals cbuffer placed at DescriptorSet #0, Binding #1

In order provide more control over the descriptor set and binding number of the $Globals cbuffer, you can use the -fvk-bind-globals B S command line option, which will place this cbuffer at descriptor set S, and binding number B.

Example 3: (compiled with -fvk-bind-globals 2 1)

Texture2D<float4> texture1;
  // texture1         placed at DescriptorSet #0, Binding #0
float4 someColors;
  // $Globals cbuffer placed at DescriptorSet #1, Binding #2

Note that if the developer chooses to use this command line option, it is their responsibility to provide proper numbers and avoid binding overlaps.

HLSL Expressions

Unless explicitly noted, matrix per-element operations will be conducted on each component vector and then collected into the result matrix. The following sections lists the SPIR-V opcodes for scalars and vectors.

Arithmetic operators

Arithmetic operators (+, -, *, /, %) are translated into their corresponding SPIR-V opcodes according to the following table.

  (Vector of) Signed Integers (Vector of) Unsigned Integers (Vector of) Floats
+ OpIAdd OpFAdd
- OpISub OpFSub
* OpIMul OpFMul
/ OpSDiv OpUDiv OpFDiv
% OpSRem OpUMod OpFRem

Note that for modulo operation, SPIR-V has two sets of instructions: Op*Rem and Op*Mod. For Op*Rem, the sign of a non-0 result comes from the first operand; while for Op*Mod, the sign of a non-0 result comes from the second operand. HLSL doc does not mandate which set of instructions modulo operations should be translated into; it only says "the % operator is defined only in cases where either both sides are positive or both sides are negative." So technically it's undefined behavior to use the modulo operation with operands of different signs. But considering HLSL's C heritage and the behavior of Clang frontend, we translate modulo operators into Op*Rem (there is no OpURem).

For multiplications of float vectors and float scalars, the dedicated SPIR-V operation OpVectorTimesScalar will be used. Similarly, for multiplications of float matrices and float scalars, OpMatrixTimesScalar will be generated.

Bitwise operators

Bitwise operators (~, &, |, ^, <<, >>) are translated into their corresponding SPIR-V opcodes according to the following table.

  (Vector of) Signed Integers (Vector of) Unsigned Integers
~ OpNot
& OpBitwiseAnd
| OpBitwiseOr
^ OpBitwiseXor
<< OpShiftLeftLogical
>> OpShiftRightArithmetic OpShiftRightLogical

Note that for <</>>, the right hand side will be culled: only the n - 1 least significant bits are considered, where n is the bitwidth of the left hand side.

Comparison operators

Comparison operators (<, <=, >, >=, ==, !=) are translated into their corresponding SPIR-V opcodes according to the following table.

  (Vector of) Signed Integers (Vector of) Unsigned Integers (Vector of) Floats
< OpSLessThan OpULessThan OpFOrdLessThan
<= OpSLessThanEqual OpULessThanEqual OpFOrdLessThanEqual
> OpSGreaterThan OpUGreaterThan OpFOrdGreaterThan
>= OpSGreaterThanEqual OpUGreaterThanEqual OpFOrdGreaterThanEqual
== OpIEqual OpFOrdEqual
!= OpINotEqual OpFOrdNotEqual

Note that for comparison of (vectors of) floats, SPIR-V has two sets of instructions: OpFOrd*, OpFUnord*. We translate into OpFOrd* ones.

Boolean math operators

Boolean match operators (&&, ||, ?:) are translated into their corresponding SPIR-V opcodes according to the following table.

  (Vector of) Booleans
&& OpLogicalAnd
|| OpLogicalOr
?: OpSelect

Please note that "unlike short-circuit evaluation of &&, ||, and ?: in C, HLSL expressions never short-circuit an evaluation because they are vector operations. All sides of the expression are always evaluated."

Unary operators

For unary operators:

  • ! is translated into OpLogicalNot. Parsing will gurantee the operands are of boolean types by inserting necessary casts.
  • + requires no additional SPIR-V instructions.
  • - is translated into OpSNegate and OpFNegate for (vectors of) integers and floats, respectively.

Casts

Casting between (vectors) of scalar types is translated according to the following table:

From \ To Bool SInt UInt Float
Bool no-op select between one and zero
SInt compare with zero no-op OpBitcast OpConvertSToF
UInt OpBitcast no-op OpConvertUToF
Float OpConvertFToS OpConvertFToU no-op

It is also feasible in HLSL to cast a float matrix to another float matrix with a smaller size. This is known as matrix truncation cast. For instance, the following code casts a 3x4 matrix into a 2x3 matrix.

float3x4 m = { 1,  2,  3, 4,
               5,  6,  7, 8,
               9, 10, 11, 12 };

float2x3 a = (float2x3)m;

Such casting takes the upper-left most corner of the original matrix to generate the result. In the above example, matrix a will have 2 rows, with 3 columns each. First row will be 1, 2, 3 and the second row will be 5, 6, 7.

Indexing operator

The [] operator can also be used to access elements in a matrix or vector. A matrix whose row and/or column count is 1 will be translated into a vector or scalar. If a variable is used as the index for the dimension whose count is 1, that variable will be ignored in the generated SPIR-V code. This is because out-of-bound indexing triggers undefined behavior anyway. For example, for a 1xN matrix mat, mat[index][0] will be translated into OpAccessChain ... %mat %uint_0. Similarly, variable index into a size 1 vector will also be ignored and the only element will be always returned.

Assignment operators

Assigning to struct object may involve decomposing the source struct object and assign each element separately and recursively. This happens when the source struct object is of different memory layout from the destination struct object. For example, for the following source code:

struct S {
  float    a;
  float2   b;
  float2x3 c;
};

    ConstantBuffer<S> cbuf;
RWStructuredBuffer<S> sbuf;

...
sbuf[0] = cbuf[0];
...

We need to assign each element because ConstantBuffer and RWStructuredBuffer has different memory layout.

HLSL Control Flows

This section lists how various HLSL control flows are mapped.

Switch statement

HLSL switch statements are translated into SPIR-V using:

  • OpSwitch: if (all case values are integer literals or constant integer variables) and (no attribute or the forcecase attribute is specified)
  • A series of if statements: for all other scenarios (e.g., when flatten, branch, or call attribute is specified)

Loops (for, while, do)

HLSL for statements, while statements, and do statements are translated into SPIR-V by constructing all necessary basic blocks and using OpLoopMerge to organize as structured loops.

The HLSL attributes for these statements are translated into SPIR-V loop control masks according to the following table:

HLSL loop attribute SPIR-V Loop Control Mask
unroll(x) Unroll
loop DontUnroll
fastopt DontUnroll
allow_uav_condition Currently Unimplemented

HLSL Functions

All functions reachable from the entry-point function will be translated into SPIR-V code. Functions not reachable from the entry-point function will be ignored.

Entry function wrapper

HLSL entry functions takes in parameters and returns values. These parameters and return values can have semantics attached or if they are struct type, the struct fields can have semantics attached. However, in Vulkan, the entry function must be of the void(void) signature. To handle this difference, for a given entry function main, we will emit a wrapper function for it.

The wrapper function will take the name of the source code entry function, while the source code entry function will have its name prefixed with "src.". The wrapper function reads in stage input/builtin variables created according to semantics and groups them into composites meeting the requirements of the source code entry point. Then the wrapper calls the source code entry point. The return value is extracted and components of it will be written to stage output/builtin variables created according to semantics. For example:

// HLSL source code

struct S {
  bool a : A;
  uint2 b: B;
  float2x3 c: C;
};

struct T {
  S x;
  int y: D;
};

T main(T input) {
  return input;
}
; SPIR-V code

%in_var_A = OpVariable %_ptr_Input_bool Input
%in_var_B = OpVariable %_ptr_Input_v2uint Input
%in_var_C = OpVariable %_ptr_Input_mat2v3float Input
%in_var_D = OpVariable %_ptr_Input_int Input

%out_var_A = OpVariable %_ptr_Output_bool Output
%out_var_B = OpVariable %_ptr_Output_v2uint Output
%out_var_C = OpVariable %_ptr_Output_mat2v3float Output
%out_var_D = OpVariable %_ptr_Output_int Output

; Wrapper function starts

%main    = OpFunction %void None ...
...      = OpLabel

%param_var_input = OpVariable %_ptr_Function_T Function

; Load stage input variables and group into the expected composite

%inA = OpLoad %bool %in_var_A
%inB = OpLoad %v2uint %in_var_B
%inC = OpLoad %mat2v3float %in_var_C
%inS = OpCompositeConstruct %S %inA %inB %inC
%inD = OpLoad %int %in_var_D
%inT = OpCompositeConstruct %T %inS %inD
       OpStore %param_var_input %inT

%ret = OpFunctionCall %T %src_main %param_var_input

; Extract component values from the composite and store into stage output variables

%outS = OpCompositeExtract %S %ret 0
%outA = OpCompositeExtract %bool %outS 0
        OpStore %out_var_A %outA
%outB = OpCompositeExtract %v2uint %outS 1
        OpStore %out_var_B %outB
%outC = OpCompositeExtract %mat2v3float %outS 2
        OpStore %out_var_C %outC
%outD = OpCompositeExtract %int %ret 1
        OpStore %out_var_D %outD

OpReturn
OpFunctionEnd

; Source code entry point starts

%src_main = OpFunction %T None ...

In this way, we can concentrate all stage input/output/builtin variable manipulation in the wrapper function and handle the source code entry function just like other nomal functions.

Function parameter

For a function f which has a parameter of type T, the generated SPIR-V signature will use type T* for the parameter. At every call site of f, additional local variables will be allocated to hold the actual arguments. The local variables are passed in as direct function arguments. For example:

// HLSL source code

float4 f(float a, int b) { ... }

void caller(...) {
  ...
  float4 result = f(...);
  ...
}
; SPIR-V code

              ...
%i32PtrType = OpTypePointer Function %int
%f32PtrType = OpTypePointer Function %float
    %fnType = OpTypeFunction %v4float %f32PtrType %i32PtrType
              ...

         %f = OpFunction %v4float None %fnType
         %a = OpFunctionParameter %f32PtrType
         %b = OpFunctionParameter %i32PtrType
              ...

    %caller = OpFunction ...
              ...
   %aAlloca = OpVariable %_ptr_Function_float Function
   %bAlloca = OpVariable %_ptr_Function_int Function
              ...
              OpStore %aAlloca ...
              OpStore %bAlloca ...
    %result = OpFunctioncall %v4float %f %aAlloca %bAlloca
              ...

This approach gives us unified handling of function parameters and local variables: both of them are accessed via load/store instructions.

Intrinsic functions

The following intrinsic HLSL functions have no direct SPIR-V opcode or GLSL extended instruction mapping, so they are handled with additional steps:

  • dot : performs dot product of two vectors, each containing floats or integers. If the two parameters are vectors of floats, we use SPIR-V's OpDot instruction to perform the translation. If the two parameters are vectors of integers, we multiply corresponding vector elements using OpIMul and accumulate the results using OpIAdd to compute the dot product.
  • mul: performs multiplications. Each argument may be a scalar, vector, or matrix. Depending on the argument type, this will be translated into one of the multiplication instructions.
  • all: returns true if all components of the given scalar, vector, or matrix are true. Performs conversions to boolean where necessary. Uses SPIR-V OpAll for scalar arguments and vector arguments. For matrix arguments, performs OpAll on each row, and then again on the vector containing the results of all rows.
  • any: returns true if any component of the given scalar, vector, or matrix is true. Performs conversions to boolean where necessary. Uses SPIR-V OpAny for scalar arguments and vector arguments. For matrix arguments, performs OpAny on each row, and then again on the vector containing the results of all rows.
  • asfloat: converts the component type of a scalar/vector/matrix from float, uint, or int into float. Uses OpBitcast. This method currently does not support taking non-float matrix arguments.
  • asint: converts the component type of a scalar/vector/matrix from float or uint into int. Uses OpBitcast. This method currently does not support conversion into integer matrices.
  • asuint: converts the component type of a scalar/vector/matrix from float or int into uint. Uses OpBitcast. This method currently does not support
  • asuint: Converts a double into two 32-bit unsigned integers. Uses SPIR-V OpBitCast.
  • asdouble: Converts two 32-bit unsigned integers into a double, or four 32-bit unsigned integers into two doubles. Uses SPIR-V OpVectorShuffle and OpBitCast. conversion into unsigned integer matrices.
  • isfinite : Determines if the specified value is finite. Since OpIsFinite requires the Kernel capability, translation is done using OpIsNan and OpIsInf. A given value is finite iff it is not NaN and not infinite.
  • clip: Discards the current pixel if the specified value is less than zero. Uses conditional control flow as well as SPIR-V OpKill.
  • rcp: Calculates a fast, approximate, per-component reciprocal. Uses SIR-V OpFDiv.
  • lit: Returns a lighting coefficient vector. This vector is a float4 with components of (ambient, diffuse, specular, 1). How diffuse and specular are calculated are explained here.
  • D3DCOLORtoUBYTE4: Converts a floating-point, 4D vector set by a D3DCOLOR to a UBYTE4. This is achieved by performing int4(input.zyxw * 255.002) using SPIR-V OpVectorShuffle, OpVectorTimesScalar, and OpConvertFToS, respectively.
  • dst: Calculates a distance vector. The resulting vector, dest, has the following specifications: dest.x = 1.0, dest.y = src0.y * src1.y, dest.z = src0.z, and dest.w = src1.w. Uses SPIR-V OpCompositeExtract and OpFMul.

Using SPIR-V opcode

The following intrinsic HLSL functions have direct SPIR-V opcodes for them:

HLSL Intrinsic Function SPIR-V Opcode
AllMemoryBarrier OpMemoryBarrier
AllMemoryBarrierWithGroupSync OpControlBarrier
countbits OpBitCount
DeviceMemoryBarrier OpMemoryBarrier
DeviceMemoryBarrierWithGroupSync OpControlBarrier
ddx OpDPdx
ddy OpDPdy
ddx_coarse OpDPdxCoarse
ddy_coarse OpDPdyCoarse
ddx_fine OpDPdxFine
ddy_fine OpDPdyFine
fmod OpFRem
fwidth OpFwidth
GroupMemoryBarrier OpMemoryBarrier
GroupMemoryBarrierWithGroupSync OpControlBarrier
InterlockedAdd OpAtomicIAdd
InterlockedAnd OpAtomicAnd
InterlockedOr OpAtomicOr
InterlockedXor OpAtomicXor
InterlockedMin OpAtomicUMin/OpAtomicSMin
InterlockedMax OpAtomicUMax/OpAtomicSMax
InterlockedExchange OpAtomicExchange
InterlockedCompareExchange OpAtomicCompareExchange
InterlockedCompareStore OpAtomicCompareExchange
isnan OpIsNan
isInf OpIsInf
reversebits OpBitReverse
transpose OpTranspose
CheckAccessFullyMapped OpImageSparseTexelsResident

Using GLSL extended instructions

The following intrinsic HLSL functions are translated using their equivalent instruction in the GLSL extended instruction set.

HLSL Intrinsic Function GLSL Extended Instruction
abs SAbs/FAbs
acos Acos
asin Asin
atan Atan
atan2 Atan2
ceil Ceil
clamp SClamp/UClamp/FClamp
cos Cos
cosh Cosh
cross Cross
degrees Degrees
distance Distance
radians Radian
determinant Determinant
exp Exp
exp2 exp2
f16tof32 UnpackHalf2x16
f32tof16 PackHalf2x16
faceforward FaceForward
firstbithigh FindSMsb / FindUMsb
firstbitlow FindILsb
floor Floor
fma Fma
frac Fract
frexp FrexpStruct
ldexp Ldexp
length Length
lerp FMix
log Log
log10 Log2 (scaled by 1/log2(10))
log2 Log2
mad Fma
max SMax/UMax/NMax/FMax
min SMin/UMin/NMin/FMin
modf ModfStruct
normalize Normalize
pow Pow
reflect Reflect
refract Refract
round RoundEven
rsqrt InverseSqrt
saturate FClamp
sign SSign/FSign
sin Sin
sincos Sin and Cos
sinh Sinh
smoothstep SmoothStep
sqrt Sqrt
step Step
tan Tan
tanh Tanh
trunc Trunc

Note on NMax,Nmin,FMax & FMin:

This compiler supports the --ffinite-math-only option, which allows assuming non-NaN parameters to some operations. min & max intrinsics will by default generate NMin & NMax instructions, but if this option is enabled, FMin & FMax can be generated instead.

Synchronization intrinsics

Synchronization intrinsics are translated into OpMemoryBarrier (for those non-WithGroupSync variants) or OpControlBarrier (for those WithGroupSync variants) instructions with parameters:

HLSL SPIR-V SPIR-V Memory Semantics
Intrinsic Memory Scope Image Uniform Workgroup AcquireRelease
AllMemoryBarrier Device ✓ ✓ ✓ ✓
DeviceMemoryBarrier Device ✓ ✓   ✓
GroupMemoryBarrier Workgroup     ✓ ✓

For the *WithGroupSync intrinsics, SPIR-V memory scope and semantics are the same as their counterparts in the above. They have an additional execution scope:

HLSL Intrinsic SPIR-V Execution Scope
AllMemoryBarrierWithGroupSync Workgroup
DeviceMemoryBarrierWithGroupSync Workgroup
GroupMemoryBarrierWithGroupSync Workgroup

HLSL OO features

A HLSL struct/class member method is translated into a normal SPIR-V function, whose signature has an additional first parameter for the struct/class called upon. Every calling site of the method is generated to pass in the object as the first argument.

HLSL struct/class static member variables are translated into SPIR-V variables in the Private storage class.

HLSL Methods

This section lists how various HLSL methods are mapped.

Buffers

Buffer

.Load()

Since Buffers are represented as OpTypeImage with Sampled set to 1 (meaning to be used with a sampler), OpImageFetch is used to perform this operation. The return value of OpImageFetch is always a four-component vector; so proper additional instructions are generated to truncate the vector and return the desired number of elements. If an output unsigned integer status argument is present, OpImageSparseFetch is used instead. The resulting SPIR-V Residency Code will be written to status.

operator[]

Handled similarly as .Load().

.GetDimensions()

Since Buffers are represented as OpTypeImage with dimension of Buffer, OpImageQuerySize is used to perform this operation.

RWBuffer

.Load()

Since RWBuffers are represented as OpTypeImage with Sampled set to 2 (meaning to be used without a sampler), OpImageRead is used to perform this operation. If an output unsigned integer status argument is present, OpImageSparseRead is used instead. The resulting SPIR-V Residency Code will be written to status.

operator[]

Using operator[] for reading is handled similarly as .Load(), while for writing, the OpImageWrite instruction is generated.

.GetDimensions()

Since RWBuffers are represented as OpTypeImage with dimension of Buffer, OpImageQuerySize is used to perform this operation.

StructuredBuffer and RWStructuredBuffer

.GetDimensions()

Since StructuredBuffers/RWStructuredBuffers are represented as a struct with one member that is a runtime array of structures, OpArrayLength is invoked on the runtime array in order to find the dimension.

ByteAddressBuffer

.GetDimensions()

Since ByteAddressBuffers are represented as a struct with one member that is a runtime array of unsigned integers, OpArrayLength is invoked on the runtime array in order to find the number of unsigned integers. This is then multiplied by 4 to find the number of bytes.

.Load(), .Load2(), .Load3(), .Load4()

ByteAddressBuffers are represented as a struct with one member that is a runtime array of unsigned integers. The address argument passed to the function is first divided by 4 in order to find the offset into the array (because each array element is 4 bytes). The SPIR-V OpAccessChain instruction is then used to access that offset, and OpLoad is used to load a 32-bit unsigned integer. For Load2, Load3, and Load4, this is done 2, 3, and 4 times, respectively. Each time the word offset is incremented by 1 before performing OpAccessChain. After all OpLoad operations are performed, a vector is constructed with all the resulting values.

RWByteAddressBuffer

.GetDimensions()

Since RWByteAddressBuffers are represented as a struct with one member that is a runtime array of unsigned integers, OpArrayLength is invoked on the runtime array in order to find the number of unsigned integers. This is then multiplied by 4 to find the number of bytes.

.Load(), .Load2(), .Load3(), .Load4()

RWByteAddressBuffers are represented as a struct with one member that is a runtime array of unsigned integers. The address argument passed to the function is first divided by 4 in order to find the offset into the array (because each array element is 4 bytes). The SPIR-V OpAccessChain instruction is then used to access that offset, and OpLoad is used to load a 32-bit unsigned integer. For Load2, Load3, and Load4, this is done 2, 3, and 4 times, respectively. Each time the word offset is incremented by 1 before performing OpAccessChain. After all OpLoad operations are performed, a vector is constructed with all the resulting values.

.Store(), .Store2(), .Store3(), .Store4()

RWByteAddressBuffers are represented as a struct with one member that is a runtime array of unsigned integers. The address argument passed to the function is first divided by 4 in order to find the offset into the array (because each array element is 4 bytes). The SPIR-V OpAccessChain instruction is then used to access that offset, and OpStore is used to store a 32-bit unsigned integer. For Store2, Store3, and Store4, this is done 2, 3, and 4 times, respectively. Each time the word offset is incremented by 1 before performing OpAccessChain.

.Interlocked*()
HLSL Intrinsic Method SPIR-V Opcode
.InterlockedAdd() OpAtomicIAdd
.InterlockedAnd() OpAtomicAnd
.InterlockedOr() OpAtomicOr
.InterlockedXor() OpAtomicXor
.InterlockedMin() OpAtomicUMin/OpAtomicSMin
.InterlockedMax() OpAtomicUMax/OpAtomicSMax
.InterlockedExchange() OpAtomicExchange
.InterlockedCompareExchange() OpAtomicCompareExchange
.InterlockedCompareStore() OpAtomicCompareExchange

AppendStructuredBuffer

.Append()

The associated counter number will be increased by 1 using OpAtomicIAdd. The return value of OpAtomicIAdd, which is the original count number, will be used as the index for storing the new element. E.g., for buf.Append(vec):

%counter = OpAccessChain %_ptr_Uniform_int %counter_var_buf %uint_0
  %index = OpAtomicIAdd %uint %counter %uint_1 %uint_0 %uint_1
    %ptr = OpAccessChain %_ptr_Uniform_v4float %buf %uint_0 %index
    %val = OpLoad %v4float %vec
           OpStore %ptr %val
.GetDimensions()

Since AppendStructuredBuffers are represented as a struct with one member that is a runtime array, OpArrayLength is invoked on the runtime array in order to find the number of elements. The stride is also calculated based on GLSL std430 as explained above.

ConsumeStructuredBuffer

.Consume()

The associated counter number will be decreased by 1 using OpAtomicISub. The return value of OpAtomicISub minus 1, which is the new count number, will be used as the index for reading the new element. E.g., for buf.Consume(vec):

%counter = OpAccessChain %_ptr_Uniform_int %counter_var_buf %uint_0
   %prev = OpAtomicISub %uint %counter %uint_1 %uint_0 %uint_1
  %index = OpISub %uint %prev %uint_1
    %ptr = OpAccessChain %_ptr_Uniform_v4float %buf %uint_0 %index
    %val = OpLoad %v4float %vec
           OpStore %ptr %val
.GetDimensions()

Since ConsumeStructuredBuffers are represented as a struct with one member that is a runtime array, OpArrayLength is invoked on the runtime array in order to find the number of elements. The stride is also calculated based on GLSL std430 as explained above.

Read-only textures

Methods common to all texture types are explained in the "common texture methods" section. Methods unique to a specific texture type is explained in the section for that texture type.

Common texture methods

.Sample(sampler, location[, offset][, clamp][, Status])

Not available to Texture2DMS and Texture2DMSArray.

The OpImageSampleImplicitLod instruction is used to translate .Sample() since texture types are represented as OpTypeImage. An OpSampledImage is created based on the sampler passed to the function. The resulting sampled image and the location passed to the function are used as arguments to OpImageSampleImplicitLod, with the optional offset tranlated into addtional SPIR-V image operands ConstOffset or Offset on it. The optional clamp argument will be translated to the MinLod image operand.

If an output unsigned integer status argument is present, OpImageSparseSampleImplicitLod is used instead. The resulting SPIR-V Residency Code will be written to status.

.SampleLevel(sampler, location, lod[, offset][, Status])

Not available to Texture2DMS and Texture2DMSArray.

The OpImageSampleExplicitLod instruction is used to translate this method. An OpSampledImage is created based on the sampler passed to the function. The resulting sampled image and the location passed to the function are used as arguments to OpImageSampleExplicitLod. The lod passed to the function is attached to the instruction as an SPIR-V image operands Lod. The optional offset is also tranlated into addtional SPIR-V image operands ConstOffset or Offset on it.

If an output unsigned integer status argument is present, OpImageSparseSampleExplicitLod is used instead. The resulting SPIR-V Residency Code will be written to status.

.SampleGrad(sampler, location, ddx, ddy[, offset][, clamp][, Status])

Not available to Texture2DMS and Texture2DMSArray.

Similarly to .SampleLevel, the ddx and ddy parameter are attached to the OpImageSampleExplicitLod instruction as an SPIR-V image operands Grad. The optional clamp argument will be translated into the MinLod image operand.

If an output unsigned integer status argument is present, OpImageSparseSampleExplicitLod is used instead. The resulting SPIR-V Residency Code will be written to status.

.SampleBias(sampler, location, bias[, offset][, clamp][, Status])

Not available to Texture2DMS and Texture2DMSArray.

The translation is similar to .Sample(), with the bias parameter attached to the OpImageSampleImplicitLod instruction as an SPIR-V image operands Bias.

If an output unsigned integer status argument is present, OpImageSparseSampleImplicitLod is used instead. The resulting SPIR-V Residency Code will be written to status.

.SampleCmp(sampler, location, comparator[, offset][, clamp][, Status])

Not available to Texture3D, Texture2DMS, and Texture2DMSArray.

The translation is similar to .Sample(), but the OpImageSampleDrefImplicitLod instruction are used.

If an output unsigned integer status argument is present, OpImageSparseSampleDrefImplicitLod is used instead. The resulting SPIR-V Residency Code will be written to status.

.SampleCmpLevelZero(sampler, location, comparator[, offset][, Status])

Not available to Texture3D, Texture2DMS, and Texture2DMSArray.

The translation is similar to .Sample(), but the OpImageSampleDrefExplicitLod instruction are used, with the additional Lod image operands set to 0.0.

If an output unsigned integer status argument is present, OpImageSparseSampleDrefExplicitLod is used instead. The resulting SPIR-V Residency Code will be written to status.

.Gather()

Available to Texture2D, Texture2DArray, TextureCube, and TextureCubeArray.

The translation is similar to .Sample(), but the OpImageGather instruction is used, with component setting to 0.

If an output unsigned integer status argument is present, OpImageSparseGather is used instead. The resulting SPIR-V Residency Code will be written to status.

.GatherRed(), .GatherGreen(), .GatherBlue(), .GatherAlpha()

Available to Texture2D, Texture2DArray, TextureCube, and TextureCubeArray.

The OpImageGather instruction is used to translate these functions, with component setting to 0, 1, 2, and 3 respectively.

There are a few overloads for these functions:

  • For those overloads taking 4 offset parameters, those offset parameters will be conveyed as an additional ConstOffsets image operands to the instruction if those offset parameters are all constants. Otherwise, 4 separate OpImageGather instructions will be emitted to get each texel from each offset, using the Offset image operands.
  • For those overloads with the status parameter, OpImageSparseGather is used instead, and the resulting SPIR-V Residency Code will be written to status.
.GatherCmp()

Available to Texture2D, Texture2DArray, TextureCube, and TextureCubeArray.

The translation is similar to .Sample(), but the OpImageDrefGather instruction is used.

For the overload with the output unsigned integer status argument, OpImageSparseDrefGather is used instead. The resulting SPIR-V Residency Code will be written to status.

.GatherCmpRed()

Available to Texture2D, Texture2DArray, TextureCube, and TextureCubeArray.

The translation is the same as .GatherCmp().

.Load(location[, sampleIndex][, offset])

The OpImageFetch instruction is used for translation because texture types are represented as OpTypeImage. The last element in the location parameter will be used as arguments to the Lod SPIR-V image operand attached to the OpImageFetch instruction, and the rest are used as the coordinate argument to the instruction. offset is handled similarly to .Sample(). The return value of OpImageFetch is always a four-component vector; so proper additional instructions are generated to truncate the vector and return the desired number of elements.

For the overload with the output unsigned integer status argument, OpImageSparseFetch is used instead. The resulting SPIR-V Residency Code will be written to status.

operator[]

Handled similarly as .Load().

.mips[lod][position]

Not available to TextureCube, TextureCubeArray, Texture2DMS, and Texture2DMSArray.

This method is translated into the OpImageFetch instruction. The lod parameter is attached to the instruction as the parameter to the Lod SPIR-V image operands. The position parameter are used as the coordinate to the instruction directly.

.CalculateLevelOfDetail() and .CalculateLevelOfDetailUnclamped()

Not available to Texture2DMS and Texture2DMSArray.

Since texture types are represented as OpTypeImage, the OpImageQueryLod instruction is used for translation. An OpSampledImage is created based on the SamplerState passed to the function. The resulting sampled image and the coordinate passed to the function are used to invoke OpImageQueryLod. The result of OpImageQueryLod is a float2. The first element contains the mipmap array layer. The second element contains the unclamped level of detail.

Texture1D

.GetDimensions(width) or .GetDimensions(MipLevel, width, NumLevels)

Since Texture1D is represented as OpTypeImage, the OpImageQuerySizeLod instruction is used for translation. If a MipLevel argument is passed to GetDimensions, it will be used as the Lod parameter of the query instruction. Otherwise, Lod of 0 be used.

Texture1DArray

.GetDimensions(width, elements) or .GetDimensions(MipLevel, width, elements, NumLevels)

Since Texture1DArray is represented as OpTypeImage, the OpImageQuerySizeLod instruction is used for translation. If a MipLevel argument is present, it will be used as the Lod parameter of the query instruction. Otherwise, Lod of 0 be used.

Texture2D

.GetDimensions(width, height) or .GetDimensions(MipLevel, width, height, NumLevels)

Since Texture2D is represented as OpTypeImage, the OpImageQuerySizeLod instruction is used for translation. If a MipLevel argument is present, it will be used as the Lod parameter of the query instruction. Otherwise, Lod of 0 be used.

Texture2DArray

.GetDimensions(width, height, elements) or .GetDimensions(MipLevel, width, height, elements, NumLevels)

Since Texture2DArray is represented as OpTypeImage, the OpImageQuerySizeLod instruction is used for translation. If a MipLevel argument is present, it will be used as the Lod parameter of the query instruction. Otherwise, Lod of 0 be used.

Texture3D

.GetDimensions(width, height, depth) or .GetDimensions(MipLevel, width, height, depth, NumLevels)

Since Texture3D is represented as OpTypeImage, the OpImageQuerySizeLod instruction is used for translation. If a MipLevel argument is present, it will be used as the Lod parameter of the query instruction. Otherwise, Lod of 0 be used.

Texture2DMS

.sample[sample][position]

This method is translated into the OpImageFetch instruction. The sample parameter is attached to the instruction as the parameter to the Sample SPIR-V image operands. The position parameter are used as the coordinate to the instruction directly.

.GetDimensions(width, height, numSamples)

Since Texture2DMS is represented as OpTypeImage with MS of 1, the OpImageQuerySize instruction is used to get the width and the height. Furthermore, OpImageQuerySamples is used to get the numSamples.

.GetSamplePosition(index)

There are no direct mapping SPIR-V instructions for this method. Right now, it is translated into the SPIR-V code for the following HLSL source code:

// count is the number of samples in the Texture2DMS(Array)
// index is the index of the sample we are trying to get the position

static const float2 pos2[] = {
    { 4.0/16.0,  4.0/16.0 }, {-4.0/16.0, -4.0/16.0 },
};

static const float2 pos4[] = {
    {-2.0/16.0, -6.0/16.0 }, { 6.0/16.0, -2.0/16.0 }, {-6.0/16.0,  2.0/16.0 }, { 2.0/16.0,  6.0/16.0 },
};

static const float2 pos8[] = {
    { 1.0/16.0, -3.0/16.0 }, {-1.0/16.0,  3.0/16.0 }, { 5.0/16.0,  1.0/16.0 }, {-3.0/16.0, -5.0/16.0 },
    {-5.0/16.0,  5.0/16.0 }, {-7.0/16.0, -1.0/16.0 }, { 3.0/16.0,  7.0/16.0 }, { 7.0/16.0, -7.0/16.0 },
};

static const float2 pos16[] = {
    { 1.0/16.0,  1.0/16.0 }, {-1.0/16.0, -3.0/16.0 }, {-3.0/16.0,  2.0/16.0 }, { 4.0/16.0, -1.0/16.0 },
    {-5.0/16.0, -2.0/16.0 }, { 2.0/16.0,  5.0/16.0 }, { 5.0/16.0,  3.0/16.0 }, { 3.0/16.0, -5.0/16.0 },
    {-2.0/16.0,  6.0/16.0 }, { 0.0/16.0, -7.0/16.0 }, {-4.0/16.0, -6.0/16.0 }, {-6.0/16.0,  4.0/16.0 },
    {-8.0/16.0,  0.0/16.0 }, { 7.0/16.0, -4.0/16.0 }, { 6.0/16.0,  7.0/16.0 }, {-7.0/16.0, -8.0/16.0 },
};

float2 position = float2(0.0f, 0.0f);

if (count == 2) {
    position = pos2[index];
} else if (count == 4) {
    position = pos4[index];
} else if (count == 8) {
    position = pos8[index];
} else if (count == 16) {
    position = pos16[index];
}

From the above, it's clear that the current implementation only supports standard sample settings, i.e., with 1, 2, 4, 8, or 16 samples. For other cases, the implementation will just return (float2)0.

Texture2DMSArray

.sample[sample][position]

This method is translated into the OpImageFetch instruction. The sample parameter is attached to the instruction as the parameter to the Sample SPIR-V image operands. The position parameter are used as the coordinate to the instruction directly.

.GetDimensions(width, height, elements, numSamples)

Since Texture2DMS is represented as OpTypeImage with MS of 1, the OpImageQuerySize instruction is used to get the width, the height, and the elements. Furthermore, OpImageQuerySamples is used to get the numSamples.

.GetSamplePosition(index)

Similar to Texture2D.

TextureCube

TextureCubeArray

Read-write textures

Methods common to all texture types are explained in the "common texture methods" section. Methods unique to a specific texture type is explained in the section for that texture type.

Common texture methods

.Load()

Since read-write texture types are represented as OpTypeImage with Sampled set to 2 (meaning to be used without a sampler), OpImageRead is used to perform this operation.

For the overload with the output unsigned integer status argument, OpImageSparseRead is used instead. The resulting SPIR-V Residency Code will be written to status.

operator[]

Using operator[] for reading is handled similarly as .Load(), while for writing, the OpImageWrite instruction is generated.

RWTexture1D

.GetDimensions(width)

The OpImageQuerySize instruction is used to find the width.

RWTexture1DArray

.GetDimensions(width, elements)

The OpImageQuerySize instruction is used to get a uint2. The first element is the width, and the second is the elements.

RWTexture2D

.GetDimensions(width, height)

The OpImageQuerySize instruction is used to get a uint2. The first element is the width, and the second element is the height.

RWTexture2DArray

.GetDimensions(width, height, elements)

The OpImageQuerySize instruction is used to get a uint3. The first element is the width, the second element is the height, and the third is the elements.

RWTexture3D

.GetDimensions(width, height, depth)

The OpImageQuerySize instruction is used to get a uint3. The first element is the width, the second element is the height, and the third element is the depth.

HLSL Shader Stages

Hull Shaders

Hull shaders corresponds to Tessellation Control Shaders (TCS) in Vulkan. This section describes how Hull shaders are translated to SPIR-V for Vulkan.

Hull Entry Point Attributes

The following HLSL attributes are attached to the main entry point of hull shaders and are translated to SPIR-V execution modes according to the table below:

HLSL Attribute value SPIR-V Execution Mode
domain quad Quads
tri Triangles
isoline Isoline
partitioning integer SpacingEqual
fractional_even SpacingFractionalEven
fractional_odd SpacingFractionalOdd
pow2 N/A
outputtopology point PointMode
line N/A
triangle_cw VertexOrderCw
triangle_ccw VertexOrderCcw
outputcontrolpoints n OutputVertices n

The patchconstfunc attribute does not have a direct equivalent in SPIR-V. It specifies the name of the Patch Constant Function. This function is run only once per patch. This is further described below.

InputPatch and OutputPatch

Both of InputPatch<T, N> and OutputPatch<T, N> are translated to an array of constant size N where each element is of type T.

InputPatch can be passed to the Hull shader main entry function as well as the patch constant function. This would include information about each of the N vertices that are input to the tessellation control shader.

OutputPatch is an array containing N elements (where N is the number of output vertices). Each element of the array is the hull shader output for each output vertex. For example, each element of OutputPatch<HSOutput, 3> is each output value of the hull shader function for each SV_OutputControlPointID. It is shared between threads i.e., in the patch constant function, threads for the same patch must see the same values for the elements of OutputPatch<HSOutput, 3>.

The SPIR-V InvocationID (SV_OutputControlPointID in HLSL) is used to index into the InputPatch and OutputPatch arrays to read/write information for the given vertex.

The hull main entry function in HLSL returns only one value (say, of type T), but that function is in fact executed once for each control point. The Vulkan spec requires that "Tessellation control shader per-vertex output variables and blocks, and tessellation control, tessellation evaluation, and geometry shader per-vertex input variables and blocks are required to be declared as arrays, with each element representing input or output values for a single vertex of a multi-vertex primitive". Therefore, we need to create a stage output variable that is an array with elements of type T. The number of elements of the array is equal to the number of output control points. Each final output control point is written into the corresponding element in the array using SV_OutputControlPointID as the index.

Patch Constant Function

As mentioned above, the patch constant function is to be invoked only once per patch. As a result, in the SPIR-V module, the entry function wrapper will first invoke the main entry function, and then use an OpControlBarrier to wait for all vertex processing to finish. After the barrier, only the first thread (with InvocationID of 0) will invoke the patch constant function. Since the first thread has to see the OutputPatch that contains output of the hull shader function for other threads, we have to use the output stage variable (with Output storage class) of the hull shader function for OutputPatch that can be an input to the patch constant function.

The information resulting from the patch constant function will also be returned as stage output variables. The output struct of the patch constant function must include SV_TessFactor and SV_InsideTessFactor fields which will translate to TessLevelOuter and TessLevelInner builtin variables, respectively. And the rest will be flattened and translated into normal stage output variables, one for each field.

Geometry Shaders

This section describes how geometry shaders are translated to SPIR-V for Vulkan.

Geometry Shader Entry Point Attributes

The following HLSL attribute is attached to the main entry point of geometry shaders and is translated to SPIR-V execution mode as follows:

HLSL Attribute value SPIR-V Execution Mode
maxvertexcount n OutputVertices n
instance n Invocations n

Translation for Primitive Types

Geometry shader vertex inputs may be qualified with primitive types. Only one primitive type is allowed to be used in a given geometry shader. The following table shows the SPIR-V execution mode that is used in order to represent the given primitive type.

HLSL Primitive Type SPIR-V Execution Mode
point InputPoints
line InputLines
triangle Triangles
lineadj InputLinesAdjacency
triangleadj InputTrianglesAdjacency

Translation of Output Stream Types

Supported output stream types in geometry shaders are: PointStream<T>, LineStream<T>, and TriangleStream<T>. These types are translated as the underlying type T, which is recursively flattened into stand-alone variables for each field.

Furthermore, output stream objects passed to geometry shader entry points are required to be annotated with inout, but the generated SPIR-V only contains stage output variables for them.

The following table shows the SPIR-V execution mode that is used in order to represent the given output stream.

HLSL Output Stream SPIR-V Execution Mode
PointStream OutputPoints
LineStream OutputLineStrip
TriangleStream OutputTriangleStrip

In other shader stages, stage output variables are only written in the entry function wrapper after calling the source code entry function. However, geometry shaders can output as many vertices as they wish, by calling the .Append() method on the output stream object. Therefore, it is incorrect to have only one flush in the entry function wrapper like other stages. Instead, each time a *Stream<T>::Append() is encountered, all stage output variables behind T will be flushed before SPIR-V OpEmitVertex instruction is generated. .RestartStrip() method calls will be translated into the SPIR-V OpEndPrimitive instruction.

Raytracing Shader Stages

DirectX Raytracing adds six new shader stages for raytracing namely ray generation, intersection, closest-hit, any-hit, miss and callable.

Flow chart for various stages in a raytracing pipeline is as follows:

 +---------------------+
 |   Ray generation    |
 +---------------------+
            |
 TraceRay() |                      +--------------+
            |      _ _ _ _ _ _ _ _ |   Any Hit    |
            |     |                +--------------+
            V     V                       ^
 +---------------------+                  |
 |    Acceleration     |           +--------------+
 |     Structure       |           | Intersection |
 |     Traversal       |           +--------------+
 +---------------------+                  ^
           |        |                     |
           |        |_ _ _ _ _ _ _ _ _ _ _|
           |
           |
           V
 +--------------------+            +-------------+
 |      Is Hit ?      |            |  Callable   |
 +--------------------+            +-------------+
     |            |
 Yes |            | No
     V            V
+---------+    +------+
| Closest |    | Miss |
|   Hit   |    |      |
+---------+    +------+
Note : DXC does not add special shader profiles for raytracing under -T option.
All raytracing shaders must be compiled as library using lib_6_3/lib_6_4 profile option.
Note : DXC now targets SPV_KHR_ray_tracing extension by default.
This extension is provisional and subject to change.
To compile for NV extension use -fspv-extension=SPV_NV_ray_tracing.

Ray Generation Stage

Ray generation shaders start ray tracing work and work on a compute-like 3D grid of threads.
Entry functions of this stage type are annotated with [shader("raygeneration")] in HLSL source.
Such entry functions must return void and do not accept any arguments.
For example:
RaytracingAccelerationStructure rs;
struct Payload
{
float4 color;
};
[shader("raygeneration")]
void main() {
  Payload myPayload = { float4(0.0f,0.0f,0.0f,0.0f) };
  RayDesc rayDesc;
  rayDesc.Origin = float3(0.0f, 0.0f, 0.0f);
  rayDesc.Direction = float3(0.0f, 0.0f, -1.0f);
  rayDesc.TMin = 0.0f;
  rayDesc.TMax = 1000.0f;
  TraceRay(rs, 0x0, 0xff, 0, 1, 0, rayDesc, myPayload);
}

Intersection Stage

Intersection shader stage is used to implement arbitrary ray-primitive intersections such spheres or axis-aligned bounding boxes (AABB). Triangle primitives do not require a custom intersection shader.
Entry functions of this stage are annotated with [shader("intersection")] in HLSL source.
Such entry functions must return void and do not accept any arguments.
For example:
struct Attribute
{
  float2 bary;
};

[shader("intersection")]
void main() {
Attribute myHitAttribute = { float2(0.0f,0.0f) };
ReportHit(0.0f, 0U, myHitAttribute);
}

Closest-Hit Stage

Hit shaders are invoked when a ray primitive intersection is found. A closest-hit shader
is invoked for the closest intersection point along a ray and can be used to compute interactions
at intersection point or spawn secondary rays.
Entry functions of this stage are annotated with [shader("closesthit")] in HLSL source.
Such entry functions must return void and accept exactly two arguments. First argument must be an inout
variable of user defined structure type and second argument must be a in variable of user defined structure type.
For example:
struct Attribute
{
  float2 bary;
};
struct Payload {
  float4 color;
};
[shader("closesthit")]
void main(inout Payload a, in Attribute b) {
  a.color = float4(0.0f,1.0f,0.0f,0.0f);
}

Any-Hit Stage

Hit shaders are invoked when a ray primitive intersection is found. An any-hit shader
is invoked for all intersections along a ray with a primitive.
Entry functions of this stage are annotated with [shader("anyhit")] in HLSL source.
Such entry functions must return void and accept exactly two arguments. First argument must be an inout
variable of user defined structure type and second argument must be an in variable of user defined structure type.
For example:
struct Attribute
{
  float2 bary;
};
struct Payload {
  float4 color;
};
[shader("anyhit")]
void main(inout Payload a, in Attribute b) {
  a.color = float4(0.0f,1.0f,0.0f,0.0f);
}

Miss Stage

Miss shaders are invoked when no intersection is found.
Entry functions of this stage are annotated with [shader("miss")] in HLSL source.
Such entry functions return void and accept exactly one argument. First argument must be an inout variable of user defined structure type.
For example:
struct Payload {
  float4 color;
};
[shader("miss")]
void main(inout Payload a) {
  a.color = float4(0.0f,1.0f,0.0f,0.0f);
}

Callable Stage

Callables are generic function calls which can be invoked from either raygeneration, closest-hit,
miss or callable shader stages.
Entry functions of this stage are annotated with [shader("callable")] in HLSL source.
Such entry functions must return void and accept exactly one argument. First argument must be an inout
variable of user defined structure type.
For example:
struct CallData {
  float4 data;
};
[shader("callable")]
void main(inout CallData a) {
  a.color = float4(0.0f,1.0f,0.0f,0.0f);
}

Mesh and Amplification Shaders

DirectX adds 2 new shader stages for using MeshShading pipeline namely Mesh and Amplification.
Amplification shaders corresponds to Task Shaders in Vulkan.

Refer to following HLSL and SPIR-V specs for details:

This section describes how Mesh and Amplification shaders are translated to SPIR-V for Vulkan.

Entry Point Attributes

The following HLSL attributes are attached to the main entry point of Mesh and/or Amplification shaders and are translated to SPIR-V execution modes according to the table below:

HLSL Attribute Value SPIR-V Execution Mode

outputtopology

(SPV_NV_mesh_shader)
 point OutputPoints
line OutputLinesNV
triangle OutputTrianglesNV

outputtopology

(SPV_EXT_mesh_shader)
 point OutputPoints
line OutputLinesEXT
triangle OutputTrianglesEXT
numthreads

X, Y, Z

(X*Y*Z <= 128)

 LocalSize X, Y, Z

Intrinsics

The following HLSL intrinsics are used in Mesh or Amplification shaders and are translated to SPIR-V intrinsics according to the table below:

HLSL Intrinsic Parameters SPIR-V Intrinsic

SetMeshOutputCounts

(Mesh shader)

numVertices

numPrimitives

 PrimitiveCountNV numPrimitives

DispatchMesh

(Amplification shader)

ThreadX

ThreadY

ThreadZ

MeshPayload

OpControlBarrier

TaskCountNV ThreadX*ThreadY*ThreadZ
HLSL Intrinsic Parameters SPIR-V Intrinsic

SetMeshOutputCounts

(Mesh shader)

numVertices

numPrimitives

 OpSetMeshOutputsEXT

DispatchMesh

(Amplification shader)

ThreadX

ThreadY

ThreadZ

MeshPayload

OpEmitMeshTasksEXT ThreadX ThreadY ThreadZ MeshPayload

TaskCountNV ThreadX*ThreadY*ThreadZ
Note : For DispatchMesh intrinsic, we also emit MeshPayload as output block with PerTaskNV decoration

Mesh Interface Variables

Interface variables are defined for Mesh shaders using HLSL modifiers.
Following table gives high level overview of the mapping:

HLSL modifier SPIR-V definition
indices

Maps to SPIR-V intrinsic PrimitiveIndicesNV

Defines SPIR-V Execution Mode OutputPrimitivesNV <array-size>
vertices

Maps to per-vertex out attributes

Defines existing SPIR-V Execution Mode OutputVertices <array-size>
primitives Maps to per-primitive out attributes with PerPrimitiveNV decoration
payload Maps to per-task in attributes with PerTaskNV decoration

Raytracing in Vulkan and SPIRV

SPIR-V codegen is currently supported for NVIDIA platforms via SPV_NV_ray_tracing extension or
on other platforms via provisional cross vendor SPV_KHR_ray_tracing extension.
SPIR-V specification for reference:

Raytracing Mapping to SPIR-V

Intrinsics

Following table provides mapping for system value intrinsics along with supported shader stages.
HLSL SPIR-V HLSL Shader Stage
System Value Intrinsic Builtin Raygen Intersection Closest Hit Any Hit Miss Callable
DispatchRaysIndex() LaunchId{NV/KHR} ✓ ✓ ✓ ✓ ✓ ✓
DispatchRaysDimensions() LaunchSize{NV/KHR} ✓ ✓ ✓ ✓ ✓ ✓
WorldRayOrigin() WorldRayOrigin{NV/KHR}   ✓ ✓ ✓ ✓  
WorldRayDirection() WorldRayDirection{NV/KHR}   ✓ ✓ ✓ ✓  
RayTMin() RayTmin{NV/KHR}   ✓ ✓ ✓ ✓  
RayTCurrent() HitT{NV/KHR}   ✓ ✓ ✓ ✓  
RayFlags() IncomingRayFlags{NV/KHR}   ✓ ✓ ✓ ✓  
InstanceIndex() InstanceId   ✓ ✓ ✓    
GeometryIndex() RayGeometryIndexKHR   ✓ ✓ ✓    
InstanceID() InstanceCustomIndex{NV/KHR}   ✓ ✓ ✓    
PrimitiveIndex() PrimitiveId   ✓ ✓ ✓    
ObjectRayOrigin() ObjectRayOrigin{NV/KHR}   ✓ ✓ ✓    
ObjectRayDirection() ObjectRayDirection{NV/KHR}   ✓ ✓ ✓    
ObjectToWorld3x4() ObjectToWorld{NV/KHR}   ✓ ✓ ✓    
ObjectToWorld4x3() ObjectToWorld{NV/KHR}   ✓ ✓ ✓    
WorldToObject3x4() WorldToObject{NV/KHR}   ✓ ✓ ✓    
WorldToObject4x3() WorldToObject{NV/KHR}   ✓ ✓ ✓    
HitKind() HitKind{NV/KHR}   ✓ ✓ ✓    
There is no separate builtin for transposed matrices ObjectToWorld3x4 and WorldToObject3x4 in SPIR-V hence we internally transpose during translation
GeometryIndex() is only supported under SPV_KHR_ray_tracing extension.
Following table provides mapping for other intrinsics along with supported shader stages.
HLSL SPIR-V HLSL Shader Stage
Intrinsic Opcode Raygen Intersection Closest Hit Any Hit Miss Callable
TraceRay OpTrace{NV/KHR} ✓   ✓   ✓  
ReportHit OpReportIntersection{NV/KHR} ✓ ✓        
IgnoreHit OpIgnoreIntersection{NV/KHR} ✓     ✓    
AcceptHitAndEndSearch OpTerminateRay{NV/KHR} ✓     ✓    
CallShader OpExecuteCallable{NV/KHR} ✓   ✓ ✓ ✓  

Resource Types

Following table provides mapping for new resource types supported in all raytracing shaders.
HLSL Type SPIR-V Opcode
RaytracingAccelerationStructure OpTypeAccelerationStructure{NV/KHR}

Interface Variables

Interface variables are created for various ray tracing storage classes based on intrinsic/shader stage
Following table gives high level overview of the mapping.
SPIR-V Storage Class Created For
RayPayload{NV/KHR} Last argument to TraceRay
IncomingRayPayload{NV/KHR} First argument of entry for AnyHit/ClosestHit & Miss stage
HitAttribute{NV/KHR} Last argument to ReportHit
CallableData{NV/KHR} Last argument to CallShader
IncomingCallableData{NV/KHR} First argument of entry for Callable stage

RayQuery

Ray Query is subfeature of the DirectX ray tracing and belongs to the DirectX ray tracing spec 1.1 (DXR 1.1). DirectX add RayQuery object type and its member TraceRayInline() to do the TraceRay() that doesn't use any seperate ray-tracing shader stages. Shaders can instantiate RayQuery objects as local variables, the RayQuery object acts as a state machine for ray query. The shader interacts with the RayQuery object's methods to advance the query through an acceleration structure and query traversal information

Refer to following pages for details: https://microsoft.github.io/DirectX-Specs/d3d/Raytracing.html

A flow chart for a simple ray query process

  +------------------------------+
  |   RayQuery<RAY_FLAG_NONE> q  |
  +------------------------------+
                 |
                 V
  +------------------------------+
  |      q.TraceRayInline()      |
  +------------------------------+
          |               — — — — — — — — — — — — —
          |              |                         |
          |              |              +------------------------+
          |              |              | Your intersection code |
          |              |              +------------------------+
          |              |                         ^
          V              V                         |
  +------------------------------+      +---------------------+
  |  q.Proceed() // AS traversal |      |  q.CandidateType()  |
  +------------------------------+      +---------------------+
       |                   |                       ^
   No  |                   | Yes                   |
       |                   |_ _ _ _ _ _ _ _ _ _ _ _|
       V
 +------------------------------+
 |     q.CommittedStatus()      |
 +------------------------------+
               |
               V
+----------------------------------+
| Your Intersection/shader code    |
+----------------------------------+

Example:

void main() {
  RayQuery<RAY_FLAG_CULL_NON_OPAQUE | RAY_FLAG_ACCEPT_FIRST_HIT_AND_END_SEARCH> q;
  q.TraceRayInline(myAccelerationStructure, 0 , 0xff, myRay);

  // Proceed() is AccelerationStructure traversal loop take places
  while(q.Proceed()) {
    switch(q.CandidateType()) {
      // retrieve intersection information/Do the shadering
    }
  }

  // AccelerationStructure traversal end
  // Get the Committed status
  switch(q.CommittedStatus()) {
    // retrieve intersection information/ Do the shadering
  }
}

Ray Query in SPIRV

RayQuery SPIR-V codegen is currently supported via SPV_KHR_ray_query extension SPIR-V specification for reference: https://github.com/KhronosGroup/SPIRV-Registry/blob/master/extensions/KHR/SPV_KHR_ray_query.asciidoc

Object Type

RayQuery<RAY_FLAGS>

RayQuery represents the state of an inline ray tracing call into an acceleration structure.

HLSL Type SPIR-V Opcode
RayQuery OpTypeRayQueryKHR

RayQuery Mapping to SPIR-V

HLSL RayQuery member Intrinsic SPIR-V Opcode
.Abort OpRayQueryTerminateKHR
.CandidateType OpRayQueryGetIntersectionTypeKHR
.CandidateProceduralPrimitiveNonOpaque OpRayQueryGetIntersectionCandidateAABBOpaqueKHR
.CandidateInstanceIndex OpRayQueryGetIntersectionInstanceIdKHR
.CandidateInstanceID OpRayQueryGetIntersectionInstanceCustomIndexKHR
.CandidateInstanceContributionToHitGroupIndex OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR
.CandidateGeometryIndex OpRayQueryGetIntersectionGeometryIndexKHR
.CandidatePrimitiveIndex OpRayQueryGetIntersectionPrimitiveIndexKHR
.CandidateObjectRayOrigin OpRayQueryGetIntersectionObjectRayOriginKHR
.CandidateObjectRayDirection OpRayQueryGetIntersectionObjectRayDirectionKHR
.CandidateObjectToWorld3x4 OpRayQueryGetIntersectionObjectToWorldKHR
.CandidateObjectToWorld4x3 OpRayQueryGetIntersectionObjectToWorldKHR
.CandidateWorldToObject3x4 OpRayQueryGetIntersectionWorldToObjectKHR
.CandidateWorldToObject4x3 OpRayQueryGetIntersectionWorldToObjectKHR
.CandidateTriangleBarycentrics OpRayQueryGetIntersectionBarycentricsKHR
.CandidateTriangleFrontFace OpRayQueryGetIntersectionFrontFaceKHR
.CommittedStatus OpRayQueryGetIntersectionTypeKHR
.CommittedInstanceIndex OpRayQueryGetIntersectionInstanceIdKHR
.CommittedInstanceID OpRayQueryGetIntersectionInstanceCustomIndexKHR
.CommittedInstanceContributionToHitGroupIndex OpRayQueryGetIntersectionInstanceShaderBindingTableRecordOffsetKHR
.CommittedGeometryIndex OpRayQueryGetIntersectionGeometryIndexKHR
.CommittedPrimitiveIndex OpRayQueryGetIntersectionPrimitiveIndexKHR
.CommittedRayT OpRayQueryGetIntersectionTKHR
.CommittedObjectRayOrigin OpRayQueryGetIntersectionObjectRayOriginKHR
.CommittedObjectRayDirection OpRayQueryGetIntersectionObjectRayDirectionKHR
.CommittedObjectToWorld3x4 OpRayQueryGetIntersectionObjectToWorldKHR
.CommittedObjectToWorld4x3 OpRayQueryGetIntersectionObjectToWorldKHR
.CommittedWorldToObject3x4 OpRayQueryGetIntersectionWorldToObjectKHR
.CommittedWorldToObject4x3 OpRayQueryGetIntersectionWorldToObjectKHR
.CommittedTriangleBarycentrics OpRayQueryGetIntersectionBarycentricsKHR
.CommittedTriangleFrontFace OpRayQueryGetIntersectionFrontFaceKHR
.CommitNonOpaqueTriangleHit OpRayQueryConfirmIntersectionKHR
.CommitProceduralPrimitiveHit OpRayQueryGenerateIntersectionKHR
.Proceed OpRayQueryProceedKHR
.RayFlags OpRayQueryGetRayFlagsKHR
.RayTMin OpRayQueryGetRayTMinKHR
.TraceRayInline OpRayQueryInitializeKHR
.WorldRayDirection OpRayQueryGetWorldRayDirectionKHR
``.WorldRayOrigin` OpRayQueryGetWorldRayOriginKHR

Shader Model 6.0 Wave Intrinsics

Note that Wave intrinsics requires SPIR-V 1.3, which is supported by Vulkan 1.1. If you use wave intrinsics in your source code, you will need to specify -fspv-target-env=vulkan1.1 via the command line to target Vulkan 1.1.

Shader model 6.0 introduces a set of wave operations. Apart from WaveGetLaneCount() and WaveGetLaneIndex(), which are translated into loading from SPIR-V builtin variable SubgroupSize and SubgroupLocalInvocationId respectively, the rest are translated into SPIR-V group operations with Subgroup scope according to the following chart:

Wave Category Wave Intrinsics SPIR-V Opcode SPIR-V Group Operation
Query WaveIsFirstLane() OpGroupNonUniformElect  
Vote WaveActiveAnyTrue() OpGroupNonUniformAny  
Vote WaveActiveAllTrue() OpGroupNonUniformAll  
Vote WaveActiveBallot() OpGroupNonUniformBallot  
Reduction WaveActiveAllEqual() OpGroupNonUniformAllEqual Reduction
Reduction WaveActiveCountBits() OpGroupNonUniformBallotBitCount Reduction
Reduction WaveActiveSum() OpGroupNonUniform*Add Reduction
Reduction WaveActiveProduct() OpGroupNonUniform*Mul Reduction
Reduction WaveActiveBitAdd() OpGroupNonUniformBitwiseAnd Reduction
Reduction WaveActiveBitOr() OpGroupNonUniformBitwiseOr Reduction
Reduction WaveActiveBitXor() OpGroupNonUniformBitwiseXor Reduction
Reduction WaveActiveMin() OpGroupNonUniform*Min Reduction
Reduction WaveActiveMax() OpGroupNonUniform*Max Reduction
Scan/Prefix WavePrefixSum() OpGroupNonUniform*Add ExclusiveScan
Scan/Prefix WavePrefixProduct() OpGroupNonUniform*Mul ExclusiveScan
Scan/Prefix WavePrefixCountBits() OpGroupNonUniformBallotBitCount ExclusiveScan
Broadcast WaveReadLaneAt() OpGroupNonUniformBroadcast  
Broadcast WaveReadLaneFirst() OpGroupNonUniformBroadcastFirst  
Quad QuadReadAcrossX() OpGroupNonUniformQuadSwap  
Quad QuadReadAcrossY() OpGroupNonUniformQuadSwap  
Quad QuadReadAcrossDiagonal() OpGroupNonUniformQuadSwap  
Quad QuadReadLaneAt() OpGroupNonUniformQuadBroadcast  

The Implicit vk Namespace

Overview

We have introduced an implicit namepace (called vk) that will be home to all Vulkan-specific functions, enums, etc. Given the similarity between HLSL and C++, developers are likely familiar with namespaces -- and implicit namespaces (e.g. std:: in C++). The vk namespace provides an interface for expressing Vulkan-specific features (core spec and KHR extensions).

The compiler will generate the proper error message ( unknown 'vk' identifier ) if vk:: is used for compiling to DXIL.

Any intrinsic function or enum in the vk namespace will be deprecated if an equivalent one is added to the default namepsace.

Current Features

The following intrinsic functions and constants are currently defined in the implicit vk namepsace.

// Implicitly defined when compiling to SPIR-V.
namespace vk {

  const uint CrossDeviceScope = 0;
  const uint DeviceScope      = 1;
  const uint WorkgroupScope   = 2;
  const uint SubgroupScope    = 3;
  const uint InvocationScope  = 4;
  const uint QueueFamilyScope = 5;

  uint64_t ReadClock(in uint scope);
  T        RawBufferLoad<T = uint>(in uint64_t deviceAddress,
                                   in uint alignment = 4);
} // end namespace

Intrinsic Constants

The following constants are currently defined:

Constant value (SPIR-V constant equivalent, if any)
vk::CrossDeviceScope 0 (CrossDevice)
vk::DeviceScope 1 (Device)
vk::WorkgroupScope 2 (Workgroup)
vk::SubgroupScope 3 (Subgroup)
vk::InvocationScope 4 (Invocation)
vk::QueueFamilyScope 5 (QueueFamily)

Intrinsic Functions

ReadClock

This intrinsic funcion has the following signature:

uint64_t ReadClock(in uint scope);

It translates to performing OpReadClockKHR defined in VK_KHR_shader_clock. One can use the predefined scopes in the vk namepsace to specify the scope argument. For example:

uint64_t clock = vk::ReadClock(vk::SubgroupScope);

RawBufferLoad and RawBufferStore

The Vulkan extension VK_KHR_buffer_device_address supports getting the 64-bit address of a buffer and passing it to SPIR-V as a Uniform buffer. SPIR-V can use the address to load and store data without a descriptor. We add the following intrinsic functions to expose a subset of the VK_KHR_buffer_device_address and SPV_KHR_physical_storage_buffer functionality to HLSL:

// RawBufferLoad and RawBufferStore use 'uint' for the default template argument.
// The default alignment is 4. Note that 'alignment' must be a constant integer.
T RawBufferLoad<T = uint>(in uint64_t deviceAddress, in uint alignment = 4);
void RawBufferStore<T = uint>(in uint64_t deviceAddress, in T value, in uint alignment = 4);

These intrinsics allow the shader program to load and store a single value with type T (int, float2, struct, etc...) from GPU accessible memory at given address, similar to ByteAddressBuffer.Load(). Additionally, these intrinsics allow users to set the memory alignment for the underlying data. We assume a 'uint' type when the template argument is missing, and we use a value of '4' for the default alignment. Note that the alignment argument must be a constant integer if it is given.

Though we do support setting the alignment of the data load and store, we do not currently support setting the memory layout for the data. Since these intrinsics are supposed to load "arbitrary" data to or from a random device address, we assume that the program loads/stores some "bytes of data", but that its format or layout is unknown. Therefore, keep in mind that these intrinsics load or store sizeof(T) bytes of data, and that loading/storing data with a struct with a custom memory alignment may yield undefined behavior due to the missing custom memory layout support. Loading data with customized memory layouts is future work.

Using either of these intrinsics adds PhysicalStorageBufferAddresses capability and SPV_KHR_physical_storage_buffer extension requirements as well as changing the addressing model to PhysicalStorageBuffer64.

Example:

uint64_t address;
[numthreads(32, 1, 1)]
void main(uint3 tid : SV_DispatchThreadID) {
  double foo = vk::RawBufferLoad<double>(address, 8);
  uint bar = vk::RawBufferLoad(address + 8);
  ...
  vk::RawBufferStore<uint>(address + tid.x, bar + tid.x);
}

Inline SPIR-V (HLSL version of GL_EXT_spirv_intrinsics)

GL_EXT_spirv_intrinsics is an extension of GLSL that allows users to embed arbitrary SPIR-V instructions in the GLSL code similar to the concept of inline assembly in the C code. We support the HLSL version of GL_EXT_spirv_intrinsics. See wiki for the details.

Supported Command-line Options

Command-line options supported by SPIR-V CodeGen are listed below. They are also recognized by the library API calls.

General options

  • -T: specifies shader profile
  • -E: specifies entry point
  • -D: Defines macro
  • -I: Adds directory to include search path
  • -O{|0|1|2|3}: Specifies optimization level
  • -enable-16bit-types: enables 16-bit types and disables min precision types
  • -Zpc: Packs matrices in column-major order by deafult
  • -Zpr: Packs matrices in row-major order by deafult
  • -Fc: outputs SPIR-V disassembly to the given file
  • -Fe: outputs warnings and errors to the given file
  • -Fo: outputs SPIR-V code to the given file
  • -Fh: outputs SPIR-V code as a header file
  • -Vn: specifies the variable name for SPIR-V code in generated header file
  • -Zi: Emits more debug information (see Debugging)
  • -Cc: colorizes SPIR-V disassembly
  • -No: adds instruction byte offsets to SPIR-V disassembly
  • -H: Shows header includes and nesting depth
  • -Vi: Shows details about the include process
  • -Vd: Disables SPIR-V verification
  • -WX: Treats warnings as errors
  • -no-warnings: Suppresses all warnings
  • -flegacy-macro-expansion: expands the operands before performing token-pasting operation (fxc behavior)

Vulkan-specific options

The following command line options are added into dxc to support SPIR-V codegen for Vulkan:

  • -spirv: Generates SPIR-V code.
  • -fvk-b-shift N M: Shifts by N the inferred binding numbers for all resources in b-type registers of space M. Specifically, for a resouce attached with :register(bX, spaceM) but not [vk::binding(...)], sets its Vulkan descriptor set to M and binding number to X + N. If you need to shift the inferred binding numbers for more than one space, provide more than one such option. If more than one such option is provided for the same space, the last one takes effect. If you need to shift the inferred binding numbers for all sets, use all as M. See HLSL register and Vulkan binding for explanation and examples.
  • -fvk-t-shift N M, similar to -fvk-b-shift, but for t-type registers.
  • -fvk-s-shift N M, similar to -fvk-b-shift, but for s-type registers.
  • -fvk-u-shift N M, similar to -fvk-b-shift, but for u-type registers.
  • -fvk-auto-shift-bindings: Automatically detects the register type for resources that are missing the :register assignment, so the above shifts can be applied to them if needed.
  • -fvk-bind-register xX Y N M (short alias: -vkbr): Binds the resouce at register(xX, spaceY) to descriptor set M and binding N. This option cannot be used together with other binding assignment options. It requires all source code resources have :register() attribute and all registers have corresponding Vulkan descriptors specified using this option. If the $Globals cbuffer resource is used, it must also be bound with -fvk-bind-globals.
  • -fvk-bind-globals N M: Places the $Globals cbuffer at descriptor set #M and binding #N. See HLSL global variables and Vulkan binding for explanation and examples.
  • -fvk-use-gl-layout: Uses strict OpenGL std140/std430 layout rules for resources.
  • -fvk-use-dx-layout: Uses DirectX layout rules for resources.
  • -fvk-invert-y: Negates (additively inverts) SV_Position.y before writing to stage output. Used to accommodate the difference between Vulkan's coordinate system and DirectX's. Only allowed in VS/DS/GS.
  • -fvk-use-dx-position-w: Reciprocates (multiplicatively inverts) SV_Position.w after reading from stage input. Used to accommodate the difference between Vulkan DirectX: the w component of SV_Position in PS is stored as 1/w in Vulkan. Only recognized in PS; applying to other stages is no-op.
  • -fvk-stage-io-order={alpha|decl}: Assigns the stage input/output variable location number according to alphabetical order or declaration order. See HLSL semantic and Vulkan Location for more details.
  • -fspv-reflect: Emits additional SPIR-V instructions to aid reflection.
  • -fspv-debug=<category>: Controls what category of debug information should be emitted. Accepted values are file, source, line, and tool. See Debugging for more details.
  • -fspv-extension=<extension>: Only allows using <extension> in CodeGen. If you want to allow multiple extensions, provide more than one such option. If you want to allow all KHR extensions, use -fspv-extension=KHR.
  • -fspv-target-env=<env>: Specifies the target environment for this compilation. The current valid options are vulkan1.0 and vulkan1.1. If no target environment is provided, vulkan1.0 is used as default.
  • -fspv-flatten-resource-arrays: Flattens arrays of textures and samplers into individual resources, each taking one binding number. For example, an array of 3 textures will become 3 texture resources taking 3 binding numbers. This makes the behavior similar to DX. Without this option, you would get 1 array object taking 1 binding number. Note that arrays of {RW|Append|Consume}StructuredBuffers are currently not supported in the SPIR-V backend. Also note that this requires the optimizer to be able to resolve all array accesses with constant indeces. Therefore, all loops using the resource arrays must be marked with [unroll].
  • -fspv-entrypoint-name=<name>: Specify the SPIR-V entry point name. Defaults to the HLSL entry point name.
  • -fspv-use-legacy-buffer-matrix-order: Assumes the legacy matrix order (row major) when accessing raw buffers (e.g., ByteAdddressBuffer).
  • -fspv-preserve-interface: Preserves all interface variables in the entry point, even when those variables are unused.
  • -Wno-vk-ignored-features: Does not emit warnings on ignored features resulting from no Vulkan support, e.g., cbuffer member initializer.

Unsupported HLSL Features

The following HLSL language features are not supported in SPIR-V codegen, either because of no Vulkan equivalents at the moment, or because of deprecation.

  • Literal/immediate sampler state: deprecated feature. The compiler will emit a warning and ignore it.
  • abort() intrinsic function: no Vulkan equivalent. The compiler will emit an error.
  • GetRenderTargetSampleCount() intrinsic function: no Vulkan equivalent. (Its GLSL counterpart is gl_NumSamples, which is not available in GLSL for Vulkan.) The compiler will emit an error.
  • GetRenderTargetSamplePosition() intrinsic function: no Vulkan equivalent. (gl_SamplePosition provides similar functionality but it's only for the sample currently being processed.) The compiler will emit an error.
  • tex*() intrinsic functions: deprecated features. The compiler will emit errors.
  • .GatherCmpGreen(), .GatherCmpBlue(), .GatherCmpAlpha() intrinsic method: no Vulkan equivalent. (SPIR-V OpImageDrefGather instruction does not take component as input.) The compiler will emit an error.
  • Since StructuredBuffer, RWStructuredBuffer, ByteAddressBuffer, and RWByteAddressBuffer are not represented as image types in SPIR-V, using the output unsigned integer status argument in their Load* methods is not supported. Using these methods with the status argument will cause a compiler error.
  • Applying row_major or column_major attributes to a stand-alone matrix will be ignored by the compiler because RowMajor and ColMajor decorations in SPIR-V are only allowed to be applied to members of structures. A warning will be issued by the compiler.
  • The Hull shader partitioning attribute may not have the pow2 value. The compiler will emit an error. Other attribute values are supported and described in the Hull Entry Point Attributes section.
  • cbuffer/tbuffer member initializer: no Vulkan equivalent. The compiler will emit an warning and ignore it.

Appendix

Appendix A. Matrix Representation

Consider a matrix in HLSL defined as float2x3 m;. Conceptually, this is a matrix with 2 rows and 3 columns. This means that you can access its elements via expressions such as m[i][j], where i can be {0, 1} and j can be {0, 1, 2}.

Now let's look how matrices are defined in SPIR-V:

%columnType = OpTypeVector %float      <number of rows>
   %matType = OpTypeMatrix %columnType <number of columns>

As you can see, SPIR-V conceptually represents matrices as a collection of vectors where each vector is a column.

Now, let's represent our float2x3 matrix in SPIR-V. If we choose a naive translation (3 columns, each of which is a vector of size 2), we get:

    %v2float = OpTypeVector %float 2
%mat3v2float = OpTypeMatrix %v2float 3

Now, let's use this naive translation to access into the matrix (e.g. m[0][2]). This is evaluated by first finding n = m[0], and then finding n[2]. Notice that in HLSL, m[0] represents a row, which is a vector of size 3. But accessing the first dimension of the SPIR-V matrix give us the first column which is a vector of size 2.

; n is a vector of size 2
%n = OpAccessChain %v2float %m %int_0

Notice that in HLSL access m[i][j], i can be {0, 1} and j can be {0, 1, 2}. But in SPIR-V OpAccessChain access, the first index (i) can be {0, 1, 2} and the second index (j) can be {1, 0}. Therefore, the naive translation does not work well with indexing.

As a result, we must translate a given HLSL float2x3 matrix (with 2 rows and 3 columns) as a SPIR-V matrix with 3 rows and 2 columns:

    %v3float = OpTypeVector %float 3
%mat2v3float = OpTypeMatrix %v3float 2

This way, all accesses into the matrix can be naturally handled correctly.

Packing

The HLSL row_major and column_major type modifiers change the way packing is done. The following table provides an example which should make our translation more clear:

Host CPU Data HLSL Variable GPU (HLSL Representation) GPU (SPIR-V Representation) SPIR-V Decoration
{1,2,3,4,5,6} float2x3

[1 3 5]

[2 4 6]

[1 2]

[3 4]

[5 6]

 RowMajor
{1,2,3,4,5,6} column_major float2x3

[1 3 5]

[2 4 6]

[1 2]

[3 4]

[5 6]

 RowMajor
{1,2,3,4,5,6} row_major float2x3

[1 2 3]

[4 5 6]

[1 4]

[2 5]

[3 6]

 ColMajor