Version:

AZSL, The Amazon Shading Language

Shaders in Atom Renderer are written in Amazon Shading Language (AZSL). AZSL is a superset of Microsoft’s high-level shader language (HLSL) . AZSL supports Vertex, Fragment and Compute shaders according to Shader Model 6.2. Ray-Tracing shaders are also supported according to Shader Model 6.3.

Typically, you write shader code in a file with the .azsl extension, also known as an AZSL source file. You can also write shader code in a file with the .azsli extension, which is an include file that multiple AZSL source files can reference. Consider writing your shaders in .azsli files to reuse them in multiple AZSL source files and reduce code redundancy.

AZSL is defined on top of HLSL with the following differences:

The shader build pipeline allows for .azsl files to include other files and define C/C++ Macros. The .azsl files are always pre-processed by the C Pre-Processor (MCPP) before being transpiled into HLSL. In other words, if a raw .azsl file has #include, #pragma or #define directives it should be run through MCPP before being compiled with Amazon Shading Language Compiler (AZSLc), otherwise AZSLc will fail to compile such files.

Note:

MCPP doesn’t support C Macros with arithmetic expressions. Examples:

    // Bad (Arithmetic Expression in C Macros are not supported by MCPP)
    #define BAD_MACRO(x, y) ((x * 0.5) - (y * 2.0))
     
    // Good
    #define HEIGHT (7.2)

Declaring Shader Constants

Shader constants are also known as shader uniforms in other shader languages, like GLSL.

AZSL allows the declaration of shader constants only in the following places:

AZSLc triggers a compilation error if an arbitrary cbuffer or any other type of resource is declared outside of the scopes mentioned above. The exception is static global variables, which shaders use internally, but aren’t exposed to the application.

Note:
You can also use option variables equivalently as Vulkan’s Specialization Constants. In other words, You can define these constants statically at compilation time to avoid branching at runtime.

Shader Resource Groups

A Shader Resource Group (SRG) is a logical namespace to declare shader constants and resources. In AZSL, the keyword to declare an SRG is ShaderResourceGroup. Shader resources are constants of the types: Shader Resource Views (SRV), Unordered Access Views (UAV), Constant Buffer Views (CBV), and Sampler States.

When compiling shaders with the --use-spaces argument, each SRG takes over a whole register space (or descriptor set, as known in Vulkan). AZSL abstracts the register and register space assignment of resource descriptors with the ShaderResourceGroupSemantic keyword, which you must specify when declaring SRGs.

Use the following syntax to declare an SRG:

    ShaderResourceGroup <Name> : <Semantic>
    {
        <Data>
    };

Alternatively, use a partial qualifier to declare an SRG. For more information on partials, refer to Partial ShaderResourceGroup Definitions.

    partial ShaderResourceGroup <Name> : <Semantic>
    {
        <Data>
    };
  • <Name> is the user defined name for the SRG.
  • <Semantic> is the name of the ShaderResourceGroupSemantic for the SRG. For more information on SRG Semantics, refer to AZSL, Shader Resource Group Semantics.
  • <Data> is one or more declarations of functions, constants, and resources for the binding data that the SRG contains. For more information on the type data that an SRG can contain, refer to Data.

Data

  • Fundamental type variables like float, float3, int2, bool, matrices, and so on. AZSLc emits a Constant Buffer View (CBV) that packs all the Fundamental type variables that the SRG declares.

  • Resource declarations like Texture2D, Texture2DMS, ByteAddressBuffer, as well as HLSL shader resource of the categories: Shader Resource Views (SRV), Unordered Access Views (UAV), Samplers and CBVs.

  • Unbounded arrays for the following resource declarations: SRVs, UAVs, CBVs and Samplers. You may need to consider some rules for unbounded arrays, depending on the command line arguments you use when you compile with AZSLc. For more information, refer to AZSL, Binding Rules fore Unbounded Arrays.

  • Functions.

  • Structs (declaration and instances).

  • Instances of classes. You can instantiate classes in an SRG, but you must declare them outside of the SRG.

Example
The following example demonstrates an SRG declaration. A Shader Resource Group Semantic (SRG Semantic) defines the space index, or descriptor set, of the constants and resources declared inside the SRG. When you declare an SRG, it must specify an SRG Semantic. For more information on SRG Semantics, refer to AZSL, Shader Resource Group Semantics.

    ShaderResourceGroupSemantic BindingPerExample
    {
        FrequencyId = 0;
    };
       
    ShaderResourceGroup ExampleSRG : BindingPerExample
    {
        float4          m_uniformColor;
        Texture2D       m_texture;
        Buffer<float>   m_buffer;
        Sampler m_dynamicSampler;
        Sampler m_staticSampler
        {
            AddressU = Wrap;
            AddressV = Wrap;
            AddressW = Wrap;
            MagFilter = Linear;
        };
        
        // You can define functions inside a ShaderResourceGroup
        float GetRedColor()
        {
            return m_uniformColor.r;
        }
    };

Scope

Every SRG defines a scope. The SRG data must be qualified when accessed from outside of the SRG’s scope in code.

Example:

    float4 color = ExampleSRG::m_uniformColor;
    color *= ExampleSRG::m_texture.Sample(ExampleSRG::m_staticSampler, float2(0.5, 0.5));

Sampler variables

A dynamic sampler declaration does not contain a body definition because they must be set at runtime. You can also declare a dynamic sampler as an array.

A static sampler declaration contains a body definition. You can declare a sampler and sampler-comparison by using AZSL’s Sampler keyword, which conveniently encompasses HLSL’s SamplerState and SamplerComparisonState keywords. AZSLc compiles sampler variables into the corresponding HLSL variables: SamplerState, or SamplerComparisonState, if it contains a ComparisonFunc definition..

The following examples demonstrate declarations for a dynamic sampler, a static sampler, and a static sampler comparison. They must be declared inside an SRG.

Example:

    // Dynamic sampler declaration
    Sampler m_dynamicSampler;
     
    // Static sampler declaration. Compiled as HLSL SamplerState
    Sampler m_staticSamplerState
    {
        AddressU = Wrap;
        AddressV = Wrap;
        AddressW = Wrap;
        MagFilter = Linear;
    };
     
    // Static sampler declaration. Compiled as HLSL SamplerComparisonState because of the ComparisonFunc declaration.
    Sampler m_staticSamplerComparisonState
    {
        AddressU = Wrap;
        AddressV = Wrap;
        AddressW = Wrap;
        MagFilter = Linear;
        ComparisonFunc = Less;
    };
    ...

Functions

Functions can be declared in classes, global scopes, or inside SRGs. Defining functions inside SRGs is unique to AZSL.

Example:

    ShaderResourceGroup PerObject : BindingPerObject
    {
        row_major float3x4 m_modelToWorld;
        row_major float3x3 m_normalMatrix;
       
        float4x4 GetWorldMatrix()
        {
            float4x4 modelToWorld = float4x4(
                float4(1, 0, 0, 0),
                float4(0, 1, 0, 0),
                float4(0, 0, 1, 0),
                float4(0, 0, 0, 1));
       
            modelToWorld[0] = m_modelToWorld[0];
            modelToWorld[1] = m_modelToWorld[1];
            modelToWorld[2] = m_modelToWorld[2];
            return modelToWorld;
        }
           
        float3x3 GetNormalMatrix()
        {
            return m_normalMatrix;
        }
    }
      
      
    VertexOutput MainVS(VertexInput input)
    {
        const float4x4 worldMatrix = PerObject::GetWorldMatrix();
       
        VertexOutput output;
        float3 worldPosition = mul(worldMatrix, float4(input.m_position,1)).xyz;
        output.m_position = mul(GetView_ViewProjectionMatrix(), float4(worldPosition, 1.0));
       
        output.m_uv = input.m_uv;
               
        output.m_positionToCamera = GetView_WorldPosition() - worldPosition;
       
        output.m_normal = mul(PerObject::GetNormalMatrix(), input.m_normal);
        output.m_tangent = mul(PerObject::GetNormalMatrix(), input.m_tangent);
       
        return output;
    }

Structs

Just like in HLSL, the following rules apply for structs:

  • Must be defined globally.
  • Contains only data, not methods.
  • Cannot have a constructor or destructor.
  • Cannot accept default member initializers.

AZSL adds an extended rule: Define and instantiate structs inside SRGS. You can instantiate globally defined structs inside SRGs, root constants, and another struct or class.

Classes

Just like in HLSL, the following rules apply for classes:

  • Must be declared and defined only in global scopes, outside of SRGs.
  • Can inherit from interfaces, which forces the class to repect a minimal method set.
  • Don’t have constructors or destructors.
  • Don’t accept default member initializers.

ASZL adds the following extended rules for classes:

  • Can instantiate class type variables in an SRG.
  • until 1.7.35: classes cannot inherit. Class inheritance is introduced in azslc 1.8.9 released on December 2022.

Matrices

AZSL follows the HLSL convention to use column major matrices by default. However, in O3DE runtime, the default convention is row major matrices. When declaring a matrix, you can specify the order by using the row_major or column_major keywords.

Example:

    ShaderResourceGroupSemantic MatrixExample
    {
        FrequencyId = 1;
    };
      
    ShaderResourceGroup MatrixExampleSRG : MatrixExample
    {
                     float2x3 defaultMatrix;
        row_major    float2x3 rowMajorMatrix;
        column_major float2x3 colMajorMatrix;
    };

Arrays And Unbounded Arrays

For array definitions, the rules and limitations are exactly the same as HLSL. .

Example:

    ShaderResourceGroup PerObject : BindingPerObject
    {
        float m_arrayOfFloats[64]; // GOOD. Arrays of fundamental types must be of constant size.
     
        struct SomeStruct
        {
            int m_integer;
            bool m_boolean;
        }
        static const int MyArraySize = 32;
        SomeStruct m_arrayOfStructs[MyArraySize]; // GOOD. Arrays of packable types (structs & fundamentals) require constant size definition.
     
        int m_unboundedArrayOfIntegers[]; // ERROR. Unbounded arrays are not supported for fundamental(packables, in general) data types.
     
        Texture2D<float4> m_textures[MyArraySize]; // GOOD. Contant size array of Resource View types is ok.
         
        ConstantBuffer<SomeStruct> m_unboundedArrayOfStructs[]; // GOOD. Unbounded array of Resource View type is ok. BUT, there are some restrictions.
    }

To learn more about the limitations when declaring unbounded arrays, go to: Binding Rules For Unbounded Arrays.

Partial ShaderResourceGroup Definitions.

Before talking about partial SRGs, it’s important to talk about SRG Semantics, which is always required when declaring a ShaderResourceGroup. Read about SRG Semantics in Shader Resource Group Semantics.

Define a partial SRG by using partial ShaderResourceGroup. The partial keyword is useful to define parts of an SRG in different blocks (for example, across different files). This is a similar concept as partial classes in C#.
There are no limits on how many partial blocks can be used to define an SRG.

There are three requirements for defining partial SRGs:

  1. When using partial to define an SRG block, all other SRG definition blocks within for the same SRG must also start with the partial keyword.
  2. For a given SRG, at least one partial SRG block must define the SRG Semantic.
  3. For a given SRG, if the SRG Semantic is defined more than once across different partial blocks, then they must refer to the same SRG Semantic.
Important:
When AZSLc finds the first partial ShaderResourceGroup block, it will use it as the unified point of emission for all the data of a given SRG across all partial blocks. The main recommendation that comes out of this rule, is that any globally defined struct, or class that may be referenced by a partial SRG block must be defined before the first partial block for the SRG in question.

Example: Using the partial keyword

    ShaderResourceGroupSemantic PerPass
    {
        FrequencyId = 1;
    }
     
    partial ShaderResourceGroup PassSrg
    {
        Texture2D m_texture; // A regular SRV resource,
    }
     
    partial ShaderResourceGroup PassSrg : PerPass // At least one partial block must define the semantic
    {
        Sampler m_sampler; // A Dynamic sampler.
    }
     
    partial ShaderResourceGroup PassSrg
    {
        float2 m_resolution;
        float4 ReadColor(float2 uv)
        {
            return m_texture.Sample(m_sampler, uv);
        }
    }
     
    PSOutput MainPS(VSOutput IN)
    {
        PSOutput OUT;
        float2 uv = IN.m_position.xy / PassSrg::m_resolution;
        OUT.m_color = PassSrg::ReadColor(uv);
        return OUT;
    }

Example (Error): partial, emission point Error

    ShaderResourceGroupSemantic PerPass
    {
        FrequencyId = 1;
    }
     
    // This is the first encounter of a "partial" block for "PassSrg".
    // The whole SRG will be emitted here in the generated HLSL.
    partial ShaderResourceGroup PassSrg : PerPass
    {
        float4 m_color;
    }
     
    class SomeClass
    {
        float4 m_colorOffset;
        float4 GetColorOffset()
        {
            return m_colorOffset;
        }
    };
     
    // This is the second encounter of a "partial" block for "PassSrg",
    // All of this code will be emitted at the first encounter.
    partial ShaderResourceGroup PassSrg : PerPass
    {
        // Because the declaration of this variable will be emitted before
        // class SomeClass is defined, it'll appear as if this SRG is referencing an undefined class.
        SomeClass m_instanceOfClass;
     
        float4 GetColor()
        {
            return m_color + m_instanceOfClass.GetColorOffset();
        }
    }
     
    PSOutput MainPS(VSOutput IN)
    {
        PSOutput OUT;
        OUT.m_color = PassSrg::GetColor();
        return OUT;
    }

Example (Solution): partial, emission point success

    ShaderResourceGroupSemantic PerPass
    {
        FrequencyId = 1;
    }
     
    class SomeClass
    {
        float4 m_colorOffset;
        float4 GetColorOffset()
        {
            return m_colorOffset;
        }
    };
     
    partial ShaderResourceGroup PassSrg : PerPass
    {
        float4 m_color;
    }
     
    partial ShaderResourceGroup PassSrg : PerPass
    {
        SomeClass m_instanceOfClass;
        float4 GetColor()
        {
            return m_color + m_instanceOfClass.GetColorOffset();
        }
    }
     
    PSOutput MainPS(VSOutput IN)
    {
        PSOutput OUT;
        OUT.m_color = PassSrg::GetColor();
        return OUT;
    }

Root Constants

AZSL supports the definition of shader constants that conform to DirectX 12’s concept for root constants .

Root constants are declared with the rootconstant keyword. AZSLc transparently embeds root constants into a ConstantBuffer. They are useful to define shader constants that you frequently access. Because they are located in the root signature they can be read instantly without requiring additional levels of indirection.

Shader constants qualified as rootconstant are always declared globally.

Example:

rootconstant float4x4 s_objectMatrix;
rootconstant uint s_materialIndex;

Consider using root constants sparingly due to limitations with root signatures, the space to define root constants. Root signatures are shared between root constants, root descriptors, and descriptor tables. In Direct3D 12, a root signature is limited to 64 DWords (256 bytes of space).

Shader Variant Options

Shader variant options are declared as global variables with the option keyword. Their ultimate purpose is to be used as statically defined flags. It is helpful to think of Shader variant options as compile time configurable C Macros.
By default, during shader asset compilation, all shader variant options are compiled as global shader constant flags that can be used for conditional branching. For the sake of performance, You can define the values of each shader variant option at compilation time, and thanks to shader compiler optimizations the conditional branches are removed in favor of static execution of some sections of the code.

Shader variant options can be of type bool, int, or enum. They cannot be of type float and struct.

int types require an attribute with the inclusive minimum and maximum range of values.

Example:

    option bool o_useIBL;
    option bool o_useShadows = true;
     
     
    option enum QualityT {Low, Medium, High} o_quality;
    option enum class SurfaceT {Metal, Plastic, Hair} o_surface = SurfaceT::Metal;
     
    enum LightContributionT {None, Diffuse, Specular, Both};
    option LightContributionT o_directLightContribution;
    option LightContributionT o_indirectLightContribution;
     
    [[range(3, 16)]] // This integer option accepts values between 3 and 16 (both ends included).  
    option int o_numberOfTaps;

For more information on shader variant options, refer to AZSL: Design Principles And Purpose Of Shader variant options..
Even though the Shader variant options are declared globally, they are actually encoded in a single array of bits as a member variable of one, and only one, SRG.
To learn more about how Shader variant options are encoded when compiled, please read: Shader variant options & The Fallback Key.

The special attribute [[pad_to(N)]]

The [[pad_to(N)]] attribute pads the data with dummy variables to guarantee that the offset of the next variable starts with the desired alignment. The N argument is a single integral literal that must be a multiple of 4.

In DX12, there can be alignment problems for variables inside struct definitions. The variable’s layout, offset, and size changes depending on if the struct is part of a ConstantBuffer or a StructuredBuffer. To prevent this problem, you can use the [[pad_to(N)]] attribute to enforce data alignment.

Example:

struct MyStruct
{
     float m_data;
     float4 m_arr[2];
};

For ConstantBuffer case you’ll get these offsets:

  • float m_data; at offset: 0
  • float4 m_arr[2]; at offset: 16

For StructuredBuffer case you’ll get:

  • float m_data; at offset: 0
  • float4 m_arr[2]; at offset: 4

To resolve the example above, use [[pad_to(N)]] attributes to pad the data. This guarantees that struct MyStruct has the same layout, regardless of whether its a ConstantBuffer or StructuredBuffer.

struct MyStruct
{
     float m_data;
     [[pad_to(16)]]
     float4 m_arr[2]; //Now, this member variable will always start at offset 16.
};

In this example, MyStruct is forced to be of size 64 bytes:

struct MyStruct
{
     float m_data;
     [[pad_to(16)]]
     float4 m_arr[2]; //Now, this member variable will always start at offset 16.
     [[pad_to(64)]] // This guarantees the sizeof(MyStruct) to be 64.
};

Class method, late declaration member access

This scenario is a limitation of the reference tracking system internal to AZSLc: the seenat more documentation .

class C
{
	void f()
	{
		a;
	}
	int a;
};

with CLI switch --dumpsym will result in:

Symbol '/C/a':
  kind: Variable
  references:
  line: 7
  type:
    core: {name: "?int", validity: found, tclass: Scalar, underlying_scalar: int}
    generic: <NA>
  storage: 
  array dim: ""
  has sampler state: no

The field references is empty, which means AZSLc didn’t detect that a line 5, is the a delcared line 7. Let’s look at the swapping of declaration sites of f and a:

class C
{
  int a;
	void f()
	{
		a;
	}
};

results in:

Symbol '/C/a':
  kind: Variable
  references:
    - {line: 6, col: 3}
  line: 3
  type:
    core: {name: "?int", validity: found, tclass: Scalar, underlying_scalar: int}
    generic: <NA>
  storage: 
  array dim: ""
  has sampler state: no

Now it does see the access, at line 6, column 3 (after 2 tabs).

The problem happens because AZSL does not parse in 2 sweeps, it has only one pass. f()’s content parsing happens before a’s declaration, so ‘a’ apparition in f() is an internal unresolved reference. You can see it, using --W3 warning level. The compiler will output:

source.azsl(5,3) : warning: undeclared sub-symbol in idexpression: a

This shouldn’t be a problem because AZSLc doesn’t change the layout of the inner section of classes, so normally doesn’t need to be successful identifying class fields. But for structures or enums it would be another story, since they can be used as constant buffer resources, or option values, they may undergo reflection and scope migration + name mutation and collision avoidance during transpilation to HLSL. So it would be a problem if such a tracking miss would happen to a symbol that undergo any of that. Normally it shouldn’t happen since structs can’t have methods in AZSL. But for safety, you can simply apply the following coding rule: always define your methods out-of-class (we call this, deported definitions). Use one-line declarations in-class, and code the body of the functions after the end of the class declaration.