Using SAS in FX and CgFX File Formats


Current Version: 1.03, 29 March 2008

For a PDF version of this document, click here.


What, Why, How

This document covers the use of SAS (that is, “Standard Annotations and Semantics,” and usually pronounced “sass”) in effect files for either of the two popular text-based “.*fx” formats – either .fx (for DirectX only, and based on HLSL) or .cgfx (for use with either DirectX or OpenGL, and based on Cg). Both formats are extremely similar, and the SAS details are identical between the two[1].

SAS support can be found in multiple programs. To date, NVIDIA FX Composer is the shading tool with the most comprehensive support of SAS, and will be used as a baseline example. This document will also cover annotations and semantics that are unique to FX Composer.

This is not a general shading tutorial. It assumes that you can be comfortable with C++ -like languages, and in general doesn’t try to teach you about color and light, nor HLSL or Cg except in the context of SAS. For a general introduction to hardware shading languages, try “The Cg Tutorial,” which is available both at booksellers and at the NVIDIA developer web site, //

Effect files (that is, shader files with the suffixes “.cgfx” or “.fx”) try to describe a complete setup for real-time rendering. That means they include not only the code for the specific vertex and pixel shader functions that will run on the GPU, but also definitions of the contexts for these shader functions: the required graphics states; the user parameters that control them (whose values may be controlled by a UI, or automatically provided by the rendering program); and in many cases, the multiple rendering passes that together draw a complete, final image. In some advanced cases, as we’ll describe, these multiple passes may be executing in varying order – and in some cases there can be some limited interactions between multiple effect files.

Control over all of these interactions is provided by SAS.

Semantics and Annotations

Just so that we are clear on what semantics and annotations are, consider this typical parameter declaration, drawn from the NVIDIA shader library’s “plasticP.fx” effect[2]:

float3 gLamp0Pos : Position <

string Object = "PointLight0";

string UIName = "Lamp 0 Position";

string Space = "World";

> = { -0.5f, 2.0f, 1.25f } ;

In this example, global variable gLamp0Pos has a semantic, three annotations, and a value. All three are optional – at a minimum, all that’s required is a data type and parameter name. In general, parameter declarations are of the form:

type name : semantic <annotation_list> = value ;

which can cover one or many lines, up to the closing semicolon.

A bare minimum declaration might be something like:

float gCurrentScore;

Where the host program (e.g., FX Composer) will insert defaults for the semantic (none), annotation_list (none) and value (in this case, probably a zero).

Why Semantics and Annotations?

In general, semantics describe specific functional parts of the hardware or graphics API. As use cases of these basic types grew, more-general annotations were added to allow more-complex specifications of ideas such as “how to represent a color in the UI?”, a “metaconcept” that goes beyond the hardware itself. To understand these ideas in practice, let’s look at how just semantics were initially used to describe hardware connections between the CPU and GPU.

During the initial development of Cg and HLSL, semantics were use to describe specific GPU hardware registers. For example, this structure declaration describes the data that an effect expects in the vertex buffer before executing a vertex shader function:

struct appdata {

float3 Position : POSITION;

float4 UV : TEXCOORD0;

float4 Normal : NORMAL;

float4 Tangent : TANGENT0;

float4 Binormal : BINORMAL0;


Each of the semantics in this structure declaration indicate a specific part of the hardware memory that’s fed to vertex shaders that use input values of type “appdata.” Here’s the vertex shader from that same file (along with the declarations of its output data).

struct vertexOutput {

float4 HPosition : POSITION;

float2 UV : TEXCOORD0;

// The following values are passed in "World"

// coordinates since it tends to be the most flexible

// and easy for handling reflections, sky lighting,

// and other "global" effects.

float3 LightVec : TEXCOORD1;

float3 WorldNormal : TEXCOORD2;

float3 WorldTangent : TEXCOORD3;

float3 WorldBinormal : TEXCOORD4;

float3 WorldView : TEXCOORD5;



vertexOutput std_VS(appdata IN,

uniform float4x4 WorldITXf,

uniform float4x4 WorldXf,

uniform float4x4 ViewIXf,

uniform float4x4 WvpXf,

uniform float3 LampPos) {

vertexOutput OUT = (vertexOutput)0;

OUT.WorldNormal = mul(IN.Normal,WorldITXf).xyz;

OUT.WorldTangent = mul(IN.Tangent,WorldITXf).xyz;

OUT.WorldBinormal = mul(IN.Binormal,WorldITXf).xyz;

float4 Po = float4(,1);

float4 Pw = mul(Po,WorldXf);

OUT.LightVec = (LampPos -;


OUT.UV = float2(IN.UV.x,(1.0-IN.UV.y));

#else /* !FLIP_TEXTURE_Y */

OUT.UV = IN.UV.xy;

#endif /* !FLIP_TEXTURE_Y */

OUT.WorldView = normalize(ViewIXf[3].xyz -;

OUT.HPosition = mul(Po,WvpXf);

return OUT;


The semantics specify hardware registers, and their usage is also defined by context. You may have noticed that both the input structure AND the output structure contain members with the POSITION semantic – the context (in this case, being compiled as a vertex shader) makes the difference clear to the compiler and host application. If you tried to compile this function as a pixel shader, it would fail.

Here is a pixel shader from the same effect file:

float4 plasticPS(vertexOutput IN,

uniform float3 SurfaceColor,

uniform float Kd,

uniform float Ks,

uniform float SpecExpon,

uniform float3 LampColor,

uniform float3 AmbiColor

) : COLOR {

float3 diffContrib;

float3 specContrib;




float3 result = specContrib +

(SurfaceColor * diffContrib);

return float4(result,1);


Here the function itself has been assigned a hardware semantic – for a pixel shader, that COLOR semantic (or the synonym COLOR0) means the final output color of the pixel as the image is delivered to the final render target (either to a render surface or the frame buffer).

There are a variety of syntactical ways to do this – some users may prefer to use “out” parameters rather than directly returning a value, or they may not like to use structs, or they may prefer global-scope variables as opposed to passing formal parameters. Here’s a version using a “void” function and returning the COLOR via an “out” parameter instead:

void plasticPS_v2(vertexOutput IN,

uniform float3 SurfaceColor,

uniform float Kd,

uniform float Ks,

uniform float SpecExpon,

uniform float3 LampColor,

uniform float3 AmbiColor,

out float4 FinalColor : COLOR0

) {

float3 diffContrib;

float3 specContrib;




float3 result = specContrib +

(SurfaceColor * diffContrib);

FinalColor = float4(result,1);


See the Appendix “Alternative Constructions” to see some possibilities. They are all, however, functionally identical. The same semantics lead to the same hardware paths.

The semantic names describing standard hardware, such as TEXCOORD0, are understood to have specific mappings to specific hardware registers, as defined by the API (DirectX or OpenGL). In 99.9% of the cases, these are the names you should use. For some very old game or CAD engines (think: 20th Century code), you may need to specify exactly the hardware register you need. In these rare cases, you can use the register semantic. The register semantic looks like a function: e.g., register(c1) means “color register 1.” This semantic’s use is now fading into oblivion.

Other Hardware Semantic Names

The sample shown has already covered the most common hardware semantics. There are a couple of other rarely-used ones, such as the polygon-facing flag (VFACE in DirectX, FACE in OpenGL) and the PSIZE attribute of particles. Both entities are well-documented for their specific APIs.

Semantics from software API’s

Semantics for hardware registers have connections defined by the overall graphics state. What about values that should be connected to the surrounding software environment? For example, how can we bind the world location of the current object to a value, or the color of the current light source?

DirectX and the CgFX runtime both provide several semantics relating to just these issues. These semantics can hint to the renderer about connecting colors and positions to existing parts of the scene, and they call tell the renderer to load data from the scene into specific parameters.

All of these semantics will be assigned to *fx global parameters – never to functions. When a parameter is bound to a scene element (either automatically or by the user), its values will then be automatically assigned once each frame – so the user will not be able to tweak those values (say, in the FX Composer properties pane) except by altering the underlying scene. Once bound, they are “untweakable.”

Some hints are stronger than others, depending on the renderer. In general, the semantics indicating transform matrices are strong and universal: they appear to bind automatically in all renderers. Semantics for light positions, surface colors, and so forth: unpredictable! Sometimes it’s difficult for a renderer to know which light you want, if you are trying to assign a color from an object or a light, etc.

Typical semantics used on parameters will correspond to the values of the old DirectX “hardware transform and lighting” days – for example the DIFFUSE semantic is intended to correspond to the “Diffuse” value in a DirectX material structure and assigned in a DirectX9 program by IDirect3DDevice9::SetMaterial(). In general, these semantics relating to old-style, pre-pixel-shader-days material and light definitions (others include SPECULAR, or AMBIENT – if you look at the member names of the D3D Material structs they will look surprisingly familiar…) are being used less and less, as they are references to an antiquated material definition scheme (if you do want to bind to such values, there is also a newer way, using annotations to bind to an object or light – in such cases, use the semantic COLOR and see the section “Light and Object Binding” – there are additional semantics associated with lights and cameras that will be described in that section).

Vectors indicating the position or direction of a lamp can be distinguished by the POSITION or DIRECTION semantics, which can also let the renderer know what sorts of lights can be successfully bound to which parameters. (Later on, see info on the special FX Composer “dirpos” use of the POSITION semantic for float4 vectors, which can dynamically manage both point and directional lamp types at run time).


Transforms never go out of style. Both fx and cgfx formats follow the DirectX definition of a matrix stack:

Object → WorldViewProjection

Points in the vertex buffer are in “Object” space, which can be transformed to a shared “World” coordinate system. Transforming from “World” to “View” space sets the points to be located relative to the camera’s location and orientation, and the final matrix transforms them into projected screen space.

Those last three transforms can be used as semantics, and can also be (in the order shown) concatenated: That is, you can have semantics WORLD, VIEW, PROJECTION, WORLDVIEW, VIEWPROJECTION, or WORLDVIEWPROJECTION. Each will provide a single (possibly concatenated) matrix.

(Why no “OBJECT” semantic? Because it will always be an identity matrix!)

In addition, the optional “operational” suffixes INVERSE and TRANSPOSE can be added to the mix: VIEWINVERSE is useful for determining camera locations in WORLD coordinates, for example (while an un-inverted VIEW matrix would move points the opposite way, from WORLD coordinates into VIEW coordinates). The ultimate automatic transform name could be the intimidating agglutination of all transforms and operations: WORLDVIEWPROJECTIONINVERSETRANSPOSE.

Generally, a small number of transforms will get you what you need for any specific effect. Here are the global transform variables declared by the example “plasticP.fx” file:

float4x4 gWorldITXf : WorldInverseTranspose;

float4x4 gWvpXf : WorldViewProjection;

float4x4 gWorldXf : World;

float4x4 gViewIXf : ViewInverse;

(In this case, we are defining the View according to the primary scene camera. What about spotlight shadow cameras? These are specified by the Object annotation. This specialized usage will be covered in the section on shadowing.)


Annotations appeared primarily because rendering applications needed a method to specify the UI for specific effect parameters. While an application could guess, either by name or by reference to a semantic, the results are unpredictable. It seems obvious that a float3 parameter with the DIFFUSE semantic is a color, but what about a float3 called, say, “Mixer”? Is it a location, a color, something else?

Annotations eliminate the guessing by providing a mechanism to specify UI details explicitly. Any type of metadata can be attached to any parameter, for whatever use you desire. If an application doesn’t recognize the specific annotations, it can just ignore them (for FX Composer users, unrecognized or user-defined attributes can still be queried from Python).

Here is another example from our sample effect:

float gKd <

string UIWidget = "slider";

float UIMin = 0.0;

float UIMax = 1.0;

float UIStep = 0.01;

string UIName = "Diffuse Strength";

> = 0.9;

As you can see annotations like UIWidget and UIMax require a type declaration as well as their own name. In this example, the annotations are purely related to the UI for the parameter “gKd” – they say that it should be displayed as a fine-grained slider with a range from zero to one, and that its displayed name should not be the cryptic-looking (to a non-shader-programmer) “gKd,” but the more human-readable “Diffuse Strength.”

Annotations can also be added to technique and pass definitions, which will be important later.

Annotations without End?

Annotations can define anything that a rendering program might want – if you want to add a string “author” or “notes” annotation to a parameter, you can – it will work fine, though as the creator of that effect file you will need to provide a way to identify and interpret those strings yourself.

While annotations let you add arbitrary (meta)data, they also invite a Babel-like explosion of arbitrary duplications of the same purposes – say, five different definitions of a “slider” UI element – to avoid this hazard, standard annotations and semantics were defined[3].

Standard Annotations and Semantics

At its core, SAS simply defines a standard list. While any rendering application can use semantics and annotations as they see fit, following the standard list gives everyone the best chance of broad compatibility.

SAS Parameter UI Annotations

For most programs, the key annotations are few, and related to UI. For most user parameters, it’s a good idea to provide them with a name and a choice of UI widget. This can be done using the obviously-named annotations UIName and UIWidget.

We saw the UIName annotation above. It provides a human-readable description for sometimes cryptic variable names.

Next is the UIWidget annotation, which we also saw in the example. The UIWidget string has three especially useful values: “Slider,” “Color,” and “None.”

We saw “Slider,” as well as the related UIMin, UIMax, and UIStep annotations that are unique to slider widgets. Here is a “Color”:

float3 gSurfaceColor : DIFFUSE <

string UIName = "Surface";

string UIWidget = "Color";

> = {1,1,1}; //white

The parameter will be be displayed with the name “Surface” (which, since it’s being displayed as a color, is obviously “Surface Color”).

The “None” value for UIWidget completely suppresses display of the parameter. It’s principally useful for parameters that are automatically-bound to an element from the scene, such as a transform. There’ no need to clutter-up the properties pane with a lot of changing but un-tweakable numbers (“number noise”). The full matrix declarations from “plasticP.fx” actually include this semantic for just that reason:

float4x4 gWorldITXf : WorldInverseTranspose <

string UIWidget="None"; >;

float4x4 gWvpXf : WorldViewProjection <

string UIWidget="None"; >;

float4x4 gWorldXf : World <

string UIWidget="None"; >;

float4x4 gViewIXf : ViewInverse <

string UIWidget="None"; >;

Any rendering application may have its own special cases – for example, the standard Maya Cg plugin only allows float4 parameters to be displayed as colors. Some applications may have special UIWidget values that are not widely recognized (say, an enumerated pulldown). But the standard UIWidget name should still be used.

Most of the other SAS parameter semantics and annotations deliver specifically-detailed information about the current graphics state back to the effect file.


Textures have a small number of semantics but can be complex in combination. The most important semantics are those that designate a texture as a writable render target: RENDERCOLORTARGET or RENDERDEPTHSTENCILTARGET. We will discuss Render Targets in greater depth as part of the discussion on SAS Scripting.

Occasionally you will find textures marked as DIFFUSE, DIFFUSEMAP, NORMAL, SPECULAR, SPECULARMAP, ENVMAP, ENVIRONMENTNORMAL, or ENVIRONMENT. These provide hints to the API about automatic binding, but these usages are disappearing over time (again, they are mostly useful in mapping fixed-function values to shader values, and are now antiquated).

For disk textures, the annotations are simple – usually just a UIName and a ResourceName (described below) to identify the disk file. But for Render Targets, the annotations define many of the most-crucial aspects of texture definition and creation.

Texture Annotations, Sampler States

FX formats require two parameters for texture use: the declaration of the texture data, and the declaration of at least one sampler object that will pass the texture data to the shader(s). Textures can have annotations, while samplers have a state declaration that happens to look like annotations – but is not.

Sampler states are enclosed in { braces }, while annotations use < angle brackets > – and more tellingly, sampler states are predefined entities – unlike annotations sampler state declarations never require a data type.

Textures, however, may have annotations. Look at the following pair, a texture declaration followed by a matching sampler declaration:

texture gEnvTexture : ENVIRONMENT <

// These are annotations

string ResourceName = "";

string UIName = "Environment";

string ResourceType = "Cube";



samplerCUBE gEnvSampler = sampler_state {

// These are not annotations
Texture = <gEnvTexture>;

MagFilter = Linear;

MinFilter = Linear;

MipFilter = Linear;

AddressU = Clamp;

AddressV = Clamp;

AddressW = Clamp;


The following annotations are supported for textures. Some may not be meaningful depending upon the context: ViewportRatio has no influence on disk-based textures, for example (while conversely ResourceName means nothing to a RenderTarget):

      • ResourceName (string): The name of a file that should be read by default, if not overwritten by material. Meaningless for RenderTargets.
      • ResourceType (string): The type of disk file to seek when loading textures. Can be “2D,” “3D,” or “CUBE.” Meaningless for RenderTargets.

The annotations below are unique to texture declarations where the texture data will itself be created by the effect, rather than loaded from disk:

      • ViewportRatio (float2): The dimension of a RenderTarget texture expressed as ratios relative to the viewport.  “float2 ViewportRatio = {1,1}; results in the same dimensions as the screen.  “float2 ViewportRatio = {0.5, 1.0};” would result in a texture ½ as wide as the render window, but just as tall. Mutually exclusive with Dimensions, Width, Height, and Depth annotations.
      • Dimensions (int 2-3): The expected width, height, & optional depth for a procedural or RenderTarget texture.  Mutually exclusive with ViewportRatio, Width, Height, and Depth annotations.
      • Width (component of dimensions) (int): Assigns the width of a procedural or RenderTarget texture. Mutually exclusive with Dimensions and ViewportRatio annotatons.
      • Height (component of dimensions) (int): Assigns the width of a procedural or RenderTarget texture. Mutually exclusive with Dimensions and ViewportRatio annotations.
      • Depth (component of dimensions) (int): Assigns the depth of a procedural texture. Mutually exclusive with Dimensions and ViewportRatio annotations.
      • Format (string): Pixel format. Ignored for disk-based textures.
      • Miplevels (int): The number of mip levels to create.
      • Levels (int): A synonym for MipLevels.
      • Function (string): The name of the virtual-machine function used to procedurally generate this texture (HLSL only). See the appendix section “Texture Shaders in the HLSL Virtual Machine” for details.

FX Composer-specific note: Render Targets are often hidden from the list of material properties (by setting a UIWidget annotation to “None”). If you want to use a shared Render Target, make sure not to hide the texture, so that it can be assigned to the shared surface by the users of your effect. For more details, see the Appendix section on Sharing Textures.

For DirectX9 effects, it is also possible to have the CPU generate texture data dynamically at effect-load time. This requires a special syntax, which you can read about in the Appendix labeled “Texture Shaders in the HLSL Virtual Machine.”

Viewport Semantics

While not directly connected to texture data, knowing the size of the current render viewport is crucial for render-to-texture effects. The VIEWPORTPIXELSIZE semantic should be attached to a float2 parameter and provides pixel-count information about the size of the current scene render window.

The “Quad.fxh” file used by many of the Shader Library FX samples defines the following global variable, used widely by full-screen library shaders:

float2 QuadScreenSize : VIEWPORTPIXELSIZE <

string UIName="Screen Size";

string UIWidget="None";


Individual pixels/texels will be of the size (float2(1.0,1.0)/QuadScreenSize) – and importantly (a fact already factored into the vertex shaders defined in the same header file), this calculation can be used to adjust for the half-texel origin offset in DirectX.

Light and Object Binding

Often, scales, colors, and transforms need to be bound to a specific lamp or other object in the current scene. Annotations and semantics, working together, provide a flexible mechanism for accomplishing this. A semantic can be used to specify the kind of data required – say, a COLOR – while an annotation (usually Object) is used to describe where to find that kind of information.

Depending on the kind of object, a semantic may make sense or not. In general, if the Object named includes the substring “light,” then the data should be appropriate for a light. If instead it includes the substring “camera,” then the data should be appropriate for a camera. Otherwise, it’s just a user-defined name & there’s no rigid guarantee that the value can be bound sensibly.

Coordinate Systems

For bound data that is spatial, getting the data in the appropriate coordinate system is important. Currently only POSITION and DIRECTION semantics support the Space annotation, which can be assigned as “World,” “View,” or “Object.”

Light Names

Typical light names might be something like “Spotlight0” – FX Composer will keep track of all such names encountered, and permit the user to bind any light to the symbolic name “Spotlight0” – with correct propagation of values to any and all parameters that use the name.

Light names can be used with either Object or Frustum annotations, as we’ll see below.

Light Semantics

The following semantics can be used sensibly with lights. Specify the light in the Object annotation and (for geometric values) be sure to specify a desired Space for coordinates.












As an example:

float gCone1 : FALLOFFANGLE <

string Object = “Spotlight1”;


The values associated with lights within FX Composer can be seen by selecting the light and its properties. They can be edited there, if not already bound by animation (say, from a 3ds Max clip).

Viewing transforms associated with shadow maps are sometimes bound using the Frustum annotation rather than Object. FX Composer supports both usages. See the section “Shadow Transforms” for an example.

Camera Semantics

Only “direct” camera transform transforms are currently supported by FX Composer – that is, you can grab the VIEW matrix, but not WORLDVIEW or VIEWINVERSE. These combinations can be coded either in the vertex shaders or (in DirectX9) as “static” declarations. See the NVIDIA Shader Library header file “nvMatrix” for an array of available matrix operations.

POSITION and DIRECTION semantics are also supported for cameras.

Scaling Info

How big is an object? The shading of a 2-inch block of wood and a 20-foot one are likely to be different. The UNITSSCALE semantic indicates to the rendering program that the indicated value can be mapped to some external or real-world measurement. Here’s an example from the Shader Library’s “cage.fx”:

float gScale : UNITSSCALE <

string units = "inches";

string UIWidget = "slider";

float uimin = 0.0;

float uimax = 20.0;

float uistep = 0.01;

string UIName = "Size of Pattern";

> = 5.1;

The host render may define its own default value of for units if it is not supplied by the annotation.


The TIME semantic provides absolute system time. In seconds. As this makes for a rather large unreadable number, such parameters are usually marked as hidden using a UIWidget annotation.

Absolute system time is best used for “self-animating” effects like fire, spinning fighter propellers, water, etc. Several NVIDIA Shader Library effects use it, such as “MrWiggle,” “Ocean,” or “post_corona.”

To grab frame-count information from animated clips, consider FX Composer’s “playblast” Python module.

Float values tagged with the ELAPSEDTIME semantic return the time since the previous frame – great for calculations such as those used in motion blur algorithms.

For FX Composer users, TIME is also often important when dealing with interactivity, via the mouse – which will be described in the next section.

FX Composer-Specific Semantics

A handful of non-SAS semantics (or beyond-SAS uses of existing semantics) are supported specifically to reflect the functionality of NVIDIA FX Composer.

Mouse Position and State

FX Composer permits querying the position of the mouse, the states of the mousebutton, and the time when those buttons were last pressed.

MOUSEPOSITION parameters should be float2. They will report the current location relative to the render window.

LEFTMOUSEDOWN & RIGHTMOUSEDOWN provide four bits of info as a float4: if the button was pressed; where the mouse was when it was pressed as an XY pair just like MOUSEPOSITION; and when the button was pressed, in the same format as used by TIME parameters. For either button semantic, the format will be:

XY: Where the mouse was when pressed.

Z: Is the button pressed? One or zero.

W: The system time when the button was last pressed.

Renderer Reset

When the RenderWindow is reset, either by being resized, a new scene has started, etc., then “bool” parameters with the FXCOMPOSER_RESETPULSE semantic will be set to true (1) – for exactly one frame. At the end of that first new frame they’ll flip back to false (0). This will be important later on when we discuss full-screen effects.

Combining Directions and Positions

Some game and rendering engines like to use the nature of homogenous data to let the same shader accept light data from either point lights or directional lights using the same vector. That is, for a float4 vector, the w component indicates if the data is a point in space (nonzero w) or a vector though space (zero w). FX Composer supports this idea, and float4 vectors can be tagged with the semantic POSITION. FX Composer will know how to bind both point and directional lights to such values, and will set the w component appropriately[4].

As a momentary shading aside, here is a typical vertex shader for such “dirpos” data:

vertexOutput std_dp_VS(appdata IN,

uniform float4x4 WorldITXf,

uniform float4x4 WorldXf,

uniform float4x4 ViewIXf,

uniform float4x4 WvpXf,

uniform float4 LampDirPos // ”dirpos” POSITION

) {

vertexOutput OUT = (vertexOutput)0;

OUT.WorldNormal = mul(IN.Normal,WorldITXf).xyz;

OUT.WorldTangent = mul(IN.Tangent,WorldITXf).xyz;

OUT.WorldBinormal = mul(IN.Binormal,WorldITXf).xyz;

float4 Po = float4(,1);

float4 Pw = mul(Po,WorldXf);

if (LampDirPos.w == 0) { // “dirpos” switch HERE

OUT.LightVec = -normalize(;

} else {

// we are still passing a (non-normalized) vector

OUT.LightVec = -;



OUT.UV = float2(IN.UV.x,(1.0-IN.UV.y));

#else /* !FLIP_TEXTURE_Y */

OUT.UV = IN.UV.xy;

#endif /* !FLIP_TEXTURE_Y */

OUT.WorldView = normalize(ViewIXf[3].xyz -;

OUT.HPosition = mul(Po,WvpXf);

return OUT;


The float3 result in “OUT.LightVec” correctly preserves the direction and magnitude for later light calculations in the pixel shader, regardless of lightsource type.

The Two SAS’s

Two different versions of SAS exist. The version 1.0 spec, which is officially supported by Microsoft in the latest DirectX9; and the 0.86 pre-release, which is the version used by FX Composer and several other applications (for DirecctX9, DirectX10, and OpenGL). In general, parameter semantics and annotations like WORLDTRANSPOSE and UIWidget are identical in both versions of SAS. The principal difference between the two? Scripting.

SAS Scripting

SAS 0.86 includes support for the scripting of techniques and passes (while DirectX9’s version 1.0 does not).

Why would you want to script techniques and passes?

Rendering occurs in a series of discrete steps. Often, only one pass is needed to render an object (or entire scene). Sometimes, several passes are accumulated or otherwise combined. By default, the passes are simply executed in order, one after another. If that’s all you need, then scripting isn’t required. But it often is! For example, you may have an iterative technique that requires looping on a single pass. Or you may want to use a shadow map, which needs to be rendered to an offscreen buffer with a different render size and transform from the regular scene camera.

Besides ordering passes, scripting allows you to set the desired priority between techniques (for effects that contain multiple techniques), and to pass data to the rendering program to let it know when to execute your effect in relationship to other effects in the scene.

In short: scripting lets effects mimic the variety of rendering patterns found in real game engines.

The Script : STANDARDSGLOBAL Parameter

In “plastic.fx,” you may notice this block:

float Script : STANDARDSGLOBAL <

string UIWidget = "none";

string ScriptClass = "object";

string ScriptOrder = "standard";

string ScriptOutput = "color";

string Script = "Technique=Technique?SimplePS:TexturedPS:SimpleQuadraticPS:TexturedQuadraticPS;";

> = 0.8;

This special “parameter” Script, when semantically tapped as STANDARDSGLOBAL, is expected by the renderer to contain annotations to describe the overall effect.

The value 0.8 specifies the SAS version.

The UIWidget hides this variable (it isn’t ever actually called by the shader functions).

The ScriptClass, ScriptOrder, and ScriptOutput annotations indicate that this is a surface material (“object”) effect, that it follows standard use and it outputs colored pixels. So far so good and the expected default state -- in fact we could completely ignore and delete the Script parameter, except for the contents of the last Script annotation (yes, the Script variable has a Script annotation).

Script <Script> and Technique Selection

By default, *fx files will by select techniques according to the order in which those techniques appear in the file. If a technique is unavailable (say, because the user’s computer isn’t capable of running the specified profile), it will be skipped and the next technique will be tried. The renderer will proceed down the list until a working technique is found. In the example above, the Script annotation is letting the shader author explicitly name which techniques should be available, and the order in which they should be evaluated (and for renderers that allow user choice of multiple techniques, the desired order of the techniques in the UI).

The Script annotation can specify either a single technique, in the form: “Technique=name;” or the annotation can indicate a prioritized sequence of techniques, using the syntax: “Technique=Technique?nameA:nameB:nameC;” – that is, as a colon-separated list prefaced by the token “Technique?” – the list will explicitly prioritize the listed techniques.

In the example case, the Script specifies that the “SimplePS” technique is the correct default. In programs like FX Composer or Maya that permit technique selection, the preferred order of the displayed lists in those programs should be


If there are other techniques in the file (say, a vertex-shaded alternative, or a version specific to a particular game platform), they will not, by default, be displayed or available to the user of this effect.

(The shader library includes Script parameters and other SAS details in most shaders, even though they are often redundant. For example, there may be a Script definition like “Technique=Main;” for a file that only has one technique! This is because shader library samples are intended to provide templates for further exploration. They’re a bit deliberately “over-engineered” – we do the typing up front so you don’t have to later on.)


Let’s get back to those other annotations: ScriptClass, ScriptOrder, and ScriptOutput. The last is the easiest: ScriptOutput should always be “color” and nothing else. This is a practical assertion – while values like “depth” were once intended to be supported, they’ve never been used.

We’ll see the value of ScriptOutput later on, when discussing a Script command called “ScriptExternal.” For now, just trust us: it’s always “color.”


ScriptClass defines the purpose of this effect. Is it a material to be applied to objects? Then it should be “Object,” which is the default expectation. Some effects, however, are complex, and may require rendering to multiple render targets. They may in fact render no objects directly at all. Such effects should designate themselves as being in the “Scene” ScriptClass.

There is also a third choice for ScriptClass, “SceneOrObject.” Its purpose has been to provide a more complete choice of techniques. By default, it’s expected that “Object” effects don’t need to write to offscreen buffers, while “Scene” effects can. “Object” effects can be bound to geometry, while “Scene” effects are typically bound only to the entire frame. But sometimes you really do need both – for example, a material for terrain may have a simple version, but also a second, more-complicated version that renders shadow maps.


The ScriptOrder is meaningful for “Scene” and “SceneOrObject” effects. It has three possible values: “standard,” “preprocess,” and “postprocess.” You should normally not assign “preprocess” or “postprocess” to “SceneOrObject” effects – these two ScriptOrder choices are specifically intended for image processing.

The three choices allow some broad-scale control over effect ordering in the overall scene. Generally, the user may have little control over the order in which effects are executed – they just get fired-off as their objects appear in the buffer. The ScriptOrder lets this amorphous mass be split into three, smaller amorphous masses:

1.     The preprocess phase, which will always be rendered first and can include the creation of background images and (for COLLADA) shared textures;

2.     the standard phase, which will render next (and unless over-ridden, will include all object rendering);

3.     and the closing postprocess phase, which renders after all others to add screen-buffer effects like motion blur, color styling, and bloom.

There may be multiple preprocess and/or postprocess effects in a single scene. They will render in the same order as they were added to the scene, and always within their own ScriptOrder group.

This is most important for image-processing-based rendering. For example, a scene may include a preprocess effect to render a background, then the scene geometry is rendered on top of that, and then a postprocess effect adds contrast control and color tinting.

Pass Ordering and Screen Clears: Technique Scripting

As mentioned, we can attach annotations to individual techniques, and to individual passes within those techniques. These annoatation will control the render process.

If we just want to execute those passes in a straight-ahead, all-pixels-to-the-framebuffer fashion, then we don’t need to do anything special. But if we want additional control, then the Script annotation is useful here, too.

We can also manipulate when and how the screen is to be cleared while executing our effect (as you can imagine, it’s very dangerous to let object materials clear the screen! Other objects may disappear).

Pass ordering and framebffer control is managed by a script in the technique’s Script annotation. Here’s one of the longest in the library, from “scene_uvds_skin.fx”:

technique Main <

string Script =











> {

// …rest of technique and pass declarations….

As we can see, the Script contains a string with multiple commands (the per-line strings are automatically concatenated up until the final semicolon). Each of the commands itself has a semicolon (forgetting these is a common mistake), and each command is essentially a “leftside=rightside” assignment. In two cases, the rightside is empty. In others, the rightside contains the names of scene global parameters, and in others the names of passes, and in others keywords telling the renderer what to do. It’s admittedly a strange language-within-a-language. Fortunately, you’ve already seen 90% of it!

Each line represents a specific action to take. First, the passes named “MakeShadow,” “HBlur,” and “VBlur” are executed, in that order.

Then, we assign the “RenderColorTarget” (for DirectX9, this is analogous to calling IDirect3DDevice9::SetRenderTarget()). Having an empty rightside means “set to the current enclosing-scope default” (e.g., the frame buffer). Passes can also assign their own “RenderColorTarget” (and these ones do!), so this command redirects the renderer back to the main scene.

In the same way, we next redirect the “RenderDepthStencilTarget” to “” (that is, to the framebuffer).

Next, we assign the Clear color and the Clear depth to the values of the indicated global parameters, which had been declared earlier in the effect file. We’re not actually clearing any buffers here, we’re just setting some “state” values to be used when we do call “Clear,” Which is next.

The following two lines clear the buffer, twice. “Clear=Color” clears the “RenderColorTarget” with our values from gClearColor, while “Clear=Depth” clears the “RenderDepthStencilTarget” with the values from gClearDepth. So after these nine commands, we’ve drawn some buffers, come back, selected the main frame buffer, and cleared it.

Finally, we execute one more pass, named “useBakedLighting.” We’re done!

If there are other passes in the technique, they’re ignored (this means the technique Script is a good tool for trying different combinations while creating and debugging, by the way).

In this sample we’ve already seen that we can assign render and depth targets, clear the frame (which constitutes a sort of “hidden pass”), and control the frame order. Technique Script annotations can perform one extra bit of magic: Looping.


Here’s the technique Script from the Shader Library “paint_brush.fx” effect (which turns FX Composer into a very minimal paint program):

< string Script =






"Pass=splat;"; >

What it does, and how it does it:

The parameter bReset is a global parameter that’s controlled by the application-specific semantic FXCOMPOSER_RESETPULSE (described earlier). Here is its declaration:


string UIName="Clear Canvas"; >;

A parameter bound to FXCOMPOSER_RESETPULSE is set to one (true) if either the user selects “Clear Canvas” or if the canvas (render window) is reset – say, by a view-size change. It will be true for that one frame – on all subsequent frames, the value will be zero (false).

Okay, so we have a value in bReset that is one or zero. Now we can use the Script commands “LoopByCount” and “LoopEnd.” To give us a more elegant frame reset than just a simple “clear the screen” (in this case, our “revert” pass redraws the screen with a user-assigned background texture, which the user can then paint over).

“LoopByCount” looks at the value of the designated parameter and executes the commands up to “LoopEnd” exactly that many times. In this case, the count is either one or zero. If it’s zero, the enclosed commands are executed zero times – that is, the “revert” pass doesn’t execute at all.

This combination of FXCOMPOSER_RESETPULSE (or really, any bool parameter) and the “LoopByCount” Script command provides us with a somewhat kludgy but effective “if { }” mechanism.

Oh, and by the way, the indentation of the sub-strings in a Script means nothing – it just makes it easier for a person to read.

Let’s look at a more traditional use of looping, this time from “scene_reaction_diffusion.fx.” As before, we declare a parameter bReset, but also one called gIterationsPerFrame. Its declaration is (note the deliberately-chunky UIStep):

float gIterationsPerFrame <

string UIName = "Iterations Per Frame";

string UIWidget = "slider";

float UIMin = 0;

float UIMax = 100;

float UIStep = 1;

> = 10;

…and the technique Script is:

< string Script =












"Pass=display;"; >

Here we still keep our revert pass (which initializes an offscreen buffer). We also let the user actively choose how many iterations of reaction and diffusion (passes “simulate01” and “simulate10”) to perform each frame.

Our third and final looping example is from “scene_EasyBake.fx” and demonstrates the ability to push the loop count back into the shader function (in this case, to place instances of a model in varying locations). We declare a global value that will define the number of iterations, and also a global value into which we will be passing-back each current loop’s count:

float gThisInstance <

string UIWidget = "none"; >; // loop counter, hidden

then in the technique Script:

< string Script =











"LoopEnd;"; >

Here we see the loop iteration count set by gNInstances, and a new Script command, “LoopGetIndex” is used to set the value of the global parameter gThisInstance. That parameter can be used in any way by any passes (in this case, “useBakedLighting”) within the loop.

Here is the vertex shader declared by the “useBakedLighting” pass declaration itself:

VertexShader = compile vs_2_0 instancesVS(gXvpXf,




The value gThisInstance is passed along like any other, and “instancesVS()” can use the value to determine the position of each different instance.

Pass Scripting and Rendering to Texture

Pass scripting, like technique scripting, can assign render targets and execute clear commands. It can be tough to decide if you should assign the targets in the technique Script or in the pass Script – in general, the method chosen for the NVIDIA Shader library has been: passes that use off-screen targets assign their own targets, while handling of the global-scope framebuffer is done by the technique. There is no hard and fast rule, but in general that has been the easiest to manage and debug.

RenderTargets are a piece of the rendering state, and there is no state stack – that is, if a pass assigns a “RenderColorTarget,” that target won’t “pop” or reset when the pass is completed – its assignment will remain, until explicitly reset by the same pass, the technique, or another pass. Are you unexpectedly rendering a black frame? Make sure you’re rendering where you think!

(Note too that if you want your Script to reassign the RenderTarget after the Geometry has drawn, that’s also okay).

Passes can also specify what to draw in each pass – either the geometry from the scene, or a screen-aligned quadrilateral that will exactly fit the render window. We use the pass Script “Draw” command to chose: either “Draw=Geometry;” for (you guessed it) geometric models, or “Draw=Buffer;” for a full-screen quad. If neither is specified, a “Draw=Geometry;” will be implied at the end of your pass Script.

Here is a sample from the “scene_uvds_skin.fx” file:

pass HBlur < string Script =




> { //…

Both the color and depth Render Targets are assigned, to textures that have been defined using the RENDERCOLORTARGET or RENDERDEPTHSTENCILTARGET semantics.

After the target assignment, the full-screen quad is drawn.

On the next pass, any sampler that has been bound to the texture gBlur1Tex will be available for reading – as long as you change the “RenderColorTarget0” first (remember, it’s illegal to read and write to the same texture in the same pass)!

Examples from the Shader Library that draw full-screen quads usually include the NVIDIA “Quad” header file (either “include/Quad.fxh” or “include/Quad.cgh,” depending on the language). These header files include sample vertex shaders and pixel shaders to handle the “Buffer” quadrilateral and map previously-rendered textures exactly to the screen coordinates (which can be tricky, because of a half-pixel offset under DirectX).


When a reset event occurs (the same events that would trigger FXCOMPOSER_RESETPULSE): all RenderTargets may be potentially reset, resized, and flushed from memory. You must always do your own clearing of render targets, or they will be filled with either random junk or the results of the previous render.

This isn’t always a bad thing! Shader library effects like “post_corona” (which displays an animated “flaming” halo around objects) or “post_trail” deliberately use this feature.

Shadow Mapping

Shadow mapping a specialized kind of RTT that will introduce yet another Script command, “RenderPort.” We will also need to get some special matrices that describe our shadowing lamp’s relationship to the rest of the scene.

Shadow Transforms

We can use the regular automatic transform semantics, with one additional feature: the Object annotation. We assign Object to the “semantic name” of the light (see also the section above on Light and Object Binding):

float4x4 gLampView : View <

string Object =”SpotLight0”; >;

float4x4 gLampProj : Projection <

string Object = “SpotLight0”; >;

float4x4 gLampViewProj : ViewProjection <

string Object = “SpotLight0”; >;

Instead of basing their values on the default camea, they will base their values on a square camera controlled by the “SpotLight0” assignment (by default, the scene’s first SpotLight object).

Shadow Bias Matrices

These you need to construct yourself. The methods are different between HLSL and Cg – the Shader Library provides and uses a header called either “include/shadowMap.fxh” or “include/shadowMap.cgh” that can manage the details of shadow bias-matrix construction in an almost invisible manner.


To complete our pass, we need to let the renderer know to switch cameras and formats. We can use the “Renderport” Script command to do just that. “Renderport” can be set to the name of a lamp e.g., “RenderPort=SpotLight0;” and can be reset (that is, re-assigned to the master scene camera) by setting back to “” – “Renderport=” alone.

Remember to reset before your final pass!


Using multiple render targets (MRTs) is straightforward for renderers that support it.

Besides “RenderColorTarget0,” you may also assign “RenderColorTarget1,” “RenderColorTarget2,” etc in your Script – up to the limits of your system.

Remember to reset all the targets after rendering to them!

In the pixel shader, MRT values can be written using “out” values with semantics COLOR0, COLOR1, etc, as in this sample with three RenderTargets:

void prepMRTPS(vertexOutput IN,

uniform float3 SurfaceColor,

uniform sampler2D ColorSampler,

out float4 ColorOutput : COLOR0,

out float4 NormalOutput : COLOR1,

out float4 ViewptOutput : COLOR2)


float3 Nn = normalize(IN.WorldNormal);

NormalOutput = float4(Nn,0);

float3 Vn = normalize(IN.WorldView);

ViewptOutput = float4(Vn,0);

float3 texC = SurfaceColor *


ColorOutput = float4(texC,1);


Preprocess and Postprocess effects

Preprocess effects are simple. They have no geometry, and thus they should just “Draw=Buffer;”

Postprocess files are somewhat trickier as they introduce our final Script command: “ScriptExternal=Color;” which is only used for postprocess effects.

Here is the entire technique from the shader library “post_negative” effect:

technique Main < string Script =









> {

pass PostP0 < string Script =




> {

VertexShader = compile vs_2_0


ZEnable = false;

ZWriteEnable = false;

AlphaBlendEnable = false;

CullMode = None;

PixelShader = compile ps_2_a




At this point you should be able to read the entire Script, save for the “ScriptExternal=Color.” This technique has a single pass, but before that pass it assigns and clears color and depth RenderTargets, calls “ScriptExternal,” then calls the Pass “PostP0” – the pass in turn immediately resets the RenderTargets to “” and draws a fullscreen quad using the Sampler associated with the RenderTarget (gSceneSampler) that was used by the technique when it called “ScriptExternal.”

Just what did “ScriptExternal” render? Everything.

That is, everything in the scene that has come before this postprocess effect: first preprocess effects, then all geometries with their object effects. “ScriptExternal” redirects the output of these effects to the postprocess effect’s own RenderTarget, rather than the frame buffer (a difference tat’s entirely transparent to those other effects).

That’s it! For a complete list of commands to use in technique and pass Script annotations, see the very end of this document.


Alternate Constructions

The choices made for the formatting of NVIDIA Shader Library were made in the interest of general clarity. They’re not the only way to do it!

Shader parameter names have been chosen to be as consistent as possible between shaders, for example, so that it’s easy to copy-and-paste between different samples when creating your own new variations.

Structs have been used to indicate hardware-register and hardware-connector inputs and outputs. These struct definitions are shared between to most of the shaders, which eliminates a lot of typing and therefore a lot of potential for error.

Similarly, except for MRTs, library pixel shaders usually return a float4 and the function is marked with the COLOR0 semantic. Some users may prefer to use void functions and enumerate the outputs explicitly using “out” variables. Both are valid choices.

As in this document, semantics are usually CAPITALIZED to make them easy to spot.

Global scope: in many older shaders, globally-scoped parameters were used. This made the code clean in some ways, but made it difficult to copy and paste such shader code into other effect files or into game engines. Newer shaders, and revisions of many old ones, have switched to using formal parameter passing, and avoiding the direct global use of the shader parameters within functions. This makes for a bulkier-looking function, but also makes it much easier to drop such a function into other non-fx-based engines, or to share snippets between shaders.

To make global parameters easier to spot, we have adopted the “Hungarian” convention of prepending a small letter “g” before parameter names, and passing the values to shader functions by using the same name without the “g”: e.g., the float4x4 parameter “gWorldViewXf” would be passed to the appropriate vertex shader as “uniform float4x4 WorldViewXf

HLSL virtual machine: some HLSL effects use the virtual machine to generate textures (see “Texture Shaders in the HLSL Virtual Machine,” immediately below). Some HLSL effects also include uses of the HLSL “static” construct. Their corresponding CgFX versions (if any) cannot, as COLLADA lacks such a virtual machine. So those calculations are usually moved to the vertex and/or fragment shaders in the Cg versions. If you intend to use Shader Library code as a starting point for your game engine, be aware that sometimes simple calculations like calculations using TIME variables (which only need to be calculated once for the entire frame) can be effectively moved to the CPU for a fractional GPU boost.

We’re always interested in improving NVIDIA’s products. If you have ideas on improving these constructions that may increase their power and usefulness, let us know at the NVIDIA Developer Forums, on the web at //

Texture Shaders in the HLSL Virtual Machine

DirectX9 includes the notion of “texture shaders” – functions run on the CPU once, when the effect is loaded, to procedurally create texture data. The parameters of these shaders have their very own special usages of the POSITION, PSIZE, and COLOR semantics.

To use a texture shader, you write an HLSL function that will be iteratively executed on each of the texture MIP levels. It is much like a pixel shader function, and should return a value with the COLOR semantic. Automatic inputs to this function will have two special semantics: POSITION will contain XY data varying from 0 to 1, while PSIZE is used indicate the pixel (texel) size (thus allowing the texture function to alter its calculations according to the MIP level).

Here is a sample used by the Shader Library “stripe_tex.fxh” header file:

#define STRIPE_TEX_SIZE 128


float4 make_stripe_tex(

float2 Pos : POSITION,

float ps : PSIZE



float v = 0;

float nx = Pos.x+ps; // last column = 1 for all MIPs

v = nx > Pos.y;

return float4(v.xxxx);



texture gStripeTexture <

string function = "make_stripe_tex"; // our function

string UIWidget = "None";

float2 Dimensions = {





sampler2D gStripeSampler = sampler_state {

Texture = <gStripeTexture>;

MinFilter = Linear;

MipFilter = Linear;

MagFilter = Linear;

AddressU = Wrap;

AddressV = Clamp;


A cautionary note: Because they are run at effect-load time, the inputs to a texture shader must be constants –user parameters will have no control over the texture shader (since those parameters haven’t been displayed yet by the time the texture shader runs). If you want to change a texture shader input, you must edit the code and recompile (build) the effect.

Things you cannot currently do with SAS Scripts

Loop per-object or per-light – that is, you can’t cause a loop that goes “once per object” or “for each light, do….” – Loops only function by numeric count, though you could, Loop by a count and then extract light data from arrays by using “LoopGetIndex.”

The Script directive “LoopByType,” described in some versions of the DXSAS Spec, is not implemented (nor have we seen it implemented in any other tool, so far).

Control object rendering order (if there is more than one object) – this is defined by the rendering application, not SAS. Order-dependant algorithms (such as transparency) need to have the data ordered properly in advance.

Specify single objects to draw (when there is more than one object). Object hiding is up to the rendering application. This means there no non-hacky way to use, say, shadow proxy objects.

Call passes from other passes.

Persistent non-texture values – while textures can persist from frame to frame, numeric parameter values do not. If you really need them, you could write them to a texture and retrieve them each frame. This has been done in FX Composer 1.8. For FX Composer 2+, why not consider a Python script? Python namespaces persist across frames. Look at the “playblast” module, distributed with FX Composer. Using “playblast()” can control the frame count and play animated clips while executing your own Python per-frame callback functions, which can load the effect parameters with the values you need – more like a regular game engine.

Share texture buffers, explicitly under SAS control – that is, SAS provides no mechanism for sharing image buffers between effects – say a shadow map that would be used by multiple material effects, dynamic refelction maps, or full-scene deferred-render buffers. There are some possibilities, however…

How to Share Textures: Two Methods

COLLADA method: FX Composer supports, for COLLADA-FX, the ability to write to a shared texture surface. Other COLLADA-FX effects (and CgFX effects) can access this texture from their properties pane. This is how the NVIDIA “Ninja Scene” demo was created, where the numerous materials on the environment and the running armored Ninja all share the same shadow maps. See the NVIDIA Developer Forums post at // for technical details.

Recent user experiments have found that this COLLADA-based method can also work for DirectX effects in FX Composer – again, check the NVIDIA developer forums for the latest details and developments.

Autodesk 3DS Max method: while not exactly controlled by SAS, 3DS Max has a slightly-different mechanism for defining preprocess, standard, and postprocess effects. The result is that Max scenes can indeed render themselves and use shared shadows, reflections, and so forth. For details, check the web for the Autodesk Presentation “Integrating High-Level Shading Effects into Autodesk 3ds Max and Maya,” which has been updated repeatedly by Autodesk and available from multiple sources.

(A “secret” third method? The HLSL specification provides an explicit “shared” storage classifier for global parameters – the intent was to provide a mechanism for sharing both texture buffers and parameter values between effects – e.g., “shared float AA;” would provide the same numeric values to any loaded effects that contained a reference to that shared “AA” parameter. Alas, this keyword has so far never appeared in a running HLSL implementation)

Semantics, Annotations, and Python

Here is a quick IronPython snippet that can be used within FX Composer to view predefined semantics recognized by your current edition of the program:

Import FXComposer
[s for s in dir(FXComposer.Core.FXSemanticID) if \
not callable(getattr(FXComposer.Core.FXSemanticID.s)) \
and not s.__contains__(’_’)]

IronPython within FX Composer has access to the complete annotation list for any parameter – including annotations not usually recognized by FX Composer itself. See the provided”” module and its method “named_annotation()” for an example of how to retrieve annotation values.

List of Common Semantics

Hardware Semantics

Register Names in vertex buffers and vertex-output connections











TEXCOORD8 (synonym for COLOR0)

TEXCOORD9 (synonym for COLOR1)

ATTR0 (synonym for TEXCOORD0 sometimes seen in Cg)

ATTR1 (synonym for TEXCOORD1)

ATTR2 (synonym for TEXCOORD2)

ATTR3 (synonym for TEXCOORD3)

ATTR4 (synonym for TEXCOORD4)

ATTR5 (synonym for TEXCOORD5)

ATTR6 (synonym for TEXCOORD6)

ATTR7 (synonym for TEXCOORD7)









Particle Size and Polygon Facing Flag Inputs for Pixel Shaders


FACE (OpenGL equivalent to VFACE)

VFACE (DirectX equivalent to FACE)

Pixel Shader Output Values

COLOR (synonym for COLOR0)






Standard Software Semantics

Common Material and Light Characteristics



























Transforms & Locations































FXComposer Specific Semantics

Resetting Views


Mouse Interactions







Other Semantics Sometimes Encountered



COS_TIME (just call cos())

If you encounter any new semantics, please send them to us on the NVIDIA forums: //

List of Scripting Commands




















Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no responsibility for the consequences of use of such information or for any infringement of patents or other rights of third parties that may result from its use. No license is granted by implication or otherwise under any patent or patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to change without notice. This publication supersedes and replaces all information previously supplied. NVIDIA Corporation products are not authorized for use as critical components in life support devices or systems without express written approval of NVIDIA Corporation.


NVIDIA and the NVIDIA logo are trademarks or registered trademarks of NVIDIA Corporation in the United States and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.

Macrovision Compliance Statement

NVIDIA Products that are Macrovision enabled can only be sold or distributed to buyers with a valid and existing authorization from Macrovision to purchase and incorporate the device into buyer’s products.

Macrovision copy protection technology is protected by U.S. patent numbers 4,631,603, 4,577,216 and 4,819,098 and other intellectual property rights. The use of Macrovision’s copy protection technology in the device must be authorized by Macrovision and is intended for home and other limited pay-per-view uses only, unless otherwise authorized in writing by Macrovision. Reverse engineering or disassembly is prohibited


© 2008 NVIDIA Corporation. All rights reserved.





[1] SAS is sometimes also called “DXSAS” to indicate its DirectX origins, but the “DX” is usually dropped since SAS is effective across graphics APIs.

[2] For OpenGL users, there is also a corresponding “plasticP.cgfx” in the library, which is the same in terms of semantics and annotations. In general we will use HLSL examples in this document, but the SAS usage applies identically to both languages, save for the use of the HLSL virtual machine as described in the Appendix on “texture shaders.”

[3] Some damage is already done! For example, effect files for some DCC applications like XSI and 3ds Max are already using their own semantics for camera location. Fortunately simple workarounds are available.

[4] Float3 data, which lacks a w component, can’t be used this way. Float3 positions will only bind to regular position data in FX Composer.