 |
OGRE
1.10.12
Object-Oriented Graphics Rendering Engine
|
|
OGRE
1.10.12
Object-Oriented Graphics Rendering Engine
|
Support for high level vertex and fragment programs is provided through plugins; this is to make sure that an application using OGRE can use as little or as much of the high-level program functionality as they like. OGRE currently supports 3 high-level program types, Cg (Cg) (an API- and card-independent, high-level language which lets you write programs for both OpenGL and DirectX for lots of cards), DirectX 9 High-Level Shader Language (HLSL), and OpenGL Shader Language (GLSL). HLSL can only be used with the DirectX rendersystem, and GLSL can only be used with the GL rendersystem. Cg can be used with both, although experience has shown that more advanced programs, particularly fragment programs which perform a lot of texture fetches, can produce better code in the rendersystem-specific shader language.
One way to support both HLSL and GLSL is to include separate techniques in the material script, each one referencing separate programs. However, if the programs are basically the same, with the same parameters, and the techniques are complex this can bloat your material scripts with duplication fairly quickly. Instead, if the only difference is the language of the vertex & fragment program you can use OGRE’s Unified High-level Programs to automatically pick a program suitable for your rendersystem whilst using a single technique.
In order to define Cg programs, you have to have to load Plugin_CgProgramManager.so/.dll at startup, either through plugins.cfg or through your own plugin loading code. They are very easy to define:
There are a few differences between this and the assembler program - to begin with, we declare that the fragment program is of type cg rather than asm, which indicates that it’s a high-level program using Cg. The source parameter is the same, except this time it’s referencing a Cg source file instead of a file of assembler. Here is where things start to change. Firstly, we need to define an entry_point, which is the name of a function in the Cg program which will be the first one called as part of the fragment program. Unlike assembler programs, which just run top-to-bottom, Cg programs can include multiple functions and as such you must specify the one which start the ball rolling. Next, instead of a fixed syntax parameter, you specify one or more profiles; profiles are how Cg compiles a program down to the low-level assembler. The profiles have the same names as the assembler syntax codes mentioned above; the main difference is that you can list more than one, thus allowing the program to be compiled down to more low-level syntaxes so you can write a single high-level program which runs on both D3D and GL. You are advised to just enter the simplest profiles under which your programs can be compiled in order to give it the maximum compatibility. The ordering also matters; if a card supports more than one syntax then the one listed first will be used.
Lastly, there is a final option called compile_arguments, where you can specify arguments exactly as you would to the cgc command-line compiler, should you wish to.
DirectX9 HLSL has a very similar language syntax to Cg but is tied to the DirectX API. The only benefit over Cg is that it only requires the DirectX 9 render system plugin, not any additional plugins. Declaring a DirectX9 HLSL program is very similar to Cg. Here’s an example:
As you can see, the main syntax is almost identical, except that instead of profiles with a list of assembler formats, you have a target parameter which allows a single assembler target to be specified - obviously this has to be a DirectX assembler format syntax code.
Important Matrix Ordering Note: One thing to bear in mind is that HLSL allows you to use 2 different ways to multiply a vector by a matrix - mul(v,m) or mul(m,v). The only difference between them is that the matrix is effectively transposed. You should use mul(m,v) with the matrices passed in from Ogre - this agrees with the shaders produced from tools like RenderMonkey, and is consistent with Cg too, but disagrees with the Dx9 SDK and FX Composer which use mul(v,m) - you will have to switch the parameters to mul() in those shaders.
Note that if you use the float3x4 / matrix3x4 type in your shader, bound to an OGRE auto-definition (such as bone matrices) you should use the column_major_matrices = false option (discussed below) in your program definition. This is because OGRE passes float3x4 as row-major to save constant space (3 float4’s rather than 4 float4’s with only the top 3 values used) and this tells OGRE to pass all matrices like this, so that you can use mul(m,v) consistently for all calculations. OGRE will also to tell the shader to compile in row-major form (you don’t have to set the /Zpr compile option or #pragma pack(row-major) option, OGRE does this for you). Note that passing bones in float4x3 form is not supported by OGRE, but you don’t need it given the above.
Advanced options
This allows you to define symbols which can be used inside the HLSL shader code to alter the behaviour (through #ifdef or #if clauses). Definitions are separated by ’;’ or ’,’ and may optionally have a ’=’ operator within them to specify a definition value. Those without an ’=’ will implicitly have a definition of 1.
The default for this option is ’true’ so that OGRE passes matrices auto-bound matrices in a form where mul(m,v) works. Setting this option to false does 2 things - it transpose auto-bound 4x4 matrices and also sets the /Zpr (row-major) option on the shader compilation. This means you can still use mul(m,v), but the matrix layout is row-major instead. This is only useful if you need to use bone matrices (float3x4) in a shader since it saves a float4 constant for every bone involved.
Set the optimisation level, which can be one of ’default’, ’none’, ’0’, ’1’, ’2’, or ’3’. This corresponds to the /O parameter of fxc.exe, except that in ’default’ mode, optimisation is disabled in debug mode and set to 1 in release mode (fxc.exe uses 1 all the time). Unsurprisingly the default value is ’default’. You may want to change this if you want to tweak the optimisation, for example if your shader gets so complex that it will not longer compile without some minimum level of optimisation.
OpenGL GLSL has a similar language syntax to HLSL but is tied to the OpenGL API. The are a few benefits over Cg in that it only requires the OpenGL render system plugin, not any additional plugins. Declaring a OpenGL GLSL program is similar to Cg but simpler. Here’s an example:
In GLSL, no entry point needs to be defined since it is always main() and there is no target definition since GLSL source is compiled into native GPU code and not intermediate assembly.
GLSL supports the use of modular shaders. This means you can write GLSL external functions that can be used in multiple shaders.
External GLSL functions are attached to the program that needs them by using attach and including the names of all external programs required on the same line separated by spaces. This can be done for both vertex and fragment programs.
To pass texture unit index values from the material script to texture samplers in glsl use int type named parameters. See the example below:
excerpt from GLSL example.frag source:
In material script:
An index value of 0 refers to the first texture unit in the pass, an index value of 1 refers to the second unit in the pass and so on.
Here are some examples of passing matrices to GLSL mat2, mat3, mat4 uniforms:
GLSL can access most of the GL states directly so you do not need to pass these states through param_named_auto in the material script. This includes lights, material state, and all the matrices used in the openGL state i.e. model view matrix, worldview projection matrix etc.
GLSL natively supports automatic binding of the most common incoming per-vertex attributes (e.g. gl_Vertex, gl_Normal, gl_MultiTexCoord0 etc). However, there are some which are not automatically bound, which must be declared in the shader using the attribute <type> <name> syntax, and the vertex data bound to it by Ogre.
In addition to the built in attributes described in section 7.3 of the GLSL manual, Ogre supports a number of automatically bound custom vertex attributes. There are some drivers that do not behave correctly when mixing built-in vertex attributes like gl_Normal and custom vertex attributes, so for maximum compatibility you should use all custom attributes
Binds Ogre::VES_POSITION, declare as ’attribute vec4 vertex;’.
Binds Ogre::VES_NORMAL, declare as ’attribute vec3 normal;’.
Binds Ogre::VES_DIFFUSE, declare as ’attribute vec4 colour;’.
Binds Ogre::VES_SPECULAR, declare as ’attribute vec4 secondary_colour;’.
Binds Ogre::VES_TEXTURE_COORDINATES, declare as ’attribute vec4 uv0;’. Note that uv6 and uv7 share attributes with tangent and binormal respectively so cannot both be present.
Binds Ogre::VES_TANGENT, declare as ’attribute vec3 tangent;’.
Binds Ogre::VES_BINORMAL, declare as ’attribute vec3 binormal;’.
Binds Ogre::VES_BLEND_INDICES, declare as ’attribute vec4 blendIndices;’.
Binds Ogre::VES_BLEND_WEIGHTS, declare as ’attribute vec4 blendWeights;’.
GLSL supports using preprocessor definitions in your code - some are defined by the implementation, but you can also define your own, say in order to use the same source code for a few different variants of the same technique. In order to use this feature, include preprocessor conditions in your GLSL code, of the kind #ifdef SYMBOL, #if SYMBOL==2 etc. Then in your program definition, use the ’preprocessor_defines’ option, following it with a string if definitions. Definitions are separated by ’;’ or ’,’ and may optionally have a ’=’ operator within them to specify a definition value. Those without an ’=’ will implicitly have a definition of 1. For example:
This way you can use the same source code but still include small variations, each one defined as a different Ogre program name but based on the same source code.
GLSL allows the same shader to run on different types of geometry primitives. In order to properly link the shaders together, you have to specify which primitives it will receive as input, which primitives it will emit and how many vertices a single run of the shader can generate. The GLSL geometry_program definition requires three additional parameters :
The operation type of the geometry that the shader will receive. Can be ’point_list’, ’line_list’, ’line_strip’, ’triangle_list’, ’triangle_strip’ or ’triangle_fan’.
The operation type of the geometry that the shader will emit. Can be ’point_list’, ’line_strip’ or ’triangle_strip’.
The maximum number of vertices that the shader can emit. There is an upper limit for this value, it is exposed in the render system capabilities.
For example:
As mentioned above, it can often be useful to write both HLSL and GLSL programs to specifically target each platform, but if you do this via multiple material techniques this can cause a bloated material definition when the only difference is the program language. Well, there is another option. You can ’wrap’ multiple programs in a ’unified’ program definition, which will automatically choose one of a series of ’delegate’ programs depending on the rendersystem and hardware support.
This works for both vertex and fragment programs, and you can list as many delegates as you like - the first one to be supported by the current rendersystem & hardware will be used as the real program. This is almost like a mini-technique system, but for a single program and with a much tighter purpose. You can only use this where the programs take all the same inputs, particularly textures and other pass / sampler state. Where the only difference between the programs is the language (or possibly the target in HLSL - you can include multiple HLSL programs with different targets in a single unified program too if you want, or indeed any number of other high-level programs), this can become a very powerful feature. For example, without this feature here’s how you’d have to define a programmable material which supported HLSL and GLSL:
And that’s a really small example. Everything you added to the HLSL technique, you’d have to duplicate in the GLSL technique too. So instead, here’s how you’d do it with unified program definitions:
At runtime, when myVertexProgram or myFragmentProgram are used, OGRE automatically picks a real program to delegate to based on what’s supported on the current hardware / rendersystem. If none of the delegates are supported, the entire technique referencing the unified program is marked as unsupported and the next technique in the material is checked fro fallback, just like normal. As your materials get larger, and you find you need to support HLSL and GLSL specifically (or need to write multiple interface-compatible versions of a program for whatever other reason), unified programs can really help reduce duplication.
Within a pass section of a material script, you can reference a vertex, geometry and / or a fragment program which is been defined in a .program script (See Declaring GPU Programs). The programs are defined separately from the usage of them in the pass, since the programs are very likely to be reused between many separate materials, probably across many different .material scripts, so this approach lets you define the program only once and use it many times.
As well as naming the program in question, you can also provide parameters to it. Here’s a simple example:
In this example, we bind a vertex program called ’myVertexProgram’ (which will be defined elsewhere) to the pass, and give it 2 parameters, one is an ’auto’ parameter, meaning we do not have to supply a value as such, just a recognised code (in this case it’s the world/view/projection matrix which is kept up to date automatically by Ogre). The second parameter is a manually specified parameter, a 4-element float. The indexes are described later.
The syntax of the link to a vertex program and a fragment or geometry program are identical, the only difference is that ’fragment_program_ref’ and ’geometry_program_ref’ are used respectively instead of ’vertex_program_ref’.
For many situations vertex, geometry and fragment programs are associated with each other in a pass but this is not cast in stone. You could have a vertex program that can be used by several different fragment programs. Another situation that arises is that you can mix fixed pipeline and programmable pipeline (shaders) together. You could use the non-programmable vertex fixed function pipeline and then provide a fragment_program_ref in a pass i.e. there would be no vertex_program_ref section in the pass. The fragment program referenced in the pass must meet the requirements as defined in the related API in order to read from the outputs of the vertex fixed pipeline. You could also just have a vertex program that outputs to the fragment fixed function pipeline.
The requirements to read from or write to the fixed function pipeline are similar between rendering API’s (DirectX and OpenGL) but how its actually done in each type of shader (vertex, geometry or fragment) depends on the shader language. For HLSL (DirectX API) and associated asm consult MSDN at http://msdn.microsoft.com/library/. For GLSL (OpenGL), consult section 7.6 of the GLSL spec 1.1 available at http://www.opengl.org/registry/. The built in varying variables provided in GLSL allow your program to read/write to the fixed function pipeline varyings. For Cg consult the Language Profiles section in CgUsersManual.pdf that comes with the Cg Toolkit available at https://developer.nvidia.com/cg-toolkit. For HLSL and Cg its the varying bindings that allow your shader programs to read/write to the fixed function pipeline varyings.
Parameters can be specified using one of 4 commands as shown below. The same syntax is used whether you are defining a parameter just for this particular use of the program, or when specifying the Default Program Parameters. Parameters set in the specific use of the program override the defaults.
This command sets the value of an indexed parameter.
format: param_indexed <index> <type> <value> example: param_indexed 0 float4 10.0 0 0 0
| index | simply a number representing the position in the parameter list which the value should be written, and you should derive this from your program definition. The index is relative to the way constants are stored on the card, which is in 4-element blocks. For example if you defined a float4 parameter at index 0, the next index would be 1. If you defined a matrix4x4 at index 0, the next usable index would be 4, since a 4x4 matrix takes up 4 indexes. |
| type | can be float4, matrix4x4, float<n>, int4, int<n>. Note that ’int’ parameters are only available on some more advanced program syntaxes, check the D3D or GL vertex / fragment program documentation for full details. Typically the most useful ones will be float4 and matrix4x4. Note that if you use a type which is not a multiple of 4, then the remaining values up to the multiple of 4 will be filled with zeroes for you (since GPUs always use banks of 4 floats per constant even if only one is used). |
| value | a space or tab-delimited list of values which can be converted into the type you have specified. |
This command tells Ogre to automatically update a given parameter with a derived value. This frees you from writing code to update program parameters every frame when they are always changing.
format: param_indexed_auto <index> <value_code> <extra_params> example: param_indexed_auto 0 worldviewproj_matrix
| index | has the same meaning as param_indexed; note this time you do not have to specify the size of the parameter because the engine knows this already. In the example, the world/view/projection matrix is being used so this is implicitly a matrix4x4. |
| value_code | is one of Ogre::GpuProgramParameters::AutoConstantType without the ACT_ prefix. E.g. ACT_WORLD_MATRIX becomes world_matrix. |
This is the same as param_indexed, but uses a named parameter instead of an index. This can only be used with high-level programs which include parameter names; if you’re using an assembler program then you have no choice but to use indexes. Note that you can use indexed parameters for high-level programs too, but it is less portable since if you reorder your parameters in the high-level program the indexes will change. format: param_named <name> <type> <value> example: param_named shininess float4 10.0 0 0 0 The type is required because the program is not compiled and loaded when the material script is parsed, so at this stage we have no idea what types the parameters are. Programs are only loaded and compiled when they are used, to save memory.
This is the named equivalent of param_indexed_auto, for use with high-level programs. Format: param_named_auto <name> <value_code> <extra_params> Example: param_named_auto worldViewProj WORLDVIEWPROJ_MATRIX
The allowed value codes and the meaning of extra_params are detailed in param_indexed_auto.
This option allows you to reference shared parameter sets as defined in Declaring Shared Parameters. Format: shared_params_ref <shared_set_name> Example: shared_params_ref mySharedParams
The only required parameter is a name, which must be the name of an already defined shared parameter set. All named parameters which are present in the program that are also present in the shared parameter set will be linked, and the shared parameters used as if you had defined them locally. This is dependent on the definitions (type and array size) matching between the shared set and the program.
When using shadows (See Shadows), the use of vertex programs can add some additional complexities, because Ogre can only automatically deal with everything when using the fixed-function pipeline. If you use vertex programs, and you are also using shadows, you may need to make some adjustments.
If you use stencil shadows, then any vertex programs which do vertex deformation can be a problem, because stencil shadows are calculated on the CPU, which does not have access to the modified vertices. If the vertex program is doing standard skeletal animation, this is ok (see section above) because Ogre knows how to replicate the effect in software, but any other vertex deformation cannot be replicated, and you will either have to accept that the shadow will not reflect this deformation, or you should turn off shadows for that object.
If you use texture shadows, then vertex deformation is acceptable; however, when rendering the object into the shadow texture (the shadow caster pass), the shadow has to be rendered in a solid colour (linked to the ambient colour). You must therefore provide an alternative vertex program, so Ogre provides you with a way of specifying one to use when rendering the caster. Basically you link an alternative vertex program, using exactly the same syntax as the original vertex program link:
When rendering a shadow caster, Ogre will automatically use the alternate program. You can bind the same or different parameters to the program - the most important thing is that you bind ambient_light_colour, since this determines the colour of the shadow in modulative texture shadows. If you don’t supply an alternate program, Ogre will fall back on a fixed-function material which will not reflect any vertex deformation you do in your vertex program.
In addition, when rendering the shadow receivers with shadow textures, Ogre needs to project the shadow texture. It does this automatically in fixed function mode, but if the receivers use vertex programs, they need to have a shadow receiver program which does the usual vertex deformation, but also generates projective texture coordinates. The additional program linked into the pass like this:
For the purposes of writing this alternate program, there is an automatic parameter binding of ’texture_viewproj_matrix’ which provides the program with texture projection parameters. The vertex program should do it’s normal vertex processing, and generate texture coordinates using this matrix and place them in texture coord sets 0 and 1, since some shadow techniques use 2 texture units. The colour of the vertices output by this vertex program must always be white, so as not to affect the final colour of the rendered shadow.
When using additive texture shadows, the shadow pass render is actually the lighting render, so if you perform any fragment program lighting you also need to pull in a custom fragment program. You use the shadow_receiver_fragment_program_ref for this:
You should pass the projected shadow coordinates from the custom vertex program. As for textures, texture unit 0 will always be the shadow texture. Any other textures which you bind in your pass will be carried across too, but will be moved up by 1 unit to make room for the shadow texture. Therefore your shadow receiver fragment program is likely to be the same as the bare lighting pass of your normal material, except that you insert an extra texture sampler at index 0, which you will use to adjust the result by (modulating diffuse and specular components).
You can implement skeletal animation in hardware by writing a vertex program which uses the per-vertex blending indices and blending weights, together with an array of world matrices (which will be provided for you by Ogre if you bind the automatic parameter ’world_matrix_array_3x4’). However, you need to communicate this support to Ogre so it does not perform skeletal animation in software for you. You do this by adding the following attribute to your vertex_program definition:
When you do this, any skeletally animated entity which uses this material will forgo the usual animation blend and will expect the vertex program to do it, for both vertex positions and normals. Note that ALL submeshes must be assigned a material which implements this, and that if you combine skeletal animation with vertex animation (See Animation) then all techniques must be hardware accelerated for any to be.
You can implement morph animation in hardware by writing a vertex program which linearly blends between the first and second position keyframes passed as positions and the first free texture coordinate set, and by binding the animation_parametric value to a parameter (which tells you how far to interpolate between the two). However, you need to communicate this support to Ogre so it does not perform morph animation in software for you. You do this by adding the following attribute to your vertex_program definition:
When you do this, any skeletally animated entity which uses this material will forgo the usual software morph and will expect the vertex program to do it. Note that if your model includes both skeletal animation and morph animation, they must both be implemented in the vertex program if either is to be hardware acceleration. Note that ALL submeshes must be assigned a material which implements this, and that if you combine skeletal animation with vertex animation (See Animation) then all techniques must be hardware accelerated for any to be.
You can implement pose animation (blending between multiple poses based on weight) in a vertex program by pulling in the original vertex data (bound to position), and as many pose offset buffers as you’ve defined in your ’includes_pose_animation’ declaration, which will be in the first free texture unit upwards. You must also use the animation_parametric parameter to define the starting point of the constants which will contain the pose weights; they will start at the parameter you define and fill ’n’ constants, where ’n’ is the max number of poses this shader can blend, i.e. the parameter to includes_pose_animation.
Note that ALL submeshes must be assigned a material which implements this, and that if you combine skeletal animation with vertex animation (See Animation) then all techniques must be hardware accelerated for any to be.
If your vertex program makes use of Vertex Texture Fetch, you should declare that with the ’uses_vertex_texture_fetch’ directive. This is enough to tell Ogre that your program uses this feature and that hardware support for it should be checked.
Some geometry programs require adjacency information from the geometry. It means that a geometry shader doesn’t only get the information of the primitive it operates on, it also has access to its neighbours (in the case of lines or triangles). This directive will tell Ogre to send the information to the geometry shader.
More recent generations of video card allow you to perform a read from a texture in the vertex program rather than just the fragment program, as is traditional. This allows you to, for example, read the contents of a texture and displace vertices based on the intensity of the colour contained within.
Since hardware support for vertex texture fetching is not ubiquitous, you should use the uses_vertex_texture_fetch (See Vertex texture fetching in vertex programs) directive when declaring your vertex programs which use vertex textures, so that if it is not supported, technique fallback can be enabled. This is not strictly necessary for DirectX-targeted shaders, since vertex texture fetching is only supported in vs_3_0, which can be stated as a required syntax in your shader definition, but for OpenGL (GLSL), there are cards which support GLSL but not vertex textures, so you should be explicit about your need for them.
Unfortunately the method for binding textures so that they are available to a vertex program is not well standardised. As at the time of writing, Shader Model 3.0 (SM3.0) hardware under DirectX9 include 4 separate sampler bindings for the purposes of vertex textures. OpenGL, on the other hand, is able to access vertex textures in GLSL (and in assembler through NV_vertex_program_3, although this is less popular), but the textures are shared with the fragment pipeline. I expect DirectX to move to the GL model with the advent of DirectX10, since a unified shader architecture implies sharing of texture resources between the two stages. As it is right now though, we’re stuck with an inconsistent situation.
To reflect this, you should use the binding_type attribute in a texture unit to indicate which unit you are targeting with your texture - ’fragment’ (the default) or ’vertex’. For render systems that don’t have separate bindings, this actually does nothing. But for those that do, it will ensure your texture gets bound to the right processing unit.
Note that whilst DirectX9 has separate bindings for the vertex and fragment pipelines, binding a texture to the vertex processing unit still uses up a ’slot’ which is then not available for use in the fragment pipeline. I didn’t manage to find this documented anywhere, but the nVidia samples certainly avoid binding a texture to the same index on both vertex and fragment units, and when I tried to do it, the texture did not appear correctly in the fragment unit, whilst it did as soon as I moved it into the next unit.
Again as at the time of writing, the types of texture you can use in a vertex program are limited to 1- or 4-component, full precision floating point formats. In code that equates to PF_FLOAT32_R or PF_FLOAT32_RGBA. No other formats are supported. In addition, the textures must be regular 2D textures (no cube or volume maps) and mipmapping and filtering is not supported, although you can perform filtering in your vertex program if you wish by sampling multiple times.
As at the time of writing (early Q3 2006), ATI do not support texture fetch in their current crop of cards (Radeon X1n00). nVidia do support it in both their 6n00 and 7n00 range. ATI support an alternative called ’Render to Vertex Buffer’, but this is not standardised at this time and is very much different in its implementation, so cannot be considered to be a drop-in replacement. This is the case even though the Radeon X1n00 cards claim to support vs_3_0 (which requires vertex texture fetch).
1.8.13