Jump to content

  • Log In with Google      Sign In   
  • Create Account


#ActualAvengerDr

Posted 25 October 2013 - 08:30 AM

In the past few months I found myself juggling between different projects  aimed at several different platforms (Windows 7,8, RT and Phone). Some of those have different capabilities so some shaders needed to be modified in order to work correctly. I know that premature optimization is just as bad but in this specific situation I thought that addressing the problem sooner than later would be the right choice.
 
To address this problem I created a little tool that allows me to dynamically generate a set of shaders through a graph like structure. Which is nothing new, as it is usually the basis for this kind of application. I did probably reinvent a lot of wheels but since I couldn't use MS's shader designer (it only works with C++ I think) nor Unity's equivalent (as I have my own puny little engine) I decided to roll my own. I am writing here to get some feedback on my architecture and if there is something I overlooked.
 
Basically I have defined classes for most of the HLSL language. Then there are nodes such as constants, math operations and special function nodes. The latter are the most important ones as they correspond to high-level functions such as Phong Lighting, shadow algorithms and so on. Each of these function nodes expose several switches that enable me to enable/disable specific features. For example if I set a Phong node's "Shadows" to true then it will generate a different signature for the function than if it had it set to false. Once the structure is complete the graph is traversed and the actual shader code is generated line by line. From my understanding I think that dynamic shader linking works similarly but I've not been able to find a lot of information on the subject.
 
Right now shaders can only be defined in code, in the future I could build a graphical engine. A classic phong lighting pixel shader looks like this and this is the generated output. It is also possible to configure the amount of "verbosity". The interesting thing is that once the shader is compiled it gets serialized to a custom format that contains other information. Variables and nodes that are part of the shader are decorated with engine references. If I add a reference to the camera position for example, that variable tells the engine that it has to look for that value when initialising the shader. Analogously for the values needed to assemble constant buffers (like world/view/projection matrices). 

Once the shader is serialised, all this metadata helps the engine to automatically assign each shader variable or cbuffer with the right values. Before in my engine, each shader class had these huge parts of code that fetched needed values from somewhere else in the engine. Now all that has been deleted and is taken care automatically as long as the shaders are loaded in this format.
 
[attachment=18541:ShaderGenV01.png]

Another neat feature is that within the tool I have built I can define different techniques; i.e: a regular Phong shader, one using a Diffuse Map, one using a Shadow Map. Each technique maps a different combination of vertex and pixel shaders. The decoration that I mentioned earlier helps the tool generate a "TechniqueKey" for each shader that is then used by the engine to fetch the right shader from the file on disk. For example the PhongDiffusePS shader is decorated with attributes defining its use of a DiffuseMap (among other things). When in the actual application I enable the DiffuseMap feature, the shader system checks whether that feature is supported by the current set of shaders assigned the material. If a suitable technique is found, then the systeme enables the relevant parameters. In this way it is also possible to check for different feature levels and so on.

Probably something like this is overkill for a lot of small projects and I reckon it is not as easy to fix something in the generated code of this tool as it is when making changes in the actual source code it self. But once it does finally work, the the automatic configuration of shader variables is something that I like (at least if compared to my previous implementation, I don't how everyone else handles that). What I am asking is how extensible or expandable this approach is (maybe it is too late to ask this kind of questions biggrin.png)? Right now I have a handful of shaders defined in the system. If you had a look at the code, what kind of problems am I likely to run into when adding nodes to deal with Geometry shaders and other advanced features?

Finally, if anyone could be interested in having a look at the tool I'm happy to share it on GitHub.

#1AvengerDr

Posted 25 October 2013 - 08:27 AM

In the past few months I found myself juggling between different projects  aimed at several different platforms (Windows 7,8, RT and Phone). Some of those have different capabilities so some shaders needed to be modified in order to work correctly. I know that premature optimization is just as bad but in this specific situation I thought that addressing the problem sooner than later would be the right choice.
 
To address this problem I created a little tool that allows me to dynamically generate a set of shaders through a graph like structure. Which is nothing new, as it is usually the basis for this kind of application. I did probably reinvent a lot of wheels but since I couldn't use MS's shader designer (it only works with C++ I think) nor Unity's equivalent (as I have my own puny little engine) I decided to roll my own. I am writing here to get some feedback on my architecture and if there is something I overlooked.
 
Basically I have defined classes for most of the HLSL language. Then there are nodes such as constants, math operations and special function nodes. The latter are the most important ones as they correspond to high-level functions such as Phong Lighting, shadow algorithms and so on. Each of these function nodes expose several switches that enable me to enable/disable specific features. For example if I set a Phong node's "Shadows" to true then it will generate a different signature for the function than if it had it set to false. Once the structure is complete the graph is traversed and the actual shader code is generated line by line. From my understanding I think that dynamic shader linking works similarly but I've not been able to find a lot of information on the subject.
 
Right now shaders can only be defined in code, in the future I could build a graphical engine. A classic phong lighting pixel shader looks like this and this is the generated output. It is also possible to configure the amount of "verbosity". The interesting thing is that once the shader is compiled it gets serialized to a custom format that contains other information. Variables and nodes that are part of the shader are decorated with engine references. If I add a reference to the camera position for example, that variable tells the engine that it has to look for that value when initialising the shader. Analogously for the values needed to assemble constant buffers (like world/view/projection matrices). 

Once the shader is serialised, all this metadata helps the engine to automatically assign each shader variable or cbuffer with the right values. Before in my engine, each shader class had these huge parts of code that fetched needed values from somewhere else in the engine. Now all that has been deleted and is taken care automatically as long as the shaders are loaded in this format.
 
[attachment=18541:ShaderGenV01.png]

Another neat feature is that within the tool I have built I can define different techniques; i.e: a regular Phong shader, one using a Diffuse Map, one using a Shadow Map. Each technique maps a different combination of vertex and pixel shaders. The decoration that I mentioned earlier helps the tool generate a "TechniqueKey" for each shader that is then used by the engine to fetch the right shader from the file on disk. For example the PhongDiffusePS shader is decorated with attributes defining its use of a DiffuseMap (among other things). When in the actual application I enable the DiffuseMap feature, the shader system checks whether that feature is supported by the current set of shaders assigned the material. If a suitable technique is found, then the systeme enables the relevant parameters. In this way it is also possible to check for different feature levels and so on.

Probably something like this is overkill for a lot of small projects and I reckon it is not as easy to fix something in the generated code of this tool as it is when making changes in the actual source code it self. But once it does finally work, the the automatic configuration of shader variables is something that I like (at least if compared to my previous implementation, I don't how everyone else handles that). What I am asking is how extensible or expandable this approach is (maybe it is too late to ask this kind of questions biggrin.png)? Right now I have a handful of shaders defined in the system. If you had a look at the code, what kind of problems am I likely to run into when adding nodes to deal with Geometry shaders and other advanced features?

Finally, if anyone could be interested in having a look at the tool I'm happy to share it on GitHub.

PARTNERS