Hi!

I'm creating a data-driven rendering engine which is loosely based on Bitsquid/Stingray slides.

Rendering pipelines are compiled from descriptions written in a JSON-like DSL.

Until now, everything was nice and clean, the system allows to quickly prototype new renderers without writing tons of C++ glue.

But things get messy when the rendering pipeline must be changed at run-time (e.g. the player wants to toggle a certain feature on/off).

For instance, in the pipeline description all render targets are fixed, but if the user enables SSAO, the ueber_postprocessing shader must now output to two render targets instead of one. Similarly, if FXAA is disabled, the postprocessing shader must write straight to the backbuffer. There's a bunch of lesser problems (e.g. changing the render target clear colors to white when I want to take screenshots for whitepapers).

Here's my actual rendering engine pipeline description (with ugly HACKs):

|	Tiled deferred rendering pipeline description.
|

; dependencies
includes = [
	'renderer_common'    ; defines common render states
	'deferred_common'    ; defines common render targets and the main depth-stencil
]

color_targets = [
	; R8G8B8 - world-space normal, A8 - specular intensity
	{
		name = 'GBufferTexture0'
		depends_on = 'BackBuffer'
		;ScaleX	=	1.0
		;ScaleY	=	1.0
		sizeX = { size = 1 relative = 1 }
		sizeY = { size = 1 relative = 1 }
		format	=	'RGBA8'
		auto_clear = 1
		;Info	=	'Normals & Specular intensity'
	}
	; R8G8B8 - albedo, A8 - specular power
	{
		name = 'GBufferTexture1'
		depends_on = 'BackBuffer'
		;ScaleX	=	1.0
		;ScaleY	=	1.0
		sizeX = { size = 1 relative = 1 }
		sizeY = { size = 1 relative = 1 }
		format	=	'RGBA8'
		auto_clear = 1
		;Info	=	'Diffuse & Specular power'
	}
]

depth_targets = [
	; CSM/PSSM
	{
		name = 'ShadowMap'
		sizeX = { size = 2048 relative = 0 }
		sizeY = { size = 2048 relative = 0 }
		format	=	'D32'
		sample = 1
	}
]

|
35. Layers Configuration:
    Defines the draw order of the visible batches in a game world 
    Layers are processed in the order they are declared 
    Shader system points out which layer to render in 
|
layers =
[
; HACK:
	{
		id = 'FrameStart'
		render_targets = [
			{ name = '$Default' clear = 1 }
		]
		depth_stencil = { name = 'MainDepthStencil' clear = 1 }
		profiling_scope = 'Clear FrameBuffer'
	}

	; Directional light
	{
		id = 'ShadowMap0'
		; depth-only rendering
		depth_stencil = { name = 'ShadowMap' clear = 1 }
		sort = 'FrontToBack'
		profiling_scope = 'ShadowMap0'
	}
	{
		id = 'ShadowMap1'
		; depth-only rendering
		depth_stencil = { name = 'ShadowMap' }
		sort = 'FrontToBack'
		profiling_scope = 'ShadowMap1'
	}
	{
		id = 'ShadowMap2'
		; depth-only rendering
		depth_stencil = { name = 'ShadowMap' }
		sort = 'FrontToBack'
		profiling_scope = 'ShadowMap2'
	}
	{
		id = 'ShadowMap3'
		; depth-only rendering
		depth_stencil = { name = 'ShadowMap' }
		sort = 'FrontToBack'
		profiling_scope = 'ShadowMap3'
	}


	; G-Buffer Stage: Render all solid objects to a very sparse G-Buffer
	{
		id = 'FillGBuffer'
		; Clear Geometry buffer.
		; Bind render targets and main depth-stencil surface.
		render_targets = [
			{ name = 'GBufferTexture0' clear = 1 }	; normals
			{ name = 'GBufferTexture1' clear = 1 }	; albedo
		]
		depth_stencil = { name = 'MainDepthStencil' clear = 1 }
		;render_state = 'Default'
		;default_shader = ?
		;filter = Solid/Opaque only
		sort = 'FrontToBack'	; Draw opaque objects front to back
		profiling_scope = 'Fill G-Buffer'
		; Fill buffers with material properties of *opaque* primitives.
	}
	; G-buffer has been filled with data.


; HACK:
{
	id = 'ClearLightingTarget'
	render_targets = [
		{ name = 'LightingTarget' clear = 1 }
	]
}


	; point lights -> 'LightingTarget'
	{
		id = 'TiledDeferred_CS'
		; Bind null color buffer and depth-stencil.
		; No need to clear, we write all pixels.
		profiling_scope = 'TiledDeferred_CS'
	}
	; Render deferred local lights.
	{
		id = 'DeferredLocalLights'
		render_targets = [
			{ name = 'LightingTarget' }
		]
		depth_stencil = { name = 'MainDepthStencil' }
		profiling_scope = 'Deferred Local Lights'
	}
	; Apply deferred directional lights.
	{
		id = 'DeferredGlobalLights'
		render_targets = [
			{ name = 'LightingTarget' }
		]
		; Cannot bind the main depth-stencil surface - we read from it.
		profiling_scope = 'Deferred Global Lights'
	}

	; Forward stage
	{
		id = 'Unlit'
		render_targets = [
			{ name = 'LightingTarget' }
		]
		depth_stencil = { name = 'MainDepthStencil' }
		profiling_scope = 'Unlit'
	}	

	{
		id = 'SkyLast'
		render_targets = [
			{ name = 'LightingTarget' }
		]
		depth_stencil = { name = 'MainDepthStencil' }
		profiling_scope = 'Sky'
	}

	; Alpha-Blended (Order-Dependent) Transparency
	{
		id = 'Translucent'
		render_targets = [
			{ name = 'LightingTarget' }
		]
		depth_stencil = { name = 'MainDepthStencil' }
	}

	; Blended Order-Independent Transparency (OIT):
	;
	{
		id = 'BlendedOIT'
		render_targets = [
			{ name = 'BlendedOIT_Accumulation' clear = 1 }
			{ name = 'BlendedOIT_Revealage' clear = 1 }
		]
		; Keep the depth buffer that was rendered for opaque surfaces
		; bound for depth testing when rendering transparent surfaces,
		; but do not write to it.
		depth_stencil = { name = 'MainDepthStencil' }
	}
	
	; Combine: LightingTarget -> FrameBuffer
	; Blended OIT: 2D Compositing Pass
	{
		id = 'DeferredCompositeLit'
		render_targets = [
;TEMP HACK:
;			{ name = '$Default' }
			{ name = 'FXAA_Input' }
		]
		profiling_scope = 'Tonemap'
	}


	; thick wireframes should be antialiased
	{
		id = 'DebugWireframe'
		render_targets = [
;TEMP HACK:
;			{ name = '$Default' }
			{ name = 'FXAA_Input' }
		]
		depth_stencil = { name = 'MainDepthStencil' }
		profiling_scope = 'DebugWireframe'
	}

	
	; Fast Approximate Anti-Aliasing (FXAA)
	{
		id = 'FXAA'
		render_targets = [
			{ name = '$Default' }
		]
	}

	; Draw debug lines and GUI above everything else
	{
		id = 'DebugLines'
		render_targets = [
			{ name = '$Default' }
		]
		depth_stencil = { name = 'MainDepthStencil' }
		profiling_scope = 'DebugLines'
	}
	{
		id = 'DebugScreenQuads'
		render_targets = [
			{ name = '$Default' }
		]
		profiling_scope = 'DebugScreenQuads'
	}

	; draw GUI above everything else
	{
		id = 'GUI'
		render_targets = [
			{ name = '$Default' }
		]
		profiling_scope = 'ImGui'
	}
]

1) What is the cleanest way to allow on-the-fly modification of the rendering pipeline in a data-driven graphics engine?

2) How should render targets be described in a flexible rendering pipeline? How to specify that a render target should be taken from a render target pool?

1) What is the cleanest way to allow on-the-fly modification of the rendering pipeline in a data-driven graphics engine?

Reload everything?

2) How should render targets be described in a flexible rendering pipeline? How to specify that a render target should be taken from a render target pool?

I put a "shared" flag in each render target. If they're shared, they go to a name->target map. All targets first look into that map to see if there is a target with the same id, or create it otherwise. If a target is created and it has the shared flag, I add it to the pool, otherwise I just leave it alone in the render pass.

* Make it code driven :lol:
One non-facetious way of doing that would be to use something like Lua to store the graph - it's good at storing JSON-like data tables, but is also a fully fledged language.
* Add a small amount of branching logic to your DSL - e.g. this branch of the graph only active when blah condition is true.
* Make a whole boatload of different data files and load the right permutation depending on the situation.


As for the pooling, we describe every target as if it's unique to its pass - e.g. Bloom declares that it requires a half-res buffer, then DOF also declares that it requires a half-res buffer. If you then walk through the graph, allocating a target as a pass needs it and returning it to a pool when the pass is done (using only size/format, not name), then you've now got a minimal pool of targets that can be reused across passes (e.g. DOF and bloom will reuse the same half-res buffer).