VTK/Depth Peeling

From KitwarePublic
Jump to navigationJump to search


Depth peeling is a multipass technique to render translucent polygonal geometry without sorting polygons.

The algorithm peels the translucent geometry from front to back until there is no more geometry to render. The iteration loop stops either if it reaches the maximum number of iterations set by the user or if the number of pixels modified by the last peel is less than some ratio of the area of the window (this ratio is set by the user, if the ratio is set to 0.0, it means the user wants the exact result. A ratio of 0.2 will render faster than a ratio of 0.1).

Date of Implementation

It was added to VTK on November 27th 2006. Texturing was added on January 9th 2007. Improvement with volumetric dataset and polygonal translucent geometry was made on March 5th 2007. Implemented by (any question to) François Bertel.


  • Translucent geometry does not have to be sorted.


  • The algorithm is inherently slow due to rendering the geometry multiple times.
  • The implementation based on OpenGL API requires a couple of extensions, restricting its usage to recent GPU. It works fine with the current CVS version of Mesa (2006/12/01)


Required OpenGL extensions

  • GL_ARB_depth_texture
  • GL_ARB_shadow
  • GL_EXT_shadow_funcs
  • GL_ARB_vertex_shader
  • GL_ARB_fragment_shader
  • GL_ARB_shader_objects
  • GL_ARB_occlusion_query
  • GL_ARB_multitexture
  • GL_ARB_texture_rectangle
  • GL_SGIS_texture_edge_clamp or GL_EXT_texture_edge_clamp or OpenGL 1.2

OpenGL context requirement

  • at least 8 alpha bits (by default vtkRenderWindow does not ask for alpha bit planes, make sure to call vtkRenderWindow::SetAlphaBitPlanes(1))

Rational: the algorithm has two main stages: first it renders peels from front to back in the framebuffer. each peel is saved in a 2D RGBA texture by grabbing the contents of the framebuffer. So the framebuffer is required to have an alpha channel in order to save it into an RGBA texture. The second part is doing compositing from back to front into the framebuffer by drawing quads on which each peel texture is applied. The compositing stage is drawing into the framebuffer but does not require alpha channel on it (the blending function used at this stage only use the alpha channel from the textures).

Linux Nvidia driver

Make sure to install a driver newer than 9500. For older drivers, there is a bug with GL_ARB_texture_rectangle and GLSL that prevents to use GLSL for depth peeling.

http://www.gpgpu.org/forums/viewtopic.php?p=11843&sid=9069e02401e2acf53dedf61349fc8431 . In order to use rectangle textures in GLSL, we are supposed to used sampler2DRectShadow in the shader and starts the shader source code with "#extension GL_ARB_texture_rectangle: enable". But with a driver like 8776, it returns:

warning C7508: extension GL_ARB_texture_rectangle not supported

VTK calls

The interface is in vtkRenderer.

  • 1. Use a render window with alpha bits (as initial value is 0 (false) ):
  • 2. Choose to use depth peeling (if supported) (initial value is 0 (false) )
  • 3. Set depth peeling parameters.
    • Set the maximum number of rendering passes (initial value is 4)
    • Set the occlusion ratio (initial value is 0.0, exact image)
  • 4. Render
  • 5. Check if depth peeling was actually used:
int depthPeelingWasUsed=renderer->GetLastRenderingUsedDepthPeeling();

Implementation notes

Modified files

in VTK/Rendering:


VTK changes

Before the changes, Render() of vtkRenderer used to iterate over all props and call RenderTranslucentGeometry on each of them. Now, it calls a virtual method DeviceRenderTranslucentGeometry(). By default this method call UpdateTranslucentGeometry(). And UpdateTranslucentGeometry() do exactly what Render() used to do.

In vtkOpenGLRenderer, DeviceRenderTranslucentGeometry() is overridden to use depth peeling flags and implement depth peeling technique. For each peel it will call UpdateTranslucentGeometry().

It uses a lot of GPU RAM because each rgba buffer of each peel have to be saved. The last step perfoms compositing with each layer, back-to-front order ( if it was possible to do front-to-back compositing, we could perform compositing on the fly and we would not need to store each peel.) Compositing is done by rendering each layer as a texture on a quad.

As we first render opaque geometry, we use it to cull transparent geometry, limiting the number of required layers.

Comments on references

  • The original depth peeling algorithm is described in [1].
  • [2] first render opaque geometry and use depth peeling only for none-opaque geometry which can reduce drastically the number of depth peeling passes.
  • To stop the iteration, follow [2], page 9, section 4. Transparency: "an occlusion query test on each pass indicates whether that layer was empty; if it was, the algorithm halts.". This kind of query is possible with openGL with BeginQuery/EndQuery (section 4.1.7 Occlusion Queries in page 207/221 spec 2.1). This is part of core OpenGL 1.5, promoted from extension GL_ARB_occlusion_query http://www.opengl.org/registry/specs/ARB/occlusion_query.txt.
  • More on query: as the extension can count the number of pixels that changed, the stop condition can be a percentage of the screen size instead of a complete unmodified screen ([5])
  • GLSL and Nvidia specific notes about shadow textures: [8]



  • OIT: Order Independent Transparency
  • MRT: Multiple Rendering Targets