🔒 Locked

soft shadow algorithms

Started by maxgpgpu Jun 9, 2009 at 6:11 AM 0 replies 1.3k views

Original Post

Jun 9, 2009 at 6:11 AM

What "soft shadow" algorithm gives the best combination of speed and appearance [with 3 to 8 lights] and OpenGL310/GLSL140 support? I read a few HTML and PDF soft-shadow articles recently, but most focus on their favorite approach, and none feel like an objective, unbiased review and comparison. Also, many articles assume older version APIs and shaders. I am guessing that floating-point depth buffers and other recent GPU capabilities might help some approaches considerably. Personal experiences and links to PDFs and HTML are welcome. Thanks. A related request for comments: Reading the OpenGL310/GLSL140 documents last night gave me the following idea that might speed some shadow algorithms considerably --- IF --- I understand what I've been reading about soft shadow algorithms. This idea is inspired by new features like blocks of uniform variables. What if 4 sets of transformation matrices are supplied to the GPU shaders in a large block of uniform variables - one set of transformation matrices for each of 4 light-sources (treating each light like a camera/viewpoint)? Is there any reason the shaders cannot create 4 separate depth buffers (in 4 separate FBO renderbuffers), one for each light? If so, wouldn't this be faster than a separate pass for each light? Something tells me this won't work, but I can't see why off hand.

LtJax

Jun 11, 2009 at 10:22 AM

Quote:

This idea is inspired by new features like blocks of uniform variables. What if 4 sets of transformation matrices are supplied to the GPU shaders in a large block of uniform variables - one set of transformation matrices for each of 4 light-sources (treating each light like a camera/viewpoint)? Is there any reason the shaders cannot create 4 separate depth buffers (in 4 separate FBO renderbuffers), one for each light? If so, wouldn't this be faster than a separate pass for each light? Something tells me this won't work, but I can't see why off hand.

That is possible with GPUs that support geometry shaders and texture arrays, and optionally instancing. See the DX10-style acceleration in this GPU Gems 3 Article

They are using it to split the frustum of one light, but that's essentially the same as doing it for totally disjoint lights. This is also a good technique for point lights with cube-shadowmaps. The problem is that you can't do individual culling for each light on the CPU.

Topic Locked

This topic has been locked by a moderator. New replies are not allowed.