Speed up lighting in deferred shading

Question

I implemented a simple deferred shading renderer. I use 3 G-Buffer for storing position (R32F), normal (G16R16F) and albedo (ARGB8). I use sphere map algorithm to store normals in world space. Currently I use inverse of view * projection matrix to calculate the position of each pixel from stored depth value.

1. First I want to avoid per pixel matrix multiplication for calculating the position. Is there another way to store and calculate position in G-Buffer without the need of matrix multiplication
2. Store the normal in view space

Every lighting in my engine is in world space and I want do the lighting in view space to speed up my lighting pass.

I want an optimized lighting pass for my deferred engine.

Answer 1

For pixel positions, as mentioned in the comments, it's generally considered best to use the depth buffer and the pixel's 2D position on screen to calculate its view space or world space position. If you're using an older D3D/OpenGL API version you may need to store an extra copy of the depth in an R32F buffer (it's unclear to me if this is what you're doing now); but in more recent APIs you can read the actual depth buffer directly, saving yourself one G-buffer slot.

In any case, most of the position reconstruction math can be moved into the vertex shader or the CPU, with only a reciprocal and a madd in the fragment shader (the reciprocal is not needed if you're using a linear depth buffer). The details of how to do this are in this blog post.