Conclusions
Optimal solution
- better than 1/W (and W at 32 bpp); no problems at f/d-ɭ
- simple: 1-Z in the driver or glDepthRange(zNear > zFar)
- low-cost: needs only floating-point storage format(s)
Next steps
- 3D hardware with variable size of exponent, exposed to an application
- Application-specific format selection