Hi. I'm not 'new' to opengl nor programming but I've only recently started work on independent projects outside of messing around in textbooks with fixed pipeline and programmable pipeline projects, glsl, etc. Basically the basics are basic to me ;)

However, I've only started experimenting with the texture filter modes. On a heightmap terrain I had the mag filter as GL_LINEAR and the min filter set as the same, and my app was crushingly slow. Now I set the min filter as GL_NEAREST_MIPMAP_LINEAR and the performance shot through the roof!

I was under the false assumption that using mipmaps would slow things down somewhat. So what gives?

Mipmapping selects the mipmap level where the size of the texels matches the size of the pixels the best. This is good for the cache where rasterinzing adjacent pixels also means sampling adjacent texels. If, instead, the much larger base-level texture is selected (which is effectively what happens when you have a non-mipmapping filter), then the texels are much smaller than the pixels on the screen and sampling the texture for adjacent pixels means sampling very distant texels; this is very bad for the cache.

Mipmapping, especially full linear filtering on all dimensions, is computationally more expensive. But that is only a problem if you are computationally saturated. In your case, you are probably saturated by the memory transfer and so reducing that and increasing the computational complexity is an overall win.

Brother Bob. That is a very informative answer! I have always thought about the cache with regards to Data-Oriented programming and trying to 'sort' virtual function calls in arrays but that's the first time I ever thought about memory look ups on texels.

JTippetts, thank you also for your explanation. There can never be too many points of view of the same problem, even if they are similar, as the different wording of each answer can convey different aspects with more clarity.

Hodgman, do you know of a good resource for learning more about the hardware architecture of video cards as it relates to how data flows through the rendering pipeline?

CDProp, this might not be EXACTLY what you're looking for but Coursera.org has a great course on CUDA and GPU, you learn a lot about bandwidth issues like Hodgman has been mentioning. His input made me think about it actually.