On Demand Loading of MipMap Layers

Question

I have heard that in OGL and in DX9 D3DPOOL_MANAGED mipmap layers are pulled to the GPU memory on demand. As in If I pull from layer 4 of a mipmap, that layer and all smaller layers are pulled into the GPU memory, but larger mipmap layers are not.

I have tried to test this by creating a scene which requires more textures than the GPU has room for, but never getting close enough to any object such that it's larger mipmaps are required.
In OGL I don't hit a performance cliff, and as best I can tell with gDebugger I don't fill the GPU memory.
In D3DPOOL_DEFAULT I actually get a DX9 error cause I run out of GPU memory space.

At such a high level of abstraction I can't conclusively say what's happening. But I was hoping someone here could confirm or deny the partial mipmap rumor I've heard, as that would make my tests make sense.

Answer 1

If you want on-demand loading of "mipmap layers" there is actually a new hardware feature in DX 11.2 class hardware for this exact purpose. AMD introduced it first to OpenGL as sparse textures, it virtualizes the memory used by textures and allows individual LODs to be demand-paged the first time they are actually referenced in a fragment/pixel shader. D3D 11.2 standardizes this feature and calls it "tiled resources." This is honestly the only time you will ever (portably) have fine-grained knowledge of how the driver is managing your texture's memory at the LOD level.

Using this hardware-level feature, you can even tell inside your shader, whether an LOD is resident or not. This is very handy because demand-loading any resource introduces quite a bit of latency, and you want to be able to fallback to a lower LOD while you load the LOD into memory rather than suddenly increasing the time taken to render 1 frame by a factor of 10.