Chalnoth said:
Martillo1 said:
So, in the end, the method that ATI is using can be correct since it adjusts the aggressiveness of the filtering depending of the "gradient" of the mipmaps.
No, that's not true. The texture filtering equations are invariant to texture data.
I think you're right with regard to "trilinear texture filtering" and what has been shown about the behavior, but not necessarily to "image quality for mip map transitions" which is likely what Martillo1 had in mind. Trilinear is quite often done for "image quality", so that can be a criteria for correctness in those cases. However, it seems like there would be cases where invariance might be important.
For example, if this "adaptation" is done to pre-generated mip map levels (I simply don't know how the mip map levels in the established cases where there are differences were auto-generated or not, except for mip map levels, so this is still a question to me), it seems like it could be "incorrect" simply due to there being apparent usefulness to the developer in invariance with regard to trilinear filtering for special applications of custom mip maps. Regardless of image quality impact, this case would seem then be "incorrect", and the colored mip map level behavior is actually encouraging indication for that behavior. It would be important to test this, but not just for evaluating if it depended upon a high degree of changes between mip map levels...I'd think any pre-generated/uploaded mip map sets should have invariant trilinear filtering to allow developers this control. I.e., I think image quality equivalence probably shouldn't matter for this case, and absolute invariance from "trilinear filtering" is important to allows this type of usage.
One interesting question, however, is whether trilinear filtering is defined for invariance, for OpenGL and Direct3D, for
auto-generated mip map levels? It makes sense to me that it would not be, as the cases where trilinear would be for other than image quality (where equivalent image quality could still be "correct" even without invariance) would seem to go perfectly with the developer controlling the mip map level data specifically. I.e., I don't see how absolute invariance would be important to the developer when they use auto-generating mip map levels, whereas if they
are using them, achieving a performance threshold for their game, with sufficient image quality, would be (as it typically is for game developers). Also, if it is, uploading mip map levels would seem to still allow (
if the implementation is invariant when handling these, as it seems it should be) the developer such control when they
do want it. If the API allows this distinction, both possible priorities for a developer are addressed under developer control...win/win.
All this, of course, hinging on whether the image quality is actually determined to be reasonable be called "equivalent" so that this latter case isn't still a "cheat" for the performance priority.
