Game Development Reference
In-Depth Information
Border Size
Resolution
1
2
4
4
×
4
56/56%
75/75%
89/89%
8
×
8
36/41%
56/62%
75/80%
16
×
16
21/24%
36/41%
56/62%
32
×
32
11/13%
21/23%
36/40%
64
×
64
6.0/6.4%
11/12%
21/23%
128
×
128
3.1/3.2%
6.0/6.2%
11/12%
Tab l e 2 . 2 . This table shows memory overhead for borders. The first percentage in each
pair is for a single resolution, and the second is for mip chains down to 4 × 4. These
values should be weighed against the waste inherent in a conventional texture atlas,
such as the 37% illustrated in Figure 2.4(b) .
2.7
Alternatives and Future Work
One way to avoid having a separate texture per face is to put each per-face texture
in its own texture array slice [McDonald and Burley 11, McDonald 12]. This
simplifies the texture addressing to some extent. However, there are limitations
in the number of texture array slices, and resolutions cannot be mixed within a
single texture array. Therefore, what would be a single texture in the conventional
or packed Ptex approach would be split into multiple textures, one for each
resolution, with further splitting as required for texture array limits. The amount
of texture data used in the shader does not increase, excepting perhaps due to
alignment or other per-texture costs, but the amount of conditional reads is
significantly higher.
Rather than computing per-face texture information, we could store it in a
resource indexed by face ID, and possibly by mip level [McDonald and Burley 11,
McDonald 12].
Ptex takes the extreme approach of assigning an individual texture map to
each primitive. The paper by B. Purnomo, et al. describes similar solutions
to what is described here, but they group multiple primitives into rectangular
patches in texture space for packing and handling seams [Purnomo et al. 04].
This reduces the overhead for borders, which would make larger filter kernels
feasible. A next step might be to integrate some of the ideas from that paper.
2.8
Conclusion
Packed Ptex enables the main advantages of the original Ptex method while en-
abling real-time use. Authoring effort is saved first by eliminating the need for
explicit UV assignment and second by naturally avoiding seaming issues that nor-
mally arise when trying to unwrap a 3D surface into at 2D rectangular domain.
It does, however, require modeling with quads in its current implementation.
 
 
Search Nedrilad ::




Custom Search