Graphics Reference
In-Depth Information
Single Draw Call
Multiple Draw Calls
Draw calls
1
1 per tile
VS data
dynamic VBO
static VBO
+ uniforms
Texcoord transforms
on CPU
on GPU
Number of textures
1(upto8forwholegrid) 8 per tile
Unused layers
dynamic branching
different programs
Interpolators usage
high
low to high depending on aggregated
layer count
Tab l e 5. 1.
Single versus multiple draw calls summary.
5.4.3 Stencil
Stencil test is another common optimization for effects that affect only portions
of the screen [Weber and Quayle 11].
5.5 Results
The savings achieved with this technique depend greatly on the scene complex-
ity, the tile layout and, first and foremost, the underlying architecture. We can,
however, estimate a best case scenario by comparing the cost of rendering sev-
eral fullscreen quads using regular blending versus rendering the same amount
of fullscreen quads using deferred tile blending. The results are presented in
Figure 5.6 and in Table 5.2.
We can observe a linear relationship between the number of fullscreen layers
to blend and the time it takes to render the frame. As expected, we did not
notice any improvement on TBR architectures and the technique was even a bit
slower than simple blending on some devices. However, on IMR GPUs such as the
Tegra 3 equipping the Nexus 7, rendering time was approximately 35% shorter
than without using tile-based blending.
Fullscreen
layer count
SB
(Nexus7)
TDB
(Nexus7)
SB
(iPhone4S)
TDB
(iPhone4S)
8
23
16
7
8
16
45
30
13
14
24
66
44
19
21
32
87.5
58
24
28
40
109
72
32
34
Tab l e 5. 2.
Tiled deferred blending (TDB) and simple blending (SB) rendering times
(in ms).
