Graphics Reference
In-Depth Information
900
Set 1
Set 2
Set 3
Set 4
800
700
600
500
400
300
200
100
0
Scene 1
Scene 2
Scene 3
Scene 4
Scene 5
Scene 6
Figure 4.6.
The number of non-empty cells for each scene.
shows the number of non-empty cells for each benchmark scene. Although scene
6 has about 800 non-empty cells for each set, scene 1 has less than 200 cells.
Thus, all the SIMDs on the GPU are active. When the number of non-empty
cells is small, we could increase the resolution of the spatial split. However, this
results in a reduction of the number of constraints in a cell. At the same time,
we do not want to reduce the number of constraints in a cell to fill SIMD lanes
of each SIMD.
4.6.2 Computation Time
Computation times for pipelined local batching and constraint solvers are shown
in Table 4.1. Pipelined local batching takes longer for a simulation with more
constraints. Because constraints are processed by the pipeline one by one, the
time taken for pipelined local batching is linear to the number of constraints in
Scene 1
Scene 2
Scene 3
Scene 4
Scene 5
Scene 6
Number of bodies
1,200
6,240
12,640
19,040
9,600
38,400
Number of constraint pairs
1,763
13,203
21,904
33,209
16,684
67,263
Number of global batches
11
16
14
16
13
25
Pipelined local batching time
0.105
0.357
0.524
0.735
0.376
1.16
Global solver time (1 iter.)
0.264
0.415
0.45
0.552
0.367
0.999
Two-level solver time (1 iter.)
0.162
0.387
0.433
0.540
0.359
0.839
Max batch count in set1
6
12
11
16
13
17
Max batch count in set2
5
14
16
16
11
12
Max batch count in set3
5
12
14
14
11
13
Max batch count in set4
5
12
12
14
12
14
Tab l e 4. 1.
Benchmark data (time in milliseconds).
