Linear Model Analysis of Real-Time Rendering - Real-Time Rendering: Computer Graphics with Control Engineering - page 41

Measured Output

380

370

360

350

340

330

0 1000 2000 3000 4000 5000

Ve rtex Count (Input)

×10 4

4.5

4.4

4.3

4.2

4.1

4

0 1000 2000 3000 4000 5000

Frame

FIGURE 3.17 Reference tracking with actual rendering application.

and keyboard input. Furthermore, in our analysis of the dynamics of the rendering
process, we observed various step-like interruptions in the frame rate changes that
persisted for a certain period. We attribute this phenomenon to hardware or software
operations such as driver wait states and memory transport delays. Furthermore,
current tools available to us do not allow access to sufficiently low-level operations
to identify these interruptions and irregularities.
As indicated in previous research [32,33,34], common performance metrics in
interactive rendering affect mainly frame rate and image quality. In this research,
frame rate was selected over image quality because it is well known that computation
at image level is resource intensive [29] and assessment of image quality metrics may
be subjective. Furthermore, research indicates that frame rates exert greater impact
on user perception and response [35,36,37] in many applications.
As noted in previous research [24,27] that reactive rendering time control tech-
niques cannot deal well with abrupt changes in scene load because the resultant
oscillations in frame rates may negatively affect the user experience. We have shown
in this research that our data-driven modelling framework provides an important
basis for application of a control strategy that generates stable frame rates without
noticeable oscillation. These are the benefits of employing system modelling and
control techniques in the real-time computer graphics domain.

3.7.1 c omPaRison with o theR e stimation t echniQues
Many application developers may intuitively perceive the real-time rendering process
as linear and thus use simplistic mathematical models to estimate its performance.
In this section, we validate the accuracy of our data-driven modelling framework by
comparing its outputs with those of two other intuitive estimation models based on a