Linear Model Analysis of Real-Time Rendering - Real-Time Rendering: Computer Graphics with Control Engineering - page 24

TABLE 3.1
Performance Counters in DirectX
Direct3D Counter Description Official Name

FPS (#) D3D FPS

Frame time in milliseconds D3D frame time

Driver time in milliseconds D3D time in driver

Triangle count (#) D3D triangle count

Triangle count instanced (#) D3D triangle count instanced

Batch count (#) D3D batch count

Locked render targets count (#) D3D locked render targets count

AGP/PCIE memory used in integer MB (#) D3D agpmem MB

AGP/PCIE memory used in bytes (#) D3D agpmem bytes

Video memory used in integer MB (#) D3D vidmem MB

Video memory used in bytes (#) D3D vidmem bytes

Total video memory available in bytes (#) D3D vidmem total bytes

Total video memory available in integer MB (#) D3D vidmem total MB

Source: NVPerfKit documentation from www.nvidia.com

selected category. To illustrate, the input geometry to the rendering pipeline may
include lines, triangle fans, strips, and polygons. These are different input formats
that share the same basis—3D geometry data. Hence the natural choice as the input
variable of a rendering system should be the vertex count.
In addition to finding the appropriate variable by using its simplified form, another very
important characteristic that determines suitability is whether a variable can be changed
easily. For example, the batch counts and batch sizes of indexed buffers can impact the
performance of a rendering system. However, little can be done to control these variables
during an application runtime because these batches of vertices are predefined.
Finally, the resolution at which the selected variable may be adjusted affects the
quality of the system model as well. The ideal case would involve a variable that
allows fine resolution changes. For example, since the number of vertices is used
as an input variable of a rendering system, it may be difficult to obtain an accurate
model when this number can be varied only in limited steps.
One reason for this limitation is the underlying geometry LoD mechanism that
controls the resolution of a 3D object with a certain topological objective and algo-
rithm. The discrete LoD technique is an example of such a mechanism. Figure 3.1
illustrates the progressive variation (in steps) in the number of vertices that describe
a 3D object. Conversely, other techniques such as progressive meshes and geometry
tessellation allow 3D geometry variation at fine resolution levels. These techniques
are preferred in comparison to the approaches cited earlier.
So far we have discussed guidelines for inputs to the rendering system. As for the
output of the rendering system, the performance metric of primary concern to a user
of real-time computer graphics is widely accepted as the frame rate (inverse of the
time required to render one frame or image in a sequence) and quality of the gener-
ated imagery. The frame rate has a significant impact on the quality of the visual