Game Development Reference
In-Depth Information
Fill-rate without Blending
10
8
6
4
2
0
8-bit
8.28
8.18
16-bit
4.61
8.07
32-bit
Uncompressed
Compressed
2.42
4.56
Fill-rate with Blending
10
8
6
4
2
0
8-bit
16-bit
32-bit
Uncompressed
Compressed
8.28
4.56
1.1
8.18
4.52
1.99
Figure 4.6. Fill-rate measurements (in gigapixels per second) when using the compressed
framebuffer to perform forward rendering.
can rasterize pixels, and it is directly proportional to the available framebuffer
bandwidth and the number of bits emitted for each pixel. Figure 4.6 shows the
actual measurements on an Nvidia GTX 460 with 768 MB of memory (192-bit
bus). The compressed framebuffer uses two output channels, while the uncom-
pressed one uses four channels, since three-channel formats are not valid render
targets due to memory alignment restrictions.
Our measurements indicate an impressive 75% fill-rate increase when render-
ing with half-float precision without blending, confirming our theoretical analysis.
On the other hand, when blending is enabled on a half-float render target, we
could not measure any gain, perhaps indicating some limitation in the flexibility
of the render-output units in this specific GPU architecture. When using a 32-bit
floating-point render target, our measurements demonstrate the expected increase
in the fill rate when blending is both enabled and disabled, but this format is
rarely used to store color information. In the 8-bit case we did not measure any
Search Nedrilad ::




Custom Search