Graphics Reference
In-Depth Information
Using extensions
Multiple render targets
71,19
39,55
31,64
7,91
7,91
0
Write
Read
Total
Figure 3.4.
External bandwidth comparison for deferred shading implementations. As-
suming 1080p resolution with 32-bit framebuffer and D24S8 depth/stencil.
Extensions
MRT
Read
Write
Read
Write
Fill G-buffer
0MB
0MB
0 MB
23.73 MB
Light accumulation
0MB
0MB
15.82 MB
7.91 MB
Resolve
0MB 7.91MB
15.82 MB
7.91 MB
Totals
0MB 7.91MB
31.64 MB
39.55 MB
Tab l e 3. 5.
Breakdown of graph data showing a 9
×
bandwidth reduction.
Note:
Read data for lighting is approximated since this depends on coverage
and overdraw. This is isolated to only framebuffer bandwidth and does not
take into account texture lookups for diffuse and normal when filling the G-
buffer. It does not take into account bandwidth consumed by vertex shader
reads/writes and fragment shader varying reads since that will be similar on
both implementations.
3.5 Conclusion
We introduced extensions that allow applications to reduce the need to access
external memory. Reducing external memory access can not only increase per-
formance, but also increase battery life. Use-cases such as programmable blending
and reading of depth/stencil can enable applications to allow more work to be
done in on-chip memory. The novel Shader Pixel Local Storage extension gives
applications even more control of the per-pixel data in tile-enabling use-cases
such as deferred shading.
