GPU: A closer look
Geometry primitive is represented by groups of vertex, application provides (x,y,z) to VG stage
VP stage chage 3D position into 2D position
PG group vertex together to into order stream to form primitives
PP application dependent, generate new order stream
FG sample each primitive, output fragment record
about the image position on the surface, the distance to the virtual camera
FP give surfaces realistic appearances
PO apply fragment contribution to output image pixel value
shader functions define the behavior of application programmable stage (vp, pp, fp). They are compiled into bytecode offline and tranlated into binary code specific to a GPU in runtime. It does not execute parallel but only serially process inputs ( I don't understand here...)
Graphics processing present large amount of data and task level parallelism, but the PO is a serial point where there are a lot of find grain dependence.
The bandwidth and computational demand vary significant for different load, so it is necessary to map the processing and storage resource to pipeline stage dynamically.
阅读(447) | 评论(0) | 转发(0) |