PCI Express Transactions
 – Relaxed Ordering
• Allocate USWC buffers to permit the GPU to optimize
requests by setting the Relaxed Ordering attribute
– permits Read completions to pass downstream Posted
Writes which are potentially draining slowly
– lowers GPU read latency permitting a stalled engine to proceed
doing useful work more quickly
• GPU driver will use an appropriate programming
model so that relaxed ordering does not cause
coherency issues
– a GPU engine’s Read Completion can not be dependent on
GPU context that may be in flight from the CPU issued as
Posted Writes