In this exercise you will learn how to create a fixed data flow using an
in-order queue
.
Your old coworker is lazy, and has enqueued all the kernels in the same in-order queue. Performance is poor as kernels cannot execute concurrently. Your goal is to improve it:
- Using one out-of-order queue
- Using multiple in-order queues
Using the application from exercise 10 which uses the buffer/accessor model
convert the queue
to in-order using the property::queue::in_order
property,
converting the data flow graph to a fixed execution order.
Feel free to have the kernel functions execute in any order you like providing the necessary dependencies described in exercise 10 are still met.
Now do the same using the application from exercise 10 which uses the USM model,
again converting the queue
to in-order, converting the data flow graph to a
fixed execution order.
Again feel free to have the kernel functions execute in any order you like providing the necessary dependencies described in exercise 10 are still met.
Note that in the USM model when using an in-order queue
it is no longer
necessary to chain commands using event
s.
For DevCloud via JupiterLab follow these instructions.
For DPC++: instructions.
For AdaptiveCpp: instructions.