In this exercise you will learn how to use the USM API to write a SYCL application which performs a vector add.
Allocate memory on the host for the input and out data just as you did when using the buffer/accessor model.
Create a queue using the USM device selector from the “Introduction to USM”
exercise, remember to handle errors.
When using the USM model the first thing you need to do is allocate the USM device memory.
To do this call malloc_device to allocate memory for the two inputs and the
output.
Before you can perform any computation on the data you must copy it to the device.
To do this call the queue member function memcpy for each of the two inputs,
remember to call wait on the event that is returned.
Now you can define the kernel function itself, which is largely the same as in the “Data Parallelism” exercise.
This can be done differently from the buffer/accessor model, by calling the
shortcut member function parallel_for on the queue rather than creating a
command group.
Note that as you are accessing a pointer rather than an accessor you must
retrieve an integral index, which can be done by calling the subscript operator
on the id passed into the kernel function, with index 0.
Remember to name your kernel function, and to call wait on the event that is
returned.
Once the kernel function has completed you can copy the result back to the host.
As you did when copying to the device, you can do this by calling the queue
member function memcpy, again remember to call wait on the event that is
returned.
Finally once you have copied the data back from the device you can free that memory.
To do this call free on each of the USM device allocations, note this is the
SYCL API free and not the standard C free.
For DevCloud via JupiterLab follow these instructions.
For DPC++: instructions.
For AdaptiveCpp: instructions.