[SYCL][CUDA] Add test for a multi-device context #1102
Conversation
| sycl::event e2 = q2.memcpy(usm2, &d, 1); | ||
|
|
||
| // test combination of usm and buffers in a kernel | ||
| sycl::event e3 = add(q1, buff1, usm1, e1); |
There was a problem hiding this comment.
What is the main purpose of the test? Can we use explict q.wait() to eliminate the need for all the events?
There was a problem hiding this comment.
To check that common operations work when using two devices within the same context, including transferring buffer data between the devices and waiting on event , associated with an operation on another device. So while we could change each event synchronization with a wait on one or both of the queues, that would reduce the scope of the test.
There was a problem hiding this comment.
In this case, would the test somewhat reliably fail if we didn't pass the events around?
In other words, would race condition manifest itself on a reasonable HW with such small workloads?
There was a problem hiding this comment.
I believe a race condition would be relatively unlikely to manifest in such a test. But I still prefer events, as a race condition is not the only possible mode of failure.
There was a problem hiding this comment.
Also using waits would not make code much more readable, as there would need to be a wait call after every operation.
There was a problem hiding this comment.
Every other, I think. Anyway, I'm fine with events.
as a race condition is not the only possible mode of failure
For my education, what are the others here?
There was a problem hiding this comment.
Something completely unexpected is also always an option.
Co-authored-by: aelovikov-intel <[email protected]>
Test for intel/llvm#6446.