Check for support of specific type of USM in tests #420
Conversation
dmitriy-sobolev
requested review from
MikeDvorskiy,
dparanic,
KseniiaBakina and
SergeyKopienko
|
@SergeyKopienko, could you review this patch? |
|
@dmitriy-sobolev, please rebase from main this branch. |
| @@ -117,6 +117,10 @@ void test2_with_buffers() | |||
| void test_with_usm() | |||
| { | |||
| cl::sycl::queue q; | |||
|
|
|||
| if(!TestUtils::has_aspect(q.get_device(), sycl::aspect::usm_shared_allocations)) | |||
| return; | |||
There was a problem hiding this comment.
I think that this function should should become boolean. And when we return from this function, we should return as in skip test vase from main.
There was a problem hiding this comment.
The buffers will be tested even If test_with_usm is skipped. Given that modifying return seems unnecessary.
Do you think there should be a notification that the test case with USM has been skipped (e.g. via some message in the output)?
| @@ -119,6 +119,10 @@ void test_with_buffers() | |||
| void test_with_usm() | |||
| { | |||
| sycl::queue q; | |||
|
|
|||
| if(!TestUtils::has_aspect(q.get_device(), sycl::aspect::usm_shared_allocations)) | |||
There was a problem hiding this comment.
I think that this function should should become boolean. And when we return from this function, we should return as in skip test vase from main.
There was a problem hiding this comment.
Let's continue discussion on the same issue in this thread: #420 (comment)
dmitriy-sobolev
force-pushed
the
dev/dmitriy-sobolev/check-usm-support
branch
from
e681327 to
a03ba23
Compare
Signed-off-by: Sobolev, Dmitriy <[email protected]>
Signed-off-by: Sobolev, Dmitriy <[email protected]>
* Rebase * Add const to sycl::exception and log it * Map sycl::usm::alloc to usm sycl::aspect Signed-off-by: Sobolev, Dmitriy <[email protected]>
dmitriy-sobolev
force-pushed
the
dev/dmitriy-sobolev/check-usm-support
branch
from
a03ba23 to
af3d153
Compare
Signed-off-by: Sobolev, Dmitriy <[email protected]>
* Apply fixes required after rebasing Signed-off-by: Sobolev, Dmitriy <[email protected]>
Remove extra line
Signed-off-by: Sobolev, Dmitriy <[email protected]>
|
The PR is outdated, and I do not deem it to be worth of revival. More tests have been added which rely on USM only: they need to be rewritten using SYCL buffers. That's a lot of work not connected to the changes here, hence it is better to start from scratch. I've created an issue for it: #1948 (comment) |
Signed-off-by: Sobolev, Dmitriy [email protected]
Some devices do not support shared allocations, e.g. some FPGA cards/configurations. Using not supported allocation will result in getting nullptr and then accessing it.
2021.1.2 compiler throws "sycl::runtime_error" and returns "This device aspect has not been implemented yet".
The patch which adds support of USM aspects: intel/llvm@ce3ac09