[SYCL] Reflect updates to get_pointer return type to relevant tests. #1670
Conversation
| @@ -122,22 +122,22 @@ template <typename T> int test(size_t Stride) { | |||
| size_t Offset = GrId * WorkGroupSize; | |||
| if (Stride == 1) { // Check the version without stride arg. | |||
| auto E = NDId.async_work_group_copy( | |||
| Local.get_pointer(), In.get_pointer() + Offset, NElemsToCopy); | |||
| local_ptr<T>(Local), In.get_pointer() + Offset, NElemsToCopy); | |||
There was a problem hiding this comment.
IIUC, this uses
template <typename ElementType,
access::decorated IsDecorated = access::decorated::legacy>
using local_ptr
deprecated value (legacy) for access::decorated template param. Is that possible to use either yes or no instead?
There was a problem hiding this comment.
@aelovikov-intel, Sounds reasonable, I would suggest doing so in another PR, though as it's not directly related to this one.
There was a problem hiding this comment.
Well, one can argue that such a change would be independent of your intel/llvm PR and consequently make this one redundant as well.
Also, the spec for async_work_group_copy says it only accepts decorated_local_ptr so I'm not even sure what exactly happens here. Do we have an implicit conversion happening?
I guess I can give you an LGTM, but still...
| @@ -122,22 +122,22 @@ template <typename T> int test(size_t Stride) { | |||
| size_t Offset = GrId * WorkGroupSize; | |||
| if (Stride == 1) { // Check the version without stride arg. | |||
| auto E = NDId.async_work_group_copy( | |||
| Local.get_pointer(), In.get_pointer() + Offset, NElemsToCopy); | |||
| local_ptr<T>(Local), In.get_pointer() + Offset, NElemsToCopy); | |||
There was a problem hiding this comment.
Well, one can argue that such a change would be independent of your intel/llvm PR and consequently make this one redundant as well.
Also, the spec for async_work_group_copy says it only accepts decorated_local_ptr so I'm not even sure what exactly happens here. Do we have an implicit conversion happening?
I guess I can give you an LGTM, but still...
|
@aelovikov-intel, thanks for the comments. An argument of async_work_group_copy is local_ptr which is a multi_ptr. In multi_ptr class there is an implicit conversion from local_accessor to multi_ptr. |
In the specification or in the implementation? Because per the spec at https://registry.khronos.org/SYCL/specs/sycl-2020/html/sycl-2020.html#table.members.nditem, all of the overloads have some kind of |
…ntel/llvm-test-suite#1670) Reflect updates to get_pointer return type to relevant tests.
Reflect updates to get_pointer return type to relevant tests.
intel/llvm#8493