[SYCL][DOC] Add spec and design for "if_device" #8917
Conversation
Add a proposed extension specification and design for `if_device`, an alternative to `__SYCL_DEVICE_ONLY__`, which is implementable in a 1-pass compiler. This commit also renames the proposed extension "sycl_ext_oneapi_device_if" to "sycl_ext_oneapi_if_device_has" in order to avoid confusion with this new extension.
sycl/doc/extensions/proposed/sycl_ext_oneapi_if_device.asciidoc
Outdated
Show resolved
Hide resolved
Allow the `fn` parameter to be a function pointer for the cases that run on the host.
Specification PR is at intel#8917.
sycl/doc/extensions/proposed/sycl_ext_oneapi_if_device.asciidoc
Outdated
Show resolved
Hide resolved
Pass the callable object using perfect forwarding, rather than by
value. This enables (weird) cases where the callable object is not
copy constructible:
```
struct move_only {
move_only();
move_only(move_only &&other);
void operator()() {/* code that runs only on device */}
};
void foo() {
syclex::if_device(move_only{});
}
```
sycl/doc/design/IfDevice.md
Outdated
|
|
||
| template<typename T> | ||
| static auto if_device(T &&fn) { | ||
| detail::call_if_on_device(fn); |
There was a problem hiding this comment.
I think I do not need to use std::forward here, correct?
There was a problem hiding this comment.
Without the std::forward, if a r-value-ref gets passed to if_device then call_if_on_device gets passed a l-value-ref. If the callable only works with an r-value-ref this break. Always use forward for any universal reference (if you don't use it anymore afterwards).
There was a problem hiding this comment.
Thanks. Added in f3e15a7.
Use `std::forward` when passing callable object in code snippet.
Update the overview section and add an example showing how to interpret data differently on host vs. device.
|
Ping to reviewers (@rolandschulz @steffenlarsen). If there are no more comments, I'd like to merge this PR. Merging is not a commitment to implement this feature. We can decide that separately. |
|
This feature looks really useful, in particular we need in it in ESIMD component/development. |
No, it's not blocked by anything in particular; our focus just turned to other things. I think we just need approvals from @intel/dpcpp-doc-reviewers and @intel/llvm-reviewers-runtime, and I need to fix the merge conflicts. |
|
This pull request is stale because it has been open 180 days with no activity. Remove stale label or comment or this will be automatically closed in 30 days. |
Add a proposed extension specification and design for
if_device, an alternative to__SYCL_DEVICE_ONLY__, which is implementable in a 1-pass compiler.This commit also renames the proposed extension
"sycl_ext_oneapi_device_if" to "sycl_ext_oneapi_if_device_has" in order to avoid confusion with this new extension.