[SYCL] Use custom type to pass CGF around instead of std::function
#16668
Conversation
aelovikov-intel
requested review from
cperkinsintel,
AlexeySachkov,
sergey-semenov and
steffenlarsen
aelovikov-intel
had a problem deploying
to
WindowsCILock
GitHub Actions
Error
aelovikov-intel
had a problem deploying
to
WindowsCILock
GitHub Actions
Error
| // UseFallBackAssert as template param vs `#if` in function body is necessary | ||
| // to prevent ODR-violation between TUs built with different fallback assert | ||
| // modes. |
There was a problem hiding this comment.
@steffenlarsen , any idea if
llvm/sycl/include/sycl/queue.hpp
Lines 792 to 793 in a87ac06
There was a problem hiding this comment.
The "fallback kernel" in this case is not referring to fallback asserts. It's because the implementation of ext_oneapi_memcpy2d has a kernel it uses if the operation isn't natively supported, but we don't want that kernel to pop up unless the user calls ext_oneapi_memcpy2d in their code, so we make it dependent.
aelovikov-intel
added
the
run-perf-tests
* `std::function` is affecting compile-time too much * our `queue::submit` is a *synchronous* operation, so we don't need to make any copy. Maybe `std::function_ref` would be a good choice here, but that's C++26. * Try to convert typed CGFO into type-erased version as soon as possible to limit number of template instantiations FE needs to perform
aelovikov-intel
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
aelovikov-intel
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
| // TODO: Is that true? | ||
| // As such, we know that it can't be a [member] function pointer. |
There was a problem hiding this comment.
@gmlueck , is this true?
| Q.submit_without_event< | ||
| #if __SYCL_USE_FALLBACK_ASSERT | ||
| true | ||
| #else | ||
| false | ||
| #endif | ||
| >(Props, detail::type_erased_cgfo_ty{CGF}, CodeLoc); |
There was a problem hiding this comment.
Looks like we always have __SYCL_USE_FALLBACK_ASSERT set to a value, could we not just use it directly here?
| Q.submit_without_event< | |
| #if __SYCL_USE_FALLBACK_ASSERT | |
| true | |
| #else | |
| false | |
| #endif | |
| >(Props, detail::type_erased_cgfo_ty{CGF}, CodeLoc); | |
| Q.submit_without_event<__SYCL_USE_FALLBACK_ASSERT>(Props, detail::type_erased_cgfo_ty{CGF}, CodeLoc); |
|
|
||
| void operator()(sycl::handler &cgh) const { invoker_f(object, cgh); } | ||
| }; | ||
|
|
There was a problem hiding this comment.
When I first saw this I was wondering if we could use std::invokable instead, but I think that'd change the ABI, plus it might have the same too-much-templating problem as std::function
There was a problem hiding this comment.
I don't see how it's applicable here. std::invokable is a concept, not a data type; so more like std::is_invokable_r trait that we are already using.
std::functionis affecting compile-time too muchqueue::submitis a synchronous operation, so we don't need to make any copy. Maybestd::function_refwould be a good choice here, but that's C++26.