[SYCL][Matrix spec] Add joint_matrix_prefetch and overloads of load/store with annotated_ptr

[dkhaldi]

No description provided.

[dkhaldi]

@gmlueck, gentle ping to review the spec of overloads of load/store on annotated pointer and prefetch.

[gmlueck]

It seems like we should have some documentation about which properties can be used with joint_matrix_load and joint_matrix_store. The same goes for joint_matrix_prefetch. That function takes a property list, but there's no documentation about which properties you can pass.

[dkhaldi]

It seems like we should have some documentation about which properties can be used with joint_matrix_load and joint_matrix_store. The same goes for joint_matrix_prefetch. That function takes a property list, but there's no documentation about which properties you can pass.

I added some documentation as we will use existing properties for cache control and cache levels. Is what I added enough?

[dkhaldi]

The current spec might change as Row and Cols should be literals or const expr like in the load instructions otherwise the IGC 2d load instructions cannot be generated.
@fda0, please confirm your findings here as well.

[dkhaldi]

@gmlueck, as we discussed on Monday, I put the prefetch size as template arguments since they must be constexpr for the implementation.
Can you please review/approve?

[gmlueck]

Suggested change

	ext::oneapi::experimental::annotated_ptr<T2, PropertyListT> src,
	annotated_ptr<T2, PropertyListT> src,

No need for this namespace because the function is also defined in that namespace.

[dkhaldi]

Well, this function is defined in ext::oneapi::experimental::matrix namespace
annotated_ptr is defined in ext::oneapi::experimental namespace
Do you still think I should remove it?

[gmlueck]

Yes. This function declaration has visibility to identifiers defined in any of the parent namespaces, which includes ext::oneapi::experimental.

[gmlueck]

Suggested change

	syclex::properties{syclintelex::read_assertion<
	syclintelex::cache_control<syclintelex::cache_mode::constant,
	syclex::cache_level::L2>>}};
	syclex::properties{syclintelex::read_hint<
	syclintelex::cache_control<syclintelex::cache_mode::cached,
	syclex::cache_level::L2>>}};

The comment below says "this load will be cached to L2". Assuming that's what you want the example to show, then you want read_hint here, not read_assertion.

[gmlueck]

Suggested change

	template <typename Group, typename T1, typename T2, size_t Rows, size_t Cols,
	typename PropertyListT>
	void joint_matrix_store(Group g,
	const joint_matrix<Group, T1, use::accumulator, Rows, Cols, layout::dynamic> &res,
	ext::oneapi::experimental::annotated_ptr<T2, PropertyListT> dest,
	size_t stride, layout Layout);
	template <typename Group, typename T, size_t Rows, size_t Cols,
	typename PropertyListT>
	void joint_matrix_store(Group g,
	const joint_matrix<Group, T, use::accumulator, Rows, Cols, layout::dynamic> &res,
	ext::oneapi::experimental::annotated_ptr<T, PropertyListT> dest,
	size_t stride, layout Layout);

The other overload of joint_matrix_store has only one T template parameter (not two), so I presume you want the same here.

Note, however, in another PR I suggested changing joint_matrix_store to have two template parameters T1 and T2. So, I do think it makes sense for both overloads to have these two template parameters. However, I think you should make that change in a separate PR, which changes both overloads.

[dkhaldi]

You are right. Let's add T1 and T2 in the other PR #12276
Whichever PR gets merged first, the other one will follow.

[gmlueck]

Delete this line. syclintelex isn't used in this example.

[gmlueck]

Suggested change

	link:../../proposed/sycl_ext_oneapi_prefetch.asciidoc[sycl_ext_oneapi_prefetch.asciidoc#specifying-cache-levels]
	link:../../proposed/sycl_ext_oneapi_prefetch.asciidoc[sycl_ext_oneapi_prefetch]

I'd suggest pointing just at the extension as a whole, rather than at a specification section, as I have here. Note that the previous link wasn't right anyway, so it wasn't actually pointing at the section.

[gmlueck]

I think you should also describe the behavior when there are no properties. For example:

... as illustrated in the example below. When no cache levels are specified, the default behavior is to prefetch into the lowest level cache (i.e. L1).

[Example]

[dkhaldi]

@intel/llvm-gatekeepers, can you please merge?