Skip to content
Merged
Show file tree
Hide file tree
Changes from 6 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Original file line number Diff line number Diff line change
Expand Up @@ -233,6 +233,9 @@ namespace sycl::ext::oneapi::experimental {
template <typename CommandGroupFunc>
void submit(sycl::queue q, CommandGroupFunc&& cgf);

template <typename CommandGroupFunc, typename Properties>
void submit(sycl::queue q, Properties properties, CommandGroupFunc&& cgf);

}
----
!====
Expand All @@ -250,6 +253,10 @@ namespace sycl::ext::oneapi::experimental {
template <typename CommandGroupFunc>
sycl::event submit_with_event(sycl::queue q, CommandGroupFunc&& cgf);

template <typename CommandGroupFunc, typename Properties>
sycl::event submit_with_event(sycl::queue q, Properties properties,
CommandGroupFunc&& cgf);

}
----
!====
Expand Down
165 changes: 165 additions & 0 deletions sycl/doc/extensions/proposed/sycl_ext_intel_event_mode.asciidoc
Original file line number Diff line number Diff line change
@@ -0,0 +1,165 @@
= sycl_ext_intel_event_mode

:source-highlighter: coderay
:coderay-linenums-mode: table

// This section needs to be after the document title.
:doctype: book
:toc2:
:toc: left
:encoding: utf-8
:lang: en
:dpcpp: pass:[DPC++]
:endnote: &#8212;{nbsp}end{nbsp}note

// Set the default source code type in this document to C++,
// for syntax highlighting purposes. This is needed because
// docbook uses c++ and html5 uses cpp.
:language: {basebackend@docbook:c++:cpp}

:common_ref_sem: https://registry.khronos.org/SYCL/specs/sycl-2020/html/sycl-2020.html#sec:reference-semantics

== Notice

[%hardbreaks]
Copyright (C) 2024 Intel Corporation. All rights reserved.

Khronos(R) is a registered trademark and SYCL(TM) and SPIR(TM) are trademarks
of The Khronos Group Inc. OpenCL(TM) is a trademark of Apple Inc. used by
permission by Khronos.


== Contact

To report problems with this extension, please open a new issue at:

https://github.com/intel/llvm/issues


== Dependencies

This extension is written against the SYCL 2020 revision 9 specification. All
references below to the "core SYCL specification" or to section numbers in the
SYCL specification refer to that revision.

This extension also depends on the following other SYCL extensions:

* link:../experimental/sycl_ext_oneapi_enqueue_functions.asciidoc[
sycl_ext_oneapi_enqueue_functions]
* link:../experimental/sycl_ext_oneapi_properties.asciidoc[
sycl_ext_oneapi_properties]


== Status

This is a proposed extension specification, intended to gather community
feedback. Interfaces defined in this specification may not be implemented yet
or may be in a preliminary state. The specification itself may also change in
incompatible ways before it is finalized. *Shipping software products should
not rely on APIs defined in this specification.*


== Overview

On some backends, calling `wait()` on an `event` will synchronize using a
busy-waiting implementation. Though this comes at a low latency for the
synchronization of the event, it has the downside of consuming high amounts of
CPU time for no meaningful work. This extension introduces a new property for
SYCL commands that allow users to pick modes for the associated events, one of
these modes being a "low-power" event. These new low-power events will, if
possible, yield the thread that the `wait()` member function is called on and
only wake up occasionally to check if the event has finished. This reduces the
time the CPU spends checking finish condition of the wait, at the cost of
latency.


== Specification

=== Feature test macro

This extension provides a feature-test macro as described in the core SYCL
specification. An implementation supporting this extension must predefine the
macro `SYCL_EXT_INTEL_EVENT_MODE` to one of the values defined in the table
below. Applications can test for the existence of this macro to determine if
the implementation supports this feature, or applications can test the macro's
value to determine which of the extension's features the implementation
supports.

[%header,cols="1,5"]
|===
|Value
|Description

|1
|The APIs of this experimental extension are not versioned, so the
feature-test macro always has this value.
|===


=== Event mode property

This extension adds a new property `event_mode` which can be used with the
`submit_with_event` free function from
link:../experimental/sycl_ext_oneapi_enqueue_functions.asciidoc[sycl_ext_oneapi_enqueue_functions],
allowing the user some control over how the resulting event is created and
managed.

```
namespace sycl::ext::intel::experimental {

enum class event_modes { none, low_power };
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@Pennycook: Do you like this naming convention where the enum name is plural and the property name is singular?

The other alternative we discussed is to name the property event_mode and name the enum event_mode_enum.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A quick search of the ISO C++ specification suggests that their convention is to use a plural when the enum is a bitmask (e.g., std::filesystem::copy_options) but a singular when it is not (e.g., std::filesystem::file_type).

If we want to keep the plural here, then I think we should define this as a bitmask. But that also suggests that the property name should also be a plural. Didn't we recently discuss reintroducing a nested namespace for properties? That would allow sycl::event_modes to be a bitmask of event modes, and sycl::property::event_modes to be a property for setting an event's modes.

I don't really like the idea of encoding _enum in a name, because of the redundancy there.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think we should make this a bitmask.

We have already used the _enum suffix for some other properties, so we have a precedent.

If we did add the namespace, why wouldn't we put the property and the enum in the same namespace? The enum is only used in conjunction with the property.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@steffenlarsen, I saw that you re-requested my review. I think the naming issue in this thread is not yet resolved.

If we can't think of a better solution, I suggest we go with the name event_mode_enum, which is consistent with existing properties that take an enum.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah, apologies. I glanced over this one. event_mode_enum is fine with me. 👍


struct event_mode {
event_mode(event_modes mode);

event_modes value;
};

using event_mode_key = event_mode;

} // namespace sycl::ext::intel::experimental
```


=== Low power event mode

Passing the `event_mode` property with `event_modes::low_power` to
`submit_with_event` will act as a hint to the `event` created from the
corresponding commands to do low-power synchronization. If the backend is able
to handle low-power events, calling `event::wait()` or `event::wait_and_throw()`
will cause the thread to yield and only do occasional wake-ups to check the
event progress.

[_Note:_ The low-power event mode currently only has an effect on `barrier` and
`partial_barrier` commands enqueued on queues that return
`backend::ext_oneapi_level_zero` from `queue::get_backend()`.
_{endnote}_]


=== New property usage example

As an example of how to use the new `low_power_event` property, see the
following code:

```
#include <sycl/sycl.hpp>

namespace oneapiex = sycl::ext::oneapi::experimental;
namespace intelex = sycl::ext::intel::experimental;

int main() {
sycl::queue Q;

// Submit some work to the queue.
oneapiex::submit(Q, [&](sycl::handler &CGH) {...});

// Submit a command with the low-power event property.
oneapiex::properties Props{intelex::low_power_event};
sycl::event E = oneapiex::submit_with_event(Q, Props, [&](sycl::handler &CGH) {
...
});

// Waiting for the resulting event will use low-power waiting if possible.
E.wait();
}
```
23 changes: 20 additions & 3 deletions sycl/include/sycl/ext/oneapi/experimental/enqueue_functions.hpp
Original file line number Diff line number Diff line change
Expand Up @@ -95,21 +95,38 @@ void submit_impl(queue &Q, CommandGroupFunc &&CGF,
}
} // namespace detail

template <typename CommandGroupFunc>
void submit(queue Q, CommandGroupFunc &&CGF,
template <typename CommandGroupFunc, typename PropertiesT>
void submit(queue Q, PropertiesT Props, CommandGroupFunc &&CGF,
const sycl::detail::code_location &CodeLoc =
sycl::detail::code_location::current()) {
std::ignore = Props;
sycl::ext::oneapi::experimental::detail::submit_impl(
Q, std::forward<CommandGroupFunc>(CGF), CodeLoc);
}

template <typename CommandGroupFunc>
event submit_with_event(queue Q, CommandGroupFunc &&CGF,
void submit(queue Q, CommandGroupFunc &&CGF,
const sycl::detail::code_location &CodeLoc =
sycl::detail::code_location::current()) {
submit(Q, empty_properties_t{}, std::forward<CommandGroupFunc>(CGF), CodeLoc);
}

template <typename CommandGroupFunc, typename PropertiesT>
event submit_with_event(queue Q, PropertiesT Props, CommandGroupFunc &&CGF,
const sycl::detail::code_location &CodeLoc =
sycl::detail::code_location::current()) {
std::ignore = Props;
return Q.submit(std::forward<CommandGroupFunc>(CGF), CodeLoc);
}

template <typename CommandGroupFunc>
event submit_with_event(queue Q, CommandGroupFunc &&CGF,
const sycl::detail::code_location &CodeLoc =
sycl::detail::code_location::current()) {
return submit_with_event(Q, empty_properties_t{},
std::forward<CommandGroupFunc>(CGF), CodeLoc);
}

template <typename KernelName = sycl::detail::auto_name, typename KernelType>
void single_task(handler &CGH, const KernelType &KernelObj) {
CGH.single_task<KernelName>(KernelObj);
Expand Down
Loading