SYCL Command Graph Extensions

This fork is mostly stale since the main development has been moved to the Intel staging area for llvm.org contributions.

This fork has been the collaboration space for the oneAPI vendor Command Graph extension for SYCL2020 until September 2023. SYCL Graph provides an API for defining a graph of operations and their dependencies once and submitting this graph repeatedly for execution.

Specification

A draft of our Command Graph extension proposal can be found here: sycl_ext_oneapi_graph.

Implementation

Our implementation can be found here: https://github.com/intel/llvm.

Backends

An application can query the SYCL library for the level of support it provides for using the extension with a device by using ext::oneapi::experimental::info::device::graph_support, which returns one of:

• Native - Backend command-buffer construct is used to implement graphs.
• Emulated - Graphs support is emulated by reissuing commands to the backend.
• Unsupported - Extension is not supported on the device.

Currently the Level Zero backend is the only supported SYCL backend for the sycl_ext_oneapi_graph extension. As the focus of the current prototype is good Level Zero support to prove the value of the extension, rather than emulated support for many backends. However, broadening the number of backends supported is something we are interested in expanding on.

Backend	Implementation Support
Level Zero	Native using command-lists
CUDA	Native support is a work-in-progress
OpenCL	Unsupported
HIP	Unsupported
Others	Unsupported

Implementation Status

Feature	Implementation Status (- error type)
Adding a command-group node with command_graph::add()	Implemented
Begin & end queue recording to a graph to create nodes	Implemented
Edges created from buffer accessor dependencies	Implemented
Edges created from handler::depends_on dependencies	Implemented
Edges created using make_edge()	Implemented
Edges created by passing a property list to command_graph::add()	Implemented
Empty node	Implemented
Queue ext_oneapi_get_state() query	Implemented
Vendor test macro	Implemented
Ability to add a graph as a node of another graph (Sub-graphs)	Implemented
Using all capabilities of USM in a graph node	Implemented
Extending lifetime of buffers used in a graph, as defined by the "Storage Lifetimes" specification section	Not implemented - Throws an exception that feature is not supported yet
Buffer taking a copy of underlying host data when buffer is used in a graph, as defined by the "Storage Lifetimes" specification section	Not implemented - Throws an exception that feature is not supported yet
Executable graph update()	Not implemented - Exception "Method not yet implemented"
Recording an in-order queue preserves linear dependencies	Implemented
Using handler::parallel_for in a graph node	Implemented
Using handler::single_task in a graph node	Implemented
Using handler::memcpy in a graph node	Implemented
Using handler::copy in a graph node	Implemented
Using handler::host_task in a graph node	Not implemented - Assert: "getCGCopy() const: Assertion `false' failed."
Using handler::fill in a graph node	Implemented for USM, not implemented for buffer accessors - Exception: "CG type not implemented for command buffers"
Using handler::memset in a graph node	Not implemented - Exception: "CG type not implemented for command buffers"
Using handler::prefetch in a graph node	Not implemented - Exception: "CG type not implemented for command buffers"
Using handler::memadvise in a graph node	Not implemented - Exception: "CG type not implemented for command buffers"
Using specialization constants in a graph node	Not implemented - Throws an exception that feature is not supported yet
Using reductions in a graph node	Not implemented - Throws an exception that feature is not supported yet
Using kernel bundles in a graph node	Not implemented - Throws an exception that feature is not supported yet
Using sycl streams in a graph node	Not implemented - Exception: "Failed to add kernel to PI command-buffer"
Thread safety of new methods	Implemented
Profiling an event returned from graph submission with event::get_profiling_info()	Not implemented - Throws an exception that feature is not supported yet
Querying the state of an event returned from graph submission with event::get_info<info::event::command_execution_status>()	Implemented

Other Material

This extension was presented at the oneAPI Technical Advisory board (Sept'22 meeting). Slides: https://github.com/oneapi-src/oneAPI-tab/blob/main/language/presentations/2022-09-28-TAB-SYCL-Graph.pdf.

Extension was presented at IWOCL 2023, and the talk can be found on Youtube.

Intel Project for LLVM* technology

We target a contribution through the origin of this fork: Intel staging area for llvm.org contributions.

How to use DPC++

Releases

TDB

Build from sources

See Get Started Guide.

Report a problem

Submit an issue or initiate a discussion.

How to contribute to DPC++

This project welcomes contributions from the community. Please refer to CONTRIBUTING for general guidelines around contributing to this project. You can then see ContributeToDPCPP for DPC++ specific guidelines.

License

See LICENSE for details.

_{*Other names and brands may be claimed as the property of others.}