[SYCL][UR][Graph] Move L0 simultaneous graph synchronization from SYCL-RT to L0 adapter #17734
Assignees
Labels
Comments
kbenzie
pushed a commit
that referenced
this issue
Apr 3, 2025
To support the SYCL-Graph extension on an OpenCL backend, we currently only require the presence of the `cl_khr_command_buffer` extension. This PR introduces an extra requirement on the [CL_COMMAND_BUFFER_SIMULTANEOUS_USE_KHR](https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_API.html#CL_COMMAND_BUFFER_SIMULTANEOUS_USE_KHR) capability being present. This is based on the [graph execution wording](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc#765-new-handler-member-functions) on the definition of `handler::ext_oneapi_graph()` that: > Only one instance of graph will execute at any time. If graph is submitted multiple times, dependencies are automatically added by the runtime to prevent concurrent executions of an identical graph. Such usage results in multiple calls by the SYCL runtime to `urEnqueueCommandBufferExp` with the same UR command-buffer and event dependencies to prevent concurrent execution. Without support for simultaneous-use the OpenCL adapter code cannot guarantee that the first command-buffer submission has finished execution before it makes following `clEnqueueCommandBufferKHR` calls with the `cl_event` decencies. If the first submission is still executing, then an error will be reported. Workarounds like adding blocking host waits to the OpenCL UR adapter are possible, but requiring simultaneous use reflects the vendor requirements as they are for the currently implementation. I've tried to document this all in the UR spec and SYCL-Graph design docs, which also includes a couple of cleanups I found along the way. Note that the new CTS test fails for Level-Zero adapter, which I've created #17734 to resolve. --------- Co-authored-by: Mikołaj Komar <[email protected]>
kbenzie
pushed a commit
to oneapi-src/unified-runtime
that referenced
this issue
Apr 3, 2025
To support the SYCL-Graph extension on an OpenCL backend, we currently only require the presence of the `cl_khr_command_buffer` extension. This PR introduces an extra requirement on the [CL_COMMAND_BUFFER_SIMULTANEOUS_USE_KHR](https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_API.html#CL_COMMAND_BUFFER_SIMULTANEOUS_USE_KHR) capability being present. This is based on the [graph execution wording](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc#765-new-handler-member-functions) on the definition of `handler::ext_oneapi_graph()` that: > Only one instance of graph will execute at any time. If graph is submitted multiple times, dependencies are automatically added by the runtime to prevent concurrent executions of an identical graph. Such usage results in multiple calls by the SYCL runtime to `urEnqueueCommandBufferExp` with the same UR command-buffer and event dependencies to prevent concurrent execution. Without support for simultaneous-use the OpenCL adapter code cannot guarantee that the first command-buffer submission has finished execution before it makes following `clEnqueueCommandBufferKHR` calls with the `cl_event` decencies. If the first submission is still executing, then an error will be reported. Workarounds like adding blocking host waits to the OpenCL UR adapter are possible, but requiring simultaneous use reflects the vendor requirements as they are for the currently implementation. I've tried to document this all in the UR spec and SYCL-Graph design docs, which also includes a couple of cleanups I found along the way. Note that the new CTS test fails for Level-Zero adapter, which I've created intel/llvm#17734 to resolve. --------- Co-authored-by: Mikołaj Komar <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
In the
graph_impl.cppSYCL-RT code we currently have a special case for the Level Zero backend that blocks on previous executions of a graph - https://github.com/intel/llvm/blob/sycl/sycl/source/detail/graph_impl.cpp#L1005This code is backend specific so should be moved into the Unified Runtime Level-Zero v1 adapter in order to be conformance with the UR spec wording from #17658. Defining that a UR command-buffer should be able to be submitted while a previous submission is still executing.
Note that the Level-Zero v2 adapter will implemented this behavior in #17709.
The text was updated successfully, but these errors were encountered: