[SYCL][Graph] Implement dynamic command-groups #15700
Draft
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
EwanC
force-pushed
the
ewan/dyn_cgf_impl
branch
from
9ddbe49
to
266d60c
Compare
EwanC
changed the title
EwanC
force-pushed
the
ewan/dyn_cgf_impl
branch
from
266d60c
to
70652e0
Compare
EwanC
force-pushed
the
ewan/dyn_cgf_impl
branch
from
70652e0
to
6f38259
Compare
EwanC
force-pushed
the
ewan/dyn_cgf_impl
branch
from
6f38259
to
62c50d3
Compare
EwanC
force-pushed
the
ewan/dyn_cgf_impl
branch
from
62c50d3
to
c1dfee9
Compare
EwanC
force-pushed
the
ewan/dyn_cgf_impl
branch
from
c1dfee9
to
1c16b5b
Compare
EwanC
force-pushed
the
ewan/dyn_cgf_impl
branch
from
1c16b5b
to
bd67c82
Compare
EwanC
force-pushed
the
ewan/dyn_cgf_impl
branch
from
bd67c82
to
54db889
Compare
EwanC
force-pushed
the
ewan/dyn_cgf_impl
branch
from
54db889
to
b8bbdd6
Compare
Implement Dynamic Command-Group feature specified in PR [[SYCL][Graph] Add specification for kernel binary updates](intel#14896) This feature enables updating `ur_kernel_handle_t` objects in graph nodes between executions as well as parameters and execution range of nodes. This functionality is currently supported on CUDA & HIP which are used for testing in the new E2E tests. Level Zero support will follow shortly, resulting in the removal of the `XFAIL` labels from the E2E tests. The code for adding nodes to a graph has been refactored to split out verification of edges, and marking memory objects used in a node, as separate helper functions. This allows path for adding a command-group node to do this functions over each CG in the list before creating the node itself. The `dynamic_parameter_impl` code has also been refactored so the code is shared for updating a dynamic parameter used in both a regular kernel node and a dynamic command-group node. See the addition to the design doc for further details on the implementation.
Create helper function for getting UR details out of CG.
EwanC
force-pushed
the
ewan/dyn_cgf_impl
branch
from
b8bbdd6
to
c838513
Compare
EwanC
force-pushed
the
ewan/dyn_cgf_impl
branch
from
12decab
to
319c43e
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Implement Dynamic Command-Group feature specified in PR [SYCL][Graph] Add specification for kernel binary updates. This feature enables updating
ur_kernel_handle_tobjects in graph nodes between executions as well as parameters and execution range of nodes.Points to note in this change:
The functionality is currently supported on CUDA & HIP which are used for testing in the new E2E tests. Level Zero support will follow shortly, resulting in the removal of the
XFAILlabels with tracker number from the E2E tests.The code for adding nodes to a graph has been refactored to split out verification of edges, and marking memory objects used in a node, as separate helper functions. This allows path for adding a command-group node to do this functions over each CG in the list before creating the node itself.
The
dynamic_parameter_implcode has also been refactored so the code is shared for updating a dynamic parameter used in both a regular kernel node and a dynamic command-group node.There is now no need for the
handler::setNDRangeUsed()API now that graph kernel nodes can update between kernels usingsycl::nd_rangeandsycl::range. The functionality in this method has be turned into a no-op, however removing the method is an ABI breaking change, so it remains guarded by the__INTEL_PREVIEW_BREAKING_CHANGESmacro.See the addition to the design doc for further details on the implementation.