[SYCL][Graph] Fix issues with sycl_ext_oneapi_graph subgraphs #10822
Conversation
This file contains hidden or 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
* [SYCL][Graph] Duplicate sub-graph nodes Duplicates the nodes of a sub-graph when added to its parents to enable multiple parents. As a result nodes of the initial sub-graph and the one of the main graph are no longer the exact same (shared_pointer) but two different nodes with similar content. A few unitests have been updated accordingly and an equal operator has been added to node_impl to ease comparison between nodes.
This was referenced Aug 15, 2023
* [SYCL][Graph] Enable empty nodes in Subgraphs The implementation uses the list of scheduled nodes to add a subgraph to a main graph. However, since empty nodes are not scheduled, empty nodes were not listed in this list, resulting in inconsistent graphs when subgraph with empty node(s) were added to a main graph. This PR fixes this issue by forcing to list empty nodes when creating the list for inserting subgraph. It also adds unitests to check that subgraphs with empty nodes are correctly added to a main graph. * [SYCL][Graph] Enable empty nodes in Subgraphs Changes the definition of MSchedule as it now contains all types of nodes (including empty nodes). Empty nodes are now filtered out when creating the commandbuffer and/or enqueuing nodes to only keep their dependencies. This PR updates a few unitests to make them compliant to the new schedule list definition. * Update sycl/source/detail/graph_impl.cpp Co-authored-by: Ben Tracy <[email protected]> --------- Co-authored-by: Ben Tracy <[email protected]>
Bensuo
approved these changes
Aug 17, 2023
…location policy (#303)
reble
changed the title
|
CI fail for |
steffenlarsen
approved these changes
Aug 24, 2023
| auto Successor = NodesMap[NextNode]; | ||
| NodeCopy->registerSuccessor(Successor, NodeCopy); | ||
| } else { | ||
| assert("Node duplication failed. A duplicated node is missing."); |
There was a problem hiding this comment.
This caused MacOS build failure in post commit. Fixing in #10977.
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.
This PR addresses some issues with the sycl_ext_oneapi_graph extension regarding sub-graphs. A sub-graph is when an executable
command_graphis added as a node in another graph.Currently, adding an executable graph object as a subgraph permanently connects the child graph's root nodes to the parent graph. This affects subsequent independent submissions of the executable child graph object and later additions as a subgraph node. This is incorrect behaviour, as the previous uses as a sub-graph should have no effect on the state of the executable graph object.
Fixed in this PR by duplicating the nodes of a subgraph when added to a parent, enabling multiple parents. As a result, nodes of the initial subgraph and the one of the main graph are no longer the exact same (shared_pointer) but two different nodes with similar content.
Authors
Co-authored-by: Maxime France-Pillois [email protected]