marcoabreu commented on a change in pull request #11325: Added TensorRT runtime
integration
URL: https://github.com/apache/incubator-mxnet/pull/11325#discussion_r202307264
##########
File path: src/executor/exec_pass.h
##########
@@ -198,6 +198,27 @@ Graph InferStorageType(Graph&& graph,
StorageTypeVector&& storage_type_inputs =
StorageTypeVector(),
const std::string& storage_type_attr_key = "");
+/*! \brief The default storage type inference function, which assigns all
undefined
+ * storage types to kDefaultStorage. If all of input and output
storage types
+ * are kDefaultStorage, DispatchMode::kFCompute is assigned to
dispatch_mode. Otherwise,
+ * DispatchMode::kFComputeFallback is assigned to dispatch_mode.
+ */
+bool DefaultStorageType(const nnvm::NodeAttrs& attrs,
+ const int dev_mask,
+ DispatchMode* dispatch_mode,
+ std::vector<int> *iattr,
+ std::vector<int> *oattr);
+
+/*!
+ * \brief Replace subgraphs by TRT (forward only)
+ */
+Graph ReplaceSubgraph(Graph&& g,
+ const std::unordered_set<nnvm::Node*>& set_subgraph,
+ std::unordered_map<std::string, NDArray>* const
params_map);
+
+std::vector<std::unordered_set<nnvm::Node*>> GetTrtCompatibleSubsets(const
Graph& g,
Review comment:
Yeah that's a good idea.
It might be out of scope, but I want to highlight that this technical debt
that's going to be increasingly harder to fix the more the code evolves. Right
now, it's still in a state that allows us to redesign it as we wish. Later on
we might be limited due to increasing complexity or compatibility constraints.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
With regards,
Apache Git Services
