You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The suggested graph update mechanism seems computationally complex: an implementation has to traverse both the modifiable and executable graphs to identify any changes to either inputs or outputs.
I better like the OpenCL approach, where one directly updates a cl_mutable_command_khr handle. We can add an update method to class node. We can also add graph::get_node(size_t index) for obtaining the nodes created during queue recording.
The text was updated successfully, but these errors were encountered:
Hi Aharon, thanks for your input. The current draft update mechanism is aimed at library use cases, where an application creates an graphs from recording a library call, where they don't have access to SYCL objects like buffers/USM to update.
A more fine grained update mechanism is something we're definitely looking at adding for future versions of the spec, so will keep a design like this in mind.
The suggested graph update mechanism seems computationally complex: an implementation has to traverse both the modifiable and executable graphs to identify any changes to either inputs or outputs.
I better like the OpenCL approach, where one directly updates a cl_mutable_command_khr handle. We can add an update method to
class node
. We can also addgraph::get_node(size_t index)
for obtaining the nodes created during queue recording.The text was updated successfully, but these errors were encountered: