When possible, data tasks should be elided just before generating the function code. This means that instead of generating calls to identity for data tasks that do not move data between devices, the subsequent uses of the variable should be renamed. For scheduling on a single device, this would reduce the number of lines and variables in a function by about 50%, which should speedup compilation.
I'm not yet sure where the best place is to do this elision. The DAG itself should not be changed for this.
When possible, data tasks should be elided just before generating the function code. This means that instead of generating calls to
identityfor data tasks that do not move data between devices, the subsequent uses of the variable should be renamed. For scheduling on a single device, this would reduce the number of lines and variables in a function by about 50%, which should speedup compilation.I'm not yet sure where the best place is to do this elision. The DAG itself should not be changed for this.