Created by: dzhwinter

This is an experiment branch which I am doing some experiment on the variable reuse is based on reference counting and dependency analysis.

Now we go two branches of memory optimize, one of them is runtime memory optimize. The MXNet does go this way, and do excellent job on memory consumption reducing. https://mxnet.incubator.apache.org/architecture/note_engine.html https://mxnet.incubator.apache.org/architecture/note_memory.html However, we have two questions.

• The GPU is an asynchronize device, is it possible to release the memory in time.
• How is the fancy memory reuse technique fit in Fluid.