How graph compiler determines the execution order operations on TPC and MME

General Questions
1

From the profiling figure, we find execution order of operations on TPC and MME are not determined by code order if operations have data dependency. I think graph compiler tries to improve the utilization of MME so graph compiler will change execution order operations.

Question1:
But what are rules graph compiler uses to determine the execution order of operations?

Question2:
Are there program rules for users to improve the utilization of both MME and TPC.

2

The rules graph compiler uses for MME/TPC ordering are complex and might change from release to release. There are no knobs from user (pytorch) code to control this.

3

Firstly, thank you so much for your reply.

Is any general rule graph compiler of the current version used to determine the execution order of operations?

4

These are internal synapse graph compiler implementations and are not open sourced