alibaba / heterogeneity-aware-lowering-and-optimization

heterogeneity-aware-lowering-and-optimization
Apache License 2.0
253 stars 76 forks source link

[ODLA/TRT] Support multiple device instances and concurrent computation #958

Closed weimingzha0 closed 2 years ago

weimingzha0 commented 2 years ago

Since cuda current device is a TLS variable, if ODLA apis are called in different thread context, we need to set current device again.