An hexagon autoscheduler problems about halide. #7001
-
|
When I directly use the conv3x3 code automatically generated by halide 13 autoscheduler in offload mode(target=arm-64-linux-hvx-hvx_128), the vector length is always wrong(use arm cpu vector length). However, the vector length generated by halide software using Qualcomm's official SDK is correct and has better performance. |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments
-
|
I don't believe any of the Halide autoschedulers are trained for HVX, so you should not expect them to work. Using Halide with HVX requires writing manual schedules, and I believe that's what the Qualcomm SDK does. |
Beta Was this translation helpful? Give feedback.
-
|
Hi @pengxuan2022 - depending on the version of the Qualcomm SDK you have we do have an experimental autoscheduler which will use the right vector length. We believe it still needs some work before it is in a state worthy of submitting upstream - at the moment it is a little hacky. Like Andrew said, the community autoscheduler will not schedule for Hexagon. |
Beta Was this translation helpful? Give feedback.
Hi @pengxuan2022 - depending on the version of the Qualcomm SDK you have we do have an experimental autoscheduler which will use the right vector length. We believe it still needs some work before it is in a state worthy of submitting upstream - at the moment it is a little hacky. Like Andrew said, the community autoscheduler will not schedule for Hexagon.