liuzf1988

I have some problems when compiling centerpoint with TVM, both for "llvm" backend and "cuda" backend. I'm not sure if the problem is caused by TVM or improper configuration file...

> Your configuration file looks correct. I think the issue comes from TVM not having good support for dynamic parameters. > > I'm not sure what is the exact effect...

> Although it seems we can't compile this model with dynamic parameters with the current state of TVM, I can compile it just fine with fixed parameters. The code for...

> @liuzf1988 Are there any updates on this task? @xmfcx So busy these days, sorry for not being able to reply to the message in time. If there is no...

@ambroise-arm Referring to the official TVM documentation and the existing script (tvm_cli.py) of autoware model zoo, I wrote the independent script for compiling the neuron network model. However, there are...

> @liuzf1988 Do you have any progress with this issue? @angry-crab might be able to help you if there are any other blocking items. A ROS2 package named `lidar_centerpoint_tvm` was...

@angry-crab Yes, it take some time to rewrite the `cuda` code (currently it is changed to `std::thread` implementation). Can I submit the code to a pull request tomorrow (I need...

> @liuzf1988 Thanks for the update. I had a look into the current `centerpoint` implementation. It is using some hand-crafted `cuda` kernels for processing intermediate data. Did you convert these...

> The [QoS setting](https://docs.ros.org/en/humble/Concepts/About-Quality-of-Service-Settings.html) in ROS2 is a little complicated. In brief, QoS profiles can be specified for publishers, subscriptions, service servers and clients. A QoS profile can be applied...