search

nvidia-holoscan / holohub

Central repository for applications and operators for Holoscan
Apache License 2.0
104 stars 65 forks source link

VTK Operator: Add Performance Benchmarks #276

Closed tbirdso closed 1 month ago

tbirdso commented 5 months ago

Add information on performance benchmarks for the VTK Operator contributed in https://github.com/nvidia-holoscan/holohub/pull/261.

having some benchmarkings compared to holoviz would be beneficial. VTK brings a huge convenience given the breadth of capabilities and the existing user base. At the same time, it would be great to also ensure we're setting it up for optimal performance as that is Holoscan's mission.

Yes, in order to pass the data to VTK we need to first copy to host(cpu). AFAIK, there is no easy way of rendering directly from CUDA to the GPU framebuffer with VTK. Perhaps something can be done in the next iterations of this task and with changes in upstream VTK. Perhaps the best course of action is to have this first version of the vtk_renderer operator merged and later explore (if possible) how to further optimize this, this is to make VTK both: Accept GPU memory, Render to the GPU directly.

tbirdso commented 5 months ago

You can refer to the holoscan_flow_benchmarking guide for convenient HoloHub benchmarking tools: https://github.com/nvidia-holoscan/holohub/tree/main/benchmarks/holoscan_flow_benchmarking

tbirdso commented 5 months ago

cc @vicentebolea

tbirdso commented 1 month ago

Closing as not a bug. Future benchmark contributions are welcome.