CUDA dependency in autoware.auto

edit

For parallel processing:

there is only currently a limited amount of code in Autoware.auto that requires parallel acceleration. Previously OpenMP has been used https://gitlab.com/autowarefoundation.... There has not been a need for massive parallel acceleration in Autoware just yet. In the future when a need arises, we will choose an appropriate API to use.

Avoiding CUDA is deliberate because it is a proprietary API with vendor hardware lock-in. This is highly undesirable in an open-source project such as Autoware.

For Machine Learning:

We are currently working on adding machine learning facilities in Autoware via TVM https://gitlab.com/autowarefoundation.... This allows the abstraction of backends. We have chosen libvulkan for our reference implementation. This means there is no vendor lockin for hardware acceleration.

Autoware.Auto does not depend on CUDA. We will be adding machine-learning components but the back-ends which will be supported first are Vulkan and OpenCL due to them being more open and hardware-agnostic than CUDA. Please see https://gitlab.com/autowarefoundation... for tracking progress.