Frequently Asked Questions (FAQ)

Q. Does Barracuda work on iPhone / Android / Magic Leap / Switch / PS4 / Xbox?

• A. Yes. Barracuda supports all platforms that Unity supports. Some restrictions apply however: for more information see Supported platforms.

Q. Can I import my model from TensorFlow / Pytorch / Keras?

• A. Yes. You must first convert it to ONNX and then load it into Unity. For more information see Model exporting.

Q. What operators does Barracuda support?

• A. See the list of supported operators.

Q. Does Barracuda support training?

• A. No. Barracuda currently only supports inference.

Q. Can Barracuda run on a GPU?

• A. Yes. You must specify the GPU backend when creating a Worker.

Q. Can Barracuda run on a CPU?

• A. Yes. You must specify the CPU backend when creating a Worker.

Q. What is the difference between ComputeRef, Compute and ComputePrecompiled?

• A. A Compute<***> WorkerType is a worker that operates on the GPU.
  • ComputeRef is a slow reference implementation. It is useful for debugging but uses more memory.
  • Compute is a fast GPU implementation. It has some CPU overhead but is flexible to a change of input dimensions.
  • ComputePrecompiled strips out all CPU overhead, but optimizes the execution for a given input dimension. If the input dimensions change, there is a compile overhead.

Q. What is the difference between CSharpRef , CSharp andCSharpBurst?

• A. A CSharp<***> WorkerType is a worker that operates on the CPU.
  • CSharpRef is a very slow reference implementation. It is useful for debugging.
  • CSharp is a fast C# synchronous implementation.
  • CSharpBurst is a very fast asynchronous CPU implementation.

Q. My operator is not supported, what do I do?

• A. Please post the issue on the Barracuda GitHub repository and the Unity Barracuda team will do their best to help.

Q. I found a bug. What do I do?

• A. Please post the issue on the Barracuda GitHub repository and the Unity Barracuda team will do their best to help.

Q. Execution is slow.

• A.
  • GPU execution can be slower than CPU inference on mobile.
  • When reading back data from the GPU, there needs to be synchronization, which can induce lag.
  • Keep all data on the same pipe as much as possible.
  • Please post the issue on the Barracuda GitHub repository and the Unity Barracuda team will do their best to help.

Q. I get a lot of garbage collection warnings when using Barracuda.

• A. Ensure that you Dispose all allocated Tensor and Worker objects. For more information see Memory management.