What's the relationship between oneDNN and caffe2 in pytorch

C++
1

Hello everyone, recently I am learning the source code of pytorch. From python to the c/c++ implementation, and I saw one quantized layer goes to oneDNN library.
I have thought pytorch use caffe2 as it’s c/c++ implementation, so why there is a oneDNN?bz this library is faster? bz this library support onnx better?
thanks in advance.

2

PyTorch uses different backends to accelerate its workloads such as e.g. cuDNN for NVIDIA GPUs and oneDNN for Intel CPUs.
Caffe2 was merged into the PyTorch backend a while ago to share common backend implementations but is deprecated now.

1 Like
3

Thanks for replying , very clear.