I came across the concept of "deep learning primitives" from the Nvidia talk Jetson AGX Xavier New Era Autonomous Machines (on slide 44).
There doesn't seem to be a lot of articles in the community on this concept. I was able to find one definition from here, where it defined deep learning primitives as the "fundamental building blocks of deep networks" like fully connected layers, convolutions layers, etc.
I was curious to find out if a self-attention layer is a primitive, I came across this OpenDNN issue and one person explained that self-attention layers can be built by other primitives like inner product, concat, etc.
So my question is what exactly are primitives in deep learning? What makes a convolution layer a primitive and a self-attention layer not a primitive?