Why TensorFlow XLA needs many new xla op kernels?

112 Views Asked by At

In TensorFlow code about XLA, I see kernels about many OPs like compiler/tf2xla/kernels/concat_op. It seems like a repetition of core/kernels/concat_op. Why Ops like compiler/tf2xla/kernels/concat_op are needed? And why not just replace it with core/kernels/concat_op to save code?

0

There are 0 best solutions below