Right now, some parts of the code use ATen atomic functions from headers like <THC/THCAtomics.cuh> and <ATen/native/cuda/KernelUtils.cuh>. PyTorch is planning to migrate to <ATen/cuda/Atomics.cuh> as the unified header for this in the future, so this is a somewhat breaking change.
We should plan on writing a macro to deal with this.
Right now, some parts of the code use ATen atomic functions from headers like
<THC/THCAtomics.cuh>and<ATen/native/cuda/KernelUtils.cuh>. PyTorch is planning to migrate to<ATen/cuda/Atomics.cuh>as the unified header for this in the future, so this is a somewhat breaking change.We should plan on writing a macro to deal with this.