Now that we have "embedded" benchmarks which take torch::Tensor arguments, this gives us a convenient way to write benchmarks which use ATen C++ functions. So there is no need for the separate "ATen" benchmark in conftest.py. The new method requires much less boilerplate.
Now that we have "embedded" benchmarks which take
torch::Tensor
arguments, this gives us a convenient way to write benchmarks which use ATen C++ functions. So there is no need for the separate "ATen" benchmark inconftest.py
. The new method requires much less boilerplate.