xref: /aosp_15_r20/external/pytorch/aten/src/ATen/LegacyBatchedFallback.h (revision da0073e96a02ea20f0ac840b70461e3646d07c45)
1 #pragma once
2 #include <ATen/ATen.h>
3 #include <ATen/core/op_registration/op_registration.h>
4 #include <torch/library.h>
5 
6 namespace at {
7 
8 // If an operator doesn't have a batching rule implemented then we fallback
9 // to this implementation. The fallback only works on out-of-place operators
10 // that return only tensors with new memory. (e.g., no in-place operators, no
11 // view operations).
12 //
13 // The fallback effectively takes all of the BatchedTensors in `stack`, slices
14 // them, and runs `op` on all of the corresponding slices to produce slices
15 // of the outputs. The output slices then get `torch.stack`ed to create the
16 // final returns.
17 //
18 // The performance of the fallback is not very good because it introduces an
19 // extra copy from stacking the sliced outputs. Because of this, we prefer to
20 // write batching rules for operators whenever possible.
21 void batchedTensorForLoopFallback(
22     const c10::OperatorHandle& op,
23     torch::jit::Stack* stack);
24 
25 } // namespace at
26