1. 1. 1. 🚀 The feature, motivation and pitch

1. 1. 1. Feature
         Teach TorchInductor to generate native ldexp code for CUDA and CPU backends instead of relying on the decomposition x * pow(2, n).

1. 1. 1. Motivation

Currently, `torch.ldexp` is decomposed into:
```python
x * torch.pow(2.0, n)
```

This decomposition is suboptimal because:

• It currently requires computing `pow(2, n)` followed by a multiplication, while native `ldexp` implementations are significantly faster as they operate directly on the floating-point exponent bits

1. 1. 1. Proposed Solution

Add Inductor codegen support for `aten.ldexp.Tensor` to emit native calls:

• *CUDA*: Use `__nv_ldexp` from libdevice
• *CPU*: Use `std::ldexp` from `<cmath>`

1. 1. 1. Implementation Notes

• libdevice provides `_nv_ldexp(double, int)` and `_nv_ldexpf(float, int)`
• Similar patterns exist for other math ops in `torch/_inductor/codegen/`
• May need to handle the non-integer exponent input type differently

1. 1. 1. cc
         @ngimel

cc @chauhang @penguinwu