Namespace c10

Defines the Struct Half type (half-precision floating-point) including conversions to standard C types and basic arithmetic operations.

Detailed Description

Note that arithmetic operations are implemented by converting to floating point and performing the operation in float32, instead of using CUDA half intrinsics. Most uses of this type within ATen are memory bound, including the element-wise kernels, and the half intrinsics aren’t efficient on all GPUs. If you are writing a compute bound kernel, you can use the CUDA half intrinsics directly on the Struct Half type from device code.



Access comprehensive developer documentation for PyTorch

View Docs


Get in-depth tutorials for beginners and advanced developers

View Tutorials


Find development resources and get your questions answered

View Resources