15.30. __nv_tf32

struct __nv_tf32

This structure implements the datatype for storing Tensor Float 32 numbers. This format uses at least 19 bits to represent its value: 1 sign bit, 8 bits for the exponent, and 10 or more bits for the mantissa. The number of precision bits is unspecified and depends on the hardware target. The storage size of this type is 32 bits in total and it requires a 4-byte alignment. The arrangement of the value bits within the 32 bits of storage is unspecified.

This type is implemented in the cuda_tf32.h header. Currently, the __nv_tf32 type may only be used in tile code and there are no APIs that directly operate on this type. Instead, __nv_tf32 should be used as the element of a tile (e.g. cuda::tiles::tile<__nv_tf32, cuda::tiles::shape<4>>) and passed to CUDA Tile C++ APIs. The supported CUDA Tile C++ APIs include conversions to other numeric types, memory operations, and matrix multiplication. For details on these operations, please see the CUDA Tile C++ API Reference.