Is there any way to use TF32 with Thrust? E.g. thrust::device_vector<tf32> or something similar?
TF32 is mostly associated with Tensor Core usage. AFAIK thrust has no mechanisms currently that use tensorcore. With respect to the data type, in terms of storage format there is no difference between TF32 and float. A thrust device vector of float is equivalent and in some sense indistinguishable in memory from a vector of TF32. If you really want to write contorted code, it is probably possible to write a thrust functor that issues tensor core ops, but this is not something I expect people are asking about. And with respect to the format of storage, there wouldn’t be any difference between that and float storage in terms of storage size, order, or meaning of bit positions.