Module ml::element

tensor element types.

A DenseTensor<E> (tensor.h) stores its values as E and computes in float32: an op loads its operands to float32, runs the float32 kernel and stores the result as E, and every gradient is float32. E is 'float' itself for FloatTensor, whose values are already float32 (no conversion, no copy), or one of the 16-bit storage types below for HalfTensor and BF16Tensor. A storage type provides two static conversions, which generic code reaches through the type parameter ('E::load_n(...)'):

load_n   n stored values -> float32
store_n  n float32 values -> stored values, rounded to nearest even

Both use float16.h's bulk conversions (hardware where the CPU has it).