Skip to contents

tengen

Provides S3 generics for common tensor operations.

Installation 

pak::pak("r-xla/tengen")

Available Generics

  • shape(): Different from dim() by also supporting 0-dimensional tensors.
  • dtype(): Returns the data type of the tensor.
  • as_array(): Converts the tensor to an R array.
  • as_raw(): Converts the tensor to a raw() vector.
  • device(): Returns the device of the tensor.

Other functions:

  • naxes(): Returns the number of axes of the tensor. Is implemented as length(shape(x)).
  • nelts(): Returns the number of elements of the tensor. Is implemented as prod(shape(x)).

Tensor Data Types

tengen provides DataType, an enum-style S3 class representing a tensor element type. There is one singleton object per supported dtype, wrapping its canonical string. The members are:

  • boolean: bool (aliases i1, pred)
  • signed integers: i8, i16, i32, i64
  • unsigned integers: ui8, ui16, ui32, ui64
  • floats: f16, bf16, f32, f64
  • complex: c64, c128

Construct and inspect data types with:

  • as_dtype(x): Convert a string (e.g. "f32") or DataType to a DataType object.
  • is_dtype(x): Check whether an object is a DataType.
  • assert_dtype(x): Assert that an object is a DataType.
  • dtype_width(x): The width of one element in bits.
  • dtype_category(x): The category, one of "bool", "int", "uint", "float", or "complex".
  • is_dtype_float(x), is_dtype_int(x), is_dtype_uint(x), is_dtype_bool(x), is_dtype_complex(x): Per-category predicates.

Data types support equality comparison with == and !=.