Dtype policy#

Each Context stores a backend-normalized dtype. Array construction through ctx.asarray(...) and ctx.assparse(...) uses that dtype.

Dtype resolution is a separate concern from context resolution. Context resolution determines which context will be assigned to a new object. Once that context is fixed, input objects carrying their own contexts may need to be converted to it. During that conversion, the backend is determined by the resolved context, but dtype handling is controlled by the dtype policy.

For example, if an input object has dtype float32 and is converted to a context on another backend, SpaceCore must decide whether to preserve the original dtype or cast the object to the dtype supplied by the resolved context. Conceptually, the dtype policy controls whether conversion prioritizes dtype preservation or dtype unification.

Dtype resolution#

When a constructor infers several compatible contexts, SpaceCore joins their dtypes using NumPy result-type promotion and then sanitizes the result for the selected backend.

import spacecore as sc

X = sc.VectorSpace((3,), ctx=sc.Context(sc.NumpyOps(), dtype="float32"))
Y = sc.VectorSpace((3,), ctx=sc.Context(sc.NumpyOps(), dtype="float64"))

Z = sc.ProductSpace((X, Y))

The product space resolves to one backend family and a dtype capable of representing both inputs.

Preserve or convert#

The dtype_resolution_policy regulates how dtype is chosen when a new context is normalized relative to an existing one.

Use spacecore.set_dtype_resolution_policy(...) to set it and spacecore.get_dtype_resolution_policy() to inspect the active value.

dtype_resolution_policy types

  • convert: when an object is converted to the resolved context, use the dtype that the resolved context provides. This prioritizes unifying values under the target context.

  • keep_native: when an object is converted to another backend, preserve the object’s dtype by converting it to the equivalent dtype in the target backend. This prioritizes preserving the source object’s numerical dtype.

The default is keep_native.

Dtype policy checkpoints

  • Context creation sanitizes dtype through the selected backend.

  • Array construction uses the context dtype.

  • Multi-input constructors join compatible inferred dtypes.

  • Conversion either preserves the native dtype or follows target context resolution, depending on policy.

Use an explicit Context when dtype is part of the numerical contract of an algorithm.

Backend defaults#

Every BackendOps implementation defines its own dtype normalization rule through sanitize_dtype(dtype). That method both normalizes dtype objects into backend-native dtype representations and determines the backend default when dtype=None.

For NumpyOps, sanitize_dtype(None) currently resolves to numpy.float64. For JaxOps, the default depends on JAX configuration: if jax_enable_x64=True the default is float64; otherwise the default is float32.