MATRIX OPERATIONS

hyperfold.matmul

From version 0.1.1

Multiply two rank-two matrices.

float16 · bfloat16 · float32 · float64 · complex32 · complex64 · complex128

AVX2 · AVX-512 · ARM NEON · Apple Silicon

CUDA

Signature

hyperfold.matmul(left, right, *, out=None, context=None, copy=None)

Parameters

Returns

array

An array in the input framework.

Example

PythonNumPy

Notes and errors

SHARED ARRAY CONTRACT

Framework, output and copy behaviour

NumPy inputs run on the CPU. CuPy and PyTorch inputs must be CUDA arrays and preserve their framework and device. Inputs from different frameworks, devices or dtypes are rejected.

out= must match the result's framework, dtype, device and exact shape; it must be writable, C-contiguous and must not overlap an input.

copy=None borrows compatible storage and otherwise creates a same-device contiguous input. copy=False forbids boundary copies; copy=True always creates fresh input storage.

float16 and bfloat16 compute in float32; complex32 computes in complex64. Ordinary array results retain their input storage dtype. Numeric scalar reductions, solves and specialised solver results use compute precision. General eigensolvers and LU log-determinants return the matching complex precision. Predicates return booleans.

Invalid Python objects and incompatible dtypes raise TypeError; invalid ranks, shapes, values and output buffers raise ValueError. Numerical factorisation or convergence failures raise numpy.linalg.LinAlgError where stated above. Dispatch, device and native execution failures raise BackendError.

See Arrays and dtypes for packed complex storage, DLPack ordering and the complete precision table.