warp.utils.array_inner#

warp.utils.array_inner(a, b, out=None, count=None, axis=None)[source]#

Compute the inner product of two arrays.

This function computes the dot product between two arrays with the same shape, optionally along a specified axis. When axis is None, it computes a single inner product over the flattened arrays. Partial reductions and inputs with different layouts use logical row-major order. Full reductions of inputs with the same compact layout may use physical storage order.

During CPU graph capture, or CUDA graph capture with apic=True, non-empty calls require an explicit out array so replay can store the current result. Existing whole-array, explicit-count, composite-dtype, and positive-stride axis reductions are recorded. Non-empty whole-array reductions that require contiguous copies are unsupported during API Capture (APIC) recording; make inputs contiguous before capture or specify an axis for strided reductions. Negative strides raise NotImplementedError during APIC recording. Counts and reduction strides must fit signed 32-bit integers, and participating addresses and strides must be aligned to the scalar type.

Parameters:

  • a (array) – First input array.

  • b (array) – Second input array. Must have the same shape and data type as a.

  • out (array | None) – Output array to store results. If None, a new array is created.

  • count (int | None) – Number of elements to process. Must not exceed the input size, or the selected axis length when axis is specified. If None, processes the entire reduction.

  • axis (int | None) – Axis along which to compute inner product. Negative values count from the last dimension. If None, computes a single inner product over the flattened arrays.

Returns:

The inner product result. Returns a float if axis is None and out is None, otherwise returns the out array.

Raises:

  • ValueError – If a and b have different shapes.

  • IndexError – If axis is outside the valid range for a.

  • RuntimeError – If count exceeds the reduction size, array storage devices or data types are incompatible, or an APIC-recorded call uses an unsupported count or memory layout.

  • NotImplementedError – If a non-empty call during APIC recording omits out. Also raised if any input or output array has a negative stride during APIC recording, including for a zero-count call, or a non-empty whole-array reduction requires a contiguous copy during APIC recording.

Return type:

array | float

On this page