Module cutensor 

Local sys-level wrappers around cudarc::cutensor::sys for the cuTENSOR entry points the safe cudarc::cutensor::result module does not expose (Reduce/ElementwiseBinary/ElementwiseTrinary create+execute, Permutation create+execute, predefined compute descriptors).

Every function here is unsafe and takes the raw cuTENSOR enum types from cudarc::cutensor::sys. The entire crate’s actor layer drives these through kernel/tensor/.

All wrappers convert cutensorStatus_t into a thin CutensorError that mirrors what cudarc::cutensor::result emits; callers that already use cudarc::cutensor::result can interleave both freely.

Structs

CutensorError

Error wrapper around a cutensorStatus_t. Mirrors cudarc::cutensor::result::CutensorError so error messages are consistent across the safe/sys boundary.

Functions

c_32f

create_elementwise_binary^⚠

Create a binary elementwise operation descriptor. Wraps cutensorCreateElementwiseBinary.

create_elementwise_trinary^⚠

Create a trinary elementwise operation descriptor. Wraps cutensorCreateElementwiseTrinary.

create_permutation^⚠

Create a permutation operation descriptor. Wraps cutensorCreatePermutation.

elementwise_binary_execute^⚠

Execute a previously-planned binary elementwise op.

elementwise_trinary_execute^⚠

Execute a previously-planned trinary elementwise op.

permute^⚠

Execute a previously-planned permutation.

plan_preference_set_algo^⚠

Set the pinned algorithm on a plan-preference object. Used by the contraction autotune to probe a specific cutensorAlgo_t value.

r_32f

r_64f

r_min_16bf

r_min_16f

r_min_32f

r_min_64f

r_min_tf32

reduce^⚠

Wraps cutensorReduce (post-plan execution).