Tensor Mapping APIs

Warning

Experimental features:

• May be replaced, updated, or removed at any time.

• Do not require maintaining API/ABI stability of their own additions over time.

• Do not require conformance testing of their own additions.

Motivation

Used to target the CUDA entry points cuTensorMapEncodeIm2col and cuTensorMapEncodeTiled.

For some tensor core operations on sm_90+ Nvidia devices, a tensor descriptor must be built on the host and passed to the kernel. The interfaces mentioned above, and mapped to UR in this extension, provide the APIs necessary to create these tensor descriptor objects, that can then be passed to the kernels.

API

Enums

ur_exp_tensor_map_data_type_flags_t ur_exp_tensor_map_interleave_flags_t ur_exp_tensor_map_l2_promotion_flags_t ur_exp_tensor_map_swizzle_flags_t ur_exp_tensor_map_oob_fill_flags_t

Types

ur_exp_tensor_map_handle_t

Functions

• urTensorMapEncodeIm2ColExp

• urTensorMapEncodeTiledExp

Changelog

Revision	Changes
1.0	Initial Draft

Support

This extension is only supported on the UR_PLATFORM_BACKEND_CUDA backend.

Contributors

• Hugh Delaney hugh.delaney@codeplay.com