src/uni20/backend/cusolver

This directory owns cuSOLVER provider state and checked error handling. The Tensor-facing exact SVD lowering lives in the linalg backend layer.

Contents

• cusolver_error.hpp/.cpp: checked cuSOLVER status handling.
• execution.hpp/.cpp: device-local handle pools and stream-paired execution leases.
• CMakeLists.txt: cuSOLVER backend target setup.

Notes

• UNI20_BACKEND_CUSOLVER is an optional CUDA-dependent build feature.
• Each DeviceResources lazily owns one cuSOLVER execution pool. The pool defaults to two exclusive handles, capped by the CUDA stream count.
• cusolver::execution_pool(resources, count) configures another count before first provider use. Later explicit configuration must match, and the count cannot exceed stream capacity.
• A leased handle is rebound to the operation stream and returned through a callback at that stream's completion boundary.
• The first operation is blocking real float/double exact SVD through cusolverDnSgesvd/cusolverDnDgesvd for supported column-major CUDA-buffer descriptors.
• Higher-level entry points belong in the linalg backend layer.

Related Documentation

• Backend source layer
• cuSOLVER architecture
• CUDA runtime resolution