src/uni20/linalg/backends/cusolver

This directory implements the blocking exact SVD operation-tag adapter in svd.hpp, using cuSOLVER gesvd on CUDA-buffer descriptors. The current slice supports real float and double, nonempty matrices with rows >= columns, column-major provider-compatible storage, and separate factor outputs. Unsupported shapes, layouts, or overwrite modes are declined before execution.

Execution acquires buffer access on the cuSOLVER stream and synchronizes before returning, including checking the device convergence status. This is a blocking adapter, not a coroutine-native provider task. See the cuSOLVER architecture for the execution and ownership model.

Notes

• Gate operations through kernel_accepts_types and try_kernel before committing to a cuSOLVER path.
• Keep CUDA library ABI details in the cuSOLVER provider layer.

Related Documentation

• Linalg backend source map
• CUDA kernel dispatch and provider scheduling
• cuSOLVER architecture
• CUDA runtime resolution