src/uni20/linalg/backends/direct_gemm

This directory contains the operation-specific backend that lowers a tensor contraction to one rank-two GEMM dispatch.

DirectGemmContractionBackend retains an execution backend selector. It plans M/N/K grouping over mdspec metadata, projects each fixed operand without acquiring a data handle, and dispatches gemm_op through that retained selector. The selected GEMM backend therefore owns host or CUDA access acquisition and provider execution.

The backend cleanly declines before nested dispatch when the contraction does not admit one direct rank-two projection. A nested GEMM decline is also clean; once its single GEMM succeeds, the contraction is complete.

Related Documentation

• Linalg backend source map
• Tensor contraction
• Kernel dispatch