This directory contains the BLAS-compatible dense linalg adapter layer. It sits above the raw BLAS and LAPACK provider facades, and below tensor-level linalg front ends.

blas_matrix.hpp: provider-ready readable and writable BLAS matrix operands plus backend-independent transform algebra.
blas_vector.hpp: provider-ready readable and writable BLAS vector operands.
mdspan_access.hpp: shared direct-accessor eligibility for ranked BLAS mdspan operands.
mdspan_matrix.hpp: mdspan-axis staging descriptor construction and lowering to provider-ready operands.
mdspan_vector.hpp: rank-one stride and accessor staging for BLAS increments.
gemm.hpp: direct no-copy GEMM wrappers over the configured BLAS provider.
gemv.hpp: direct no-copy GEMV wrappers over matrix and vector operands.
blas.hpp: include point for this adapter layer.

Notes

Keep raw provider ABI details in the BLAS and LAPACK source layers.
Keep tensor allocation, symmetry policy, and high-level fallback policy above this layer.
try_* direct wrappers decline layouts and ABI dimensions that cannot be represented without copies. They also decline complex conjugate-only GEMM matrix operands and conjugating GEMV input vectors because the generic provider path is currently gated to the portable N/T/C matrix transpose opcodes and GEMV has no vector-conjugation opcode. The transform helper can spell conjugate-only as R; OpenBLAS exposes this through CblasConjNoTrans and its develop-branch Fortran GEMM dispatcher, while MKL Fortran GEMM rejected R in local testing. Checked wrappers treat inconsistent operation dimensions or unsupported direct transforms as logic errors and abort through Uni20 checks. Prepared wrappers that allocate temporaries, or explicit provider extensions such as OpenBLAS conjugate-no-transpose, should make that contract explicit in their names and documentation. Prepared wrappers may also use internal output-storage conjugation as a non-user-visible workaround for unsupported readable transform combinations.
Direct mdspan wrappers must not treat pointer-like data_handle_type as enough for raw BLAS access. Writable outputs require default-accessor views. Readable inputs require default accessors or explicitly lowerable accessors, currently Uni20's conjugated_accessor over a default accessor.
GEMV explicitly scales its output when alpha == 0 or the logical input is empty. This preserves y = alpha*A*x + beta*y semantics without depending on provider-specific quick-return behavior, and beta == 0 does not read the previous output values.
Direct provider GEMM returns success without staging operands when the output is empty, and cleanly declines a zero inner extent before staging. The host backend list falls through from BlasBackend to CpuReferenceBackend, which implements the complete C = alpha*A*B + beta*C semantics for that case.
Direct rank-one lowering accepts positive mdspan strides and canonicalizes an unobserved empty or singleton stride to increment 1. A negative stride on a multi-element view declines: the Fortran BLAS negative-increment convention adjusts its starting element and cannot use the mdspan logical-origin handle unchanged.
Direct rank-two lowering likewise ignores an empty or singleton axis's reported stride. A singleton axis may supply the provider unit-stride axis, while the other axis's observed stride becomes the leading dimension.

Contents

Notes

Related Documentation