This directory contains generic CUDA runtime implementations of linalg operation tags. CudaReferenceBackend follows provider backends such as cuBLAS in CudaStorage's ordered selector and handles operations that do not require a provider library.

The copy_op implementation keeps runtime-copy fast paths for compatible raw contiguous mappings. Same-device positive-strided mappings whose compact rank is at most eight, including padding and differing physical order, use the reference backend's typed elementwise CUDA kernel. elementwise_plan.hpp lowers the same backend-neutral affine plan used by the CPU executor to a device POD, preferring 32-bit logical indices and offsets when every reachable offset fits and using a 64-bit payload otherwise. The compiled executor is arity-generic and separates overwrite operations, which do not read the output, from update operations, which do.

Copy is the first one-input overwrite instantiation. conjugate_inplace_op is the first zero-input update instantiation: it reads and writes each output element through one exclusive CUDA access. Real and integer conjugation succeeds without acquiring a stream or launching because it is semantically the identity. Complex conjugation supports persistent uni20::cfloat and uni20::cdouble storage through CUDA execution proxies.

The CUDA reference backend explicitly registers stateless unary negate, square, and reciprocal transforms; stateful scale<Factor>; and stateless binary add, subtract, multiply, and divide transforms. Binary operations carry two independently strided input operands. Supported real and complex combinations are compiled into the library. Other callable types are not assumed to be available inside the precompiled library and therefore remain ineligible for this backend.

The kernels publish stream completion through the same buffer ledgers as the runtime copy path. Nonpositive strides on active axes cleanly decline for nontrivial elementwise work; supporting a future negative-stride Uni20 layout requires descriptor rebasing and signed traversal rather than raw runtime copying. Distinct-offset views into one CUDA buffer use a single exclusive access and rely on the C++ copy precondition that the operands do not destructively overlap. Nontrivial same-offset transformations decline.

Related Documentation