Back to Eqiora docs Skip to main content

Module cuda

Module cuda 

Source
Description

Dynamically loaded single-device CUDA adapter.

Structs§

CudaComputeCapability
CUDA compute capability reported by the selected physical device.
CudaCsrActionEvidence
Auditable evidence for one CUDA CSR action accepted against a host reference vector.
CudaCsrActionResult
CUDA values plus complete action evidence.
CudaCsrTransferEvidence
Explicit matrix, input, and output transfer evidence for one CSR action.
CudaLibraryVersions
CUDA driver, bindings, and sparse-library versions used by one action.
CudaLinearSolveEvidence
Auditable device evidence surrounding the backend-neutral solve report.
CudaLinearSolveResult
Independently accepted solution and CUDA-specific execution evidence.
CudaLinearSolver
One CUDA device selected for CSR Krylov execution.
CudaLinearTransferEvidence
Explicit matrix, problem-vector, preconditioner, and result transfers for one CUDA linear solve.
CudaRuntime
Dynamically loaded cudarc runtime adapter.

Constants§

CUDARC_VERSION
Exact dynamically loaded driver-binding dependency version.
CUDA_ADAPTER_VERSION
Exact Eqiora CUDA adapter package version compiled into this binary.
CUDA_BINDING_TOOLKIT
CUDA toolkit ABI selected for the generated driver bindings.
CUDA_LINEAR_EXECUTION
Stable CUDA execution identity, separate from the solver algorithm.
CUDA_LINEAR_EXECUTION_PROVIDER
Exact declared CUDA execution release and binding versions.
CUDA_LINEAR_SOLVER_BACKEND
Stable CUDA Krylov backend identity.
CUDA_LINEAR_SOLVER_PROVIDER
Exact declared CUDA Krylov solver release.
CUDA_RUNTIME_ID
Stable runtime identity even when the optional adapter is not compiled.

Functions§

verify_csr_action
Execute one finalized f64 CSR action on CUDA and compare it with the independent host operator under an explicit tolerance.
verify_csr_action_against
Execute one finalized f64 CSR action on CUDA and compare it with an explicitly supplied independent reference vector.