Doxygen 1.9.8
Toolkit for Adaptive Stochastic Modeling and Non-Intrusive ApproximatioN: Tasmanian v8.3
 
Loading...
Searching...
No Matches

Wrapper class around calls GPU accelerated linear algebra libraries. More...

#include <tsgAcceleratedDataStructures.hpp>

Public Member Functions

 GpuEngine ()=default
 Default construction, does not use any GPU acceleration, will set handles on-the-fly.
 
 GpuEngine (GpuEngine const &)=delete
 Do not copy, even if no GPU backend is enabled.
 
GpuEngine & operator= (GpuEngine const &)=delete
 Do not copy, even if no GPU backend is enabled.
 
 GpuEngine (GpuEngine &&)=default
 Moving the engine is allowed, unique pointers will be moved.
 
GpuEngine & operator= (GpuEngine &&)=default
 Moving the engine is allowed, unique pointers will be moved.
 
void setCuBlasHandle (void *handle)
 Manually sets the cuBlas handle, handle must be a valid cublasHandle_t associated with this CUDA device.
 
void setCuSparseHandle (void *handle)
 Manually sets the cuSparse handle, handle must be a valid cusparseHandle_t associated with this CUDA device.
 
void setCuSolverDnHandle (void *handle)
 Manually sets the cuSparse handle, handle must be a valid cusolverDnHandle_t associated with this CUDA device.
 

Public Attributes

std::unique_ptr< int, HandleDeleter< AccHandle::Cublas > > cublas_handle
 Holds the cuBlas handle.
 
std::unique_ptr< int, HandleDeleter< AccHandle::Cusparse > > cusparse_handle
 Holds the cuSparse handle.
 
std::unique_ptr< int, HandleDeleter< AccHandle::Cusolver > > cusolver_handle
 Holds the cuSolver handle.
 
bool called_magma_init = false
 Avoids an empty engine when no acceleration is enabled, allows for default constructor/move/copy, skips extraneous calls to MAGMA init.
 

Detailed Description

Wrapper class around calls GPU accelerated linear algebra libraries.

The class also manages the required handles and queues and holds the context of the active GPU device.


The documentation for this struct was generated from the following file: