TasGrid::AccelerationContext Struct Reference

Wrapper class around GPU device ID, acceleration type and GpuEngine. More...

#include <tsgAcceleratedDataStructures.hpp>

Public Types
enum	AlgorithmPreference { algorithm_dense , algorithm_sparse , algorithm_autoselect }
	Defines the sparse-dense algorithm flavors, whenever applicable. More...
enum	ChangeType {   change_none , change_gpu_device , change_gpu_enabled , change_cpu_blas ,   change_sparse_dense }
	Defines the types of acceleration context updates so they can be linked to acceleration cache updates. More...

Public Member Functions
	AccelerationContext ()
	Creates a default context, the device id is set to 0 and acceleration is BLAS (if available) or none.
ChangeType	favorSparse (bool favor)
	Sets algorithm affinity in the direction of sparse.
bool	blasCompatible () const
	Returns true if BLAS is enabled and the current mode is not none.
bool	useKernels () const
	Returns true if the current mode implies the use of custom GPU kernels.
ChangeType	testEnable (TypeAcceleration acc, int new_gpu_id) const
	Returns the ChangeType if enable() is called, but does not change the acceleration.
void	enable (TypeAcceleration acc, int new_gpu_id)
	Accepts parameters directly from TasmanianSparseGrid::enableAcceleration()
void	setDevice () const
	Set default device.
	operator GpuEngine * () const
	Custom convert to GpuEngine.
bool	on_gpu () const
	Returns true if any of the GPU-based acceleration modes have been enabled.

Static Public Member Functions
static constexpr TypeAcceleration	getDefaultAccMode ()
	Returns the default acceleration mode, cpu_blas if BLAS is enabled and none otherwise.
static constexpr int	getDefaultAccDevice ()
	Returns the default acceleration device, CUDA/HIP use GPU 0, SYCL uses -1 which uses sycl::default_selector_v.

Public Attributes
TypeAcceleration	mode
	The current active acceleration mode.
AlgorithmPreference	algorithm_select
	The preference to use dense or sparse algorithms.
int	device
	If using a GPU acceleration mode, holds the active device.
std::unique_ptr< GpuEngine >	engine
	Holds the context to the GPU TPL handles, e.g., MAGMA queue.

Detailed Description

Wrapper class around GPU device ID, acceleration type and GpuEngine.

Single acceleration context held by the TasmanianSparseGrid class and aliased into each of the sub-classes. The context is modified through the main class and is persistent on move, copy and read. The sub-classes hold only a const pointer and can read and use the associated variables and the engine.

Member Enumeration Documentation

AlgorithmPreference

enum TasGrid::AccelerationContext::AlgorithmPreference

Defines the sparse-dense algorithm flavors, whenever applicable.

Enumerator
algorithm_dense	Use dense algorithm.
algorithm_sparse	Use sparse algorithm.
algorithm_autoselect	Use automatically select based on heuristics.

ChangeType

enum TasGrid::AccelerationContext::ChangeType

Defines the types of acceleration context updates so they can be linked to acceleration cache updates.

The update methods will generate a change event that may require an update of the acceleration cache. For example, switching to a different GPU device means that the cache from the old device must be cleared.

Enumerator
change_none	No change, do nothing.
change_gpu_device	Change the associated GPU device.
change_gpu_enabled	Change from BLAS or none to a GPU acceleration mode.
change_cpu_blas	Change BLAS to none or none to BLAS.
change_sparse_dense	Change the sparse-dense AlgorithmPreference.

---

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

• SparseGrids/tsgAcceleratedDataStructures.hpp