gmx::detail Namespace Reference

Description

Client software developers do not interact directly with the contents of gmx::detail, but some declarations and definitions are necessary in the public headers, such as forward declarations of implementation classes and definitions of custom deleters.

Classes
struct	PaddingTraits
	Traits classes for handling padding for types used with PaddedVector. More...
struct	SimulationInputHandleImplDeleter
	Explicit deleter details for SimulationInputHolderImpl. More...
struct	extents_analyse< R, E0, StaticExtents...>
	Enable querying extent of specific rank by splitting a static extents off the variadic template arguments. More...
struct	extents_analyse< R, dynamic_extent, StaticExtents...>
	Enable querying extent of specific rank by splitting a dynamic extent off the variadic template arguments. More...
struct	extents_analyse< 0 >
	Specialisation for rank 0 extents analysis. Ends recursive rank analysis. More...

Functions
template<std::size_t alignment>
bool	is_sufficiently_aligned (const void *ptr)
void	mtopFromMolType (gmx_mtop_t *mtop, const gmx_moltype_t &sourceMolType)
	Create a topology from a molecule type. More...
template<typename T >
Index	computePaddedSize (Index numElements)
	Returns the allocation size for PaddedVector that contains numElements elements plus padding for SIMD operations. More...
template<typename T , typename AllocatorType >
void	insertPaddingElements (std::vector< T, AllocatorType > *v, Index newPaddedSize)
	Helper function to insert padding elements for most T.
template<typename T , typename AllocatorType >
void	insertPaddingElements (std::vector< BasicVector< T >, AllocatorType > *v, Index newPaddedSize)
	Specialization of helper function to insert padding elements, used for BasicVector<T>.
static void	coordinateMpiRanks (MPI_Comm communicator, const bool thisRankThrew, const std::exception_ptr exceptionPtr)
	If any rank threw, make sure all throw. More...

Variables
constexpr double	c_largeNegativeExponent = -10000.0
	A value that can be passed to exp() with result 0, also with SIMD.
constexpr double	c_largePositiveExponent = 700.0
	The largest acceptable positive exponent for variables that are passed to exp().
template<typename T >
constexpr bool	is_resizable_v = is_resizable<T>::value
	Type has a resize member function callable with size_t argument.
constexpr int	c_nbnxnGpuClusterSize = 8
	The i- and j-cluster size for GPU lists, 8 atoms for CUDA, set at configure time for OpenCL and SYCL.
static constexpr int	c_nbnxnGpuClusterpairSplit = 2
	The number of sub-parts used for data storage for a GPU cluster pair. More...
constexpr int	c_nBoxZ = 2 * gmx::c_dBoxZ + 1
constexpr int	c_nBoxY = 2 * gmx::c_dBoxY + 1
constexpr int	c_nBoxX = 2 * gmx::c_dBoxX + 1
constexpr int	c_numIvecs = detail::c_nBoxZ * detail::c_nBoxY * detail::c_nBoxX
constexpr int	c_TabulatedNormalDistributionDefaultBits = 14
	Number of bits that determines the resolution of the lookup table for the normal distribution.
constexpr int	c_gpuNumClusterPerCellZ = GMX_GPU_NB_NUM_CLUSTER_PER_CELL_Z
	The number of clusters along a direction in a pair-search grid cell for GPU lists. More...
constexpr int	c_gpuNumClusterPerCellY = GMX_GPU_NB_NUM_CLUSTER_PER_CELL_Y
constexpr int	c_gpuNumClusterPerCellX = GMX_GPU_NB_NUM_CLUSTER_PER_CELL_X

Function Documentation

template<typename T >

Index gmx::detail::computePaddedSize

(

Index

numElements

)

Returns the allocation size for PaddedVector that contains numElements elements plus padding for SIMD operations.

Parameters

[in]

numElements

The number of T elements for which data will be stored.

Returns

The number of T elements that must be allocated (ie >= numElements).

static void gmx::detail::coordinateMpiRanks ( MPI_Comm  communicator, const bool  thisRankThrew, const std::exception_ptr  exceptionPtr  )

static

If any rank threw, make sure all throw.

Any rank that threw an exception will re-throw it. If so, any other rank will throw ParallelConsistencyError. Otherwise, does not throw.

Does global MPI communication on communicator.

Parameters

[in]	communicator	The MPI communicator over which exception-throwing behaviour should be coordinated. Can be MPI_COMM_NULL only for thread-MPI before rank launch.
[in]	thisRankThrew	Describes whether the current MPI rank threw an exception
[in]	exceptionPtr	If the current MPI rank threw an exception, points to that exception, otherwise nullptr.

void gmx::detail::mtopFromMolType	(	gmx_mtop_t *	mtop,
		const gmx_moltype_t &	sourceMolType
	)

Create a topology from a molecule type.

This function is designed for obtaining the AtomRange of the desired molecule type for writing atomic properties and residue information.

Parameters

[out]	mtop	The pointer to the topology to create.
[in]	sourceMolType	The source molecule type to create the topology from.

Variable Documentation

constexpr int gmx::detail::c_gpuNumClusterPerCellZ = GMX_GPU_NB_NUM_CLUSTER_PER_CELL_Z

The number of clusters along a direction in a pair-search grid cell for GPU lists.

Typically all 2, but X can be 1 when targeting Intel Ponte Vecchio

constexpr int gmx::detail::c_nbnxnGpuClusterpairSplit = 2

static

The number of sub-parts used for data storage for a GPU cluster pair.

In CUDA the number of threads in a warp is 32 and we have cluster pairs of 8*8=64 atoms, so it's convenient to store data for cluster pair halves, i.e. split in 2.

On architectures with 64-wide execution however it is better to avoid splitting (e.g. AMD GCN, CDNA and later).