Skip to content

Kokkos::DynRankView

jeffmiles63 edited this page Jun 1, 2020 · 5 revisions

Kokkos::DynRankView

Header File: Kokkos_DynRankView.hpp

Usage:

DynRankView is a potentially reference counted multi dimensional array with compile time layouts and memory space. Its semantics are similar to that of std::shared_ptr. The DynRankView differs from the View in that its rank is not provided with the DataType template parameter; it is determined dynamically based on the number of extent arguments passed to the constructor. The rank has an upper bound of 7 dimensions.

Interface

template <class DataType [, class LayoutType] [, class MemorySpace] [, class MemoryTraits]>
class DynRankView;

Parameters

Template parameters other than DataType are optional, but ordering is enforced. That means for example that LayoutType can be omitted but if both MemorySpace and MemoryTraits are specified, MemorySpace must come before MemoryTraits.

  • DataType: Defines the fundamental scalar type of the DynRankView. The basic structure is ScalarType. Examples:
    • double: a DynRankView of double, dimensions are passed as arguments to the constructor, the number of which determine the rank.
  • LayoutType: Determines the mapping of indices into the underlying 1D memory storage. Custom Layouts can be implemented, but Kokkos comes with some built-in ones:
    • LayoutRight: Strides increase from the right most to the left most dimension. The last dimension has a stride of one. This corresponds to how C multi dimensional arrays ([][][]) are laid out in memory.
    • LayoutLeft: Strides increase from the left most to the right most dimension. The first dimension has a stride of one. This is the layout Fortran uses for its arrays.
    • LayoutStride: Strides can be arbitrary for each dimension.
  • MemorySpace: Controls the storage location. If omitted, the default memory space of the default execution space is used (i.e. Kokkos::DefaultExecutionSpace::memory_space)
  • [[MemoryTraits|Kokkos::MemoryTraits]]: Sets access properties via enum parameters for the templated Kokkos::MemoryTraits<> class. Enums can be bit combined. Possible values:
    • Unmanaged: The DynRankView will not be reference counted. The allocation has to be provided to the constructor.
    • Atomic: All accesses to the view will use atomic operations.
    • RandomAccess: Hint that the view is used in a random access manner. If the view is also const, this will trigger special load operations on GPUs (i.e. texture fetches).
    • Restrict: There is no aliasing of the view by other data structures in the current scope.

Requirements:

Public Class Members

Enums

  • rank: rank of the view (i.e. the dimensionality).
  • rank_dynamic: number of runtime determined dimensions.
  • reference_type_is_lvalue_reference: whether the reference type is a C++ lvalue reference.

Typedefs

Data Types

  • data_type: the DataType of the DynRankView.

  • const_data_type: const version of DataType, same as data_type if that is already const.

  • non_const_data_type: non-const version of DataType, same as data_type if that is already non-const.

  • scalar_array_type: If DataType represents some properly specialised array data type such as Sacado FAD types, scalar_array_type is the underlying fundamental scalar type.

  • const_scalar_array_type: const version of scalar_array_type, same as scalar_array_type if that is already const

  • non_const_scalar_array_type: non-Const version of scalar_array_type, same as scalar_array_type if that is already non-const.

Scalar Types

  • value_type: the data_type stripped of its array specifiers, i.e. the scalar type of the data the view is referencing (e.g. if data_type is const int*******, value_type is const int.
  • const_value_type: const version of value_type.
  • non_const_value_type: non-const version of value_type.

Spaces

  • execution_space: execution space associated with the view, will be used for performing view initialization, and certain deep_copy operations.
  • memory_space: data storage location type.
  • device_type: the compound type defined by Device<execution_space,memory_space>
  • memory_traits: the memory traits of the view.
  • host_mirror_space: host accessible memory space used in HostMirror.

ViewTypes

  • non_const_type: this view type with all template parameters explicitly defined.
  • const_type: this view type with all template parameters explicitly defined using a const data type.
  • HostMirror: compatible view type with the same DataType and LayoutType stored in host accessible memory space.

Data Handles

  • reference_type: return type of the view access operators.
  • pointer_type: pointer to scalar type.

Other

  • array_layout: the layout of the DynRankView.
  • size_type: index type associated with the memory space of this view.
  • dimension: an integer array like type, able to represent the extents of the view.
  • specialize: a specialization tag used for partial specialization of the mapping construct underlying a Kokkos DynRankView.

Constructors

  • DynRankView(): default constructor. No allocations are made, no reference counting happens. All extents are zero and its data pointer is nullptr and its rank is set to 0.
  • DynRankView( const DynRankView<DT, Prop...>& rhs): Copy constructor with compatible DynRankViews. Follows DynRankView assignment rules.
  • DynRankView( DynRankView&& rhs): move constructor
  • DynRankView( const View<RT,RP...> & rhs ): copy constructor taking View as imput.
  • DynRankView( const std::string& name, const IntType& ... indices): standard allocating constructor.
    • name: a user provided label, which is used for profiling and debugging purposes. Names are not required to be unique,
    • indices: runtime dimensions of the view.
    • Requires: array_layout::is_regular == true.
  • DynRankView( const std::string& name, const array_layout& layout): standard allocating constructor.
    • name: a user provided label, which is used for profiling and debugging purposes. Names are not required to be unique,
    • layout: an instance of a layout class.
  • DynRankView( const AllocProperties& prop, , const IntType& ... indices): allocating constructor with allocation properties.
    • An allocation properties object is returned by the view_alloc function.
    • indices: runtime dimensions of the view.
    • Requires: array_layout::is_regular == true.
  • DynRankView( const AllocProperties& prop, const array_layout& layout): allocating constructor with allocation properties and a layout object.
    • layout: an instance of a layout class.
  • DynRankView( const pointer_type& ptr, const IntType& ... indices): unmanaged data wrapping constructor.
    • ptr: pointer to a user provided memory allocation. Must provide storage of size DynRankView::required_allocation_size(n0,...,nR)
    • indices: runtime dimensions of the view.
    • Requires: array_layout::is_regular == true.
  • DynRankView( const std::string& name, const array_layout& layout): unmanaged data wrapper constructor.
    • ptr: pointer to a user provided memory allocation. Must provide storage of size DynRankView::required_allocation_size(layout) (NEEDS TO BE IMPLEMENTED)
    • layout: an instance of a layout class.
  • DynRankView( const ScratchSpace& space, const IntType& ... indices): constructor which acquires memory from a Scratch Memory handle.
    • space: scratch memory handle. Typically returned from team_handles in TeamPolicy kernels.
    • indices: runtime dimensions of the view.
    • Requires: sizeof(IntType...)==rank_dynamic()
    • Requires: array_layout::is_regular == true.
  • DynRankView( const ScratchSpace& space, const array_layout& layout): constructor which acquires memory from a Scratch Memory handle.
    • space: scratch memory handle. Typically returned from team_handles in TeamPolicy kernels.
    • layout: an instance of a layout class.
  • DynRankView( const DynRankView<DT, Prop...>& rhs, Args ... args): subview constructor. See subview function for arguments.

Data Access Functions

  • reference_type operator() (const IntType& ... indices) const

    Returns a value of reference_type which may or not be referenceable itself. The number of index arguments must match the rank of the view. See notes on reference_type for properties of the return type.

  • reference_type access (const IntType& i0=0, ... , const IntType& i6=0) const

    Returns a value of reference_type which may or not be referenceable itself. The number of index arguments must be equal or larger than the rank of the view. Index arguments beyond rank must be 0, which will be enforced if KOKKOS_DEBUG is defined. See notes on reference_type for properties of the return type.

Data Layout, Dimensions, Strides

  • constexpr array_layout layout() const
    Returns the layout object. Can be used to to construct other views with the same dimensions.
  • template<class iType>
    constexpr size_t extent( const iType& dim) const
    Return the extent of the specified dimension. iType must be an integral type, and dim must be smaller than rank.
  • template<class iType>
    constexpr int extent_int( const iType& dim) const
    Return the extent of the specified dimension as an int. iType must be an integral type, and dim must be smaller than rank. Compared to extent this function can be useful on architectures where int operations are more efficient than size_t. It also may eliminate the need for type casts in applications which otherwise perform all index operations with int.
  • template<class iType>
    constexpr size_t stride(const iType& dim) const
    Return the stride of the specified dimension. iType must be an integral type, and dim must be smaller than rank. Example: a.stride(3) == (&a(i0,i1,i2,i3+1,i4)-&a(i0,i1,i2,i3,i4))
  • constexpr size_t stride_0() const
    Return the stride of dimension 0.
  • constexpr size_t stride_1() const
    Return the stride of dimension 1.
  • constexpr size_t stride_2() const
    Return the stride of dimension 2.
  • constexpr size_t stride_3() const
    Return the stride of dimension 3.
  • constexpr size_t stride_4() const
    Return the stride of dimension 4.
  • constexpr size_t stride_5() const
    Return the stride of dimension 5.
  • constexpr size_t stride_6() const
    Return the stride of dimension 6.
  • constexpr size_t stride_7() const
    Return the stride of dimension 7.
  • constexpr size_t span() const
    Returns the memory span in elements between the element with the lowest and the highest address. This can be larger than the product of extents due to padding, and or non-contiguous data layout as for example LayoutStride allows.
  • constexpr pointer_type data() const
    Return the pointer to the underlying data allocation.
  • bool span_is_contiguous() const
    Whether the span is contiguous (i.e. whether every memory location between in span belongs to the index space covered by the view).
  • static constexpr size_t required_allocation_size(size_t N0 = 0, ..., size_t N8 = 0);
    Returns the number of bytes necessary for an unmanaged view of the provided dimensions. This function is only valid if array_layout::is_regular == true.
  • static constexpr size_t required_allocation_size(const array_layout& layout);
    Returns the number of bytes necessary for an unmanaged view of the provided layout.

Other

  • int use_count() const

    Returns the current reference count of the underlying allocation.

  • const char* label() const

    Returns the label of the DynRankView.

  • constexpr unsigned rank() const

    Returns the dynamic rank of the DynRankView

  • constexpr bool is_allocated() const;

    Returns true if the view points to a valid memory location. This function works for both managed and unmanaged views. With the unmanaged view, there is no guarantee that referenced address is valid, only that it is a non-null pointer.

Assignment Rules

Assignment rules cover the assignment operator as well as copy constructors. We aim at making all logically legal assignments possible, while intercepting illegal assignments if possible at compile time, otherwise at runtime. In the following, we use DstType and SrcType as the type of the destination view and source view respectively. dst_view and src_view refer to the runtime instances of the destination and source views, i.e.:

ScrType src_view(...);
DstType dst_view(src_view);
dst_view = src_view;

The following conditions must be met at and are evaluated at compile time:

  • DstType::rank == SrcType::rank
  • DstType::non_const_value_type is the same as SrcType::non_const_value_type
  • If std::is_const<SrcType::value_type>::value == true than std::is_const<DstType::value_type>::value == true.
  • MemorySpaceAccess<DstType::memory_space,SrcType::memory_space>::assignable == true

Furthermore there are rules which must be met if DstType::array_layout is not the same as SrcType::array_layout. These rules only cover cases where both layouts are one of LayoutLeft, LayoutRight or LayoutStride

  • If neither DstType::array_layout nor SrcType::array_layout is LayoutStride:
    • If DstType::rank > 1 than DstType::array_layout must be the same as SrcType::array_layout.
  • If either DstType::array_layout or SrcType::array_layout is LayoutStride
    • For each dimension k it must hold that dst_view.extent(k) == src_view.extent(k)

Examples

#include<Kokkos_Core.hpp>
#include<cstdio> 

int main(int argc, char* argv[]) {
   Kokkos::initialize(argc,argv);

   int N0 = atoi(argv[1]);
   int N1 = atoi(argv[2]);

   Kokkos::DynRankView<double> a("A",N0);
   Kokkos::DynRankView<double> b("B",N1);

   Kokkos::parallel_for("InitA", N0, KOKKOS_LAMBDA (const int& i) {
     a(i) = i;
   });

   Kokkos::parallel_for("InitB", N1, KOKKOS_LAMBDA (const int& i) {
     b(i) = i;
   });

   Kokkos::DynRankView<double,Kokkos::LayoutLeft> c("C",N0,N1);
   {
     Kokkos::DynRankView<const double> const_a(a);
     Kokkos::DynRankView<const double> const_b(b);
     Kokkos::parallel_for("SetC", Kokkos::MDRangePolicy<Kokkos::Rank<2,Kokkos::Iterate::Left>>({0,0},{N0,N1}),
       KOKKOS_LAMBDA (const int& i0, const int& i1) {
       c(i0,i1) = a(i0) * b(i1);
     });
   }

   Kokkos::finalize();
}
Clone this wiki locally