-
Notifications
You must be signed in to change notification settings - Fork 0
View
After reading this chapter, you should understand the following:
- A Kokkos View is an array of zero or more dimensions
- How to use View's first template parameter to specify the type of entries, the number of dimensions, and whether the dimensions are determined at run time or compile time
- Kokkos handles array deallocation automatically
- Kokkos chooses array layout at compile time for best overall performance as a function of the computer architecture
- How to set optional template parameters of View for low-level control of execution space, layout, and how Kokkos accesses array entries
In all code examples in this chapter, we assume that all classes in the Kokkos namespace have been imported into the working namespace.
Many scientific and engineering codes spend a lot of time computing with arrays of data and programmers invest a lot of effort making these array computations as fast as possible. This effort is often intimately bound to details of the computer architecture, run-time environment, language, and programming model. For example, optimal array layout may differ based on the architecture, with a large integer factor penalty if wrong. Low-level issues like pointer alignment, array layout, indexing overhead, and initialization all affect performance. This is true even for sequential codes but thread parallelism adds even more pitfalls, like first-touch allocation and false sharing.
For best performance, coders need to tie details of how they manage arrays to details of how parallel code accesses and manages those arrays. Programmers who write architecture-specific code then need to mix low-level features of the architecture and programming model into high-level algorithms. This makes it hard to port codes between architectures, or to write a code that still performs well as architectures evolve.
Kokkos aims to relieve some of this burden by optimizing array management and access for the specific architecture. Tying arrays to shared-memory parallelism lets Kokkos optimize the former to the latter. For example, Kokkos can easily do first-touch allocation because it controls threads that it can use to initialize arrays. Kokkos' architecture-awareness lets it pick optimal layout and pad allocations for good alignment. Expert coders can also use Kokkos to access low-level or more architecture-specific optimizations in a more user-friendly way. For instance, Kokkos makes it easy to experiment with different array layouts.
A View is an array of zero or more dimensions. Programmers set both the type of entries and the number of dimensions at compile time as part of the type of the View. For example, the following specifies and allocates a View with four dimensions for entries that have type int:
const size_t N0 = ...;
const size_t N1 = ...;
const size_t N2 = ...;
const size_t N3 = ...;
Kokkos::View<int****> a ("some label", N0, N1, N2, N3);The string argument is a label which Kokkos uses for debugging. Different Views may have the same label. The ellipses indicate some integer dimensions specified at run time. Users may also set some dimensions at compile time. For example, the following View has two dimensions where the first (represented by the asterisk) is a run-time dimension and the second (represented by [3]) is a compile-time dimension. Thus, the View is an N by 3 array of type double, where N is specified at run time in the View's constructor.
const size_t N = ...;
Kokkos::View<double*[3]> b ("another label", N);Views may have up to (at most) 8 dimensions and any number of these may be run-time or compile-time. The only limitation is that the run-time dimensions (if any) must appear first followed by all the compile-time dimensions (if any). For example, the following are valid three-dimensional View types:
-
View<int***>(3 run-time dimensions) -
View<int**[8]>(2 run-time, 1 compile-time) -
View<int*[3][8]>(1 run-time, 2 compile-time) -
View<int[4][3][8]>(3 compile-time)
and the following are not valid three-dimensional View types:
View<int[4]**>View<int[4][3]*>View<int[4]*[8]>View<int*[3]*>
This limitation comes from the implementation of View using C++ templates. View's first template parameter must be a valid C++ type.
Note that the above used constructor is not necessarily available for all view types; specific Layouts or Memory Spaces may require more specialized allocators. This is discussed later.
Another important thing to keep in mind is that a View handle is a stateful object. It is not legal to create a View handle from raw memory by typecasting a pointer. To call any operator on a View, including the assignment operator, its constructor must have been called before. If it is necessary to initialize raw memory with a View handle, one can legally do so using placement new. The above has nothing to do with the data a View is referencing. It is completely legal to give a typecast pointer to the constructor of an unmanaged View.
Kokkos::View<int*> *a_ptr = (Kokkos::View<int*>*) malloc(10*sizeof(View<int*);
a_ptr[0] = Kokkos::View<int*>("A0",1000); // This is illegal
new(&a_ptr[1]) Kokkos::View<int*>("A1",10000); // This is legal C++ lets users construct data types that may "look like" numbers in terms of syntax but do arbitrarily complicated things inside. Some of those things may not be thread safe, like unprotected updates to global state. Others may perform badly in parallel, like fine-grained dynamic memory allocation. Therefore, it is strongly advised to use only simple data types inside Kokkos Views. Users may always construct a View whose entries are
- built-in data types ("plain old data"), like
intordouble, or - structs of built-in data types.
While it is in principle possible to have Kokkos Views of arbitrary objects, Kokkos imposes restrictions on the set of types T for which one can construct a View<T*>. For example:
-
Tmust have a default constructor and destructor. -
Tmust not have virtual methods. -
T's default constructor and destructor must not allocate or deallocate data, and must be thread safe. -
T's assignment operators as well as its default constructor and deconstructor must be marked with theKOKKOS_INLINE_FUNCTIONorKOKKOS_FUNCTIONmacro.
All restrictions but the first come from the requirement that View<T*> work with every execution and memory space,
including those that use CUDA. The constructor of View<T*> does not just allocate memory; it also initializes the allocation by iterating over it using a Kokkos::parallel_for, and invoking T's default constructor for each entry. If the View's execution space is Cuda, then that means T's default constructor needs to be correct to call on device. Keep in mind that the semantics of the resulting View are a combination of the Views 'view' semantics and the behavior of the element type.
The requirement that the destructor of T not deallocate memory technically disallows T being a managed View, or a structure which directly or indirectly contains a managed View. In extreme cases we do allow users to have managed Views in their type T, so long as a non-parallel loop is used to safely deallocate the Views contained in each T prior to the deallocation of the View<T> itself. This can be done by assigning to each contained View a default-constructed View of the same type. Having managed Views in T is not recommended.
Finally, note that virtual functions are technically allowed, but calling them is subject to further restrictions; developers should consult the discussions in Chapter 13, Kokkos and Virtual Functions (under development).
NOVICES: THE ANSWER FOR YOU IS "NO." PLEASE SKIP THIS SECTION.
A "View of Views" is a special case of View, where the type of each entry is itself a View. It is possible to make this, but before you try, please see below.
If you really just want a multidimensional array, please don't do this. Instead, see Section 6.2.1 above for the correct syntax.
If you want to represent an array of arrays, and the inner arrays have fixed length or a fixed upper bound on length, consider instead using a ``compressed sparse row'' data structure. Kokkos' Containers subpackage has a StaticCrsGraph class that you may use for this purpose.
If you want a hash table, Kokkos' Containers subpackage has an UnorderedMap class that you may use for this purpose.
One reason you might actually want a View of Views is because you need a representation of a "ragged" array of arrays -- where the inner arrays have widely varying length -- and you need to be able to reallocate the inner arrays dynamically.
You might also want a View of some class that itself contains Views. If you want this, first think about how to reorganize your data structures for better efficiency.
A View of Views would have an "outer View," with zero or more "inner Views." Section 6.2.2 above explains how the outer View's constructor would work. The outer View's constructor does not just allocate memory; it also initializes the allocation by iterating over it using a Kokkos::parallel_for, and invoking the entry type's default constructor for each entry. If the View's execution space is Cuda, then that means the entry type's default constructor needs to be correct to call on device. That is a problem, because the entry type in this case is itself View. View's constructor wants to allocate memory, and thus does not work on device. Kokkos parallel regions generally forbid memory allocation.
You could create the outer View without initializing, like this:
using Kokkos::View;
using Kokkos::view_alloc;
using Kokkos::WithoutInitializing;
// Need an std::string here, because the compiler may get confused
// if you pass view_alloc a char* as its first argument.
const std::string label ("v_outer");
View<View<int*>> v_outer (view_alloc (label, WithoutInitializing));However, that leaves the inner Views in an undefined state. You can't legally assign to them or call their destructors. (Remember that View assignment updates the assignee's reference count.) You'll need to do more than just this in order to create valid inner Views and ensure their safe deallocation.
You'll have worse problems if the outer View's memory space is CudaSpace. Allocating View construction must run on host in order to allocate memory, but you won't be able to assign the resulting constructed View to any element of the outer View.
Another issue is that View construction in a Kokkos parallel region does not update the View's reference count. Thus, the inner Views must be created in sequential host code, not inside of a Kokkos::parallel_*.
Here is how to create a View of Views, where each inner View has a separate owning allocation:
- The outer View must have a memory space that is both host and device accessible, such as
CudaUVMSpace. - Create the outer View without initializing it.
- Create inner Views using placement new, in a sequential host loop. (Prefer creating the inner Views uninitialized. Creating the inner Views initialized launches one device kernel per inner View. This is likely much slower than just initializing them all yourself from a single kernel over the outer View.)
- At this point, you may access the outer and inner Views on device.
- Before deallocating inner Views, fence to ensure all device kernels that access them have finished.
- Destroy the inner Views explicitly. (Otherwise, Step 7 will leak the inner Views' memory.)
- Get rid of the outer View as you normally would.
Here is an example:
using Kokkos::Cuda;
using Kokkos::CudaSpace;
using Kokkos::CudaUVMSpace;
using Kokkos::View;
using Kokkos::view_alloc;
using Kokkos::WithoutInitializing;
using inner_view_type = View<double*, CudaSpace>;
using outer_view_type = View<inner_view_type*, CudaUVMSpace>;
const int numOuter = 5;
const int numInner = 4;
outer_view_type outer (view_alloc (std::string ("Outer"), WithoutInitializing), numOuter);
// Create inner Views on host, outside of a parallel region, uninitialized
for (int k = 0; k < numOuter; ++k) {
const std::string label = std::string ("Inner ") + std::to_string (k);
new (&outer[k]) inner_view_type (view_alloc (label, WithoutInitializing), numInner);
}
// Outer and inner views are now ready for use on device
Kokkos::RangePolicy<Cuda, int> range (0, numOuter);
Kokkos::parallel_for ("my kernel label", range,
KOKKOS_LAMBDA (const int i) {
for (int j = 0; j < numInner; ++j) {
device_outer[i][j] = 10.0 * double (i) + double (j);
}
}
});
// Fence before deallocation on host, to make sure
// that the device kernel is done first.
// Note the new fence syntax that requires an instance.
// This will work with other CUDA streams, etc.
Cuda ().fence ();
// Destroy inner Views, again on host, outside of a parallel region.
for (int k = 0; k < 5; ++k) {
outer[k].~inner_view_type ();
}
// You're better off disposing of outer immediately.
outer = outer_view_type ();Another approach is to create the inner Views as nonowning, from a single pool of memory. This makes it unnecessary to invoke their destructors.
A view can have const data semantics (i.e. its entries are read-only) by specifying a const data type. It is a compile-time error to assign to an entry of a "const View". Assignment semantics are equivalent to a pointer to const data. A const View means the entries are const; you may still assign to a const View. View<const double*> corresponds exactly to const double*, and const View<double*> to double* const. Therefore, it does not make sense to allocate a const View since you could not obtain a non-const view of the same data and you can not assign to it. You can however assign a non-const view to a const view. Here is an example:
const size_t N0 = ...;
Kokkos::View<double*> a_nonconst ("a_nonconst", N0);
// Assign a nonconst View to a const View
Kokkos::View<const double*> a_const = a_nonconst;
// Pass the const View to some read-only function.
const double result = readOnlyFunction (a_const);Const Views often enables the compiler to optimize more aggressively by allowing it to reason about possible write conflicts and data aliasing. For example, in a vector update a(i+1)+=b(i) with skewed indexing, it is safe to vectorize if b is a View of const data.
You may access an entry of a View using parentheses enclosing a comma-delimited list of integer indices. This looks just like a Fortran multidimensional array access. For example:
const size_t N = ...;
Kokkos::View<double*[3][4]> a ("some label", N);
// KOKKOS_LAMBDA macro includes capture-by-value specifier [=].
Kokkos::parallel_for (N, KOKKOS_LAMBDA (const ptrdiff_t i) {
const size_t j = ...;
const size_t k = ...;
const double a_ijk = a(i,j,k);
/* rest of the loop body */
});Note how in the above example, we only access the View's entries in a parallel loop body. In general, you may only access a View's entries in an execution space which is allowed to access that View's memory space. For example, if the default execution space is Cuda, a View for which no specific Memory Space was given may not be accessed in host code1.
Furthermore, access costs (e.g., latency and bandwidth) may vary, depending on the View's "native" memory and execution spaces and the execution space from which you access it. CUDA UVM may work, but it may also be slow, depending on your access pattern and performance requirements. Thus, best practice is to access the View only in a Kokkos parallel for, reduce, or scan, using the same execution space as the View. This also ensures that access to the View's entries respect first-touch allocation. The first (leftmost) dimension of the View is the parallel dimension over which it is most efficient to do parallel array access if the default memory layout is used (e.g. if no specific memory layout is specified).
1 An exemption is if you specified for CUDA compilation that the default memory space is CudaUVMSpace, which can be accessed from the host.
Kokkos automatically manages deallocation of Views through a reference-counting mechanism. Otherwise, Views behave like raw pointers. Copying or assigning a View does a shallow copy, and changes the reference count. (The View copied has its reference count incremented, and the assigned-to View has its reference count decremented.) A View's destructor (called when the View falls out of scope or during a stack unwind due to an exception) decrements the reference count. Once the reference count reaches zero, Kokkos may deallocate the View.
For example, the following code allocates two Views, then assigns one to the other. That assignment may deallocate the first View, since it reduces its reference count to zero. It then increases the reference count of the second View, since now both Views point to it.
Kokkos::View<int*> a ("a", 10);
Kokkos::View<int*> b ("b", 10);
a = b; // assignment does shallow copyFor efficiency, View allocation and reference counting turn off inside of Kokkos' parallel for, reduce, and scan operations. This affects what you can do with Views inside of Kokkos' parallel operations.
The lifetime of an allocation begins when a View is constructed by an allocating constructor such as
Kokkos::View<int*> b("b", 10);The lifetime of an allocation ends when there are no more Views which reference that allocation (see reference counting above).
Kokkos requires that the lifetime of all allocations ends before the call to Kokkos::finalize!
For example, the following is incorrect usage of Kokkos:
int main() {
Kokkos::initialize();
Kokkos::View<double*> p("constructed view", 100);
Kokkos::finalize();
// p is destroyed here, after Kokkos::finalize
}Kokkos Views can be resized using the resize non-member function. It takes an existing view as its input by reference and the new dimension information corresponding to the constructor arguments. A new view with the new dimensions will be created and a kernel will be run in the views execution space to copy the data element by element from the old view to the new one. Note that the old allocation is only deleted if the view to be resized was the only view referencing the underlying allocation.
// Allocate a view with 100x50x4 elements
Kokkos::View<int**[4]> a( "a", 100,50);
// Resize a to 200x50x4 elements; the original allocation is freed
Kokkos::resize(a, 200,50);
// Create a second view b viewing the same data as a
Kokkos::View<int**[4]> b = a;
// Resize a again to 300x60x4 elements; b is still 200x50x4
Kokkos::resize(a,300,60);Layout refers to the mapping from a logical multidimensional index (i, j, k, . . .) to a physical memory offset. Different programming languages may have different layout conventions. For example, Fortran uses column-major or "left" layout, where consecutive entries in the same column of a 2-D array are contiguous in memory. Kokkos calls this LayoutLeft. C, C++, and Java use row-major or "right" layout, where consecutive entries in the same row of a 2-D array are contiguous in memory. Kokkos calls this LayoutRight.
The generalization of both left and right layouts is "strided." For a strided layout, each dimension has a stride. The stride for that dimension determines how far apart in memory two array entries are, whose indices in that dimension differ only by one, and whose other indices are all the same. For example, with a 3-D strided view with strides (s_1, s_2, s_3), entries (i, j, k) and (i, j+1, k) are s_2 entries (not bytes) apart in memory. Kokkos calls this LayoutStride.
Strides may differ from dimensions. For example, Kokkos reserves the right to pad each dimension for cache or vector alignment. You may access the dimensions of a View using the (ISO/C++ form) extent method, which takes the index of the dimension.
Strides are accessed using the stride method. It takes a raw integer array, and only fills in as many entries as the rank of the View. For example:
const size_t N0 = ...;
const size_t N1 = ...;
const size_t N2 = ...;
Kokkos::View<int***> a ("a", N0, N1, N2);
int dim1 = a.extent (1); // returns dimension 1
size_t strides[3]
a.stride (strides); // fill 'strides' with stridesYou may also refer to specific dimensions without a runtime parameter:
const size_t n0 = a.extent_0 ();
const size_t n2 = a.extent_2 ();Note the return type of extent_N() is the size_type of the views memory space. This causes some issues if warning-free compilation should be achieved since it will typically be necessary to cast the return value. In particular, in cases where the size_type is more conservative than required, it can be beneficial to cast the value to int since signed 32 bit integers typically give the best performance when used as index types. In index heavy codes, this performance difference can be significant compared to using size_t since the vector length on many architectures is twice as long for 32 bit values as for 64 bit values and signed integers have less stringent overflow testing requirements than unsigned integers.
Users of the BLAS and LAPACK libraries may be familiar with the ideas of layout and stride. These libraries only accept matrices in column-major format. The stride between consecutive entries in the same column is 1, and the stride between consecutive entries in the same row is LDA ("leading dimension of the matrix A"). The number of rows may be less than LDA, but may not be greater.
Other layouts are possible. For example, Kokkos has a "tiled" layout, where a tile's entries are stored contiguously (in either row- or column-major order) and tiles have compile-time dimensions. One may also use Kokkos to implement Morton ordering or variants thereof. In order to write a custom layout one has to define a new layout class and specialise the ViewMapping class for that layout. The ViewMapping class implements the offset operator as well as stride calculation for regular layouts. A good way to start such a customization is by copying the implementation of LayoutLeft and its associated ViewMapping specialization, renaming the layout and then change the offset operator.
Kokkos selects a View's default layout for optimal parallel access over the leftmost dimension based on its execution space. For example, View<int**, Cuda> has LayoutLeft, so that consecutive threads in the same warp access consecutive entries in memory. This coalesced access gives the code better memory bandwidth.
In contrast, View<int**, OpenMP> has LayoutRight, so that a single thread accesses contiguous entries of the array. This avoids wasting cache lines and helps prevent false sharing of a cache line between threads. In Section 6.4 more details will be discussed.
We prefer that users let Kokkos determine a View's layout, based on its execution space. However, sometimes you really need to specify the layout. For example, the BLAS and LAPACK libraries only accept column-major arrays. If you want to give a View to the BLAS or LAPACK library, that View must be LayoutLeft. You may specify the layout as a template parameter of View. For example:
const size_t N0 = ...;
const size_t N1 = ...;
Kokkos::View<double**, Kokkos::LayoutLeft> A ("A", N0, N1);
// Get 'LDA' for BLAS / LAPACK
int strides[2]; // any integer type works in stride()
A.stride (strides);
const int LDA = strides[1];You may ask a View for its layout via its array_layout typedef. This can be helpful for C++ template metaprogramming. For example:
template<class ViewType>
void callBlas (const ViewType& A) {
typedef typename ViewType::array_layout array_layout;
if (std::is_same<array_layout, LayoutLeft>::value) {
callSomeBlasFunction (A.data(), ...);
} else {
throw std::invalid_argument ("A is not LayoutLeft");
}
}Views are allocated by default in the default execution space's default memory space. You may access the View's execution space via its execution_space typedef, and its memory space via its memory_space typedef. You may also specify the memory space explicitly as a template parameter. For example, the following allocates a View in CUDA device memory:
Kokkos::View<int*, Kokkos::CudaSpace> a ("a", 100000);and the following allocates a View in "host" memory, using the default host execution space for first-touch initialization:
Kokkos::View<int*, Kokkos::HostSpace> a ("a", 100000);Since there is no bijective association between execution spaces and memory spaces, Kokkos provides a way to explicitly provide both to a View as a Device.
Kokkos::View<int*, Kokkos::Device<Kokkos::Cuda,Kokkos::CudaUVMSpace> > a ("a", 100000);
Kokkos::View<int*, Kokkos::Device<Kokkos::OpenMP,Kokkos::CudaUVMSpace> > b ("b", 100000);In this case a and b will live in the same memory space, but a will be initialized on the GPU while b will be
initialized on the host. The Device type can be accessed as a views device_type typedef. A Device has only three typedef members: device_type, execution_space and memory_space. The execution_space and memory_space typedefs are the same for a view as the device_type typedef.
It is important to understand that accessibility of a View does not depend on its execution space directly. It is only determined by its memory space. Therefore both a and b have the same access properties. They differ only in how they are initialized and in where parallel kernels associated with operations such as resizing or deep copies are run.
The following is the accessibility matrix for execution and memory spaces:
| Serial | OpenMP | Threads | Cuda | ROCm | |
|---|---|---|---|---|---|
| HostSpace | x | x | x | - | - |
| HBWSpace | x | x | x | - | - |
| CudaSpace | - | - | - | x | - |
| CudaUVMSpace | x | x | x | x | - |
| CudaHostPinnedSpace | x | x | x | x | - |
| ROCmSpace | - | - | - | - | x |
| ROCmHostPinnedSpace | x | x | x | - | x |
This relationship can be queried via the SpaceAccessibility class:
template< typename AccessSpace , typename MemorySpace >
struct SpaceAccessibility {
enum { accessible }; // AccessSpace can access MemorySpace
enum { assignable }; // Can assign View<...,AccessSpace,...> = View<...,MemorySpace,...>
enum { deep_copy }; // Can deep copy to AccessSpace::memory_space from MemorySpace
};A typical use case would be:
if(SpaceAccessibility<ExecSpace, ViewType::memory_space>::accessible) {
parallel_for(RangePolicy<ExecSpace>, functor);
}A View's entries are initialized to zero by default. Initialization happens in parallel for first-touch allocation over the first (leftmost) dimension of the View using the execution space of the View.
You may allocate a View without initializing. For example:
Kokkos::View<int*> x (Kokkos::ViewAllocateWithoutInitializing (label), 100000);This is mainly useful in cases when the initial values of the view are not important because they will be overwritten without ever being read. It is still important that the first write to each location be done within a parallel kernel in a way that reflects how first-touch affinity to threads is desired. Typically it is sufficient to use the parallel iteration index as the index of the location in the view to write to.
Copying data from one view to another, in particular between views in different memory spaces, is called deep copy.
Kokkos never performs a hidden deep copy. To do so a user has to call the deep_copy function. For example:
Kokkos::View<int*> a ("a", 10);
Kokkos::View<int*> b ("b", 10);
Kokkos::deep_copy (a, b); // copy contents of b into aDeep copies can only be performed between views with an identical memory layout and padding. For example the following two operations are not valid:
Kokkos::View<int*[3], Kokkos::CudaSpace> a ("a", 10);
Kokkos::View<int*[3], Kokkos::HostSpace> b ("b", 10);
Kokkos::deep_copy (a, b); // This will give a compiler error
Kokkos::View<int*[3], Kokkos::LayoutLeft, Kokkos::CudaSpace> c ("c", 10);
Kokkos::View<int*[3], Kokkos::LayoutLeft, Kokkos::HostSpace> d ("d", 10);
Kokkos::deep_copy (c, d); // This might give a runtime errorThe first one will not work because the default layouts of CudaSpace and HostSpace are different. The compiler will catch that since no overload of the deep_copy function exists to copy view from one layout to another. The second case will fail at runtime if padding settings are different for the two memory spaces. This would result in different allocation sizes and thus prevent a direct memcopy.
The reasoning for allowing only direct bitwise copies is that a deep copy between different memory spaces would otherwise require a temporary copy of the data to which a bitwise copy is performed followed by a parallel kernel to transfer the data element by element.
Kokkos provides the following way to work around those limitations. First views have a HostMirror typedef which is a view type with compatible layout inside the HostSpace. Additionally there is a create_mirror and create_mirror_view function which allocate views of the HostMirror type of a view. The difference between the two is that create_mirror will always allocate a new view, while create_mirror_view will only create a new view if the original one is not in HostSpace.
Kokkos::View<int*[3], MemorySpace> a ("a", 10);
// Allocate a view in HostSpace with the layout and padding of a
typename Kokkos::View<int*[3], MemorySpace>::HostMirror b =
create_mirror(a);
// This is always a memcopy
Kokkos::deep_copy (b, a);
typename Kokkos::View<int*[3]>::HostMirror c =
Kokkos::create_mirror_view(a);
// This is a no-op if MemorySpace is HostSpace
Kokkos::deep_copy (c, a)We discourage access to a View's "raw" pointer. This circumvents reference counting, that is, the memory may be deallocated once the View's reference count goes to zero so holding on to a raw pointer may result in invalid memory access. Furthermore, it may not even be possible to access the View's memory from a given execution space. For example, a View in the Cuda space points to CUDA device memory. Also using raw pointers would normally defeat the usability of polymorpic layouts and automatic padding. Nevertheless, for instances you really need access to the pointer, we provide the data() method. For example:
// Legacy function that takes a raw pointer.
extern void legacyFunction (double* x_raw, const size_t len);
// Your function that takes a View.
void myFunction (const Kokkos::View<double*>& x) {
// DON'T DO THIS UNLESS YOU MUST
double* x_raw = x.data();
const size_t N = x.extent(0);
legacyFunction (x_raw, N);
}A user is in most cases also allowed to obtain a pointer to a specific element via the usual & operator. For example
// Legacy function that takes a raw pointer.
void someLibraryFunction (double* x_raw);
KOKKOS_INLINE_FUNCTION
void foo(const Kokkos::View<double*>& x) {
someLibraryFunction(&x(3));
}This is only valid if a Views reference type is an lvalue. That property can be queried statically at compile time from the view through its reference_is_lvalue member.
Another way to get optimized data accesses is to specify a memory trait. These traits are used to declare intended use of the particular view of an allocation. For example, a particular kernel might use a view only for streaming writes. By declaring that intention, Kokkos can insert the appropriate store intrinsics on each architecture if available. Access traits are specified through an optional template parameter which comes last in the list of parameters. Multiple traits can be combined with binary "or" operators:
Kokkos::View<double*, Kokkos::MemoryTraits<SomeTrait> > a;
Kokkos::View<const double*, Kokkos::MemoryTraits<SomeTrait | SomeOtherTrait> > b;
Kokkos::View<int*, Kokkos::LayoutLeft, Kokkos::MemoryTraits<SomeTrait | SomeOtherTrait> > c;
Kokkos::View<int*, MemorySpace, Kokkos::MemoryTraits<SomeTrait | SomeOtherTrait> > d;
Kokkos::View<int*, Kokkos::LayoutLeft, MemorySpace, Kokkos::MemoryTraits<SomeTrait> > e;The Atomic access trait lets you create a View of data such that every read or write to any entry uses an atomic update. Kokkos supports atomics for all data types independent of size. Restrictions are that you are
- not allowed to alias data for which atomic operations are performed, and
- the results of non atomic accesses (including read) to data which is at the same time atomically accessed is not defined.
Performance characteristics of atomic operations depend on the data type. Some types (in particular integer types) are natively supported and might even provide asynchronous atomic operations. Others (such as 32 bit and 64 bit atomics for non-integer types) are often implemented using CAS loops of integers. Everything else is implemented with a locking approach where an atomic operation acquires a lock based on a hash of the pointer value of the data element.
Types for which atomic access are performed must support the necessary operators such as =, +=, -=, +, - etc. as well as have a number of volatile overloads of functions such as assign and copy constructors defined.
Kokkos::View<int*> a("a" , 100);
Kokkos::View<int*, Kokkos::MemoryTraits<Kokkos::Atomic> > a_atomic = a;
a_atomic(1) += 1; // This access will do an atomic additionThe RandomAccess trait declares the intent to access a View irregularly (in particular non consecutively). If used for a const View in the CudaSpace or CudaUVMSpace, Kokkos will use texture fetches for accesses when executing in the Cuda execution space. For example:
const size_t N0 = ...;
Kokkos::View<int*> a_nonconst ("a", N0); // allocate nonconst View
// Assign to const, RandomAccess View
Kokkos::View<const int*, Kokkos::MemoryTraits<Kokkos::RandomAccess>> a_ra = a_nonconst;If the default execution space is Cuda, access to a RandomAccess View may use CUDA texture fetches. Texture fetches are not cache-coherent with respect to writes so you must use read-only access. The texture cache is optimized for noncontiguous access since it has a shorter cache line than the regular cache.
While RandomAccess is valid for other execution spaces, currently no specific optimizations are performed. But in the future a view allocated with the RandomAccess attribute might for example, use a larger page size, and thus reduce page faults in the memory system.
The standard idiom for View is to pass it around using as few template parameters as possible. Then, assign to a View with the desired access traits only at the "last moment" when those access traits are needed just before entering a computational kernel. This lets you template C++ classes and functions on the View type without proliferating instantiations. Here is an example:
// Compute a sparse matrix-vector product, for a sparse
// matrix stored in compressed sparse row (CSR) format.
void spmatvec (const Kokkos::View<double*>& y,
const Kokkos::View<const size_t*>& ptr,
const Kokkos::View<const int*>& ind,
const Kokkos::View<const double*>& val,
const Kokkos::View<const double*>& x)
{
// Access to x has less locality than access to y.
Kokkos::View<const double*, Kokkos::MemoryTraits<Kokkos::RandomAccess>> x_ra = x;
typedef Kokkos::View<const size_t*>::size_type size_type;
Kokkos::parallel_for (y.extent_0 (), KOKKOS_LAMBDA (const size_type i) {
double y_i = y(i);
for (size_t k = ptr(i); k < ptr(i+1); ++k) {
y_i += val(k) * x_ra(ind(k));
}
y(i) = y_i;
});
}It's always better to let Kokkos control memory allocation but sometimes you don't have a choice. You might have to work with an application or an interface that returns a raw pointer, for example. Kokkos lets you wrap raw pointers in an unmanaged View. "Unmanaged" means that Kokkos does not do reference counting or automatic deallocation for those Views. The following example shows how to create an unmanaged View of host memory. You may do this for CUDA device memory too, or indeed for memory allocated in any memory space, by specifying the View's execution or memory space accordingly.
// Sometimes other code gives you a raw pointer, ...
const size_t N0 = ...;
double* x_raw = malloc (N0 * sizeof (double));
{
// ... but you want to access it with Kokkos.
//
// malloc() returns host memory, so we use the host memory space HostSpace.
// Unmanaged Views have no label because labels work with the reference counting system.
Kokkos::View<double*, Kokkos::HostSpace, Kokkos::MemoryTraits<Kokkos::Unmanaged> >
x_view (x_raw, N0);
functionThatTakesKokkosView (x_view);
// It's safest for unmanaged Views to fall out of scope before freeing their memory.
}
free (x_raw);Not all view types can be assigned to each other. Requirements are:
- the data type and dimension have to match,
- the layout must be compatible and
- the memory space has to match.
Examples illustrating the rules are:
- Data Type and Rank has to Match
-
int* -> int*/ ok -
int* -> const int*/ ok -
const int* -> int*/ not ok, const violation -
int** -> int*/ not ok, rank mismatch -
int*[3] -> int**/ ok -
int** -> int*[3]/ ok if runtime dimension check matches -
int* -> long*/ not ok, type mismatch
-
- Layouts must be compatible
-
LayoutRight -> LayoutRight/ ok -
LayoutLeft -> LayoutRight/ not ok except for 1D Views -
LayoutLeft -> LayoutStride/ ok -
LayoutStride -> LayoutLeft/ ok if runtime dimensions allow assignment
-
- Memory Spaces must match
-
Kokkos::View<int*> -> Kokkos::View<int*,HostSpace>/ ok if default memory space is HostSpace
-
- Memory Traits
Home:
- Introduction
- Machine Model
- Programming Model
- Compiling
- Initialization
- View
- Parallel Dispatch
- Hierarchical Parallelism
- Custom Reductions
- Atomic Operations
- Subviews
- Interoperability
- Kokkos and Virtual Functions
- Initialization and Finalization
- View
- Data Parallelism
- Execution Policies
- Spaces
- Task Parallelism
- Utilities
- STL Compatibility
- Numerics
- Detection Idiom