Skip to main content Link Menu Expand (external link) Document Search Copy Copied

Parallel Execution Policies

template <typename DerivedPolicy> struct thrust::host_execution_policy;
template <typename DerivedPolicy> struct thrust::device_execution_policy;
static const detail::host_t thrust::host;
THRUST_INLINE_CONSTANT detail::device_t thrust::device;

Member Classes

Struct thrust::host_execution_policy

Inherits From: thrust::system::__THRUST_HOST_SYSTEM_NAMESPACE::execution_policy< DerivedPolicy >

Struct thrust::device_execution_policy

Inherits From: thrust::system::__THRUST_DEVICE_SYSTEM_NAMESPACE::execution_policy< DerivedPolicy >

Variables

Variable thrust::host

static const detail::host_t host; thrust::host is the default parallel execution policy associated with Thrust’s host backend system configured by the THRUST_HOST_SYSTEM macro.

Instead of relying on implicit algorithm dispatch through iterator system tags, users may directly target algorithm dispatch at Thrust’s host system by providing thrust::host as an algorithm parameter.

Explicit dispatch can be useful in avoiding the introduction of data copies into containers such as thrust::host_vector.

Note that even though thrust::host targets the host CPU, it is a parallel execution policy. That is, the order that an algorithm invokes functors or dereferences iterators is not defined.

The type of thrust::host is implementation-defined.

The following code snippet demonstrates how to use thrust::host to explicitly dispatch an invocation of thrust::for_each to the host backend system:

#include <thrust/for_each.h>
#include <thrust/execution_policy.h>
#include <cstdio>

struct printf_functor
{
  __host__ __device__
  void operator()(int x)
  {
    printf("%d\n", x);
  }
};
...
int vec(3);
vec[0] = 0; vec[1] = 1; vec[2] = 2;

thrust::for_each(thrust::host, vec.begin(), vec.end(), printf_functor());

// 0 1 2 is printed to standard output in some unspecified order

See:

Variable thrust::device

THRUST_INLINE_CONSTANT detail::device_t device; thrust::device is the default parallel execution policy associated with Thrust’s device backend system configured by the THRUST_DEVICE_SYSTEM macro.

Instead of relying on implicit algorithm dispatch through iterator system tags, users may directly target algorithm dispatch at Thrust’s device system by providing thrust::device as an algorithm parameter.

Explicit dispatch can be useful in avoiding the introduction of data copies into containers such as thrust::device_vector or to avoid wrapping e.g. raw pointers allocated by the CUDA API with types such as thrust::device_ptr.

The user must take care to guarantee that the iterators provided to an algorithm are compatible with the device backend system. For example, raw pointers allocated by std::malloc typically cannot be dereferenced by a GPU. For this reason, raw pointers allocated by host APIs should not be mixed with a thrust::device algorithm invocation when the device backend is CUDA.

The type of thrust::device is implementation-defined.

The following code snippet demonstrates how to use thrust::device to explicitly dispatch an invocation of thrust::for_each to the device backend system:

#include <thrust/for_each.h>
#include <thrust/device_vector.h>
#include <thrust/execution_policy.h>
#include <cstdio>

struct printf_functor
{
  __host__ __device__
  void operator()(int x)
  {
    printf("%d\n", x);
  }
};
...
thrust::device_vector<int> vec(3);
vec[0] = 0; vec[1] = 1; vec[2] = 2;

thrust::for_each(thrust::device, vec.begin(), vec.end(), printf_functor());

// 0 1 2 is printed to standard output in some unspecified order

See: