Skip to content
This repository

Introduction

Building a stand alone Thrust system is similar to customizing an existing one. However, because a stand alone system has no derived algorithm implementations to rely on, all primitives must be implemented. There are two major classes of primitive functionality which a Thrust system must provide: memory model and algorithms. On this page, we'll demonstrate a stand alone system similar to Thrust's standard C++ system. Our novel system will be roughly equivalent in functionality to thrust::cpp, but otherwise unrelated.

Memory Model Primitives

The first task when building a stand alone Thrust system is to build the protocol which defines its memory model. Special primitives are required to support systems which may be implemented with distributed, or remote memory architectures. These primitives allow the programmer to manipulate logical scalar objects in memory.

Let's choose a name for our system. How about standalone? Let's create its system policy tag in that namespace:

#include <thrust/execution_policy.h>

namespace standalone
{

// standalone's tag is just an empty struct derived from thrust::execution_policy.
struct tag : thrust::execution_policy<tag> {};

}

Next, we'll define how to allocate and deallocate the standalone system's storage. Since we're just using C++, we'll implement standalone::malloc and standalone::free with std::malloc and std::free.

#include <cstdlib>
#include <thrust/memory.h>

namespace standalone
{

void *malloc(policy_tag, std::size_t n)
{
  return std::malloc(n);
}

template<typename Pointer>
void free(tag, Pointer ptr)
{
  std::free(thrust::raw_pointer_cast(ptr));
}

}

In our example, standalone::malloc's semantics are identical to std::malloc, with the addition of the standalone::tag parameter. standalone::free needs to be a template, because it is responsible for deallocating any pointer-like thing tagged with standalone::tag. We use thrust::raw_pointer_cast to get ptr's raw pointer to give to std::free.

Next, we'll define the protocol for manipulating pointers tagged with standalone::tag. There are three of these primitives. assign_value assigns the value pointed to by one pointer to that of another. get_value dereferences a pointer and returns a copy of its pointee's value. iter_swap exchanges the values of two pointees.

#include <thrust/memory.h>
#include <thrust/iterator/iterator_traits.h>
#include <thrust/swap.h>

namespace standalone
{

template<typename Pointer1, typename Pointer2>
__host__ __device__
void assign_value(tag, Pointer1 dst, Pointer2 src)
{
  *thrust::raw_pointer_cast(dst) = *thrust::raw_pointer_cast(src);
}

template<typename Pointer>
__host__ __device__
typename thrust::iterator_value<Pointer>::type
get_value(tag, Pointer ptr)
{
  return *thrust::raw_pointer_cast(ptr);
}

template<typename Pointer1, typename Pointer2>
__host__ __device__
void iter_swap(tag, Pointer1 a, Pointer2 b)
{
  using thrust::swap;

  swap(*thrust::raw_pointer_cast(a), *thrust::raw_pointer_cast(b));
}

}

In this example, assign_value can simply dereference the raw pointers wrapped by dst and src. However, systems such as Thrust's CUDA system must implement this primitive carefully to avoid dereferencing a remote pointer.

Likewise, get_value simply returns the result of dereferencing ptr's raw pointer. This primitive may seem superfluous and implementable with assign_value, but this is not the case for heterogeneous memories like CUDA's.

Finally, iter_swapis necessary to exchange the values of two potentially remote objects without introducing a temporary. The simple standalone system can implement this primitive withraw_pointer_castandswap`.

Even though these primitives are unrelated to CUDA, they require __host__ __device__ decoration at this time because they may be invoked by a __host__ __device__ function in Thrust's implementation.

Tagged Pointer and Reference Types

Now that we've defined how to allocate and interact with objects in memory, we can wrap up the protocol so that it's less verbose to use. This is the job of thrust::pointer.

We want our standalone system to be interoperable with Thrust's standard C++ system so that we can write expressions like *standalone_ptr = 13. To enable this to work, we need to define what happens when the Thrust dispatch process encounters a standalone::tag paired with a thrust::cpp::tag.

We do this by providing an overload of select_system:

#include <thrust/system/cpp/memory.h>

namespace standalone
{

__host__ __device__
tag select_system(tag, thrust::cpp::tag)
{
  return tag();
}

}

Let's try it all out:

#include <thrust/memory.h>

int main()
{
  typedef thrust::pointer<int, standalone::tag> int_ptr_t;

  int_ptr_t ptr(static_cast<int*>(malloc(standalone::tag(), sizeof(int))));

  *ptr = 13;

  std::cout << "value is " << *ptr << std::endl;

  free(standalone::tag(), ptr);
}
$ nvcc standalone_memory.cu -o standalone_memory
./standalone_memory
value is 13

It works!

Algorithmic Primitives

Now that we've implemented our standalone system's model primitives, we can move on to implementing the algorithmic primitives. There are several of these, and to build a fully-conforming system we'd need to implement all of them. In this example we'll demonstrate how to implement just a couple of them.

Perhaps the simplest algorithmic primitive to implement is for_each. Let's build a straightforward serial implementation with a for loop:

#include <thrust/memory.h>

namespace standalone
{

template<typename Iterator, typename Function>
  Iterator for_each(tag, Iterator first, Iterator last, Function f)
{
  for(; first != last; ++first)
  {
    f(thrust::raw_reference_cast(*first));
  }

  return first;
}

}

Let's deconstruct what we see here. First, we've provided an overloaded form of for_each which takes standalone::tag as its first parameter. Importantly, we've provided our for_each in the same namespace as the tag. The body of the for loop looks normal, except that we use thrust::raw_reference_cast to transform the tagged reference returned by *first into a raw reference suitable to pass to for_each's function object parameter.

Let's see if Thrust can dispatch our for_each with the pointers we built before:

#include <thrust/for_each.h>
#include <thrust/memory.h>

struct printer
{
  void operator()(int x)
  {
    std::cout << "x" << std::endl;
  }
};

int main()
{
  typedef thrust::pointer<int,standalone::tag> int_ptr_t;

  int_ptr_t ptr(static_cast<int*>(malloc(standalone::tag(), 3 * sizeof(int))));

  ptr[0] = 7;
  ptr[1] = 13;
  ptr[2] = 42;

  thrust::for_each(ptr, ptr + 3, printer());

  free(standalone::tag(), ptr);
}

Here's the result:

$ nvcc standalone_for_each.cu -o standalone_for_each
$ ./standalone_for_each
7
13
42

Let's continue with reduce:

#include <thrust/iterator/iterator_traits.h>
#include <thrust/memory.h>

namespace standalone
{

template<typename InputIterator, typename T, typename BinaryFunction>
  typename thrust::iterator_value<InputIterator>::type
    reduce(tag, InputIterator first, InputIterator last, T init, BinaryFunction binary_op)
{
  T result = init;

  for(; first != last; ++first)
  {
    result = binary_op(result, thrust::raw_reference_cast(*first));
  }

  return result;
}

}

That's one version of reduce, but what about its two other overloads? Our version of reduce is its most general form; we get the other two for free. Let's try it:

#include <thrust/memory.h>
#include <thrust/functional.h>
#include <iostream>

int main()
{
  typedef thrust::pointer<int,standalone::tag> int_ptr_t;

  int_ptr_t ptr(static_cast<int*>(malloc(standalone::tag(), 3 * sizeof(int))));

  ptr[0] = 7;
  ptr[1] = 13;
  ptr[2] = 42;

  int result1 = thrust::reduce(ptr, ptr + 3);

  std::cout << "result1 is " << result1 << std::endl;

  int result2 = thrust::reduce(ptr, ptr + 3, 1);

  std::cout << "result2 is " << result2 << std::endl;

  int result3 = thrust::reduce(ptr, ptr + 3, 1, thrust::multiplies<int>());

  std::cout << "result3 is " << result3 << std::endl1;

  free(standalone::tag(), ptr);
}

Here's the result:

$ nvcc standalone_reduce.cu -o standalone_reduce
$ ./standalone_reduce
result1 is 62
result2 is 63
result3 is 3822

Only the most general form of reduce is a primitive; Thrust knows how to implement the other two in terms of the more primitive form. Of course, we can always provide overloads for those other forms if we wish.

System Interoperability

What happens when an algorithm has many iterator parameters and their system tags differ? This is where the function select_system comes in. Remember that Thrust dispatches algorithms using the result of the special function select_system. In general, select_system may receive many tag arguments during dispatch:

template<typename Iterator1, typename Iterator2, ..., typename IteratorN>
void some_algorithm(Iterator1 first1, Iterator1 last1, Iterator2 first2, ..., IteratorN firstN)
{
  // get the iterators' system tags
  typedef typename thrust::iterator_system<Iterator1>::type tag1;
  typedef typename thrust::iterator_system<Iterator2>::type tag2;
  ...
  typedef typename thrust::iterator_system<Iterator4>::type tagN;

  return some_algorithm(select_system(tag1(),tag2(), ..., tagN()), first1, first2, ..., firstN);
}

The default version of select_system performs a reduction over its tag arguments and returns the "minimal" tag, if it finds one. Here, the "minimal" tag is the tag whose type can be converted to from every other tag in select_system's parameter list. For example, the result of select_system(thrust::cuda::tag(), thrust::any_system_tag()) is thrust::cuda::tag, since thrust::any_system_tag is convertible to thrust::cuda::tag. It is an error if no such system tag exists.

We might wish to customize the behavior of select_system, especially when we want two systems to interoperate. A good example is when overloading Thrust's copy algorithm. Recall that our standalone system has already provided the overload select_system(standalone::tag, thrust::cpp::tag), which allows us to store values into standalone-tagged pointers using expressions like *ptr = 13.

To enable thrust::copy to allow copying between standard C++ and our standalone system, we need to provide two overloads of select_system, each corresponding to a direction of the copy:

namespace standalone
{

__host__ __device__
tag select_system(tag, thrust::cpp::tag)
{
  return tag();
}

__host__ __device-_
tag select_system(thrust::cpp::tag, tag)
{
  return tag();
}

template<typename InputIterator, typename OutputIterator>
OutputIterator copy(tag, InputIterator first, InputIterator last, OutputIterator result)
{
  for(; first != last; ++first, ++result)
  {
    thrust::raw_reference_cast(*result) = thrust::raw_reference_cast(*first);
  }

  return result;
}

}

Now we should be able to copy in and out of our standalone backend.

#include <thrust/copy.h>
#include <vector>
#include <iterator>

int main()
{
  typedef thrust::pointer<int,standalone::tag> int_ptr_t;

  int_ptr_t ptr(static_cast<int*>(malloc(standalone::tag(), 3 * sizeof(int))));

  std::vector<int> vec(3);
  vec[0] = 7;
  vec[1] = 13;
  vec[2] = 42;

  thrust::copy(vec.begin(), vec.end(), ptr);

  std::cout << "copied in ";
  thrust::copy(ptr, ptr + 3, std::ostream_iterator<int>(std::cout, " "));
  std::cout << std::endl;

  ptr[0] = 77;
  ptr[1] = 1313;
  ptr[2] = 4242;

  thrust::copy(ptr, ptr + 3, vec.begin());

  std::cout << "copied out ";
  thrust::copy(vec.begin(), vec.end(), std::ostream_iterator<int>(std::cout, " "));
  std::cout << std::endl;

  free(standalone::tag(), ptr);
}
$ nvcc standalone_copy.cu -o standalone_copy
$ ./standalone_copy
copied in 7 13 42
copied out 77 13 4242

If necessary, we can encode the directionality of the copy by introducing new tags, and specialize standalone's versions of select_system and copy further:

#include <thrust/system/cpp/memory.h>
#include <iostream>

namespace standalone
{

struct cpp_to_standalone {};

struct standalone_to_cpp {};

__host__ __device__
cpp_to_standalone select_system(thrust::cpp::tag, tag)
{
  return cpp_to_standalone();
}

__host__ __device__
standalone_to_cpp select_system(tag, thrust::cpp::tag)
{
  return standalone_to_cpp();
}

template<typename InputIterator, typename OutputIterator>
  OutputIterator copy(cpp_to_standalone, InputIterator first, InputIterator last, OutputIterator result)
{
  std::cout << "copying from cpp to standalone" << std::endl;

  ...
}

template<typename InputIterator, typename OutputIterator>
  OutputIterator copy(standalone_to_cpp, InputIterator first, InputIterator last, OutputIterator result)
{
  std::cout << "copying from standalone to cpp" << std::endl;

  ...
}

}

Thrust's CUDA system uses this idiom to implement copies between system and GPU memory.

In general, it is at each system's discretion how (if at all) to handle iterator parameters with heterogeneous system tags.

Further Reading

Something went wrong with that request. Please try again.