我有一个指向迭代器元组的zip迭代器。我想提供一个push :: transform with functor,它将使用元组获取元素并生成标量输出。
我的程序不起作用,我不知道为什么。
我认为这可能与某些事情有关: CUDA thrust zip_iterator tuple transform_reduce,但更改仿函数的模板参数并没有成功。
以下代码编译:
#include <iostream>
#include <thrust/transform.h>
#include <thrust/functional.h>
#include <thrust/device_vector.h>
#include <thrust/device_ptr.h>
#include <thrust/iterator/zip_iterator.h>
#include <thrust/tuple.h>
typedef thrust::device_vector<double>::iterator realIter;
typedef thrust::tuple<realIter,realIter> Tup;
typedef thrust::zip_iterator<Tup> Zip;
typedef thrust::tuple<double,double> Tup2; //I tried replacing Tup with this in the functor
struct dummyOp : public thrust::unary_function<Tup, double> {
__host__ __device__ double operator()(Tup &tup){
double result = *thrust::get<0>(tup);
return result;
}
};
int main(){
thrust::device_vector<double> A(4);
thrust::device_vector<double> B(4);
thrust::device_vector<double> C(4);
A[0] = 1.; A[1] = 2.;
A[2] = 3.; A[3] = 4.;
B[0] = 4.; B[1] = 3.;
B[2] = 2.; B[3] = 1.;
Tup tup = thrust::tuple<realIter,realIter>(A.begin(),B.begin());
Zip zippy = thrust::zip_iterator<Tup>(tup);
dummyOp f;
// The following does not work:
//thrust::transform(zippy, zippy + 4, C.begin(), f);
std::cout << "A:\n";
thrust::copy(A.begin(), A.end(), std::ostream_iterator<double>(std::cout, " "));
std::cout << "\nB:\n";
thrust::copy(B.begin(), B.end(), std::ostream_iterator<double>(std::cout, " "));
std::cout << "\nC:\n";
thrust::copy(C.begin(), C.end(), std::ostream_iterator<double>(std::cout, " "));
std::cout << std::endl;
std::cout <<"get<0>(zippy[0]) returns:\n" << thrust::get<0>(zippy[0]) << std::endl;
std::cout <<"get<1>(zippy[1]) returns:\n" << thrust::get<1>(zippy[1]) << std::endl;
return 0;
}
运行它会给出:
$ ./so2
A:
1 2 3 4
B:
4 3 2 1
C:
0 0 0 0
get<0>(zippy[0]) returns:
1
get<1>(zippy[1]) returns:
3
如果违反了违规行,我们会:
$ nvcc -arch=compute_35 so2.cu -o so2
/usr/local/cuda/bin/../targets/x86_64-linux/include/thrust/detail/internal_functional.h(322): error: function "dummyOp::operator()" cannot be called with the given argument list
argument types are: (thrust::detail::tuple_of_iterator_references<double &, double &, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type>)
object type is: dummyOp
detected during:
instantiation of "thrust::detail::enable_if_non_const_reference_or_tuple_of_iterator_references<thrust::tuple_element<1, Tuple>::type>::type thrust::detail::unary_transform_functor<UnaryFunction>::operator()(Tuple) [with UnaryFunction=dummyOp, Tuple=thrust::detail::tuple_of_iterator_references<thrust::detail::tuple_of_iterator_references<double &, double &, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type>, double &, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type>]"
/usr/local/cuda/bin/../targets/x86_64-linux/include/thrust/detail/function.h(60): here
instantiation of "Result thrust::detail::wrapped_function<Function, Result>::operator()(const Argument &) const [with Function=thrust::detail::unary_transform_functor<dummyOp>, Result=void, Argument=thrust::detail::tuple_of_iterator_references<thrust::detail::tuple_of_iterator_references<thrust::device_reference<double>, thrust::device_reference<double>, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type>, thrust::device_reference<double>, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type>]"
/usr/local/cuda/bin/../targets/x86_64-linux/include/thrust/system/cuda/detail/for_each.inl(57): here
instantiation of "void thrust::system::cuda::detail::for_each_n_detail::for_each_kernel::operator()(thrust::system::cuda::detail::bulk_::parallel_group<thrust::system::cuda::detail::bulk_::concurrent_group<thrust::system::cuda::detail::bulk_::agent<1UL>, 0UL>, 0UL> &, Iterator, Function, Size) [with Iterator=thrust::zip_iterator<thrust::tuple<Zip, thrust::detail::normal_iterator<thrust::device_ptr<double>>, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type>>, Function=thrust::detail::wrapped_function<thrust::detail::unary_transform_functor<dummyOp>, void>, Size=unsigned int]"
/usr/local/cuda/bin/../targets/x86_64-linux/include/thrust/system/cuda/detail/bulk/detail/apply_from_tuple.hpp(71): here
instantiation of "void thrust::system::cuda::detail::bulk_::detail::apply_from_tuple(Function, const thrust::tuple<Arg1, Arg2, Arg3, Arg4, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type> &) [with Function=thrust::system::cuda::detail::for_each_n_detail::for_each_kernel, Arg1=thrust::system::cuda::detail::bulk_::parallel_group<thrust::system::cuda::detail::bulk_::concurrent_group<thrust::system::cuda::detail::bulk_::agent<1UL>, 0UL>, 0UL> &, Arg2=thrust::zip_iterator<thrust::tuple<Zip, thrust::detail::normal_iterator<thrust::device_ptr<double>>, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type>>, Arg3=thrust::detail::wrapped_function<thrust::detail::unary_transform_functor<dummyOp>, void>, Arg4=unsigned int]"
/usr/local/cuda/bin/../targets/x86_64-linux/include/thrust/system/cuda/detail/bulk/detail/closure.hpp(50): here
instantiation of "void thrust::system::cuda::detail::bulk_::detail::closure<Function, Tuple>::operator()() [with Function=thrust::system::cuda::detail::for_each_n_detail::for_each_kernel, Tuple=thrust::tuple<thrust::system::cuda::detail::bulk_::parallel_group<thrust::system::cuda::detail::bulk_::concurrent_group<thrust::system::cuda::detail::bulk_::agent<1UL>, 0UL>, 0UL> &, thrust::zip_iterator<thrust::tuple<Zip, thrust::detail::normal_iterator<thrust::device_ptr<double>>, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type>>, thrust::detail::wrapped_function<thrust::detail::unary_transform_functor<dummyOp>, void>, unsigned int, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type>]"
/usr/local/cuda/bin/../targets/x86_64-linux/include/thrust/system/cuda/detail/bulk/detail/cuda_task.hpp(58): here
[ 9 instantiation contexts not shown ]
instantiation of "RandomAccessIterator thrust::system::cuda::detail::for_each(thrust::system::cuda::detail::execution_policy<DerivedPolicy> &, RandomAccessIterator, RandomAccessIterator, UnaryFunction) [with DerivedPolicy=thrust::system::cuda::detail::tag, RandomAccessIterator=thrust::zip_iterator<thrust::tuple<Zip, thrust::detail::normal_iterator<thrust::device_ptr<double>>, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type>>, UnaryFunction=thrust::detail::unary_transform_functor<dummyOp>]"
/usr/local/cuda/bin/../targets/x86_64-linux/include/thrust/detail/for_each.inl(44): here
instantiation of "InputIterator thrust::for_each(const thrust::detail::execution_policy_base<DerivedPolicy> &, InputIterator, InputIterator, UnaryFunction) [with DerivedPolicy=thrust::system::cuda::detail::tag, InputIterator=thrust::zip_iterator<thrust::tuple<Zip, thrust::detail::normal_iterator<thrust::device_ptr<double>>, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type, thrust::null_type>>, UnaryFunction=thrust::detail::unary_transform_functor<dummyOp>]"
/usr/local/cuda/bin/../targets/x86_64-linux/include/thrust/system/detail/generic/transform.inl(57): here
instantiation of "OutputIterator thrust::system::detail::generic::transform(thrust::execution_policy<DerivedPolicy> &, InputIterator, InputIterator, OutputIterator, UnaryFunction) [with DerivedPolicy=thrust::system::cuda::detail::tag, InputIterator=Zip, OutputIterator=thrust::detail::normal_iterator<thrust::device_ptr<double>>, UnaryFunction=dummyOp]"
/usr/local/cuda/bin/../targets/x86_64-linux/include/thrust/detail/transform.inl(44): here
instantiation of "OutputIterator thrust::transform(const thrust::detail::execution_policy_base<DerivedPolicy> &, InputIterator, InputIterator, OutputIterator, UnaryFunction) [with DerivedPolicy=thrust::system::cuda::detail::tag, InputIterator=Zip, OutputIterator=thrust::detail::normal_iterator<thrust::device_ptr<double>>, UnaryFunction=dummyOp]"
/usr/local/cuda/bin/../targets/x86_64-linux/include/thrust/detail/transform.inl(142): here
instantiation of "OutputIterator thrust::transform(InputIterator, InputIterator, OutputIterator, UnaryFunction) [with InputIterator=Zip, OutputIterator=thrust::detail::normal_iterator<thrust::device_ptr<double>>, UnaryFunction=dummyOp]"
so2.cu(36): here
更新: 虽然我仍然非常希望帮助理解上面代码中的问题,但以下工作原理如下:
struct dummyOp {
template <typename Tuple>
__host__ __device__ double operator()(Tuple tup){
double result = thrust::get<0>(tup);
return result;
}
};
这个想法是从这里偷来的:https://github.com/thrust/thrust/blob/master/examples/arbitrary_transformation.cu
这不能解释我自己的问题,因为我仍然不明白原始代码中的类型有什么问题。
另外:它不能使参数成为显式引用(Tuple&amp;)。这是否意味着tup是按值传递的?
答案 0 :(得分:2)
一般来说,当一个zip迭代器作为推力算法的一部分被解引用时,它会创建一个基本类型的元组(即不是迭代器或指针)传递给有问题的仿函数。
当我们使用“违规行”分析代码中的编译器输出时:
argument types are: (thrust::detail::tuple_of_iterator_references<double &, double &,
我们观察到当推力解引用zippy
zip迭代器时,它会生成引用到double
项的元组。我们可以使用它来告诉我们编码器所期望的输入类型,即:
thrust::tuple<double &, double &>
由于这些是对基本类型的引用,因此我们不必在函子中取消引用这些(就像它们是指针或迭代器一样)来获取它们的值。
以下修改后的代码包含了这些想法并编译而没有错误:
$ cat t4.cu
#include <iostream>
#include <thrust/transform.h>
#include <thrust/functional.h>
#include <thrust/device_vector.h>
#include <thrust/device_ptr.h>
#include <thrust/iterator/zip_iterator.h>
#include <thrust/tuple.h>
typedef thrust::device_vector<double>::iterator realIter;
typedef thrust::tuple<realIter,realIter> Tup;
typedef thrust::zip_iterator<Tup> Zip;
typedef thrust::tuple<double &,double &> Tup2; //I tried replacing Tup with this in the functor
struct dummyOp : public thrust::unary_function<Tup2, double> {
__host__ __device__ double operator()(Tup2 &tup){
double result = thrust::get<0>(tup);
return result;
}
};
int main(){
thrust::device_vector<double> A(4);
thrust::device_vector<double> B(4);
thrust::device_vector<double> C(4);
A[0] = 1.; A[1] = 2.;
A[2] = 3.; A[3] = 4.;
B[0] = 4.; B[1] = 3.;
B[2] = 2.; B[3] = 1.;
Tup tup = thrust::tuple<realIter,realIter>(A.begin(),B.begin());
Zip zippy = thrust::zip_iterator<Tup>(tup);
dummyOp f;
// The following does not work:
thrust::transform(zippy, zippy + 4, C.begin(), f);
std::cout << "A:\n";
thrust::copy(A.begin(), A.end(), std::ostream_iterator<double>(std::cout, " "));
std::cout << "\nB:\n";
thrust::copy(B.begin(), B.end(), std::ostream_iterator<double>(std::cout, " "));
std::cout << "\nC:\n";
thrust::copy(C.begin(), C.end(), std::ostream_iterator<double>(std::cout, " "));
std::cout << std::endl;
std::cout <<"get<0>(zippy[0]) returns:\n" << thrust::get<0>(zippy[0]) << std::endl;
std::cout <<"get<1>(zippy[1]) returns:\n" << thrust::get<1>(zippy[1]) << std::endl;
return 0;
}
$ nvcc -arch=sm_61 -o t4 t4.cu
$ ./t4
A:
1 2 3 4
B:
4 3 2 1
C:
1 2 3 4
get<0>(zippy[0]) returns:
1
get<1>(zippy[1]) returns:
3
$
作为替代方案,在元组类型上模板化运算符操作符当然是有效的,因为编译器推导出必要的类型细节并实例化了函数运算符的适当版本。