CUDA 推力 zip_iterator 元组 transform_reduce
CUDA thrust zip_iterator tuple transform_reduce
我想计算 for vectors and , where denotes the magnitude of the vector 。由于这涉及对两个向量的每个对应分量之间的差的平方和求平方根,因此它应该是一个高度可并行化的任务。我在 Windows 10 上通过 Cygwin 使用 Cuda 和 Thrust。Cuda 和 Thrust 都在正常工作。
下面的代码编译并运行(使用 nvcc),但这只是因为我在 main
的底部注释掉了三行,我认为每一行都应该工作,但没有。 func::operator()(tup t)
认为我传递给它的参数实际上不是 tup 类型。
我还注释掉了运算符的实际主体,以使其至少更有可能编译。运算符应该找到输入 tup 的元素之间的平方差。 transform_reduce
的缩减 unary_op
(在本例中为 func()
)然后将这些相加,得到向量差的范数平方。
#include <iostream>
#include <stdlib.h>
#include <thrust/device_vector.h>
#include <thrust/transform.h>
#include <thrust/tuple.h>
#include <thrust/transform_reduce.h>
#include <thrust/iterator/zip_iterator.h>
typedef thrust::device_vector<float> dvec;
typedef dvec::iterator iter;
typedef thrust::tuple<iter, iter> tup;
struct func: public thrust::unary_function<tup, float>
{
__device__ float operator()(tup t) //difsq
{
// I've commented out these two lines for testing purposes:
// float f = thrust::get<0>(t) - thrust::get<1>(t);
// return f*f;
return 3.14;
}
};
int main()
{
dvec a(40, 4.f);
dvec b(40, 3.f);
auto begin = thrust::make_zip_iterator(thrust::make_tuple(a.begin(), b.begin()));
auto end = thrust::make_zip_iterator(thrust::make_tuple(a.end(), b.end()));
//these two lines work
thrust::get<0>(begin[0]);
std::cout << thrust::get<0>(begin[0]) - thrust::get<1>(begin[0]);
//these three lines do not
//thrust::transform_reduce(begin, end, func(), 0.0f, thrust::plus<float>());
//func()(begin[0]);
//thrust::transform(begin, end, begin, func());
std::cout << "done" << std::endl;
return 0;
}
我收到此错误:(我的程序名为 sandbox.cu
)
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\bin/../include\thrust/detail/tuple.inl(310): error: no instance of constructor "thrust::detail::normal_iterator<Pointer>::normal_iterator [with Pointer=thrust::device_ptr<float>]" matches the argument list
argument types are: (const thrust::device_reference<float>)
detected during:
instantiation of "thrust::detail::cons<HT, TT>::cons(const thrust::detail::cons<HT2, TT2> &) [with HT=iter, TT=thrust::detail::cons<iter, thrust::null_type>, HT2=thrust::device_reference<float>, TT2=thrust::detail::cons<thrust::device_reference<float>, thrust::null_type>]"
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\bin/../include\thrust/tuple.h(361): here
instantiation of "thrust::tuple<T0, T1, T2, T3, T4, T5, T6, T7, T8, T9>::tuple(const thrust::detail::cons<U1, U2> &) [with T0=iter, T1=iter, T2=thrust::null_type, T3=thrust::null_type, T4=thrust::null_type, T5=thrust::null_type, T6=thrust::null_type, T7=thrust::null_type, T8=thrust::null_type, T9=thrust::null_type, U1=thrust::device_reference<float>, U2=thrust::detail::cons<thrust::device_reference<float>, thrust::null_type>]"
sandbox.cu(37): here
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\bin/../include\thrust/detail/tuple.inl(411): error: no instance of constructor "thrust::detail::normal_iterator<Pointer>::normal_iterator [with Pointer=thrust::device_ptr<float>]" matches the argument list
argument types are: (const thrust::device_reference<float>)
detected during:
instantiation of "thrust::detail::cons<HT, thrust::null_type>::cons(const thrust::detail::cons<HT2, thrust::null_type> &) [with HT=iter, HT2=thrust::device_reference<float>]"
(310): here
instantiation of "thrust::detail::cons<HT, TT>::cons(const thrust::detail::cons<HT2, TT2> &) [with HT=iter, TT=thrust::detail::cons<iter, thrust::null_type>, HT2=thrust::device_reference<float>, TT2=thrust::detail::cons<thrust::device_reference<float>, thrust::null_type>]"
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\bin/../include\thrust/tuple.h(361): here
instantiation of "thrust::tuple<T0, T1, T2, T3, T4, T5, T6, T7, T8, T9>::tuple(const thrust::detail::cons<U1, U2> &) [with T0=iter, T1=iter, T2=thrust::null_type, T3=thrust::null_type, T4=thrust::null_type, T5=thrust::null_type, T6=thrust::null_type, T7=thrust::null_type, T8=thrust::null_type, T9=thrust::null_type, U1=thrust::device_reference<float>, U2=thrust::detail::cons<thrust::device_reference<float>, thrust::null_type>]"
sandbox.cu(37): here
2 errors detected in the compilation of "C:/cygwin64/tmp/tmpxft_00001a90_00000000-10_sandbox.cpp1.ii".
已解决! tup
应该是 thrust::tuple<float, float>
,而不是 thrust::tuple<iter, iter>
。完整解决方案:
#include <iostream>
#include <stdlib.h>
#include <thrust/device_vector.h>
#include <thrust/transform.h>
#include <thrust/tuple.h>
#include <thrust/transform_reduce.h>
#include <thrust/iterator/zip_iterator.h>
typedef thrust::device_vector<float> dvec;
typedef thrust::tuple<float, float> tup;
struct func
{
__device__ float operator()(tup t) //difsq
{
float f = thrust::get<0>(t) - thrust::get<1>(t);
return f*f;
}
};
int main()
{
dvec a(4, 3.f);
dvec b(4, 2.f);
auto begin = thrust::make_zip_iterator(thrust::make_tuple(a.begin(), b.begin()));
auto end = thrust::make_zip_iterator(thrust::make_tuple(a.end(), b.end()));
std::cout << thrust::transform_reduce(begin, end, func(), 0.0f, thrust::plus<float>()) << std::endl;
std::cout << "done" << std::endl;
return 0;
}
我想计算
下面的代码编译并运行(使用 nvcc),但这只是因为我在 main
的底部注释掉了三行,我认为每一行都应该工作,但没有。 func::operator()(tup t)
认为我传递给它的参数实际上不是 tup 类型。
我还注释掉了运算符的实际主体,以使其至少更有可能编译。运算符应该找到输入 tup 的元素之间的平方差。 transform_reduce
的缩减 unary_op
(在本例中为 func()
)然后将这些相加,得到向量差的范数平方。
#include <iostream>
#include <stdlib.h>
#include <thrust/device_vector.h>
#include <thrust/transform.h>
#include <thrust/tuple.h>
#include <thrust/transform_reduce.h>
#include <thrust/iterator/zip_iterator.h>
typedef thrust::device_vector<float> dvec;
typedef dvec::iterator iter;
typedef thrust::tuple<iter, iter> tup;
struct func: public thrust::unary_function<tup, float>
{
__device__ float operator()(tup t) //difsq
{
// I've commented out these two lines for testing purposes:
// float f = thrust::get<0>(t) - thrust::get<1>(t);
// return f*f;
return 3.14;
}
};
int main()
{
dvec a(40, 4.f);
dvec b(40, 3.f);
auto begin = thrust::make_zip_iterator(thrust::make_tuple(a.begin(), b.begin()));
auto end = thrust::make_zip_iterator(thrust::make_tuple(a.end(), b.end()));
//these two lines work
thrust::get<0>(begin[0]);
std::cout << thrust::get<0>(begin[0]) - thrust::get<1>(begin[0]);
//these three lines do not
//thrust::transform_reduce(begin, end, func(), 0.0f, thrust::plus<float>());
//func()(begin[0]);
//thrust::transform(begin, end, begin, func());
std::cout << "done" << std::endl;
return 0;
}
我收到此错误:(我的程序名为 sandbox.cu
)
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\bin/../include\thrust/detail/tuple.inl(310): error: no instance of constructor "thrust::detail::normal_iterator<Pointer>::normal_iterator [with Pointer=thrust::device_ptr<float>]" matches the argument list
argument types are: (const thrust::device_reference<float>)
detected during:
instantiation of "thrust::detail::cons<HT, TT>::cons(const thrust::detail::cons<HT2, TT2> &) [with HT=iter, TT=thrust::detail::cons<iter, thrust::null_type>, HT2=thrust::device_reference<float>, TT2=thrust::detail::cons<thrust::device_reference<float>, thrust::null_type>]"
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\bin/../include\thrust/tuple.h(361): here
instantiation of "thrust::tuple<T0, T1, T2, T3, T4, T5, T6, T7, T8, T9>::tuple(const thrust::detail::cons<U1, U2> &) [with T0=iter, T1=iter, T2=thrust::null_type, T3=thrust::null_type, T4=thrust::null_type, T5=thrust::null_type, T6=thrust::null_type, T7=thrust::null_type, T8=thrust::null_type, T9=thrust::null_type, U1=thrust::device_reference<float>, U2=thrust::detail::cons<thrust::device_reference<float>, thrust::null_type>]"
sandbox.cu(37): here
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\bin/../include\thrust/detail/tuple.inl(411): error: no instance of constructor "thrust::detail::normal_iterator<Pointer>::normal_iterator [with Pointer=thrust::device_ptr<float>]" matches the argument list
argument types are: (const thrust::device_reference<float>)
detected during:
instantiation of "thrust::detail::cons<HT, thrust::null_type>::cons(const thrust::detail::cons<HT2, thrust::null_type> &) [with HT=iter, HT2=thrust::device_reference<float>]"
(310): here
instantiation of "thrust::detail::cons<HT, TT>::cons(const thrust::detail::cons<HT2, TT2> &) [with HT=iter, TT=thrust::detail::cons<iter, thrust::null_type>, HT2=thrust::device_reference<float>, TT2=thrust::detail::cons<thrust::device_reference<float>, thrust::null_type>]"
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\bin/../include\thrust/tuple.h(361): here
instantiation of "thrust::tuple<T0, T1, T2, T3, T4, T5, T6, T7, T8, T9>::tuple(const thrust::detail::cons<U1, U2> &) [with T0=iter, T1=iter, T2=thrust::null_type, T3=thrust::null_type, T4=thrust::null_type, T5=thrust::null_type, T6=thrust::null_type, T7=thrust::null_type, T8=thrust::null_type, T9=thrust::null_type, U1=thrust::device_reference<float>, U2=thrust::detail::cons<thrust::device_reference<float>, thrust::null_type>]"
sandbox.cu(37): here
2 errors detected in the compilation of "C:/cygwin64/tmp/tmpxft_00001a90_00000000-10_sandbox.cpp1.ii".
已解决! tup
应该是 thrust::tuple<float, float>
,而不是 thrust::tuple<iter, iter>
。完整解决方案:
#include <iostream>
#include <stdlib.h>
#include <thrust/device_vector.h>
#include <thrust/transform.h>
#include <thrust/tuple.h>
#include <thrust/transform_reduce.h>
#include <thrust/iterator/zip_iterator.h>
typedef thrust::device_vector<float> dvec;
typedef thrust::tuple<float, float> tup;
struct func
{
__device__ float operator()(tup t) //difsq
{
float f = thrust::get<0>(t) - thrust::get<1>(t);
return f*f;
}
};
int main()
{
dvec a(4, 3.f);
dvec b(4, 2.f);
auto begin = thrust::make_zip_iterator(thrust::make_tuple(a.begin(), b.begin()));
auto end = thrust::make_zip_iterator(thrust::make_tuple(a.end(), b.end()));
std::cout << thrust::transform_reduce(begin, end, func(), 0.0f, thrust::plus<float>()) << std::endl;
std::cout << "done" << std::endl;
return 0;
}