[英]Using cuda thrust with arrays instead vectors to inclusive_scan
I have a code given by @ms: 我有一个@ms给出的代码:
#include <thrust/device_vector.h>
#include <thrust/scan.h>
#include <thrust/iterator/transform_iterator.h>
#include <thrust/iterator/counting_iterator.h>
#include <iostream>
struct omit_negative : public thrust::unary_function<int, int>
{
__host__ __device__
int operator()(int value)
{
if (value<0)
{
value = 0;
}
return value;
}
};
int main()
{
int array[] = {2,1,-1,3,-1,2};
const int array_size = sizeof(array)/sizeof(array[0]);
thrust::device_vector<int> d_array(array, array + array_size);
thrust::device_vector<int> d_result(array_size);
std::cout << "input data" << std::endl;
thrust::copy(d_array.begin(), d_array.end(), std::ostream_iterator<int>(std::cout, " "));
thrust::inclusive_scan(thrust::make_transform_iterator(d_array.begin(), omit_negative()),
thrust::make_transform_iterator(d_array.end(), omit_negative()),
d_result.begin());
std::cout << std::endl << "after inclusive_scan" << std::endl;
thrust::copy(d_result.begin(), d_result.end(), std::ostream_iterator<int>(std::cout, " "));
using namespace thrust::placeholders;
thrust::scatter_if(d_array.begin(),
d_array.end(),
thrust::make_counting_iterator(0),
d_array.begin(),
d_result.begin(),
_1<0
);
std::cout << std::endl << "after scatter_if" << std::endl;
thrust::copy(d_result.begin(), d_result.end(), std::ostream_iterator<int>(std::cout, " "));
std::cout << std::endl;
}
It refers to previous question . 它指的是先前的问题 。
I didn't know about thrust, but now I guess I'm going to quit idea of writing own code. 我不知道推力,但是现在我想我会放弃编写自己的代码的想法。 I'd rather use thrust.
我宁愿使用推力。 I modified my algorithm: instead -1 there are 0's (so make_transform is not necessary).
我修改了算法:取而代之的是-1,而是0(因此不必make_transform)。 Also your example creates array on host.
同样,您的示例在主机上创建数组。 But actually I have prepared array stored on device, and I' like to use it (instead of vectors) to avoid creating redundant memory and to avoid copying memory (it costs time - minimal time cost is my goal).
但是实际上我已经准备好了存储在设备上的数组,并且我想使用它(而不是向量)来避免创建冗余内存并避免复制内存(这会花费时间-我的目标是花费最少的时间)。 I'm not sure how to use arrays instead of vectors.
我不确定如何使用数组而不是向量。 Here is what I've written:
这是我写的:
int* dev_l_set = 0;
cudaMalloc((void**)&dev_l_set, actualVerticesRowCount * sizeof(int));
...prepare array in kernel...
thrust::device_vector<int> d_result(actualVerticesRowCount);
thrust::inclusive_scan(dev_l_set, dev_l_set + actualVerticesRowCount, dev_l_set);
using namespace thrust::placeholders;
thrust::scatter_if(dev_l_set, dev_l_set + actualVerticesRowCount, thrust::make_counting_iterator(0), dev_l_set, d_result.begin(), _1 <= 0);
cudaFree(dev_l_set);
dev_l_set = thrust::raw_pointer_cast(d_result.data());
I can't cast from device_vector to int*, but I'd like to store result of scanning in initial dev_l_set
array. 我无法从device_vector转换为int *,但是我想将扫描结果存储在初始
dev_l_set
数组中。 Also it'd be great to do it in place, is it necessary to use d_result
in scatter_if? 此外,最好在适当的位置执行此操作,是否有必要在
d_result
中使用d_result?
Actual Input (stored on int* - device side): (example) 实际输入(存储在int *-设备端):(示例)
dev_l_set[0] = 0
dev_l_set[1] = 2
dev_l_set[2] = 0
dev_l_set[3] = 3
dev_l_set[4] = 0
dev_l_set[5] = 1
Desired output to the above input: 所需的输出到上述输入:
dev_l_set[0] = 0
dev_l_set[1] = 2
dev_l_set[2] = 0
dev_l_set[3] = 5
dev_l_set[4] = 0
dev_l_set[5] = 6
dev_l_set
should store input, then do scan in place and in the end it should store output. dev_l_set
应存储输入,然后进行扫描,最后应存储输出。
It could be something like this. 可能是这样的。
int* dev_l_set = 0;
cudaMalloc((void**)&dev_l_set, actualVerticesRowCount * sizeof(int));
...prepare array in kernel... (see input data)
thrust::inclusive_scan(dev_l_set, dev_l_set + actualVerticesRowCount, dev_l_set);
using namespace thrust::placeholders;
thrust::scatter_if(dev_l_set, dev_l_set + actualVerticesRowCount, thrust::make_counting_iterator(0), dev_l_set, dev_l_set, _1 <= 0);
My Cuda version (minimal that app should work) is 5.5 (Tesla M2070) and unfortunately I can't use c++11. 我的Cuda版本(最低应能运行该应用程序)是5.5(Tesla M2070),不幸的是我无法使用c ++ 11。
You can do the inclusive scan as well as the scatter step in place without an additional result vector. 您可以进行包含性扫描以及分散步骤,而无需其他结果向量。
The following example directly uses the data from a raw device pointer without thrust::device_vector
. 以下示例直接使用原始设备指针中的数据,而没有
thrust::device_vector
。 After the inclusive scan, the previously 0
elements are restored. 包含性扫描之后,将还原先前的
0
元素。
As @JaredHoberock pointed out, one should not rely on code residing in thrust::detail
. 正如@JaredHoberock指出的那样,不应依赖于
thrust::detail
驻留的代码。 I therefore edited my answer and copied part of the code from thrust::detail::head_flags
directly into this example. 因此,我编辑了我的答案并将部分代码从
thrust::detail::head_flags
直接复制到此示例中。
#include <thrust/scan.h>
#include <thrust/scatter.h>
#include <thrust/device_ptr.h>
#include <thrust/iterator/constant_iterator.h>
#include <iostream>
// the following code is copied from <thrust/detail/range/head_flags.h>
#include <thrust/detail/config.h>
#include <thrust/iterator/transform_iterator.h>
#include <thrust/iterator/zip_iterator.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/tuple.h>
#include <thrust/functional.h>
template<typename RandomAccessIterator,
typename BinaryPredicate = thrust::equal_to<typename thrust::iterator_value<RandomAccessIterator>::type>,
typename ValueType = bool,
typename IndexType = typename thrust::iterator_difference<RandomAccessIterator>::type>
class head_flags
{
public:
struct head_flag_functor
{
BinaryPredicate binary_pred; // this must be the first member for performance reasons
IndexType n;
typedef ValueType result_type;
__host__ __device__
head_flag_functor(IndexType n)
: binary_pred(), n(n)
{}
__host__ __device__
head_flag_functor(IndexType n, BinaryPredicate binary_pred)
: binary_pred(binary_pred), n(n)
{}
template<typename Tuple>
__host__ __device__ __thrust_forceinline__
result_type operator()(const Tuple &t)
{
const IndexType i = thrust::get<0>(t);
// note that we do not dereference the tuple's 2nd element when i <= 0
// and therefore do not dereference a bad location at the boundary
return (i == 0 || !binary_pred(thrust::get<1>(t), thrust::get<2>(t)));
}
};
typedef thrust::counting_iterator<IndexType> counting_iterator;
public:
typedef thrust::transform_iterator<
head_flag_functor,
thrust::zip_iterator<thrust::tuple<counting_iterator,RandomAccessIterator,RandomAccessIterator> >
> iterator;
__host__ __device__
head_flags(RandomAccessIterator first, RandomAccessIterator last)
: m_begin(thrust::make_transform_iterator(thrust::make_zip_iterator(thrust::make_tuple(thrust::counting_iterator<IndexType>(0), first, first - 1)),
head_flag_functor(last - first))),
m_end(m_begin + (last - first))
{}
__host__ __device__
head_flags(RandomAccessIterator first, RandomAccessIterator last, BinaryPredicate binary_pred)
: m_begin(thrust::make_transform_iterator(thrust::make_zip_iterator(thrust::make_tuple(thrust::counting_iterator<IndexType>(0), first, first - 1)),
head_flag_functor(last - first, binary_pred))),
m_end(m_begin + (last - first))
{}
__host__ __device__
iterator begin() const
{
return m_begin;
}
__host__ __device__
iterator end() const
{
return m_end;
}
template<typename OtherIndex>
__host__ __device__
typename iterator::reference operator[](OtherIndex i)
{
return *(begin() + i);
}
private:
iterator m_begin, m_end;
};
template<typename RandomAccessIterator>
__host__ __device__
head_flags<RandomAccessIterator>
make_head_flags(RandomAccessIterator first, RandomAccessIterator last)
{
return head_flags<RandomAccessIterator>(first, last);
}
int main()
{
// copy data to device, this will be produced by your kernel
int array[] = {0,2,0,3,0,1};
const int array_size = sizeof(array)/sizeof(array[0]);
int* dev_l_set;
cudaMalloc((void**)&dev_l_set, array_size * sizeof(int));
cudaMemcpy(dev_l_set, array, array_size * sizeof(int), cudaMemcpyHostToDevice);
// wrap raw pointer in a thrust::device_ptr so thrust knows that this memory is located on the GPU
thrust::device_ptr<int> dev_ptr = thrust::device_pointer_cast(dev_l_set);
thrust::inclusive_scan(dev_ptr,
dev_ptr+array_size,
dev_ptr);
// copy result back to host for printing
cudaMemcpy(array, dev_l_set, array_size * sizeof(int), cudaMemcpyDeviceToHost);
std::cout << "after inclusive_scan" << std::endl;
thrust::copy(array, array+array_size, std::ostream_iterator<int>(std::cout, " "));
std::cout << std::endl;
using namespace thrust::placeholders;
thrust::scatter_if(thrust::make_constant_iterator(0),
thrust::make_constant_iterator(0)+array_size,
thrust::make_counting_iterator(0),
make_head_flags(dev_ptr, dev_ptr+array_size).begin(),
dev_ptr,
!_1
);
// copy result back to host for printing
cudaMemcpy(array, dev_l_set, array_size * sizeof(int), cudaMemcpyDeviceToHost);
std::cout << "after scatter_if" << std::endl;
thrust::copy(array, array+array_size, std::ostream_iterator<int>(std::cout, " "));
std::cout << std::endl;
}
output 产量
after inclusive_scan
0 2 2 5 5 6
after scatter_if
0 2 0 5 0 6
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.