建设银行网站打不开别的网站可以wordpress更新后台反应慢

接上篇文章，可以发现使用CUDA提供的API进行前缀和扫描时，第一次运行的时间不如共享内存访问，猜测是使用到了全局内存。
首先看调用逻辑：

thrust::inclusive_scan(thrust::device, d_x, d_x + N, d_x);

第一个参数指定了设备，根据实参数量和类型找到对应的函数，是scan.h中的如下函数：

template <typename DerivedPolicy, typename InputIterator, typename OutputIterator>
_CCCL_HOST_DEVICE OutputIterator inclusive_scan(const thrust::detail::execution_policy_base<DerivedPolicy>& exec,InputIterator first,InputIterator last,OutputIterator result);

其实现位于thrust\thrust\system\cuda\detail\scan.h
注意：路径可能与实际有偏差，可以在/usr/local/下使用find . -name xx查找对应的文件

template <typename Derived, typename InputIt, typename OutputIt>
_CCCL_HOST_DEVICE OutputIt
inclusive_scan(thrust::cuda_cub::execution_policy<Derived>& policy, InputIt first, InputIt last, OutputIt result)
{return thrust::cuda_cub::inclusive_scan(policy, first, last, result, thrust::plus<>{});
}

将操作指定为plus，
然后执行同一文件下的此函数：

template <typename Derived, typename InputIt, typename OutputIt, typename ScanOp>
_CCCL_HOST_DEVICE OutputIt inclusive_scan(thrust::cuda_cub::execution_policy<Derived>& policy, InputIt first, InputIt last, OutputIt result, ScanOp scan_op)
{using diff_t           = typename thrust::iterator_traits<InputIt>::difference_type;diff_t const num_items = thrust::distance(first, last);return thrust::cuda_cub::inclusive_scan_n(policy, first, num_items, result, scan_op);
}

最终找到主要的执行逻辑：

_CCCL_EXEC_CHECK_DISABLE
template <typename Derived, typename InputIt, typename Size, typename OutputIt, typename ScanOp>
_CCCL_HOST_DEVICE OutputIt inclusive_scan_n_impl(thrust::cuda_cub::execution_policy<Derived>& policy, InputIt first, Size num_items, OutputIt result, ScanOp scan_op)
{using AccumT     = typename thrust::iterator_traits<InputIt>::value_type;using Dispatch32 = cub::DispatchScan<InputIt, OutputIt, ScanOp, cub::NullType, std::int32_t, AccumT>;using Dispatch64 = cub::DispatchScan<InputIt, OutputIt, ScanOp, cub::NullType, std::int64_t, AccumT>;cudaStream_t stream = thrust::cuda_cub::stream(policy);cudaError_t status;// Determine temporary storage requirements:size_t tmp_size = 0;{THRUST_INDEX_TYPE_DISPATCH2(status,Dispatch32::Dispatch,Dispatch64::Dispatch,num_items,(nullptr, tmp_size, first, result, scan_op, cub::NullType{}, num_items_fixed, stream));thrust::cuda_cub::throw_on_error(status,"after determining tmp storage ""requirements for inclusive_scan");}// Run scan:{// Allocate temporary storage:thrust::detail::temporary_array<std::uint8_t, Derived> tmp{policy, tmp_size};THRUST_INDEX_TYPE_DISPATCH2(status,Dispatch32::Dispatch,Dispatch64::Dispatch,num_items,(tmp.data().get(), tmp_size, first, result, scan_op, cub::NullType{}, num_items_fixed, stream));thrust::cuda_cub::throw_on_error(status, "after dispatching inclusive_scan kernel");thrust::cuda_cub::throw_on_error(thrust::cuda_cub::synchronize_optional(policy), "inclusive_scan failed to synchronize");}return result + num_items;
}

可以看到，此处thrust调用了cub的Dispatchscan操作，而cub中是使用全局内存的，因此造成了效率还不如手动编写使用共享内存的算法。