CUDA:较慢的推力：：在调用推力：：for\u之后减少
我正在尝试使用CUDA:较慢的推力：：在调用推力：：for\u之后减少,cuda,gpgpu,thrust,Cuda,Gpgpu,Thrust,我正在尝试使用推力和GK107[GeForce GTX 650]获取一个数字的总和。在内存上初始化设备向量之后，推力：：减少的执行时间显著增加，这让我感到困惑 以下是示例代码： #include <iostream> #include <stack> #include <ctime> #include <thrust/device_vector.h> #include <thrust/iterator/counting_iterator.h
推力GK107[GeForce GTX 650]设备向量之后，推力：：减少的执行时间显著增加，这让我感到困惑
以下是示例代码：
#include <iostream>
#include <stack>
#include <ctime>
#include <thrust/device_vector.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/transform.h>
#include <thrust/for_each.h>
#include <curand.h>
#include <curand_kernel.h>
struct tic_toc{
std::stack<clock_t> tictoc_stack;
inline void tic() { tictoc_stack.push(clock());}
inline void toc() {
std::cout << "Time elapsed: "
<< ((double)(clock() - tictoc_stack.top())) / CLOCKS_PER_SEC << "s"
<< std::endl;
tictoc_stack.pop();
}
};
struct curand_setup{
using init_tuple = thrust::tuple<int, curandState &>;
const unsigned long long seed;
curand_setup(unsigned long long _seed) : seed(_seed) {}
__device__ void operator()(init_tuple t){
curandState s;
int id = thrust::get<0>(t);
curand_init(seed, id, 0, &s);
thrust::get<1>(t) = s;
}
};
int main(int argc, char** argv){
int N = 1<<18;
std::cout << "N " << N << std::endl;
tic_toc tt;
thrust::device_vector<float> val(N,1);
tt.tic();
float mean=thrust::reduce(val.begin(),val.end(),0.f,thrust::plus<float>())/N;
tt.toc();
thrust::device_vector<curandState> rand_state(N);
auto rand_init_it = thrust::make_zip_iterator(
thrust::make_tuple(thrust::counting_iterator<int>(0),rand_state.begin()));
thrust::for_each_n(rand_init_it, N, curand_setup(0));
tt.tic();
mean=thrust::reduce(val.begin(),val.end(),0.f,thrust::plus<float>())/N;
tt.toc();
tt.tic();
mean=thrust::reduce(val.begin(),val.end(),0.f,thrust::plus<float>())/N;
tt.toc();
return 0;
}
当我为求和编写自己的内核时，或者我将数据复制到asch:：host_vector
并减少它们时，情况并没有改变
为什么在初始化推力：：设备_向量
之后，推力：：减少
的速度如此之慢，有没有办法避免这个问题？我将非常感谢你的帮助
我的系统是LinuxMint18.3
，内核4.15.0-23-generic
nvcc的输出--version
：
nvcc:NVIDIA（R）Cuda编译器驱动程序
版权所有（c）2005-2015英伟达公司
建于2015年8月11日星期二14:27:32
Cuda编译工具，7.5版，V7.5.17
为什么在初始化之后，struch:：reduce
的速度如此之慢
推力：设备矢量
事实并非如此。你困惑的根源是你的时间测量，这是不正确的
通常，在设备上运行的推力API调用在主机上是异步的。唯一的例外是返回值的调用（而推力：：减少就是其中之一）。因此，代码中的中间调用不仅测量了stress:：reduce
的执行时间，而且还测量了每个\u n
调用之前的stress:：for，而之前的调用要慢得多
您可以通过两种方式向自己确认这一点。如果您这样修改推力代码：
tt.tic();
float mean=thrust::reduce(val.begin(),val.end(),0.f,thrust::plus<float>())/N;
tt.toc();
thrust::device_vector<curandState> rand_state(N);
auto rand_init_it = thrust::make_zip_iterator(
thrust::make_tuple(thrust::counting_iterator<int>(0),rand_state.begin()));
thrust::for_each_n(rand_init_it, N, curand_setup(0));
cudaDeviceSynchronize(); // wait until for_each is complete
tt.tic();
mean=thrust::reduce(val.begin(),val.end(),0.f,thrust::plus<float>())/N;
tt.toc();
tt.tic();
mean=thrust::reduce(val.begin(),val.end(),0.f,thrust::plus<float>())/N;
tt.toc();
$ nvcc -arch=sm_52 -std=c++11 -o slow_thrust slow_thrust.cu
$ ./slow_thrust
N 262144
Time elapsed: 0.000471s
Time elapsed: 0.000621s
Time elapsed: 0.000448s
i、 e.使用cudaDeviceSynchronize（）
捕获前一个调用的运行时时，所有reduce调用的运行时大致相同。或者，您可以在原始代码上使用分析工具，例如：
$ nvprof --print-gpu-trace ./slow_thrust
N 262144
==7870== NVPROF is profiling process 7870, command: ./slow_thrust
Time elapsed: 0.000521s
Time elapsed: 0.06983s
Time elapsed: 0.000538s
==7870== Profiling application: ./slow_thrust
==7870== Profiling result:
Start Duration Grid Size Block Size Regs* SSMem* DSMem* Size Throughput SrcMemType DstMemType Device Context Stream Name
214.30ms 7.6800us (512 1 1) (256 1 1) 8 0B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::core::_kernel_agent<thrust::cuda_cub::__parallel_for::ParallelForAgent<thrust::cuda_cub::__uninitialized_fill::functor<thrust::device_ptr<float>, float>, unsigned long>, thrust::cuda_cub::__uninitialized_fill::functor<thrust::device_ptr<float>, float>, unsigned long>(thrust::device_ptr<float>, float) [109]
214.56ms 5.8550us (52 1 1) (256 1 1) 29 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, thrust::detail::normal_iterator<thrust::device_ptr<float>>, float*, int, thrust::plus<float>>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::GridEvenShare<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600) [128]
214.58ms 2.7200us (1 1 1) (256 1 1) 27 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceSingleTileKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*, thrust::detail::normal_iterator<thrust::pointer<float, thrust::cuda_cub::tag, thrust::use_default, thrust::use_default>>, int, thrust::plus<float>, float>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*) [136]
214.60ms 1.1840us - - - - - 4B 3.2219MB/s Device Pageable GeForce GTX 970 1 7 [CUDA memcpy DtoH]
214.98ms 221.27us (512 1 1) (256 1 1) 20 0B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::core::_kernel_agent<thrust::cuda_cub::__parallel_for::ParallelForAgent<thrust::cuda_cub::__uninitialized_fill::functor<thrust::device_ptr<curandStateXORWOW>, curandStateXORWOW>, unsigned long>, thrust::cuda_cub::__uninitialized_fill::functor<thrust::device_ptr<curandStateXORWOW>, curandStateXORWOW>, unsigned long>(thrust::device_ptr<curandStateXORWOW>, curandStateXORWOW) [151]
219.51ms 69.492ms (512 1 1) (256 1 1) 108 0B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::core::_kernel_agent<thrust::cuda_cub::__parallel_for::ParallelForAgent<thrust::cuda_cub::for_each_f<thrust::zip_iterator<thrust::tuple<thrust::counting_iterator<int, thrust::use_default, thrust::use_default, thrust::use_default>, thrust::detail::normal_iterator<thrust::device_ptr<curandStateXORWOW>>, 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<curand_setup, void>>, int>, thrust::cuda_cub::for_each_f<thrust::zip_iterator<thrust::tuple<thrust::counting_iterator<int, thrust::use_default, thrust::use_default, thrust::use_default>, thrust::detail::normal_iterator<thrust::device_ptr<curandStateXORWOW>>, 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<curand_setup, void>>, int>(thrust::use_default, thrust::use_default) [160]
289.00ms 9.5360us (52 1 1) (256 1 1) 29 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, thrust::detail::normal_iterator<thrust::device_ptr<float>>, float*, int, thrust::plus<float>>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::GridEvenShare<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600) [179]
289.01ms 3.4880us (1 1 1) (256 1 1) 27 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceSingleTileKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*, thrust::detail::normal_iterator<thrust::pointer<float, thrust::cuda_cub::tag, thrust::use_default, thrust::use_default>>, int, thrust::plus<float>, float>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*) [187]
289.07ms 1.3120us - - - - - 4B 2.9075MB/s Device Pageable GeForce GTX 970 1 7 [CUDA memcpy DtoH]
289.66ms 9.9520us (52 1 1) (256 1 1) 29 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, thrust::detail::normal_iterator<thrust::device_ptr<float>>, float*, int, thrust::plus<float>>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::GridEvenShare<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600) [211]
289.68ms 3.3280us (1 1 1) (256 1 1) 27 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceSingleTileKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*, thrust::detail::normal_iterator<thrust::pointer<float, thrust::cuda_cub::tag, thrust::use_default, thrust::use_default>>, int, thrust::plus<float>, float>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*) [219]
289.69ms 1.3120us - - - - - 4B 2.9075MB/s Device Pageable GeForce GTX 970 1 7 [CUDA memcpy DtoH]
$nvprof——打印gpu跟踪。/slow\u
N 262144
==7870==NVPROF正在分析过程7870，命令：./slow\u推力
所用时间：0.000521s
所用时间：0.06983s
所用时间：0.000538s
==7870==仿形应用：./slow\u推力
==7870==分析结果：
开始持续时间网格大小块大小Regs*SSMem*DSMem*大小吞吐量SrcMemType DstMemType设备上下文流名称
214.30ms 7.6800us（512 11）（256 11）8 0B 0B--GeForce GTX 970 1 7无效推力：：cuda_cub：：核心：：_内核_代理（推力：：设备_ptr，浮点）[109]
214.56ms 5.8550us（52 1 1）（256 1 1）29 44B 0B-----GeForce GTX 970 1 7无效推力：：cuda_cub：：cub：：DeviceReduceKernel（int，float，推力：：plus，推力：：cuda_cub：：GridEvenShare，推力：：cuda_cub：：cub：：DeviceReducePolicy：：policy 600）[128]
214.58ms 2.7200us（11）（256 11）27 44B 0B----GeForce GTX 970 1 7无效推力：：cuda_cub：：cub：：DeviceReduceSingleTileKernel（int，float，推力：：plus，推力：：cuda_cub：：cub：：DeviceReducePolicy：：Policy 600，float*）[136]
214.60ms 1.1840us----4B 3.2219MB/s设备可分页GeForce GTX 970 1 7[CUDA memcpy DtoH]
214.98ms 221.27us（512 1 1）（256 1 1 1）20 0B 0B--GeForce GTX 970 1 7无效推力：：cuda_cub：：核心：：_内核_代理（推力：：设备_ptr，CurandStateXorow）[151]
219.51ms 69.492ms（512 11）（256 11）108 0B 0B-----GeForce GTX 970 1 7无效推力：：cuda_cub：：内核：：内核代理（推力：：使用默认值，推力：：使用默认值）[160]
289.00ms 9.5360us（52 1 1）（256 1 1）29 44B 0B-----GeForce GTX 970 1 7无效推力：：cuda_cub：：cub：：DeviceReduceKernel（int，float，推力：：plus，推力：：cuda_cub：：GridEvenShare，推力：：cuda_cub：：cub：：DeviceReducePolicy：：policy 600）[179
289.01ms 3.4880us（11）（256 11）27 44B 0B----GeForce GTX 970 1 7无效推力：：cuda_cub：：cub：：DeviceDuceSingleTilekernel（int，float，推力：：plus，推力：：cuda_cub：：cub：：DeviceDucePolicy：：Policy 600，float*）[187]
289.07ms 1.3120us----4B 2.9075MB/s设备可分页GeForce GTX 970 1 7[CUDA memcpy DtoH]
289.66ms 9.9520us（52 1 1）（256 1 1）29 44B 0B----GeForce GTX 970 1 7无效推力：：cuda_cub：：cub：：DeviceReduceKernel（int，float，推力：：plus，推力：：cuda_cub：：GridEvenShare，推力：：cuda_cub：：cub：：DeviceReducePolicy：：policy 600）[211
289.68ms 3.3280us（1 1 1）（256 1 1）27 44B 0B----GeForce GTX 970 1 7无效推力：：cuda_cub：：cub：：DeviceDuceSingleTilekernel（int，float，推力：：plus，推力：：cuda_cub：：cub：：DeviceDucePolicy：：Policy 600，float*）[219]
289.69ms 1.3120us----4B 2.9075MB/s设备可分页GeForce GTX 970 1 7[CUDA memcpy DtoH]
在这里，您可以看到，组成reduce操作的三个调用每次累计需要8-13微秒，而每个调用的需要69毫秒才能完成。现在我知道发生了什么，谢谢
$ nvprof --print-gpu-trace ./slow_thrust
N 262144
==7870== NVPROF is profiling process 7870, command: ./slow_thrust
Time elapsed: 0.000521s
Time elapsed: 0.06983s
Time elapsed: 0.000538s
==7870== Profiling application: ./slow_thrust
==7870== Profiling result:
Start Duration Grid Size Block Size Regs* SSMem* DSMem* Size Throughput SrcMemType DstMemType Device Context Stream Name
214.30ms 7.6800us (512 1 1) (256 1 1) 8 0B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::core::_kernel_agent<thrust::cuda_cub::__parallel_for::ParallelForAgent<thrust::cuda_cub::__uninitialized_fill::functor<thrust::device_ptr<float>, float>, unsigned long>, thrust::cuda_cub::__uninitialized_fill::functor<thrust::device_ptr<float>, float>, unsigned long>(thrust::device_ptr<float>, float) [109]
214.56ms 5.8550us (52 1 1) (256 1 1) 29 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, thrust::detail::normal_iterator<thrust::device_ptr<float>>, float*, int, thrust::plus<float>>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::GridEvenShare<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600) [128]
214.58ms 2.7200us (1 1 1) (256 1 1) 27 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceSingleTileKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*, thrust::detail::normal_iterator<thrust::pointer<float, thrust::cuda_cub::tag, thrust::use_default, thrust::use_default>>, int, thrust::plus<float>, float>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*) [136]
214.60ms 1.1840us - - - - - 4B 3.2219MB/s Device Pageable GeForce GTX 970 1 7 [CUDA memcpy DtoH]
214.98ms 221.27us (512 1 1) (256 1 1) 20 0B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::core::_kernel_agent<thrust::cuda_cub::__parallel_for::ParallelForAgent<thrust::cuda_cub::__uninitialized_fill::functor<thrust::device_ptr<curandStateXORWOW>, curandStateXORWOW>, unsigned long>, thrust::cuda_cub::__uninitialized_fill::functor<thrust::device_ptr<curandStateXORWOW>, curandStateXORWOW>, unsigned long>(thrust::device_ptr<curandStateXORWOW>, curandStateXORWOW) [151]
219.51ms 69.492ms (512 1 1) (256 1 1) 108 0B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::core::_kernel_agent<thrust::cuda_cub::__parallel_for::ParallelForAgent<thrust::cuda_cub::for_each_f<thrust::zip_iterator<thrust::tuple<thrust::counting_iterator<int, thrust::use_default, thrust::use_default, thrust::use_default>, thrust::detail::normal_iterator<thrust::device_ptr<curandStateXORWOW>>, 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<curand_setup, void>>, int>, thrust::cuda_cub::for_each_f<thrust::zip_iterator<thrust::tuple<thrust::counting_iterator<int, thrust::use_default, thrust::use_default, thrust::use_default>, thrust::detail::normal_iterator<thrust::device_ptr<curandStateXORWOW>>, 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<curand_setup, void>>, int>(thrust::use_default, thrust::use_default) [160]
289.00ms 9.5360us (52 1 1) (256 1 1) 29 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, thrust::detail::normal_iterator<thrust::device_ptr<float>>, float*, int, thrust::plus<float>>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::GridEvenShare<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600) [179]
289.01ms 3.4880us (1 1 1) (256 1 1) 27 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceSingleTileKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*, thrust::detail::normal_iterator<thrust::pointer<float, thrust::cuda_cub::tag, thrust::use_default, thrust::use_default>>, int, thrust::plus<float>, float>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*) [187]
289.07ms 1.3120us - - - - - 4B 2.9075MB/s Device Pageable GeForce GTX 970 1 7 [CUDA memcpy DtoH]
289.66ms 9.9520us (52 1 1) (256 1 1) 29 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, thrust::detail::normal_iterator<thrust::device_ptr<float>>, float*, int, thrust::plus<float>>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::GridEvenShare<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600) [211]
289.68ms 3.3280us (1 1 1) (256 1 1) 27 44B 0B - - - - GeForce GTX 970 1 7 void thrust::cuda_cub::cub::DeviceReduceSingleTileKernel<thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*, thrust::detail::normal_iterator<thrust::pointer<float, thrust::cuda_cub::tag, thrust::use_default, thrust::use_default>>, int, thrust::plus<float>, float>(int, float, thrust::plus<float>, thrust::cuda_cub::cub::DeviceReducePolicy<float, int, thrust::plus<float>>::Policy600, float*) [219]
289.69ms 1.3120us - - - - - 4B 2.9075MB/s Device Pageable GeForce GTX 970 1 7 [CUDA memcpy DtoH]