简单的 Thrust 代码执行速度大约是我的原始 cuda 内核的一半。我使用推力错了吗?
Simple Thrust code performs about half as fast as my naive cuda kernel. Am I using Thrust wrong?
我对 Cuda 和 Thrust 很陌生,但我的印象是,如果使用得当,Thrust 应该会提供比单纯编写的 Cuda 内核更好的性能。我是否以次优方式使用 Thrust?下面是一个完整的最小示例,它采用长度为 N+2
的数组 u
,并且对于 1
和 N
之间的每个 i
计算平均值 0.5*(u[i-1] + u[i+1])
并将结果放入 uNew[i]
。 (uNew[0]
设置为 u[0]
,u[N+1]
设置为 u[N+1]
,这样边界项就不会改变)。该代码多次执行此平均操作以获得用于计时测试的合理时间。在我的硬件上,Thrust 计算花费的时间大约是原始代码的两倍。有没有办法改进我的 Thrust 代码?
#include <iostream>
#include <thrust/device_vector.h>
#include <boost/timer.hpp>
#include <thrust/device_malloc.h>
typedef double numtype;
template <typename T> class NeighborAverageFunctor{
int N;
public:
NeighborAverageFunctor(int _N){
N = _N;
}
template <typename Tuple>
__host__ __device__ void operator()(Tuple t){
T uL = thrust::get<0>(t);
T uR = thrust::get<1>(t);
thrust::get<2>(t) = 0.5*(uL + uR);
}
int getN(){
return N;
}
};
template <typename T> void thrust_sweep(thrust::device_ptr<T> u, thrust::device_ptr<T> uNew, NeighborAverageFunctor<T>& op){
int N = op.getN();
thrust::for_each(thrust::make_zip_iterator(thrust::make_tuple(u, u + 2, uNew + 1)), thrust::make_zip_iterator(thrust::make_tuple(u + N, u + N+2, uNew + N+1)), op);
// Propagate boundary values without changing them
uNew[0] = u[0];
uNew[N+1] = u[N+1];
}
template <typename T> __global__ void initialization_kernel(int n, T* u){
const int i = blockIdx.x * blockDim.x + threadIdx.x;
if(i < n+2){
if(i == 0){
u[i] = 1.0;
}
else{
u[i] = 0.0;
}
}
}
template <typename T> __global__ void sweep_kernel(int n, T, T* u, T* uNew){
const int i = blockDim.x * blockIdx.x + threadIdx.x;
if (i >= 1 && i < n-1){
uNew[i] = 0.5*(u[i+1] + u[i-1]);
}
else if(i == 0 || i == n+1){
uNew[i] = u[i];
}
}
int main(void){
int sweeps = 2000;
int N = 4096*2048;
numtype h = 1.0/N;
numtype hSquared = pow(h, 2);
NeighborAverageFunctor<numtype> op(N);
thrust::device_ptr<numtype> u_d = thrust::device_malloc<numtype>(N+2);
thrust::device_ptr<numtype> uNew_d = thrust::device_malloc<numtype>(N+2);
thrust::device_ptr<numtype> uTemp_d;
thrust::fill(u_d, u_d + (N+2), 0.0);
u_d[0] = 1.0;
boost::timer::timer timer1;
for(int k = 0; k < sweeps; k++){
thrust_sweep<numtype>(u_d, uNew_d, op);
uTemp_d = u_d;
u_d = uNew_d;
uNew_d = uTemp_d;
}
double thrust_time = timer1.elapsed();
thrust::host_vector<numtype> u_h(N+2);
thrust::copy(u_d, u_d + N+2, u_h.begin());
for(int i = 0; i < 10; i++){
std::cout << u_h[i] << " ";
}
std::cout << std::endl;
thrust::device_free(u_d);
thrust::device_free(uNew_d);
numtype * u_raw_d, * uNew_raw_d, * uTemp_raw_d;
cudaMalloc(&u_raw_d, (N+2)*sizeof(numtype));
cudaMalloc(&uNew_raw_d, (N+2)*sizeof(numtype));
numtype * u_raw_h = (numtype*)malloc((N+2)*sizeof(numtype));
int block_size = 256;
int grid_size = ((N+2) + block_size - 1) / block_size;
initialization_kernel<numtype><<<grid_size, block_size>>>(N, u_raw_d);
boost::timer::timer timer2;
for(int k = 0; k < sweeps; k++){
sweep_kernel<numtype><<<grid_size, block_size>>>(N+2, hSquared, u_raw_d, uNew_raw_d);
uTemp_raw_d = u_raw_d;
u_raw_d = uNew_raw_d;
uNew_raw_d = uTemp_raw_d;
}
double raw_time = timer2.elapsed();
cudaMemcpy(u_raw_h, u_raw_d, (N+2)*sizeof(numtype), cudaMemcpyDeviceToHost);
for(int i = 0; i < 10; i++){
std::cout << u_raw_h[i] << " ";
}
std::cout << std::endl;
std::cout << "Thrust: " << thrust_time << " s" << std::endl;
std::cout << "Raw: " << raw_time << " s" << std::endl;
free(u_raw_h);
cudaFree(u_raw_d);
cudaFree(uNew_raw_d);
return 0;
}
根据我的测试,这些行:
uNew[0] = u[0];
uNew[N+1] = u[N+1];
相对于内核方法,正在扼杀您的推力性能。当我消除它们时,结果似乎没有任何不同。与内核处理边界情况的方式相比,推力代码使用非常昂贵的方法(cudaMemcpy
操作,在幕后)执行边界处理。
由于推力函子实际上从未写入边界位置,因此只写入这些值一次就足够了,而不是循环写入。
您可以通过更好地处理边界情况来显着加快推力性能。
我对 Cuda 和 Thrust 很陌生,但我的印象是,如果使用得当,Thrust 应该会提供比单纯编写的 Cuda 内核更好的性能。我是否以次优方式使用 Thrust?下面是一个完整的最小示例,它采用长度为 N+2
的数组 u
,并且对于 1
和 N
之间的每个 i
计算平均值 0.5*(u[i-1] + u[i+1])
并将结果放入 uNew[i]
。 (uNew[0]
设置为 u[0]
,u[N+1]
设置为 u[N+1]
,这样边界项就不会改变)。该代码多次执行此平均操作以获得用于计时测试的合理时间。在我的硬件上,Thrust 计算花费的时间大约是原始代码的两倍。有没有办法改进我的 Thrust 代码?
#include <iostream>
#include <thrust/device_vector.h>
#include <boost/timer.hpp>
#include <thrust/device_malloc.h>
typedef double numtype;
template <typename T> class NeighborAverageFunctor{
int N;
public:
NeighborAverageFunctor(int _N){
N = _N;
}
template <typename Tuple>
__host__ __device__ void operator()(Tuple t){
T uL = thrust::get<0>(t);
T uR = thrust::get<1>(t);
thrust::get<2>(t) = 0.5*(uL + uR);
}
int getN(){
return N;
}
};
template <typename T> void thrust_sweep(thrust::device_ptr<T> u, thrust::device_ptr<T> uNew, NeighborAverageFunctor<T>& op){
int N = op.getN();
thrust::for_each(thrust::make_zip_iterator(thrust::make_tuple(u, u + 2, uNew + 1)), thrust::make_zip_iterator(thrust::make_tuple(u + N, u + N+2, uNew + N+1)), op);
// Propagate boundary values without changing them
uNew[0] = u[0];
uNew[N+1] = u[N+1];
}
template <typename T> __global__ void initialization_kernel(int n, T* u){
const int i = blockIdx.x * blockDim.x + threadIdx.x;
if(i < n+2){
if(i == 0){
u[i] = 1.0;
}
else{
u[i] = 0.0;
}
}
}
template <typename T> __global__ void sweep_kernel(int n, T, T* u, T* uNew){
const int i = blockDim.x * blockIdx.x + threadIdx.x;
if (i >= 1 && i < n-1){
uNew[i] = 0.5*(u[i+1] + u[i-1]);
}
else if(i == 0 || i == n+1){
uNew[i] = u[i];
}
}
int main(void){
int sweeps = 2000;
int N = 4096*2048;
numtype h = 1.0/N;
numtype hSquared = pow(h, 2);
NeighborAverageFunctor<numtype> op(N);
thrust::device_ptr<numtype> u_d = thrust::device_malloc<numtype>(N+2);
thrust::device_ptr<numtype> uNew_d = thrust::device_malloc<numtype>(N+2);
thrust::device_ptr<numtype> uTemp_d;
thrust::fill(u_d, u_d + (N+2), 0.0);
u_d[0] = 1.0;
boost::timer::timer timer1;
for(int k = 0; k < sweeps; k++){
thrust_sweep<numtype>(u_d, uNew_d, op);
uTemp_d = u_d;
u_d = uNew_d;
uNew_d = uTemp_d;
}
double thrust_time = timer1.elapsed();
thrust::host_vector<numtype> u_h(N+2);
thrust::copy(u_d, u_d + N+2, u_h.begin());
for(int i = 0; i < 10; i++){
std::cout << u_h[i] << " ";
}
std::cout << std::endl;
thrust::device_free(u_d);
thrust::device_free(uNew_d);
numtype * u_raw_d, * uNew_raw_d, * uTemp_raw_d;
cudaMalloc(&u_raw_d, (N+2)*sizeof(numtype));
cudaMalloc(&uNew_raw_d, (N+2)*sizeof(numtype));
numtype * u_raw_h = (numtype*)malloc((N+2)*sizeof(numtype));
int block_size = 256;
int grid_size = ((N+2) + block_size - 1) / block_size;
initialization_kernel<numtype><<<grid_size, block_size>>>(N, u_raw_d);
boost::timer::timer timer2;
for(int k = 0; k < sweeps; k++){
sweep_kernel<numtype><<<grid_size, block_size>>>(N+2, hSquared, u_raw_d, uNew_raw_d);
uTemp_raw_d = u_raw_d;
u_raw_d = uNew_raw_d;
uNew_raw_d = uTemp_raw_d;
}
double raw_time = timer2.elapsed();
cudaMemcpy(u_raw_h, u_raw_d, (N+2)*sizeof(numtype), cudaMemcpyDeviceToHost);
for(int i = 0; i < 10; i++){
std::cout << u_raw_h[i] << " ";
}
std::cout << std::endl;
std::cout << "Thrust: " << thrust_time << " s" << std::endl;
std::cout << "Raw: " << raw_time << " s" << std::endl;
free(u_raw_h);
cudaFree(u_raw_d);
cudaFree(uNew_raw_d);
return 0;
}
根据我的测试,这些行:
uNew[0] = u[0];
uNew[N+1] = u[N+1];
相对于内核方法,正在扼杀您的推力性能。当我消除它们时,结果似乎没有任何不同。与内核处理边界情况的方式相比,推力代码使用非常昂贵的方法(cudaMemcpy
操作,在幕后)执行边界处理。
由于推力函子实际上从未写入边界位置,因此只写入这些值一次就足够了,而不是循环写入。
您可以通过更好地处理边界情况来显着加快推力性能。