问题描述
我有两个空隙,它们的作用相同:向量与数字的乘积。一个是简单的cvoid,另一个是全局的。但是在测量时间之后,我发现常规c函数的运行速度比全局函数快得多。他们在这里:
preg_replace('/\s+/u',' ',$arr_return ['name']);
为什么会这样? 预先谢谢你。
解决方法
有很多缺陷:不使用pointerToU
,因为v
不能使用dim3 grid,threads
。无论如何,我知道有N
个块只有一个线程,因此您的内核无法受益于合并的内存访问,这可能是cuda版本比cpu版本慢的主要原因
试试
VectorOnNumber<<<N/32+1,32>>>(pointerToVector,10,pointerToVector);
,
这是我的代码: GPU内核:
void VectorOnNumber(double *vector1,double number,double *resultVector,int N)
{
dim3 grid(N/256+1),threads(256);
VectorOnNumber_K<<<grid,threads>>>(vector1,number,resultVector,N);
}
__global__
void VectorOnNumber_K(double *vector1,int N)
{
int tid = threadIdx.x + blockIdx.x*blockDim.x;
if(tid < N){
resultVector[tid] = vector1[tid]*number;
}
}
void VectorOnNumberf(float *vector1,float number,float *resultVector,threads(256);
VectorOnNumberf_K<<<grid,N);
}
__global__
void VectorOnNumberf_K(float *vector1,int N)
{
int tid = threadIdx.x + blockIdx.x*blockDim.x;
if(tid < N){
resultVector[tid] = vector1[tid]*number;
}
}
CPU:
void Stack_von(double *vec,double n,double *res,int N)
{
int i;
for(i = 0; i < N; i++){
res[i] = vec[i]*n;
}
}
void Stack_vonf(float *vec,float n,float *res,int N)
{
int i;
for(i = 0; i < N; i++){
res[i] = vec[i]*n;
}
}
全面测试:
void Stack()
{
int i,N;
double *x,*u,*dx,*du;
float *fx,*fu,*dfx,*dfu;
N=1000000;
x=new double[N];
u=new double[N];
fx=new float[N];
fu=new float[N];
for(i = 0; i < N; i++){
x[i] = i*i;
fx[i] = i*i;
}
// cpu
printf("start\n");
clock_t start = clock();
for(int k=0; k < 1000; k++) Stack_von(x,u,N);
clock_t end = clock();
float seconds = (float)(end - start) / CLOCKS_PER_SEC;
printf("host double %f ms\n",seconds);
start = clock();
for(int k=0; k < 1000; k++) Stack_vonf(fx,fu,N);
end = clock();
seconds = (float)(end - start) / CLOCKS_PER_SEC;
printf("host float %f ms\n",seconds); // 0.03 ms
// gpu
cudaMalloc(&dfx,N*sizeof(float));
cudaMalloc(&dfu,N*sizeof(float));
cudaMemcpy(dfx,fx,N*sizeof(float),cudaMemcpyHostToDevice);
cudaMalloc(&dx,N*sizeof(double));
cudaMalloc(&du,N*sizeof(double));
cudaMemcpy(dx,x,N*sizeof(double),cudaMemcpyHostToDevice);
cudaEvent_t dstart,dstop;
float elapsedTime;
cudaEventCreate(&dstart);
cudaEventCreate(&dstop);
cudaEventRecord(dstart,0);
VectorOnNumber(dx,du,N);
cudaEventRecord(dstop,0);
cudaEventSynchronize(dstop);
cudaEventElapsedTime(&elapsedTime,dstart,dstop);
printf("device double %f ms\n",elapsedTime);
cudaEventRecord(dstart,0);
VectorOnNumberf(dfx,dfu,dstop);
printf("device float %f ms\n",elapsedTime);
cudaFree(dx);
cudaFree(du);
cudaFree(dfx);
cudaFree(dfu);
delete [] x;
delete [] u;
delete [] fx;
delete [] fu;
}
结果 主机倍数:1.35ms浮点0.45ms 设备两倍0.067ms浮点0.037ms
设备(GTX1080)比主机(XEON 3.50GHz 8核)快10倍 我将N设置为10 ^ 6以使其可测量
如果每个块只有一个线程,则设备上的时间为1.37毫秒!