我正在尝试编写一个在内核内部执行多个向量点积的代码。我正在使用cublas库中的cublasSdot函数来执行矢量点积。这是我的代码:
using namespace std;
__global__ void ker(float * a, float * c,long long result_size,int n, int m)
{
float *sum;
int id = blockIdx.x*blockDim.x+threadIdx.x;
float *out1,*out2;
int k;
if(id<result_size)
{
cublasHandle_t handle;
cublasCreate(&handle);
out1 = a + id*m;
for(k=0;k<n;k++)
{
out2 =a + k*m;
cublasSdot(handle, m,out1,1,out2,1,sum);
c[id*n + k]= *sum;
}
}
}
int main()
{
int n=70000,m=100;
long result_size=n;
result_size*=n;
float * dev_data,*dev_result;
float * data = new float [n*m];
float * result = new float [result_size];
for (int i = 0; i< n; i++)
for(int j = 0; j <m;j++)
{
data[i*m+j]=rand();
}
cudaMalloc ((void**)&dev_data,sizeof(float)*m*n);
cudaMalloc ((void**)&dev_result,sizeof(float)*result_size);
cudaMemcpy( dev_data, data, sizeof(float) * m* n, cudaMemcpyHostToDevice);
int block_size=1024;
int grid_size=ceil((float)result_size/(float)block_size);
ker<<<grid_size,block_size>>>(dev_data,dev_result,result_size,n,m);
cudaDeviceSynchronize();
cudaMemcpy(result, dev_result, sizeof(float)*(result_size), cudaMemcpyDeviceToHost);
return 0;
}
我已包含 cublas_v2 库并使用以下命令编译代码:
nvcc -lcublas_device -arch=sm_35 -rdc=true askstack.cu -o askstack
但我收到以下消息:
ptxas info : 'device-function-maxrregcount' is a BETA feature
谁能让我知道我应该怎么做?