579
社区成员
发帖
与我相关
我的任务
分享
cublasHandle_t cnpHandle;
cublasStatus_t status;
float Sigma_W[d*d];
for (int iters = 0; iters <numIter; iters ++)
{
if (iters/50 == 0)
{
cout<< iters <<endl;
}
status = cublasCreate(&cnpHandle);
if (status != CUBLAS_STATUS_SUCCESS)
{
//*returnValue = status;
cout<<"1"<<endl;
return false;
}
memset(Sigma_W,0,d*d*sizeof(float));
int k = 0;
for (int i = 0; i<d; i++)
{
for( int j=0; j<d; j++)
{
if (i==j)
{
Sigma_W[k] = Zd[i];
}
k++;
}
}
float b = 1;
status = cublasSgemm(cnpHandle, CUBLAS_OP_N, CUBLAS_OP_T, d, d, N, &E_v, X,d , X,d,&b,Sigma_W,d);
cublasDestroy(cnpHandle);
cudaThreadSynchronize();
status = cublasCreate(&cnpHandle);
if (status != CUBLAS_STATUS_SUCCESS)
{
cout<<"2"<<endl;
return false;
}
float *Sigma_W_inv = (float*)malloc(d*d);
int info;
const float *Sigma_Ws[] = {Sigma_W};
float *Sigma_W_invs[] = {Sigma_W_inv};
status = cublasSmatinvBatched(cnpHandle,
d,
Sigma_Ws,
d,
Sigma_W_invs,
d,
&info,
1
);
cublasDestroy(cnpHandle);
cudaThreadSynchronize();
status = cublasCreate(&cnpHandle);
if (status != CUBLAS_STATUS_SUCCESS)
{
//*returnValue = status;
cout<<"3"<<endl;
return false;
}
status = cublasScopy(cnpHandle, d*d, Sigma_W_inv,1,Sigma_W,1);
cublasDestroy(cnpHandle);
cudaThreadSynchronize();
float *E_W_before = (float*)malloc(d*sizeof(float));
status = cublasCreate(&cnpHandle);
if (status != CUBLAS_STATUS_SUCCESS)
{
cout<<"4"<<endl;
//*returnValue = status;
return false;
}
status = cublasScopy(cnpHandle, d, W, 1, E_W_before, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
cout<<"6"<<endl;
//*returnValue = status;
return false;
}
cublasDestroy(cnpHandle);
cudaThreadSynchronize();
status = cublasCreate(&cnpHandle);
if (status != CUBLAS_STATUS_SUCCESS)
{
cout<<"5"<<endl;
//*returnValue = status;
return false;
}
cublasDestroy(cnpHandle);
cudaThreadSynchronize();