[CUBLAS] How to use cublasCgerc_v2?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello dear CUDA community,

I am quiet new to CUDA programming, but I a making progress. However, I tried to include the CUBLAS function cublasCgerc_v2 into my program, but the results are more or less total rubbish.

Assume a matrix A with zeros only, a vector x = [1 … 8] and a vector y = [1 … 7]. Everything is complex valued, but with imaginary part 0.

Then, A should be updated with

A = A + x*y**H

Updating the matrix A with cublasCgerc_v2 should give me a matrix:

1     2     3     4     5     6     7

2     4     6     8    10    12    14

3     6     9    12    15    18    21

4     8    12    16    20    24    28

5    10    15    20    25    30    35

6    12    18    24    30    36    42

7    14    21    28    35    42    49

8    16    24    32    40    48    56

However, I a not able to get there.

An excerpt of my code is:

size_t M = 8;

    size_t N = 7;

CudaAECComplex alpha;

    alpha.x = 1.0;

    alpha.y = 0.0;

CudaAECComplex *x,*y,*A;

    CudaAECComplex *dx,*dy,*dA;

x = new CudaAECComplex[M];

    y = new CudaAECComplex[N];

    A = new CudaAECComplex[M*N];

cudaMalloc((void**)&dx, M*sizeof(CudaAECComplex));

    cudaMalloc((void**)&dy, N*sizeof(CudaAECComplex));

    cudaMalloc((void**)&dA, M*N*sizeof(CudaAECComplex));

for(unsigned int i=0;i<M;++i) {

        x[i].x = i+1;

        x[i].y = 0;

    }

    for(unsigned int i=0;i<N;++i) {

        y[i].x = i+1;

        y[i].y = 0;

    }

    for(unsigned int i=0;i<M*N;++i) {

        A[i].x = 0;

        A[i].y = 0;

    }

cudaMemcpy(dx, (void**)x, M*sizeof(CudaAECComplex), cudaMemcpyHostToDevice);

    cudaMemcpy(dy, (void**)y, N*sizeof(CudaAECComplex), cudaMemcpyHostToDevice);

    cudaMemcpy(dA, (void**)A, M*N*sizeof(CudaAECComplex), cudaMemcpyHostToDevice);

size_t status = cublasCgerc_v2(m_cublasHandle, M, N, &alpha, dx, 1, dy, 1, dA, M);

std::cout << "Status: " << status << std::endl;

cudaMemcpy((void**)A, dA, M*N*sizeof(CudaAECComplex), cudaMemcpyDeviceToHost);

for(unsigned int i=0;i<M;++i) {

        for(unsigned int j=0;j<N;++j) {

            std::cout << "("<< A[i+j*N].x << "," << A[i+j*N].y << ") ";

        }

        std::cout << std::endl;

    }

cudaFree(dx);

    cudaFree(dy);

    cudaFree(dA);

delete[] x;

    delete[] y;

    delete[] A;

I guess there is a very dumb error somewhere in there, but I was not able to locate it. Any help is deepply appreciated.

Best regards, floxtrott

2

The program above works pretty good in case you have

M=N=8

But I need it for

M!=N

Is there anybody with a helpful hint?

Best regards,

floxtrott

3

CUBLAS is expecting data in column-major order.

4

Thanks for your reply. However, the only matrix is initialized with zeros and later printed out in column-major mode. Anyway, it doesn’t matter, since it now works. I am not sure what exactly I changed, but my unit tests runs without complains. Must have been a typo or something.

Greetings and thank you

floxtrott