[petsc-users] issue of MatCreateDense in the CUDA codes

[Jose E. Roman]

Does it work if you declare everything as PetscScalar instead of cuDoubleComplex?

> El 2 oct 2024, a las 11:23, 刘浪天 <langtian.liu at icloud.com> escribió:
> 
> Hi Jose,
> 
> Since my matrix is two large, I cannot create the Mat on GPU. So I still want to create and compute the eigenvalues of this matrix on CPU using SLEPc.
> 
> Best,
> -------------------- Langtian Liu Institute for Theorectical Physics, Justus-Liebig-University Giessen Heinrich-Buff-Ring 16, 35392 Giessen Germany email: langtian.liu at icloud.com Tel: (+49)641 99 33342
> 
>> On Oct 2, 2024, at 11:18 AM, Jose E. Roman <jroman at dsic.upv.es> wrote:
>> 
>> 
>> For the CUDA case you should use MatCreateDenseCUDA() instead of MatCreateDense(). With this you pass a pointer with the data on the GPU memory. But I guess "new cuDoubleComplex[dim*dim]" is allocating on the CPU, you should use cudaMalloc() instead.
>> 
>> Jose
>> 
>> 
>>> El 2 oct 2024, a las 10:56, 刘浪天 via petsc-users <petsc-users at mcs.anl.gov> escribió:
>>> 
>>> Hi all,
>>> 
>>> I am using the PETSc and SLEPc to solve the Faddeev equation of baryons. I encounter a problem of function MatCreateDense when changing from CPU to CPU-GPU computations.
>>> At first, I write the codes in purely CPU computation in the following way and it works.
>>> ```
>>> Eigen::MatrixXcd H_KER;
>>> Eigen::MatrixXcd G0;
>>> printf("\nCompute the propagator matrix.\n");
>>> prop_matrix_nucleon_sc_av(Mn, pp_nodes, cos1_nodes);
>>> printf("\nCompute the propagator matrix done.\n");
>>> printf("\nCompute the kernel matrix.\n");
>>> bse_kernel_nucleon_sc_av(Mn, pp_nodes, pp_weights, cos1_nodes, cos1_weights);
>>> printf("\nCompute the kernel matrix done.\n");
>>> printf("\nCompute the full kernel matrix by multiplying kernel and propagator matrix.\n");
>>> MatrixXcd kernel_temp = H_KER * G0;
>>> printf("\nCompute the full kernel matrix done.\n");
>>> 
>>> // Solve the eigen system with SLEPc
>>> printf("\nSolve the eigen system in the rest frame.\n");
>>> // Get the size of the Eigen matrix
>>> int nRows = (int) kernel_temp.rows();
>>> int nCols = (int) kernel_temp.cols();
>>> // Create PETSc matrix and share the data of kernel_temp
>>> Mat kernel;
>>> PetscCall(MatCreateDense(PETSC_COMM_WORLD, PETSC_DECIDE, PETSC_DECIDE, nRows, nCols, kernel_temp.data(), &kernel));
>>> PetscCall(MatAssemblyBegin(kernel, MAT_FINAL_ASSEMBLY));
>>> PetscCall(MatAssemblyEnd(kernel, MAT_FINAL_ASSEMBLY));
>>> ```
>>> Now I change to compute the propagator and kernel matrices in GPU and then compute the largest eigenvalues in CPU using SLEPc in the ways below.
>>> ```
>>> cuDoubleComplex *h_propmat;
>>> cuDoubleComplex *h_kernelmat;
>>> int dim = EIGHT * NP * NZ;
>>> printf("\nCompute the propagator matrix.\n");
>>> prop_matrix_nucleon_sc_av_cuda(Mn, pp_nodes.data(), cos1_nodes.data());
>>> printf("\nCompute the propagator matrix done.\n");
>>> printf("\nCompute the kernel matrix.\n");
>>> kernel_matrix_nucleon_sc_av_cuda(Mn, pp_nodes.data(), pp_weights.data(), cos1_nodes.data(), cos1_weights.data());
>>> printf("\nCompute the kernel matrix done.\n");
>>> printf("\nCompute the full kernel matrix by multiplying kernel and propagator matrix.\n");
>>> // Map the raw arrays to Eigen matrices (column-major order)
>>> auto *h_kernel_temp = new cuDoubleComplex [dim*dim];
>>> matmul_cublas_cuDoubleComplex(h_kernelmat,h_propmat,h_kernel_temp,dim,dim,dim);
>>> printf("\nCompute the full kernel matrix done.\n");
>>> 
>>> // Solve the eigen system with SLEPc
>>> printf("\nSolve the eigen system in the rest frame.\n");
>>> int nRows = dim;
>>> int nCols = dim;
>>> // Create PETSc matrix and share the data of kernel_temp
>>> Mat kernel;
>>> auto* h_kernel = (std::complex<double>*)(h_kernel_temp);
>>> PetscCall(MatCreateDense(PETSC_COMM_WORLD, PETSC_DECIDE, PETSC_DECIDE, nRows, nCols, h_kernel_temp, &kernel));
>>> PetscCall(MatAssemblyBegin(kernel, MAT_FINAL_ASSEMBLY));
>>> PetscCall(MatAssemblyEnd(kernel, MAT_FINAL_ASSEMBLY));
>>> But in this case, the compiler told me that the MatCreateDense function uses the data pointer as type of "thrust::complex<double>" instead of "std::complex<double>".
>>> I am sure I only configured and install PETSc in purely CPU without GPU and this codes are written in the host function.
>>> Why the function changes its behavior? Did you also meet this problem when writing the cuda codes and how to solve this problem.
>>> I tried to copy the data to a new thrust::complex<double> matrix but this is very time consuming since my matrix is very big. Is there a way to create the Mat from the original data without changing the data type to thrust::complex<double> in the cuda applications? Any response will be appreciated. Thank you!
>>> 
>>> Best wishes,
>>> Langtian Liu
>>> 
>>> ------
>>> Institute for Theorectical Physics, Justus-Liebig-University Giessen
>>> Heinrich-Buff-Ring 16, 35392 Giessen Germany
>