Using rocSOLVER Library

Once installed, rocSOLVER can be used just like any other library with a C API. The header file will need to be included in the user code, and both the rocBLAS and rocSOLVER shared libraries will become link-time and run-time dependencies for the user application.

Next, some examples are used to illustrate the basic use of rocSOLVER API and rocSOLVER batched API.

QR factorization of a single matrix

The following code snippet uses rocSOLVER to compute the QR factorization of a general m-by-n real matrix in double precision. For a full description of the used rocSOLVER routine, see the API documentation here: rocsolver_dgeqrf().

#include <hip/hip_runtime_api.h> // for hip functions
#include <rocsolver/rocsolver.h> // for all the rocsolver C interfaces and type declarations
#include <stdio.h>   // for printf
#include <stdlib.h>  // for malloc

// Example: Compute the QR Factorization of a matrix on the GPU

double *create_example_matrix(rocblas_int *M_out,
                              rocblas_int *N_out,
                              rocblas_int *lda_out) {
  // a *very* small example input; not a very efficient use of the API
  const double A[3][3] = { {  12, -51,   4},
                           {   6, 167, -68},
                           {  -4,  24, -41} };
  const rocblas_int M = 3;
  const rocblas_int N = 3;
  const rocblas_int lda = 3;
  *M_out = M;
  *N_out = N;
  *lda_out = lda;
  // note: rocsolver matrices must be stored in column major format,
  //       i.e. entry (i,j) should be accessed by hA[i + j*lda]
  double *hA = (double*)malloc(sizeof(double)*lda*N);
  for (size_t i = 0; i < M; ++i) {
    for (size_t j = 0; j < N; ++j) {
      // copy A (2D array) into hA (1D array, column-major)
      hA[i + j*lda] = A[i][j];
    }
  }
  return hA;
}

// We use rocsolver_dgeqrf to factor a real M-by-N matrix, A.
// See https://rocm.docs.amd.com/projects/rocSOLVER/en/latest/api/lapack.html#rocsolver-type-geqrf
int main() {
  rocblas_int M;          // rows
  rocblas_int N;          // cols
  rocblas_int lda;        // leading dimension
  double *hA = create_example_matrix(&M, &N, &lda); // input matrix on CPU

  // let's print the input matrix, just to see it
  printf("A = [\n");
  for (size_t i = 0; i < M; ++i) {
    printf("  ");
    for (size_t j = 0; j < N; ++j) {
      printf("% .3f ", hA[i + j*lda]);
    }
    printf(";\n");
  }
  printf("]\n");

  // initialization
  rocblas_handle handle;
  rocblas_create_handle(&handle);

  // Some rocsolver functions may trigger rocblas to load its GEMM kernels.
  // You can preload the kernels by explicitly invoking rocblas_initialize
  // (e.g., to exclude one-time initialization overhead from benchmarking).

  // preload rocBLAS GEMM kernels (optional)
  // rocblas_initialize();

  // calculate the sizes of our arrays
  size_t size_A = lda * (size_t)N;   // count of elements in matrix A
  size_t size_piv = (M < N) ? M : N; // count of Householder scalars

  // allocate memory on GPU
  double *dA, *dIpiv;
  hipMalloc((void**)&dA, sizeof(double)*size_A);
  hipMalloc((void**)&dIpiv, sizeof(double)*size_piv);

  // copy data to GPU
  hipMemcpy(dA, hA, sizeof(double)*size_A, hipMemcpyHostToDevice);

  // compute the QR factorization on the GPU
  rocsolver_dgeqrf(handle, M, N, dA, lda, dIpiv);

  // copy the results back to CPU
  double *hIpiv = (double*)malloc(sizeof(double)*size_piv); // householder scalars on CPU
  hipMemcpy(hA, dA, sizeof(double)*size_A, hipMemcpyDeviceToHost);
  hipMemcpy(hIpiv, dIpiv, sizeof(double)*size_piv, hipMemcpyDeviceToHost);

  // the results are now in hA and hIpiv
  // we can print some of the results if we want to see them
  printf("R = [\n");
  for (size_t i = 0; i < M; ++i) {
    printf("  ");
    for (size_t j = 0; j < N; ++j) {
      printf("% .3f ", (i <= j) ? hA[i + j*lda] : 0);
    }
    printf(";\n");
  }
  printf("]\n");

  // clean up
  free(hIpiv);
  hipFree(dA);
  hipFree(dIpiv);
  free(hA);
  rocblas_destroy_handle(handle);
}

The exact command used to compile the example above may vary depending on the system environment, but here is a typical example:

/opt/rocm/bin/hipcc -I/opt/rocm/include -c example.c
/opt/rocm/bin/hipcc -o example -L/opt/rocm/lib -lrocsolver -lrocblas example.o

QR factorization of a batch of matrices

One of the advantages of using GPUs is the ability to execute in parallel many operations of the same type but on different data sets. Based on this idea, rocSOLVER and rocBLAS provide a batch version of most of their routines. These batch versions allow the user to execute the same operation on a set of different matrices and/or vectors with a single library call.

Strided_batched version

The following code snippet uses rocSOLVER to compute the QR factorization of a series of general m-by-n real matrices in double precision. The matrices must be stored in contiguous memory locations on the GPU and, in this example, are accessed by a pointer to the first matrix and a stride value that gives the separation between one matrix and the next. For a full description of the used rocSOLVER routine, see the API documentation of rocsolver_dgeqrf_strided_batched().

#include <hip/hip_runtime_api.h> // for hip functions
#include <rocsolver/rocsolver.h> // for all the rocsolver C interfaces and type declarations
#include <stdio.h>   // for printf
#include <stdlib.h> // for malloc

// Example: Compute the QR Factorizations of an array of matrices on the GPU

double *create_example_matrices(rocblas_int *M_out,
                                rocblas_int *N_out,
                                rocblas_int *lda_out,
                                rocblas_stride *strideA_out,
                                rocblas_int *batch_count_out) {
  const double A[2][3][3] = {
    // First input matrix
    { { 12, -51,   4},
      {  6, 167, -68},
      { -4,  24, -41} },

    // Second input matrix
    { {  3, -12,  11},
      {  4, -46,  -2},
      {  0,   5,  15} } };

  const rocblas_int M = 3;
  const rocblas_int N = 3;
  const rocblas_int lda = 3;
  const rocblas_stride strideA = lda * N;
  const rocblas_int batch_count = 2;
  *M_out = M;
  *N_out = N;
  *lda_out = lda;
  *strideA_out = strideA;
  *batch_count_out = batch_count;

  // allocate space for input matrix data on CPU
  double *hA = (double*)malloc(sizeof(double)*strideA*batch_count);

  // copy A (3D array) into hA (1D array, column-major)
  for (size_t b = 0; b < batch_count; ++b)
    for (size_t i = 0; i < M; ++i)
      for (size_t j = 0; j < N; ++j)
        hA[i + j*lda + b*strideA] = A[b][i][j];

  return hA;
}

// Use rocsolver_dgeqrf_strided_batched to factor an array of real M-by-N matrices.
int main() {
  rocblas_int M;           // rows
  rocblas_int N;           // cols
  rocblas_int lda;         // leading dimension
  rocblas_stride strideA;  // stride from start of one matrix to the next
  rocblas_int batch_count; // number of matricies
  double *hA = create_example_matrices(&M, &N, &lda, &strideA, &batch_count);

  // print the input matrices
  for (size_t b = 0; b < batch_count; ++b) {
    printf("A[%zu] = [\n", b);
    for (size_t i = 0; i < M; ++i) {
      printf("  ");
      for (size_t j = 0; j < N; ++j) {
        printf("% 4.f ", hA[i + j*lda + strideA*b]);
      }
      printf(";\n");
    }
    printf("]\n");
  }

  // initialization
  rocblas_handle handle;
  rocblas_create_handle(&handle);

  // preload rocBLAS GEMM kernels (optional)
  // rocblas_initialize();

  // calculate the sizes of our arrays
  size_t size_A = strideA * (size_t)batch_count;   // elements in array for matrices
  rocblas_stride strideP = (M < N) ? M : N;        // stride of Householder scalar sets
  size_t size_piv = strideP * (size_t)batch_count; // elements in array for Householder scalars

  // allocate memory on GPU
  double *dA, *dIpiv;
  hipMalloc((void**)&dA, sizeof(double)*size_A);
  hipMalloc((void**)&dIpiv, sizeof(double)*size_piv);

  // copy data to GPU
  hipMemcpy(dA, hA, sizeof(double)*size_A, hipMemcpyHostToDevice);

  // compute the QR factorizations on the GPU
  rocsolver_dgeqrf_strided_batched(handle, M, N, dA, lda, strideA, dIpiv, strideP, batch_count);

  // copy the results back to CPU
  double *hIpiv = (double*)malloc(sizeof(double)*size_piv); // householder scalars on CPU
  hipMemcpy(hA, dA, sizeof(double)*size_A, hipMemcpyDeviceToHost);
  hipMemcpy(hIpiv, dIpiv, sizeof(double)*size_piv, hipMemcpyDeviceToHost);

  // the results are now in hA and hIpiv
  // print some of the results
  for (size_t b = 0; b < batch_count; ++b) {
    printf("R[%zu] = [\n", b);
    for (size_t i = 0; i < M; ++i) {
      printf("  ");
      for (size_t j = 0; j < N; ++j) {
        printf("% 4.f ", (i <= j) ? hA[i + j*lda + strideA*b] : 0);
      }
      printf(";\n");
    }
    printf("]\n");
  }

  // clean up
  free(hIpiv);
  hipFree(dA);
  hipFree(dIpiv);
  free(hA);
  rocblas_destroy_handle(handle);
}

Batched version

The following code snippet uses rocSOLVER to compute the QR factorization of a series of general m-by-n real matrices in double precision. In this case, the matrices do not need to be in contiguous memory locations on the GPU, and will be accessed by an array of pointers. For a full description of the used rocSOLVER routine, see the API documentation of rocsolver_dgeqrf_batched.

#include <hip/hip_runtime_api.h> // for hip functions
#include <rocsolver/rocsolver.h> // for all the rocsolver C interfaces and type declarations
#include <stdio.h> // for printf
#include <stdlib.h> // for malloc

// Example: Compute the QR Factorizations of a batch of matrices on the GPU

double **create_example_matrices(rocblas_int *M_out,
                                 rocblas_int *N_out,
                                 rocblas_int *lda_out,
                                 rocblas_int *batch_count_out) {
  // a small example input
  const double A[2][3][3] = {
    // First input matrix
    { { 12, -51,   4},
      {  6, 167, -68},
      { -4,  24, -41} },

    // Second input matrix
    { {  3, -12,  11},
      {  4, -46,  -2},
      {  0,   5,  15} } };

  const rocblas_int M = 3;
  const rocblas_int N = 3;
  const rocblas_int lda = 3;
  const rocblas_int batch_count = 2;
  *M_out = M;
  *N_out = N;
  *lda_out = lda;
  *batch_count_out = batch_count;

  // allocate space for input matrix data on CPU
  double **hA = (double**)malloc(sizeof(double*)*batch_count);
  hA[0] = (double*)malloc(sizeof(double)*lda*N);
  hA[1] = (double*)malloc(sizeof(double)*lda*N);

  for (size_t b = 0; b < batch_count; ++b)
    for (size_t i = 0; i < M; ++i)
      for (size_t j = 0; j < N; ++j)
        hA[b][i + j*lda] = A[b][i][j];

  return hA;
}

// Use rocsolver_dgeqrf_batched to factor a batch of real M-by-N matrices.
int main() {
  rocblas_int M;           // rows
  rocblas_int N;           // cols
  rocblas_int lda;         // leading dimension
  rocblas_int batch_count; // number of matricies
  double **hA = create_example_matrices(&M, &N, &lda, &batch_count);

  // print the input matrices
  for (size_t b = 0; b < batch_count; ++b) {
    printf("A[%zu] = [\n", b);
    for (size_t i = 0; i < M; ++i) {
      printf("  ");
      for (size_t j = 0; j < N; ++j) {
        printf("% 4.f ", hA[b][i + j*lda]);
      }
      printf(";\n");
    }
    printf("]\n");
  }

  // initialization
  rocblas_handle handle;
  rocblas_create_handle(&handle);

  // preload rocBLAS GEMM kernels (optional)
  // rocblas_initialize();

  // calculate the sizes of the arrays
  size_t size_A = lda * (size_t)N;          // count of elements in each matrix A
  rocblas_stride strideP = (M < N) ? M : N; // stride of Householder scalar sets
  size_t size_piv = strideP * (size_t)batch_count; // elements in array for Householder scalars

  // allocate memory on the CPU for an array of pointers,
  // then allocate memory for each matrix on the GPU.
  double **A = (double**)malloc(sizeof(double*)*batch_count);
  for (rocblas_int b = 0; b < batch_count; ++b)
    hipMalloc((void**)&A[b], sizeof(double)*size_A);

  // allocate memory on GPU for the array of pointers and Householder scalars
  double **dA, *dIpiv;
  hipMalloc((void**)&dA, sizeof(double*)*batch_count);
  hipMalloc((void**)&dIpiv, sizeof(double)*size_piv);

  // copy each matrix to the GPU
  for (rocblas_int b = 0; b < batch_count; ++b)
    hipMemcpy(A[b], hA[b], sizeof(double)*size_A, hipMemcpyHostToDevice);

  // copy the array of pointers to the GPU
  hipMemcpy(dA, A, sizeof(double*)*batch_count, hipMemcpyHostToDevice);

  // compute the QR factorizations on the GPU
  rocsolver_dgeqrf_batched(handle, M, N, dA, lda, dIpiv, strideP, batch_count);

  // copy the results back to CPU
  double *hIpiv = (double*)malloc(sizeof(double)*size_piv); // householder scalars on CPU
  hipMemcpy(hIpiv, dIpiv, sizeof(double)*size_piv, hipMemcpyDeviceToHost);
  for (rocblas_int b = 0; b < batch_count; ++b)
    hipMemcpy(hA[b], A[b], sizeof(double)*size_A, hipMemcpyDeviceToHost);

  // the results are now in hA and hIpiv
  // print some of the results
  for (size_t b = 0; b < batch_count; ++b) {
    printf("R[%zu] = [\n", b);
    for (size_t i = 0; i < M; ++i) {
      printf("  ");
      for (size_t j = 0; j < N; ++j) {
        printf("% 4.f ", (i <= j) ? hA[b][i + j*lda] : 0);
      }
      printf(";\n");
    }
    printf("]\n");
  }

  // clean up
  free(hIpiv);
  for (rocblas_int b = 0; b < batch_count; ++b)
    free(hA[b]);
  free(hA);
  for (rocblas_int b = 0; b < batch_count; ++b)
    hipFree(A[b]);
  free(A);
  hipFree(dA);
  hipFree(dIpiv);
  rocblas_destroy_handle(handle);
}