Commit 35cd635f authored by Mulas, Giacomo's avatar Mulas, Giacomo
Browse files

use less memory for identity subtraction in iterative refinement

parent 7b044c50

src/libnptm/cublas_calls.cpp

+6 −6
Original line number Diff line number Diff line
@@ -143,12 +143,12 @@ void cublas_zinvert_and_refine(dcomplex **mat, np_int n, int &maxiters, double & 
    cudacall(cudaMalloc<cuDoubleComplex>(&d_a_residual, m*m*sizeof(cuDoubleComplex)));

    cudacall(cudaMalloc<cuDoubleComplex>(&d_a_refine, m*m*sizeof(cuDoubleComplex)));

    // allocate memory for the temporary matrix products

    dcomplex *native_id = new dcomplex[m];

    for (np_int i=0; i<m; i++) native_id[i] = 1;

    dcomplex *native_id = new dcomplex[1];

    native_id[0] = 1;

    cuDoubleComplex *m_id = (cuDoubleComplex *) &(native_id[0]);

    // fill it with 1

    cudacall(cudaMalloc<cuDoubleComplex>(&d_id, m*sizeof(cuDoubleComplex)));

    cudacall(cudaMemcpy(d_id, m_id, m*sizeof(cuDoubleComplex),cudaMemcpyHostToDevice)); // copy identity to device vector

    cudacall(cudaMalloc<cuDoubleComplex>(&d_id, sizeof(cuDoubleComplex)));

    cudacall(cudaMemcpy(d_id, m_id, sizeof(cuDoubleComplex),cudaMemcpyHostToDevice)); // copy identity to device vector

    delete[] native_id; // free identity vector on host

  }

  bool iteraterefine = true;
@@ -167,7 +167,7 @@ void cublas_zinvert_and_refine(dcomplex **mat, np_int n, int &maxiters, double & 
    // Fortran-style arrays, whereas our arrays are C-style row-major.

    cublascall(CUZGEMM(handle, CUBLAS_OP_N, CUBLAS_OP_N, m, m, m, &cu_mone, d_c, m, d_a, m, &cu_zero, d_a_residual, m));

    // add the identity to the product

    cublascall(CUZAXPY(handle, m, &cu_one, d_id, 1, d_a_residual, m+1));

    cublascall(CUZAXPY(handle, m, &cu_one, d_id, 0, d_a_residual, m+1));

    double oldmax=0;

    if (refinemode >0) {

      np_int maxindex;
@@ -191,7 +191,7 @@ void cublas_zinvert_and_refine(dcomplex **mat, np_int n, int &maxiters, double & 
      // multiply minus the original matrix times the new inverse matrix

      cublascall(CUZGEMM(handle, CUBLAS_OP_N, CUBLAS_OP_N, m, m, m, &cu_mone, d_c, m, d_a, m, &cu_zero, d_a_residual, m));

      // add the identity to the product

      cublascall(CUZAXPY(handle, m, &cu_one, d_id, 1, d_a_residual, m+1));

      cublascall(CUZAXPY(handle, m, &cu_one, d_id, 0, d_a_residual, m+1));

      if ((refinemode==2) || ((refinemode==1) && (iter == (maxiters-1)))) {

	// find the maximum absolute value of the residual

	np_int maxindex;

src/libnptm/lapack_calls.cpp

+8 −9
Original line number Diff line number Diff line
@@ -92,6 +92,7 @@ void zinvert_and_refine(dcomplex **mat, np_int n, int &jer, int &maxiters, doubl 
#endif

  np_int nn = n*n;

  np_int incx = 1;

  np_int incx0 = 0;

#ifdef USE_MKL

  MKL_Complex16 *arr_orig = NULL;

  MKL_Complex16 *arr_residual = NULL;
@@ -108,17 +109,15 @@ void zinvert_and_refine(dcomplex **mat, np_int n, int &jer, int &maxiters, doubl 
    arr_orig = new MKL_Complex16[nn];

    arr_residual = new MKL_Complex16[nn];

    arr_refine = new MKL_Complex16[nn];

    id = new MKL_Complex16[n];

    for (np_int i=0; i<n ; i++) {

      id[i].real =  1;

      id[i].imag =  0;

    }

    id = new MKL_Complex16[1];

    id[0].real =  1;

    id[0].imag =  0;

#else

    arr_orig = new dcomplex[nn];

    arr_residual = new dcomplex[nn];

    arr_refine = new dcomplex[nn];

    id = new dcomplex[n];

    for (np_int i=0; i<n ; i++) id[i] = (dcomplex) 1;

    id = new dcomplex[1];

    id[0] = (dcomplex) 1;

#endif

    zcopy_(&nn, arr, &incx, arr_orig, &incx);

  }
@@ -153,7 +152,7 @@ void zinvert_and_refine(dcomplex **mat, np_int n, int &jer, int &maxiters, doubl 
    // Fortran-style arrays, whereas our arrays are C-style row-major.

    zgemm_(&transa, &transa, &n, &n, &n, &dcmone, arr, &n, arr_orig, &n, &dczero, arr_residual, &n);

    np_int incy = n+1;

    zaxpy_(&n, &dcone, id, &incx, arr_residual, &incy);

    zaxpy_(&n, &dcone, id, &incx0, arr_residual, &incy);

    double oldmax = 0;

    if (refinemode >0) {

      np_int maxindex = izamax_(&nn, arr_residual, &incx);
@@ -171,7 +170,7 @@ void zinvert_and_refine(dcomplex **mat, np_int n, int &jer, int &maxiters, doubl 
      zaxpy_(&nn, &dcone, arr_refine, &incx, arr, &incx);

	// zcopy_(&nn, arr_refine, &incx, arr, &incx);

      zgemm_(&transa, &transa, &n, &n, &n, &dcmone, arr, &n, arr_orig, &n, &dczero, arr_residual, &n);

      zaxpy_(&n, &dcone, id, &incx, arr_residual, &incy);

      zaxpy_(&n, &dcone, id, &incx0, arr_residual, &incy);

      if ((refinemode==2) || ((refinemode==1) && (iter == (maxiters-1)))) {

	np_int maxindex = izamax_(&nn, arr_residual, &incx);

#ifdef USE_MKL

src/libnptm/magma_calls.cpp

+6 −6
Original line number Diff line number Diff line
@@ -132,16 +132,16 @@ void magma_zinvert_and_refine(dcomplex **mat, np_int n, int &jer, int &maxiters, 
    }

    // allocate memory for the identity vector on the host

    {

      dcomplex *native_id = new dcomplex[m];

      for (magma_int_t i=0; i<m; i++) native_id[i] = 1;

      dcomplex *native_id = new dcomplex[1];

      native_id[0] = 1;

      magmaDoubleComplex *id = (magmaDoubleComplex *) &(native_id[0]);

      // fill it with 1

      err = magma_zmalloc(&d_id, m);

      err = magma_zmalloc(&d_id, 1);

      if (err != MAGMA_SUCCESS) {

	printf("Error allocating d_id\n");

	exit(1);

      }

      magma_zsetvector(m, id, 1, d_id, 1, queue); // copy identity to device vector

      magma_zsetvector(1, id, 1, d_id, 1, queue); // copy identity to device vector

      delete[] native_id; // free identity vector on host

    }

  }
@@ -161,7 +161,7 @@ void magma_zinvert_and_refine(dcomplex **mat, np_int n, int &jer, int &maxiters, 
    // Fortran-style arrays, whereas our arrays are C-style row-major.

    magma_zgemm(MagmaNoTrans, MagmaNoTrans, m, m, m,  magma_mone, d_a, m, d_a_orig, m, magma_zero, d_a_residual, m, queue);

    // add the identity to the product

    magma_zaxpy (m, magma_one, d_id, 1, d_a_residual, m+1, queue);

    magma_zaxpy (m, magma_one, d_id, 0, d_a_residual, m+1, queue);

    double oldmax=0;

    if (refinemode >0) {

      // find the maximum absolute value of the residual
@@ -184,7 +184,7 @@ void magma_zinvert_and_refine(dcomplex **mat, np_int n, int &jer, int &maxiters, 
      // multiply minus the original matrix times the new inverse matrix

      magma_zgemm(MagmaNoTrans, MagmaNoTrans, m, m, m, magma_mone, d_a, m, d_a_orig, m, magma_zero, d_a_residual, m, queue);

      // add the identity to the product

      magma_zaxpy (m, magma_one, d_id, 1, d_a_residual, m+1, queue);

      magma_zaxpy (m, magma_one, d_id, 0, d_a_residual, m+1, queue);

      if ((refinemode==2) || ((refinemode==1) && (iter == (maxiters-1)))) {

	// find the maximum absolute value of the residual

	magma_int_t maxindex = magma_izamax(mm, d_a_residual, 1, queue);