int
hypre_CSRMatrixMatvecT( double alpha,
hypre_CSRMatrix *A,
hypre_Vector *x,
double beta,
hypre_Vector *y )
{
double *A_data = hypre_CSRMatrixData(A);
int *A_i = hypre_CSRMatrixI(A);
int *A_j = hypre_CSRMatrixJ(A);
int num_rows = hypre_CSRMatrixNumRows(A);
int num_cols = hypre_CSRMatrixNumCols(A);
double *x_data = hypre_VectorData(x);
double *y_data = hypre_VectorData(y);
int x_size = hypre_VectorSize(x);
int y_size = hypre_VectorSize(y);
int num_vectors = hypre_VectorNumVectors(x);
int idxstride_y = hypre_VectorIndexStride(y);
int vecstride_y = hypre_VectorVectorStride(y);
int idxstride_x = hypre_VectorIndexStride(x);
int vecstride_x = hypre_VectorVectorStride(x);
double temp;
int i, i1, j, jv, jj, ns, ne, size, rest;
int num_threads;
int ierr = 0;
/*---------------------------------------------------------------------
* Check for size compatibility. MatvecT returns ierr = 1 if
* length of X doesn't equal the number of rows of A,
* ierr = 2 if the length of Y doesn't equal the number of
* columns of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in MatvecT, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
hypre_assert( num_vectors == hypre_VectorNumVectors(y) );
if (num_rows != x_size)
ierr = 1;
if (num_cols != y_size)
ierr = 2;
if (num_rows != x_size && num_cols != y_size)
ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= beta;
return ierr;
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
if (temp != 1.0)
{
if (temp == 0.0)
{
#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] = 0.0;
}
else
{
#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A^T*x
*-----------------------------------------------------------------*/
num_threads = hypre_NumThreads();
if (num_threads > 1)
{
#define HYPRE_SMP_PRIVATE i, i1,jj,j,ns,ne,size,rest
#include "../utilities/hypre_smp_forloop.h"
for (i1 = 0; i1 < num_threads; i1++)
{
size = num_cols/num_threads;
rest = num_cols - size*num_threads;
if (i1 < rest)
{
ns = i1*size+i1-1;
ne = (i1+1)*size+i1+1;
}
else
{
ns = i1*size+rest-1;
ne = (i1+1)*size+rest;
}
if ( num_vectors==1 )
{
for (i = 0; i < num_rows; i++)
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
if (j > ns && j < ne)
y_data[j] += A_data[jj] * x_data[i];
}
}
}
else
{
for (i = 0; i < num_rows; i++)
{
for ( jv=0; jv<num_vectors; ++jv )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
if (j > ns && j < ne)
y_data[ j*idxstride_y + jv*vecstride_y ] +=
A_data[jj] * x_data[ i*idxstride_x + jv*vecstride_x];
}
}
}
}
}
}
else
{
for (i = 0; i < num_rows; i++)
{
if ( num_vectors==1 )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[j] += A_data[jj] * x_data[i];
}
}
else
{
for ( jv=0; jv<num_vectors; ++jv )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[ j*idxstride_y + jv*vecstride_y ] +=
A_data[jj] * x_data[ i*idxstride_x + jv*vecstride_x ];
}
}
}
}
}
/*-----------------------------------------------------------------
* y = alpha*y
*-----------------------------------------------------------------*/
if (alpha != 1.0)
{
#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= alpha;
}
return ierr;
}
int
hypre_CSRMatrixMatvec( double alpha,
hypre_CSRMatrix *A,
hypre_Vector *x,
double beta,
hypre_Vector *y )
{
double *A_data = hypre_CSRMatrixData(A);
int *A_i = hypre_CSRMatrixI(A);
int *A_j = hypre_CSRMatrixJ(A);
int num_rows = hypre_CSRMatrixNumRows(A);
int num_cols = hypre_CSRMatrixNumCols(A);
int *A_rownnz = hypre_CSRMatrixRownnz(A);
int num_rownnz = hypre_CSRMatrixNumRownnz(A);
double *x_data = hypre_VectorData(x);
double *y_data = hypre_VectorData(y);
int x_size = hypre_VectorSize(x);
int y_size = hypre_VectorSize(y);
int num_vectors = hypre_VectorNumVectors(x);
int idxstride_y = hypre_VectorIndexStride(y);
int vecstride_y = hypre_VectorVectorStride(y);
int idxstride_x = hypre_VectorIndexStride(x);
int vecstride_x = hypre_VectorVectorStride(x);
double temp, tempx;
int i, j, jj;
int m;
double xpar=0.7;
int ierr = 0;
/*---------------------------------------------------------------------
* Check for size compatibility. Matvec returns ierr = 1 if
* length of X doesn't equal the number of columns of A,
* ierr = 2 if the length of Y doesn't equal the number of rows
* of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in Matvec, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
hypre_assert( num_vectors == hypre_VectorNumVectors(y) );
if (num_cols != x_size)
ierr = 1;
if (num_rows != y_size)
ierr = 2;
if (num_cols != x_size && num_rows != y_size)
ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
/*#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h" */
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] *= beta;
return ierr;
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
if (temp != 1.0)
{
if (temp == 0.0)
{
#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] = 0.0;
}
else
{
#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A*x
*-----------------------------------------------------------------*/
if (num_rownnz < xpar*(num_rows))
{
#define HYPRE_SMP_PRIVATE i,jj
#include "../utilities/hypre_smp_forloop.h"
/* use rownnz pointer to do the A*x multiplication when num_rownnz is smaller than num_rows */
for (i = 0; i < num_rownnz; i++)
{
m = A_rownnz[i];
/*
* for (jj = A_i[m]; jj < A_i[m+1]; jj++)
* {
* j = A_j[jj];
* y_data[m] += A_data[jj] * x_data[j];
* } */
if ( num_vectors==1 )
{
tempx = y_data[m];
for (jj = A_i[m]; jj < A_i[m+1]; jj++)
tempx += A_data[jj] * x_data[A_j[jj]];
y_data[m] = tempx;
}
else
for ( j=0; j<num_vectors; ++j )
{
tempx = y_data[ j*vecstride_y + m*idxstride_y ];
for (jj = A_i[m]; jj < A_i[m+1]; jj++)
tempx += A_data[jj] * x_data[ j*vecstride_x + A_j[jj]*idxstride_x ];
y_data[ j*vecstride_y + m*idxstride_y] = tempx;
}
}
}
else
{
#define HYPRE_SMP_PRIVATE i,jj,temp
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_rows; i++)
{
if ( num_vectors==1 )
{
temp = y_data[i];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
temp += A_data[jj] * x_data[A_j[jj]];
y_data[i] = temp;
}
else
for ( j=0; j<num_vectors; ++j )
{
temp = y_data[ j*vecstride_y + i*idxstride_y ];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
temp += A_data[jj] * x_data[ j*vecstride_x + A_j[jj]*idxstride_x ];
}
y_data[ j*vecstride_y + i*idxstride_y ] = temp;
}
}
}
/*-----------------------------------------------------------------
* y = alpha*y
*-----------------------------------------------------------------*/
if (alpha != 1.0)
{
#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] *= alpha;
}
return ierr;
}
HYPRE_Int
hypre_CSRMatrixMatvecT( HYPRE_Complex alpha,
hypre_CSRMatrix *A,
hypre_Vector *x,
HYPRE_Complex beta,
hypre_Vector *y )
{
HYPRE_Complex *A_data = hypre_CSRMatrixData(A);
HYPRE_Int *A_i = hypre_CSRMatrixI(A);
HYPRE_Int *A_j = hypre_CSRMatrixJ(A);
HYPRE_Int num_rows = hypre_CSRMatrixNumRows(A);
HYPRE_Int num_cols = hypre_CSRMatrixNumCols(A);
HYPRE_Complex *x_data = hypre_VectorData(x);
HYPRE_Complex *y_data = hypre_VectorData(y);
HYPRE_Int x_size = hypre_VectorSize(x);
HYPRE_Int y_size = hypre_VectorSize(y);
HYPRE_Int num_vectors = hypre_VectorNumVectors(x);
HYPRE_Int idxstride_y = hypre_VectorIndexStride(y);
HYPRE_Int vecstride_y = hypre_VectorVectorStride(y);
HYPRE_Int idxstride_x = hypre_VectorIndexStride(x);
HYPRE_Int vecstride_x = hypre_VectorVectorStride(x);
HYPRE_Complex temp;
HYPRE_Complex *y_data_expand;
HYPRE_Int my_thread_num = 0, offset = 0;
HYPRE_Int i, j, jv, jj;
HYPRE_Int num_threads;
HYPRE_Int ierr = 0;
hypre_Vector *x_tmp = NULL;
/*---------------------------------------------------------------------
* Check for size compatibility. MatvecT returns ierr = 1 if
* length of X doesn't equal the number of rows of A,
* ierr = 2 if the length of Y doesn't equal the number of
* columns of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in MatvecT, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
hypre_assert( num_vectors == hypre_VectorNumVectors(y) );
if (num_rows != x_size)
ierr = 1;
if (num_cols != y_size)
ierr = 2;
if (num_rows != x_size && num_cols != y_size)
ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= beta;
return ierr;
}
if (x == y)
{
x_tmp = hypre_SeqVectorCloneDeep(x);
x_data = hypre_VectorData(x_tmp);
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
if (temp != 1.0)
{
if (temp == 0.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] = 0.0;
}
else
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A^T*x
*-----------------------------------------------------------------*/
num_threads = hypre_NumThreads();
if (num_threads > 1)
{
y_data_expand = hypre_CTAlloc(HYPRE_Complex, num_threads*y_size);
if ( num_vectors==1 )
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel private(i,jj,j,my_thread_num,offset)
#endif
{
my_thread_num = hypre_GetThreadNum();
offset = y_size*my_thread_num;
#ifdef HYPRE_USING_OPENMP
#pragma omp for HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data_expand[offset + j] += A_data[jj] * x_data[i];
}
}
/* implied barrier (for threads)*/
#ifdef HYPRE_USING_OPENMP
#pragma omp for HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < y_size; i++)
{
for (j = 0; j < num_threads; j++)
{
y_data[i] += y_data_expand[j*y_size + i];
}
}
} /* end parallel threaded region */
}
else
{
/* multiple vector case is not threaded */
for (i = 0; i < num_rows; i++)
{
for ( jv=0; jv<num_vectors; ++jv )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[ j*idxstride_y + jv*vecstride_y ] +=
A_data[jj] * x_data[ i*idxstride_x + jv*vecstride_x];
}
}
}
}
hypre_TFree(y_data_expand);
}
else
{
for (i = 0; i < num_rows; i++)
{
if ( num_vectors==1 )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[j] += A_data[jj] * x_data[i];
}
}
else
{
for ( jv=0; jv<num_vectors; ++jv )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[ j*idxstride_y + jv*vecstride_y ] +=
A_data[jj] * x_data[ i*idxstride_x + jv*vecstride_x ];
}
}
}
}
}
/*-----------------------------------------------------------------
* y = alpha*y
*-----------------------------------------------------------------*/
if (alpha != 1.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= alpha;
}
if (x == y) hypre_SeqVectorDestroy(x_tmp);
return ierr;
}
HYPRE_Int
hypre_CSRMatrixMatvec_FF( HYPRE_Complex alpha,
hypre_CSRMatrix *A,
hypre_Vector *x,
HYPRE_Complex beta,
hypre_Vector *y,
HYPRE_Int *CF_marker_x,
HYPRE_Int *CF_marker_y,
HYPRE_Int fpt )
{
HYPRE_Complex *A_data = hypre_CSRMatrixData(A);
HYPRE_Int *A_i = hypre_CSRMatrixI(A);
HYPRE_Int *A_j = hypre_CSRMatrixJ(A);
HYPRE_Int num_rows = hypre_CSRMatrixNumRows(A);
HYPRE_Int num_cols = hypre_CSRMatrixNumCols(A);
HYPRE_Complex *x_data = hypre_VectorData(x);
HYPRE_Complex *y_data = hypre_VectorData(y);
HYPRE_Int x_size = hypre_VectorSize(x);
HYPRE_Int y_size = hypre_VectorSize(y);
HYPRE_Complex temp;
HYPRE_Int i, jj;
HYPRE_Int ierr = 0;
/*---------------------------------------------------------------------
* Check for size compatibility. Matvec returns ierr = 1 if
* length of X doesn't equal the number of columns of A,
* ierr = 2 if the length of Y doesn't equal the number of rows
* of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in Matvec, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
if (num_cols != x_size)
ierr = 1;
if (num_rows != y_size)
ierr = 2;
if (num_cols != x_size && num_rows != y_size)
ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
if (CF_marker_x[i] == fpt) y_data[i] *= beta;
return ierr;
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
if (temp != 1.0)
{
if (temp == 0.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
if (CF_marker_x[i] == fpt) y_data[i] = 0.0;
}
else
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
if (CF_marker_x[i] == fpt) y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A*x
*-----------------------------------------------------------------*/
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i,jj) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
{
if (CF_marker_x[i] == fpt)
{
temp = y_data[i];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
if (CF_marker_y[A_j[jj]] == fpt) temp += A_data[jj] * x_data[A_j[jj]];
y_data[i] = temp;
}
}
/*-----------------------------------------------------------------
* y = alpha*y
*-----------------------------------------------------------------*/
if (alpha != 1.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
if (CF_marker_x[i] == fpt) y_data[i] *= alpha;
}
return ierr;
}
HYPRE_Int
hypre_CSRBlockMatrixMatvec(double alpha, hypre_CSRBlockMatrix *A,
hypre_Vector *x, double beta, hypre_Vector *y)
{
double *A_data = hypre_CSRBlockMatrixData(A);
HYPRE_Int *A_i = hypre_CSRBlockMatrixI(A);
HYPRE_Int *A_j = hypre_CSRBlockMatrixJ(A);
HYPRE_Int num_rows = hypre_CSRBlockMatrixNumRows(A);
HYPRE_Int num_cols = hypre_CSRBlockMatrixNumCols(A);
HYPRE_Int blk_size = hypre_CSRBlockMatrixBlockSize(A);
double *x_data = hypre_VectorData(x);
double *y_data = hypre_VectorData(y);
HYPRE_Int x_size = hypre_VectorSize(x);
HYPRE_Int y_size = hypre_VectorSize(y);
HYPRE_Int i, b1, b2, jj, bnnz=blk_size*blk_size;
HYPRE_Int ierr = 0;
double temp;
/*---------------------------------------------------------------------
* Check for size compatibility. Matvec returns ierr = 1 if
* length of X doesn't equal the number of columns of A,
* ierr = 2 if the length of Y doesn't equal the number of rows
* of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in Matvec, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
if (num_cols*blk_size != x_size) ierr = 1;
if (num_rows*blk_size != y_size) ierr = 2;
if (num_cols*blk_size != x_size && num_rows*blk_size != y_size) ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_rows*blk_size; i++) y_data[i] *= beta;
return ierr;
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
if (temp != 1.0)
{
if (temp == 0.0)
{
#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_rows*blk_size; i++)
y_data[i] = 0.0;
}
else
{
#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_rows*blk_size; i++)
y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A*x
*-----------------------------------------------------------------*/
#define HYPRE_SMP_PRIVATE i,jj,b1,b2,temp
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_rows; i++)
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
for (b1 = 0; b1 < blk_size; b1++)
{
temp = y_data[i*blk_size+b1];
for (b2 = 0; b2 < blk_size; b2++)
temp += A_data[jj*bnnz+b1*blk_size+b2] * x_data[A_j[jj]*blk_size+b2];
y_data[i*blk_size+b1] = temp;
}
}
}
/*-----------------------------------------------------------------
* y = alpha*y
*-----------------------------------------------------------------*/
if (alpha != 1.0)
{
#define HYPRE_SMP_PRIVATE i
#include "../utilities/hypre_smp_forloop.h"
for (i = 0; i < num_rows*blk_size; i++)
y_data[i] *= alpha;
}
return ierr;
}
/* y[offset:end] = alpha*A[offset:end,:]*x + beta*b[offset:end] */
HYPRE_Int
hypre_CSRMatrixMatvecOutOfPlace( HYPRE_Complex alpha,
hypre_CSRMatrix *A,
hypre_Vector *x,
HYPRE_Complex beta,
hypre_Vector *b,
hypre_Vector *y,
HYPRE_Int offset )
{
#ifdef HYPRE_PROFILE
HYPRE_Real time_begin = hypre_MPI_Wtime();
#endif
HYPRE_Complex *A_data = hypre_CSRMatrixData(A);
HYPRE_Int *A_i = hypre_CSRMatrixI(A) + offset;
HYPRE_Int *A_j = hypre_CSRMatrixJ(A);
HYPRE_Int num_rows = hypre_CSRMatrixNumRows(A) - offset;
HYPRE_Int num_cols = hypre_CSRMatrixNumCols(A);
/*HYPRE_Int num_nnz = hypre_CSRMatrixNumNonzeros(A);*/
HYPRE_Int *A_rownnz = hypre_CSRMatrixRownnz(A);
HYPRE_Int num_rownnz = hypre_CSRMatrixNumRownnz(A);
HYPRE_Complex *x_data = hypre_VectorData(x);
HYPRE_Complex *b_data = hypre_VectorData(b) + offset;
HYPRE_Complex *y_data = hypre_VectorData(y);
HYPRE_Int x_size = hypre_VectorSize(x);
HYPRE_Int b_size = hypre_VectorSize(b) - offset;
HYPRE_Int y_size = hypre_VectorSize(y) - offset;
HYPRE_Int num_vectors = hypre_VectorNumVectors(x);
HYPRE_Int idxstride_y = hypre_VectorIndexStride(y);
HYPRE_Int vecstride_y = hypre_VectorVectorStride(y);
/*HYPRE_Int idxstride_b = hypre_VectorIndexStride(b);
HYPRE_Int vecstride_b = hypre_VectorVectorStride(b);*/
HYPRE_Int idxstride_x = hypre_VectorIndexStride(x);
HYPRE_Int vecstride_x = hypre_VectorVectorStride(x);
HYPRE_Complex temp, tempx;
HYPRE_Int i, j, jj;
HYPRE_Int m;
HYPRE_Real xpar=0.7;
HYPRE_Int ierr = 0;
hypre_Vector *x_tmp = NULL;
/*---------------------------------------------------------------------
* Check for size compatibility. Matvec returns ierr = 1 if
* length of X doesn't equal the number of columns of A,
* ierr = 2 if the length of Y doesn't equal the number of rows
* of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in Matvec, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
hypre_assert( num_vectors == hypre_VectorNumVectors(y) );
hypre_assert( num_vectors == hypre_VectorNumVectors(b) );
if (num_cols != x_size)
ierr = 1;
if (num_rows != y_size || num_rows != b_size)
ierr = 2;
if (num_cols != x_size && (num_rows != y_size || num_rows != b_size))
ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] *= beta;
#ifdef HYPRE_PROFILE
hypre_profile_times[HYPRE_TIMER_ID_MATVEC] += hypre_MPI_Wtime() - time_begin;
#endif
return ierr;
}
if (x == y)
{
x_tmp = hypre_SeqVectorCloneDeep(x);
x_data = hypre_VectorData(x_tmp);
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
/* use rownnz pointer to do the A*x multiplication when num_rownnz is smaller than num_rows */
if (num_rownnz < xpar*(num_rows) || num_vectors > 1)
{
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
if (temp != 1.0)
{
if (temp == 0.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] = 0.0;
}
else
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] = b_data[i]*temp;
}
}
/*-----------------------------------------------------------------
* y += A*x
*-----------------------------------------------------------------*/
if (num_rownnz < xpar*(num_rows))
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i,j,jj,m,tempx) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rownnz; i++)
{
m = A_rownnz[i];
/*
* for (jj = A_i[m]; jj < A_i[m+1]; jj++)
* {
* j = A_j[jj];
* y_data[m] += A_data[jj] * x_data[j];
* } */
if ( num_vectors==1 )
{
tempx = 0;
for (jj = A_i[m]; jj < A_i[m+1]; jj++)
tempx += A_data[jj] * x_data[A_j[jj]];
y_data[m] += tempx;
}
else
for ( j=0; j<num_vectors; ++j )
{
tempx = 0;
for (jj = A_i[m]; jj < A_i[m+1]; jj++)
tempx += A_data[jj] * x_data[ j*vecstride_x + A_j[jj]*idxstride_x ];
y_data[ j*vecstride_y + m*idxstride_y] += tempx;
}
}
}
else // num_vectors > 1
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i,j,jj,tempx) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
{
for (j = 0; j < num_vectors; ++j)
{
tempx = 0;
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx += A_data[jj] * x_data[ j*vecstride_x + A_j[jj]*idxstride_x ];
}
y_data[ j*vecstride_y + i*idxstride_y ] += tempx;
}
}
}
/*-----------------------------------------------------------------
* y = alpha*y
*-----------------------------------------------------------------*/
if (alpha != 1.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] *= alpha;
}
}
else
{ // JSP: this is currently the only path optimized
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel private(i,jj,tempx)
#endif
{
HYPRE_Int iBegin = hypre_CSRMatrixGetLoadBalancedPartitionBegin(A);
HYPRE_Int iEnd = hypre_CSRMatrixGetLoadBalancedPartitionEnd(A);
hypre_assert(iBegin <= iEnd);
hypre_assert(iBegin >= 0 && iBegin <= num_rows);
hypre_assert(iEnd >= 0 && iEnd <= num_rows);
if (0 == temp)
{
if (1 == alpha) // JSP: a common path
{
for (i = iBegin; i < iEnd; i++)
{
tempx = 0.0;
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx += A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = tempx;
}
} // y = A*x
else if (-1 == alpha)
{
for (i = iBegin; i < iEnd; i++)
{
tempx = 0.0;
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx -= A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = tempx;
}
} // y = -A*x
else
{
for (i = iBegin; i < iEnd; i++)
{
tempx = 0.0;
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx += A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = alpha*tempx;
}
} // y = alpha*A*x
} // temp == 0
else if (-1 == temp) // beta == -alpha
{
if (1 == alpha) // JSP: a common path
{
for (i = iBegin; i < iEnd; i++)
{
tempx = -b_data[i];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx += A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = tempx;
}
} // y = A*x - y
else if (-1 == alpha) // JSP: a common path
{
for (i = iBegin; i < iEnd; i++)
{
tempx = b_data[i];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx -= A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = tempx;
}
} // y = -A*x + y
else
{
for (i = iBegin; i < iEnd; i++)
{
tempx = -b_data[i];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx += A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = alpha*tempx;
}
} // y = alpha*(A*x - y)
} // temp == -1
else if (1 == temp)
{
if (1 == alpha) // JSP: a common path
{
for (i = iBegin; i < iEnd; i++)
{
tempx = b_data[i];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx += A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = tempx;
}
} // y = A*x + y
else if (-1 == alpha)
{
for (i = iBegin; i < iEnd; i++)
{
tempx = -b_data[i];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx -= A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = tempx;
}
} // y = -A*x - y
else
{
for (i = iBegin; i < iEnd; i++)
{
tempx = b_data[i];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx += A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = alpha*tempx;
}
} // y = alpha*(A*x + y)
}
else
{
if (1 == alpha) // JSP: a common path
{
for (i = iBegin; i < iEnd; i++)
{
tempx = b_data[i]*temp;
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx += A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = tempx;
}
} // y = A*x + temp*y
else if (-1 == alpha)
{
for (i = iBegin; i < iEnd; i++)
{
tempx = -b_data[i]*temp;
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx -= A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = tempx;
}
} // y = -A*x - temp*y
else
{
for (i = iBegin; i < iEnd; i++)
{
tempx = b_data[i]*temp;
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
tempx += A_data[jj] * x_data[A_j[jj]];
}
y_data[i] = alpha*tempx;
}
} // y = alpha*(A*x + temp*y)
} // temp != 0 && temp != -1 && temp != 1
} // omp parallel
}
if (x == y) hypre_SeqVectorDestroy(x_tmp);
#ifdef HYPRE_PROFILE
hypre_profile_times[HYPRE_TIMER_ID_MATVEC] += hypre_MPI_Wtime() - time_begin;
#endif
return ierr;
}
HYPRE_Int
hypre_BoomerAMGCycle( void *amg_vdata,
hypre_ParVector **F_array,
hypre_ParVector **U_array )
{
hypre_ParAMGData *amg_data = amg_vdata;
HYPRE_Solver *smoother;
/* Data Structure variables */
hypre_ParCSRMatrix **A_array;
hypre_ParCSRMatrix **P_array;
hypre_ParCSRMatrix **R_array;
hypre_ParVector *Utemp;
hypre_ParVector *Vtemp;
hypre_ParVector *Rtemp;
hypre_ParVector *Ptemp;
hypre_ParVector *Ztemp;
hypre_ParVector *Aux_U;
hypre_ParVector *Aux_F;
hypre_ParCSRBlockMatrix **A_block_array;
hypre_ParCSRBlockMatrix **P_block_array;
hypre_ParCSRBlockMatrix **R_block_array;
HYPRE_Real *Ztemp_data;
HYPRE_Real *Ptemp_data;
HYPRE_Int **CF_marker_array;
/* HYPRE_Int **unknown_map_array;
HYPRE_Int **point_map_array;
HYPRE_Int **v_at_point_array; */
HYPRE_Real cycle_op_count;
HYPRE_Int cycle_type;
HYPRE_Int num_levels;
HYPRE_Int max_levels;
HYPRE_Real *num_coeffs;
HYPRE_Int *num_grid_sweeps;
HYPRE_Int *grid_relax_type;
HYPRE_Int **grid_relax_points;
HYPRE_Int block_mode;
HYPRE_Real *max_eig_est;
HYPRE_Real *min_eig_est;
HYPRE_Int cheby_order;
HYPRE_Real cheby_fraction;
/* Local variables */
HYPRE_Int *lev_counter;
HYPRE_Int Solve_err_flag;
HYPRE_Int k;
HYPRE_Int i, j, jj;
HYPRE_Int level;
HYPRE_Int cycle_param;
HYPRE_Int coarse_grid;
HYPRE_Int fine_grid;
HYPRE_Int Not_Finished;
HYPRE_Int num_sweep;
HYPRE_Int cg_num_sweep = 1;
HYPRE_Int relax_type;
HYPRE_Int relax_points;
HYPRE_Int relax_order;
HYPRE_Int relax_local;
HYPRE_Int old_version = 0;
HYPRE_Real *relax_weight;
HYPRE_Real *omega;
HYPRE_Real alfa, beta, gammaold;
HYPRE_Real gamma = 1.0;
HYPRE_Int local_size;
/* HYPRE_Int *smooth_option; */
HYPRE_Int smooth_type;
HYPRE_Int smooth_num_levels;
HYPRE_Int num_threads, my_id;
HYPRE_Real alpha;
HYPRE_Real **l1_norms = NULL;
HYPRE_Real *l1_norms_level;
HYPRE_Int seq_cg = 0;
MPI_Comm comm;
#if 0
HYPRE_Real *D_mat;
HYPRE_Real *S_vec;
#endif
/* Acquire data and allocate storage */
num_threads = hypre_NumThreads();
A_array = hypre_ParAMGDataAArray(amg_data);
P_array = hypre_ParAMGDataPArray(amg_data);
R_array = hypre_ParAMGDataRArray(amg_data);
CF_marker_array = hypre_ParAMGDataCFMarkerArray(amg_data);
Vtemp = hypre_ParAMGDataVtemp(amg_data);
Rtemp = hypre_ParAMGDataRtemp(amg_data);
Ptemp = hypre_ParAMGDataPtemp(amg_data);
Ztemp = hypre_ParAMGDataZtemp(amg_data);
num_levels = hypre_ParAMGDataNumLevels(amg_data);
max_levels = hypre_ParAMGDataMaxLevels(amg_data);
cycle_type = hypre_ParAMGDataCycleType(amg_data);
A_block_array = hypre_ParAMGDataABlockArray(amg_data);
P_block_array = hypre_ParAMGDataPBlockArray(amg_data);
R_block_array = hypre_ParAMGDataRBlockArray(amg_data);
block_mode = hypre_ParAMGDataBlockMode(amg_data);
num_grid_sweeps = hypre_ParAMGDataNumGridSweeps(amg_data);
grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data);
grid_relax_points = hypre_ParAMGDataGridRelaxPoints(amg_data);
relax_order = hypre_ParAMGDataRelaxOrder(amg_data);
relax_weight = hypre_ParAMGDataRelaxWeight(amg_data);
omega = hypre_ParAMGDataOmega(amg_data);
smooth_type = hypre_ParAMGDataSmoothType(amg_data);
smooth_num_levels = hypre_ParAMGDataSmoothNumLevels(amg_data);
l1_norms = hypre_ParAMGDataL1Norms(amg_data);
/* smooth_option = hypre_ParAMGDataSmoothOption(amg_data); */
max_eig_est = hypre_ParAMGDataMaxEigEst(amg_data);
min_eig_est = hypre_ParAMGDataMinEigEst(amg_data);
cheby_order = hypre_ParAMGDataChebyOrder(amg_data);
cheby_fraction = hypre_ParAMGDataChebyFraction(amg_data);
cycle_op_count = hypre_ParAMGDataCycleOpCount(amg_data);
lev_counter = hypre_CTAlloc(HYPRE_Int, num_levels);
if (hypre_ParAMGDataParticipate(amg_data)) seq_cg = 1;
/* Initialize */
Solve_err_flag = 0;
if (grid_relax_points) old_version = 1;
num_coeffs = hypre_CTAlloc(HYPRE_Real, num_levels);
num_coeffs[0] = hypre_ParCSRMatrixDNumNonzeros(A_array[0]);
comm = hypre_ParCSRMatrixComm(A_array[0]);
hypre_MPI_Comm_rank(comm,&my_id);
if (block_mode)
{
for (j = 1; j < num_levels; j++)
num_coeffs[j] = hypre_ParCSRBlockMatrixNumNonzeros(A_block_array[j]);
}
else
{
for (j = 1; j < num_levels; j++)
num_coeffs[j] = hypre_ParCSRMatrixDNumNonzeros(A_array[j]);
}
/*---------------------------------------------------------------------
* Initialize cycling control counter
*
* Cycling is controlled using a level counter: lev_counter[k]
*
* Each time relaxation is performed on level k, the
* counter is decremented by 1. If the counter is then
* negative, we go to the next finer level. If non-
* negative, we go to the next coarser level. The
* following actions control cycling:
*
* a. lev_counter[0] is initialized to 1.
* b. lev_counter[k] is initialized to cycle_type for k>0.
*
* c. During cycling, when going down to level k, lev_counter[k]
* is set to the max of (lev_counter[k],cycle_type)
*---------------------------------------------------------------------*/
Not_Finished = 1;
lev_counter[0] = 1;
for (k = 1; k < num_levels; ++k)
{
lev_counter[k] = cycle_type;
}
level = 0;
cycle_param = 1;
smoother = hypre_ParAMGDataSmoother(amg_data);
if (smooth_num_levels > 0)
{
if (smooth_type == 7 || smooth_type == 8
|| smooth_type == 17 || smooth_type == 18
|| smooth_type == 9 || smooth_type == 19)
{
HYPRE_Int actual_local_size = hypre_ParVectorActualLocalSize(Vtemp);
Utemp = hypre_ParVectorCreate(comm,hypre_ParVectorGlobalSize(Vtemp),
hypre_ParVectorPartitioning(Vtemp));
hypre_ParVectorOwnsPartitioning(Utemp) = 0;
local_size
= hypre_VectorSize(hypre_ParVectorLocalVector(Vtemp));
if (local_size < actual_local_size)
{
hypre_VectorData(hypre_ParVectorLocalVector(Utemp)) =
hypre_CTAlloc(HYPRE_Complex, actual_local_size);
hypre_ParVectorActualLocalSize(Utemp) = actual_local_size;
}
else
hypre_ParVectorInitialize(Utemp);
}
}
/*---------------------------------------------------------------------
* Main loop of cycling
*--------------------------------------------------------------------*/
while (Not_Finished)
{
if (num_levels > 1)
{
local_size
= hypre_VectorSize(hypre_ParVectorLocalVector(F_array[level]));
hypre_VectorSize(hypre_ParVectorLocalVector(Vtemp)) = local_size;
if (smooth_num_levels <= level)
{
cg_num_sweep = 1;
num_sweep = num_grid_sweeps[cycle_param];
Aux_U = U_array[level];
Aux_F = F_array[level];
}
else if (smooth_type > 9)
{
hypre_VectorSize(hypre_ParVectorLocalVector(Ztemp)) = local_size;
hypre_VectorSize(hypre_ParVectorLocalVector(Rtemp)) = local_size;
hypre_VectorSize(hypre_ParVectorLocalVector(Ptemp)) = local_size;
Ztemp_data = hypre_VectorData(hypre_ParVectorLocalVector(Ztemp));
Ptemp_data = hypre_VectorData(hypre_ParVectorLocalVector(Ptemp));
hypre_ParVectorSetConstantValues(Ztemp,0);
alpha = -1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[level],
U_array[level], beta, F_array[level], Rtemp);
cg_num_sweep = hypre_ParAMGDataSmoothNumSweeps(amg_data);
num_sweep = num_grid_sweeps[cycle_param];
Aux_U = Ztemp;
Aux_F = Rtemp;
}
else
{
cg_num_sweep = 1;
num_sweep = hypre_ParAMGDataSmoothNumSweeps(amg_data);
Aux_U = U_array[level];
Aux_F = F_array[level];
}
relax_type = grid_relax_type[cycle_param];
}
else /* AB: 4/08: removed the max_levels > 1 check - should do this when max-levels = 1 also */
{
/* If no coarsening occurred, apply a simple smoother once */
Aux_U = U_array[level];
Aux_F = F_array[level];
num_sweep = 1;
/* TK: Use the user relax type (instead of 0) to allow for setting a
convergent smoother (e.g. in the solution of singular problems). */
relax_type = hypre_ParAMGDataUserRelaxType(amg_data);
}
if (l1_norms != NULL)
l1_norms_level = l1_norms[level];
else
l1_norms_level = NULL;
if (cycle_param == 3 && seq_cg)
{
hypre_seqAMGCycle(amg_data, level, F_array, U_array);
}
else
{
/*------------------------------------------------------------------
* Do the relaxation num_sweep times
*-----------------------------------------------------------------*/
for (jj = 0; jj < cg_num_sweep; jj++)
{
if (smooth_num_levels > level && smooth_type > 9)
hypre_ParVectorSetConstantValues(Aux_U,0);
for (j = 0; j < num_sweep; j++)
{
if (num_levels == 1 && max_levels > 1)
{
relax_points = 0;
relax_local = 0;
}
else
{
if (old_version)
relax_points = grid_relax_points[cycle_param][j];
relax_local = relax_order;
}
/*-----------------------------------------------
* VERY sloppy approximation to cycle complexity
*-----------------------------------------------*/
if (old_version && level < num_levels -1)
{
switch (relax_points)
{
case 1:
cycle_op_count += num_coeffs[level+1];
break;
case -1:
cycle_op_count += (num_coeffs[level]-num_coeffs[level+1]);
break;
}
}
else
{
cycle_op_count += num_coeffs[level];
}
/*-----------------------------------------------
Choose Smoother
-----------------------------------------------*/
if (smooth_num_levels > level &&
(smooth_type == 7 || smooth_type == 8 ||
smooth_type == 9 || smooth_type == 19 ||
smooth_type == 17 || smooth_type == 18))
{
hypre_VectorSize(hypre_ParVectorLocalVector(Utemp)) = local_size;
alpha = -1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[level],
U_array[level], beta, Aux_F, Vtemp);
if (smooth_type == 8 || smooth_type == 18)
HYPRE_ParCSRParaSailsSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Vtemp,
(HYPRE_ParVector) Utemp);
else if (smooth_type == 7 || smooth_type == 17)
HYPRE_ParCSRPilutSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Vtemp,
(HYPRE_ParVector) Utemp);
else if (smooth_type == 9 || smooth_type == 19)
HYPRE_EuclidSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Vtemp,
(HYPRE_ParVector) Utemp);
hypre_ParVectorAxpy(relax_weight[level],Utemp,Aux_U);
}
else if (smooth_num_levels > level &&
(smooth_type == 6 || smooth_type == 16))
{
HYPRE_SchwarzSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Aux_F,
(HYPRE_ParVector) Aux_U);
}
/*else if (relax_type == 99)*/
else if (relax_type == 9 || relax_type == 99)
{ /* Gaussian elimination */
hypre_GaussElimSolve(amg_data, level, relax_type);
}
else if (relax_type == 18)
{ /* L1 - Jacobi*/
if (relax_order == 1 && cycle_param < 3)
{
/* need to do CF - so can't use the AMS one */
HYPRE_Int i;
HYPRE_Int loc_relax_points[2];
if (cycle_type < 2)
{
loc_relax_points[0] = 1;
loc_relax_points[1] = -1;
}
else
{
loc_relax_points[0] = -1;
loc_relax_points[1] = 1;
}
for (i=0; i < 2; i++)
hypre_ParCSRRelax_L1_Jacobi(A_array[level],
Aux_F,
CF_marker_array[level],
loc_relax_points[i],
relax_weight[level],
l1_norms[level],
Aux_U,
Vtemp);
}
else /* not CF - so use through AMS */
{
if (num_threads == 1)
hypre_ParCSRRelax(A_array[level],
Aux_F,
1,
1,
l1_norms_level,
relax_weight[level],
omega[level],0,0,0,0,
Aux_U,
Vtemp,
Ztemp);
else
hypre_ParCSRRelaxThreads(A_array[level],
Aux_F,
1,
1,
l1_norms_level,
relax_weight[level],
omega[level],
Aux_U,
Vtemp,
Ztemp);
}
}
else if (relax_type == 15)
{ /* CG */
if (j ==0) /* do num sweep iterations of CG */
hypre_ParCSRRelax_CG( smoother[level],
A_array[level],
Aux_F,
Aux_U,
num_sweep);
}
else if (relax_type == 16)
{ /* scaled Chebyshev */
HYPRE_Int scale = 1;
HYPRE_Int variant = 0;
hypre_ParCSRRelax_Cheby(A_array[level],
Aux_F,
max_eig_est[level],
min_eig_est[level],
cheby_fraction, cheby_order, scale,
variant, Aux_U, Vtemp, Ztemp );
}
else if (relax_type ==17)
{
hypre_BoomerAMGRelax_FCFJacobi(A_array[level],
Aux_F,
CF_marker_array[level],
relax_weight[level],
Aux_U,
Vtemp);
}
else if (old_version)
{
Solve_err_flag = hypre_BoomerAMGRelax(A_array[level],
Aux_F,
CF_marker_array[level],
relax_type, relax_points,
relax_weight[level],
omega[level],
l1_norms_level,
Aux_U,
Vtemp,
Ztemp);
}
else
{
/* smoother than can have CF ordering */
if (block_mode)
{
Solve_err_flag = hypre_BoomerAMGBlockRelaxIF(A_block_array[level],
Aux_F,
CF_marker_array[level],
relax_type,
relax_local,
cycle_param,
relax_weight[level],
omega[level],
Aux_U,
Vtemp);
}
else
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(A_array[level],
Aux_F,
CF_marker_array[level],
relax_type,
relax_local,
cycle_param,
relax_weight[level],
omega[level],
l1_norms_level,
Aux_U,
Vtemp,
Ztemp);
}
}
if (Solve_err_flag != 0)
return(Solve_err_flag);
}
if (smooth_num_levels > level && smooth_type > 9)
{
gammaold = gamma;
gamma = hypre_ParVectorInnerProd(Rtemp,Ztemp);
if (jj == 0)
hypre_ParVectorCopy(Ztemp,Ptemp);
else
{
beta = gamma/gammaold;
for (i=0; i < local_size; i++)
Ptemp_data[i] = Ztemp_data[i] + beta*Ptemp_data[i];
}
hypre_ParCSRMatrixMatvec(1.0,A_array[level],Ptemp,0.0,Vtemp);
alfa = gamma /hypre_ParVectorInnerProd(Ptemp,Vtemp);
hypre_ParVectorAxpy(alfa,Ptemp,U_array[level]);
hypre_ParVectorAxpy(-alfa,Vtemp,Rtemp);
}
}
}
/*------------------------------------------------------------------
* Decrement the control counter and determine which grid to visit next
*-----------------------------------------------------------------*/
--lev_counter[level];
if (lev_counter[level] >= 0 && level != num_levels-1)
{
/*---------------------------------------------------------------
* Visit coarser level next.
* Compute residual using hypre_ParCSRMatrixMatvec.
* Perform restriction using hypre_ParCSRMatrixMatvecT.
* Reset counters and cycling parameters for coarse level
*--------------------------------------------------------------*/
fine_grid = level;
coarse_grid = level + 1;
hypre_ParVectorSetConstantValues(U_array[coarse_grid], 0.0);
alpha = -1.0;
beta = 1.0;
if (block_mode)
{
hypre_ParVectorCopy(F_array[fine_grid],Vtemp);
hypre_ParCSRBlockMatrixMatvec(alpha, A_block_array[fine_grid], U_array[fine_grid],
beta, Vtemp);
}
else
{
// JSP: avoid unnecessary copy using out-of-place version of SpMV
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[fine_grid], U_array[fine_grid],
beta, F_array[fine_grid], Vtemp);
}
alpha = 1.0;
beta = 0.0;
if (block_mode)
{
hypre_ParCSRBlockMatrixMatvecT(alpha,R_block_array[fine_grid],Vtemp,
beta,F_array[coarse_grid]);
}
else
{
hypre_ParCSRMatrixMatvecT(alpha,R_array[fine_grid],Vtemp,
beta,F_array[coarse_grid]);
}
++level;
lev_counter[level] = hypre_max(lev_counter[level],cycle_type);
cycle_param = 1;
if (level == num_levels-1) cycle_param = 3;
}
else if (level != 0)
{
/*---------------------------------------------------------------
* Visit finer level next.
* Interpolate and add correction using hypre_ParCSRMatrixMatvec.
* Reset counters and cycling parameters for finer level.
*--------------------------------------------------------------*/
fine_grid = level - 1;
coarse_grid = level;
alpha = 1.0;
beta = 1.0;
if (block_mode)
{
hypre_ParCSRBlockMatrixMatvec(alpha, P_block_array[fine_grid],
U_array[coarse_grid],
beta, U_array[fine_grid]);
}
else
{
hypre_ParCSRMatrixMatvec(alpha, P_array[fine_grid],
U_array[coarse_grid],
beta, U_array[fine_grid]);
}
--level;
cycle_param = 2;
}
else
{
Not_Finished = 0;
}
}
hypre_ParAMGDataCycleOpCount(amg_data) = cycle_op_count;
hypre_TFree(lev_counter);
hypre_TFree(num_coeffs);
if (smooth_num_levels > 0)
{
if (smooth_type == 7 || smooth_type == 8 || smooth_type == 9 ||
smooth_type == 17 || smooth_type == 18 || smooth_type == 19 )
hypre_ParVectorDestroy(Utemp);
}
return(Solve_err_flag);
}
HYPRE_Int
hypre_BoomerAMGAdditiveCycle( void *amg_vdata)
{
hypre_ParAMGData *amg_data = amg_vdata;
/* Data Structure variables */
hypre_ParCSRMatrix **A_array;
hypre_ParCSRMatrix **P_array;
hypre_ParCSRMatrix **R_array;
hypre_ParCSRMatrix *Lambda;
hypre_ParVector **F_array;
hypre_ParVector **U_array;
hypre_ParVector *Vtemp;
hypre_ParVector *Ztemp;
hypre_ParVector *Xtilde, *Rtilde;
HYPRE_Int **CF_marker_array;
HYPRE_Int num_levels;
HYPRE_Int addlvl;
HYPRE_Int additive;
HYPRE_Int mult_additive;
HYPRE_Int simple;
HYPRE_Int i, num_rows;
HYPRE_Int n_global;
HYPRE_Int rlx_order;
/* Local variables */
HYPRE_Int Solve_err_flag = 0;
HYPRE_Int level;
HYPRE_Int coarse_grid;
HYPRE_Int fine_grid;
HYPRE_Int relax_type;
HYPRE_Int rlx_down;
HYPRE_Int rlx_up;
HYPRE_Int *grid_relax_type;
HYPRE_Real **l1_norms;
HYPRE_Real alpha, beta;
HYPRE_Int num_threads;
HYPRE_Real *u_data;
HYPRE_Real *f_data;
HYPRE_Real *v_data;
HYPRE_Real *l1_norms_lvl;
HYPRE_Real *D_inv;
HYPRE_Real *x_global;
HYPRE_Real *r_global;
HYPRE_Real *relax_weight;
HYPRE_Real *omega;
#if 0
HYPRE_Real *D_mat;
HYPRE_Real *S_vec;
#endif
/* Acquire data and allocate storage */
num_threads = hypre_NumThreads();
A_array = hypre_ParAMGDataAArray(amg_data);
F_array = hypre_ParAMGDataFArray(amg_data);
U_array = hypre_ParAMGDataUArray(amg_data);
P_array = hypre_ParAMGDataPArray(amg_data);
R_array = hypre_ParAMGDataRArray(amg_data);
CF_marker_array = hypre_ParAMGDataCFMarkerArray(amg_data);
Vtemp = hypre_ParAMGDataVtemp(amg_data);
Ztemp = hypre_ParAMGDataZtemp(amg_data);
num_levels = hypre_ParAMGDataNumLevels(amg_data);
additive = hypre_ParAMGDataAdditive(amg_data);
mult_additive = hypre_ParAMGDataMultAdditive(amg_data);
simple = hypre_ParAMGDataSimple(amg_data);
grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data);
Lambda = hypre_ParAMGDataLambda(amg_data);
Xtilde = hypre_ParAMGDataXtilde(amg_data);
Rtilde = hypre_ParAMGDataRtilde(amg_data);
l1_norms = hypre_ParAMGDataL1Norms(amg_data);
D_inv = hypre_ParAMGDataDinv(amg_data);
grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data);
relax_weight = hypre_ParAMGDataRelaxWeight(amg_data);
omega = hypre_ParAMGDataOmega(amg_data);
rlx_order = hypre_ParAMGDataRelaxOrder(amg_data);
/* Initialize */
addlvl = hypre_max(additive, mult_additive);
addlvl = hypre_max(addlvl, simple);
Solve_err_flag = 0;
/*---------------------------------------------------------------------
* Main loop of cycling --- multiplicative version --- V-cycle
*--------------------------------------------------------------------*/
/* down cycle */
relax_type = grid_relax_type[1];
rlx_down = grid_relax_type[1];
rlx_up = grid_relax_type[2];
for (level = 0; level < num_levels-1; level++)
{
fine_grid = level;
coarse_grid = level + 1;
u_data = hypre_VectorData(hypre_ParVectorLocalVector(U_array[fine_grid]));
f_data = hypre_VectorData(hypre_ParVectorLocalVector(F_array[fine_grid]));
v_data = hypre_VectorData(hypre_ParVectorLocalVector(Vtemp));
l1_norms_lvl = l1_norms[level];
hypre_ParVectorSetConstantValues(U_array[coarse_grid], 0.0);
if (level < addlvl) /* multiplicative version */
{
/* smoothing step */
if (rlx_down == 0)
{
HYPRE_Real *A_data = hypre_CSRMatrixData(hypre_ParCSRMatrixDiag(A_array[fine_grid]));
HYPRE_Int *A_i = hypre_CSRMatrixI(hypre_ParCSRMatrixDiag(A_array[fine_grid]));
hypre_ParVectorCopy(F_array[fine_grid],Vtemp);
num_rows = hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(A_array[fine_grid]));
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
u_data[i] = relax_weight[level]*v_data[i] / A_data[A_i[i]];
}
else if (rlx_down != 18)
{
/*hypre_BoomerAMGRelax(A_array[fine_grid],F_array[fine_grid],NULL,rlx_down,0,*/
hypre_BoomerAMGRelaxIF(A_array[fine_grid],F_array[fine_grid],
CF_marker_array[fine_grid], rlx_down,rlx_order,1,
relax_weight[fine_grid], omega[fine_grid],
l1_norms_lvl, U_array[fine_grid], Vtemp, Ztemp);
hypre_ParVectorCopy(F_array[fine_grid],Vtemp);
}
else
{
hypre_ParVectorCopy(F_array[fine_grid],Vtemp);
num_rows = hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(A_array[fine_grid]));
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
u_data[i] += v_data[i] / l1_norms_lvl[i];
}
alpha = -1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvec(alpha, A_array[fine_grid], U_array[fine_grid],
beta, Vtemp);
alpha = 1.0;
beta = 0.0;
hypre_ParCSRMatrixMatvecT(alpha,R_array[fine_grid],Vtemp,
beta,F_array[coarse_grid]);
}
else /* additive version */
{
hypre_ParVectorCopy(F_array[fine_grid],Vtemp);
if (level == 0) /* compute residual */
{
hypre_ParVectorCopy(Vtemp, Rtilde);
hypre_ParVectorCopy(U_array[fine_grid],Xtilde);
}
alpha = 1.0;
beta = 0.0;
hypre_ParCSRMatrixMatvecT(alpha,R_array[fine_grid],Vtemp,
beta,F_array[coarse_grid]);
}
}
/* solve coarse grid */
if (addlvl < num_levels)
{
if (simple > -1)
{
x_global = hypre_VectorData(hypre_ParVectorLocalVector(Xtilde));
r_global = hypre_VectorData(hypre_ParVectorLocalVector(Rtilde));
n_global = hypre_VectorSize(hypre_ParVectorLocalVector(Xtilde));
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i=0; i < n_global; i++)
x_global[i] += D_inv[i]*r_global[i];
}
else
hypre_ParCSRMatrixMatvec(1.0, Lambda, Rtilde, 1.0, Xtilde);
if (addlvl == 0) hypre_ParVectorCopy(Xtilde, U_array[0]);
}
else
{
fine_grid = num_levels -1;
hypre_ParCSRRelax(A_array[fine_grid], F_array[fine_grid],
1, 1, l1_norms[fine_grid],
1.0, 1.0 ,0,0,0,0,
U_array[fine_grid], Vtemp, Ztemp);
}
/* up cycle */
relax_type = grid_relax_type[2];
for (level = num_levels-1; level > 0; level--)
{
fine_grid = level - 1;
coarse_grid = level;
if (level <= addlvl) /* multiplicative version */
{
alpha = 1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvec(alpha, P_array[fine_grid],
U_array[coarse_grid],
beta, U_array[fine_grid]);
if (rlx_up != 18)
/*hypre_BoomerAMGRelax(A_array[fine_grid],F_array[fine_grid],NULL,rlx_up,0,*/
hypre_BoomerAMGRelaxIF(A_array[fine_grid],F_array[fine_grid],
CF_marker_array[fine_grid],
rlx_up,rlx_order,2,
relax_weight[fine_grid], omega[fine_grid],
l1_norms[fine_grid], U_array[fine_grid], Vtemp, Ztemp);
else if (rlx_order)
{
HYPRE_Int loc_relax_points[2];
loc_relax_points[0] = -1;
loc_relax_points[1] = 1;
for (i=0; i < 2; i++)
hypre_ParCSRRelax_L1_Jacobi(A_array[fine_grid],F_array[fine_grid],
CF_marker_array[fine_grid],
loc_relax_points[i],
1.0, l1_norms[fine_grid],
U_array[fine_grid], Vtemp);
}
else
hypre_ParCSRRelax(A_array[fine_grid], F_array[fine_grid],
1, 1, l1_norms[fine_grid],
1.0, 1.0 ,0,0,0,0,
U_array[fine_grid], Vtemp, Ztemp);
}
else /* additive version */
{
alpha = 1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvec(alpha, P_array[fine_grid],
U_array[coarse_grid],
beta, U_array[fine_grid]);
}
}
return(Solve_err_flag);
}
HYPRE_Int
main( HYPRE_Int argc,
char *argv[] )
{
hypre_ParVector *vector1;
hypre_ParVector *vector2;
hypre_ParVector *tmp_vector;
HYPRE_Int num_procs, my_id;
HYPRE_Int global_size = 20;
HYPRE_Int local_size;
HYPRE_Int first_index;
HYPRE_Int i;
HYPRE_Int *partitioning;
HYPRE_Complex prod;
HYPRE_Complex *data, *data2;
hypre_Vector *vector;
hypre_Vector *local_vector;
hypre_Vector *local_vector2;
/* Initialize MPI */
hypre_MPI_Init(&argc, &argv);
hypre_MPI_Comm_size(hypre_MPI_COMM_WORLD, &num_procs );
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &my_id );
hypre_printf(" my_id: %d num_procs: %d\n", my_id, num_procs);
partitioning = NULL;
vector1 = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_size,partitioning);
partitioning = hypre_ParVectorPartitioning(vector1);
hypre_ParVectorInitialize(vector1);
local_vector = hypre_ParVectorLocalVector(vector1);
data = hypre_VectorData(local_vector);
local_size = hypre_VectorSize(local_vector);
first_index = partitioning[my_id];
for (i=0; i < local_size; i++)
data[i] = first_index+i;
/*
hypre_ParVectorPrint(vector1, "Vector");
*/
local_vector2 = hypre_SeqVectorCreate(global_size);
hypre_SeqVectorInitialize(local_vector2);
data2 = hypre_VectorData(local_vector2);
for (i=0; i < global_size; i++)
data2[i] = i+1;
/* partitioning = hypre_CTAlloc(HYPRE_Int,4);
partitioning[0] = 0;
partitioning[1] = 10;
partitioning[2] = 10;
partitioning[3] = 20;
*/
vector2 = hypre_VectorToParVector(hypre_MPI_COMM_WORLD,local_vector2,partitioning);
hypre_ParVectorSetPartitioningOwner(vector2,0);
hypre_ParVectorPrint(vector2, "Convert");
vector = hypre_ParVectorToVectorAll(vector2);
/*-----------------------------------------------------------
* Copy the vector into tmp_vector
*-----------------------------------------------------------*/
tmp_vector = hypre_ParVectorRead(hypre_MPI_COMM_WORLD, "Convert");
/*
tmp_vector = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_size,partitioning);
hypre_ParVectorSetPartitioningOwner(tmp_vector,0);
hypre_ParVectorInitialize(tmp_vector);
hypre_ParVectorCopy(vector1, tmp_vector);
hypre_ParVectorPrint(tmp_vector,"Copy");
*/
/*-----------------------------------------------------------
* Scale tmp_vector
*-----------------------------------------------------------*/
hypre_ParVectorScale(2.0, tmp_vector);
/*
hypre_ParVectorPrint(tmp_vector,"Scale");
*/
/*-----------------------------------------------------------
* Do an Axpy (2*vector - vector) = vector
*-----------------------------------------------------------*/
hypre_ParVectorAxpy(-1.0, vector1, tmp_vector);
/*
hypre_ParVectorPrint(tmp_vector,"Axpy");
*/
/*-----------------------------------------------------------
* Do an inner product vector* tmp_vector
*-----------------------------------------------------------*/
prod = hypre_ParVectorInnerProd(vector1, tmp_vector);
hypre_printf (" prod: %8.2f \n", prod);
/*-----------------------------------------------------------
* Finalize things
*-----------------------------------------------------------*/
hypre_ParVectorDestroy(vector1);
hypre_ParVectorDestroy(vector2);
hypre_ParVectorDestroy(tmp_vector);
hypre_SeqVectorDestroy(local_vector2);
if (vector) hypre_SeqVectorDestroy(vector);
/* Finalize MPI */
hypre_MPI_Finalize();
return 0;
}
HYPRE_Int
hypre_seqAMGCycle( hypre_ParAMGData *amg_data,
HYPRE_Int p_level,
hypre_ParVector **Par_F_array,
hypre_ParVector **Par_U_array )
{
hypre_ParVector *Aux_U;
hypre_ParVector *Aux_F;
/* Local variables */
HYPRE_Int Solve_err_flag = 0;
HYPRE_Int n;
HYPRE_Int i;
hypre_Vector *u_local;
HYPRE_Real *u_data;
HYPRE_Int first_index;
/* Acquire seq data */
MPI_Comm new_comm = hypre_ParAMGDataNewComm(amg_data);
HYPRE_Solver coarse_solver = hypre_ParAMGDataCoarseSolver(amg_data);
hypre_ParCSRMatrix *A_coarse = hypre_ParAMGDataACoarse(amg_data);
hypre_ParVector *F_coarse = hypre_ParAMGDataFCoarse(amg_data);
hypre_ParVector *U_coarse = hypre_ParAMGDataUCoarse(amg_data);
HYPRE_Int redundant = hypre_ParAMGDataRedundant(amg_data);
Aux_U = Par_U_array[p_level];
Aux_F = Par_F_array[p_level];
first_index = hypre_ParVectorFirstIndex(Aux_U);
u_local = hypre_ParVectorLocalVector(Aux_U);
u_data = hypre_VectorData(u_local);
n = hypre_VectorSize(u_local);
/*if (A_coarse)*/
if (hypre_ParAMGDataParticipate(amg_data))
{
HYPRE_Real *f_data;
hypre_Vector *f_local;
hypre_Vector *tmp_vec;
HYPRE_Int nf;
HYPRE_Int local_info;
HYPRE_Real *recv_buf = NULL;
HYPRE_Int *displs = NULL;
HYPRE_Int *info = NULL;
HYPRE_Int size;
HYPRE_Int new_num_procs, my_id;
hypre_MPI_Comm_size(new_comm, &new_num_procs);
hypre_MPI_Comm_rank(new_comm, &my_id);
f_local = hypre_ParVectorLocalVector(Aux_F);
f_data = hypre_VectorData(f_local);
nf = hypre_VectorSize(f_local);
/* first f */
info = hypre_CTAlloc(HYPRE_Int, new_num_procs);
local_info = nf;
if (redundant)
hypre_MPI_Allgather(&local_info, 1, HYPRE_MPI_INT, info, 1, HYPRE_MPI_INT, new_comm);
else
hypre_MPI_Gather(&local_info, 1, HYPRE_MPI_INT, info, 1, HYPRE_MPI_INT, 0, new_comm);
if (redundant || my_id ==0)
{
displs = hypre_CTAlloc(HYPRE_Int, new_num_procs+1);
displs[0] = 0;
for (i=1; i < new_num_procs+1; i++)
displs[i] = displs[i-1]+info[i-1];
size = displs[new_num_procs];
if (F_coarse)
{
tmp_vec = hypre_ParVectorLocalVector(F_coarse);
recv_buf = hypre_VectorData(tmp_vec);
}
}
if (redundant)
hypre_MPI_Allgatherv ( f_data, nf, HYPRE_MPI_REAL,
recv_buf, info, displs,
HYPRE_MPI_REAL, new_comm );
else
hypre_MPI_Gatherv ( f_data, nf, HYPRE_MPI_REAL,
recv_buf, info, displs,
HYPRE_MPI_REAL, 0, new_comm );
if (redundant || my_id ==0)
{
tmp_vec = hypre_ParVectorLocalVector(U_coarse);
recv_buf = hypre_VectorData(tmp_vec);
}
/*then u */
if (redundant)
{
hypre_MPI_Allgatherv ( u_data, n, HYPRE_MPI_REAL,
recv_buf, info, displs,
HYPRE_MPI_REAL, new_comm );
hypre_TFree(displs);
hypre_TFree(info);
}
else
hypre_MPI_Gatherv ( u_data, n, HYPRE_MPI_REAL,
recv_buf, info, displs,
HYPRE_MPI_REAL, 0, new_comm );
/* clean up */
if (redundant || my_id ==0)
{
hypre_BoomerAMGSolve(coarse_solver, A_coarse, F_coarse, U_coarse);
}
/*copy my part of U to parallel vector */
if (redundant)
{
HYPRE_Real *local_data;
local_data = hypre_VectorData(hypre_ParVectorLocalVector(U_coarse));
for (i = 0; i < n; i++)
{
u_data[i] = local_data[first_index+i];
}
}
else
{
HYPRE_Real *local_data=NULL;
if (my_id == 0)
local_data = hypre_VectorData(hypre_ParVectorLocalVector(U_coarse));
hypre_MPI_Scatterv ( local_data, info, displs, HYPRE_MPI_REAL,
u_data, n, HYPRE_MPI_REAL, 0, new_comm );
/*if (my_id == 0)
local_data = hypre_VectorData(hypre_ParVectorLocalVector(F_coarse));
hypre_MPI_Scatterv ( local_data, info, displs, HYPRE_MPI_REAL,
f_data, n, HYPRE_MPI_REAL, 0, new_comm );*/
if (my_id == 0) hypre_TFree(displs);
hypre_TFree(info);
}
}
return(Solve_err_flag);
}
HYPRE_Int
main( HYPRE_Int argc,
char *argv[] )
{
hypre_ParVector *vector1;
hypre_ParVector *vector2;
hypre_ParVector *tmp_vector;
HYPRE_Int num_procs, my_id;
HYPRE_Int global_size = 20;
HYPRE_Int local_size;
HYPRE_Int first_index;
HYPRE_Int num_vectors, vecstride, idxstride;
HYPRE_Int i, j;
HYPRE_Int *partitioning;
double prod;
double *data, *data2;
hypre_Vector *vector;
hypre_Vector *local_vector;
hypre_Vector *local_vector2;
/* Initialize MPI */
hypre_MPI_Init(&argc, &argv);
hypre_MPI_Comm_size(hypre_MPI_COMM_WORLD, &num_procs );
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &my_id );
hypre_printf(" my_id: %d num_procs: %d\n", my_id, num_procs);
partitioning = NULL;
num_vectors = 3;
vector1 = hypre_ParMultiVectorCreate
( hypre_MPI_COMM_WORLD, global_size, partitioning, num_vectors );
partitioning = hypre_ParVectorPartitioning(vector1);
hypre_ParVectorInitialize(vector1);
local_vector = hypre_ParVectorLocalVector(vector1);
data = hypre_VectorData(local_vector);
local_size = hypre_VectorSize(local_vector);
vecstride = hypre_VectorVectorStride(local_vector);
idxstride = hypre_VectorIndexStride(local_vector);
first_index = partitioning[my_id];
hypre_printf("vecstride=%i idxstride=%i local_size=%i num_vectors=%i",
vecstride, idxstride, local_size, num_vectors );
for (j=0; j<num_vectors; ++j )
for (i=0; i < local_size; i++)
data[ j*vecstride + i*idxstride ] = first_index+i + 100*j;
hypre_ParVectorPrint(vector1, "Vector");
local_vector2 = hypre_SeqMultiVectorCreate( global_size, num_vectors );
hypre_SeqVectorInitialize(local_vector2);
data2 = hypre_VectorData(local_vector2);
vecstride = hypre_VectorVectorStride(local_vector2);
idxstride = hypre_VectorIndexStride(local_vector2);
for (j=0; j<num_vectors; ++j )
for (i=0; i < global_size; i++)
data2[ j*vecstride + i*idxstride ] = i + 100*j;
/* partitioning = hypre_CTAlloc(HYPRE_Int,4);
partitioning[0] = 0;
partitioning[1] = 10;
partitioning[2] = 10;
partitioning[3] = 20;
*/
partitioning = hypre_CTAlloc(HYPRE_Int,1+num_procs);
hypre_GeneratePartitioning( global_size, num_procs, &partitioning );
vector2 = hypre_VectorToParVector(hypre_MPI_COMM_WORLD,local_vector2,partitioning);
hypre_ParVectorSetPartitioningOwner(vector2,0);
hypre_ParVectorPrint(vector2, "Convert");
vector = hypre_ParVectorToVectorAll(vector2);
/*-----------------------------------------------------------
* Copy the vector into tmp_vector
*-----------------------------------------------------------*/
/* Read doesn't work for multivectors yet...
tmp_vector = hypre_ParVectorRead(hypre_MPI_COMM_WORLD, "Convert");*/
tmp_vector = hypre_ParMultiVectorCreate
( hypre_MPI_COMM_WORLD, global_size, partitioning, num_vectors );
hypre_ParVectorInitialize( tmp_vector );
hypre_ParVectorCopy( vector2, tmp_vector );
/*
tmp_vector = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_size,partitioning);
hypre_ParVectorSetPartitioningOwner(tmp_vector,0);
hypre_ParVectorInitialize(tmp_vector);
hypre_ParVectorCopy(vector1, tmp_vector);
hypre_ParVectorPrint(tmp_vector,"Copy");
*/
/*-----------------------------------------------------------
* Scale tmp_vector
*-----------------------------------------------------------*/
hypre_ParVectorScale(2.0, tmp_vector);
hypre_ParVectorPrint(tmp_vector,"Scale");
/*-----------------------------------------------------------
* Do an Axpy (2*vector - vector) = vector
*-----------------------------------------------------------*/
hypre_ParVectorAxpy(-1.0, vector1, tmp_vector);
hypre_ParVectorPrint(tmp_vector,"Axpy");
/*-----------------------------------------------------------
* Do an inner product vector* tmp_vector
*-----------------------------------------------------------*/
prod = hypre_ParVectorInnerProd(vector1, tmp_vector);
hypre_printf (" prod: %8.2f \n", prod);
/*-----------------------------------------------------------
* Finalize things
*-----------------------------------------------------------*/
hypre_ParVectorDestroy(vector1);
hypre_ParVectorDestroy(vector2);
hypre_ParVectorDestroy(tmp_vector);
hypre_SeqVectorDestroy(local_vector2);
if (vector) hypre_SeqVectorDestroy(vector);
/* Finalize MPI */
hypre_MPI_Finalize();
return 0;
}
