HYPRE_Int
hypre_ParCSRMatrixMatvecT( HYPRE_Complex alpha,
hypre_ParCSRMatrix *A,
hypre_ParVector *x,
HYPRE_Complex beta,
hypre_ParVector *y )
{
hypre_ParCSRCommHandle **comm_handle;
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg(A);
hypre_CSRMatrix *diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix *offd = hypre_ParCSRMatrixOffd(A);
hypre_Vector *x_local = hypre_ParVectorLocalVector(x);
hypre_Vector *y_local = hypre_ParVectorLocalVector(y);
hypre_Vector *y_tmp;
HYPRE_Int vecstride = hypre_VectorVectorStride( y_local );
HYPRE_Int idxstride = hypre_VectorIndexStride( y_local );
HYPRE_Complex *y_tmp_data, **y_buf_data;
HYPRE_Complex *y_local_data = hypre_VectorData(y_local);
HYPRE_Int num_rows = hypre_ParCSRMatrixGlobalNumRows(A);
HYPRE_Int num_cols = hypre_ParCSRMatrixGlobalNumCols(A);
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols(offd);
HYPRE_Int x_size = hypre_ParVectorGlobalSize(x);
HYPRE_Int y_size = hypre_ParVectorGlobalSize(y);
HYPRE_Int num_vectors = hypre_VectorNumVectors(y_local);
HYPRE_Int i, j, jv, index, start, num_sends;
HYPRE_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.
*--------------------------------------------------------------------*/
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;
/*-----------------------------------------------------------------------
*-----------------------------------------------------------------------*/
comm_handle = hypre_CTAlloc(hypre_ParCSRCommHandle*,num_vectors);
if ( num_vectors==1 )
{
y_tmp = hypre_SeqVectorCreate(num_cols_offd);
}
else
{
y_tmp = hypre_SeqMultiVectorCreate(num_cols_offd,num_vectors);
}
hypre_SeqVectorInitialize(y_tmp);
/*---------------------------------------------------------------------
* If there exists no CommPkg for A, a CommPkg is generated using
* equally load balanced partitionings
*--------------------------------------------------------------------*/
if (!comm_pkg)
{
hypre_MatvecCommPkgCreate(A);
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
}
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
y_buf_data = hypre_CTAlloc( HYPRE_Complex*, num_vectors );
for ( jv=0; jv<num_vectors; ++jv )
y_buf_data[jv] = hypre_CTAlloc(HYPRE_Complex, hypre_ParCSRCommPkgSendMapStart
(comm_pkg, num_sends));
y_tmp_data = hypre_VectorData(y_tmp);
y_local_data = hypre_VectorData(y_local);
hypre_assert( idxstride==1 ); /* only 'column' storage of multivectors
* implemented so far */
if (num_cols_offd) hypre_CSRMatrixMatvecT(alpha, offd, x_local, 0.0, y_tmp);
for ( jv=0; jv<num_vectors; ++jv )
{
/* this is where we assume multivectors are 'column' storage */
comm_handle[jv] = hypre_ParCSRCommHandleCreate
( 2, comm_pkg, &(y_tmp_data[jv*num_cols_offd]), y_buf_data[jv] );
}
hypre_CSRMatrixMatvecT(alpha, diag, x_local, beta, y_local);
for ( jv=0; jv<num_vectors; ++jv )
{
hypre_ParCSRCommHandleDestroy(comm_handle[jv]);
comm_handle[jv] = NULL;
}
hypre_TFree(comm_handle);
if ( num_vectors==1 )
{
index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
y_local_data[hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j)]
+= y_buf_data[0][index++];
}
}
else
for ( jv=0; jv<num_vectors; ++jv )
{
index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
y_local_data[ jv*vecstride +
idxstride*hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j) ]
+= y_buf_data[jv][index++];
}
}
hypre_SeqVectorDestroy(y_tmp);
y_tmp = NULL;
for ( jv=0; jv<num_vectors; ++jv ) hypre_TFree(y_buf_data[jv]);
hypre_TFree(y_buf_data);
return ierr;
}
HYPRE_Int
hypre_ParCSRMatrixMatvec( HYPRE_Complex alpha,
hypre_ParCSRMatrix *A,
hypre_ParVector *x,
HYPRE_Complex beta,
hypre_ParVector *y )
{
hypre_ParCSRCommHandle **comm_handle;
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg(A);
hypre_CSRMatrix *diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix *offd = hypre_ParCSRMatrixOffd(A);
hypre_Vector *x_local = hypre_ParVectorLocalVector(x);
hypre_Vector *y_local = hypre_ParVectorLocalVector(y);
HYPRE_Int num_rows = hypre_ParCSRMatrixGlobalNumRows(A);
HYPRE_Int num_cols = hypre_ParCSRMatrixGlobalNumCols(A);
hypre_Vector *x_tmp;
HYPRE_Int x_size = hypre_ParVectorGlobalSize(x);
HYPRE_Int y_size = hypre_ParVectorGlobalSize(y);
HYPRE_Int num_vectors = hypre_VectorNumVectors(x_local);
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols(offd);
HYPRE_Int ierr = 0;
HYPRE_Int num_sends, i, j, jv, index, start;
HYPRE_Int vecstride = hypre_VectorVectorStride( x_local );
HYPRE_Int idxstride = hypre_VectorIndexStride( x_local );
HYPRE_Complex *x_tmp_data, **x_buf_data;
HYPRE_Complex *x_local_data = hypre_VectorData(x_local);
/*---------------------------------------------------------------------
* Check for size compatibility. ParMatvec returns ierr = 11 if
* length of X doesn't equal the number of columns of A,
* ierr = 12 if the length of Y doesn't equal the number of rows
* of A, and ierr = 13 if both are true.
*
* Because temporary vectors are often used in ParMatvec, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
hypre_assert( idxstride>0 );
if (num_cols != x_size)
ierr = 11;
if (num_rows != y_size)
ierr = 12;
if (num_cols != x_size && num_rows != y_size)
ierr = 13;
hypre_assert( hypre_VectorNumVectors(y_local)==num_vectors );
if ( num_vectors==1 )
x_tmp = hypre_SeqVectorCreate( num_cols_offd );
else
{
hypre_assert( num_vectors>1 );
x_tmp = hypre_SeqMultiVectorCreate( num_cols_offd, num_vectors );
}
hypre_SeqVectorInitialize(x_tmp);
x_tmp_data = hypre_VectorData(x_tmp);
comm_handle = hypre_CTAlloc(hypre_ParCSRCommHandle*,num_vectors);
/*---------------------------------------------------------------------
* If there exists no CommPkg for A, a CommPkg is generated using
* equally load balanced partitionings
*--------------------------------------------------------------------*/
if (!comm_pkg)
{
hypre_MatvecCommPkgCreate(A);
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
}
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
x_buf_data = hypre_CTAlloc( HYPRE_Complex*, num_vectors );
for ( jv=0; jv<num_vectors; ++jv )
x_buf_data[jv] = hypre_CTAlloc(HYPRE_Complex, hypre_ParCSRCommPkgSendMapStart
(comm_pkg, num_sends));
if ( num_vectors==1 )
{
index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
x_buf_data[0][index++]
= x_local_data[hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j)];
}
}
else
for ( jv=0; jv<num_vectors; ++jv )
{
index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
x_buf_data[jv][index++]
= x_local_data[
jv*vecstride +
idxstride*hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j) ];
}
}
hypre_assert( idxstride==1 );
/* ... The assert is because the following loop only works for 'column'
storage of a multivector. This needs to be fixed to work more generally,
at least for 'row' storage. This in turn, means either change CommPkg so
num_sends is no.zones*no.vectors (not no.zones) or, less dangerously, put
a stride in the logic of CommHandleCreate (stride either from a new arg or
a new variable inside CommPkg). Or put the num_vector iteration inside
CommHandleCreate (perhaps a new multivector variant of it).
*/
for ( jv=0; jv<num_vectors; ++jv )
{
comm_handle[jv] = hypre_ParCSRCommHandleCreate
( 1, comm_pkg, x_buf_data[jv], &(x_tmp_data[jv*num_cols_offd]) );
}
hypre_CSRMatrixMatvec( alpha, diag, x_local, beta, y_local);
for ( jv=0; jv<num_vectors; ++jv )
{
hypre_ParCSRCommHandleDestroy(comm_handle[jv]);
comm_handle[jv] = NULL;
}
hypre_TFree(comm_handle);
if (num_cols_offd) hypre_CSRMatrixMatvec( alpha, offd, x_tmp, 1.0, y_local);
hypre_SeqVectorDestroy(x_tmp);
x_tmp = NULL;
for ( jv=0; jv<num_vectors; ++jv ) hypre_TFree(x_buf_data[jv]);
hypre_TFree(x_buf_data);
return ierr;
}
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)
{
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)
{
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] = 0.0;
}
else
{
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)
{
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)
{
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)
{
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)
{
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] = 0.0;
}
else
{
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A*x
*-----------------------------------------------------------------*/
/* use rownnz pointer to do the A*x multiplication when num_rownnz is smaller than num_rows */
if (num_rownnz < xpar*(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
{
#pragma omp parallel for private(i,jj,temp) schedule(static)
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)
{
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] *= alpha;
}
return ierr;
}
hypre_Vector *
hypre_ParVectorToVectorAll (hypre_ParVector *par_v)
{
MPI_Comm comm = hypre_ParVectorComm(par_v);
HYPRE_Int global_size = hypre_ParVectorGlobalSize(par_v);
#ifndef HYPRE_NO_GLOBAL_PARTITION
HYPRE_Int *vec_starts = hypre_ParVectorPartitioning(par_v);
#endif
hypre_Vector *local_vector = hypre_ParVectorLocalVector(par_v);
HYPRE_Int num_procs, my_id;
HYPRE_Int num_vectors = hypre_ParVectorNumVectors(par_v);
hypre_Vector *vector;
double *vector_data;
double *local_data;
HYPRE_Int local_size;
hypre_MPI_Request *requests;
hypre_MPI_Status *status;
HYPRE_Int i, j;
HYPRE_Int *used_procs;
HYPRE_Int num_types, num_requests;
HYPRE_Int vec_len, proc_id;
#ifdef HYPRE_NO_GLOBAL_PARTITION
HYPRE_Int *new_vec_starts;
HYPRE_Int num_contacts;
HYPRE_Int contact_proc_list[1];
HYPRE_Int contact_send_buf[1];
HYPRE_Int contact_send_buf_starts[2];
HYPRE_Int max_response_size;
HYPRE_Int *response_recv_buf=NULL;
HYPRE_Int *response_recv_buf_starts = NULL;
hypre_DataExchangeResponse response_obj;
hypre_ProcListElements send_proc_obj;
HYPRE_Int *send_info = NULL;
hypre_MPI_Status status1;
HYPRE_Int count, tag1 = 112, tag2 = 223;
HYPRE_Int start;
#endif
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm, &my_id);
#ifdef HYPRE_NO_GLOBAL_PARTITION
local_size = hypre_ParVectorLastIndex(par_v) -
hypre_ParVectorFirstIndex(par_v) + 1;
/* determine procs which hold data of par_v and store ids in used_procs */
/* we need to do an exchange data for this. If I own row then I will contact
processor 0 with the endpoint of my local range */
if (local_size > 0)
{
num_contacts = 1;
contact_proc_list[0] = 0;
contact_send_buf[0] = hypre_ParVectorLastIndex(par_v);
contact_send_buf_starts[0] = 0;
contact_send_buf_starts[1] = 1;
}
else
{
num_contacts = 0;
contact_send_buf_starts[0] = 0;
contact_send_buf_starts[1] = 0;
}
/*build the response object*/
/*send_proc_obj will be for saving info from contacts */
send_proc_obj.length = 0;
send_proc_obj.storage_length = 10;
send_proc_obj.id = hypre_CTAlloc(HYPRE_Int, send_proc_obj.storage_length);
send_proc_obj.vec_starts = hypre_CTAlloc(HYPRE_Int, send_proc_obj.storage_length + 1);
send_proc_obj.vec_starts[0] = 0;
send_proc_obj.element_storage_length = 10;
send_proc_obj.elements = hypre_CTAlloc(HYPRE_Int, send_proc_obj.element_storage_length);
max_response_size = 0; /* each response is null */
response_obj.fill_response = hypre_FillResponseParToVectorAll;
response_obj.data1 = NULL;
response_obj.data2 = &send_proc_obj; /*this is where we keep info from contacts*/
hypre_DataExchangeList(num_contacts,
contact_proc_list, contact_send_buf,
contact_send_buf_starts, sizeof(HYPRE_Int),
sizeof(HYPRE_Int), &response_obj,
max_response_size, 1,
comm, (void**) &response_recv_buf,
&response_recv_buf_starts);
/* now processor 0 should have a list of ranges for processors that have rows -
these are in send_proc_obj - it needs to create the new list of processors
and also an array of vec starts - and send to those who own row*/
if (my_id)
{
if (local_size)
{
/* look for a message from processor 0 */
hypre_MPI_Probe(0, tag1, comm, &status1);
hypre_MPI_Get_count(&status1, HYPRE_MPI_INT, &count);
send_info = hypre_CTAlloc(HYPRE_Int, count);
hypre_MPI_Recv(send_info, count, HYPRE_MPI_INT, 0, tag1, comm, &status1);
/* now unpack */
num_types = send_info[0];
used_procs = hypre_CTAlloc(HYPRE_Int, num_types);
new_vec_starts = hypre_CTAlloc(HYPRE_Int, num_types+1);
for (i=1; i<= num_types; i++)
{
used_procs[i-1] = send_info[i];
}
for (i=num_types+1; i< count; i++)
{
new_vec_starts[i-num_types-1] = send_info[i] ;
}
}
else /* clean up and exit */
{
hypre_TFree(send_proc_obj.vec_starts);
hypre_TFree(send_proc_obj.id);
hypre_TFree(send_proc_obj.elements);
if(response_recv_buf) hypre_TFree(response_recv_buf);
if(response_recv_buf_starts) hypre_TFree(response_recv_buf_starts);
return NULL;
}
}
else /* my_id ==0 */
{
num_types = send_proc_obj.length;
used_procs = hypre_CTAlloc(HYPRE_Int, num_types);
new_vec_starts = hypre_CTAlloc(HYPRE_Int, num_types+1);
new_vec_starts[0] = 0;
for (i=0; i< num_types; i++)
{
used_procs[i] = send_proc_obj.id[i];
new_vec_starts[i+1] = send_proc_obj.elements[i]+1;
}
qsort0(used_procs, 0, num_types-1);
qsort0(new_vec_starts, 0, num_types);
/*now we need to put into an array to send */
count = 2*num_types+2;
send_info = hypre_CTAlloc(HYPRE_Int, count);
send_info[0] = num_types;
for (i=1; i<= num_types; i++)
{
send_info[i] = used_procs[i-1];
}
for (i=num_types+1; i< count; i++)
{
send_info[i] = new_vec_starts[i-num_types-1];
}
requests = hypre_CTAlloc(hypre_MPI_Request, num_types);
status = hypre_CTAlloc(hypre_MPI_Status, num_types);
/* don't send to myself - these are sorted so my id would be first*/
start = 0;
if (used_procs[0] == 0)
{
start = 1;
}
for (i=start; i < num_types; i++)
{
hypre_MPI_Isend(send_info, count, HYPRE_MPI_INT, used_procs[i], tag1, comm, &requests[i-start]);
}
hypre_MPI_Waitall(num_types-start, requests, status);
hypre_TFree(status);
hypre_TFree(requests);
}
/* clean up */
hypre_TFree(send_proc_obj.vec_starts);
hypre_TFree(send_proc_obj.id);
hypre_TFree(send_proc_obj.elements);
hypre_TFree(send_info);
if(response_recv_buf) hypre_TFree(response_recv_buf);
if(response_recv_buf_starts) hypre_TFree(response_recv_buf_starts);
/* now proc 0 can exit if it has no rows */
if (!local_size) {
hypre_TFree(used_procs);
hypre_TFree(new_vec_starts);
return NULL;
}
/* everyone left has rows and knows: new_vec_starts, num_types, and used_procs */
/* this vector should be rather small */
local_data = hypre_VectorData(local_vector);
vector = hypre_SeqVectorCreate(global_size);
hypre_VectorNumVectors(vector) = num_vectors;
hypre_SeqVectorInitialize(vector);
vector_data = hypre_VectorData(vector);
num_requests = 2*num_types;
requests = hypre_CTAlloc(hypre_MPI_Request, num_requests);
status = hypre_CTAlloc(hypre_MPI_Status, num_requests);
/* initialize data exchange among used_procs and generate vector - here we
send to ourself also*/
j = 0;
for (i = 0; i < num_types; i++)
{
proc_id = used_procs[i];
vec_len = new_vec_starts[i+1] - new_vec_starts[i];
hypre_MPI_Irecv(&vector_data[new_vec_starts[i]], num_vectors*vec_len, hypre_MPI_DOUBLE,
proc_id, tag2, comm, &requests[j++]);
}
for (i = 0; i < num_types; i++)
{
hypre_MPI_Isend(local_data, num_vectors*local_size, hypre_MPI_DOUBLE, used_procs[i],
tag2, comm, &requests[j++]);
}
hypre_MPI_Waitall(num_requests, requests, status);
if (num_requests)
{
hypre_TFree(requests);
hypre_TFree(status);
hypre_TFree(used_procs);
}
hypre_TFree(new_vec_starts);
#else
local_size = vec_starts[my_id+1] - vec_starts[my_id];
/* if my_id contains no data, return NULL */
if (!local_size)
return NULL;
local_data = hypre_VectorData(local_vector);
vector = hypre_SeqVectorCreate(global_size);
hypre_VectorNumVectors(vector) = num_vectors;
hypre_SeqVectorInitialize(vector);
vector_data = hypre_VectorData(vector);
/* determine procs which hold data of par_v and store ids in used_procs */
num_types = -1;
for (i=0; i < num_procs; i++)
if (vec_starts[i+1]-vec_starts[i])
num_types++;
num_requests = 2*num_types;
used_procs = hypre_CTAlloc(HYPRE_Int, num_types);
j = 0;
for (i=0; i < num_procs; i++)
if (vec_starts[i+1]-vec_starts[i] && i-my_id)
used_procs[j++] = i;
requests = hypre_CTAlloc(hypre_MPI_Request, num_requests);
status = hypre_CTAlloc(hypre_MPI_Status, num_requests);
/* initialize data exchange among used_procs and generate vector */
j = 0;
for (i = 0; i < num_types; i++)
{
proc_id = used_procs[i];
vec_len = vec_starts[proc_id+1] - vec_starts[proc_id];
hypre_MPI_Irecv(&vector_data[vec_starts[proc_id]], num_vectors*vec_len, hypre_MPI_DOUBLE,
proc_id, 0, comm, &requests[j++]);
}
for (i = 0; i < num_types; i++)
{
hypre_MPI_Isend(local_data, num_vectors*local_size, hypre_MPI_DOUBLE, used_procs[i],
0, comm, &requests[j++]);
}
for (i=0; i < num_vectors*local_size; i++)
vector_data[vec_starts[my_id]+i] = local_data[i];
hypre_MPI_Waitall(num_requests, requests, status);
if (num_requests)
{
hypre_TFree(used_procs);
hypre_TFree(requests);
hypre_TFree(status);
}
#endif
return vector;
}
hypre_ParVector *
hypre_VectorToParVector (MPI_Comm comm, hypre_Vector *v, HYPRE_Int *vec_starts)
{
HYPRE_Int global_size;
HYPRE_Int local_size;
HYPRE_Int num_vectors;
HYPRE_Int num_procs, my_id;
HYPRE_Int global_vecstride, vecstride, idxstride;
hypre_ParVector *par_vector;
hypre_Vector *local_vector;
double *v_data;
double *local_data;
hypre_MPI_Request *requests;
hypre_MPI_Status *status, status0;
HYPRE_Int i, j, k, p;
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
if (my_id == 0)
{
global_size = hypre_VectorSize(v);
v_data = hypre_VectorData(v);
num_vectors = hypre_VectorNumVectors(v); /* for multivectors */
global_vecstride = hypre_VectorVectorStride(v);
}
hypre_MPI_Bcast(&global_size,1,HYPRE_MPI_INT,0,comm);
hypre_MPI_Bcast(&num_vectors,1,HYPRE_MPI_INT,0,comm);
hypre_MPI_Bcast(&global_vecstride,1,HYPRE_MPI_INT,0,comm);
if ( num_vectors==1 )
par_vector = hypre_ParVectorCreate(comm, global_size, vec_starts);
else
par_vector = hypre_ParMultiVectorCreate(comm, global_size, vec_starts, num_vectors);
vec_starts = hypre_ParVectorPartitioning(par_vector);
local_size = vec_starts[my_id+1] - vec_starts[my_id];
hypre_ParVectorInitialize(par_vector);
local_vector = hypre_ParVectorLocalVector(par_vector);
local_data = hypre_VectorData(local_vector);
vecstride = hypre_VectorVectorStride(local_vector);
idxstride = hypre_VectorIndexStride(local_vector);
hypre_assert( idxstride==1 ); /* <<< so far only the only implemented multivector StorageMethod is 0 <<< */
if (my_id == 0)
{
requests = hypre_CTAlloc(hypre_MPI_Request,num_vectors*(num_procs-1));
status = hypre_CTAlloc(hypre_MPI_Status,num_vectors*(num_procs-1));
k = 0;
for ( p=1; p<num_procs; p++)
for ( j=0; j<num_vectors; ++j )
{
hypre_MPI_Isend( &v_data[vec_starts[p]]+j*global_vecstride,
(vec_starts[p+1]-vec_starts[p]),
hypre_MPI_DOUBLE, p, 0, comm, &requests[k++] );
}
if ( num_vectors==1 )
{
for (i=0; i < local_size; i++)
local_data[i] = v_data[i];
}
else
for ( j=0; j<num_vectors; ++j )
{
for (i=0; i < local_size; i++)
local_data[i+j*vecstride] = v_data[i+j*global_vecstride];
}
hypre_MPI_Waitall(num_procs-1,requests, status);
hypre_TFree(requests);
hypre_TFree(status);
}
else
{
for ( j=0; j<num_vectors; ++j )
hypre_MPI_Recv( local_data+j*vecstride, local_size, hypre_MPI_DOUBLE, 0, 0, comm,&status0 );
}
return par_vector;
}
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;
double *y_data_expand = NULL;
int offset = 0;
#ifdef HYPRE_USING_OPENMP
int my_thread_num = 0;
#endif
int i, j, jv, jj;
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)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) schedule(static)
#endif
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)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) schedule(static)
#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) schedule(static)
#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(double, num_threads*y_size);
if ( num_vectors==1 )
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel private(i,jj,j, my_thread_num, offset)
{
my_thread_num = omp_get_thread_num();
offset = y_size*my_thread_num;
#pragma omp for schedule(static)
#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];
}
}
#ifdef HYPRE_USING_OPENMP
/* implied barrier */
#pragma omp for schedule(static)
#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];
/*y_data_expand[j*y_size + i] = 0; //zero out for next time */
}
}
#ifdef HYPRE_USING_OPENMP
} /* end parallel region */
#endif
hypre_TFree(y_data_expand);
}
else
{
/* MULTIPLE VECTORS NOT THREADED YET */
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);
}
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;
}
/* 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;
}
