MatrixViewType
get_cache_block (MatrixViewType A,
const typename MatrixViewType::ordinal_type cache_block_index,
const bool contiguous_cache_blocks) const
{
typedef typename MatrixViewType::ordinal_type ordinal_type;
typedef typename MatrixViewType::scalar_type scalar_type;
// Total number of cache blocks.
const ordinal_type num_cache_blocks =
strategy_.num_cache_blocks (A.nrows(), A.ncols(), nrows_cache_block());
if (cache_block_index >= num_cache_blocks)
return MatrixViewType (0, 0, NULL, 0); // empty
// result[0] = starting row index of the cache block
// result[1] = number of rows in the cache block
// result[2] = pointer offset (A.get() + result[2])
// result[3] = leading dimension (a.k.a. stride) of the cache block
std::vector<Ordinal> result =
strategy_.cache_block_details (cache_block_index, A.nrows(), A.ncols(),
A.lda(), nrows_cache_block(),
contiguous_cache_blocks);
if (result[1] == 0)
// For some reason, the cache block is empty.
return MatrixViewType (0, 0, NULL, 0);
// We expect that ordinal_type is signed, so adding signed
// (ordinal_type) to unsigned (pointer) may raise compiler
// warnings.
return MatrixViewType (result[1], A.ncols(),
A.get() + static_cast<size_t>(result[2]),
result[3]);
}
Matrix (const MatrixViewType& in) :
nrows_ (in.nrows()),
ncols_ (in.ncols()),
A_ (verified_alloc_size (in.nrows(), in.ncols()))
{
if (A_.size() != 0)
copy_matrix (nrows(), ncols(), get(), lda(), in.get(), in.lda());
}
static void
scaleMatrix (MatrixViewType& A,
const typename MatrixViewType::scalar_type& denom)
{
typedef typename MatrixViewType::ordinal_type ordinal_type;
typedef typename MatrixViewType::scalar_type scalar_type;
const ordinal_type nrows = A.nrows();
const ordinal_type ncols = A.ncols();
const ordinal_type lda = A.lda();
if (nrows == lda) { // A is stored contiguously.
const ordinal_type nelts = nrows * ncols;
scalar_type* const A_ptr = A.get ();
for (ordinal_type k = 0; k < nelts; ++k) {
A_ptr[k] /= denom;
}
}
else { // Each column of A is stored contiguously.
for (ordinal_type j = 0; j < ncols; ++j) {
scalar_type* const A_j = &A(0,j);
for (ordinal_type i = 0; i < nrows; ++i) {
A_j[i] /= denom;
}
}
}
}
bool operator== (const MatrixViewType& B) const
{
if (get() != B.get() || nrows() != B.nrows() || ncols() != B.ncols() || lda() != B.lda()) {
return false;
} else {
return true;
}
}
static
std::vector<typename Teuchos::ScalarTraits<typename MatrixViewType::scalar_type>::magnitudeType>
localVerify (const MatrixViewType& A,
const MatrixViewType& Q,
const MatrixViewType& R)
{
return local_verify (A.nrows(), A.ncols(), A.get(), A.lda(),
Q.get(), Q.lda(), R.get(), R.lda());
}
MatrixViewType
split_top_block (MatrixViewType& A, const bool contiguous_cache_blocks) const
{
typedef typename MatrixViewType::ordinal_type ordinal_type;
const ordinal_type nrows_top =
strategy_.top_block_split_nrows (A.nrows(), ncols(),
nrows_cache_block());
// split_top() sets A to A_rest, and returns A_top.
return A.split_top (nrows_top, contiguous_cache_blocks);
}
MatrixViewType
split_bottom_block (MatrixViewType& A, const bool contiguous_cache_blocks) const
{
typedef typename MatrixViewType::ordinal_type ordinal_type;
const ordinal_type nrows_bottom =
strategy_.bottom_block_split_nrows (A.nrows(), ncols(),
nrows_cache_block());
// split_bottom() modifies A
return A.split_bottom (nrows_bottom, contiguous_cache_blocks);
}
MatrixViewType
top_block (const MatrixViewType& A, const bool contiguous_cache_blocks) const
{
typedef typename MatrixViewType::ordinal_type ordinal_type;
const ordinal_type nrows_top =
strategy_.top_block_split_nrows (A.nrows(), ncols(),
nrows_cache_block());
MatrixViewType A_copy (A);
return A_copy.split_top (nrows_top, contiguous_cache_blocks);
}
MatrixViewType
split_bottom_block (MatrixViewType& A, const bool contiguous_cache_blocks) const
{
typedef typename MatrixViewType::ordinal_type ordinal_type;
// Ignore the number of columns in A, since we want to block all
// matrices using the same cache blocking strategy.
const ordinal_type nrows_bottom =
strategy_.bottom_block_split_nrows (A.nrows(), ncols(),
nrows_cache_block());
// split_bottom() sets A to A_rest, and returns A_bot.
return A.split_bottom (nrows_bottom, contiguous_cache_blocks);
}
MatrixViewType
top_block (const MatrixViewType& A, const bool contiguous_cache_blocks) const
{
typedef typename MatrixViewType::ordinal_type ordinal_type;
// Ignore the number of columns in A, since we want to block all
// matrices using the same cache blocking strategy.
const ordinal_type nrows_top =
strategy_.top_block_split_nrows (A.nrows(), ncols(),
nrows_cache_block());
MatrixViewType A_copy (A);
return A_copy.split_top (nrows_top, contiguous_cache_blocks);
}
void
implicit_Q (MatrixViewType& Q,
typename MatrixViewType::scalar_type tau[])
{
implicit_Q (Q.nrows(), Q.ncols(), Q.get(), Q.lda(), tau);
}
static void
printMatrix (std::ostream& out,
const MatrixViewType& A)
{
print_local_matrix (out, A.nrows(), A.ncols(), A.get(), A.lda());
}
void
randomGlobalMatrix (Generator* const pGenerator,
MatrixViewType& A_local,
const typename Teuchos::ScalarTraits< typename MatrixViewType::scalar_type >::magnitudeType singular_values[],
MessengerBase< typename MatrixViewType::ordinal_type >* const ordinalMessenger,
MessengerBase< typename MatrixViewType::scalar_type >* const scalarMessenger)
{
using Teuchos::NO_TRANS;
using std::vector;
typedef typename MatrixViewType::ordinal_type ordinal_type;
typedef typename MatrixViewType::scalar_type scalar_type;
const bool b_local_debug = false;
const int rootProc = 0;
const int nprocs = ordinalMessenger->size();
const int myRank = ordinalMessenger->rank();
Teuchos::BLAS<ordinal_type, scalar_type> blas;
const ordinal_type nrowsLocal = A_local.nrows();
const ordinal_type ncols = A_local.ncols();
// Theory: Suppose there are P processors. Proc q wants an m_q by n
// component of the matrix A, which we write as A_q. On Proc 0, we
// generate random m_q by n orthogonal matrices Q_q (in explicit
// form), and send Q_q to Proc q. The m by n matrix [Q_0; Q_1; ...;
// Q_{P-1}] is not itself orthogonal. However, the m by n matrix
// Q = [Q_0 / P; Q_1 / P; ...; Q_{P-1} / P] is orthogonal:
//
// \sum_{q = 0}^{P-1} (Q_q^T * Q_q) / P = I.
if (myRank == rootProc)
{
typedef Random::MatrixGenerator< ordinal_type, scalar_type, Generator > matgen_type;
matgen_type matGen (*pGenerator);
// Generate a random ncols by ncols upper triangular matrix
// R with the given singular values.
Matrix< ordinal_type, scalar_type > R (ncols, ncols, scalar_type(0));
matGen.fill_random_R (ncols, R.get(), R.lda(), singular_values);
// Broadcast R to all the processors.
scalarMessenger->broadcast (R.get(), ncols*ncols, rootProc);
// Generate (for myself) a random nrowsLocal x ncols
// orthogonal matrix, stored in explicit form.
Matrix< ordinal_type, scalar_type > Q_local (nrowsLocal, ncols);
matGen.explicit_Q (nrowsLocal, ncols, Q_local.get(), Q_local.lda());
// Scale the (local) orthogonal matrix by the number of
// processors P, to make the columns of the global matrix Q
// orthogonal. (Otherwise the norm of each column will be P
// instead of 1.)
const scalar_type P = static_cast< scalar_type > (nprocs);
// Do overflow check. If casting P back to scalar_type
// doesn't produce the same value as nprocs, the cast
// overflowed. We take the real part, because scalar_type
// might be complex.
if (nprocs != static_cast<int> (Teuchos::ScalarTraits<scalar_type>::real (P)))
throw std::runtime_error ("Casting nprocs to Scalar failed");
scaleMatrix (Q_local, P);
// A_local := Q_local * R
blas.GEMM (NO_TRANS, NO_TRANS, nrowsLocal, ncols, ncols,
scalar_type(1), Q_local.get(), Q_local.lda(),
R.get(), R.lda(),
scalar_type(0), A_local.get(), A_local.lda());
for (int recvProc = 1; recvProc < nprocs; ++recvProc)
{
// Ask the receiving processor how big (i.e., how many rows)
// its local component of the matrix is.
ordinal_type nrowsRemote = 0;
ordinalMessenger->recv (&nrowsRemote, 1, recvProc, 0);
if (b_local_debug)
{
std::ostringstream os;
os << "For Proc " << recvProc << ": local block is "
<< nrowsRemote << " by " << ncols << std::endl;
std::cerr << os.str();
}
// Make sure Q_local is big enough to hold the data for
// the current receiver proc.
Q_local.reshape (nrowsRemote, ncols);
// Compute a random nrowsRemote * ncols orthogonal
// matrix Q_local, for the current receiving processor.
matGen.explicit_Q (nrowsRemote, ncols, Q_local.get(), Q_local.lda());
// Send Q_local to the current receiving processor.
scalarMessenger->send (Q_local.get(), nrowsRemote*ncols, recvProc, 0);
}
}
else
{
// Receive the R factor from Proc 0. There's only 1 R
// factor for all the processes.
Matrix< ordinal_type, scalar_type > R (ncols, ncols, scalar_type (0));
scalarMessenger->broadcast (R.get(), ncols*ncols, rootProc);
// Q_local (nrows_local by ncols, random orthogonal matrix)
// will be received from Proc 0, where it was generated.
const ordinal_type recvSize = nrowsLocal * ncols;
Matrix< ordinal_type, scalar_type > Q_local (nrowsLocal, ncols);
// Tell Proc 0 how many rows there are in the random orthogonal
// matrix I want to receive from Proc 0.
ordinalMessenger->send (&nrowsLocal, 1, rootProc, 0);
// Receive the orthogonal matrix from Proc 0.
scalarMessenger->recv (Q_local.get(), recvSize, rootProc, 0);
// Scale the (local) orthogonal matrix by the number of
// processors, to make the global matrix Q orthogonal.
const scalar_type P = static_cast< scalar_type > (nprocs);
// Do overflow check. If casting P back to scalar_type
// doesn't produce the same value as nprocs, the cast
// overflowed. We take the real part, because scalar_type
// might be complex.
if (nprocs != static_cast<int> (Teuchos::ScalarTraits<scalar_type>::real (P)))
throw std::runtime_error ("Casting nprocs to Scalar failed");
scaleMatrix (Q_local, P);
// A_local := Q_local * R
blas.GEMM (NO_TRANS, NO_TRANS, nrowsLocal, ncols, ncols,
scalar_type(1), Q_local.get(), Q_local.lda(),
R.get(), R.lda(),
scalar_type(0), A_local.get(), A_local.lda());
}
}