/*
* Distributes A in such a way that
* Layer 0 <- A(:, 0:(n/h - 1))
* Layer 1 <- A(:, (n/h):(2n/h - 1))
* .
* .
* .
* Layer h-1 <- A(:, ((h-1)n/h):n)
*/
void DistributeCols
( const mpi::Comm& depthComm,
const DistMatrix<double,MC,MR>& A,
DistMatrix<double,MC,MR>& B )
{
const Grid& meshGrid = A.Grid();
const int meshSize = meshGrid.Size();
const int depthSize = mpi::CommSize( depthComm );
const int depthRank = mpi::CommRank( depthComm );
const int sendCount = A.LocalHeight()*A.LocalWidth();
const int recvCount = sendCount / depthSize;
// For now, we will make B as large as A...
// TODO: NOT DO THIS
if( A.LocalHeight() != A.LocalLDim() )
throw std::logic_error("Local height did not match local ldim");
B.Empty();
B.AlignWith( A );
Zeros( A.Height(), A.Width(), B );
// Scatter
const int localColOffset = (A.LocalWidth()/depthSize)*depthRank;
mpi::Scatter
( A.LockedLocalBuffer(), recvCount,
B.LocalBuffer(0,localColOffset), recvCount, 0, depthComm );
}
inline void
BinaryFlat
( DistMatrix<T,U,V>& A, Int height, Int width, const std::string filename )
{
DEBUG_ONLY(CallStackEntry cse("read::BinaryFlat"))
std::ifstream file( filename.c_str(), std::ios::binary );
if( !file.is_open() )
RuntimeError("Could not open ",filename);
const Int numBytes = FileSize( file );
const Int numBytesExp = height*width*sizeof(T);
if( numBytes != numBytesExp )
RuntimeError
("Expected file to be ",numBytesExp," bytes but found ",numBytes);
A.Resize( height, width );
if( U == A.UGath && V == A.VGath )
{
if( A.CrossRank() == A.Root() )
{
if( A.Height() == A.LDim() )
file.read( (char*)A.Buffer(), height*width*sizeof(T) );
else
for( Int j=0; j<width; ++j )
file.read( (char*)A.Buffer(0,j), height*sizeof(T) );
}
}
else if( U == A.UGath )
{
const Int localWidth = A.LocalWidth();
for( Int jLoc=0; jLoc<localWidth; ++jLoc )
{
const Int j = A.GlobalCol(jLoc);
const Int localIndex = j*height;
const std::streamoff pos = localIndex*sizeof(T);
file.seekg( pos );
file.read( (char*)A.Buffer(0,jLoc), height*sizeof(T) );
}
}
else
{
const Int localHeight = A.LocalHeight();
const Int localWidth = A.LocalWidth();
for( Int jLoc=0; jLoc<localWidth; ++jLoc )
{
const Int j = A.GlobalCol(jLoc);
for( Int iLoc=0; iLoc<localHeight; ++iLoc )
{
const Int i = A.GlobalRow(iLoc);
const Int localIndex = i+j*height;
const std::streamoff pos = localIndex*sizeof(T);
file.seekg( pos );
file.read( (char*)A.Buffer(iLoc,jLoc), sizeof(T) );
}
}
}
}
inline void
MakeHilbert( DistMatrix<F,U,V>& A )
{
#ifndef RELEASE
PushCallStack("MakeHilbert");
#endif
const int m = A.Height();
const int n = A.Width();
if( m != n )
throw std::logic_error("Cannot make a non-square matrix Hilbert");
const F one = static_cast<F>(1);
const int localHeight = A.LocalHeight();
const int localWidth = A.LocalWidth();
const int colShift = A.ColShift();
const int rowShift = A.RowShift();
const int colStride = A.ColStride();
const int rowStride = A.RowStride();
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
const int j = rowShift + jLocal*rowStride;
for( int iLocal=0; iLocal<localHeight; ++iLocal )
{
const int i = colShift + iLocal*colStride;
A.SetLocalEntry( iLocal, jLocal, one/(i+j+1) );
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
Hankel( int m, int n, const std::vector<T>& a, DistMatrix<T,U,V>& A )
{
#ifndef RELEASE
PushCallStack("Hankel");
#endif
const int length = m+n-1;
if( a.size() != (unsigned)length )
throw std::logic_error("a was the wrong size");
A.ResizeTo( m, n );
const int localHeight = A.LocalHeight();
const int localWidth = A.LocalWidth();
const int colShift = A.ColShift();
const int rowShift = A.RowShift();
const int colStride = A.ColStride();
const int rowStride = A.RowStride();
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
const int j = rowShift + jLocal*rowStride;
for( int iLocal=0; iLocal<localHeight; ++iLocal )
{
const int i = colShift + iLocal*colStride;
A.SetLocal( iLocal, jLocal, a[i+j] );
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
MakeJordan( DistMatrix<T,U,V>& J, T lambda )
{
DEBUG_ONLY(CallStackEntry cse("MakeJordan"))
Zero( J.Matrix() );
const Int localHeight = J.LocalHeight();
const Int localWidth = J.LocalWidth();
const Int colShift = J.ColShift();
const Int rowShift = J.RowShift();
const Int colStride = J.ColStride();
const Int rowStride = J.RowStride();
for( Int jLoc=0; jLoc<localWidth; ++jLoc )
{
const Int j = rowShift + jLoc*rowStride;
for( Int iLoc=0; iLoc<localHeight; ++iLoc )
{
const Int i = colShift + iLoc*colStride;
if( i == j )
J.SetLocal( iLoc, jLoc, lambda );
else if( i == j-1 )
J.SetLocal( iLoc, jLoc, T(1) );
}
}
}
inline void
MakeLegendre( DistMatrix<F,U,V>& A )
{
#ifndef RELEASE
CallStackEntry entry("MakeLegendre");
#endif
if( A.Height() != A.Width() )
LogicError("Cannot make a non-square matrix Legendre");
MakeZeros( A );
const Int localHeight = A.LocalHeight();
const Int localWidth = A.LocalWidth();
const Int colShift = A.ColShift();
const Int rowShift = A.RowShift();
const Int colStride = A.ColStride();
const Int rowStride = A.RowStride();
for( Int jLoc=0; jLoc<localWidth; ++jLoc )
{
const Int j = rowShift + jLoc*rowStride;
for( Int iLoc=0; iLoc<localHeight; ++iLoc )
{
const Int i = colShift + iLoc*colStride;
if( j == i+1 || j == i-1 )
{
const Int k = Max( i, j );
const F gamma = F(1) / Pow( F(2)*k, F(2) );
const F beta = F(1) / (2*Sqrt(F(1)-gamma));
A.SetLocal( iLoc, jLoc, beta );
}
}
}
}
inline typename Base<F>::type
HermitianEntrywiseOneNorm( UpperOrLower uplo, const DistMatrix<F>& A )
{
#ifndef RELEASE
PushCallStack("HermitianEntrywiseOneNorm");
#endif
if( A.Height() != A.Width() )
throw std::logic_error("Hermitian matrices must be square.");
const int r = A.Grid().Height();
const int c = A.Grid().Width();
const int colShift = A.ColShift();
const int rowShift = A.RowShift();
typedef typename Base<F>::type R;
R localSum = 0;
const int localWidth = A.LocalWidth();
if( uplo == UPPER )
{
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
int j = rowShift + jLocal*c;
int numUpperRows = Length(j+1,colShift,r);
for( int iLocal=0; iLocal<numUpperRows; ++iLocal )
{
int i = colShift + iLocal*r;
const R alpha = Abs(A.GetLocal(iLocal,jLocal));
if( i ==j )
localSum += alpha;
else
localSum += 2*alpha;
}
}
}
else
{
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
int j = rowShift + jLocal*c;
int numStrictlyUpperRows = Length(j,colShift,r);
for( int iLocal=numStrictlyUpperRows;
iLocal<A.LocalHeight(); ++iLocal )
{
int i = colShift + iLocal*r;
const R alpha = Abs(A.GetLocal(iLocal,jLocal));
if( i ==j )
localSum += alpha;
else
localSum += 2*alpha;
}
}
}
R norm;
mpi::AllReduce( &localSum, &norm, 1, mpi::SUM, A.Grid().VCComm() );
#ifndef RELEASE
PopCallStack();
#endif
return norm;
}
// Broadcast a matrix from the root grid to the others
void DepthBroadcast
( const mpi::Comm& depthComm,
const DistMatrix<double,MC,MR>& A,
DistMatrix<double,MC,MR>& B )
{
const int rank = mpi::CommRank(mpi::COMM_WORLD);
const Grid& meshGrid = A.Grid();
const int meshSize = meshGrid.Size();
const int depthRank = rank / meshSize;
const int localSize = A.LocalHeight()*A.LocalWidth();
if( A.LocalHeight() != A.LocalLDim() )
throw std::logic_error("Leading dimension did not match local height");
B.Empty();
B.AlignWith( A );
B.ResizeTo( A.Height(), A.Width() );
// Have the root pack the broadcast data
if( depthRank == 0 )
MemCopy( B.LocalBuffer(), A.LockedLocalBuffer(), localSize );
// Broadcast from the root
mpi::Broadcast( B.LocalBuffer(), localSize, 0, depthComm );
}
int Corrupt( DistMatrix<F>& A, double probCorrupt )
{
#ifndef RELEASE
CallStackEntry entry("Corrupt");
#endif
typedef BASE(F) Real;
Int numLocalCorrupt = 0;
const Int localHeight = A.LocalHeight();
const Int localWidth = A.LocalWidth();
for( Int jLocal=0; jLocal<localWidth; ++jLocal )
{
for( Int iLocal=0; iLocal<localHeight; ++iLocal )
{
if( Uniform<Real>() <= probCorrupt )
{
++numLocalCorrupt;
const F perturb = SampleBall<F>();
A.SetLocal( iLocal, jLocal, A.GetLocal(iLocal,jLocal)+perturb );
}
}
}
Int numCorrupt;
mpi::AllReduce
( &numLocalCorrupt, &numCorrupt, 1, mpi::SUM, A.Grid().VCComm() );
return numCorrupt;
}
inline void
Riemann( DistMatrix<T,U,V>& R, int n )
{
#ifndef RELEASE
CallStackEntry entry("Riemann");
#endif
R.ResizeTo( n, n );
const int localHeight = R.LocalHeight();
const int localWidth = R.LocalWidth();
const int colShift = R.ColShift();
const int rowShift = R.RowShift();
const int colStride = R.ColStride();
const int rowStride = R.RowStride();
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
const int j = rowShift + jLocal*rowStride;
for( int iLocal=0; iLocal<localHeight; ++iLocal )
{
const int i = colShift + iLocal*colStride;
if( ((j+2)%(i+2))==0 )
R.SetLocal( iLocal, jLocal, T(i+1) );
else
R.SetLocal( iLocal, jLocal, T(-1) );
}
}
}
static void Func
( DistMatrix<T,STAR,MR>& A, T center, typename Base<T>::type radius )
{
const Grid& grid = A.Grid();
const int m = A.Height();
const int localWidth = A.LocalWidth();
const int bufSize = m*localWidth;
std::vector<T> buffer( bufSize );
// Create random matrix on process row 0, then broadcast
if( grid.Row() == 0 )
{
for( int j=0; j<localWidth; ++j )
for( int i=0; i<m; ++i )
buffer[i+j*m] = center+radius*SampleUnitBall<T>();
}
mpi::Broadcast( &buffer[0], bufSize, 0, grid.ColComm() );
// Unpack
T* localBuffer = A.LocalBuffer();
const int ldim = A.LocalLDim();
#ifdef HAVE_OPENMP
#pragma omp parallel for
#endif
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
const T* bufferCol = &buffer[jLocal*m];
T* col = &localBuffer[jLocal*ldim];
MemCopy( col, bufferCol, m );
}
}
inline void
MakeDiscreteFourier( DistMatrix<Complex<R>,U,V>& A )
{
#ifndef RELEASE
CallStackEntry entry("MakeDiscreteFourier");
#endif
typedef Complex<R> F;
const int m = A.Height();
const int n = A.Width();
if( m != n )
throw std::logic_error("Cannot make a non-square DFT matrix");
const R pi = 4*Atan( R(1) );
const F nSqrt = Sqrt( R(n) );
const int localHeight = A.LocalHeight();
const int localWidth = A.LocalWidth();
const int colShift = A.ColShift();
const int rowShift = A.RowShift();
const int colStride = A.ColStride();
const int rowStride = A.RowStride();
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
const int j = rowShift + jLocal*rowStride;
for( int iLocal=0; iLocal<localHeight; ++iLocal )
{
const int i = colShift + iLocal*colStride;
A.SetLocal( iLocal, jLocal, Exp(-2*pi*i*j/n)/nSqrt );
const R theta = -2*pi*i*j/n;
const Complex<R> alpha( Cos(theta), Sin(theta) );
A.SetLocal( iLocal, jLocal, alpha/nSqrt );
}
}
}
inline void
BinaryFlat
( DistMatrix<T,STAR,V>& A, Int height, Int width, const std::string filename )
{
DEBUG_ONLY(CallStackEntry cse("read::BinaryFlat"))
std::ifstream file( filename.c_str(), std::ios::binary );
if( !file.is_open() )
RuntimeError("Could not open ",filename);
const Int numBytes = FileSize( file );
const Int numBytesExp = height*width*sizeof(T);
if( numBytes != numBytesExp )
RuntimeError
("Expected file to be ",numBytesExp," bytes but found ",numBytes);
A.Resize( height, width );
const Int localWidth = A.LocalWidth();
const Int rowShift = A.RowShift();
const Int rowStride = A.RowStride();
for( Int jLoc=0; jLoc<localWidth; ++jLoc )
{
const Int j = rowShift + jLoc*rowStride;
const Int localIndex = j*height;
const std::streamoff pos = localIndex*sizeof(T);
file.seekg( pos );
file.read( (char*)A.Buffer(0,jLoc), height*sizeof(T) );
}
}
inline void
Hankel( DistMatrix<T,U,V>& A, Int m, Int n, const std::vector<T>& a )
{
#ifndef RELEASE
CallStackEntry entry("Hankel");
#endif
const Int length = m+n-1;
if( a.size() != (Unsigned)length )
LogicError("a was the wrong size");
A.ResizeTo( m, n );
const Int localHeight = A.LocalHeight();
const Int localWidth = A.LocalWidth();
const Int colShift = A.ColShift();
const Int rowShift = A.RowShift();
const Int colStride = A.ColStride();
const Int rowStride = A.RowStride();
for( Int jLoc=0; jLoc<localWidth; ++jLoc )
{
const Int j = rowShift + jLoc*rowStride;
for( Int iLoc=0; iLoc<localHeight; ++iLoc )
{
const Int i = colShift + iLoc*colStride;
A.SetLocal( iLoc, jLoc, a[i+j] );
}
}
}
inline void
Wilkinson( DistMatrix<T,U,V>& A, int k )
{
#ifndef RELEASE
CallStackEntry entry("Wilkinson");
#endif
const int n = 2*k+1;
A.ResizeTo( n, n );
MakeZeros( A );
const int localHeight = A.LocalHeight();
const int localWidth = A.LocalWidth();
const int colShift = A.ColShift();
const int rowShift = A.RowShift();
const int colStride = A.ColStride();
const int rowStride = A.RowStride();
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
const int j = rowShift + jLocal*rowStride;
for( int iLocal=0; iLocal<localHeight; ++iLocal )
{
const int i = colShift + iLocal*colStride;
if( i == j )
{
if( j <= k )
A.SetLocal( iLocal, jLocal, T(k-j) );
else
A.SetLocal( iLocal, jLocal, T(j-k) );
}
else if( i == j-1 || i == j+1 )
A.SetLocal( iLocal, jLocal, T(1) );
}
}
}
InfinityNorm( const DistMatrix<F,U,V>& A )
{
#ifndef RELEASE
CallStackEntry entry("InfinityNorm");
#endif
// Compute the partial row sums defined by our local matrix, A[U,V]
typedef BASE(F) R;
const Int localHeight = A.LocalHeight();
const Int localWidth = A.LocalWidth();
std::vector<R> myPartialRowSums( localHeight );
for( Int iLoc=0; iLoc<localHeight; ++iLoc )
{
myPartialRowSums[iLoc] = 0;
for( Int jLoc=0; jLoc<localWidth; ++jLoc )
myPartialRowSums[iLoc] += Abs(A.GetLocal(iLoc,jLoc));
}
// Sum our partial row sums to get the row sums over A[U,* ]
std::vector<R> myRowSums( localHeight );
mpi::Comm rowComm = ReduceRowComm<U,V>( A.Grid() );
mpi::AllReduce( &myPartialRowSums[0], &myRowSums[0], localHeight, rowComm );
// Find the maximum out of the row sums
R myMaxRowSum = 0;
for( Int iLoc=0; iLoc<localHeight; ++iLoc )
myMaxRowSum = std::max( myMaxRowSum, myRowSums[iLoc] );
// Find the global maximum row sum by searching over the U team
mpi::Comm colComm = ReduceColComm<U,V>( A.Grid() );
return mpi::AllReduce( myMaxRowSum, mpi::MAX, colComm );
}
inline void
MakeOneTwoOne( DistMatrix<T,U,V>& A )
{
#ifndef RELEASE
PushCallStack("MakeOneTwoOne");
#endif
if( A.Height() != A.Width() )
throw std::logic_error("Cannot make a non-square matrix 1-2-1");
MakeZeros( A );
const int localHeight = A.LocalHeight();
const int localWidth = A.LocalWidth();
const int colShift = A.ColShift();
const int rowShift = A.RowShift();
const int colStride = A.ColStride();
const int rowStride = A.RowStride();
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
const int j = rowShift + jLocal*rowStride;
for( int iLocal=0; iLocal<localHeight; ++iLocal )
{
const int i = colShift + iLocal*colStride;
if( i == j )
A.SetLocal( iLocal, jLocal, T(2) );
else if( i == j-1 || i == j+1 )
A.SetLocal( iLocal, jLocal, T(1) );
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
Diagonal( const std::vector<T>& d, DistMatrix<T,U,V>& D )
{
#ifndef RELEASE
PushCallStack("Diagonal");
#endif
const int n = d.size();
D.ResizeTo( n, n );
MakeZeros( D );
const int localWidth = D.LocalWidth();
const int colShift = D.ColShift();
const int rowShift = D.RowShift();
const int colStride = D.ColStride();
const int rowStride = D.RowStride();
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
const int j = rowShift + jLocal*rowStride;
if( (j-colShift+colStride) % colStride == 0 )
{
const int iLocal = (j-colShift) / colStride;
D.SetLocal( iLocal, jLocal, d[j] );
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
Redheffer( DistMatrix<T,U,V>& R, Int n )
{
#ifndef RELEASE
CallStackEntry entry("Redheffer");
#endif
R.ResizeTo( n, n );
const Int localHeight = R.LocalHeight();
const Int localWidth = R.LocalWidth();
const Int colShift = R.ColShift();
const Int rowShift = R.RowShift();
const Int colStride = R.ColStride();
const Int rowStride = R.RowStride();
for( Int jLoc=0; jLoc<localWidth; ++jLoc )
{
const Int j = rowShift + jLoc*rowStride;
for( Int iLoc=0; iLoc<localHeight; ++iLoc )
{
const Int i = colShift + iLoc*colStride;
if( j==0 || ((j+1)%(i+1))==0 )
R.SetLocal( iLoc, jLoc, T(1) );
else
R.SetLocal( iLoc, jLoc, T(0) );
}
}
}
inline void
Pei( DistMatrix<T,U,V>& P, Int n, T alpha )
{
#ifndef RELEASE
CallStackEntry entry("MakeIdentity");
#endif
P.ResizeTo( n, n );
const Int localHeight = P.LocalHeight();
const Int localWidth = P.LocalWidth();
const Int colShift = P.ColShift();
const Int rowShift = P.RowShift();
const Int colStride = P.ColStride();
const Int rowStride = P.RowStride();
for( Int jLoc=0; jLoc<localWidth; ++jLoc )
{
const Int j = rowShift + jLoc*rowStride;
for( Int iLoc=0; iLoc<localHeight; ++iLoc )
{
const Int i = colShift + iLoc*colStride;
P.SetLocal( iLoc, jLoc, T(1) );
if( i == j )
P.UpdateLocal( iLoc, jLoc, alpha );
}
}
}
inline void
MakeHilbert( DistMatrix<F,U,V>& A )
{
#ifndef RELEASE
CallStackEntry entry("MakeHilbert");
#endif
const Int m = A.Height();
const Int n = A.Width();
if( m != n )
LogicError("Cannot make a non-square matrix Hilbert");
const F one = F(1);
const Int localHeight = A.LocalHeight();
const Int localWidth = A.LocalWidth();
const Int colShift = A.ColShift();
const Int rowShift = A.RowShift();
const Int colStride = A.ColStride();
const Int rowStride = A.RowStride();
for( Int jLoc=0; jLoc<localWidth; ++jLoc )
{
const Int j = rowShift + jLoc*rowStride;
for( Int iLoc=0; iLoc<localHeight; ++iLoc )
{
const Int i = colShift + iLoc*colStride;
A.SetLocal( iLoc, jLoc, one/(i+j+1) );
}
}
}
void calc_x_else(const R alpha, const R beta,const R c, const R delta, DistMatrix<R> &x, DistMatrix<R> &phi,DistMatrix<R> &w){
R bphi;
R temp;
R temp2;
const R pi = 4*atan(1);
const int colShift = x.ColShift(); // first row we own
const int rowShift = x.RowShift(); // first col we own
const int colStride = x.ColStride();
const int rowStride = x.RowStride();
const int localHeight = x.LocalHeight();
const int localWidth = x.LocalWidth();
for( int iLocal=0; iLocal<localHeight; ++iLocal ){
for( int jLocal=0; jLocal<localWidth; ++jLocal ){
const int i = colShift + iLocal*colStride;
const int j = rowShift + jLocal*rowStride;
temp = phi.GetLocal(0,jLocal);//phi
temp2 = w.GetLocal(0,jLocal); //w
bphi = (pi/2) + beta * temp;
//cout<< (delta + c * (((2/pi) * (bphi * tan(temp) - beta * log((pi/2) * temp2 * cos(temp) / bphi))) + (beta * tan (pi *alpha/2))))<<endl;
x.SetLocal(iLocal, jLocal, (-1* sqrt(abs(delta + c * (((2/pi) * (bphi * tan(temp) - beta * log((pi/2) * temp2 * cos(temp) / bphi))) + (beta * tan (pi *alpha/2)))))));
}
}
}
inline void
MakeGKS( DistMatrix<F,U,V>& A )
{
#ifndef RELEASE
CallStackEntry entry("MakeGKS");
#endif
const Int m = A.Height();
const Int n = A.Width();
if( m != n )
LogicError("Cannot make a non-square matrix GKS");
const Int localHeight = A.LocalHeight();
const Int localWidth = A.LocalWidth();
const Int colShift = A.ColShift();
const Int rowShift = A.RowShift();
const Int colStride = A.ColStride();
const Int rowStride = A.RowStride();
for( Int jLoc=0; jLoc<localWidth; ++jLoc )
{
const Int j = rowShift + jLoc*rowStride;
const F jDiag = F(1)/Sqrt(F(j));
for( Int iLoc=0; iLoc<localHeight; ++iLoc )
{
const Int i = colShift + iLoc*colStride;
if( i < j )
A.SetLocal( iLoc, jLoc, -jDiag );
else if( i == j )
A.SetLocal( iLoc, jLoc, jDiag );
else
A.SetLocal( iLoc, jLoc, 0 );
}
}
}
//phi = (rand(m,n)-.5)*pi;
void phi_uniform(DistMatrix<R> &U){
boost::random::uniform_01<> dist;
boost::random::mt19937 rng(static_cast<boost::uint32_t>(commRank));
R temp;
R pi = 4*atan(1);
const int colShift = U.ColShift(); // first row we own
const int rowShift = U.RowShift(); // first col we own
const int colStride = U.ColStride();
const int rowStride = U.RowStride();
const int localHeight = U.LocalHeight();
const int localWidth = U.LocalWidth();
for( int iLocal=0; iLocal<localHeight; ++iLocal ){
for( int jLocal=0; jLocal<localWidth; ++jLocal ){
const int i = colShift + iLocal*colStride;
const int j = rowShift + jLocal*rowStride;
temp = dist(rng);
U.SetLocal(iLocal, jLocal, ((temp-0.5)*pi) );
}
}
}
void calc_x_if(const R alpha,const R beta,const R c, const R delta,DistMatrix<R> &x, DistMatrix<R> &phi,DistMatrix<R> &w){
R temp; // phi
R temp2; // w
R aphi;
R a1phi;
const R pi = 4*atan(1);
const R zeta = beta * tan(pi * alpha/2);
const int colShift = x.ColShift(); // first row we own
const int rowShift = x.RowShift(); // first col we own
const int colStride = x.ColStride();
const int rowStride = x.RowStride();
const int localHeight = x.LocalHeight();
const int localWidth = x.LocalWidth();
for( int iLocal=0; iLocal<localHeight; ++iLocal ){
for( int jLocal=0; jLocal<localWidth; ++jLocal ){
const int i = colShift + iLocal*colStride;
const int j = rowShift + jLocal*rowStride;
temp = phi.GetLocal(iLocal,jLocal);//phi
temp2 = w.GetLocal(iLocal,jLocal); //w
aphi = alpha * temp;
a1phi = (1-alpha)*temp;
x.SetLocal(iLocal, jLocal, (-1 * sqrt(abs(delta + c *( ( (sin(aphi)+zeta * cos(aphi))/ cos(temp)) * -1 * pow( abs( ((cos(a1phi) + zeta * sin(a1phi))/ (temp2 * cos(temp))) ) ,((1-alpha)/alpha) ) + beta * tan(pi * alpha/2) )))) );
}
}
}
//w = -log(rand(m,n));
void log_uniform(DistMatrix<R> &U){
//boost::random::gamma_distribution<> dist(df/2.0,2.0);
boost::random::uniform_01<> dist;
boost::random::mt19937 rng(static_cast<boost::uint32_t>(commRank));
double temp;
const int colShift = U.ColShift(); // first row we own
const int rowShift = U.RowShift(); // first col we own
const int colStride = U.ColStride();
const int rowStride = U.RowStride();
const int localHeight = U.LocalHeight();
const int localWidth = U.LocalWidth();
for( int iLocal=0; iLocal<localHeight; ++iLocal ){
for( int jLocal=0; jLocal<localWidth; ++jLocal ){
const int i = colShift + iLocal*colStride;
const int j = rowShift + jLocal*rowStride;
temp = dist(rng);
U.SetLocal(iLocal, jLocal, log(temp));
}
}
}
void dotproduct(DistMatrix<R> &A,DistMatrix<R> &B){
if(A.Height() != B.Height()){
//ERROR!
}
if(A.Width() != B.Width()){
//ERROR!
}
double temp,temp1;
const int colShift = A.ColShift(); // first row we own
const int rowShift = A.RowShift(); // first col we own
const int colStride = A.ColStride();
const int rowStride = A.RowStride();
const int localHeight = A.LocalHeight();
const int localWidth = A.LocalWidth();
for( int iLocal=0; iLocal<localHeight; ++iLocal ){
for( int jLocal=0; jLocal<localWidth; ++jLocal ){
const int i = colShift + iLocal*colStride;
const int j = rowShift + jLocal*rowStride;
temp = A.GetLocal(iLocal, jLocal);
temp1 = B.GetLocal(iLocal,jLocal);
B.SetLocal(iLocal, jLocal, (temp*temp1));
}
}
}
inline void
MakeTriangular( UpperOrLower uplo, DistMatrix<T,U,V>& A )
{
#ifndef RELEASE
PushCallStack("MakeTriangular");
#endif
const int height = A.Height();
const int localHeight = A.LocalHeight();
const int localWidth = A.LocalWidth();
const int colShift = A.ColShift();
const int rowShift = A.RowShift();
const int colStride = A.ColStride();
const int rowStride = A.RowStride();
T* localBuffer = A.LocalBuffer();
const int ldim = A.LocalLDim();
if( uplo == LOWER )
{
#ifdef HAVE_OPENMP
#pragma omp parallel for
#endif
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
const int j = rowShift + jLocal*rowStride;
const int lastZeroRow = j-1;
if( lastZeroRow >= 0 )
{
const int boundary = std::min( lastZeroRow+1, height );
const int numZeroRows =
RawLocalLength( boundary, colShift, colStride );
MemZero( &localBuffer[jLocal*ldim], numZeroRows );
}
}
}
else
{
#ifdef HAVE_OPENMP
#pragma omp parallel for
#endif
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
const int j = rowShift + jLocal*rowStride;
const int firstZeroRow = j+1;
const int numNonzeroRows =
RawLocalLength(firstZeroRow,colShift,colStride);
if( numNonzeroRows < localHeight )
{
T* col = &localBuffer[numNonzeroRows+jLocal*ldim];
MemZero( col, localHeight-numNonzeroRows );
}
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
static void Func
( DistMatrix<T,MC,MR>& A, T center, typename Base<T>::type radius )
{
const int localHeight = A.LocalHeight();
const int localWidth = A.LocalWidth();
for( int jLocal=0; jLocal<localWidth; ++jLocal )
for( int iLocal=0; iLocal<localHeight; ++iLocal )
A.SetLocal( iLocal, jLocal, center+radius*SampleUnitBall<T>() );
}
inline typename Base<F>::type
FrobeniusNorm( const DistMatrix<F,U,V>& A )
{
#ifndef RELEASE
PushCallStack("internal::FrobeniusNorm");
#endif
typedef typename Base<F>::type R;
mpi::Comm comm = NormComm( A );
R localScale = 0;
R localScaledSquare = 1;
const int localHeight = A.LocalHeight();
const int localWidth = A.LocalWidth();
for( int jLocal=0; jLocal<localWidth; ++jLocal )
{
for( int iLocal=0; iLocal<localHeight; ++iLocal )
{
const R alphaAbs = Abs(A.GetLocal(iLocal,jLocal));
if( alphaAbs != 0 )
{
if( alphaAbs <= localScale )
{
const R relScale = alphaAbs/localScale;
localScaledSquare += relScale*relScale;
}
else
{
const R relScale = localScale/alphaAbs;
localScaledSquare = localScaledSquare*relScale*relScale + 1;
localScale = alphaAbs;
}
}
}
}
// Find the maximum relative scale
R scale;
mpi::AllReduce( &localScale, &scale, 1, mpi::MAX, comm );
R norm = 0;
if( scale != 0 )
{
// Equilibrate our local scaled sum to the maximum scale
R relScale = localScale/scale;
localScaledSquare *= relScale*relScale;
// The scaled square is now simply the sum of the local contributions
R scaledSquare;
mpi::AllReduce( &localScaledSquare, &scaledSquare, 1, mpi::SUM, comm );
norm = scale*Sqrt(scaledSquare);
}
#ifndef RELEASE
PopCallStack();
#endif
return norm;
}