void
LSLocateCircularInterface::setLevelSetPatchData(int D_idx,
Pointer<HierarchyMathOps> hier_math_ops,
double /*time*/,
bool /*initial_time*/)
{
// In this version of this class, the initial level set location is set to be
// exact since we always know the radius of the ball
Pointer<PatchHierarchy<NDIM> > patch_hierarchy = hier_math_ops->getPatchHierarchy();
const int coarsest_ln = 0;
const int finest_ln = patch_hierarchy->getFinestLevelNumber();
// Set the initial condition for locating the interface
const double& R = d_circle->R;
const IBTK::Vector& X0 = d_circle->X0;
for (int ln = coarsest_ln; ln <= finest_ln; ++ln)
{
Pointer<PatchLevel<NDIM> > level = patch_hierarchy->getPatchLevel(ln);
for (PatchLevel<NDIM>::Iterator p(level); p; p++)
{
Pointer<Patch<NDIM> > patch = level->getPatch(p());
const Box<NDIM>& patch_box = patch->getBox();
Pointer<CellData<NDIM, double> > D_data = patch->getPatchData(D_idx);
for (Box<NDIM>::Iterator it(patch_box); it; it++)
{
CellIndex<NDIM> ci(it());
// Get physical coordinates
IBTK::Vector coord = IBTK::Vector::Zero();
Pointer<CartesianPatchGeometry<NDIM> > patch_geom = patch->getPatchGeometry();
const double* patch_X_lower = patch_geom->getXLower();
const hier::Index<NDIM>& patch_lower_idx = patch_box.lower();
const double* const patch_dx = patch_geom->getDx();
for (int d = 0; d < NDIM; ++d)
{
coord[d] = patch_X_lower[d] + patch_dx[d] * (static_cast<double>(ci(d) - patch_lower_idx(d)) + 0.5);
}
const double distance = std::sqrt(std::pow((coord[0] - X0(0)), 2.0) + std::pow((coord[1] - X0(1)), 2.0)
#if (NDIM == 3)
+ std::pow((coord[2] - X0(2)), 2.0)
#endif
);
(*D_data)(ci) = distance - R;
}
}
}
return;
} // setLevelSetPatchData
template <class cCRNode> void cTplCoxRoyAlgo<cCRNode>::SetStdCostRegul(double aCoeff,double aCste,int aVmin)
{
for (int anX=X0(); anX<X1() ; anX++)
for (int anY=Y0(); anY<Y1() ; anY++)
for (int aZ = ZMin(anX,anY); aZ< ZMax(anX,anY) ; aZ++)
{
cRoyPt aP1 (anX,anY,aZ);
cCRNode & aS1 = NodeOfP(aP1);
int aC1 = aS1.ResidualFlow(mDirZPlus);
for (int anEdg=0; anEdg<NbEdges() ; anEdg++)
{
if (aS1.EdgeIsValide(anEdg) && (!tabCRIsVertical[anEdg]))
{
cRoyPt aP2(aP1,anEdg);
cCRNode & aS2 = NodeOfP(aP2);
int aC2 = aS2.ResidualFlow(mDirZPlus);
double aCost = aCste + aCoeff*(aC1+aC2)/2.0;
if (Cnx8())
aCost *= tabCRIsArcV8[anEdg] ? 0.2928 : 0.4142 ;
int iCost = int(aCost+0.5);
if (iCost<aVmin)
iCost = aVmin;
aS1.SetResidualFlow(anEdg,iCost);
}
}
}
}
/** An sample for taylor expansion of logdet(X). */
void taylorSample() {
std::string ans;
char rowChar[5];
int rowTmp = ROW;
sprintf(rowChar, "%d", rowTmp);
std::string row = rowChar;
// Initialize the matrices.
symbolic_matrix_type X("X", ROW, COL);
symbolic_matrix_type X0("X0", ROW, COL);
symbolic_matrix_type Delta("(X-X0)", ROW, COL);
AMD::SymbolicScalarMatlab a2("1/2!");
AMD::SymbolicScalarMatlab a3("1/3!");
SymbolicSMFunc r2(a2,ROW,COL);
SymbolicSMFunc r3(a3,ROW, COL);
// Initialize MatrixMatrixFunction.
SymbolicMMFunc fX(X, false);
SymbolicMMFunc fX0(X0, false);
SymbolicMMFunc fDelta(Delta, true);
// Compute Taylor series iteratively.
SymbolicSMFunc f0 = logdet(fX0);
SymbolicSMFunc f1 = trace(fDelta * transpose(*f0.derivativeFuncVal));
SymbolicSMFunc f2 = trace(fDelta * transpose(*f1.derivativeFuncVal));
SymbolicSMFunc f3 = trace(fDelta * transpose(*f2.derivativeFuncVal));
// Taylor Expansion.
SymbolicSMFunc func = f0 + f1 + r2*f2 + r3*f3;
std::cout<<"The first 4 terms of Taylor Expansion for logdet(X) around X0 is:";
std::cout << std::endl;
std::cout << func.functionVal.getString() << std::endl;
}
AR_Process::AR_Process(void)
{
// State Equation
F.resize(1,1);
F(0,0) = 0.8;
f.resize(1);
f(0) = 0.;
G.resize(1,1);
G.identity();
Qw.resize(1);
Qw(0,0)= 0.1;
// Observation noise
H.resize(1,1);
H(0,0) = 1;
h.resize(1);
h(0) = 0.;
Qv.resize(1);
Qv(0,0)=1;
// Init state
X0.resize(1);
X0(0) = 10.;
R0.resize(1);
R0.zero();
}
inline void
TrmmLLNA
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmLLNA");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( L.Height() != L.Width() || L.Width() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmLLNA: \n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = L.Grid();
DistMatrix<T>
XL(g), XR(g),
X0(g), X1(g), X2(g);
DistMatrix<T,VR, STAR> X1_VR_STAR(g);
DistMatrix<T,STAR,MR > X1Trans_STAR_MR(g);
DistMatrix<T,MC, STAR> Z1_MC_STAR(g);
X1_VR_STAR.AlignWith( L );
X1Trans_STAR_MR.AlignWith( L );
Z1_MC_STAR.AlignWith( L );
PartitionRight( X, XL, XR, 0 );
while( XL.Width() < X.Width() )
{
RepartitionRight
( XL, /**/ XR,
X0, /**/ X1, X2 );
Zeros( X1.Height(), X1.Width(), Z1_MC_STAR );
//--------------------------------------------------------------------//
X1_VR_STAR = X1;
X1Trans_STAR_MR.TransposeFrom( X1_VR_STAR );
LocalTrmmAccumulateLLN
( TRANSPOSE, diag, alpha, L, X1Trans_STAR_MR, Z1_MC_STAR );
X1.SumScatterFrom( Z1_MC_STAR );
//--------------------------------------------------------------------//
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
std::pair<int,double> slice_sample_multi(double x0, vector<slice_function*>& g, double w, int m)
{
int N = g.size();
vector<double> X0(N,x0);
return slice_sample_multi(X0,g,w,m);
}
inline void
TrmmRUNA
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& U,
DistMatrix<T>& X )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrmmRUNA");
if( U.Grid() != X.Grid() )
throw std::logic_error("{U,X} must be distributed over the same grid");
#endif
const Grid& g = U.Grid();
DistMatrix<T>
XT(g), X0(g),
XB(g), X1(g),
X2(g);
DistMatrix<T,STAR,VC > X1_STAR_VC(g);
DistMatrix<T,STAR,MC > X1_STAR_MC(g);
DistMatrix<T,MR, STAR> Z1Trans_MR_STAR(g);
DistMatrix<T,MR, MC > Z1Trans_MR_MC(g);
X1_STAR_VC.AlignWith( U );
X1_STAR_MC.AlignWith( U );
Z1Trans_MR_STAR.AlignWith( U );
PartitionDown
( X, XT,
XB, 0 );
while( XT.Height() < X.Height() )
{
RepartitionDown
( XT, X0,
/**/ /**/
X1,
XB, X2 );
Z1Trans_MR_MC.AlignWith( X1 );
//--------------------------------------------------------------------//
X1_STAR_VC = X1;
X1_STAR_MC = X1_STAR_VC;
Zeros( Z1Trans_MR_STAR, X1.Width(), X1.Height() );
LocalTrmmAccumulateRUN
( TRANSPOSE, diag, alpha, U, X1_STAR_MC, Z1Trans_MR_STAR );
Z1Trans_MR_MC.SumScatterFrom( Z1Trans_MR_STAR );
Transpose( Z1Trans_MR_MC.Matrix(), X1.Matrix() );
//--------------------------------------------------------------------//
Z1Trans_MR_MC.FreeAlignments();
SlidePartitionDown
( XT, X0,
X1,
/**/ /**/
XB, X2 );
}
}
Van_Der_Pol::Van_Der_Pol(void)
{
lambda = 3.;
Qw.resize(1);
Qw(0,0)= 1.;
Qv.resize(1);
Qv(0,0)=0.1;
X0.resize(2);
X0(0) = 0.5;
X0(1) = 0.5;
R0.resize(2);
R0.zero();
R0(0,0)=0.;
R0(1,1)=.1;
Ts=.1;
}
bool ACovarianceFunction::check_covariance_Kgrad_x(ACovarianceFunction& covar,mfloat_t relchange,mfloat_t threshold,bool check_diag)
{
mfloat_t RV=0;
//copy inputs for which we calculate gradients
CovarInput X = covar.getX();
CovarInput X0 = X;
for (int ic=0;ic<X.cols();ic++)
{
//analytical gradient is per columns all in one go:
MatrixXd Kgrad_x = covar.Kgrad_X(ic);
MatrixXd Kgrad_x_diag = covar.Kdiag_grad_X(ic);
for (int ir=0;ir<X.rows();ir++)
{
mfloat_t change = relchange*X0(ir,ic);
change = std::max(change,1E-5);
X(ir,ic) = X0(ir,ic) + change;
covar.setX(X);
MatrixXd Lplus = covar.K();
X(ir,ic) = X0(ir,ic) - change;
covar.setX(X);
MatrixXd Lminus = covar.K();
X(ir,ic) = X0(ir,ic);
covar.setX(X);
//numerical gradient
MatrixXd diff_numerical = (Lplus-Lminus)/(2.*change);
//build analytical gradient matrix
MatrixXd diff_analytical = MatrixXd::Zero(X.rows(),X.rows());
diff_analytical.row(ir) = Kgrad_x.row(ir);
diff_analytical.col(ir) += Kgrad_x.row(ir);
RV+= (diff_numerical-diff_analytical).squaredNorm();
//difference
if (check_diag)
{
double delta =(diff_numerical(ir,ir)-Kgrad_x_diag(ir));
RV+= delta*delta;
}
} //end for ir
}
return (RV < threshold);
}
Foam::tmp<Foam::pointField> Foam::RBD::rigidBodyMotion::transformPoints
(
const label bodyID,
const pointField& initialPoints
) const
{
// Calculate the transform from the initial state in the global frame
// to the current state in the global frame
spatialTransform X(X0(bodyID).inv() & X00(bodyID));
tmp<pointField> tpoints(new pointField(initialPoints.size()));
pointField& points = tpoints.ref();
forAll(points, i)
{
points[i] = X.transformPoint(initialPoints[i]);
}
inline void
TrsmLUNSmall
( UnitOrNonUnit diag,
F alpha, const DistMatrix<F,VC,STAR>& U, DistMatrix<F,VC,STAR>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmLUNSmall");
if( U.Grid() != X.Grid() )
throw std::logic_error
("U and X must be distributed over the same grid");
if( U.Height() != U.Width() || U.Width() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrsmLUN: \n"
<< " U ~ " << U.Height() << " x " << U.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str() );
}
if( U.ColAlignment() != X.ColAlignment() )
throw std::logic_error("U and X are assumed to be aligned");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<F,VC,STAR>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
DistMatrix<F,VC,STAR> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> X1_STAR_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionUp
( X, XT,
XB, 0 );
while( XT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionUp
( XT, X0,
X1,
/**/ /**/
XB, X2 );
//--------------------------------------------------------------------//
U11_STAR_STAR = U11; // U11[* ,* ] <- U11[VC,* ]
X1_STAR_STAR = X1; // X1[* ,* ] <- X1[VC,* ]
// X1[* ,* ] := U11^-1[* ,* ] X1[* ,* ]
LocalTrsm
( LEFT, UPPER, NORMAL, diag,
F(1), U11_STAR_STAR, X1_STAR_STAR, checkIfSingular );
X1 = X1_STAR_STAR;
// X0[VC,* ] -= U01[VC,* ] X1[* ,* ]
LocalGemm( NORMAL, NORMAL, F(-1), U01, X1_STAR_STAR, F(1), X0 );
//--------------------------------------------------------------------//
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrsmLUNLarge
( UnitOrNonUnit diag,
F alpha, const DistMatrix<F>& U, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmLUNLarge");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<F>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
DistMatrix<F> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<F,MC, STAR> U01_MC_STAR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,MR > X1_STAR_MR(g);
DistMatrix<F,STAR,VR > X1_STAR_VR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionUp
( X, XT,
XB, 0 );
while( XT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionUp
( XT, X0,
X1,
/**/ /**/
XB, X2 );
U01_MC_STAR.AlignWith( X0 );
X1_STAR_MR.AlignWith( X0 );
//--------------------------------------------------------------------//
U11_STAR_STAR = U11; // U11[* ,* ] <- U11[MC,MR]
X1_STAR_VR = X1; // X1[* ,VR] <- X1[MC,MR]
// X1[* ,VR] := U11^-1[* ,* ] X1[* ,VR]
LocalTrsm
( LEFT, UPPER, NORMAL, diag, F(1), U11_STAR_STAR, X1_STAR_VR,
checkIfSingular );
X1_STAR_MR = X1_STAR_VR; // X1[* ,MR] <- X1[* ,VR]
X1 = X1_STAR_MR; // X1[MC,MR] <- X1[* ,MR]
U01_MC_STAR = U01; // U01[MC,* ] <- U01[MC,MR]
// X0[MC,MR] -= U01[MC,* ] X1[* ,MR]
LocalGemm( NORMAL, NORMAL, F(-1), U01_MC_STAR, X1_STAR_MR, F(1), X0 );
//--------------------------------------------------------------------//
U01_MC_STAR.FreeAlignments();
X1_STAR_MR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
bool
ICBinaryArith_Int32::Compiler::generateStubCode(MacroAssembler& masm)
{
// Guard that R0 is an integer and R1 is an integer.
Label failure;
masm.branchTestInt32(Assembler::NotEqual, R0, &failure);
masm.branchTestInt32(Assembler::NotEqual, R1, &failure);
// Add R0 and R1. Don't need to explicitly unbox, just use R2.
Register Rscratch = R2_;
ARMRegister Wscratch = ARMRegister(Rscratch, 32);
#ifdef MERGE
// DIV and MOD need an extra non-volatile ValueOperand to hold R0.
AllocatableGeneralRegisterSet savedRegs(availableGeneralRegs(2));
savedRegs.set() = GeneralRegisterSet::Intersect(GeneralRegisterSet::NonVolatile(), savedRegs);
#endif
// get some more ARM-y names for the registers
ARMRegister W0(R0_, 32);
ARMRegister X0(R0_, 64);
ARMRegister W1(R1_, 32);
ARMRegister X1(R1_, 64);
ARMRegister WTemp(ExtractTemp0, 32);
ARMRegister XTemp(ExtractTemp0, 64);
Label maybeNegZero, revertRegister;
switch(op_) {
case JSOP_ADD:
masm.Adds(WTemp, W0, Operand(W1));
// Just jump to failure on overflow. R0 and R1 are preserved, so we can
// just jump to the next stub.
masm.j(Assembler::Overflow, &failure);
// Box the result and return. We know R0 already contains the
// integer tag, so we just need to move the payload into place.
masm.movePayload(ExtractTemp0, R0_);
break;
case JSOP_SUB:
masm.Subs(WTemp, W0, Operand(W1));
masm.j(Assembler::Overflow, &failure);
masm.movePayload(ExtractTemp0, R0_);
break;
case JSOP_MUL:
masm.mul32(R0.valueReg(), R1.valueReg(), Rscratch, &failure, &maybeNegZero);
masm.movePayload(Rscratch, R0_);
break;
case JSOP_DIV:
case JSOP_MOD: {
// Check for INT_MIN / -1, it results in a double.
Label check2;
masm.Cmp(W0, Operand(INT_MIN));
masm.B(&check2, Assembler::NotEqual);
masm.Cmp(W1, Operand(-1));
masm.j(Assembler::Equal, &failure);
masm.bind(&check2);
Label no_fail;
// Check for both division by zero and 0 / X with X < 0 (results in -0).
masm.Cmp(W1, Operand(0));
// If x > 0, then it can't be bad.
masm.B(&no_fail, Assembler::GreaterThan);
// if x == 0, then ignore any comparison, and force
// it to fail, if x < 0 (the only other case)
// then do the comparison, and fail if y == 0
masm.Ccmp(W0, Operand(0), vixl::ZFlag, Assembler::NotEqual);
masm.B(&failure, Assembler::Equal);
masm.bind(&no_fail);
masm.Sdiv(Wscratch, W0, W1);
// Start calculating the remainder, x - (x / y) * y.
masm.mul(WTemp, W1, Wscratch);
if (op_ == JSOP_DIV) {
// Result is a double if the remainder != 0, which happens
// when (x/y)*y != x.
masm.branch32(Assembler::NotEqual, R0.valueReg(), ExtractTemp0, &revertRegister);
masm.movePayload(Rscratch, R0_);
} else {
// Calculate the actual mod. Set the condition code, so we can see if it is non-zero.
masm.Subs(WTemp, W0, WTemp);
// If X % Y == 0 and X < 0, the result is -0.
masm.Ccmp(W0, Operand(0), vixl::NoFlag, Assembler::Equal);
masm.branch(Assembler::LessThan, &revertRegister);
masm.movePayload(ExtractTemp0, R0_);
}
break;
}
// ORR, EOR, AND can trivially be coerced int
// working without affecting the tag of the dest..
case JSOP_BITOR:
masm.Orr(X0, X0, Operand(X1));
break;
case JSOP_BITXOR:
masm.Eor(X0, X0, Operand(W1, vixl::UXTW));
break;
case JSOP_BITAND:
masm.And(X0, X0, Operand(X1));
break;
// LSH, RSH and URSH can not.
case JSOP_LSH:
// ARM will happily try to shift by more than 0x1f.
masm.Lsl(Wscratch, W0, W1);
masm.movePayload(Rscratch, R0.valueReg());
break;
case JSOP_RSH:
masm.Asr(Wscratch, W0, W1);
masm.movePayload(Rscratch, R0.valueReg());
break;
case JSOP_URSH:
masm.Lsr(Wscratch, W0, W1);
if (allowDouble_) {
Label toUint;
// Testing for negative is equivalent to testing bit 31
masm.Tbnz(Wscratch, 31, &toUint);
// Move result and box for return.
masm.movePayload(Rscratch, R0_);
EmitReturnFromIC(masm);
masm.bind(&toUint);
masm.convertUInt32ToDouble(Rscratch, ScratchDoubleReg);
masm.boxDouble(ScratchDoubleReg, R0, ScratchDoubleReg);
} else {
// Testing for negative is equivalent to testing bit 31
masm.Tbnz(Wscratch, 31, &failure);
// Move result for return.
masm.movePayload(Rscratch, R0_);
}
break;
default:
MOZ_CRASH("Unhandled op for BinaryArith_Int32.");
}
EmitReturnFromIC(masm);
switch (op_) {
case JSOP_MUL:
masm.bind(&maybeNegZero);
// Result is -0 if exactly one of lhs or rhs is negative.
masm.Cmn(W0, W1);
masm.j(Assembler::Signed, &failure);
// Result is +0, so use the zero register.
masm.movePayload(rzr, R0_);
EmitReturnFromIC(masm);
break;
case JSOP_DIV:
case JSOP_MOD:
masm.bind(&revertRegister);
break;
default:
break;
}
// Failure case - jump to next stub.
masm.bind(&failure);
EmitStubGuardFailure(masm);
return true;
}
inline void
TrmmLLTA
( Orientation orientation,
UnitOrNonUnit diag,
T alpha,
const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmLLTA");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( orientation == NORMAL )
throw std::logic_error
("TrmmLLTA expects a (Conjugate)Transpose option");
if( L.Height() != L.Width() || L.Height() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmLLTA: \n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<T>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<T>
XL(g), XR(g),
X0(g), X1(g), X2(g);
DistMatrix<T,MC,STAR> X1_MC_STAR(g);
DistMatrix<T,MR,STAR> Z1_MR_STAR(g);
DistMatrix<T,MR,MC > Z1_MR_MC(g);
X1_MC_STAR.AlignWith( L );
Z1_MR_STAR.AlignWith( L );
PartitionRight( X, XL, XR, 0 );
while( XL.Width() < X.Width() )
{
RepartitionRight
( XL, /**/ XR,
X0, /**/ X1, X2 );
Zeros( X1.Height(), X1.Width(), Z1_MR_STAR );
//--------------------------------------------------------------------//
X1_MC_STAR = X1;
LocalTrmmAccumulateLLT
( orientation, diag, alpha, L, X1_MC_STAR, Z1_MR_STAR );
Z1_MR_MC.SumScatterFrom( Z1_MR_STAR );
X1 = Z1_MR_MC;
//--------------------------------------------------------------------//
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void benchmark_sort(
const vex::Context &ctx, vex::profiler<> &prof
)
{
const size_t N = 16 * 1024 * 1024;
const size_t M = 16;
typedef typename std::conditional<
std::is_same<float, real>::value, cl_uint, cl_ulong
>::type key_type;
std::default_random_engine rng( std::rand() );
std::uniform_int_distribution<key_type> rnd;
std::vector<key_type> x0(N);
std::vector<key_type> x1(N);
std::generate(x0.begin(), x0.end(), [&]() { return rnd(rng); });
vex::vector<key_type> X0(ctx, x0);
vex::vector<key_type> X1(ctx, N);
X1 = X0;
vex::sort(X1);
double tot_time = 0;
for(size_t i = 0; i < M; i++) {
X1 = X0;
ctx.finish();
prof.tic_cpu("VexCL");
vex::sort(X1);
ctx.finish();
tot_time += prof.toc("VexCL");
}
std::cout
<< "Sort (" << vex::type_name<key_type>() << ")\n"
<< " VexCL: " << N * M / tot_time << " keys/sec\n";
#ifdef HAVE_BOOST_COMPUTE
X1 = X0;
vex::compute::sort(X1);
tot_time = 0;
for(size_t i = 0; i < M; i++) {
X1 = X0;
ctx.finish();
prof.tic_cpu("Boost.Compute");
vex::compute::sort(X1);
ctx.finish();
tot_time += prof.toc("Boost.Compute");
}
std::cout
<< " Boost.Compute: " << N * M / tot_time << " keys/sec\n";
#endif
#ifdef HAVE_CLOGS
X1 = X0;
vex::clogs::sort(X1);
tot_time = 0;
for(size_t i = 0; i < M; i++) {
X1 = X0;
ctx.finish();
prof.tic_cpu("CLOGS");
vex::clogs::sort(X1);
ctx.finish();
tot_time += prof.toc("CLOGS");
}
std::cout
<< " CLOGS: " << N * M / tot_time << " keys/sec\n";
#endif
if (options.bm_cpu) {
tot_time = 0;
for(size_t i = 0; i < M; i++) {
std::copy(x0.begin(), x0.end(), x1.begin());
prof.tic_cpu("STL");
std::sort(x1.begin(), x1.end());
tot_time += prof.toc("STL");
}
std::cout << " STL: " << N * M / tot_time << " keys/sec\n";
}
std::cout << std::endl;
}
static void
ec_mulm (gcry_mpi_t w, gcry_mpi_t u, gcry_mpi_t v, mpi_ec_t ctx)
{
#if 0
/* NOTE: This code works only for limb sizes of 32 bit. */
mpi_limb_t *wp, *sp;
if (ctx->nist_nbits == 192)
{
mpi_mul (w, u, v);
mpi_resize (w, 12);
wp = w->d;
sp = ctx->s[0]->d;
sp[0*2+0] = wp[0*2+0];
sp[0*2+1] = wp[0*2+1];
sp[1*2+0] = wp[1*2+0];
sp[1*2+1] = wp[1*2+1];
sp[2*2+0] = wp[2*2+0];
sp[2*2+1] = wp[2*2+1];
sp = ctx->s[1]->d;
sp[0*2+0] = wp[3*2+0];
sp[0*2+1] = wp[3*2+1];
sp[1*2+0] = wp[3*2+0];
sp[1*2+1] = wp[3*2+1];
sp[2*2+0] = 0;
sp[2*2+1] = 0;
sp = ctx->s[2]->d;
sp[0*2+0] = 0;
sp[0*2+1] = 0;
sp[1*2+0] = wp[4*2+0];
sp[1*2+1] = wp[4*2+1];
sp[2*2+0] = wp[4*2+0];
sp[2*2+1] = wp[4*2+1];
sp = ctx->s[3]->d;
sp[0*2+0] = wp[5*2+0];
sp[0*2+1] = wp[5*2+1];
sp[1*2+0] = wp[5*2+0];
sp[1*2+1] = wp[5*2+1];
sp[2*2+0] = wp[5*2+0];
sp[2*2+1] = wp[5*2+1];
ctx->s[0]->nlimbs = 6;
ctx->s[1]->nlimbs = 6;
ctx->s[2]->nlimbs = 6;
ctx->s[3]->nlimbs = 6;
mpi_add (ctx->c, ctx->s[0], ctx->s[1]);
mpi_add (ctx->c, ctx->c, ctx->s[2]);
mpi_add (ctx->c, ctx->c, ctx->s[3]);
while ( mpi_cmp (ctx->c, ctx->p ) >= 0 )
mpi_sub ( ctx->c, ctx->c, ctx->p );
mpi_set (w, ctx->c);
}
else if (ctx->nist_nbits == 384)
{
int i;
mpi_mul (w, u, v);
mpi_resize (w, 24);
wp = w->d;
#define NEXT(a) do { ctx->s[(a)]->nlimbs = 12; \
sp = ctx->s[(a)]->d; \
i = 0; } while (0)
#define X(a) do { sp[i++] = wp[(a)];} while (0)
#define X0(a) do { sp[i++] = 0; } while (0)
NEXT(0);
X(0);X(1);X(2);X(3);X(4);X(5);X(6);X(7);X(8);X(9);X(10);X(11);
NEXT(1);
X0();X0();X0();X0();X(21);X(22);X(23);X0();X0();X0();X0();X0();
NEXT(2);
X(12);X(13);X(14);X(15);X(16);X(17);X(18);X(19);X(20);X(21);X(22);X(23);
NEXT(3);
X(21);X(22);X(23);X(12);X(13);X(14);X(15);X(16);X(17);X(18);X(19);X(20);
NEXT(4);
X0();X(23);X0();X(20);X(12);X(13);X(14);X(15);X(16);X(17);X(18);X(19);
NEXT(5);
X0();X0();X0();X0();X(20);X(21);X(22);X(23);X0();X0();X0();X0();
NEXT(6);
X(20);X0();X0();X(21);X(22);X(23);X0();X0();X0();X0();X0();X0();
NEXT(7);
X(23);X(12);X(13);X(14);X(15);X(16);X(17);X(18);X(19);X(20);X(21);X(22);
NEXT(8);
X0();X(20);X(21);X(22);X(23);X0();X0();X0();X0();X0();X0();X0();
NEXT(9);
X0();X0();X0();X(23);X(23);X0();X0();X0();X0();X0();X0();X0();
#undef X0
#undef X
#undef NEXT
mpi_add (ctx->c, ctx->s[0], ctx->s[1]);
mpi_add (ctx->c, ctx->c, ctx->s[1]);
mpi_add (ctx->c, ctx->c, ctx->s[2]);
mpi_add (ctx->c, ctx->c, ctx->s[3]);
mpi_add (ctx->c, ctx->c, ctx->s[4]);
mpi_add (ctx->c, ctx->c, ctx->s[5]);
mpi_add (ctx->c, ctx->c, ctx->s[6]);
mpi_sub (ctx->c, ctx->c, ctx->s[7]);
mpi_sub (ctx->c, ctx->c, ctx->s[8]);
mpi_sub (ctx->c, ctx->c, ctx->s[9]);
while ( mpi_cmp (ctx->c, ctx->p ) >= 0 )
mpi_sub ( ctx->c, ctx->c, ctx->p );
while ( ctx->c->sign )
mpi_add ( ctx->c, ctx->c, ctx->p );
mpi_set (w, ctx->c);
}
else
#endif /*0*/
mpi_mulm (w, u, v, ctx->p);
}
inline void
TrsmLLTLarge
( Orientation orientation, UnitOrNonUnit diag,
F alpha, const DistMatrix<F>& L, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmLLTLarge");
if( orientation == NORMAL )
throw std::logic_error("TrsmLLT expects a (Conjugate)Transpose option");
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<F>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<F> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<F,STAR,MC > L10_STAR_MC(g);
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,MR > X1_STAR_MR(g);
DistMatrix<F,STAR,VR > X1_STAR_VR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionUp
( X, XT,
XB, 0 );
while( XT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
RepartitionUp
( XT, X0,
X1,
/**/ /**/
XB, X2 );
L10_STAR_MC.AlignWith( X0 );
X1_STAR_MR.AlignWith( X0 );
//--------------------------------------------------------------------//
L11_STAR_STAR = L11; // L11[* ,* ] <- L11[MC,MR]
X1_STAR_VR = X1; // X1[* ,VR] <- X1[MC,MR]
// X1[* ,VR] := L11^-[T/H][* ,* ] X1[* ,VR]
LocalTrsm
( LEFT, LOWER, orientation, diag, F(1), L11_STAR_STAR, X1_STAR_VR,
checkIfSingular );
X1_STAR_MR = X1_STAR_VR; // X1[* ,MR] <- X1[* ,VR]
X1 = X1_STAR_MR; // X1[MC,MR] <- X1[* ,MR]
L10_STAR_MC = L10; // L10[* ,MC] <- L10[MC,MR]
// X0[MC,MR] -= (L10[* ,MC])^(T/H) X1[* ,MR]
// = L10^[T/H][MC,* ] X1[* ,MR]
LocalGemm
( orientation, NORMAL, F(-1), L10_STAR_MC, X1_STAR_MR, F(1), X0 );
//--------------------------------------------------------------------//
L10_STAR_MC.FreeAlignments();
X1_STAR_MR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void OpticalFlow::baseCalculate(cv::Mat& Im1, cv::Mat& Im2, flowUV& UV, const OpticalFlowParams& params){
int rows = Im1.rows;
int cols = Im1.cols;
FlowOperator flowOp(rows, cols);
FArray X0(2 * rows * cols, false);
FArray dUdV(2 * rows * cols, true, 0);
cv::Mat Ix1(rows, cols, OPTFLOW_TYPE);
cv::Mat Iy1(rows, cols, OPTFLOW_TYPE);
cv::Mat Ix(rows, cols, OPTFLOW_TYPE);
cv::Mat Iy(rows, cols, OPTFLOW_TYPE);
getDXsCV(Im1, Ix1, Iy1);
for (int i = 0; i < params.getIters(); ++i){
cv::Mat Ix2(rows, cols, OPTFLOW_TYPE);
cv::Mat Iy2(rows, cols, OPTFLOW_TYPE);
cv::Mat It(rows, cols, OPTFLOW_TYPE);
cv::Mat im2Warpped(rows, cols, Im1.type());
WarpImage(Im2, UV.getU(), UV.getV(), im2Warpped);
getDXsCV(im2Warpped, Ix2, Iy2);
Ix = params.getWeightedDeriveFactor() * (Ix1 + Ix2);
Iy = params.getWeightedDeriveFactor() * (Iy1 + Iy2);
cv::subtract(im2Warpped, Im1, It);
if (params.getDisplayDerivativs()){
cv::imshow("Derivative Ix", Ix);
cv::imshow("Derivative Iy", Iy);
cv::waitKey(1);
}
cv::Mat Du(rows, cols, OPTFLOW_TYPE, cv::Scalar(0));
cv::Mat Dv(rows, cols, OPTFLOW_TYPE, cv::Scalar(0));
for (int j = 0; j < params.getLinearIters(); ++j){
#if OPTFLOW_VERBOSE
cout << "solving Ax=b with SOR ";
clock_t start = std::clock();
#endif
flowOp.construct(UV, Du, Dv, Ix, Iy, It, params);
memcpy(X0.ptr, UV.getU().data, rows * cols * sizeof(float));
memcpy(X0.ptr + (rows * cols), UV.getV().data, rows * cols * sizeof(float));
//UtilsDebug::printCRSSparseMat(flowOp.getA(), "aaaa.txt");
if (params.getCheckResidualTolerance()){
LinearSolver::sparseMatSor(flowOp.getA(), X0 ,dUdV, flowOp.getb(), params.getOverRelaxation(), params.getSorIters(), params.getResidualTolerance());
}else{
//LinearSolver::multigrid(10,10,flowOp.getA(),flowOp.getb(), params.getResidualTolerance(), dUdV, 20, 20, LinearSolver::vCycle);
LinearSolver::sparseMatSorNoResidual(flowOp.getA(), X0 ,dUdV, flowOp.getb(), params.getOverRelaxation(), params.getSorIters());
}
#if OPTFLOW_VERBOSE
std::cout<<" --- "<< (std::clock() - start) / (double)CLOCKS_PER_SEC <<'\n';
#endif
#if OPTFLOW_DEBUG
for(int i = 0; i < dUdV.size(); ++i){
if (!(dUdV.ptr[i] == dUdV.ptr[i])){
cout << "ERROR - NAN";
}
}
#endif
UtilsMat::clamp(dUdV, -1, 1);
memcpy(Du.data, dUdV.ptr, rows * cols * sizeof(float));
memcpy(Dv.data, dUdV.ptr + (rows * cols), rows * cols * sizeof(float));
flowUV UV0(UV);
UV.getU() += Du;
UV.getV() += Dv;
cv::Mat tmpU, tmpV;
UV.getU().copyTo(tmpU);
UV.getV().copyTo(tmpV);
WeightedMedianFilter::computeMedianFilter(UV.getU(), UV.getV(), Im1, Im2, params.getMedianFilterRadius());
Du = UV.getU() - UV0.getU();
Dv = UV.getV() - UV0.getV();
UV0.getU().copyTo(UV.getU());
UV0.getV().copyTo(UV.getV());
UV0.getU() += Du;
UV0.getV() += Dv;
if (params.isDisplay())
UtilsFlow::DrawFlow(UV0.getU(), UV0.getV(), "Flow");
}
UV.getU() += Du;
UV.getV() += Dv;
}
}
static
int f0_val(int base)
{
int id = X0(base);
return id2val(id);
} /* f0_val */
inline void
TrsmRUT
( Orientation orientation, UnitOrNonUnit diag,
F alpha, const DistMatrix<F>& U, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmRUT");
if( orientation == NORMAL )
throw std::logic_error("TrsmRUT expects a (Conjugate)Transpose option");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<F>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
DistMatrix<F> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<F,VR, STAR> U01_VR_STAR(g);
DistMatrix<F,STAR,MR > U01AdjOrTrans_STAR_MR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,VC, STAR> X1_VC_STAR(g);
DistMatrix<F,STAR,MC > X1Trans_STAR_MC(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionLeft( X, XL, XR, 0 );
while( XL.Width() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionLeft
( XL, /**/ XR,
X0, X1, /**/ X2 );
X1_VC_STAR.AlignWith( X0 );
X1Trans_STAR_MC.AlignWith( X0 );
U01_VR_STAR.AlignWith( X0 );
U01AdjOrTrans_STAR_MR.AlignWith( X0 );
//--------------------------------------------------------------------//
U11_STAR_STAR = U11;
X1_VC_STAR = X1;
LocalTrsm
( RIGHT, UPPER, orientation, diag,
F(1), U11_STAR_STAR, X1_VC_STAR, checkIfSingular );
X1Trans_STAR_MC.TransposeFrom( X1_VC_STAR );
X1.TransposeFrom( X1Trans_STAR_MC );
U01_VR_STAR = U01;
if( orientation == ADJOINT )
U01AdjOrTrans_STAR_MR.AdjointFrom( U01_VR_STAR );
else
U01AdjOrTrans_STAR_MR.TransposeFrom( U01_VR_STAR );
// X0[MC,MR] -= X1[MC,* ] (U01[MR,* ])^(T/H)
// = X1^T[* ,MC] (U01^(T/H))[* ,MR]
LocalGemm
( TRANSPOSE, NORMAL,
F(-1), X1Trans_STAR_MC, U01AdjOrTrans_STAR_MR, F(1), X0 );
//--------------------------------------------------------------------//
X1_VC_STAR.FreeAlignments();
X1Trans_STAR_MC.FreeAlignments();
U01_VR_STAR.FreeAlignments();
U01AdjOrTrans_STAR_MR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionLeft
( XL, /**/ XR,
X0, /**/ X1, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrmmLLNC
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrmmLLNC");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( L.Height() != L.Width() || L.Width() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmLLNC: \n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<T>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<T> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<T,MC, STAR> L21_MC_STAR(g);
DistMatrix<T,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<T,STAR,VR > X1_STAR_VR(g);
DistMatrix<T,MR, STAR> X1Trans_MR_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionUp
( X, XT,
XB, 0 );
while( XT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
RepartitionUp
( XT, X0,
X1,
/**/ /**/
XB, X2 );
L21_MC_STAR.AlignWith( X2 );
X1Trans_MR_STAR.AlignWith( X2 );
X1_STAR_VR.AlignWith( X1 );
//--------------------------------------------------------------------//
L21_MC_STAR = L21;
X1Trans_MR_STAR.TransposeFrom( X1 );
LocalGemm
( NORMAL, TRANSPOSE, T(1), L21_MC_STAR, X1Trans_MR_STAR, T(1), X2 );
L11_STAR_STAR = L11;
X1_STAR_VR.TransposeFrom( X1Trans_MR_STAR );
LocalTrmm( LEFT, LOWER, NORMAL, diag, T(1), L11_STAR_STAR, X1_STAR_VR );
X1 = X1_STAR_VR;
//--------------------------------------------------------------------//
L21_MC_STAR.FreeAlignments();
X1Trans_MR_STAR.FreeAlignments();
X1_STAR_VR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
}
inline void
TrmmRUNC
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& U,
DistMatrix<T>& X )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrmmRUNC");
if( U.Grid() != X.Grid() )
throw std::logic_error
("U and X must be distributed over the same grid");
if( U.Height() != U.Width() || X.Width() != U.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmRUNC: \n"
<< " U ~ " << U.Height() << " x " << U.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<T>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
DistMatrix<T> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<T,MR, STAR> U12Trans_MR_STAR(g);
DistMatrix<T,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<T,VC, STAR> X1_VC_STAR(g);
DistMatrix<T,MC, STAR> X1_MC_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionLeft( X, XL, XR, 0 );
while( XL.Width() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionLeft
( XL, /**/ XR,
X0, X1, /**/ X2 );
X1_MC_STAR.AlignWith( X2 );
U12Trans_MR_STAR.AlignWith( X2 );
X1_VC_STAR.AlignWith( X1 );
//--------------------------------------------------------------------//
X1_MC_STAR = X1;
U12Trans_MR_STAR.TransposeFrom( U12 );
LocalGemm
( NORMAL, TRANSPOSE, T(1), X1_MC_STAR, U12Trans_MR_STAR, T(1), X2 );
U11_STAR_STAR = U11;
X1_VC_STAR = X1_MC_STAR;
LocalTrmm
( RIGHT, UPPER, NORMAL, diag, T(1), U11_STAR_STAR, X1_VC_STAR );
X1 = X1_VC_STAR;
//--------------------------------------------------------------------//
X1_MC_STAR.FreeAlignments();
U12Trans_MR_STAR.FreeAlignments();
X1_VC_STAR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionLeft
( XL, /**/ XR,
X0, /**/ X1, X2 );
}
}
inline void
TrsmLLTSmall
( Orientation orientation, UnitOrNonUnit diag,
F alpha, const DistMatrix<F,STAR,VR>& L, DistMatrix<F,VR,STAR>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmLLTSmall");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( orientation == NORMAL )
throw std::logic_error("TrsmLLT expects a (Conjugate)Transpose option");
if( L.Height() != L.Width() || L.Height() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrsmLLT: \n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
if( L.RowAlignment() != X.ColAlignment() )
throw std::logic_error("L and X must be aligned");
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<F,STAR,VR>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<F,VR,STAR> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> X1_STAR_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionUp
( X, XT,
XB, 0 );
while( XT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
RepartitionUp
( XT, X0,
X1,
/**/ /**/
XB, X2 );
//--------------------------------------------------------------------//
L11_STAR_STAR = L11; // L11[* ,* ] <- L11[* ,VR]
X1_STAR_STAR = X1; // X1[* ,* ] <- X1[VR,* ]
// X1[* ,* ] := L11^-[T/H][* ,* ] X1[* ,* ]
LocalTrsm
( LEFT, LOWER, orientation, diag,
F(1), L11_STAR_STAR, X1_STAR_STAR, checkIfSingular );
X1 = X1_STAR_STAR;
// X0[VR,* ] -= L10[* ,VR]^(T/H) X1[* ,* ]
LocalGemm( orientation, NORMAL, F(-1), L10, X1_STAR_STAR, F(1), X0 );
//--------------------------------------------------------------------//
SlideLockedPartitionUpDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrmmRLNCOld
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmRLNCOld");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( L.Height() != L.Width() || X.Width() != L.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmRLNC: \n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<T>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<T> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<T,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<T,MR, STAR> L21_MR_STAR(g);
DistMatrix<T,VC, STAR> X1_VC_STAR(g);
DistMatrix<T,MC, STAR> D1_MC_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionRight( X, XL, XR, 0 );
while( XR.Width() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionRight
( XL, /**/ XR,
X0, /**/ X1, X2 );
L21_MR_STAR.AlignWith( X2 );
D1_MC_STAR.AlignWith( X1 );
Zeros( X1.Height(), X1.Width(), D1_MC_STAR );
//--------------------------------------------------------------------//
X1_VC_STAR = X1;
L11_STAR_STAR = L11;
LocalTrmm
( RIGHT, LOWER, NORMAL, diag, T(1), L11_STAR_STAR, X1_VC_STAR );
X1 = X1_VC_STAR;
L21_MR_STAR = L21;
LocalGemm( NORMAL, NORMAL, T(1), X2, L21_MR_STAR, T(0), D1_MC_STAR );
X1.SumScatterUpdate( T(1), D1_MC_STAR );
//--------------------------------------------------------------------//
L21_MR_STAR.FreeAlignments();
D1_MC_STAR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void StereogramWidget::paintEvent(QPaintEvent* event)
{
m_radius = 0; //means that nothing has been drawn (yet ;)
QLabel::paintEvent(event);
QPainter painter(this);
painter.setRenderHints(QPainter::Antialiasing | QPainter::TextAntialiasing, true);
//pen
QPen pen;
pen.setStyle(Qt::SolidLine);
pen.setBrush(QColor(Qt::black));
int diameter = std::min(width(),height());
int halfW = width()/2;
int halfH = height()/2;
QPoint center(halfW,halfH);
int hsvThickness = 0;
if (m_showHSVRing)
{
int newDiameter = static_cast<int>(ceil(0.9*static_cast<double>(diameter)));
hsvThickness = diameter - newDiameter;
//TODO
if (hsvThickness > 0)
{
QRect rectangle(center.x()-diameter/2+1,center.y()-diameter/2+1,diameter-2,diameter-2);
int angle_span = static_cast<int>(m_angularStep_deg * 16.0); //see QPainter::drawPie
QBrush brush;
brush.setStyle(Qt::SolidPattern);
painter.setPen(Qt::NoPen);
//dip direction steps (dip dir. in [0,360])
unsigned ddSteps = static_cast<unsigned>(ceil(360.0 / std::max(m_angularStep_deg,1.0)));
for (unsigned j=0; j<ddSteps; ++j)
{
double dipDir_deg = static_cast<double>(j) * m_angularStep_deg;
//set family color
ccColor::Rgb col;
FacetsClassifier::GenerateSubfamilyColor(col, 90.0, dipDir_deg + 0.5 * m_angularStep_deg, 0, 1);
brush.setColor(QColor( static_cast<int>(col.r),
static_cast<int>(col.g),
static_cast<int>(col.b),
255));
painter.setBrush(brush);
int angle_start = static_cast<int>((360.0 - dipDir_deg - m_angularStep_deg + 90.0) * 16.0); //see QPainter::drawPie
painter.drawPie(rectangle,angle_start,angle_span);
}
}
diameter = newDiameter;
}
//outer circle
pen.setWidth(2);
painter.setPen(pen);
painter.setBrush(Qt::white);
int radius = diameter/2 - 2;
painter.drawEllipse(center,radius,radius);
painter.setBrush(Qt::NoBrush);
//keep track of the circle position
m_radius = radius;
m_center = center;
//main axes
painter.drawLine(center-QPoint(radius,0),center+QPoint(radius,0));
painter.drawLine(center-QPoint(0,radius),center+QPoint(0,radius));
//draw circles
if (m_angularStep_deg > 0)
{
//dip steps (dip in [0,90])
unsigned dSteps = static_cast<unsigned>(ceil(90.0 / m_angularStep_deg));
//dip direction steps (dip dir. in [0,360])
unsigned ddSteps = static_cast<unsigned>(ceil(360.0 / m_angularStep_deg));
//draw inner circles
pen.setWidth(1);
pen.setColor(Qt::gray);
painter.setPen(pen);
for (unsigned i=1; i<dSteps; ++i)
{
double dip_deg = static_cast<double>(i) * m_angularStep_deg;
if (dip_deg < 90.0)
{
int R = static_cast<int>(static_cast<double>(radius) * (dip_deg/90.0));
if (R > 1)
painter.drawEllipse(center,R-1,R-1);
}
}
//draw rays (+ 'm_ticksFreq' times more ticks)
int ticksFreq = std::max(m_ticksFreq,1);
for (unsigned j=1; j<=ddSteps*ticksFreq; ++j)
{
double dipDir_deg = static_cast<double>(j) * m_angularStep_deg / static_cast<double>(ticksFreq);
if (dipDir_deg < 360.0)
{
QPoint X( static_cast<int>(sin(dipDir_deg * CC_DEG_TO_RAD) * static_cast<double>(radius)),
-static_cast<int>(cos(dipDir_deg * CC_DEG_TO_RAD) * static_cast<double>(radius)) );
if ((j % ticksFreq) == 0) //long ticks
painter.drawLine(center,center+X);
else
painter.drawLine(center+X*0.93,center+X);
}
}
}
//draw density map
if (m_densityGrid && m_densityColorScale && m_densityGrid->grid && m_densityGrid->minMaxDensity[1] != 0)
{
assert(m_densityColorScale);
assert(m_densityGrid->grid);
QBrush brush;
brush.setStyle(Qt::SolidPattern);
painter.setPen(Qt::NoPen);
QPolygon poly(4);
const double* d = m_densityGrid->grid;
for (unsigned j=0; j<m_densityGrid->ddSteps; ++j)
{
double dipDir0_rad = static_cast<double>(j) * m_densityGrid->step_deg * CC_DEG_TO_RAD;
double dipDir1_rad = static_cast<double>(j+1) * m_densityGrid->step_deg * CC_DEG_TO_RAD;
double cos_dipDir0 = cos(dipDir0_rad);
double sin_dipDir0 = sin(dipDir0_rad);
double cos_dipDir1 = cos(dipDir1_rad);
double sin_dipDir1 = sin(dipDir1_rad);
for (unsigned i=0; i<m_densityGrid->rSteps; ++i, ++d)
{
if (*d != 0)
{
double relPos = static_cast<double>(*d)/static_cast<double>(m_densityGrid->minMaxDensity[1]);
const colorType* col = m_densityColorScale->getColorByRelativePos(relPos,m_densityColorScaleSteps);
brush.setColor(QColor( static_cast<int>(col[0]),
static_cast<int>(col[1]),
static_cast<int>(col[2]),
255));
painter.setBrush(brush);
//stereographic projection
//double dip0_rad = static_cast<double>(i) * m_densityGrid->step_deg * CC_DEG_TO_RAD;
//double dip1_rad = static_cast<double>(i+1) * m_densityGrid->step_deg * CC_DEG_TO_RAD;
//double R0 = static_cast<double>(radius) * cos(dip0_rad) / (1.0 + sin(dip0_rad));
//double R1 = static_cast<double>(radius) * cos(dip1_rad) / (1.0 + sin(dip1_rad));
double R0 = static_cast<double>(radius) * static_cast<double>(i) * m_densityGrid->step_R;
double R1 = static_cast<double>(radius) * static_cast<double>(i+1) * m_densityGrid->step_R;
poly.setPoint(0,center+QPoint(static_cast<int>(sin_dipDir0 * R0),-static_cast<int>(cos_dipDir0 * R0)));
poly.setPoint(1,center+QPoint(static_cast<int>(sin_dipDir0 * R1),-static_cast<int>(cos_dipDir0 * R1)));
poly.setPoint(2,center+QPoint(static_cast<int>(sin_dipDir1 * R1),-static_cast<int>(cos_dipDir1 * R1)));
poly.setPoint(3,center+QPoint(static_cast<int>(sin_dipDir1 * R0),-static_cast<int>(cos_dipDir1 * R0)));
painter.drawPolygon(poly);
}
}
}
}
//draw main 'dip direction'
if (m_meanDipDir_deg >= 0)
{
pen.setWidth(2);
pen.setColor(Qt::red);
painter.setPen(pen);
//draw main direction
QPoint X( static_cast<int>(sin(m_meanDipDir_deg * CC_DEG_TO_RAD) * static_cast<double>(radius)),
-static_cast<int>(cos(m_meanDipDir_deg * CC_DEG_TO_RAD) * static_cast<double>(radius)) );
pen.setStyle(Qt::DashLine);
painter.setPen(pen);
painter.drawLine(center,center+X);
//draw orthogonal to main direction
QPoint Y( static_cast<int>(cos(m_meanDipDir_deg * CC_DEG_TO_RAD) * static_cast<double>(radius)),
static_cast<int>(sin(m_meanDipDir_deg * CC_DEG_TO_RAD) * static_cast<double>(radius)) );
pen.setStyle(Qt::SolidLine);
painter.setPen(pen);
painter.drawLine(center-Y,center+Y);
}
//draw filter window around last cliked point
if (m_trackMouseClick)
{
pen.setWidth(2);
pen.setColor(Qt::magenta);
painter.setPen(pen);
//QBrush brush;
//brush.setStyle(Qt::Dense6Pattern);
//brush.setColor(Qt::red);
//painter.setBrush(brush);
painter.setBrush(Qt::NoBrush);
double R0 = static_cast<double>(radius) * (std::max(0.0,m_clickDip_deg-m_clickDipSpan_deg/2)/90.0);
double R1 = static_cast<double>(radius) * (std::min(90.0,m_clickDip_deg+m_clickDipSpan_deg/2)/90.0);
//draw radial limits
{
QPoint X0( static_cast<int>(sin((m_clickDipDir_deg-m_clickDipDirSpan_deg/2) * CC_DEG_TO_RAD) * R0),
-static_cast<int>(cos((m_clickDipDir_deg-m_clickDipDirSpan_deg/2) * CC_DEG_TO_RAD) * R0) );
QPoint X1( static_cast<int>(sin((m_clickDipDir_deg-m_clickDipDirSpan_deg/2) * CC_DEG_TO_RAD) * R1),
-static_cast<int>(cos((m_clickDipDir_deg-m_clickDipDirSpan_deg/2) * CC_DEG_TO_RAD) * R1) );
painter.drawLine(center+X0,center+X1);
}
{
QPoint X0( static_cast<int>(sin((m_clickDipDir_deg+m_clickDipDirSpan_deg/2) * CC_DEG_TO_RAD) * R0),
-static_cast<int>(cos((m_clickDipDir_deg+m_clickDipDirSpan_deg/2) * CC_DEG_TO_RAD) * R0) );
QPoint X1( static_cast<int>(sin((m_clickDipDir_deg+m_clickDipDirSpan_deg/2) * CC_DEG_TO_RAD) * R1),
-static_cast<int>(cos((m_clickDipDir_deg+m_clickDipDirSpan_deg/2) * CC_DEG_TO_RAD) * R1) );
painter.drawLine(center+X0,center+X1);
}
//draw concentric limits
{
int angle_start = static_cast<int>((360.0 - m_clickDipDir_deg - m_clickDipDirSpan_deg/2 + 90.0) * 16.0); //see QPainter::drawPie
int angle_span = static_cast<int>(m_clickDipDirSpan_deg * 16.0); //see QPainter::drawPie
QRectF rect0(static_cast<double>(center.x()) - R0,
static_cast<double>(center.y()) - R0,
2*R0, 2*R0);
painter.drawArc(rect0,angle_start,angle_span);
QRectF rect1(static_cast<double>(center.x()) - R1,
static_cast<double>(center.y()) - R1,
2*R1, 2*R1);
painter.drawArc(rect1,angle_start,angle_span);
}
}
}
inline void
TrsmRUN
( UnitOrNonUnit diag, F alpha, const DistMatrix<F>& U, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmRUN");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<F>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
DistMatrix<F> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,MR > U12_STAR_MR(g);
DistMatrix<F,VC, STAR> X1_VC_STAR(g);
DistMatrix<F,STAR,MC > X1Trans_STAR_MC(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionRight( X, XL, XR, 0 );
while( XR.Width() > 0 )
{
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
RepartitionRight
( XL, /**/ XR,
X0, /**/ X1, X2 );
X1_VC_STAR.AlignWith( X2 );
X1Trans_STAR_MC.AlignWith( X2 );
U12_STAR_MR.AlignWith( X2 );
//--------------------------------------------------------------------//
U11_STAR_STAR = U11;
X1_VC_STAR = X1;
LocalTrsm
( RIGHT, UPPER, NORMAL, diag, F(1), U11_STAR_STAR, X1_VC_STAR,
checkIfSingular );
X1Trans_STAR_MC.TransposeFrom( X1_VC_STAR );
X1.TransposeFrom( X1Trans_STAR_MC );
U12_STAR_MR = U12;
// X2[MC,MR] -= X1[MC,* ] U12[* ,MR]
// = X1^T[* ,MC] U12[* ,MR]
LocalGemm
( TRANSPOSE, NORMAL, F(-1), X1Trans_STAR_MC, U12_STAR_MR, F(1), X2 );
//--------------------------------------------------------------------//
X1_VC_STAR.FreeAlignments();
X1Trans_STAR_MC.FreeAlignments();
U12_STAR_MR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void benchmark_scan(
const vex::Context &ctx, vex::profiler<> &prof
)
{
const size_t N = 16 * 1024 * 1024;
const size_t M = 16;
typedef typename std::conditional<
std::is_same<float, real>::value, cl_uint, cl_ulong
>::type key_type;
std::default_random_engine rng( std::rand() );
std::uniform_int_distribution<key_type> rnd;
std::vector<key_type> x0(N);
std::vector<key_type> x1(N);
std::generate(x0.begin(), x0.end(), [&]() { return rnd(rng); });
vex::vector<key_type> X0(ctx, x0);
vex::vector<key_type> X1(ctx, N);
vex::exclusive_scan(X0, X1);
ctx.finish();
prof.tic_cpu("VexCL");
for(size_t i = 0; i < M; i++)
vex::exclusive_scan(X0, X1);
ctx.finish();
double tot_time = prof.toc("VexCL");
std::cout
<< "Scan (" << vex::type_name<key_type>() << ")\n"
<< " VexCL: " << N * M / tot_time << " keys/sec\n";
#ifdef HAVE_BOOST_COMPUTE
vex::compute::exclusive_scan(X0, X1);
ctx.finish();
prof.tic_cpu("Boost.Compute");
for(size_t i = 0; i < M; i++)
vex::compute::exclusive_scan(X0, X1);
ctx.finish();
tot_time = prof.toc("Boost.Compute");
std::cout
<< " Boost.Compute: " << N * M / tot_time << " keys/sec\n";
#endif
#ifdef HAVE_CLOGS
vex::clogs::exclusive_scan(X0, X1);
ctx.finish();
prof.tic_cpu("CLOGS");
for(size_t i = 0; i < M; i++)
vex::clogs::exclusive_scan(X0, X1);
ctx.finish();
tot_time = prof.toc("CLOGS");
std::cout
<< " CLOGS: " << N * M / tot_time << " keys/sec\n";
#endif
if (options.bm_cpu) {
prof.tic_cpu("CPU");
for(size_t i = 0; i < M; i++) {
key_type sum = key_type();
for(size_t j = 0; j < N; ++j) {
key_type next = sum + x0[j];
x1[j] = sum;
sum = next;
}
}
tot_time = prof.toc("CPU");
std::cout << " CPU: " << N * M / tot_time << " keys/sec\n";
}
std::cout << std::endl;
}
inline void
TrsmRLT
( Orientation orientation, UnitOrNonUnit diag,
F alpha, const DistMatrix<F>& L, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrsmRLT");
if( orientation == NORMAL )
LogicError("TrsmRLT expects a (Conjugate)Transpose option");
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<F>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<F> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,VR, STAR> L21_VR_STAR(g);
DistMatrix<F,STAR,MR > L21AdjOrTrans_STAR_MR(g);
DistMatrix<F,VC, STAR> X1_VC_STAR(g);
DistMatrix<F,STAR,MC > X1Trans_STAR_MC(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionRight( X, XL, XR, 0 );
while( XR.Width() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionRight
( XL, /**/ XR,
X0, /**/ X1, X2 );
X1_VC_STAR.AlignWith( X2 );
X1Trans_STAR_MC.AlignWith( X2 );
L21_VR_STAR.AlignWith( X2 );
L21AdjOrTrans_STAR_MR.AlignWith( X2 );
//--------------------------------------------------------------------//
L11_STAR_STAR = L11;
X1_VC_STAR = X1;
LocalTrsm
( RIGHT, LOWER, orientation, diag,
F(1), L11_STAR_STAR, X1_VC_STAR, checkIfSingular );
X1Trans_STAR_MC.TransposeFrom( X1_VC_STAR );
X1.TransposeFrom( X1Trans_STAR_MC );
L21_VR_STAR = L21;
if( orientation == ADJOINT )
L21AdjOrTrans_STAR_MR.AdjointFrom( L21_VR_STAR );
else
L21AdjOrTrans_STAR_MR.TransposeFrom( L21_VR_STAR );
// X2[MC,MR] -= X1[MC,*] (L21[MR,*])^(T/H)
// = X1^T[* ,MC] (L21^(T/H))[*,MR]
LocalGemm
( TRANSPOSE, NORMAL,
F(-1), X1Trans_STAR_MC, L21AdjOrTrans_STAR_MR, F(1), X2 );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
}
inline void
TrmmLLTCOld
( Orientation orientation,
UnitOrNonUnit diag,
T alpha,
const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmLLTCOld");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( orientation == NORMAL )
throw std::logic_error("TrmmLLT expects a (Conjugate)Transpose option");
if( L.Height() != L.Width() || L.Height() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmLLTC: \n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<T>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<T> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<T,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<T,MC, STAR> L21_MC_STAR(g);
DistMatrix<T,STAR,VR > X1_STAR_VR(g);
DistMatrix<T,MR, STAR> D1AdjOrTrans_MR_STAR(g);
DistMatrix<T,MR, MC > D1AdjOrTrans_MR_MC(g);
DistMatrix<T,MC, MR > D1(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionDown
( X, XT,
XB, 0 );
while( XB.Height() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionDown
( XT, X0,
/**/ /**/
X1,
XB, X2 );
L21_MC_STAR.AlignWith( X2 );
D1AdjOrTrans_MR_STAR.AlignWith( X1 );
D1AdjOrTrans_MR_MC.AlignWith( X1 );
D1.AlignWith( X1 );
Zeros( X1.Width(), X1.Height(), D1AdjOrTrans_MR_STAR );
Zeros( X1.Height(), X1.Width(), D1 );
//--------------------------------------------------------------------//
X1_STAR_VR = X1;
L11_STAR_STAR = L11;
LocalTrmm
( LEFT, LOWER, orientation, diag, T(1), L11_STAR_STAR, X1_STAR_VR );
X1 = X1_STAR_VR;
L21_MC_STAR = L21;
LocalGemm
( orientation, NORMAL,
T(1), X2, L21_MC_STAR, T(0), D1AdjOrTrans_MR_STAR );
D1AdjOrTrans_MR_MC.SumScatterFrom( D1AdjOrTrans_MR_STAR );
if( orientation == TRANSPOSE )
Transpose( D1AdjOrTrans_MR_MC.LocalMatrix(), D1.LocalMatrix() );
else
Adjoint( D1AdjOrTrans_MR_MC.LocalMatrix(), D1.LocalMatrix() );
Axpy( T(1), D1, X1 );
//--------------------------------------------------------------------//
D1.FreeAlignments();
D1AdjOrTrans_MR_MC.FreeAlignments();
D1AdjOrTrans_MR_STAR.FreeAlignments();
L21_MC_STAR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionDown
( XT, X0,
X1,
/**/ /**/
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrsmLUNMedium
( UnitOrNonUnit diag, F alpha, const DistMatrix<F>& U, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrsmLUNMedium");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<F>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
DistMatrix<F> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<F,MC, STAR> U01_MC_STAR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,MR, STAR> X1Trans_MR_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionUp
( X, XT,
XB, 0 );
while( XT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionUp
( XT, X0,
X1,
/**/ /**/
XB, X2 );
U01_MC_STAR.AlignWith( X0 );
X1Trans_MR_STAR.AlignWith( X0 );
//--------------------------------------------------------------------//
U11_STAR_STAR = U11; // U11[* ,* ] <- U11[MC,MR]
X1Trans_MR_STAR.TransposeFrom( X1 ); // X1[* ,MR] <- X1[MC,MR]
// X1[* ,MR] := U11^-1[* ,* ] X1[* ,MR]
//
// X1^T[MR,* ] := X1^T[MR,* ] U11^-T[* ,* ]
LocalTrsm
( RIGHT, UPPER, TRANSPOSE, diag,
F(1), U11_STAR_STAR, X1Trans_MR_STAR, checkIfSingular );
X1.TransposeFrom( X1Trans_MR_STAR );
U01_MC_STAR = U01; // U01[MC,* ] <- U01[MC,MR]
// X0[MC,MR] -= U01[MC,* ] X1[* ,MR]
LocalGemm
( NORMAL, TRANSPOSE, F(-1), U01_MC_STAR, X1Trans_MR_STAR, F(1), X0 );
//--------------------------------------------------------------------//
U01_MC_STAR.FreeAlignments();
X1Trans_MR_STAR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
}
