// Radau5 step function func radau5_step_mpi(o *Solver, y0 []float64, x0 float64, args ...interface{}) (rerr float64, err error) { // factors α := r5.α_ / o.h β := r5.β_ / o.h γ := r5.γ_ / o.h // Jacobian and decomposition if o.reuseJdec { o.reuseJdec = false } else { // calculate only first Jacobian for all iterations (simple/modified Newton's method) if o.reuseJ { o.reuseJ = false } else if !o.jacIsOK { // Jacobian triplet if o.jac == nil { // numerical //if x0 == 0.0 { io.Pfgrey(" > > > > > > > > . . . numerical Jacobian . . . < < < < < < < < <\n") } err = num.JacobianMpi(&o.dfdyT, func(fy, y []float64) (e error) { e = o.fcn(fy, o.h, x0, y, args...) return }, y0, o.f0, o.w[0], o.Distr) // w works here as workspace variable } else { // analytical //if x0 == 0.0 { io.Pfgrey(" > > > > > > > > . . . analytical Jacobian . . . < < < < < < < < <\n") } err = o.jac(&o.dfdyT, o.h, x0, y0, args...) } if err != nil { return } // create M matrix if o.doinit && !o.hasM { o.mTri = new(la.Triplet) if o.Distr { id, sz := mpi.Rank(), mpi.Size() start, endp1 := (id*o.ndim)/sz, ((id+1)*o.ndim)/sz o.mTri.Init(o.ndim, o.ndim, endp1-start) for i := start; i < endp1; i++ { o.mTri.Put(i, i, 1.0) } } else { o.mTri.Init(o.ndim, o.ndim, o.ndim) for i := 0; i < o.ndim; i++ { o.mTri.Put(i, i, 1.0) } } } o.njeval += 1 o.jacIsOK = true } // initialise triplets if o.doinit { o.rctriR = new(la.Triplet) o.rctriC = new(la.TripletC) o.rctriR.Init(o.ndim, o.ndim, o.mTri.Len()+o.dfdyT.Len()) xzmono := o.Distr o.rctriC.Init(o.ndim, o.ndim, o.mTri.Len()+o.dfdyT.Len(), xzmono) } // update triplets la.SpTriAdd(o.rctriR, γ, o.mTri, -1, &o.dfdyT) // rctriR := γ*M - dfdy la.SpTriAddR2C(o.rctriC, α, β, o.mTri, -1, &o.dfdyT) // rctriC := (α+βi)*M - dfdy // initialise solver if o.doinit { err = o.lsolR.InitR(o.rctriR, false, false, false) if err != nil { return } err = o.lsolC.InitC(o.rctriC, false, false, false) if err != nil { return } } // perform factorisation o.lsolR.Fact() o.lsolC.Fact() o.ndecomp += 1 } // updated u[i] o.u[0] = x0 + r5.c[0]*o.h o.u[1] = x0 + r5.c[1]*o.h o.u[2] = x0 + r5.c[2]*o.h // (trial/initial) updated z[i] and w[i] if o.first || o.ZeroTrial { for m := 0; m < o.ndim; m++ { o.z[0][m], o.w[0][m] = 0.0, 0.0 o.z[1][m], o.w[1][m] = 0.0, 0.0 o.z[2][m], o.w[2][m] = 0.0, 0.0 } } else { c3q := o.h / o.hprev c1q := r5.μ1 * c3q c2q := r5.μ2 * c3q for m := 0; m < o.ndim; m++ { o.z[0][m] = c1q * (o.ycol[0][m] + (c1q-r5.μ4)*(o.ycol[1][m]+(c1q-r5.μ3)*o.ycol[2][m])) o.z[1][m] = c2q * (o.ycol[0][m] + (c2q-r5.μ4)*(o.ycol[1][m]+(c2q-r5.μ3)*o.ycol[2][m])) o.z[2][m] = c3q * (o.ycol[0][m] + (c3q-r5.μ4)*(o.ycol[1][m]+(c3q-r5.μ3)*o.ycol[2][m])) o.w[0][m] = r5.Ti[0][0]*o.z[0][m] + r5.Ti[0][1]*o.z[1][m] + r5.Ti[0][2]*o.z[2][m] o.w[1][m] = r5.Ti[1][0]*o.z[0][m] + r5.Ti[1][1]*o.z[1][m] + r5.Ti[1][2]*o.z[2][m] o.w[2][m] = r5.Ti[2][0]*o.z[0][m] + r5.Ti[2][1]*o.z[1][m] + r5.Ti[2][2]*o.z[2][m] } } // iterations o.nit = 0 o.η = math.Pow(max(o.η, o.ϵ), 0.8) o.θ = o.θmax o.diverg = false var Lδw, oLδw, thq, othq, iterr, itRerr, qnewt float64 var it int for it = 0; it < o.NmaxIt; it++ { // max iterations ? o.nit = it + 1 if o.nit > o.nitmax { o.nitmax = o.nit } // evaluate f(x,y) at (u[i],v[i]=y0+z[i]) for i := 0; i < 3; i++ { for m := 0; m < o.ndim; m++ { o.v[i][m] = y0[m] + o.z[i][m] } o.nfeval += 1 err = o.fcn(o.f[i], o.h, o.u[i], o.v[i], args...) if err != nil { return } } // calc rhs if o.hasM { // using δw as workspace here la.SpMatVecMul(o.δw[0], 1, o.mMat, o.w[0]) // δw0 := M * w0 la.SpMatVecMul(o.δw[1], 1, o.mMat, o.w[1]) // δw1 := M * w1 la.SpMatVecMul(o.δw[2], 1, o.mMat, o.w[2]) // δw2 := M * w2 if o.Distr { mpi.AllReduceSum(o.δw[0], o.v[0]) // v is used as workspace here mpi.AllReduceSum(o.δw[1], o.v[1]) // v is used as workspace here mpi.AllReduceSum(o.δw[2], o.v[2]) // v is used as workspace here } for m := 0; m < o.ndim; m++ { o.v[0][m] = r5.Ti[0][0]*o.f[0][m] + r5.Ti[0][1]*o.f[1][m] + r5.Ti[0][2]*o.f[2][m] - γ*o.δw[0][m] o.v[1][m] = r5.Ti[1][0]*o.f[0][m] + r5.Ti[1][1]*o.f[1][m] + r5.Ti[1][2]*o.f[2][m] - α*o.δw[1][m] + β*o.δw[2][m] o.v[2][m] = r5.Ti[2][0]*o.f[0][m] + r5.Ti[2][1]*o.f[1][m] + r5.Ti[2][2]*o.f[2][m] - β*o.δw[1][m] - α*o.δw[2][m] } } else { for m := 0; m < o.ndim; m++ { o.v[0][m] = r5.Ti[0][0]*o.f[0][m] + r5.Ti[0][1]*o.f[1][m] + r5.Ti[0][2]*o.f[2][m] - γ*o.w[0][m] o.v[1][m] = r5.Ti[1][0]*o.f[0][m] + r5.Ti[1][1]*o.f[1][m] + r5.Ti[1][2]*o.f[2][m] - α*o.w[1][m] + β*o.w[2][m] o.v[2][m] = r5.Ti[2][0]*o.f[0][m] + r5.Ti[2][1]*o.f[1][m] + r5.Ti[2][2]*o.f[2][m] - β*o.w[1][m] - α*o.w[2][m] } } // solve linear system o.nlinsol += 1 var errR, errC error if !o.Distr && o.Pll { wg := new(sync.WaitGroup) wg.Add(2) go func() { errR = o.lsolR.SolveR(o.δw[0], o.v[0], false) wg.Done() }() go func() { errC = o.lsolC.SolveC(o.δw[1], o.δw[2], o.v[1], o.v[2], false) wg.Done() }() wg.Wait() } else { errR = o.lsolR.SolveR(o.δw[0], o.v[0], false) errC = o.lsolC.SolveC(o.δw[1], o.δw[2], o.v[1], o.v[2], false) } // check for errors from linear solution if errR != nil || errC != nil { var errmsg string if errR != nil { errmsg += errR.Error() } if errC != nil { if errR != nil { errmsg += "\n" } errmsg += errC.Error() } err = errors.New(errmsg) return } // update w and z for m := 0; m < o.ndim; m++ { o.w[0][m] += o.δw[0][m] o.w[1][m] += o.δw[1][m] o.w[2][m] += o.δw[2][m] o.z[0][m] = r5.T[0][0]*o.w[0][m] + r5.T[0][1]*o.w[1][m] + r5.T[0][2]*o.w[2][m] o.z[1][m] = r5.T[1][0]*o.w[0][m] + r5.T[1][1]*o.w[1][m] + r5.T[1][2]*o.w[2][m] o.z[2][m] = r5.T[2][0]*o.w[0][m] + r5.T[2][1]*o.w[1][m] + r5.T[2][2]*o.w[2][m] } // rms norm of δw Lδw = 0.0 for m := 0; m < o.ndim; m++ { Lδw += math.Pow(o.δw[0][m]/o.scal[m], 2.0) + math.Pow(o.δw[1][m]/o.scal[m], 2.0) + math.Pow(o.δw[2][m]/o.scal[m], 2.0) } Lδw = math.Sqrt(Lδw / float64(3*o.ndim)) // check convergence if it > 0 { thq = Lδw / oLδw if it == 1 { o.θ = thq } else { o.θ = math.Sqrt(thq * othq) } othq = thq if o.θ < 0.99 { o.η = o.θ / (1.0 - o.θ) iterr = Lδw * math.Pow(o.θ, float64(o.NmaxIt-o.nit)) / (1.0 - o.θ) itRerr = iterr / o.fnewt if itRerr >= 1.0 { // diverging qnewt = max(1.0e-4, min(20.0, itRerr)) o.dvfac = 0.8 * math.Pow(qnewt, -1.0/(4.0+float64(o.NmaxIt)-1.0-float64(o.nit))) o.diverg = true break } } else { // diverging badly (unexpected step-rejection) o.dvfac = 0.5 o.diverg = true break } } // save old norm oLδw = Lδw // converged if o.η*Lδw < o.fnewt { break } } // did not converge if it == o.NmaxIt-1 { chk.Panic("radau5_step failed with it=%d", it) } // diverging => stop if o.diverg { rerr = 2.0 // must leave state intact, any rerr is OK return } // error estimate if o.LerrStrat == 1 { // simple strategy => HW-VII p123 Eq.(8.17) (not good for stiff problems) for m := 0; m < o.ndim; m++ { o.ez[m] = r5.e0*o.z[0][m] + r5.e1*o.z[1][m] + r5.e2*o.z[2][m] o.lerr[m] = r5.γ0*o.h*o.f0[m] + o.ez[m] rerr += math.Pow(o.lerr[m]/o.scal[m], 2.0) } rerr = max(math.Sqrt(rerr/float64(o.ndim)), 1.0e-10) } else { // common if o.hasM { for m := 0; m < o.ndim; m++ { o.ez[m] = r5.e0*o.z[0][m] + r5.e1*o.z[1][m] + r5.e2*o.z[2][m] o.rhs[m] = o.f0[m] } if o.Distr { la.SpMatVecMul(o.δw[0], γ, o.mMat, o.ez) // δw[0] = γ * M * ez (δw[0] is workspace) mpi.AllReduceSumAdd(o.rhs, o.δw[0], o.δw[1]) // rhs += join_with_sum(δw[0]) (δw[1] is workspace) } else { la.SpMatVecMulAdd(o.rhs, γ, o.mMat, o.ez) // rhs += γ * M * ez } } else { for m := 0; m < o.ndim; m++ { o.ez[m] = r5.e0*o.z[0][m] + r5.e1*o.z[1][m] + r5.e2*o.z[2][m] o.rhs[m] = o.f0[m] + γ*o.ez[m] } } // HW-VII p123 Eq.(8.19) if o.LerrStrat == 2 { o.lsolR.SolveR(o.lerr, o.rhs, false) rerr = o.rms_norm(o.lerr) // HW-VII p123 Eq.(8.20) } else { o.lsolR.SolveR(o.lerr, o.rhs, false) rerr = o.rms_norm(o.lerr) if !(rerr < 1.0) { if o.first || o.reject { for m := 0; m < o.ndim; m++ { o.v[0][m] = y0[m] + o.lerr[m] // y0perr } o.nfeval += 1 err = o.fcn(o.f[0], o.h, x0, o.v[0], args...) // f0perr if err != nil { return } if o.hasM { la.VecCopy(o.rhs, 1, o.f[0]) // rhs := f0perr if o.Distr { la.SpMatVecMul(o.δw[0], γ, o.mMat, o.ez) // δw[0] = γ * M * ez (δw[0] is workspace) mpi.AllReduceSumAdd(o.rhs, o.δw[0], o.δw[1]) // rhs += join_with_sum(δw[0]) (δw[1] is workspace) } else { la.SpMatVecMulAdd(o.rhs, γ, o.mMat, o.ez) // rhs += γ * M * ez } } else { la.VecAdd2(o.rhs, 1, o.f[0], γ, o.ez) // rhs = f0perr + γ * ez } o.lsolR.SolveR(o.lerr, o.rhs, false) rerr = o.rms_norm(o.lerr) } } } } return }
// Solve solves linear programming problem func (o *LinIpm) Solve(verbose bool) (err error) { // starting point AAt := la.MatAlloc(o.Nl, o.Nl) // A*Aᵀ d := make([]float64, o.Nl) // inv(AAt) * b e := make([]float64, o.Nl) // A * c la.SpMatMatTrMul(AAt, 1, o.A) // AAt := A*Aᵀ la.SpMatVecMul(e, 1, o.A, o.C) // e := A * c la.SPDsolve2(d, o.L, AAt, o.B, e) // d := inv(AAt) * b and L := inv(AAt) * e la.SpMatTrVecMul(o.X, 1, o.A, d) // x := Aᵀ * d la.VecCopy(o.S, 1, o.C) // s := c la.SpMatTrVecMulAdd(o.S, -1, o.A, o.L) // s -= Aᵀλ xmin := o.X[0] smin := o.S[0] for i := 1; i < o.Nx; i++ { xmin = min(xmin, o.X[i]) smin = min(smin, o.S[i]) } δx := max(-1.5*xmin, 0) δs := max(-1.5*smin, 0) var xdots, xsum, ssum float64 for i := 0; i < o.Nx; i++ { o.X[i] += δx o.S[i] += δs xdots += o.X[i] * o.S[i] xsum += o.X[i] ssum += o.S[i] } δx = 0.5 * xdots / ssum δs = 0.5 * xdots / xsum for i := 0; i < o.Nx; i++ { o.X[i] += δx o.S[i] += δs } // constants for linear solver symmetric := false timing := false // auxiliary I := o.Nx + o.Nl // control variables var μ, σ float64 // μ and σ var xrmin float64 // min{ x_i / (-Δx_i) } (x_ratio_minimum) var srmin float64 // min{ s_i / (-Δs_i) } (s_ratio_minimum) var αpa float64 // α_prime_affine var αda float64 // α_dual_affine var μaff float64 // μ_affine var ctx, btl float64 // cᵀx and bᵀl // message if verbose { io.Pf("%3s%16s%16s\n", "it", "f(x)", "error") } // perform iterations it := 0 for it = 0; it < o.NmaxIt; it++ { // compute residual la.SpMatTrVecMul(o.Rx, 1, o.A, o.L) // rx := Aᵀλ la.SpMatVecMul(o.Rl, 1, o.A, o.X) // rλ := A x ctx, btl, μ = 0, 0, 0 for i := 0; i < o.Nx; i++ { o.Rx[i] += o.S[i] - o.C[i] o.Rs[i] = o.X[i] * o.S[i] ctx += o.C[i] * o.X[i] μ += o.X[i] * o.S[i] } for i := 0; i < o.Nl; i++ { o.Rl[i] -= o.B[i] btl += o.B[i] * o.L[i] } μ /= float64(o.Nx) // check convergence lerr := math.Abs(ctx-btl) / (1.0 + math.Abs(ctx)) if verbose { fx := la.VecDot(o.C, o.X) io.Pf("%3d%16.8e%16.8e\n", it, fx, lerr) } if lerr < o.Tol { break } // assemble Jacobian o.J.Start() o.J.PutCCMatAndMatT(o.A) for i := 0; i < o.Nx; i++ { o.J.Put(i, I+i, 1.0) o.J.Put(I+i, i, o.S[i]) o.J.Put(I+i, I+i, o.X[i]) } // solve linear system if it == 0 { err = o.Lis.InitR(o.J, symmetric, false, timing) if err != nil { return } } err = o.Lis.Fact() if err != nil { return } err = o.Lis.SolveR(o.Mdy, o.R, false) // mdy := inv(J) * R if err != nil { return } // control variables xrmin, srmin = o.calc_min_ratios() αpa = min(1, xrmin) αda = min(1, srmin) μaff = 0 for i := 0; i < o.Nx; i++ { μaff += (o.X[i] - αpa*o.Mdx[i]) * (o.S[i] - αda*o.Mds[i]) } μaff /= float64(o.Nx) σ = math.Pow(μaff/μ, 3) // update residual for i := 0; i < o.Nx; i++ { o.Rs[i] += o.Mdx[i]*o.Mds[i] - σ*μ } // solve linear system again err = o.Lis.SolveR(o.Mdy, o.R, false) // mdy := inv(J) * R if err != nil { return } // step lengths xrmin, srmin = o.calc_min_ratios() αpa = min(1, 0.99*xrmin) αda = min(1, 0.99*srmin) // update for i := 0; i < o.Nx; i++ { o.X[i] -= αpa * o.Mdx[i] o.S[i] -= αda * o.Mds[i] } for i := 0; i < o.Nl; i++ { o.L[i] -= αda * o.Mdl[i] } } // check convergence if it == o.NmaxIt { err = chk.Err("iterations did not converge") } return }