func (gp *gpConvexProg) F1(x *matrix.FloatMatrix) (f, Df *matrix.FloatMatrix, err error) {
f = nil
Df = nil
err = nil
f = matrix.FloatZeros(gp.mnl+1, 1)
Df = matrix.FloatZeros(gp.mnl+1, gp.n)
y := gp.g.Copy()
blas.GemvFloat(gp.F, x, y, 1.0, 1.0)
for i, s := range gp.ind {
start := s[0]
stop := s[1]
// yi := exp(yi) = exp(Fi*x+gi)
ymax := maxvec(y.FloatArray()[start:stop])
// ynew = exp(y[start:stop] - ymax)
ynew := matrix.Exp(matrix.FloatVector(y.FloatArray()[start:stop]).Add(-ymax))
y.SetIndexesFromArray(ynew.FloatArray(), matrix.Indexes(start, stop)...)
// fi = log sum yi = log sum exp(Fi*x+gi)
ysum := blas.AsumFloat(y, &la.IOpt{"n", stop - start}, &la.IOpt{"offset", start})
f.SetIndex(i, ymax+math.Log(ysum))
blas.ScalFloat(y, 1.0/ysum, &la.IOpt{"n", stop - start}, &la.IOpt{"offset", start})
blas.GemvFloat(gp.F, y, Df, 1.0, 0.0, la.OptTrans, &la.IOpt{"m", stop - start},
&la.IOpt{"incy", gp.mnl + 1}, &la.IOpt{"offseta", start},
&la.IOpt{"offsetx", start}, &la.IOpt{"offsety", i})
}
return
}
func _TestMultMV(t *testing.T) {
bM := 100 * M
bN := 100 * N
A := matrix.FloatNormal(bM, bN)
X := matrix.FloatNormal(bN, 1)
Y1 := matrix.FloatZeros(bM, 1)
Y0 := matrix.FloatZeros(bM, 1)
Ar := A.FloatArray()
Xr := X.FloatArray()
Y1r := Y1.FloatArray()
blas.GemvFloat(A, X, Y0, 1.0, 1.0)
DMultMV(Y1r, Ar, Xr, 1.0, 1.0, NOTRANS, 1, A.LeadingIndex(), 1, 0, bN, 0, bM, 32, 32)
t.Logf("Y0 == Y1: %v\n", Y0.AllClose(Y1))
/*
if ! Y0.AllClose(Y1) {
y0 := Y0.SubMatrix(0, 0, 2, 1)
y1 := Y1.SubMatrix(0, 0, 2, 1)
t.Logf("y0=\n%v\n", y0)
t.Logf("y1=\n%v\n", y1)
}
*/
}
func _TestMultMVTransA(t *testing.T) {
bM := 1000 * M
bN := 1000 * N
A := matrix.FloatNormal(bN, bM)
X := matrix.FloatWithValue(bN, 1, 1.0)
Y1 := matrix.FloatZeros(bM, 1)
Y0 := matrix.FloatZeros(bM, 1)
Ar := A.FloatArray()
Xr := X.FloatArray()
Y1r := Y1.FloatArray()
blas.GemvFloat(A, X, Y0, 1.0, 1.0, linalg.OptTrans)
DMultMV(Y1r, Ar, Xr, 1.0, 1.0, TRANSA, 1, A.LeadingIndex(), 1, 0, bN, 0, bM, 4, 4)
ok := Y0.AllClose(Y1)
t.Logf("Y0 == Y1: %v\n", ok)
if !ok {
var y1, y0 matrix.FloatMatrix
Y1.SubMatrix(&y1, 0, 0, 5, 1)
t.Logf("Y1[0:5]:\n%v\n", y1)
Y0.SubMatrix(&y0, 0, 0, 5, 1)
t.Logf("Y0[0:5]:\n%v\n", y0)
}
}
/*
Matrix-vector multiplication.
A is a matrix or spmatrix of size (m, n) where
N = dims['l'] + sum(dims['q']) + sum( k**2 for k in dims['s'] )
representing a mapping from R^n to S.
If trans is 'N':
y := alpha*A*x + beta * y (trans = 'N').
x is a vector of length n. y is a vector of length N.
If trans is 'T':
y := alpha*A'*x + beta * y (trans = 'T').
x is a vector of length N. y is a vector of length n.
The 's' components in S are stored in unpacked 'L' storage.
*/
func sgemv(A, x, y *matrix.FloatMatrix, alpha, beta float64, dims *sets.DimensionSet, opts ...la_.Option) error {
m := dims.Sum("l", "q") + dims.SumSquared("s")
n := la_.GetIntOpt("n", -1, opts...)
if n == -1 {
n = A.Cols()
}
trans := la_.GetIntOpt("trans", int(la_.PNoTrans), opts...)
offsetX := la_.GetIntOpt("offsetx", 0, opts...)
offsetY := la_.GetIntOpt("offsety", 0, opts...)
offsetA := la_.GetIntOpt("offseta", 0, opts...)
if trans == int(la_.PTrans) && alpha != 0.0 {
trisc(x, dims, offsetX)
//fmt.Printf("trisc x=\n%v\n", x.ConvertToString())
}
//fmt.Printf("alpha=%.4f beta=%.4f m=%d n=%d\n", alpha, beta, m, n)
//fmt.Printf("A=\n%v\nx=\n%v\ny=\n%v\n", A, x.ConvertToString(), y.ConvertToString())
err := blas.GemvFloat(A, x, y, alpha, beta, &la_.IOpt{"trans", trans},
&la_.IOpt{"n", n}, &la_.IOpt{"m", m}, &la_.IOpt{"offseta", offsetA},
&la_.IOpt{"offsetx", offsetX}, &la_.IOpt{"offsety", offsetY})
//fmt.Printf("gemv y=\n%v\n", y.ConvertToString())
if trans == int(la_.PTrans) && alpha != 0.0 {
triusc(x, dims, offsetX)
}
return err
}
func CTestGemv(m, n, p int) (fnc func(), A, X, Y *matrix.FloatMatrix) {
A = matrix.FloatNormal(m, n)
X = matrix.FloatNormal(n, 1)
Y = matrix.FloatZeros(m, 1)
fnc = func() {
blas.GemvFloat(A, X, Y, 1.0, 1.0)
}
return
}
func (gp *gpConvexProg) F2(x, z *matrix.FloatMatrix) (f, Df, H *matrix.FloatMatrix, err error) {
err = nil
f = matrix.FloatZeros(gp.mnl+1, 1)
Df = matrix.FloatZeros(gp.mnl+1, gp.n)
H = matrix.FloatZeros(gp.n, gp.n)
y := gp.g.Copy()
Fsc := matrix.FloatZeros(gp.maxK, gp.n)
blas.GemvFloat(gp.F, x, y, 1.0, 1.0)
//fmt.Printf("y=\n%v\n", y.ToString("%.3f"))
for i, s := range gp.ind {
start := s[0]
stop := s[1]
// yi := exp(yi) = exp(Fi*x+gi)
ymax := maxvec(y.FloatArray()[start:stop])
ynew := matrix.Exp(matrix.FloatVector(y.FloatArray()[start:stop]).Add(-ymax))
y.SetIndexesFromArray(ynew.FloatArray(), matrix.Indexes(start, stop)...)
// fi = log sum yi = log sum exp(Fi*x+gi)
ysum := blas.AsumFloat(y, &la.IOpt{"n", stop - start}, &la.IOpt{"offset", start})
f.SetIndex(i, ymax+math.Log(ysum))
blas.ScalFloat(y, 1.0/ysum, &la.IOpt{"n", stop - start}, &la.IOpt{"offset", start})
blas.GemvFloat(gp.F, y, Df, 1.0, 0.0, la.OptTrans, &la.IOpt{"m", stop - start},
&la.IOpt{"incy", gp.mnl + 1}, &la.IOpt{"offseta", start},
&la.IOpt{"offsetx", start}, &la.IOpt{"offsety", i})
Fsc.SetSubMatrix(0, 0, gp.F.GetSubMatrix(start, 0, stop-start))
for k := start; k < stop; k++ {
blas.AxpyFloat(Df, Fsc, -1.0, &la.IOpt{"n", gp.n},
&la.IOpt{"incx", gp.mnl + 1}, &la.IOpt{"incy", Fsc.Rows()},
&la.IOpt{"offsetx", i}, &la.IOpt{"offsety", k - start})
blas.ScalFloat(Fsc, math.Sqrt(y.GetIndex(k)),
&la.IOpt{"inc", Fsc.Rows()}, &la.IOpt{"offset", k - start})
}
// H += z[i]*Hi = z[i] *Fisc' * Fisc
blas.SyrkFloat(Fsc, H, z.GetIndex(i), 1.0, la.OptTrans,
&la.IOpt{"k", stop - start})
}
return
}
func CTestGemvTransA(m, n, p int) (fnc func(), A, X, Y *matrix.FloatMatrix) {
A = matrix.FloatNormal(n, m)
X = matrix.FloatNormal(n, 1)
Y = matrix.FloatZeros(m, 1)
A = A.Transpose()
fnc = func() {
blas.GemvFloat(A, X, Y, 1.0, 1.0, linalg.OptTrans)
}
return
}
func (a *epMatrixA) Af(u, v MatrixVariable, alpha, beta float64, trans la.Option) (err error) {
err = nil
if trans.Equal(la.OptNoTrans) {
ue, u_ok := u.(*epigraph)
ve := v.Matrix()
if !u_ok {
return errors.New("'u' not a epigraph")
}
err = blas.GemvFloat(a.A, ue.m(), ve, alpha, beta)
} else {
ve, v_ok := v.(*epigraph)
ue := u.Matrix()
if !v_ok {
return errors.New("'v' not a epigraph")
}
err = blas.GemvFloat(a.A, ue, ve.m(), alpha, beta, trans)
ve.set(ve.t() * beta)
}
return
}
func (d *epigraphDf) Df(u, v MatrixVariable, alpha, beta float64, trans la.Option) (err error) {
err = nil
if trans.Equal(la.OptNoTrans) {
u_e, u_ok := u.(*epigraph)
v_e := v.Matrix()
if !u_ok {
fmt.Printf("Df: not a epigraph\n")
return errors.New("'u' not a epigraph")
}
err = blas.GemvFloat(d.df, u_e.m(), v_e, alpha, beta, la.OptNoTrans)
v_e.Add(-alpha*u_e.t(), 0)
} else {
v_e, v_ok := v.(*epigraph)
u_e := u.Matrix()
if !v_ok {
fmt.Printf("Df: not a epigraph\n")
return errors.New("'v' not a epigraph")
}
err = blas.GemvFloat(d.df, u_e, v_e.m(), alpha, beta, la.OptTrans)
v_e.set(-alpha*u_e.GetIndex(0) + beta*v_e.t())
}
return
}
func _TestMultMVSmall(t *testing.T) {
bM := 5
bN := 4
A := matrix.FloatNormal(bM, bN)
X := matrix.FloatVector([]float64{1.0, 2.0, 3.0, 4.0})
Y1 := matrix.FloatZeros(bM, 1)
Y0 := matrix.FloatZeros(bM, 1)
Ar := A.FloatArray()
Xr := X.FloatArray()
Y1r := Y1.FloatArray()
blas.GemvFloat(A, X, Y0, 1.0, 1.0)
DMultMV(Y1r, Ar, Xr, 1.0, 1.0, NOTRANS, 1, A.LeadingIndex(), 1, 0, bN, 0, bM, 4, 4)
ok := Y0.AllClose(Y1)
t.Logf("Y0 == Y1: %v\n", ok)
if !ok {
t.Logf("blas: Y=A*X\n%v\n", Y0)
t.Logf("Y1: Y1 = A*X\n%v\n", Y1)
}
}
func (a *matrixVarA) Af(u, v MatrixVariable, alpha, beta float64, trans la.Option) (err error) {
err = blas.GemvFloat(a.mA, u.Matrix(), v.Matrix(), alpha, beta, trans)
return
}
func kktChol2(G *matrix.FloatMatrix, dims *sets.DimensionSet, A *matrix.FloatMatrix, mnl int) (kktFactor, error) {
if len(dims.At("q")) > 0 || len(dims.At("s")) > 0 {
return nil, errors.New("'chol2' solver only for problems with no second-order or " +
"semidefinite cone constraints")
}
p, n := A.Size()
ml := dims.At("l")[0]
F := &chol2Data{firstcall: true, singular: false, A: A, G: G, dims: dims}
factor := func(W *sets.FloatMatrixSet, H, Df *matrix.FloatMatrix) (KKTFunc, error) {
var err error = nil
minor := 0
if !checkpnt.MinorEmpty() {
minor = checkpnt.MinorTop()
}
if F.firstcall {
F.Gs = matrix.FloatZeros(F.G.Size())
if mnl > 0 {
F.Dfs = matrix.FloatZeros(Df.Size())
}
F.S = matrix.FloatZeros(n, n)
F.K = matrix.FloatZeros(p, p)
checkpnt.AddMatrixVar("Gs", F.Gs)
checkpnt.AddMatrixVar("Dfs", F.Dfs)
checkpnt.AddMatrixVar("S", F.S)
checkpnt.AddMatrixVar("K", F.K)
}
if mnl > 0 {
dnli := matrix.FloatZeros(mnl, mnl)
dnli.SetIndexesFromArray(W.At("dnli")[0].FloatArray(), matrix.DiagonalIndexes(dnli)...)
blas.GemmFloat(dnli, Df, F.Dfs, 1.0, 0.0)
}
checkpnt.Check("02factor_chol2", minor)
di := matrix.FloatZeros(ml, ml)
di.SetIndexesFromArray(W.At("di")[0].FloatArray(), matrix.DiagonalIndexes(di)...)
err = blas.GemmFloat(di, G, F.Gs, 1.0, 0.0)
checkpnt.Check("06factor_chol2", minor)
if F.firstcall {
blas.SyrkFloat(F.Gs, F.S, 1.0, 0.0, la.OptTrans)
if mnl > 0 {
blas.SyrkFloat(F.Dfs, F.S, 1.0, 1.0, la.OptTrans)
}
if H != nil {
F.S.Plus(H)
}
checkpnt.Check("10factor_chol2", minor)
err = lapack.Potrf(F.S)
if err != nil {
err = nil // reset error
F.singular = true
// original code recreates F.S as dense if it is sparse and
// A is dense, we don't do it as currently no sparse matrices
//F.S = matrix.FloatZeros(n, n)
//checkpnt.AddMatrixVar("S", F.S)
blas.SyrkFloat(F.Gs, F.S, 1.0, 0.0, la.OptTrans)
if mnl > 0 {
blas.SyrkFloat(F.Dfs, F.S, 1.0, 1.0, la.OptTrans)
}
checkpnt.Check("14factor_chol2", minor)
blas.SyrkFloat(F.A, F.S, 1.0, 1.0, la.OptTrans)
if H != nil {
F.S.Plus(H)
}
lapack.Potrf(F.S)
}
F.firstcall = false
checkpnt.Check("20factor_chol2", minor)
} else {
blas.SyrkFloat(F.Gs, F.S, 1.0, 0.0, la.OptTrans)
if mnl > 0 {
blas.SyrkFloat(F.Dfs, F.S, 1.0, 1.0, la.OptTrans)
}
if H != nil {
F.S.Plus(H)
}
checkpnt.Check("40factor_chol2", minor)
if F.singular {
blas.SyrkFloat(F.A, F.S, 1.0, 1.0, la.OptTrans)
}
lapack.Potrf(F.S)
checkpnt.Check("50factor_chol2", minor)
}
// Asct := L^{-1}*A'. Factor K = Asct'*Asct.
Asct := F.A.Transpose()
blas.TrsmFloat(F.S, Asct, 1.0)
blas.SyrkFloat(Asct, F.K, 1.0, 0.0, la.OptTrans)
lapack.Potrf(F.K)
checkpnt.Check("90factor_chol2", minor)
solve := func(x, y, z *matrix.FloatMatrix) (err error) {
// Solve
//
// [ H A' GG'*W^{-1} ] [ ux ] [ bx ]
// [ A 0 0 ] * [ uy ] = [ by ]
// [ W^{-T}*GG 0 -I ] [ W*uz ] [ W^{-T}*bz ]
//
// and return ux, uy, W*uz.
//
// If not F['singular']:
//
// K*uy = A * S^{-1} * ( bx + GG'*W^{-1}*W^{-T}*bz ) - by
// S*ux = bx + GG'*W^{-1}*W^{-T}*bz - A'*uy
// W*uz = W^{-T} * ( GG*ux - bz ).
//
// If F['singular']:
//
// K*uy = A * S^{-1} * ( bx + GG'*W^{-1}*W^{-T}*bz + A'*by )
// - by
// S*ux = bx + GG'*W^{-1}*W^{-T}*bz + A'*by - A'*y.
// W*uz = W^{-T} * ( GG*ux - bz ).
minor := 0
if !checkpnt.MinorEmpty() {
minor = checkpnt.MinorTop()
}
// z := W^{-1} * z = W^{-1} * bz
scale(z, W, true, true)
checkpnt.Check("10solve_chol2", minor)
// If not F['singular']:
// x := L^{-1} * P * (x + GGs'*z)
// = L^{-1} * P * (x + GG'*W^{-1}*W^{-T}*bz)
//
// If F['singular']:
// x := L^{-1} * P * (x + GGs'*z + A'*y))
// = L^{-1} * P * (x + GG'*W^{-1}*W^{-T}*bz + A'*y)
if mnl > 0 {
blas.GemvFloat(F.Dfs, z, x, 1.0, 1.0, la.OptTrans)
}
blas.GemvFloat(F.Gs, z, x, 1.0, 1.0, la.OptTrans, &la.IOpt{"offsetx", mnl})
//checkpnt.Check("20solve_chol2", minor)
if F.singular {
blas.GemvFloat(F.A, y, x, 1.0, 1.0, la.OptTrans)
}
checkpnt.Check("30solve_chol2", minor)
blas.TrsvFloat(F.S, x)
//checkpnt.Check("50solve_chol2", minor)
// y := K^{-1} * (Asc*x - y)
// = K^{-1} * (A * S^{-1} * (bx + GG'*W^{-1}*W^{-T}*bz) - by)
// (if not F['singular'])
// = K^{-1} * (A * S^{-1} * (bx + GG'*W^{-1}*W^{-T}*bz +
// A'*by) - by)
// (if F['singular']).
blas.GemvFloat(Asct, x, y, 1.0, -1.0, la.OptTrans)
//checkpnt.Check("55solve_chol2", minor)
lapack.Potrs(F.K, y)
//checkpnt.Check("60solve_chol2", minor)
// x := P' * L^{-T} * (x - Asc'*y)
// = S^{-1} * (bx + GG'*W^{-1}*W^{-T}*bz - A'*y)
// (if not F['singular'])
// = S^{-1} * (bx + GG'*W^{-1}*W^{-T}*bz + A'*by - A'*y)
// (if F['singular'])
blas.GemvFloat(Asct, y, x, -1.0, 1.0)
blas.TrsvFloat(F.S, x, la.OptTrans)
checkpnt.Check("70solve_chol2", minor)
// W*z := GGs*x - z = W^{-T} * (GG*x - bz)
if mnl > 0 {
blas.GemvFloat(F.Dfs, x, z, 1.0, -1.0)
}
blas.GemvFloat(F.Gs, x, z, 1.0, -1.0, &la.IOpt{"offsety", mnl})
checkpnt.Check("90solve_chol2", minor)
return nil
}
return solve, err
}
return factor, nil
}
func TestMultMVTransASmall(t *testing.T) {
data6 := [][]float64{
[]float64{-1.59e+00, 6.56e-02, 2.14e-01, 6.79e-01, 2.93e-01, 5.24e-01},
[]float64{4.28e-01, 1.57e-01, 3.81e-01, 2.19e-01, 2.97e-01, 2.83e-02},
[]float64{3.02e-01, 9.70e-02, 3.18e-01, 2.03e-01, 7.53e-01, 1.58e-01},
[]float64{1.99e-01, 3.01e-01, 4.69e-01, 3.61e-01, 2.07e-01, 6.07e-01},
[]float64{1.93e-01, 5.15e-01, 2.83e-01, 5.71e-01, 8.65e-01, 9.75e-01},
[]float64{3.13e-01, 8.14e-01, 2.93e-01, 8.62e-01, 6.97e-01, 7.95e-02}}
data5 := [][]float64{
[]float64{1.57e-01, 3.81e-01, 2.19e-01, 2.97e-01, 2.83e-02},
[]float64{9.70e-02, 3.18e-01, 2.03e-01, 7.53e-01, 1.58e-01},
[]float64{3.01e-01, 4.69e-01, 3.61e-01, 2.07e-01, 6.07e-01},
[]float64{5.15e-01, 2.83e-01, 5.71e-01, 8.65e-01, 9.75e-01},
[]float64{8.14e-01, 2.93e-01, 8.62e-01, 6.97e-01, 7.95e-02}}
data2 := []float64{4.28e-01, 3.02e-01, 1.99e-01, 1.93e-01, 3.13e-01}
bM := 5
bN := 4
nb := 2
//A := matrix.FloatNormal(bN, bM)
//X := matrix.FloatWithValue(bN, 1, 1.0)
A := matrix.FloatMatrixFromTable(data5, matrix.RowOrder)
X := matrix.FloatNew(5, 1, data2)
bM = A.Rows()
bN = A.Cols()
Ym := matrix.FloatZeros(3, bM)
Y1 := matrix.FloatZeros(bM, 1)
Y0 := matrix.FloatZeros(bM, 1)
Ar := A.FloatArray()
Xr := X.FloatArray()
Y1r := Y1.FloatArray()
blas.GemvFloat(A, X, Y0, 1.0, 1.0, linalg.OptTrans)
DMultMV(Y1r, Ar, Xr, 1.0, 1.0, TRANSA, 1, A.LeadingIndex(), 1, 0, bN, 0, bM, nb, nb)
ok := Y0.AllClose(Y1)
t.Logf("Y0 == Y1: %v\n", ok)
if ok || !ok {
t.Logf("blas: Y=A.T*X\n%v\n", Y0)
t.Logf("Y1: Y1 = A*X\n%v\n", Y1)
}
// zero Y0, Y1
Y0.Scale(0.0)
Y1.Scale(0.0)
// test with matrix view; A is view
var A0 matrix.FloatMatrix
A6 := matrix.FloatMatrixFromTable(data6, matrix.RowOrder)
A0.SubMatrixOf(A6, 1, 1)
blas.GemvFloat(&A0, X, Y0, 1.0, 1.0, linalg.OptTrans)
Ar = A0.FloatArray()
DMultMV(Y1r, Ar, Xr, 1.0, 1.0, TRANSA, 1, A0.LeadingIndex(), 1, 0, bN, 0, bM, nb, nb)
ok = Y0.AllClose(Y1)
t.Logf("lda>rows: Y0 == Y1: %v\n", ok)
if ok || !ok {
t.Logf("blas: Y=A.T*X\n%v\n", Y0)
t.Logf("Y1: Y1 = A*X\n%v\n", Y1)
}
// Y is view too.
Y1.SubMatrixOf(Ym, 0, 0, 1, bM)
Y1r = Y1.FloatArray()
DMultMV(Y1r, Ar, Xr, 1.0, 1.0, TRANSA, Y1.LeadingIndex(), A0.LeadingIndex(), 1, 0, bN, 0, bM, nb, nb)
ok = Y0.AllClose(Y1.Transpose())
t.Logf("Y0 == Y1 row: %v\n", ok)
t.Logf("row Y1: %v\n", Y1)
}
func (d *matrixVarDf) Df(u, v MatrixVariable, alpha, beta float64, trans la.Option) error {
return blas.GemvFloat(d.df, u.Matrix(), v.Matrix(), alpha, beta, trans)
}
// Solution of KKT equations by reduction to a 2 x 2 system, a QR
// factorization to eliminate the equality constraints, and a dense
// Cholesky factorization of order n-p.
//
// Computes the QR factorization
//
// A' = [Q1, Q2] * [R; 0]
//
// and returns a function that (1) computes the Cholesky factorization
//
// Q_2^T * (H + GG^T * W^{-1} * W^{-T} * GG) * Q2 = L * L^T,
//
// given H, Df, W, where GG = [Df; G], and (2) returns a function for
// solving
//
// [ H A' GG' ] [ ux ] [ bx ]
// [ A 0 0 ] * [ uy ] = [ by ].
// [ GG 0 -W'*W ] [ uz ] [ bz ]
//
// H is n x n, A is p x n, Df is mnl x n, G is N x n where
// N = dims['l'] + sum(dims['q']) + sum( k**2 for k in dims['s'] ).
//
func kktChol(G *matrix.FloatMatrix, dims *sets.DimensionSet, A *matrix.FloatMatrix, mnl int) (kktFactor, error) {
p, n := A.Size()
cdim := mnl + dims.Sum("l", "q") + dims.SumSquared("s")
cdim_pckd := mnl + dims.Sum("l", "q") + dims.SumPacked("s")
QA := A.Transpose()
tauA := matrix.FloatZeros(p, 1)
lapack.Geqrf(QA, tauA)
Gs := matrix.FloatZeros(cdim, n)
K := matrix.FloatZeros(n, n)
bzp := matrix.FloatZeros(cdim_pckd, 1)
yy := matrix.FloatZeros(p, 1)
checkpnt.AddMatrixVar("tauA", tauA)
checkpnt.AddMatrixVar("Gs", Gs)
checkpnt.AddMatrixVar("K", K)
factor := func(W *sets.FloatMatrixSet, H, Df *matrix.FloatMatrix) (KKTFunc, error) {
// Compute
//
// K = [Q1, Q2]' * (H + GG' * W^{-1} * W^{-T} * GG) * [Q1, Q2]
//
// and take the Cholesky factorization of the 2,2 block
//
// Q_2' * (H + GG^T * W^{-1} * W^{-T} * GG) * Q2.
var err error = nil
minor := 0
if !checkpnt.MinorEmpty() {
minor = checkpnt.MinorTop()
}
// Gs = W^{-T} * GG in packed storage.
if mnl > 0 {
Gs.SetSubMatrix(0, 0, Df)
}
Gs.SetSubMatrix(mnl, 0, G)
checkpnt.Check("00factor_chol", minor)
scale(Gs, W, true, true)
pack2(Gs, dims, mnl)
//checkpnt.Check("10factor_chol", minor)
// K = [Q1, Q2]' * (H + Gs' * Gs) * [Q1, Q2].
blas.SyrkFloat(Gs, K, 1.0, 0.0, la.OptTrans, &la.IOpt{"k", cdim_pckd})
if H != nil {
K.SetSubMatrix(0, 0, matrix.Plus(H, K.GetSubMatrix(0, 0, H.Rows(), H.Cols())))
}
//checkpnt.Check("20factor_chol", minor)
symm(K, n, 0)
lapack.Ormqr(QA, tauA, K, la.OptLeft, la.OptTrans)
lapack.Ormqr(QA, tauA, K, la.OptRight)
//checkpnt.Check("30factor_chol", minor)
// Cholesky factorization of 2,2 block of K.
lapack.Potrf(K, &la.IOpt{"n", n - p}, &la.IOpt{"offseta", p * (n + 1)})
checkpnt.Check("40factor_chol", minor)
solve := func(x, y, z *matrix.FloatMatrix) (err error) {
// Solve
//
// [ 0 A' GG'*W^{-1} ] [ ux ] [ bx ]
// [ A 0 0 ] * [ uy ] = [ by ]
// [ W^{-T}*GG 0 -I ] [ W*uz ] [ W^{-T}*bz ]
//
// and return ux, uy, W*uz.
//
// On entry, x, y, z contain bx, by, bz. On exit, they contain
// the solution ux, uy, W*uz.
//
// If we change variables ux = Q1*v + Q2*w, the system becomes
//
// [ K11 K12 R ] [ v ] [Q1'*(bx+GG'*W^{-1}*W^{-T}*bz)]
// [ K21 K22 0 ] * [ w ] = [Q2'*(bx+GG'*W^{-1}*W^{-T}*bz)]
// [ R^T 0 0 ] [ uy ] [by ]
//
// W*uz = W^{-T} * ( GG*ux - bz ).
minor := 0
if !checkpnt.MinorEmpty() {
minor = checkpnt.MinorTop()
}
// bzp := W^{-T} * bz in packed storage
scale(z, W, true, true)
pack(z, bzp, dims, &la.IOpt{"mnl", mnl})
// x := [Q1, Q2]' * (x + Gs' * bzp)
// = [Q1, Q2]' * (bx + Gs' * W^{-T} * bz)
blas.GemvFloat(Gs, bzp, x, 1.0, 1.0, la.OptTrans, &la.IOpt{"m", cdim_pckd})
lapack.Ormqr(QA, tauA, x, la.OptLeft, la.OptTrans)
// y := x[:p]
// = Q1' * (bx + Gs' * W^{-T} * bz)
blas.Copy(y, yy)
blas.Copy(x, y, &la.IOpt{"n", p})
// x[:p] := v = R^{-T} * by
blas.Copy(yy, x)
lapack.Trtrs(QA, x, la.OptUpper, la.OptTrans, &la.IOpt{"n", p})
// x[p:] := K22^{-1} * (x[p:] - K21*x[:p])
// = K22^{-1} * (Q2' * (bx + Gs' * W^{-T} * bz) - K21*v)
blas.GemvFloat(K, x, x, -1.0, 1.0, &la.IOpt{"m", n - p}, &la.IOpt{"n", p},
&la.IOpt{"offseta", p}, &la.IOpt{"offsety", p})
lapack.Potrs(K, x, &la.IOpt{"n", n - p}, &la.IOpt{"offseta", p * (n + 1)},
&la.IOpt{"offsetb", p})
// y := y - [K11, K12] * x
// = Q1' * (bx + Gs' * W^{-T} * bz) - K11*v - K12*w
blas.GemvFloat(K, x, y, -1.0, 1.0, &la.IOpt{"m", p}, &la.IOpt{"n", n})
// y := R^{-1}*y
// = R^{-1} * (Q1' * (bx + Gs' * W^{-T} * bz) - K11*v
// - K12*w)
lapack.Trtrs(QA, y, la.OptUpper, &la.IOpt{"n", p})
// x := [Q1, Q2] * x
lapack.Ormqr(QA, tauA, x, la.OptLeft)
// bzp := Gs * x - bzp.
// = W^{-T} * ( GG*ux - bz ) in packed storage.
// Unpack and copy to z.
blas.GemvFloat(Gs, x, bzp, 1.0, -1.0, &la.IOpt{"m", cdim_pckd})
unpack(bzp, z, dims, &la.IOpt{"mnl", mnl})
checkpnt.Check("90solve_chol", minor)
return nil
}
return solve, err
}
return factor, nil
}
func CheckNoTrans(A, X, Y *matrix.FloatMatrix) {
blas.GemvFloat(A, X, Y, 1.0, 1.0)
}
/*
Applies Nesterov-Todd scaling or its inverse.
Computes
x := W*x (trans is false 'N', inverse = false 'N')
x := W^T*x (trans is true 'T', inverse = false 'N')
x := W^{-1}*x (trans is false 'N', inverse = true 'T')
x := W^{-T}*x (trans is true 'T', inverse = true 'T').
x is a dense float matrix.
W is a MatrixSet with entries:
- W['dnl']: positive vector
- W['dnli']: componentwise inverse of W['dnl']
- W['d']: positive vector
- W['di']: componentwise inverse of W['d']
- W['v']: lists of 2nd order cone vectors with unit hyperbolic norms
- W['beta']: list of positive numbers
- W['r']: list of square matrices
- W['rti']: list of square matrices. rti[k] is the inverse transpose
of r[k].
The 'dnl' and 'dnli' entries are optional, and only present when the
function is called from the nonlinear solver.
*/
func scale(x *matrix.FloatMatrix, W *sets.FloatMatrixSet, trans, inverse bool) (err error) {
/*DEBUGGED*/
var wl []*matrix.FloatMatrix
var w *matrix.FloatMatrix
ind := 0
err = nil
// var minor int = 0
//if ! checkpnt.MinorEmpty() {
// minor = checkpnt.MinorTop()
//}
//fmt.Printf("\n%d.%04d scaling x=\n%v\n", checkpnt.Major(), minor, x.ToString("%.17f"))
// Scaling for nonlinear component xk is xk := dnl .* xk; inverse
// scaling is xk ./ dnl = dnli .* xk, where dnl = W['dnl'],
// dnli = W['dnli'].
if wl = W.At("dnl"); wl != nil {
if inverse {
w = W.At("dnli")[0]
} else {
w = W.At("dnl")[0]
}
for k := 0; k < x.Cols(); k++ {
err = blas.TbmvFloat(w, x, &la_.IOpt{"n", w.Rows()}, &la_.IOpt{"k", 0},
&la_.IOpt{"lda", 1}, &la_.IOpt{"offsetx", k * x.Rows()})
if err != nil {
//fmt.Printf("1. TbmvFloat: %v\n", err)
return
}
}
ind += w.Rows()
}
//if ! checkpnt.MinorEmpty() {
// checkpnt.Check("000scale", minor)
//}
// Scaling for linear 'l' component xk is xk := d .* xk; inverse
// scaling is xk ./ d = di .* xk, where d = W['d'], di = W['di'].
if inverse {
w = W.At("di")[0]
} else {
w = W.At("d")[0]
}
for k := 0; k < x.Cols(); k++ {
err = blas.TbmvFloat(w, x, &la_.IOpt{"n", w.Rows()}, &la_.IOpt{"k", 0},
&la_.IOpt{"lda", 1}, &la_.IOpt{"offsetx", k*x.Rows() + ind})
if err != nil {
//fmt.Printf("2. TbmvFloat: %v\n", err)
return
}
}
ind += w.Rows()
//if ! checkpnt.MinorEmpty() {
// checkpnt.Check("010scale", minor)
//}
// Scaling for 'q' component is
//
// xk := beta * (2*v*v' - J) * xk
// = beta * (2*v*(xk'*v)' - J*xk)
//
// where beta = W['beta'][k], v = W['v'][k], J = [1, 0; 0, -I].
//
//Inverse scaling is
//
// xk := 1/beta * (2*J*v*v'*J - J) * xk
// = 1/beta * (-J) * (2*v*((-J*xk)'*v)' + xk).
//wf := matrix.FloatZeros(x.Cols(), 1)
w = matrix.FloatZeros(x.Cols(), 1)
for k, v := range W.At("v") {
m := v.Rows()
if inverse {
blas.ScalFloat(x, -1.0, &la_.IOpt{"offset", ind}, &la_.IOpt{"inc", x.Rows()})
}
err = blas.GemvFloat(x, v, w, 1.0, 0.0, la_.OptTrans, &la_.IOpt{"m", m},
&la_.IOpt{"n", x.Cols()}, &la_.IOpt{"offsetA", ind},
&la_.IOpt{"lda", x.Rows()})
if err != nil {
//fmt.Printf("3. GemvFloat: %v\n", err)
return
}
err = blas.ScalFloat(x, -1.0, &la_.IOpt{"offset", ind}, &la_.IOpt{"inc", x.Rows()})
if err != nil {
return
}
err = blas.GerFloat(v, w, x, 2.0, &la_.IOpt{"m", m},
&la_.IOpt{"n", x.Cols()}, &la_.IOpt{"lda", x.Rows()},
&la_.IOpt{"offsetA", ind})
if err != nil {
//fmt.Printf("4. GerFloat: %v\n", err)
return
}
var a float64
if inverse {
blas.ScalFloat(x, -1.0,
&la_.IOpt{"offset", ind}, &la_.IOpt{"inc", x.Rows()})
// a[i,j] := 1.0/W[i,j]
a = 1.0 / W.At("beta")[0].GetIndex(k)
} else {
a = W.At("beta")[0].GetIndex(k)
}
for i := 0; i < x.Cols(); i++ {
blas.ScalFloat(x, a, &la_.IOpt{"n", m}, &la_.IOpt{"offset", ind + i*x.Rows()})
}
ind += m
}
//if ! checkpnt.MinorEmpty() {
// checkpnt.Check("020scale", minor)
//}
// Scaling for 's' component xk is
//
// xk := vec( r' * mat(xk) * r ) if trans = 'N'
// xk := vec( r * mat(xk) * r' ) if trans = 'T'.
//
// r is kth element of W['r'].
//
// Inverse scaling is
//
// xk := vec( rti * mat(xk) * rti' ) if trans = 'N'
// xk := vec( rti' * mat(xk) * rti ) if trans = 'T'.
//
// rti is kth element of W['rti'].
maxn := 0
for _, r := range W.At("r") {
if r.Rows() > maxn {
maxn = r.Rows()
}
}
a := matrix.FloatZeros(maxn, maxn)
for k, v := range W.At("r") {
t := trans
var r *matrix.FloatMatrix
if !inverse {
r = v
t = !trans
} else {
r = W.At("rti")[k]
}
n := r.Rows()
for i := 0; i < x.Cols(); i++ {
// scale diagonal of xk by 0.5
blas.ScalFloat(x, 0.5, &la_.IOpt{"offset", ind + i*x.Rows()},
&la_.IOpt{"inc", n + 1}, &la_.IOpt{"n", n})
// a = r*tril(x) (t is 'N') or a = tril(x)*r (t is 'T')
blas.Copy(r, a)
if !t {
err = blas.TrmmFloat(x, a, 1.0, la_.OptRight, &la_.IOpt{"m", n},
&la_.IOpt{"n", n}, &la_.IOpt{"lda", n}, &la_.IOpt{"ldb", n},
&la_.IOpt{"offsetA", ind + i*x.Rows()})
if err != nil {
//fmt.Printf("5. TrmmFloat: %v\n", err)
return
}
// x := (r*a' + a*r') if t is 'N'
err = blas.Syr2kFloat(r, a, x, 1.0, 0.0, la_.OptNoTrans, &la_.IOpt{"n", n},
&la_.IOpt{"k", n}, &la_.IOpt{"ldb", n}, &la_.IOpt{"ldc", n},
&la_.IOpt{"offsetC", ind + i*x.Rows()})
if err != nil {
//fmt.Printf("6. Syr2kFloat: %v\n", err)
return
}
} else {
err = blas.TrmmFloat(x, a, 1.0, la_.OptLeft, &la_.IOpt{"m", n},
&la_.IOpt{"n", n}, &la_.IOpt{"lda", n}, &la_.IOpt{"ldb", n},
&la_.IOpt{"offsetA", ind + i*x.Rows()})
if err != nil {
//fmt.Printf("7. TrmmFloat: %v\n", err)
return
}
// x := (r'*a + a'*r) if t is 'T'
err = blas.Syr2kFloat(r, a, x, 1.0, 0.0, la_.OptTrans, &la_.IOpt{"n", n},
&la_.IOpt{"k", n}, &la_.IOpt{"ldb", n}, &la_.IOpt{"ldc", n},
&la_.IOpt{"offsetC", ind + i*x.Rows()})
if err != nil {
//fmt.Printf("8. Syr2kFloat: %v\n", err)
return
}
}
}
ind += n * n
}
//if ! checkpnt.MinorEmpty() {
// checkpnt.Check("030scale", minor)
//}
return
}
func (a *matrixA) Af(x, y *matrix.FloatMatrix, alpha, beta float64, trans la.Option) error {
return blas.GemvFloat(a.mA, x, y, alpha, beta, trans)
}
func CheckTransA(A, X, Y *matrix.FloatMatrix) {
blas.GemvFloat(A, X, Y, 1.0, 1.0, linalg.OptTrans)
}
// Solution of KKT equations with zero 1,1 block, by eliminating the
// equality constraints via a QR factorization, and solving the
// reduced KKT system by another QR factorization.
//
// Computes the QR factorization
//
// A' = [Q1, Q2] * [R1; 0]
//
// and returns a function that (1) computes the QR factorization
//
// W^{-T} * G * Q2 = Q3 * R3
//
// (with columns of W^{-T}*G in packed storage), and (2) returns a function for solving
//
// [ 0 A' G' ] [ ux ] [ bx ]
// [ A 0 0 ] * [ uy ] = [ by ].
// [ G 0 -W'*W ] [ uz ] [ bz ]
//
// A is p x n and G is N x n where N = dims['l'] + sum(dims['q']) +
// sum( k**2 for k in dims['s'] ).
//
func kktQr(G *matrix.FloatMatrix, dims *sets.DimensionSet, A *matrix.FloatMatrix, mnl int) (kktFactor, error) {
p, n := A.Size()
cdim := dims.Sum("l", "q") + dims.SumSquared("s")
cdim_pckd := dims.Sum("l", "q") + dims.SumPacked("s")
QA := A.Transpose()
tauA := matrix.FloatZeros(p, 1)
lapack.Geqrf(QA, tauA)
Gs := matrix.FloatZeros(cdim, n)
tauG := matrix.FloatZeros(n-p, 1)
u := matrix.FloatZeros(cdim_pckd, 1)
vv := matrix.FloatZeros(n, 1)
w := matrix.FloatZeros(cdim_pckd, 1)
checkpnt.AddMatrixVar("tauA", tauA)
checkpnt.AddMatrixVar("tauG", tauG)
checkpnt.AddMatrixVar("Gs", Gs)
checkpnt.AddMatrixVar("qr_u", u)
checkpnt.AddMatrixVar("qr_vv", vv)
factor := func(W *sets.FloatMatrixSet, H, Df *matrix.FloatMatrix) (KKTFunc, error) {
var err error = nil
minor := 0
if !checkpnt.MinorEmpty() {
minor = checkpnt.MinorTop()
}
// Gs = W^{-T}*G, in packed storage.
blas.Copy(G, Gs)
//checkpnt.Check("00factor_qr", minor)
scale(Gs, W, true, true)
//checkpnt.Check("01factor_qr", minor)
pack2(Gs, dims, 0)
//checkpnt.Check("02factor_qr", minor)
// Gs := [ Gs1, Gs2 ]
// = Gs * [ Q1, Q2 ]
lapack.Ormqr(QA, tauA, Gs, la.OptRight, &la.IOpt{"m", cdim_pckd})
//checkpnt.Check("03factor_qr", minor)
// QR factorization Gs2 := [ Q3, Q4 ] * [ R3; 0 ]
lapack.Geqrf(Gs, tauG, &la.IOpt{"n", n - p}, &la.IOpt{"m", cdim_pckd},
&la.IOpt{"offseta", Gs.Rows() * p})
checkpnt.Check("10factor_qr", minor)
solve := func(x, y, z *matrix.FloatMatrix) (err error) {
// On entry, x, y, z contain bx, by, bz. On exit, they
// contain the solution x, y, W*z of
//
// [ 0 A' G'*W^{-1} ] [ x ] [bx ]
// [ A 0 0 ] * [ y ] = [by ].
// [ W^{-T}*G 0 -I ] [ W*z ] [W^{-T}*bz]
//
// The system is solved in five steps:
//
// w := W^{-T}*bz - Gs1*R1^{-T}*by
// u := R3^{-T}*Q2'*bx + Q3'*w
// W*z := Q3*u - w
// y := R1^{-1} * (Q1'*bx - Gs1'*(W*z))
// x := [ Q1, Q2 ] * [ R1^{-T}*by; R3^{-1}*u ]
minor := 0
if !checkpnt.MinorEmpty() {
minor = checkpnt.MinorTop()
}
// w := W^{-T} * bz in packed storage
scale(z, W, true, true)
pack(z, w, dims)
//checkpnt.Check("00solve_qr", minor)
// vv := [ Q1'*bx; R3^{-T}*Q2'*bx ]
blas.Copy(x, vv)
lapack.Ormqr(QA, tauA, vv, la.OptTrans)
lapack.Trtrs(Gs, vv, la.OptUpper, la.OptTrans, &la.IOpt{"n", n - p},
&la.IOpt{"offseta", Gs.Rows() * p}, &la.IOpt{"offsetb", p})
//checkpnt.Check("10solve_qr", minor)
// x[:p] := R1^{-T} * by
blas.Copy(y, x)
lapack.Trtrs(QA, x, la.OptUpper, la.OptTrans, &la.IOpt{"n", p})
//checkpnt.Check("20solve_qr", minor)
// w := w - Gs1 * x[:p]
// = W^{-T}*bz - Gs1*by
blas.GemvFloat(Gs, x, w, -1.0, 1.0, &la.IOpt{"n", p}, &la.IOpt{"m", cdim_pckd})
//checkpnt.Check("30solve_qr", minor)
// u := [ Q3'*w + v[p:]; 0 ]
// = [ Q3'*w + R3^{-T}*Q2'*bx; 0 ]
blas.Copy(w, u)
lapack.Ormqr(Gs, tauG, u, la.OptTrans, &la.IOpt{"k", n - p},
&la.IOpt{"offseta", Gs.Rows() * p}, &la.IOpt{"m", cdim_pckd})
blas.AxpyFloat(vv, u, 1.0, &la.IOpt{"offsetx", p}, &la.IOpt{"n", n - p})
blas.ScalFloat(u, 0.0, &la.IOpt{"offset", n - p})
//checkpnt.Check("40solve_qr", minor)
// x[p:] := R3^{-1} * u[:n-p]
blas.Copy(u, x, &la.IOpt{"offsety", p}, &la.IOpt{"n", n - p})
lapack.Trtrs(Gs, x, la.OptUpper, &la.IOpt{"n", n - p},
&la.IOpt{"offset", Gs.Rows() * p}, &la.IOpt{"offsetb", p})
//checkpnt.Check("50solve_qr", minor)
// x is now [ R1^{-T}*by; R3^{-1}*u[:n-p] ]
// x := [Q1 Q2]*x
lapack.Ormqr(QA, tauA, x)
//checkpnt.Check("60solve_qr", minor)
// u := [Q3, Q4] * u - w
// = Q3 * u[:n-p] - w
lapack.Ormqr(Gs, tauG, u, &la.IOpt{"k", n - p}, &la.IOpt{"m", cdim_pckd},
&la.IOpt{"offseta", Gs.Rows() * p})
blas.AxpyFloat(w, u, -1.0)
//checkpnt.Check("70solve_qr", minor)
// y := R1^{-1} * ( v[:p] - Gs1'*u )
// = R1^{-1} * ( Q1'*bx - Gs1'*u )
blas.Copy(vv, y, &la.IOpt{"n", p})
blas.GemvFloat(Gs, u, y, -1.0, 1.0, &la.IOpt{"m", cdim_pckd},
&la.IOpt{"n", p}, la.OptTrans)
lapack.Trtrs(QA, y, la.OptUpper, &la.IOpt{"n", p})
//checkpnt.Check("80solve_qr", minor)
unpack(u, z, dims)
checkpnt.Check("90solve_qr", minor)
return nil
}
return solve, err
}
return factor, nil
}
