func updateBlas(t *testing.T, Y1, Y2, C1, C2, T, W *matrix.FloatMatrix) {
if W.Rows() != C1.Cols() {
panic("W.Rows != C1.Cols")
}
// W = C1.T
ScalePlus(W, C1, 0.0, 1.0, TRANSB)
//fmt.Printf("W = C1.T:\n%v\n", W)
// W = C1.T*Y1
blas.TrmmFloat(Y1, W, 1.0, linalg.OptLower, linalg.OptUnit, linalg.OptRight)
t.Logf("W = C1.T*Y1:\n%v\n", W)
// W = W + C2.T*Y2
blas.GemmFloat(C2, Y2, W, 1.0, 1.0, linalg.OptTransA)
t.Logf("W = W + C2.T*Y2:\n%v\n", W)
// --- here: W == C.T*Y ---
// W = W*T
blas.TrmmFloat(T, W, 1.0, linalg.OptUpper, linalg.OptRight)
t.Logf("W = C.T*Y*T:\n%v\n", W)
// --- here: W == C.T*Y*T ---
// C2 = C2 - Y2*W.T
blas.GemmFloat(Y2, W, C2, -1, 1.0, linalg.OptTransB)
t.Logf("C2 = C2 - Y2*W.T:\n%v\n", C2)
// W = Y1*W.T ==> W.T = W*Y1.T
blas.TrmmFloat(Y1, W, 1.0, linalg.OptLower,
linalg.OptUnit, linalg.OptRight, linalg.OptTrans)
t.Logf("W.T = W*Y1.T:\n%v\n", W)
// C1 = C1 - W.T
ScalePlus(C1, W, 1.0, -1.0, TRANSB)
//fmt.Printf("C1 = C1 - W.T:\n%v\n", C1)
// --- here: C = (I - Y*T*Y.T).T * C ---
}
func CTestGemm(m, n, p int) (fnc func(), A, B, C *matrix.FloatMatrix) {
A, B, C = mperf.MakeData(m, n, p, randomData, false)
fnc = func() {
blas.GemmFloat(A, B, C, 1.0, 1.0)
}
return fnc, A, B, C
}
func CTestGemmTransB(m, n, p int) (fnc func(), A, B, C *matrix.FloatMatrix) {
A, B, C = mperf.MakeData(m, n, p, randomData, false)
B = B.Transpose()
fnc = func() {
blas.GemmFloat(A, B, C, 1.0, 1.0, linalg.OptTransB)
}
return fnc, A, B, C
}
func _TestMultTransABig(t *testing.T) {
bM := 100*M + 3
bN := 100*N + 3
bP := 100*P + 3
D := matrix.FloatNormal(bM, bP)
E := matrix.FloatNormal(bP, bN)
C0 := matrix.FloatZeros(bM, bN)
C1 := matrix.FloatZeros(bM, bN)
Dt := D.Transpose()
Dr := Dt.FloatArray()
Er := E.FloatArray()
C1r := C1.FloatArray()
blas.GemmFloat(Dt, E, C0, 1.0, 1.0, linalg.OptTransA)
DMult(C1r, Dr, Er, 1.0, 1.0, TRANSA, bM, bM, bP, bP, 0, bN, 0, bM, 32, 32, 32)
t.Logf("C0 == C1: %v\n", C0.AllClose(C1))
}
func _TestMultTransASmall(t *testing.T) {
bM := 7
bN := 7
bP := 7
D := matrix.FloatNormal(bM, bP)
E := matrix.FloatNormal(bP, bN)
C0 := matrix.FloatWithValue(bM, bN, 0.0)
C1 := C0.Copy()
Dt := D.Transpose()
Dr := Dt.FloatArray()
Er := E.FloatArray()
C1r := C1.FloatArray()
blas.GemmFloat(Dt, E, C0, 1.0, 1.0, linalg.OptTransA)
t.Logf("blas: C=D*E\n%v\n", C0)
DMult(C1r, Dr, Er, 1.0, 1.0, TRANSA, bM, bM, bP, bP, 0, bN, 0, bM, 4, 4, 4)
t.Logf("C0 == C1: %v\n", C0.AllClose(C1))
t.Logf("C1: C1=D*E\n%v\n", C1)
}
func _TestMultBig(t *testing.T) {
bM := 100*M + 3
bN := 100*N + 3
bP := 100*P + 3
D := matrix.FloatNormal(bM, bP)
E := matrix.FloatNormal(bP, bN)
C0 := matrix.FloatZeros(bM, bN)
C1 := matrix.FloatZeros(bM, bN)
Dr := D.FloatArray()
Er := E.FloatArray()
C1r := C1.FloatArray()
blas.GemmFloat(D, E, C0, 1.0, 1.0)
//t.Logf("blas: C=D*E\n%v\n", C0)
DMult(C1r, Dr, Er, 1.0, 1.0, NOTRANS, bM, bM, bP, bP, 0, bN, 0, bM, 32, 32, 32)
res := C0.AllClose(C1)
t.Logf("C0 == C1: %v\n", res)
}
func _TestMultSmall(t *testing.T) {
bM := 6
bN := 6
bP := 6
D := matrix.FloatNormal(bM, bP)
E := matrix.FloatNormal(bP, bN)
C0 := matrix.FloatWithValue(bM, bN, 1.0)
C1 := C0.Copy()
Dr := D.FloatArray()
Er := E.FloatArray()
C1r := C1.FloatArray()
blas.GemmFloat(D, E, C0, 1.0, 1.0)
t.Logf("blas: C=D*E\n%v\n", C0)
DMult(C1r, Dr, Er, 1.0, 1.0, NOTRANS, bM, bM, bP, bP, 0, bN, 0, bM, 4, 4, 4)
t.Logf("C0 == C1: %v\n", C0.AllClose(C1))
t.Logf("C1: C1=D*E\n%v\n", C1)
}
func _TestMultTransABSmall(t *testing.T) {
bM := 7
bN := 7
bP := 7
D := matrix.FloatNormal(bM, bP)
E := matrix.FloatNormal(bP, bN)
C0 := matrix.FloatZeros(bM, bN)
C1 := matrix.FloatZeros(bM, bN)
Dt := D.Transpose()
Et := E.Transpose()
Dr := Dt.FloatArray()
Er := Et.FloatArray()
C1r := C1.FloatArray()
blas.GemmFloat(Dt, Et, C0, 1.0, 1.0, linalg.OptTransA, linalg.OptTransB)
t.Logf("blas: C=D.T*E.T\n%v\n", C0)
DMult(C1r, Dr, Er, 1.0, 1.0, TRANSA|TRANSB, bM, bM, bP, bP, 0, bN, 0, bM, 4, 4, 4)
t.Logf("C0 == C1: %v\n", C0.AllClose(C1))
t.Logf("C1: C1=D.T*E.T\n%v\n", C1)
}
func mcsdp(w *matrix.FloatMatrix) (*Solution, error) {
//
// Returns solution x, z to
//
// (primal) minimize sum(x)
// subject to w + diag(x) >= 0
//
// (dual) maximize -tr(w*z)
// subject to diag(z) = 1
// z >= 0.
//
n := w.Rows()
G := &matrixFs{n}
cngrnc := func(r, x *matrix.FloatMatrix, alpha float64) (err error) {
// Congruence transformation
//
// x := alpha * r'*x*r.
//
// r and x are square matrices.
//
err = nil
// tx = matrix(x, (n,n)) is copying and reshaping
// scale diagonal of x by 1/2, (x is (n,n))
tx := x.Copy()
matrix.Reshape(tx, n, n)
tx.Diag().Scale(0.5)
// a := tril(x)*r
// (python: a = +r is really making a copy of r)
a := r.Copy()
err = blas.TrmmFloat(tx, a, 1.0, linalg.OptLeft)
// x := alpha*(a*r' + r*a')
err = blas.Syr2kFloat(r, a, tx, alpha, 0.0, linalg.OptTrans)
// x[:] = tx[:]
tx.CopyTo(x)
return
}
Fkkt := func(W *sets.FloatMatrixSet) (KKTFunc, error) {
// Solve
// -diag(z) = bx
// -diag(x) - inv(rti*rti') * z * inv(rti*rti') = bs
//
// On entry, x and z contain bx and bs.
// On exit, they contain the solution, with z scaled
// (inv(rti)'*z*inv(rti) is returned instead of z).
//
// We first solve
//
// ((rti*rti') .* (rti*rti')) * x = bx - diag(t*bs*t)
//
// and take z = -rti' * (diag(x) + bs) * rti.
var err error = nil
rti := W.At("rti")[0]
// t = rti*rti' as a nonsymmetric matrix.
t := matrix.FloatZeros(n, n)
err = blas.GemmFloat(rti, rti, t, 1.0, 0.0, linalg.OptTransB)
if err != nil {
return nil, err
}
// Cholesky factorization of tsq = t.*t.
tsq := matrix.Mul(t, t)
err = lapack.Potrf(tsq)
if err != nil {
return nil, err
}
f := func(x, y, z *matrix.FloatMatrix) (err error) {
// tbst := t * zs * t = t * bs * t
tbst := z.Copy()
matrix.Reshape(tbst, n, n)
cngrnc(t, tbst, 1.0)
// x := x - diag(tbst) = bx - diag(rti*rti' * bs * rti*rti')
diag := tbst.Diag().Transpose()
x.Minus(diag)
// x := (t.*t)^{-1} * x = (t.*t)^{-1} * (bx - diag(t*bs*t))
err = lapack.Potrs(tsq, x)
// z := z + diag(x) = bs + diag(x)
// z, x are really column vectors here
z.AddIndexes(matrix.MakeIndexSet(0, n*n, n+1), x.FloatArray())
// z := -rti' * z * rti = -rti' * (diag(x) + bs) * rti
cngrnc(rti, z, -1.0)
return nil
}
return f, nil
}
c := matrix.FloatWithValue(n, 1, 1.0)
// initial feasible x: x = 1.0 - min(lmbda(w))
lmbda := matrix.FloatZeros(n, 1)
wp := w.Copy()
lapack.Syevx(wp, lmbda, nil, 0.0, nil, []int{1, 1}, linalg.OptRangeInt)
x0 := matrix.FloatZeros(n, 1).Add(-lmbda.GetAt(0, 0) + 1.0)
s0 := w.Copy()
s0.Diag().Plus(x0.Transpose())
matrix.Reshape(s0, n*n, 1)
// initial feasible z is identity
z0 := matrix.FloatIdentity(n)
matrix.Reshape(z0, n*n, 1)
dims := sets.DSetNew("l", "q", "s")
dims.Set("s", []int{n})
primalstart := sets.FloatSetNew("x", "s")
dualstart := sets.FloatSetNew("z")
primalstart.Set("x", x0)
primalstart.Set("s", s0)
dualstart.Set("z", z0)
var solopts SolverOptions
solopts.MaxIter = 30
solopts.ShowProgress = false
h := w.Copy()
matrix.Reshape(h, h.NumElements(), 1)
return ConeLpCustomMatrix(c, G, h, nil, nil, dims, Fkkt, &solopts, primalstart, dualstart)
}
func updateScaling(W *sets.FloatMatrixSet, lmbda, s, z *matrix.FloatMatrix) (err error) {
err = nil
var stmp, ztmp *matrix.FloatMatrix
/*
Nonlinear and 'l' blocks
d := d .* sqrt( s ./ z )
lmbda := lmbda .* sqrt(s) .* sqrt(z)
*/
mnl := 0
dnlset := W.At("dnl")
dnliset := W.At("dnli")
dset := W.At("d")
diset := W.At("di")
beta := W.At("beta")[0]
if dnlset != nil && dnlset[0].NumElements() > 0 {
mnl = dnlset[0].NumElements()
}
ml := dset[0].NumElements()
m := mnl + ml
//fmt.Printf("ml=%d, mnl=%d, m=%d'n", ml, mnl, m)
stmp = matrix.FloatVector(s.FloatArray()[:m])
stmp.Apply(math.Sqrt)
s.SetIndexesFromArray(stmp.FloatArray(), matrix.MakeIndexSet(0, m, 1)...)
ztmp = matrix.FloatVector(z.FloatArray()[:m])
ztmp.Apply(math.Sqrt)
z.SetIndexesFromArray(ztmp.FloatArray(), matrix.MakeIndexSet(0, m, 1)...)
// d := d .* s .* z
if len(dnlset) > 0 {
blas.TbmvFloat(s, dnlset[0], &la_.IOpt{"n", mnl}, &la_.IOpt{"k", 0}, &la_.IOpt{"lda", 1})
blas.TbsvFloat(z, dnlset[0], &la_.IOpt{"n", mnl}, &la_.IOpt{"k", 0}, &la_.IOpt{"lda", 1})
//dnliset[0].Apply(dnlset[0], func(a float64)float64 { return 1.0/a})
//--dnliset[0] = matrix.Inv(dnlset[0])
matrix.Set(dnliset[0], dnlset[0])
dnliset[0].Inv()
}
blas.TbmvFloat(s, dset[0], &la_.IOpt{"n", ml},
&la_.IOpt{"k", 0}, &la_.IOpt{"lda", 1}, &la_.IOpt{"offseta", mnl})
blas.TbsvFloat(z, dset[0], &la_.IOpt{"n", ml},
&la_.IOpt{"k", 0}, &la_.IOpt{"lda", 1}, &la_.IOpt{"offseta", mnl})
//diset[0].Apply(dset[0], func(a float64)float64 { return 1.0/a})
//--diset[0] = matrix.Inv(dset[0])
matrix.Set(diset[0], dset[0])
diset[0].Inv()
// lmbda := s .* z
blas.CopyFloat(s, lmbda, &la_.IOpt{"n", m})
blas.TbmvFloat(z, lmbda, &la_.IOpt{"n", m}, &la_.IOpt{"k", 0}, &la_.IOpt{"lda", 1})
// 'q' blocks.
// Let st and zt be the new variables in the old scaling:
//
// st = s_k, zt = z_k
//
// and a = sqrt(st' * J * st), b = sqrt(zt' * J * zt).
//
// 1. Compute the hyperbolic Householder transformation 2*q*q' - J
// that maps st/a to zt/b.
//
// c = sqrt( (1 + st'*zt/(a*b)) / 2 )
// q = (st/a + J*zt/b) / (2*c).
//
// The new scaling point is
//
// wk := betak * sqrt(a/b) * (2*v[k]*v[k]' - J) * q
//
// with betak = W['beta'][k].
//
// 3. The scaled variable:
//
// lambda_k0 = sqrt(a*b) * c
// lambda_k1 = sqrt(a*b) * ( (2vk*vk' - J) * (-d*q + u/2) )_1
//
// where
//
// u = st/a - J*zt/b
// d = ( vk0 * (vk'*u) + u0/2 ) / (2*vk0 *(vk'*q) - q0 + 1).
//
// 4. Update scaling
//
// v[k] := wk^1/2
// = 1 / sqrt(2*(wk0 + 1)) * (wk + e).
// beta[k] *= sqrt(a/b)
ind := m
for k, v := range W.At("v") {
m = v.NumElements()
// ln = sqrt( lambda_k' * J * lambda_k ) !! NOT USED!!
jnrm2(lmbda, m, ind) // ?? NOT USED ??
// a = sqrt( sk' * J * sk ) = sqrt( st' * J * st )
// s := s / a = st / a
aa := jnrm2(s, m, ind)
blas.ScalFloat(s, 1.0/aa, &la_.IOpt{"n", m}, &la_.IOpt{"offset", ind})
// b = sqrt( zk' * J * zk ) = sqrt( zt' * J * zt )
// z := z / a = zt / b
bb := jnrm2(z, m, ind)
blas.ScalFloat(z, 1.0/bb, &la_.IOpt{"n", m}, &la_.IOpt{"offset", ind})
// c = sqrt( ( 1 + (st'*zt) / (a*b) ) / 2 )
cc := blas.DotFloat(s, z, &la_.IOpt{"offsetx", ind}, &la_.IOpt{"offsety", ind},
&la_.IOpt{"n", m})
cc = math.Sqrt((1.0 + cc) / 2.0)
// vs = v' * st / a
vs := blas.DotFloat(v, s, &la_.IOpt{"offsety", ind}, &la_.IOpt{"n", m})
// vz = v' * J *zt / b
vz := jdot(v, z, m, 0, ind)
// vq = v' * q where q = (st/a + J * zt/b) / (2 * c)
vq := (vs + vz) / 2.0 / cc
// vq = v' * q where q = (st/a + J * zt/b) / (2 * c)
vu := vs - vz
// lambda_k0 = c
lmbda.SetIndex(ind, cc)
// wk0 = 2 * vk0 * (vk' * q) - q0
wk0 := 2.0*v.GetIndex(0)*vq - (s.GetIndex(ind)+z.GetIndex(ind))/2.0/cc
// d = (v[0] * (vk' * u) - u0/2) / (wk0 + 1)
dd := (v.GetIndex(0)*vu - s.GetIndex(ind)/2.0 + z.GetIndex(ind)/2.0) / (wk0 + 1.0)
// lambda_k1 = 2 * v_k1 * vk' * (-d*q + u/2) - d*q1 + u1/2
blas.CopyFloat(v, lmbda, &la_.IOpt{"offsetx", 1}, &la_.IOpt{"offsety", ind + 1},
&la_.IOpt{"n", m - 1})
blas.ScalFloat(lmbda, (2.0 * (-dd*vq + 0.5*vu)),
&la_.IOpt{"offsetx", ind + 1}, &la_.IOpt{"offsety", ind + 1}, &la_.IOpt{"n", m - 1})
blas.AxpyFloat(s, lmbda, 0.5*(1.0-dd/cc),
&la_.IOpt{"offsetx", ind + 1}, &la_.IOpt{"offsety", ind + 1}, &la_.IOpt{"n", m - 1})
blas.AxpyFloat(z, lmbda, 0.5*(1.0+dd/cc),
&la_.IOpt{"offsetx", ind + 1}, &la_.IOpt{"offsety", ind + 1}, &la_.IOpt{"n", m - 1})
// Scale so that sqrt(lambda_k' * J * lambda_k) = sqrt(aa*bb).
blas.ScalFloat(lmbda, math.Sqrt(aa*bb), &la_.IOpt{"offset", ind}, &la_.IOpt{"n", m})
// v := (2*v*v' - J) * q
// = 2 * (v'*q) * v' - (J* st/a + zt/b) / (2*c)
blas.ScalFloat(v, 2.0*vq)
v.SetIndex(0, v.GetIndex(0)-(s.GetIndex(ind)/2.0/cc))
blas.AxpyFloat(s, v, 0.5/cc, &la_.IOpt{"offsetx", ind + 1}, &la_.IOpt{"offsety", 1},
&la_.IOpt{"n", m - 1})
blas.AxpyFloat(z, v, -0.5/cc, &la_.IOpt{"offsetx", ind}, &la_.IOpt{"n", m})
// v := v^{1/2} = 1/sqrt(2 * (v0 + 1)) * (v + e)
v0 := v.GetIndex(0) + 1.0
v.SetIndex(0, v0)
blas.ScalFloat(v, 1.0/math.Sqrt(2.0*v0))
// beta[k] *= ( aa / bb )**1/2
bk := beta.GetIndex(k)
beta.SetIndex(k, bk*math.Sqrt(aa/bb))
ind += m
}
//fmt.Printf("-- end of q:\nz=\n%v\nlmbda=\n%v\n", z.ConvertToString(), lmbda.ConvertToString())
//fmt.Printf("beta=\n%v\n", beta.ConvertToString())
// 's' blocks
//
// Let st, zt be the updated variables in the old scaling:
//
// st = Ls * Ls', zt = Lz * Lz'.
//
// where Ls and Lz are the 's' components of s, z.
//
// 1. SVD Lz'*Ls = Uk * lambda_k^+ * Vk'.
//
// 2. New scaling is
//
// r[k] := r[k] * Ls * Vk * diag(lambda_k^+)^{-1/2}
// rti[k] := r[k] * Lz * Uk * diag(lambda_k^+)^{-1/2}.
//
maxr := 0
for _, m := range W.At("r") {
if m.Rows() > maxr {
maxr = m.Rows()
}
}
work := matrix.FloatZeros(maxr*maxr, 1)
vlensum := 0
for _, m := range W.At("v") {
vlensum += m.NumElements()
}
ind = mnl + ml + vlensum
ind2 := ind
ind3 := 0
rset := W.At("r")
rtiset := W.At("rti")
for k, _ := range rset {
r := rset[k]
rti := rtiset[k]
m = r.Rows()
//fmt.Printf("m=%d, r=\n%v\nrti=\n%v\n", m, r.ConvertToString(), rti.ConvertToString())
// r := r*sk = r*Ls
blas.GemmFloat(r, s, work, 1.0, 0.0, &la_.IOpt{"m", m}, &la_.IOpt{"n", m},
&la_.IOpt{"k", m}, &la_.IOpt{"ldb", m}, &la_.IOpt{"ldc", m},
&la_.IOpt{"offsetb", ind2})
//fmt.Printf("1 work=\n%v\n", work.ConvertToString())
blas.CopyFloat(work, r, &la_.IOpt{"n", m * m})
// rti := rti*zk = rti*Lz
blas.GemmFloat(rti, z, work, 1.0, 0.0, &la_.IOpt{"m", m}, &la_.IOpt{"n", m},
&la_.IOpt{"k", m}, &la_.IOpt{"ldb", m}, &la_.IOpt{"ldc", m},
&la_.IOpt{"offsetb", ind2})
//fmt.Printf("2 work=\n%v\n", work.ConvertToString())
blas.CopyFloat(work, rti, &la_.IOpt{"n", m * m})
// SVD Lz'*Ls = U * lmbds^+ * V'; store U in sk and V' in zk. '
blas.GemmFloat(z, s, work, 1.0, 0.0, la_.OptTransA, &la_.IOpt{"m", m},
&la_.IOpt{"n", m}, &la_.IOpt{"k", m}, &la_.IOpt{"lda", m}, &la_.IOpt{"ldb", m},
&la_.IOpt{"ldc", m}, &la_.IOpt{"offseta", ind2}, &la_.IOpt{"offsetb", ind2})
//fmt.Printf("3 work=\n%v\n", work.ConvertToString())
// U = s, Vt = z
lapack.GesvdFloat(work, lmbda, s, z, la_.OptJobuAll, la_.OptJobvtAll,
&la_.IOpt{"m", m}, &la_.IOpt{"n", m}, &la_.IOpt{"lda", m}, &la_.IOpt{"ldu", m},
&la_.IOpt{"ldvt", m}, &la_.IOpt{"offsets", ind}, &la_.IOpt{"offsetu", ind2},
&la_.IOpt{"offsetvt", ind2})
// r := r*V
blas.GemmFloat(r, z, work, 1.0, 0.0, la_.OptTransB, &la_.IOpt{"m", m},
&la_.IOpt{"n", m}, &la_.IOpt{"k", m}, &la_.IOpt{"ldb", m}, &la_.IOpt{"ldc", m},
&la_.IOpt{"offsetb", ind2})
//fmt.Printf("4 work=\n%v\n", work.ConvertToString())
blas.CopyFloat(work, r, &la_.IOpt{"n", m * m})
// rti := rti*U
blas.GemmFloat(rti, s, work, 1.0, 0.0, &la_.IOpt{"m", m}, &la_.IOpt{"n", m},
&la_.IOpt{"k", m}, &la_.IOpt{"ldb", m}, &la_.IOpt{"ldc", m},
&la_.IOpt{"offsetb", ind2})
//fmt.Printf("5 work=\n%v\n", work.ConvertToString())
blas.CopyFloat(work, rti, &la_.IOpt{"n", m * m})
for i := 0; i < m; i++ {
a := 1.0 / math.Sqrt(lmbda.GetIndex(ind+i))
blas.ScalFloat(r, a, &la_.IOpt{"n", m}, &la_.IOpt{"offset", m * i})
blas.ScalFloat(rti, a, &la_.IOpt{"n", m}, &la_.IOpt{"offset", m * i})
}
ind += m
ind2 += m * m
ind3 += m // !!NOT USED: ind3!!
}
//fmt.Printf("-- end of s:\nz=\n%v\nlmbda=\n%v\n", z.ConvertToString(), lmbda.ConvertToString())
return
}
func CheckTransAB(A, B, C *matrix.FloatMatrix) {
blas.GemmFloat(A, B, C, 1.0, 1.0, linalg.OptTransA, linalg.OptTransB)
}
func CheckNoTrans(A, B, C *matrix.FloatMatrix) {
blas.GemmFloat(A, B, C, 1.0, 1.0)
}
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
}