/*
* Applies a real elementary reflector H to a real m by n matrix A,
* from either the left or the right. H is represented in the form
*
* H = I - tau * ( 1 ) * ( 1 v.T )
* ( v )
*
* where tau is a real scalar and v is a real vector.
*
* If tau = 0, then H is taken to be the unit cmat.
*
* A is /a1\ a1 := a1 - w1
* \A2/ A2 := A2 - v*w1
* w1 := tau*(a1 + A2.T*v) if side == LEFT
* := tau*(a1 + A2*v) if side == RIGHT
*
* Intermediate work space w1 required as parameter, no allocation.
*/
func applyHouseholder2x1(tau, v, a1, A2, w1 *cmat.FloatMatrix, flags int) *gomas.Error {
var err *gomas.Error = nil
tval := tau.Get(0, 0)
if tval == 0.0 {
return err
}
// shape oblivious vector copy.
blasd.Axpby(w1, a1, 1.0, 0.0)
if flags&gomas.LEFT != 0 {
// w1 = a1 + A2.T*v
err = blasd.MVMult(w1, A2, v, 1.0, 1.0, gomas.TRANSA)
} else {
// w1 = a1 + A2*v
err = blasd.MVMult(w1, A2, v, 1.0, 1.0, gomas.NONE)
}
// w1 = tau*w1
blasd.Scale(w1, tval)
// a1 = a1 - w1
blasd.Axpy(a1, w1, -1.0)
// A2 = A2 - v*w1
if flags&gomas.LEFT != 0 {
err = blasd.MVUpdate(A2, v, w1, -1.0)
} else {
err = blasd.MVUpdate(A2, w1, v, -1.0)
}
return err
}
func TestDSyr1(t *testing.T) {
const N = 911
A := cmat.NewMatrix(N, N)
X := cmat.NewMatrix(N, 1)
B := cmat.NewMatrix(N, N)
//ones := cmat.NewFloatConstSource(1.0)
zeromean := cmat.NewFloatUniformSource(0.5, 2.0)
A.SetFrom(zeromean, cmat.LOWER)
X.SetFrom(zeromean)
B.Copy(A)
// B = A*B
blasd.MVUpdate(B, X, X, 1.0)
cmat.TriL(B, cmat.NONE)
blasd.MVUpdateSym(A, X, 1.0, gomas.LOWER)
ok := B.AllClose(A)
t.Logf("MVUpdateSym(A, X, L) == TriL(MVUpdate(A, X, X)) : %v\n", ok)
A.SetFrom(zeromean, cmat.UPPER)
cmat.TriU(A, cmat.NONE)
B.Copy(A)
blasd.MVUpdate(B, X, X, 1.0)
cmat.TriU(B, cmat.NONE)
blasd.MVUpdateSym(A, X, 1.0, gomas.UPPER)
ok = B.AllClose(A)
t.Logf("MVUpdateSym(A, X, U) == TriU(MVUpdate(A, X, X)) : %v\n", ok)
}
func TestDSyrOther(t *testing.T) {
const N = 911
var vec, As, Bs cmat.FloatMatrix
P := N / 3
A := cmat.NewMatrix(P, P)
X := cmat.NewMatrix(P, 1)
B := cmat.NewMatrix(P, P)
//ones := cmat.NewFloatConstSource(1.0)
zeromean := cmat.NewFloatUniformSource(0.5, 2.0)
A.SetFrom(zeromean, cmat.UPPER)
X.SetFrom(zeromean)
B.Copy(A)
// update submatrices
for i := 1; i < P; i++ {
vec.SubMatrix(A, i-1, i, 1, P-i)
As.SubMatrix(A, i, i)
Bs.SubMatrix(B, i, i)
// update with normal and symmetric
blasd.MVUpdate(&Bs, &vec, &vec, 1.0)
blasd.MVUpdateSym(&As, &vec, 1.0, gomas.UPPER)
}
// make normal update triangular and compare
cmat.TriU(B, cmat.NONE)
ok := B.AllClose(A)
t.Logf("submatrix updates on upper triangular : %v\n", ok)
A.SetFrom(zeromean, cmat.LOWER)
cmat.TriL(A, cmat.NONE)
B.Copy(A)
// update submatrices
for i := 1; i < P; i++ {
vec.SubMatrix(A, i-1, i, 1, P-i)
As.SubMatrix(A, i, i)
Bs.SubMatrix(B, i, i)
// update with normal and symmetric
blasd.MVUpdate(&Bs, &vec, &vec, 1.0)
blasd.MVUpdateSym(&As, &vec, 1.0, gomas.LOWER)
}
// make normal update triangular and compare
cmat.TriL(B, cmat.NONE)
ok = B.AllClose(A)
t.Logf("submatrix updates on lower triangular : %v\n", ok)
}
// unblocked LU decomposition w/o pivots, FLAME LU nopivots variant 5
func unblockedLUnoPiv(A *cmat.FloatMatrix, conf *gomas.Config) *gomas.Error {
var ATL, ATR, ABL, ABR cmat.FloatMatrix
var A00, a01, A02, a10, a11, a12, A20, a21, A22 cmat.FloatMatrix
var err *gomas.Error = nil
util.Partition2x2(
&ATL, &ATR,
&ABL, &ABR, A, 0, 0, util.PTOPLEFT)
for m(&ATL) < m(A) {
util.Repartition2x2to3x3(&ATL,
&A00, &a01, &A02,
&a10, &a11, &a12,
&A20, &a21, &A22, A, 1, util.PBOTTOMRIGHT)
// a21 = a21/a11
blasd.InvScale(&a21, a11.Get(0, 0))
// A22 = A22 - a21*a12
blasd.MVUpdate(&A22, &a21, &a12, -1.0)
util.Continue3x3to2x2(
&ATL, &ATR,
&ABL, &ABR, &A00, &a11, &A22, A, util.PBOTTOMRIGHT)
}
return err
}
func TestDSyr2(t *testing.T) {
const N = 911
A := cmat.NewMatrix(N, N)
X := cmat.NewMatrix(N, 1)
Y := cmat.NewMatrix(N, 1)
B := cmat.NewMatrix(N, N)
ones := cmat.NewFloatConstSource(1.0)
twos := cmat.NewFloatConstSource(2.0)
zeromean := cmat.NewFloatUniformSource(0.5, 2.0)
A.SetFrom(zeromean, cmat.LOWER)
X.SetFrom(ones)
Y.SetFrom(twos)
B.Copy(A)
// B = A*B
blasd.MVUpdate(B, X, Y, 1.0)
blasd.MVUpdate(B, Y, X, 1.0)
cmat.TriL(B, cmat.NONE)
blasd.MVUpdate2Sym(A, X, Y, 1.0, gomas.LOWER)
ok := B.AllClose(A)
if N < 10 {
t.Logf("A:\n%v\n", A)
t.Logf("B:\n%v\n", B)
}
t.Logf("MVUpdate2Sym(A, X, Y, L) == TriL(MVUpdate(A, X, Y);MVUpdate(A, Y, X)) : %v\n", ok)
A.SetFrom(zeromean, cmat.UPPER)
cmat.TriU(A, cmat.NONE)
B.Copy(A)
blasd.MVUpdate(B, X, Y, 1.0)
blasd.MVUpdate(B, Y, X, 1.0)
cmat.TriU(B, cmat.NONE)
blasd.MVUpdate2Sym(A, X, Y, 1.0, gomas.UPPER)
ok = B.AllClose(A)
if N < 10 {
t.Logf("A:\n%v\n", A)
t.Logf("B:\n%v\n", B)
}
t.Logf("MVUpdate2Sym(A, X, Y, U) == TriU(MVUpdate(A, X, Y);MVUpdate(A, Y, X)) : %v\n", ok)
}
/*
* Apply elementary Householder reflector v to matrix A2.
*
* H = I - tau*v*v.t;
*
* RIGHT: A = A*H = A - tau*A*v*v.T = A - tau*w1*v.T
* LEFT: A = H*A = A - tau*v*v.T*A = A - tau*v*A.T*v = A - tau*v*w1
*/
func applyHouseholder1x1(tau, v, A2, w1 *cmat.FloatMatrix, flags int) *gomas.Error {
var err *gomas.Error = nil
tval := tau.Get(0, 0)
if tval == 0.0 {
return nil
}
if flags&gomas.LEFT != 0 {
// w1 = A2.T*v
err = blasd.MVMult(w1, A2, v, 1.0, 0.0, gomas.TRANSA)
if err == nil {
// A2 = A2 - tau*v*w1; m(A2) == len(v) && n(A2) == len(w1)
err = blasd.MVUpdate(A2, v, w1, -tval)
}
} else {
// w1 = A2*v
err = blasd.MVMult(w1, A2, v, 1.0, 0.0, gomas.NONE)
if err == nil {
// A2 = A2 - tau*w1*v; m(A2) == len(w1) && n(A2) == len(v)
err = blasd.MVUpdate(A2, w1, v, -tval)
}
}
return err
}
/* From LAPACK/dlarf.f
*
* Applies a real elementary reflector H to a real m by n matrix A,
* from either the left or the right. H is represented in the form
*
* H = I - tau * ( 1 ) * ( 1 v.T )
* ( v )
*
* where tau is a real scalar and v is a real vector.
*
* If tau = 0, then H is taken to be the unit cmat.
*
* A is /a1\ a1 := a1 - w1
* \A2/ A2 := A2 - v*w1
* w1 := tau*(a1 + A2.T*v) if side == LEFT
* := tau*(a1 + A2*v) if side == RIGHT
*
* Allocates/frees intermediate work space matrix w1.
*/
func applyHouseholder(tau, v, a1, A2 *cmat.FloatMatrix, flags int) {
tval := tau.Get(0, 0)
if tval == 0.0 {
return
}
w1 := cmat.NewCopy(a1)
if flags&gomas.LEFT != 0 {
// w1 = a1 + A2.T*v
blasd.MVMult(w1, A2, v, 1.0, 1.0, gomas.TRANSA)
} else {
// w1 = a1 + A2*v
blasd.MVMult(w1, A2, v, 1.0, 1.0, gomas.NONE)
}
// w1 = tau*w1
blasd.Scale(w1, tval)
// a1 = a1 - w1
blasd.Axpy(a1, w1, -1.0)
// A2 = A2 - v*w1
blasd.MVUpdate(A2, v, w1, -1.0)
}
/*
* Computes upper Hessenberg reduction of N-by-N matrix A using unblocked
* algorithm as described in (1).
*
* Hessengerg reduction: A = Q.T*B*Q, Q unitary, B upper Hessenberg
* Q = H(0)*H(1)*...*H(k) where H(k) is k'th Householder reflector.
*
* Compatible with lapack.DGEHD2.
*/
func unblkHessGQvdG(A, Tvec, W *cmat.FloatMatrix, row int) {
var ATL, ATR, ABL, ABR cmat.FloatMatrix
var A00, a11, a21, A22 cmat.FloatMatrix
var AL, AR, A0, a1, A2 cmat.FloatMatrix
var tT, tB cmat.FloatMatrix
var t0, tau1, t2, w12, v1 cmat.FloatMatrix
util.Partition2x2(
&ATL, &ATR,
&ABL, &ABR, A, row, 0, util.PTOPLEFT)
util.Partition1x2(
&AL, &AR, A, 0, util.PLEFT)
util.Partition2x1(
&tT,
&tB, Tvec, 0, util.PTOP)
v1.SubMatrix(W, 0, 0, m(A), 1)
for m(&ABR) > 1 && n(&ABR) > 0 {
util.Repartition2x2to3x3(&ATL,
&A00, nil, nil,
nil, &a11, nil,
nil, &a21, &A22, A, 1, util.PBOTTOMRIGHT)
util.Repartition1x2to1x3(&AL,
&A0, &a1, &A2, A, 1, util.PRIGHT)
util.Repartition2x1to3x1(&tT,
&t0,
&tau1,
&t2, Tvec, 1, util.PBOTTOM)
// ------------------------------------------------------
// a21 = [beta; H(k)].T
computeHouseholderVec(&a21, &tau1)
tauval := tau1.Get(0, 0)
beta := a21.Get(0, 0)
a21.Set(0, 0, 1.0)
// v1 := A2*a21
blasd.MVMult(&v1, &A2, &a21, 1.0, 0.0, gomas.NONE)
// A2 := A2 - tau*v1*a21 (A2 := A2*H(k))
blasd.MVUpdate(&A2, &v1, &a21, -tauval)
w12.SubMatrix(W, 0, 0, n(&A22), 1)
// w12 := a21.T*A22 = A22.T*a21
blasd.MVMult(&w12, &A22, &a21, 1.0, 0.0, gomas.TRANS)
// A22 := A22 - tau*a21*w12 (A22 := H(k)*A22)
blasd.MVUpdate(&A22, &a21, &w12, -tauval)
a21.Set(0, 0, beta)
// ------------------------------------------------------
util.Continue3x3to2x2(
&ATL, &ATR,
&ABL, &ABR, &A00, &a11, &A22, A, util.PBOTTOMRIGHT)
util.Continue1x3to1x2(
&AL, &AR, &A0, &a1, A, util.PRIGHT)
util.Continue3x1to2x1(
&tT,
&tB, &t0, &tau1, Tvec, util.PBOTTOM)
}
}
/*
* Unblocked solve A*X = B for Bunch-Kauffman factorized symmetric real matrix.
*/
func unblkSolveBKUpper(B, A *cmat.FloatMatrix, p Pivots, phase int, conf *gomas.Config) *gomas.Error {
var err *gomas.Error = nil
var ATL, ATR, ABL, ABR cmat.FloatMatrix
var A00, a01, A02, a11, a12t, A22 cmat.FloatMatrix
var Aref *cmat.FloatMatrix
var BT, BB, B0, b1, B2, Bx cmat.FloatMatrix
var pT, pB, p0, p1, p2 Pivots
var aStart, aDir, bStart, bDir util.Direction
var nc int
np := 0
if phase == 2 {
aStart = util.PTOPLEFT
aDir = util.PBOTTOMRIGHT
bStart = util.PTOP
bDir = util.PBOTTOM
nc = 1
Aref = &ABR
} else {
aStart = util.PBOTTOMRIGHT
aDir = util.PTOPLEFT
bStart = util.PBOTTOM
bDir = util.PTOP
nc = m(A)
Aref = &ATL
}
util.Partition2x2(
&ATL, &ATR,
&ABL, &ABR, A, 0, 0, aStart)
util.Partition2x1(
&BT,
&BB, B, 0, bStart)
partitionPivot2x1(
&pT,
&pB, p, 0, bStart)
// phase 1:
// - solve U*D*X = B, overwriting B with X
// - looping from BOTTOM to TOP
// phase 1:
// - solve U*X = B, overwriting B with X
// - looping from TOP to BOTTOM
for n(Aref) > 0 {
// see if next diagonal block is 1x1 or 2x2
np = 1
if p[nc-1] < 0 {
np = 2
}
// repartition according the pivot size
util.Repartition2x2to3x3(&ATL,
&A00, &a01, &A02,
nil, &a11, &a12t,
nil, nil, &A22 /**/, A, np, aDir)
util.Repartition2x1to3x1(&BT,
&B0,
&b1,
&B2 /**/, B, np, bDir)
repartPivot2x1to3x1(&pT,
&p0,
&p1,
&p2 /**/, p, np, bDir)
// ------------------------------------------------------------
switch phase {
case 1:
// computes D.-1*(U.-1*B);
// b1 is current row, last row of BT
if np == 1 {
if p1[0] != nc {
// swap rows on top part of B
swapRows(&BT, m(&BT)-1, p1[0]-1)
}
// B0 = B0 - a01*b1
blasd.MVUpdate(&B0, &a01, &b1, -1.0)
// b1 = b1/d1
blasd.InvScale(&b1, a11.Get(0, 0))
nc -= 1
} else if np == 2 {
if p1[0] != -nc {
// swap rows on top part of B
swapRows(&BT, m(&BT)-2, -p1[0]-1)
}
b := a11.Get(0, 1)
apb := a11.Get(0, 0) / b
dpb := a11.Get(1, 1) / b
// (a/b)*(d/b)-1.0 == (a*d - b^2)/b^2
scale := apb*dpb - 1.0
scale *= b
// B0 = B0 - a01*b1
blasd.Mult(&B0, &a01, &b1, -1.0, 1.0, gomas.NONE, conf)
// b1 = a11.-1*b1.T
//(2x2 block, no subroutine for doing this in-place)
for k := 0; k < n(&b1); k++ {
s0 := b1.Get(0, k)
s1 := b1.Get(1, k)
b1.Set(0, k, (dpb*s0-s1)/scale)
b1.Set(1, k, (apb*s1-s0)/scale)
}
nc -= 2
}
case 2:
// compute X = U.-T*B
if np == 1 {
blasd.MVMult(&b1, &B0, &a01, -1.0, 1.0, gomas.TRANS)
if p1[0] != nc {
// swap rows on bottom part of B
util.Merge2x1(&Bx, &B0, &b1)
swapRows(&Bx, m(&Bx)-1, p1[0]-1)
}
nc += 1
} else if np == 2 {
blasd.Mult(&b1, &a01, &B0, -1.0, 1.0, gomas.TRANSA, conf)
if p1[0] != -nc {
// swap rows on bottom part of B
util.Merge2x1(&Bx, &B0, &b1)
swapRows(&Bx, m(&Bx)-2, -p1[0]-1)
}
nc += 2
}
}
// ------------------------------------------------------------
util.Continue3x3to2x2(
&ATL, &ATR,
&ABL, &ABR, &A00, &a11, &A22, A, aDir)
util.Continue3x1to2x1(
&BT,
&BB, &B0, &b1, B, bDir)
contPivot3x1to2x1(
&pT,
&pB, p0, p1, p, bDir)
}
return err
}
/*
* Compute unblocked bidiagonal reduction for A when M >= N
*
* Diagonal and first super/sub diagonal are overwritten with the
* upper/lower bidiagonal matrix B.
*
* This computing (1-tauq*v*v.T)*A*(1-taup*u.u.T) from left to right.
*/
func unblkReduceBidiagLeft(A, tauq, taup, W *cmat.FloatMatrix) {
var ATL, ABR cmat.FloatMatrix
var A00, a11, a12t, a21, A22 cmat.FloatMatrix
var tqT, tqB, tq0, tauq1, tq2 cmat.FloatMatrix
var tpT, tpB, tp0, taup1, tp2 cmat.FloatMatrix
var y21, z21 cmat.FloatMatrix
var v0 float64
util.Partition2x2(
&ATL, nil,
nil, &ABR, A, 0, 0, util.PTOPLEFT)
util.Partition2x1(
&tqT,
&tqB, tauq, 0, util.PTOP)
util.Partition2x1(
&tpT,
&tpB, taup, 0, util.PTOP)
for m(&ABR) > 0 && n(&ABR) > 0 {
util.Repartition2x2to3x3(&ATL,
&A00, nil, nil,
nil, &a11, &a12t,
nil, &a21, &A22, A, 1, util.PBOTTOMRIGHT)
util.Repartition2x1to3x1(&tqT,
&tq0,
&tauq1,
&tq2, tauq, 1, util.PBOTTOM)
util.Repartition2x1to3x1(&tpT,
&tp0,
&taup1,
&tp2, taup, 1, util.PBOTTOM)
// set temp vectors for this round
y21.SetBuf(n(&a12t), 1, n(&a12t), W.Data())
z21.SetBuf(m(&a21), 1, m(&a21), W.Data()[y21.Len():])
// ------------------------------------------------------
// Compute householder to zero subdiagonal entries
computeHouseholder(&a11, &a21, &tauq1)
// y21 := a12 + A22.T*a21
blasd.Axpby(&y21, &a12t, 1.0, 0.0)
blasd.MVMult(&y21, &A22, &a21, 1.0, 1.0, gomas.TRANSA)
// a12t := a12t - tauq*y21
tauqv := tauq1.Get(0, 0)
blasd.Axpy(&a12t, &y21, -tauqv)
// Compute householder to zero elements above 1st superdiagonal
computeHouseholderVec(&a12t, &taup1)
v0 = a12t.Get(0, 0)
a12t.Set(0, 0, 1.0)
taupv := taup1.Get(0, 0)
// [v == a12t, u == a21]
beta := blasd.Dot(&y21, &a12t)
// z21 := tauq*beta*u
blasd.Axpby(&z21, &a21, tauqv*beta, 0.0)
// z21 := A22*v - z21
blasd.MVMult(&z21, &A22, &a12t, 1.0, -1.0, gomas.NONE)
// A22 := A22 - tauq*u*y21
blasd.MVUpdate(&A22, &a21, &y21, -tauqv)
// A22 := A22 - taup*z21*v
blasd.MVUpdate(&A22, &z21, &a12t, -taupv)
a12t.Set(0, 0, v0)
// ------------------------------------------------------
util.Continue3x3to2x2(
&ATL, nil,
nil, &ABR, &A00, &a11, &A22, A, util.PBOTTOMRIGHT)
util.Continue3x1to2x1(
&tqT,
&tqB, &tq0, &tauq1, tauq, util.PBOTTOM)
util.Continue3x1to2x1(
&tpT,
&tpB, &tp0, &taup1, taup, util.PBOTTOM)
}
}
