func Mult0(C, A, B *matrix.FloatMatrix, alpha, beta float64, flags Flags) error {
if A.Cols() != B.Rows() {
return errors.New("A.cols != B.rows: size mismatch")
}
psize := int64(C.NumElements()) * int64(A.Cols())
Ar := A.FloatArray()
ldA := A.LeadingIndex()
Br := B.FloatArray()
ldB := B.LeadingIndex()
Cr := C.FloatArray()
ldC := C.LeadingIndex()
if nWorker <= 1 || psize <= limitOne {
calgo.DMult0(Cr, Ar, Br, alpha, beta, calgo.Flags(flags), ldC, ldA, ldB, B.Rows(),
0, C.Cols(), 0, C.Rows(),
vpLen, nB, mB)
return nil
}
// here we have more than one worker available
worker := func(cstart, cend, rstart, rend int, ready chan int) {
calgo.DMult0(Cr, Ar, Br, alpha, beta, calgo.Flags(flags), ldC, ldA, ldB, B.Rows(),
cstart, cend, rstart, rend, vpLen, nB, mB)
ready <- 1
}
colworks, rowworks := divideWork(C.Rows(), C.Cols(), nWorker)
scheduleWork(colworks, rowworks, C.Cols(), C.Rows(), worker)
return nil
}
// Calculate C = alpha*A*B.T + beta*C, C is M*N, A is M*P and B is N*P
func MMMultTransB(C, A, B *matrix.FloatMatrix, alpha, beta float64) error {
psize := int64(C.NumElements() * A.Cols())
Ar := A.FloatArray()
ldA := A.LeadingIndex()
Br := B.FloatArray()
ldB := B.LeadingIndex()
Cr := C.FloatArray()
ldC := C.LeadingIndex()
if nWorker <= 1 || psize <= limitOne {
calgo.DMult(Cr, Ar, Br, alpha, beta, calgo.TRANSB, ldC, ldA, ldB,
B.Rows(), 0, C.Cols(), 0, C.Rows(), vpLen, nB, mB)
return nil
}
// here we have more than one worker available
worker := func(cstart, cend, rstart, rend int, ready chan int) {
calgo.DMult(Cr, Ar, Br, alpha, beta, calgo.TRANSB, ldC, ldA, ldB, B.Rows(),
cstart, cend, rstart, rend, vpLen, nB, mB)
ready <- 1
}
colworks, rowworks := divideWork(C.Rows(), C.Cols(), nWorker)
scheduleWork(colworks, rowworks, C.Cols(), C.Rows(), worker)
//scheduleWork(colworks, rowworks, worker)
return nil
}
// Calculate C = alpha*A*B + beta*C, C is M*N, A is M*M and B is M*N
func MMSymmUpper(C, A, B *matrix.FloatMatrix, alpha, beta float64) error {
if A.Rows() != A.Cols() {
return errors.New("A matrix not square matrix.")
}
psize := int64(C.NumElements() * A.Cols())
Ar := A.FloatArray()
ldA := A.LeadingIndex()
Br := B.FloatArray()
ldB := B.LeadingIndex()
Cr := C.FloatArray()
ldC := C.LeadingIndex()
if nWorker <= 1 || psize <= limitOne {
calgo.DMultSymm(Cr, Ar, Br, alpha, beta, calgo.LEFT|calgo.UPPER, ldC, ldA, ldB,
A.Cols(), 0, C.Cols(), 0, C.Rows(), vpLen, nB, mB)
return nil
}
// here we have more than one worker available
worker := func(cstart, cend, rstart, rend int, ready chan int) {
calgo.DMultSymm(Cr, Ar, Br, alpha, beta, calgo.LEFT|calgo.UPPER, ldC, ldA, ldB,
A.Cols(), cstart, cend, rstart, rend, vpLen, nB, mB)
ready <- 1
}
colworks, rowworks := divideWork(C.Rows(), C.Cols(), nWorker)
scheduleWork(colworks, rowworks, C.Cols(), C.Rows(), worker)
return nil
}
// Generic matrix-matrix multpily. (blas.GEMM). Calculates
// C = beta*C + alpha*A*B (default)
// C = beta*C + alpha*A.T*B flags&TRANSA
// C = beta*C + alpha*A*B.T flags&TRANSB
// C = beta*C + alpha*A.T*B.T flags&(TRANSA|TRANSB)
//
// C is M*N, A is M*P or P*M if flags&TRANSA. B is P*N or N*P if flags&TRANSB.
//
func Mult(C, A, B *matrix.FloatMatrix, alpha, beta float64, flags Flags) error {
var ok, empty bool
// error checking must take in account flag values!
ar, ac := A.Size()
br, bc := B.Size()
cr, cc := C.Size()
switch flags & (TRANSA | TRANSB) {
case TRANSA | TRANSB:
empty = ac == 0 || br == 0
ok = cr == ac && cc == br && ar == bc
case TRANSA:
empty = ac == 0 || bc == 0
ok = cr == ac && cc == bc && ar == br
case TRANSB:
empty = ar == 0 || br == 0
ok = cr == ar && cc == br && ac == bc
default:
empty = ar == 0 || bc == 0
ok = cr == ar && cc == bc && ac == br
}
if empty {
return nil
}
if !ok {
return errors.New("Mult: size mismatch")
}
psize := int64(C.NumElements()) * int64(A.Cols())
Ar := A.FloatArray()
ldA := A.LeadingIndex()
Br := B.FloatArray()
ldB := B.LeadingIndex()
Cr := C.FloatArray()
ldC := C.LeadingIndex()
// matrix A, B common dimension
P := A.Cols()
if flags&TRANSA != 0 {
P = A.Rows()
}
if nWorker <= 1 || psize <= limitOne {
calgo.DMult(Cr, Ar, Br, alpha, beta, calgo.Flags(flags), ldC, ldA, ldB, P,
0, C.Cols(), 0, C.Rows(),
vpLen, nB, mB)
return nil
}
// here we have more than one worker available
worker := func(cstart, cend, rstart, rend int, ready chan int) {
calgo.DMult(Cr, Ar, Br, alpha, beta, calgo.Flags(flags), ldC, ldA, ldB, P,
cstart, cend, rstart, rend, vpLen, nB, mB)
ready <- 1
}
colworks, rowworks := divideWork(C.Rows(), C.Cols(), nWorker)
scheduleWork(colworks, rowworks, C.Cols(), C.Rows(), worker)
return nil
}
/*
Returns x' * J * y, where J = [1, 0; 0, -I].
*/
func jdot(x, y *matrix.FloatMatrix, n, offsetx, offsety int) float64 {
if n <= 0 {
n = x.NumElements()
}
a := blas.DotFloat(x, y, &la_.IOpt{"n", n - 1}, &la_.IOpt{"offsetx", offsetx + 1},
&la_.IOpt{"offsety", offsety + 1})
return x.GetIndex(offsetx)*y.GetIndex(offsety) - a
}
func (g *matrixFs) Gf(x, y *matrix.FloatMatrix, alpha, beta float64, trans linalg.Option) error {
//
minor := 0
if !checkpnt.MinorEmpty() {
minor = checkpnt.MinorTop()
} else {
loopg += 1
minor = loopg
}
checkpnt.Check("00-Gfunc", minor)
m, n := g.A.Size()
y.Scale(beta)
// x_n = x[:n]
//x_n := matrix.FloatVector(x.FloatArray()[:n])
x_n := x.SubMatrix(0, 0, n, 1).Copy()
// x_n_2n = x[n:2*n]
//x_n_2n := matrix.FloatVector(x.FloatArray()[n : 2*n])
x_n_2n := x.SubMatrix(n, 0, n, 1).Copy()
if linalg.Equal(trans, linalg.OptNoTrans) {
// y += alpha * G * x
// y[:n] += alpha * (x[:n] - x[n:2*n])
y_n := matrix.Minus(x_n, x_n_2n).Scale(alpha)
y.SubMatrix(0, 0, n, 1).Plus(y_n)
//y.AddIndexes(matrix.Indexes(n), y_n.FloatArray())
// y[n:2*n] += alpha * (-x[:n] - x[n:2*n]) = -alpha * (x[:n]+x[n:2*n])
y_n = matrix.Plus(x_n, x_n_2n).Scale(-alpha)
y.SubMatrix(n, 0, n, 1).Plus(y_n)
//y.AddIndexes(matrix.Indexes(n, 2*n), y_n.FloatArray())
// y[2*n+1:] += -alpha * A * x[:n]
y_2n := matrix.Times(g.A, x_n).Scale(-alpha)
//y.AddIndexes(matrix.Indexes(2*n+1, y.NumElements()), y_2n.FloatArray())
y.SubMatrix(2*n+1, 0, y_2n.NumElements(), 1).Plus(y_2n)
} else {
// x_m = x[-m:]
//x_m := matrix.FloatVector(x.FloatArray()[x.NumElements()-m:])
x_m := x.SubMatrix(x.NumElements()-m, 0)
// x_tmp = (x[:n] - x[n:2*n] - A.T * x[-m:])
x_tmp := matrix.Minus(x_n, x_n_2n, matrix.Times(g.A.Transpose(), x_m))
// y[:n] += alpha * (x[:n] - x[n:2*n] - A.T * x[-m:])
//y.AddIndexes(matrix.Indexes(n), x_tmp.Scale(alpha).FloatArray())
y.SubMatrix(0, 0, n, 1).Plus(x_tmp.Scale(alpha))
x_tmp = matrix.Plus(x_n, x_n_2n).Scale(-alpha)
//y.AddIndexes(matrix.Indexes(n, y.NumElements()), x_tmp.FloatArray())
y.SubMatrix(n, 0).Plus(x_tmp)
}
checkpnt.Check("10-Gfunc", minor)
return nil
}
// Symmetric matrix multiply. (blas.SYMM)
// C = beta*C + alpha*A*B (default)
// C = beta*C + alpha*A.T*B flags&TRANSA
// C = beta*C + alpha*A*B.T flags&TRANSB
// C = beta*C + alpha*A.T*B.T flags&(TRANSA|TRANSB)
//
// C is N*P, A is N*N symmetric matrix. B is N*P or P*N if flags&TRANSB.
//
func MultSym(C, A, B *matrix.FloatMatrix, alpha, beta float64, flags Flags) error {
var ok, empty bool
ar, ac := A.Size()
br, bc := B.Size()
cr, cc := C.Size()
switch flags & (TRANSA | TRANSB) {
case TRANSA | TRANSB:
empty = ac == 0 || br == 0
ok = ar == ac && cr == ac && cc == br && ar == bc
case TRANSA:
empty = ac == 0 || bc == 0
ok = ar == ac && cr == ac && cc == bc && ar == br
case TRANSB:
empty = ar == 0 || br == 0
ok = ar == ac && cr == ar && cc == br && ac == bc
default:
empty = ar == 0 || bc == 0
ok = ar == ac && cr == ar && cc == bc && ac == br
}
if empty {
return nil
}
if !ok {
return errors.New("MultSym: size mismatch")
}
/*
if A.Rows() != A.Cols() {
return errors.New("A matrix not square matrix.");
}
if A.Cols() != B.Rows() {
return errors.New("A.cols != B.rows: size mismatch")
}
*/
psize := int64(C.NumElements()) * int64(A.Cols())
Ar := A.FloatArray()
ldA := A.LeadingIndex()
Br := B.FloatArray()
ldB := B.LeadingIndex()
Cr := C.FloatArray()
ldC := C.LeadingIndex()
if nWorker <= 1 || psize <= limitOne {
calgo.DMultSymm(Cr, Ar, Br, alpha, beta, calgo.Flags(flags), ldC, ldA, ldB,
A.Cols(), 0, C.Cols(), 0, C.Rows(), vpLen, nB, mB)
return nil
}
// here we have more than one worker available
worker := func(cstart, cend, rstart, rend int, ready chan int) {
calgo.DMultSymm(Cr, Ar, Br, alpha, beta, calgo.Flags(flags), ldC, ldA, ldB,
A.Cols(), cstart, cend, rstart, rend, vpLen, nB, mB)
ready <- 1
}
colworks, rowworks := divideWork(C.Rows(), C.Cols(), nWorker)
scheduleWork(colworks, rowworks, C.Cols(), C.Rows(), worker)
return nil
}
/*
Returns sqrt(x' * J * x) where J = [1, 0; 0, -I], for a vector
x in a second order cone.
*/
func jnrm2(x *matrix.FloatMatrix, n, offset int) float64 {
/*DEBUGGED*/
if n <= 0 {
n = x.NumElements()
}
if offset < 0 {
offset = 0
}
a := blas.Nrm2Float(x, &la_.IOpt{"n", n - 1}, &la_.IOpt{"offset", offset + 1})
fst := x.GetIndex(offset)
return math.Sqrt(fst-a) * math.Sqrt(fst+a)
}
func _TestPartition2D(t *testing.T) {
var ATL, ATR, ABL, ABR, As matrix.FloatMatrix
var A00, a01, A02, a10, a11, a12, A20, a21, A22 matrix.FloatMatrix
A := matrix.FloatZeros(6, 6)
As.SubMatrixOf(A, 1, 1, 4, 4)
As.SetIndexes(1.0)
partition2x2(&ATL, &ATR, &ABL, &ABR, &As, 0)
t.Logf("ATL:\n%v\n", &ATL)
for ATL.Rows() < As.Rows() {
repartition2x2to3x3(&ATL,
&A00, &a01, &A02,
&a10, &a11, &a12,
&A20, &a21, &A22, &As, 1)
t.Logf("m(a12)=%d [%d], m(a11)=%d\n", a12.Cols(), a12.NumElements(), a11.NumElements())
a11.Add(1.0)
a21.Add(-2.0)
continue3x3to2x2(&ATL, &ATR, &ABL, &ABR, &A00, &a11, &A22, &As)
}
t.Logf("A:\n%v\n", A)
}
/*
* Generic rank update of diagonal matrix.
* diag(D) = diag(D) + alpha * x * y.T
*
* Arguments:
* D N element column or row vector or N-by-N matrix
*
* x, y N element vectors
*
* alpha scalar
*/
func MVUpdateDiag(D, x, y *matrix.FloatMatrix, alpha float64) error {
var d *matrix.FloatMatrix
var dvec matrix.FloatMatrix
if !isVector(x) || !isVector(y) {
return errors.New("x, y not vectors")
}
if D.Rows() > 0 && D.Cols() == D.Rows() {
D.Diag(&dvec)
d = &dvec
} else if isVector(D) {
d = D
} else {
return errors.New("D not a diagonal")
}
N := d.NumElements()
for k := 0; k < N; k++ {
val := d.GetIndex(k)
val += x.GetIndex(k) * y.GetIndex(k) * alpha
d.SetIndex(k, val)
}
return nil
}