// Gets
func GetMaxTorque() float64 {
torque, recycle := Torque.Slice()
if recycle {
defer cuda.Recycle(torque)
}
return cuda.MaxVecNorm(torque)
}
func (mini *Minimizer) Step() {
m := M.Buffer()
size := m.Size()
k := mini.k
h := mini.h
// save original magnetization
m0 := cuda.Buffer(3, size)
defer cuda.Recycle(m0)
data.Copy(m0, m)
// make descent
cuda.Minimize(m, m0, k, h)
// calculate new torque for next step
k0 := cuda.Buffer(3, size)
defer cuda.Recycle(k0)
data.Copy(k0, k)
torqueFn(k)
setMaxTorque(k) // report to user
// just to make the following readable
dm := m0
dk := k0
// calculate step difference of m and k
cuda.Madd2(dm, m, m0, 1., -1.)
cuda.Madd2(dk, k, k0, -1., 1.) // reversed due to LLNoPrecess sign
// get maxdiff and add to list
max_dm := cuda.MaxVecNorm(dm)
mini.lastDm.Add(max_dm)
setLastErr(mini.lastDm.Max()) // report maxDm to user as LastErr
// adjust next time step
var nom, div float32
if NSteps%2 == 0 {
nom = cuda.Dot(dm, dm)
div = cuda.Dot(dm, dk)
} else {
nom = cuda.Dot(dm, dk)
div = cuda.Dot(dk, dk)
}
if div != 0. {
mini.h = nom / div
} else { // in case of division by zero
mini.h = 1e-4
}
M.normalize()
// as a convention, time does not advance during relax
NSteps++
}
func setMaxTorque(τ *data.Slice) {
LastTorque = cuda.MaxVecNorm(τ)
}
func (rk *RK23) Step() {
m := M.Buffer()
size := m.Size()
if FixDt != 0 {
Dt_si = FixDt
}
// upon resize: remove wrongly sized k1
if rk.k1.Size() != m.Size() {
rk.Free()
}
// first step ever: one-time k1 init and eval
if rk.k1 == nil {
rk.k1 = cuda.NewSlice(3, size)
torqueFn(rk.k1)
}
// FSAL cannot be used with temperature
if !Temp.isZero() {
torqueFn(rk.k1)
}
t0 := Time
// backup magnetization
m0 := cuda.Buffer(3, size)
defer cuda.Recycle(m0)
data.Copy(m0, m)
k2, k3, k4 := cuda.Buffer(3, size), cuda.Buffer(3, size), cuda.Buffer(3, size)
defer cuda.Recycle(k2)
defer cuda.Recycle(k3)
defer cuda.Recycle(k4)
h := float32(Dt_si * GammaLL) // internal time step = Dt * gammaLL
// there is no explicit stage 1: k1 from previous step
// stage 2
Time = t0 + (1./2.)*Dt_si
cuda.Madd2(m, m, rk.k1, 1, (1./2.)*h) // m = m*1 + k1*h/2
M.normalize()
torqueFn(k2)
// stage 3
Time = t0 + (3./4.)*Dt_si
cuda.Madd2(m, m0, k2, 1, (3./4.)*h) // m = m0*1 + k2*3/4
M.normalize()
torqueFn(k3)
// 3rd order solution
madd4(m, m0, rk.k1, k2, k3, 1, (2./9.)*h, (1./3.)*h, (4./9.)*h)
M.normalize()
// error estimate
Time = t0 + Dt_si
torqueFn(k4)
Err := k2 // re-use k2 as error
// difference of 3rd and 2nd order torque without explicitly storing them first
madd4(Err, rk.k1, k2, k3, k4, (7./24.)-(2./9.), (1./4.)-(1./3.), (1./3.)-(4./9.), (1. / 8.))
// determine error
err := cuda.MaxVecNorm(Err) * float64(h)
// adjust next time step
if err < MaxErr || Dt_si <= MinDt || FixDt != 0 { // mindt check to avoid infinite loop
// step OK
setLastErr(err)
setMaxTorque(k4)
NSteps++
Time = t0 + Dt_si
adaptDt(math.Pow(MaxErr/err, 1./3.))
data.Copy(rk.k1, k4) // FSAL
} else {
// undo bad step
//util.Println("Bad step at t=", t0, ", err=", err)
util.Assert(FixDt == 0)
Time = t0
data.Copy(m, m0)
NUndone++
adaptDt(math.Pow(MaxErr/err, 1./4.))
}
}
func (rk *RK45DP) Step() {
m := M.Buffer()
size := m.Size()
if FixDt != 0 {
Dt_si = FixDt
}
// upon resize: remove wrongly sized k1
if rk.k1.Size() != m.Size() {
rk.Free()
}
// first step ever: one-time k1 init and eval
if rk.k1 == nil {
rk.k1 = cuda.NewSlice(3, size)
torqueFn(rk.k1)
}
// FSAL cannot be used with finite temperature
if !Temp.isZero() {
torqueFn(rk.k1)
}
t0 := Time
// backup magnetization
m0 := cuda.Buffer(3, size)
defer cuda.Recycle(m0)
data.Copy(m0, m)
k2, k3, k4, k5, k6 := cuda.Buffer(3, size), cuda.Buffer(3, size), cuda.Buffer(3, size), cuda.Buffer(3, size), cuda.Buffer(3, size)
defer cuda.Recycle(k2)
defer cuda.Recycle(k3)
defer cuda.Recycle(k4)
defer cuda.Recycle(k5)
defer cuda.Recycle(k6)
// k2 will be re-used as k7
h := float32(Dt_si * GammaLL) // internal time step = Dt * gammaLL
// there is no explicit stage 1: k1 from previous step
// stage 2
Time = t0 + (1./5.)*Dt_si
cuda.Madd2(m, m, rk.k1, 1, (1./5.)*h) // m = m*1 + k1*h/5
M.normalize()
torqueFn(k2)
// stage 3
Time = t0 + (3./10.)*Dt_si
cuda.Madd3(m, m0, rk.k1, k2, 1, (3./40.)*h, (9./40.)*h)
M.normalize()
torqueFn(k3)
// stage 4
Time = t0 + (4./5.)*Dt_si
madd4(m, m0, rk.k1, k2, k3, 1, (44./45.)*h, (-56./15.)*h, (32./9.)*h)
M.normalize()
torqueFn(k4)
// stage 5
Time = t0 + (8./9.)*Dt_si
madd5(m, m0, rk.k1, k2, k3, k4, 1, (19372./6561.)*h, (-25360./2187.)*h, (64448./6561.)*h, (-212./729.)*h)
M.normalize()
torqueFn(k5)
// stage 6
Time = t0 + (1.)*Dt_si
madd6(m, m0, rk.k1, k2, k3, k4, k5, 1, (9017./3168.)*h, (-355./33.)*h, (46732./5247.)*h, (49./176.)*h, (-5103./18656.)*h)
M.normalize()
torqueFn(k6)
// stage 7: 5th order solution
Time = t0 + (1.)*Dt_si
// no k2
madd6(m, m0, rk.k1, k3, k4, k5, k6, 1, (35./384.)*h, (500./1113.)*h, (125./192.)*h, (-2187./6784.)*h, (11./84.)*h) // 5th
M.normalize()
k7 := k2 // re-use k2
torqueFn(k7) // next torque if OK
// error estimate
Err := cuda.Buffer(3, size) //k3 // re-use k3 as error estimate
defer cuda.Recycle(Err)
madd6(Err, rk.k1, k3, k4, k5, k6, k7, (35./384.)-(5179./57600.), (500./1113.)-(7571./16695.), (125./192.)-(393./640.), (-2187./6784.)-(-92097./339200.), (11./84.)-(187./2100.), (0.)-(1./40.))
// determine error
err := cuda.MaxVecNorm(Err) * float64(h)
// adjust next time step
if err < MaxErr || Dt_si <= MinDt || FixDt != 0 { // mindt check to avoid infinite loop
// step OK
setLastErr(err)
setMaxTorque(k7)
NSteps++
Time = t0 + Dt_si
adaptDt(math.Pow(MaxErr/err, 1./5.))
data.Copy(rk.k1, k7) // FSAL
} else {
// undo bad step
//util.Println("Bad step at t=", t0, ", err=", err)
util.Assert(FixDt == 0)
Time = t0
data.Copy(m, m0)
NUndone++
adaptDt(math.Pow(MaxErr/err, 1./6.))
}
}