// shift dst by shx cells (positive or negative) along X-axis.
// new edge value is clampL at left edge or clampR at right edge.
func ShiftX(dst, src *data.Slice, shiftX int, clampL, clampR float32) {
util.Argument(dst.NComp() == 1 && src.NComp() == 1)
util.Assert(dst.Len() == src.Len())
N := dst.Size()
cfg := make3DConf(N)
k_shiftx_async(dst.DevPtr(0), src.DevPtr(0), N[X], N[Y], N[Z], shiftX, clampL, clampR, cfg)
}
// dst += LUT[region], for vectors. Used to add terms to excitation.
func RegionAddV(dst *data.Slice, lut LUTPtrs, regions *Bytes) {
util.Argument(dst.NComp() == 3)
N := dst.Len()
cfg := make1DConf(N)
k_regionaddv_async(dst.DevPtr(X), dst.DevPtr(Y), dst.DevPtr(Z),
lut[X], lut[Y], lut[Z], regions.Ptr, N, cfg)
}
// Sets vector dst to zero where mask != 0.
func ZeroMask(dst *data.Slice, mask LUTPtr, regions *Bytes) {
N := dst.Len()
cfg := make1DConf(N)
for c := 0; c < dst.NComp(); c++ {
k_zeromask_async(dst.DevPtr(c), unsafe.Pointer(mask), regions.Ptr, N, cfg)
}
}
// Landau-Lifshitz torque with precession disabled.
// Used by engine.Relax().
func LLNoPrecess(torque, m, B *data.Slice) {
N := torque.Len()
cfg := make1DConf(N)
k_llnoprecess_async(torque.DevPtr(X), torque.DevPtr(Y), torque.DevPtr(Z),
m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z),
B.DevPtr(X), B.DevPtr(Y), B.DevPtr(Z), N, cfg)
}
// Set Bth to thermal noise (Brown).
// see temperature.cu
func SetTemperature(Bth, noise *data.Slice, temp_red LUTPtr, k2mu0_VgammaDt float64, regions *Bytes) {
util.Argument(Bth.NComp() == 1 && noise.NComp() == 1)
N := Bth.Len()
cfg := make1DConf(N)
k_settemperature_async(Bth.DevPtr(0), noise.DevPtr(0), float32(k2mu0_VgammaDt), unsafe.Pointer(temp_red),
regions.Ptr, N, cfg)
}
// m = 1 / (4 + τ²(m x H)²) [{4 - τ²(m x H)²} m - 4τ(m x m x H)]
// note: torque from LLNoPrecess has negative sign
func Minimize(m, m0, torque *data.Slice, dt float32) {
N := m.Len()
cfg := make1DConf(N)
k_minimize_async(m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z),
m0.DevPtr(X), m0.DevPtr(Y), m0.DevPtr(Z),
torque.DevPtr(X), torque.DevPtr(Y), torque.DevPtr(Z),
dt, N, cfg)
}
// select the part of src within the specified region, set 0's everywhere else.
func RegionSelect(dst, src *data.Slice, regions *Bytes, region byte) {
util.Argument(dst.NComp() == src.NComp())
N := dst.Len()
cfg := make1DConf(N)
for c := 0; c < dst.NComp(); c++ {
k_regionselect_async(dst.DevPtr(c), src.DevPtr(c), regions.Ptr, region, N, cfg)
}
}
// Landau-Lifshitz torque divided by gamma0:
// - 1/(1+α²) [ m x B + α m x (m x B) ]
// torque in Tesla
// m normalized
// B in Tesla
// see lltorque.cu
func LLTorque(torque, m, B *data.Slice, alpha LUTPtr, regions *Bytes) {
N := torque.Len()
cfg := make1DConf(N)
k_lltorque_async(torque.DevPtr(X), torque.DevPtr(Y), torque.DevPtr(Z),
m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z),
B.DevPtr(X), B.DevPtr(Y), B.DevPtr(Z),
unsafe.Pointer(alpha), regions.Ptr, N, cfg)
}
// Landau-Lifshitz torque divided by gamma0:
// - 1/(1+α²) [ m x B + α m x (m x B) ]
// torque in Tesla
// m normalized
// B in Tesla
// see lltorque.cu
func LLTorque(torque, m, B *data.Slice, alpha MSlice) {
N := torque.Len()
cfg := make1DConf(N)
k_lltorque2_async(torque.DevPtr(X), torque.DevPtr(Y), torque.DevPtr(Z),
m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z),
B.DevPtr(X), B.DevPtr(Y), B.DevPtr(Z),
alpha.DevPtr(0), alpha.Mul(0), N, cfg)
}
// Copies src (larger) into dst (smaller).
// Used to extract demag field after convolution on padded m.
func copyUnPad(dst, src *data.Slice, dstsize, srcsize [3]int) {
util.Argument(dst.NComp() == 1 && src.NComp() == 1)
util.Argument(dst.Len() == prod(dstsize) && src.Len() == prod(srcsize))
cfg := make3DConf(dstsize)
k_copyunpad_async(dst.DevPtr(0), dstsize[X], dstsize[Y], dstsize[Z],
src.DevPtr(0), srcsize[X], srcsize[Y], srcsize[Z], cfg)
}
// Copies src into dst, which is larger, and multiplies by vol*Bsat.
// The remainder of dst is not filled with zeros.
// Used to zero-pad magnetization before convolution and in the meanwhile multiply m by its length.
func copyPadMul(dst, src, vol *data.Slice, dstsize, srcsize [3]int, Bsat LUTPtr, regions *Bytes) {
util.Argument(dst.NComp() == 1 && src.NComp() == 1)
util.Assert(dst.Len() == prod(dstsize) && src.Len() == prod(srcsize))
cfg := make3DConf(srcsize)
k_copypadmul_async(dst.DevPtr(0), dstsize[X], dstsize[Y], dstsize[Z],
src.DevPtr(0), vol.DevPtr(0), srcsize[X], srcsize[Y], srcsize[Z],
unsafe.Pointer(Bsat), regions.Ptr, cfg)
}
// Dot product.
func Dot(a, b *data.Slice) float32 {
nComp := a.NComp()
util.Argument(nComp == b.NComp())
out := reduceBuf(0)
// not async over components
for c := 0; c < nComp; c++ {
k_reducedot_async(a.DevPtr(c), b.DevPtr(c), out, 0, a.Len(), reducecfg) // all components add to out
}
return copyback(out)
}
// Copies src into dst, which is larger, and multiplies by vol*Bsat.
// The remainder of dst is not filled with zeros.
// Used to zero-pad magnetization before convolution and in the meanwhile multiply m by its length.
func copyPadMul(dst, src, vol *data.Slice, dstsize, srcsize [3]int, Msat MSlice) {
util.Argument(dst.NComp() == 1 && src.NComp() == 1)
util.Assert(dst.Len() == prod(dstsize) && src.Len() == prod(srcsize))
cfg := make3DConf(srcsize)
k_copypadmul2_async(dst.DevPtr(0), dstsize[X], dstsize[Y], dstsize[Z],
src.DevPtr(0), srcsize[X], srcsize[Y], srcsize[Z],
Msat.DevPtr(0), Msat.Mul(0), vol.DevPtr(0), cfg)
}
// dst += prefactor * dot(a, b), as used for energy density
func AddDotProduct(dst *data.Slice, prefactor float32, a, b *data.Slice) {
util.Argument(dst.NComp() == 1 && a.NComp() == 3 && b.NComp() == 3)
util.Argument(dst.Len() == a.Len() && dst.Len() == b.Len())
N := dst.Len()
cfg := make1DConf(N)
k_dotproduct_async(dst.DevPtr(0), prefactor,
a.DevPtr(X), a.DevPtr(Y), a.DevPtr(Z),
b.DevPtr(X), b.DevPtr(Y), b.DevPtr(Z),
N, cfg)
}
// Add uniaxial magnetocrystalline anisotropy field to Beff.
// see uniaxialanisotropy.cu
func AddUniaxialAnisotropy(Beff, m *data.Slice, k1_red, k2_red LUTPtr, u LUTPtrs, regions *Bytes) {
util.Argument(Beff.Size() == m.Size())
N := Beff.Len()
cfg := make1DConf(N)
k_adduniaxialanisotropy_async(Beff.DevPtr(X), Beff.DevPtr(Y), Beff.DevPtr(Z),
m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z),
unsafe.Pointer(k1_red), unsafe.Pointer(k2_red),
u[X], u[Y], u[Z],
regions.Ptr, N, cfg)
}
// Set Bth to thermal noise (Brown).
// see temperature.cu
func SetTemperature(Bth, noise *data.Slice, k2mu0_Mu0VgammaDt float64, Msat, Temp, Alpha MSlice) {
util.Argument(Bth.NComp() == 1 && noise.NComp() == 1)
N := Bth.Len()
cfg := make1DConf(N)
k_settemperature2_async(Bth.DevPtr(0), noise.DevPtr(0), float32(k2mu0_Mu0VgammaDt),
Msat.DevPtr(0), Msat.Mul(0),
Temp.DevPtr(0), Temp.Mul(0),
Alpha.DevPtr(0), Alpha.Mul(0),
N, cfg)
}
// Add Slonczewski ST torque to torque (Tesla).
// see slonczewski.cu
func AddSlonczewskiTorque(torque, m, J, fixedP *data.Slice, Msat, alpha, pol, λ, ε_prime LUTPtr, regions *Bytes, mesh *data.Mesh) {
N := torque.Len()
cfg := make1DConf(N)
thickness := float32(mesh.WorldSize()[Z])
k_addslonczewskitorque_async(torque.DevPtr(X), torque.DevPtr(Y), torque.DevPtr(Z),
m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z), J.DevPtr(Z),
fixedP.DevPtr(X), fixedP.DevPtr(Y), fixedP.DevPtr(Z),
unsafe.Pointer(Msat), unsafe.Pointer(alpha),
thickness, unsafe.Pointer(pol),
unsafe.Pointer(λ), unsafe.Pointer(ε_prime),
regions.Ptr, N, cfg)
}
// Memset sets the Slice's components to the specified values.
// To be carefully used on unified slice (need sync)
func Memset(s *data.Slice, val ...float32) {
if Synchronous { // debug
Sync()
timer.Start("memset")
}
util.Argument(len(val) == s.NComp())
for c, v := range val {
cu.MemsetD32Async(cu.DevicePtr(uintptr(s.DevPtr(c))), math.Float32bits(v), int64(s.Len()), stream0)
}
if Synchronous { //debug
Sync()
timer.Stop("memset")
}
}
// Adds cubic anisotropy field to Beff.
// see cubicanisotropy.cu
func AddCubicAnisotropy(Beff, m *data.Slice, k1_red, k2_red, k3_red LUTPtr, c1, c2 LUTPtrs, regions *Bytes) {
util.Argument(Beff.Size() == m.Size())
N := Beff.Len()
cfg := make1DConf(N)
k_addcubicanisotropy_async(
Beff.DevPtr(X), Beff.DevPtr(Y), Beff.DevPtr(Z),
m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z),
unsafe.Pointer(k1_red), unsafe.Pointer(k2_red), unsafe.Pointer(k3_red),
c1[X], c1[Y], c1[Z],
c2[X], c2[Y], c2[Z],
regions.Ptr, N, cfg)
}
// Returns a buffer obtained from GetBuffer to the pool.
func Recycle(s *data.Slice) {
if Synchronous {
Sync()
}
N := s.Len()
pool := buf_pool[N]
// put each component buffer back on the stack
for i := 0; i < s.NComp(); i++ {
ptr := s.DevPtr(i)
if _, ok := buf_check[ptr]; !ok {
log.Panic("recyle: was not obtained with getbuffer")
}
pool = append(pool, ptr)
}
s.Disable() // make it unusable, protect against accidental use after recycle
buf_pool[N] = pool
}
// multiply-add: dst[i] = src1[i] * factor1 + src2[i] * factor2 + src3 * factor3
func Madd3(dst, src1, src2, src3 *data.Slice, factor1, factor2, factor3 float32) {
N := dst.Len()
nComp := dst.NComp()
util.Assert(src1.Len() == N && src2.Len() == N && src3.Len() == N)
util.Assert(src1.NComp() == nComp && src2.NComp() == nComp && src3.NComp() == nComp)
cfg := make1DConf(N)
for c := 0; c < nComp; c++ {
k_madd3_async(dst.DevPtr(c), src1.DevPtr(c), factor1,
src2.DevPtr(c), factor2, src3.DevPtr(c), factor3, N, cfg)
}
}
// Add uniaxial magnetocrystalline anisotropy field to Beff.
// see uniaxialanisotropy.cu
func AddUniaxialAnisotropy2(Beff, m *data.Slice, Msat, k1, k2, u MSlice) {
util.Argument(Beff.Size() == m.Size())
checkSize(Beff, m, k1, k2, u, Msat)
N := Beff.Len()
cfg := make1DConf(N)
k_adduniaxialanisotropy2_async(
Beff.DevPtr(X), Beff.DevPtr(Y), Beff.DevPtr(Z),
m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z),
Msat.DevPtr(0), Msat.Mul(0),
k1.DevPtr(0), k1.Mul(0),
k2.DevPtr(0), k2.Mul(0),
u.DevPtr(X), u.Mul(X),
u.DevPtr(Y), u.Mul(Y),
u.DevPtr(Z), u.Mul(Z),
N, cfg)
}
// Add uniaxial magnetocrystalline anisotropy field to Beff.
// see uniaxialanisotropy.cu
func AddCubicAnisotropy2(Beff, m *data.Slice, Msat, k1, k2, k3, c1, c2 MSlice) {
util.Argument(Beff.Size() == m.Size())
N := Beff.Len()
cfg := make1DConf(N)
k_addcubicanisotropy2_async(
Beff.DevPtr(X), Beff.DevPtr(Y), Beff.DevPtr(Z),
m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z),
Msat.DevPtr(0), Msat.Mul(0),
k1.DevPtr(0), k1.Mul(0),
k2.DevPtr(0), k2.Mul(0),
k3.DevPtr(0), k3.Mul(0),
c1.DevPtr(X), c1.Mul(X),
c1.DevPtr(Y), c1.Mul(Y),
c1.DevPtr(Z), c1.Mul(Z),
c2.DevPtr(X), c2.Mul(X),
c2.DevPtr(Y), c2.Mul(Y),
c2.DevPtr(Z), c2.Mul(Z),
N, cfg)
}
// Add Slonczewski ST torque to torque (Tesla).
// see slonczewski.cu
func AddSlonczewskiTorque2(torque, m *data.Slice, Msat, J, fixedP, alpha, pol, λ, ε_prime MSlice, mesh *data.Mesh) {
N := torque.Len()
cfg := make1DConf(N)
flt := float32(mesh.WorldSize()[Z])
k_addslonczewskitorque2_async(
torque.DevPtr(X), torque.DevPtr(Y), torque.DevPtr(Z),
m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z),
Msat.DevPtr(0), Msat.Mul(0),
J.DevPtr(Z), J.Mul(Z),
fixedP.DevPtr(X), fixedP.Mul(X),
fixedP.DevPtr(Y), fixedP.Mul(Y),
fixedP.DevPtr(Z), fixedP.Mul(Z),
alpha.DevPtr(0), alpha.Mul(0),
pol.DevPtr(0), pol.Mul(0),
λ.DevPtr(0), λ.Mul(0),
ε_prime.DevPtr(0), ε_prime.Mul(0),
unsafe.Pointer(uintptr(0)), flt,
N, cfg)
}
// Maximum of the norms of the difference between all vectors (x1,y1,z1) and (x2,y2,z2)
// (dx, dy, dz) = (x1, y1, z1) - (x2, y2, z2)
// max_i sqrt( dx[i]*dx[i] + dy[i]*dy[i] + dz[i]*dz[i] )
func MaxVecDiff(x, y *data.Slice) float64 {
util.Argument(x.Len() == y.Len())
out := reduceBuf(0)
k_reducemaxvecdiff2_async(x.DevPtr(0), x.DevPtr(1), x.DevPtr(2),
y.DevPtr(0), y.DevPtr(1), y.DevPtr(2),
out, 0, x.Len(), reducecfg)
return math.Sqrt(float64(copyback(out)))
}
// multiply: dst[i] = a[i] * b[i]
// a and b must have the same number of components
func Mul(dst, a, b *data.Slice) {
N := dst.Len()
nComp := dst.NComp()
util.Assert(a.Len() == N && a.NComp() == nComp && b.Len() == N && b.NComp() == nComp)
cfg := make1DConf(N)
for c := 0; c < nComp; c++ {
k_mul_async(dst.DevPtr(c), a.DevPtr(c), b.DevPtr(c), N, cfg)
}
}
// Execute the FFT plan, asynchronous.
// src and dst are 3D arrays stored 1D arrays.
func (p *fft3DC2RPlan) ExecAsync(src, dst *data.Slice) {
if Synchronous {
Sync()
timer.Start("fft")
}
oksrclen := p.InputLenFloats()
if src.Len() != oksrclen {
panic(fmt.Errorf("fft size mismatch: expecting src len %v, got %v", oksrclen, src.Len()))
}
okdstlen := p.OutputLenFloats()
if dst.Len() != okdstlen {
panic(fmt.Errorf("fft size mismatch: expecting dst len %v, got %v", okdstlen, dst.Len()))
}
p.handle.ExecC2R(cu.DevicePtr(uintptr(src.DevPtr(0))), cu.DevicePtr(uintptr(dst.DevPtr(0))))
if Synchronous {
Sync()
timer.Stop("fft")
}
}
// Execute the FFT plan, asynchronous.
// src and dst are 3D arrays stored 1D arrays.
func (p *fft3DR2CPlan) ExecAsync(src, dst *data.Slice) {
if Synchronous {
Sync()
timer.Start("fft")
}
util.Argument(src.NComp() == 1 && dst.NComp() == 1)
oksrclen := p.InputLen()
if src.Len() != oksrclen {
log.Panicf("fft size mismatch: expecting src len %v, got %v", oksrclen, src.Len())
}
okdstlen := p.OutputLen()
if dst.Len() != okdstlen {
log.Panicf("fft size mismatch: expecting dst len %v, got %v", okdstlen, dst.Len())
}
p.handle.ExecR2C(cu.DevicePtr(uintptr(src.DevPtr(0))), cu.DevicePtr(uintptr(dst.DevPtr(0))))
if Synchronous {
Sync()
timer.Stop("fft")
}
}
// Extract real parts, copy them from src to dst.
// In the meanwhile, check if imaginary parts are nearly zero
// and scale the kernel to compensate for unnormalized FFTs.
// scale = 1/N, with N the FFT logical size.
func scaleRealParts(dst, src *data.Slice, scale float32) {
util.Argument(2*dst.Len() == src.Len())
util.Argument(dst.NComp() == 1 && src.NComp() == 1)
srcList := src.Host()[0]
dstList := dst.Host()[0]
// Normally, the FFT'ed kernel is purely real because of symmetry,
// so we only store the real parts...
maximg := float32(0.)
for i := 0; i < src.Len()/2; i++ {
dstList[i] = srcList[2*i] * scale
if fabs(srcList[2*i+1]) > maximg {
maximg = fabs(srcList[2*i+1])
}
}
maximg *= float32(math.Sqrt(float64(scale))) // after 1 FFT, normalization is sqrt(N)
// ...however, we check that the imaginary parts are nearly zero,
// just to be sure we did not make a mistake during kernel creation.
if maximg > FFT_IMAG_TOLERANCE {
log.Fatalf("FFT kernel imaginary part: %v\n", maximg)
}
}
// Normalize vec to unit length, unless length or vol are zero.
func Normalize(vec, vol *data.Slice) {
util.Argument(vol == nil || vol.NComp() == 1)
N := vec.Len()
cfg := make1DConf(N)
k_normalize_async(vec.DevPtr(X), vec.DevPtr(Y), vec.DevPtr(Z), vol.DevPtr(0), N, cfg)
}