// 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)
}
// kernel multiplication for 2D demag convolution on X and Y, exploiting full kernel symmetry.
func kernMulRSymm2Dxy_async(fftMx, fftMy, Kxx, Kyy, Kxy *data.Slice, Nx, Ny int) {
util.Argument(fftMy.NComp() == 1 && Kxx.NComp() == 1)
cfg := make3DConf([3]int{Nx, Ny, 1})
k_kernmulRSymm2Dxy_async(fftMx.DevPtr(0), fftMy.DevPtr(0),
Kxx.DevPtr(0), Kyy.DevPtr(0), Kxy.DevPtr(0),
Nx, Ny, 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)
}
}
// 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)
}
// 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)
}
}
// 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)
}
// 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)
}
// Finds the average exchange strength around each cell, for debugging.
func ExchangeDecode(dst *data.Slice, Aex_red SymmLUT, regions *Bytes, mesh *data.Mesh) {
c := mesh.CellSize()
wx := float32(2 * 1e-18 / (c[X] * c[X]))
wy := float32(2 * 1e-18 / (c[Y] * c[Y]))
wz := float32(2 * 1e-18 / (c[Z] * c[Z]))
N := mesh.Size()
pbc := mesh.PBC_code()
cfg := make3DConf(N)
k_exchangedecode_async(dst.DevPtr(0), unsafe.Pointer(Aex_red), regions.Ptr, wx, wy, wz, N[X], N[Y], N[Z], pbc, 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)
}
// 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")
}
}
// Crop stores in dst a rectangle cropped from src at given offset position.
// dst size may be smaller than src.
func Crop(dst, src *data.Slice, offX, offY, offZ int) {
D := dst.Size()
S := src.Size()
util.Argument(dst.NComp() == src.NComp())
util.Argument(D[X]+offX <= S[X] && D[Y]+offY <= S[Y] && D[Z]+offZ <= S[Z])
cfg := make3DConf(D)
for c := 0; c < dst.NComp(); c++ {
k_crop_async(dst.DevPtr(c), D[X], D[Y], D[Z],
src.DevPtr(c), S[X], S[Y], S[Z],
offX, offY, offZ, cfg)
}
}
// Select and resize one layer for interactive output
func Resize(dst, src *data.Slice, layer int) {
dstsize := dst.Size()
srcsize := src.Size()
util.Assert(dstsize[Z] == 1)
util.Assert(dst.NComp() == 1 && src.NComp() == 1)
scalex := srcsize[X] / dstsize[X]
scaley := srcsize[Y] / dstsize[Y]
util.Assert(scalex > 0 && scaley > 0)
cfg := make3DConf(dstsize)
k_resize_async(dst.DevPtr(0), dstsize[X], dstsize[Y], dstsize[Z],
src.DevPtr(0), srcsize[X], srcsize[Y], srcsize[Z], layer, scalex, scaley, cfg)
}
// SetMaxAngle sets dst to the maximum angle of each cells magnetization with all of its neighbors,
// provided the exchange stiffness with that neighbor is nonzero.
func SetMaxAngle(dst, m *data.Slice, Aex_red SymmLUT, regions *Bytes, mesh *data.Mesh) {
N := mesh.Size()
pbc := mesh.PBC_code()
cfg := make3DConf(N)
k_setmaxangle_async(dst.DevPtr(0),
m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z),
unsafe.Pointer(Aex_red), regions.Ptr,
N[X], N[Y], N[Z], pbc, 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
}
// Set s to the toplogogical charge density s = m · (m/∂x ❌ ∂m/∂y)
// See topologicalcharge.cu
func SetTopologicalCharge(s *data.Slice, m *data.Slice, mesh *data.Mesh) {
cellsize := mesh.CellSize()
N := s.Size()
util.Argument(m.Size() == N)
cfg := make3DConf(N)
icxcy := float32(1.0 / (cellsize[X] * cellsize[Y]))
k_settopologicalcharge_async(s.DevPtr(X),
m.DevPtr(X), m.DevPtr(Y), m.DevPtr(Z),
icxcy, N[X], N[Y], N[Z], mesh.PBC_code(), cfg)
}
// 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)
}
}
// 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")
}
}
// 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)
}
// 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)
}
}
// 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)
}
// 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)
}
// Add Zhang-Li ST torque (Tesla) to torque.
// see zhangli.cu
func AddZhangLiTorque(torque, m *data.Slice, Msat, J, alpha, xi, pol MSlice, mesh *data.Mesh) {
c := mesh.CellSize()
N := mesh.Size()
cfg := make3DConf(N)
k_addzhanglitorque2_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(X), J.Mul(X),
J.DevPtr(Y), J.Mul(Y),
J.DevPtr(Z), J.Mul(Z),
alpha.DevPtr(0), alpha.Mul(0),
xi.DevPtr(0), xi.Mul(0),
pol.DevPtr(0), pol.Mul(0),
float32(c[X]), float32(c[Y]), float32(c[Z]),
N[X], N[Y], N[Z], mesh.PBC_code(), cfg)
}
// 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)
}
// kernel multiplication for general 1D convolution. Does not assume any symmetry.
// Used for MFM images.
func kernMulC_async(fftM, K *data.Slice, Nx, Ny int) {
util.Argument(fftM.NComp() == 1 && K.NComp() == 1)
cfg := make3DConf([3]int{Nx, Ny, 1})
k_kernmulC_async(fftM.DevPtr(0), K.DevPtr(0), Nx, Ny, cfg)
}
// kernel multiplication for 2D demag convolution on Z, exploiting full kernel symmetry.
func kernMulRSymm2Dz_async(fftMz, Kzz *data.Slice, Nx, Ny int) {
util.Argument(fftMz.NComp() == 1 && Kzz.NComp() == 1)
cfg := make3DConf([3]int{Nx, Ny, 1})
k_kernmulRSymm2Dz_async(fftMz.DevPtr(0), Kzz.DevPtr(0), Nx, Ny, cfg)
}
// kernel multiplication for 3D demag convolution, exploiting full kernel symmetry.
func kernMulRSymm3D_async(fftM [3]*data.Slice, Kxx, Kyy, Kzz, Kyz, Kxz, Kxy *data.Slice, Nx, Ny, Nz int) {
util.Argument(fftM[X].NComp() == 1 && Kxx.NComp() == 1)
cfg := make3DConf([3]int{Nx, Ny, Nz})
k_kernmulRSymm3D_async(fftM[X].DevPtr(0), fftM[Y].DevPtr(0), fftM[Z].DevPtr(0),
Kxx.DevPtr(0), Kyy.DevPtr(0), Kzz.DevPtr(0), Kyz.DevPtr(0), Kxz.DevPtr(0), Kxy.DevPtr(0),
Nx, Ny, Nz, cfg)
}
// zero 1-component slice
func zero1_async(dst *data.Slice) {
cu.MemsetD32Async(cu.DevicePtr(uintptr(dst.DevPtr(0))), 0, int64(dst.Len()), stream0)
}
// 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 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)
}