void rk_combine_result(
double h, ButcherTableau &tab,
const VectorMatrix &k0, const VectorMatrix &k1, const VectorMatrix &k2,
const VectorMatrix &k3, const VectorMatrix &k4, const VectorMatrix &k5,
VectorMatrix &y, VectorMatrix &y_error)
{
const int s = y.size();
if ( s != y_error.size()
|| s != k1.size()
|| s != k2.size()
|| s != k3.size()
|| s != k4.size()
|| s != k5.size()) throw std::runtime_error("rk_combine_result: Input matrix size mismatch.");
if (!tab.num_steps == 6) throw std::runtime_error("Need num_steps == 6 in rk_combine_result");
if (isCudaEnabled()) {
#ifdef HAVE_CUDA
rk_combine_result_cuda(h, tab, k0, k1, k2, k3, k4, k5, y, y_error, isCuda64Enabled());
#else
assert(0);
#endif
} else {
rk_combine_result_cpu(h, tab, k0, k1, k2, k3, k4, k5, y, y_error);
}
}
double cubic_anisotropy(
const VectorMatrix &axis1,
const VectorMatrix &axis2,
const Matrix &k,
const Matrix &Ms,
const VectorMatrix &M,
VectorMatrix &H)
{
const bool use_cuda = isCudaEnabled();
double energy_sum = 0.0;
if (use_cuda) {
#ifdef HAVE_CUDA
CUTIC("cubic_anisotropy");
energy_sum = cubic_anisotropy_cuda(axis1, axis2, k, Ms, M, H, isCuda64Enabled());
CUTOC("cubic_anisotropy");
#else
assert(0);
#endif
} else {
TIC("cubic_anisotropy");
energy_sum = cubic_anisotropy_cpu(axis1, axis2, k, Ms, M, H);
TOC("cubic_anisotropy");
}
return energy_sum;
}
void minimize(
const Matrix &f, const double h,
const VectorMatrix &M,
const VectorMatrix &H,
VectorMatrix &M2)
{
const bool use_cuda = isCudaEnabled();
if (use_cuda) {
#ifdef HAVE_CUDA
CUTIC("minimize");
#ifdef HAVE_CUDA_64
if (isCuda64Enabled())
minimize_cu64(f, h, M, H, M2);
else
#endif
minimize_cu32(f, h, M, H, M2);
CUTOC("minimize");
#else
assert(0);
#endif
} else {
TIC("minimize");
minimize_cpu(f, h, M, H, M2);
TOC("minimize");
}
}
void Transposer_CUDA::copy_unpad(const float *in_x, const float *in_y, const float *in_z, VectorMatrix &H)
{
// Ifdef HAVE_CUDA_64 and isCuda64Enabled(), we directly store output matrices on the GPU with 64 bit precision.
#ifdef HAVE_CUDA_64
if (isCuda64Enabled()) {
// xyz, s1 -> H
VectorMatrix::cu64_accessor H_acc(H);
cuda_copy_unpad_r2r(exp_x, dim_y, dim_z, dim_x, in_x, in_y, in_z, H_acc.ptr_x(), H_acc.ptr_y(), H_acc.ptr_z());
}
else
#endif
{
// xyz, s1 -> H
VectorMatrix::cu32_accessor H_acc(H);
cuda_copy_unpad_r2r(exp_x, dim_y, dim_z, dim_x, in_x, in_y, in_z, H_acc.ptr_x(), H_acc.ptr_y(), H_acc.ptr_z());
}
}
void rk_prepare_step(
int step, double h, ButcherTableau &tab,
const VectorMatrix &k0, const VectorMatrix &k1, const VectorMatrix &k2,
const VectorMatrix &k3, const VectorMatrix &k4, const VectorMatrix &k5,
const VectorMatrix &y,
VectorMatrix &ytmp)
{
if (isCudaEnabled()) {
#ifdef HAVE_CUDA
rk_prepare_step_cuda(step, h, tab, k0, k1, k2, k3, k4, k5, y, ytmp, isCuda64Enabled());
#else
assert(0);
#endif
} else {
rk_prepare_step_cpu(step, h, tab, k0, k1, k2, k3, k4, k5, y, ytmp);
}
}
double exchange(
int dim_x, int dim_y, int dim_z,
double delta_x, double delta_y, double delta_z,
bool periodic_x, bool periodic_y, bool periodic_z,
const Matrix &Ms,
const Matrix &A,
const VectorMatrix &M,
VectorMatrix &H)
{
const bool use_cuda = isCudaEnabled();
double res = 0;
if (use_cuda) {
#ifdef HAVE_CUDA
CUTIC("exchange");
res = exchange_cuda(dim_x, dim_y, dim_z, delta_x, delta_y, delta_z, periodic_x, periodic_y, periodic_z, Ms, A, M, H, isCuda64Enabled());
CUTOC("exchange");
#else
assert(0);
#endif
} else {
TIC("exchange");
res = exchange_cpu(dim_x, dim_y, dim_z, delta_x, delta_y, delta_z, periodic_x, periodic_y, periodic_z, Ms, A, M, H);
TOC("exchange");
}
return res;
}
