void HPCoarsenTest::add_projection(const System & system,
const Elem * elem,
unsigned int var)
{
// If we have children, we need to add their projections instead
if (!elem->active())
{
libmesh_assert(!elem->subactive());
for (unsigned int c = 0; c != elem->n_children(); ++c)
this->add_projection(system, elem->child(c), var);
return;
}
// The DofMap for this system
const DofMap & dof_map = system.get_dof_map();
// The type of finite element to use for this variable
const FEType & fe_type = dof_map.variable_type (var);
const FEContinuity cont = fe->get_continuity();
fe->reinit(elem);
dof_map.dof_indices(elem, dof_indices, var);
const unsigned int n_dofs =
cast_int<unsigned int>(dof_indices.size());
FEInterface::inverse_map (system.get_mesh().mesh_dimension(),
fe_type, coarse, *xyz_values, coarse_qpoints);
fe_coarse->reinit(coarse, &coarse_qpoints);
const unsigned int n_coarse_dofs =
cast_int<unsigned int>(phi_coarse->size());
if (Uc.size() == 0)
{
Ke.resize(n_coarse_dofs, n_coarse_dofs);
Ke.zero();
Fe.resize(n_coarse_dofs);
Fe.zero();
Uc.resize(n_coarse_dofs);
Uc.zero();
}
libmesh_assert_equal_to (Uc.size(), phi_coarse->size());
// Loop over the quadrature points
for (unsigned int qp=0; qp<qrule->n_points(); qp++)
{
// The solution value at the quadrature point
Number val = libMesh::zero;
Gradient grad;
Tensor hess;
for (unsigned int i=0; i != n_dofs; i++)
{
dof_id_type dof_num = dof_indices[i];
val += (*phi)[i][qp] *
system.current_solution(dof_num);
if (cont == C_ZERO || cont == C_ONE)
grad.add_scaled((*dphi)[i][qp],system.current_solution(dof_num));
// grad += (*dphi)[i][qp] *
// system.current_solution(dof_num);
if (cont == C_ONE)
hess.add_scaled((*d2phi)[i][qp], system.current_solution(dof_num));
// hess += (*d2phi)[i][qp] *
// system.current_solution(dof_num);
}
// The projection matrix and vector
for (unsigned int i=0; i != Fe.size(); ++i)
{
Fe(i) += (*JxW)[qp] *
(*phi_coarse)[i][qp]*val;
if (cont == C_ZERO || cont == C_ONE)
Fe(i) += (*JxW)[qp] *
(grad*(*dphi_coarse)[i][qp]);
if (cont == C_ONE)
Fe(i) += (*JxW)[qp] *
hess.contract((*d2phi_coarse)[i][qp]);
// Fe(i) += (*JxW)[qp] *
// (*d2phi_coarse)[i][qp].contract(hess);
for (unsigned int j=0; j != Fe.size(); ++j)
{
Ke(i,j) += (*JxW)[qp] *
(*phi_coarse)[i][qp]*(*phi_coarse)[j][qp];
if (cont == C_ZERO || cont == C_ONE)
Ke(i,j) += (*JxW)[qp] *
(*dphi_coarse)[i][qp]*(*dphi_coarse)[j][qp];
if (cont == C_ONE)
Ke(i,j) += (*JxW)[qp] *
((*d2phi_coarse)[i][qp].contract((*d2phi_coarse)[j][qp]));
}
}
}
}
void MeshFunction::hessian (const Point& p,
const Real,
std::vector<Tensor>& output,
const std::set<subdomain_id_type>* subdomain_ids)
{
libmesh_assert (this->initialized());
const Elem* element = this->find_element(p,subdomain_ids);
if (!element)
{
output.resize(0);
}
else
{
// resize the output vector to the number of output values
// that the user told us
output.resize (this->_system_vars.size());
{
const unsigned int dim = element->dim();
/*
* Get local coordinates to feed these into compute_data().
* Note that the fe_type can safely be used from the 0-variable,
* since the inverse mapping is the same for all FEFamilies
*/
const Point mapped_point (FEInterface::inverse_map (dim,
this->_dof_map.variable_type(0),
element,
p));
std::vector<Point> point_list (1, mapped_point);
// loop over all vars
for (unsigned int index=0; index < this->_system_vars.size(); index++)
{
/*
* the data for this variable
*/
const unsigned int var = _system_vars[index];
const FEType& fe_type = this->_dof_map.variable_type(var);
UniquePtr<FEBase> point_fe (FEBase::build(dim, fe_type));
const std::vector<std::vector<RealTensor> >& d2phi =
point_fe->get_d2phi();
point_fe->reinit(element, &point_list);
// where the solution values for the var-th variable are stored
std::vector<dof_id_type> dof_indices;
this->_dof_map.dof_indices (element, dof_indices, var);
// interpolate the solution
Tensor hess;
for (unsigned int i=0; i<dof_indices.size(); i++)
hess.add_scaled(d2phi[i][0], this->_vector(dof_indices[i]));
output[index] = hess;
}
}
}
// all done
return;
}
void Tensor::CreateLinearSystem(vector<double>& B_vec, Matrix& A_matrix,
Tensor& X, Tensor& A, Tensor& B,
vector<int>& mult_modesX, vector<int>& mult_modesA)
{
// fake multiply x and A together to create B, creating the linear system in the process
assert(mult_modesX.size() == mult_modesA.size());
if (X.Order() == mult_modesX.size() && A.Order() == mult_modesA.size())
{
assert(0);
}
int numMultElements = 1;
vector<int> mult_dims(mult_modesX.size(), 0);
for (int i = 0; i < mult_modesX.size(); ++i)
{
assert(X.Dim(mult_modesX[i]) == A.Dim(mult_modesA[i]));
mult_dims[i] = X.Dim(mult_modesX[i]);
numMultElements = numMultElements * mult_dims[i];
}
vector<int> mult_offsets;
ComputeOffsets(mult_offsets, mult_dims);
int result_order = X.Order() + A.Order() - mult_modesX.size() - mult_modesA.size();
if (result_order == 0)
assert(0);
vector<int> result_dims;
vector<int> free_modesX;
vector<int> free_modesA;
// find free indices from X
for (int i = 0; i < X.Order(); ++i)
{
if (!VectorPlus::Contains(mult_modesX, i))
{
free_modesX.push_back(i);
}
}
// find free indices from A
for (int i = 0; i < A.Order(); ++i)
{
if (!VectorPlus::Contains(mult_modesA, i))
{
free_modesA.push_back(i);
}
}
vector<int> a_mat_dims = VectorPlus::CreatePair(B.NumElements(), X.NumElements());
A_matrix.Initialize(a_mat_dims);
B_vec.reserve(B.NumElements());
// fill in elements from result tensor
FastIndexer B_indexer(B.Dims());
for (int n = 0; n < B.NumElements(); ++n)
{
B_vec.push_back(B.At(n));
vector<int>& indices = B_indexer.GetNext();
vector<int> free_indicesX;
vector<int> free_indicesA;
// B.ComputeIndexArray(indices, n);
for (int i = 0; i < B.Order(); ++i)
{
if (!VectorPlus::Contains(mult_modesX, i))
free_indicesX.push_back(indices[i]);
else
free_indicesA.push_back(indices[i]);
}
// sum over elementwise products of mult-mode elements
double temp_sum = 0;
FastIndexer mult_indexer(mult_dims);
for (int k = 0; k < numMultElements; ++k)
{
vector<int>& mult_indices = mult_indexer.GetNext();
// ComputeIndexArray(mult_indices, mult_offsets, k);
vector<int> indicesX;
vector<int> indicesA;
MergeIndices(indicesX, mult_modesX, free_modesX, mult_indices, free_indicesX);
MergeIndices(indicesA, mult_modesA, free_modesA, mult_indices, free_indicesA);
A_matrix.Set(n, X.ComputeIndex(indicesX), A.At(indicesA));
}
}
}
void TrigonometricPathVessel::finish( const std::vector<double>& buffer ) {
// Store the data calculated during mpi loop
StoreDataVessel::finish( buffer );
// Get current value of all arguments
for(unsigned i=0; i<cargs.size(); ++i) cargs[i]=mymap->getArgument(i);
// Determine closest and second closest point to current position
double lambda=mymap->getLambda();
std::vector<double> dist( getNumberOfComponents() ), dist2( getNumberOfComponents() );;
retrieveSequentialValue( 0, false, dist );
retrieveSequentialValue( 1, false, dist2 );
iclose1=getStoreIndex(0); iclose2=getStoreIndex(1);
double mindist1=dist[0], mindist2=dist2[0];
if( lambda>0.0 ) {
mindist1=-std::log( dist[0] ) / lambda;
mindist2=-std::log( dist2[0] ) / lambda;
}
if( mindist2<mindist1 ) {
double tmp=mindist1; mindist1=mindist2; mindist2=tmp;
iclose1=getStoreIndex(1); iclose2=getStoreIndex(0);
}
for(unsigned i=2; i<getNumberOfStoredValues(); ++i) {
retrieveSequentialValue( i, false, dist );
double ndist=dist[0];
if( lambda>0.0 ) ndist=-std::log( dist[0] ) / lambda;
if( ndist<mindist1 ) {
mindist2=mindist1; iclose2=iclose1;
mindist1=ndist; iclose1=getStoreIndex(i);
} else if( ndist<mindist2 ) {
mindist2=ndist; iclose2=getStoreIndex(i);
}
}
// And find third closest point
int isign = iclose1 - iclose2;
if( isign>1 ) isign=1; else if( isign<-1 ) isign=-1;
int iclose3 = iclose1 + isign; double v2v2;
// We now have to compute vectors connecting the three closest points to the
// new point
double v1v1 = (mymap->getReferenceConfiguration( iclose1 ))->calculate( mymap->getPositions(), mymap->getPbc(), mymap->getArguments(), mypack1, true );
double v3v3 = (mymap->getReferenceConfiguration( iclose2 ))->calculate( mymap->getPositions(), mymap->getPbc(), mymap->getArguments(), mypack3, true );
if( iclose3<0 || iclose3>=mymap->getFullNumberOfTasks() ) {
ReferenceConfiguration* conf2=mymap->getReferenceConfiguration( iclose1 );
v2v2=(mymap->getReferenceConfiguration( iclose2 ))->calc( conf2->getReferencePositions(), mymap->getPbc(), mymap->getArguments(),
conf2->getReferenceArguments(), mypack2, true );
(mymap->getReferenceConfiguration( iclose2 ))->extractDisplacementVector( conf2->getReferencePositions(), mymap->getArguments(),
conf2->getReferenceArguments(), false, projdir );
} else {
ReferenceConfiguration* conf2=mymap->getReferenceConfiguration( iclose3 );
v2v2=(mymap->getReferenceConfiguration( iclose1 ))->calc( conf2->getReferencePositions(), mymap->getPbc(), mymap->getArguments(),
conf2->getReferenceArguments(), mypack2, true );
(mymap->getReferenceConfiguration( iclose1 ))->extractDisplacementVector( conf2->getReferencePositions(), mymap->getArguments(),
conf2->getReferenceArguments(), false, projdir );
}
// Stash derivatives of v1v1
for(unsigned i=0; i<mymap->getNumberOfArguments(); ++i) mypack1_stashd_args[i]=mypack1.getArgumentDerivative(i);
if( mymap->getNumberOfAtoms()>0 ) {
ReferenceAtoms* at = dynamic_cast<ReferenceAtoms*>( mymap->getReferenceConfiguration( iclose1 ) );
const std::vector<double> & displace( at->getDisplace() );
for(unsigned i=0; i<mymap->getNumberOfAtoms(); ++i) {
mypack1_stashd_atoms[i]=mypack1.getAtomDerivative(i); mypack1.getAtomsDisplacementVector()[i] /= displace[i];
}
}
// Calculate the dot product of v1 with v2
double v1v2 = (mymap->getReferenceConfiguration(iclose1))->projectDisplacementOnVector( projdir, mymap->getArguments(), cargs, mypack1 );
// This computes s value
double spacing = mymap->getPropertyValue( iclose1, (mymap->property.begin())->first ) - mymap->getPropertyValue( iclose2, (mymap->property.begin())->first );
double root = sqrt( v1v2*v1v2 - v2v2 * ( v1v1 - v3v3) );
dx = 0.5 * ( (root + v1v2) / v2v2 - 1.);
double path_s = mymap->getPropertyValue(iclose1, (mymap->property.begin())->first ) + spacing * dx; sp->set( path_s );
double fact = 0.25*spacing / v2v2;
// Derivative of s wrt arguments
for(unsigned i=0; i<mymap->getNumberOfArguments(); ++i) {
sp->setDerivative( i, fact*( mypack2.getArgumentDerivative(i) + (v2v2 * (-mypack1_stashd_args[i] + mypack3.getArgumentDerivative(i))
+ v1v2*mypack2.getArgumentDerivative(i) )/root ) );
}
// Derivative of s wrt atoms
unsigned narg=mymap->getNumberOfArguments(); Tensor vir; vir.zero(); fact = 0.5*spacing / v2v2;
if( mymap->getNumberOfAtoms()>0 ) {
for(unsigned i=0; i<mymap->getNumberOfAtoms(); ++i) {
Vector ader = fact*(( v1v2*mypack1.getAtomDerivative(i) + 0.5*v2v2*(-mypack1_stashd_atoms[i] + mypack3.getAtomDerivative(i) ) )/root + mypack1.getAtomDerivative(i) );
for(unsigned k=0; k<3; ++k) sp->setDerivative( narg+3*i+k, ader[k] );
vir-=Tensor( mymap->getPosition(i), ader );
}
// Set the virial
unsigned nbase=narg+3*mymap->getNumberOfAtoms();
for(unsigned i=0; i<3; ++i) for(unsigned j=0; j<3; ++j) sp->setDerivative( nbase+3*i+j, vir(i,j) );
}
// Now compute z value
ReferenceConfiguration* conf2=mymap->getReferenceConfiguration( iclose1 );
double v4v4=(mymap->getReferenceConfiguration( iclose2 ))->calc( conf2->getReferencePositions(), mymap->getPbc(), mymap->getArguments(),
conf2->getReferenceArguments(), mypack2, true );
// Extract vector connecting frames
(mymap->getReferenceConfiguration( iclose2 ))->extractDisplacementVector( conf2->getReferencePositions(), mymap->getArguments(),
conf2->getReferenceArguments(), false, projdir );
// Calculate projection of vector on line connnecting frames
double proj = (mymap->getReferenceConfiguration(iclose1))->projectDisplacementOnVector( projdir, mymap->getArguments(), cargs, mypack1 );
double path_z = v1v1 + dx*dx*v4v4 - 2*dx*proj;
// Derivatives for z path
path_z = sqrt(path_z); zp->set( path_z ); vir.zero();
for(unsigned i=0; i<mymap->getNumberOfArguments(); ++i) zp->setDerivative( i, (mypack1_stashd_args[i] - 2*dx*mypack1.getArgumentDerivative(i))/(2.0*path_z) );
// Derivative wrt atoms
if( mymap->getNumberOfAtoms()>0 ) {
for(unsigned i=0; i<mymap->getNumberOfAtoms(); ++i) {
Vector dxder; for(unsigned k=0; k<3; ++k) dxder[k] = ( 2*v4v4*dx - 2*proj )*spacing*sp->getDerivative( narg + 3*i+k );
Vector ader = ( mypack1_stashd_atoms[i] - 2.*dx*mypack1.getAtomDerivative(i) + dxder )/ (2.0*path_z);
for(unsigned k=0; k<3; ++k) zp->setDerivative( narg+3*i+k, ader[k] );
vir-=Tensor( mymap->getPosition(i), ader );
}
// Set the virial
unsigned nbase=narg+3*mymap->getNumberOfAtoms();
for(unsigned i=0; i<3; ++i) for(unsigned j=0; j<3; ++j) zp->setDerivative( nbase+3*i+j, vir(i,j) );
}
}
std::tuple<Tensor, Tensor, Tensor, Tensor>
embedding_bag_cpu(const Tensor &weight, const Tensor &indices__,
const Tensor &offsets__, const bool scale_grad_by_freq,
const int64_t mode, bool sparse) {
auto indices_arg = TensorArg(indices__, "indices__", 1);
checkScalarType("embedding_bag", indices_arg, kLong);
auto offsets_arg = TensorArg(offsets__, "offsets__", 1);
checkScalarType("embedding_bag", offsets_arg, kLong);
Tensor indices = indices__.contiguous();
Tensor offsets = offsets__.contiguous();
auto weight_arg = TensorArg(weight, "weight", 1);
checkScalarTypes("embedding_bag", weight_arg, {kFloat, kDouble});
auto bag_size = at::zeros(offsets.sizes(), indices.type());
make_bag_size(offsets, indices, mode, bag_size);
// If the last entries are empty, that the last offsets are irrelevant as they
// won't change anything in the assignment of ID -> bag, but index_add would
// throw out of bounds error. So to keep it simple we just add one more
// entry to the end then get rid of it after make_offset2bag.
auto offset2bag = at::zeros(
{indices.sizes()[0] + 1}, indices__.type()); // offset2bag = [0 0 0 0 0]
make_offset2bag(offsets, indices, offset2bag);
offset2bag.resize_({indices.sizes()[0]});
auto output = at::zeros({offsets.size(0), weight.size(1)}, weight.type());
if (mode == MODE_MEAN || mode == MODE_SUM) {
if (weight.type().scalarType() == kFloat) {
index_select_add<float>(indices, offset2bag, weight, output);
} else if (weight.type().scalarType() == kDouble) {
index_select_add<double>(indices, offset2bag, weight, output);
}
auto ret = apply_bag_size(offsets, indices, mode, output, bag_size);
return std::tuple<Tensor, Tensor, Tensor, Tensor>(ret, offset2bag, bag_size, bag_size);
} else { // MODE_MAX
return AT_DISPATCH_FLOATING_TYPES_AND_HALF(
weight.type(), "embedding_bag_cpu_max", [&]() {
return embedding_bag_cpu_max<scalar_t>(weight, indices, offset2bag, output, bag_size, offsets);
}
);
}
}
inline void MapPlan(Tensor<gpu,dim> _dst, const expr::Plan<E> &plan){
cuda::MapPlan<Saver>( _dst.FlatTo2D(), plan );
}
void Tensor<Dtype>::ShareMem(const Tensor& other) {
ASSERT(count_ == other.count(), "");
mem_ = other.mem();
}
void Tensor::copyTo(Tensor& dst)
{
void* p = map();
dst.reshape((const char*)p, shape_, format_);
unMap();
}
void MeshFunction::hessian (const Point& p,
const Real,
std::vector<Tensor>& output)
{
libmesh_assert (this->initialized());
/* Ensure that in the case of a master mesh function, the
out-of-mesh mode is enabled either for both or for none. This is
important because the out-of-mesh mode is also communicated to
the point locator. Since this is time consuming, enable it only
in debug mode. */
#ifdef DEBUG
if (this->_master != NULL)
{
const MeshFunction* master =
cast_ptr<const MeshFunction*>(this->_master);
if(_out_of_mesh_mode!=master->_out_of_mesh_mode)
libmesh_error_msg("ERROR: If you use out-of-mesh-mode in connection with master mesh " \
<< "functions, you must enable out-of-mesh mode for both the master and the slave mesh function.");
}
#endif
// locate the point in the other mesh
const Elem* element = this->_point_locator->operator()(p);
// If we have an element, but it's not a local element, then we
// either need to have a serialized vector or we need to find a
// local element sharing the same point.
if (element &&
(element->processor_id() != this->processor_id()) &&
_vector.type() != SERIAL)
{
// look for a local element containing the point
std::set<const Elem*> point_neighbors;
element->find_point_neighbors(p, point_neighbors);
element = NULL;
std::set<const Elem*>::const_iterator it = point_neighbors.begin();
const std::set<const Elem*>::const_iterator end = point_neighbors.end();
for (; it != end; ++it)
{
const Elem* elem = *it;
if (elem->processor_id() == this->processor_id())
{
element = elem;
break;
}
}
}
if (!element)
{
output.resize(0);
}
else
{
// resize the output vector to the number of output values
// that the user told us
output.resize (this->_system_vars.size());
{
const unsigned int dim = this->_eqn_systems.get_mesh().mesh_dimension();
/*
* Get local coordinates to feed these into compute_data().
* Note that the fe_type can safely be used from the 0-variable,
* since the inverse mapping is the same for all FEFamilies
*/
const Point mapped_point (FEInterface::inverse_map (dim,
this->_dof_map.variable_type(0),
element,
p));
std::vector<Point> point_list (1, mapped_point);
// loop over all vars
for (unsigned int index=0; index < this->_system_vars.size(); index++)
{
/*
* the data for this variable
*/
const unsigned int var = _system_vars[index];
const FEType& fe_type = this->_dof_map.variable_type(var);
AutoPtr<FEBase> point_fe (FEBase::build(dim, fe_type));
const std::vector<std::vector<RealTensor> >& d2phi =
point_fe->get_d2phi();
point_fe->reinit(element, &point_list);
// where the solution values for the var-th variable are stored
std::vector<dof_id_type> dof_indices;
this->_dof_map.dof_indices (element, dof_indices, var);
// interpolate the solution
Tensor hess;
for (unsigned int i=0; i<dof_indices.size(); i++)
hess.add_scaled(d2phi[i][0], this->_vector(dof_indices[i]));
output[index] = hess;
}
}
}
// all done
return;
}
bool
test_fundamentals(Index const dimension)
{
bool
passed = true;
Index const
number_components = integer_power(dimension, Tensor::ORDER);
std::vector<Scalar> const
X = generate_sequence<Scalar>(number_components, 1.0, 1.0);
// Test constructor with pointer
Tensor const
A(dimension, &X[0]);
// Test copy constructor
Tensor
B = A;
Tensor
C;
// Test copy assignment
C = B - A;
Scalar
error = norm_f(C);
bool const
copy_assigned = error <= machine_epsilon<Scalar>();
passed = passed && copy_assigned;
// Test fill with pointer
B.fill(&X[0]);
C = B - A;
error = norm_f(C);
bool const
filled_pointer = error <= machine_epsilon<Scalar>();
passed = passed && filled_pointer;
std::vector<Scalar> const
Y = generate_sequence<Scalar>(number_components, -1.0, -1.0);
C.fill(&Y[0]);
// Test increment
C += A;
error = norm_f(C);
bool const
incremented = error <= machine_epsilon<Scalar>();
passed = passed && incremented;
C.fill(&X[0]);
// Test decrement
C -= A;
error = norm_f(C);
bool const
decremented = error <= machine_epsilon<Scalar>();
passed = passed && decremented;
#ifdef HAVE_INTREPID_KOKKOSCORE
//test Tensor fill and create for Kokkos data types
Kokkos::View<Scalar *, Kokkos::DefaultExecutionSpace>
X1("X1_kokkos", dimension);
Kokkos::View<Scalar **, Kokkos::DefaultExecutionSpace>
X2("X2_kokkos", dimension, dimension);
Kokkos::View<Scalar ***, Kokkos::DefaultExecutionSpace>
X3("X3_kokkos", dimension, dimension, dimension);
Kokkos::View<Scalar ****, Kokkos::DefaultExecutionSpace>
X4("X4_kokkos", dimension, dimension, dimension, dimension);
Kokkos::deep_copy(X1, 3.1);
Kokkos::deep_copy(X2, 3.2);
Kokkos::deep_copy(X3, 3.3);
Kokkos::deep_copy(X4, 3.4);
Tensor
Z(dimension); //(X1_k,0);
Index
rank = 0;
Index
temp = number_components;
while (temp != 1) {
temp = temp / dimension;
rank = rank + 1;
assert(temp > 0);
}
switch (rank) {
default:
assert(false);
break;
case 1:
Z.fill(X1, 0);
break;
case 2:
Z.fill(X2, 0, 0);
break;
case 3:
Z.fill(X3, 0, 0, 0);
break;
case 4:
Z.fill(X4, 0, 0, 0, 0);
break;
}
// Test copy constructor.
Tensor const
U = Z;
// Test copy assignment.
Tensor
V;
V = U - Z;
error = norm_f(V);
bool const
tensor_create_from_1d_kokkos = error <= machine_epsilon<Scalar>();
passed = passed && tensor_create_from_1d_kokkos;
#endif
return passed;
}
#define CATCH_CONFIG_MAIN
#include "catch.hpp"
#include "ATen/ATen.h"
#include "ATen/DLConvertor.h"
#include <iostream>
#include <string.h>
#include <sstream>
#include "test_seed.h"
using namespace at;
TEST_CASE( "parallel", "[cpu]" ) {
manual_seed(123, at::Backend::CPU);
set_num_threads(1);
Tensor a = rand(CPU(at::kFloat), {1,3});
a[0][0] = 1;
a[0][1] = 0;
a[0][2] = 0;
Tensor as = rand(CPU(at::kFloat), {3});
as[0] = 1;
as[1] = 0;
as[2] = 0;
REQUIRE(a.sum(0).equal(as));
}
Tensor upsample_nearest1d_cpu(const Tensor& input, IntArrayRef output_size) {
auto output = at::empty({0}, input.options());
upsample_nearest1d_out_cpu_template(output, input, output_size);
return output;
}
static void test_resize()
{
Tensor<int, 3> epsilon;
epsilon.resize(2,3,7);
VERIFY_IS_EQUAL(epsilon.dimension(0), 2);
VERIFY_IS_EQUAL(epsilon.dimension(1), 3);
VERIFY_IS_EQUAL(epsilon.dimension(2), 7);
VERIFY_IS_EQUAL(epsilon.dimensions().TotalSize(), 2*3*7);
const int* old_data = epsilon.data();
epsilon.resize(3,2,7);
VERIFY_IS_EQUAL(epsilon.dimension(0), 3);
VERIFY_IS_EQUAL(epsilon.dimension(1), 2);
VERIFY_IS_EQUAL(epsilon.dimension(2), 7);
VERIFY_IS_EQUAL(epsilon.dimensions().TotalSize(), 2*3*7);
VERIFY_IS_EQUAL(epsilon.data(), old_data);
epsilon.resize(3,5,7);
VERIFY_IS_EQUAL(epsilon.dimension(0), 3);
VERIFY_IS_EQUAL(epsilon.dimension(1), 5);
VERIFY_IS_EQUAL(epsilon.dimension(2), 7);
VERIFY_IS_EQUAL(epsilon.dimensions().TotalSize(), 3*5*7);
VERIFY_IS_NOT_EQUAL(epsilon.data(), old_data);
}
void ERMSD::calcMat(const std::vector<Vector> & positions,const Pbc& pbc, std::vector<Vector4d> &mat, std::vector<TensorGeneric<4,3> > &Gderi) {
std::vector<Vector3d> pos;
pos.resize(3*nresidues);
std::vector<Tensor3d> deri;
deri.resize(nresidues*9);
std::vector<Vector> centers;
centers.resize(nresidues);
unsigned idx_deri = 0;
Tensor da_dxa = (2./3.)*Tensor::identity();
Tensor da_dxb = -(1./3.)*Tensor::identity();
Tensor da_dxc = -(1./3.)*Tensor::identity();
Tensor db_dxa = -(1./3.)*Tensor::identity();
Tensor db_dxb = (2./3.)*Tensor::identity();
Tensor db_dxc = -(1./3.)*Tensor::identity();
// Form factors - should this be somewhere else?
double w = 1./3.;
Vector form_factor = Vector(2.0,2.0,1.0/0.3);
for(unsigned res_idx=0; res_idx<natoms/3; res_idx++) {
const unsigned at_idx = 3*res_idx;
//center
for (unsigned j=0; j<3; j++) {
centers[res_idx] += w*positions[at_idx+j];
}
Vector3d a = delta(centers[res_idx],positions[at_idx]);
Vector3d b = delta(centers[res_idx],positions[at_idx+1]);
Vector3d d = crossProduct(a,b);
double ianorm = 1./a.modulo();
double idnorm = 1./d.modulo();
// X vector: COM-C2
pos[at_idx] = a*ianorm;
// Z versor: C2 x (COM-C4/C6)
pos[at_idx+2] = d*idnorm;
// Y versor: Z x Y
pos[at_idx+1] = crossProduct(pos[at_idx+2],pos[at_idx]);
// Derivatives ////////
Tensor3d t1 = ianorm*(Tensor::identity()-extProduct(pos[at_idx],pos[at_idx]));
// dv1/dxa
deri[idx_deri] = (2./3. )*t1;
// dv1/dxb
deri[idx_deri+3] = -(1./3.)*t1;
// dv1/dxc
deri[idx_deri+6] = -(1./3.)*t1;
Tensor dd_dxa = VcrossTensor(a,db_dxa) -VcrossTensor(b,da_dxa);
Tensor dd_dxb = VcrossTensor(a,db_dxb)-VcrossTensor(b,da_dxb);
Tensor dd_dxc = VcrossTensor(a,db_dxc)-VcrossTensor(b,da_dxc);
// dv3/dxa
deri[idx_deri+2] = deriNorm(d,dd_dxa);
// dv3/dxb
deri[idx_deri+5] = deriNorm(d,dd_dxb);
// dv3/dxc
deri[idx_deri+8] = deriNorm(d,dd_dxc);
// dv2/dxa = dv3/dxa cross v1 + v3 cross dv1/dxa
deri[idx_deri+1] = (VcrossTensor(deri[idx_deri+2],pos[at_idx]) + \
VcrossTensor(pos[at_idx+2],deri[idx_deri]));
// dv2/dxb
deri[idx_deri+4] = (VcrossTensor(deri[idx_deri+5],pos[at_idx]) + \
VcrossTensor(pos[at_idx+2],deri[idx_deri+3]));
// dv2/dxc
deri[idx_deri+7] = (VcrossTensor(deri[idx_deri+8],pos[at_idx]) + \
VcrossTensor(pos[at_idx+2],deri[idx_deri+6]));
idx_deri += 9;
// End derivatives ///////
}
// Initialization (unnecessary?)
for (unsigned i1=0; i1<nresidues*nresidues; i1++) {
for (unsigned i2=0; i2<4; i2++) {
mat[i1][i2] = 0.0;
}
}
double maxdist = cutoff/form_factor[0];
double gamma = pi/cutoff;
unsigned idx;
unsigned idx1 = 0;
// Calculate mat
for (unsigned i=0; i<nresidues; i++) {
for (unsigned j=0; j<nresidues; j++) {
// skip i==j
if(inPair(i,j) and i != j) {
//if(i!=j){
// Calculate normal distance first
Vector diff = delta(centers[i],centers[j]);
double d1 = diff.modulo();
//std::cout << inPair(i,j) << " " << i << " " << j << " "<< d1 <<"\n";
//std::cout << inPair(i,j) << " " << i << " " << j << " "<< d1 <<"\n";
if(d1<maxdist) {
// calculate r_tilde_ij
Vector3d rtilde;
for (unsigned k=0; k<3; k++) {
for (unsigned l=0; l<3; l++) {
rtilde[l] += pos[3*i+l][k]*diff[k]*form_factor[l];
}
}
double rtilde_norm = rtilde.modulo();
double irnorm = 1./rtilde_norm;
// ellipsoidal cutoff
if(rtilde_norm < cutoff) {
idx = i*nresidues + j;
//std::cout << i << " " << j << " " << rtilde_norm << " " << idx <<"\n";
// fill 4d matrix
double dummy = sin(gamma*rtilde_norm)/(rtilde_norm*gamma);
mat[idx][0] = dummy*rtilde[0];
mat[idx][1] = dummy*rtilde[1];
mat[idx][2] = dummy*rtilde[2];
mat[idx][3] = (1.+ cos(gamma*rtilde_norm))/gamma;
// Derivative (drtilde_dx)
std::vector<Tensor3d> drtilde_dx;
drtilde_dx.resize(6);
unsigned pos_idx = 3*i;
unsigned deri_idx = 9*i;
for (unsigned at=0; at<3; at++) {
for (unsigned l=0; l<3; l++) {
Vector3d rvec = form_factor[l]*((pos[pos_idx+l])/3.);
Vector3d vvec = form_factor[l]*(matmul(deri[deri_idx+3*at+l],diff));
drtilde_dx[at].setRow(l,vvec-rvec);
drtilde_dx[at+3].setRow(l,rvec);
}
}
//std::vector<TensorGeneric<4,3> > dG_dx;
//dG_dx.resize(6);
double dummy1 = (cos(gamma*rtilde_norm) - dummy);
idx1 = i*nresidues*6 + j*6;
for (unsigned l=0; l<6; l++) {
//std::cout << i << " " << j << " " << idx1 << " " << idx1+l << "\n";
// components 1,2,3
// sin(gamma*|rtilde|)/gamma*|rtilde|*d_rtilde +
// + ((d_rtilde*r_tilde/r_tilde^2) out r_tilde)*
// (cos(gamma*|rtilde| - sin(gamma*|rtilde|)/gamma*|rtilde|))
Vector3d rdr = matmul(rtilde,drtilde_dx[l]);
Tensor tt = dummy*drtilde_dx[l] + (dummy1*irnorm*irnorm)*Tensor(rtilde,rdr);
for (unsigned m=0; m<3; m++) {
// Transpose here
//dG_dx[l].setRow(m,tt.getRow(m));
Gderi[idx1+l].setRow(m,tt.getRow(m));
}
// component 4
// - sin(gamma*|rtilde|)/|rtilde|*(r_tilde*d_rtilde)
//dG_dx[l].setRow(3,-dummy*gamma*rdr);
Gderi[idx1+l].setRow(3,-dummy*gamma*rdr);
}
}
}
}
}
}
}
void Tensor<XPU, DT>::blas_vsqrt(const Tensor<XPU, DT> &in)
{ const int N = size(); CHECK_EQ (N, in.size());
unary_vexpr_kernel<opsqrt<DT>><<<cuda_get_blocks(N), CUDA_NUM_THREADS, 0, get_calc_stream()>>>
(N, in.dptr, dptr);
cuda_sync_check ("cublas_vsqrt");
};
TEST(TestScalar, TestScalar) {
manual_seed(123);
Scalar what = 257;
Scalar bar = 3.0;
Half h = bar.toHalf();
Scalar h2 = h;
cout << "H2: " << h2.toDouble() << " " << what.toFloat() << " "
<< bar.toDouble() << " " << what.isIntegral() << "\n";
Generator& gen = at::globalContext().defaultGenerator(at::kCPU);
ASSERT_NO_THROW(gen.seed());
auto&& C = at::globalContext();
if (at::hasCUDA()) {
auto t2 = zeros({4, 4}, at::kCUDA);
cout << &t2 << "\n";
}
auto t = ones({4, 4});
auto wha2 = zeros({4, 4}).add(t).sum();
ASSERT_EQ(wha2.item<double>(), 16.0);
ASSERT_EQ(t.sizes()[0], 4);
ASSERT_EQ(t.sizes()[1], 4);
ASSERT_EQ(t.strides()[0], 4);
ASSERT_EQ(t.strides()[1], 1);
TensorOptions options = dtype(kFloat);
Tensor x = randn({1, 10}, options);
Tensor prev_h = randn({1, 20}, options);
Tensor W_h = randn({20, 20}, options);
Tensor W_x = randn({20, 10}, options);
Tensor i2h = at::mm(W_x, x.t());
Tensor h2h = at::mm(W_h, prev_h.t());
Tensor next_h = i2h.add(h2h);
next_h = next_h.tanh();
ASSERT_ANY_THROW(Tensor{}.item());
test_overflow();
if (at::hasCUDA()) {
auto r = next_h.to(at::Device(kCUDA), kFloat, /*non_blocking=*/ false, /*copy=*/ true);
ASSERT_TRUE(r.to(at::Device(kCPU), kFloat, /*non_blocking=*/ false, /*copy=*/ true).equal(next_h));
}
ASSERT_NO_THROW(randn({10, 10, 2}, options));
// check Scalar.toTensor on Scalars backed by different data types
ASSERT_EQ(scalar_to_tensor(bar).scalar_type(), kDouble);
ASSERT_EQ(scalar_to_tensor(what).scalar_type(), kLong);
ASSERT_EQ(scalar_to_tensor(ones({}).item()).scalar_type(), kDouble);
if (x.scalar_type() != ScalarType::Half) {
AT_DISPATCH_ALL_TYPES(x.scalar_type(), "foo", [&] {
scalar_t s = 1;
std::stringstream ss;
ASSERT_NO_THROW(
ss << "hello, dispatch" << x.dispatch_type().toString() << s << "\n");
auto data = (scalar_t*)x.data_ptr();
(void)data;
});
}
// test direct C-scalar type conversions
{
auto x = ones({1, 2}, options);
ASSERT_ANY_THROW(x.item<float>());
}
auto float_one = ones({}, options);
ASSERT_EQ(float_one.item<float>(), 1);
ASSERT_EQ(float_one.item<int32_t>(), 1);
ASSERT_EQ(float_one.item<at::Half>(), 1);
}
void Tensor<XPU, DT>::blas_vdiv (const Tensor<XPU, DT> &A, const Tensor<XPU, DT> &B)
{ const int N = size(); CHECK_EQ (A.size(), B.size());
binary_vexpr_kernel<opdiv <DT>><<<cuda_get_blocks(N), CUDA_NUM_THREADS, 0, get_calc_stream()>>>
(N, A.dptr, B.dptr, dptr);
cuda_sync_check ("cublas_vdiv");
};
void SpatialDivisiveNormalization::init(std::shared_ptr<TorchData> input) {
RASSERT(input->type() == TorchDataType::TENSOR_DATA);
Tensor<float>* in = TO_TENSOR_PTR(input.get());
RASSERT(in->dim() == 3);
if (output != nullptr) {
if (!in->isSameSizeAs(*TO_TENSOR_PTR(output.get()))) {
// Input dimension has changed!
cleanup();
}
}
if (output == nullptr) {
output.reset(new Tensor<float>(in->dim(), in->size()));
std_pass1_.reset(new Tensor<float>(in->dim(), in->size()));
std_pass2_.reset(new Tensor<float>(in->dim(), in->size()));
}
if (kernel_norm_ == nullptr) {
bool onedim_kernel = kernel_->dim() == 1;
const float n_feats = (float)in->size()[2];
// Clone and normalize the input kernel
kernel_norm_.reset(Tensor<float>::clone(*kernel_));
float sum = Tensor<float>::slowSum(*kernel_norm_);
float div_val = onedim_kernel ? (sum * sqrtf(n_feats)) : (sum * n_feats);
Tensor<float>::div(*kernel_norm_, div_val);
}
if (std_coef_ == nullptr) {
uint32_t std_coeff_size[2];
std_coeff_size[0] = TO_TENSOR_PTR(output.get())->size()[0];
std_coeff_size[1] = TO_TENSOR_PTR(output.get())->size()[1];
std_coef_.reset(new Tensor<float>(2, std_coeff_size));
std::unique_ptr<float[]> std_coef_cpu(new float[std_coef_->nelems()]);
std::unique_ptr<float[]> kernel_norm_cpu(new float[kernel_norm_->nelems()]);
kernel_norm_->getData(kernel_norm_cpu.get());
bool onedim_kernel = kernel_->dim() == 1;
// Filter an image of all 1 values to create the normalization constants
// See norm_test.lua for proof that this works as well as:
// https://github.com/andresy/torch/blob/master/extra/nn/SpatialDivisiveNormalization.lua
int32_t n_feats = TO_TENSOR_PTR(output.get())->size()[2];
int32_t height = TO_TENSOR_PTR(output.get())->size()[1];
int32_t width = TO_TENSOR_PTR(output.get())->size()[0];
if (onedim_kernel) {
// 1D case - The filter is seperable, but we'll just do the dumb 2D
// version since we only do this once on startup. --> O(n * m)
int32_t kernel_size = kernel_norm_->size()[0];
int32_t filt_rad = (kernel_size - 1) / 2;
for (int32_t v = 0; v < height; v++) {
for (int32_t u = 0; u < width; u++) {
float tmp = 0.0f;
for (int32_t v_filt = -filt_rad; v_filt <= filt_rad; v_filt++) {
for (int32_t u_filt = -filt_rad; u_filt <= filt_rad; u_filt++) {
int32_t u_in = u + u_filt;
int32_t v_in = v + v_filt;
if (u_in >= 0 && u_in < width && v_in >= 0 && v_in < height) {
// Pixel is inside --> We'll effectively clamp zeros elsewhere.
tmp += (kernel_norm_cpu[v_filt + filt_rad] *
kernel_norm_cpu[u_filt + filt_rad]);
}
}
}
std_coef_cpu[v * width + u] = tmp / n_feats;
}
}
} else {
// 2D case
int32_t kernel_size_u = kernel_norm_->size()[0];
int32_t kernel_size_v = kernel_norm_->size()[1];
int32_t filt_rad_u = (kernel_size_u - 1) / 2;
int32_t filt_rad_v = (kernel_size_v - 1) / 2;
for (int32_t v = 0; v < height; v++) {
for (int32_t u = 0; u < width; u++) {
float tmp = 0.0f;
for (int32_t v_filt = -filt_rad_v; v_filt <= filt_rad_v; v_filt++) {
for (int32_t u_filt = -filt_rad_u; u_filt <= filt_rad_u; u_filt++) {
int32_t u_in = u + u_filt;
int32_t v_in = v + v_filt;
if (u_in >= 0 && u_in < width && v_in >= 0 && v_in < height) {
// Pixel is inside --> We'll effectively clamp zeros elsewhere.
tmp += kernel_norm_cpu[(v_filt + filt_rad_v) * kernel_size_u +
(u_filt + filt_rad_u)];
}
}
}
std_coef_cpu[v * width + u] = tmp / n_feats;
}
}
}
std_coef_->setData(std_coef_cpu.get());
}
if (std_ == nullptr) {
uint32_t std_coeff_size[2];
std_coeff_size[0] = TO_TENSOR_PTR(output.get())->size()[0];
std_coeff_size[1] = TO_TENSOR_PTR(output.get())->size()[1];
std_.reset(new Tensor<float>(2, std_coeff_size));
}
}
void Tensor<Dtype>::ReshapeLike(const Tensor<Dtype>& other) {
Reshape(other.shape());
}
void SpatialDivisiveNormalization::forwardProp(
std::shared_ptr<TorchData> input) {
init(input);
bool onedim_kernel = kernel_->dim() == 1;
Tensor<float>* in = TO_TENSOR_PTR(input.get());
Tensor<float>* out = TO_TENSOR_PTR(output.get());
if (onedim_kernel) {
int32_t filt_rad = ((int32_t)kernel_norm_->size()[0] - 1) / 2;
// Perform horizontal filter pass
cl_context->useKernelCStr(kSpatialDivisiveNormalizationKernel,
"SpatialDivisiveNormalizationHoriz");
cl_context->setArg(0, in->storage());
cl_context->setArg(1, std_pass1_->storage());
cl_context->setArg(2, kernel_norm_->storage());
cl_context->setArg(3, filt_rad);
cl_context->runKernel(jtorch::deviceid, std_pass1_->dim(),
std_pass1_->size(), false);
// Perform vertical filter pass
cl_context->useKernelCStr(kSpatialDivisiveNormalizationKernel,
"SpatialDivisiveNormalizationVert");
cl_context->setArg(0, std_pass1_->storage());
cl_context->setArg(1, std_pass2_->storage());
cl_context->setArg(2, kernel_norm_->storage());
cl_context->setArg(3, filt_rad);
cl_context->runKernel(jtorch::deviceid, std_pass2_->dim(),
std_pass2_->size(), false);
} else {
int32_t filt_rad_u = ((int32_t)kernel_norm_->size()[0] - 1) / 2;
int32_t filt_rad_v = ((int32_t)kernel_norm_->size()[1] - 1) / 2;
// Perform vertical filter pass
cl_context->useKernelCStr(kSpatialDivisiveNormalizationKernel,
"SpatialDivisiveNormalization2D");
cl_context->setArg(0, in->storage());
cl_context->setArg(1, std_pass2_->storage());
cl_context->setArg(2, kernel_norm_->storage());
cl_context->setArg(3, filt_rad_u);
cl_context->setArg(4, filt_rad_v);
cl_context->runKernel(jtorch::deviceid, std_pass2_->dim(),
std_pass2_->size(), false);
}
// Perform accumulation and division pass
cl_context->useKernelCStr(kSpatialDivisiveNormalizationKernel,
"SpatialDivisiveNormalizationAccumDiv");
cl_context->setArg(0, std_pass2_->storage());
cl_context->setArg(1, std_->storage());
cl_context->setArg(2, std_coef_->storage());
cl_context->setArg(3, (int)out->size()[2]);
cl_context->setArg(4, threshold_);
cl_context->runKernel(jtorch::deviceid, std_->dim(), std_->size(), false);
// Perform normalization pass
cl_context->useKernelCStr(kSpatialDivisiveNormalizationKernel,
"SpatialDivisiveNormalization");
cl_context->setArg(0, in->storage());
cl_context->setArg(1, out->storage());
cl_context->setArg(2, std_->storage());
cl_context->runKernel(jtorch::deviceid, out->dim(), out->size(), false);
}
RTensor
do_block_svd(const Tensor &A, Tensor *pU, Tensor *pVT, bool economic)
{
index rows = A.rows();
index cols = A.columns();
if (rows != cols && !economic)
return svd(A, pU, pVT, economic);
index minrc = std::min(rows, cols);
index nblocks;
Indices *block_rows, *block_cols;
if (!find_blocks<Tensor>(A, &nblocks, &block_rows, &block_cols)) {
return svd(A, pU, pVT, economic);
}
if ((nblocks == 1) &&
(block_rows[0].size() >= rows/2) &&
(block_cols[0].size() >= cols/2)) {
RTensor s = svd(A, pU, pVT, economic);
delete[] block_rows;
delete[] block_cols;
return s;
}
RTensor s(minrc);
s.fill_with_zeros();
if (pU) {
*pU = Tensor::zeros(rows, economic? minrc : rows);
}
if (pVT) {
*pVT = Tensor::zeros(economic? minrc : cols, cols);
}
RTensor stemp;
Tensor Utemp, Vtemp;
Tensor *pUtemp = pU? &Utemp : 0;
Tensor *pVtemp = pVT? &Vtemp : 0;
for (index b = 0, sndx = 0; b < nblocks; b++) {
Tensor m = A(range(block_rows[b]), range(block_cols[b]));
index n = m.size();
if (m.size() > 1) {
stemp = svd(m, pUtemp, pVtemp, economic);
index slast = sndx + stemp.size() - 1;
s.at(range(sndx, slast)) = stemp;
if (pU) {
(*pU).at(range(block_rows[b]), range(sndx, slast)) = Utemp;
}
if (pVT) {
(*pVT).at(range(sndx, slast), range(block_cols[b])) = Vtemp;
}
sndx = slast + 1;
} else {
index row = block_rows[b][0];
index col = block_cols[b][0];
double aux = abs(m[0]);
s.at(sndx) = aux;
if (pU) {
(*pU).at(row,sndx) = 1.0;
}
if (pVT) {
(*pVT).at(sndx,col) = m[0]/aux;
}
++sndx;
}
}
delete[] block_rows;
delete[] block_cols;
Indices ndx = sort_indices(s, true);
s = s(range(ndx));
if (pU)
*pU = (*pU)(range(), range(ndx));
if (pVT)
*pVT = (*pVT)(range(ndx), range());
return s;
}
std::tuple<Tensor, Tensor>
_unique_cpu(const Tensor& self, const bool sorted, const bool return_inverse) {
return AT_DISPATCH_ALL_TYPES(self.type(), "unique", [&] {
return _unique_cpu_template<scalar_t>(self, sorted, return_inverse);
});
}
void SpatialBatchNormalization::init(std::shared_ptr<TorchData> input) {
RASSERT(input->type() == TorchDataType::TENSOR_DATA);
Tensor<float>* in = TO_TENSOR_PTR(input.get());
Tensor<float>* out = TO_TENSOR_PTR(output.get());
RASSERT(in->dim() >= 3);
RASSERT(in->size()[2] == nfeats_);
if (output != nullptr && in->dim() != out->dim()) {
output = nullptr;
}
// Check that the input and output size are the same.
if (output != nullptr) {
if (in->size()[0] != out->size()[0] ||
in->size()[1] != out->size()[1] ||
in->size()[2] != out->size()[2]) {
output = nullptr;
}
}
if (output == nullptr) {
output.reset(new Tensor<float>(in->dim(), in->size()));
}
}
static void test_simple_reductions() {
Tensor<float, 4, DataLayout> tensor(2, 3, 5, 7);
tensor.setRandom();
array<ptrdiff_t, 2> reduction_axis2;
reduction_axis2[0] = 1;
reduction_axis2[1] = 3;
Tensor<float, 2, DataLayout> result = tensor.sum(reduction_axis2);
VERIFY_IS_EQUAL(result.dimension(0), 2);
VERIFY_IS_EQUAL(result.dimension(1), 5);
for (int i = 0; i < 2; ++i) {
for (int j = 0; j < 5; ++j) {
float sum = 0.0f;
for (int k = 0; k < 3; ++k) {
for (int l = 0; l < 7; ++l) {
sum += tensor(i, k, j, l);
}
}
VERIFY_IS_APPROX(result(i, j), sum);
}
}
{
Tensor<float, 0, DataLayout> sum1 = tensor.sum();
VERIFY_IS_EQUAL(sum1.rank(), 0);
array<ptrdiff_t, 4> reduction_axis4;
reduction_axis4[0] = 0;
reduction_axis4[1] = 1;
reduction_axis4[2] = 2;
reduction_axis4[3] = 3;
Tensor<float, 0, DataLayout> sum2 = tensor.sum(reduction_axis4);
VERIFY_IS_EQUAL(sum2.rank(), 0);
VERIFY_IS_APPROX(sum1(), sum2());
}
reduction_axis2[0] = 0;
reduction_axis2[1] = 2;
result = tensor.prod(reduction_axis2);
VERIFY_IS_EQUAL(result.dimension(0), 3);
VERIFY_IS_EQUAL(result.dimension(1), 7);
for (int i = 0; i < 3; ++i) {
for (int j = 0; j < 7; ++j) {
float prod = 1.0f;
for (int k = 0; k < 2; ++k) {
for (int l = 0; l < 5; ++l) {
prod *= tensor(k, i, l, j);
}
}
VERIFY_IS_APPROX(result(i, j), prod);
}
}
{
Tensor<float, 0, DataLayout> prod1 = tensor.prod();
VERIFY_IS_EQUAL(prod1.rank(), 0);
array<ptrdiff_t, 4> reduction_axis4;
reduction_axis4[0] = 0;
reduction_axis4[1] = 1;
reduction_axis4[2] = 2;
reduction_axis4[3] = 3;
Tensor<float, 0, DataLayout> prod2 = tensor.prod(reduction_axis4);
VERIFY_IS_EQUAL(prod2.rank(), 0);
VERIFY_IS_APPROX(prod1(), prod2());
}
reduction_axis2[0] = 0;
reduction_axis2[1] = 2;
result = tensor.maximum(reduction_axis2);
VERIFY_IS_EQUAL(result.dimension(0), 3);
VERIFY_IS_EQUAL(result.dimension(1), 7);
for (int i = 0; i < 3; ++i) {
for (int j = 0; j < 7; ++j) {
float max_val = std::numeric_limits<float>::lowest();
for (int k = 0; k < 2; ++k) {
for (int l = 0; l < 5; ++l) {
max_val = (std::max)(max_val, tensor(k, i, l, j));
}
}
VERIFY_IS_APPROX(result(i, j), max_val);
}
}
{
Tensor<float, 0, DataLayout> max1 = tensor.maximum();
VERIFY_IS_EQUAL(max1.rank(), 0);
array<ptrdiff_t, 4> reduction_axis4;
reduction_axis4[0] = 0;
reduction_axis4[1] = 1;
reduction_axis4[2] = 2;
reduction_axis4[3] = 3;
Tensor<float, 0, DataLayout> max2 = tensor.maximum(reduction_axis4);
VERIFY_IS_EQUAL(max2.rank(), 0);
VERIFY_IS_APPROX(max1(), max2());
}
reduction_axis2[0] = 0;
reduction_axis2[1] = 1;
result = tensor.minimum(reduction_axis2);
VERIFY_IS_EQUAL(result.dimension(0), 5);
VERIFY_IS_EQUAL(result.dimension(1), 7);
for (int i = 0; i < 5; ++i) {
for (int j = 0; j < 7; ++j) {
float min_val = (std::numeric_limits<float>::max)();
for (int k = 0; k < 2; ++k) {
for (int l = 0; l < 3; ++l) {
min_val = (std::min)(min_val, tensor(k, l, i, j));
}
}
VERIFY_IS_APPROX(result(i, j), min_val);
}
}
{
Tensor<float, 0, DataLayout> min1 = tensor.minimum();
VERIFY_IS_EQUAL(min1.rank(), 0);
array<ptrdiff_t, 4> reduction_axis4;
reduction_axis4[0] = 0;
reduction_axis4[1] = 1;
reduction_axis4[2] = 2;
reduction_axis4[3] = 3;
Tensor<float, 0, DataLayout> min2 = tensor.minimum(reduction_axis4);
VERIFY_IS_EQUAL(min2.rank(), 0);
VERIFY_IS_APPROX(min1(), min2());
}
reduction_axis2[0] = 0;
reduction_axis2[1] = 1;
result = tensor.mean(reduction_axis2);
VERIFY_IS_EQUAL(result.dimension(0), 5);
VERIFY_IS_EQUAL(result.dimension(1), 7);
for (int i = 0; i < 5; ++i) {
for (int j = 0; j < 7; ++j) {
float sum = 0.0f;
int count = 0;
for (int k = 0; k < 2; ++k) {
for (int l = 0; l < 3; ++l) {
sum += tensor(k, l, i, j);
++count;
}
}
VERIFY_IS_APPROX(result(i, j), sum / count);
}
}
{
Tensor<float, 0, DataLayout> mean1 = tensor.mean();
VERIFY_IS_EQUAL(mean1.rank(), 0);
array<ptrdiff_t, 4> reduction_axis4;
reduction_axis4[0] = 0;
reduction_axis4[1] = 1;
reduction_axis4[2] = 2;
reduction_axis4[3] = 3;
Tensor<float, 0, DataLayout> mean2 = tensor.mean(reduction_axis4);
VERIFY_IS_EQUAL(mean2.rank(), 0);
VERIFY_IS_APPROX(mean1(), mean2());
}
{
Tensor<int, 1> ints(10);
std::iota(ints.data(), ints.data() + ints.dimension(0), 0);
TensorFixedSize<bool, Sizes<> > all;
all = ints.all();
VERIFY(!all());
all = (ints >= ints.constant(0)).all();
VERIFY(all());
TensorFixedSize<bool, Sizes<> > any;
any = (ints > ints.constant(10)).any();
VERIFY(!any());
any = (ints < ints.constant(1)).any();
VERIFY(any());
}
}
Tensor embedding_bag_backward_cpu(const Tensor &grad_, const Tensor &indices__,
const Tensor &offsets__,
const Tensor &offset2bag__,
const Tensor &bag_size_,
const Tensor& max_indices_, int64_t num_weights,
bool scale_grad_by_freq, int64_t mode) {
auto grad = grad_.contiguous();
auto grad_arg = TensorArg(grad, "grad_", 1);
checkScalarTypes("embedding_bag", grad_arg, {kFloat, kDouble});
auto indices_arg = TensorArg(indices__, "indices__", 1);
checkScalarType("embedding_bag", indices_arg, kLong);
auto offsets_arg = TensorArg(offsets__, "offsets__", 1);
checkScalarType("embedding_bag", offsets_arg, kLong);
auto offset2bag_arg = TensorArg(offset2bag__, "offset2bag__", 1);
checkScalarType("embedding_bag", offset2bag_arg, kLong);
checkContiguous("embedding_bag", offset2bag_arg);
Tensor indices_ = indices__.contiguous();
Tensor offsets_ = offsets__.contiguous();
Tensor &offset2bag_ = const_cast<Tensor &>(offset2bag__);
auto ind_sort_ = indices_.sort();
auto indices = std::get<0>(ind_sort_);
auto ind_sort = std::get<1>(ind_sort_);
auto offset2bag = offset2bag_.index_select(0, ind_sort);
auto indices_data = indices.data<int64_t>();
auto offsets_data = offsets_.data<int64_t>();
auto offset2bag_data = offset2bag.data<int64_t>();
int64_t numel = indices.numel();
std::vector<int64_t> counts(num_weights);
for (int i = 0; i < numel; i++) {
counts[indices_data[i]] = 0;
}
for (int i = 0; i < numel; i++) {
counts[indices_data[i]]++;
}
auto index_grad_weight =
at::zeros({num_weights, grad.size(1)}, grad.type()).contiguous();
std::vector<int64_t> counts_uniq;
counts_uniq.reserve(num_weights);
int64_t o = 0;
for (int64_t i = 0; i < numel; i += counts[indices_data[i]]) {
counts_uniq.push_back(counts[indices_data[i]]);
if (o > 0) {
counts_uniq[o] += counts_uniq[o - 1];
}
o++;
}
if (mode == MODE_MEAN || mode == MODE_SUM) {
#pragma omp parallel for if (numel > 1000)
for (int64_t i = 0; i < (int64_t)counts_uniq.size(); i++) {
int64_t start = i == 0 ? 0 : counts_uniq[i - 1];
int64_t index = indices_data[start];
for (int64_t j = start; j < counts_uniq[i]; j++) {
int64_t source = offset2bag_data[j];
double scale = 1.0;
if (scale_grad_by_freq) {
scale /= counts[indices_data[i]];
}
if (mode == 1) { // MODE_MEAN
if (offsets_.size(0) == 1) {
auto bag_size = indices.size(0);
scale /= bag_size;
} else {
if (source == offsets_.size(0) - 1) {
scale /= indices.size(0) - offsets_data[offsets_.size(0) - 1];
} else {
scale /= offsets_data[source + 1] - offsets_data[source];
}
}
}
int64_t ddim = grad.size(1);
if (grad.type().scalarType() == kFloat) {
auto igwd = index_grad_weight.data<float>();
auto gd = grad.data<float>();
THBlas_axpy<float>(ddim, (float)scale, gd + ddim * source, 1,
igwd + ddim * index, 1);
} else if (grad.type().scalarType() == kDouble) {
auto igwd = index_grad_weight.data<double>();
auto gd = grad.data<double>();
THBlas_axpy<double>(ddim, (double)scale, gd + ddim * source, 1,
igwd + ddim * index, 1);
}
}
}
} else if (mode == MODE_MAX) {
auto nonempty_max_indices = max_indices_.index_select(0, bag_size_.nonzero().view(-1));
auto nonempty_grad = grad_.index_select(0, bag_size_.nonzero().view(-1));
for (int64_t dim = 0; dim < grad.size(1); dim++) {
index_grad_weight.select(1, dim).index_add_(
0, nonempty_max_indices.select(1, dim), nonempty_grad.select(1, dim));
}
}
return index_grad_weight;
}
void MDAtomsTyped<T>::getBox(Tensor&box)const{
if(this->box) for(int i=0;i<3;i++)for(int j=0;j<3;j++) box(i,j)=this->box[3*i+j]*scaleb;
else box.zero();
}
void CH3Shifts::calculate()
{
double energy=0.;
Tensor virial;
virial.zero();
vector<Vector> deriv(getNumberOfAtoms());
int N = getNumberOfAtoms();
Coor<double> coor(N);
Coor<double> forces(N);
forces.clear();
for(int i=0; i<numResidues; i++) for(unsigned j=0; j<6; j++) sh[i][j]=0.;
for (int i = 0; i < N; i++) {
int ipos = 4 * i;
Vector Pos = getPosition(i);
coor.coor[ipos] = len_pl2alm*Pos[0];
coor.coor[ipos+1] = len_pl2alm*Pos[1];
coor.coor[ipos+2] = len_pl2alm*Pos[2];
}
double fact=1.0;
if(!ensemble) {
energy = meth_list[0]->calc_cs_force(coor, forces);
bool printout=false;
if(pperiod>0&&comm.Get_rank()==0) printout = (!(getStep()%pperiod));
if(printout) {
string csfile;
char tmps1[21];
// add to the name the label of the cv in such a way to have different files
// when there is more than one defined variable
sprintf(tmps1, "%li", getStep());
csfile = string("cs")+tmps1+string(".dat");
meth_list[0]->write_cs(csfile.c_str());
}
} else {
meth_list[0]->calc_cs(coor);
bool printout=false;
if(pperiod>0&&comm.Get_rank()==0) printout = (!(getStep()%pperiod));
if(printout) {
string csfile;
char tmps1[21], tmps2[21];
// add to the name the label of the cv in such a way to have different files
// when there is more than one defined variable
sprintf(tmps1, "%li", getStep());
sprintf(tmps2, "%i", multi_sim_comm.Get_rank());
csfile = string("cs")+tmps2+"-"+tmps1+string(".dat");
meth_list[0]->write_cs(csfile.c_str());
}
unsigned size = meth_list[0]->ala_calc_hb.size();
for(unsigned j=0;j<size;j++) sh[0][j] = meth_list[0]->ala_calc_hb[j];
size = meth_list[0]->ile_calc_hd.size();
for(unsigned j=0;j<size;j++) sh[1][j] = meth_list[0]->ile_calc_hd[j];
size = meth_list[0]->ile_calc_hg2.size();
for(unsigned j=0;j<size;j++) sh[2][j] = meth_list[0]->ile_calc_hg2[j];
size = meth_list[0]->leu_calc_hd1.size();
for(unsigned j=0;j<size;j++) sh[3][j] = meth_list[0]->leu_calc_hd1[j];
size = meth_list[0]->leu_calc_hd2.size();
for(unsigned j=0;j<size;j++) sh[4][j] = meth_list[0]->leu_calc_hd2[j];
size = meth_list[0]->thr_calc_hg2.size();
for(unsigned j=0;j<size;j++) sh[5][j] = meth_list[0]->thr_calc_hg2[j];
size = meth_list[0]->val_calc_hg1.size();
for(unsigned j=0;j<size;j++) sh[6][j] = meth_list[0]->val_calc_hg1[j];
size = meth_list[0]->val_calc_hg2.size();
for(unsigned j=0;j<size;j++) sh[7][j] = meth_list[0]->val_calc_hg2[j];
fact = 1./((double) ens_dim);
if(comm.Get_rank()==0) { // I am the master of my replica
// among replicas
multi_sim_comm.Sum(&sh[0][0], numResidues*8);
multi_sim_comm.Barrier();
for(unsigned i=0;i<8;i++) for(int j=0;j<numResidues;j++) sh[j][i] *= fact;
} else for(unsigned i=0;i<8;i++) for(int j=0;j<numResidues;j++) sh[j][i] = 0.;
// inside each replica
comm.Sum(&sh[0][0], numResidues*8);
// now send the averaged shifts back to almost
size = meth_list[0]->ala_calc_hb.size();
for(unsigned j=0;j<size;j++) meth_list[0]->ala_calc_hb[j] = sh[0][j];
size = meth_list[0]->ile_calc_hd.size();
for(unsigned j=0;j<size;j++) meth_list[0]->ile_calc_hd[j] = sh[1][j];
size = meth_list[0]->ile_calc_hg2.size();
for(unsigned j=0;j<size;j++) meth_list[0]->ile_calc_hg2[j] = sh[2][j];
size = meth_list[0]->leu_calc_hd1.size();
for(unsigned j=0;j<size;j++) meth_list[0]->leu_calc_hd1[j] = sh[3][j];
size = meth_list[0]->leu_calc_hd2.size();
for(unsigned j=0;j<size;j++) meth_list[0]->leu_calc_hd2[j] = sh[4][j];
size = meth_list[0]->thr_calc_hg2.size();
for(unsigned j=0;j<size;j++) meth_list[0]->thr_calc_hg2[j] = sh[5][j];
size = meth_list[0]->val_calc_hg1.size();
for(unsigned j=0;j<size;j++) meth_list[0]->val_calc_hg1[j] = sh[6][j];
size = meth_list[0]->val_calc_hg2.size();
for(unsigned j=0;j<size;j++) meth_list[0]->val_calc_hg2[j] = sh[7][j];
// calculate all the forces now
energy = meth_list[0]->ens_calc_cs_force(coor, forces);
}
for (int i = 0; i < N; i++)
{
Vector For;
int ipos = 4 * i;
For[0] = forces.coor[ipos];
For[1] = forces.coor[ipos+1];
For[2] = forces.coor[ipos+2];
deriv[i] = fact*for_pl2alm*For;
virial=virial+(-1.*Tensor(getPosition(i),deriv[i]));
}
for(unsigned i=0;i<getNumberOfAtoms();++i) setAtomsDerivatives(i,deriv[i]);
setValue (ene_pl2alm*energy);
setBoxDerivatives (virial);
}
Tensor< S > Tensor< S >::apply(
const Tensor &x,
const ::std::function< void(value_type*) > &lambda) {
return x.clone().apply(lambda);
}
bool Tensor::ComputeJointSVD(Tensor& Umat, vector<Tensor*>& extra_tensor_list, vector<int>& mult_modes, int nonzerovals)
{
int num_sing_vals = (int)pow((double)nonzerovals, (double)mult_modes.size());
vector<int> free_modes;
vector<int> mult_dims;
vector<int> free_dims;
vector<int> mult_offsets;
vector<int> free_offsets;
Tensor& temp_tensor = *(extra_tensor_list.at(0));
VectorPlus::SetDiff(free_modes, temp_tensor.Modes(), mult_modes);
VectorPlus::Subset(mult_dims, temp_tensor.Dims(), mult_modes);
VectorPlus::CSubset(free_dims, temp_tensor.Dims(), mult_modes);
ComputeOffsets(mult_offsets, mult_dims);
ComputeOffsets(free_offsets, free_dims);
vector<int> usmalldims(free_modes.size(), nonzerovals);
vector<int> udims;
vector<int> usmall_offsets;
ComputeOffsets(usmall_offsets, usmalldims);
int numMultElements = VectorPlus::Product(mult_dims);
int numFreeElements = VectorPlus::Product(free_dims);
assert(numMultElements == numFreeElements);
Eigen::MatrixXd matricized_tensor(numFreeElements,extra_tensor_list.size() * numMultElements);
//cout << "copy start 1\n";
for (int z = 0; z < extra_tensor_list.size(); ++z)
{
int z_offset = z * numMultElements;
FastIndexer i_indexer(free_dims);
for (int i = 0; i < numFreeElements; ++i)
{
vector<int>& free_indices = i_indexer.GetNext();
// ComputeIndexArray(free_indices, free_offsets, i);
FastIndexer j_indexer(mult_dims);
for (int j = 0; j < numMultElements; ++j)
{
vector<int>& mult_indices = j_indexer.GetNext();
// ComputeIndexArray(mult_indices, mult_offsets, j);
vector<int> total_indices;
VectorPlus::Concat(total_indices, free_indices, mult_indices);
matricized_tensor(i,z_offset + j) = extra_tensor_list.at(z)->At(total_indices);
}
}
}
// cout << "copy end 1\n";
// MatrixXd matricized_inverse = matricized_tensor.inverse();
// cout << matricized_inverse;
// cout << "\n";
//compute pseudoinverse
//cout << "svd start 1\n";
Eigen::JacobiSVD<Eigen::MatrixXd> svd(matricized_tensor, Eigen::ComputeFullU);
Eigen::MatrixXd U = svd.matrixU();
// cout << "svd end 1\n";
Eigen::MatrixXd thinU = U.leftCols(num_sing_vals);
VectorPlus::Concat(udims, free_dims, usmalldims);
Umat.Initialize(udims);
vector<int> semi_dims;
semi_dims.push_back(thinU.rows());
semi_dims.push_back(thinU.cols());
// cout << "copy start 2 \n";
FastIndexer i_indexer(free_dims);
for (int i = 0; i < thinU.rows(); ++i)
{
vector<int>& left_indices = i_indexer.GetNext();
// ComputeIndexArray(left_indices, free_offsets, i);
FastIndexer j_indexer(usmalldims);
for (int j = 0; j < thinU.cols(); ++j)
{
vector<int>& right_indices = j_indexer.GetNext();
// ComputeIndexArray(right_indices, usmall_offsets, j);
vector<int> indices;
VectorPlus::Concat(indices, left_indices, right_indices);
// Umat.Set(indices, rand_matrix.At(i,j));
Umat.Set(indices, thinU(i,j));
if (thinU.rows() == thinU.cols())
{
if (VectorPlus::Equals(left_indices, right_indices))
{
Umat.Set(indices, 1);
}
else
{
Umat.Set(indices, 0);
}
}
}
}
// cout << "copy end 2 \n";
return true;
}
void HPCoarsenTest::select_refinement (System & system)
{
START_LOG("select_refinement()", "HPCoarsenTest");
// The current mesh
MeshBase & mesh = system.get_mesh();
// The dimensionality of the mesh
const unsigned int dim = mesh.mesh_dimension();
// The number of variables in the system
const unsigned int n_vars = system.n_vars();
// The DofMap for this system
const DofMap & dof_map = system.get_dof_map();
// The system number (for doing bad hackery)
const unsigned int sys_num = system.number();
// Check for a valid component_scale
if (!component_scale.empty())
{
if (component_scale.size() != n_vars)
libmesh_error_msg("ERROR: component_scale is the wrong size:\n" \
<< " component_scale.size()=" \
<< component_scale.size() \
<< "\n n_vars=" \
<< n_vars);
}
else
{
// No specified scaling. Scale all variables by one.
component_scale.resize (n_vars, 1.0);
}
// Resize the error_per_cell vectors to handle
// the number of elements, initialize them to 0.
std::vector<ErrorVectorReal> h_error_per_cell(mesh.max_elem_id(), 0.);
std::vector<ErrorVectorReal> p_error_per_cell(mesh.max_elem_id(), 0.);
// Loop over all the variables in the system
for (unsigned int var=0; var<n_vars; var++)
{
// Possibly skip this variable
if (!component_scale.empty())
if (component_scale[var] == 0.0) continue;
// The type of finite element to use for this variable
const FEType & fe_type = dof_map.variable_type (var);
// Finite element objects for a fine (and probably a coarse)
// element will be needed
fe = FEBase::build (dim, fe_type);
fe_coarse = FEBase::build (dim, fe_type);
// Any cached coarse element results have expired
coarse = libmesh_nullptr;
unsigned int cached_coarse_p_level = 0;
const FEContinuity cont = fe->get_continuity();
libmesh_assert (cont == DISCONTINUOUS || cont == C_ZERO ||
cont == C_ONE);
// Build an appropriate quadrature rule
qrule = fe_type.default_quadrature_rule(dim);
// Tell the refined finite element about the quadrature
// rule. The coarse finite element need not know about it
fe->attach_quadrature_rule (qrule.get());
// We will always do the integration
// on the fine elements. Get their Jacobian values, etc..
JxW = &(fe->get_JxW());
xyz_values = &(fe->get_xyz());
// The shape functions
phi = &(fe->get_phi());
phi_coarse = &(fe_coarse->get_phi());
// The shape function derivatives
if (cont == C_ZERO || cont == C_ONE)
{
dphi = &(fe->get_dphi());
dphi_coarse = &(fe_coarse->get_dphi());
}
#ifdef LIBMESH_ENABLE_SECOND_DERIVATIVES
// The shape function second derivatives
if (cont == C_ONE)
{
d2phi = &(fe->get_d2phi());
d2phi_coarse = &(fe_coarse->get_d2phi());
}
#endif // defined (LIBMESH_ENABLE_SECOND_DERIVATIVES)
// Iterate over all the active elements in the mesh
// that live on this processor.
MeshBase::const_element_iterator elem_it =
mesh.active_local_elements_begin();
const MeshBase::const_element_iterator elem_end =
mesh.active_local_elements_end();
for (; elem_it != elem_end; ++elem_it)
{
const Elem * elem = *elem_it;
// We're only checking elements that are already flagged for h
// refinement
if (elem->refinement_flag() != Elem::REFINE)
continue;
const dof_id_type e_id = elem->id();
// Find the projection onto the parent element,
// if necessary
if (elem->parent() &&
(coarse != elem->parent() ||
cached_coarse_p_level != elem->p_level()))
{
Uc.resize(0);
coarse = elem->parent();
cached_coarse_p_level = elem->p_level();
unsigned int old_parent_level = coarse->p_level();
(const_cast<Elem *>(coarse))->hack_p_level(elem->p_level());
this->add_projection(system, coarse, var);
(const_cast<Elem *>(coarse))->hack_p_level(old_parent_level);
// Solve the h-coarsening projection problem
Ke.cholesky_solve(Fe, Uc);
}
fe->reinit(elem);
// Get the DOF indices for the fine element
dof_map.dof_indices (elem, dof_indices, var);
// The number of quadrature points
const unsigned int n_qp = qrule->n_points();
// The number of DOFS on the fine element
const unsigned int n_dofs =
cast_int<unsigned int>(dof_indices.size());
// The number of nodes on the fine element
const unsigned int n_nodes = elem->n_nodes();
// The average element value (used as an ugly hack
// when we have nothing p-coarsened to compare to)
// Real average_val = 0.;
Number average_val = 0.;
// Calculate this variable's contribution to the p
// refinement error
if (elem->p_level() == 0)
{
unsigned int n_vertices = 0;
for (unsigned int n = 0; n != n_nodes; ++n)
if (elem->is_vertex(n))
{
n_vertices++;
const Node * const node = elem->get_node(n);
average_val += system.current_solution
(node->dof_number(sys_num,var,0));
}
average_val /= n_vertices;
}
else
{
unsigned int old_elem_level = elem->p_level();
(const_cast<Elem *>(elem))->hack_p_level(old_elem_level - 1);
fe_coarse->reinit(elem, &(qrule->get_points()));
const unsigned int n_coarse_dofs =
cast_int<unsigned int>(phi_coarse->size());
(const_cast<Elem *>(elem))->hack_p_level(old_elem_level);
Ke.resize(n_coarse_dofs, n_coarse_dofs);
Ke.zero();
Fe.resize(n_coarse_dofs);
Fe.zero();
// Loop over the quadrature points
for (unsigned int qp=0; qp<qrule->n_points(); qp++)
{
// The solution value at the quadrature point
Number val = libMesh::zero;
Gradient grad;
Tensor hess;
for (unsigned int i=0; i != n_dofs; i++)
{
dof_id_type dof_num = dof_indices[i];
val += (*phi)[i][qp] *
system.current_solution(dof_num);
if (cont == C_ZERO || cont == C_ONE)
grad.add_scaled((*dphi)[i][qp], system.current_solution(dof_num));
// grad += (*dphi)[i][qp] *
// system.current_solution(dof_num);
if (cont == C_ONE)
hess.add_scaled((*d2phi)[i][qp], system.current_solution(dof_num));
// hess += (*d2phi)[i][qp] *
// system.current_solution(dof_num);
}
// The projection matrix and vector
for (unsigned int i=0; i != Fe.size(); ++i)
{
Fe(i) += (*JxW)[qp] *
(*phi_coarse)[i][qp]*val;
if (cont == C_ZERO || cont == C_ONE)
Fe(i) += (*JxW)[qp] *
grad * (*dphi_coarse)[i][qp];
if (cont == C_ONE)
Fe(i) += (*JxW)[qp] *
hess.contract((*d2phi_coarse)[i][qp]);
for (unsigned int j=0; j != Fe.size(); ++j)
{
Ke(i,j) += (*JxW)[qp] *
(*phi_coarse)[i][qp]*(*phi_coarse)[j][qp];
if (cont == C_ZERO || cont == C_ONE)
Ke(i,j) += (*JxW)[qp] *
(*dphi_coarse)[i][qp]*(*dphi_coarse)[j][qp];
if (cont == C_ONE)
Ke(i,j) += (*JxW)[qp] *
((*d2phi_coarse)[i][qp].contract((*d2phi_coarse)[j][qp]));
}
}
}
// Solve the p-coarsening projection problem
Ke.cholesky_solve(Fe, Up);
}
// loop over the integration points on the fine element
for (unsigned int qp=0; qp<n_qp; qp++)
{
Number value_error = 0.;
Gradient grad_error;
Tensor hessian_error;
for (unsigned int i=0; i<n_dofs; i++)
{
const dof_id_type dof_num = dof_indices[i];
value_error += (*phi)[i][qp] *
system.current_solution(dof_num);
if (cont == C_ZERO || cont == C_ONE)
grad_error.add_scaled((*dphi)[i][qp], system.current_solution(dof_num));
// grad_error += (*dphi)[i][qp] *
// system.current_solution(dof_num);
if (cont == C_ONE)
hessian_error.add_scaled((*d2phi)[i][qp], system.current_solution(dof_num));
// hessian_error += (*d2phi)[i][qp] *
// system.current_solution(dof_num);
}
if (elem->p_level() == 0)
{
value_error -= average_val;
}
else
{
for (unsigned int i=0; i<Up.size(); i++)
{
value_error -= (*phi_coarse)[i][qp] * Up(i);
if (cont == C_ZERO || cont == C_ONE)
grad_error.subtract_scaled((*dphi_coarse)[i][qp], Up(i));
// grad_error -= (*dphi_coarse)[i][qp] * Up(i);
if (cont == C_ONE)
hessian_error.subtract_scaled((*d2phi_coarse)[i][qp], Up(i));
// hessian_error -= (*d2phi_coarse)[i][qp] * Up(i);
}
}
p_error_per_cell[e_id] += static_cast<ErrorVectorReal>
(component_scale[var] *
(*JxW)[qp] * TensorTools::norm_sq(value_error));
if (cont == C_ZERO || cont == C_ONE)
p_error_per_cell[e_id] += static_cast<ErrorVectorReal>
(component_scale[var] *
(*JxW)[qp] * grad_error.norm_sq());
if (cont == C_ONE)
p_error_per_cell[e_id] += static_cast<ErrorVectorReal>
(component_scale[var] *
(*JxW)[qp] * hessian_error.norm_sq());
}
// Calculate this variable's contribution to the h
// refinement error
if (!elem->parent())
{
// For now, we'll always start with an h refinement
h_error_per_cell[e_id] =
std::numeric_limits<ErrorVectorReal>::max() / 2;
}
else
{
FEInterface::inverse_map (dim, fe_type, coarse,
*xyz_values, coarse_qpoints);
unsigned int old_parent_level = coarse->p_level();
(const_cast<Elem *>(coarse))->hack_p_level(elem->p_level());
fe_coarse->reinit(coarse, &coarse_qpoints);
(const_cast<Elem *>(coarse))->hack_p_level(old_parent_level);
// The number of DOFS on the coarse element
unsigned int n_coarse_dofs =
cast_int<unsigned int>(phi_coarse->size());
// Loop over the quadrature points
for (unsigned int qp=0; qp<n_qp; qp++)
{
// The solution difference at the quadrature point
Number value_error = libMesh::zero;
Gradient grad_error;
Tensor hessian_error;
for (unsigned int i=0; i != n_dofs; ++i)
{
const dof_id_type dof_num = dof_indices[i];
value_error += (*phi)[i][qp] *
system.current_solution(dof_num);
if (cont == C_ZERO || cont == C_ONE)
grad_error.add_scaled((*dphi)[i][qp], system.current_solution(dof_num));
// grad_error += (*dphi)[i][qp] *
// system.current_solution(dof_num);
if (cont == C_ONE)
hessian_error.add_scaled((*d2phi)[i][qp], system.current_solution(dof_num));
// hessian_error += (*d2phi)[i][qp] *
// system.current_solution(dof_num);
}
for (unsigned int i=0; i != n_coarse_dofs; ++i)
{
value_error -= (*phi_coarse)[i][qp] * Uc(i);
if (cont == C_ZERO || cont == C_ONE)
// grad_error -= (*dphi_coarse)[i][qp] * Uc(i);
grad_error.subtract_scaled((*dphi_coarse)[i][qp], Uc(i));
if (cont == C_ONE)
hessian_error.subtract_scaled((*d2phi_coarse)[i][qp], Uc(i));
// hessian_error -= (*d2phi_coarse)[i][qp] * Uc(i);
}
h_error_per_cell[e_id] += static_cast<ErrorVectorReal>
(component_scale[var] *
(*JxW)[qp] * TensorTools::norm_sq(value_error));
if (cont == C_ZERO || cont == C_ONE)
h_error_per_cell[e_id] += static_cast<ErrorVectorReal>
(component_scale[var] *
(*JxW)[qp] * grad_error.norm_sq());
if (cont == C_ONE)
h_error_per_cell[e_id] += static_cast<ErrorVectorReal>
(component_scale[var] *
(*JxW)[qp] * hessian_error.norm_sq());
}
}
}
}
// Now that we've got our approximations for p_error and h_error, let's see
// if we want to switch any h refinement flags to p refinement
// Iterate over all the active elements in the mesh
// that live on this processor.
MeshBase::element_iterator elem_it =
mesh.active_local_elements_begin();
const MeshBase::element_iterator elem_end =
mesh.active_local_elements_end();
for (; elem_it != elem_end; ++elem_it)
{
Elem * elem = *elem_it;
// We're only checking elements that are already flagged for h
// refinement
if (elem->refinement_flag() != Elem::REFINE)
continue;
const dof_id_type e_id = elem->id();
unsigned int dofs_per_elem = 0, dofs_per_p_elem = 0;
// Loop over all the variables in the system
for (unsigned int var=0; var<n_vars; var++)
{
// The type of finite element to use for this variable
const FEType & fe_type = dof_map.variable_type (var);
// FIXME: we're overestimating the number of DOFs added by h
// refinement
FEType elem_fe_type = fe_type;
elem_fe_type.order =
static_cast<Order>(fe_type.order + elem->p_level());
dofs_per_elem +=
FEInterface::n_dofs(dim, elem_fe_type, elem->type());
elem_fe_type.order =
static_cast<Order>(fe_type.order + elem->p_level() + 1);
dofs_per_p_elem +=
FEInterface::n_dofs(dim, elem_fe_type, elem->type());
}
const unsigned int new_h_dofs = dofs_per_elem *
(elem->n_children() - 1);
const unsigned int new_p_dofs = dofs_per_p_elem -
dofs_per_elem;
/*
libMesh::err << "Cell " << e_id << ": h = " << elem->hmax()
<< ", p = " << elem->p_level() + 1 << "," << std::endl
<< " h_error = " << h_error_per_cell[e_id]
<< ", p_error = " << p_error_per_cell[e_id] << std::endl
<< " new_h_dofs = " << new_h_dofs
<< ", new_p_dofs = " << new_p_dofs << std::endl;
*/
const Real p_value =
std::sqrt(p_error_per_cell[e_id]) * p_weight / new_p_dofs;
const Real h_value =
std::sqrt(h_error_per_cell[e_id]) /
static_cast<Real>(new_h_dofs);
if (p_value > h_value)
{
elem->set_p_refinement_flag(Elem::REFINE);
elem->set_refinement_flag(Elem::DO_NOTHING);
}
}
STOP_LOG("select_refinement()", "HPCoarsenTest");
}
