void RevJacSweep(
bool nz_compare,
size_t n,
size_t numvar,
player<Base> *play,
Vector_set& var_sparsity
)
{
OpCode op;
size_t i_op;
size_t i_var;
const addr_t* arg = CPPAD_NULL;
size_t i, j, k;
// length of the parameter vector (used by CppAD assert macros)
const size_t num_par = play->num_par_rec();
// check numvar argument
CPPAD_ASSERT_UNKNOWN( numvar > 0 );
CPPAD_ASSERT_UNKNOWN( play->num_var_rec() == numvar );
CPPAD_ASSERT_UNKNOWN( var_sparsity.n_set() == numvar );
// upper limit (exclusive) for elements in the set
size_t limit = var_sparsity.end();
// vecad_sparsity contains a sparsity pattern for each VecAD object.
// vecad_ind maps a VecAD index (beginning of the VecAD object)
// to the index of the corresponding set in vecad_sparsity.
size_t num_vecad_ind = play->num_vec_ind_rec();
size_t num_vecad_vec = play->num_vecad_vec_rec();
Vector_set vecad_sparsity;
vecad_sparsity.resize(num_vecad_vec, limit);
pod_vector<size_t> vecad_ind;
if( num_vecad_vec > 0 )
{ size_t length;
vecad_ind.extend(num_vecad_ind);
j = 0;
for(i = 0; i < num_vecad_vec; i++)
{ // length of this VecAD
length = play->GetVecInd(j);
// set to proper index for this VecAD
vecad_ind[j] = i;
for(k = 1; k <= length; k++)
vecad_ind[j+k] = num_vecad_vec; // invalid index
// start of next VecAD
j += length + 1;
}
CPPAD_ASSERT_UNKNOWN( j == play->num_vec_ind_rec() );
}
// work space used by UserOp.
typedef std::set<size_t> size_set;
size_set::iterator set_itr; // iterator for a standard set
size_set::iterator set_end; // end of iterator sequence
vector< size_set > set_r; // set sparsity pattern for the argument x
vector< size_set > set_s; // set sparisty pattern for the result y
//
vector<bool> bool_r; // bool sparsity pattern for the argument x
vector<bool> bool_s; // bool sparisty pattern for the result y
//
const size_t user_q = limit; // maximum element plus one
size_t user_index = 0; // indentifier for this atomic operation
size_t user_id = 0; // user identifier for this call to operator
size_t user_i = 0; // index in result vector
size_t user_j = 0; // index in argument vector
size_t user_m = 0; // size of result vector
size_t user_n = 0; // size of arugment vector
//
atomic_base<Base>* user_atom = CPPAD_NULL; // user's atomic op calculator
bool user_bool = false; // use bool or set sparsity ?
# ifndef NDEBUG
bool user_ok = false; // atomic op return value
# endif
// next expected operator in a UserOp sequence
enum { user_start, user_arg, user_ret, user_end } user_state = user_end;
// Initialize
play->reverse_start(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( op == EndOp );
# if CPPAD_REV_JAC_SWEEP_TRACE
std::cout << std::endl;
CppAD::vectorBool z_value(limit);
# endif
bool more_operators = true;
while(more_operators)
{
// next op
play->reverse_next(op, arg, i_op, i_var);
# ifndef NDEBUG
if( i_op <= n )
{ CPPAD_ASSERT_UNKNOWN((op == InvOp) | (op == BeginOp));
}
else CPPAD_ASSERT_UNKNOWN((op != InvOp) & (op != BeginOp));
# endif
// rest of information depends on the case
switch( op )
{
case AbsOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case AddvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1);
reverse_sparse_jacobian_binary_op(
i_var, arg, var_sparsity
);
break;
// -------------------------------------------------
case AddpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1);
reverse_sparse_jacobian_unary_op(
i_var, arg[1], var_sparsity
);
break;
// -------------------------------------------------
case AcosOp:
// sqrt(1 - x * x), acos(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case AsinOp:
// sqrt(1 - x * x), asin(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case AtanOp:
// 1 + x * x, atan(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case BeginOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1);
more_operators = false;
break;
// -------------------------------------------------
case CSkipOp:
// CSkipOp has a variable number of arguments and
// reverse_next thinks it one has one argument.
// We must inform reverse_next of this special case.
play->reverse_cskip(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case CSumOp:
// CSumOp has a variable number of arguments and
// reverse_next thinks it one has one argument.
// We must inform reverse_next of this special case.
play->reverse_csum(op, arg, i_op, i_var);
reverse_sparse_jacobian_csum_op(
i_var, arg, var_sparsity
);
break;
// -------------------------------------------------
case CExpOp:
reverse_sparse_jacobian_cond_op(
nz_compare, i_var, arg, num_par, var_sparsity
);
break;
// ---------------------------------------------------
case CosOp:
// sin(x), cos(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// ---------------------------------------------------
case CoshOp:
// sinh(x), cosh(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case DisOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1);
// derivative is identically zero
break;
// -------------------------------------------------
case DivvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1);
reverse_sparse_jacobian_binary_op(
i_var, arg, var_sparsity
);
break;
// -------------------------------------------------
case DivpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1);
reverse_sparse_jacobian_unary_op(
i_var, arg[1], var_sparsity
);
break;
// -------------------------------------------------
case DivvpOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case ErfOp:
// arg[1] is always the parameter 0
// arg[0] is always the parameter 2 / sqrt(pi)
CPPAD_ASSERT_NARG_NRES(op, 3, 5);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case ExpOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case InvOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 1);
break;
// -------------------------------------------------
case LdpOp:
reverse_sparse_jacobian_load_op(
op,
i_var,
arg,
num_vecad_ind,
vecad_ind.data(),
var_sparsity,
vecad_sparsity
);
break;
// -------------------------------------------------
case LdvOp:
reverse_sparse_jacobian_load_op(
op,
i_var,
arg,
num_vecad_ind,
vecad_ind.data(),
var_sparsity,
vecad_sparsity
);
break;
// -------------------------------------------------
case EqpvOp:
case EqvvOp:
case LtpvOp:
case LtvpOp:
case LtvvOp:
case LepvOp:
case LevpOp:
case LevvOp:
case NepvOp:
case NevvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 0);
break;
// -------------------------------------------------
case LogOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case MulvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1);
reverse_sparse_jacobian_binary_op(
i_var, arg, var_sparsity
);
break;
// -------------------------------------------------
case MulpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1);
reverse_sparse_jacobian_unary_op(
i_var, arg[1], var_sparsity
);
break;
// -------------------------------------------------
case ParOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1);
break;
// -------------------------------------------------
case PowvpOp:
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case PowpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3);
reverse_sparse_jacobian_unary_op(
i_var, arg[1], var_sparsity
);
break;
// -------------------------------------------------
case PowvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3);
reverse_sparse_jacobian_binary_op(
i_var, arg, var_sparsity
);
break;
// -------------------------------------------------
case PriOp:
CPPAD_ASSERT_NARG_NRES(op, 5, 0);
break;
// -------------------------------------------------
case SignOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1);
// derivative is identically zero
break;
// -------------------------------------------------
case SinOp:
// cos(x), sin(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case SinhOp:
// cosh(x), sinh(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case SqrtOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case StppOp:
// sparsity cannot proagate through a parameter
CPPAD_ASSERT_NARG_NRES(op, 3, 0);
break;
// -------------------------------------------------
case StpvOp:
reverse_sparse_jacobian_store_op(
op,
arg,
num_vecad_ind,
vecad_ind.data(),
var_sparsity,
vecad_sparsity
);
break;
// -------------------------------------------------
case StvpOp:
CPPAD_ASSERT_NARG_NRES(op, 3, 0);
break;
// -------------------------------------------------
case StvvOp:
reverse_sparse_jacobian_store_op(
op,
arg,
num_vecad_ind,
vecad_ind.data(),
var_sparsity,
vecad_sparsity
);
break;
// -------------------------------------------------
case SubvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1);
reverse_sparse_jacobian_binary_op(
i_var, arg, var_sparsity
);
break;
// -------------------------------------------------
case SubpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1);
reverse_sparse_jacobian_unary_op(
i_var, arg[1], var_sparsity
);
break;
// -------------------------------------------------
case SubvpOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case TanOp:
// tan(x)^2, tan(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case TanhOp:
// tanh(x)^2, tanh(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2);
reverse_sparse_jacobian_unary_op(
i_var, arg[0], var_sparsity
);
break;
// -------------------------------------------------
case UserOp:
// start or end atomic operation sequence
CPPAD_ASSERT_UNKNOWN( NumRes( UserOp ) == 0 );
CPPAD_ASSERT_UNKNOWN( NumArg( UserOp ) == 4 );
if( user_state == user_end )
{ user_index = arg[0];
user_id = arg[1];
user_n = arg[2];
user_m = arg[3];
user_atom = atomic_base<Base>::class_object(user_index);
# ifndef NDEBUG
if( user_atom == CPPAD_NULL )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base function has been deleted";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
user_bool = user_atom->sparsity() ==
atomic_base<Base>::bool_sparsity_enum;
if( user_bool )
{ if( bool_r.size() != user_m * user_q )
bool_r.resize( user_m * user_q );
if( bool_s.size() != user_n * user_q )
bool_s.resize( user_n * user_q );
for(i = 0; i < user_m; i++)
for(j = 0; j < user_q; j++)
bool_r[ i * user_q + j] = false;
}
else
{ if(set_r.size() != user_m )
set_r.resize(user_m);
if(set_s.size() != user_n )
set_s.resize(user_n);
for(i = 0; i < user_m; i++)
set_r[i].clear();
}
user_j = user_n;
user_i = user_m;
user_state = user_ret;
}
else
{ CPPAD_ASSERT_UNKNOWN( user_state == user_start );
CPPAD_ASSERT_UNKNOWN( user_index == size_t(arg[0]) );
CPPAD_ASSERT_UNKNOWN( user_id == size_t(arg[1]) );
CPPAD_ASSERT_UNKNOWN( user_n == size_t(arg[2]) );
CPPAD_ASSERT_UNKNOWN( user_m == size_t(arg[3]) );
# ifndef NDEBUG
if( ! user_ok )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base.rev_sparse_jac: returned false";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
user_state = user_end;
}
break;
case UsrapOp:
// parameter argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( 0 < user_j && user_j <= user_n );
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
--user_j;
if( user_j == 0 )
user_state = user_start;
break;
case UsravOp:
// variable argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( 0 < user_j && user_j <= user_n );
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) <= i_var );
CPPAD_ASSERT_UNKNOWN( 0 < arg[0] );
--user_j;
// It might be faster if we add set union to var_sparsity
// where one of the sets is not in var_sparsity.
if( user_bool )
{ for(j = 0; j < user_q; j++)
if( bool_s[ user_j * user_q + j ] )
var_sparsity.add_element(arg[0], j);
}
else
{ set_itr = set_s[user_j].begin();
set_end = set_s[user_j].end();
while( set_itr != set_end )
var_sparsity.add_element(arg[0], *set_itr++);
}
if( user_j == 0 )
user_state = user_start;
break;
case UsrrpOp:
// parameter result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( 0 < user_i && user_i <= user_m );
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
--user_i;
if( user_i == 0 )
{ // call users function for this operation
user_atom->set_id(user_id);
if( user_bool)
CPPAD_ATOMIC_CALL(
user_q, bool_r, bool_s
);
else
CPPAD_ATOMIC_CALL(
user_q, set_r, set_s
);
user_state = user_arg;
}
break;
case UsrrvOp:
// variable result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( 0 < user_i && user_i <= user_m );
--user_i;
var_sparsity.begin(i_var);
i = var_sparsity.next_element();
while( i < user_q )
{ if( user_bool )
bool_r[ user_i * user_q + i ] = true;
else
set_r[user_i].insert(i);
i = var_sparsity.next_element();
}
if( user_i == 0 )
{ // call users function for this operation
user_atom->set_id(user_id);
if( user_bool)
CPPAD_ATOMIC_CALL(
user_q, bool_r, bool_s
);
else
CPPAD_ATOMIC_CALL(
user_q, set_r, set_s
);
user_state = user_arg;
}
break;
// -------------------------------------------------
default:
CPPAD_ASSERT_UNKNOWN(0);
}
# if CPPAD_REV_JAC_SWEEP_TRACE
for(j = 0; j < limit; j++)
z_value[j] = false;
var_sparsity.begin(i_var);
j = var_sparsity.next_element();
while( j < limit )
{ z_value[j] = true;
j = var_sparsity.next_element();
}
printOp(
std::cout,
play,
i_op,
i_var,
op,
arg
);
if( NumRes(op) > 0 && op != BeginOp ) printOpResult(
std::cout,
0,
(CppAD::vectorBool *) CPPAD_NULL,
1,
&z_value
);
std::cout << std::endl;
}
std::cout << std::endl;
# else
}
void optimize(
size_t n ,
CppAD::vector<size_t>& dep_taddr ,
player<Base>* play ,
recorder<Base>* rec )
{
// temporary indices
size_t i, j, k;
// temporary variables
OpCode op; // current operator
const size_t *arg; // operator arguments
size_t i_var; // index of first result for current operator
// range and domain dimensions for F
size_t m = dep_taddr.size();
// number of variables in the player
const size_t num_var = play->num_rec_var();
// number of VecAD indices
size_t num_vecad_ind = play->num_rec_vecad_ind();
// number of VecAD vectors
size_t num_vecad_vec = play->num_rec_vecad_vec();
// -------------------------------------------------------------
// data structure that maps variable index in original operation
// sequence to corresponding operator information
CppAD::vector<struct optimize_old_variable> tape(num_var);
// -------------------------------------------------------------
// Determine how each variable is connected to the dependent variables
// initialize all variables has having no connections
for(i = 0; i < num_var; i++)
tape[i].connect = not_connected;
for(j = 0; j < m; j++)
{ // mark dependent variables as having one or more connections
tape[ dep_taddr[j] ].connect = yes_connected;
}
// vecad_connect contains a value for each VecAD object.
// vecad maps a VecAD index (which corresponds to the beginning of the
// VecAD object) to the vecad_connect falg for the VecAD object.
CppAD::vector<optimize_connection> vecad_connect(num_vecad_vec);
CppAD::vector<size_t> vecad(num_vecad_ind);
j = 0;
for(i = 0; i < num_vecad_vec; i++)
{ vecad_connect[i] = not_connected;
// length of this VecAD
size_t length = play->GetVecInd(j);
// set to proper index for this VecAD
vecad[j] = i;
for(k = 1; k <= length; k++)
vecad[j+k] = num_vecad_vec; // invalid index
// start of next VecAD
j += length + 1;
}
CPPAD_ASSERT_UNKNOWN( j == num_vecad_ind );
// Initialize a reverse mode sweep through the operation sequence
size_t i_op;
play->start_reverse(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( op == EndOp );
size_t mask;
while(op != BeginOp)
{ // next op
play->next_reverse(op, arg, i_op, i_var);
// This if is not necessary becasue last assignment
// with this value of i_var will have NumRes(op) > 0
if( NumRes(op) > 0 )
{ tape[i_var].op = op;
tape[i_var].arg = arg;
}
# ifndef NDEBUG
if( i_op <= n )
{ CPPAD_ASSERT_UNKNOWN((op == InvOp) | (op == BeginOp));
}
else CPPAD_ASSERT_UNKNOWN((op != InvOp) & (op != BeginOp));
# endif
switch( op )
{
// Unary operator where operand is arg[0]
case AbsOp:
case AcosOp:
case AsinOp:
case AtanOp:
case CosOp:
case CoshOp:
case DisOp:
case DivvpOp:
case ExpOp:
case LogOp:
case PowvpOp:
case SinOp:
case SinhOp:
case SqrtOp:
if( tape[i_var].connect != not_connected )
tape[arg[0]].connect = yes_connected;
break; // --------------------------------------------
// Unary operator where operand is arg[1]
case DivpvOp:
case MulpvOp:
case PowpvOp:
case PrivOp:
if( tape[i_var].connect != not_connected )
tape[arg[1]].connect = yes_connected;
break; // --------------------------------------------
// Special case for SubvpOp
case SubvpOp:
if( tape[i_var].connect != not_connected )
{
if( tape[arg[0]].connect == not_connected )
tape[arg[0]].connect = sum_connected;
else
tape[arg[0]].connect = yes_connected;
if( tape[i_var].connect == sum_connected )
tape[i_var].connect = csum_connected;
}
break; // --------------------------------------------
// Special case for AddpvOp and SubpvOp
case AddpvOp:
case SubpvOp:
if( tape[i_var].connect != not_connected )
{
if( tape[arg[1]].connect == not_connected )
tape[arg[1]].connect = sum_connected;
else
tape[arg[1]].connect = yes_connected;
if( tape[i_var].connect == sum_connected )
tape[i_var].connect = csum_connected;
}
break; // --------------------------------------------
// Special case for AddvvOp and SubvvOp
case AddvvOp:
case SubvvOp:
if( tape[i_var].connect != not_connected )
{
if( tape[arg[0]].connect == not_connected )
tape[arg[0]].connect = sum_connected;
else
tape[arg[0]].connect = yes_connected;
if( tape[arg[1]].connect == not_connected )
tape[arg[1]].connect = sum_connected;
else
tape[arg[1]].connect = yes_connected;
if( tape[i_var].connect == sum_connected )
tape[i_var].connect = csum_connected;
}
break; // --------------------------------------------
// Other binary operators
// where operands are arg[0], arg[1]
case DivvvOp:
case MulvvOp:
case PowvvOp:
if( tape[i_var].connect != not_connected )
{
tape[arg[0]].connect = yes_connected;
tape[arg[1]].connect = yes_connected;
}
break; // --------------------------------------------
// Conditional expression operators
case CExpOp:
CPPAD_ASSERT_UNKNOWN( NumArg(CExpOp) == 6 );
if( tape[i_var].connect != not_connected )
{
mask = 1;
for(i = 2; i < 6; i++)
{ if( arg[1] & mask )
{ CPPAD_ASSERT_UNKNOWN( arg[i] < i_var );
tape[arg[i]].connect = yes_connected;
}
mask = mask << 1;
}
}
break; // --------------------------------------------
// Operations where there is noting to do
case BeginOp:
case ComOp:
case EndOp:
case InvOp:
case ParOp:
case PripOp:
case StppOp:
break; // --------------------------------------------
// Load using a parameter index
case LdpOp:
if( tape[i_var].connect != not_connected )
{
i = vecad[ arg[0] - 1 ];
vecad_connect[i] = yes_connected;
}
break; // --------------------------------------------
// Load using a variable index
case LdvOp:
if( tape[i_var].connect != not_connected )
{
i = vecad[ arg[0] - 1 ];
vecad_connect[i] = yes_connected;
tape[arg[1]].connect = yes_connected;
}
break; // --------------------------------------------
// Store a variable using a parameter index
case StpvOp:
i = vecad[ arg[0] - 1 ];
if( vecad_connect[i] != not_connected )
tape[arg[2]].connect = yes_connected;
break; // --------------------------------------------
// Store a variable using a variable index
case StvvOp:
i = vecad[ arg[0] - 1 ];
if( vecad_connect[i] )
{ tape[arg[1]].connect = yes_connected;
tape[arg[2]].connect = yes_connected;
}
break; // --------------------------------------------
// all cases should be handled above
default:
CPPAD_ASSERT_UNKNOWN(0);
}
}
// values corresponding to BeginOp
CPPAD_ASSERT_UNKNOWN( i_op == 0 && i_var == 0 && op == BeginOp );
tape[i_var].op = op;
// -------------------------------------------------------------
// Erase all information in the recording
rec->Erase();
// Initilaize table mapping hash code to variable index in tape
// as pointing to the BeginOp at the beginning of the tape
CppAD::vector<size_t> hash_table_var(CPPAD_HASH_TABLE_SIZE);
for(i = 0; i < CPPAD_HASH_TABLE_SIZE; i++)
hash_table_var[i] = 0;
CPPAD_ASSERT_UNKNOWN( tape[0].op == BeginOp );
// initialize mapping from old variable index to new variable index
for(i = 0; i < num_var; i++)
tape[i].new_var = num_var; // invalid index
// initialize mapping from old VecAD index to new VecAD index
CppAD::vector<size_t> new_vecad_ind(num_vecad_ind);
for(i = 0; i < num_vecad_ind; i++)
new_vecad_ind[i] = num_vecad_ind; // invalid index
j = 0; // index into the old set of indices
for(i = 0; i < num_vecad_vec; i++)
{ // length of this VecAD
size_t length = play->GetVecInd(j);
if( vecad_connect[i] != not_connected )
{ // Put this VecAD vector in new recording
CPPAD_ASSERT_UNKNOWN(length < num_vecad_ind);
new_vecad_ind[j] = rec->PutVecInd(length);
for(k = 1; k <= length; k++) new_vecad_ind[j+k] =
rec->PutVecInd(
rec->PutPar(
play->GetPar(
play->GetVecInd(j+k)
) ) );
}
// start of next VecAD
j += length + 1;
}
CPPAD_ASSERT_UNKNOWN( j == num_vecad_ind );
// start playing the operations in the forward direction
play->start_forward(op, arg, i_op, i_var);
// playing forward skips BeginOp at the beginning, but not EndOp at
// the end. Put BeginOp at beginning of recording
CPPAD_ASSERT_UNKNOWN( op == BeginOp );
CPPAD_ASSERT_NARG_NRES(BeginOp, 0, 1);
tape[i_var].new_var = rec->PutOp(BeginOp);
// temporary buffer for new argument values
size_t new_arg[6];
// temporary work space used by optimize_record_csum
// (decalared here to avoid realloaction of memory)
optimize_csum_stacks csum_work;
while(op != EndOp)
{ // next op
play->next_forward(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( (i_op > n) | (op == InvOp) );
CPPAD_ASSERT_UNKNOWN( (i_op <= n) | (op != InvOp) );
// determine if we should keep this operation in the new
// operation sequence
bool keep;
switch( op )
{ case ComOp:
case PripOp:
case PrivOp:
keep = false;
break;
case InvOp:
case EndOp:
keep = true;
break;
case StppOp:
case StvpOp:
case StpvOp:
case StvvOp:
CPPAD_ASSERT_UNKNOWN( NumRes(op) == 0 );
i = vecad[ arg[0] - 1 ];
keep = vecad_connect[i] != not_connected;
break;
case AddpvOp:
case AddvvOp:
case SubpvOp:
case SubvpOp:
case SubvvOp:
keep = tape[i_var].connect != not_connected;
keep &= tape[i_var].connect != csum_connected;
break;
default:
keep = tape[i_var].connect != not_connected;
break;
}
size_t match_var = 0;
unsigned short code = 0;
bool replace_hash = false;
if( keep ) switch( op )
{
// Unary operator where operand is arg[0]
case AbsOp:
case AcosOp:
case AsinOp:
case AtanOp:
case CosOp:
case CoshOp:
case DisOp:
case ExpOp:
case LogOp:
case SinOp:
case SinhOp:
case SqrtOp:
match_var = optimize_unary_match(
tape , // inputs
i_var ,
play->num_rec_par() ,
play->GetPar() ,
hash_table_var ,
code // outputs
);
if( match_var > 0 )
tape[i_var].new_var = match_var;
else
{
replace_hash = true;
new_arg[0] = tape[ arg[0] ].new_var;
rec->PutArg( new_arg[0] );
i = rec->PutOp(op);
tape[i_var].new_var = i;
CPPAD_ASSERT_UNKNOWN( new_arg[0] < i );
}
break;
// ---------------------------------------------------
// Binary operators where
// left is a variable and right is a parameter
case SubvpOp:
if( tape[arg[0]].connect == csum_connected )
{
// convert to a sequence of summation operators
tape[i_var].new_var = optimize_record_csum(
tape , // inputs
i_var ,
play->num_rec_par() ,
play->GetPar() ,
rec ,
csum_work
);
// abort rest of this case
break;
}
case DivvpOp:
case PowvpOp:
match_var = optimize_binary_match(
tape , // inputs
i_var ,
play->num_rec_par() ,
play->GetPar() ,
hash_table_var ,
code // outputs
);
if( match_var > 0 )
tape[i_var].new_var = match_var;
else
{ tape[i_var].new_var = optimize_record_vp(
tape , // inputs
i_var ,
play->num_rec_par() ,
play->GetPar() ,
rec ,
op ,
arg
);
replace_hash = true;
}
break;
// ---------------------------------------------------
// Binary operators where
// left is a parameter and right is a variable
case SubpvOp:
case AddpvOp:
if( tape[arg[1]].connect == csum_connected )
{
// convert to a sequence of summation operators
tape[i_var].new_var = optimize_record_csum(
tape , // inputs
i_var ,
play->num_rec_par() ,
play->GetPar() ,
rec ,
csum_work
);
// abort rest of this case
break;
}
case DivpvOp:
case MulpvOp:
case PowpvOp:
match_var = optimize_binary_match(
tape , // inputs
i_var ,
play->num_rec_par() ,
play->GetPar() ,
hash_table_var ,
code // outputs
);
if( match_var > 0 )
tape[i_var].new_var = match_var;
else
{ tape[i_var].new_var = optimize_record_pv(
tape , // inputs
i_var ,
play->num_rec_par() ,
play->GetPar() ,
rec ,
op ,
arg
);
replace_hash = true;
}
break;
// ---------------------------------------------------
// Binary operator where
// both operators are variables
case AddvvOp:
case SubvvOp:
if( (tape[arg[0]].connect == csum_connected) |
(tape[arg[1]].connect == csum_connected)
)
{
// convert to a sequence of summation operators
tape[i_var].new_var = optimize_record_csum(
tape , // inputs
i_var ,
play->num_rec_par() ,
play->GetPar() ,
rec ,
csum_work
);
// abort rest of this case
break;
}
case DivvvOp:
case MulvvOp:
case PowvvOp:
match_var = optimize_binary_match(
tape , // inputs
i_var ,
play->num_rec_par() ,
play->GetPar() ,
hash_table_var ,
code // outputs
);
if( match_var > 0 )
tape[i_var].new_var = match_var;
else
{ tape[i_var].new_var = optimize_record_vv(
tape , // inputs
i_var ,
play->num_rec_par() ,
play->GetPar() ,
rec ,
op ,
arg
);
replace_hash = true;
}
break;
// ---------------------------------------------------
// Conditional expression operators
case CExpOp:
CPPAD_ASSERT_NARG_NRES(op, 6, 1);
new_arg[0] = arg[0];
new_arg[1] = arg[1];
mask = 1;
for(i = 2; i < 6; i++)
{ if( arg[1] & mask )
{ new_arg[i] = tape[arg[i]].new_var;
CPPAD_ASSERT_UNKNOWN(
new_arg[i] < num_var
);
}
else new_arg[i] = rec->PutPar(
play->GetPar( arg[i] )
);
mask = mask << 1;
}
rec->PutArg(
new_arg[0] ,
new_arg[1] ,
new_arg[2] ,
new_arg[3] ,
new_arg[4] ,
new_arg[5]
);
tape[i_var].new_var = rec->PutOp(op);
break;
// ---------------------------------------------------
// Operations with no arguments and no results
case EndOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 0);
rec->PutOp(op);
break;
// ---------------------------------------------------
// Operations with no arguments and one result
case InvOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 1);
tape[i_var].new_var = rec->PutOp(op);
break;
// ---------------------------------------------------
// Operations with one argument that is a parameter
case ParOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1);
new_arg[0] = rec->PutPar( play->GetPar(arg[0] ) );
rec->PutArg( new_arg[0] );
tape[i_var].new_var = rec->PutOp(op);
break;
// ---------------------------------------------------
// Load using a parameter index
case LdpOp:
CPPAD_ASSERT_NARG_NRES(op, 3, 1);
new_arg[0] = new_vecad_ind[ arg[0] ];
new_arg[1] = arg[1];
CPPAD_ASSERT_UNKNOWN( new_arg[0] < num_vecad_ind );
rec->PutArg(
new_arg[0],
new_arg[1],
0
);
tape[i_var].new_var = rec->PutOp(op);
break;
// ---------------------------------------------------
// Load using a variable index
case LdvOp:
CPPAD_ASSERT_NARG_NRES(op, 3, 1);
new_arg[0] = new_vecad_ind[ arg[0] ];
new_arg[1] = tape[arg[1]].new_var;
CPPAD_ASSERT_UNKNOWN( new_arg[0] < num_vecad_ind );
CPPAD_ASSERT_UNKNOWN( new_arg[1] < num_var );
rec->PutArg(
new_arg[0],
new_arg[1],
0
);
tape[i_var].new_var = rec->PutOp(op);
break;
// ---------------------------------------------------
// Store a parameter using a parameter index
case StppOp:
CPPAD_ASSERT_NARG_NRES(op, 3, 0);
new_arg[0] = new_vecad_ind[ arg[0] ];
new_arg[1] = rec->PutPar( play->GetPar(arg[1]) );
new_arg[2] = rec->PutPar( play->GetPar(arg[2]) );
CPPAD_ASSERT_UNKNOWN( new_arg[0] < num_vecad_ind );
rec->PutArg(
new_arg[0],
new_arg[1],
new_arg[2]
);
rec->PutOp(op);
break;
// ---------------------------------------------------
// Store a parameter using a variable index
case StvpOp:
CPPAD_ASSERT_NARG_NRES(op, 3, 0);
new_arg[0] = new_vecad_ind[ arg[0] ];
new_arg[1] = tape[arg[1]].new_var;
new_arg[2] = rec->PutPar( play->GetPar(arg[2]) );
CPPAD_ASSERT_UNKNOWN( new_arg[0] < num_vecad_ind );
CPPAD_ASSERT_UNKNOWN( new_arg[1] < num_var );
rec->PutArg(
new_arg[0],
new_arg[1],
new_arg[2]
);
rec->PutOp(op);
break;
// ---------------------------------------------------
// Store a variable using a parameter index
case StpvOp:
CPPAD_ASSERT_NARG_NRES(op, 3, 0);
new_arg[0] = new_vecad_ind[ arg[0] ];
new_arg[1] = rec->PutPar( play->GetPar(arg[1]) );
new_arg[2] = tape[arg[2]].new_var;
CPPAD_ASSERT_UNKNOWN( new_arg[0] < num_vecad_ind );
CPPAD_ASSERT_UNKNOWN( new_arg[1] < num_var );
CPPAD_ASSERT_UNKNOWN( new_arg[2] < num_var );
rec->PutArg(
new_arg[0],
new_arg[1],
new_arg[2]
);
rec->PutOp(op);
break;
// ---------------------------------------------------
// Store a variable using a variable index
case StvvOp:
CPPAD_ASSERT_NARG_NRES(op, 3, 0);
new_arg[0] = new_vecad_ind[ arg[0] ];
new_arg[1] = tape[arg[1]].new_var;
new_arg[2] = tape[arg[2]].new_var;
CPPAD_ASSERT_UNKNOWN( new_arg[0] < num_vecad_ind );
CPPAD_ASSERT_UNKNOWN( new_arg[1] < num_var );
CPPAD_ASSERT_UNKNOWN( new_arg[2] < num_var );
rec->PutArg(
new_arg[0],
new_arg[1],
new_arg[2]
);
rec->PutOp(op);
break;
// ---------------------------------------------------
// all cases should be handled above
default:
CPPAD_ASSERT_UNKNOWN(false);
}
if( replace_hash )
{ // The old variable index i_var corresponds to the
// new variable index tape[i_var].new_var. In addition
// this is the most recent variable that has this code.
hash_table_var[code] = i_var;
}
}
// modify the dependent variable vector to new indices
for(i = 0; i < dep_taddr.size(); i++ )
{ CPPAD_ASSERT_UNKNOWN( tape[ dep_taddr[i] ].new_var < num_var );
dep_taddr[i] = tape[ dep_taddr[i] ].new_var;
}
}
inline void reverse_asin_op(
size_t d ,
size_t i_z ,
size_t i_x ,
size_t nc_taylor ,
const Base* taylor ,
size_t nc_partial ,
Base* partial )
{
// check assumptions
CPPAD_ASSERT_UNKNOWN( NumArg(AsinOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumRes(AsinOp) == 2 );
CPPAD_ASSERT_UNKNOWN( i_x + 1 < i_z );
CPPAD_ASSERT_UNKNOWN( d < nc_taylor );
CPPAD_ASSERT_UNKNOWN( d < nc_partial );
// Taylor coefficients and partials corresponding to argument
const Base* x = taylor + i_x * nc_taylor;
Base* px = partial + i_x * nc_partial;
// Taylor coefficients and partials corresponding to first result
const Base* z = taylor + i_z * nc_taylor;
Base* pz = partial + i_z * nc_partial;
// Taylor coefficients and partials corresponding to auxillary result
const Base* b = z - nc_taylor; // called y in documentation
Base* pb = pz - nc_partial;
// number of indices to access
size_t j = d;
size_t k;
while(j)
{
// scale partials w.r.t b[j] by 1 / b[0]
pb[j] /= b[0];
// scale partials w.r.t z[j] by 1 / b[0]
pz[j] /= b[0];
// update partials w.r.t b^0
pb[0] -= pz[j] * z[j] + pb[j] * b[j];
// update partial w.r.t. x^0
px[0] -= pb[j] * x[j];
// update partial w.r.t. x^j
px[j] += pz[j] - pb[j] * x[0];
// further scale partial w.r.t. z[j] by 1 / j
pz[j] /= Base(j);
for(k = 1; k < j; k++)
{ // update partials w.r.t b^(j-k)
pb[j-k] -= Base(k) * pz[j] * z[k] + pb[j] * b[k];
// update partials w.r.t. x^k
px[k] -= pb[j] * x[j-k];
// update partials w.r.t. z^k
pz[k] -= pz[j] * Base(k) * b[j-k];
}
--j;
}
// j == 0 case
px[0] += ( pz[0] - pb[0] * x[0]) / b[0];
}
void ADTape<Base>::RecordCondExp(
enum CompareOp cop ,
AD<Base> &returnValue ,
const AD<Base> &left ,
const AD<Base> &right ,
const AD<Base> &if_true ,
const AD<Base> &if_false )
{ size_t ind0, ind1, ind2, ind3, ind4, ind5;
size_t returnValue_taddr;
// taddr_ of this variable
CPPAD_ASSERT_UNKNOWN( NumRes(CExpOp) == 1 );
returnValue_taddr = Rec_.PutOp(CExpOp);
// ind[0] = cop
ind0 = addr_t( cop );
// ind[1] = base 2 representaion of the value
// [Var(left), Var(right), Var(if_true), Var(if_false)]
ind1 = 0;
// Make sure returnValue is in the list of variables and set its taddr
if( Parameter(returnValue) )
returnValue.make_variable(id_, returnValue_taddr );
else returnValue.taddr_ = returnValue_taddr;
// ind[2] = left address
if( Parameter(left) )
ind2 = Rec_.PutPar(left.value_);
else
{ ind1 += 1;
ind2 = left.taddr_;
}
// ind[3] = right address
if( Parameter(right) )
ind3 = Rec_.PutPar(right.value_);
else
{ ind1 += 2;
ind3 = right.taddr_;
}
// ind[4] = if_true address
if( Parameter(if_true) )
ind4 = Rec_.PutPar(if_true.value_);
else
{ ind1 += 4;
ind4 = if_true.taddr_;
}
// ind[5] = if_false address
if( Parameter(if_false) )
ind5 = Rec_.PutPar(if_false.value_);
else
{ ind1 += 8;
ind5 = if_false.taddr_;
}
CPPAD_ASSERT_UNKNOWN( NumArg(CExpOp) == 6 );
CPPAD_ASSERT_UNKNOWN( ind1 > 0 );
Rec_.PutArg(ind0, ind1, ind2, ind3, ind4, ind5);
// check that returnValue is a dependent variable
CPPAD_ASSERT_UNKNOWN( Variable(returnValue) );
}
void forward0sweep(
std::ostream& s_out,
bool print,
size_t n,
size_t numvar,
local::player<Base>* play,
size_t J,
Base* taylor,
bool* cskip_op,
pod_vector<addr_t>& var_by_load_op,
size_t compare_change_count,
size_t& compare_change_number,
size_t& compare_change_op_index
)
{ CPPAD_ASSERT_UNKNOWN( J >= 1 );
CPPAD_ASSERT_UNKNOWN( play->num_var_rec() == numvar );
// use p, q, r so other forward sweeps can use code defined here
size_t p = 0;
size_t q = 0;
size_t r = 1;
/*
<!-- define forward0sweep_code_define -->
*/
// op code for current instruction
OpCode op;
// index for current instruction
size_t i_op;
// next variables
size_t i_var;
// operation argument indices
const addr_t* arg = CPPAD_NULL;
// initialize the comparision operator counter
if( p == 0 )
{ compare_change_number = 0;
compare_change_op_index = 0;
}
// If this includes a zero calculation, initialize this information
pod_vector<bool> isvar_by_ind;
pod_vector<size_t> index_by_ind;
if( p == 0 )
{ size_t i;
// this includes order zero calculation, initialize vector indices
size_t num = play->num_vec_ind_rec();
if( num > 0 )
{ isvar_by_ind.extend(num);
index_by_ind.extend(num);
for(i = 0; i < num; i++)
{ index_by_ind[i] = play->GetVecInd(i);
isvar_by_ind[i] = false;
}
}
// includes zero order, so initialize conditional skip flags
num = play->num_op_rec();
for(i = 0; i < num; i++)
cskip_op[i] = false;
}
// work space used by UserOp.
vector<bool> user_vx; // empty vecotor
vector<bool> user_vy; // empty vecotor
vector<Base> user_tx; // argument vector Taylor coefficients
vector<Base> user_ty; // result vector Taylor coefficients
//
atomic_base<Base>* user_atom = CPPAD_NULL; // user's atomic op calculator
# ifndef NDEBUG
bool user_ok = false; // atomic op return value
# endif
//
// information defined by forward_user
size_t user_old=0, user_m=0, user_n=0, user_i=0, user_j=0;
enum_user_state user_state = start_user; // proper initialization
// length of the parameter vector (used by CppAD assert macros)
const size_t num_par = play->num_par_rec();
// pointer to the beginning of the parameter vector
const Base* parameter = CPPAD_NULL;
if( num_par > 0 )
parameter = play->GetPar();
// length of the text vector (used by CppAD assert macros)
const size_t num_text = play->num_text_rec();
// pointer to the beginning of the text vector
const char* text = CPPAD_NULL;
if( num_text > 0 )
text = play->GetTxt(0);
/*
<!-- end forward0sweep_code_define -->
*/
# if CPPAD_FORWARD0SWEEP_TRACE
// flag as to when to trace user function values
bool user_trace = false;
// variable indices for results vector
// (done differently for order zero).
vector<size_t> user_iy;
# endif
// skip the BeginOp at the beginning of the recording
play->forward_start(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( op == BeginOp );
# if CPPAD_FORWARD0SWEEP_TRACE
std::cout << std::endl;
# endif
bool flag; // a temporary flag to use in switch cases
bool more_operators = true;
while(more_operators)
{
// this op
play->forward_next(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( (i_op > n) | (op == InvOp) );
CPPAD_ASSERT_UNKNOWN( (i_op <= n) | (op != InvOp) );
CPPAD_ASSERT_UNKNOWN( i_op < play->num_op_rec() );
CPPAD_ASSERT_ARG_BEFORE_RESULT(op, arg, i_var);
// check if we are skipping this operation
while( cskip_op[i_op] )
{ switch(op)
{ case CSumOp:
// CSumOp has a variable number of arguments
play->forward_csum(op, arg, i_op, i_var);
break;
case CSkipOp:
// CSkip has a variable number of arguments
play->forward_cskip(op, arg, i_op, i_var);
break;
case UserOp:
{ // skip all operations in this user atomic call
CPPAD_ASSERT_UNKNOWN( user_state == start_user );
play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
size_t n_skip = user_m + user_n + 1;
for(size_t i = 0; i < n_skip; i++)
{ play->forward_next(op, arg, i_op, i_var);
play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
}
CPPAD_ASSERT_UNKNOWN( user_state == start_user );
}
break;
default:
break;
}
play->forward_next(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( i_op < play->num_op_rec() );
}
// action to take depends on the case
switch( op )
{
case AbsOp:
forward_abs_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case AddvvOp:
forward_addvv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case AddpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_addpv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case AcosOp:
// sqrt(1 - x * x), acos(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_acos_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case AcoshOp:
// sqrt(x * x - 1), acosh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_acosh_op_0(i_var, arg[0], J, taylor);
break;
# endif
// -------------------------------------------------
case AsinOp:
// sqrt(1 - x * x), asin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_asin_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case AsinhOp:
// sqrt(1 + x * x), asinh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_asinh_op_0(i_var, arg[0], J, taylor);
break;
# endif
// -------------------------------------------------
case AtanOp:
// 1 + x * x, atan(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_atan_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case AtanhOp:
// 1 - x * x, atanh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_atanh_op_0(i_var, arg[0], J, taylor);
break;
# endif
// -------------------------------------------------
case CExpOp:
// Use the general case with d == 0
// (could create an optimzied verison for this case)
forward_cond_op_0(
i_var, arg, num_par, parameter, J, taylor
);
break;
// ---------------------------------------------------
case CosOp:
// sin(x), cos(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_cos_op_0(i_var, arg[0], J, taylor);
break;
// ---------------------------------------------------
case CoshOp:
// sinh(x), cosh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_cosh_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case CSkipOp:
// CSkipOp has a variable number of arguments and
// forward_next thinks it has no arguments.
// we must inform forward_next of this special case.
forward_cskip_op_0(
i_var, arg, num_par, parameter, J, taylor, cskip_op
);
play->forward_cskip(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case CSumOp:
// CSumOp has a variable number of arguments and
// forward_next thinks it has no arguments.
// we must inform forward_next of this special case.
forward_csum_op(
0, 0, i_var, arg, num_par, parameter, J, taylor
);
play->forward_csum(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case DisOp:
forward_dis_op(p, q, r, i_var, arg, J, taylor);
break;
// -------------------------------------------------
case DivvvOp:
forward_divvv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case DivpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_divpv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case DivvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_divvp_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case EndOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 0);
more_operators = false;
break;
// -------------------------------------------------
case EqpvOp:
if( compare_change_count )
{ forward_eqpv_op_0(
compare_change_number, arg, parameter, J, taylor
);
{ if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
}
break;
// -------------------------------------------------
case EqvvOp:
if( compare_change_count )
{ forward_eqvv_op_0(
compare_change_number, arg, parameter, J, taylor
);
{ if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
}
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case ErfOp:
forward_erf_op_0(i_var, arg, parameter, J, taylor);
break;
# endif
// -------------------------------------------------
case ExpOp:
forward_exp_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case Expm1Op:
forward_expm1_op_0(i_var, arg[0], J, taylor);
break;
# endif
// -------------------------------------------------
case InvOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 1);
break;
// ---------------------------------------------------
case LdpOp:
forward_load_p_op_0(
play,
i_var,
arg,
parameter,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data(),
var_by_load_op.data()
);
break;
// -------------------------------------------------
case LdvOp:
forward_load_v_op_0(
play,
i_var,
arg,
parameter,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data(),
var_by_load_op.data()
);
break;
// -------------------------------------------------
case LepvOp:
if( compare_change_count )
{ forward_lepv_op_0(
compare_change_number, arg, parameter, J, taylor
);
{ if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
}
break;
// -------------------------------------------------
case LevpOp:
if( compare_change_count )
{ forward_levp_op_0(
compare_change_number, arg, parameter, J, taylor
);
{ if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
}
break;
// -------------------------------------------------
case LevvOp:
if( compare_change_count )
{ forward_levv_op_0(
compare_change_number, arg, parameter, J, taylor
);
{ if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
}
break;
// -------------------------------------------------
case LogOp:
forward_log_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case Log1pOp:
forward_log1p_op_0(i_var, arg[0], J, taylor);
break;
# endif
// -------------------------------------------------
case LtpvOp:
if( compare_change_count )
{ forward_ltpv_op_0(
compare_change_number, arg, parameter, J, taylor
);
{ if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
}
break;
// -------------------------------------------------
case LtvpOp:
if( compare_change_count )
{ forward_ltvp_op_0(
compare_change_number, arg, parameter, J, taylor
);
{ if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
}
break;
// -------------------------------------------------
case LtvvOp:
if( compare_change_count )
{ forward_ltvv_op_0(
compare_change_number, arg, parameter, J, taylor
);
{ if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
}
break;
// -------------------------------------------------
case MulpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_mulpv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case MulvvOp:
forward_mulvv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case NepvOp:
if( compare_change_count )
{ forward_nepv_op_0(
compare_change_number, arg, parameter, J, taylor
);
{ if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
}
break;
// -------------------------------------------------
case NevvOp:
if( compare_change_count )
{ forward_nevv_op_0(
compare_change_number, arg, parameter, J, taylor
);
{ if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
}
break;
// -------------------------------------------------
case ParOp:
forward_par_op_0(
i_var, arg, num_par, parameter, J, taylor
);
break;
// -------------------------------------------------
case PowvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_powvp_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case PowpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_powpv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case PowvvOp:
forward_powvv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case PriOp:
if( print ) forward_pri_0(s_out,
arg, num_text, text, num_par, parameter, J, taylor
);
break;
// -------------------------------------------------
case SignOp:
// cos(x), sin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sign_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case SinOp:
// cos(x), sin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sin_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case SinhOp:
// cosh(x), sinh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sinh_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case SqrtOp:
forward_sqrt_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case StppOp:
forward_store_pp_op_0(
i_var,
arg,
num_par,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data()
);
break;
// -------------------------------------------------
case StpvOp:
forward_store_pv_op_0(
i_var,
arg,
num_par,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data()
);
break;
// -------------------------------------------------
case StvpOp:
forward_store_vp_op_0(
i_var,
arg,
num_par,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data()
);
break;
// -------------------------------------------------
case StvvOp:
forward_store_vv_op_0(
i_var,
arg,
num_par,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data()
);
break;
// -------------------------------------------------
case SubvvOp:
forward_subvv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case SubpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_subpv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case SubvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_subvp_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case TanOp:
// tan(x)^2, tan(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_tan_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case TanhOp:
// tanh(x)^2, tanh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_tanh_op_0(i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case UserOp:
// start or end an atomic operation sequence
flag = user_state == start_user;
user_atom = play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
if( flag )
{ user_tx.resize(user_n);
user_ty.resize(user_m);
# if CPPAD_FORWARD0SWEEP_TRACE
user_iy.resize(user_m);
# endif
}
else
{
# ifndef NDEBUG
if( ! user_ok )
{ std::string msg =
user_atom->afun_name()
+ ": atomic_base.forward: returned false";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
# if CPPAD_FORWARD0SWEEP_TRACE
user_trace = true;
# endif
}
break;
case UsrapOp:
// parameter argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( size_t( arg[0] ) < num_par );
user_tx[user_j] = parameter[ arg[0] ];
play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
if( user_j == user_n )
{ // call users function for this operation
user_atom->set_old(user_old);
CPPAD_ATOMIC_CALL(p, q,
user_vx, user_vy, user_tx, user_ty
);
}
break;
case UsravOp:
// variable argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) <= i_var );
user_tx[user_j] = taylor[ arg[0] * J + 0 ];
play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
if( user_j == user_n )
{ // call users function for this operation
user_atom->set_old(user_old);
CPPAD_ATOMIC_CALL(p, q,
user_vx, user_vy, user_tx, user_ty
);
}
break;
case UsrrpOp:
// parameter result in an atomic operation sequence
# if CPPAD_FORWARD0SWEEP_TRACE
user_iy[user_i] = 0;
# endif
play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
break;
case UsrrvOp:
// variable result in an atomic operation sequence
# if CPPAD_FORWARD0SWEEP_TRACE
user_iy[user_i] = i_var;
# endif
taylor[ i_var * J + 0 ] = user_ty[user_i];
play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
break;
// -------------------------------------------------
case ZmulpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_zmulpv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case ZmulvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_zmulvp_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case ZmulvvOp:
forward_zmulvv_op_0(i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
default:
CPPAD_ASSERT_UNKNOWN(false);
}
# if CPPAD_FORWARD0SWEEP_TRACE
size_t d = 0;
if( user_trace )
{ user_trace = false;
CPPAD_ASSERT_UNKNOWN( op == UserOp );
CPPAD_ASSERT_UNKNOWN( NumArg(UsrrvOp) == 0 );
for(size_t i = 0; i < user_m; i++) if( user_iy[i] > 0 )
{ size_t i_tmp = (i_op + i) - user_m;
printOp(
std::cout,
play,
i_tmp,
user_iy[i],
UsrrvOp,
CPPAD_NULL
);
Base* Z_tmp = taylor + user_iy[i] * J;
printOpResult(
std::cout,
d + 1,
Z_tmp,
0,
(Base *) CPPAD_NULL
);
std::cout << std::endl;
}
}
Base* Z_tmp = taylor + i_var * J;
const addr_t* arg_tmp = arg;
if( op == CSumOp )
arg_tmp = arg - arg[-1] - 4;
if( op == CSkipOp )
arg_tmp = arg - arg[-1] - 7;
if( op != UsrrvOp )
{
printOp(
std::cout,
play,
i_op,
i_var,
op,
arg_tmp
);
if( NumRes(op) > 0 ) printOpResult(
std::cout,
d + 1,
Z_tmp,
0,
(Base *) CPPAD_NULL
);
std::cout << std::endl;
}
}
std::cout << std::endl;
# else
}
inline void reverse_cond_op(
size_t d ,
size_t i_z ,
const addr_t* arg ,
size_t num_par ,
const Base* parameter ,
size_t nc_taylor ,
const Base* taylor ,
size_t nc_partial ,
Base* partial )
{ Base y_0, y_1;
Base zero(0);
Base* pz;
Base* py_2;
Base* py_3;
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < static_cast<size_t> (CompareNe) );
CPPAD_ASSERT_UNKNOWN( NumArg(CExpOp) == 6 );
CPPAD_ASSERT_UNKNOWN( NumRes(CExpOp) == 1 );
CPPAD_ASSERT_UNKNOWN( arg[1] != 0 );
pz = partial + i_z * nc_partial + 0;
if( arg[1] & 1 )
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[2]) < i_z );
y_0 = taylor[ arg[2] * nc_taylor + 0 ];
}
else
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[2]) < num_par );
y_0 = parameter[ arg[2] ];
}
if( arg[1] & 2 )
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[3]) < i_z );
y_1 = taylor[ arg[3] * nc_taylor + 0 ];
}
else
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[3]) < num_par );
y_1 = parameter[ arg[3] ];
}
if( arg[1] & 4 )
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[4]) < i_z );
py_2 = partial + arg[4] * nc_partial;
size_t j = d + 1;
while(j--)
{ py_2[j] += CondExpOp(
CompareOp( arg[0] ),
y_0,
y_1,
pz[j],
zero
);
}
}
if( arg[1] & 8 )
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[5]) < i_z );
py_3 = partial + arg[5] * nc_partial;
size_t j = d + 1;
while(j--)
{ py_3[j] += CondExpOp(
CompareOp( arg[0] ),
y_0,
y_1,
zero,
pz[j]
);
}
}
return;
}
void for_hes(
const local::player<Base>* play,
size_t n,
size_t numvar,
const Vector_set& for_jac_sparse,
const Vector_set& rev_jac_sparse,
Vector_set& for_hes_sparse,
const RecBase& not_used_rec_base
)
{
// length of the parameter vector (used by CppAD assert macros)
const size_t num_par = play->num_par_rec();
size_t i, j, k;
// check numvar argument
size_t limit = n+1;
CPPAD_ASSERT_UNKNOWN( play->num_var_rec() == numvar );
CPPAD_ASSERT_UNKNOWN( for_jac_sparse.n_set() == numvar );
CPPAD_ASSERT_UNKNOWN( for_hes_sparse.n_set() == limit );
CPPAD_ASSERT_UNKNOWN( numvar > 0 );
// upper limit exclusive for set elements
CPPAD_ASSERT_UNKNOWN( for_jac_sparse.end() == limit );
CPPAD_ASSERT_UNKNOWN( for_hes_sparse.end() == limit );
// vecad_sparsity contains a sparsity pattern for each VecAD object.
// vecad_ind maps a VecAD index (beginning of the VecAD object)
// to the index for the corresponding set in vecad_sparsity.
size_t num_vecad_ind = play->num_vec_ind_rec();
size_t num_vecad_vec = play->num_vecad_vec_rec();
Vector_set vecad_sparse;
pod_vector<size_t> vecad_ind;
pod_vector<bool> vecad_jac;
if( num_vecad_vec > 0 )
{ size_t length;
vecad_sparse.resize(num_vecad_vec, limit);
vecad_ind.extend(num_vecad_ind);
vecad_jac.extend(num_vecad_vec);
j = 0;
for(i = 0; i < num_vecad_vec; i++)
{ // length of this VecAD
length = play->GetVecInd(j);
// set vecad_ind to proper index for this VecAD
vecad_ind[j] = i;
// make all other values for this vector invalid
for(k = 1; k <= length; k++)
vecad_ind[j+k] = num_vecad_vec;
// start of next VecAD
j += length + 1;
// initialize this vector's reverse jacobian value
vecad_jac[i] = false;
}
CPPAD_ASSERT_UNKNOWN( j == play->num_vec_ind_rec() );
}
// ------------------------------------------------------------------------
// work space used by AFunOp.
vector<Base> atom_x; //// value of parameter arguments to function
vector<ad_type_enum> type_x; // argument types
pod_vector<size_t> atom_ix; // variable index (on tape) for each argument
pod_vector<size_t> atom_iy; // variable index (on tape) for each result
//
// information set by atomic forward (initialization to avoid warnings)
size_t atom_index=0, atom_old=0, atom_m=0, atom_n=0, atom_i=0, atom_j=0;
// information set by atomic forward (necessary initialization)
enum_atom_state atom_state = start_atom;
// -------------------------------------------------------------------------
//
// pointer to the beginning of the parameter vector
// (used by atomic functions)
const Base* parameter = CPPAD_NULL;
if( num_par > 0 )
parameter = play->GetPar();
//
// which parametes are dynamic
const pod_vector<bool>& dyn_par_is( play->dyn_par_is() );
//
// skip the BeginOp at the beginning of the recording
play::const_sequential_iterator itr = play->begin();
// op_info
OpCode op;
size_t i_var;
const Addr* arg;
itr.op_info(op, arg, i_var);
CPPAD_ASSERT_UNKNOWN( op == BeginOp );
# if CPPAD_FOR_HES_TRACE
vector<size_t> atom_funrp; // parameter index for FunrpOp operators
std::cout << std::endl;
CppAD::vectorBool zf_value(limit);
CppAD::vectorBool zh_value(limit * limit);
# endif
bool flag; // temporary for use in switch cases below
bool more_operators = true;
while(more_operators)
{
// next op
(++itr).op_info(op, arg, i_var);
// does the Hessian in question have a non-zero derivative
// with respect to this variable
bool include = NumRes(op) > 0;
if( include )
include = rev_jac_sparse.is_element(i_var, 0);
//
// operators to include even if derivative is zero
include |= op == EndOp;
include |= op == CSkipOp;
include |= op == CSumOp;
include |= op == AFunOp;
include |= op == FunapOp;
include |= op == FunavOp;
include |= op == FunrpOp;
include |= op == FunrvOp;
//
if( include ) switch( op )
{ // operators that should not occurr
// case BeginOp
// -------------------------------------------------
// operators that do not affect hessian
case AbsOp:
case AddvvOp:
case AddpvOp:
case CExpOp:
case DisOp:
case DivvpOp:
case InvOp:
case LdpOp:
case LdvOp:
case MulpvOp:
case ParOp:
case PriOp:
case SignOp:
case StppOp:
case StpvOp:
case StvpOp:
case StvvOp:
case SubvvOp:
case SubpvOp:
case SubvpOp:
case ZmulpvOp:
case ZmulvpOp:
break;
// -------------------------------------------------
// nonlinear unary operators
case AcosOp:
case AsinOp:
case AtanOp:
case CosOp:
case CoshOp:
case ExpOp:
case LogOp:
case SinOp:
case SinhOp:
case SqrtOp:
case TanOp:
case TanhOp:
# if CPPAD_USE_CPLUSPLUS_2011
case AcoshOp:
case AsinhOp:
case AtanhOp:
case Expm1Op:
case Log1pOp:
# endif
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 )
forward_sparse_hessian_nonlinear_unary_op(
size_t(arg[0]), for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case CSkipOp:
itr.correct_before_increment();
break;
// -------------------------------------------------
case CSumOp:
itr.correct_before_increment();
break;
// -------------------------------------------------
case DivvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_div_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case DivpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_nonlinear_unary_op(
size_t(arg[1]), for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case EndOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 0);
more_operators = false;
break;
// -------------------------------------------------
case ErfOp:
// arg[1] is always the parameter 0
// arg[2] is always the parameter 2 / sqrt(pi)
CPPAD_ASSERT_NARG_NRES(op, 3, 5);
forward_sparse_hessian_nonlinear_unary_op(
size_t(arg[0]), for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
// -------------------------------------------------
// logical comparision operators
case EqppOp:
case EqpvOp:
case EqvvOp:
case LtppOp:
case LtpvOp:
case LtvpOp:
case LtvvOp:
case LeppOp:
case LepvOp:
case LevpOp:
case LevvOp:
case NepvOp:
case NeppOp:
case NevvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 0);
break;
// -------------------------------------------------
case MulvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_mul_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case PowpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
forward_sparse_hessian_nonlinear_unary_op(
size_t(arg[1]), for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case PowvpOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
forward_sparse_hessian_nonlinear_unary_op(
size_t(arg[0]), for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case PowvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
forward_sparse_hessian_pow_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case AFunOp:
// start or end an atomic function call
CPPAD_ASSERT_UNKNOWN(
atom_state == start_atom || atom_state == end_atom
);
flag = atom_state == start_atom;
play::atom_op_info<RecBase>(
op, arg, atom_index, atom_old, atom_m, atom_n
);
if( flag )
{ atom_state = arg_atom;
atom_i = 0;
atom_j = 0;
//
atom_x.resize( atom_n );
type_x.resize( atom_n );
atom_ix.resize( atom_n );
atom_iy.resize( atom_m );
# if CPPAD_FOR_HES_TRACE
atom_funrp.resize( atom_m );
# endif
}
else
{ CPPAD_ASSERT_UNKNOWN( atom_i == atom_m );
CPPAD_ASSERT_UNKNOWN( atom_j == atom_n );
atom_state = start_atom;
//
call_atomic_for_hes_sparsity<Base,RecBase>(
atom_index, atom_old, atom_x, type_x, atom_ix, atom_iy,
for_jac_sparse, rev_jac_sparse, for_hes_sparse
);
}
break;
case FunapOp:
// parameter argument for a atomic function
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
CPPAD_ASSERT_UNKNOWN( atom_state == arg_atom );
CPPAD_ASSERT_UNKNOWN( atom_i == 0 );
CPPAD_ASSERT_UNKNOWN( atom_j < atom_n );
CPPAD_ASSERT_UNKNOWN( size_t( arg[0] ) < num_par );
//
atom_x[atom_j] = parameter[arg[0]];
// argument type
if( dyn_par_is[arg[0]] )
type_x[atom_j] = dynamic_enum;
else
type_x[atom_j] = constant_enum;
atom_ix[atom_j] = 0; // special variable used for parameters
//
++atom_j;
if( atom_j == atom_n )
atom_state = ret_atom;
break;
case FunavOp:
// variable argument for a atomic function
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
CPPAD_ASSERT_UNKNOWN( atom_state == arg_atom );
CPPAD_ASSERT_UNKNOWN( atom_i == 0 );
CPPAD_ASSERT_UNKNOWN( atom_j < atom_n );
//
// arguemnt variables not avaialbe during sparisty calculations
atom_x[atom_j] = CppAD::numeric_limits<Base>::quiet_NaN();
type_x[atom_j] = variable_enum;
atom_ix[atom_j] = size_t(arg[0]); // variable for this argument
//
++atom_j;
if( atom_j == atom_n )
atom_state = ret_atom;
break;
case FunrpOp:
// parameter result for a atomic function
CPPAD_ASSERT_NARG_NRES(op, 1, 0);
CPPAD_ASSERT_UNKNOWN( atom_state == ret_atom );
CPPAD_ASSERT_UNKNOWN( atom_i < atom_m );
CPPAD_ASSERT_UNKNOWN( atom_j == atom_n );
CPPAD_ASSERT_UNKNOWN( size_t( arg[0] ) < num_par );
//
atom_iy[atom_i] = 0; // special variable used for parameters
# if CPPAD_FOR_HES_TRACE
// remember argument for delayed tracing
atom_funrp[atom_i] = arg[0];
# endif
++atom_i;
if( atom_i == atom_m )
atom_state = end_atom;
break;
case FunrvOp:
// variable result for a atomic function
CPPAD_ASSERT_NARG_NRES(op, 0, 1);
CPPAD_ASSERT_UNKNOWN( atom_state == ret_atom );
CPPAD_ASSERT_UNKNOWN( atom_i < atom_m );
CPPAD_ASSERT_UNKNOWN( atom_j == atom_n );
//
atom_iy[atom_i] = i_var; // variable index for this result
//
++atom_i;
if( atom_i == atom_m )
atom_state = end_atom;
break;
// -------------------------------------------------
case ZmulvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_mul_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
default:
CPPAD_ASSERT_UNKNOWN(0);
}
# if CPPAD_FOR_HES_TRACE
typedef typename Vector_set::const_iterator const_iterator;
if( op == AFunOp && atom_state == start_atom )
{ // print operators that have been delayed
CPPAD_ASSERT_UNKNOWN( atom_m == atom_iy.size() );
CPPAD_ASSERT_UNKNOWN( itr.op_index() > atom_m );
CPPAD_ASSERT_NARG_NRES(FunrpOp, 1, 0);
CPPAD_ASSERT_NARG_NRES(FunrvOp, 0, 1);
addr_t arg_tmp[1];
for(k = 0; k < atom_m; k++)
{ size_t k_var = atom_iy[k];
// value for this variable
for(i = 0; i < limit; i++)
{ zf_value[i] = false;
for(j = 0; j < limit; j++)
zh_value[i * limit + j] = false;
}
const_iterator itr_1(for_jac_sparse, i_var);
j = *itr_1;
while( j < limit )
{ zf_value[j] = true;
j = *(++itr_1);
}
for(i = 0; i < limit; i++)
{ const_iterator itr_2(for_hes_sparse, i);
j = *itr_2;
while( j < limit )
{ zh_value[i * limit + j] = true;
j = *(++itr_2);
}
}
OpCode op_tmp = FunrvOp;
if( k_var == 0 )
{ op_tmp = FunrpOp;
arg_tmp[0] = atom_funrp[k];
}
// k_var is zero when there is no result
printOp<Base, RecBase>(
std::cout,
play,
itr.op_index() - atom_m + k,
k_var,
op_tmp,
arg_tmp
);
if( k_var > 0 ) printOpResult(
std::cout,
1,
&zf_value,
1,
&zh_value
);
std::cout << std::endl;
}
}
for(i = 0; i < limit; i++)
{ zf_value[i] = false;
for(j = 0; j < limit; j++)
zh_value[i * limit + j] = false;
}
const_iterator itr_1(for_jac_sparse, i_var);
j = *itr_1;
while( j < limit )
{ zf_value[j] = true;
j = *(++itr_1);
}
for(i = 0; i < limit; i++)
{ const_iterator itr_2(for_hes_sparse, i);
j = *itr_2;
while( j < limit )
{ zh_value[i * limit + j] = true;
j = *(++itr_2);
}
}
// must delay print for these cases till after atomic function call
bool delay_print = op == FunrpOp;
delay_print |= op == FunrvOp;
if( ! delay_print )
{ printOp<Base, RecBase>(
std::cout,
play,
itr.op_index(),
i_var,
op,
arg
);
if( NumRes(op) > 0 && (! delay_print) ) printOpResult(
std::cout,
1,
&zf_value,
1,
&zh_value
);
std::cout << std::endl;
}
}
std::cout << std::endl;
# else
}
void ForHesSweep(
size_t n,
size_t numvar,
player<Base> *play,
Vector_set& for_jac_sparse, // should be const
Vector_set& rev_jac_sparse, // should be const
Vector_set& for_hes_sparse
)
{
OpCode op;
size_t i_op;
size_t i_var;
const addr_t* arg = CPPAD_NULL;
// length of the parameter vector (used by CppAD assert macros)
const size_t num_par = play->num_par_rec();
size_t i, j, k;
// check numvar argument
size_t limit = n+1;
CPPAD_ASSERT_UNKNOWN( play->num_var_rec() == numvar );
CPPAD_ASSERT_UNKNOWN( for_jac_sparse.n_set() == numvar );
CPPAD_ASSERT_UNKNOWN( for_hes_sparse.n_set() == limit );
CPPAD_ASSERT_UNKNOWN( numvar > 0 );
// upper limit exclusive for set elements
CPPAD_ASSERT_UNKNOWN( for_jac_sparse.end() == limit );
CPPAD_ASSERT_UNKNOWN( for_hes_sparse.end() == limit );
// vecad_sparsity contains a sparsity pattern for each VecAD object.
// vecad_ind maps a VecAD index (beginning of the VecAD object)
// to the index for the corresponding set in vecad_sparsity.
size_t num_vecad_ind = play->num_vec_ind_rec();
size_t num_vecad_vec = play->num_vecad_vec_rec();
Vector_set vecad_sparse;
vecad_sparse.resize(num_vecad_vec, limit);
pod_vector<size_t> vecad_ind;
pod_vector<bool> vecad_jac;
if( num_vecad_vec > 0 )
{ size_t length;
vecad_ind.extend(num_vecad_ind);
vecad_jac.extend(num_vecad_vec);
j = 0;
for(i = 0; i < num_vecad_vec; i++)
{ // length of this VecAD
length = play->GetVecInd(j);
// set vecad_ind to proper index for this VecAD
vecad_ind[j] = i;
// make all other values for this vector invalid
for(k = 1; k <= length; k++)
vecad_ind[j+k] = num_vecad_vec;
// start of next VecAD
j += length + 1;
// initialize this vector's reverse jacobian value
vecad_jac[i] = false;
}
CPPAD_ASSERT_UNKNOWN( j == play->num_vec_ind_rec() );
}
// work space used by UserOp.
vector<size_t> user_ix; // variable indices for argument vector x
vector<Base> user_x; // parameters in x as integers
//
//
typedef std::set<size_t> size_set;
vector< size_set > set_h; // forward Hessian sparsity
vector<bool> bool_h; // forward Hessian sparsity
vectorBool pack_h; // forward Hessian sparstiy
//
vector<bool> user_vx; // which components of x are variables
vector<bool> user_r; // forward Jacobian sparsity for x
vector<bool> user_s; // reverse Jacobian sparsity for y
//
size_t user_index = 0; // indentifier for this atomic operation
size_t user_id = 0; // user identifier for this call to operator
size_t user_i = 0; // index in result vector
size_t user_j = 0; // index in argument vector
size_t user_m = 0; // size of result vector
size_t user_n = 0; // size of arugment vector
//
atomic_base<Base>* user_atom = CPPAD_NULL; // user's atomic op calculator
bool user_pack = false; // sparsity pattern type is pack
bool user_bool = false; // sparsity pattern type is bool
bool user_set = false; // sparsity pattern type is set
bool user_ok = false; // atomic op return value
// next expected operator in a UserOp sequence
enum { user_start, user_arg, user_ret, user_end } user_state = user_start;
//
// pointer to the beginning of the parameter vector
// (used by user atomic functions)
const Base* parameter = CPPAD_NULL;
if( num_par > 0 )
parameter = play->GetPar();
// Initialize
play->forward_start(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( op == BeginOp );
bool more_operators = true;
# if CPPAD_FOR_HES_SWEEP_TRACE
std::cout << std::endl;
CppAD::vectorBool zf_value(limit);
CppAD::vectorBool zh_value(limit * limit);
# endif
while(more_operators)
{
// next op
play->forward_next(op, arg, i_op, i_var);
# ifndef NDEBUG
if( i_op <= n )
{ CPPAD_ASSERT_UNKNOWN((op == InvOp) | (op == BeginOp));
}
else CPPAD_ASSERT_UNKNOWN((op != InvOp) & (op != BeginOp));
# endif
// does the Hessian in question have a non-zero derivative
// with respect to this variable
bool include = rev_jac_sparse.is_element(i_var, 0);
//
// operators to include even if derivative is zero
include |= op == EndOp;
include |= op == CSkipOp;
include |= op == UserOp;
include |= op == UsrapOp;
include |= op == UsravOp;
include |= op == UsrrpOp;
include |= op == UsrrvOp;
//
if( include ) switch( op )
{ // operators that should not occurr
// case BeginOp
// -------------------------------------------------
// operators that do not affect hessian
case AbsOp:
case AddvvOp:
case AddpvOp:
case CExpOp:
case DisOp:
case DivvpOp:
case InvOp:
case LdpOp:
case LdvOp:
case MulpvOp:
case ParOp:
case PriOp:
case SignOp:
case StppOp:
case StpvOp:
case StvpOp:
case StvvOp:
case SubvvOp:
case SubpvOp:
case SubvpOp:
case ZmulpvOp:
case ZmulvpOp:
break;
// -------------------------------------------------
// nonlinear unary operators
case AcosOp:
case AsinOp:
case AtanOp:
case CosOp:
case CoshOp:
case ExpOp:
case LogOp:
case SinOp:
case SinhOp:
case SqrtOp:
case TanOp:
case TanhOp:
# if CPPAD_USE_CPLUSPLUS_2011
case AcoshOp:
case AsinhOp:
case AtanhOp:
case Expm1Op:
case Log1pOp:
# endif
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 )
forward_sparse_hessian_nonlinear_unary_op(
arg[0], for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case CSkipOp:
// CSkipOp has a variable number of arguments and
// reverse_next thinks it one has one argument.
// We must inform reverse_next of this special case.
play->reverse_cskip(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case CSumOp:
// CSumOp has a variable number of arguments and
// reverse_next thinks it one has one argument.
// We must inform reverse_next of this special case.
play->reverse_csum(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case DivvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_div_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case DivpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_nonlinear_unary_op(
arg[1], for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case EndOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 0);
more_operators = false;
break;
// -------------------------------------------------
case ErfOp:
// arg[1] is always the parameter 0
// arg[2] is always the parameter 2 / sqrt(pi)
CPPAD_ASSERT_NARG_NRES(op, 3, 5);
forward_sparse_hessian_nonlinear_unary_op(
arg[0], for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
// -------------------------------------------------
// logical comparision operators
case EqpvOp:
case EqvvOp:
case LtpvOp:
case LtvpOp:
case LtvvOp:
case LepvOp:
case LevpOp:
case LevvOp:
case NepvOp:
case NevvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 0);
break;
// -------------------------------------------------
case MulvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_mul_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case PowpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
forward_sparse_hessian_nonlinear_unary_op(
arg[1], for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case PowvpOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
forward_sparse_hessian_nonlinear_unary_op(
arg[0], for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case PowvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
forward_sparse_hessian_pow_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case UserOp:
// start or end an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( NumRes( UserOp ) == 0 );
CPPAD_ASSERT_UNKNOWN( NumArg( UserOp ) == 4 );
if( user_state == user_start )
{ user_index = arg[0];
user_id = arg[1];
user_n = arg[2];
user_m = arg[3];
user_atom = atomic_base<Base>::class_object(user_index);
# ifndef NDEBUG
if( user_atom == CPPAD_NULL )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base function has been deleted";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
user_x.resize( user_n );
//
user_pack = user_atom->sparsity() ==
atomic_base<Base>::pack_sparsity_enum;
user_bool = user_atom->sparsity() ==
atomic_base<Base>::bool_sparsity_enum;
user_set = user_atom->sparsity() ==
atomic_base<Base>::set_sparsity_enum;
CPPAD_ASSERT_UNKNOWN( user_pack || user_bool || user_set );
user_ix.resize(user_n);
user_vx.resize(user_n);
user_r.resize(user_n);
user_s.resize(user_m);
// simpler to initialize sparsity patterns as empty
for(i = 0; i < user_m; i++)
user_s[i] = false;
for(i = 0; i < user_n; i++)
user_r[i] = false;
if( user_pack )
{ pack_h.resize(user_n * user_n);
for(i = 0; i < user_n; i++)
{ for(j = 0; j < user_n; j++)
pack_h[ i * user_n + j] = false;
}
}
if( user_bool )
{ bool_h.resize(user_n * user_n);
for(i = 0; i < user_n; i++)
{ for(j = 0; j < user_n; j++)
bool_h[ i * user_n + j] = false;
}
}
if( user_set )
{ set_h.resize(user_n);
for(i = 0; i < user_n; i++)
set_h[i].clear();
}
user_j = 0;
user_i = 0;
user_state = user_arg;
}
else
{ CPPAD_ASSERT_UNKNOWN( user_state == user_end );
CPPAD_ASSERT_UNKNOWN( user_index == size_t(arg[0]) );
CPPAD_ASSERT_UNKNOWN( user_id == size_t(arg[1]) );
CPPAD_ASSERT_UNKNOWN( user_n == size_t(arg[2]) );
CPPAD_ASSERT_UNKNOWN( user_m == size_t(arg[3]) );
// call users function for this operation
user_atom->set_id(user_id);
if( user_pack )
{ user_ok = user_atom->for_sparse_hes(
user_vx, user_r, user_s, pack_h, user_x
);
if( ! user_ok ) user_ok = user_atom->for_sparse_hes(
user_vx, user_r, user_s, pack_h
);
}
if( user_bool )
{ user_ok = user_atom->for_sparse_hes(
user_vx, user_r, user_s, bool_h, user_x
);
if( ! user_ok ) user_ok = user_atom->for_sparse_hes(
user_vx, user_r, user_s, bool_h
);
}
if( user_set )
{ user_ok = user_atom->for_sparse_hes(
user_vx, user_r, user_s, set_h, user_x
);
if( ! user_ok ) user_ok = user_atom->for_sparse_hes(
user_vx, user_r, user_s, set_h
);
}
if( ! user_ok )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base.for_sparse_hes: returned false\n";
if( user_pack )
msg += "sparsity = pack_sparsity_enum";
if( user_bool )
msg += "sparsity = bool_sparsity_enum";
if( user_set )
msg += "sparsity = set_sparsity_enum";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
for(i = 0; i < user_n; i++) for(j = 0; j < user_n; j++)
{ if( user_ix[i] > 0 && user_ix[j] > 0 )
{ bool flag = false;
if( user_pack )
flag = pack_h[i * user_n + j];
if( user_bool )
flag = bool_h[i * user_n + j];
if( user_set )
flag = set_h[i].find(j) != set_h[i].end();
if( flag )
{ size_t i_x = user_ix[i];
size_t j_x = user_ix[j];
for_hes_sparse.binary_union(
i_x, i_x, j_x, for_jac_sparse
);
for_hes_sparse.binary_union(
j_x, j_x, i_x, for_jac_sparse
);
}
}
}
//
user_state = user_start;
}
break;
case UsrapOp:
// parameter argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( user_j < user_n );
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
user_ix[user_j] = 0;
user_vx[user_j] = false;
//
// parameters as integers
user_x[user_j] = parameter[arg[0]];
//
++user_j;
if( user_j == user_n )
user_state = user_ret;
break;
case UsravOp:
// variable argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( user_j < user_n );
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) <= i_var );
CPPAD_ASSERT_UNKNOWN( 0 < arg[0] );
user_ix[user_j] = arg[0];
user_vx[user_j] = true;
// variables as integers
user_x[user_j] = CppAD::numeric_limits<Base>::quiet_NaN();
//
for_jac_sparse.begin(arg[0]);
i = for_jac_sparse.next_element();
if( i < for_jac_sparse.end() )
user_r[user_j] = true;
++user_j;
if( user_j == user_n )
user_state = user_ret;
break;
case UsrrpOp:
// parameter result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( user_i < user_m );
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
++user_i;
if( user_i == user_m )
user_state = user_end;
break;
case UsrrvOp:
// variable result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( user_i < user_m );
if( rev_jac_sparse.is_element(i_var, 0) )
user_s[user_i] = true;
++user_i;
if( user_i == user_m )
user_state = user_end;
break;
// -------------------------------------------------
case ZmulvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_mul_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
default:
CPPAD_ASSERT_UNKNOWN(0);
}
# if CPPAD_FOR_HES_SWEEP_TRACE
if( include )
{ for(i = 0; i < limit; i++)
{ zf_value[i] = false;
for(j = 0; j < limit; j++)
zh_value[i * limit + j] = false;
}
for_jac_sparse.begin(i_var);;
j = for_jac_sparse.next_element();;
while( j < limit )
{ zf_value[j] = true;
j = for_jac_sparse.next_element();
}
for(i = 0; i < limit; i++)
{ for_hes_sparse.begin(i);;
j = for_hes_sparse.next_element();;
while( j < limit )
{ zh_value[i * limit + j] = true;
j = for_hes_sparse.next_element();
}
}
printOp(
std::cout,
play,
i_op,
i_var,
op,
arg
);
// should also print RevJac[i_var], but printOpResult does not
// yet allow for this
if( NumRes(op) > 0 && op != BeginOp ) printOpResult(
std::cout,
1,
&zf_value,
1,
&zh_value
);
std::cout << std::endl;
}
}
std::cout << std::endl;
# else
}
AD<Base>& AD<Base>::operator -= (const AD<Base> &right)
{ ADTape<Base> *tape = AD<Base>::tape_ptr();
size_t tape_id = 0;
if( tape != CPPAD_NULL )
tape_id = tape->id_;
// id_ setting for parameters cannot match 0
bool var_left = id_ == tape_id;
bool var_right = right.id_ == tape_id;
CPPAD_ASSERT_KNOWN(
Parameter(*this) || var_left ,
"-=: left operand is a variable for a different thread"
);
CPPAD_ASSERT_KNOWN(
Parameter(right) || var_right ,
"-=: right operand is a variable for a different thread"
);
Base left;
left = value_;
value_ -= right.value_;
if( var_left )
{ if( var_right )
{ // this = variable - variable
CPPAD_ASSERT_UNKNOWN( NumRes(SubvvOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(SubvvOp) == 2 );
// put operand addresses in tape
tape->Rec_.PutArg(taddr_, right.taddr_);
// put operator in the tape
taddr_ = tape->Rec_.PutOp(SubvvOp);
// make this a variable
CPPAD_ASSERT_UNKNOWN( id_ == tape_id );
}
else if( IdenticalZero( right.value_ ) )
{ // this = variable - 0
}
else
{ // this = variable - parameter
CPPAD_ASSERT_UNKNOWN( NumRes(SubvpOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(SubvpOp) == 2 );
// put operand addresses in tape
addr_t p = tape->Rec_.PutPar(right.value_);
tape->Rec_.PutArg(taddr_, p);
// put operator in the tape
taddr_ = tape->Rec_.PutOp(SubvpOp);
// make this a variable
CPPAD_ASSERT_UNKNOWN( id_ == tape_id );
}
}
else if( var_right )
{ // this = parameter - variable
CPPAD_ASSERT_UNKNOWN( NumRes(SubpvOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(SubpvOp) == 2 );
// put operand addresses in tape
addr_t p = tape->Rec_.PutPar(left);
tape->Rec_.PutArg(p, right.taddr_);
// put operator in the tape
taddr_ = tape->Rec_.PutOp(SubpvOp);
// make this a variable
id_ = tape_id;
}
return *this;
}
void ForHesSweep(
size_t n,
size_t numvar,
local::player<Base>* play,
const Vector_set& for_jac_sparse,
const Vector_set& rev_jac_sparse,
Vector_set& for_hes_sparse
)
{
OpCode op;
size_t i_op;
size_t i_var;
const addr_t* arg = CPPAD_NULL;
// length of the parameter vector (used by CppAD assert macros)
const size_t num_par = play->num_par_rec();
size_t i, j, k;
// check numvar argument
size_t limit = n+1;
CPPAD_ASSERT_UNKNOWN( play->num_var_rec() == numvar );
CPPAD_ASSERT_UNKNOWN( for_jac_sparse.n_set() == numvar );
CPPAD_ASSERT_UNKNOWN( for_hes_sparse.n_set() == limit );
CPPAD_ASSERT_UNKNOWN( numvar > 0 );
// upper limit exclusive for set elements
CPPAD_ASSERT_UNKNOWN( for_jac_sparse.end() == limit );
CPPAD_ASSERT_UNKNOWN( for_hes_sparse.end() == limit );
// vecad_sparsity contains a sparsity pattern for each VecAD object.
// vecad_ind maps a VecAD index (beginning of the VecAD object)
// to the index for the corresponding set in vecad_sparsity.
size_t num_vecad_ind = play->num_vec_ind_rec();
size_t num_vecad_vec = play->num_vecad_vec_rec();
Vector_set vecad_sparse;
vecad_sparse.resize(num_vecad_vec, limit);
pod_vector<size_t> vecad_ind;
pod_vector<bool> vecad_jac;
if( num_vecad_vec > 0 )
{ size_t length;
vecad_ind.extend(num_vecad_ind);
vecad_jac.extend(num_vecad_vec);
j = 0;
for(i = 0; i < num_vecad_vec; i++)
{ // length of this VecAD
length = play->GetVecInd(j);
// set vecad_ind to proper index for this VecAD
vecad_ind[j] = i;
// make all other values for this vector invalid
for(k = 1; k <= length; k++)
vecad_ind[j+k] = num_vecad_vec;
// start of next VecAD
j += length + 1;
// initialize this vector's reverse jacobian value
vecad_jac[i] = false;
}
CPPAD_ASSERT_UNKNOWN( j == play->num_vec_ind_rec() );
}
// ------------------------------------------------------------------------
// user's atomic op calculator
atomic_base<Base>* user_atom = CPPAD_NULL; // user's atomic op calculator
//
// work space used by UserOp.
vector<Base> user_x; // value of parameter arguments to function
vector<size_t> user_ix; // variable index (on tape) for each argument
vector<size_t> user_iy; // variable index (on tape) for each result
//
// information set by forward_user (initialization to avoid warnings)
size_t user_old=0, user_m=0, user_n=0, user_i=0, user_j=0;
// information set by forward_user (necessary initialization)
enum_user_state user_state = start_user;
// -------------------------------------------------------------------------
//
// pointer to the beginning of the parameter vector
// (used by user atomic functions)
const Base* parameter = CPPAD_NULL;
if( num_par > 0 )
parameter = play->GetPar();
// Initialize
play->forward_start(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( op == BeginOp );
bool more_operators = true;
# if CPPAD_FOR_HES_SWEEP_TRACE
vector<size_t> user_usrrp; // parameter index for UsrrpOp operators
CppAD::vectorBool zf_value(limit);
CppAD::vectorBool zh_value(limit * limit);
# endif
bool flag; // temporary for use in switch cases below
while(more_operators)
{
// next op
play->forward_next(op, arg, i_op, i_var);
# ifndef NDEBUG
if( i_op <= n )
{ CPPAD_ASSERT_UNKNOWN((op == InvOp) | (op == BeginOp));
}
else CPPAD_ASSERT_UNKNOWN((op != InvOp) & (op != BeginOp));
# endif
// does the Hessian in question have a non-zero derivative
// with respect to this variable
bool include = rev_jac_sparse.is_element(i_var, 0);
//
// operators to include even if derivative is zero
include |= op == EndOp;
include |= op == CSkipOp;
include |= op == CSumOp;
include |= op == UserOp;
include |= op == UsrapOp;
include |= op == UsravOp;
include |= op == UsrrpOp;
include |= op == UsrrvOp;
//
if( include ) switch( op )
{ // operators that should not occurr
// case BeginOp
// -------------------------------------------------
// operators that do not affect hessian
case AbsOp:
case AddvvOp:
case AddpvOp:
case CExpOp:
case DisOp:
case DivvpOp:
case InvOp:
case LdpOp:
case LdvOp:
case MulpvOp:
case ParOp:
case PriOp:
case SignOp:
case StppOp:
case StpvOp:
case StvpOp:
case StvvOp:
case SubvvOp:
case SubpvOp:
case SubvpOp:
case ZmulpvOp:
case ZmulvpOp:
break;
// -------------------------------------------------
// nonlinear unary operators
case AcosOp:
case AsinOp:
case AtanOp:
case CosOp:
case CoshOp:
case ExpOp:
case LogOp:
case SinOp:
case SinhOp:
case SqrtOp:
case TanOp:
case TanhOp:
# if CPPAD_USE_CPLUSPLUS_2011
case AcoshOp:
case AsinhOp:
case AtanhOp:
case Expm1Op:
case Log1pOp:
# endif
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 )
forward_sparse_hessian_nonlinear_unary_op(
arg[0], for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case CSkipOp:
// CSkipOp has a variable number of arguments and
// reverse_next thinks it one has one argument.
// We must inform reverse_next of this special case.
play->forward_cskip(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case CSumOp:
// CSumOp has a variable number of arguments and
// reverse_next thinks it one has one argument.
// We must inform reverse_next of this special case.
play->forward_csum(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case DivvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_div_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case DivpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_nonlinear_unary_op(
arg[1], for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case EndOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 0);
more_operators = false;
break;
// -------------------------------------------------
case ErfOp:
// arg[1] is always the parameter 0
// arg[2] is always the parameter 2 / sqrt(pi)
CPPAD_ASSERT_NARG_NRES(op, 3, 5);
forward_sparse_hessian_nonlinear_unary_op(
arg[0], for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
// -------------------------------------------------
// logical comparision operators
case EqpvOp:
case EqvvOp:
case LtpvOp:
case LtvpOp:
case LtvvOp:
case LepvOp:
case LevpOp:
case LevvOp:
case NepvOp:
case NevvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 0);
break;
// -------------------------------------------------
case MulvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_mul_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case PowpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
forward_sparse_hessian_nonlinear_unary_op(
arg[1], for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case PowvpOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
forward_sparse_hessian_nonlinear_unary_op(
arg[0], for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case PowvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
forward_sparse_hessian_pow_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
case UserOp:
CPPAD_ASSERT_UNKNOWN(
user_state == start_user || user_state == end_user
);
flag = user_state == start_user;
user_atom = play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
if( flag )
{ // start of user atomic operation sequence
user_x.resize( user_n );
user_ix.resize( user_n );
user_iy.resize( user_m );
# if CPPAD_FOR_HES_SWEEP_TRACE
user_usrrp.resize( user_m );
# endif
}
else
{ // end of user atomic operation sequence
user_atom->set_old(user_old);
user_atom->for_sparse_hes(
user_x, user_ix, user_iy,
for_jac_sparse, rev_jac_sparse, for_hes_sparse
);
}
break;
case UsrapOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
// argument parameter value
user_x[user_j] = parameter[arg[0]];
// special variable user for parameters
user_ix[user_j] = 0;
//
play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
break;
case UsravOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) <= i_var );
CPPAD_ASSERT_UNKNOWN( 0 < arg[0] );
// arguemnt varialbes not avaialbe during sparisty calculations
user_x[user_j] = CppAD::numeric_limits<Base>::quiet_NaN();
// varialbe index for this argument
user_ix[user_j] = arg[0];
//
play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
break;
case UsrrpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
// special variable index user for parameters
user_iy[user_i] = 0;
# if CPPAD_FOR_HES_SWEEP_TRACE
// remember argument for delayed tracing
user_usrrp[user_i] = arg[0];
# endif
//
play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
break;
case UsrrvOp:
// variable index for this result
user_iy[user_i] = i_var;
//
play->forward_user(op, user_state,
user_old, user_m, user_n, user_i, user_j
);
break;
// -------------------------------------------------
case ZmulvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
forward_sparse_hessian_mul_op(
arg, for_jac_sparse, for_hes_sparse
);
break;
// -------------------------------------------------
default:
CPPAD_ASSERT_UNKNOWN(0);
}
# if CPPAD_FOR_HES_SWEEP_TRACE
typedef typename Vector_set::const_iterator const_iterator;
if( op == UserOp && user_state == start_user )
{ // print operators that have been delayed
CPPAD_ASSERT_UNKNOWN( user_m == user_iy.size() );
CPPAD_ASSERT_UNKNOWN( i_op > user_m );
CPPAD_ASSERT_NARG_NRES(UsrrpOp, 1, 0);
CPPAD_ASSERT_NARG_NRES(UsrrvOp, 0, 1);
addr_t arg_tmp[1];
for(k = 0; k < user_m; k++)
{ size_t k_var = user_iy[k];
// value for this variable
for(i = 0; i < limit; i++)
{ zf_value[i] = false;
for(j = 0; j < limit; j++)
zh_value[i * limit + j] = false;
}
const_iterator itr_1(for_jac_sparse, i_var);
j = *itr_1;
while( j < limit )
{ zf_value[j] = true;
j = *(++itr_1);
}
for(i = 0; i < limit; i++)
{ const_iterator itr_2(for_hes_sparse, i);
j = *itr_2;
while( j < limit )
{ zh_value[i * limit + j] = true;
j = *(++itr_2);
}
}
OpCode op_tmp = UsrrvOp;
if( k_var == 0 )
{ op_tmp = UsrrpOp;
arg_tmp[0] = user_usrrp[k];
}
// k_var is zero when there is no result
//printOp(
// std::cout,
// play,
// i_op - user_m + k,
// k_var,
// op_tmp,
// arg_tmp
//);
if( k_var > 0 ){
// printOpResult(
// std::cout,
// 1,
// &zf_value,
// 1,
// &zh_value
//);
}
}
}
const addr_t* arg_tmp = arg;
if( op == CSumOp )
arg_tmp = arg - arg[-1] - 4;
if( op == CSkipOp )
arg_tmp = arg - arg[-1] - 7;
for(i = 0; i < limit; i++)
{ zf_value[i] = false;
for(j = 0; j < limit; j++)
zh_value[i * limit + j] = false;
}
const_iterator itr_1(for_jac_sparse, i_var);
j = *itr_1;
while( j < limit )
{ zf_value[j] = true;
j = *(++itr_1);
}
for(i = 0; i < limit; i++)
{ const_iterator itr_2(for_hes_sparse, i);
j = *itr_2;
while( j < limit )
{ zh_value[i * limit + j] = true;
j = *(++itr_2);
}
}
// must delay print for these cases till after atomic user call
bool delay_print = op == UsrrpOp;
delay_print |= op == UsrrvOp;
if( ! delay_print )
{ // printOp(
// std::cout,
// play,
// i_op,
// i_var,
// op,
// arg_tmp
//);
if( NumRes(op) > 0 && (! delay_print) ){
// printOpResult(
// std::cout,
// 1,
// &zf_value,
// 1,
// &zh_value
// );
}
}
}
# else
}
size_t forward_sweep(
std::ostream& s_out,
bool print,
size_t d,
size_t n,
size_t numvar,
player<Base> *Rec,
size_t J,
Base *Taylor
)
{ CPPAD_ASSERT_UNKNOWN( J >= d + 1 );
// op code for current instruction
OpCode op;
// index for current instruction
size_t i_op;
// next variables
size_t i_var;
CPPAD_ASSERT_UNKNOWN( d == 0 || ! print );
# if CPPAD_USE_FORWARD0SWEEP
CPPAD_ASSERT_UNKNOWN( d > 0 );
# else
addr_t* non_const_arg;
# endif
const addr_t* arg = 0;
// temporary indices
size_t i, ell;
// initialize the comparision operator (ComOp) counter
size_t compareCount = 0;
// if this is an order zero calculation, initialize vector indices
pod_vector<size_t> VectorInd; // address for each element
pod_vector<bool> VectorVar; // is element a variable
i = Rec->num_rec_vecad_ind();
if( i > 0 )
{ VectorInd.extend(i);
VectorVar.extend(i);
while(i--)
{ VectorInd[i] = Rec->GetVecInd(i);
VectorVar[i] = false;
}
}
// work space used by UserOp.
const size_t user_k = d; // order of this forward mode calculation
const size_t user_k1 = d+1; // number of orders for this calculation
vector<Base> user_tx; // argument vector Taylor coefficients
vector<Base> user_ty; // result vector Taylor coefficients
vector<size_t> user_iy; // variable indices for results vector
size_t user_index = 0; // indentifier for this user_atomic operation
size_t user_id = 0; // user identifier for this call to operator
size_t user_i = 0; // index in result vector
size_t user_j = 0; // index in argument vector
size_t user_m = 0; // size of result vector
size_t user_n = 0; // size of arugment vector
// next expected operator in a UserOp sequence
enum { user_start, user_arg, user_ret, user_end } user_state = user_start;
// check numvar argument
CPPAD_ASSERT_UNKNOWN( Rec->num_rec_var() == numvar );
// length of the parameter vector (used by CppAD assert macros)
const size_t num_par = Rec->num_rec_par();
// pointer to the beginning of the parameter vector
const Base* parameter = 0;
if( num_par > 0 )
parameter = Rec->GetPar();
# if ! CPPAD_USE_FORWARD0SWEEP
// length of the text vector (used by CppAD assert macros)
const size_t num_text = Rec->num_rec_text();
// pointer to the beginning of the text vector
const char* text = 0;
if( num_text > 0 )
text = Rec->GetTxt(0);
# endif
// skip the BeginOp at the beginning of the recording
Rec->start_forward(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( op == BeginOp );
# if CPPAD_FORWARD_SWEEP_TRACE
std::cout << std::endl;
# endif
while(op != EndOp)
{
// this op
Rec->next_forward(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( (i_op > n) | (op == InvOp) );
CPPAD_ASSERT_UNKNOWN( (i_op <= n) | (op != InvOp) );
// action depends on the operator
switch( op )
{
case AbsOp:
forward_abs_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case AddvvOp:
forward_addvv_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case AddpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_addpv_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case AcosOp:
// sqrt(1 - x * x), acos(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_acos_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case AsinOp:
// sqrt(1 - x * x), asin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_asin_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case AtanOp:
// 1 + x * x, atan(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_atan_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case CSumOp:
// CSumOp has a variable number of arguments and
// next_forward thinks it one has one argument.
// we must inform next_forward of this special case.
Rec->forward_csum(op, arg, i_op, i_var);
forward_csum_op(
d, i_var, arg, num_par, parameter, J, Taylor
);
break;
// -------------------------------------------------
case CExpOp:
forward_cond_op(
d, i_var, arg, num_par, parameter, J, Taylor
);
break;
// ---------------------------------------------------
case ComOp:
# if ! USE_FORWARD0SWEEP
if( d == 0 ) forward_comp_op_0(
compareCount, arg, num_par, parameter, J, Taylor
);
# endif
break;
// ---------------------------------------------------
case CosOp:
// sin(x), cos(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_cos_op(d, i_var, arg[0], J, Taylor);
break;
// ---------------------------------------------------
case CoshOp:
// sinh(x), cosh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_cosh_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case DisOp:
# if ! CPPAD_USE_FORWARD0SWEEP
if( d == 0 )
forward_dis_op_0(i_var, arg, J, Taylor);
else
# endif
{ Taylor[ i_var * J + d] = Base(0);
}
break;
// -------------------------------------------------
case DivvvOp:
forward_divvv_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case DivpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_divpv_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case DivvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_divvp_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case EndOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 0);
break;
// -------------------------------------------------
case ExpOp:
forward_exp_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case InvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 0 );
break;
// -------------------------------------------------
case LdpOp:
# if ! CPPAD_USE_FORWARD0SWEEP
if( d == 0 )
{ non_const_arg = Rec->forward_non_const_arg();
forward_load_p_op_0(
i_var,
non_const_arg,
num_par,
parameter,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
}
else
# endif
{ forward_load_op( op, d, i_var, arg, J, Taylor);
}
break;
// -------------------------------------------------
case LdvOp:
# if ! CPPAD_USE_FORWARD0SWEEP
if( d == 0 )
{ non_const_arg = Rec->forward_non_const_arg();
forward_load_v_op_0(
i_var,
non_const_arg,
num_par,
parameter,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
}
else
# endif
{ forward_load_op( op, d, i_var, arg, J, Taylor);
}
break;
// -------------------------------------------------
case LogOp:
forward_log_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case MulvvOp:
forward_mulvv_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case MulpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_mulpv_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case ParOp:
# if ! CPPAD_USE_FORWARD0SWEEP
if( d == 0 ) forward_par_op_0(
i_var, arg, num_par, parameter, J, Taylor
);
else
# endif
{ Taylor[ i_var * J + d] = Base(0);
}
break;
// -------------------------------------------------
case PowvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_powvp_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case PowpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_powpv_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case PowvvOp:
forward_powvv_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case PriOp:
# if ! CPPAD_USE_FORWARD0SWEEP
if( print ) forward_pri_0(s_out,
i_var, arg, num_text, text, num_par, parameter, J, Taylor
);
# endif
break;
// -------------------------------------------------
case SignOp:
// cos(x), sin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sign_op(d, i_var, arg[0], J, Taylor);
break;
case SinOp:
// cos(x), sin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sin_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case SinhOp:
// cosh(x), sinh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sinh_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case SqrtOp:
forward_sqrt_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case StppOp:
# if ! CPPAD_USE_FORWARD0SWEEP
if( d == 0 )
{ forward_store_pp_op_0(
i_var,
arg,
num_par,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
}
# endif
break;
// -------------------------------------------------
case StpvOp:
# if ! CPPAD_USE_FORWARD0SWEEP
if( d == 0 )
{ forward_store_pv_op_0(
i_var,
arg,
num_par,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
}
# endif
break;
// -------------------------------------------------
case StvpOp:
# if ! CPPAD_USE_FORWARD0SWEEP
if( d == 0 )
{ forward_store_vp_op_0(
i_var,
arg,
num_par,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
}
# endif
break;
// -------------------------------------------------
case StvvOp:
# if ! CPPAD_USE_FORWARD0SWEEP
if( d == 0 )
{ forward_store_vv_op_0(
i_var,
arg,
num_par,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
}
# endif
break;
// -------------------------------------------------
case SubvvOp:
forward_subvv_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case SubpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_subpv_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case SubvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_subvp_op(d, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case TanOp:
// tan(x)^2, tan(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_tan_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case TanhOp:
// tanh(x)^2, tanh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_tanh_op(d, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case UserOp:
// start or end an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( NumRes( UserOp ) == 0 );
CPPAD_ASSERT_UNKNOWN( NumArg( UserOp ) == 4 );
if( user_state == user_start )
{ user_index = arg[0];
user_id = arg[1];
user_n = arg[2];
user_m = arg[3];
if(user_tx.size() < user_n * user_k1)
user_tx.resize(user_n * user_k1);
if(user_ty.size() < user_m * user_k1)
user_ty.resize(user_m * user_k1);
if(user_iy.size() < user_m)
user_iy.resize(user_m);
user_j = 0;
user_i = 0;
user_state = user_arg;
}
else
{ CPPAD_ASSERT_UNKNOWN( user_state == user_end );
CPPAD_ASSERT_UNKNOWN( user_index == size_t(arg[0]) );
CPPAD_ASSERT_UNKNOWN( user_id == size_t(arg[1]) );
CPPAD_ASSERT_UNKNOWN( user_n == size_t(arg[2]) );
CPPAD_ASSERT_UNKNOWN( user_m == size_t(arg[3]) );
user_state = user_start;
// call users function for this operation
user_atomic<Base>::forward(user_index, user_id,
user_k, user_n, user_m, user_tx, user_ty
);
for(i = 0; i < user_m; i++) if( user_iy[i] > 0 )
Taylor[ user_iy[i] * J + user_k ] =
user_ty[ i * user_k1 + user_k ];
}
break;
case UsrapOp:
// parameter argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( user_j < user_n );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
user_tx[user_j * user_k1 + 0] = parameter[ arg[0]];
for(ell = 1; ell < user_k1; ell++)
user_tx[user_j * user_k1 + ell] = Base(0);
++user_j;
if( user_j == user_n )
user_state = user_ret;
break;
case UsravOp:
// variable argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( user_j < user_n );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) <= i_var );
for(ell = 0; ell < user_k1; ell++)
user_tx[user_j * user_k1 + ell] = Taylor[ arg[0] * J + ell];
++user_j;
if( user_j == user_n )
user_state = user_ret;
break;
case UsrrpOp:
// parameter result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( user_i < user_m );
user_iy[user_i] = 0;
user_ty[user_i * user_k1 + 0] = parameter[ arg[0]];
for(ell = 1; ell < user_k; ell++)
user_ty[user_i * user_k1 + ell] = Base(0);
user_i++;
if( user_i == user_m )
user_state = user_end;
break;
case UsrrvOp:
// variable result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( user_i < user_m );
user_iy[user_i] = i_var;
for(ell = 0; ell < user_k; ell++)
user_ty[user_i * user_k1 + ell] = Taylor[ i_var * J + ell];
user_i++;
if( user_i == user_m )
user_state = user_end;
break;
// -------------------------------------------------
default:
CPPAD_ASSERT_UNKNOWN(0);
}
# if CPPAD_FORWARD_SWEEP_TRACE
size_t i_tmp = i_var;
Base* Z_tmp = Taylor + J * i_var;
printOp(
std::cout,
Rec,
i_tmp,
op,
arg,
d + 1,
Z_tmp,
0,
(Base *) CPPAD_NULL
);
}
std::cout << std::endl;
# else
}
AD<Base> operator + (const AD<Base> &left , const AD<Base> &right)
{
// compute the Base part of this AD object
AD<Base> result;
result.value_ = left.value_ + right.value_;
CPPAD_ASSERT_UNKNOWN( Parameter(result) );
// check if there is a recording in progress
ADTape<Base>* tape = AD<Base>::tape_ptr();
if( tape == CPPAD_NULL )
return result;
tape_id_t tape_id = tape->id_;
// tape_id cannot match the default value for tape_id_; i.e., 0
CPPAD_ASSERT_UNKNOWN( tape_id > 0 );
bool var_left = left.tape_id_ == tape_id;
bool var_right = right.tape_id_ == tape_id;
if( var_left )
{ if( var_right )
{ // result = variable + variable
CPPAD_ASSERT_UNKNOWN( NumRes(AddvvOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(AddvvOp) == 2 );
// put operand addresses in tape
tape->Rec_.PutArg(left.taddr_, right.taddr_);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(AddvvOp);
// make result a variable
result.tape_id_ = tape_id;
}
else if( IdenticalZero(right.value_) )
{ // result = variable + 0
result.make_variable(left.tape_id_, left.taddr_);
}
else
{ // result = variable + parameter
// = parameter + variable
CPPAD_ASSERT_UNKNOWN( NumRes(AddpvOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(AddpvOp) == 2 );
// put operand addresses in tape
addr_t p = tape->Rec_.PutPar(right.value_);
tape->Rec_.PutArg(p, left.taddr_);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(AddpvOp);
// make result a variable
result.tape_id_ = tape_id;
}
}
else if( var_right )
{ if( IdenticalZero(left.value_) )
{ // result = 0 + variable
result.make_variable(right.tape_id_, right.taddr_);
}
else
{ // result = parameter + variable
CPPAD_ASSERT_UNKNOWN( NumRes(AddpvOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(AddpvOp) == 2 );
// put operand addresses in tape
addr_t p = tape->Rec_.PutPar(left.value_);
tape->Rec_.PutArg(p, right.taddr_);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(AddpvOp);
// make result a variable
result.tape_id_ = tape_id;
}
}
return result;
}
void printOp(
std::ostream &os ,
const player<Base> *Rec ,
size_t i_var ,
OpCode op ,
const addr_t *ind ,
size_t nfz ,
const Value *fz ,
size_t nrz ,
const Value *rz )
{ size_t i;
CPPAD_ASSERT_KNOWN(
! thread_alloc::in_parallel() ,
"cannot print trace of AD operations in parallel mode"
);
static const char *CompareOpName[] =
{ "Lt", "Le", "Eq", "Ge", "Gt", "Ne" };
static const char *OpName[] = {
"Abs" ,
"Acos" ,
"Addpv" ,
"Addvv" ,
"Asin" ,
"Atan" ,
"Begin" ,
"CExp" ,
"Com" ,
"Cos" ,
"Cosh" ,
"CSum" ,
"Dis" ,
"Divpv" ,
"Divvp" ,
"Divvv" ,
"End" ,
"Exp" ,
"Inv" ,
"Ldp" ,
"Ldv" ,
"Log" ,
"Mulpv" ,
"Mulvv" ,
"Par" ,
"Powpv" ,
"Powvp" ,
"Powvv" ,
"Pri" ,
"Sign" ,
"Sin" ,
"Sinh" ,
"Sqrt" ,
"Stpp" ,
"Stpv" ,
"Stvp" ,
"Stvv" ,
"Subpv" ,
"Subvp" ,
"Subvv" ,
"Tan" ,
"Tanh" ,
"User" ,
"Usrap" ,
"Usrav" ,
"Usrrp" ,
"Usrrv"
};
CPPAD_ASSERT_UNKNOWN(
size_t(NumberOp) == sizeof(OpName) / sizeof(OpName[0])
);
// print operator
printOpField(os, "i=", i_var, 5);
if( op == CExpOp )
{ printOpField(os, "op=", OpName[op], 4);
printOpField(os, "", CompareOpName[ ind[0] ], 3);
}
else if( op == ComOp )
{ printOpField(os, "op=", OpName[op], 3);
printOpField(os, "", CompareOpName[ ind[0] ], 4);
}
else printOpField(os, "op=", OpName[op], 7);
// print other fields
size_t ncol = 5;
switch( op )
{
case CSumOp:
/*
ind[0] = number of addition variables in summation
ind[1] = number of subtraction variables in summation
ind[2] = index of parameter that initializes summation
ind[3], ... , ind[2+ind[0]] = index for positive variables
ind[3+ind[0]], ..., ind[2+ind[0]+ind[1]] = negative variables
ind[3+ind[0]+ind[1]] = ind[0] + ind[1]
*/
CPPAD_ASSERT_UNKNOWN( ind[3+ind[0]+ind[1]] == ind[0]+ind[1] );
printOpField(os, " pr=", Rec->GetPar(ind[2]), ncol);
for(i = 0; i < ind[0]; i++)
printOpField(os, " +v=", ind[3+i], ncol);
for(i = 0; i < ind[1]; i++)
printOpField(os, " -v=", ind[3+ind[0]+i], ncol);
break;
case LdpOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", ind[0], ncol);
printOpField(os, "idx=", ind[1], ncol);
break;
case LdvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", ind[0], ncol);
printOpField(os, " v=", ind[1], ncol);
break;
case StppOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", ind[0], ncol);
printOpField(os, "idx=", ind[1], ncol);
printOpField(os, " pr=", Rec->GetPar(ind[2]), ncol);
break;
case StpvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", ind[0], ncol);
printOpField(os, "idx=", ind[1], ncol);
printOpField(os, " vr=", ind[2], ncol);
break;
case StvpOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", ind[0], ncol);
printOpField(os, " vl=", ind[1], ncol);
printOpField(os, " pr=", Rec->GetPar(ind[2]), ncol);
break;
case StvvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", ind[0], ncol);
printOpField(os, " vl=", ind[1], ncol);
printOpField(os, " vr=", ind[2], ncol);
break;
case AddvvOp:
case DivvvOp:
case MulvvOp:
case PowvvOp:
case SubvvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 2 );
printOpField(os, " vl=", ind[0], ncol);
printOpField(os, " vr=", ind[1], ncol);
break;
case AddpvOp:
case SubpvOp:
case MulpvOp:
case PowpvOp:
case DivpvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 2 );
printOpField(os, " pl=", Rec->GetPar(ind[0]), ncol);
printOpField(os, " vr=", ind[1], ncol);
break;
case DivvpOp:
case PowvpOp:
case SubvpOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 2 );
printOpField(os, " vl=", ind[0], ncol);
printOpField(os, " pr=", Rec->GetPar(ind[1]), ncol);
break;
case AbsOp:
case AcosOp:
case AsinOp:
case AtanOp:
case CosOp:
case CoshOp:
case ExpOp:
case LogOp:
case SignOp:
case SinOp:
case SinhOp:
case SqrtOp:
case UsravOp:
case TanOp:
case TanhOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
printOpField(os, " v=", ind[0], ncol);
break;
case ParOp:
case UsrapOp:
case UsrrpOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
printOpField(os, " p=", Rec->GetPar(ind[0]), ncol);
break;
case UserOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 4 );
{ const char* name = user_atomic<Base>::name(ind[0]);
printOpField(os, " f=", name, ncol);
printOpField(os, " i=", ind[1], ncol);
printOpField(os, " n=", ind[2], ncol);
printOpField(os, " m=", ind[3], ncol);
}
break;
case PriOp:
CPPAD_ASSERT_NARG_NRES(op, 5, 0);
if( ind[0] & 1 )
printOpField(os, " v=", ind[1], ncol);
else printOpField(os, " p=", Rec->GetPar(ind[1]), ncol);
os << "before=\"" << Rec->GetTxt(ind[2]) << "\"";
if( ind[0] & 2 )
printOpField(os, " v=", ind[3], ncol);
else printOpField(os, " p=", Rec->GetPar(ind[3]), ncol);
os << "after=\"" << Rec->GetTxt(ind[4]) << "\"";
break;
case BeginOp:
case EndOp:
case InvOp:
case UsrrvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 0 );
break;
case DisOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 2 );
{ const char* name = discrete<Base>::name(ind[0]);
printOpField(os, " f=", name, ncol);
printOpField(os, " x=", ind[1], ncol);
}
break;
case CExpOp:
CPPAD_ASSERT_UNKNOWN(ind[1] != 0);
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 6 );
if( ind[1] & 1 )
printOpField(os, " vl=", ind[2], ncol);
else printOpField(os, " pl=", Rec->GetPar(ind[2]), ncol);
if( ind[1] & 2 )
printOpField(os, " vr=", ind[3], ncol);
else printOpField(os, " pr=", Rec->GetPar(ind[3]), ncol);
if( ind[1] & 4 )
printOpField(os, " vt=", ind[4], ncol);
else printOpField(os, " pt=", Rec->GetPar(ind[4]), ncol);
if( ind[1] & 8 )
printOpField(os, " vf=", ind[5], ncol);
else printOpField(os, " pf=", Rec->GetPar(ind[5]), ncol);
break;
case ComOp:
CPPAD_ASSERT_UNKNOWN(ind[1] != 0);
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 4 );
if( ind[1] & 1 )
printOpField(os, "res=", 1, ncol);
else printOpField(os, "res=", 0, ncol);
if( ind[1] & 2 )
printOpField(os, " vl=", ind[2], ncol);
else printOpField(os, " pl=", Rec->GetPar(ind[2]), ncol);
if( ind[1] & 4 )
printOpField(os, " vr=", ind[3], ncol);
else printOpField(os, " pr=", Rec->GetPar(ind[3]), ncol);
break;
default:
CPPAD_ASSERT_UNKNOWN(0);
}
size_t k;
if( NumRes(op) > 0 && (op != BeginOp) )
{
for(k = 0; k < nfz; k++)
std::cout << "| fz[" << k << "]=" << fz[k];
for(k = 0; k < nrz; k++)
std::cout << "| rz[" << k << "]=" << rz[k];
}
std::cout << std::endl;
}
void my_init(vector<bool> keepcol){
Partial.extend(num_var_tape_ * 1);
arg_mark_.resize(play_.op_arg_rec_.size());
for(size_t i=0;i<arg_mark_.size();i++)arg_mark_[i]=false;
/* Run a reverse test-sweep to store pointers once */
tape_point tp;
play_.reverse_start(tp.op, tp.op_arg, tp.op_index, tp.var_index);
tp_.resize(tp.op_index+1);
var2op_.resize(tp.var_index+1);
op_mark_.resize(tp.op_index+1);
for(size_t i=0;i<op_mark_.size();i++)op_mark_[i]=0;
user_region_mark_.resize(tp.op_index+1);
for(size_t i=0;i<user_region_mark_.size();i++)user_region_mark_[i]=0;
tp_[tp.op_index]=tp;
/* 1. We need to be able to find out, for a given variable, what operator created
the variable. This is easiest done by looping through the _operators_ because for a
given op we have access to all the resulting variables it creates.
2. We precompute the a vector of "tape_points" so that instead of calling
"reverse_next", we simply get the next tape entry by tp_[i-1].
*/
while(tp.op != BeginOp ){ /* tp.op_index is decremented by one in each iteration ... */
// printTP(tp); /* For debugging */
play_.reverse_next(tp.op, tp.op_arg, tp.op_index, tp.var_index);
/* Csum is special case - see remarks in player.hpp and reverse_sweep.hpp */
if(tp.op == CSumOp)play_.reverse_csum(tp.op, tp.op_arg, tp.op_index, tp.var_index);
for(size_t i=0;i<NumRes(tp.op);i++)var2op_[tp.var_index-i]=tp.op_index;
tp_[tp.op_index]=tp;
markArgs(tp);
}
/* Lookup table: is tape_point within a UserOp region? */
bool user_within=false;
user_region_.resize(tp_.size());
for(size_t i=0;i<tp_.size();i++){
if(tp_[i].op==UserOp){
user_region_[i]=true;
user_within=!user_within;
} else {
user_region_[i]=user_within;
}
}
/* Lookup table: is tape_point a constant (=only fixed effect dependent) ? */
constant_tape_point_.resize(tp_.size());
int indep_var_number=0;
for(size_t i=0;i<tp_.size();i++){
if(tp_[i].op==InvOp){ /* All independent variables are marked according to being
random or fixed effect */
constant_tape_point_[i]=!keepcol[indep_var_number];
indep_var_number++;
} else { /* Mark operator as constant if _all_ arguments are constant */
constant_tape_point_[i] = is_tape_point_constant(i);
}
//std::cout << constant_tape_point_[i] << " "; printTP(tp_[i]);
}
// std::cout << "Total: " << constant_tape_point_.size() << "\n";
// int sum=0; for(int i=0;i<constant_tape_point_.size();i++)sum+=constant_tape_point_[i];
// std::cout << "Constant:" << sum << "\n";
// Calculate pattern
int m=Range();
colpattern.resize(m);
for(int i=0;i<m;i++)my_pattern(i);
for(size_t i=0;i<op_mark_.size();i++)op_mark_[i]=0; /* remember to reset marks */
for(size_t i=0;i<user_region_mark_.size();i++)user_region_mark_[i]=0; /* remember to reset marks */
}
AD<Base> operator - (const AD<Base> &left , const AD<Base> &right)
{ ADTape<Base> *tape = AD<Base>::tape_ptr();
bool var_left, var_right;
# ifdef NDEBUG
if( tape == CPPAD_NULL )
{ var_left = false;
var_right = false;
}
else
{
var_left = left.id_ == tape->id_;
var_right = right.id_ == tape->id_;
}
# else
var_left = Variable(left);
var_right = Variable(right);
CPPAD_ASSERT_KNOWN(
(! var_left) || left.id_ == tape->id_ ,
"- left operand is a variable for a different thread"
);
CPPAD_ASSERT_KNOWN(
(! var_right) || right.id_ == tape->id_ ,
"- right operand is a variable for a different thread"
);
# endif
AD<Base> result;
result.value_ = left.value_ - right.value_;
CPPAD_ASSERT_UNKNOWN( Parameter(result) );
if( var_left )
{ if( var_right )
{ // result = variable - variable
CPPAD_ASSERT_UNKNOWN( NumRes(SubvvOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(SubvvOp) == 2 );
// put operand addresses in tape
tape->Rec_.PutArg(left.taddr_, right.taddr_);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(SubvvOp);
// make result a variable
result.id_ = tape->id_;
}
else if( IdenticalZero(right.value_) )
{ // result = variable - 0
result.make_variable(left.id_, left.taddr_);
}
else
{ // result = variable - parameter
CPPAD_ASSERT_UNKNOWN( NumRes(SubvpOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(SubvpOp) == 2 );
// put operand addresses in tape
size_t p = tape->Rec_.PutPar(right.value_);
tape->Rec_.PutArg(left.taddr_, p);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(SubvpOp);
// make result a variable
result.id_ = tape->id_;
}
}
else if( var_right )
{ // result = parameter - variable
CPPAD_ASSERT_UNKNOWN( NumRes(SubpvOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(SubpvOp) == 2 );
// put operand addresses in tape
size_t p = tape->Rec_.PutPar(left.value_);
tape->Rec_.PutArg(p, right.taddr_);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(SubpvOp);
// make result a variable
result.id_ = tape->id_;
}
return result;
}
void random_setup(
size_t num_var ,
const pod_vector<opcode_t>& op_vec ,
const pod_vector<addr_t>& arg_vec ,
pod_vector<Addr>* op2arg_vec ,
pod_vector<Addr>* op2var_vec ,
pod_vector<Addr>* var2op_vec )
{
if( op2arg_vec->size() != 0 )
{ CPPAD_ASSERT_UNKNOWN( op2arg_vec->size() == op_vec.size() );
CPPAD_ASSERT_UNKNOWN( op2var_vec->size() == op_vec.size() );
CPPAD_ASSERT_UNKNOWN( var2op_vec->size() == num_var );
return;
}
CPPAD_ASSERT_UNKNOWN( op2var_vec->size() == 0 );
CPPAD_ASSERT_UNKNOWN( op2var_vec->size() == 0 );
CPPAD_ASSERT_UNKNOWN( var2op_vec->size() == 0 );
CPPAD_ASSERT_UNKNOWN( OpCode( op_vec[0] ) == BeginOp );
CPPAD_ASSERT_NARG_NRES(BeginOp, 1, 1);
//
size_t num_op = op_vec.size();
size_t var_index = 0;
size_t arg_index = 0;
//
op2arg_vec->resize( num_op );
op2var_vec->resize( num_op );
var2op_vec->resize( num_var );
# ifndef NDEBUG
// value of var2op for auxillary variables is num_op (invalid)
for(size_t i_var = 0; i_var < num_var; ++i_var)
(*var2op_vec)[i_var] = Addr( num_op );
// value of op2var is num_var (invalid) when NumRes(op) = 0
for(size_t i_op = 0; i_op < num_op; ++i_op)
(*op2var_vec)[i_op] = Addr( num_var );
# endif
for(size_t i_op = 0; i_op < num_op; ++i_op)
{ OpCode op = OpCode( op_vec[i_op] );
//
// index of first argument for this operator
(*op2arg_vec)[i_op] = Addr( arg_index );
arg_index += NumArg(op);
//
// index of first result for next operator
var_index += NumRes(op);
if( NumRes(op) > 0 )
{ // index of last (primary) result for this operator
(*op2var_vec)[i_op] = Addr( var_index - 1 );
//
// mapping from primary variable to its operator
(*var2op_vec)[var_index - 1] = Addr( i_op );
}
// CSumOp
if( op == CSumOp )
{ CPPAD_ASSERT_UNKNOWN( NumArg(CSumOp) == 0 );
//
// pointer to first argument for this operator
const addr_t* op_arg = arg_vec.data() + arg_index;
//
// The actual number of arugments for this operator is
// op_arg[4] + 1
// Correct index of first argument for next operator
arg_index += size_t(op_arg[4] + 1);
}
//
// CSkip
if( op == CSkipOp )
{ CPPAD_ASSERT_UNKNOWN( NumArg(CSumOp) == 0 );
//
// pointer to first argument for this operator
const addr_t* op_arg = arg_vec.data() + arg_index;
//
// The actual number of arugments for this operator is
// 7 + op_arg[4] + op_arg[5].
// Correct index of first argument for next operator.
arg_index += size_t(7 + op_arg[4] + op_arg[5]);
}
}
}
inline void forward_cond_op(
size_t q ,
size_t p ,
size_t i_z ,
const addr_t* arg ,
size_t num_par ,
const Base* parameter ,
size_t nc_taylor ,
Base* taylor )
{ Base y_0, y_1, y_2, y_3;
Base zero(0);
Base* z = taylor + i_z * nc_taylor;
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < static_cast<size_t> (CompareNe) );
CPPAD_ASSERT_UNKNOWN( NumArg(CExpOp) == 6 );
CPPAD_ASSERT_UNKNOWN( NumRes(CExpOp) == 1 );
CPPAD_ASSERT_UNKNOWN( arg[1] != 0 );
if( arg[1] & 1 )
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[2]) < i_z );
y_0 = taylor[ arg[2] * nc_taylor + 0 ];
}
else
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[2]) < num_par );
y_0 = parameter[ arg[2] ];
}
if( arg[1] & 2 )
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[3]) < i_z );
y_1 = taylor[ arg[3] * nc_taylor + 0 ];
}
else
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[3]) < num_par );
y_1 = parameter[ arg[3] ];
}
if( q == 0 )
{ if( arg[1] & 4 )
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[4]) < i_z );
y_2 = taylor[ arg[4] * nc_taylor + 0 ];
}
else
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[4]) < num_par );
y_2 = parameter[ arg[4] ];
}
if( arg[1] & 8 )
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[5]) < i_z );
y_3 = taylor[ arg[5] * nc_taylor + 0 ];
}
else
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[5]) < num_par );
y_3 = parameter[ arg[5] ];
}
z[0] = CondExpOp(
CompareOp( arg[0] ),
y_0,
y_1,
y_2,
y_3
);
q++;
}
for(size_t d = q; d <= p; d++)
{ if( arg[1] & 4 )
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[4]) < i_z );
y_2 = taylor[ arg[4] * nc_taylor + d];
}
else y_2 = zero;
if( arg[1] & 8 )
{ CPPAD_ASSERT_UNKNOWN( size_t(arg[5]) < i_z );
y_3 = taylor[ arg[5] * nc_taylor + d];
}
else y_3 = zero;
z[d] = CondExpOp(
CompareOp( arg[0] ),
y_0,
y_1,
y_2,
y_3
);
}
return;
}
inline void reverse_erf_op(
size_t d ,
size_t i_z ,
const addr_t* arg ,
const Base* parameter ,
size_t cap_order ,
const Base* taylor ,
size_t nc_partial ,
Base* partial )
{
// check assumptions
CPPAD_ASSERT_UNKNOWN( NumArg(ErfOp) == 3 );
CPPAD_ASSERT_UNKNOWN( NumRes(ErfOp) == 5 );
CPPAD_ASSERT_UNKNOWN( d < cap_order );
// array used to pass parameter values for sub-operations
addr_t addr[2];
// If pz is zero, make sure this operation has no effect
// (zero times infinity or nan would be non-zero).
Base* pz = partial + i_z * nc_partial;
bool skip(true);
for(size_t i_d = 0; i_d <= d; i_d++)
skip &= IdenticalZero(pz[i_d]);
if( skip )
return;
// convert from final result to first result
i_z -= 4; // 4 = NumRes(ErfOp) - 1;
// Taylor coefficients and partials corresponding to x
const Base* x = taylor + arg[0] * cap_order;
Base* px = partial + arg[0] * nc_partial;
// Taylor coefficients and partials corresponding to z_3
const Base* z_3 = taylor + (i_z+3) * cap_order;
Base* pz_3 = partial + (i_z+3) * nc_partial;
// Taylor coefficients and partials corresponding to z_4
Base* pz_4 = partial + (i_z+4) * nc_partial;
// Reverse z_4
size_t j = d;
while(j)
{ pz_4[j] /= Base(j);
for(size_t k = 1; k <= j; k++)
{ px[k] += pz_4[j] * z_3[j-k] * Base(k);
pz_3[j-k] += pz_4[j] * x[k] * Base(k);
}
j--;
}
px[0] += pz_4[0] * z_3[0];
// z_3 = (2 / sqrt(pi)) * exp( - x * x )
addr[0] = arg[2]; // 2 / sqrt(pi)
addr[1] = i_z + 2; // z_2
reverse_mulpv_op(
d, i_z+3, addr, parameter, cap_order, taylor, nc_partial, partial
);
// z_2 = exp( - x * x )
reverse_exp_op(
d, i_z+2, i_z+1, cap_order, taylor, nc_partial, partial
);
// z_1 = - x * x
addr[0] = arg[1]; // zero
addr[1] = i_z; // z_0
reverse_subpv_op(
d, i_z+1, addr, parameter, cap_order, taylor, nc_partial, partial
);
// z_0 = x * x
addr[0] = arg[0]; // x
addr[1] = arg[0]; // x
reverse_mulvv_op(
d, i_z+0, addr, parameter, cap_order, taylor, nc_partial, partial
);
}
void RevHesSweep(
size_t n,
size_t numvar,
player<Base> *play,
Vector_set& for_jac_sparse, // should be const
bool* RevJac,
Vector_set& rev_hes_sparse
)
{
OpCode op;
size_t i_op;
size_t i_var;
const addr_t* arg = CPPAD_NULL;
// length of the parameter vector (used by CppAD assert macros)
const size_t num_par = play->num_par_rec();
size_t i, j, k;
// check numvar argument
CPPAD_ASSERT_UNKNOWN( play->num_var_rec() == numvar );
CPPAD_ASSERT_UNKNOWN( for_jac_sparse.n_set() == numvar );
CPPAD_ASSERT_UNKNOWN( rev_hes_sparse.n_set() == numvar );
CPPAD_ASSERT_UNKNOWN( numvar > 0 );
// upper limit exclusive for set elements
size_t limit = rev_hes_sparse.end();
CPPAD_ASSERT_UNKNOWN( for_jac_sparse.end() == limit );
// check number of sets match
CPPAD_ASSERT_UNKNOWN(
for_jac_sparse.n_set() == rev_hes_sparse.n_set()
);
// vecad_sparsity contains a sparsity pattern for each VecAD object.
// vecad_ind maps a VecAD index (beginning of the VecAD object)
// to the index for the corresponding set in vecad_sparsity.
size_t num_vecad_ind = play->num_vec_ind_rec();
size_t num_vecad_vec = play->num_vecad_vec_rec();
Vector_set vecad_sparse;
vecad_sparse.resize(num_vecad_vec, limit);
pod_vector<size_t> vecad_ind;
pod_vector<bool> vecad_jac;
if( num_vecad_vec > 0 )
{ size_t length;
vecad_ind.extend(num_vecad_ind);
vecad_jac.extend(num_vecad_vec);
j = 0;
for(i = 0; i < num_vecad_vec; i++)
{ // length of this VecAD
length = play->GetVecInd(j);
// set vecad_ind to proper index for this VecAD
vecad_ind[j] = i;
// make all other values for this vector invalid
for(k = 1; k <= length; k++)
vecad_ind[j+k] = num_vecad_vec;
// start of next VecAD
j += length + 1;
// initialize this vector's reverse jacobian value
vecad_jac[i] = false;
}
CPPAD_ASSERT_UNKNOWN( j == play->num_vec_ind_rec() );
}
// work space used by UserOp.
vector<size_t> user_ix; // variable indices for argument vector x
typedef std::set<size_t> size_set;
size_set::iterator set_itr; // iterator for a standard set
size_set::iterator set_end; // end of iterator sequence
vector< size_set > set_r; // forward Jacobian sparsity for x
vector< size_set > set_u; // reverse Hessian sparsity for y
vector< size_set > set_v; // reverse Hessian sparsity for x
//
vector<bool> bool_r; // bool forward Jacobian sparsity for x
vector<bool> bool_u; // bool reverse Hessian sparsity for y
vector<bool> bool_v; // bool reverse Hessian sparsity for x
//
vectorBool pack_r; // pack forward Jacobian sparsity for x
vectorBool pack_u; // pack reverse Hessian sparsity for y
vectorBool pack_v; // pack reverse Hessian sparsity for x
//
vector<bool> user_vx; // which components of x are variables
vector<bool> user_s; // reverse Jacobian sparsity for y
vector<bool> user_t; // reverse Jacobian sparsity for x
const size_t user_q = limit; // maximum element plus one
size_t user_index = 0; // indentifier for this atomic operation
size_t user_id = 0; // user identifier for this call to operator
size_t user_i = 0; // index in result vector
size_t user_j = 0; // index in argument vector
size_t user_m = 0; // size of result vector
size_t user_n = 0; // size of arugment vector
//
atomic_base<Base>* user_atom = CPPAD_NULL; // user's atomic op calculator
bool user_pack = false; // sparsity pattern type is pack
bool user_bool = false; // sparsity pattern type is bool
bool user_set = false; // sparsity pattern type is set
# ifndef NDEBUG
bool user_ok = false; // atomic op return value
# endif
// next expected operator in a UserOp sequence
enum { user_start, user_arg, user_ret, user_end } user_state = user_end;
// Initialize
play->reverse_start(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( op == EndOp );
# if CPPAD_REV_HES_SWEEP_TRACE
std::cout << std::endl;
CppAD::vectorBool zf_value(limit);
CppAD::vectorBool zh_value(limit);
# endif
bool more_operators = true;
while(more_operators)
{
// next op
play->reverse_next(op, arg, i_op, i_var);
# ifndef NDEBUG
if( i_op <= n )
{ CPPAD_ASSERT_UNKNOWN((op == InvOp) | (op == BeginOp));
}
else CPPAD_ASSERT_UNKNOWN((op != InvOp) & (op != BeginOp));
# endif
// rest of information depends on the case
switch( op )
{
case AbsOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1)
reverse_sparse_hessian_linear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case AddvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_addsub_op(
i_var, arg, RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case AddpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_linear_unary_op(
i_var, arg[1], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case AcosOp:
// sqrt(1 - x * x), acos(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case AcoshOp:
// sqrt(x * x - 1), acosh(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
# endif
// -------------------------------------------------
case AsinOp:
// sqrt(1 - x * x), asin(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case AsinhOp:
// sqrt(1 + x * x), asinh(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
# endif
// -------------------------------------------------
case AtanOp:
// 1 + x * x, atan(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case AtanhOp:
// 1 - x * x, atanh(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
# endif
// -------------------------------------------------
case BeginOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1)
more_operators = false;
break;
// -------------------------------------------------
case CSkipOp:
// CSkipOp has a variable number of arguments and
// reverse_next thinks it one has one argument.
// We must inform reverse_next of this special case.
play->reverse_cskip(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case CSumOp:
// CSumOp has a variable number of arguments and
// reverse_next thinks it one has one argument.
// We must inform reverse_next of this special case.
play->reverse_csum(op, arg, i_op, i_var);
reverse_sparse_hessian_csum_op(
i_var, arg, RevJac, rev_hes_sparse
);
break;
// -------------------------------------------------
case CExpOp:
reverse_sparse_hessian_cond_op(
i_var, arg, num_par, RevJac, rev_hes_sparse
);
break;
// ---------------------------------------------------
case CosOp:
// sin(x), cos(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// ---------------------------------------------------
case CoshOp:
// sinh(x), cosh(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case DisOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
// derivativve is identically zero
break;
// -------------------------------------------------
case DivvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_div_op(
i_var, arg, RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case DivpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[1], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case DivvpOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_linear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case ErfOp:
// arg[1] is always the parameter 0
// arg[2] is always the parameter 2 / sqrt(pi)
CPPAD_ASSERT_NARG_NRES(op, 3, 5);
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case ExpOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case Expm1Op:
CPPAD_ASSERT_NARG_NRES(op, 1, 1)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
# endif
// -------------------------------------------------
case InvOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 1)
// Z is already defined
break;
// -------------------------------------------------
case LdpOp:
reverse_sparse_hessian_load_op(
op,
i_var,
arg,
num_vecad_ind,
vecad_ind.data(),
rev_hes_sparse,
vecad_sparse,
RevJac,
vecad_jac.data()
);
break;
// -------------------------------------------------
case LdvOp:
reverse_sparse_hessian_load_op(
op,
i_var,
arg,
num_vecad_ind,
vecad_ind.data(),
rev_hes_sparse,
vecad_sparse,
RevJac,
vecad_jac.data()
);
break;
// -------------------------------------------------
case EqpvOp:
case EqvvOp:
case LtpvOp:
case LtvpOp:
case LtvvOp:
case LepvOp:
case LevpOp:
case LevvOp:
case NepvOp:
case NevvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 0);
break;
// -------------------------------------------------
case LogOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case Log1pOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
# endif
// -------------------------------------------------
case MulpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_linear_unary_op(
i_var, arg[1], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case MulvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_mul_op(
i_var, arg, RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case ParOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1)
break;
// -------------------------------------------------
case PowpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[1], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case PowvpOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case PowvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 3)
reverse_sparse_hessian_pow_op(
i_var, arg, RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case PriOp:
CPPAD_ASSERT_NARG_NRES(op, 5, 0);
break;
// -------------------------------------------------
case SignOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1);
// Derivative is identiaclly zero
break;
// -------------------------------------------------
case SinOp:
// cos(x), sin(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case SinhOp:
// cosh(x), sinh(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case SqrtOp:
CPPAD_ASSERT_NARG_NRES(op, 1, 1)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case StppOp:
// sparsity cannot propagate through a parameter
CPPAD_ASSERT_NARG_NRES(op, 3, 0)
break;
// -------------------------------------------------
case StpvOp:
reverse_sparse_hessian_store_op(
op,
arg,
num_vecad_ind,
vecad_ind.data(),
rev_hes_sparse,
vecad_sparse,
RevJac,
vecad_jac.data()
);
break;
// -------------------------------------------------
case StvpOp:
// sparsity cannot propagate through a parameter
CPPAD_ASSERT_NARG_NRES(op, 3, 0)
break;
// -------------------------------------------------
case StvvOp:
reverse_sparse_hessian_store_op(
op,
arg,
num_vecad_ind,
vecad_ind.data(),
rev_hes_sparse,
vecad_sparse,
RevJac,
vecad_jac.data()
);
break;
// -------------------------------------------------
case SubvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_addsub_op(
i_var, arg, RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case SubpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_linear_unary_op(
i_var, arg[1], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case SubvpOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_linear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case TanOp:
// tan(x)^2, tan(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case TanhOp:
// tanh(x)^2, tanh(x)
CPPAD_ASSERT_NARG_NRES(op, 1, 2)
reverse_sparse_hessian_nonlinear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case UserOp:
// start or end an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( NumRes( UserOp ) == 0 );
CPPAD_ASSERT_UNKNOWN( NumArg( UserOp ) == 4 );
if( user_state == user_end )
{ user_index = arg[0];
user_id = arg[1];
user_n = arg[2];
user_m = arg[3];
user_atom = atomic_base<Base>::class_object(user_index);
# ifndef NDEBUG
if( user_atom == CPPAD_NULL )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base function has been deleted";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
user_pack = user_atom->sparsity() ==
atomic_base<Base>::pack_sparsity_enum;
user_bool = user_atom->sparsity() ==
atomic_base<Base>::bool_sparsity_enum;
user_set = user_atom->sparsity() ==
atomic_base<Base>::set_sparsity_enum;
CPPAD_ASSERT_UNKNOWN( user_pack || user_bool || user_set );
user_ix.resize(user_n);
user_vx.resize(user_n);
user_s.resize(user_m);
user_t.resize(user_n);
// simpler to initialize all sparsity patterns as empty
for(i = 0; i < user_m; i++)
user_s[i] = false;
for(i = 0; i < user_n; i++)
user_t[i] = false;
if( user_pack )
{ pack_r.resize(user_n * user_q);
pack_u.resize(user_m * user_q);
pack_v.resize(user_n * user_q);
// simpler to initialize all patterns as empty
for(i = 0; i < user_m; i++)
{
for(j = 0; j < user_q; j++)
pack_u[ i * user_q + j] = false;
}
for(i = 0; i < user_n; i++)
{
for(j = 0; j < user_q; j++)
{ pack_r[ i * user_q + j] = false;
pack_v[ i * user_q + j] = false;
}
}
}
if( user_bool )
{ bool_r.resize(user_n * user_q);
bool_u.resize(user_m * user_q);
bool_v.resize(user_n * user_q);
// simpler to initialize all patterns as empty
for(i = 0; i < user_m; i++)
{
for(j = 0; j < user_q; j++)
bool_u[ i * user_q + j] = false;
}
for(i = 0; i < user_n; i++)
{
for(j = 0; j < user_q; j++)
{ bool_r[ i * user_q + j] = false;
bool_v[ i * user_q + j] = false;
}
}
}
if( user_set )
{ set_r.resize(user_n);
set_u.resize(user_m);
set_v.resize(user_n);
for(i = 0; i < user_m; i++)
set_u[i].clear();
for(i = 0; i < user_n; i++)
{ set_r[i].clear();
set_v[i].clear();
}
}
user_j = user_n;
user_i = user_m;
user_state = user_ret;
}
else
{ CPPAD_ASSERT_UNKNOWN( user_state == user_start );
CPPAD_ASSERT_UNKNOWN( user_index == size_t(arg[0]) );
CPPAD_ASSERT_UNKNOWN( user_id == size_t(arg[1]) );
CPPAD_ASSERT_UNKNOWN( user_n == size_t(arg[2]) );
CPPAD_ASSERT_UNKNOWN( user_m == size_t(arg[3]) );
user_state = user_end;
// call users function for this operation
user_atom->set_id(user_id);
# ifdef NDEBUG
if( user_pack )
user_atom->rev_sparse_hes(user_vx,
user_s, user_t, user_q, pack_r, pack_u, pack_v
);
if( user_bool )
user_atom->rev_sparse_hes(user_vx,
user_s, user_t, user_q, bool_r, bool_u, bool_v
);
if( user_set )
user_atom->rev_sparse_hes(user_vx,
user_s, user_t, user_q, set_r, set_u, set_v
);
# else
if( user_pack )
user_ok = user_atom->rev_sparse_hes(user_vx,
user_s, user_t, user_q, pack_r, pack_u, pack_v
);
if( user_bool )
user_ok = user_atom->rev_sparse_hes(user_vx,
user_s, user_t, user_q, bool_r, bool_u, bool_v
);
if( user_set )
user_ok = user_atom->rev_sparse_hes(user_vx,
user_s, user_t, user_q, set_r, set_u, set_v
);
if( ! user_ok )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base.rev_sparse_hes: returned false\n";
if( user_pack )
msg += "sparsity = pack_sparsity_enum";
if( user_bool )
msg += "sparsity = bool_sparsity_enum";
if( user_set )
msg += "sparsity = set_sparsity_enum";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
for(i = 0; i < user_n; i++) if( user_ix[i] > 0 )
{
size_t i_x = user_ix[i];
if( user_t[i] )
RevJac[i_x] = true;
if( user_pack )
{
for(j = 0; j < user_q; j++)
if( pack_v[ i * user_q + j ] )
rev_hes_sparse.add_element(i_x, j);
}
if( user_bool )
{
for(j = 0; j < user_q; j++)
if( bool_v[ i * user_q + j ] )
rev_hes_sparse.add_element(i_x, j);
}
if( user_set )
{
set_itr = set_v[i].begin();
set_end = set_v[i].end();
while( set_itr != set_end )
rev_hes_sparse.add_element(i_x, *set_itr++);
}
}
}
break;
case UsrapOp:
// parameter argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( 0 < user_j && user_j <= user_n );
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
--user_j;
user_ix[user_j] = 0;
user_vx[user_j] = false;
if( user_j == 0 )
user_state = user_start;
break;
case UsravOp:
// variable argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( 0 < user_j && user_j <= user_n );
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) <= i_var );
CPPAD_ASSERT_UNKNOWN( 0 < arg[0] );
--user_j;
user_ix[user_j] = arg[0];
user_vx[user_j] = true;
for_jac_sparse.begin(arg[0]);
i = for_jac_sparse.next_element();
while( i < user_q )
{ if( user_pack )
pack_r[ user_j * user_q + i ] = true;
if( user_bool )
bool_r[ user_j * user_q + i ] = true;
if( user_set )
set_r[user_j].insert(i);
i = for_jac_sparse.next_element();
}
if( user_j == 0 )
user_state = user_start;
break;
case UsrrpOp:
// parameter result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( 0 < user_i && user_i <= user_m );
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
--user_i;
if( user_i == 0 )
user_state = user_arg;
break;
case UsrrvOp:
// variable result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( 0 < user_i && user_i <= user_m );
--user_i;
if( RevJac[i_var] )
{
user_s[user_i] = true;
}
rev_hes_sparse.begin(i_var);
j = rev_hes_sparse.next_element();
while( j < user_q )
{ if( user_pack )
pack_u[user_i * user_q + j] = true;
if( user_bool )
bool_u[user_i * user_q + j] = true;
if( user_set )
set_u[user_i].insert(j);
j = rev_hes_sparse.next_element();
}
if( user_i == 0 )
user_state = user_arg;
break;
// -------------------------------------------------
case ZmulpvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_linear_unary_op(
i_var, arg[1], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case ZmulvpOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_linear_unary_op(
i_var, arg[0], RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
case ZmulvvOp:
CPPAD_ASSERT_NARG_NRES(op, 2, 1)
reverse_sparse_hessian_mul_op(
i_var, arg, RevJac, for_jac_sparse, rev_hes_sparse
);
break;
// -------------------------------------------------
default:
CPPAD_ASSERT_UNKNOWN(0);
}
# if CPPAD_REV_HES_SWEEP_TRACE
for(j = 0; j < limit; j++)
{ zf_value[j] = false;
zh_value[j] = false;
}
for_jac_sparse.begin(i_var);;
j = for_jac_sparse.next_element();;
while( j < limit )
{ zf_value[j] = true;
j = for_jac_sparse.next_element();
}
rev_hes_sparse.begin(i_var);;
j = rev_hes_sparse.next_element();;
while( j < limit )
{ zh_value[j] = true;
j = rev_hes_sparse.next_element();
}
printOp(
std::cout,
play,
i_op,
i_var,
op,
arg
);
// should also print RevJac[i_var], but printOpResult does not
// yet allow for this
if( NumRes(op) > 0 && op != BeginOp ) printOpResult(
std::cout,
1,
&zf_value,
1,
&zh_value
);
std::cout << std::endl;
}
std::cout << std::endl;
# else
}
// case where x and y are AD<Base> -------------------------------------------
template <class Base> AD<Base>
azmul(const AD<Base>& x, const AD<Base>& y)
{
// compute the Base part
AD<Base> result;
result.value_ = azmul(x.value_, y.value_);
// check if there is a recording in progress
ADTape<Base>* tape = AD<Base>::tape_ptr();
if( tape == CPPAD_NULL )
return result;
tape_id_t tape_id = tape->id_;
// tape_id cannot match the default value for tape_id_; i.e., 0
CPPAD_ASSERT_UNKNOWN( tape_id > 0 );
bool var_x = x.tape_id_ == tape_id;
bool var_y = y.tape_id_ == tape_id;
if( var_x )
{ if( var_y )
{ // result = azmul(variable, variable)
CPPAD_ASSERT_UNKNOWN( NumRes(ZmulvvOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(ZmulvvOp) == 2 );
// put operand addresses in tape
tape->Rec_.PutArg(x.taddr_, y.taddr_);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(ZmulvvOp);
// make result a variable
result.tape_id_ = tape_id;
}
else if( IdenticalZero( y.value_ ) )
{ // result = variable * 0
}
else if( IdenticalOne( y.value_ ) )
{ // result = variable * 1
result.make_variable(x.tape_id_, x.taddr_);
}
else
{ // result = zmul(variable, parameter)
CPPAD_ASSERT_UNKNOWN( NumRes(ZmulvpOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(ZmulvpOp) == 2 );
// put operand addresses in tape
addr_t p = tape->Rec_.PutPar(y.value_);
tape->Rec_.PutArg(x.taddr_, p);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(ZmulvpOp);
// make result a variable
result.tape_id_ = tape_id;
}
}
else if( var_y )
{ if( IdenticalZero(x.value_) )
{ // result = 0 * variable
}
else if( IdenticalOne( x.value_ ) )
{ // result = 1 * variable
result.make_variable(y.tape_id_, y.taddr_);
}
else
{ // result = zmul(parameter, variable)
CPPAD_ASSERT_UNKNOWN( NumRes(ZmulpvOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(ZmulpvOp) == 2 );
// put operand addresses in tape
addr_t p = tape->Rec_.PutPar(x.value_);
tape->Rec_.PutArg(p, y.taddr_);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(ZmulpvOp);
// make result a variable
result.tape_id_ = tape_id;
}
}
return result;
}
size_t forward_sweep(
std::ostream& s_out,
const bool print,
const size_t q,
const size_t p,
const size_t n,
const size_t numvar,
player<Base> *Rec,
const size_t J,
Base *Taylor,
CppAD::vector<bool>& cskip_op
)
{ CPPAD_ASSERT_UNKNOWN( J >= p + 1 );
CPPAD_ASSERT_UNKNOWN( q <= p );
// op code for current instruction
OpCode op;
// index for current instruction
size_t i_op;
// next variables
size_t i_var;
// arg (not as a constant)
addr_t* non_const_arg = CPPAD_NULL;
// arg (as a constant)
const addr_t* arg = CPPAD_NULL;
// temporary indices
size_t i, ell;
// initialize the comparision operator (ComOp) counter
size_t compareCount = 0;
pod_vector<size_t> VectorInd; // address for each element
pod_vector<bool> VectorVar; // is element a variable
if( q == 0 )
{
// this includes order zero calculation, initialize vector indices
i = Rec->num_rec_vecad_ind();
if( i > 0 )
{ VectorInd.extend(i);
VectorVar.extend(i);
while(i--)
{ VectorInd[i] = Rec->GetVecInd(i);
VectorVar[i] = false;
}
}
// includes zero order, so initialize conditional skip flags
for(i = 0; i < Rec->num_rec_op(); i++)
cskip_op[i] = false;
}
// Work space used by UserOp. Note User assumes q = p.
const size_t user_p1 = p+1; // number of orders for this user calculation
vector<bool> user_vx; // empty vector
vector<bool> user_vy; // empty vector
vector<Base> user_tx; // argument vector Taylor coefficients
vector<Base> user_ty; // result vector Taylor coefficients
vector<size_t> user_iy; // variable indices for results vector
size_t user_index = 0; // indentifier for this atomic operation
size_t user_id = 0; // user identifier for this call to operator
size_t user_i = 0; // index in result vector
size_t user_j = 0; // index in argument vector
size_t user_m = 0; // size of result vector
size_t user_n = 0; // size of arugment vector
//
atomic_base<Base>* user_atom = CPPAD_NULL; // user's atomic op calculator
# ifndef NDEBUG
bool user_ok = false; // atomic op return value
# endif
//
// next expected operator in a UserOp sequence
enum { user_start, user_arg, user_ret, user_end } user_state = user_start;
// check numvar argument
CPPAD_ASSERT_UNKNOWN( Rec->num_rec_var() == numvar );
// length of the parameter vector (used by CppAD assert macros)
const size_t num_par = Rec->num_rec_par();
// pointer to the beginning of the parameter vector
const Base* parameter = CPPAD_NULL;
if( num_par > 0 )
parameter = Rec->GetPar();
// length of the text vector (used by CppAD assert macros)
const size_t num_text = Rec->num_rec_text();
// pointer to the beginning of the text vector
const char* text = CPPAD_NULL;
if( num_text > 0 )
text = Rec->GetTxt(0);
// skip the BeginOp at the beginning of the recording
Rec->start_forward(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( op == BeginOp );
# if CPPAD_FORWARD_SWEEP_TRACE
std::cout << std::endl;
# endif
bool more_operators = true;
while(more_operators)
{
// this op
Rec->next_forward(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( (i_op > n) | (op == InvOp) );
CPPAD_ASSERT_UNKNOWN( (i_op <= n) | (op != InvOp) );
// check if we are skipping this operation
while( cskip_op[i_op] )
{ if( op == CSumOp )
{ // CSumOp has a variable number of arguments
Rec->forward_csum(op, arg, i_op, i_var);
}
Rec->next_forward(op, arg, i_op, i_var);
}
// action depends on the operator
switch( op )
{
case AbsOp:
forward_abs_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case AddvvOp:
forward_addvv_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case AddpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_addpv_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case AcosOp:
// sqrt(1 - x * x), acos(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_acos_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case AsinOp:
// sqrt(1 - x * x), asin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_asin_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case AtanOp:
// 1 + x * x, atan(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_atan_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case CExpOp:
forward_cond_op(
q, p, i_var, arg, num_par, parameter, J, Taylor
);
break;
// ---------------------------------------------------
case ComOp:
if( q == 0 ) forward_comp_op_0(
compareCount, arg, num_par, parameter, J, Taylor
);
break;
// ---------------------------------------------------
case CosOp:
// sin(x), cos(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_cos_op(q, p, i_var, arg[0], J, Taylor);
break;
// ---------------------------------------------------
case CoshOp:
// sinh(x), cosh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_cosh_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case CSkipOp:
// CSkipOp has a variable number of arguments and
// next_forward thinks it one has one argument.
// we must inform next_forward of this special case.
Rec->forward_cskip(op, arg, i_op, i_var);
if( q == 0 )
{ forward_cskip_op_0(
i_var, arg, num_par, parameter, J, Taylor, cskip_op
);
}
break;
// -------------------------------------------------
case CSumOp:
// CSumOp has a variable number of arguments and
// next_forward thinks it one has one argument.
// we must inform next_forward of this special case.
Rec->forward_csum(op, arg, i_op, i_var);
forward_csum_op(
q, p, i_var, arg, num_par, parameter, J, Taylor
);
break;
// -------------------------------------------------
case DisOp:
i = q;
if( i == 0 )
{ forward_dis_op_0(i_var, arg, J, Taylor);
i++;
}
while(i <= p)
{ Taylor[ i_var * J + i] = Base(0);
i++;
}
break;
// -------------------------------------------------
case DivvvOp:
forward_divvv_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case DivpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_divpv_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case DivvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_divvp_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case EndOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 0);
more_operators = false;
break;
// -------------------------------------------------
case ExpOp:
forward_exp_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case InvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 0 );
break;
// -------------------------------------------------
case LdpOp:
if( q == 0 )
{ non_const_arg = Rec->forward_non_const_arg();
forward_load_p_op_0(
i_var,
non_const_arg,
num_par,
parameter,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
if( q < p )
forward_load_op( op, q+1, p, i_var, arg, J, Taylor);
}
else
{ forward_load_op( op, q, p, i_var, arg, J, Taylor);
}
break;
// -------------------------------------------------
case LdvOp:
if( q == 0 )
{ non_const_arg = Rec->forward_non_const_arg();
forward_load_v_op_0(
i_var,
non_const_arg,
num_par,
parameter,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
if( q < p )
forward_load_op( op, q+1, p, i_var, arg, J, Taylor);
}
else
{ forward_load_op( op, q, p, i_var, arg, J, Taylor);
}
break;
// -------------------------------------------------
case LogOp:
forward_log_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case MulvvOp:
forward_mulvv_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case MulpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_mulpv_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case ParOp:
i = q;
if( i == 0 )
{ forward_par_op_0(
i_var, arg, num_par, parameter, J, Taylor
);
i++;
}
while(i <= p)
{ Taylor[ i_var * J + i] = Base(0);
i++;
}
break;
// -------------------------------------------------
case PowvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_powvp_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case PowpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_powpv_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case PowvvOp:
forward_powvv_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case PriOp:
if( (q == 0) & print ) forward_pri_0(s_out,
i_var, arg, num_text, text, num_par, parameter, J, Taylor
);
break;
// -------------------------------------------------
case SignOp:
// sign(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sign_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case SinOp:
// cos(x), sin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sin_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case SinhOp:
// cosh(x), sinh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sinh_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case SqrtOp:
forward_sqrt_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case StppOp:
if( q == 0 )
{ forward_store_pp_op_0(
i_var,
arg,
num_par,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
}
break;
// -------------------------------------------------
case StpvOp:
if( q == 0 )
{ forward_store_pv_op_0(
i_var,
arg,
num_par,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
}
break;
// -------------------------------------------------
case StvpOp:
if( q == 0 )
{ forward_store_vp_op_0(
i_var,
arg,
num_par,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
}
break;
// -------------------------------------------------
case StvvOp:
if( q == 0 )
{ forward_store_vv_op_0(
i_var,
arg,
num_par,
J,
Taylor,
Rec->num_rec_vecad_ind(),
VectorVar.data(),
VectorInd.data()
);
}
break;
// -------------------------------------------------
case SubvvOp:
forward_subvv_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case SubpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_subpv_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case SubvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_subvp_op(q, p, i_var, arg, parameter, J, Taylor);
break;
// -------------------------------------------------
case TanOp:
// tan(x)^2, tan(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_tan_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case TanhOp:
// tanh(x)^2, tanh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_tanh_op(q, p, i_var, arg[0], J, Taylor);
break;
// -------------------------------------------------
case UserOp:
// start or end an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( NumRes( UserOp ) == 0 );
CPPAD_ASSERT_UNKNOWN( NumArg( UserOp ) == 4 );
if( user_state == user_start )
{ user_index = arg[0];
user_id = arg[1];
user_n = arg[2];
user_m = arg[3];
user_atom = atomic_base<Base>::class_object(user_index);
# ifndef NDEBUG
if( user_atom == CPPAD_NULL )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base function has been deleted";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
if(user_tx.size() != user_n * user_p1)
user_tx.resize(user_n * user_p1);
if(user_ty.size() != user_m * user_p1)
user_ty.resize(user_m * user_p1);
if(user_iy.size() != user_m)
user_iy.resize(user_m);
user_j = 0;
user_i = 0;
user_state = user_arg;
}
else
{ CPPAD_ASSERT_UNKNOWN( user_state == user_end );
CPPAD_ASSERT_UNKNOWN( user_index == size_t(arg[0]) );
CPPAD_ASSERT_UNKNOWN( user_id == size_t(arg[1]) );
CPPAD_ASSERT_UNKNOWN( user_n == size_t(arg[2]) );
CPPAD_ASSERT_UNKNOWN( user_m == size_t(arg[3]) );
// call users function for this operation
user_atom->set_id(user_id);
CPPAD_ATOMIC_CALL(
q, p, user_vx, user_vy, user_tx, user_ty
);
# ifndef NDEBUG
if( ! user_ok )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base.forward: returned false";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
for(i = 0; i < user_m; i++)
if( user_iy[i] > 0 )
for(ell = q; ell <= p; ell++)
Taylor[ user_iy[i] * J + ell ] =
user_ty[ i * user_p1 + ell ];
user_state = user_start;
}
break;
case UsrapOp:
// parameter argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( user_j < user_n );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
user_tx[user_j * user_p1 + 0] = parameter[ arg[0]];
for(ell = 1; ell < user_p1; ell++)
user_tx[user_j * user_p1 + ell] = Base(0);
++user_j;
if( user_j == user_n )
user_state = user_ret;
break;
case UsravOp:
// variable argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( user_j < user_n );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) <= i_var );
for(ell = 0; ell < user_p1; ell++)
user_tx[user_j * user_p1 + ell] = Taylor[ arg[0] * J + ell];
++user_j;
if( user_j == user_n )
user_state = user_ret;
break;
case UsrrpOp:
// parameter result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( user_i < user_m );
user_iy[user_i] = 0;
user_ty[user_i * user_p1 + 0] = parameter[ arg[0]];
for(ell = 1; ell < q; ell++)
user_ty[user_i * user_p1 + ell] = Base(0);
user_i++;
if( user_i == user_m )
user_state = user_end;
break;
case UsrrvOp:
// variable result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( user_i < user_m );
user_iy[user_i] = i_var;
for(ell = 0; ell < q; ell++)
user_ty[user_i * user_p1 + ell] = Taylor[ i_var * J + ell];
user_i++;
if( user_i == user_m )
user_state = user_end;
break;
// -------------------------------------------------
default:
CPPAD_ASSERT_UNKNOWN(0);
}
# if CPPAD_FORWARD_SWEEP_TRACE
size_t i_tmp = i_var;
Base* Z_tmp = Taylor + J * i_var;
printOp(
std::cout,
Rec,
i_op,
i_tmp,
op,
arg,
p + 1,
Z_tmp,
0,
(Base *) CPPAD_NULL
);
}
std::cout << std::endl;
# else
}
void subgraph_info::get_rev(
const play::const_random_iterator<Addr>& random_itr ,
const pod_vector<size_t>& dep_taddr ,
addr_t i_dep ,
pod_vector<addr_t>& subgraph )
{ // check sizes
CPPAD_ASSERT_UNKNOWN( map_user_op_.size() == n_op_ );
// process_range_
CPPAD_ASSERT_UNKNOWN( process_range_[i_dep] == false );
process_range_[i_dep] = true;
// special value; see init_rev_in_subgraph
addr_t depend_yes = addr_t( n_dep_ );
// assumption on i_dep
CPPAD_ASSERT_UNKNOWN( i_dep < depend_yes );
// start with an empty subgraph for this dependent variable
subgraph.resize(0);
// tape index corresponding to this dependent variable
size_t i_var = dep_taddr[i_dep];
// operator corresponding to this dependent variable
size_t i_op = random_itr.var2op(i_var);
i_op = size_t( map_user_op_[i_op] );
// if this variable depends on the selected indepent variables
// process its subgraph
CPPAD_ASSERT_UNKNOWN( in_subgraph_[i_op] != i_dep )
if( in_subgraph_[i_op] <= depend_yes )
{ subgraph.push_back( addr_t(i_op) );
in_subgraph_[i_op] = i_dep;
}
// space used to return set of arguments that are variables
pod_vector<size_t> argument_variable;
// temporary space used by get_argument_variable
pod_vector<bool> work;
// scan all the operators in this subgraph
size_t sub_index = 0;
while(sub_index < subgraph.size() )
{ // this operator connected to this dependent and selected independent
i_op = size_t( subgraph[sub_index] );
CPPAD_ASSERT_UNKNOWN( in_subgraph_[i_op] == i_dep );
//
// There must be a result for this operator
# ifndef NDEBUG
OpCode op = random_itr.get_op(i_op);
CPPAD_ASSERT_UNKNOWN(op == AFunOp || NumRes(op) > 0 );
# endif
//
// which variables are connected to this operator
get_argument_variable(random_itr, i_op, argument_variable, work);
for(size_t j = 0; j < argument_variable.size(); ++j)
{ // add the corresponding operators to the subgraph
size_t j_var = argument_variable[j];
size_t j_op = random_itr.var2op(j_var);
j_op = size_t( map_user_op_[j_op] );
bool add = in_subgraph_[j_op] <= depend_yes;
add &= in_subgraph_[j_op] != i_dep;
if( random_itr.get_op(j_op) == InvOp )
{ CPPAD_ASSERT_UNKNOWN( j_op == j_var );
add &= select_domain_[j_var - 1];
}
if( add )
{ subgraph.push_back( addr_t(j_op) );
in_subgraph_[j_op] = i_dep;
}
}
// we are done scaning this subgraph operator
++sub_index;
}
}
void forward1sweep(
std::ostream& s_out,
const bool print,
const size_t p,
const size_t q,
const size_t n,
const size_t numvar,
player<Base>* play,
const size_t J,
Base* taylor,
bool* cskip_op,
pod_vector<addr_t>& var_by_load_op,
size_t compare_change_count,
size_t& compare_change_number,
size_t& compare_change_op_index
)
{
// number of directions
const size_t r = 1;
CPPAD_ASSERT_UNKNOWN( p <= q );
CPPAD_ASSERT_UNKNOWN( J >= q + 1 );
CPPAD_ASSERT_UNKNOWN( play->num_var_rec() == numvar );
/*
<!-- replace forward0sweep_code_define -->
*/
// op code for current instruction
OpCode op;
// index for current instruction
size_t i_op;
// next variables
size_t i_var;
// operation argument indices
const addr_t* arg = CPPAD_NULL;
// initialize the comparision operator counter
if( p == 0 )
{ compare_change_number = 0;
compare_change_op_index = 0;
}
// If this includes a zero calculation, initialize this information
pod_vector<bool> isvar_by_ind;
pod_vector<size_t> index_by_ind;
if( p == 0 )
{ size_t i;
// this includes order zero calculation, initialize vector indices
size_t num = play->num_vec_ind_rec();
if( num > 0 )
{ isvar_by_ind.extend(num);
index_by_ind.extend(num);
for(i = 0; i < num; i++)
{ index_by_ind[i] = play->GetVecInd(i);
isvar_by_ind[i] = false;
}
}
// includes zero order, so initialize conditional skip flags
num = play->num_op_rec();
for(i = 0; i < num; i++)
cskip_op[i] = false;
}
// work space used by UserOp.
vector<bool> user_vx; // empty vecotor
vector<bool> user_vy; // empty vecotor
vector<Base> user_tx; // argument vector Taylor coefficients
vector<Base> user_ty; // result vector Taylor coefficients
size_t user_index = 0; // indentifier for this atomic operation
size_t user_id = 0; // user identifier for this call to operator
size_t user_i = 0; // index in result vector
size_t user_j = 0; // index in argument vector
size_t user_m = 0; // size of result vector
size_t user_n = 0; // size of arugment vector
//
atomic_base<Base>* user_atom = CPPAD_NULL; // user's atomic op calculator
# ifndef NDEBUG
bool user_ok = false; // atomic op return value
# endif
//
// next expected operator in a UserOp sequence
enum { user_start, user_arg, user_ret, user_end, user_trace }
user_state = user_start;
// length of the parameter vector (used by CppAD assert macros)
const size_t num_par = play->num_par_rec();
// pointer to the beginning of the parameter vector
const Base* parameter = CPPAD_NULL;
if( num_par > 0 )
parameter = play->GetPar();
// length of the text vector (used by CppAD assert macros)
const size_t num_text = play->num_text_rec();
// pointer to the beginning of the text vector
const char* text = CPPAD_NULL;
if( num_text > 0 )
text = play->GetTxt(0);
/*
<!-- end forward0sweep_code_define -->
*/
// temporary indices
size_t i, k;
// number of orders for this user calculation
// (not needed for order zero)
const size_t user_q1 = q+1;
// variable indices for results vector
// (done differently for order zero).
vector<size_t> user_iy;
// skip the BeginOp at the beginning of the recording
play->forward_start(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( op == BeginOp );
# if CPPAD_FORWARD1SWEEP_TRACE
std::cout << std::endl;
# endif
bool more_operators = true;
while(more_operators)
{
// this op
play->forward_next(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( (i_op > n) | (op == InvOp) );
CPPAD_ASSERT_UNKNOWN( (i_op <= n) | (op != InvOp) );
CPPAD_ASSERT_UNKNOWN( i_op < play->num_op_rec() );
CPPAD_ASSERT_ARG_BEFORE_RESULT(op, arg, i_var);
// check if we are skipping this operation
while( cskip_op[i_op] )
{ if( op == CSumOp )
{ // CSumOp has a variable number of arguments
play->forward_csum(op, arg, i_op, i_var);
}
CPPAD_ASSERT_UNKNOWN( op != CSkipOp );
// if( op == CSkipOp )
// { // CSkip has a variable number of arguments
// play->forward_cskip(op, arg, i_op, i_var);
// }
play->forward_next(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( i_op < play->num_op_rec() );
}
// action depends on the operator
switch( op )
{
case AbsOp:
forward_abs_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case AddvvOp:
forward_addvv_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case AddpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_addpv_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case AcosOp:
// sqrt(1 - x * x), acos(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_acos_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case AsinOp:
// sqrt(1 - x * x), asin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_asin_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case AtanOp:
// 1 + x * x, atan(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_atan_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case CExpOp:
forward_cond_op(
p, q, i_var, arg, num_par, parameter, J, taylor
);
break;
// ---------------------------------------------------
case CosOp:
// sin(x), cos(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_cos_op(p, q, i_var, arg[0], J, taylor);
break;
// ---------------------------------------------------
case CoshOp:
// sinh(x), cosh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_cosh_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case CSkipOp:
// CSkipOp has a variable number of arguments and
// forward_next thinks it has no arguments.
// we must inform forward_next of this special case.
if( p == 0 )
{ forward_cskip_op_0(
i_var, arg, num_par, parameter, J, taylor, cskip_op
);
}
play->forward_cskip(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case CSumOp:
// CSumOp has a variable number of arguments and
// forward_next thinks it has no arguments.
// we must inform forward_next of this special case.
forward_csum_op(
p, q, i_var, arg, num_par, parameter, J, taylor
);
play->forward_csum(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case DisOp:
forward_dis_op(p, q, r, i_var, arg, J, taylor);
break;
// -------------------------------------------------
case DivvvOp:
forward_divvv_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case DivpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_divpv_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case DivvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_divvp_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case EndOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 0);
more_operators = false;
break;
// -------------------------------------------------
case EqpvOp:
if( ( p == 0 ) & ( compare_change_count > 0 ) )
{ forward_eqpv_op_0(
compare_change_number, arg, parameter, J, taylor
);
if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
break;
// -------------------------------------------------
case EqvvOp:
if( ( p == 0 ) & ( compare_change_count > 0 ) )
{ forward_eqvv_op_0(
compare_change_number, arg, parameter, J, taylor
);
if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
break;
// -------------------------------------------------
# if CPPAD_COMPILER_HAS_ERF
case ErfOp:
CPPAD_ASSERT_UNKNOWN( CPPAD_COMPILER_HAS_ERF );
// 2DO: implement zero order version of this function
forward_erf_op(p, q, i_var, arg, parameter, J, taylor);
break;
# endif
// -------------------------------------------------
case ExpOp:
forward_exp_op(p, q, i_var, arg[0], J, taylor);
break;
// ---------------------------------------------------
case InvOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 1);
break;
// -------------------------------------------------
case LdpOp:
if( p == 0 )
{ forward_load_p_op_0(
play,
i_var,
arg,
parameter,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data(),
var_by_load_op.data()
);
if( p < q ) forward_load_op(
play,
op,
p+1,
q,
r,
J,
i_var,
arg,
var_by_load_op.data(),
taylor
);
}
else forward_load_op(
play,
op,
p,
q,
r,
J,
i_var,
arg,
var_by_load_op.data(),
taylor
);
break;
// -------------------------------------------------
case LdvOp:
if( p == 0 )
{ forward_load_v_op_0(
play,
i_var,
arg,
parameter,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data(),
var_by_load_op.data()
);
if( p < q ) forward_load_op(
play,
op,
p+1,
q,
r,
J,
i_var,
arg,
var_by_load_op.data(),
taylor
);
}
else forward_load_op(
play,
op,
p,
q,
r,
J,
i_var,
arg,
var_by_load_op.data(),
taylor
);
break;
// -------------------------------------------------
case LepvOp:
if( ( p == 0 ) & ( compare_change_count > 0 ) )
{ forward_lepv_op_0(
compare_change_number, arg, parameter, J, taylor
);
if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
break;
case LevpOp:
if( ( p == 0 ) & ( compare_change_count > 0 ) )
{ forward_levp_op_0(
compare_change_number, arg, parameter, J, taylor
);
if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
break;
// -------------------------------------------------
case LevvOp:
if( ( p == 0 ) & ( compare_change_count > 0 ) )
{ forward_levv_op_0(
compare_change_number, arg, parameter, J, taylor
);
if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
break;
// -------------------------------------------------
case LogOp:
forward_log_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case LtpvOp:
if( ( p == 0 ) & ( compare_change_count > 0 ) )
{ forward_ltpv_op_0(
compare_change_number, arg, parameter, J, taylor
);
if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
break;
case LtvpOp:
if( ( p == 0 ) & ( compare_change_count > 0 ) )
{ forward_ltvp_op_0(
compare_change_number, arg, parameter, J, taylor
);
if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
break;
// -------------------------------------------------
case LtvvOp:
if( ( p == 0 ) & ( compare_change_count > 0 ) )
{ forward_ltvv_op_0(
compare_change_number, arg, parameter, J, taylor
);
if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
break;
// -------------------------------------------------
case MulvvOp:
forward_mulvv_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case MulpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_mulpv_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case NepvOp:
if( ( p == 0 ) & ( compare_change_count > 0 ) )
{ forward_nepv_op_0(
compare_change_number, arg, parameter, J, taylor
);
if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
break;
// -------------------------------------------------
case NevvOp:
if( ( p == 0 ) & ( compare_change_count > 0 ) )
{ forward_nevv_op_0(
compare_change_number, arg, parameter, J, taylor
);
if( compare_change_count == compare_change_number )
compare_change_op_index = i_op;
}
break;
// -------------------------------------------------
case ParOp:
i = p;
if( i == 0 )
{ forward_par_op_0(
i_var, arg, num_par, parameter, J, taylor
);
i++;
}
while(i <= q)
{ taylor[ i_var * J + i] = Base(0);
i++;
}
break;
// -------------------------------------------------
case PowvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_powvp_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case PowpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_powpv_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case PowvvOp:
forward_powvv_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case PriOp:
if( (p == 0) & print ) forward_pri_0(s_out,
arg, num_text, text, num_par, parameter, J, taylor
);
break;
// -------------------------------------------------
case SignOp:
// sign(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sign_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case SinOp:
// cos(x), sin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sin_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case SinhOp:
// cosh(x), sinh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sinh_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case SqrtOp:
forward_sqrt_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case StppOp:
if( p == 0 )
{ forward_store_pp_op_0(
i_var,
arg,
num_par,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data()
);
}
break;
// -------------------------------------------------
case StpvOp:
if( p == 0 )
{ forward_store_pv_op_0(
i_var,
arg,
num_par,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data()
);
}
break;
// -------------------------------------------------
case StvpOp:
if( p == 0 )
{ forward_store_vp_op_0(
i_var,
arg,
num_par,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data()
);
}
break;
// -------------------------------------------------
case StvvOp:
if( p == 0 )
{ forward_store_vv_op_0(
i_var,
arg,
num_par,
J,
taylor,
isvar_by_ind.data(),
index_by_ind.data()
);
}
break;
// -------------------------------------------------
case SubvvOp:
forward_subvv_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case SubpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_subpv_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case SubvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_subvp_op(p, q, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case TanOp:
// tan(x)^2, tan(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_tan_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case TanhOp:
// tanh(x)^2, tanh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_tanh_op(p, q, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case UserOp:
// start or end an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( NumRes( UserOp ) == 0 );
CPPAD_ASSERT_UNKNOWN( NumArg( UserOp ) == 4 );
if( user_state == user_start )
{ user_index = arg[0];
user_id = arg[1];
user_n = arg[2];
user_m = arg[3];
user_atom = atomic_base<Base>::class_object(user_index);
# ifndef NDEBUG
if( user_atom == CPPAD_NULL )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base function has been deleted";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
if(user_tx.size() != user_n * user_q1)
user_tx.resize(user_n * user_q1);
if(user_ty.size() != user_m * user_q1)
user_ty.resize(user_m * user_q1);
if(user_iy.size() != user_m)
user_iy.resize(user_m);
user_j = 0;
user_i = 0;
user_state = user_arg;
}
else
{ CPPAD_ASSERT_UNKNOWN( user_state == user_end );
CPPAD_ASSERT_UNKNOWN( user_index == size_t(arg[0]) );
CPPAD_ASSERT_UNKNOWN( user_id == size_t(arg[1]) );
CPPAD_ASSERT_UNKNOWN( user_n == size_t(arg[2]) );
CPPAD_ASSERT_UNKNOWN( user_m == size_t(arg[3]) );
// call users function for this operation
user_atom->set_id(user_id);
CPPAD_ATOMIC_CALL(
p, q, user_vx, user_vy, user_tx, user_ty
);
# ifndef NDEBUG
if( ! user_ok )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base.forward: returned false";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
for(i = 0; i < user_m; i++)
if( user_iy[i] > 0 )
for(k = p; k <= q; k++)
taylor[ user_iy[i] * J + k ] =
user_ty[ i * user_q1 + k ];
# if CPPAD_FORWARD1SWEEP_TRACE
user_state = user_trace;
# else
user_state = user_start;
# endif
}
break;
case UsrapOp:
// parameter argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( user_j < user_n );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
user_tx[user_j * user_q1 + 0] = parameter[ arg[0]];
for(k = 1; k < user_q1; k++)
user_tx[user_j * user_q1 + k] = Base(0);
++user_j;
if( user_j == user_n )
user_state = user_ret;
break;
case UsravOp:
// variable argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( user_j < user_n );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) <= i_var );
for(k = 0; k < user_q1; k++)
user_tx[user_j * user_q1 + k] = taylor[ arg[0] * J + k];
++user_j;
if( user_j == user_n )
user_state = user_ret;
break;
case UsrrpOp:
// parameter result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( user_i < user_m );
user_iy[user_i] = 0;
user_ty[user_i * user_q1 + 0] = parameter[ arg[0]];
for(k = 1; k < p; k++)
user_ty[user_i * user_q1 + k] = Base(0);
user_i++;
if( user_i == user_m )
user_state = user_end;
break;
case UsrrvOp:
// variable result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( user_i < user_m );
user_iy[user_i] = i_var;
for(k = 0; k < p; k++)
user_ty[user_i * user_q1 + k] = taylor[ i_var * J + k];
user_i++;
if( user_i == user_m )
user_state = user_end;
break;
// -------------------------------------------------
default:
CPPAD_ASSERT_UNKNOWN(0);
}
# if CPPAD_FORWARD1SWEEP_TRACE
if( user_state == user_trace )
{ user_state = user_start;
CPPAD_ASSERT_UNKNOWN( op == UserOp );
CPPAD_ASSERT_UNKNOWN( NumArg(UsrrvOp) == 0 );
for(i = 0; i < user_m; i++) if( user_iy[i] > 0 )
{ size_t i_tmp = (i_op + i) - user_m;
printOp(
std::cout,
play,
i_tmp,
user_iy[i],
UsrrvOp,
CPPAD_NULL
);
Base* Z_tmp = taylor + user_iy[i] * J;
printOpResult(
std::cout,
q + 1,
Z_tmp,
0,
(Base *) CPPAD_NULL
);
std::cout << std::endl;
}
}
Base* Z_tmp = taylor + J * i_var;
const addr_t* arg_tmp = arg;
if( op == CSumOp )
arg_tmp = arg - arg[-1] - 4;
if( op == CSkipOp )
arg_tmp = arg - arg[-1] - 7;
if( op != UsrrvOp )
{
printOp(
std::cout,
play,
i_op,
i_var,
op,
arg_tmp
);
if( NumRes(op) > 0 ) printOpResult(
std::cout,
q + 1,
Z_tmp,
0,
(Base *) CPPAD_NULL
);
std::cout << std::endl;
}
}
std::cout << std::endl;
# else
}
inline void reverse_acos_op(
size_t d ,
size_t i_z ,
size_t i_x ,
size_t cap_order ,
const Base* taylor ,
size_t nc_partial ,
Base* partial )
{
// check assumptions
CPPAD_ASSERT_UNKNOWN( NumArg(AcosOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumRes(AcosOp) == 2 );
CPPAD_ASSERT_UNKNOWN( d < cap_order );
CPPAD_ASSERT_UNKNOWN( d < nc_partial );
// Taylor coefficients and partials corresponding to argument
const Base* x = taylor + i_x * cap_order;
Base* px = partial + i_x * nc_partial;
// Taylor coefficients and partials corresponding to first result
const Base* z = taylor + i_z * cap_order;
Base* pz = partial + i_z * nc_partial;
// Taylor coefficients and partials corresponding to auxillary result
const Base* b = z - cap_order; // called y in documentation
Base* pb = pz - nc_partial;
// If pz is zero, make sure this operation has no effect
// (zero times infinity or nan would be non-zero).
bool skip(true);
for(size_t i_d = 0; i_d <= d; i_d++)
skip &= IdenticalZero(pz[i_d]);
if( skip )
return;
// number of indices to access
size_t j = d;
size_t k;
while(j)
{
// scale partials w.r.t b[j] by 1 / b[0]
pb[j] /= b[0];
// scale partials w.r.t z[j] by 1 / b[0]
pz[j] /= b[0];
// update partials w.r.t b^0
pb[0] -= pz[j] * z[j] + pb[j] * b[j];
// update partial w.r.t. x^0
px[0] -= pb[j] * x[j];
// update partial w.r.t. x^j
px[j] -= pz[j] + pb[j] * x[0];
// further scale partial w.r.t. z[j] by 1 / j
pz[j] /= Base(j);
for(k = 1; k < j; k++)
{ // update partials w.r.t b^(j-k)
pb[j-k] -= Base(k) * pz[j] * z[k] + pb[j] * b[k];
// update partials w.r.t. x^k
px[k] -= pb[j] * x[j-k];
// update partials w.r.t. z^k
pz[k] -= pz[j] * Base(k) * b[j-k];
}
--j;
}
// j == 0 case
px[0] -= ( pz[0] + pb[0] * x[0]) / b[0];
}
inline size_t optimize_binary_match(
const CppAD::vector<struct optimize_old_variable>& tape ,
size_t current ,
size_t npar ,
const Base* par ,
const CppAD::vector<size_t>& hash_table_var ,
unsigned short& code )
{ OpCode op = tape[current].op;
const size_t* arg = tape[current].arg;
size_t new_arg[2];
bool parameter[2];
// initialize return value
size_t match_var = 0;
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 2 );
CPPAD_ASSERT_UNKNOWN( NumRes(op) > 0 );
switch(op)
{ // parameter op variable ----------------------------------
case AddpvOp:
case MulpvOp:
case DivpvOp:
case PowpvOp:
case SubpvOp:
// arg[0]
parameter[0] = true;
new_arg[0] = arg[0];
CPPAD_ASSERT_UNKNOWN( arg[0] < npar );
// arg[1]
parameter[1] = false;
new_arg[1] = tape[arg[1]].new_var;
CPPAD_ASSERT_UNKNOWN( arg[1] < current );
break;
// variable op parameter -----------------------------------
case DivvpOp:
case PowvpOp:
case SubvpOp:
// arg[0]
parameter[0] = false;
new_arg[0] = tape[arg[0]].new_var;
CPPAD_ASSERT_UNKNOWN( arg[0] < current );
// arg[1]
parameter[1] = true;
new_arg[1] = arg[1];
CPPAD_ASSERT_UNKNOWN( arg[1] < npar );
break;
// variable op variable -----------------------------------
case AddvvOp:
case MulvvOp:
case DivvvOp:
case PowvvOp:
case SubvvOp:
// arg[0]
parameter[0] = false;
new_arg[0] = tape[arg[0]].new_var;
CPPAD_ASSERT_UNKNOWN( arg[0] < current );
// arg[1]
parameter[1] = false;
new_arg[1] = tape[arg[1]].new_var;
CPPAD_ASSERT_UNKNOWN( arg[1] < current );
break;
// must be one of the cases above
default:
CPPAD_ASSERT_UNKNOWN(false);
}
code = hash_code(
op ,
new_arg ,
npar ,
par
);
size_t i = hash_table_var[code];
CPPAD_ASSERT_UNKNOWN( i < current );
if( op == tape[i].op )
{ bool match = true;
size_t j;
for(j = 0; j < 2; j++)
{ size_t k = tape[i].arg[j];
if( parameter[j] )
{ CPPAD_ASSERT_UNKNOWN( k < npar );
match &= IdenticalEqualPar(
par[ arg[j] ], par[k]
);
}
else
{ CPPAD_ASSERT_UNKNOWN( k < i );
match &= (new_arg[j] == tape[k].new_var);
}
}
if( match )
match_var = tape[i].new_var;
}
if( (match_var > 0) | ( (op != AddvvOp) & (op != MulvvOp ) ) )
return match_var;
// check for match with argument order switched ----------------------
CPPAD_ASSERT_UNKNOWN( op == AddvvOp || op == MulvvOp );
i = new_arg[0];
new_arg[0] = new_arg[1];
new_arg[1] = i;
unsigned short code_switch = hash_code(
op ,
new_arg ,
npar ,
par
);
i = hash_table_var[code_switch];
CPPAD_ASSERT_UNKNOWN( i < current );
if( op == tape[i].op )
{ bool match = true;
size_t j;
for(j = 0; j < 2; j++)
{ size_t k = tape[i].arg[j];
CPPAD_ASSERT_UNKNOWN( k < i );
match &= (new_arg[j] == tape[k].new_var);
}
if( match )
match_var = tape[i].new_var;
}
return match_var;
}
AD<Base> operator / (const AD<Base> &left , const AD<Base> &right)
{
// compute the Base part
AD<Base> result;
result.value_ = left.value_ / right.value_;
CPPAD_ASSERT_UNKNOWN( Parameter(result) );
// check if there is a recording in progress
ADTape<Base>* tape = AD<Base>::tape_ptr();
if( tape == CPPAD_NULL )
return result;
tape_id_t tape_id = tape->id_;
// tape_id cannot match the default value for tape_id_; i.e., 0
CPPAD_ASSERT_UNKNOWN( tape_id > 0 );
bool var_left = left.tape_id_ == tape_id;
bool var_right = right.tape_id_ == tape_id;
if( var_left )
{ if( var_right )
{ // result = variable / variable
CPPAD_ASSERT_KNOWN(
left.tape_id_ == right.tape_id_,
"Dividing AD objects that are"
" variables on different tapes."
);
CPPAD_ASSERT_UNKNOWN( NumRes(DivvvOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(DivvvOp) == 2 );
// put operand addresses in tape
tape->Rec_.PutArg(left.taddr_, right.taddr_);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(DivvvOp);
// make result a variable
result.tape_id_ = tape_id;
}
else if( IdenticalOne(right.value_) )
{ // result = variable / 1
result.make_variable(left.tape_id_, left.taddr_);
}
else
{ // result = variable / parameter
CPPAD_ASSERT_UNKNOWN( NumRes(DivvpOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(DivvpOp) == 2 );
// put operand addresses in tape
addr_t p = tape->Rec_.PutPar(right.value_);
tape->Rec_.PutArg(left.taddr_, p);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(DivvpOp);
// make result a variable
result.tape_id_ = tape_id;
}
}
else if( var_right )
{ if( IdenticalZero(left.value_) )
{ // result = 0 / variable
}
else
{ // result = parameter / variable
CPPAD_ASSERT_UNKNOWN( NumRes(DivpvOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumArg(DivpvOp) == 2 );
// put operand addresses in tape
addr_t p = tape->Rec_.PutPar(left.value_);
tape->Rec_.PutArg(p, right.taddr_);
// put operator in the tape
result.taddr_ = tape->Rec_.PutOp(DivpvOp);
// make result a variable
result.tape_id_ = tape_id;
}
}
return result;
}
inline void reverse_acosh_op(
size_t d ,
size_t i_z ,
size_t i_x ,
size_t cap_order ,
const Base* taylor ,
size_t nc_partial ,
Base* partial )
{
// check assumptions
CPPAD_ASSERT_UNKNOWN( NumArg(AcoshOp) == 1 );
CPPAD_ASSERT_UNKNOWN( NumRes(AcoshOp) == 2 );
CPPAD_ASSERT_UNKNOWN( d < cap_order );
CPPAD_ASSERT_UNKNOWN( d < nc_partial );
// Taylor coefficients and partials corresponding to argument
const Base* x = taylor + i_x * cap_order;
Base* px = partial + i_x * nc_partial;
// Taylor coefficients and partials corresponding to first result
const Base* z = taylor + i_z * cap_order;
Base* pz = partial + i_z * nc_partial;
// Taylor coefficients and partials corresponding to auxillary result
const Base* b = z - cap_order; // called y in documentation
Base* pb = pz - nc_partial;
Base inv_b0 = Base(1) / b[0];
// number of indices to access
size_t j = d;
size_t k;
while(j)
{
// scale partials w.r.t b[j] by 1 / b[0]
pb[j] = azmul(pb[j], inv_b0);
// scale partials w.r.t z[j] by 1 / b[0]
pz[j] = azmul(pz[j], inv_b0);
// update partials w.r.t b^0
pb[0] -= azmul(pz[j], z[j]) + azmul(pb[j], b[j]);
// update partial w.r.t. x^0
px[0] += azmul(pb[j], x[j]);
// update partial w.r.t. x^j
px[j] += pz[j] + azmul(pb[j], x[0]);
// further scale partial w.r.t. z[j] by 1 / j
pz[j] /= Base(j);
for(k = 1; k < j; k++)
{ // update partials w.r.t b^(j-k)
pb[j-k] -= Base(k) * azmul(pz[j], z[k]) + azmul(pb[j], b[k]);
// update partials w.r.t. x^k
px[k] += azmul(pb[j], x[j-k]);
// update partials w.r.t. z^k
pz[k] -= Base(k) * azmul(pz[j], b[j-k]);
}
--j;
}
// j == 0 case
px[0] += azmul(pz[0] + azmul(pb[0], x[0]), inv_b0);
}
void forward2sweep(
const size_t q,
const size_t r,
const size_t n,
const size_t numvar,
player<Base>* play,
const size_t J,
Base* taylor,
const bool* cskip_op,
const pod_vector<addr_t>& var_by_load_op
)
{
CPPAD_ASSERT_UNKNOWN( q > 0 );
CPPAD_ASSERT_UNKNOWN( J >= q + 1 );
CPPAD_ASSERT_UNKNOWN( play->num_var_rec() == numvar );
// used to avoid compiler errors until all operators are implemented
size_t p = q;
// op code for current instruction
OpCode op;
// index for current instruction
size_t i_op;
// next variables
size_t i_var;
// operation argument indices
const addr_t* arg = CPPAD_NULL;
// work space used by UserOp.
vector<bool> user_vx; // empty vecotor
vector<bool> user_vy; // empty vecotor
vector<Base> user_tx_one; // argument vector Taylor coefficients
vector<Base> user_tx_all;
vector<Base> user_ty_one; // result vector Taylor coefficients
vector<Base> user_ty_all;
size_t user_index = 0; // indentifier for this atomic operation
size_t user_id = 0; // user identifier for this call to operator
size_t user_i = 0; // index in result vector
size_t user_j = 0; // index in argument vector
size_t user_m = 0; // size of result vector
size_t user_n = 0; // size of arugment vector
//
atomic_base<Base>* user_atom = CPPAD_NULL; // user's atomic op calculator
# ifndef NDEBUG
bool user_ok = false; // atomic op return value
# endif
//
// next expected operator in a UserOp sequence
enum { user_start, user_arg, user_ret, user_end, user_trace }
user_state = user_start;
// length of the parameter vector (used by CppAD assert macros)
const size_t num_par = play->num_par_rec();
// pointer to the beginning of the parameter vector
const Base* parameter = CPPAD_NULL;
if( num_par > 0 )
parameter = play->GetPar();
// temporary indices
size_t i, j, k, ell;
// number of orders for this user calculation
// (not needed for order zero)
const size_t user_q1 = q+1;
// variable indices for results vector
// (done differently for order zero).
vector<size_t> user_iy;
// skip the BeginOp at the beginning of the recording
play->forward_start(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( op == BeginOp );
# if CPPAD_FORWARD2SWEEP_TRACE
std::cout << std::endl;
CppAD::vector<Base> Z_vec(q+1);
# endif
bool more_operators = true;
while(more_operators)
{
// this op
play->forward_next(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( (i_op > n) | (op == InvOp) );
CPPAD_ASSERT_UNKNOWN( (i_op <= n) | (op != InvOp) );
CPPAD_ASSERT_UNKNOWN( i_op < play->num_op_rec() );
CPPAD_ASSERT_ARG_BEFORE_RESULT(op, arg, i_var);
// check if we are skipping this operation
while( cskip_op[i_op] )
{ if( op == CSumOp )
{ // CSumOp has a variable number of arguments
play->forward_csum(op, arg, i_op, i_var);
}
CPPAD_ASSERT_UNKNOWN( op != CSkipOp );
// if( op == CSkipOp )
// { // CSkip has a variable number of arguments
// play->forward_cskip(op, arg, i_op, i_var);
// }
play->forward_next(op, arg, i_op, i_var);
CPPAD_ASSERT_UNKNOWN( i_op < play->num_op_rec() );
}
// action depends on the operator
switch( op )
{
case AbsOp:
forward_abs_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case AddvvOp:
forward_addvv_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case AddpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_addpv_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case AcosOp:
// sqrt(1 - x * x), acos(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_acos_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case AcoshOp:
// sqrt(x * x - 1), acosh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_acosh_op_dir(q, r, i_var, arg[0], J, taylor);
break;
# endif
// -------------------------------------------------
case AsinOp:
// sqrt(1 - x * x), asin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_asin_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case AsinhOp:
// sqrt(1 + x * x), asinh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_asinh_op_dir(q, r, i_var, arg[0], J, taylor);
break;
# endif
// -------------------------------------------------
case AtanOp:
// 1 + x * x, atan(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_atan_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case AtanhOp:
// 1 - x * x, atanh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_atanh_op_dir(q, r, i_var, arg[0], J, taylor);
break;
# endif
// -------------------------------------------------
case CExpOp:
forward_cond_op_dir(
q, r, i_var, arg, num_par, parameter, J, taylor
);
break;
// ---------------------------------------------------
case CosOp:
// sin(x), cos(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_cos_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// ---------------------------------------------------
case CoshOp:
// sinh(x), cosh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_cosh_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case CSkipOp:
// CSkipOp has a variable number of arguments and
// forward_next thinks it has no arguments.
// we must inform forward_next of this special case.
play->forward_cskip(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case CSumOp:
// CSumOp has a variable number of arguments and
// forward_next thinks it has no arguments.
// we must inform forward_next of this special case.
forward_csum_op_dir(
q, r, i_var, arg, num_par, parameter, J, taylor
);
play->forward_csum(op, arg, i_op, i_var);
break;
// -------------------------------------------------
case DisOp:
forward_dis_op(p, q, r, i_var, arg, J, taylor);
break;
// -------------------------------------------------
case DivvvOp:
forward_divvv_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case DivpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_divpv_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case DivvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_divvp_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case EndOp:
// needed for sparse_jacobian test
CPPAD_ASSERT_NARG_NRES(op, 0, 0);
more_operators = false;
break;
// -------------------------------------------------
case ExpOp:
forward_exp_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case Expm1Op:
forward_expm1_op_dir(q, r, i_var, arg[0], J, taylor);
break;
# endif
// -------------------------------------------------
case InvOp:
CPPAD_ASSERT_NARG_NRES(op, 0, 1);
break;
// -------------------------------------------------
case LdpOp:
case LdvOp:
forward_load_op(
play,
op,
p,
q,
r,
J,
i_var,
arg,
var_by_load_op.data(),
taylor
);
break;
// ---------------------------------------------------
case EqpvOp:
case EqvvOp:
case LtpvOp:
case LtvpOp:
case LtvvOp:
case LepvOp:
case LevpOp:
case LevvOp:
case NepvOp:
case NevvOp:
CPPAD_ASSERT_UNKNOWN(q > 0 );
break;
// -------------------------------------------------
case LogOp:
forward_log_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// ---------------------------------------------------
# if CPPAD_USE_CPLUSPLUS_2011
case Log1pOp:
forward_log1p_op_dir(q, r, i_var, arg[0], J, taylor);
break;
# endif
// ---------------------------------------------------
case MulvvOp:
forward_mulvv_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case MulpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_mulpv_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case ParOp:
k = i_var*(J-1)*r + i_var + (q-1)*r + 1;
for(ell = 0; ell < r; ell++)
taylor[k + ell] = Base(0);
break;
// -------------------------------------------------
case PowpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_powpv_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case PowvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_powvp_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case PowvvOp:
forward_powvv_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case PriOp:
CPPAD_ASSERT_UNKNOWN(q > 0);
break;
// -------------------------------------------------
case SignOp:
// sign(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sign_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case SinOp:
// cos(x), sin(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sin_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case SinhOp:
// cosh(x), sinh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_sinh_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case SqrtOp:
forward_sqrt_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case StppOp:
case StpvOp:
case StvpOp:
case StvvOp:
CPPAD_ASSERT_UNKNOWN(q > 0 );
break;
// -------------------------------------------------
case SubvvOp:
forward_subvv_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case SubpvOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
forward_subpv_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case SubvpOp:
CPPAD_ASSERT_UNKNOWN( size_t(arg[1]) < num_par );
forward_subvp_op_dir(q, r, i_var, arg, parameter, J, taylor);
break;
// -------------------------------------------------
case TanOp:
// tan(x)^2, tan(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_tan_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case TanhOp:
// tanh(x)^2, tanh(x)
CPPAD_ASSERT_UNKNOWN( i_var < numvar );
forward_tanh_op_dir(q, r, i_var, arg[0], J, taylor);
break;
// -------------------------------------------------
case UserOp:
// start or end an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( NumRes( UserOp ) == 0 );
CPPAD_ASSERT_UNKNOWN( NumArg( UserOp ) == 4 );
if( user_state == user_start )
{ user_index = arg[0];
user_id = arg[1];
user_n = arg[2];
user_m = arg[3];
user_atom = atomic_base<Base>::class_object(user_index);
# ifndef NDEBUG
if( user_atom == CPPAD_NULL )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base function has been deleted";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
if(user_tx_one.size() != user_n * user_q1)
user_tx_one.resize(user_n * user_q1);
if( user_tx_all.size() != user_n * (q * r + 1) )
user_tx_all.resize(user_n * (q * r + 1));
//
if(user_ty_one.size() != user_m * user_q1)
user_ty_one.resize(user_m * user_q1);
if( user_ty_all.size() != user_m * (q * r + 1) )
user_ty_all.resize(user_m * (q * r + 1));
//
if(user_iy.size() != user_m)
user_iy.resize(user_m);
user_j = 0;
user_i = 0;
user_state = user_arg;
}
else
{ CPPAD_ASSERT_UNKNOWN( user_state == user_end );
CPPAD_ASSERT_UNKNOWN( user_index == size_t(arg[0]) );
CPPAD_ASSERT_UNKNOWN( user_id == size_t(arg[1]) );
CPPAD_ASSERT_UNKNOWN( user_n == size_t(arg[2]) );
CPPAD_ASSERT_UNKNOWN( user_m == size_t(arg[3]) );
// call users function for this operation
user_atom->set_id(user_id);
for(ell = 0; ell < r; ell++)
{ // set user_tx
for(j = 0; j < user_n; j++)
{ size_t j_all = j * (q * r + 1);
size_t j_one = j * user_q1;
user_tx_one[j_one+0] = user_tx_all[j_all+0];
for(k = 1; k < user_q1; k++)
{ size_t k_all = j_all + (k-1)*r+1+ell;
size_t k_one = j_one + k;
user_tx_one[k_one] = user_tx_all[k_all];
}
}
// set user_ty
for(i = 0; i < user_m; i++)
{ size_t i_all = i * (q * r + 1);
size_t i_one = i * user_q1;
user_ty_one[i_one+0] = user_ty_all[i_all+0];
for(k = 1; k < q; k++)
{ size_t k_all = i_all + (k-1)*r+1+ell;
size_t k_one = i_one + k;
user_ty_one[k_one] = user_ty_all[k_all];
}
}
CPPAD_ATOMIC_CALL(
q, q, user_vx, user_vy, user_tx_one, user_ty_one
);
# ifndef NDEBUG
if( ! user_ok )
{ std::string msg =
atomic_base<Base>::class_name(user_index)
+ ": atomic_base.forward: returned false";
CPPAD_ASSERT_KNOWN(false, msg.c_str() );
}
# endif
for(i = 0; i < user_m; i++)
{ if( user_iy[i] > 0 )
{ size_t i_taylor = user_iy[i]*((J-1)*r+1);
size_t q_taylor = i_taylor + (q-1)*r+1+ell;
size_t q_one = i * user_q1 + q;
taylor[q_taylor] = user_ty_one[q_one];
}
}
}
# if CPPAD_FORWARD2SWEEP_TRACE
user_state = user_trace;
# else
user_state = user_start;
# endif
}
break;
case UsrapOp:
// parameter argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( user_j < user_n );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) < num_par );
user_tx_all[user_j*(q*r+1) + 0] = parameter[ arg[0]];
for(ell = 0; ell < r; ell++)
for(k = 1; k < user_q1; k++)
user_tx_all[user_j*(q*r+1) + (k-1)*r+1+ell] = Base(0);
++user_j;
if( user_j == user_n )
user_state = user_ret;
break;
case UsravOp:
// variable argument in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_arg );
CPPAD_ASSERT_UNKNOWN( user_j < user_n );
CPPAD_ASSERT_UNKNOWN( size_t(arg[0]) <= i_var );
user_tx_all[user_j*(q*r+1)+0] = taylor[arg[0]*((J-1)*r+1)+0];
for(ell = 0; ell < r; ell++)
{ for(k = 1; k < user_q1; k++)
{ user_tx_all[user_j*(q*r+1) + (k-1)*r+1+ell] =
taylor[arg[0]*((J-1)*r+1) + (k-1)*r+1+ell];
}
}
++user_j;
if( user_j == user_n )
user_state = user_ret;
break;
case UsrrpOp:
// parameter result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( user_i < user_m );
user_iy[user_i] = 0;
user_ty_all[user_i*(q*r+1) + 0] = parameter[ arg[0]];
for(ell = 0; ell < r; ell++)
for(k = 1; k < user_q1; k++)
user_ty_all[user_i*(q*r+1) + (k-1)*r+1+ell] = Base(0);
user_i++;
if( user_i == user_m )
user_state = user_end;
break;
case UsrrvOp:
// variable result in an atomic operation sequence
CPPAD_ASSERT_UNKNOWN( user_state == user_ret );
CPPAD_ASSERT_UNKNOWN( user_i < user_m );
user_iy[user_i] = i_var;
user_ty_all[user_i*(q*r+1)+0] = taylor[i_var*((J-1)*r+1)+0];
for(ell = 0; ell < r; ell++)
{ for(k = 1; k < user_q1; k++)
{ user_ty_all[user_i*(q*r+1) + (k-1)*r+1+ell] =
taylor[i_var*((J-1)*r+1) + (k-1)*r+1+ell];
}
}
user_i++;
if( user_i == user_m )
user_state = user_end;
break;
// -------------------------------------------------
default:
CPPAD_ASSERT_UNKNOWN(0);
}
# if CPPAD_FORWARD2SWEEP_TRACE
if( user_state == user_trace )
{ user_state = user_start;
CPPAD_ASSERT_UNKNOWN( op == UserOp );
CPPAD_ASSERT_UNKNOWN( NumArg(UsrrvOp) == 0 );
for(i = 0; i < user_m; i++) if( user_iy[i] > 0 )
{ size_t i_tmp = (i_op + i) - user_m;
printOp(
std::cout,
play,
i_tmp,
user_iy[i],
UsrrvOp,
CPPAD_NULL
);
Base* Z_tmp = taylor + user_iy[i]*((J-1) * r + 1);
{ Z_vec[0] = Z_tmp[0];
for(ell = 0; ell < r; ell++)
{ std::cout << std::endl << " ";
for(size_t p_tmp = 1; p_tmp <= q; p_tmp++)
Z_vec[p_tmp] = Z_tmp[(p_tmp-1)*r+ell+1];
printOpResult(
std::cout,
q + 1,
Z_vec.data(),
0,
(Base *) CPPAD_NULL
);
}
}
std::cout << std::endl;
}
}
const addr_t* arg_tmp = arg;
if( op == CSumOp )
arg_tmp = arg - arg[-1] - 4;
if( op == CSkipOp )
arg_tmp = arg - arg[-1] - 7;
if( op != UsrrvOp )
{ printOp(
std::cout,
play,
i_op,
i_var,
op,
arg_tmp
);
Base* Z_tmp = CPPAD_NULL;
if( op == UsravOp )
Z_tmp = taylor + arg[0]*((J-1) * r + 1);
else if( NumRes(op) > 0 )
Z_tmp = taylor + i_var*((J-1)*r + 1);
if( Z_tmp != CPPAD_NULL )
{ Z_vec[0] = Z_tmp[0];
for(ell = 0; ell < r; ell++)
{ std::cout << std::endl << " ";
for(size_t p_tmp = 1; p_tmp <= q; p_tmp++)
Z_vec[p_tmp] = Z_tmp[ (p_tmp-1)*r + ell + 1];
printOpResult(
std::cout,
q + 1,
Z_vec.data(),
0,
(Base *) CPPAD_NULL
);
}
}
std::cout << std::endl;
}
}
std::cout << std::endl;
# else
}
void ADFun<Base>::Dependent(ADTape<Base> *tape, const ADvector &y)
{
size_t m = y.size();
size_t n = tape->size_independent_;
size_t i, j;
size_t y_taddr;
// check ADvector is Simple Vector class with AD<Base> elements
CheckSimpleVector< AD<Base>, ADvector>();
CPPAD_ASSERT_KNOWN(
y.size() > 0,
"ADFun operation sequence dependent variable size is zero size"
);
// ---------------------------------------------------------------------
// Begin setting ad_fun.hpp private member data
// ---------------------------------------------------------------------
// dep_parameter_, dep_taddr_
CPPAD_ASSERT_UNKNOWN( NumRes(ParOp) == 1 );
dep_parameter_.resize(m);
dep_taddr_.resize(m);
for(i = 0; i < m; i++)
{ dep_parameter_[i] = CppAD::Parameter(y[i]);
if( dep_parameter_[i] )
{ // make a tape copy of dependent variables that are parameters,
y_taddr = tape->RecordParOp( y[i].value_ );
}
else y_taddr = y[i].taddr_;
CPPAD_ASSERT_UNKNOWN( y_taddr > 0 );
dep_taddr_[i] = y_taddr;
}
// put an EndOp at the end of the tape
tape->Rec_.PutOp(EndOp);
// some size_t values in ad_fun.hpp
has_been_optimized_ = false;
compare_change_count_ = 1;
compare_change_number_ = 0;
compare_change_op_index_ = 0;
num_order_taylor_ = 0;
num_direction_taylor_ = 0;
cap_order_taylor_ = 0;
// num_var_tape_
// Now that all the variables are in the tape, we can set this value.
num_var_tape_ = tape->Rec_.num_var_rec();
// taylor_
taylor_.erase();
// cskip_op_
cskip_op_.erase();
cskip_op_.extend( tape->Rec_.num_op_rec() );
// load_op_
load_op_.erase();
load_op_.extend( tape->Rec_.num_load_op_rec() );
// play_
// Now that each dependent variable has a place in the tape,
// and there is a EndOp at the end of the tape, we can transfer the
// recording to the player and and erase the tape.
play_.get(tape->Rec_);
// ind_taddr_
// Note that play_ has been set, we can use it to check operators
ind_taddr_.resize(n);
CPPAD_ASSERT_UNKNOWN( n < num_var_tape_);
for(j = 0; j < n; j++)
{ CPPAD_ASSERT_UNKNOWN( play_.GetOp(j+1) == InvOp );
ind_taddr_[j] = j+1;
}
// for_jac_sparse_pack_, for_jac_sparse_set_
for_jac_sparse_pack_.resize(0, 0);
for_jac_sparse_set_.resize(0,0);
// ---------------------------------------------------------------------
// End set ad_fun.hpp private member data
// ---------------------------------------------------------------------
// now we can delete the tape
AD<Base>::tape_manage(tape_manage_delete);
// total number of varables in this recording
CPPAD_ASSERT_UNKNOWN( num_var_tape_ == play_.num_var_rec() );
// used to determine if there is an operation sequence in *this
CPPAD_ASSERT_UNKNOWN( num_var_tape_ > 0 );
}
void printOp(
std::ostream& os ,
const local::player<Base>* play,
size_t i_op ,
size_t i_var ,
OpCode op ,
const addr_t* arg )
{
CPPAD_ASSERT_KNOWN(
! thread_alloc::in_parallel() ,
"cannot print trace of AD operations in parallel mode"
);
static const char *CompareOpName[] =
{ "Lt", "Le", "Eq", "Ge", "Gt", "Ne" };
// print operator
printOpField(os, "o=", i_op, 5);
if( NumRes(op) > 0 && op != BeginOp )
printOpField(os, "v=", i_var, 5);
else
printOpField(os, "v=", "", 5);
if( op == CExpOp || op == CSkipOp )
{ printOpField(os, "", OpName(op), 5);
printOpField(os, "", CompareOpName[ arg[0] ], 3);
}
else
printOpField(os, "", OpName(op), 8);
// print other fields
size_t ncol = 5;
switch( op )
{
case CSkipOp:
/*
arg[0] = the Rel operator: Lt, Le, Eq, Ge, Gt, or Ne
arg[1] & 1 = is left a variable
arg[1] & 2 = is right a variable
arg[2] = index correspoding to left
arg[3] = index correspoding to right
arg[4] = number of operations to skip if CExpOp comparision is true
arg[5] = number of operations to skip if CExpOp comparision is false
arg[6] -> arg[5+arg[4]] = skip operations if true
arg[6+arg[4]] -> arg[5+arg[4]+arg[5]] = skip operations if false
arg[6+arg[4]+arg[5]] = arg[4] + arg[5]
*/
CPPAD_ASSERT_UNKNOWN( arg[6+arg[4]+arg[5]] == arg[4]+arg[5] );
CPPAD_ASSERT_UNKNOWN(arg[1] != 0);
if( arg[1] & 1 )
printOpField(os, " vl=", arg[2], ncol);
else
printOpField(os, " pl=", play->GetPar(arg[2]), ncol);
if( arg[1] & 2 )
printOpField(os, " vr=", arg[3], ncol);
else
printOpField(os, " pr=", play->GetPar(arg[3]), ncol);
if( size_t(arg[4]) < 3 )
{ for(addr_t i = 0; i < arg[4]; i++)
printOpField(os, " ot=", arg[6+i], ncol);
}
else
{ printOpField(os, "\n\tot=", arg[6+0], ncol);
for(addr_t i = 1; i < arg[4]; i++)
printOpField(os, " ot=", arg[6+i], ncol);
}
if( size_t(arg[5]) < 3 )
{ for(addr_t i = 0; i < arg[5]; i++)
printOpField(os, " of=", arg[6+arg[4]+i], ncol);
}
else
{ printOpField(os, "\n\tof=", arg[6+arg[4]+0], ncol);
{ for(addr_t i = 1; i < arg[5]; i++)
printOpField(os, " of=", arg[6+arg[4]+i], ncol);
}
}
break;
case CSumOp:
/*
arg[0] = index of parameter that initializes summation
arg[1] = end in arg of addition variables in summation
arg[2] = end in arg of subtraction variables in summation
arg[3] = end in arg of addition dynamic parameters in summation
arg[4] = end in arg of subtraction dynamic parameters in summation
arg[5], ... , arg[arg[1]-1]: indices for addition variables
arg[arg[1]], ... , arg[arg[2]-1]: indices for subtraction variables
arg[arg[2]], ... , arg[arg[3]-1]: indices for additon dynamics
arg[arg[3]], ... , arg[arg[4]-1]: indices for subtraction dynamics
arg[arg[4]] = arg[4]
*/
CPPAD_ASSERT_UNKNOWN( arg[arg[4]] == arg[4] );
printOpField(os, " pr=", play->GetPar(arg[0]), ncol);
for(addr_t i = 5; i < arg[1]; i++)
printOpField(os, " +v=", arg[i], ncol);
for(addr_t i = arg[1]; i < arg[2]; i++)
printOpField(os, " -v=", arg[i], ncol);
for(addr_t i = arg[2]; i < arg[3]; i++)
printOpField(os, " +d=", play->GetPar(arg[i]), ncol);
for(addr_t i = arg[3]; i < arg[4]; i++)
printOpField(os, " -d=", play->GetPar(arg[i]), ncol);
break;
case LdpOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", arg[0], ncol);
printOpField(os, "idx=", arg[1], ncol);
break;
case LdvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", arg[0], ncol);
printOpField(os, " v=", arg[1], ncol);
break;
case StppOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", arg[0], ncol);
printOpField(os, "idx=", arg[1], ncol);
printOpField(os, " pr=", play->GetPar(arg[2]), ncol);
break;
case StpvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", arg[0], ncol);
printOpField(os, "idx=", arg[1], ncol);
printOpField(os, " vr=", arg[2], ncol);
break;
case StvpOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", arg[0], ncol);
printOpField(os, " vl=", arg[1], ncol);
printOpField(os, " pr=", play->GetPar(arg[2]), ncol);
break;
case StvvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
printOpField(os, "off=", arg[0], ncol);
printOpField(os, " vl=", arg[1], ncol);
printOpField(os, " vr=", arg[2], ncol);
break;
case AddvvOp:
case DivvvOp:
case EqvvOp:
case LevvOp:
case LtvvOp:
case NevvOp:
case MulvvOp:
case PowvvOp:
case SubvvOp:
case ZmulvvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 2 );
printOpField(os, " vl=", arg[0], ncol);
printOpField(os, " vr=", arg[1], ncol);
break;
case AddpvOp:
case EqpvOp:
case DivpvOp:
case LepvOp:
case LtpvOp:
case NepvOp:
case SubpvOp:
case MulpvOp:
case PowpvOp:
case ZmulpvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 2 );
printOpField(os, " pl=", play->GetPar(arg[0]), ncol);
printOpField(os, " vr=", arg[1], ncol);
break;
case DivvpOp:
case LevpOp:
case LtvpOp:
case PowvpOp:
case SubvpOp:
case ZmulvpOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 2 );
printOpField(os, " vl=", arg[0], ncol);
printOpField(os, " pr=", play->GetPar(arg[1]), ncol);
break;
case AbsOp:
case AcosOp:
case AcoshOp:
case AsinOp:
case AsinhOp:
case AtanOp:
case AtanhOp:
case CosOp:
case CoshOp:
case ExpOp:
case Expm1Op:
case LogOp:
case Log1pOp:
case SignOp:
case SinOp:
case SinhOp:
case SqrtOp:
case FunavOp:
case TanOp:
case TanhOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
printOpField(os, " v=", arg[0], ncol);
break;
case ErfOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 3 );
// arg[1] points to the parameter 0
// arg[2] points to the parameter 2 / sqrt(pi)
printOpField(os, " v=", arg[0], ncol);
break;
case ParOp:
case FunapOp:
case FunrpOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
printOpField(os, " p=", play->GetPar(arg[0]), ncol);
break;
case AFunOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 4 );
{
// get the name of this atomic function
bool set_null = false;
size_t atom_index = size_t( arg[0] );
size_t type = 0; // set to avoid warning
std::string name;
void* v_ptr = CPPAD_NULL; // set to avoid warning
atomic_index<RecBase>(set_null, atom_index, type, &name, v_ptr);
printOpField(os, " f=", name.c_str(), ncol);
printOpField(os, " i=", arg[1], ncol);
printOpField(os, " n=", arg[2], ncol);
printOpField(os, " m=", arg[3], ncol);
}
break;
case PriOp:
CPPAD_ASSERT_NARG_NRES(op, 5, 0);
if( arg[0] & 1 )
printOpField(os, " v=", arg[1], ncol);
else
printOpField(os, " p=", play->GetPar(arg[1]), ncol);
os << "before=\"" << play->GetTxt(arg[2]) << "\"";
if( arg[0] & 2 )
printOpField(os, " v=", arg[3], ncol);
else
printOpField(os, " p=", play->GetPar(arg[3]), ncol);
os << "after=\"" << play->GetTxt(arg[4]) << "\"";
break;
case BeginOp:
// argument not used (created by independent)
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 1 );
break;
case EndOp:
case InvOp:
case FunrvOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 0 );
break;
case DisOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 2 );
{ const char* name = discrete<Base>::name(arg[0]);
printOpField(os, " f=", name, ncol);
printOpField(os, " x=", arg[1], ncol);
}
break;
case CExpOp:
CPPAD_ASSERT_UNKNOWN(arg[1] != 0);
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 6 );
if( arg[1] & 1 )
printOpField(os, " vl=", arg[2], ncol);
else
printOpField(os, " pl=", play->GetPar(arg[2]), ncol);
if( arg[1] & 2 )
printOpField(os, " vr=", arg[3], ncol);
else
printOpField(os, " pr=", play->GetPar(arg[3]), ncol);
if( arg[1] & 4 )
printOpField(os, " vt=", arg[4], ncol);
else
printOpField(os, " pt=", play->GetPar(arg[4]), ncol);
if( arg[1] & 8 )
printOpField(os, " vf=", arg[5], ncol);
else
printOpField(os, " pf=", play->GetPar(arg[5]), ncol);
break;
case EqppOp:
case LeppOp:
case LtppOp:
case NeppOp:
CPPAD_ASSERT_UNKNOWN( NumArg(op) == 2 );
printOpField(os, " pl=", play->GetPar(arg[0]), ncol);
printOpField(os, " pr=", play->GetPar(arg[1]), ncol);
break;
default:
CPPAD_ASSERT_UNKNOWN(0);
}
}
