// FIXME
void Duration::endRange(const QLoggingCategory& category, uint64_t rangeId) {
#ifdef NSIGHT_TRACING
if (tracingEnabled() && category.isDebugEnabled()) {
nvtxRangeEnd(rangeId);
}
#endif
}
void ProfileRange::endRange(uint64_t rangeId) {
nvtxRangeEnd(rangeId);
// nvtxRangePop();
}
void
cg_solve(OperatorType& A,
const VectorType& b,
VectorType& x,
Matvec matvec,
typename OperatorType::LocalOrdinalType max_iter,
typename TypeTraits<typename OperatorType::ScalarType>::magnitude_type& tolerance,
typename OperatorType::LocalOrdinalType& num_iters,
typename TypeTraits<typename OperatorType::ScalarType>::magnitude_type& normr,
timer_type* my_cg_times)
{
typedef typename OperatorType::ScalarType ScalarType;
typedef typename OperatorType::GlobalOrdinalType GlobalOrdinalType;
typedef typename OperatorType::LocalOrdinalType LocalOrdinalType;
typedef typename TypeTraits<ScalarType>::magnitude_type magnitude_type;
timer_type t0 = 0, tWAXPY = 0, tDOT = 0, tMATVEC = 0, tMATVECDOT = 0;
timer_type total_time = mytimer();
int myproc = 0;
#ifdef HAVE_MPI
MPI_Comm_rank(MPI_COMM_WORLD, &myproc);
#endif
if (!A.has_local_indices) {
std::cerr << "miniFE::cg_solve ERROR, A.has_local_indices is false, needs to be true. This probably means "
<< "miniFE::make_local_matrix(A) was not called prior to calling miniFE::cg_solve."
<< std::endl;
return;
}
size_t nrows = A.rows.size();
LocalOrdinalType ncols = A.num_cols;
nvtxRangeId_t r1=nvtxRangeStartA("Allocation of Temporary Vectors");
VectorType r(b.startIndex, nrows);
VectorType p(0, ncols);
VectorType Ap(b.startIndex, nrows);
nvtxRangeEnd(r1);
#ifdef HAVE_MPI
#ifndef GPUDIRECT
//TODO move outside?
cudaHostRegister(&p.coefs[0],ncols*sizeof(typename VectorType::ScalarType),0);
cudaCheckError();
if(A.send_buffer.size()>0) cudaHostRegister(&A.send_buffer[0],A.send_buffer.size()*sizeof(typename VectorType::ScalarType),0);
cudaCheckError();
#endif
#endif
normr = 0;
magnitude_type rtrans = 0;
magnitude_type oldrtrans = 0;
LocalOrdinalType print_freq = max_iter/10;
if (print_freq>50) print_freq = 50;
if (print_freq<1) print_freq = 1;
ScalarType one = 1.0;
ScalarType zero = 0.0;
TICK(); waxpby(one, x, zero, x, p); TOCK(tWAXPY);
TICK();
matvec(A, p, Ap);
TOCK(tMATVEC);
TICK(); waxpby(one, b, -one, Ap, r); TOCK(tWAXPY);
TICK(); rtrans = dot(r, r); TOCK(tDOT);
normr = std::sqrt(rtrans);
if (myproc == 0) {
std::cout << "Initial Residual = "<< normr << std::endl;
}
magnitude_type brkdown_tol = std::numeric_limits<magnitude_type>::epsilon();
#ifdef MINIFE_DEBUG
std::ostream& os = outstream();
os << "brkdown_tol = " << brkdown_tol << std::endl;
#endif
for(LocalOrdinalType k=1; k <= max_iter && normr > tolerance; ++k) {
if (k == 1) {
TICK(); waxpby(one, r, zero, r, p); TOCK(tWAXPY);
}
else {
oldrtrans = rtrans;
TICK(); rtrans = dot(r, r); TOCK(tDOT);
magnitude_type beta = rtrans/oldrtrans;
TICK(); waxpby(one, r, beta, p, p); TOCK(tWAXPY);
}
normr = std::sqrt(rtrans);
if (myproc == 0 && (k%print_freq==0 || k==max_iter)) {
std::cout << "Iteration = "<<k<<" Residual = "<<normr<<std::endl;
}
magnitude_type alpha = 0;
magnitude_type p_ap_dot = 0;
TICK(); matvec(A, p, Ap); TOCK(tMATVEC);
TICK(); p_ap_dot = dot(Ap, p); TOCK(tDOT);
#ifdef MINIFE_DEBUG
os << "iter " << k << ", p_ap_dot = " << p_ap_dot;
os.flush();
#endif
//TODO remove false below
if (false && p_ap_dot < brkdown_tol) {
if (p_ap_dot < 0 || breakdown(p_ap_dot, Ap, p)) {
std::cerr << "miniFE::cg_solve ERROR, numerical breakdown!"<<std::endl;
#ifdef MINIFE_DEBUG
os << "ERROR, numerical breakdown!"<<std::endl;
#endif
//update the timers before jumping out.
my_cg_times[WAXPY] = tWAXPY;
my_cg_times[DOT] = tDOT;
my_cg_times[MATVEC] = tMATVEC;
my_cg_times[TOTAL] = mytimer() - total_time;
return;
}
else brkdown_tol = 0.1 * p_ap_dot;
}
alpha = rtrans/p_ap_dot;
#ifdef MINIFE_DEBUG
os << ", rtrans = " << rtrans << ", alpha = " << alpha << std::endl;
#endif
TICK(); waxpby(one, x, alpha, p, x);
waxpby(one, r, -alpha, Ap, r); TOCK(tWAXPY);
num_iters = k;
}
#ifdef HAVE_MPI
#ifndef GPUDIRECT
//TODO move outside?
cudaHostUnregister(&p.coefs[0]);
cudaCheckError();
if(A.send_buffer.size()>0) cudaHostUnregister(&A.send_buffer[0]);
cudaCheckError();
#endif
#endif
my_cg_times[WAXPY] = tWAXPY;
my_cg_times[DOT] = tDOT;
my_cg_times[MATVEC] = tMATVEC;
my_cg_times[MATVECDOT] = tMATVECDOT;
my_cg_times[TOTAL] = mytimer() - total_time;
}
