void traceProperties_NDFSM( NDFSM_TestEngineData* ndfsm_state )
{
char buff[64];
// Deferred Reactions Enabled
traceTestEngineProperty( "Deferred Reactions Enabled", areDeferredReactionsEnabled() ? "true" : "false" );
if (areDeferredReactionsEnabled())
{// WTime
if (getWTimeMSec() % 1000 == 0)
sprintf(buff,"%d secs",getWTime());
else
sprintf(buff,"%d millisecs",getWTimeMSec());
traceTestEngineProperty( "WTime", buff );
// Is Find First Series Only
traceTestEngineProperty( "Find First Series Only", isFindFirstSeriesOnly() ? "true" : "false" );
}
// Time Model
switch (getTSTimeModel())
{case NotUseTSTime :
traceTestEngineProperty( "Time Model", "not used" );
break;
case LinearTSTime :
traceTestEngineProperty( "Time Model", "linear" );
break;
case DistributedTSTime :
traceTestEngineProperty( "Time Model", "distributed" );
break;
default :
assertion( false, "Unexpected Time Model" );
}
// Finish Mode
switch (getFinishMode())
{case UNTIL_ERROR :
if (tsNumberOfErrorsToFinish == 1)
traceTestEngineProperty( "Finish Mode", "until an error" );
else
{
sprintf( buff, "until %d errors", tsNumberOfErrorsToFinish );
traceTestEngineProperty( "Finish Mode", buff );
}
break;
case UNTIL_END :
traceTestEngineProperty( "Finish Mode", "until end" );
break;
default :
assertion( false, "Unexpected Finish Mode" );
}
}
/***************************************
* Conjugate Gradient *
* This function will do the CG *
* algorithm without preconditioning. *
* For optimiziation you must not *
* change the algorithm. *
***************************************
r(0) = b - Ax(0)
p(0) = r(0)
rho(0) = <r(0),r(0)>
***************************************
for k=0,1,2,...,n-1
q(k) = A * p(k)
dot_pq = <p(k),q(k)>
alpha = rho(k) / dot_pq
x(k+1) = x(k) + alpha*p(k)
r(k+1) = r(k) - alpha*q(k)
check convergence ||r(k+1)||_2 < eps
rho(k+1) = <r(k+1), r(k+1)>
beta = rho(k+1) / rho(k)
p(k+1) = r(k+1) + beta*p(k)
***************************************/
void cg(const int n, const int nnz, const int maxNNZ, const floatType* data, const int* indices, const int* length, const floatType* b, floatType* x, struct SolverConfig* sc){
floatType* r, *p, *q;
floatType alpha, beta, rho, rho_old, dot_pq, bnrm2;
int iter;
double timeMatvec_s;
double timeMatvec=0;
int i;
floatType temp;
/* allocate memory */
r = (floatType*)malloc(n * sizeof(floatType));
p = (floatType*)malloc(n * sizeof(floatType));
q = (floatType*)malloc(n * sizeof(floatType));
#pragma acc data copyin(data[0:n*maxNNZ], indices[0:n*maxNNZ], length[0:n], n, nnz, maxNNZ, b[0:n]) copy(x[0:n]) create(alpha, beta, r[0:n], p[0:n], q[0:n], i, temp) //eigentlich auch copy(x[0:n]) aber error: not found on device???
{
DBGMAT("Start matrix A = ", n, nnz, maxNNZ, data, indices, length)
DBGVEC("b = ", b, n);
DBGVEC("x = ", x, n);
/* r(0) = b - Ax(0) */
timeMatvec_s = getWTime();
matvec(n, nnz, maxNNZ, data, indices, length, x, r);
//hier inline ausprobieren
/*int i, j, k;
#pragma acc parallel loop present(data, indices, length, x)
for (i = 0; i < n; i++) {
r[i] = 0;
for (j = 0; j < length[i]; j++) {
k = j * n + i;
r[i] += data[k] * x[indices[k]];
}
}*/
timeMatvec += getWTime() - timeMatvec_s;
xpay(b, -1.0, n, r);
DBGVEC("r = b - Ax = ", r, n);
/* Calculate initial residuum */
nrm2(r, n, &bnrm2);
bnrm2 = 1.0 /bnrm2;
/* p(0) = r(0) */
memcpy(p, r, n*sizeof(floatType));
DBGVEC("p = r = ", p, n);
/* rho(0) = <r(0),r(0)> */
vectorDot(r, r, n, &rho);
printf("rho_0=%e\n", rho);
for(iter = 0; iter < sc->maxIter; iter++){
DBGMSG("=============== Iteration %d ======================\n", iter);
/* q(k) = A * p(k) */
timeMatvec_s = getWTime();
matvec(n, nnz, maxNNZ, data, indices, length, p, q);
timeMatvec += getWTime() - timeMatvec_s;
DBGVEC("q = A * p= ", q, n);
/* dot_pq = <p(k),q(k)> */
vectorDot(p, q, n, &dot_pq);
DBGSCA("dot_pq = <p, q> = ", dot_pq);
/* alpha = rho(k) / dot_pq */
alpha = rho / dot_pq;
DBGSCA("alpha = rho / dot_pq = ", alpha);
/* x(k+1) = x(k) + alpha*p(k) */
axpy(alpha, p, n, x);
#pragma acc update host(x[0:n])
DBGVEC("x = x + alpha * p= ", x, n);
/* r(k+1) = r(k) - alpha*q(k) */
axpy(-alpha, q, n, r);
DBGVEC("r = r - alpha * q= ", r, n);
rho_old = rho;
DBGSCA("rho_old = rho = ", rho_old);
/* rho(k+1) = <r(k+1), r(k+1)> */
vectorDot(r, r, n, &rho);
DBGSCA("rho = <r, r> = ", rho);
/* Normalize the residual with initial one */
sc->residual= sqrt(rho) * bnrm2;
/* Check convergence ||r(k+1)||_2 < eps
* If the residual is smaller than the CG
* tolerance specified in the CG_TOLERANCE
* environment variable our solution vector
* is good enough and we can stop the
* algorithm. */
printf("res_%d=%e\n", iter+1, sc->residual);
if(sc->residual <= sc->tolerance)
break;
/* beta = rho(k+1) / rho(k) */
beta = rho / rho_old;
DBGSCA("beta = rho / rho_old= ", beta);
/* p(k+1) = r(k+1) + beta*p(k) */
xpay(r, beta, n, p);
DBGVEC("p = r + beta * p> = ", p, n);
}
/* Store the number of iterations and the
* time for the sparse matrix vector
* product which is the most expensive
* function in the whole CG algorithm. */
sc->iter = iter;
sc->timeMatvec = timeMatvec;
/* Clean up */
free(r);
free(p);
free(q);
}//ende data region
}