You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 
 

369 lines
11 KiB

//==============================================================================
//
// Copyright (c) 2002-
// Authors:
// * Dave Parker <david.parker@comlab.ox.ac.uk> (University of Oxford, formerly University of Birmingham)
//
//------------------------------------------------------------------------------
//
// This file is part of PRISM.
//
// PRISM is free software; you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation; either version 2 of the License, or
// (at your option) any later version.
//
// PRISM is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with PRISM; if not, write to the Free Software Foundation,
// Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
//
//==============================================================================
// includes
#include "PrismSparse.h"
#include <math.h>
#include <util.h>
#include <cudd.h>
#include <dd.h>
#include <odd.h>
#include <dv.h>
#include <prism.h>
#include "sparse.h"
#include "PrismSparseGlob.h"
#include "jnipointer.h"
#include "Measures.h"
#include <new>
//------------------------------------------------------------------------------
JNIEXPORT jlong __jlongpointer JNICALL Java_sparse_PrismSparse_PS_1StochCumulReward
(
JNIEnv *env,
jclass cls,
jlong __jlongpointer tr, // trans matrix
jlong __jlongpointer sr, // state rewards
jlong __jlongpointer trr, // transition rewards
jlong __jlongpointer od, // odd
jlong __jlongpointer rv, // row vars
jint num_rvars,
jlong __jlongpointer cv, // col vars
jint num_cvars,
jdouble time // time bound
)
{
// cast function parameters
DdNode *trans = jlong_to_DdNode(tr); // trans matrix
DdNode *state_rewards = jlong_to_DdNode(sr); // state rewards
DdNode *trans_rewards = jlong_to_DdNode(trr); // transition rewards
ODDNode *odd = jlong_to_ODDNode(od); // odd
DdNode **rvars = jlong_to_DdNode_array(rv); // row vars
DdNode **cvars = jlong_to_DdNode_array(cv); // col vars
// mtbdds
DdNode *tmp = NULL;
// model stats
int n;
long nnz;
// flags
bool compact_tr, compact_d;
// sparse matrix
RMSparseMatrix *rmsm = NULL;
CMSRSparseMatrix *cmsrsm = NULL;
// vectors
double *diags = NULL, *soln = NULL, *soln2 = NULL, *tmpsoln = NULL, *sum = NULL;
DistVector *diags_dist = NULL;
// fox glynn stuff
FoxGlynnWeights fgw;
// timing stuff
long start1, start2, start3, stop;
double time_taken, time_for_setup, time_for_iters;
// misc
bool done;
int j, l, h;
long i, iters, num_iters;
double d, max_diag, weight, kb, kbt, unif, term_crit_param_unif;
// measure for convergence termination check
MeasureSupNorm measure(term_crit == TERM_CRIT_RELATIVE);
// exception handling around whole function
try {
// start clocks
start1 = start2 = util_cpu_time();
// get number of states
n = odd->eoff + odd->toff;
// build sparse matrix
PS_PrintToMainLog(env, "\nBuilding sparse matrix... ");
// if requested, try and build a "compact" version
compact_tr = true;
cmsrsm = NULL;
if (compact) cmsrsm = build_cmsr_sparse_matrix(ddman, trans, rvars, cvars, num_rvars, odd);
if (cmsrsm != NULL) {
nnz = cmsrsm->nnz;
kb = cmsrsm->mem;
}
// if not or if it wasn't possible, built a normal one
else {
compact_tr = false;
rmsm = build_rm_sparse_matrix(ddman, trans, rvars, cvars, num_rvars, odd);
nnz = rmsm->nnz;
kb = rmsm->mem;
}
kbt = kb;
// print some info
PS_PrintToMainLog(env, "[n=%d, nnz=%d%s] ", n, nnz, compact_tr?", compact":"");
PS_PrintMemoryToMainLog(env, "[", kb, "]\n");
// get vector of diagonals
PS_PrintToMainLog(env, "Creating vector for diagonals... ");
diags = compact_tr ? cmsr_negative_row_sums(cmsrsm) : rm_negative_row_sums(rmsm);
// try and convert to compact form if required
compact_d = false;
if (compact) {
if ((diags_dist = double_vector_to_dist(diags, n))) {
compact_d = true;
delete diags; diags = NULL;
}
}
kb = (!compact_d) ? n*8.0/1024.0 : (diags_dist->num_dist*8.0+n*2.0)/1024.0;
kbt += kb;
if (compact_d) PS_PrintToMainLog(env, "[dist=%d, compact] ", diags_dist->num_dist);
PS_PrintMemoryToMainLog(env, "[", kb, "]\n");
// find max diagonal element
if (!compact_d) {
max_diag = diags[0];
for (i = 1; i < n; i++) if (diags[i] < max_diag) max_diag = diags[i];
} else {
max_diag = diags_dist->dist[0];
for (i = 1; i < diags_dist->num_dist; i++) if (diags_dist->dist[i] < max_diag) max_diag = diags_dist->dist[i];
}
max_diag = -max_diag;
// constant for uniformization
unif = 1.02*max_diag;
// modify diagonals
if (!compact_d) {
for (i = 0; i < n; i++) diags[i] = diags[i] / unif + 1;
} else {
for (i = 0; i < diags_dist->num_dist; i++) diags_dist->dist[i] = diags_dist->dist[i] / unif + 1;
}
// uniformization
if (!compact_tr) {
for (i = 0; i < nnz; i++) rmsm->non_zeros[i] /= unif;
} else {
for (i = 0; i < cmsrsm->dist_num; i++) cmsrsm->dist[i] /= unif;
}
// combine state/transition rewards into a single vector
// new state rewards = c + (R.C)1
// first, multiply transition rates by transition rewards and sum rows
// = (R.C)1
Cudd_Ref(trans);
Cudd_Ref(trans_rewards);
tmp = DD_Apply(ddman, APPLY_TIMES, trans, trans_rewards);
tmp = DD_SumAbstract(ddman, tmp, cvars, num_cvars);
// then add state rewards
// = c + (R.C)1
Cudd_Ref(state_rewards);
tmp = DD_Apply(ddman, APPLY_PLUS, tmp, state_rewards);
soln = mtbdd_to_double_vector(ddman, tmp, rvars, num_rvars, odd);
Cudd_RecursiveDeref(ddman, tmp);
// create solution/iteration vectors
PS_PrintToMainLog(env, "Allocating iteration vectors... ");
// soln has already been created and initialised to rewards vector as required
// need to create soln2 and sum
soln2 = new double[n];
sum = new double[n];
kb = n*8.0/1024.0;
kbt += 3*kb;
PS_PrintMemoryToMainLog(env, "[3 x ", kb, "]\n");
// print total memory usage
PS_PrintMemoryToMainLog(env, "TOTAL: [", kbt, "]\n");
// compute new termination criterion parameter (epsilon/8)
term_crit_param_unif = term_crit_param / 8.0;
// compute poisson probabilities (fox/glynn)
PS_PrintToMainLog(env, "\nUniformisation: q.t = %f x %f = %f\n", unif, time, unif * time);
fgw = fox_glynn(unif * time, 1.0e-300, 1.0e+300, term_crit_param_unif);
if (fgw.right < 0) throw "Overflow in Fox-Glynn computation (time bound too big?)";
for (i = fgw.left; i <= fgw.right; i++) {
fgw.weights[i-fgw.left] /= fgw.total_weight;
}
PS_PrintToMainLog(env, "Fox-Glynn: left = %ld, right = %ld\n", fgw.left, fgw.right);
// modify the poisson probabilities to what we need for this computation
// first make the kth value equal to the sum of the values for 0...k
for (i = fgw.left+1; i <= fgw.right; i++) {
fgw.weights[i-fgw.left] += fgw.weights[i-1-fgw.left];
}
// then subtract from 1 and divide by uniformisation constant (q) to give mixed poisson probabilities
for (i = fgw.left; i <= fgw.right; i++) {
fgw.weights[i-fgw.left] = (1 - fgw.weights[i-fgw.left]) / unif;
}
// set up vectors
for (i = 0; i < n; i++) {
sum[i] = 0.0;
}
// get setup time
stop = util_cpu_time();
time_for_setup = (double)(stop - start2)/1000;
start2 = stop;
start3 = stop;
// start transient analysis
done = false;
num_iters = -1;
PS_PrintToMainLog(env, "\nStarting iterations...\n");
// do 0th element of summation (doesn't require any matrix powers)
if (fgw.left == 0) {
for (i = 0; i < n; i++) sum[i] += fgw.weights[0] * soln[i];
} else {
for (i = 0; i < n; i++) sum[i] += soln[i] / unif;
}
// note that we ignore max_iters as we know how any iterations _should_ be performed
for (iters = 1; (iters <= fgw.right) && !done; iters++) {
// store local copies of stuff
double *non_zeros;
unsigned char *row_counts;
int *row_starts;
bool use_counts;
unsigned int *cols;
double *dist;
int dist_shift;
int dist_mask;
if (!compact_tr) {
non_zeros = rmsm->non_zeros;
row_counts = rmsm->row_counts;
row_starts = (int *)rmsm->row_counts;
use_counts = rmsm->use_counts;
cols = rmsm->cols;
} else {
row_counts = cmsrsm->row_counts;
row_starts = (int *)cmsrsm->row_counts;
use_counts = cmsrsm->use_counts;
cols = cmsrsm->cols;
dist = cmsrsm->dist;
dist_shift = cmsrsm->dist_shift;
dist_mask = cmsrsm->dist_mask;
}
// do matrix vector multiply bit
h = 0;
for (i = 0; i < n; i++) {
d = (!compact_d) ? (diags[i] * soln[i]) : (diags_dist->dist[diags_dist->ptrs[i]] * soln[i]);
if (!use_counts) { l = row_starts[i]; h = row_starts[i+1]; }
else { l = h; h += row_counts[i]; }
// "row major" version
if (!compact_tr) {
for (j = l; j < h; j++) {
d += non_zeros[j] * soln[cols[j]];
}
// "compact msr" version
} else {
for (j = l; j < h; j++) {
d += dist[(int)(cols[j] & dist_mask)] * soln[(int)(cols[j] >> dist_shift)];
}
}
// set vector element
soln2[i] = d;
}
// check for steady state convergence
if (do_ss_detect) {
measure.reset();
measure.measure(soln, soln2, n);
if (measure.value() < term_crit_param_unif) {
done = true;
}
}
// special case when finished early (steady-state detected)
if (done) {
// work out sum of remaining poisson probabilities
if (iters <= fgw.left) {
weight = time - iters/unif;
} else {
weight = 0.0;
for (i = iters; i <= fgw.right; i++) {
weight += fgw.weights[i-fgw.left];
}
}
// add to sum
for (i = 0; i < n; i++) sum[i] += weight * soln2[i];
PS_PrintToMainLog(env, "\nSteady state detected at iteration %ld\n", iters);
num_iters = iters;
break;
}
// print occasional status update
if ((util_cpu_time() - start3) > UPDATE_DELAY) {
PS_PrintToMainLog(env, "Iteration %d (of %d): ", iters, fgw.right);
if (do_ss_detect) PS_PrintToMainLog(env, "max %sdiff=%f, ", measure.isRelative()?"relative ":"", measure.value());
PS_PrintToMainLog(env, "%.2f sec so far\n", ((double)(util_cpu_time() - start2)/1000));
start3 = util_cpu_time();
}
// prepare for next iteration
tmpsoln = soln;
soln = soln2;
soln2 = tmpsoln;
// add to sum
if (iters < fgw.left) {
for (i = 0; i < n; i++) sum[i] += soln[i] / unif;
} else {
for (i = 0; i < n; i++) sum[i] += fgw.weights[iters-fgw.left] * soln[i];
}
}
// stop clocks
stop = util_cpu_time();
time_for_iters = (double)(stop - start2)/1000;
time_taken = (double)(stop - start1)/1000;
// print iters/timing info
if (num_iters == -1) num_iters = fgw.right;
PS_PrintToMainLog(env, "\nIterative method: %ld iterations in %.2f seconds (average %.6f, setup %.2f)\n", num_iters, time_taken, time_for_iters/num_iters, time_for_setup);
// catch exceptions: register error, free memory
} catch (std::bad_alloc e) {
PS_SetErrorMessage("Out of memory");
if (sum) delete[] sum;
sum = 0;
} catch (const char *err) {
PS_SetErrorMessage(err);
if (sum) delete sum;
sum = 0;
}
// free memory
if (rmsm) delete rmsm;
if (cmsrsm) delete cmsrsm;
if (diags) delete[] diags;
if (diags_dist) delete diags_dist;
if (soln) delete[] soln;
if (soln2) delete[] soln2;
return ptr_to_jlong(sum);
}
//------------------------------------------------------------------------------