double Forwarder::mr_pthread_forward(const Matrix &pi, const Matrix &A, const Matrix &B, const DeviceDescriptor &device_descriptor) const {
if(pi.get_width() != 1 || pi.get_height() != A.get_width() || A.get_height() != A.get_width() ||
B.get_height() != A.get_width() || B.get_width() != orig_alphabet_size) {
std::cerr << "Dimensions of input matrices do not match:" << std::endl;
std::cerr << "\t" << "pi width: " << pi.get_width() << std::endl;
std::cerr << "\t" << "pi height: " << pi.get_height() << std::endl;
std::cerr << "\t" << "A width: " << A.get_width() << std::endl;
std::cerr << "\t" << "A height: " << A.get_height() << std::endl;
std::cerr << "\t" << "B width: " << B.get_width() << std::endl;
std::cerr << "\t" << "B height: " << B.get_height() << std::endl;
std::exit(-1);
}
double *symbol2scale;
Matrix *symbol2matrix;
const std::vector<std::vector< unsigned> > *sequences;
std::vector<ProcessingDevice*> devices;
size_t numberOfDevices;
double loglikelihood;
// find alphabet and seqs for given number of states
size_t no_states = A.get_width();
size_t alphabet_size = 0;
for(std::map<size_t, std::vector<std::vector<unsigned> > >::const_iterator it = nStates2seqs.begin(); it != nStates2seqs.end(); ++it) {
if(it->first >= no_states) {
sequences = &(it->second);
alphabet_size = nStates2alphabet_size.find(it->first)->second;
break;
}
}
if(sequences == 0) {
sequences = &(nStates2seqs.rbegin()->second);
alphabet_size = nStates2alphabet_size.rbegin()->second;
}
symbol2scale = new double[alphabet_size];
symbol2matrix = new Matrix[alphabet_size];
compute_symbol2scale_and_symbol2matrix(symbol2matrix, symbol2scale, A, B, alphabet_size);
numberOfDevices = device_descriptor.getNDevices();
for(unsigned i = 0; i < numberOfDevices; ++i) {
devices.push_back(device_descriptor.createDevice(i));
devices[i]->setParameters(&pi, &A, &B, &symbol2pair, symbol2scale, symbol2matrix);
devices[i]->setSeqs(sequences);
}
const size_t length = sequences->size();
const size_t nBlocks = size_t(std::sqrt(length));
MapReduceJobControl control(length, nBlocks);
for(unsigned i = 0; i < numberOfDevices; ++i)
devices[i]->mapReduceLoglikelihood(control);
for(unsigned i = 0; i < numberOfDevices; ++i)
devices[i]->join();
loglikelihood = 0.0;
for(size_t i = 0; i < nBlocks; ++i) {
loglikelihood += control.resultLogLikelihoods[i];
}
for(unsigned i = 0; i < devices.size(); ++i)
delete devices[i];
delete[] symbol2scale;
delete[] symbol2matrix;
return loglikelihood;
}
double Forwarder::pthread_forward(const Matrix &pi, const Matrix &A, const Matrix &B, const DeviceDescriptor &device_descriptor) const {
if(pi.get_width() != 1 || pi.get_height() != A.get_width() || A.get_height() != A.get_width() ||
B.get_height() != A.get_width() || B.get_width() != orig_alphabet_size) {
std::cerr << "Dimensions of input matrices do not match:" << std::endl;
std::cerr << "\t" << "pi width: " << pi.get_width() << std::endl;
std::cerr << "\t" << "pi height: " << pi.get_height() << std::endl;
std::cerr << "\t" << "A width: " << A.get_width() << std::endl;
std::cerr << "\t" << "A height: " << A.get_height() << std::endl;
std::cerr << "\t" << "B width: " << B.get_width() << std::endl;
std::cerr << "\t" << "B height: " << B.get_height() << std::endl;
std::exit(-1);
}
double *symbol2scale;
Matrix *symbol2matrix;
std::vector<ProcessingDevice*> devices;
size_t numberOfDevices;
Matrix *result;
Matrix *temp;
double loglikelihood;
size_t head;
// find alphabet and seqs for given number of states
size_t no_states = A.get_width();
size_t alphabet_size = 0;
const std::vector<std::vector< unsigned> > *sequences;
for(std::map<size_t, std::vector<std::vector<unsigned> > >::const_iterator it = nStates2seqs.begin(); it != nStates2seqs.end(); ++it) {
if(it->first >= no_states) {
sequences = &(it->second);
alphabet_size = nStates2alphabet_size.find(it->first)->second;
break;
}
}
if(sequences == 0) {
sequences = &(nStates2seqs.rbegin()->second);
alphabet_size = nStates2alphabet_size.rbegin()->second;
}
symbol2scale = new double[alphabet_size];
symbol2matrix = new Matrix[alphabet_size];
compute_symbol2scale_and_symbol2matrix(symbol2matrix, symbol2scale, A, B, alphabet_size);
result = new Matrix();
temp = new Matrix();
numberOfDevices = device_descriptor.getNDevices();
for(unsigned i = 0; i < numberOfDevices; ++i) {
devices.push_back(device_descriptor.createDevice(i));
devices[i]->setParameters(&pi, &A, &B, &symbol2pair, symbol2scale, symbol2matrix);
}
loglikelihood = 0.0;
for(std::vector<std::vector<unsigned> >::const_iterator it = sequences->begin(); it != sequences->end(); ++it) {
for(unsigned i = 0; i < numberOfDevices; ++i) {
devices[i]->setSeq(&(*it));
}
const size_t length = it->size();
const size_t nBlocks = size_t(std::sqrt(length));
Stage2JobControl control(length, nBlocks);
devices[0]->likelihoodVector(control);
for(unsigned i = 1; i < numberOfDevices; ++i)
devices[i]->likelihoodMatrix(control);
for(unsigned i = 0; i < numberOfDevices; ++i)
devices[i]->join();
head = control.headBlock - 1;
Matrix::copy(*control.resultMatrices[head], *result);
loglikelihood += control.resultLogLikelihoods[head];
for(size_t i = head + 1; i < nBlocks; ++i) {
Matrix::mult(*control.resultMatrices[i], *result, *temp);
std::swap(result, temp);
loglikelihood += LinearSpace::toLogSpace(result->normalize()) + control.resultLogLikelihoods[i];
}
}
for(unsigned i = 0; i < devices.size(); ++i)
delete devices[i];
delete result;
delete temp;
delete[] symbol2scale;
delete[] symbol2matrix;
return loglikelihood;
}
