void multiply(DistSlicedMat & mat, int curoffset, double a){
assert(a>0);
DistVec current = mat[curoffset];
assert(mat.start_offset <= current.offset);
vec result = zeros(curoffset);
if (curoffset > 0){
#pragma omp parallel for
for (int i=mat.start_offset; i< current.offset; i++){
for (int k=info.get_start_node(!current.transpose); k< info.get_end_node(!current.transpose); k++){
result[i-mat.start_offset] += latent_factors_inmem[k].pvec[i] * latent_factors_inmem[k].pvec[current.offset];
}
}
#pragma omp parallel for
for (int k=info.get_start_node(!current.transpose); k< info.get_end_node(!current.transpose); k++){
latent_factors_inmem[k].pvec[curoffset] /= a;
}
for (int i=mat.start_offset; i< current.offset; i++){
#pragma omp parallel for
for (int k=info.get_start_node(!current.transpose); k< info.get_end_node(!current.transpose); k++){
latent_factors_inmem[k].pvec[current.offset] -= result[i-mat.start_offset]/a * latent_factors_inmem[k].pvec[i];
}
}
}
current.debug_print(current.name);
}
void orthogonalize_vs_all(DistSlicedMat & mat, int curoffset, double &alpha){
assert(mi.ortho_repeats >=1 && mi.ortho_repeats <= 3);
bool old_debug = debug;
debug = false;
DistVec current = mat[curoffset];
assert(mat.start_offset <= current.offset);
double * alphas = new double[curoffset];
//DistDouble * alphas = new DistDouble[curoffset];
//cout<<current.to_vec().transpose() << endl;
if (curoffset > 0){
for (int j=0; j < mi.ortho_repeats; j++){
memset(alphas, 0, sizeof(double)*curoffset);
#pragma omp parallel for
for (int i=mat.start_offset; i< current.offset; i++){
for (int k=info.get_start_node(!current.transpose); k< info.get_end_node(!current.transpose); k++){
assert(i-mat.start_offset>=0 && i-mat.start_offset < curoffset);
assert(i < latent_factors_inmem[k].pvec.size());
assert(k < (int)latent_factors_inmem.size());
assert(current.offset < latent_factors_inmem[k].pvec.size());
alphas[i-mat.start_offset] += latent_factors_inmem[k].pvec[i] * latent_factors_inmem[k].pvec[current.offset];
}
}
for (int i=mat.start_offset; i< current.offset; i++){
#pragma omp parallel for
for (int k=info.get_start_node(!current.transpose); k< info.get_end_node(!current.transpose); k++){
latent_factors_inmem[k].pvec[current.offset] -= alphas[i-mat.start_offset] * latent_factors_inmem[k].pvec[i];
}
}
} //for ortho_repeast
}
delete [] alphas;
debug = old_debug;
current.debug_print(current.name);
// alpha = 0;
double sum = 0;
int k;
//#pragma omp parallel for private(k) reduction(+: sum)
for (k=info.get_start_node(!current.transpose); k< info.get_end_node(!current.transpose); k++){
sum = sum + pow(latent_factors_inmem[k].pvec[current.offset],2);
}
alpha = sqrt(sum);
if (alpha >= 1e-10 ){
#pragma omp parallel for
for (int k=info.get_start_node(!current.transpose); k< info.get_end_node(!current.transpose); k++){
latent_factors_inmem[k].pvec[current.offset]/=alpha;
}
}
}
double orthogonalize_vs_all(DistSlicedMat & mat, int _curoffset, double &alpha){
assert(mi.ortho_repeats >=1 && mi.ortho_repeats <= 3);
curoffset = _curoffset;
curMat = &mat;
mi.mat_offset = mat.start_offset;
INITIALIZE_TRACER(orthogonalize_vs_alltrace, "orthogonalization step - optimized");
BEGIN_TRACEPOINT(orthogonalize_vs_alltrace);
bool old_debug = debug;
debug = false;
DistVec current = mat[curoffset];
pcurrent =¤t;
mi.vec_offset = pcurrent->offset;
assert(mat.start_offset <= current.offset);
vertex_set nodes = pgraph->select(selected_node);
if (curoffset > 0){
for (int j=0; j < mi.ortho_repeats; j++){
INITIALIZE_TRACER(orth1, "map reduce in ortho");
BEGIN_TRACEPOINT(orth1);
alphas = pgraph->map_reduce_vertices<gather_type>(map_reduce_ortho, nodes);
END_TRACEPOINT(orth1);
//pgraph->transform_vertices(transform_ortho, nodes);
mat[_curoffset] = mat[_curoffset].orthogonalize();
} //for ortho_repeast
}
debug = old_debug;
current.debug_print(current.name);
INITIALIZE_TRACER(orth2, "map reduce in ortho2");
BEGIN_TRACEPOINT(orth2);
sum_alpha = pgraph->map_reduce_vertices<gather_type>(map_reduce_sum_power, nodes);
END_TRACEPOINT(orth2);
sum_alpha.training_rmse = sqrt(sum_alpha.training_rmse);
alpha = sum_alpha.training_rmse;
if (alpha >= 1e-10 ){
INITIALIZE_TRACER(orth3, "transform_vertices in ortho3");
BEGIN_TRACEPOINT(orth3);
//pgraph->transform_vertices(divide_by_sum, nodes);
mat[_curoffset] = mat[_curoffset] / alpha;
END_TRACEPOINT(orth3);
}
END_TRACEPOINT(orthogonalize_vs_alltrace);
return alpha;
}
void orthogonalize_vs_all(DistSlicedMat & mat, int curoffset){
assert(mi.ortho_repeats >=1 && mi.ortho_repeats <= 3);
INITIALIZE_TRACER(orthogonalize_vs_alltrace, "orthogonalization step");
BEGIN_TRACEPOINT(orthogonalize_vs_alltrace);
bool old_debug = debug;
debug = false;
DistVec current = mat[curoffset];
//DistDouble * alphas = new DistDouble[curoffset];
//cout<<current.to_vec().transpose() << endl;
for (int j=0; j < mi.ortho_repeats; j++){
for (int i=0; i< curoffset; i++){
DistDouble alpha = mat[i]*current;
// //cout<<mat[i].to_vec().transpose()<<endl;
// //cout<<"alpha is: " <<alpha.toDouble()<<endl;
if (alpha.toDouble() > 1e-10)
current = current - mat[i]*alpha;
}
}
END_TRACEPOINT(orthogonalize_vs_alltrace);
debug = old_debug;
current.debug_print(current.name);
}
