X-Git-Url: https://git.auder.net/?p=morpheus.git;a=blobdiff_plain;f=pkg%2Fsrc%2Ffunctions.c;h=79deff46c702b718b4b9bebeb8d24720d98a3bc2;hp=94ea12dd5e52036f45b6792160944d851329a874;hb=a27d53c381a88e0f8c8788a772575e35464294ef;hpb=5af71d43f3f2dba21c6667939fcff88923af3b7b

diff --git a/pkg/src/functions.c b/pkg/src/functions.c
index 94ea12d..79deff4 100644
--- a/pkg/src/functions.c
+++ b/pkg/src/functions.c
@@ -64,8 +64,9 @@ void Compute_Omega(double* X, int* Y, double* M, int* pnc, int* pn, int* pd, dou
       W[j*dim+k] = 0.0;
   }
   double* g = (double*)malloc(dim*sizeof(double));
-  omp_set_num_threads(nc >= 1 ? nc : omp_get_num_procs());
-  #pragma omp parallel for
+  // TODO: stabilize this (for now, random result)
+//  omp_set_num_threads(nc >= 1 ? nc : omp_get_num_procs());
+//  #pragma omp parallel for
   for (int i=0; i<n; i++)
   {
     // g == gi:
@@ -95,21 +96,27 @@ void Compute_Omega(double* X, int* Y, double* M, int* pnc, int* pn, int* pd, dou
       g[j] += Y[i] * X[mi(i,idx1,n,d)]*X[mi(i,idx2,n,d)]*X[mi(i,idx3,n,d)] - M[j];
     }
     // Add 1/n t(gi) %*% gi to W
-    for (int j=dim-1; j>=0; j--)
+    for (int j=0; j<dim; j++)
     {
       // This final nested loop is very costly. Some basic optimisations:
       double gj = g[j];
       int baseIdx = j * dim;
-      #pragma GCC unroll 100
-      for (int k=dim-1; k>=0; k--)
+      #pragma GCC unroll 32
+      for (int k=j; k>=0; k--)
         W[baseIdx+k] += gj * g[k];
     }
   }
   // Normalize W: x 1/n
   for (int j=0; j<dim; j++)
   {
-    for (int k=0; k<dim; k++)
+    for (int k=j; k<dim; k++)
       W[mi(j,k,dim,dim)] /= n;
   }
+  // Symmetrize W: W[k,j] = W[j,k] for k > j
+  for (int j=0; j<dim; j++)
+  {
+    for (int k=j+1; k<dim; k++)
+      W[mi(k,j,dim,dim)] = W[mi(j,k,dim,dim)];
+  }
   free(g);
 }