X-Git-Url: https://git.auder.net/images/pieces/%22%20%20VariantRules.getPpath%28board%5Bi%5D%5Bj%5D%29%20%20%20%22.svg?a=blobdiff_plain;f=pkg%2Fsrc%2Ffunctions.c;h=89e84ef7518a7e3e686e098dabe4706af8175b27;hb=f47183deb252f12d532af81a491083f1411cacc8;hp=94ea12dd5e52036f45b6792160944d851329a874;hpb=5af71d43f3f2dba21c6667939fcff88923af3b7b;p=morpheus.git

diff --git a/pkg/src/functions.c b/pkg/src/functions.c
index 94ea12d..89e84ef 100644
--- a/pkg/src/functions.c
+++ b/pkg/src/functions.c
@@ -1,5 +1,5 @@
 #include <stdlib.h>
-#include <omp.h>
+//#include <omp.h>
 
 // Index matrix (by columns)
 #define mi(i, j, d1, d2) (j*d1 + i)
@@ -53,7 +53,7 @@ void Moments_M3(double* X, double* Y, int* pn, int* pd, double* M3)
 // with g(Zi, theta) = i-th contribution to all moments (size dim) - real moments
 void Compute_Omega(double* X, int* Y, double* M, int* pnc, int* pn, int* pd, double* W)
 {
-  int nc=*pnc, n=*pn, d=*pd;
+  int n=*pn, d=*pd; //,nc=*pnc
   int dim = d + d*d + d*d*d;
   //double* W = (double*)malloc(dim*dim*sizeof(double));
 
@@ -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);
 }