X-Git-Url: https://git.auder.net/?p=morpheus.git;a=blobdiff_plain;f=pkg%2Fsrc%2Ffunctions.c;h=6444b8649875ef2022fb5151aa77207743c48740;hp=feea3ade92def6f860e0a88f12f7ac0c802f8b54;hb=9fdd3e5f57aa1fd0f839214d9cbfac022a8739a8;hpb=074c721aa587d45c4f7ef10f8a28fdc09273d31a

diff --git a/pkg/src/functions.c b/pkg/src/functions.c
index feea3ad..6444b86 100644
--- a/pkg/src/functions.c
+++ b/pkg/src/functions.c
@@ -1,16 +1,11 @@
 #include <stdlib.h>
+#include <omp.h>
 
 // Index matrix (by columns)
-int mi(int i, int j, int d1, int d2)
-{
-  return j*d1 + i;
-}
+#define mi(i, j, d1, d2) (j*d1 + i)
 
 // Index 3-tensor (by columns, matrices ordered by last dim)
-int ti(int i, int j, int k, int d1, int d2, int d3)
-{
-  return k*d1*d2 + j*d1 + i;
-}
+#define ti(i, j, k, d1, d2, d3) (k*d1*d2 + j*d1 + i)
 
 // Empirical cross-moment of order 2 between X size nxd and Y size n
 void Moments_M2(double* X, double* Y, int* pn, int* pd, double* M2)
@@ -54,65 +49,11 @@ void Moments_M3(double* X, double* Y, int* pn, int* pd, double* M3)
   }
 }
 
-#include <stdio.h>
-
 // W = 1/N sum( t(g(Zi,theta)) g(Zi,theta) )
 // with g(Zi, theta) = i-th contribution to all moments (size dim) - real moments
-//void Compute_Omega(double* X, int* Y, double* M, int* pn, int* pd, double* W)
-//{
-//  int n=*pn, d=*pd;
-//  int dim = d + d*d + d*d*d;
-//  //double* W = (double*)malloc(dim*dim*sizeof(double));
-//
-//  // (Re)Initialize W:
-//  for (int j=0; j<dim; j++)
-//  {
-//    for (int k=0; k<dim; k++)
-//      W[j*dim+k] = 0.0;
-//  }
-//  double* g = (double*)malloc(dim*sizeof(double));
-//  for (int i=0; i<n; i++)
-//  {
-//    // g == gi:
-//    for (int j=0; j<d; j++)
-//      g[j] = Y[i] * X[mi(i,j,n,d)] - M[j];
-//    for (int j=d; j<d+(d*d); j++)
-//    {
-//      int idx1 = (j-d) % d; //num row
-//      int idx2 = ((j-d) - idx1) / d; //num col
-//      g[j] = 0.0;
-//      if (idx1 == idx2)
-//        g[j] -= Y[i];
-//      g[j] += Y[i] * X[mi(i,idx1,n,d)]*X[mi(i,idx2,n,d)] - M[j];
-//    }
-//    for (int j=d+d*d; j<dim; j++)
-//    {
-//      int idx1 = (j-d-d*d) % d; //num row
-//      int idx2 = ((j-d-d*d - idx1) / d) %d; //num col
-//      int idx3 = (((j-d-d*d - idx1) / d) - idx2) / d; //num "depth"
-//      g[j] = 0.0;
-//      if (idx1 == idx2)
-//        g[j] -= Y[i] * X[mi(i,idx3,n,d)];
-//      if (idx1 == idx3)
-//        g[j] -= Y[i] * X[mi(i,idx2,n,d)];
-//      if (idx2 == idx3)
-//        g[j] -= Y[i] * X[mi(i,idx1,n,d)];
-//      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=0; j<dim; j++)
-//    {
-//      for (int k=0; k<dim; k++)
-//        W[j*dim+k] += g[j] * g[k] / n;
-//    }
-//  }
-//  free(g);
-//}
-
-// Optimisation attempt:
-void Compute_Omega(double* X, int* Y, double* M, int* pn, int* pd, double* W)
+void Compute_Omega(double* X, int* Y, double* M, int* pnc, int* pn, int* pd, double* W)
 {
-  int n=*pn, d=*pd;
+  int nc=*pnc, n=*pn, d=*pd;
   int dim = d + d*d + d*d*d;
   //double* W = (double*)malloc(dim*dim*sizeof(double));
 
@@ -123,23 +64,21 @@ void Compute_Omega(double* X, int* Y, double* M, int* pn, int* pd, double* W)
       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
   for (int i=0; i<n; i++)
   {
-    printf("i: %i\n",i);
     // g == gi:
     for (int j=0; j<d; j++)
-      g[j] = (Y[i] ? X[mi(i,j,n,d)] - M[j] : 0.0);
+      g[j] = Y[i] * X[mi(i,j,n,d)] - M[j];
     for (int j=d; j<d+(d*d); j++)
     {
       int idx1 = (j-d) % d; //num row
       int idx2 = ((j-d) - idx1) / d; //num col
       g[j] = 0.0;
-      if (Y[i])
-      {
-        if (idx1 == idx2)
-          g[j]--;
-        g[j] += X[mi(i,idx1,n,d)]*X[mi(i,idx2,n,d)] - M[j];
-      }
+      if (idx1 == idx2)
+        g[j] -= Y[i];
+      g[j] += Y[i] * X[mi(i,idx1,n,d)]*X[mi(i,idx2,n,d)] - M[j];
     }
     for (int j=d+d*d; j<dim; j++)
     {
@@ -147,23 +86,36 @@ void Compute_Omega(double* X, int* Y, double* M, int* pn, int* pd, double* W)
       int idx2 = ((j-d-d*d - idx1) / d) %d; //num col
       int idx3 = (((j-d-d*d - idx1) / d) - idx2) / d; //num "depth"
       g[j] = 0.0;
-      if (Y[i])
-      {
-        if (idx1 == idx2)
-          g[j] -= X[mi(i,idx3,n,d)];
-        if (idx1 == idx3)
-          g[j] -= X[mi(i,idx2,n,d)];
-        if (idx2 == idx3)
-          g[j] -= X[mi(i,idx1,n,d)];
-        g[j] += X[mi(i,idx1,n,d)]*X[mi(i,idx2,n,d)]*X[mi(i,idx3,n,d)] - M[j];
-      }
+      if (idx1 == idx2)
+        g[j] -= Y[i] * X[mi(i,idx3,n,d)];
+      if (idx1 == idx3)
+        g[j] -= Y[i] * X[mi(i,idx2,n,d)];
+      if (idx2 == idx3)
+        g[j] -= Y[i] * X[mi(i,idx1,n,d)];
+      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=0; j<dim; j++)
     {
-      for (int k=0; k<dim; k++)
-        W[j*dim+k] += g[j] * g[k] / n;
+      // This final nested loop is very costly. Some basic optimisations:
+      double gj = g[j];
+      int baseIdx = j * dim;
+      #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<=j; k++)
+      W[mi(j,k,dim,dim)] /= n;
+  }
+  // Symmetrize W: W[j,k] = W[k,j] for k > j
+  for (int j=0; j<dim; j++)
+  {
+    for (int k=j+1: k<dim; k++)
+      W[mi(j,k,dim,dim)] = W[mi(k,j,dim,dim)];
+  }
   free(g);
 }