greedy cpu without degree computation: results in O(E) runtime :(

aten/TH/THGreedy.c

@@ -4,36 +4,36 @@
 
 #define THGreedy_ TH_CONCAT_3(TH,Real,Greedy)
 
-#define TH_GREEDY_CLUSTER(cluster, row, col, deg, SELECT) { \
+#define TH_GREEDY_CLUSTER(cluster, row, col, SELECT) { \
   THLongTensor_fill(cluster, -1); \
   int64_t *clusterData = THTensor_getData(cluster); \
   int64_t *rowData = THTensor_getData(row); \
   int64_t *colData = THTensor_getData(col); \
-  int64_t *degData = THTensor_getData(deg); \
-  ptrdiff_t rowIdx = 0, neighborIdx; \
-  int64_t rowValue, colValue, clusterValue, tmp; \
-  while(rowIdx < THLongTensor_nElement(row)) { \
-    rowValue = rowData[rowIdx]; \
+  ptrdiff_t idx = 0, nEdges = THLongTensor_nElement(row); \
+  int64_t rowValue, pairValue, colValue, clusterValue; \
+  while(idx < nEdges) { \
+    rowValue = rowData[idx]; \
+    pairValue = rowValue; \
     if (clusterData[rowValue] < 0) { \
-      colValue = rowValue; \
-      SELECT \
-      clusterValue = rowValue < colValue ? rowValue : colValue; \
-      clusterData[rowValue] = clusterValue; \
-      clusterData[colValue] = clusterValue; \
+      while(idx < nEdges && rowData[idx] == rowValue) { \
+        colValue = colData[idx]; \
+        if (clusterData[colValue] < 0) { \
+          SELECT \
+        } \
+        idx++; \
+      } \
     } \
-    rowIdx += degData[rowValue]; \
+    clusterValue = rowValue < pairValue ? rowValue : pairValue; \
+    clusterData[rowValue] = clusterValue; \
+    clusterData[pairValue] = clusterValue; \
+    while(idx < nEdges && rowData[idx] == rowValue) idx++; \
   } \
 }
 
-void THGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THLongTensor *deg) {
-  TH_GREEDY_CLUSTER(cluster, row, col, deg,
-    for (neighborIdx = rowIdx; neighborIdx < rowIdx + degData[rowValue]; neighborIdx++) {
-      tmp = colData[neighborIdx];
-      if (clusterData[tmp] < 0) {
-        colValue = tmp;
-        break;
-      }
-    }
+void THGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col) {
+  TH_GREEDY_CLUSTER(cluster, row, col,
+    pairValue = colValue;
+    break;
   )
 }
 

aten/TH/THGreedy.h

-void THGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THLongTensor *deg);
+void THGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col);
 
-void   THByteGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THLongTensor *deg,   THByteTensor *weight);
-void   THCharGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THLongTensor *deg,   THCharTensor *weight);
-void  THShortGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THLongTensor *deg,  THShortTensor *weight);
-void    THIntGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THLongTensor *deg,    THIntTensor *weight);
-void   THLongGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THLongTensor *deg,   THLongTensor *weight);
-void  THFloatGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THLongTensor *deg,  THFloatTensor *weight);
-void THDoubleGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THLongTensor *deg, THDoubleTensor *weight);
+void   THByteGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col,   THByteTensor *weight);
+void   THCharGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col,   THCharTensor *weight);
+void  THShortGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col,  THShortTensor *weight);
+void    THIntGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col,    THIntTensor *weight);
+void   THLongGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col,   THLongTensor *weight);
+void  THFloatGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col,  THFloatTensor *weight);
+void THDoubleGreedy(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THDoubleTensor *weight);

aten/TH/generic/THGreedy.c

@@ -2,18 +2,14 @@
 #define TH_GENERIC_FILE "generic/THGreedy.c"
 #else
 
-void THGreedy_(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THLongTensor *deg,
-               THTensor *weight) {
+void THGreedy_(THLongTensor *cluster, THLongTensor *row, THLongTensor *col, THTensor *weight) {
   real *weightData = THTensor_getData(weight);
   real maxWeight = 0, tmpWeight;
-  TH_GREEDY_CLUSTER(cluster, row, col, deg,
-    for (neighborIdx = rowIdx; neighborIdx < rowIdx + degData[rowValue]; neighborIdx++) {
-      tmp = colData[neighborIdx];
-      tmpWeight = weightData[neighborIdx];
-      if (clusterData[tmp] < 0 && tmpWeight > maxWeight) {
-        colValue = tmp;
-        maxWeight = tmpWeight;
-      }
+  TH_GREEDY_CLUSTER(cluster, row, col,
+    tmpWeight = weightData[idx];
+    if (tmpWeight > maxWeight) {
+      pairValue = colValue;
+      maxWeight = tmpWeight;
     }
   )
 }