Merge remote-tracking branch 'origin/fix_desc_leak' into model_training

Author: solomonik

src/interface/matrix.cxx

@@ -1,6 +1,7 @@
 /*Copyright (c) 2011, Edgar Solomonik, all rights reserved.*/
 
 #include "common.h"
+#include "world.h"
 #include "../shared/blas_symbs.h"
 #include "../shared/lapack_symbs.h"
 #include <stdlib.h>
@@ -236,7 +237,7 @@ namespace CTF {
         }
       }
       this->write(nmyr*nmyc, pairs);
-      CTF_int::cdealloc(pairs);
+      delete [] pairs;
     }
   }
 
@@ -296,7 +297,17 @@ namespace CTF {
     int ctxt;
     IASSERT(this->wrld->comm == MPI_COMM_WORLD);
     CTF_SCALAPACK::Cblacs_get(-1, 0, &ctxt);
-    CTF_SCALAPACK::Cblacs_gridinit(&ctxt, &C, pr, pc);
+    CTF_int::grid_wrapper gw;
+    gw.pr = pr;
+    gw.pc = pc;
+    std::set<CTF_int::grid_wrapper>::iterator s = CTF_int::scalapack_grids.find(gw);
+    if (s != CTF_int::scalapack_grids.end()){
+      ctxt = s->ctxt;
+    } else {
+      CTF_SCALAPACK::Cblacs_gridinit(&ctxt, &C, pr, pc);
+      gw.ctxt = ctxt;
+      CTF_int::scalapack_grids.insert(gw);
+    }
     ictxt = ctxt;
 
     desc = (int*)malloc(sizeof(int)*9);
@@ -439,6 +450,8 @@ namespace CTF {
     CTF_SCALAPACK::porgqr<dtype>(m,n,n,dQ,1,1,desca,tau,work,lwork,&info);
     Q = Matrix<dtype>(desca, dQ, (*(this->wrld)));
     free(work);
+    free(tau);
+    free(desca);
     //make upper-tri
     int syns[] = {SY, NS};
     Tensor<dtype> tR(R,syns);

src/interface/partition.cxx

@@ -41,6 +41,32 @@ namespace CTF {
 
   Idx_Partition::Idx_Partition(Partition const & part_, char const * idx_){
     part = part_;
-    idx = idx_;
+    idx = (char*)malloc(part.order*sizeof(char));
+    memcpy(idx, idx_, part.order*sizeof(char));
   }
+
+  Idx_Partition::~Idx_Partition(){
+    if (idx != NULL){
+      free(idx);
+      idx = NULL;
+    }
+  }
+
+  Idx_Partition Idx_Partition::reduce_order() const {
+    int * new_lens = (int*)malloc(part.order*sizeof(int));
+    int new_order = 0;
+    char * new_idx = (char*)malloc(part.order);
+    for (int i=0; i<part.order; i++){
+      if (part.lens[i] != 1){
+        new_lens[new_order] = part.lens[i];
+        new_idx[new_order] = idx[i];
+        new_order++;
+      }
+    }
+    Idx_Partition p = Partition(new_order, new_lens)[new_idx];
+    free(new_idx);
+    free(new_lens);
+    return p;
+  }
+
 }

src/interface/partition.h

@@ -28,9 +28,16 @@ namespace CTF {
   class Idx_Partition {
     public:
       Partition part;
-      char const * idx;
+      char * idx;
       Idx_Partition();
+      ~Idx_Partition();
       Idx_Partition(Partition const & part, char const * idx);
+
+      /**
+       * \brief extracts non-trivial part of partition by ommitting unit dimensions
+       * \return new partition with all dimensions non-unit
+       */
+      Idx_Partition reduce_order() const;
   };
 
 /**

src/interface/world.cxx

@@ -2,6 +2,7 @@
 
 #include "common.h"
 #include "world.h"
+#include "../shared/lapack_symbs.h"
 #include "../tensor/algstrct.h"
 #include "../shared/util.h"
 #include "../shared/memcontrol.h"
@@ -14,6 +15,18 @@ extern "C"
 
 using namespace CTF_int;
 
+namespace CTF_int {
+  bool grid_wrapper::operator<(grid_wrapper const & other) const {
+    if (this->pr == other.pr)
+      return this->pc < other.pc;
+    else
+      return this->pr < other.pr;
+  }
+
+  /** \brief index for ScaLAPACK processor grids */
+  std::set<grid_wrapper> scalapack_grids;
+}
+
 namespace CTF {
   bool universe_exists = false;
   World universe("");
@@ -96,6 +109,10 @@ namespace CTF {
     initialized = 0;
     mem_exit(rank);
     if (get_num_instances() == 0){
+      for (std::set<grid_wrapper>::iterator it=scalapack_grids.begin(); it!=scalapack_grids.end(); it++){
+        //printf("HERE %d %d %d\n",it->pr,it->pc,it->ctxt);
+        CTF_SCALAPACK::Cblacs_gridexit(it->ctxt);
+      }
 #ifdef OFFLOAD
       offload_exit();
 #endif

src/interface/world.h

@@ -2,6 +2,7 @@
 #define __WORLD_H__
 
 #include "common.h"
+#include <set>
 #include "../mapping/topology.h"
 
 namespace CTF {
@@ -144,6 +145,19 @@ namespace CTF {
   /**
    * @}
    */
+
 }
 
+namespace CTF_int {
+  class grid_wrapper {
+    public:
+      int pr;
+      int pc;
+      int ctxt;
+
+      bool operator<(grid_wrapper const & other) const;
+  }; 
+  extern std::set<grid_wrapper> scalapack_grids;
+
+}
 #endif

src/tensor/untyped_tensor.cxx

@@ -1224,8 +1224,9 @@ namespace CTF_int {
   }
 
   void tensor::set_distribution(char const *          idx,
-                                Idx_Partition const & prl,
+                                Idx_Partition const & prl_,
                                 Idx_Partition const & blk){
+    Idx_Partition prl = prl_.reduce_order();
     topology * top = new topology(prl.part.order, prl.part.lens, wrld->cdt);
     int itopo = find_topology(top, wrld->topovec);
 /*    if (wrld->rank == 0){
@@ -1239,7 +1240,7 @@ namespace CTF_int {
     if (itopo == -1){
       itopo = wrld->topovec.size();
       wrld->topovec.push_back(top);
-    }
+    } else delete top;
     ASSERT(itopo != -1);
     assert(itopo != -1);
 
@@ -1284,7 +1285,7 @@ namespace CTF_int {
       ASSERT(0);
       assert(0);
     }
-
+    cdealloc(idx_A);
   }