Made redistribution kernel use 64-bit counts/offsets and implemented …

…a custom all-to-allv to do this
solomonik · Apr 3, 2014 · a754cbf · a754cbf
1 parent cf539cd
commit a754cbf
Show file tree

Hide file tree

Showing 5 changed files with 228 additions and 136 deletions.
diff --git a/examples/gemm.cxx b/examples/gemm.cxx
@@ -60,7 +60,7 @@ int  gemm(int const     m,
   free(indices);
 
   C["ij"] += A["ik"]*B["kj"];
-  C["ij"] += (.3*i)*A["ik"]*B["kj"];
+  C["ij"] += .3*A["ik"]*B["kj"];
 #ifndef TEST_SUITE
   double t;
 

diff --git a/src/dist_tensor/dist_tensor_op.cxx b/src/dist_tensor/dist_tensor_op.cxx
@@ -2475,19 +2475,24 @@ int dist_tensor<dtype>::
     ntsr_C->is_data_aliased = 1;
     del_tsr(ntype.tid_C);
   }
-  if (was_home_A && !ntsr_A->is_home){
-    ntsr_A->has_home = 0;
-    del_tsr(ntype.tid_A);
-  } else if (was_home_A) {
-    ntsr_A->is_data_aliased = 1;
-    del_tsr(ntype.tid_A);
-  }
-  if (was_home_B && stype->tid_A != stype->tid_B && !ntsr_B->is_home){
-    ntsr_B->has_home = 0;
-    del_tsr(ntype.tid_B);
-  } else if (was_home_B && stype->tid_A != stype->tid_B) {
-    ntsr_B->is_data_aliased = 1;
-    del_tsr(ntype.tid_B);
+  if (ntype.tid_A != ntype.tid_C){
+    if (was_home_A && !ntsr_A->is_home){
+      ntsr_A->has_home = 0;
+      del_tsr(ntype.tid_A);
+    } else if (was_home_A) {
+      ntsr_A->is_data_aliased = 1;
+      del_tsr(ntype.tid_A);
+    }
+  }
+  if (ntype.tid_B != ntype.tid_A &&
+      ntype.tid_B != ntype.tid_C){
+    if (was_home_B && stype->tid_A != stype->tid_B && !ntsr_B->is_home){
+      ntsr_B->has_home = 0;
+      del_tsr(ntype.tid_B);
+    } else if (was_home_B && stype->tid_A != stype->tid_B) {
+      ntsr_B->is_data_aliased = 1;
+      del_tsr(ntype.tid_B);
+    }
   }
   return DIST_TENSOR_SUCCESS;
 #endif