Merge pull request #6952 from jsquyres/pr/v4.0.x/ddt-opt-and-fix

v4.0.x: Datatype optimization and fix

opal/datatype/opal_datatype_optimize.c

@@ -163,6 +163,12 @@ opal_datatype_optimize_short( opal_datatype_t* pData,
             if( 0 == last.count ) {  /* first data of the datatype */
                 last = *current;
                 continue;  /* next data */
+            } else {  /* can we merge it in order to decrease count */
+                if( (ptrdiff_t)last.blocklen * (ptrdiff_t)opal_datatype_basicDatatypes[last.common.type]->size == last.extent ) {
+                    last.extent *= last.count;
+                    last.blocklen *= last.count;
+                    last.count = 1;
+                }
             }
 
             /* are the two elements compatible: aka they have very similar values and they
@@ -176,6 +182,16 @@ opal_datatype_optimize_short( opal_datatype_t* pData,
                     last.common.type  = OPAL_DATATYPE_UINT1;
                 }
 
+                if( (last.extent * (ptrdiff_t)last.count + last.disp) == current->disp ) {
+                    if( 1 == current->count ) {
+                        last.count++;
+                        continue;
+                    }
+                    if( last.extent == current->extent ) {
+                        last.count += current->count;
+                        continue;
+                    }
+                }
                 if( 1 == last.count ) {
                     /* we can ignore the extent of the element with count == 1 and merge them together if their displacements match */
                     if( 1 == current->count ) {
@@ -186,17 +202,7 @@ opal_datatype_optimize_short( opal_datatype_t* pData,
                     /* can we compute a matching displacement ? */
                     if( (last.disp + current->extent) == current->disp ) {
                         last.extent = current->extent;
-                        last.count = current->count + 1;
-                        continue;
-                    }
-                }
-                if( (last.extent * (ptrdiff_t)last.count + last.disp) == current->disp ) {
-                    if( 1 == current->count ) {
-                        last.count++;
-                        continue;
-                    }
-                    if( last.extent == current->extent ) {
-                        last.count += current->count;
+                        last.count = current->count + last.count;
                         continue;
                     }
                 }

opal/datatype/opal_datatype_pack.c

@@ -121,6 +121,7 @@ opal_pack_homogeneous_contig_with_gaps_function( opal_convertor_t* pConv,
      * how much we should jump between elements.
      */
     assert( (pData->flags & OPAL_DATATYPE_FLAG_CONTIGUOUS) && ((ptrdiff_t)pData->size != extent) );
+    assert( pData->opt_desc.used <= 1 );
     DO_DEBUG( opal_output( 0, "pack_homogeneous_contig( pBaseBuf %p, iov_count %d )\n",
                            (void*)pConv->pBaseBuf, *out_size ); );
     if( stack[1].type != opal_datatype_uint1.id ) {

test/datatype/to_self.c

@@ -24,6 +24,26 @@ extern void ompi_datatype_dump( MPI_Datatype ddt );
 #define MPI_DDT_DUMP(ddt)
 #endif  /* OPEN_MPI */
 
+static MPI_Datatype
+create_merged_contig_with_gaps(int count)  /* count of the basic datatype */
+{
+    int array_of_blocklengths[] = {1, 1, 1};
+    MPI_Aint array_of_displacements[] = {0, 8, 16};
+    MPI_Datatype array_of_types[] = {MPI_DOUBLE, MPI_LONG, MPI_CHAR};
+    MPI_Datatype type;
+
+    MPI_Type_create_struct(3, array_of_blocklengths,
+                           array_of_displacements, array_of_types,
+                           &type);
+    if( 1 < count ) {
+        MPI_Datatype temp = type;
+        MPI_Type_contiguous(count, temp, &type);
+    }
+    MPI_Type_commit(&type);
+    MPI_DDT_DUMP( type );
+    return type;
+}
+
 /* Create a non-contiguous resized datatype */
 struct structure {
     double not_transfered;
@@ -183,11 +203,12 @@ create_indexed_gap_optimized_ddt( void )
 /********************************************************************
  *******************************************************************/
 
-#define DO_CONTIG                       0x00000001
-#define DO_CONSTANT_GAP                 0x00000002
-#define DO_INDEXED_GAP                  0x00000004
-#define DO_OPTIMIZED_INDEXED_GAP        0x00000008
-#define DO_STRUCT_CONSTANT_GAP_RESIZED  0x00000010
+#define DO_CONTIG                         0x00000001
+#define DO_CONSTANT_GAP                   0x00000002
+#define DO_INDEXED_GAP                    0x00000004
+#define DO_OPTIMIZED_INDEXED_GAP          0x00000008
+#define DO_STRUCT_CONSTANT_GAP_RESIZED    0x00000010
+#define DO_STRUCT_MERGED_WITH_GAP_RESIZED 0x00000020
 
 #define DO_PACK                         0x01000000
 #define DO_UNPACK                       0x02000000
@@ -483,7 +504,7 @@ static int do_test_for_ddt( int doop, MPI_Datatype sddt, MPI_Datatype rddt, int
 
 int main( int argc, char* argv[] )
 {
-    int run_tests = 0xffff;  /* do all datatype tests by default */
+    int run_tests = DO_STRUCT_MERGED_WITH_GAP_RESIZED;  /* do all datatype tests by default */
     int rank, size;
     MPI_Datatype ddt;
 
@@ -544,6 +565,14 @@ int main( int argc, char* argv[] )
         MPI_Type_free( &ddt );
     }
 
+    if( run_tests & DO_STRUCT_MERGED_WITH_GAP_RESIZED ) {
+        printf( "\nstruct constant gap resized\n\n" );
+        ddt = create_merged_contig_with_gaps( 1 );
+        MPI_DDT_DUMP( ddt );
+        do_test_for_ddt( run_tests, ddt, ddt, MAX_LENGTH );
+        MPI_Type_free( &ddt );
+    }
+
     MPI_Finalize ();
     exit(0);
 }