fix issue boostorg#774 by force swapping keys when they are identical

Author: rosenrodt

include/boost/compute/algorithm/detail/merge_sort_on_gpu.hpp

@@ -170,8 +170,12 @@ inline size_t bitonic_block_sort(KeyIterator keys_first,
             k.decl<bool>("compare") << " = " <<
                 compare(k.var<key_type>("sibling_key"),
                         k.var<key_type>("my_key")) << ";\n" <<
+            k.decl<bool>("equal") << " = !(compare || " <<
+                compare(k.var<key_type>("my_key"),
+                        k.var<key_type>("sibling_key")) << ");\n" <<
             k.decl<bool>("swap") <<
                 " = compare ^ (sibling_idx < lid) ^ direction;\n" <<
+            "swap |= equal;\n" <<
             "my_key = swap ? sibling_key : my_key;\n";
     if(sort_by_key)
     {
@@ -220,8 +224,12 @@ inline size_t bitonic_block_sort(KeyIterator keys_first,
             k.decl<bool>("compare") << " = " <<
                 compare(k.var<key_type>("sibling_key"),
                         k.var<key_type>("my_key")) << ";\n" <<
+            k.decl<bool>("equal") << " = !(compare || " <<
+                compare(k.var<key_type>("my_key"),
+                        k.var<key_type>("sibling_key")) << ");\n" <<
             k.decl<bool>("swap") <<
                 " = compare ^ (sibling_idx < lid);\n" <<
+            "swap |= equal;\n" <<
             "my_key = swap ? sibling_key : my_key;\n";
     if(sort_by_key)
     {