Changes from fix_sorting

src/include/migraphx/module.hpp

@@ -222,12 +222,17 @@ struct MIGRAPHX_EXPORT module
     void annotate(std::ostream& os, std::function<void(instruction_ref)> a) const;
 
     std::vector<module_ref> get_sub_modules(bool shallow = false) const;
-    // sorts the module in reverse-post order DFS order
+    /* sorts the module in topological order aka reverse-post order (RPO) DFS order
+       it takes last instruction or @return as the root and walks back the graph and moves inputs
+       of the each instruction such that it appears before the instruction itself.
+    */
     module& sort();
-    // if the instruction has the module arguments then all the parameters/instructions used by that
-    // module from the main/parent module must be calculated before the instruction can be executed.
-    // therefore, apart from the input instructions, those other instructions are implicit
-    // dependencies
+    /* Any instruction "X" can have module arguments and those modules inside them can use any other
+     * instruction "Y" from predecessor modules of the instruction "X". Such instruction "Y" inside
+     * module args are not listed as input instructions to "X". But those instructions "Y" must be
+     * evaluted before the instruction "X" can. Therefore such "Y" instructions are considered
+     * implicit dependency to "X".
+     */
     ins_dep_map calc_implicit_deps() const;
 
     MIGRAPHX_EXPORT friend std::ostream& operator<<(std::ostream& os, const module& m);

src/program.cpp

@@ -41,12 +41,12 @@
 #include <migraphx/marker.hpp>
 #include <migraphx/supported_segments.hpp>
 #include <iostream>
+#include <queue>
 #include <sstream>
 #include <algorithm>
 #include <set>
 #include <unordered_map>
 #include <utility>
-
 #include <unordered_set>
 #include <map>
 #include <cassert>
@@ -1198,11 +1198,19 @@ void program::remove_unused_modules()
 
 program& program::sort()
 {
-    for(auto& pp : this->impl->modules)
+    std::queue<migraphx::module_ref> mqueue;
+    mqueue.push(get_main_module());
+    while(!mqueue.empty())
+    {
+        module_ref current_mod = mqueue.front();
+        current_mod->sort();
+        mqueue.pop();
+        auto child_mods = current_mod->get_sub_modules(true);
+        for(auto& sub_mod : child_mods)
         {
-        pp.second.sort();
+            mqueue.push(sub_mod);
+        }
     }
-
     return *this;
 }