make separated literal coloring a default

src/opt/memory_coloring_impl.cpp

@@ -83,13 +83,6 @@ bool memory_coloring_impl::allocate(interval_ptr interval)
             offset += (element_size - (offset % element_size));
         conflict_queue.pop();
     }
-#if 1
-    if (interval->get_end() == latest_end_point) {
-        offset = required_bytes;
-        if((offset % element_size) != 0)
-            offset += (element_size - (offset % element_size));
-    }
-#endif    
     segment.offset = offset;
     MIGRAPH_DEBUG(segment.dump());
     required_bytes = std::max(required_bytes, offset + segment.size);
@@ -124,6 +117,7 @@ void memory_coloring_impl::build()
                 live_range& range        = def_interval->segment;
                 def_interval->result     = iter->result;
                 def_interval->is_literal = is_lit;
+                if(!is_lit || unify_literals)
                     alloc_queue.push(def_interval);
                 range.begin             = cur_points;
                 def_interval->def_point = cur_points;
@@ -172,7 +166,7 @@ void memory_coloring_impl::build()
                 live_set.insert(max_value_number);
                 live_ranges[max_value_number] = &(interval->segment);
                 earliest_end_point            = cur_points;
-                if (latest_end_point == -1)
+                if(latest_end_point == -1)
                     latest_end_point = cur_points;
             }
             else
@@ -219,6 +213,9 @@ void memory_coloring_impl::rewrite()
             if(interval->get_begin() == invalid_offset)
                 continue;
 
+            if(!unify_literals && interval->is_literal)
+                continue;
+
             std::size_t offset = 0;
             if(interval->get_offset() == invalid_offset)
             {

src/opt/memory_coloring_impl.hpp

@@ -50,7 +50,8 @@ using interval_ptr = live_interval*;
 
 struct memory_coloring_impl
 {
-    memory_coloring_impl(program* p, std::string alloc_op) : p_program(p), allocation_op(std::move(alloc_op))
+    memory_coloring_impl(program* p, std::string alloc_op)
+        : p_program(p), allocation_op(std::move(alloc_op))
     {
         instr2_live.clear();
         live_ranges.clear();
@@ -61,6 +62,7 @@ struct memory_coloring_impl
         operand_alias.clear();
         earliest_end_point = -1;
         latest_end_point   = -1;
+        unify_literals     = false;
     }
     bool allocate(interval_ptr);
     void add_conflicts(std::set<int>& live_set, int val)
@@ -171,6 +173,8 @@ struct memory_coloring_impl
     int earliest_end_point;
     // The latest program point where an live interval ends.
     int latest_end_point;
+    // Whether to unify literals into coloring.
+    bool unify_literals;
     std::string allocation_op{};
 };
 } // namespace migraph

src/targets/gpu/target.cpp

@@ -31,7 +31,6 @@ std::vector<pass> target::get_passes(migraph::context& gctx) const
         dead_code_elimination{},
         lowering{ctx},
         memory_coloring{"hip::allocate"},
-        lowering_memory_coloring{&ctx},
         fuse_ops{},
         dead_code_elimination{},
         eliminate_workspace{},

src/targets/gpu/write_literals.cpp

@@ -26,9 +26,6 @@ struct hip_load_literal
 
 void write_literals::apply(program& p) const
 {
-    if(!enabled(MIGRAPH_DISABLE_MEMORY_COLORING{}))
-        return;
-
     assert(ctx != nullptr);
     for(auto ins : iterator_for(p))
     {