Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/AMDMIGraphX into lstm_operator

src/opt/memory_coloring_impl.cpp

@@ -118,11 +118,11 @@ void memory_coloring_impl::build()
                 live_range& range        = def_interval->segment;
                 def_interval->result     = iter->get_shape();
                 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;
                 range.size               = (iter->get_shape()).bytes();
+                if(!is_lit || unify_literals)
+                    alloc_queue.push(def_interval);
                 live_set.erase(range.vn);
             }
         }
@@ -233,9 +233,8 @@ void memory_coloring_impl::verify()
 
             if(segment.begin == invalid_offset)
             {
-                // TODO: This check breaks on the tests
-                // if(!interval.is_live_on_entry)
-                // MIGRAPHX_THROW("interval is not live on entry");
+                if(!interval.is_live_on_entry)
+                    MIGRAPHX_THROW("interval is not live on entry");
                 continue;
             }
 

test/memory_coloring_test.cpp

@@ -78,7 +78,7 @@ TEST_CASE(test3)
     auto p3 = add_alloc(p, {migraphx::shape::float_type, {40}});
     p.add_instruction(pass_op{}, p3, p1);
     p.compile(memory_coloring_target{});
-    CHECK(p.get_parameter_shape("scratch").bytes() == 704); // The optimal solution is actually 672
+    CHECK(p.get_parameter_shape("scratch").bytes() == 672);
     CHECK(no_allocate(p));
 }
 
@@ -487,7 +487,7 @@ TEST_CASE(test33)
     auto a5 = add_alloc(p, {migraphx::shape::float_type, {40}});
     p.add_instruction(pass_op{}, a5, p1);
     p.compile(memory_coloring_target{});
-    CHECK(p.get_parameter_shape("scratch").bytes() == 224);
+    CHECK(p.get_parameter_shape("scratch").bytes() == 192);
     CHECK(no_allocate(p));
 }
 
@@ -594,7 +594,7 @@ TEST_CASE(test38)
     auto p83    = p.add_instruction(pass_op{}, p78, p77);
     p.add_instruction(pass_op{}, output, p83, p63);
     p.compile(memory_coloring_target{});
-    CHECK(p.get_parameter_shape("scratch").bytes() == 6422528);
+    CHECK(p.get_parameter_shape("scratch").bytes() == 7225344); // Optimal solution is 6422528
     CHECK(no_allocate(p));
 }