IntelGPU backend: Add graph dump ability (#1925)

src/ngraph/runtime/intelgpu/CMakeLists.txt

@@ -25,6 +25,7 @@ set(SRC
     intelgpu_op_softmax.cpp
     intelgpu_op_custom_func_call.cpp
     code_writer.cpp
+    visualize_tree.cpp
 )
 
 if (NGRAPH_INTELGPU_ENABLE)

src/ngraph/runtime/intelgpu/intelgpu_backend.cpp

@@ -54,6 +54,7 @@
 #include "ngraph/runtime/intelgpu/intelgpu_op_custom_kernels.hpp"
 #include "ngraph/runtime/intelgpu/intelgpu_op_softmax.hpp"
 #include "ngraph/runtime/intelgpu/intelgpu_tensor_view.hpp"
+#include "ngraph/runtime/intelgpu/visualize_tree.hpp"
 
 #include "ngraph/function.hpp"
 #include "ngraph/node.hpp"
@@ -328,6 +329,12 @@ runtime::intelgpu::IntelGPUBackend::IntelGPUBackend()
         m_disable_backend_optimizations = true;
     }
 
+    // Dumps the input Function into Graphviz format
+    if (getenv("NGRAPH_INTELGPU_DUMP_FUNCTION") != nullptr)
+    {
+        m_dump_graph_enable = true;
+    }
+
     cldnn::engine_configuration cldnn_configuration(profiling);
     ocl_engine = make_shared<cldnn::engine>(cldnn_configuration);
 }
@@ -356,6 +363,11 @@ bool runtime::intelgpu::IntelGPUBackend::compile(shared_ptr<Function> func)
 
     cldnn::topology topology;
 
+    if (m_dump_graph_enable)
+    {
+        visualize_tree(func, "intelgpu_", "_orig");
+    }
+
     if (!m_disable_backend_optimizations)
     {
         ngraph::pass::Manager pass_manager;
@@ -369,6 +381,11 @@ bool runtime::intelgpu::IntelGPUBackend::compile(shared_ptr<Function> func)
         pass_manager.register_pass<ngraph::pass::GetOutputElementElimination>();
 
         pass_manager.run_passes(func);
+
+        if (m_dump_graph_enable)
+        {
+            visualize_tree(func, "intelgpu_", "_opt");
+        }
     }
 
     for (shared_ptr<Node> op : func->get_ops())

src/ngraph/runtime/intelgpu/intelgpu_backend.hpp

@@ -82,5 +82,6 @@ private:
 
     bool m_profile_enable = false;
     long m_profile_lines_limit_count = 10;
+    bool m_dump_graph_enable = false;
     std::string delim = std::string(":");
 };

src/ngraph/runtime/intelgpu/visualize_tree.hpp

+//*****************************************************************************
+// Copyright 2017-2018 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//*****************************************************************************
+
+#pragma once
+
+#include "ngraph/function.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace intelgpu
+        {
+            // This function writes the input func into file in Graphviz format.
+            // On large graphs, the "dot" utility requires lot of time to visualize the input.
+            // Example: dot -Tpdf intelgpu_Function_0_orig.dot -o intelgpu_Function_0_orig.pdf
+            void visualize_tree(const std::shared_ptr<Function>& func,
+                                const std::string& file_prefix,
+                                const std::string& file_suffix);
+        }
+    }
+}