[GPU] New debug config: dry_run (#7689)

Enabling dry_run will shorten runtime of any application by exiting just
after graph compilation.

inference-engine/src/cldnn_engine/cldnn_graph.cpp

@@ -4,6 +4,7 @@
 
 #include <cldnn/graph/network.hpp>
 #include <cldnn/runtime/profiling.hpp>
+#include "cldnn/runtime/debug_configuration.hpp"
 
 #include "cldnn_graph.h"
 #include "simple_math.h"
@@ -82,6 +83,13 @@ void CLDNNGraph::Build() {
     }
 
     UpdateImplementationsMap();
+
+    GPU_DEBUG_GET_INSTANCE(debug_config);
+    GPU_DEBUG_IF(!debug_config->dry_run_path.empty()) {
+        CNNNetwork net(GetExecGraphInfo());
+        net.serialize(debug_config->dry_run_path);
+        exit(0);
+    }
 }
 
 std::shared_ptr<cldnn::network> CLDNNGraph::BuildNetwork(std::shared_ptr<cldnn::program> program) {

inference-engine/thirdparty/clDNN/api/cldnn/runtime/debug_configuration.hpp

@@ -32,6 +32,7 @@ public:
     std::string dump_sources;       // Dump opencl sources
     std::string dump_layers_path;   // Enable dumping intermediate buffers and set the dest path
     std::string dump_layers;        // Dump intermediate buffers of specified layers only, separated by space
+    std::string dry_run_path;       // Dry run and serialize execution graph into the specified path
     int dump_layers_dst_only;       // Dump only output of layers
     static const debug_configuration *get_instance();
 };

inference-engine/thirdparty/clDNN/runtime/debug_configuration.cpp

@@ -103,6 +103,7 @@ debug_configuration::debug_configuration()
         , dump_layers_path(std::string())
         , dump_layers(std::string())
         , dump_layers_dst_only(0)
+        , dry_run_path(std::string())
         , disable_onednn(0) {
 #ifdef GPU_DEBUG_CONFIG
     get_common_debug_env_var("Verbose", verbose);
@@ -114,6 +115,7 @@ debug_configuration::debug_configuration()
     get_gpu_debug_env_var("DumpLayers", dump_layers);
     get_gpu_debug_env_var("DumpLayersDstOnly", dump_layers_dst_only);
     get_gpu_debug_env_var("DisableOnednn", disable_onednn);
+    get_gpu_debug_env_var("DryRunPath", dry_run_path);
 
     if (dump_layers_path.length() > 0 && !disable_usm) {
         disable_usm = 1;

inference-engine/thirdparty/clDNN/src/impls/ocl/primitive_base.hpp

@@ -9,7 +9,6 @@
 #include "primitive_inst.h"
 #include "cldnn/graph/program.hpp"
 #include "cldnn/runtime/error_handler.hpp"
-#include "cldnn/runtime/debug_configuration.hpp"
 #include "kernel_selector_helper.h"
 #include "cldnn/graph/network.hpp"
 #include "register.hpp"

inference-engine/thirdparty/clDNN/src/program.cpp

@@ -423,9 +423,17 @@ void program::build_program(bool is_internal) {
     { pre_optimize_graph(is_internal); }
     run_graph_compilation();
     { post_optimize_graph(is_internal); }
+
+    GPU_DEBUG_GET_INSTANCE(debug_config);
+#ifdef GPU_DEBUG_CONFIG
+    if (debug_config->dry_run_path.empty()) {
+#else
+    {
+#endif
         prepare_memory_dependencies();
         compile();
         init_kernels();
+    }
 
     if (!is_internal) {
         prim_info = get_current_stage_info();