Change compiler action to EmitLLVMOnly. (#3127)

Wait until performance counter data is set to release function for CODEGEN.

src/ngraph/codegen/compiler.cpp

@@ -343,7 +343,7 @@ std::unique_ptr<codegen::Module>
     preprocessor_options.RemappedFileBuffers.push_back({m_source_name, buffer.get()});
 
     // Create and execute action
-    m_compiler_action.reset(new EmitCodeGenOnlyAction());
+    m_compiler_action.reset(new EmitLLVMOnlyAction());
     std::unique_ptr<llvm::Module> rc;
     bool reinitialize = false;
     if (m_compiler->ExecuteAction(*m_compiler_action) == true)

src/ngraph/runtime/cpu/cpu_emitter.cpp

@@ -1639,23 +1639,12 @@ namespace ngraph
                 auto type_name = embed->get_element_type().c_type_string();
                 auto element_count = shape_size(embed->get_argument(0)->get_shape());
 
-                // FIXME
-                // clang generates 16 bytes aligned store with unaligned address,
-                // which results in segmentation fault.
-                // Workaround for now: Use push_back to avoid generating such store.
-                auto arg1_shape = args[1].get_shape();
-                writer << "ngraph::Shape shape;\n";
-                for (auto i = 0; i < arg1_shape.size(); i++)
-                {
-                    writer << "shape.push_back(" << std::to_string(arg1_shape[i]) << ");\n";
-                }
-
                 writer << "reference::embedding<" << type_name << "," << index_type_name << ">(";
                 writer << "            " << args[0].get_name() << ",\n";
                 writer << "            " << args[1].get_name() << ",\n";
                 writer << "            " << out[0].get_name() << ",\n";
                 writer << "            " << element_count << ",\n";
-                writer << "             shape);\n";
+                writer << "           {" << join(args[1].get_shape()) << "});\n";
                 writer.block_end();
             }
 

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -1138,7 +1138,7 @@ using namespace ngraph::runtime;
     }
 
     m_is_compiled = true;
-    if (m_release_function)
+    if (m_release_function && !m_emit_timing)
     {
         release_function();
     }
@@ -2023,6 +2023,10 @@ const vector<runtime::PerformanceCounter>& runtime::cpu::CPU_ExternalFunction::g
                 }
             }
         }
+        if (m_release_function)
+        {
+            release_function();
+        }
     }
 #endif
     return m_perf_counters;

src/ngraph/runtime/cpu/pass/cpu_mkldnn_primitive_build.cpp

@@ -55,17 +55,7 @@
 #include "ngraph/runtime/cpu/op/sigmoid.hpp"
 #include "ngraph/runtime/cpu/op/update_slice.hpp"
 
-#define WRITE_MKLDNN_DIMS(X)                                                                       \
-    writer << "mkldnn::memory::dims{";                                                             \
-    if (X.size() > 1)                                                                              \
-    {                                                                                              \
-        for (auto i = 0; i < X.size() - 1; i++)                                                    \
-        {                                                                                          \
-            writer << std::to_string(X[i]) << ", ";                                                \
-        }                                                                                          \
-    }                                                                                              \
-    writer << std::to_string(X[X.size() - 1]);                                                     \
-    writer << "}, \n"
+#define WRITE_MKLDNN_DIMS(X) writer << "mkldnn::memory::dims{" << join(X) << "}, \n";
 
 using namespace ngraph;
 using namespace ngraph::op;
@@ -780,21 +770,10 @@ namespace ngraph
                     serialize_memory_descs(desc_file, descs, deps[0]);
 
                     writer << "\n// create memory primitive descriptors\n";
-                    // FIXME
-                    // clang generates 16 bytes aligned store with unaligned address,
-                    // which results in segmentation fault.
-                    // Workaround for now: Use push_back to avoid generating such store.
-                    writer << "mkldnn::memory::dims dims1;\n";
-                    for (auto i = 0; i < result_shape.size(); i++)
-                    {
-                        writer << "dims1.push_back(" << std::to_string(result_shape[i]) << ");\n";
-                    }
 
-                    writer << "\nmkldnn::memory::dims dims2;\n";
-                    for (auto i = 0; i < lower_bounds.size(); i++)
-                    {
-                        writer << "dims2.push_back(" << std::to_string(lower_bounds[i]) << ");\n";
-                    }
+                    writer << "mkldnn::memory::dims dims1{" << join(result_shape) << "};\n";
+                    writer << "mkldnn::memory::dims dims2{" << join(lower_bounds) << "};\n";
+                    ;
 
                     writer << "mkldnn::memory::primitive_desc input_pd = "
                               "mkldnn::memory::primitive_desc(*cg_ctx->mkldnn_descriptors["