Merge pull request #243 from NervanaSystems/bob/static_compiler

Bob/static compiler

src/ngraph/CMakeLists.txt

@@ -106,12 +106,14 @@ if (NGRAPH_CPU_ENABLE AND LLVM_INCLUDE_DIR AND
     set(SRC ${SRC}
         codegen/code_writer.cpp
         codegen/compiler.cpp
+        codegen/execution_engine.cpp
         runtime/cpu/call_frame.cpp
         runtime/cpu/cpu_backend.cpp
         runtime/cpu/cpu_manager.cpp
         runtime/cpu/cpu_kernels.cpp
         runtime/cpu/emitter.cpp
         runtime/cpu/external_function.cpp
+        runtime/cpu/memory_handler.cpp
         runtime/cpu/tensor_view.cpp
         )
     # LLVM binary builds are typically built without RTTI

src/ngraph/codegen/compiler.cpp

@@ -45,13 +45,15 @@
 #include <clang/Frontend/TextDiagnosticPrinter.h>
 #include <llvm/Support/TargetSelect.h>
 
-#include <llvm/ExecutionEngine/ExecutionEngine.h>
-#include <llvm/ExecutionEngine/SectionMemoryManager.h>
-
 #include "ngraph/codegen/compiler.hpp"
+#include "ngraph/file_util.hpp"
+#include "ngraph/log.hpp"
+#include "ngraph/util.hpp"
 
 // TODO: Fix leaks
 
+// #define USE_CACHE
+
 using namespace clang;
 using namespace llvm;
 using namespace llvm::opt;
@@ -59,26 +61,36 @@ using namespace std;
 
 using namespace ngraph::codegen;
 
-static std::string GetExecutablePath(const char* Argv0)
+static HeaderCache s_header_cache;
+static StaticCompiler s_static_compiler;
+static std::mutex m_mutex;
+
+Compiler::Compiler()
 {
-    // This just needs to be some symbol in the binary; C++ doesn't
-    // allow taking the address of ::main however.
-    void* MainAddr = reinterpret_cast<void*>(GetExecutablePath);
-    return llvm::sys::fs::getMainExecutable(Argv0, MainAddr);
 }
 
-execution_state::execution_state()
-    : m_execution_engine{nullptr}
-    , precompiled_headers_enabled(false)
-    , debuginfo_enabled(false)
+Compiler::~Compiler()
 {
 }
 
-execution_state::~execution_state()
+std::unique_ptr<llvm::Module> Compiler::compile(const std::string& source)
 {
+    lock_guard<mutex> lock(m_mutex);
+    return s_static_compiler.compile(source);
 }
 
-std::unique_ptr<llvm::Module> execution_state::compile(const string& source, const string& name)
+static std::string GetExecutablePath(const char* Argv0)
+{
+    // This just needs to be some symbol in the binary; C++ doesn't
+    // allow taking the address of ::main however.
+    void* MainAddr = reinterpret_cast<void*>(GetExecutablePath);
+    return llvm::sys::fs::getMainExecutable(Argv0, MainAddr);
+}
+
+StaticCompiler::StaticCompiler()
+    : m_precompiled_headers_enabled(false)
+    , m_debuginfo_enabled(false)
+    , m_source_name("code.cpp")
 {
     llvm::InitializeAllTargets();
     llvm::InitializeAllTargetMCs();
@@ -87,7 +99,7 @@ std::unique_ptr<llvm::Module> execution_state::compile(const string& source, con
 
     // Prepare compilation arguments
     vector<const char*> args;
-    args.push_back(name.c_str());
+    args.push_back(m_source_name.c_str());
 
     // Prepare DiagnosticEngine
     DiagnosticOptions DiagOpts;
@@ -97,45 +109,76 @@ std::unique_ptr<llvm::Module> execution_state::compile(const string& source, con
         new DiagnosticsEngine(pDiagIDs, &DiagOpts, textDiagPrinter);
 
     // Create and initialize CompilerInstance
-    std::unique_ptr<CompilerInstance> Clang(new CompilerInstance());
-    Clang->createDiagnostics();
+    m_compiler = std::unique_ptr<CompilerInstance>(new CompilerInstance());
+    m_compiler->createDiagnostics();
 
     // Initialize CompilerInvocation
     CompilerInvocation::CreateFromArgs(
-        Clang->getInvocation(), &args[0], &args[0] + args.size(), *pDiagnosticsEngine);
+        m_compiler->getInvocation(), &args[0], &args[0] + args.size(), *pDiagnosticsEngine);
 
     // Infer the builtin include path if unspecified.
-    if (Clang->getHeaderSearchOpts().UseBuiltinIncludes &&
-        Clang->getHeaderSearchOpts().ResourceDir.empty())
+    if (m_compiler->getHeaderSearchOpts().UseBuiltinIncludes &&
+        m_compiler->getHeaderSearchOpts().ResourceDir.empty())
     {
         void* MainAddr = reinterpret_cast<void*>(GetExecutablePath);
         auto path = CompilerInvocation::GetResourcesPath(args[0], MainAddr);
-        Clang->getHeaderSearchOpts().ResourceDir = path;
+        m_compiler->getHeaderSearchOpts().ResourceDir = path;
     }
 
-    auto& HSO = Clang->getInvocation().getHeaderSearchOpts();
-    // Add base toolchain-supplied header paths
-    // Ideally one would use the Linux toolchain definition in clang/lib/Driver/ToolChains.h
-    // But that's a private header and isn't part of the public libclang API
-    // Instead of re-implementing all of that functionality in a custom toolchain
-    // just hardcode the paths relevant to frequently used build/test machines for now
-    HSO.AddPath(CLANG_BUILTIN_HEADERS_PATH, clang::frontend::System, false, false);
-    HSO.AddPath("/usr/include/x86_64-linux-gnu", clang::frontend::System, false, false);
-    HSO.AddPath("/usr/include", clang::frontend::System, false, false);
-    // Add C++ standard library headers
-    // Debian-like + GCC 4.8 libstdc++
-    HSO.AddPath("/usr/include/x86_64-linux-gnu/c++/4.8", clang::frontend::System, false, false);
-    HSO.AddPath("/usr/include/c++/4.8", clang::frontend::System, false, false);
-    // Debian-like + GCC 5 libstdc++
-    HSO.AddPath("/usr/include/x86_64-linux-gnu/c++/5", clang::frontend::System, false, false);
-    HSO.AddPath("/usr/include/c++/5", clang::frontend::System, false, false);
-    HSO.AddPath(EIGEN_HEADERS_PATH, clang::frontend::System, false, false);
-    HSO.AddPath(NGRAPH_HEADERS_PATH, clang::frontend::System, false, false);
+    if (s_header_cache.is_valid() == false)
+    {
+        // Add base toolchain-supplied header paths
+        // Ideally one would use the Linux toolchain definition in clang/lib/Driver/ToolChains.h
+        // But that's a private header and isn't part of the public libclang API
+        // Instead of re-implementing all of that functionality in a custom toolchain
+        // just hardcode the paths relevant to frequently used build/test machines for now
+        add_header_search_path(CLANG_BUILTIN_HEADERS_PATH);
+        add_header_search_path("/usr/include/x86_64-linux-gnu");
+        add_header_search_path("/usr/include");
+
+        // Search for headers in
+        //    /usr/include/x86_64-linux-gnu/c++/N.N
+        //    /usr/include/c++/N.N
+        // and add them to the header search path
+
+        file_util::iterate_files("/usr/include/x86_64-linux-gnu/c++/",
+                                 [&](const std::string& file, bool is_dir) {
+                                     if (is_dir)
+                                     {
+                                         string dir_name = file_util::get_file_name(file);
+                                         if (is_version_number(dir_name))
+                                         {
+                                             add_header_search_path(file);
+                                         }
+                                     }
+                                 });
+
+        file_util::iterate_files("/usr/include/c++/", [&](const std::string& file, bool is_dir) {
+            if (is_dir)
+            {
+                string dir_name = file_util::get_file_name(file);
+                if (is_version_number(dir_name))
+                {
+                    add_header_search_path(file);
+                }
+            }
+        });
+
+        add_header_search_path(EIGEN_HEADERS_PATH);
+        add_header_search_path(NGRAPH_HEADERS_PATH);
+#ifdef USE_CACHE
+        s_header_cache.set_valid();
+#endif
+    }
+
+#ifdef USE_CACHE
+    use_cached_files(m_compiler);
+#endif
 
     // Language options
     // These are the C++ features needed to compile ngraph headers
     // and any dependencies like Eigen
-    auto LO = Clang->getInvocation().getLangOpts();
+    auto LO = m_compiler->getInvocation().getLangOpts();
     LO->CPlusPlus = 1;
     LO->CPlusPlus11 = 1;
     LO->Bool = 1;
@@ -148,7 +191,7 @@ std::unique_ptr<llvm::Module> execution_state::compile(const string& source, con
     LO->OpenMPUseTLS = 1;
 
     // CodeGen options
-    auto& CGO = Clang->getInvocation().getCodeGenOpts();
+    auto& CGO = m_compiler->getInvocation().getCodeGenOpts();
     CGO.OptimizationLevel = 3;
     CGO.RelocationModel = "static";
     CGO.ThreadModel = "posix";
@@ -158,22 +201,22 @@ std::unique_ptr<llvm::Module> execution_state::compile(const string& source, con
     CGO.VectorizeSLP = 1;
     CGO.CXAAtExit = 0;
 
-    if (debuginfo_enabled)
+    if (m_debuginfo_enabled)
     {
         CGO.setDebugInfo(codegenoptions::FullDebugInfo);
     }
 
-    if (precompiled_headers_enabled)
+    if (m_precompiled_headers_enabled)
     {
         // Preprocessor options
-        auto& PPO = Clang->getInvocation().getPreprocessorOpts();
+        auto& PPO = m_compiler->getInvocation().getPreprocessorOpts();
         PPO.ImplicitPCHInclude = "ngcpu.pch";
         PPO.DisablePCHValidation = 1;
     }
 
     // Enable various target features
     // Most of these are for Eigen
-    auto& TO = Clang->getInvocation().getTargetOpts();
+    auto& TO = m_compiler->getInvocation().getTargetOpts();
     // TODO: This needs to be configurable and selected carefully
     TO.CPU = "broadwell";
     TO.FeaturesAsWritten.emplace_back("+sse");
@@ -185,62 +228,121 @@ std::unique_ptr<llvm::Module> execution_state::compile(const string& source, con
     TO.FeaturesAsWritten.emplace_back("+avx");
     TO.FeaturesAsWritten.emplace_back("+avx2");
     TO.FeaturesAsWritten.emplace_back("+fma");
+}
 
-    // Map code filename to a memoryBuffer
-    StringRef source_ref(source);
-    unique_ptr<MemoryBuffer> buffer = MemoryBuffer::getMemBufferCopy(source_ref);
-    Clang->getInvocation().getPreprocessorOpts().addRemappedFile(name, buffer.get());
-
-    // Create and execute action
-    CodeGenAction* compilerAction = new EmitCodeGenOnlyAction();
-    std::unique_ptr<llvm::Module> rc;
-    if (Clang->ExecuteAction(*compilerAction) == true)
-    {
-        rc = compilerAction->takeModule();
-    }
-
-    buffer.release();
-
-    return rc;
+StaticCompiler::~StaticCompiler()
+{
 }
 
-bool execution_state::add_module(std::unique_ptr<llvm::Module>& module)
+bool StaticCompiler::is_version_number(const string& path)
 {
-    if (module)
+    bool rc = true;
+    vector<string> tokens = ngraph::split(path, '.');
+    for (string s : tokens)
     {
-        if (!m_execution_engine)
+        for (char c : s)
         {
-            m_execution_engine = llvm::EngineBuilder(move(module))
-                                     .setEngineKind(llvm::EngineKind::JIT)
-                                     .setOptLevel(llvm::CodeGenOpt::Aggressive)
-                                     .setErrorStr(&jit_error)
-                                     .create();
-
-            if (!m_execution_engine)
+            if (!isdigit(c))
             {
-                return false;
+                rc = false;
             }
         }
     }
-    else
+    return rc;
+}
+
+void StaticCompiler::add_header_search_path(const string& path)
+{
+    if (!contains(m_extra_search_path_list, path))
     {
-        return false;
-    }
+        m_extra_search_path_list.push_back(path);
+        HeaderSearchOptions& hso = m_compiler->getInvocation().getHeaderSearchOpts();
+#ifdef USE_CACHE
+        static vector<string> valid_ext = {".h", ".hpp", ".tcc", ""};
+        string mapped_path = file_util::path_join("/$BUILTIN", path);
+        mapped_path = path;
+        s_header_cache.add_path(mapped_path);
+        auto func = [&](const std::string& file, bool is_dir) {
+            if (!is_dir)
+            {
+                string ext = file_util::get_file_ext(file);
+                if (contains(valid_ext, ext))
+                {
+                    // This is a header file
+                    string relative_name = file.substr(path.size() + 1);
+                    string mapped_name = file_util::path_join(mapped_path, relative_name);
+
+                    ErrorOr<unique_ptr<MemoryBuffer>> code = MemoryBuffer::getFile(file);
+                    if (error_code ec = code.getError())
+                    {
+                        // throw up
+                        throw runtime_error("could not find file '" + file + "'");
+                    }
 
-    return true;
+                    s_header_cache.add_file(mapped_name, code.get());
+                }
+            }
+        };
+        file_util::iterate_files(path, func, true);
+#else
+        hso.AddPath(path, clang::frontend::System, false, false);
+#endif
+    }
 }
 
-void execution_state::finalize()
+void StaticCompiler::use_cached_files()
 {
-    if (m_execution_engine)
+    HeaderSearchOptions& hso = m_compiler->getInvocation().getHeaderSearchOpts();
+    for (const string& path : s_header_cache.get_include_paths())
     {
-        m_execution_engine->finalizeObject();
-        m_execution_engine->runStaticConstructorsDestructors(false);
+        hso.AddPath(path, clang::frontend::System, false, false);
     }
-    else
+    for (auto& header : s_header_cache.get_header_map())
     {
-        throw std::runtime_error(
-            "Error in finalize: " +
-            (jit_error.empty() ? "Could not create an execution engine" : jit_error));
+        m_compiler->getPreprocessorOpts().addRemappedFile(header.first, header.second.get());
     }
 }
+
+std::unique_ptr<llvm::Module> StaticCompiler::compile(const string& source)
+{
+    // Map code filename to a memoryBuffer
+    StringRef source_ref(source);
+    unique_ptr<MemoryBuffer> buffer = MemoryBuffer::getMemBufferCopy(source_ref);
+    m_compiler->getInvocation().getPreprocessorOpts().addRemappedFile(m_source_name, buffer.get());
+
+    // Create and execute action
+    CodeGenAction* compilerAction = new EmitCodeGenOnlyAction();
+    std::unique_ptr<llvm::Module> rc;
+    if (m_compiler->ExecuteAction(*compilerAction) == true)
+    {
+        rc = compilerAction->takeModule();
+    }
+
+    buffer.release();
+
+    m_compiler->getInvocation().getPreprocessorOpts().clearRemappedFiles();
+
+    return rc;
+}
+
+// std::unique_ptr<llvm::Module> StaticCompiler::generate_pch(const string& source)
+// {
+//     // Map code filename to a memoryBuffer
+//     StringRef source_ref(source);
+//     unique_ptr<MemoryBuffer> buffer = MemoryBuffer::getMemBufferCopy(source_ref);
+//     m_compiler->getInvocation().getPreprocessorOpts().addRemappedFile(m_source_name, buffer.get());
+
+//     // Create and execute action
+//     CodeGenAction* compilerAction = new GeneratePCHAction();
+//     std::unique_ptr<llvm::Module> rc;
+//     if (m_compiler->ExecuteAction(*compilerAction) == true)
+//     {
+//         rc = compilerAction->takeModule();
+//     }
+
+//     buffer.release();
+
+//     m_compiler->getInvocation().getPreprocessorOpts().clearRemappedFiles();
+
+//     return rc;
+// }

src/ngraph/codegen/compiler.hpp

@@ -27,10 +27,18 @@ namespace ngraph
     namespace codegen
     {
         class module;
-        class execution_state;
+        class Compiler;
+        class StaticCompiler;
+        class HeaderCache;
     }
 }
 
+namespace clang
+{
+    class HeaderSearchOptions;
+    class CompilerInstance;
+}
+
 class ngraph::codegen::module
 {
 public:
@@ -38,39 +46,57 @@ private:
     std::unique_ptr<llvm::Module> m_module;
 };
 
-class ngraph::codegen::execution_state : public llvm::SectionMemoryManager
+class ngraph::codegen::Compiler
 {
 public:
-    execution_state();
-    ~execution_state();
-
-    void set_precompiled_headers_enabled(bool state) { precompiled_headers_enabled = state; }
-    bool is_precompiled_headers_enabled() { return precompiled_headers_enabled; }
-    void set_debuginfo_enabled(bool state) { debuginfo_enabled = state; }
-    bool is_debuginfo_enabled() { return debuginfo_enabled; }
-    std::unique_ptr<llvm::Module> compile(const std::string& source, const std::string& name = "");
-
-    bool add_module(std::unique_ptr<llvm::Module>&);
+    Compiler();
+    ~Compiler();
+    std::unique_ptr<llvm::Module> compile(const std::string& source);
 
-    void finalize();
+private:
+};
 
-    template <typename ftype>
-    std::function<ftype> find_function(const std::string& func_name)
-    {
-        auto f = m_execution_engine->getPointerToNamedFunction(func_name);
+class ngraph::codegen::StaticCompiler : public llvm::SectionMemoryManager
+{
+public:
+    StaticCompiler();
+    ~StaticCompiler();
 
-        return f_cast<ftype>(f);
-    }
+    void set_precompiled_headers_enabled(bool state) { m_precompiled_headers_enabled = state; }
+    bool is_precompiled_headers_enabled() { return m_precompiled_headers_enabled; }
+    void set_debuginfo_enabled(bool state) { m_debuginfo_enabled = state; }
+    bool is_debuginfo_enabled() { return m_debuginfo_enabled; }
+    void add_header_search_path(const std::string& path);
+    std::unique_ptr<llvm::Module> compile(const std::string& source);
 
 private:
-    llvm::ExecutionEngine* m_execution_engine;
-    std::string jit_error;
-    bool precompiled_headers_enabled;
-    bool debuginfo_enabled;
+    std::unique_ptr<clang::CompilerInstance> m_compiler;
+    bool m_precompiled_headers_enabled;
+    bool m_debuginfo_enabled;
+    std::string m_source_name;
+    std::vector<std::string> m_extra_search_path_list;
 
-    template <typename signature>
-    std::function<signature> f_cast(void* f)
+    bool is_version_number(const std::string& path);
+    void use_cached_files();
+};
+
+class ngraph::codegen::HeaderCache
+{
+public:
+    bool is_valid() const { return m_headers_valid; }
+    bool set_valid() { return m_headers_valid = true; }
+    void add_path(const std::string& path) { m_include_paths.push_back(path); }
+    void add_file(const std::string& path, std::unique_ptr<llvm::MemoryBuffer>& code)
+    {
+        m_headers.insert(std::make_pair(path, std::move(code)));
+    }
+    const std::map<std::string, std::unique_ptr<llvm::MemoryBuffer>>& get_header_map() const
     {
-        return static_cast<signature*>(reinterpret_cast<signature*>(f));
+        return m_headers;
     }
+    const std::vector<std::string>& get_include_paths() const { return m_include_paths; }
+private:
+    std::map<std::string, std::unique_ptr<llvm::MemoryBuffer>> m_headers;
+    std::vector<std::string> m_include_paths;
+    bool m_headers_valid;
 };

src/ngraph/codegen/execution_engine.cpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#include <llvm/ExecutionEngine/ExecutionEngine.h>
+#include <llvm/ExecutionEngine/SectionMemoryManager.h>
+
+#include "ngraph/codegen/execution_engine.hpp"
+
+using namespace ngraph;
+
+codegen::ExecutionEngine::ExecutionEngine()
+    : m_execution_engine{nullptr}
+{
+}
+
+codegen::ExecutionEngine::~ExecutionEngine()
+{
+}
+
+bool codegen::ExecutionEngine::add_module(std::unique_ptr<llvm::Module>& module)
+{
+    if (module)
+    {
+        if (!m_execution_engine)
+        {
+            m_execution_engine = llvm::EngineBuilder(move(module))
+                                     .setEngineKind(llvm::EngineKind::JIT)
+                                     .setOptLevel(llvm::CodeGenOpt::Aggressive)
+                                     .setErrorStr(&m_jit_error)
+                                     .create();
+
+            if (!m_execution_engine)
+            {
+                return false;
+            }
+        }
+    }
+    else
+    {
+        return false;
+    }
+
+    return true;
+}
+
+void codegen::ExecutionEngine::finalize()
+{
+    if (m_execution_engine)
+    {
+        m_execution_engine->finalizeObject();
+        m_execution_engine->runStaticConstructorsDestructors(false);
+    }
+    else
+    {
+        throw std::runtime_error(
+            "Error in finalize: " +
+            (m_jit_error.empty() ? "Could not create an execution engine" : m_jit_error));
+    }
+}

src/ngraph/codegen/execution_engine.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include <memory>
+
+#include <llvm/ExecutionEngine/MCJIT.h> // forces JIT to link in
+#include <llvm/ExecutionEngine/SectionMemoryManager.h>
+#include <llvm/Option/Arg.h>
+
+namespace ngraph
+{
+    namespace codegen
+    {
+        class ExecutionEngine;
+    }
+}
+
+class ngraph::codegen::ExecutionEngine
+{
+public:
+    ExecutionEngine();
+    ~ExecutionEngine();
+
+    bool add_module(std::unique_ptr<llvm::Module>& module);
+    void finalize();
+
+    template <typename ftype>
+    std::function<ftype> find_function(const std::string& func_name)
+    {
+        auto f = m_execution_engine->getPointerToNamedFunction(func_name);
+
+        return f_cast<ftype>(f);
+    }
+
+private:
+    llvm::ExecutionEngine* m_execution_engine;
+    std::string m_jit_error;
+
+    template <typename signature>
+    std::function<signature> f_cast(void* f)
+    {
+        return static_cast<signature*>(reinterpret_cast<signature*>(f));
+    }
+};

src/ngraph/file_util.cpp

-/*
- Copyright 2016 Nervana Systems Inc.
- 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.
-*/
+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
 
 #include <cassert>
 #include <dirent.h>
@@ -29,10 +28,37 @@
 #include <unistd.h>
 #include <vector>
 
-#include "file_util.hpp"
+#include "ngraph/file_util.hpp"
+#include "ngraph/log.hpp"
 
 using namespace std;
 
+std::string ngraph::file_util::get_file_name(const std::string& s)
+{
+    string rc = s;
+    auto pos = s.find_last_of('/');
+    if (pos != string::npos)
+    {
+        rc = s.substr(pos + 1);
+    }
+    return rc;
+}
+
+std::string ngraph::file_util::get_file_ext(const std::string& s)
+{
+    string rc = get_file_name(s);
+    auto pos = rc.find_last_of('.');
+    if (pos != string::npos)
+    {
+        rc = rc.substr(pos);
+    }
+    else
+    {
+        rc = "";
+    }
+    return rc;
+}
+
 string ngraph::file_util::path_join(const string& s1, const string& s2)
 {
     string rc;
@@ -79,20 +105,22 @@ size_t ngraph::file_util::get_file_size(const string& filename)
 void ngraph::file_util::remove_directory(const string& dir)
 {
     struct stat status;
-    if (stat(dir.c_str(), &status) == -1)
+    if (stat(dir.c_str(), &status) != -1)
     {
-        return;
+        iterate_files(dir,
+                      [](const string& file, bool is_dir) {
+                          if (is_dir)
+                          {
+                              rmdir(file.c_str());
+                          }
+                          else
+                          {
+                              remove(file.c_str());
+                          }
+                      },
+                      true);
+        rmdir(dir.c_str());
     }
-
-    file_util::iterate_files(dir,
-                             [](const string& file, bool is_dir) {
-                                 if (is_dir)
-                                     rmdir(file.c_str());
-                                 else
-                                     remove(file.c_str());
-                             },
-                             true);
-    rmdir(dir.c_str());
 }
 
 void ngraph::file_util::remove_file(const string& file)
@@ -117,7 +145,7 @@ bool ngraph::file_util::make_directory(const string& dir)
 string ngraph::file_util::make_temp_directory(const string& path)
 {
     string fname = path.empty() ? file_util::get_temp_directory() : path;
-    string tmp_template = file_util::path_join(fname, "aeonXXXXXX");
+    string tmp_template = file_util::path_join(fname, "ngraph_XXXXXX");
     char* tmpname = strdup(tmp_template.c_str());
 
     mkdtemp(tmpname);
@@ -129,7 +157,7 @@ string ngraph::file_util::make_temp_directory(const string& path)
 
 std::string ngraph::file_util::get_temp_directory()
 {
-    const vector<string> potential_tmps = {"NERVANA_AEON_TMP", "TMPDIR", "TMP", "TEMP", "TEMPDIR"};
+    const vector<string> potential_tmps = {"NGRAPH_TMP", "TMPDIR", "TMP", "TEMP", "TEMPDIR"};
 
     const char* path = nullptr;
     for (const string& var : potential_tmps)
@@ -161,7 +189,7 @@ vector<char> ngraph::file_util::read_file_contents(const string& path)
         char* p = data.data();
         size_t remainder = file_size;
         size_t offset = 0;
-        while (remainder > 0)
+        while (f && remainder > 0)
         {
             size_t rc = fread(&p[offset], 1, remainder, f);
             offset += rc;
@@ -197,12 +225,16 @@ void ngraph::file_util::iterate_files(const string& path,
                                         else
                                             files.push_back(file);
                                     },
-                                    true);
+                                    recurse);
 
     for (auto f : files)
+    {
         func(f, false);
+    }
     for (auto f : dirs)
+    {
         func(f, true);
+    }
 }
 
 void ngraph::file_util::iterate_files_worker(
@@ -218,18 +250,23 @@ void ngraph::file_util::iterate_files_worker(
             switch (ent->d_type)
             {
             case DT_DIR:
-                if (recurse && name != "." && name != "..")
+                if (name != "." && name != "..")
                 {
                     string dir_path = file_util::path_join(path, name);
-                    iterate_files(dir_path, func, recurse);
+                    if (recurse)
+                    {
+                        iterate_files(dir_path, func, recurse);
+                    }
                     func(dir_path, true);
                 }
                 break;
             case DT_LNK: break;
             case DT_REG:
-                name = file_util::path_join(path, name);
-                func(name, false);
+            {
+                string file_name = file_util::path_join(path, name);
+                func(file_name, false);
                 break;
+            }
             default: break;
             }
         }

src/ngraph/file_util.hpp

-/*
- Copyright 2016 Nervana Systems Inc.
- 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.
-*/
+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
 
 #pragma once
 
@@ -27,6 +26,8 @@ namespace ngraph
 class ngraph::file_util
 {
 public:
+    static std::string get_file_name(const std::string&);
+    static std::string get_file_ext(const std::string&);
     static std::string path_join(const std::string& s1, const std::string& s2);
     static size_t get_file_size(const std::string& filename);
     static void remove_directory(const std::string& dir);

src/ngraph/runtime/cpu/cpu_kernels.hpp

@@ -14,8 +14,8 @@
 
 #pragma once
 
-#include <cinttypes>
 #include <cstddef>
+#include <cstdint>
 
 // CBLAS types and wrappers
 

src/ngraph/runtime/cpu/cpu_manager.hpp

@@ -16,7 +16,7 @@
 
 #include <memory>
 
-#include "ngraph/codegen/compiler.hpp"
+#include "ngraph/codegen/execution_engine.hpp"
 #include "ngraph/runtime/manager.hpp"
 
 namespace ngraph
@@ -33,7 +33,7 @@ namespace ngraph
             class CPUManager : public Manager
             {
             protected:
-                ngraph::codegen::execution_state exec_state;
+                ngraph::codegen::ExecutionEngine exec_state;
 
             public:
                 virtual std::shared_ptr<Backend> allocate_backend() override;

src/ngraph/runtime/cpu/eigen_utils.hpp

@@ -89,7 +89,7 @@ namespace ngraph
                     };
                 }
 
-                // ET element type
+                // T element type
                 // FMT array format (fmt::V for vector, etc.)
                 // BASE select array/matrix
                 template <typename T,
@@ -128,4 +128,4 @@ namespace ngraph
             }
         }
     }
-}
+}
\ No newline at end of file

src/ngraph/runtime/cpu/emitter.cpp

@@ -66,20 +66,12 @@ void Emitter::EmitAdd(const ngraph::Node* n,
                       const std::vector<TensorViewInfo>& inputs,
                       const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-    string type = et.c_type_string();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << type << ">(" << outputs[0].get_tensor().get_name() << ", "
-       << eigen_vector_format(outputs[0]) << ") =\n";
+    TU << emit_array1d(outputs[0]) << " = \n";
     TU.indent++;
-    TU << "EigenArray1d<" << type << ">(" << inputs[0].get_tensor().get_name() << ", "
-       << eigen_vector_format(inputs[0]) << ") +\n";
-    TU << "EigenArray1d<" << type << ">(" << inputs[1].get_tensor().get_name() << ", "
-       << eigen_vector_format(inputs[1]) << ");\n";
+    TU << emit_array1d(inputs[0]) << " +\n ";
+    TU << emit_array1d(inputs[1]) << ";\n";
     TU.indent -= 2;
     TU << "}\n";
 }
@@ -111,11 +103,8 @@ void Emitter::EmitDot(const ngraph::Node* n,
 
         TU << "{   // " << n->get_name() << "\n";
         TU.indent++;
-        TU << "EigenVector<" << arg0_element_type.c_type_string() << ">("
-           << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-           << ")\n    = " << first.get_tensor().get_name() << "[0]\n    * EigenVector<"
-           << arg0_element_type.c_type_string() << ">(" << second.get_tensor().get_name() << ", "
-           << eigen_vector_format(second) << ");\n";
+        TU << "" << emit_vector(outputs[0]) << "\n    = ";
+        TU << first.get_tensor().get_name() << "[0]\n    * " << emit_vector(second) << ";\n";
         TU.indent--;
         TU << "}\n";
     }
@@ -123,15 +112,9 @@ void Emitter::EmitDot(const ngraph::Node* n,
     {
         TU << "{   // " << n->get_name() << "\n";
         TU.indent++;
-        TU << "EigenVector<" << arg0_element_type.c_type_string() << ">("
-           << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-           << ") << \n"
-           << "    EigenVector<" << arg0_element_type.c_type_string() << ">("
-           << inputs[0].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0])
-           << ").dot("
-           << "EigenVector<" << arg0_element_type.c_type_string() << ">("
-           << inputs[1].get_tensor().get_name() << ", " << eigen_vector_format(inputs[1])
-           << "));\n";
+        TU << "" << emit_vector(outputs[0]) << " << \n"
+           << "    " << emit_vector(inputs[0]) << ".dot("
+           << "" << emit_vector(inputs[1]) << ");\n";
         TU.indent--;
         TU << "}\n";
     }
@@ -141,14 +124,9 @@ void Emitter::EmitDot(const ngraph::Node* n,
 
         TU << "{   // " << n->get_name() << "\n";
         TU.indent++;
-        TU << "EigenVector<" << arg0_element_type.c_type_string() << ">("
-           << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-           << ") = \n"
-           << "    EigenMatrix<" << arg0_element_type.c_type_string() << ">("
-           << inputs[0].get_tensor().get_name() << ", "
-           << eigen_matrix_format(arg0_layout->get_shape(), arg0_layout->get_strides()) << ") * "
-           << "EigenVector<" << arg0_element_type.c_type_string() << ">("
-           << inputs[1].get_tensor().get_name() << ", " << eigen_vector_format(inputs[1]) << ");\n";
+        TU << "" << emit_vector(outputs[0]) << " = \n"
+           << "    " << emit_matrix(inputs[0]) << " * "
+           << "" << emit_vector(inputs[1]) << ";\n";
         TU.indent--;
         TU << "}\n";
     }
@@ -179,18 +157,9 @@ void Emitter::EmitDot(const ngraph::Node* n,
         {
             TU << "{   // " << n->get_name() << "\n";
             TU.indent++;
-            TU << "EigenMatrix<" << arg0_element_type.c_type_string() << ">("
-               << outputs[0].get_tensor().get_name() << ", "
-               << eigen_matrix_format(out_layout->get_shape(), out_layout->get_strides())
-               << ") = \n"
-               << "    EigenMatrix<" << arg0_element_type.c_type_string() << ">("
-               << inputs[0].get_tensor().get_name() << ", "
-               << eigen_matrix_format(arg0_layout->get_shape(), arg0_layout->get_strides())
-               << ") * "
-               << "EigenMatrix<" << arg0_element_type.c_type_string() << ">("
-               << inputs[1].get_tensor().get_name() << ", "
-               << eigen_matrix_format(arg1_layout->get_shape(), arg1_layout->get_strides())
-               << ");\n";
+            TU << "" << emit_matrix(outputs[0]) << " = \n"
+               << "    " << emit_matrix(inputs[0]) << " * "
+               << "" << emit_matrix(inputs[1]) << ";\n";
             TU.indent--;
             TU << "}\n";
         }
@@ -213,12 +182,9 @@ void Emitter::EmitMultiply(const ngraph::Node* n,
 
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << type << ">(" << outputs[0].get_tensor().get_name() << ", "
-       << eigen_vector_format(outputs[0]) << ") =\n"
-       << "   EigenArray1d<" << type << ">(" << inputs[0].get_tensor().get_name() << ", "
-       << eigen_vector_format(inputs[0]) << ") *\n"
-       << "   EigenArray1d<" << type << ">(" << inputs[1].get_tensor().get_name() << ", "
-       << eigen_vector_format(inputs[1]) << ");\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "   " << emit_array1d(inputs[0]) << " *\n"
+       << "   " << emit_array1d(inputs[1]) << ";\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -276,16 +242,10 @@ void Emitter::EmitAbs(const ngraph::Node* n,
                       const std::vector<TensorViewInfo>& inputs,
                       const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n";
-    TU << "Eigen::abs(EigenArray1d<" << et.c_type_string() << ">("
-       << inputs[0].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0]) << "));\n";
+    TU << emit_array1d(outputs[0]) << " =\n";
+    TU << "Eigen::abs(" << emit_array1d(inputs[0]) << ");\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -299,24 +259,19 @@ void Emitter::EmitConcat(const ngraph::Node* n,
     assert(result_tensor_type);
 
     auto result_shape = result_tensor_type->get_shape();
-    auto& result_element_type = result_tensor_type->get_element_type();
 
     if (result_shape.size() == 1)
     {
         TU << "{   // " << n->get_name() << "\n";
         TU.indent++;
-        TU << "EigenVector<" << result_element_type.c_type_string() << "> out_vector("
-           << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-           << ");\n";
+        TU << "" << emit_vector(outputs[0], "out_vector") << ";\n";
 
         size_t concat_pos = 0;
         for (size_t i = 0; i < inputs.size(); i++)
         {
             TU << "out_vector.segment(" << concat_pos << ", "
                << inputs[i].get_tensor_view_layout()->get_shape().at(0) << ") << "
-               << "EigenVector<" << result_element_type.c_type_string() << ">("
-               << inputs[i].get_tensor().get_name() << ", " << eigen_vector_format(inputs[i])
-               << ");\n";
+               << "" << emit_vector(inputs[i]) << ";\n";
             concat_pos += inputs[i].get_tensor_view_layout()->get_shape().at(0);
         }
         TU.indent--;
@@ -329,9 +284,7 @@ void Emitter::EmitConcat(const ngraph::Node* n,
 
         TU << "{   // " << n->get_name() << "\n";
         TU.indent++;
-        TU << "EigenMatrix<" << result_element_type.c_type_string() << "> out_matrix("
-           << outputs[0].get_tensor().get_name() << ", "
-           << eigen_matrix_format(out_layout->get_shape(), out_layout->get_strides()) << ");\n";
+        TU << "" << emit_matrix(outputs[0], "out_matrix") << ";\n";
 
         size_t concat_pos[2]{0, 0};
         for (size_t i = 0; i < inputs.size(); i++)
@@ -341,9 +294,7 @@ void Emitter::EmitConcat(const ngraph::Node* n,
 
             TU << "out_matrix.block(" << concat_pos[0] << ", " << concat_pos[1] << ", "
                << arg_shape.at(0) << ", " << arg_shape.at(1) << ") << "
-               << "EigenMatrix<" << result_element_type.c_type_string() << ">("
-               << inputs[i].get_tensor().get_name() << ", "
-               << eigen_matrix_format(arg_layout->get_shape(), arg_layout->get_strides()) << ");\n";
+               << "" << emit_matrix(inputs[i]) << ";\n";
 
             concat_pos[axis] += arg_shape.at(axis);
         }
@@ -358,18 +309,11 @@ void Emitter::EmitDivide(const ngraph::Node* n,
                          const std::vector<TensorViewInfo>& inputs,
                          const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ") /\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[1].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[1]) << ");\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << " /\n"
+       << "    " << emit_array1d(inputs[1]) << ";\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -379,18 +323,11 @@ void Emitter::EmitEqual(const ngraph::Node* n,
                         const std::vector<TensorViewInfo>& inputs,
                         const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << element::Bool::element_type().c_type_string() << ">("
-       << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    (EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ") ==\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[1].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[1]) << ")).template cast<char>();\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    (" << emit_array1d(inputs[0]) << " ==\n"
+       << "    " << emit_array1d(inputs[1]) << ").template cast<char>();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -400,18 +337,11 @@ void Emitter::EmitGreater(const ngraph::Node* n,
                           const std::vector<TensorViewInfo>& inputs,
                           const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
-    TU << "{   // " << n->get_name() << "\n";
+    TU << "{   // " << n->get_name() << " xxx\n";
     TU.indent++;
-    TU << "EigenArray1d<" << element::Bool::element_type().c_type_string() << ">("
-       << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    (EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ") >\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[1].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[1]) << ")).template cast<char>();\n";
+    TU << "" << emit_array1d(outputs[0]) << " =\n"
+       << "    (" << emit_array1d(inputs[0]) << " >\n"
+       << "    " << emit_array1d(inputs[1]) << ").template cast<char>();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -421,18 +351,11 @@ void Emitter::EmitGreaterEq(const ngraph::Node* n,
                             const std::vector<TensorViewInfo>& inputs,
                             const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << element::Bool::element_type().c_type_string() << ">("
-       << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    (EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ") >=\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[1].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[1]) << ")).template cast<char>();\n";
+    TU << "" << emit_array1d(outputs[0]) << " =\n"
+       << "    (" << emit_array1d(inputs[0]) << " >=\n"
+       << "    " << emit_array1d(inputs[1]) << ").template cast<char>();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -442,18 +365,11 @@ void Emitter::EmitLess(const ngraph::Node* n,
                        const std::vector<TensorViewInfo>& inputs,
                        const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << element::Bool::element_type().c_type_string() << ">("
-       << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    (EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ") <\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[1].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[1]) << ")).template cast<char>();\n";
+    TU << "" << emit_array1d(outputs[0]) << " =\n"
+       << "    (" << emit_array1d(inputs[0]) << " <\n"
+       << "    " << emit_array1d(inputs[1]) << ").template cast<char>();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -463,18 +379,11 @@ void Emitter::EmitLessEq(const ngraph::Node* n,
                          const std::vector<TensorViewInfo>& inputs,
                          const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << element::Bool::element_type().c_type_string() << ">("
-       << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    (EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ") <=\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[1].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[1]) << ")).template cast<char>();\n";
+    TU << "" << emit_array1d(outputs[0]) << " =\n"
+       << "    (" << emit_array1d(inputs[0]) << " <=\n"
+       << "    " << emit_array1d(inputs[1]) << ").template cast<char>();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -484,16 +393,10 @@ void Emitter::EmitLog(const ngraph::Node* n,
                       const std::vector<TensorViewInfo>& inputs,
                       const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    Eigen::log(EigenArray1d<" << et.c_type_string() << ">("
-       << inputs[0].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0]) << "));\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    Eigen::log(" << emit_array1d(inputs[0]) << ");\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -503,18 +406,11 @@ void Emitter::EmitMaximum(const ngraph::Node* n,
                           const std::vector<TensorViewInfo>& inputs,
                           const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "        EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").max(\n"
-       << "        EigenArray1d<" << et.c_type_string() << ">(" << inputs[1].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[1]) << "));\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "        " << emit_array1d(inputs[0]) << ".max(\n"
+       << "        " << emit_array1d(inputs[1]) << ");\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -524,18 +420,11 @@ void Emitter::EmitMinimum(const ngraph::Node* n,
                           const std::vector<TensorViewInfo>& inputs,
                           const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").min(\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[1].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[1]) << "));\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << ".min(\n"
+       << "    " << emit_array1d(inputs[1]) << ");\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -545,16 +434,10 @@ void Emitter::EmitNegative(const ngraph::Node* n,
                            const std::vector<TensorViewInfo>& inputs,
                            const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    -EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ");\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    -" << emit_array1d(inputs[0]) << ";\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -564,18 +447,11 @@ void Emitter::EmitNotEqual(const ngraph::Node* n,
                            const std::vector<TensorViewInfo>& inputs,
                            const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << element::Bool::element_type().c_type_string() << ">("
-       << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    (EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ") !=\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[1].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[1]) << ")).template cast<char>();\n";
+    TU << "" << emit_array1d(outputs[0]) << " =\n"
+       << "    (" << emit_array1d(inputs[0]) << " !=\n"
+       << "    " << emit_array1d(inputs[1]) << ").template cast<char>();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -585,20 +461,12 @@ void Emitter::EmitSelect(const ngraph::Node* n,
                          const std::vector<TensorViewInfo>& inputs,
                          const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(1)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "   EigenArray1d<" << element::Bool::element_type().c_type_string() << ">("
-       << inputs[0].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0]) << ")\n"
-       << "    .select(EigenArray1d<" << et.c_type_string() << ">("
-       << inputs[1].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0]) << "),\n"
-       << "       EigenArray1d<" << et.c_type_string() << ">(" << inputs[2].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[1]) << "));\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "   " << emit_array1d(inputs[0]) << "\n"
+       << "    .select(" << emit_array1d(inputs[1]) << ",\n"
+       << "       " << emit_array1d(inputs[2]) << ");\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -608,18 +476,11 @@ void Emitter::EmitSubtract(const ngraph::Node* n,
                            const std::vector<TensorViewInfo>& inputs,
                            const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ") -\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[1].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[1]) << ");\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << " -\n"
+       << "    " << emit_array1d(inputs[1]) << ";\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -808,7 +669,6 @@ void Emitter::EmitBroadcast(const ngraph::Node* n,
 
     auto arg_shape = arg_tensor_type->get_shape();
     auto result_shape = result_tensor_type->get_shape();
-    auto& result_element_type = result_tensor_type->get_element_type();
 
     if (broadcast->get_broadcast_axes().empty())
     {
@@ -826,12 +686,8 @@ void Emitter::EmitBroadcast(const ngraph::Node* n,
     {
         TU << "{   // " << n->get_name() << "\n";
         TU.indent++;
-        TU << "EigenArray1d<" << result_element_type.c_type_string() << ">("
-           << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-           << ") =\n"
-           << "    EigenArray1d<" << result_element_type.c_type_string() << ">("
-           << inputs[0].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0])
-           << ")(0, 0);\n";
+        TU << "" << emit_array1d(outputs[0]) << " =\n"
+           << "    " << emit_array1d(inputs[0]) << "(0, 0);\n";
         TU.indent--;
         TU << "}\n";
     }
@@ -843,13 +699,8 @@ void Emitter::EmitBroadcast(const ngraph::Node* n,
 
             TU << "{   // " << n->get_name() << "\n";
             TU.indent++;
-            TU << "EigenMatrix<" << result_element_type.c_type_string() << ">("
-               << outputs[0].get_tensor().get_name() << ", "
-               << eigen_matrix_format(out_layout->get_shape(), out_layout->get_strides())
-               << ").colwise() =\n"
-               << "    EigenVector<" << result_element_type.c_type_string() << ">("
-               << inputs[0].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0])
-               << ");\n";
+            TU << "" << emit_matrix(outputs[0]) << ".colwise() =\n"
+               << "    " << emit_vector(inputs[0]) << ";\n";
             TU.indent--;
             TU << "}\n";
         }
@@ -859,13 +710,8 @@ void Emitter::EmitBroadcast(const ngraph::Node* n,
 
             TU << "{   // " << n->get_name() << "\n";
             TU.indent++;
-            TU << "EigenMatrix<" << result_element_type.c_type_string() << ">("
-               << outputs[0].get_tensor().get_name() << ", "
-               << eigen_matrix_format(out_layout->get_shape(), out_layout->get_strides())
-               << ").rowwise() =\n"
-               << "    EigenVector<" << result_element_type.c_type_string() << ">("
-               << inputs[0].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0])
-               << ").transpose();\n";
+            TU << "" << emit_matrix(outputs[0]) << ".rowwise() =\n"
+               << "    " << emit_vector(inputs[0]) << ".transpose();\n";
             TU.indent--;
             TU << "}\n";
         }
@@ -892,8 +738,6 @@ void Emitter::EmitConvert(const ngraph::Node* n,
     auto arg_tensor_type = dynamic_pointer_cast<const TensorViewType>(arg->get_value_type());
     assert(arg_tensor_type);
 
-    auto& arg_element_type = arg_tensor_type->get_element_type();
-
     auto result_tensor_type = dynamic_pointer_cast<const TensorViewType>(n->get_value_type());
     assert(result_tensor_type);
 
@@ -901,10 +745,8 @@ void Emitter::EmitConvert(const ngraph::Node* n,
 
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << result_element_type.c_type_string() << ">("
-       << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << arg_element_type.c_type_string() << ">("
-       << inputs[0].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0]) << ")\n"
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << "\n"
        << "    .template cast<" << result_element_type.c_type_string() << ">();\n";
     TU.indent--;
     TU << "}\n";
@@ -961,7 +803,8 @@ void Emitter::EmitReshape(const ngraph::Node* n,
         result_shape_product *= i;
     }
 
-    // If there is no layout change or we are just going from 1^n to 1^m or a zero-size tensor, we can just copy.
+    // If there is no layout change or we are just going from 1^n to 1^m or a zero-size tensor,
+    //  we can just copy.
     if (same_layout || result_shape_product < 2)
     {
         TU << "{   // " << n->get_name() << " 1\n";
@@ -1000,13 +843,8 @@ void Emitter::EmitReshape(const ngraph::Node* n,
         {
             TU << "{   // " << n->get_name() << " 3\n";
             TU.indent++;
-            TU << "EigenMatrix<" << result_element_type.c_type_string() << ">("
-               << outputs[0].get_tensor().get_name() << ", "
-               << eigen_matrix_format(out_layout->get_shape(), out_layout->get_strides()) << ") =\n"
-               << "        EigenMatrix<" << result_element_type.c_type_string() << ">("
-               << inputs[0].get_tensor().get_name() << ", "
-               << eigen_matrix_format(arg0_layout->get_shape(), arg0_layout->get_strides())
-               << ").transpose();\n";
+            TU << "" << emit_matrix(outputs[0]) << " =\n"
+               << "        " << emit_matrix(inputs[0]) << ".transpose();\n";
             TU.indent--;
             TU << "}\n";
         }
@@ -1029,19 +867,7 @@ void Emitter::EmitFunctionCall(const ngraph::Node* n,
 
     TU << "{   // Call " << function->get_name() << "\n";
     TU.indent++;
-    TU << "std::vector<void*> inputs;\n";
-    for (const TensorViewInfo& input : inputs)
-    {
-        TU << "inputs.push_back(" << input.get_tensor().get_name() << ");\n";
-    }
-    TU << "\n";
-    TU << "std::vector<void*> outputs;\n";
-    for (const TensorViewInfo& output : outputs)
-    {
-        TU << "outputs.push_back(" << output.get_tensor().get_name() << ");\n";
-    }
-    TU << "\n";
-    TU << function->get_name() << "(inputs, outputs);\n";
+    generate_call(inputs, outputs, function);
     TU.indent--;
     TU << "}\n";
 }
@@ -1137,24 +963,18 @@ void Emitter::EmitReduce(const ngraph::Node* n,
             TU << "{   // " << n->get_name() << " 3\n";
             TU.indent++;
             string type = f_result_element_type.c_type_string();
-            TU << "auto f = [](" << type << " x, " << type << " y) -> " << type << " {\n";
+            TU << "auto f = [](" << type << " x, " << type << " y) -> " << type << "\n{";
             TU.indent++;
-            TU << "std::vector<void*> inputs;\n";
-            TU << "inputs.push_back(&x);\n";
-            TU << "inputs.push_back(&y);\n\n";
+            TU << "\n";
             TU << type << " result;\n";
-            TU << "std::vector<void*> outputs;\n";
-            TU << "outputs.push_back(&result);\n";
+            TU << "std::vector<void*> inputs = {&x, &y};\n";
+            TU << "std::vector<void*> outputs = {&result};\n";
             TU << reduction_function->get_name() << "(inputs, outputs);\n";
             TU << "return result;\n";
             TU.indent--;
             TU << "};\n";
-            TU << "EigenArray1d<" << f_result_element_type.c_type_string() << ">("
-               << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-               << ") =\n"
-               << "    EigenArray1d<" << f_result_element_type.c_type_string() << ">("
-               << inputs[0].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0])
-               << ").redux(f);\n";
+            TU << "" << emit_array1d(outputs[0]) << " =\n"
+               << "    " << emit_array1d(inputs[0]) << ".redux(f);\n";
             TU.indent--;
             TU << "}\n";
         }
@@ -1165,12 +985,8 @@ void Emitter::EmitReduce(const ngraph::Node* n,
         {
             TU << "{   // " << n->get_name() << " 4\n";
             TU.indent++;
-            TU << "EigenArray1d<" << f_result_element_type.c_type_string() << ">("
-               << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-               << ") =\n"
-               << "    EigenArray1d<" << f_result_element_type.c_type_string() << ">("
-               << inputs[1].get_tensor().get_name() << ", " << eigen_vector_format(inputs[1])
-               << ")(0, 0);\n";
+            TU << "" << emit_array1d(outputs[0]) << " =\n"
+               << "    " << emit_array1d(inputs[1]) << "(0, 0);\n";
             TU.indent--;
             TU << "}\n";
         }
@@ -1183,25 +999,18 @@ void Emitter::EmitReduce(const ngraph::Node* n,
             TU << "{   // " << n->get_name() << " 5\n";
             TU.indent++;
             string type = f_result_element_type.c_type_string();
-            TU << "auto f = [](" << type << " x, " << type << " y) -> " << type << " {\n";
+            TU << "auto f = [](" << type << " x, " << type << " y) -> " << type << "\n{";
             TU.indent++;
-            TU << "std::vector<void*> inputs;\n";
-            TU << "inputs.push_back(&x);\n";
-            TU << "inputs.push_back(&y);\n\n";
+            TU << "\n";
             TU << type << " result;\n";
-            TU << "std::vector<void*> outputs;\n";
-            TU << "outputs.push_back(&result);\n";
+            TU << "std::vector<void*> inputs = {&x, &y};\n";
+            TU << "std::vector<void*> outputs = {&result};\n";
             TU << reduction_function->get_name() << "(inputs, outputs);\n";
             TU << "return result;\n";
             TU.indent--;
             TU << "};\n";
-            TU << "EigenVector<" << f_result_element_type.c_type_string() << ">("
-               << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-               << ") =\n"
-               << "        EigenMatrix<" << f_result_element_type.c_type_string() << ">("
-               << inputs[0].get_tensor().get_name() << ", "
-               << eigen_matrix_format(arg0_layout->get_shape(), arg0_layout->get_strides())
-               << ").rowwise().redux(f);\n";
+            TU << "" << emit_vector(outputs[0]) << " =\n"
+               << "        " << emit_matrix(inputs[0]) << ".rowwise().redux(f);\n";
             TU.indent--;
             TU << "}\n";
         }
@@ -1212,12 +1021,8 @@ void Emitter::EmitReduce(const ngraph::Node* n,
         {
             TU << "{   // " << n->get_name() << " 6\n";
             TU.indent++;
-            TU << "EigenArray1d<" << f_result_element_type.c_type_string() << ">("
-               << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-               << ") =\n"
-               << "    EigenArray1d<" << f_result_element_type.c_type_string() << ">("
-               << inputs[1].get_tensor().get_name() << ", " << eigen_vector_format(inputs[1])
-               << ")(0, 0);\n";
+            TU << "" << emit_array1d(outputs[0]) << " =\n"
+               << "    " << emit_array1d(inputs[1]) << "(0, 0);\n";
             TU.indent--;
             TU << "}\n";
         }
@@ -1226,25 +1031,18 @@ void Emitter::EmitReduce(const ngraph::Node* n,
             TU << "{   // " << n->get_name() << " 7\n";
             TU.indent++;
             string type = f_result_element_type.c_type_string();
-            TU << "auto f = [](" << type << " x, " << type << " y) -> " << type << " {\n";
+            TU << "auto f = [](" << type << " x, " << type << " y) -> " << type << "\n{";
             TU.indent++;
-            TU << "std::vector<void*> inputs;\n";
-            TU << "inputs.push_back(&x);\n";
-            TU << "inputs.push_back(&y);\n\n";
+            TU << "\n";
             TU << type << " result;\n";
-            TU << "std::vector<void*> outputs;\n";
-            TU << "outputs.push_back(&result);\n";
+            TU << "std::vector<void*> inputs = {&x, &y};\n";
+            TU << "std::vector<void*> outputs = {&result};\n";
             TU << reduction_function->get_name() << "(inputs, outputs);\n";
             TU << "return result;\n";
             TU.indent--;
             TU << "};\n";
-            TU << "EigenVector<" << f_result_element_type.c_type_string() << ">("
-               << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-               << ") =\n"
-               << "    EigenMatrix<" << f_result_element_type.c_type_string() << ">("
-               << inputs[0].get_tensor().get_name() << ", "
-               << eigen_matrix_format(arg0_layout->get_shape(), arg0_layout->get_strides())
-               << ").colwise().redux(f);\n";
+            TU << "" << emit_vector(outputs[0]) << " =\n"
+               << "    " << emit_matrix(inputs[0]) << ".colwise().redux(f);\n";
             TU.indent--;
             TU << "}\n";
         }
@@ -1260,16 +1058,10 @@ void Emitter::EmitSign(const ngraph::Node* n,
                        const std::vector<TensorViewInfo>& inputs,
                        const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").sign();\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << ".sign();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -1294,7 +1086,6 @@ void Emitter::EmitSlice(const ngraph::Node* n,
     assert(arg_tensor_view_type);
     auto arg_shape = arg_tensor_view_type->get_shape();
     auto arg_rank = arg_shape.size();
-    auto& arg_element_type = arg_tensor_view_type->get_element_type();
 
     auto& lower_bounds = slice->get_lower_bounds();
     auto& upper_bounds = slice->get_upper_bounds();
@@ -1316,12 +1107,8 @@ void Emitter::EmitSlice(const ngraph::Node* n,
     {
         TU << "{   // " << n->get_name() << " 2\n";
         TU.indent++;
-        TU << "EigenVector<" << arg_element_type.c_type_string() << ">("
-           << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-           << ") =\n"
-           << "    EigenVector<" << arg_element_type.c_type_string() << ">("
-           << inputs[0].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0])
-           << ").segment(\n"
+        TU << "" << emit_vector(outputs[0]) << " =\n"
+           << "    " << emit_vector(inputs[0]) << ".segment(\n"
            << "        " << to_string(lower_bounds[0]) << ", "
            << to_string(upper_bounds[0] - lower_bounds[0]) << ");\n";
         TU.indent--;
@@ -1334,14 +1121,9 @@ void Emitter::EmitSlice(const ngraph::Node* n,
 
         TU << "{   // " << n->get_name() << " 3\n";
         TU.indent++;
-        TU << "EigenMatrix<" << arg_element_type.c_type_string() << ">("
-           << outputs[0].get_tensor().get_name() << ", "
-           << eigen_matrix_format(out_layout->get_shape(), out_layout->get_strides()) << ") = \n"
-           << "        EigenMatrix<" << arg_element_type.c_type_string() << ">("
-           << inputs[0].get_tensor().get_name() << ", "
-           << eigen_matrix_format(arg0_layout->get_shape(), arg0_layout->get_strides())
-           << ").block(" << to_string(lower_bounds[0]) << ", " << to_string(lower_bounds[1])
-           << ",\n"
+        TU << "" << emit_matrix(outputs[0]) << " = \n"
+           << "        " << emit_matrix(inputs[0]) << ".block(" << to_string(lower_bounds[0])
+           << ", " << to_string(lower_bounds[1]) << ",\n"
            << "        " << to_string(upper_bounds[0] - lower_bounds[0]) << ",\n"
            << "        " << to_string(upper_bounds[1] - lower_bounds[1]) << ");\n";
         TU.indent--;
@@ -1362,7 +1144,6 @@ void Emitter::EmitSum(const ngraph::Node* n,
     auto s = static_cast<const op::Sum*>(n);
     auto s_tensor_view_type = dynamic_pointer_cast<const TensorViewType>(s->get_value_type());
     assert(s_tensor_view_type);
-    auto& s_element_type = s_tensor_view_type->get_element_type();
     auto s_shape = s_tensor_view_type->get_shape();
 
     auto arg = s->get_arguments().at(0);
@@ -1393,12 +1174,8 @@ void Emitter::EmitSum(const ngraph::Node* n,
     {
         TU << "{   // " << n->get_name() << "\n";
         TU.indent++;
-        TU << "EigenArray1d<" << s_element_type.c_type_string() << ">("
-           << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-           << ") =\n"
-           << "    EigenArray1d<" << s_element_type.c_type_string() << ">("
-           << inputs[0].get_tensor().get_name() << ", " << eigen_vector_format(inputs[0])
-           << ").sum();\n";
+        TU << "" << emit_array1d(outputs[0]) << " =\n"
+           << "    " << emit_array1d(inputs[0]) << ".sum();\n";
         TU.indent--;
         TU << "}\n";
     }
@@ -1408,13 +1185,8 @@ void Emitter::EmitSum(const ngraph::Node* n,
 
         TU << "{   // " << n->get_name() << "\n";
         TU.indent++;
-        TU << "EigenVector<" << s_element_type.c_type_string() << ">("
-           << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-           << ") =\n"
-           << "    EigenMatrix<" << s_element_type.c_type_string() << ">("
-           << inputs[0].get_tensor().get_name() << ", "
-           << eigen_matrix_format(arg0_layout->get_shape(), arg0_layout->get_strides())
-           << ").rowwise().sum();\n";
+        TU << "" << emit_vector(outputs[0]) << " =\n"
+           << "    " << emit_matrix(inputs[0]) << ".rowwise().sum();\n";
         TU.indent--;
         TU << "}\n";
     }
@@ -1424,13 +1196,8 @@ void Emitter::EmitSum(const ngraph::Node* n,
 
         TU << "{   // " << n->get_name() << "\n";
         TU.indent++;
-        TU << "EigenVector<" << s_element_type.c_type_string() << ">("
-           << outputs[0].get_tensor().get_name() << ", " << eigen_vector_format(outputs[0])
-           << ") =\n"
-           << "    EigenMatrix<" << s_element_type.c_type_string() << ">("
-           << inputs[0].get_tensor().get_name() << ", "
-           << eigen_matrix_format(arg0_layout->get_shape(), arg0_layout->get_strides())
-           << ").colwise().sum();\n";
+        TU << "" << emit_vector(outputs[0]) << " =\n"
+           << "    " << emit_matrix(inputs[0]) << ".colwise().sum();\n";
         TU.indent--;
         TU << "}\n";
     }
@@ -1445,16 +1212,10 @@ void Emitter::EmitExp(const ngraph::Node* n,
                       const std::vector<TensorViewInfo>& inputs,
                       const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").exp();\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << ".exp();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -1464,16 +1225,10 @@ void Emitter::EmitSin(const ngraph::Node* n,
                       const std::vector<TensorViewInfo>& inputs,
                       const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").sin();\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << ".sin();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -1483,16 +1238,10 @@ void Emitter::EmitSinh(const ngraph::Node* n,
                        const std::vector<TensorViewInfo>& inputs,
                        const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").sinh();\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << ".sinh();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -1502,16 +1251,10 @@ void Emitter::EmitCos(const ngraph::Node* n,
                       const std::vector<TensorViewInfo>& inputs,
                       const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").cos();\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << ".cos();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -1521,16 +1264,10 @@ void Emitter::EmitCosh(const ngraph::Node* n,
                        const std::vector<TensorViewInfo>& inputs,
                        const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").cosh();\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << ".cosh();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -1540,16 +1277,10 @@ void Emitter::EmitTan(const ngraph::Node* n,
                       const std::vector<TensorViewInfo>& inputs,
                       const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").tan();\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << ".tan();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -1579,16 +1310,10 @@ void Emitter::EmitAsin(const ngraph::Node* n,
                        const std::vector<TensorViewInfo>& inputs,
                        const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").asin();\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << ".asin();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -1598,16 +1323,10 @@ void Emitter::EmitAcos(const ngraph::Node* n,
                        const std::vector<TensorViewInfo>& inputs,
                        const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").acos();\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << ".acos();\n";
     TU.indent--;
     TU << "}\n";
 }
@@ -1617,16 +1336,89 @@ void Emitter::EmitAtan(const ngraph::Node* n,
                        const std::vector<TensorViewInfo>& inputs,
                        const std::vector<TensorViewInfo>& outputs)
 {
-    const element::Type& et =
-        (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
-            ->get_element_type();
-
     TU << "{   // " << n->get_name() << "\n";
     TU.indent++;
-    TU << "EigenArray1d<" << et.c_type_string() << ">(" << outputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(outputs[0]) << ") =\n"
-       << "    EigenArray1d<" << et.c_type_string() << ">(" << inputs[0].get_tensor().get_name()
-       << ", " << eigen_vector_format(inputs[0]) << ").atan();\n";
+    TU << emit_array1d(outputs[0]) << " =\n"
+       << "    " << emit_array1d(inputs[0]) << ".atan();\n";
     TU.indent--;
     TU << "}\n";
 }
+
+//------------------------------------------------------------------------------------------------
+// Utility methods
+//------------------------------------------------------------------------------------------------
+
+void Emitter::generate_call(const std::vector<TensorViewInfo>& inputs,
+                            const std::vector<TensorViewInfo>& outputs,
+                            shared_ptr<Function> function)
+{
+    vector<string> input_names;
+    vector<string> output_names;
+
+    for (const TensorViewInfo& input : inputs)
+    {
+        input_names.push_back(input.get_tensor().get_name());
+    }
+
+    for (const TensorViewInfo& output : outputs)
+    {
+        output_names.push_back(output.get_tensor().get_name());
+    }
+
+    TU << "std::vector<void*> inputs =\n{";
+    TU.indent++;
+    TU << "\n" << join(input_names, ",\n");
+    TU.indent--;
+    TU << "\n};\n";
+
+    TU << "std::vector<void*> outputs =\n{";
+    TU.indent++;
+    TU << "\n" << join(output_names, ",\n");
+    TU.indent--;
+    TU << "\n};\n";
+
+    TU << "\n";
+    TU << function->get_name() << "(inputs, outputs);\n";
+}
+
+static string format_name(const string& name)
+{
+    string rc;
+    if (name.empty())
+    {
+        rc = " " + name;
+    }
+    return rc;
+}
+
+string Emitter::emit_vector(const TensorViewInfo& tvi, const string& name)
+{
+    stringstream ss;
+
+    const element::Type& et = tvi.get_tensor_view()->get_value_type()->get_element_type();
+    ss << "EigenVector<" << et.c_type_string() << ">" << format_name(name) << "("
+       << tvi.get_tensor().get_name() << ", " << eigen_vector_format(tvi) << ")";
+    return ss.str();
+}
+
+string Emitter::emit_array1d(const TensorViewInfo& tvi, const string& name)
+{
+    stringstream ss;
+
+    const element::Type& et = tvi.get_tensor_view()->get_value_type()->get_element_type();
+    ss << "EigenArray1d<" << et.c_type_string() << ">" << format_name(name) << "("
+       << tvi.get_tensor().get_name() << ", " << eigen_vector_format(tvi) << ")";
+    return ss.str();
+}
+
+string Emitter::emit_matrix(const TensorViewInfo& tvi, const string& name)
+{
+    stringstream ss;
+    auto layout = tvi.get_layout<DenseTensorViewLayout>();
+
+    const element::Type& et = tvi.get_tensor_view()->get_value_type()->get_element_type();
+    ss << "EigenMatrix<" << et.c_type_string() << ">" << format_name(name) << "("
+       << tvi.get_tensor().get_name() << ", "
+       << eigen_matrix_format(layout->get_shape(), layout->get_strides()) << ")";
+    return ss.str();
+}

src/ngraph/runtime/cpu/emitter.hpp

@@ -94,6 +94,15 @@ namespace ngraph
                 void EMITTER_DECL(EmitAsin);
                 void EMITTER_DECL(EmitAcos);
                 void EMITTER_DECL(EmitAtan);
+
+            private:
+                void generate_call(const std::vector<TensorViewInfo>& inputs,
+                                   const std::vector<TensorViewInfo>& outputs,
+                                   std::shared_ptr<Function> function);
+
+                std::string emit_vector(const TensorViewInfo&, const std::string& name = "");
+                std::string emit_array1d(const TensorViewInfo&, const std::string& name = "");
+                std::string emit_matrix(const TensorViewInfo&, const std::string& name = "");
             };
         }
     }

src/ngraph/runtime/cpu/external_function.cpp

@@ -22,6 +22,7 @@
 
 #include "ngraph/codegen/code_writer.hpp"
 #include "ngraph/codegen/compiler.hpp"
+#include "ngraph/codegen/execution_engine.hpp"
 #include "ngraph/descriptor/input.hpp"
 #include "ngraph/descriptor/layout/dense_tensor_view_layout.hpp"
 #include "ngraph/descriptor/output.hpp"
@@ -74,6 +75,7 @@
 #include "ngraph/pass/memory_layout.hpp"
 #include "ngraph/pass/topological_sort.hpp"
 #include "ngraph/runtime/cpu/call_frame.hpp"
+#include "ngraph/runtime/cpu/cpu_backend.hpp"
 #include "ngraph/runtime/cpu/emitter.hpp"
 #include "ngraph/runtime/cpu/external_function.hpp"
 #include "ngraph/runtime/utils.hpp"
@@ -81,26 +83,17 @@
 using namespace std;
 using namespace ngraph::runtime::cpu;
 
-using ngraph::descriptor::layout::DenseTensorViewLayout;
+static const std::string s_output_dir = "cpu_codegen";
 
-extern "C" void
-    allocate_aligned_buffer(size_t size, size_t alignment, char** allocated, char** aligned_ptr)
+class StaticInitializers
 {
-    size_t allocation_size = size + alignment;
-    *allocated = static_cast<char*>(malloc(allocation_size));
-    *aligned_ptr = *allocated;
-    size_t mod = size_t(*aligned_ptr) % alignment;
+public:
+    StaticInitializers() { ngraph::file_util::remove_directory(s_output_dir); }
+};
 
-    if (mod != 0)
-    {
-        (*aligned_ptr) += (alignment - mod);
-    }
-}
+static StaticInitializers s_static_initializers;
 
-extern "C" void free_aligned_buffer(void* allocated)
-{
-    free(allocated);
-}
+using ngraph::descriptor::layout::DenseTensorViewLayout;
 
 #define TI(x) type_index(typeid(x))
 
@@ -192,37 +185,25 @@ void ExternalFunction::compile()
 
     TU +=
         R"(// Generated by the NGraph CPU backend
-#include <algorithm>
 #include <cmath>
-#include <memory>
 #include <vector>
 
 #include <Eigen/Dense>
 
 #include "ngraph/runtime/cpu/cpu_kernels.hpp"
 #include "ngraph/runtime/cpu/eigen_utils.hpp"
+#include "ngraph/runtime/cpu/memory_handler.hpp"
 
 using namespace ngraph::runtime::cpu::eigen;
 
-extern "C" void allocate_aligned_buffer(
-    size_t size,
-    size_t alignment,
-    char** allocated,
-    char** aligned_ptr);
-
-extern "C" void free_aligned_buffer(void* allocated);
-
 )";
 
-    TU << "// Declare any functions that are not main\n";
+    TU << "// Declare all functions\n";
     for (shared_ptr<Function> f : pass_manager.get_state().get_functions())
     {
-        if (f != m_function)
-        {
-            TU << "extern \"C\" void " << f->get_name() << "(\n";
-            TU << "    const std::vector<void*>& inputs,\n";
-            TU << "    const std::vector<void*>& outputs);\n";
-        }
+        TU << "extern \"C\" void " << f->get_name() << "(\n";
+        TU << "    const std::vector<void*>& inputs,\n";
+        TU << "    const std::vector<void*>& outputs);\n";
     }
     TU << "\n";
 
@@ -232,27 +213,37 @@ extern "C" void free_aligned_buffer(void* allocated);
         TU << "    const std::vector<void*>& inputs,\n";
         TU << "    const std::vector<void*>& outputs)\n";
         TU << "{\n";
-
         TU.indent++;
-        TU << "// Allocate the memory pool\n";
-        size_t temp_pool_size = pass_manager.get_state().get_temporary_pool_size();
-        TU << "char* allocated_buffer_pool;\n";
-        TU << "char* aligned_buffer_pool;\n";
-        TU << "allocate_aligned_buffer(" << temp_pool_size << ", 64"
-           << ", &allocated_buffer_pool, &aligned_buffer_pool);\n";
-        TU << "\n";
 
-        TU << "// Define temporary tensors\n";
+        bool temporaries_used = false;
         for (shared_ptr<Node> node : current_function->get_ordered_ops())
         {
-            for (descriptor::Tensor* tensor : node->liveness_new_list)
+            if (node->liveness_new_list.size() > 0)
             {
-                TU << tensor->get_element_type() << "* " << tensor->get_name() << " = ("
-                   << tensor->get_element_type() << "*)(aligned_buffer_pool + "
-                   << tensor->get_pool_offset() << ");\n";
+                temporaries_used = true;
+                break;
             }
         }
-        TU << "\n";
+        if (temporaries_used)
+        {
+            size_t temp_pool_size = pass_manager.get_state().get_temporary_pool_size();
+            TU << "// Allocate the memory pool\n";
+            TU << "ngraph::runtime::cpu::MemoryHandler memory_handler(" << temp_pool_size << ", "
+               << ngraph::runtime::cpu::alignment << ");\n";
+            TU << "\n";
+
+            TU << "// Define temporary tensors\n";
+            for (shared_ptr<Node> node : current_function->get_ordered_ops())
+            {
+                for (descriptor::Tensor* tensor : node->liveness_new_list)
+                {
+                    TU << tensor->get_element_type() << "* " << tensor->get_name() << " = ("
+                       << tensor->get_element_type() << "*)(memory_handler.get_ptr("
+                       << tensor->get_pool_offset() << "));\n";
+                }
+            }
+            TU << "\n";
+        }
 
         TU << "// Define inputs\n";
         size_t arg_index = 0;
@@ -283,13 +274,10 @@ extern "C" void free_aligned_buffer(void* allocated);
         }
         TU << "\n";
 
-        TU << "// Define tensor views\n";
-        TU << "\n";
-
         for (shared_ptr<Node> node : current_function->get_ordered_ops())
         {
             auto& n = *node; // Work around a compiler warning (*node inside typeid may have effects
-                             // with shared pointers, which is fine here but clang doesn't like it.)
+            // with shared pointers, which is fine here but clang doesn't like it.)
             auto handler = dispatcher.find(type_index(typeid(n)));
             if (handler == dispatcher.end())
             {
@@ -311,8 +299,6 @@ extern "C" void free_aligned_buffer(void* allocated);
             handler->second(&emitter, node.get(), this, in, out);
         }
 
-        TU << "\nfree_aligned_buffer(allocated_buffer_pool);\n";
-
         TU.indent--;
 
         // End TU
@@ -321,34 +307,33 @@ extern "C" void free_aligned_buffer(void* allocated);
 
     // TODO: Cleanup and make this a utility function
 
-    string output_dir = "cpu_codegen";
     string function_name = m_function->get_name();
-    file_util::remove_directory(output_dir);
-    file_util::make_directory(output_dir);
-    string filename = file_util::path_join(output_dir, function_name + "_codegen.cpp");
+    file_util::make_directory(s_output_dir);
+    string filename = file_util::path_join(s_output_dir, function_name + "_codegen.cpp");
     ofstream out(filename);
     string code = TU.get_code();
     out << code;
     out.close();
 
-    ngraph::codegen::execution_state estate;
+    codegen::Compiler compiler;
+    codegen::ExecutionEngine execution_engine;
 
 #if NGCPU_PCH
-    estate.set_precompiled_headers_enabled(true);
+    compiler.set_precompiled_headers_enabled(true);
 #endif
 
 #if NGCPU_DEBUGINFO
-    estate.set_debuginfo_enabled(true);
+    compiler.set_debuginfo_enabled(true);
 #endif
 
-    auto llvm_module = estate.compile(code, function_name + "_codegen.cpp");
+    auto llvm_module = compiler.compile(code);
     if (llvm_module == nullptr)
     {
         throw runtime_error("function failed to compile");
     }
-    estate.add_module(llvm_module);
-    estate.finalize();
-    m_compiled_function = estate.find_function<EntryPoint_t>(function_name);
+    execution_engine.add_module(llvm_module);
+    execution_engine.finalize();
+    m_compiled_function = execution_engine.find_function<EntryPoint_t>(function_name);
     assert(m_compiled_function);
 
     m_is_compiled = true;

src/ngraph/runtime/cpu/memory_handler.cpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#include "ngraph/runtime/cpu/memory_handler.hpp"
+
+using namespace ngraph;
+
+runtime::cpu::MemoryHandler::MemoryHandler(size_t byte_size, size_t alignment)
+    : m_allocated_buffer_pool(nullptr)
+    , m_aligned_buffer_pool(nullptr)
+{
+    if (byte_size > 0)
+    {
+        size_t allocation_size = byte_size + alignment;
+        m_allocated_buffer_pool = static_cast<char*>(malloc(allocation_size));
+        m_aligned_buffer_pool = m_allocated_buffer_pool;
+        size_t mod = size_t(m_aligned_buffer_pool) % alignment;
+
+        if (mod != 0)
+        {
+            m_aligned_buffer_pool += (alignment - mod);
+        }
+    }
+}
+
+runtime::cpu::MemoryHandler::~MemoryHandler()
+{
+    if (m_allocated_buffer_pool != nullptr)
+    {
+        free(m_allocated_buffer_pool);
+    }
+}

src/ngraph/runtime/cpu/memory_handler.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include <cstddef>
+#include <memory>
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace cpu
+        {
+            class MemoryHandler;
+        }
+    }
+}
+
+class ngraph::runtime::cpu::MemoryHandler
+{
+public:
+    MemoryHandler(size_t pool_size, size_t alignment);
+    ~MemoryHandler();
+
+    void* get_ptr(size_t offset) const { return m_aligned_buffer_pool + offset; }
+private:
+    char* m_allocated_buffer_pool;
+    char* m_aligned_buffer_pool;
+};

src/ngraph/runtime/cpu/tensor_view.cpp

@@ -20,35 +20,47 @@
 using namespace ngraph;
 using namespace std;
 
-extern "C" void
-    allocate_aligned_buffer(size_t size, size_t alignment, char** allocated, char** aligned_ptr);
-extern "C" void free_aligned_buffer(void* allocated);
-
 runtime::cpu::CPUTensorView::CPUTensorView(const ngraph::element::Type& element_type,
                                            const Shape& shape)
     : runtime::TensorView(std::make_shared<ngraph::descriptor::PrimaryTensorView>(
           std::make_shared<ngraph::TensorViewType>(element_type, shape), "external", true, true))
+    , m_allocated_buffer_pool(nullptr)
+    , m_aligned_buffer_pool(nullptr)
+
 {
     m_descriptor->set_tensor_view_layout(
         std::make_shared<ngraph::descriptor::layout::DenseTensorViewLayout>(*m_descriptor));
 
     m_buffer_size = m_descriptor->get_tensor_view_layout()->get_size() * element_type.size();
-    allocate_aligned_buffer(m_buffer_size, runtime::cpu::alignment, &m_allocated, &m_buffer);
+    if (m_buffer_size > 0)
+    {
+        size_t allocation_size = m_buffer_size + runtime::cpu::alignment;
+        m_allocated_buffer_pool = static_cast<char*>(malloc(allocation_size));
+        m_aligned_buffer_pool = m_allocated_buffer_pool;
+        size_t mod = size_t(m_aligned_buffer_pool) % alignment;
+        if (mod != 0)
+        {
+            m_aligned_buffer_pool += (alignment - mod);
+        }
+    }
 }
 
 runtime::cpu::CPUTensorView::~CPUTensorView()
 {
-    free_aligned_buffer(m_allocated);
+    if (m_allocated_buffer_pool != nullptr)
+    {
+        free(m_allocated_buffer_pool);
+    }
 }
 
 char* runtime::cpu::CPUTensorView::get_data_ptr()
 {
-    return m_buffer;
+    return m_aligned_buffer_pool;
 }
 
 const char* runtime::cpu::CPUTensorView::get_data_ptr() const
 {
-    return m_buffer;
+    return m_aligned_buffer_pool;
 }
 
 void runtime::cpu::CPUTensorView::write(const void* source, size_t tensor_offset, size_t n)

src/ngraph/runtime/cpu/tensor_view.hpp

@@ -52,7 +52,7 @@ public:
     void read(void* p, size_t tensor_offset, size_t n) const override;
 
 private:
-    char* m_allocated;
-    char* m_buffer;
+    char* m_allocated_buffer_pool;
+    char* m_aligned_buffer_pool;
     size_t m_buffer_size;
 };

test/codegen.cpp

@@ -19,23 +19,26 @@
 #include "gtest/gtest.h"
 
 #include "ngraph/codegen/compiler.hpp"
+#include "ngraph/codegen/execution_engine.hpp"
 
 using namespace std;
+using namespace ngraph;
 
 TEST(codegen, simple_return)
 {
-    constexpr auto name = "test.cpp";
     constexpr auto source = R"(extern "C" int test() { return 2+5; })";
 
-    ngraph::codegen::execution_state estate;
-    auto module = estate.compile(source, name);
+    codegen::Compiler compiler;
+    codegen::ExecutionEngine execution_engine;
+
+    auto module = compiler.compile(source);
     ASSERT_NE(nullptr, module);
 
-    estate.add_module(module);
+    execution_engine.add_module(module);
 
-    estate.finalize();
+    execution_engine.finalize();
 
-    auto func = estate.find_function<int()>("test");
+    auto func = execution_engine.find_function<int()>("test");
     ASSERT_NE(nullptr, func);
 
     int result = func();
@@ -44,18 +47,19 @@ TEST(codegen, simple_return)
 
 TEST(codegen, pass_args)
 {
-    constexpr auto name = "test.cpp";
     constexpr auto source = R"(extern "C" int test(int a, int b) { return a+b; })";
 
-    ngraph::codegen::execution_state estate;
-    auto module = estate.compile(source, name);
+    codegen::Compiler compiler;
+    codegen::ExecutionEngine execution_engine;
+
+    auto module = compiler.compile(source);
     ASSERT_NE(nullptr, module);
 
-    estate.add_module(module);
+    execution_engine.add_module(module);
 
-    estate.finalize();
+    execution_engine.finalize();
 
-    auto func = estate.find_function<int(int, int)>("test");
+    auto func = execution_engine.find_function<int(int, int)>("test");
     ASSERT_NE(nullptr, func);
 
     int result = func(20, 22);
@@ -64,7 +68,6 @@ TEST(codegen, pass_args)
 
 TEST(codegen, include)
 {
-    constexpr auto name = "test.cpp";
     constexpr auto source =
         R"(
         #include <cmath>
@@ -74,15 +77,17 @@ TEST(codegen, include)
         }
     )";
 
-    ngraph::codegen::execution_state estate;
-    auto module = estate.compile(source, name);
+    codegen::Compiler compiler;
+    codegen::ExecutionEngine execution_engine;
+
+    auto module = compiler.compile(source);
     ASSERT_NE(nullptr, module);
 
-    estate.add_module(module);
+    execution_engine.add_module(module);
 
-    estate.finalize();
+    execution_engine.finalize();
 
-    auto func = estate.find_function<int(int, int)>("test");
+    auto func = execution_engine.find_function<int(int, int)>("test");
     ASSERT_NE(nullptr, func);
 
     int result = func(20, 2);