Jmenon/eigen opt (#326)

* CPU: Optimize Eigen based rowwise vector broadcast

* CPU: Remove the need for transposing the broadcast vector

* CPU: Optimize to a replicate expression

* CPU: Change code model to medium and compile for the host CPU
instead of hardcoding BDW

cmake/external_eigen.cmake

@@ -16,27 +16,27 @@ include(ExternalProject)
 
 set(EIGEN_INSTALL_DIR ${EXTERNAL_INSTALL_DIR}/eigen)
 set(EIGEN_PROJECT eigen)
-set(EIGEN_SHA1_HASH dd238ca6c6b5d2ce2e7e2e9ded4c59bad77ce6d0)
-set(EIGEN_URL http://bitbucket.org/eigen/eigen/get/3.3.3.zip)
+set(EIGEN_GIT_TAG d608d9f3f577118981acbdd40da9dcf6b514668a)
+set(EIGEN_GIT_URL https://github.com/jmenon/eigen)
 
 #----------------------------------------------------------------------------------------------------------
-# Download and install GoogleTest ...
+# Download and install Eigen
 #----------------------------------------------------------------------------------------------------------
 
 # The 'BUILD_BYPRODUCTS' argument was introduced in CMake 3.2.
 if (${CMAKE_VERSION} VERSION_LESS 3.2)
     ExternalProject_Add(
         ${EIGEN_PROJECT}
-        URL ${EIGEN_URL}
-        URL_HASH SHA1=${EIGEN_SHA1_HASH}
+        GIT_REPOSITORY ${EIGEN_GIT_URL}
+        GIT_TAG ${EIGEN_GIT_TAG}
         UPDATE_COMMAND ""
         CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${EIGEN_INSTALL_DIR}
         )
 else()
     ExternalProject_Add(
         ${EIGEN_PROJECT}
-        URL ${EIGEN_URL}
-        URL_HASH SHA1=${EIGEN_SHA1_HASH}
+        GIT_REPOSITORY ${EIGEN_GIT_URL}
+        GIT_TAG ${EIGEN_GIT_TAG}
         UPDATE_COMMAND ""
         CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${EIGEN_INSTALL_DIR}
         BUILD_BYPRODUCTS "${EIGEN_INSTALL_DIR}/include/eigen3"

src/ngraph/codegen/compiler.cpp

@@ -111,6 +111,12 @@ StaticCompiler::StaticCompiler()
     vector<const char*> args;
     args.push_back(m_source_name.c_str());
 
+    // Inlining thresholds are forced to a very high value
+    // to ensure all Eigen code gets properly inlined
+    // This is for both Eigen strong and weak inlines
+    args.push_back("-mllvm");
+    args.push_back("-inline-threshold=1000000");
+
     // Prepare DiagnosticEngine
     IntrusiveRefCntPtr<DiagnosticOptions> DiagOpts = new DiagnosticOptions();
     TextDiagnosticPrinter* textDiagPrinter = new clang::TextDiagnosticPrinter(errs(), &*DiagOpts);
@@ -203,6 +209,7 @@ StaticCompiler::StaticCompiler()
     auto& CGO = m_compiler->getInvocation().getCodeGenOpts();
     CGO.OptimizationLevel = 3;
     CGO.RelocationModel = "static";
+    CGO.CodeModel = "medium";
     CGO.ThreadModel = "posix";
     CGO.FloatABI = "hard";
     CGO.OmitLeafFramePointer = 1;
@@ -218,8 +225,8 @@ StaticCompiler::StaticCompiler()
     // Enable various target features
     // Most of these are for Eigen
     auto& TO = m_compiler->getInvocation().getTargetOpts();
-    // TODO: This needs to be configurable and selected carefully
-    TO.CPU = "broadwell";
+
+    TO.CPU = sys::getHostCPUName();
     TO.FeaturesAsWritten.emplace_back("+sse");
     TO.FeaturesAsWritten.emplace_back("+sse2");
     TO.FeaturesAsWritten.emplace_back("+sse3");

src/ngraph/codegen/execution_engine.cpp

@@ -41,6 +41,8 @@ bool codegen::ExecutionEngine::add_module(std::unique_ptr<llvm::Module>& module)
             m_execution_engine.reset(llvm::EngineBuilder(move(module))
                                          .setEngineKind(llvm::EngineKind::JIT)
                                          .setOptLevel(llvm::CodeGenOpt::Aggressive)
+                                         .setMCPU(llvm::sys::getHostCPUName())
+                                         .setCodeModel(llvm::CodeModel::Medium)
                                          .setErrorStr(&m_jit_error)
                                          .create());
 

src/ngraph/runtime/cpu/cpu_emitter.cpp

@@ -523,8 +523,17 @@ void runtime::cpu::CPU_Emitter::EmitBroadcast(const ngraph::Node* n,
         {
             m_out << "{   // " << n->get_name() << "\n";
             m_out.indent++;
-            m_out << emit_matrix(out[0]) << ".rowwise() =\n"
-                  << "    " << emit_vector(args[0]) << ".transpose();\n";
+
+            m_out << "Eigen::Map<Eigen::Matrix<" << out[0].get_element_type().c_type_string()
+                  << ", " << join(out[0].get_shape())
+                  << ", Eigen::RowMajor>, Eigen::Aligned64, Eigen::Stride<"
+                  << join(out[0].get_strides()) << ">> out(" << out[0].get_name() << ");\n";
+            m_out << "Eigen::Map<Eigen::Matrix<" << args[0].get_element_type().c_type_string()
+                  << ", 1, " << args[0].get_size()
+                  << ", Eigen::RowMajor>, Eigen::Aligned64, Eigen::Stride<" << args[0].get_size()
+                  << ", 1>> arg0(" << args[0].get_name() << ");\n";
+            m_out << "out = arg0.replicate<" << out[0].get_shape().at(0) << ", 1>();\n";
+
             m_out.indent--;
             m_out << "}\n";
         }