Allow building nGraph CPU without TBB - r0.25.1 (#3652)

* Allow TBB to be excluded from CPU build.

* Change some CMake scope.

* m_use_tbb can be removed if TBB is not enabled.

* Fix some rebase errors

* Some files do not get definition.

* CPU codegen: Use different pregenerated code if TBB is not built.

* Fix codegen issues.

* Fix bug.

* Use NGRAPH_TBB_ENABLE only for CPU.

CMakeLists.txt

@@ -189,6 +189,9 @@ option(NGRAPH_STATIC_LIB_ENABLE "Enable build nGraph static library" FALSE)
 option(NGRAPH_INTERPRETER_STATIC_LIB_ENABLE "Enable build INTERPRETER backend static library" FALSE)
 option(NGRAPH_CPU_STATIC_LIB_ENABLE "Enable build CPU backend static library" FALSE)
 option(NGRAPH_DYNAMIC_COMPONENTS_ENABLE "Enable dynamic loading of components" TRUE)
+if (NGRAPH_CPU_ENABLE)
+    option(NGRAPH_TBB_ENABLE "Control usage of TBB for CPU backend" TRUE)
+endif()
 
 if (NGRAPH_CPU_ENABLE
         AND
@@ -268,6 +271,9 @@ NORMALIZE_BOOL(NGRAPH_STATIC_LIB_ENABLE)
 NORMALIZE_BOOL(NGRAPH_INTERPRETER_STATIC_LIB_ENABLE)
 NORMALIZE_BOOL(NGRAPH_CPU_STATIC_LIB_ENABLE)
 NORMALIZE_BOOL(NGRAPH_DYNAMIC_COMPONENTS_ENABLE)
+if (NGRAPH_CPU_ENABLE)
+    NORMALIZE_BOOL(NGRAPH_TBB_ENABLE)
+endif()
 
 # Turn off GPU build in 0.25.1
 set(NGRAPH_GPU_ENABLE FALSE)
@@ -298,6 +304,9 @@ message(STATUS "NGRAPH_STATIC_LIB_ENABLE:             ${NGRAPH_STATIC_LIB_ENABLE
 message(STATUS "NGRAPH_INTERPRETER_STATIC_LIB_ENABLE: ${NGRAPH_INTERPRETER_STATIC_LIB_ENABLE}")
 message(STATUS "NGRAPH_CPU_STATIC_LIB_ENABLE:         ${NGRAPH_CPU_STATIC_LIB_ENABLE}")
 message(STATUS "NGRAPH_DYNAMIC_COMPONENTS_ENABLE:     ${NGRAPH_DYNAMIC_COMPONENTS_ENABLE}")
+if (NGRAPH_CPU_ENABLE)
+    message(STATUS "NGRAPH_TBB_ENABLE:                    ${NGRAPH_TBB_ENABLE}")
+endif()
 
 #-----------------------------------------------------------------------------------------------
 # Installation logic...
@@ -433,10 +442,6 @@ if (NGRAPH_DISTRIBUTED_ENABLE)
     endif()
 endif()
 
-if (NOT DEFINED NGRAPH_TBB_ENABLE)
-    set(NGRAPH_TBB_ENABLE ${NGRAPH_CPU_ENABLE})
-endif()
-
 # Since UNIX and APPLE support Bash we can use a Bash script to do the clang-format functions
 # This is much faster than the cmake method
 if (UNIX OR APPLE)

src/ngraph/codegen/CMakeLists.txt

@@ -60,6 +60,7 @@ endif()
 if(NGRAPH_TBB_ENABLE)
     get_target_property(TBB_INCLUDE_DIR libtbb INTERFACE_INCLUDE_DIRECTORIES)
     list(APPEND HEADER_SEARCH_DEFINES TBB_HEADERS_PATH="${TBB_INCLUDE_DIR}")
+    set_source_files_properties(compiler.cpp PROPERTIES COMPILE_DEFINITIONS "NGRAPH_TBB_ENABLE")
 endif()
 
 set_source_files_properties(compiler.cpp PROPERTIES COMPILE_DEFINITIONS "${HEADER_SEARCH_DEFINES}")

src/ngraph/codegen/compiler.cpp

@@ -190,6 +190,11 @@ void codegen::CompilerCore::initialize()
     // Prevent Eigen from using any LGPL3 code
     args.push_back("-DEIGEN_MPL2_ONLY");
 
+#if defined(NGRAPH_TBB_ENABLE)
+    // Enable TBB
+    args.push_back("-DNGRAPH_TBB_ENABLE");
+#endif
+
     // Prepare DiagnosticEngine
     IntrusiveRefCntPtr<DiagnosticOptions> diag_options = new DiagnosticOptions();
     diag_options->ErrorLimit = 20;

src/ngraph/runtime/cpu/CMakeLists.txt

@@ -178,8 +178,7 @@ if (NGRAPH_CPU_ENABLE)
         target_compile_definitions(cpu_backend PRIVATE "NGRAPH_ENABLE_CPU_CONV_AUTO")
     endif()
     if(NGRAPH_TBB_ENABLE)
-        set_source_files_properties(cpu_external_function.cpp
-            PROPERTIES COMPILE_DEFINITIONS "NGRAPH_TBB_ENABLE")
+        target_compile_definitions(cpu_backend PRIVATE "NGRAPH_TBB_ENABLE")
     endif()
     if (NGRAPH_HALIDE)
         target_compile_definitions(cpu_backend PRIVATE "NGRAPH_HALIDE")
@@ -210,14 +209,16 @@ if (NGRAPH_CPU_ENABLE)
     target_compile_definitions(cpu_backend PRIVATE CPU_BACKEND_DLL_EXPORTS)
 
     add_dependencies(cpu_backend libmkldnn ext_eigen)
-    target_link_libraries(cpu_backend PUBLIC ngraph libmkldnn libmkl libeigen libtbb)
+    target_link_libraries(cpu_backend PUBLIC ngraph libmkldnn libmkl libeigen)
     if (NGRAPH_JSON_ENABLE)
         target_link_libraries(cpu_backend PUBLIC libjson)
     endif()
+    if (NGRAPH_TBB_ENABLE)
+        target_link_libraries(cpu_backend PUBLIC libtbb)
+    endif()
     if (NOT NGRAPH_DEX_ONLY)
         target_link_libraries(cpu_backend PUBLIC codegen)
     endif()
-    target_include_directories(cpu_backend SYSTEM PUBLIC libmkldnn)
 
     if (NOT APPLE AND NOT MSVS)
         # CPU backend uses third-party libraries like Eigen that might be linked in and

src/ngraph/runtime/cpu/cpu_backend.cpp

@@ -14,7 +14,9 @@
 // limitations under the License.
 //*****************************************************************************
 
+#if defined(NGRAPH_TBB_ENABLE)
 #include <tbb/tbb_stddef.h>
+#endif
 
 #include "cpu_backend_visibility.h"
 #include "ngraph/graph_util.hpp"
@@ -40,8 +42,10 @@ runtime::BackendConstructor* runtime::cpu::get_backend_constructor_pointer()
     public:
         std::shared_ptr<runtime::Backend> create(const std::string& /* config */) override
         {
+#if defined(NGRAPH_TBB_ENABLE)
             // Force TBB to link to the backend
             tbb::TBB_runtime_interface_version();
+#endif
             return make_shared<runtime::cpu::CPU_Backend>();
         }
     };

src/ngraph/runtime/cpu/cpu_call_frame.cpp

@@ -221,6 +221,7 @@ void runtime::cpu::CPU_CallFrame::setup_runtime_context(Allocator* allocator)
 
         ctx->states = m_external_function->m_states.data();
 
+#if defined(NGRAPH_TBB_ENABLE)
         if (m_external_function->is_direct_execution() &&
             std::getenv("NGRAPH_CPU_USE_TBB") != nullptr)
         {
@@ -232,6 +233,7 @@ void runtime::cpu::CPU_CallFrame::setup_runtime_context(Allocator* allocator)
             ctx->c =
                 new tbb::global_control(tbb::global_control::max_allowed_parallelism, parallelism);
         }
+#endif
     }
     m_num_ctx_available = m_num_ctx;
 }
@@ -253,6 +255,7 @@ void runtime::cpu::CPU_CallFrame::cleanup_runtime_context()
         {
             delete buffer;
         }
+#if defined(NGRAPH_TBB_ENABLE)
         if (m_external_function->is_direct_execution() &&
             std::getenv("NGRAPH_CPU_USE_TBB") != nullptr)
         {
@@ -273,6 +276,7 @@ void runtime::cpu::CPU_CallFrame::cleanup_runtime_context()
             }
             delete ctx->c;
         }
+#endif
         delete ctx;
     }
     m_num_ctx_available = 0;

src/ngraph/runtime/cpu/cpu_executor.cpp

@@ -109,10 +109,13 @@ namespace ngraph
                         m_thread_pool_devices.push_back(
                             std::unique_ptr<Eigen::ThreadPoolDevice>(new Eigen::ThreadPoolDevice(
                                 m_thread_pools[i].get(), num_threads_per_pool)));
+#if defined(NGRAPH_TBB_ENABLE)
                         m_tbb_arenas.emplace_back(1);
+#endif
                     }
                 }
 
+#if defined(NGRAPH_TBB_ENABLE)
                 void CPUExecutor::execute(CPUKernelFunctor& f,
                                           CPURuntimeContext* ctx,
                                           CPUExecutionContext* ectx,
@@ -128,6 +131,14 @@ namespace ngraph
                         f(ctx, ectx);
                     }
                 }
+#else
+                void CPUExecutor::execute(CPUKernelFunctor& f,
+                                          CPURuntimeContext* ctx,
+                                          CPUExecutionContext* ectx)
+                {
+                    f(ctx, ectx);
+                }
+#endif
 
                 CPUExecutor& GetCPUExecutor()
                 {

src/ngraph/runtime/cpu/cpu_executor.hpp

@@ -26,7 +26,9 @@
 #define EIGEN_USE_THREADS
 #include <unsupported/Eigen/CXX11/Tensor>
 
+#if defined(NGRAPH_TBB_ENABLE)
 #include "tbb/task_arena.h"
+#endif
 
 namespace ngraph
 {
@@ -49,16 +51,24 @@ namespace ngraph
                         return *m_thread_pool_devices[id].get();
                     }
 
+#if defined(NGRAPH_TBB_ENABLE)
                     void execute(CPUKernelFunctor& f,
                                  CPURuntimeContext* ctx,
                                  CPUExecutionContext* ectx,
                                  bool use_tbb = false);
+#else
+                    void execute(CPUKernelFunctor& f,
+                                 CPURuntimeContext* ctx,
+                                 CPUExecutionContext* ectx);
+#endif
                     int get_num_thread_pools() { return m_num_thread_pools; }
                     int get_num_cores() { return m_num_cores; }
                 private:
                     std::vector<std::unique_ptr<Eigen::ThreadPool>> m_thread_pools;
                     std::vector<std::unique_ptr<Eigen::ThreadPoolDevice>> m_thread_pool_devices;
+#if defined(NGRAPH_TBB_ENABLE)
                     std::vector<tbb::task_arena> m_tbb_arenas;
+#endif
                     int m_num_thread_pools;
                     int m_num_cores;
                 };

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -23,9 +23,11 @@
 #include <typeinfo>
 #include <unordered_map>
 
+#if defined(NGRAPH_TBB_ENABLE)
 #define TBB_PREVIEW_FLOW_GRAPH_TRACE 1
 
 #include <tbb/flow_graph.h>
+#endif
 
 #if !defined(NGRAPH_DEX_ONLY)
 #include "ngraph/code_writer.hpp"
@@ -236,7 +238,9 @@ runtime::cpu::CPU_ExternalFunction::CPU_ExternalFunction(
     : m_function(function)
     , m_release_function(release_function)
     , m_emit_timing(false)
+#if defined(NGRAPH_TBB_ENABLE)
     , m_use_tbb(std::getenv("NGRAPH_CPU_USE_TBB") != nullptr)
+#endif
 #if !defined(NGRAPH_DEX_ONLY)
     , m_is_compiled(false)
     , m_direct_execution((std::getenv("NGRAPH_CODEGEN") == nullptr) ||
@@ -508,6 +512,7 @@ void runtime::cpu::CPU_ExternalFunction::compile(ngraph::pass::PassConfig& pass_
     CodeWriter writer;
 
     writer << "// Generated by the nGraph CPU backend\n";
+#if defined(NGRAPH_TBB_ENABLE)
     if (m_use_tbb)
     {
         if (runtime::cpu::IsTracingEnabled() || m_emit_timing)
@@ -518,6 +523,8 @@ void runtime::cpu::CPU_ExternalFunction::compile(ngraph::pass::PassConfig& pass_
         }
         writer << "#include <tbb/flow_graph.h>";
     }
+#endif
+
     writer +=
         R"(
 #include <cmath>
@@ -769,6 +776,7 @@ using namespace ngraph::runtime;
 
     writer << "bool* t_en = (bool*)" << m_function->get_name() << "_t_en;\n";
 
+#if defined(NGRAPH_TBB_ENABLE)
     if (m_use_tbb)
     {
         writer << "\n";
@@ -779,6 +787,7 @@ using namespace ngraph::runtime;
                << " = new tbb::flow::continue_node<tbb::flow::continue_msg> "
                   "(*(cg_ctx->tbb_graph), [&](const tbb::flow::continue_msg &msg)\n{});\n";
     }
+#endif
 
     // Add inputs to the variable name map
     size_t arg_index = 0;
@@ -887,6 +896,7 @@ using namespace ngraph::runtime;
                                         t_out_attrs,
                                         t_in_attrs);
             }
+#if defined(NGRAPH_TBB_ENABLE)
             if (m_use_tbb)
             {
                 writer << "tbb::flow::continue_node<tbb::flow::continue_msg>* "
@@ -896,6 +906,7 @@ using namespace ngraph::runtime;
                           "(*(cg_ctx->tbb_graph), [&](const tbb::flow::continue_msg &msg)\n{\n";
                 writer.indent++;
             }
+#endif
             if (runtime::cpu::IsTracingEnabled() && m_function->get_name() == m_function_name)
             {
                 writer << "start_ts = cpu::Clock::now();\n";
@@ -1015,14 +1026,17 @@ using namespace ngraph::runtime;
                        << "(std::chrono::duration_cast<cpu::Timescale>(cpu::Clock::now() - "
                           "start_ts)).count();\n";
             }
+#if defined(NGRAPH_TBB_ENABLE)
             if (m_use_tbb)
             {
                 writer.indent--;
                 writer << "});\n";
             }
+#endif
         }
     }
 
+#if defined(NGRAPH_TBB_ENABLE)
     if (m_use_tbb)
     {
         writer << "\n";
@@ -1058,6 +1072,7 @@ using namespace ngraph::runtime;
                << "->try_put(tbb::flow::continue_msg());\n";
         writer << "try { cg_ctx->tbb_graph->wait_for_all(); } catch(...) { throw; }\n";
     }
+#endif
     writer << "ctx->first_iteration = false;\n";
 
     writer.indent--;
@@ -1350,12 +1365,14 @@ void runtime::cpu::CPU_ExternalFunction::build(ngraph::pass::PassConfig& pass_co
         return;
     }
 
+#if defined(NGRAPH_TBB_ENABLE)
     if (m_use_tbb && (runtime::cpu::IsTracingEnabled() || m_emit_timing))
     {
         throw ngraph_error(
             "CPU Backend: Tracing and performance breakdowns might not be accurate with TBB "
             "enabled due to concurrent graph execution");
     }
+#endif
 
 // reference all the builders for static library
 #ifdef NGRAPH_CPU_STATIC_LIB_ENABLE
@@ -1717,6 +1734,7 @@ void runtime::cpu::CPU_ExternalFunction::build(ngraph::pass::PassConfig& pass_co
         }
 
         auto functor = functors.begin();
+#if defined(NGRAPH_TBB_ENABLE)
         if (m_use_tbb)
         {
             // Build the flow graph
@@ -1824,6 +1842,7 @@ void runtime::cpu::CPU_ExternalFunction::build(ngraph::pass::PassConfig& pass_co
             }
         }
         else
+#endif
         {
             static const auto ddebug = std::getenv("NGRAPH_DEX_DEBUG");
             if (ddebug != nullptr)
@@ -1930,7 +1949,11 @@ void runtime::cpu::CPU_ExternalFunction::build(ngraph::pass::PassConfig& pass_co
 
     m_is_built = true;
 
+#if defined(NGRAPH_TBB_ENABLE)
     if (m_release_function && !m_use_tbb)
+#else
+    if (m_release_function)
+#endif
     {
         release_function();
     }

src/ngraph/runtime/cpu/cpu_external_function.hpp

@@ -262,7 +262,9 @@ namespace ngraph
                 bool m_release_function;
                 bool m_emit_timing;
 
+#if defined(NGRAPH_TBB_ENABLE)
                 bool m_use_tbb;
+#endif
 #if !defined(NGRAPH_DEX_ONLY)
                 bool m_is_compiled;
 #endif

src/ngraph/runtime/cpu/cpu_runtime_context.hpp

@@ -20,11 +20,13 @@
 #include <cstdint>
 #include <set>
 
+#if defined(NGRAPH_TBB_ENABLE)
 #define TBB_PREVIEW_GLOBAL_CONTROL 1
 #define TBB_PREVIEW_FLOW_GRAPH_TRACE 1
 #include <tbb/flow_graph.h>
 #include <tbb/global_control.h>
 #include <tbb/task_scheduler_init.h>
+#endif
 #include "ngraph/op/experimental/compiled_kernel.hpp"
 
 #ifdef NGRAPH_MLIR_ENABLE
@@ -66,8 +68,10 @@ namespace ngraph
                 std::vector<mkldnn::primitive*> mkldnn_primitives;
                 std::vector<AlignedBuffer*> memory_buffers;
                 std::vector<char*> mkldnn_workspaces;
+#if defined(NGRAPH_TBB_ENABLE)
                 tbb::flow::graph* G;
                 tbb::global_control* c;
+#endif
                 State* const* states;
                 std::set<size_t> breakpoints;
                 size_t pc;

src/ngraph/runtime/cpu/pregenerated_src/cpu_cg_runtime_context.hpp

@@ -23,11 +23,16 @@
 R"(
 struct CPURuntimeContextCG
 {
+#if defined(NGRAPH_TBB_ENABLE)
     std::unique_ptr<tbb::flow::graph> tbb_graph;
     std::unique_ptr<tbb::global_control> tbb_gcontrol;
 
     CPURuntimeContextCG() { init_tbb(); init_mkldnn_primitives();}
     ~CPURuntimeContextCG() { cleanup_tbb(); cleanup_mkldnn_primitives();}
+#else
+    CPURuntimeContextCG() { init_mkldnn_primitives();}
+    ~CPURuntimeContextCG() { cleanup_mkldnn_primitives();}
+#endif
 
     std::vector<mkldnn::primitive*> mkldnn_primitives;
     std::vector<char*> mkldnn_workspaces;
@@ -57,6 +62,7 @@ struct CPURuntimeContextCG
 
 
 private:
+#if defined(NGRAPH_TBB_ENABLE)
     inline void init_tbb()
     {
         if (std::getenv("NGRAPH_CPU_USE_TBB"))
@@ -86,6 +92,7 @@ private:
             }
         }
     }
+#endif
 
     void init_mkldnn_primitives();
 

test/CMakeLists.txt

@@ -382,7 +382,7 @@ if (NGRAPH_PLAIDML_ENABLE)
 endif()
 
 if (NGRAPH_TBB_ENABLE)
-    target_compile_definitions(unit-test PRIVATE NGRAPH_TBB_ENABLE)
+    target_compile_definitions(unit-test PRIVATE "NGRAPH_TBB_ENABLE")
 endif()
 
 if (NGRAPH_HALIDE)