[FORK][FEATURE] TBB_AUTO was enabled

Author: alexey-varyzgin

cmake/TBB.cmake

@@ -59,7 +59,7 @@ macro(handle_tbb_target)
     add_definitions(-DTBB_PREVIEW_TASK_ARENA_CONSTRAINTS_EXTENSION=1)
 endmacro()
 
-if(NOT DNNL_CPU_THREADING_RUNTIME STREQUAL "TBB")
+if(NOT "${DNNL_CPU_THREADING_RUNTIME}" MATCHES "^(TBB|TBB_AUTO)$")
     return()
 endif()
 

cmake/options.cmake

@@ -262,7 +262,7 @@ set(DNNL_CPU_RUNTIME "OMP" CACHE STRING
     To use Threading Building Blocks (TBB) one should also
     set TBBROOT (either environment variable or CMake option) to the library
     location.")
-if(NOT "${DNNL_CPU_RUNTIME}" MATCHES "^(NONE|OMP|TBB|SEQ|THREADPOOL|DPCPP|SYCL)$")
+if(NOT "${DNNL_CPU_RUNTIME}" MATCHES "^(NONE|OMP|TBB|TBB_AUTO|SEQ|THREADPOOL|DPCPP|SYCL)$")
     message(FATAL_ERROR "Unsupported CPU runtime: ${DNNL_CPU_RUNTIME}")
 endif()
 

include/oneapi/dnnl/dnnl_config.h.in

@@ -70,6 +70,9 @@
 /// TBB runtime (CPU only)
 #define DNNL_RUNTIME_TBB 4u
 
+/// TBB runtime with auto partitioning (CPU only)
+#define DNNL_RUNTIME_TBB_AUTO 5u
+
 /// Threadpool runtime (CPU only)
 #define DNNL_RUNTIME_THREADPOOL 8u
 

src/CMakeLists.txt

@@ -273,7 +273,7 @@ if(UNIX AND DNNL_INSTALL_MODE STREQUAL "BUNDLE_V2")
 endif()
 
 # Install custom find modules for transitive dependencies
-if(DNNL_CPU_THREADING_RUNTIME STREQUAL "TBB")
+if("${DNNL_CPU_THREADING_RUNTIME}" MATCHES "^(TBB|TBB_AUTO)$")
     if(WIN32)
         install(FILES "../cmake/win/TBBConfig.cmake" RENAME "FindTBB.cmake"
             DESTINATION ${LIB_CONFIG_INSTALL_DIR})

src/common/c_types_map.hpp

@@ -1914,6 +1914,7 @@ enum runtime_kind_t {
     dnnl_runtime_seq,
     dnnl_runtime_omp,
     dnnl_runtime_tbb,
+    dnnl_runtime_tbb_auto,
     dnnl_runtime_threadpool,
     dnnl_runtime_ocl,
     dnnl_runtime_sycl,
@@ -1924,6 +1925,7 @@ const runtime_kind_t none = dnnl_runtime_none;
 const runtime_kind_t seq = dnnl_runtime_seq;
 const runtime_kind_t omp = dnnl_runtime_omp;
 const runtime_kind_t tbb = dnnl_runtime_tbb;
+const runtime_kind_t tbb_auto = dnnl_runtime_tbb_auto;
 const runtime_kind_t threadpool = dnnl_runtime_threadpool;
 const runtime_kind_t ocl = dnnl_runtime_ocl;
 const runtime_kind_t sycl = dnnl_runtime_sycl;

src/common/dnnl_debug.cpp

@@ -33,6 +33,7 @@ const char *dnnl_runtime2str(unsigned runtime) {
         case DNNL_RUNTIME_SEQ: return "sequential";
         case DNNL_RUNTIME_OMP: return "OpenMP";
         case DNNL_RUNTIME_TBB: return "TBB";
+        case DNNL_RUNTIME_TBB_AUTO: return "TBB_AUTO";
         case DNNL_RUNTIME_OCL: return "OpenCL";
         case DNNL_RUNTIME_THREADPOOL: return "threadpool";
 #ifdef DNNL_WITH_SYCL

src/common/dnnl_thread.hpp

@@ -75,7 +75,7 @@ inline void dnnl_thr_barrier() {
 #pragma omp barrier
 }
 
-#elif DNNL_CPU_THREADING_RUNTIME == DNNL_RUNTIME_TBB
+#elif (DNNL_CPU_THREADING_RUNTIME == DNNL_RUNTIME_TBB || DNNL_CPU_THREADING_RUNTIME == DNNL_RUNTIME_TBB_AUTO)
 #include "tbb/parallel_for.h"
 #include "tbb/task_arena.h"
 #define DNNL_THR_SYNC 0

src/common/engine.hpp

@@ -187,6 +187,8 @@ inline runtime_kind_t get_default_runtime(engine_kind_t kind) {
     return runtime_kind::omp;
 #elif DNNL_CPU_RUNTIME == DNNL_RUNTIME_TBB
     return runtime_kind::tbb;
+#elif DNNL_CPU_RUNTIME == DNNL_RUNTIME_TBB_AUTO
+    return runtime_kind::tbb_auto;
 #elif DNNL_CPU_RUNTIME == DNNL_RUNTIME_THREADPOOL
     return runtime_kind::threadpool;
 #elif DNNL_CPU_RUNTIME == DNNL_RUNTIME_SYCL

src/common/utils.hpp

@@ -752,7 +752,7 @@ struct set_once_before_first_get_setting_t {
 
 inline bool is_native_runtime(runtime_kind_t kind) {
     return utils::one_of(kind, runtime_kind::seq, runtime_kind::omp,
-            runtime_kind::tbb, runtime_kind::threadpool);
+            runtime_kind::tbb, runtime_kind::tbb_auto, runtime_kind::threadpool);
 }
 
 // Convenience wrapper to choose at compile-time between std::unique_ptr's

src/cpu/rnn/ref_rnn.hpp

@@ -62,7 +62,7 @@ void gates_reduction(const rnn_utils::rnn_conf_t &rnn,
     // @todo block k on simd-width to enable vectorization in
     // parallel_nd path
 #if DNNL_CPU_RUNTIME == DNNL_RUNTIME_OMP && _OPENMP >= 201307 \
-        && (!defined(__INTEL_COMPILER) || __INTEL_COMPILER < 1910)
+        && defined __INTEL_COMPILER && __INTEL_COMPILER < 1910
 #pragma omp parallel for simd collapse(2)
     for (int i = 0; i < rnn.n_gates; i++)
         for (int k = 0; k < rnn.dhc; k++)