CPU pinning on Windows (#19405)

* add cpu pinning on windows

* remove pinning limitation on windows

* only support the machine with one numa node

* fix code style

* fix build error on macos

* set mask initial value

* fix test failure on window

* fix build failure on macos, add limitation on windows machine with two sockets

* fix test failure on windows

* fix test failure

* fix comments

src/inference/src/dev/threading/cpu_streams_executor_internal.cpp

@@ -51,8 +51,12 @@ void get_cur_stream_info(const int stream_id,
         max_threads_per_core = 2;
     }
 
-#if defined(_WIN32) || defined(__APPLE__)
+#if defined(__APPLE__)
     cpu_reserve = false;
+#elif defined(_WIN32)
+    if (proc_type_table.size() > 1) {
+        cpu_reserve = false;
+    }
 #endif
     if (cpu_reserve) {
         stream_type = STREAM_WITH_OBSERVE;

src/inference/src/dev/threading/thread_affinity.cpp

@@ -110,7 +110,33 @@ bool pin_current_thread_to_socket(int socket) {
     }
     return res;
 }
-#else   // no threads pinning/binding on Win/MacOS
+#elif defined(_WIN32)
+std::tuple<CpuSet, int> get_process_mask() {
+    DWORD_PTR pro_mask, sys_mask;
+    if (0 != GetProcessAffinityMask(GetCurrentProcess(), &pro_mask, &sys_mask)) {
+        CpuSet mask(new DWORD_PTR(pro_mask));
+        return std::make_tuple(std::move(mask), 0);
+    }
+    return std::make_tuple(nullptr, 0);
+}
+void release_process_mask(cpu_set_t*) {}
+
+bool pin_thread_to_vacant_core(int thrIdx,
+                               int hyperthreads,
+                               int ncores,
+                               const CpuSet& procMask,
+                               const std::vector<int>& cpu_ids,
+                               int cpuIdxOffset) {
+    return 0 != SetThreadAffinityMask(GetCurrentThread(), DWORD_PTR(1) << cpu_ids[thrIdx]);
+}
+bool pin_current_thread_by_mask(int ncores, const CpuSet& procMask) {
+    DWORD_PTR mask = static_cast<DWORD_PTR>(*procMask.get());
+    return 0 != SetThreadAffinityMask(GetCurrentThread(), mask);
+}
+bool pin_current_thread_to_socket(int socket) {
+    return false;
+}
+#else   // no threads pinning/binding on MacOS
 std::tuple<CpuSet, int> get_process_mask() {
     return std::make_tuple(nullptr, 0);
 }

src/inference/src/dev/threading/thread_affinity.hpp

@@ -11,13 +11,20 @@
 #if !(defined(__APPLE__) || defined(__EMSCRIPTEN__) || defined(_WIN32))
 #    include <sched.h>
 #endif
+#if defined(_WIN32)
+#    include <windows.h>
+
+#    include <thread>
+#endif
 
 namespace ov {
 namespace threading {
 
-#if (defined(__APPLE__) || defined(__EMSCRIPTEN__) || defined(_WIN32))
+#if (defined(__APPLE__) || defined(__EMSCRIPTEN__))
 using cpu_set_t = void;
-#endif  // (defined(__APPLE__) || defined(__EMSCRIPTEN__) || defined(_WIN32))
+#elif defined(_WIN32)
+using cpu_set_t = DWORD_PTR;
+#endif
 
 /**
  * @brief      Release the cores affinity mask for the current process

src/inference/tests/unit/cpu_stream_info_test.cpp

@@ -357,21 +357,12 @@ LinuxCpuStreamTypeCase _1sockets_4cores_binding = {
     },
     {{8, 4, 0, 4, 0, 0}},
     {{4, MAIN_CORE_PROC, 1, 0, 0}},
-#    if defined(__linux__)
     {
         STREAM_WITH_OBSERVE,
         STREAM_WITH_OBSERVE,
         STREAM_WITH_OBSERVE,
         STREAM_WITH_OBSERVE,
     },
-#    else
-    {
-        STREAM_WITHOUT_PARAM,
-        STREAM_WITHOUT_PARAM,
-        STREAM_WITHOUT_PARAM,
-        STREAM_WITHOUT_PARAM,
-    },
-#    endif
     {1, 1, 1, 1},
     {
         MAIN_CORE_PROC,
@@ -423,17 +414,10 @@ LinuxCpuStreamTypeCase _1sockets_12cores_pcore_binding = {
     },
     {{20, 8, 4, 8, 0, 0}},
     {{2, MAIN_CORE_PROC, 4, 0, 0}},
-#    if defined(__linux__)
     {
         STREAM_WITH_OBSERVE,
         STREAM_WITH_OBSERVE,
     },
-#    else
-    {
-        STREAM_WITH_CORE_TYPE,
-        STREAM_WITH_CORE_TYPE,
-    },
-#    endif
     {4, 4},
     {
         MAIN_CORE_PROC,
@@ -488,21 +472,12 @@ LinuxCpuStreamTypeCase _1sockets_12cores_ecore_binding = {
     },
     {{20, 8, 4, 8, 0, 0}},
     {{4, EFFICIENT_CORE_PROC, 1, 0, 0}},
-#    if defined(__linux__)
     {
         STREAM_WITH_OBSERVE,
         STREAM_WITH_OBSERVE,
         STREAM_WITH_OBSERVE,
         STREAM_WITH_OBSERVE,
     },
-#    else
-    {
-        STREAM_WITH_CORE_TYPE,
-        STREAM_WITH_CORE_TYPE,
-        STREAM_WITH_CORE_TYPE,
-        STREAM_WITH_CORE_TYPE,
-    },
-#    endif
     {1, 1, 1, 1},
     {
         EFFICIENT_CORE_PROC,

src/plugins/intel_cpu/src/cpu_map_scheduling.cpp

@@ -79,19 +79,29 @@ bool get_cpu_pinning(bool& input_value,
     int num_sockets = get_default_latency_streams(latency_threading_mode);
     bool latency = num_streams <= num_sockets && num_streams > 0;
 
+    if (input_changed) {
+        result_value = input_value;
+    } else {
         if (proc_type_table[0][EFFICIENT_CORE_PROC] > 0 &&
             proc_type_table[0][EFFICIENT_CORE_PROC] < proc_type_table[0][ALL_PROC]) {
             result_value =
-            input_changed
+                ((latency || bind_type == threading::IStreamsExecutor::ThreadBindingType::NUMA) ? false : true);
-                ? input_value
-                : ((latency || bind_type == threading::IStreamsExecutor::ThreadBindingType::NUMA) ? false : true);
         } else {
-        result_value = input_changed
+            result_value = (bind_type == threading::IStreamsExecutor::ThreadBindingType::NUMA ? false : true);
-                           ? input_value
-                           : (bind_type == threading::IStreamsExecutor::ThreadBindingType::NUMA ? false : true);
         }
 #if (IE_THREAD == IE_THREAD_TBB || IE_THREAD == IE_THREAD_TBB_AUTO)
-#    if defined(__APPLE__) || defined(_WIN32)
+#    if defined(_WIN32)
+        result_value = false;
+#    endif
+#endif
+    }
+#if (IE_THREAD == IE_THREAD_TBB || IE_THREAD == IE_THREAD_TBB_AUTO)
+#    if defined(_WIN32)
+    if (proc_type_table.size() > 1) {
+        result_value = false;
+    }
+#    endif
+#    if defined(__APPLE__)
     result_value = false;
 #    endif
 #endif

src/plugins/intel_cpu/tests/unit/streams_info/streams_e2e_test.cpp

@@ -578,7 +578,7 @@ StreamGenerateionTestCase generation_tput_2sockets_48cores_9 = {
     {{24, MAIN_CORE_PROC, 1, 0, 0}, {24, MAIN_CORE_PROC, 1, 1, 1}},
 };
 
-#ifdef __linux__
+#if defined (__linux__) || defined(_WIN32)
 INSTANTIATE_TEST_SUITE_P(smoke_StreamsGeneration,
                          StreamGenerationTests,
                          ::testing::Values(generation_latency_1sockets_14cores_3,