[CPU] HOT FIX: allow latency mode (streams = 1) to use hyper threading processors (#17592)

src/plugins/intel_cpu/src/cpu_map_scheduling.cpp

@@ -42,18 +42,25 @@ std::vector<std::vector<int>> apply_scheduling_core_type(const ov::hint::Schedul
     return result_table;
 }
 
-std::vector<std::vector<int>> apply_hyper_threading(bool& input_value,
+std::vector<std::vector<int>> apply_hyper_threading(bool& input_ht_hint,
-                                                    const bool input_changed,
+                                                    const bool input_ht_changed,
+                                                    const std::string input_pm_hint,
                                                     const std::vector<std::vector<int>>& proc_type_table) {
     std::vector<std::vector<int>> result_table = proc_type_table;
 
-    if ((proc_type_table[0][HYPER_THREADING_PROC] > 0) &&
+    if (proc_type_table[0][HYPER_THREADING_PROC] > 0) {
-        (((!input_value) && input_changed) || ((!input_changed) && (proc_type_table.size() > 1)))) {
+        if (((!input_ht_hint) && input_ht_changed) || ((!input_ht_changed) && (input_pm_hint == "LATENCY")) ||
+            ((!input_ht_changed) && (input_pm_hint == "THROUGHPUT") && (proc_type_table.size() > 1))) {
             for (auto& i : result_table) {
                 i[ALL_PROC] -= i[HYPER_THREADING_PROC];
                 i[HYPER_THREADING_PROC] = 0;
             }
-        input_value = false;
+            input_ht_hint = false;
+        } else {
+            input_ht_hint = true;
+        }
+    } else {
+        input_ht_hint = false;
     }
 
     return result_table;

src/plugins/intel_cpu/src/cpu_map_scheduling.hpp

@@ -28,13 +28,15 @@ std::vector<std::vector<int>> apply_scheduling_core_type(const ov::hint::Schedul
 
 /**
  * @brief      Limit available CPU resource in processors type table according to hyper threading property
- * @param[in]  input_type indicate value of property enable_hyper_threading.
+ * @param[in]  input_ht_hint indicate value of property enable_hyper_threading.
- * @param[in]  input_changed indicate if value is set by user.
+ * @param[in]  input_ht_changed indicate if value is set by user.
+ * @param[in]  input_pm_hint indicate value of property performance_mode.
  * @param[in]  proc_type_table candidate processors available at this time
  * @return     updated proc_type_table which removed unmatched processors
  */
-std::vector<std::vector<int>> apply_hyper_threading(bool& input_type,
+std::vector<std::vector<int>> apply_hyper_threading(bool& input_ht_hint,
-                                                    const bool input_changed,
+                                                    const bool input_ht_changed,
+                                                    const std::string input_pm_hint,
                                                     const std::vector<std::vector<int>>& proc_type_table);
 
 /**

src/plugins/intel_cpu/src/cpu_streams_calculation.cpp

@@ -30,26 +30,10 @@ std::vector<std::vector<int>> get_streams_info_table(const int input_streams,
     std::vector<int> stream_info(CPU_STREAMS_TABLE_SIZE);
     std::vector<std::vector<int>> streams_info_table;
 
-    if (1 == input_streams) {
+    auto UpdateMixStreamInfo = [&]() {
-        stream_info[NUMBER_OF_STREAMS] = 1;
-        int limit_threads = (input_threads == 0) ? model_prefer_threads : input_threads;
-        if (proc_type_table[0][ALL_PROC] == proc_type_table[0][EFFICIENT_CORE_PROC]) {
-            stream_info[PROC_TYPE] = EFFICIENT_CORE_PROC;
-            stream_info[THREADS_PER_STREAM] = (input_threads == 0)
-                                                  ? proc_type_table[0][EFFICIENT_CORE_PROC]
-                                                  : std::min(proc_type_table[0][EFFICIENT_CORE_PROC], limit_threads);
-            streams_info_table.push_back(stream_info);
-        } else if ((proc_type_table[0][EFFICIENT_CORE_PROC] > 0) &&
-                   ((limit_threads == 0) || (limit_threads > proc_type_table[0][MAIN_CORE_PROC]))) {
-            stream_info[PROC_TYPE] = ALL_PROC;
-            int n_threads = std::accumulate(proc_type_table[0].begin() + MAIN_CORE_PROC,
-                                            proc_type_table[0].begin() + HYPER_THREADING_PROC,
-                                            0);
-            stream_info[THREADS_PER_STREAM] = (limit_threads == 0) ? n_threads : std::min(n_threads, limit_threads);
-            streams_info_table.push_back(stream_info);
         stream_info[NUMBER_OF_STREAMS] = 0;
-            n_threads = stream_info[THREADS_PER_STREAM];
+        int n_threads = stream_info[THREADS_PER_STREAM];
-            for (int n = MAIN_CORE_PROC; n < HYPER_THREADING_PROC; n++) {
+        for (int n = MAIN_CORE_PROC; n <= HYPER_THREADING_PROC; n++) {
             if (0 != proc_type_table[0][n]) {
                 stream_info[PROC_TYPE] = n;
                 if (n_threads <= proc_type_table[0][n]) {
@@ -63,28 +47,58 @@ std::vector<std::vector<int>> get_streams_info_table(const int input_streams,
                 }
             }
         }
-        } else {
+    };
+
+    if (1 == input_streams) {
+        stream_info[NUMBER_OF_STREAMS] = 1;
+        if (input_threads > 0) {
+            stream_info[THREADS_PER_STREAM] = std::min(proc_type_table[0][ALL_PROC], input_threads);
+            if ((stream_info[THREADS_PER_STREAM] > proc_type_table[0][MAIN_CORE_PROC]) &&
+                (proc_type_table[0][MAIN_CORE_PROC] > 0) && (proc_type_table[0][EFFICIENT_CORE_PROC] > 0)) {
+                stream_info[PROC_TYPE] = ALL_PROC;
+                streams_info_table.push_back(stream_info);
+                UpdateMixStreamInfo();
+            } else if ((stream_info[THREADS_PER_STREAM] <= proc_type_table[0][MAIN_CORE_PROC]) ||
+                       (proc_type_table[0][EFFICIENT_CORE_PROC] == 0)) {
                 stream_info[PROC_TYPE] = MAIN_CORE_PROC;
-            stream_info[THREADS_PER_STREAM] = (limit_threads == 0)
+                streams_info_table.push_back(stream_info);
-                                                  ? proc_type_table[0][MAIN_CORE_PROC]
+            } else {
-                                                  : std::min(proc_type_table[0][MAIN_CORE_PROC], limit_threads);
+                stream_info[PROC_TYPE] = EFFICIENT_CORE_PROC;
                 streams_info_table.push_back(stream_info);
             }
+        } else {
+            if (proc_type_table[0][ALL_PROC] == proc_type_table[0][EFFICIENT_CORE_PROC]) {
+                stream_info[PROC_TYPE] = EFFICIENT_CORE_PROC;
+                stream_info[THREADS_PER_STREAM] =
+                    (model_prefer_threads == 0)
+                        ? proc_type_table[0][EFFICIENT_CORE_PROC]
+                        : std::min(proc_type_table[0][EFFICIENT_CORE_PROC], model_prefer_threads);
+                streams_info_table.push_back(stream_info);
+            } else if ((proc_type_table[0][EFFICIENT_CORE_PROC] > 0) &&
+                       ((model_prefer_threads == 0) || (model_prefer_threads > proc_type_table[0][MAIN_CORE_PROC]))) {
+                stream_info[PROC_TYPE] = ALL_PROC;
+                stream_info[THREADS_PER_STREAM] =
+                    (model_prefer_threads == 0 || model_prefer_threads > proc_type_table[0][MAIN_CORE_PROC])
+                        ? proc_type_table[0][ALL_PROC]
+                        : proc_type_table[0][MAIN_CORE_PROC] + proc_type_table[0][HYPER_THREADING_PROC];
+                streams_info_table.push_back(stream_info);
+                UpdateMixStreamInfo();
+            } else {
+                stream_info[PROC_TYPE] = MAIN_CORE_PROC;
+                stream_info[THREADS_PER_STREAM] =
+                    proc_type_table[0][MAIN_CORE_PROC] + proc_type_table[0][HYPER_THREADING_PROC];
+                streams_info_table.push_back(stream_info);
+            }
+        }
         return streams_info_table;
-
     } else {
         int n_streams = 0;
         int n_threads = 0;
         int n_threads_per_stream = 0;
         int base_type = MAIN_CORE_PROC;
 
-        if (proc_type_table.size() == 1) {
+        n_threads =
-            n_threads = (0 == input_threads) ? proc_type_table[0][ALL_PROC]
+            (0 == input_threads) ? proc_type_table[0][ALL_PROC] : std::min(proc_type_table[0][ALL_PROC], input_threads);
-                                             : std::min(proc_type_table[0][ALL_PROC], input_threads);
-        } else {
-            n_threads = (0 == input_threads) ? proc_type_table[0][MAIN_CORE_PROC]
-                                             : std::min(proc_type_table[0][MAIN_CORE_PROC], input_threads);
-        }
 
         if (0 != input_streams) {
             base_type = (proc_type_table[0][MAIN_CORE_PROC] == 0) ? EFFICIENT_CORE_PROC : MAIN_CORE_PROC;
@@ -108,7 +122,8 @@ std::vector<std::vector<int>> get_streams_info_table(const int input_streams,
         } else {
             base_type = (proc_type_table[0][MAIN_CORE_PROC] == 0) ? EFFICIENT_CORE_PROC : MAIN_CORE_PROC;
             if (0 == model_prefer_threads) {
-                int n_proc = std::min(n_threads, proc_type_table[0][base_type]);
+                int n_proc = (proc_type_table.size() == 1) ? std::min(n_threads, proc_type_table[0][base_type])
+                                                           : std::min(n_threads, proc_type_table[1][base_type]);
                 if (0 == n_proc % 4) {
                     n_threads_per_stream = 4;
                 } else if (0 == n_proc % 5) {
@@ -126,18 +141,14 @@ std::vector<std::vector<int>> get_streams_info_table(const int input_streams,
                     n_threads_per_stream =
                         std::min(static_cast<int>(n_threads / n_streams), proc_type_table[0][base_type]);
                 } else {
-                    while (n_streams < n_threads_per_stream) {
+                    while (n_streams * 2 <= n_threads_per_stream) {
-                        if (1 == n_threads_per_stream) {
+                        n_threads_per_stream = static_cast<int>(n_threads_per_stream / 2);
-                            break;
-                        } else {
-                            n_threads_per_stream = static_cast<int>((n_threads_per_stream * 2 - 1) / 2);
                         n_threads_per_stream = static_cast<int>(
                             proc_type_table[0][base_type] /
                             ((proc_type_table[0][base_type] + n_threads_per_stream - 1) / n_threads_per_stream));
                         n_streams = static_cast<int>(n_threads / n_threads_per_stream);
                     }
                 }
-                }
             } else if ((1 == model_prefer_threads) && (proc_type_table[0][EFFICIENT_CORE_PROC] > 0) &&
                        (proc_type_table[0][MAIN_CORE_PROC] > 0) && (n_threads > proc_type_table[0][MAIN_CORE_PROC])) {
                 n_streams = (n_threads >= proc_type_table[0][MAIN_CORE_PROC] + proc_type_table[0][EFFICIENT_CORE_PROC])
@@ -270,7 +281,10 @@ void get_num_streams(const int streams,
     executor_config._orig_proc_type_table = orig_proc_type_table;
     std::vector<std::vector<int>> proc_type_table =
         apply_scheduling_core_type(config.schedulingCoreType, orig_proc_type_table);
-    proc_type_table = apply_hyper_threading(config.enableHyperThreading, config.changedHyperThreading, proc_type_table);
+    proc_type_table = apply_hyper_threading(config.enableHyperThreading,
+                                            config.changedHyperThreading,
+                                            config.perfHintsConfig.ovPerfHint,
+                                            proc_type_table);
     executor_config._proc_type_table = proc_type_table;
     executor_config._cpu_pinning = get_cpu_pinning(config.enableCpuPinning,
                                                    config.changedCpuPinning,

src/plugins/intel_cpu/tests/functional/shared_tests_instances/behavior/ov_plugin/properties_tests.cpp

@@ -201,10 +201,10 @@ const std::vector<ov::AnyMap> configsDeviceProperties = {
     {ov::device::properties(ov::AnyMap{{"CPU", ov::AnyMap{ov::num_streams(3)}}})}};
 
 const std::vector<ov::AnyMap> configsDevicePropertiesDouble = {
-    {ov::device::properties("CPU", ov::num_streams(5)), ov::num_streams(3)},
+    {ov::device::properties("CPU", ov::num_streams(3)), ov::num_streams(5)},
-    {ov::device::properties("CPU", ov::num_streams(5)),
+    {ov::device::properties("CPU", ov::num_streams(3)),
      ov::device::properties(ov::AnyMap{{"CPU", ov::AnyMap{ov::num_streams(7)}}}),
-     ov::num_streams(3)},
+     ov::num_streams(5)},
     {ov::device::properties("CPU", ov::num_streams(3)), ov::device::properties("CPU", ov::num_streams(5))},
     {ov::device::properties("CPU", ov::num_streams(3)),
      ov::device::properties(ov::AnyMap{{"CPU", ov::AnyMap{ov::num_streams(5)}}})},

src/plugins/intel_cpu/tests/unit/streams_info_table_test.cpp

@@ -83,10 +83,12 @@ INSTANTIATE_TEST_SUITE_P(SchedulingCoreTypeTable,
                                          _1sockets_E_CORE_ONLY));
 
 struct UseHTTestCase {
-    bool use_ht_value;
+    bool input_ht_value;
-    bool use_ht_changed;
+    bool input_ht_changed;
+    std::string input_pm_hint;
     std::vector<std::vector<int>> proc_type_table;
     std::vector<std::vector<int>> result_table;
+    bool output_ht_value;
 };
 
 class UseHTTests : public CommonTestUtils::TestsCommon, public testing::WithParamInterface<std::tuple<UseHTTestCase>> {
@@ -95,80 +97,260 @@ public:
         auto test_data = std::get<0>(GetParam());
 
         std::vector<std::vector<int>> test_result_table =
-            ov::intel_cpu::apply_hyper_threading(test_data.use_ht_value, test_data.use_ht_changed, test_data.proc_type_table);
+            ov::intel_cpu::apply_hyper_threading(test_data.input_ht_value,
+                                                 test_data.input_ht_changed,
+                                                 test_data.input_pm_hint,
+                                                 test_data.proc_type_table);
 
         ASSERT_EQ(test_data.result_table, test_result_table);
+        ASSERT_EQ(test_data.input_ht_value, test_data.output_ht_value);
     }
 };
 
-UseHTTestCase _2sockets_false = {
+UseHTTestCase _2sockets_false_latency = {
     false,
     true,
+    "LATENCY",
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
     {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    false,
 };
 
-UseHTTestCase _2sockets_true = {
+UseHTTestCase _2sockets_false_throughput = {
-    true,
-    true,
-    {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
-    {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
-};
-
-UseHTTestCase _2sockets_default_1 = {
-    false,
     false,
+    true,
+    "THROUGHPUT",
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
     {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    false,
 };
 
-UseHTTestCase _2sockets_default_2 = {
+UseHTTestCase _2sockets_true_latency = {
     true,
+    true,
+    "LATENCY",
+    {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
+    {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
+    true,
+};
+
+UseHTTestCase _2sockets_true_throughput = {
+    true,
+    true,
+    "THROUGHPUT",
+    {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
+    {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
+    true,
+};
+
+UseHTTestCase _2sockets_default_1_latency = {
     false,
+    false,
+    "LATENCY",
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
     {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    false,
 };
 
-UseHTTestCase _1sockets_false = {
+UseHTTestCase _2sockets_default_1_throughput = {
+    false,
+    false,
+    "THROUGHPUT",
+    {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    false,
+};
+
+UseHTTestCase _2sockets_default_2_latency = {
+    true,
+    false,
+    "LATENCY",
+    {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    false,
+};
+
+UseHTTestCase _2sockets_default_2_throughput = {
+    true,
+    false,
+    "THROUGHPUT",
+    {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    false,
+};
+
+UseHTTestCase _1sockets_1_false_latency = {
     false,
     true,
+    "LATENCY",
     {{20, 6, 8, 6}},
     {{14, 6, 8, 0}},
+    false,
 };
 
-UseHTTestCase _1sockets_true = {
+UseHTTestCase _1sockets_1_false_throughput = {
-    true,
+    false,
     true,
+    "THROUGHPUT",
     {{20, 6, 8, 6}},
-    {{20, 6, 8, 6}},
+    {{14, 6, 8, 0}},
+    false,
 };
 
-UseHTTestCase _1sockets_default_1 = {
+UseHTTestCase _1sockets_1_true_latency = {
-    false,
+    true,
-    false,
+    true,
+    "LATENCY",
     {{20, 6, 8, 6}},
     {{20, 6, 8, 6}},
+    true,
 };
 
-UseHTTestCase _1sockets_default_2 = {
+UseHTTestCase _1sockets_1_true_throughput = {
+    true,
+    true,
+    "THROUGHPUT",
+    {{20, 6, 8, 6}},
+    {{20, 6, 8, 6}},
+    true,
+};
+
+UseHTTestCase _1sockets_1_default_1_latency = {
+    false,
+    false,
+    "LATENCY",
+    {{20, 6, 8, 6}},
+    {{14, 6, 8, 0}},
+    false,
+};
+
+UseHTTestCase _1sockets_1_default_1_throughput = {
+    false,
+    false,
+    "THROUGHPUT",
+    {{20, 6, 8, 6}},
+    {{20, 6, 8, 6}},
+    true,
+};
+
+UseHTTestCase _1sockets_1_default_2_latency = {
     true,
     false,
+    "LATENCY",
+    {{20, 6, 8, 6}},
+    {{14, 6, 8, 0}},
+    false,
+};
+
+UseHTTestCase _1sockets_1_default_2_throughput = {
+    true,
+    false,
+    "THROUGHPUT",
     {{20, 6, 8, 6}},
     {{20, 6, 8, 6}},
+    true,
+};
+
+UseHTTestCase _1sockets_2_false_latency = {
+    false,
+    true,
+    "LATENCY",
+    {{12, 6, 0, 6}},
+    {{6, 6, 0, 0}},
+    false,
+};
+
+UseHTTestCase _1sockets_2_false_throughput = {
+    false,
+    true,
+    "THROUGHPUT",
+    {{12, 6, 0, 6}},
+    {{6, 6, 0, 0}},
+    false,
+};
+
+UseHTTestCase _1sockets_2_true_latency = {
+    true,
+    true,
+    "LATENCY",
+    {{12, 6, 0, 6}},
+    {{12, 6, 0, 6}},
+    true,
+};
+
+UseHTTestCase _1sockets_2_true_throughput = {
+    true,
+    true,
+    "THROUGHPUT",
+    {{12, 6, 0, 6}},
+    {{12, 6, 0, 6}},
+    true,
+};
+
+UseHTTestCase _1sockets_2_default_1_latency = {
+    false,
+    false,
+    "LATENCY",
+    {{12, 6, 0, 6}},
+    {{6, 6, 0, 0}},
+    false,
+};
+
+UseHTTestCase _1sockets_2_default_1_throughput = {
+    false,
+    false,
+    "THROUGHPUT",
+    {{12, 6, 0, 6}},
+    {{12, 6, 0, 6}},
+    true,
+};
+
+UseHTTestCase _1sockets_2_default_2_latency = {
+    true,
+    false,
+    "LATENCY",
+    {{12, 6, 0, 6}},
+    {{6, 6, 0, 0}},
+    false,
+};
+
+UseHTTestCase _1sockets_2_default_2_throughput = {
+    true,
+    false,
+    "THROUGHPUT",
+    {{12, 6, 0, 6}},
+    {{12, 6, 0, 6}},
+    true,
 };
 
 TEST_P(UseHTTests, UseHT) {}
 
 INSTANTIATE_TEST_SUITE_P(UseHTTable,
                          UseHTTests,
-                         testing::Values(_2sockets_false,
+                         testing::Values(_2sockets_false_latency,
-                                         _2sockets_true,
+                                         _2sockets_true_latency,
-                                         _2sockets_default_1,
+                                         _2sockets_default_1_latency,
-                                         _2sockets_default_2,
+                                         _2sockets_default_2_latency,
-                                         _1sockets_false,
+                                         _1sockets_1_false_latency,
-                                         _1sockets_true,
+                                         _1sockets_1_true_latency,
-                                         _1sockets_default_1,
+                                         _1sockets_1_default_1_latency,
-                                         _1sockets_default_2));
+                                         _1sockets_1_default_2_latency,
+                                         _1sockets_2_false_latency,
+                                         _1sockets_2_true_latency,
+                                         _1sockets_2_default_1_latency,
+                                         _1sockets_2_default_2_latency,
+                                         _2sockets_false_throughput,
+                                         _2sockets_true_throughput,
+                                         _2sockets_default_1_throughput,
+                                         _2sockets_default_2_throughput,
+                                         _1sockets_1_false_throughput,
+                                         _1sockets_1_true_throughput,
+                                         _1sockets_1_default_1_throughput,
+                                         _1sockets_1_default_2_throughput,
+                                         _1sockets_2_false_throughput,
+                                         _1sockets_2_true_throughput,
+                                         _1sockets_2_default_1_throughput,
+                                         _1sockets_2_default_2_throughput));
 
 struct StreamsCalculationTestCase {
     int input_streams;
@@ -202,7 +384,7 @@ StreamsCalculationTestCase _2sockets_104cores_latency_1 = {
     0,
     0,
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
-    {{1, MAIN_CORE_PROC, 104}},
+    {{1, MAIN_CORE_PROC, 208}},
 };
 
 StreamsCalculationTestCase _2sockets_104cores_latency_2 = {
@@ -223,13 +405,31 @@ StreamsCalculationTestCase _2sockets_104cores_latency_3 = {
     {{1, MAIN_CORE_PROC, 20}},
 };
 
+StreamsCalculationTestCase _2sockets_104cores_latency_4 = {
+    1,
+    208,
+    0,
+    0,
+    {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
+    {{1, MAIN_CORE_PROC, 208}},
+};
+
+StreamsCalculationTestCase _2sockets_104cores_latency_5 = {
+    1,
+    0,
+    0,
+    0,
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    {{1, MAIN_CORE_PROC, 104}},
+};
+
 StreamsCalculationTestCase _2sockets_104cores_tput_1 = {
     0,
     0,
     0,
     0,
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
-    {{26, MAIN_CORE_PROC, 4}},
+    {{52, MAIN_CORE_PROC, 4}},
 };
 
 StreamsCalculationTestCase _2sockets_104cores_tput_2 = {
@@ -238,7 +438,7 @@ StreamsCalculationTestCase _2sockets_104cores_tput_2 = {
     0,
     0,
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
-    {{2, MAIN_CORE_PROC, 52}},
+    {{2, MAIN_CORE_PROC, 104}},
 };
 
 StreamsCalculationTestCase _2sockets_104cores_tput_3 = {
@@ -265,7 +465,7 @@ StreamsCalculationTestCase _2sockets_104cores_tput_5 = {
     0,
     1,
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
-    {{104, MAIN_CORE_PROC, 1}},
+    {{208, MAIN_CORE_PROC, 1}},
 };
 
 StreamsCalculationTestCase _2sockets_104cores_tput_6 = {
@@ -274,7 +474,7 @@ StreamsCalculationTestCase _2sockets_104cores_tput_6 = {
     0,
     2,
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
-    {{52, MAIN_CORE_PROC, 2}},
+    {{104, MAIN_CORE_PROC, 2}},
 };
 
 StreamsCalculationTestCase _2sockets_104cores_tput_7 = {
@@ -283,7 +483,7 @@ StreamsCalculationTestCase _2sockets_104cores_tput_7 = {
     0,
     8,
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
-    {{13, MAIN_CORE_PROC, 8}},
+    {{26, MAIN_CORE_PROC, 8}},
 };
 
 StreamsCalculationTestCase _2sockets_104cores_tput_8 = {
@@ -310,7 +510,7 @@ StreamsCalculationTestCase _2sockets_104cores_tput_10 = {
     2,
     0,
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
-    {{2, MAIN_CORE_PROC, 52}},
+    {{2, MAIN_CORE_PROC, 104}},
 };
 
 StreamsCalculationTestCase _2sockets_104cores_tput_11 = {
@@ -319,7 +519,7 @@ StreamsCalculationTestCase _2sockets_104cores_tput_11 = {
     5,
     0,
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
-    {{2, MAIN_CORE_PROC, 52}},
+    {{2, MAIN_CORE_PROC, 104}},
 };
 
 StreamsCalculationTestCase _2sockets_104cores_tput_12 = {
@@ -328,6 +528,78 @@ StreamsCalculationTestCase _2sockets_104cores_tput_12 = {
     2,
     2,
     {{208, 104, 0, 104}, {104, 52, 0, 52}, {104, 52, 0, 52}},
+    {{2, MAIN_CORE_PROC, 104}},
+};
+
+StreamsCalculationTestCase _2sockets_104cores_tput_13 = {
+    0,
+    0,
+    0,
+    0,
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    {{26, MAIN_CORE_PROC, 4}},
+};
+
+StreamsCalculationTestCase _2sockets_104cores_tput_14 = {
+    2,
+    0,
+    0,
+    0,
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    {{2, MAIN_CORE_PROC, 52}},
+};
+
+StreamsCalculationTestCase _2sockets_104cores_tput_15 = {
+    0,
+    0,
+    0,
+    1,
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    {{104, MAIN_CORE_PROC, 1}},
+};
+
+StreamsCalculationTestCase _2sockets_104cores_tput_16 = {
+    0,
+    0,
+    0,
+    2,
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    {{52, MAIN_CORE_PROC, 2}},
+};
+
+StreamsCalculationTestCase _2sockets_104cores_tput_17 = {
+    0,
+    0,
+    0,
+    8,
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    {{13, MAIN_CORE_PROC, 8}},
+};
+
+StreamsCalculationTestCase _2sockets_104cores_tput_18 = {
+    0,
+    0,
+    2,
+    0,
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    {{2, MAIN_CORE_PROC, 52}},
+};
+
+StreamsCalculationTestCase _2sockets_104cores_tput_19 = {
+    2,
+    0,
+    5,
+    0,
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
+    {{2, MAIN_CORE_PROC, 52}},
+};
+
+StreamsCalculationTestCase _2sockets_104cores_tput_20 = {
+    0,
+    0,
+    2,
+    2,
+    {{104, 104, 0, 0}, {52, 52, 0, 0}, {52, 52, 0, 0}},
     {{2, MAIN_CORE_PROC, 52}},
 };
 
@@ -381,7 +653,7 @@ StreamsCalculationTestCase _1sockets_14cores_latency_1 = {
     0,
     0,
     0,
-    {{20, 6, 8, 6}},
+    {{14, 6, 8, 0}},
     {{1, ALL_PROC, 14}, {0, MAIN_CORE_PROC, 6}, {0, EFFICIENT_CORE_PROC, 8}},
 };
 
@@ -400,7 +672,7 @@ StreamsCalculationTestCase _1sockets_14cores_latency_3 = {
     0,
     6,
     {{20, 6, 8, 6}},
-    {{1, MAIN_CORE_PROC, 6}},
+    {{1, MAIN_CORE_PROC, 12}},
 };
 
 StreamsCalculationTestCase _1sockets_14cores_latency_4 = {
@@ -409,7 +681,7 @@ StreamsCalculationTestCase _1sockets_14cores_latency_4 = {
     0,
     14,
     {{20, 6, 8, 6}},
-    {{1, ALL_PROC, 14}, {0, MAIN_CORE_PROC, 6}, {0, EFFICIENT_CORE_PROC, 8}},
+    {{1, ALL_PROC, 20}, {0, MAIN_CORE_PROC, 6}, {0, EFFICIENT_CORE_PROC, 8}, {0, HYPER_THREADING_PROC, 6}},
 };
 
 StreamsCalculationTestCase _1sockets_14cores_latency_5 = {
@@ -418,6 +690,42 @@ StreamsCalculationTestCase _1sockets_14cores_latency_5 = {
     2,
     14,
     {{20, 6, 8, 6}},
+    {{1, ALL_PROC, 20}, {0, MAIN_CORE_PROC, 6}, {0, EFFICIENT_CORE_PROC, 8}, {0, HYPER_THREADING_PROC, 6}},
+};
+
+StreamsCalculationTestCase _1sockets_14cores_latency_6 = {
+    1,
+    0,
+    0,
+    0,
+    {{20, 6, 8, 6}},
+    {{1, ALL_PROC, 20}, {0, MAIN_CORE_PROC, 6}, {0, EFFICIENT_CORE_PROC, 8}, {0, HYPER_THREADING_PROC, 6}},
+};
+
+StreamsCalculationTestCase _1sockets_14cores_latency_7 = {
+    1,
+    0,
+    0,
+    6,
+    {{14, 6, 8, 0}},
+    {{1, MAIN_CORE_PROC, 6}},
+};
+
+StreamsCalculationTestCase _1sockets_14cores_latency_8 = {
+    1,
+    0,
+    0,
+    14,
+    {{14, 6, 8, 0}},
+    {{1, ALL_PROC, 14}, {0, MAIN_CORE_PROC, 6}, {0, EFFICIENT_CORE_PROC, 8}},
+};
+
+StreamsCalculationTestCase _1sockets_14cores_latency_9 = {
+    1,
+    0,
+    2,
+    14,
+    {{14, 6, 8, 0}},
     {{1, ALL_PROC, 14}, {0, MAIN_CORE_PROC, 6}, {0, EFFICIENT_CORE_PROC, 8}},
 };
 
@@ -565,14 +873,13 @@ StreamsCalculationTestCase _1sockets_14cores_tput_16 = {
     {{6, MAIN_CORE_PROC, 1}, {4, EFFICIENT_CORE_PROC, 2}, {1, HYPER_THREADING_PROC, 1}},
 };
 
-
 StreamsCalculationTestCase _1sockets_10cores_latency_1 = {
     1,
     0,
     0,
     0,
     {{12, 2, 8, 2}},
-    {{1, ALL_PROC, 10}, {0, MAIN_CORE_PROC, 2}, {0, EFFICIENT_CORE_PROC, 8}},
+    {{1, ALL_PROC, 12}, {0, MAIN_CORE_PROC, 2}, {0, EFFICIENT_CORE_PROC, 8}, {0, HYPER_THREADING_PROC, 2}},
 };
 
 StreamsCalculationTestCase _1sockets_10cores_latency_2 = {
@@ -590,7 +897,7 @@ StreamsCalculationTestCase _1sockets_10cores_latency_3 = {
     0,
     2,
     {{12, 2, 8, 2}},
-    {{1, MAIN_CORE_PROC, 2}},
+    {{1, MAIN_CORE_PROC, 4}},
 };
 
 StreamsCalculationTestCase _1sockets_10cores_latency_4 = {
@@ -599,6 +906,33 @@ StreamsCalculationTestCase _1sockets_10cores_latency_4 = {
     0,
     10,
     {{12, 2, 8, 2}},
+    {{1, ALL_PROC, 12}, {0, MAIN_CORE_PROC, 2}, {0, EFFICIENT_CORE_PROC, 8}, {0, HYPER_THREADING_PROC, 2}},
+};
+
+StreamsCalculationTestCase _1sockets_10cores_latency_5 = {
+    1,
+    0,
+    0,
+    0,
+    {{10, 2, 8, 0}},
+    {{1, ALL_PROC, 10}, {0, MAIN_CORE_PROC, 2}, {0, EFFICIENT_CORE_PROC, 8}},
+};
+
+StreamsCalculationTestCase _1sockets_10cores_latency_6 = {
+    1,
+    0,
+    0,
+    2,
+    {{10, 2, 8, 0}},
+    {{1, MAIN_CORE_PROC, 2}},
+};
+
+StreamsCalculationTestCase _1sockets_10cores_latency_7 = {
+    1,
+    0,
+    0,
+    10,
+    {{10, 2, 8, 0}},
     {{1, ALL_PROC, 10}, {0, MAIN_CORE_PROC, 2}, {0, EFFICIENT_CORE_PROC, 8}},
 };
 
@@ -662,7 +996,7 @@ StreamsCalculationTestCase _1sockets_8cores_latency_1 = {
     0,
     0,
     {{12, 4, 4, 4}},
-    {{1, ALL_PROC, 8}, {0, MAIN_CORE_PROC, 4}, {0, EFFICIENT_CORE_PROC, 4}},
+    {{1, ALL_PROC, 12}, {0, MAIN_CORE_PROC, 4}, {0, EFFICIENT_CORE_PROC, 4}, {0, HYPER_THREADING_PROC, 4}},
 };
 
 StreamsCalculationTestCase _1sockets_8cores_latency_2 = {
@@ -671,7 +1005,7 @@ StreamsCalculationTestCase _1sockets_8cores_latency_2 = {
     0,
     0,
     {{12, 4, 4, 4}},
-    {{1, ALL_PROC, 8}, {0, MAIN_CORE_PROC, 4}, {0, EFFICIENT_CORE_PROC, 4}},
+    {{1, ALL_PROC, 12}, {0, MAIN_CORE_PROC, 4}, {0, EFFICIENT_CORE_PROC, 4}, {0, HYPER_THREADING_PROC, 4}},
 };
 
 StreamsCalculationTestCase _1sockets_8cores_latency_3 = {
@@ -680,7 +1014,7 @@ StreamsCalculationTestCase _1sockets_8cores_latency_3 = {
     0,
     4,
     {{12, 4, 4, 4}},
-    {{1, MAIN_CORE_PROC, 4}},
+    {{1, MAIN_CORE_PROC, 8}},
 };
 
 StreamsCalculationTestCase _1sockets_8cores_latency_4 = {
@@ -689,6 +1023,33 @@ StreamsCalculationTestCase _1sockets_8cores_latency_4 = {
     0,
     8,
     {{12, 4, 4, 4}},
+    {{1, ALL_PROC, 12}, {0, MAIN_CORE_PROC, 4}, {0, EFFICIENT_CORE_PROC, 4}, {0, HYPER_THREADING_PROC, 4}},
+};
+
+StreamsCalculationTestCase _1sockets_8cores_latency_5 = {
+    1,
+    0,
+    0,
+    0,
+    {{8, 4, 4, 0}},
+    {{1, ALL_PROC, 8}, {0, MAIN_CORE_PROC, 4}, {0, EFFICIENT_CORE_PROC, 4}},
+};
+
+StreamsCalculationTestCase _1sockets_8cores_latency_6 = {
+    1,
+    0,
+    0,
+    4,
+    {{8, 4, 4, 0}},
+    {{1, MAIN_CORE_PROC, 4}},
+};
+
+StreamsCalculationTestCase _1sockets_8cores_latency_7 = {
+    1,
+    0,
+    0,
+    8,
+    {{8, 4, 4, 0}},
     {{1, ALL_PROC, 8}, {0, MAIN_CORE_PROC, 4}, {0, EFFICIENT_CORE_PROC, 4}},
 };
 
@@ -698,7 +1059,7 @@ StreamsCalculationTestCase _1sockets_8cores_tput_1 = {
     0,
     0,
     {{12, 4, 4, 4}},
-    {{2, MAIN_CORE_PROC, 2}, {2, EFFICIENT_CORE_PROC, 2}, {2, HYPER_THREADING_PROC, 2}},
+    {{1, MAIN_CORE_PROC, 4}, {1, EFFICIENT_CORE_PROC, 4}, {1, HYPER_THREADING_PROC, 4}},
 };
 
 StreamsCalculationTestCase _1sockets_8cores_tput_2 = {
@@ -755,13 +1116,22 @@ StreamsCalculationTestCase _1sockets_8cores_tput_7 = {
     {{4, MAIN_CORE_PROC, 1}, {2, EFFICIENT_CORE_PROC, 2}, {4, HYPER_THREADING_PROC, 1}},
 };
 
+StreamsCalculationTestCase _1sockets_8cores_tput_8 = {
+    0,
+    0,
+    0,
+    0,
+    {{8, 4, 4, 0}},
+    {{2, MAIN_CORE_PROC, 2}, {2, EFFICIENT_CORE_PROC, 2}},
+};
+
 StreamsCalculationTestCase _1sockets_6cores_latency_1 = {
     1,
     0,
     0,
     0,
     {{12, 6, 0, 6}},
-    {{1, MAIN_CORE_PROC, 6}},
+    {{1, MAIN_CORE_PROC, 12}},
 };
 
 StreamsCalculationTestCase _1sockets_6cores_latency_2 = {
@@ -770,6 +1140,15 @@ StreamsCalculationTestCase _1sockets_6cores_latency_2 = {
     0,
     0,
     {{12, 6, 0, 6}},
+    {{1, MAIN_CORE_PROC, 12}},
+};
+
+StreamsCalculationTestCase _1sockets_6cores_latency_3 = {
+    1,
+    0,
+    0,
+    0,
+    {{6, 6, 0, 0}},
     {{1, MAIN_CORE_PROC, 6}},
 };
 
@@ -797,7 +1176,7 @@ StreamsCalculationTestCase _1sockets_6cores_tput_3 = {
     0,
     0,
     {{12, 6, 0, 6}},
-    {{3, MAIN_CORE_PROC, 2}, {1, HYPER_THREADING_PROC, 2}},
+    {{2, MAIN_CORE_PROC, 3}},
 };
 
 StreamsCalculationTestCase _1sockets_6cores_tput_4 = {
@@ -842,7 +1221,7 @@ StreamsCalculationTestCase _1sockets_ecores_latency_4 = {
     0,
     4,
     {{16, 0, 16, 0}},
-    {{1, EFFICIENT_CORE_PROC, 16}},
+    {{1, EFFICIENT_CORE_PROC, 4}},
 };
 
 StreamsCalculationTestCase _1sockets_ecores_tput_1 = {
@@ -906,6 +1285,8 @@ INSTANTIATE_TEST_SUITE_P(StreamsInfoTable,
                          testing::Values(_2sockets_104cores_latency_1,
                                          _2sockets_104cores_latency_2,
                                          _2sockets_104cores_latency_3,
+                                         _2sockets_104cores_latency_4,
+                                         _2sockets_104cores_latency_5,
                                          _2sockets_104cores_tput_1,
                                          _2sockets_104cores_tput_2,
                                          _2sockets_104cores_tput_3,
@@ -918,6 +1299,14 @@ INSTANTIATE_TEST_SUITE_P(StreamsInfoTable,
                                          _2sockets_104cores_tput_10,
                                          _2sockets_104cores_tput_11,
                                          _2sockets_104cores_tput_12,
+                                         _2sockets_104cores_tput_13,
+                                         _2sockets_104cores_tput_14,
+                                         _2sockets_104cores_tput_15,
+                                         _2sockets_104cores_tput_16,
+                                         _2sockets_104cores_tput_17,
+                                         _2sockets_104cores_tput_18,
+                                         _2sockets_104cores_tput_19,
+                                         _2sockets_104cores_tput_20,
                                          _2sockets_48cores_latency_1,
                                          _2sockets_48cores_tput_1,
                                          _2sockets_48cores_tput_2,
@@ -928,6 +1317,10 @@ INSTANTIATE_TEST_SUITE_P(StreamsInfoTable,
                                          _1sockets_14cores_latency_3,
                                          _1sockets_14cores_latency_4,
                                          _1sockets_14cores_latency_5,
+                                         _1sockets_14cores_latency_6,
+                                         _1sockets_14cores_latency_7,
+                                         _1sockets_14cores_latency_8,
+                                         _1sockets_14cores_latency_9,
                                          _1sockets_14cores_tput_1,
                                          _1sockets_14cores_tput_2,
                                          _1sockets_14cores_tput_3,
@@ -948,6 +1341,9 @@ INSTANTIATE_TEST_SUITE_P(StreamsInfoTable,
                                          _1sockets_10cores_latency_2,
                                          _1sockets_10cores_latency_3,
                                          _1sockets_10cores_latency_4,
+                                         _1sockets_10cores_latency_5,
+                                         _1sockets_10cores_latency_6,
+                                         _1sockets_10cores_latency_7,
                                          _1sockets_10cores_tput_1,
                                          _1sockets_10cores_tput_2,
                                          _1sockets_10cores_tput_3,
@@ -958,6 +1354,9 @@ INSTANTIATE_TEST_SUITE_P(StreamsInfoTable,
                                          _1sockets_8cores_latency_2,
                                          _1sockets_8cores_latency_3,
                                          _1sockets_8cores_latency_4,
+                                         _1sockets_8cores_latency_5,
+                                         _1sockets_8cores_latency_6,
+                                         _1sockets_8cores_latency_7,
                                          _1sockets_8cores_tput_1,
                                          _1sockets_8cores_tput_2,
                                          _1sockets_8cores_tput_3,
@@ -965,8 +1364,10 @@ INSTANTIATE_TEST_SUITE_P(StreamsInfoTable,
                                          _1sockets_8cores_tput_5,
                                          _1sockets_8cores_tput_6,
                                          _1sockets_8cores_tput_7,
+                                         _1sockets_8cores_tput_8,
                                          _1sockets_6cores_latency_1,
                                          _1sockets_6cores_latency_2,
+                                         _1sockets_6cores_latency_3,
                                          _1sockets_6cores_tput_1,
                                          _1sockets_6cores_tput_2,
                                          _1sockets_6cores_tput_3,