using single numa node for CNN model

openvinotoolkit · Dec 23, 2024 · 13ce056 · 13ce056
1 parent cc357c5
commit 13ce056
Show file tree

Hide file tree

Showing 3 changed files with 307 additions and 11 deletions.
diff --git a/src/plugins/intel_cpu/src/cpu_streams_calculation.cpp b/src/plugins/intel_cpu/src/cpu_streams_calculation.cpp
@@ -46,6 +46,7 @@ std::vector<std::vector<int>> get_streams_info_table(
     const int input_threads,
     const int input_infer_requests,
     const int model_prefer_threads,
+    const Config::ModelType model_type,
     const std::string input_perf_hint,
     const std::set<ov::hint::ModelDistributionPolicy> hint_model_distribution_policy,
     const std::vector<std::vector<int>>& proc_type_table) {
@@ -252,13 +253,20 @@ std::vector<std::vector<int>> get_streams_info_table(
                 n_threads_per_stream = proc_type_table[0][ALL_PROC];
             }
         } else {
-            size_t socket_index = 0;
-            for (socket_index = 0; socket_index < proc_socket_table.size(); socket_index++) {
-                if (proc_socket_table[socket_index][PROC_SOCKET_ID] == current_socket_id) {
-                    break;
+            std::vector<int> current_socket_info;
+
+            if (model_type == Config::ModelType::CNN) {
+                current_socket_info = proc_type_table.size() == 1 ? proc_type_table[0] : proc_type_table[1];
+            } else {
+                size_t socket_index = 0;
+                for (socket_index = 0; socket_index < proc_socket_table.size(); socket_index++) {
+                    if (proc_socket_table[socket_index][PROC_SOCKET_ID] == current_socket_id) {
+                        break;
+                    }
                 }
+                current_socket_info = socket_index < proc_socket_table.size() ? proc_socket_table[socket_index] : proc_socket_table[0];
             }
-            const std::vector<int>& current_socket_info = proc_socket_table[socket_index];
+
             n_threads_per_stream = model_prefer_threads == 0
                                        ? current_socket_info[ALL_PROC]
                                        : std::min(current_socket_info[ALL_PROC], model_prefer_threads);
@@ -271,11 +279,19 @@ std::vector<std::vector<int>> get_streams_info_table(
                 update_streams_per_node(EFFICIENT_CORE_PROC, current_socket_info);
             } else {
                 stream_info[PROC_TYPE] = ALL_PROC;
-                update_mix_stream_info(current_socket_info,
-                                       proc_type_table,
-                                       n_threads_per_stream,
-                                       IStreamsExecutor::Config::StreamsMode::SUB_STREAMS_NULL,
-                                       ALL_PROC);
+                if (model_type == Config::ModelType::CNN) {
+                    update_mix_stream_info(current_socket_info,
+                                           {current_socket_info},
+                                           n_threads_per_stream,
+                                           IStreamsExecutor::Config::StreamsMode::SUB_STREAMS_NULL,
+                                           ALL_PROC);
+                } else {
+                    update_mix_stream_info(current_socket_info,
+                                           proc_type_table,
+                                           n_threads_per_stream,
+                                           IStreamsExecutor::Config::StreamsMode::SUB_STREAMS_NULL,
+                                           ALL_PROC);
+                }
             }
             update_ids_method(current_socket_info);
         }
@@ -702,10 +718,10 @@ std::vector<std::vector<int>> generate_stream_info(const int streams,
                                                      config.threads,
                                                      config.hintNumRequests,
                                                      model_prefer_threads,
+                                                     config.modelType,
                                                      ov::util::to_string(config.hintPerfMode),
                                                      config.modelDistributionPolicy,
                                                      proc_type_table);
-    // streams_info_table = {{1, 1, 56, 1, 1}, {-1, 1, 28, 1, 1}, {-1, 1, 28, 0, 0}};
     if (config.modelDistributionPolicy.find(ov::hint::ModelDistributionPolicy::TENSOR_PARALLEL) !=
         config.modelDistributionPolicy.end()) {
         config.streamsRankTable =

diff --git a/src/plugins/intel_cpu/src/cpu_streams_calculation.hpp b/src/plugins/intel_cpu/src/cpu_streams_calculation.hpp
@@ -35,6 +35,7 @@ namespace intel_cpu {
  * function.
  *               - input "0" indicates that the function generates the optimal number of threads per stream based on
  * processors type information.
+ * @param[in]  model_type is model type of current model.
  * @param[in]  input_perf_hint is performance hint set by user via ov::hint::performance_mode or the default value.
  * @param[in]  hint_llm_distribution_policy is the distribution policy for Large language models
  * @param[in]  proc_type_table is currently available candidate processors.
@@ -48,6 +49,7 @@ std::vector<std::vector<int>> get_streams_info_table(
     const int input_threads,
     const int input_infer_requests,
     const int model_prefer_threads,
+    const Config::ModelType model_type,
     const std::string input_perf_hint,
     const std::set<ov::hint::ModelDistributionPolicy> hint_llm_distribution_policy,
     const std::vector<std::vector<int>>& proc_type_table);