Merge branch 'perf_config' into 'master'

fix cpu config

See merge request !341

mace/benchmark/benchmark_model.cc

@@ -214,7 +214,7 @@ DEFINE_string(model_data_file, "",
               "model data file name, used when EMBED_MODEL_DATA set to 0");
 DEFINE_int32(gpu_perf_hint, 2, "0:DEFAULT/1:LOW/2:NORMAL/3:HIGH");
 DEFINE_int32(gpu_priority_hint, 1, "0:DEFAULT/1:LOW/2:NORMAL/3:HIGH");
-DEFINE_int32(omp_num_threads, 8, "num of openmp threads");
+DEFINE_int32(omp_num_threads, 4, "num of openmp threads");
 DEFINE_int32(cpu_power_option, 0,
              "0:DEFAULT/1:HIGH_PERFORMANCE/2:BATTERY_SAVE");
 
@@ -269,14 +269,13 @@ int Main(int argc, char **argv) {
   }
 
   // config runtime
+  mace::ConfigOmpThreads(FLAGS_omp_num_threads);
+  mace::ConfigCPUPowerOption(
+      static_cast<CPUPowerOption>(FLAGS_cpu_power_option));
   if (device_type == OPENCL) {
     mace::ConfigOpenCLRuntime(
         static_cast<GPUPerfHint>(FLAGS_gpu_perf_hint),
         static_cast<GPUPriorityHint>(FLAGS_gpu_priority_hint));
-  } else if (device_type == CPU || device_type == NEON) {
-    mace::ConfigOmpThreadsAndAffinity(
-        FLAGS_omp_num_threads,
-        static_cast<CPUPowerOption>(FLAGS_cpu_power_option));
   }
 
   std::vector<std::string> input_names =

mace/core/mace_runtime.cc

@@ -15,11 +15,14 @@ void ConfigOpenCLRuntime(GPUPerfHint gpu_perf_hint,
   OpenCLRuntime::Configure(gpu_perf_hint, gpu_priority_hint);
 }
 
-void ConfigOmpThreadsAndAffinity(int omp_num_threads,
-                                 CPUPowerOption power_option) {
-  VLOG(1) << "Config CPU Runtime: omp_num_threads: " << omp_num_threads
-          << ", cpu_power_option: " << power_option;
-  SetOmpThreadsAndAffinity(omp_num_threads, power_option);
+void ConfigOmpThreads(int omp_num_threads) {
+  VLOG(1) << "Config CPU omp_num_threads: " << omp_num_threads;
+  SetOmpThreads(omp_num_threads);
+}
+
+void ConfigCPUPowerOption(CPUPowerOption power_option) {
+  VLOG(1) << "Config CPU power option" << power_option;
+  SetThreadsAffinity(power_option);
 }
 
 };  // namespace mace

mace/core/runtime/cpu/cpu_runtime.cc

@@ -20,7 +20,10 @@ int GetCPUMaxFreq(int cpu_id) {
           "/sys/devices/system/cpu/cpu%d/cpufreq/cpuinfo_max_freq",
           cpu_id);
   FILE *fp = fopen(path, "rb");
-  MACE_CHECK(fp, "File: ", path, " not exists");
+  if (!fp) {
+    LOG(WARNING) << "File: " << path << " not exists.";
+    return 0;
+  }
 
   int freq = 0;
   fscanf(fp, "%d", &freq);
@@ -73,8 +76,21 @@ void SetThreadAffinity(cpu_set_t mask) {
 
 }  // namespace
 
-void SetOmpThreadsAndAffinity(int omp_num_threads,
-                              CPUPowerOption power_option) {
+void SetOmpThreads(int omp_num_threads) {
+  int cpu_count = omp_get_num_procs();
+  if (omp_num_threads > cpu_count) {
+    LOG(WARNING) << "set omp num threads greater than num of cpus can use: "
+                 << cpu_count;
+  }
+  omp_set_num_threads(omp_num_threads);
+}
+
+void SetThreadsAffinity(CPUPowerOption power_option) {
+  // There is no need to set affinity in default mode
+  if (power_option == CPUPowerOption::DEFAULT) {
+    return;
+  }
+
   int cpu_count = omp_get_num_procs();
   std::vector<int> sorted_cpu_ids;
   sorted_cpu_ids.resize(cpu_count);
@@ -82,9 +98,7 @@ void SetOmpThreadsAndAffinity(int omp_num_threads,
   SortCPUIdsByMaxFreqAsc(&sorted_cpu_ids, &big_core_offset);
 
   std::vector<int> use_cpu_ids;
-  if (power_option == CPUPowerOption::DEFAULT) {
-    use_cpu_ids = sorted_cpu_ids;
-  } else if (power_option == CPUPowerOption::HIGH_PERFORMANCE) {
+  if (power_option == CPUPowerOption::HIGH_PERFORMANCE) {
     use_cpu_ids = std::vector<int>(sorted_cpu_ids.begin() + big_core_offset,
                                    sorted_cpu_ids.end());
   } else {
@@ -96,12 +110,6 @@ void SetOmpThreadsAndAffinity(int omp_num_threads,
     }
   }
 
-  if (omp_num_threads > use_cpu_ids.size()) {
-    LOG(WARNING) << "set omp num threads greater than num of cpus can use: "
-                 << use_cpu_ids.size();
-  }
-  omp_set_num_threads(omp_num_threads);
-
   // compute mask
   cpu_set_t mask;
   CPU_ZERO(&mask);
@@ -110,6 +118,7 @@ void SetOmpThreadsAndAffinity(int omp_num_threads,
   }
   VLOG(3) << "Set cpu affinity with mask: " << mask.__bits[0];
 
+  int omp_num_threads = omp_get_max_threads();
 #pragma omp parallel for
   for (int i = 0; i < omp_num_threads; ++i) {
     SetThreadAffinity(mask);

mace/core/runtime/cpu/cpu_runtime.h

@@ -10,8 +10,9 @@
 
 namespace mace {
 
-void SetOmpThreadsAndAffinity(int omp_num_threads,
-                              CPUPowerOption power_option);
+void SetOmpThreads(int omp_num_threads);
+
+void SetThreadsAffinity(CPUPowerOption power_option);
 
 }
 

mace/core/testing/test_benchmark_main.cc

@@ -21,12 +21,12 @@ int main(int argc, char **argv) {
   gflags::ParseCommandLineFlags(&argc, &argv, true);
 
   // config runtime
+  mace::ConfigOmpThreads(FLAGS_omp_num_threads);
+  mace::ConfigCPUPowerOption(
+      static_cast<mace::CPUPowerOption>(FLAGS_cpu_power_option));
   mace::ConfigOpenCLRuntime(
       static_cast<mace::GPUPerfHint>(FLAGS_gpu_perf_hint),
       static_cast<mace::GPUPriorityHint>(FLAGS_gpu_priority_hint));
-  mace::ConfigOmpThreadsAndAffinity(
-      FLAGS_omp_num_threads,
-      static_cast<mace::CPUPowerOption>(FLAGS_cpu_power_option));
 
   mace::testing::Benchmark::Run(FLAGS_pattern.c_str());
   return 0;

mace/examples/mace_run.cc

@@ -196,14 +196,13 @@ bool RunModel(const std::vector<std::string> &input_names,
   LOG(INFO) << "Runing with device type: " << device_type;
 
   // config runtime
+  mace::ConfigOmpThreads(FLAGS_omp_num_threads);
+  mace::ConfigCPUPowerOption(
+      static_cast<CPUPowerOption>(FLAGS_cpu_power_option));
   if (device_type == DeviceType::OPENCL) {
     mace::ConfigOpenCLRuntime(
         static_cast<GPUPerfHint>(FLAGS_gpu_perf_hint),
         static_cast<GPUPriorityHint>(FLAGS_gpu_priority_hint));
-  } else if (device_type == DeviceType::CPU) {
-    mace::ConfigOmpThreadsAndAffinity(
-        FLAGS_omp_num_threads,
-        static_cast<CPUPowerOption>(FLAGS_cpu_power_option));
   }
 
   // Init model

mace/public/mace_runtime.h

@@ -27,8 +27,8 @@ enum GPUPriorityHint {
 enum CPUPowerOption { DEFAULT = 0, HIGH_PERFORMANCE = 1, BATTERY_SAVE = 2 };
 
 void ConfigOpenCLRuntime(GPUPerfHint, GPUPriorityHint);
-void ConfigOmpThreadsAndAffinity(int omp_num_threads,
-                                 CPUPowerOption power_option);
+void ConfigOmpThreads(int omp_num_threads);
+void ConfigCPUPowerOption(CPUPowerOption power_option);
 
 }  // namespace mace