Merge branch 'perf_config' into 'master'

fix runtime config

See merge request !298

.gitlab-ci.yml

@@ -23,4 +23,4 @@ ops_benchmark:
   only:
     - master
   script:
-    - FAILURE_PATTERN="Aborted" tools/bazel-adb-run.sh //mace/ops:ops_benchmark .*CONV.*
+    - FAILURE_PATTERN="Aborted" tools/bazel-adb-run.sh //mace/ops:ops_benchmark --pattern=.*CONV.*

mace/benchmark/benchmark_model.cc

@@ -205,8 +205,8 @@ DEFINE_bool(show_flops, true, "whether to estimate the model's FLOPs");
 DEFINE_int32(warmup_runs, 1, "how many runs to initialize model");
 DEFINE_string(model_data_file, "", "model data file name, used when EMBED_MODEL_DATA set to 0");
 DEFINE_string(gpu_type, "ADRENO", "ADRENO/MALI");
-DEFINE_int32(gpu_perf_hint, 2, "0:NA/1:LOW/2:NORMAL/3:HIGH");
-DEFINE_int32(gpu_priority_hint, 1, "0:NA/1:LOW/2:NORMAL/3:HIGH");
+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(cpu_power_option, 0, "0:DEFAULT/1:HIGH_PERFORMANCE/2:BATTERY_SAVE");
 
@@ -266,7 +266,7 @@ int Main(int argc, char **argv) {
         static_cast<GPUPriorityHint>(FLAGS_gpu_priority_hint));
   }
   else if (device_type == CPU) {
-    mace::ConfigCPURuntime(
+    mace::ConfigOmpThreadsAndAffinity(
         FLAGS_omp_num_threads,
         static_cast<CPUPowerOption>(FLAGS_cpu_power_option));
   }

mace/core/BUILD

@@ -85,6 +85,7 @@ cc_library(
     ],
     deps = [
         ":core",
+        "//external:gflags_nothreads",
         "//mace/utils",
     ],
     alwayslink = 1,

mace/core/mace.cc

@@ -360,10 +360,11 @@ void ConfigOpenCLRuntime(GPUType gpu_type,
   OpenCLRuntime::CreateGlobal(gpu_type, gpu_perf_hint, gpu_priority_hint);
 }
 
-void ConfigCPURuntime(int omp_num_threads, CPUPowerOption power_option) {
+void ConfigOmpThreadsAndAffinity(int omp_num_threads,
+                                 CPUPowerOption power_option) {
   LOG(INFO) << "Config CPU Runtime: omp_num_threads: " << omp_num_threads
             << ", cpu_power_option: " << power_option;
-  SetCPURuntime(omp_num_threads, power_option);
+  SetOmpThreadsAndAffinity(omp_num_threads, power_option);
 }
 
 // Mace Engine

mace/core/runtime/cpu/cpu_runtime.cc

@@ -20,7 +20,7 @@ int GetCPUMaxFreq(int cpu_id) {
           "/sys/devices/system/cpu/cpu%d/cpufreq/cpuinfo_max_freq",
           cpu_id);
   FILE *fp = fopen(path, "rb");
-  if (!fp) return 0;
+  MACE_CHECK(fp, "File: ", path, " not exists");
 
   int freq = 0;
   fscanf(fp, "%d", &freq);
@@ -28,7 +28,7 @@ int GetCPUMaxFreq(int cpu_id) {
   return freq;
 }
 
-void SortCPUIdsByMaxFreqAsc(std::vector<int> *cpu_ids) {
+void SortCPUIdsByMaxFreqAsc(std::vector<int> *cpu_ids, int *big_core_offset) {
   MACE_CHECK_NOTNULL(cpu_ids);
   int cpu_count = cpu_ids->size();
   std::vector<int> cpu_max_freq;
@@ -54,51 +54,61 @@ void SortCPUIdsByMaxFreqAsc(std::vector<int> *cpu_ids) {
       }
     }
   }
+
+  *big_core_offset = 0;
+  for (int i = 1; i < cpu_count; ++i) {
+    if (cpu_max_freq[i] > cpu_max_freq[i - 1]) {
+      *big_core_offset = i;
+      break;
+    }
+  }
 }
 
 void SetThreadAffinity(cpu_set_t mask) {
   int sys_call_res;
   pid_t pid = gettid();
-
-  // TODO(chenghui): when set omp num threads to 1,
-  // sometiomes return EINVAL(22) error.
-  // https://linux.die.net/man/2/sched_setaffinity
-  sys_call_res = syscall(__NR_sched_setaffinity, pid, sizeof(mask), &mask);
-  if (sys_call_res != 0) {
-    LOG(FATAL) << "syscall setaffinity error: " << sys_call_res << ' ' << errno;
-  }
+  int err = sched_setaffinity(pid, sizeof(mask), &mask);
+  MACE_CHECK(err == 0, "set affinity error: ", errno);
 }
 
 }  // namespace
 
-void SetCPURuntime(int omp_num_threads, CPUPowerOption power_option) {
+void SetOmpThreadsAndAffinity(int omp_num_threads,
+                              CPUPowerOption power_option) {
   int cpu_count = omp_get_num_procs();
-  LOG(INFO) << "cpu_count: " << cpu_count;
   std::vector<int> sorted_cpu_ids;
   sorted_cpu_ids.resize(cpu_count);
-  SortCPUIdsByMaxFreqAsc(&sorted_cpu_ids);
+  int big_core_offset;
+  SortCPUIdsByMaxFreqAsc(&sorted_cpu_ids, &big_core_offset);
 
   std::vector<int> use_cpu_ids;
-  if (power_option == CPUPowerOption::DEFAULT || omp_num_threads >= cpu_count) {
+  if (power_option == CPUPowerOption::DEFAULT) {
     use_cpu_ids = sorted_cpu_ids;
-    omp_num_threads = cpu_count;
   } else if (power_option == CPUPowerOption::HIGH_PERFORMANCE) {
-    use_cpu_ids =
-        std::vector<int>(sorted_cpu_ids.begin() + cpu_count - omp_num_threads,
+    use_cpu_ids = std::vector<int>(sorted_cpu_ids.begin() + big_core_offset,
                                    sorted_cpu_ids.end());
   } else {
+    if (big_core_offset > 0) {
       use_cpu_ids = std::vector<int>(sorted_cpu_ids.begin(),
-                                   sorted_cpu_ids.begin() + omp_num_threads);
+                                     sorted_cpu_ids.begin() + big_core_offset);
+    } else {
+      use_cpu_ids = sorted_cpu_ids;
+    }
   }
 
+  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);
   for (auto cpu_id : use_cpu_ids) {
     CPU_SET(cpu_id, &mask);
   }
-  LOG(INFO) << "use cpus mask: " << mask.__bits[0];
+  VLOG(3) << "Set cpu affinity with mask: " << mask.__bits[0];
 
 #pragma omp parallel for
   for (int i = 0; i < omp_num_threads; ++i) {

mace/core/runtime/cpu/cpu_runtime.h

@@ -10,7 +10,8 @@
 
 namespace mace {
 
-void SetCPURuntime(int omp_num_threads, CPUPowerOption power_option);
+void SetOmpThreadsAndAffinity(int omp_num_threads,
+                              CPUPowerOption power_option);
 
 }
 

mace/core/runtime/opencl/opencl_runtime.cc

@@ -65,20 +65,24 @@ void OpenCLProfilingTimer::ClearTiming() {
   accumulated_micros_ = 0;
 }
 
+std::unique_ptr<OpenCLRuntime> OpenCLRuntime::runtime_instance_ = nullptr;
+
 OpenCLRuntime *OpenCLRuntime::Global() {
-  if (opencl_runtime_instance == nullptr) {
-    return CreateGlobal(GPUType::ADRENO, GPUPerfHint::PERF_NORMAL,
-                        GPUPriorityHint::PRIORITY_LOW);
+  // FIXME: not thread safe
+  if (runtime_instance_ == nullptr) {
+    return CreateGlobal(GPUType::ADRENO, GPUPerfHint::PERF_DEFAULT,
+                        GPUPriorityHint::PRIORITY_DEFAULT);
   }
-  return opencl_runtime_instance;
+  return runtime_instance_.get();
 }
 
 OpenCLRuntime *OpenCLRuntime::CreateGlobal(GPUType gpu_type,
                                            GPUPerfHint gpu_perf_hint,
                                            GPUPriorityHint gpu_priority_hint) {
-  opencl_runtime_instance = new OpenCLRuntime(gpu_type, gpu_perf_hint,
-                                              gpu_priority_hint);
-  return opencl_runtime_instance;
+  runtime_instance_ =
+      std::unique_ptr<OpenCLRuntime>(new OpenCLRuntime(gpu_type, gpu_perf_hint,
+                                                       gpu_priority_hint));
+  return runtime_instance_.get();
 }
 
 void ParseOpenCLRuntimeConfig(std::vector<cl_context_properties> *properties,

mace/core/runtime/opencl/opencl_runtime.h

@@ -52,10 +52,10 @@ class OpenCLRuntime {
   cl::Kernel BuildKernel(const std::string &program_name,
                          const std::string &kernel_name,
                          const std::set<std::string> &build_options);
+  ~OpenCLRuntime();
 
  private:
   OpenCLRuntime(GPUType, GPUPerfHint, GPUPriorityHint);
-  ~OpenCLRuntime();
   OpenCLRuntime(const OpenCLRuntime &) = delete;
   OpenCLRuntime &operator=(const OpenCLRuntime &) = delete;
 
@@ -74,9 +74,9 @@ class OpenCLRuntime {
   std::map<std::string, cl::Program> built_program_map_;
   std::mutex program_build_mutex_;
   std::string kernel_path_;
+  static std::unique_ptr<OpenCLRuntime> runtime_instance_;
 };
 
-static OpenCLRuntime *opencl_runtime_instance = nullptr;
 }  // namespace mace
 
 #endif  // MACE_CORE_RUNTIME_OPENCL_OPENCL_RUNTIME_H_

mace/core/testing/test_benchmark_main.cc

@@ -4,20 +4,42 @@
 
 #include <iostream>
 
+#include "gflags/gflags.h"
 #include "mace/core/testing/test_benchmark.h"
 #include "mace/public/mace.h"
 
-int main(int argc, char **argv) {
-  std::cout << "Running main() from test_main.cc\n";
-
-  mace::ConfigCPURuntime(4, mace::CPUPowerOption::HIGH_PERFORMANCE);
-  mace::ConfigOpenCLRuntime(mace::GPUType::ADRENO, mace::GPUPerfHint::PERF_HIGH,
-                            mace::GPUPriorityHint::PRIORITY_HIGH);
+DEFINE_string(pattern, "all", "op benchmark pattern, eg:.*CONV.*");
+DEFINE_string(gpu_type, "ADRENO", "ADRENO/MALI");
+DEFINE_int32(gpu_perf_hint, 3, "0:DEFAULT/1:LOW/2:NORMAL/3:HIGH");
+DEFINE_int32(gpu_priority_hint, 3, "0:DEFAULT/1:LOW/2:NORMAL/3:HIGH");
+DEFINE_int32(omp_num_threads, 1, "num of openmp threads");
+DEFINE_int32(cpu_power_option, 1,
+             "0:DEFAULT/1:HIGH_PERFORMANCE/2:BATTERY_SAVE");
 
-  if (argc == 2) {
-    mace::testing::Benchmark::Run(argv[1]);
+mace::GPUType ParseGPUType(const std::string &gpu_type_str) {
+  if (gpu_type_str.compare("ADRENO") == 0) {
+    return mace::GPUType::ADRENO;
+  } else if (gpu_type_str.compare("MALI") == 0) {
+    return mace::GPUType::MALI;
   } else {
-    mace::testing::Benchmark::Run("all");
+    return mace::GPUType::ADRENO;
   }
+}
+
+int main(int argc, char **argv) {
+  gflags::SetUsageMessage("some usage message");
+  gflags::ParseCommandLineFlags(&argc, &argv, true);
+
+  // config runtime
+  mace::GPUType gpu_type = ParseGPUType(FLAGS_gpu_type);
+  mace::ConfigOpenCLRuntime(
+      gpu_type,
+      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

@@ -171,8 +171,8 @@ DEFINE_int32(round, 1, "round");
 DEFINE_int32(restart_round, 1, "restart round");
 DEFINE_int32(malloc_check_cycle, -1, "malloc debug check cycle, -1 to disable");
 DEFINE_string(gpu_type, "ADRENO", "ADRENO/MALI");
-DEFINE_int32(gpu_perf_hint, 2, "0:NA/1:LOW/2:NORMAL/3:HIGH");
-DEFINE_int32(gpu_priority_hint, 1, "0:NA/1:LOW/2:NORMAL/3:HIGH");
+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(cpu_power_option, 0, "0:DEFAULT/1:HIGH_PERFORMANCE/2:BATTERY_SAVE");
 
@@ -199,7 +199,7 @@ bool SingleInputAndOutput(const std::vector<int64_t> &input_shape,
         static_cast<GPUPriorityHint>(FLAGS_gpu_priority_hint));
   }
   else if (device_type == DeviceType::CPU) {
-    mace::ConfigCPURuntime(
+    mace::ConfigOmpThreadsAndAffinity(
         FLAGS_omp_num_threads,
         static_cast<CPUPowerOption>(FLAGS_cpu_power_option));
   }
@@ -304,7 +304,7 @@ bool MultipleInputOrOutput(const std::vector<std::string> &input_names,
         static_cast<GPUPriorityHint>(FLAGS_gpu_priority_hint));
   }
   else if (device_type == DeviceType::CPU) {
-    mace::ConfigCPURuntime(
+    mace::ConfigOmpThreadsAndAffinity(
         FLAGS_omp_num_threads,
         static_cast<CPUPowerOption>(FLAGS_cpu_power_option));
   }

mace/public/mace.h

@@ -62,9 +62,14 @@ enum DataType {
 };
 
 enum GPUType { ADRENO = 0, MALI = 1 };
-enum GPUPerfHint { PERF_NA = 0, PERF_LOW = 1, PERF_NORMAL = 2, PERF_HIGH = 3 };
+enum GPUPerfHint {
+  PERF_DEFAULT = 0,
+  PERF_LOW = 1,
+  PERF_NORMAL = 2,
+  PERF_HIGH = 3
+};
 enum GPUPriorityHint {
-  PRIORITY_NA = 0,
+  PRIORITY_DEFAULT = 0,
   PRIORITY_LOW = 1,
   PRIORITY_NORMAL = 2,
   PRIORITY_HIGH = 3
@@ -381,7 +386,8 @@ struct MaceInputInfo {
 };
 
 void ConfigOpenCLRuntime(GPUType, GPUPerfHint, GPUPriorityHint);
-void ConfigCPURuntime(int omp_num_threads, CPUPowerOption power_option);
+void ConfigOmpThreadsAndAffinity(int omp_num_threads,
+                                 CPUPowerOption power_option);
 
 class MaceEngine {
  public: