Merge branch 'return_threadaffinity' into 'master'

return SetThreadAffinity

See merge request !639

mace/benchmark/benchmark_model.cc

@@ -253,9 +253,12 @@ int Main(int argc, char **argv) {
   mace::DeviceType device_type = ParseDeviceType(FLAGS_device);
 
   // config runtime
-  mace::SetOpenMPThreadPolicy(
+  MaceStatus ret = mace::SetOpenMPThreadPolicy(
       FLAGS_omp_num_threads,
       static_cast<CPUAffinityPolicy >(FLAGS_cpu_affinity_policy));
+  if (ret != MACE_SUCCESS) {
+    LOG(WARNING) << "Set openmp or cpu affinity failed.";
+  }
 #ifdef MACE_ENABLE_OPENCL
   if (device_type == DeviceType::GPU) {
     mace::SetGPUHints(

mace/core/runtime/cpu/cpu_runtime.cc

@@ -77,14 +77,19 @@ int GetCPUMaxFreq(int cpu_id) {
   return freq;
 }
 
-void SetThreadAffinity(cpu_set_t mask) {
+MaceStatus SetThreadAffinity(cpu_set_t mask) {
 #if defined(__ANDROID__)
   pid_t pid = gettid();
 #else
   pid_t pid = syscall(SYS_gettid);
 #endif
   int err = sched_setaffinity(pid, sizeof(mask), &mask);
-  MACE_CHECK(err == 0, "set affinity error: ", strerror(errno));
+  if (err) {
+    LOG(WARNING) << "set affinity error: " << strerror(errno);
+    return MACE_INVALID_ARGS;
+  } else {
+    return MACE_SUCCESS;
+  }
 }
 
 }  // namespace
@@ -129,8 +134,8 @@ MaceStatus GetCPUBigLittleCoreIDs(std::vector<int> *big_core_ids,
   return MACE_SUCCESS;
 }
 
-void SetOpenMPThreadsAndAffinityCPUs(int omp_num_threads,
-                                     const std::vector<int> &cpu_ids) {
+MaceStatus SetOpenMPThreadsAndAffinityCPUs(int omp_num_threads,
+                                           const std::vector<int> &cpu_ids) {
 #ifdef MACE_ENABLE_OPENMP
   VLOG(1) << "Set OpenMP threads number: " << omp_num_threads
           << ", CPU core IDs: " << MakeString(cpu_ids);
@@ -146,17 +151,23 @@ void SetOpenMPThreadsAndAffinityCPUs(int omp_num_threads,
   for (auto cpu_id : cpu_ids) {
     CPU_SET(cpu_id, &mask);
   }
-
 #ifdef MACE_ENABLE_OPENMP
+  std::vector<MaceStatus> status(omp_num_threads);
 #pragma omp parallel for
   for (int i = 0; i < omp_num_threads; ++i) {
     VLOG(1) << "Set affinity for OpenMP thread " << omp_get_thread_num()
             << "/" << omp_get_num_threads();
-    SetThreadAffinity(mask);
+    status[i] = SetThreadAffinity(mask);
+  }
+  for (int i = 0; i < omp_num_threads; ++i) {
+    if (status[i] != MACE_SUCCESS)
+      return MACE_INVALID_ARGS;
   }
+  return MACE_SUCCESS;
 #else
-  SetThreadAffinity(mask);
+  MaceStatus status = SetThreadAffinity(mask);
   VLOG(1) << "Set affinity without OpenMP: " << mask.__bits[0];
+  return status;
 #endif
 }
 
@@ -191,8 +202,8 @@ MaceStatus SetOpenMPThreadsAndAffinityPolicy(int omp_num_threads_hint,
       omp_num_threads_hint > static_cast<int>(use_cpu_ids.size())) {
     omp_num_threads_hint = use_cpu_ids.size();
   }
-  SetOpenMPThreadsAndAffinityCPUs(omp_num_threads_hint, use_cpu_ids);
-  return MACE_SUCCESS;
+
+  return SetOpenMPThreadsAndAffinityCPUs(omp_num_threads_hint, use_cpu_ids);
 }
 
 MaceStatus SetOpenMPThreadPolicy(int num_threads_hint,
@@ -202,7 +213,8 @@ MaceStatus SetOpenMPThreadPolicy(int num_threads_hint,
   return SetOpenMPThreadsAndAffinityPolicy(num_threads_hint, policy);
 }
 
-void SetOpenMPThreadAffinity(int num_threads, const std::vector<int> &cpu_ids) {
+MaceStatus SetOpenMPThreadAffinity(int num_threads,
+                                   const std::vector<int> &cpu_ids) {
   return SetOpenMPThreadsAndAffinityCPUs(num_threads, cpu_ids);
 }
 

mace/core/runtime/cpu/cpu_runtime.h

@@ -25,8 +25,8 @@ namespace mace {
 MaceStatus GetCPUBigLittleCoreIDs(std::vector<int> *big_core_ids,
                                   std::vector<int> *little_core_ids);
 
-void SetOpenMPThreadsAndAffinityCPUs(int omp_num_threads,
-                                     const std::vector<int> &cpu_ids);
+MaceStatus SetOpenMPThreadsAndAffinityCPUs(int omp_num_threads,
+                                           const std::vector<int> &cpu_ids);
 
 MaceStatus SetOpenMPThreadsAndAffinityPolicy(int omp_num_threads_hint,
                                              CPUAffinityPolicy policy);

mace/core/testing/test_benchmark_main.cc

@@ -18,6 +18,7 @@
 #include "mace/core/testing/test_benchmark.h"
 #include "mace/public/mace.h"
 #include "mace/public/mace_runtime.h"
+#include "mace/utils/logging.h"
 
 DEFINE_string(filter, "all", "op benchmark regex filter, eg:.*CONV.*");
 DEFINE_int32(gpu_perf_hint, 3, "0:DEFAULT/1:LOW/2:NORMAL/3:HIGH");
@@ -33,9 +34,12 @@ int main(int argc, char **argv) {
   gflags::ParseCommandLineFlags(&argc, &argv, true);
 
   // config runtime
-  mace::SetOpenMPThreadPolicy(
+  mace::MaceStatus status = mace::SetOpenMPThreadPolicy(
       FLAGS_omp_num_threads,
       static_cast<mace::CPUAffinityPolicy >(FLAGS_cpu_affinity_policy));
+  if (status != mace::MACE_SUCCESS) {
+    LOG(WARNING) << "Set openmp or cpu affinity failed.";
+  }
   mace::SetGPUHints(
       static_cast<mace::GPUPerfHint>(FLAGS_gpu_perf_hint),
       static_cast<mace::GPUPriorityHint>(FLAGS_gpu_priority_hint));

mace/examples/android/app/build.gradle

@@ -4,7 +4,7 @@ android {
     compileSdkVersion 26
     defaultConfig {
         applicationId "com.xiaomi.mace.demo"
-        minSdkVersion 23
+        minSdkVersion 19
         targetSdkVersion 26
         versionCode 1
         versionName "1.0"

mace/examples/android/app/src/main/java/com/xiaomi/mace/demo/result/InitData.java

@@ -34,7 +34,7 @@ public class InitData {
     public InitData() {
         model = MODELS[0];
         ompNumThreads = 2;
-        cpuAffinityPolicy = 1;
+        cpuAffinityPolicy = 0;
         gpuPerfHint = 3;
         gpuPriorityHint = 3;
         device = DEVICES[0];

mace/examples/android/build.sh

@@ -5,8 +5,8 @@ set -e -u -o pipefail
 pushd ../../../
 python tools/converter.py build --config=docs/getting_started/models/demo_app_models.yaml
 
-cp -r builds/mobilenet/include/ mace/examples/android/macelibrary/src/main/cpp/
-cp -r builds/mobilenet/lib/ mace/examples/android/macelibrary/src/main/cpp/
+cp -r builds/mobilenet/include mace/examples/android/macelibrary/src/main/cpp/
+cp -r builds/mobilenet/lib mace/examples/android/macelibrary/src/main/cpp/
 
 popd
 

mace/examples/android/macelibrary/build.gradle

@@ -9,7 +9,7 @@ android {
     }
 
     defaultConfig {
-        minSdkVersion 23
+        minSdkVersion 19
         targetSdkVersion 26
         versionCode 1
         versionName "1.0"

mace/public/mace_runtime.h

@@ -125,7 +125,8 @@ MaceStatus SetOpenMPThreadPolicy(int num_threads_hint,
 // affinity to offline cores may run very slow or unexpectedly. In such cases,
 // please use SetOpenMPThreadPolicy with default policy instead.
 __attribute__((visibility("default")))
-void SetOpenMPThreadAffinity(int num_threads, const std::vector<int> &cpu_ids);
+MaceStatus SetOpenMPThreadAffinity(int num_threads,
+                                   const std::vector<int> &cpu_ids);
 
 // Get ARM big.LITTLE configuration.
 //

mace/tools/validation/mace_run.cc

@@ -198,9 +198,12 @@ bool RunModel(const std::string &model_name,
               const std::vector<std::vector<int64_t>> &output_shapes) {
   DeviceType device_type = ParseDeviceType(FLAGS_device);
   // config runtime
-  mace::SetOpenMPThreadPolicy(
+  MaceStatus status = mace::SetOpenMPThreadPolicy(
       FLAGS_omp_num_threads,
       static_cast<CPUAffinityPolicy >(FLAGS_cpu_affinity_policy));
+  if (status != MACE_SUCCESS) {
+    LOG(WARNING) << "Set openmp or cpu affinity failed.";
+  }
 #ifdef MACE_ENABLE_OPENCL
   if (device_type == DeviceType::GPU) {
     mace::SetGPUHints(