Make init device on all gpu by default (#7345)

* "init use all default devices"

* "fix init test"

paddle/framework/device_data_transform_test.cu

@@ -105,8 +105,7 @@ static void BuildVar(const std::string& param_name,
 TEST(Operator, CPUtoGPU) {
   using namespace paddle::framework;
   using namespace paddle::platform;
-
-  ASSERT_EQ(InitDevices({"CPU", "GPU:0"}), true);
+  InitDevices();
 
   paddle::framework::Scope scope;
   paddle::platform::CPUPlace cpu_place;

paddle/framework/init.cc

@@ -40,40 +40,23 @@ void InitGflags(std::vector<std::string> &argv) {
   });
 }
 
-bool InitDevices(const std::vector<std::string> &devices) {
-  // device format
-  // CPU
-  // GPU:1
-  // TODO(dzhwinter) : add device format annotation for users.
+void InitDevices() {
+  /*Init all avaiable devices by default */
+
   std::vector<platform::Place> places;
-  for (auto &device : devices) {
-    auto p = string::Piece(device);
-    if (string::HasPrefix(p, "CPU")) {
   places.emplace_back(platform::CPUPlace());
-    } else if (string::HasPrefix(p, "GPU")) {
+
 #ifdef PADDLE_WITH_CUDA
-      auto pos = string::RFind(p, ':', string::Piece::npos);
-      auto number = device.substr(pos + 1);
-      places.emplace_back(platform::CUDAPlace(std::stoi(number)));
+  int count = platform::GetCUDADeviceCount();
+  for (int i = 0; i < count; ++i) {
+    places.emplace_back(platform::CUDAPlace(i));
+  }
 #else
   LOG(WARNING)
       << "'GPU' is not supported, Please re-compile with WITH_GPU option";
 #endif
-    } else {
-      return false;
-    }
-  }
 
-  if (std::find_if(places.begin(), places.end(),
-                   [&](const platform::Place &place) {
-                     return platform::is_cpu_place(place);
-                   }) == places.end()) {
-    places.emplace_back(platform::CPUPlace());
-    LOG(WARNING) << "Not specified CPU device, create CPU by Default.";
-  }
   platform::DeviceContextPool::Init(places);
-  // framework::UseALL();
-  return true;
 }
 
 void InitGLOG(const std::string &prog_name) {

paddle/framework/init.h

@@ -24,7 +24,7 @@ void InitGflags(std::vector<std::string> &argv);
 
 void InitGLOG(const std::string &prog_name);
 
-bool InitDevices(const std::vector<std::string> &devices);
+void InitDevices();
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/init_test.cc

@@ -14,18 +14,13 @@ limitations under the License. */
 #include "gtest/gtest.h"
 
 #include "paddle/framework/init.h"
+#include "paddle/platform/device_context.h"
 
-TEST(Init, InitDevices) {
+TEST(InitDevices, CPU) {
   using paddle::framework::InitDevices;
-  std::vector<std::string> ds1 = {"CPU"};
-  ASSERT_EQ(InitDevices(ds1), true);
+  using paddle::platform::DeviceContextPool;
 
-#ifdef PADDLE_WITH_CUDA
-  std::vector<std::string> ds2 = {"CPU", "GPU:0", "GPU:1"};
-  ASSERT_EQ(InitDevices(ds2), true);
-
-  // test re-init
-  std::vector<std::string> ds3 = {"GPU:0", "GPU:1"};
-  ASSERT_EQ(InitDevices(ds3), true);
-#endif
+  InitDevices();
+  DeviceContextPool& pool = DeviceContextPool::Instance();
+  ASSERT_GE(pool.size(), 1U);
 }

paddle/framework/operator_test.cc

@@ -69,7 +69,7 @@ REGISTER_OP_WITHOUT_GRADIENT(test_operator,
                              paddle::framework::OpWithoutKernelCheckerMaker);
 
 TEST(OperatorBase, all) {
-  paddle::framework::InitDevices({"CPU"});
+  paddle::framework::InitDevices();
   paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("test_operator");
   BuildVar("input", {"IN1"}, op_desc.add_inputs());
@@ -195,7 +195,7 @@ REGISTER_OP_CPU_KERNEL(op_with_kernel,
 
 // test with single input
 TEST(OpKernel, all) {
-  paddle::framework::InitDevices({"CPU"});
+  paddle::framework::InitDevices();
   paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("op_with_kernel");
   BuildVar("x", {"IN1"}, op_desc.add_inputs());
@@ -225,7 +225,7 @@ REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
 TEST(OpKernel, multi_inputs) {
   using namespace paddle::framework;
 
-  paddle::framework::InitDevices({"CPU"});
+  paddle::framework::InitDevices();
   proto::OpDesc op_desc;
 
   op_desc.set_type("op_multi_inputs_with_kernel");
@@ -264,7 +264,7 @@ class OperatorClone : public paddle::framework::OperatorBase {
 };
 
 TEST(Operator, Clone) {
-  paddle::framework::InitDevices({"CPU"});
+  paddle::framework::InitDevices();
   OperatorClone a("ABC", paddle::framework::VariableNameMap{},
                   paddle::framework::VariableNameMap{},
                   paddle::framework::AttributeMap{});

paddle/inference/inference.cc

@@ -169,7 +169,7 @@ void InferenceEngine::Execute(const std::vector<framework::LoDTensor>& feeds,
   }
 
   auto* place = new platform::CPUPlace();
-  framework::InitDevices({"CPU"});
+  framework::InitDevices();
   framework::Executor* executor = new framework::Executor(*place);
   framework::Scope* scope = new framework::Scope();
 

paddle/platform/device_context.h

@@ -185,6 +185,8 @@ class DeviceContextPool {
         const typename DefaultDeviceContextType<Place>::TYPE*>(Get(place));
   }
 
+  size_t size() const { return device_contexts_.size(); }
+
  private:
   static DeviceContextPool* pool;
   constexpr static int LEFT_SHIFT = 8;

paddle/testing/paddle_gtest_main.cc

@@ -34,11 +34,11 @@ int main(int argc, char** argv) {
   google::ParseCommandLineFlags(&new_argc, &new_argv_address, false);
   testing::InitGoogleTest(&argc, argv);
   paddle::memory::Used(paddle::platform::CPUPlace());
-  std::vector<std::string> devs = {"CPU"};
+
 #ifdef PADDLE_WITH_CUDA
   paddle::memory::Used(paddle::platform::CUDAPlace(0));
-  devs.push_back("GPU:0");
 #endif
-  paddle::framework::InitDevices(devs);
+
+  paddle::framework::InitDevices();
   return RUN_ALL_TESTS();
 }

python/paddle/v2/fluid/__init__.py

@@ -62,11 +62,7 @@ def __bootstrap__():
     core.init_gflags([sys.argv[0]] +
                      ["--tryfromenv=" + ",".join(read_env_flags)])
     core.init_glog(sys.argv[0])
-
-    if core.is_compile_gpu():
-        core.init_devices(["CPU", "GPU:0"])
-    else:
-        core.init_devices(["CPU"])
+    core.init_devices()
 
 
 __bootstrap__()

python/paddle/v2/fluid/tests/test_batch_norm_op.py

@@ -341,9 +341,6 @@ class TestBatchNormOp(OpTest):
         if core.is_compile_gpu() and core.op_support_gpu("batch_norm"):
             places.append(core.CUDAPlace(0))
 
-            core.init_devices(["CPU", "GPU:0"])
-        else:
-            core.init_devices(["CPU"])
         for place in places:
             for data_format in ["NCHW", "NHWC"]:
                 test_with_place(place, data_format, [2, 3, 4, 5])