add cuda version display (#13885)

test=develop

paddle/fluid/platform/device_context.cc

@@ -198,9 +198,9 @@ class CudnnHolder {
 CUDADeviceContext::CUDADeviceContext(CUDAPlace place)
     : place_(place), cudnn_holder_(nullptr) {
   SetDeviceId(place_.device);
-  compute_capability = GetCUDAComputeCapability(place_.device);
-  multi_process = GetCUDAMultiProcessors(place_.device);
-  max_threads_per_mp = GetCUDAMaxThreadsPerMultiProcessor(place_.device);
+  compute_capability_ = GetCUDAComputeCapability(place_.device);
+  multi_process_ = GetCUDAMultiProcessors(place_.device);
+  max_threads_per_mp_ = GetCUDAMaxThreadsPerMultiProcessor(place_.device);
   PADDLE_ENFORCE(cudaStreamCreate(&stream_));
   eigen_stream_.reset(new EigenCudaStreamDevice());
   eigen_stream_->Reinitialize(&stream_, place);
@@ -211,6 +211,16 @@ CUDADeviceContext::CUDADeviceContext(CUDAPlace place)
     cudnn_holder_.reset(new CudnnHolder(&stream_, place));
   }
 
+  driver_version_ = GetCUDADriverVersion(place_.device);
+  runtime_version_ = GetCUDARuntimeVersion(place_.device);
+
+  LOG(INFO) << "device: " << place_.device
+            << ", CUDA Capability: " << compute_capability_
+            << ", Driver Version: " << driver_version_ / 1000 << "."
+            << (driver_version_ % 100) / 10
+            << ", Runtime Version: " << runtime_version_ / 1000 << "."
+            << (runtime_version_ % 100) / 10;
+
   callback_manager_.reset(new StreamCallbackManager(stream_));
 }
 
@@ -232,11 +242,11 @@ void CUDADeviceContext::Wait() const {
 }
 
 int CUDADeviceContext::GetComputeCapability() const {
-  return compute_capability;
+  return compute_capability_;
 }
 
 int CUDADeviceContext::GetMaxPhysicalThreadCount() const {
-  return multi_process * max_threads_per_mp;
+  return multi_process_ * max_threads_per_mp_;
 }
 
 Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {

paddle/fluid/platform/device_context.h

@@ -135,9 +135,11 @@ class CUDADeviceContext : public DeviceContext {
   cudaStream_t stream_;
   cublasHandle_t cublas_handle_;
 
-  int compute_capability;
-  int multi_process;
-  int max_threads_per_mp;
+  int compute_capability_;
+  int runtime_version_;
+  int driver_version_;
+  int multi_process_;
+  int max_threads_per_mp_;
 
   mutable std::mutex mtx_;
 

paddle/fluid/platform/gpu_info.cc

@@ -46,6 +46,24 @@ int GetCUDAComputeCapability(int id) {
   return device_prop.major * 10 + device_prop.minor;
 }
 
+int GetCUDARuntimeVersion(int id) {
+  PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
+  int runtime_version = 0;
+  PADDLE_ENFORCE(cudaRuntimeGetVersion(&runtime_version),
+                 "cudaRuntimeGetVersion failed in "
+                 "paddle::platform::cudaRuntimeGetVersion");
+  return runtime_version;
+}
+
+int GetCUDADriverVersion(int id) {
+  PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
+  int driver_version = 0;
+  PADDLE_ENFORCE(cudaDriverGetVersion(&driver_version),
+                 "cudaDriverGetVersion failed in "
+                 "paddle::platform::GetCUDADriverVersion");
+  return driver_version;
+}
+
 int GetCUDAMultiProcessors(int id) {
   PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
   int count;

paddle/fluid/platform/gpu_info.h

@@ -29,6 +29,12 @@ int GetCUDADeviceCount();
 //! Get the compute capability of the ith GPU (format: major * 10 + minor)
 int GetCUDAComputeCapability(int i);
 
+//! Get the runtime version of the ith GPU
+int GetCUDARuntimeVersion(int id);
+
+//! Get the driver version of the ith GPU
+int GetCUDADriverVersion(int id);
+
 //! Get the MultiProcessors of the ith GPU.
 int GetCUDAMultiProcessors(int i);