Make use_cuda() only set to true if there is also a GPU available

Author: davisking

dlib/cuda/cuda_dlib.cu

@@ -44,6 +44,13 @@ namespace dlib
             CHECK_CUDA(cudaSetDeviceFlags(cudaDeviceScheduleBlockingSync));
         }
 
+        bool is_available(
+        ) 
+        {
+            int num_devices;
+            return cudaGetDeviceCount(&num_devices) == cudaSuccess && num_devices > 0;
+        }
+
         int get_num_devices (
         )
         {

dlib/cuda/cuda_dlib.h

@@ -25,6 +25,9 @@ namespace dlib
         int get_num_devices (
         );
 
+        bool is_available (
+        );
+
         std::string get_device_name (
             int device
         );
@@ -892,6 +895,9 @@ namespace dlib
         inline int get_num_devices (
         ) { return 1; }
 
+        inline bool is_available (
+        ) { return false; }
+
         inline std::string get_device_name (
             int device
         ) 

dlib/cuda/tensor_tools.cpp

@@ -22,7 +22,7 @@ namespace dlib
         )
         {
 #ifdef DLIB_USE_CUDA
-            thread_local bool var(true);
+            thread_local bool var(cuda::is_available());
 #else
             thread_local bool var(false);
 #endif

dlib/dnn/core_abstract.h

@@ -202,16 +202,20 @@ namespace dlib
         ensures
             - If dlib should use the CUDA implementation of a deep neural network
               then this function returns true and false otherwise.
-            - On program startup this function will return true if DLIB_USE_CUDA is defined.
+            - On program startup this function will return true if DLIB_USE_CUDA is defined and
+              there is an available GPU device to use.
             - This function always returns false if DLIB_USE_CUDA is not defined.
+            - This function sets a thread local variable.  That is, each thread has its own value
+              for use_cuda().  This means that one thread may use cuda while another thread might
+              not, depending on the setting of use_cuda().
     !*/
 
     void set_use_cuda(
         bool flag
     );
     /*!
         ensures
-            - #use_cuda() == flag
+            - #use_cuda() == flag for the calling thread.
     !*/
 
 // ----------------------------------------------------------------------------------------