Handle case of CUDA driver available but no GPUs (#399)

Summary:
Pull Request resolved: #399

Fixes #393.
Fixes pytorch/pytorch#60578.

Reviewed By: beauby

Differential Revision: D29363848

fbshipit-source-id: 9007cdbdbb76ec81842965a6c48dbf525f69f185

Author: lw

tensorpipe/common/cuda_lib.h

@@ -33,6 +33,13 @@
 
 namespace tensorpipe {
 
+class NoDevicesError final : public BaseError {
+ public:
+  std::string what() const override {
+    return "The CUDA driver failed to init because it didn't find any device";
+  }
+};
+
 // Master list of all symbols we care about from libcuda.
 
 #define TP_FORALL_CUDA_SYMBOLS(_)                               \
@@ -114,7 +121,14 @@ class CudaLib {
   }
     TP_FORALL_CUDA_SYMBOLS(TP_LOAD_SYMBOL)
 #undef TP_LOAD_SYMBOL
-    TP_CUDA_DRIVER_CHECK(lib, lib.init(0));
+    CUresult result = lib.init(0);
+    // If the driver doesn't find any devices it fails to init (beats me why)
+    // but we must support this case, by disabling the channels, rather than
+    // throwing. Hence we treat it as if we couldn't find the driver.
+    if (result == CUDA_ERROR_NO_DEVICE) {
+      return std::make_tuple(TP_CREATE_ERROR(NoDevicesError), CudaLib());
+    }
+    TP_CUDA_DRIVER_CHECK(lib, result);
     return std::make_tuple(Error::kSuccess, std::move(lib));
   }