improve cudnn conv type error msg

Summary: CuDNN conv op's type error was not very descriptive.

Reviewed By: Yangqing

Differential Revision: D5124638

fbshipit-source-id: 7d3f0afad36573cdb97d1f8ec3c60a9c6d87f926
diff --git a/caffe2/operators/conv_op_cudnn.cc b/caffe2/operators/conv_op_cudnn.cc
index 13f8490..32c2cc7 100644
--- a/caffe2/operators/conv_op_cudnn.cc
+++ b/caffe2/operators/conv_op_cudnn.cc
@@ -526,7 +526,9 @@
                          float,      // Math
                          float16>();   // Y
   } else {
-    LOG(FATAL) << "Unsupported type inputs";
+    LOG(FATAL) << "Only float (32bit) and float16 are supported by "
+               << "cudnn convolution, but input " << def().input(0) << " has ["
+               << Input(0).meta().name() << "]";
   }
   return true;
 }