cast to ArgMin/ArgMax based on reduce_op (#1851)

* add ArgMin, ArgMax

* change to scale

* format

* add exception

src/ngraph/runtime/gpu/gpu_emitter.cpp

@@ -176,15 +176,29 @@ void runtime::gpu::GPU_Emitter::emit_ArgMin(EMIT_ARGS)
         external_function, writer, node, args, out, reduce_op);
 }
 
-void runtime::gpu::GPU_Emitter::emit_ArgReduce(EMIT_ARGS, cudnnReduceTensorOp_t reduce_mode)
+void runtime::gpu::GPU_Emitter::emit_ArgReduce(EMIT_ARGS, cudnnReduceTensorOp_t reduce_op)
 {
     if (out[0].get_size() == 0)
     {
         return;
     }
-    auto argmax = static_cast<const ngraph::op::ArgMax*>(node);
-    std::vector<size_t> axes{argmax->get_reduction_axis()};
-    auto axis_set = AxisSet(axes);
+
+    size_t axis;
+    if (reduce_op == CUDNN_REDUCE_TENSOR_MIN)
+    {
+        auto argmin = static_cast<const ngraph::op::ArgMin*>(node);
+        axis = argmin->get_reduction_axis();
+    }
+    else if (reduce_op == CUDNN_REDUCE_TENSOR_MAX)
+    {
+        auto argmax = static_cast<const ngraph::op::ArgMax*>(node);
+        axis = argmax->get_reduction_axis();
+    }
+    else
+    {
+        throw std::runtime_error("Not supported. Only Min/Max op are supported by ArgReduce.");
+    }
+    auto axis_set = AxisSet{axis};
 
     std::vector<element::Type> dtypes{args[0].get_element_type(), out[0].get_element_type()};
 
@@ -192,7 +206,7 @@ void runtime::gpu::GPU_Emitter::emit_ArgReduce(EMIT_ARGS, cudnnReduceTensorOp_t
     {
         auto& cudnn_emitter = external_function->get_primitive_emitter()->get_cudnn_emitter();
 
-        auto index = cudnn_emitter->build_reduce_forward(reduce_mode,
+        auto index = cudnn_emitter->build_reduce_forward(reduce_op,
                                                          dtypes,
                                                          args[0].get_shape(),
                                                          axis_set,