Fix incorrect stride calculation in reduce_mean

src/targets/gpu/device/include/migraphx/gpu/device/reduce.hpp

@@ -282,10 +282,7 @@ void reduce(hipStream_t stream,
                   std::prev(output_shape.lens().end()),
                   input_shape.lens().begin()))
     {
-        std::size_t stride = std::accumulate(input_shape.strides().begin(),
-                                             input_shape.strides().end(),
-                                             1,
-                                             std::multiplies<size_t>());
+        std::size_t stride = input_shape.strides().at(input_shape.strides().size() - 2);
         reduce_standard_impl(stream,
                              result,
                              arg,

test/gpu/miopen.cpp

@@ -3792,6 +3792,18 @@ struct test_reduce_mean : verify_program<test_reduce_mean>
     };
 };
 
+struct test_reduce_mean2 : verify_program<test_reduce_mean2>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {1, 128, 768}};
+        auto x = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::reduce_mean{{2}}, x);
+        return p;
+    };
+};
+
 struct test_reduce_mean_int : verify_program<test_reduce_mean_int>
 {
     migraphx::program create_program() const