[GPU] Add condition which empty layout is regareded as 1 dim with value 1 at check memory

Author: kelvinchoi-intel

src/plugins/intel_gpu/src/graph/primitive_inst.cpp

@@ -199,10 +199,21 @@ kernel_impl_params primitive_impl::static_canonicalize_shapes(const kernel_impl_
 
 uint32_t primitive_inst::get_network_id() const { return _network.get_id(); }
 
-void primitive_inst::check_memory_to_set(const memory& mem, const layout& layout) const {
-    OPENVINO_ASSERT((mem.get_layout() == layout) || layout.is_dynamic(), "[GPU] Unexpected layout of input memory for ", id(), " node!\n",
-                     "Node layout: ", layout.to_short_string(), "\n",
-                     "Memory layout: ", mem.get_layout().to_short_string());
+void primitive_inst::check_memory_to_set(const memory& mem, const layout& l) const {
+    // The layout with empty tensor (scalar) is regarded as 1 dimension with value 1
+    bool single_value_layout = false;
+    if (!l.is_dynamic()) {
+        auto layout_ps = l.get_partial_shape();
+        single_value_layout = (layout_ps.size() == 1 && layout_ps[0] == 1);
+    }
+
+    auto mem_layout = mem.get_layout();
+    OPENVINO_ASSERT((mem_layout == l)
+                    || l.is_dynamic()
+                    || (mem_layout.get_partial_shape().size() == 0 && single_value_layout),
+                    "[GPU] Unexpected layout of input memory for ", id(), " node!\n",
+                    "Node layout: ", l.to_short_string(), "\n",
+                    "Memory layout: ", mem_layout.to_short_string());
 
     // check shared image/buffer compatibility, if applicable
     auto params = mem.get_internal_params();
@@ -218,11 +229,11 @@ void primitive_inst::check_memory_to_set(const memory& mem, const layout& layout
         switch (params.mem_type) {
         case shared_mem_type::shared_mem_vasurface:
         case shared_mem_type::shared_mem_image:
-            OPENVINO_ASSERT(layout.format.is_image_2d(), "Attempt to set user-supplied input or output image instead of a buffer");
+            OPENVINO_ASSERT(l.format.is_image_2d(), "Attempt to set user-supplied input or output image instead of a buffer");
             break;
         case shared_mem_type::shared_mem_buffer:
         case shared_mem_type::shared_mem_dxbuffer:
-            OPENVINO_ASSERT(!layout.format.is_image_2d(), "Attempt to set user-supplied input or output buffer instead of an image");
+            OPENVINO_ASSERT(!l.format.is_image_2d(), "Attempt to set user-supplied input or output buffer instead of an image");
             break;
         case shared_mem_type::shared_mem_usm:
             break;

src/plugins/intel_gpu/tests/unit/test_cases/gather_gpu_test.cpp

@@ -2373,3 +2373,48 @@ TEST(gather_gpu_fp32, dynamic_support_neg_ind) {
         ASSERT_EQ(expected_results[i], output_ptr[i]) << i;
     }
 }
+
+TEST(gather_gpu_fp32, dynamic_support_scalar_indice_empty_memory) {
+    auto& engine = get_test_engine();
+
+    ov::Shape data_shape = { 3, 3 };
+    int64_t axis = 1;
+
+    auto data_layout = layout{ov::PartialShape::dynamic(data_shape.size()), data_types::f32, format::bfyx};
+    auto indices_layout = layout{ov::PartialShape({1}), data_types::i32, format::bfyx};
+
+    auto data_mem = engine.allocate_memory(layout{ov::PartialShape(data_shape), data_types::f32, format::bfyx});
+    auto indices_mem = engine.allocate_memory(layout{ov::PartialShape({}), data_types::i32, format::bfyx});
+
+    set_values(data_mem, { 0.f, 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f, 8.f });
+    set_values(indices_mem, { -1 });
+
+    topology topology;
+    topology.add(input_layout("data", data_layout));
+    topology.add(input_layout("indices", indices_layout));
+    topology.add(gather("gather", input_info("data"), input_info("indices"), axis, data_shape.size(), ov::Shape{}, 0, true));
+
+    ExecutionConfig config = get_test_default_config(engine);
+    config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
+    network network(engine, topology, config);
+
+    network.set_input_data("data", data_mem);
+    network.set_input_data("indices", indices_mem);
+
+    auto inst = network.get_primitive("gather");
+    auto impl = inst->get_impl();
+    ASSERT_TRUE(impl != nullptr);
+    ASSERT_TRUE(impl->is_dynamic());
+
+    auto outputs = network.execute();
+
+    auto output = outputs.at("gather").get_memory();
+    cldnn::mem_lock<float> output_ptr(output, get_test_stream());
+
+    std::vector<float> expected_results = { 2.f, 5.f, 8.f };
+
+    ASSERT_EQ(expected_results.size(), output_ptr.size());
+    for (size_t i = 0; i < expected_results.size(); ++i) {
+        ASSERT_EQ(expected_results[i], output_ptr[i]) << i;
+    }
+}