[GPU] Fix dynamic loop's not matched issue during multiple shapes are inferenced

- Disable memory reuse for updated shape in case of dep of loop

Author: kelvinchoi-intel

src/plugins/intel_gpu/src/graph/include/loop_inst.h

@@ -298,6 +298,7 @@ class typed_primitive_inst<loop> : public typed_primitive_inst_base<loop> {
     std::vector<backedge_memory_mapping> backedge_memory_mappings;
     std::vector<concatenated_memory_mapping::ptr> concatenated_input_mem_mappings;
     std::vector<concatenated_memory_mapping::ptr> concatenated_output_mem_mappings;
+    ov::PartialShape dynamic_sliced_layout;
 
     static std::string to_string(const loop_node& node);
 

src/plugins/intel_gpu/src/graph/loop.cpp

@@ -375,12 +375,20 @@ loop_inst::concatenated_memory_mapping::ptr loop_inst::create_concat_memory_map(
     if (extern_mem_ptr != nullptr) {
         layout sliced_layout = intern_prim->get_output_layout(internal_id.idx);
         auto inter_mem_ptr = intern_prim->output_memory_ptr(internal_id.idx);
-        if (inter_mem_ptr == nullptr) {
+        if (inter_mem_ptr == nullptr || shape_changed()) {
             // if inner body intern_prim has no output memory because it has dynamic shape,
             // calculate inner body intern_prim layout using concat_mem's layout.
             auto updated_sliced_layout = sliced_layout.get_partial_shape();
             OPENVINO_ASSERT(updated_sliced_layout[io_prim_map.axis].is_static() || num_iterations > 0,
                                     "Not allowed dynamic dimension for axis when num_iteraiont is negative");
+
+            // Save or load dynamic sliced layout to update with various shapes
+            if (updated_sliced_layout.is_dynamic()) {
+                dynamic_sliced_layout = updated_sliced_layout;
+            } else if (dynamic_sliced_layout.is_dynamic()) {
+                updated_sliced_layout = dynamic_sliced_layout;
+            }
+
             auto concat_pshape = extern_prim->get_output_layout().get_partial_shape();
             const auto shape_size = concat_pshape.size();
             for (size_t i = 0; i < shape_size; i++) {

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

@@ -536,8 +536,20 @@ event::ptr primitive_inst::realloc_if_needed() {
         }
     }
 
+    bool has_loop_user = false;
+    for (const auto& user : _node->get_users()) {
+        if (user->is_type<loop>()) {
+            has_loop_user = true;
+        }
+    }
+
     // Clear out memory if if was previously reused, but now primitive can't be optimized
-    if (_node->is_type<gather>() || _node->is_type<permute>() || _node->is_type<reshape>() || _node->is_type<reorder>() || _node->is_type<strided_slice>()) {
+    if (!has_loop_user &&
+        (_node->is_type<gather>() ||
+        _node->is_type<permute>() ||
+        _node->is_type<reshape>() ||
+        _node->is_type<reorder>() ||
+        _node->is_type<strided_slice>())) {
         if (can_be_optimized()) {
             _max_output_layout_count = _deps[0].first->_max_output_layout_count;
             return ev;

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

@@ -601,3 +601,170 @@ TEST(loop_gpu, support_dynamic_tensoriterator_outer_axis) {
 
     test_loop_gpu_wo_trip_count({ 2, 1, 1, 2}, { 2, 5, 1, 2}, input_data_5_4, output_data_5_4, 1, 4);
 }
+
+static void test_loop_gpu_wo_trip_count_w_multiple_shapes(ov::PartialShape body_input_layout,
+                                        std::vector<ov::PartialShape> whole_layouts,
+                                        std::vector<std::vector<float>> input_data_list,
+                                        std::vector<float> expected_output_data,
+                                        size_t axis,
+                                        size_t exit_value,
+                                        bool is_caching_test = false) {
+    auto& engine = get_test_engine();
+
+    auto b_input_layout = cldnn::layout{ body_input_layout, data_types::f32, format::bfyx };
+
+    ov::PartialShape sliced_input_shape = body_input_layout;
+    sliced_input_shape[axis] = 1;
+    auto sliced_input_layout = cldnn::layout{ sliced_input_shape, data_types::f32, format::bfyx };
+
+    auto const_layout = cldnn::layout{ {}, data_types::i64, format::bfyx };
+
+    auto e_initial_condition_mem = engine.allocate_memory(const_layout);
+    auto e_num_iteration_mem = engine.allocate_memory(const_layout);
+    auto b_exit_value_mem = engine.allocate_memory(const_layout);
+    auto b_index_inc_mem = engine.allocate_memory(const_layout);
+
+    // initialize input buffers
+    set_values(e_initial_condition_mem, {1});
+    set_values(b_exit_value_mem, {exit_value});
+    set_values(b_index_inc_mem, {1});
+    set_values(e_num_iteration_mem, {0});
+
+    primitive_id body_current_iteration_id = "b_index";
+    primitive_id body_execution_condition_id = "b_cond_exit_value";
+
+    cldnn::topology body(
+        input_layout(body_current_iteration_id, const_layout),
+        input_layout("b_add_data", sliced_input_layout),
+        input_layout("b_mul_data", sliced_input_layout),
+        data("b_exit_value", b_exit_value_mem),
+        data("b_index_inc", b_index_inc_mem),
+        eltwise("b_index_update", input_info(body_current_iteration_id), input_info("b_index_inc"), eltwise_mode::sum),
+        reorder("b_index_cast", input_info("b_index_update"),
+                    cldnn::format::any, data_types::f32, {}, cldnn::reorder_mean_mode::subtract, cldnn::padding(), true),
+        eltwise(body_execution_condition_id, input_info("b_index"), input_info("b_exit_value"), eltwise_mode::lt),
+        eltwise("b_add", input_info("b_add_data"), input_info("b_index_cast"), eltwise_mode::sum),
+        eltwise("b_mul", input_info("b_mul_data"), input_info("b_index_cast"), eltwise_mode::prod));
+
+    primitive_id trip_count_id = "";
+    primitive_id actual_iteration_count_id = "actual_iteration_count";
+    primitive_id initial_condition_id = "initial_condition";
+    int64_t num_iterations = -1;
+
+    std::vector<loop::io_primitive_map> input_primitive_maps {
+        loop::io_primitive_map("input", "b_add_data", axis),
+        loop::io_primitive_map("input", "b_mul_data", axis),
+        loop::io_primitive_map(actual_iteration_count_id, body_current_iteration_id) };
+    std::vector<loop::io_primitive_map> output_primitive_maps {
+        loop::io_primitive_map(cldnn::input_info("loop", 0), cldnn::input_info("b_add", 0), axis),
+        loop::io_primitive_map(cldnn::input_info("loop", 1), cldnn::input_info("b_mul", 0), axis) };
+    std::vector<loop::backedge_mapping> back_edges {
+        loop::backedge_mapping("b_index_update", body_current_iteration_id) };
+
+    auto body_program = build_program(engine, body, body_execution_condition_id, output_primitive_maps, back_edges, true);
+
+    cldnn::topology topology(
+        input_layout("input", b_input_layout),
+        input_layout(initial_condition_id, e_initial_condition_mem->get_layout()),
+        mutable_data(actual_iteration_count_id, e_num_iteration_mem),
+        loop("loop", { input_info(actual_iteration_count_id), input_info(initial_condition_id), input_info("input") }, body_program,
+             trip_count_id, initial_condition_id, actual_iteration_count_id,
+             input_primitive_maps, output_primitive_maps, back_edges,
+             num_iterations, body_current_iteration_id, body_execution_condition_id, 2),
+        eltwise("out_sum", input_info("loop", 0), input_info("loop", 1), eltwise_mode::sum));
+
+    ExecutionConfig config = get_test_default_config(engine);
+    config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
+
+    cldnn::network::ptr network = get_network(engine, topology, config, get_test_stream_ptr(), is_caching_test);
+
+
+    for (size_t i = 0 ; i < whole_layouts.size(); i++) {
+        auto whole_layout = whole_layouts[i];
+        auto input_data = input_data_list[i];
+
+        // initialize input buffers
+        set_values(e_initial_condition_mem, {1});
+        set_values(b_exit_value_mem, {exit_value});
+        set_values(b_index_inc_mem, {1});
+        set_values(e_num_iteration_mem, {0});
+
+        auto e_input_layout = cldnn::layout{ whole_layout, data_types::f32, format::bfyx };
+        auto e_input_mem = engine.allocate_memory(e_input_layout); // b,f,x,y
+        auto expected_output_layout = whole_layout;
+        set_values(e_input_mem, input_data);
+        network->set_input_data("input", e_input_mem);
+
+        network->set_input_data(initial_condition_id, e_initial_condition_mem);
+
+        auto outputs = network->execute();
+        ASSERT_EQ(outputs.size(), 1);
+
+        auto expected_num_iterations = (exit_value + 1);
+        expected_output_layout[axis] = expected_num_iterations;
+        auto e_output_layout = cldnn::layout{ expected_output_layout, data_types::f32, format::bfyx };
+
+        auto num_iter_mem = network->get_output_memory(actual_iteration_count_id);
+        if (num_iter_mem != nullptr) {
+            mem_lock<int64_t> num_iter_ptr{ num_iter_mem, get_test_stream() };
+            ASSERT_EQ(num_iter_ptr.data()[0], expected_num_iterations);
+        }
+
+        std::vector<float> expected(input_data.size());
+        if (expected_output_data.size() == 0) {
+            size_t unit = 1;
+            for (size_t k = axis; k < whole_layout.size(); k++) {
+                unit *= whole_layout[k].get_length();
+            }
+
+            for (size_t j = 0; j < input_data.size(); j++) {
+                auto val = static_cast<size_t>((j % unit) / 4) + 1;
+                expected[j] = static_cast<float>(input_data[j] + val) + static_cast<float>(input_data[j] * val);
+            }
+        } else {
+            expected = expected_output_data;
+        }
+
+        auto output_mem = outputs.begin()->second.get_memory();
+        auto output_layout = output_mem->get_layout();
+        ASSERT_EQ(output_layout.batch(), e_output_layout.batch());
+        ASSERT_EQ(output_layout.feature(), e_output_layout.feature());
+        ASSERT_EQ(output_layout.spatial(0), e_output_layout.spatial(0));
+        ASSERT_EQ(output_layout.spatial(1), e_output_layout.spatial(1));
+        // value check
+        {
+            mem_lock<float> output_ptr{ output_mem, get_test_stream() };
+            for (size_t i = 0, iend = output_layout.count(); i < iend; ++i) {
+                ASSERT_FLOAT_EQ(output_ptr[i], expected.at(i));
+            }
+        }
+    }
+}
+
+std::vector<float> input_data_4_4{
+    1.0f,  2.0f, -15.f,  3.0f,
+    4.0f, -15.f, 5.0f,  6.0f,
+    -15.f, 7.0f, -15.f, 0.0f,
+    0.0f, -15.f, 0.5f, -0.5f,
+};
+
+std::vector<float> input_data_2_4_4{
+    1.0f,  2.0f, -15.f,  3.0f,
+    4.0f, -15.f, 5.0f,  6.0f,
+    -15.f, 7.0f, -15.f, 0.0f,
+    0.0f, -15.f, 0.5f, -0.5f,
+
+    1.0f,  2.0f, -15.f,  3.0f,
+    4.0f, -15.f, 5.0f,  6.0f,
+    -15.f, 7.0f, -15.f, 0.0f,
+    0.0f, -15.f, 0.5f, -0.5f,
+};
+
+TEST(loop_gpu, support_loop_w_dynamic_input_w_various_shapes) {
+    test_loop_gpu_wo_trip_count_w_multiple_shapes(
+        { 1, -1, 4, 4 },
+        {{ 1, 1, 4, 4 }, { 1, 2, 4, 4 }},   // axis value should be iter_num = (exit_value + 1)
+        {input_data_4_4, input_data_2_4_4},
+        std::vector<float>(),
+        2, 3);
+}