[GPU] Don't set output element type by force to f32 for ReduceMax, ReduceMin (#26675)

### Details:
- For ReduceMax, ReduceMin, output element types should be same to input
element type

### Tickets:
 - 151208

Author: davidsnam-intel

src/plugins/intel_gpu/src/graph/reduce.cpp

@@ -73,10 +73,13 @@ layout reduce_inst::calc_output_layout(reduce_node const& node, kernel_impl_para
     }
 
     std::vector<reduce_mode> reduce_bool_modes = {reduce_mode::logical_and, reduce_mode::logical_or};
-    if (std::find(reduce_bool_modes.begin(), reduce_bool_modes.end(), mode) != reduce_bool_modes.end())
+    if (std::find(reduce_bool_modes.begin(), reduce_bool_modes.end(), mode) != reduce_bool_modes.end()) {
         output_type = data_types::i8;
-    else if (output_type == data_types::i8 || output_type == data_types::u8)
-        output_type = data_types::f32;
+    } else if (output_type == data_types::i8 || output_type == data_types::u8) {
+        if (mode != reduce_mode::min && mode != reduce_mode::max) {
+            output_type = data_types::f32;
+        }
+    }
 
     if (desc->output_data_types[0])
         output_type = *desc->output_data_types[0];

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

@@ -594,7 +594,17 @@ INSTANTIATE_TEST_SUITE_P(reduce_gpu_b_fs_yx_fsv16_i8_i8,
                             TestParamType_general_reduce_gpu(3, 3, 1, 1, 3, 6, format::b_fs_yx_fsv16, reduce_mode::logical_or, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, true, data_types::i8),
                             TestParamType_general_reduce_gpu(3, 5, 1, 1, 3, 2, format::b_fs_yx_fsv16, reduce_mode::logical_and, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, true, data_types::i8),
                             TestParamType_general_reduce_gpu(3, 7, 1, 1, 3, 2, format::b_fs_yx_fsv16, reduce_mode::logical_or, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, true, data_types::i8),
-                            TestParamType_general_reduce_gpu(1, 3, 1, 1, 6, 12, format::b_fs_yx_fsv16, reduce_mode::logical_and, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, true, data_types::i8)
+                            TestParamType_general_reduce_gpu(1, 3, 1, 1, 6, 12, format::b_fs_yx_fsv16, reduce_mode::logical_and, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, true, data_types::i8),
+                            TestParamType_general_reduce_gpu(3, 3, 1, 1, 2, 11, format::b_fs_yx_fsv16, reduce_mode::min, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::i8),
+                            TestParamType_general_reduce_gpu(16, 4, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::max, {1, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::i8),
+                            TestParamType_general_reduce_gpu(16, 16, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::i8),
+                            TestParamType_general_reduce_gpu(17, 34, 1, 1, 18, 11, format::b_fs_yx_fsv16, reduce_mode::max, {2, 1}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::i8),
+                            TestParamType_general_reduce_gpu(16, 16, 1, 1, 17, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::i8),
+                            TestParamType_general_reduce_gpu(16, 4, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::max, {1, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::i8),
+                            TestParamType_general_reduce_gpu(16, 16, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::i8),
+                            TestParamType_general_reduce_gpu(2, 5, 1, 1, 3, 3, format::b_fs_yx_fsv16, reduce_mode::max, {3, 1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::i8),
+                            TestParamType_general_reduce_gpu(17, 34, 1, 1, 18, 11, format::b_fs_yx_fsv16, reduce_mode::max, {2, 1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::i8),
+                            TestParamType_general_reduce_gpu(16, 16, 1, 1, 17, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::i8)
                             ),
                             general_reduce_gpu::PrintToStringParamName);
 
@@ -603,33 +613,22 @@ INSTANTIATE_TEST_SUITE_P(reduce_gpu_b_fs_yx_fsv16_i8_i8,
                         ::testing::Values(
                             TestParamType_general_reduce_gpu(3, 3, 1, 1, 3, 2, format::b_fs_yx_fsv16, reduce_mode::sum, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(3, 3, 1, 1, 3, 3, format::b_fs_yx_fsv16, reduce_mode::l1, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(3, 3, 1, 1, 2, 11, format::b_fs_yx_fsv16, reduce_mode::min, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(7, 3, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(26, 12, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::sum, {3, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(17, 12, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::l1, {2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(16, 4, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::max, {1, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(17, 34, 1, 1, 13, 12, format::b_fs_yx_fsv16, reduce_mode::l2, {0, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(16, 16, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(1, 2, 1, 1, 5, 5, format::b_fs_yx_fsv16, reduce_mode::prod, {3, 1}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(16, 26, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(17, 34, 1, 1, 18, 11, format::b_fs_yx_fsv16, reduce_mode::max, {2, 1}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(16, 16, 1, 1, 17, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(17, 34, 1, 1, 16, 11, format::b_fs_yx_fsv16, reduce_mode::l1, {1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(16, 16, 1, 1, 15, 8, format::b_fs_yx_fsv16, reduce_mode::log_sum_exp, {0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(17, 34, 1, 1, 14, 11, format::b_fs_yx_fsv16, reduce_mode::l2, {1}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(16, 16, 1, 1, 12, 8, format::b_fs_yx_fsv16, reduce_mode::sum_square, {2}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(17, 34, 1, 1, 12, 11, format::b_fs_yx_fsv16, reduce_mode::log_sum, {3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
-
                             TestParamType_general_reduce_gpu(7, 3, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(26, 12, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::sum, {3, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(17, 12, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::l1, {2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(16, 4, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::max, {1, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(17, 34, 1, 1, 13, 9, format::b_fs_yx_fsv16, reduce_mode::l2, {0, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(16, 16, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(2, 5, 1, 1, 3, 3, format::b_fs_yx_fsv16, reduce_mode::max, {3, 1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(16, 26, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(17, 34, 1, 1, 18, 11, format::b_fs_yx_fsv16, reduce_mode::max, {2, 1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(16, 16, 1, 1, 17, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(17, 34, 1, 1, 16, 15, format::b_fs_yx_fsv16, reduce_mode::l1, {1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(16, 16, 1, 1, 15, 8, format::b_fs_yx_fsv16, reduce_mode::log_sum_exp, {0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(17, 3, 1, 1, 14, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
@@ -1831,6 +1830,38 @@ TEST(reduce_gpu, b_fs_yx_fsv16_max_dynamic) {
     }
 }
 
+TEST(reduce_gpu, reduce_min_max_default_output_element_type_should_be_same_to_input_element_type) {
+    auto& engine = get_test_engine();
+
+    auto input = engine.allocate_memory({data_types::i8, format::bfyx, {1, 1, 2, 2}});
+    set_values(input, {1, 1, 1, 1});
+
+    topology topology1(
+        input_layout("input", input->get_layout()),
+        reduce("reduce", input_info("input"), reduce_mode::max, {1}, 0),
+        reorder("reorder", input_info("reduce"), format::bfyx, data_types::i8)
+    );
+
+    ExecutionConfig config = get_test_default_config(engine);
+    network network1(engine, topology1, config);
+    network1.set_input_data("input", input);
+    auto output = network1.execute();
+
+    ASSERT_EQ(network1.get_program()->get_node("reduce").get_output_layout().data_type, data_types::i8);
+
+    topology topology2(
+        input_layout("input", input->get_layout()),
+        reduce("reduce", input_info("input"), reduce_mode::min, {1}, 0),
+        reorder("reorder", input_info("reduce"), format::bfyx, data_types::i8)
+    );
+
+    network network2(engine, topology2, config);
+    network2.set_input_data("input", input);
+    output = network2.execute();
+
+    ASSERT_EQ(network2.get_program()->get_node("reduce").get_output_layout().data_type, data_types::i8);
+}
+
 template <data_types InputT, data_types OutputT>
 class ReduceXYWithBigTensorTestBase : public ::testing::TestWithParam<TestParamType_general_reduce_gpu> {
 protected:
@@ -2150,16 +2181,9 @@ INSTANTIATE_TEST_SUITE_P(onednn_reduce_gpu_b_fs_yx_fsv16_i8_f32,
                         onednn_reduce_gpu_i8_f32,
                         ::testing::Values(
                             TestParamType_general_reduce_gpu(3, 3, 1, 1, 3, 2, format::b_fs_yx_fsv16, reduce_mode::sum, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(3, 3, 1, 1, 2, 11, format::b_fs_yx_fsv16, reduce_mode::min, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(7, 3, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(16, 4, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::max, {1, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(16, 16, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 3}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(17, 34, 1, 1, 18, 11, format::b_fs_yx_fsv16, reduce_mode::max, {2, 1}, "reduce_gpu_b_fs_yx_fsv16", false, data_types::i8, false, data_types::f32),
 
                             TestParamType_general_reduce_gpu(7, 3, 1, 1, 13, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {3, 2, 1, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(16, 4, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::max, {1, 2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(16, 16, 1, 1, 16, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 3}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
-                            TestParamType_general_reduce_gpu(16, 16, 1, 1, 17, 8, format::b_fs_yx_fsv16, reduce_mode::min, {2, 0}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32),
                             TestParamType_general_reduce_gpu(17, 3, 1, 1, 14, 11, format::b_fs_yx_fsv16, reduce_mode::mean, {1}, "reduce_gpu_b_fs_yx_fsv16", true, data_types::i8, false, data_types::f32)
                         ), general_reduce_gpu::PrintToStringParamName);