[GPU] Add i32 output for FC

Author: kelvinchoi-intel

src/plugins/intel_gpu/src/graph/impls/ocl/fully_connected.cpp

@@ -205,6 +205,7 @@ attach_fully_connected_impl::attach_fully_connected_impl() {
                                              typed_primitive_impl_ocl<fully_connected>::create<fully_connected_impl>, {
         std::make_tuple(data_types::f32, format::bfyx),
         std::make_tuple(data_types::f16, format::bfyx),
+        std::make_tuple(data_types::i32, format::bfyx),
         std::make_tuple(data_types::u8, format::bfyx),
         std::make_tuple(data_types::i8, format::bfyx),
     });

src/plugins/intel_gpu/src/kernel_selector/kernels/fully_connected/fully_connected_kernel_bfyx_ref.cpp

@@ -15,6 +15,7 @@ ParamsKey FullyConnected_bfyx_Ref::GetSupportedKey() const {
     k.EnableInputDataType(Datatype::F32);
     k.EnableInputDataType(Datatype::INT8);
     k.EnableInputDataType(Datatype::UINT8);
+    k.EnableOutputDataType(Datatype::INT32);
     k.EnableOutputDataType(Datatype::F16);
     k.EnableOutputDataType(Datatype::F32);
     k.EnableOutputDataType(Datatype::INT8);

src/plugins/intel_gpu/tests/functional/single_layer_tests/dynamic/matmul.cpp

@@ -157,6 +157,11 @@ const std::vector<ov::element::Type> netPRCs {
     ov::element::f32,
 };
 
+const std::vector<ov::element::Type> netPRCs_f32_i32 {
+    ov::element::f32,
+    ov::element::i32
+};
+
 
 /* ============= FullyConnected ============= */
 
@@ -309,7 +314,7 @@ const std::vector<ShapeRelatedParams> IS3D_nightly = {
 };
 
 const auto fullyConnectedParams3D_smoke = ::testing::Combine(::testing::ValuesIn(IS3D_smoke),
-                                                       ::testing::Values(ov::element::f32),
+                                                       ::testing::ValuesIn(netPRCs_f32_i32),
                                                        ::testing::Values(ov::element::undefined),
                                                        ::testing::Values(ov::element::undefined),
                                                        ::testing::Values(ov::test::utils::InputLayerType::CONSTANT),

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

@@ -316,6 +316,61 @@ TEST(fully_connected_gpu, no_biases_int8) {
     ASSERT_EQ(-52.0f, output_ptr[3]);
 }
 
+TEST(fully_connected_gpu, no_biases_fc_i32) {
+    const int32_t input_b = 1, input_f = 3,     // size of the whole input buffer
+                  weight_b = 4, weight_f = 3;   // size of the whole weights buffer
+
+    auto& engine = get_test_engine();
+
+    auto input_prim = engine.allocate_memory({ data_types::i32, format::bfyx, { input_b, input_f, 1, 1 } });
+    auto weights_prim = engine.allocate_memory({ data_types::i32, format::bfyx, { weight_b, weight_f, 1, 1 } });
+
+    set_values<int32_t>(input_prim, {
+        8, 2, -4
+    });
+
+    set_values<int32_t>(weights_prim, {
+        2, 1, 0,
+        -3, -2, 1,
+        0, -2, -4,
+        -5, 10, 8
+    });
+
+    cldnn::topology topology{
+        input_layout("input", input_prim->get_layout()),
+        data("weights", weights_prim),
+        reorder("reorder_to_i32", input_info("input"), { data_types::i32, format::bfyx, { input_b, input_f, 1, 1 } }),
+        fully_connected("fc_prim", input_info("reorder_to_i32"), "weights"),
+        reorder("reorder_to_float", input_info("fc_prim"), { data_types::f32, format::bfyx, { input_b, weight_b, 1, 1 } })
+    };
+
+    ExecutionConfig config = get_test_default_config(engine);
+    config.set_property(ov::intel_gpu::optimize_data(true));
+    config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
+
+    cldnn::network network(engine, topology, config);
+
+    network.set_input_data("input", input_prim);
+
+    auto outputs = network.execute();
+    ASSERT_EQ(outputs.size(), size_t(1));
+    ASSERT_EQ(outputs.begin()->first, "reorder_to_float");
+
+    auto output_memory = outputs.begin()->second.get_memory();
+    cldnn::mem_lock<float> output_ptr (output_memory, get_test_stream());
+    auto output_layout = output_memory->get_layout();
+
+    ASSERT_EQ(output_layout.format, format::bfyx);
+
+    VVF<float> output_ref = {
+        { 18, -32, 12, -52 },
+    };
+
+    for (int i = 0; i< weight_b; ++i) {
+        ASSERT_EQ(output_ref[0][i], output_ptr[i]);
+    }
+}
+
 TEST(fully_connected_gpu, xb_f32_batch_1) {
     //  Input  : 3x1
     //  Output : 4x1