[GPU] Changed ExecutionMode::ACCURACY mode behaviour on GPU. Previously, all floating-point data types were upconverted to f32 data type with enabled ExecutionMode::ACCURACY, now we preserve the original model's data types.

Author: sshlyapn

src/plugins/intel_gpu/src/runtime/execution_config.cpp

@@ -19,7 +19,7 @@ class InferencePrecisionValidator : public BaseValidator {
 public:
     bool is_valid(const ov::Any& v) const override {
         auto precision = v.as<ov::element::Type>();
-        return precision == ov::element::f16 || precision == ov::element::f32;
+        return precision == ov::element::f16 || precision == ov::element::f32 || precision == ov::element::undefined;
     }
 };
 
@@ -129,7 +129,7 @@ void ExecutionConfig::apply_execution_hints(const cldnn::device_info& info) {
         const auto mode = get_property(ov::hint::execution_mode);
         if (!is_set_by_user(ov::hint::inference_precision)) {
             if (mode == ov::hint::ExecutionMode::ACCURACY) {
-                set_property(ov::hint::inference_precision(ov::element::f32));
+                set_property(ov::hint::inference_precision(ov::element::undefined));
             } else if (mode == ov::hint::ExecutionMode::PERFORMANCE) {
                 if (info.supports_fp16)
                     set_property(ov::hint::inference_precision(ov::element::f16));

src/plugins/intel_gpu/tests/functional/behavior/inference_precision.cpp

@@ -37,8 +37,10 @@ TEST_P(InferencePrecisionTests, smoke_canSetInferencePrecisionAndInfer) {
 static const std::vector<params> test_params = {
     {ov::element::f16, ov::element::f32},
     {ov::element::f16, ov::element::f16},
+    {ov::element::f16, ov::element::undefined},
     {ov::element::f32, ov::element::f32},
     {ov::element::f32, ov::element::f16},
+    {ov::element::f32, ov::element::undefined},
 };
 
 INSTANTIATE_TEST_SUITE_P(smoke_GPU_BehaviorTests, InferencePrecisionTests, ::testing::ValuesIn(test_params), InferencePrecisionTests::getTestCaseName);
@@ -48,7 +50,6 @@ TEST(InferencePrecisionTests, CantSetInvalidInferencePrecision) {
 
     ASSERT_NO_THROW(core.get_property(ov::test::utils::DEVICE_GPU, ov::hint::inference_precision));
     ASSERT_ANY_THROW(core.set_property(ov::test::utils::DEVICE_GPU, ov::hint::inference_precision(ov::element::bf16)));
-    ASSERT_ANY_THROW(core.set_property(ov::test::utils::DEVICE_GPU, ov::hint::inference_precision(ov::element::undefined)));
 }
 
 TEST(ExecutionModeTest, SetCompileGetInferPrecisionAndExecMode) {
@@ -62,10 +63,26 @@ TEST(ExecutionModeTest, SetCompileGetInferPrecisionAndExecMode) {
         ASSERT_EQ(ov::element::f32, compiled_model.get_property(ov::hint::inference_precision));
     }
 
+    {
+        /* ov::hint::inference_precision has higher priority than ov::hint::execution_mode */
+        auto compiled_model = core.compile_model(model, ov::test::utils::DEVICE_GPU, ov::hint::inference_precision(ov::element::f16),
+                                                                                     ov::hint::execution_mode(ov::hint::ExecutionMode::ACCURACY));
+        ASSERT_EQ(ov::hint::ExecutionMode::ACCURACY, compiled_model.get_property(ov::hint::execution_mode));
+        ASSERT_EQ(ov::element::f16, compiled_model.get_property(ov::hint::inference_precision));
+    }
+
+    {
+        /* ov::hint::inference_precision has higher priority than ov::hint::execution_mode */
+        auto compiled_model = core.compile_model(model, ov::test::utils::DEVICE_GPU, ov::hint::inference_precision(ov::element::undefined),
+                                                                                     ov::hint::execution_mode(ov::hint::ExecutionMode::PERFORMANCE));
+        ASSERT_EQ(ov::hint::ExecutionMode::PERFORMANCE, compiled_model.get_property(ov::hint::execution_mode));
+        ASSERT_EQ(ov::element::undefined, compiled_model.get_property(ov::hint::inference_precision));
+    }
+
     {
         auto compiled_model = core.compile_model(model, ov::test::utils::DEVICE_GPU, ov::hint::execution_mode(ov::hint::ExecutionMode::ACCURACY));
         ASSERT_EQ(ov::hint::ExecutionMode::ACCURACY, compiled_model.get_property(ov::hint::execution_mode));
-        ASSERT_EQ(ov::element::f32, compiled_model.get_property(ov::hint::inference_precision));
+        ASSERT_EQ(ov::element::undefined, compiled_model.get_property(ov::hint::inference_precision));
     }
 
     {