[lite-xpu-subgraph] Fix xpu compile and test xpu ci. (#27932)

cmake/external/lite.cmake

@@ -22,7 +22,7 @@ if(XPU_SDK_ROOT)
   set(LITE_WITH_XPU ON)
   include_directories("${XPU_SDK_ROOT}/XTDK/include")
   include_directories("${XPU_SDK_ROOT}/XTCL/include")
-  add_definitions(-DPADDLE_WITH_XPU)
+  add_definitions(-DLITE_SUBGRAPH_WITH_XPU)
   LINK_DIRECTORIES("${XPU_SDK_ROOT}/XTDK/shlib/")
   LINK_DIRECTORIES("${XPU_SDK_ROOT}/XTDK/runtime/shlib/")
 endif()

paddle/fluid/inference/api/analysis_config.cc

@@ -381,7 +381,7 @@ void AnalysisConfig::Update() {
   }
 
   if (use_xpu_) {
-#ifndef PADDLE_WITH_XPU
+#ifndef LITE_SUBGRAPH_WITH_XPU
     PADDLE_THROW(platform::errors::Unavailable(
         "You tried to use an XPU device, but Paddle was not compiled "
         "with XPU-runtime."));

paddle/fluid/inference/lite/CMakeLists.txt

@@ -4,6 +4,6 @@ endif()
 
 cc_library(lite_op_teller SRCS op_teller.cc DEPS lite_full_static framework_proto device_context boost xxhash)
 cc_library(lite_engine SRCS engine.cc DEPS lite_full_static framework_proto ${XPU_DEPS})
-cc_library(lite_tensor_utils SRCS tensor_utils.cc DEPS memcpy lite_full_static framework_proto boost device_context)
+cc_library(lite_tensor_utils SRCS tensor_utils.cc DEPS memcpy lite_full_static framework_proto boost device_context ${XPU_DEPS})
 cc_test(test_lite_engine SRCS test_engine.cc DEPS lite_engine protobuf framework_proto glog gtest analysis)
 cc_test(test_lite_tensor_utils SRCS test_tensor_utils.cc DEPS lite_engine lite_tensor_utils)

paddle/fluid/inference/lite/engine.cc

@@ -16,7 +16,7 @@
 #define LITE_WITH_CUDA 1
 #endif
 
-#ifdef PADDLE_WITH_XPU
+#ifdef LITE_SUBGRAPH_WITH_XPU
 #define LITE_WITH_XPU 1
 #endif
 
@@ -59,7 +59,7 @@ paddle::lite_api::PaddlePredictor* EngineManager::Create(
       cfg.cpu_math_library_num_threads);
 #endif
 
-#ifdef PADDLE_WITH_XPU
+#ifdef LITE_SUBGRAPH_WITH_XPU
   lite_cxx_config.set_xpu_workspace_l3_size_per_thread(
       cfg.xpu_l3_workspace_size);
 #endif

paddle/fluid/inference/tests/api/lite_resnet50_test.cc

@@ -26,7 +26,11 @@ namespace inference {
 TEST(AnalysisPredictor, use_gpu) {
   std::string model_dir = FLAGS_infer_model + "/" + "model";
   AnalysisConfig config;
+#if defined(PADDLE_WITH_CUDA)
   config.EnableUseGpu(100, 0);
+#elif defined(LITE_SUBGRAPH_WITH_XPU)
+  config.EnableXpu(100);
+#endif
   config.SetModel(model_dir + "/model", model_dir + "/params");
   config.EnableLiteEngine(paddle::AnalysisConfig::Precision::kFloat32, true);
 
@@ -40,7 +44,7 @@ TEST(AnalysisPredictor, use_gpu) {
   std::vector<float> input(input_num, 1);
 
   PaddleTensor in;
-  in.shape = {1, 3, 318, 318};
+  in.shape = {batch, channel, height, width};
   in.data =
       PaddleBuf(static_cast<void*>(input.data()), input_num * sizeof(float));
   in.dtype = PaddleDType::FLOAT32;
@@ -92,7 +96,7 @@ TEST(Predictor, use_gpu) {
   auto input_names = predictor->GetInputNames();
   auto input_t = predictor->GetInputHandle(input_names[0]);
 
-  input_t->Reshape({1, 3, 318, 318});
+  input_t->Reshape({batch, channel, height, width});
   input_t->CopyFromCpu(input.data());
   predictor->Run();
 

paddle/fluid/pybind/inference_api.cc

@@ -443,6 +443,8 @@ void BindAnalysisConfig(py::module *m) {
       .def("params_file", &AnalysisConfig::params_file)
       .def("enable_use_gpu", &AnalysisConfig::EnableUseGpu,
            py::arg("memory_pool_init_size_mb"), py::arg("device_id") = 0)
+      .def("enable_xpu", &AnalysisConfig::EnableXpu,
+           py::arg("l3_workspace_size"))
       .def("disable_gpu", &AnalysisConfig::DisableGpu)
       .def("use_gpu", &AnalysisConfig::use_gpu)
       .def("gpu_device_id", &AnalysisConfig::gpu_device_id)