solve cuda_x86_xpu compile problem. test=develop (#3353)

lite/CMakeLists.txt

@@ -108,59 +108,53 @@ if (LITE_WITH_PYTHON)
     add_dependencies(publish_inference publish_inference_python_light_demo)
 endif()
 
+if (LITE_WITH_CUDA OR LITE_WITH_X86)
+    add_custom_target(publish_inference_cxx_lib ${TARGET}
+        COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
+        COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/bin"
+        COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/cxx/include"
+        COMMAND cp "${CMAKE_SOURCE_DIR}/lite/api/paddle_*.h" "${INFER_LITE_PUBLISH_ROOT}/cxx/include"
+        COMMAND cp "${CMAKE_BINARY_DIR}/libpaddle_api_full_bundled.a" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
+        COMMAND cp "${CMAKE_BINARY_DIR}/libpaddle_api_light_bundled.a" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
+        COMMAND cp "${CMAKE_BINARY_DIR}/lite/api/*.so" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
+        )
+    add_custom_target(publish_inference_third_party ${TARGET}
+            COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/third_party"
+            COMMAND cp -r "${CMAKE_BINARY_DIR}/third_party/install/*" "${INFER_LITE_PUBLISH_ROOT}/third_party")
+    add_dependencies(publish_inference_cxx_lib bundle_full_api)
+    add_dependencies(publish_inference_cxx_lib bundle_light_api)
+    add_dependencies(publish_inference_cxx_lib paddle_full_api_shared)
+    add_dependencies(publish_inference_cxx_lib paddle_light_api_shared)
+    add_dependencies(publish_inference publish_inference_cxx_lib)
+    add_dependencies(publish_inference publish_inference_third_party)
+endif()
+
 if (LITE_WITH_X86)
     add_custom_target(publish_inference_x86_cxx_lib ${TARGET}
-            COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
             COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/bin"
-            COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/cxx/include"
-            COMMAND cp "${CMAKE_SOURCE_DIR}/lite/api/paddle_*.h" "${INFER_LITE_PUBLISH_ROOT}/cxx/include"
-            COMMAND cp "${CMAKE_BINARY_DIR}/libpaddle_api_full_bundled.a" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
-            COMMAND cp "${CMAKE_BINARY_DIR}/libpaddle_api_light_bundled.a" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
-            COMMAND cp "${CMAKE_BINARY_DIR}/lite/api/*.so" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
             COMMAND cp "${CMAKE_BINARY_DIR}/lite/api/test_model_bin" "${INFER_LITE_PUBLISH_ROOT}/bin"
             )
-    add_dependencies(publish_inference_x86_cxx_lib bundle_full_api)
-    add_dependencies(publish_inference_x86_cxx_lib bundle_light_api)
     add_dependencies(publish_inference_x86_cxx_lib test_model_bin)
-    add_dependencies(publish_inference_x86_cxx_lib paddle_full_api_shared)
-    add_dependencies(publish_inference_x86_cxx_lib paddle_light_api_shared)
-    add_dependencies(publish_inference publish_inference_x86_cxx_lib)
 
     add_custom_target(publish_inference_x86_cxx_demos ${TARGET}
            COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/third_party"
-           COMMAND cp -r "${CMAKE_BINARY_DIR}/third_party/install/*" "${INFER_LITE_PUBLISH_ROOT}/third_party"
            COMMAND cp -r "${CMAKE_BINARY_DIR}/third_party/eigen3" "${INFER_LITE_PUBLISH_ROOT}/third_party"
-           COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/demo/cxx"
            )
     add_dependencies(publish_inference_x86_cxx_lib publish_inference_x86_cxx_demos)
     add_dependencies(publish_inference_x86_cxx_demos paddle_full_api_shared eigen3)
+    add_dependencies(publish_inference publish_inference_x86_cxx_lib)
+    add_dependencies(publish_inference publish_inference_x86_cxx_demos)
 endif()
 
 if(LITE_WITH_CUDA)
-    add_custom_target(publish_inference_cuda_cxx_lib ${TARGET}
-            COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
-            COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/bin"
-            COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/cxx/include"
-            COMMAND cp "${CMAKE_SOURCE_DIR}/lite/api/paddle_*.h" "${INFER_LITE_PUBLISH_ROOT}/cxx/include"
-            COMMAND cp "${CMAKE_BINARY_DIR}/libpaddle_api_full_bundled.a" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
-            COMMAND cp "${CMAKE_BINARY_DIR}/libpaddle_api_light_bundled.a" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
-            COMMAND cp "${CMAKE_BINARY_DIR}/lite/api/*.so" "${INFER_LITE_PUBLISH_ROOT}/cxx/lib"
-        )
-    add_dependencies(publish_inference_cuda_cxx_lib bundle_full_api)
-    add_dependencies(publish_inference_cuda_cxx_lib bundle_light_api)
-    add_dependencies(publish_inference_cuda_cxx_lib paddle_full_api_shared)
-    add_dependencies(publish_inference_cuda_cxx_lib paddle_light_api_shared)
-    add_dependencies(publish_inference publish_inference_cuda_cxx_lib)
-
     add_custom_target(publish_inference_cuda_cxx_demos ${TARGET}
-           COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/third_party"
-           COMMAND cp -r "${CMAKE_BINARY_DIR}/third_party/install/*" "${INFER_LITE_PUBLISH_ROOT}/third_party"
            COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/demo/cxx"
            COMMAND cp -r "${CMAKE_SOURCE_DIR}/lite/demo/cxx/cuda_demo/*" "${INFER_LITE_PUBLISH_ROOT}/demo/cxx"
            )
-    add_dependencies(publish_inference_cuda_cxx_lib publish_inference_cuda_cxx_demos)
     add_dependencies(publish_inference_cuda_cxx_demos paddle_full_api_shared)
-endif(LITE_WITH_CUDA) 
+    add_dependencies(publish_inference publish_inference_cuda_cxx_demos)
+endif(LITE_WITH_CUDA)
+
 if (LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND LITE_WITH_ARM)
     if (NOT LITE_ON_TINY_PUBLISH)
         # add cxx lib

lite/api/cxx_api_impl.cc

@@ -31,10 +31,17 @@ namespace lite {
 
 void CxxPaddleApiImpl::Init(const lite_api::CxxConfig &config) {
   config_ = config;
+  auto places = config.valid_places();
 #ifdef LITE_WITH_CUDA
-  Env<TARGET(kCUDA)>::Init();
+  // if kCUDA is included in valid places, it should be initialized first,
+  // otherwise skip this step.
+  for (auto &p : places) {
+    if (p.target == TARGET(kCUDA)) {
+      Env<TARGET(kCUDA)>::Init();
+      break;
+    }
+  }
 #endif
-  auto places = config.valid_places();
   std::vector<std::string> passes{};
   auto use_layout_preprocess_pass =
       config.model_dir().find("OPENCL_PRE_PRECESS");

lite/core/context.h

@@ -181,7 +181,11 @@ class Context<TargetType::kCUDA> {
       Env<TargetType::kCUDA>::Global();
   // NOTE: InitOnce should only be used by ContextScheduler
   void InitOnce() {
-    cublas_fp32_ = std::make_shared<lite::cuda::Blas<float>>();
+    if (devs.size() > 0) {
+      cublas_fp32_ = std::make_shared<lite::cuda::Blas<float>>();
+    } else {
+      LOG(INFO) << "No cuda device(s) found, CUDAContext init failed.";
+    }
   }
   void Init(int dev_id, int exec_stream_id = 0, int io_stream_id = 0) {
     CHECK_GT(devs.size(), 0UL)

lite/core/device_info.h

@@ -142,7 +142,7 @@ class Env {
     // Get device count
     count = API::num_devices();
     if (count == 0) {
-      CHECK(false) << "No device found!";
+      LOG(INFO) << "No " << TargetToStr(Type) << " device(s) found!";
     } else {
       LOG(INFO) << "Found " << count << " device(s)";
     }