[Paddle-TRT] Add trt runtime version check (#32443)

* add trt runtime version check

* use different wrap, and change to major version check

paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc

@@ -20,6 +20,7 @@
 #include "paddle/fluid/inference/analysis/helper.h"
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 #include "paddle/fluid/inference/tensorrt/engine.h"
+#include "paddle/fluid/inference/tensorrt/helper.h"
 #include "paddle/fluid/inference/tensorrt/op_teller.h"
 
 namespace paddle {
@@ -321,11 +322,20 @@ void TensorRtSubgraphPass::CreateTensorRTOp(
     opt_input_shape = {};
   }
 
-  if (min_input_shape.size() > 0 && TRT_VERSION > 6000) {
+  auto to_major_version = [&](int full_version) -> float {
+    return (full_version / 100) / 10.0;
+  };
+  const float compile_time_trt_version = to_major_version(TRT_VERSION);
+  const float run_time_trt_version =
+      to_major_version(tensorrt::GetInferLibVersion());
+  if (compile_time_trt_version != run_time_trt_version) {
     LOG_FIRST_N(WARNING, 1)
-        << "The Paddle lib links the " << TRT_VERSION << " version TensorRT, "
-        << "make sure the runtime TensorRT you are using is no less than this "
-           "version, otherwise, there might be Segfault!";
+        << "The Paddle Inference library is compiled with "
+        << compile_time_trt_version << " version TensorRT, "
+        << "but the runtime TensorRT you are using is " << run_time_trt_version
+        << " version. "
+           "This might cause serious compatibility issues. We strongly "
+           "recommend using the same TRT version at runtime.";
   }
 
   // Setting the disable_trt_plugin_fp16 to true means that TRT plugin will not

paddle/fluid/inference/tensorrt/helper.h

@@ -60,6 +60,9 @@ static nvinfer1::IRuntime* createInferRuntime(nvinfer1::ILogger* logger) {
 static nvinfer1::IPluginRegistry* GetPluginRegistry() {
   return static_cast<nvinfer1::IPluginRegistry*>(dy::getPluginRegistry());
 }
+static int GetInferLibVersion() {
+  return static_cast<int>(dy::getInferLibVersion());
+}
 #endif
 
 // A logger for create TensorRT infer builder.

paddle/fluid/platform/dynload/tensorrt.cc

@@ -27,7 +27,8 @@ void* tensorrt_plugin_dso_handle;
 
 #define DEFINE_WRAP(__name) DynLoad__##__name __name
 
-TENSORRT_RAND_ROUTINE_EACH(DEFINE_WRAP);
+TENSORRT_RAND_ROUTINE_EACH_POINTER(DEFINE_WRAP);
+TENSORRT_RAND_ROUTINE_EACH_NON_POINTER(DEFINE_WRAP);
 TENSORRT_PLUGIN_RAND_ROUTINE_EACH(DEFINE_WRAP);
 
 void* GetDsoHandle(const std::string& dso_name) {

paddle/fluid/platform/dynload/tensorrt.h

@@ -37,7 +37,7 @@ void* GetTensorRtPluginHandle();
 extern std::once_flag tensorrt_plugin_dso_flag;
 extern void* tensorrt_plugin_dso_handle;
 
-#define DECLARE_DYNAMIC_LOAD_TENSORRT_WRAP(__name)                            \
+#define DECLARE_DYNAMIC_LOAD_TENSORRT_POINTER_WRAP(__name)                    \
   struct DynLoad__##__name {                                                  \
     template <typename... Args>                                               \
     void* operator()(Args... args) {                                          \
@@ -55,6 +55,23 @@ extern void* tensorrt_plugin_dso_handle;
   };                                                                          \
   extern DynLoad__##__name __name
 
+#define DECLARE_DYNAMIC_LOAD_TENSORRT_NON_POINTER_WRAP(__name)                \
+  struct DynLoad__##__name {                                                  \
+    template <typename... Args>                                               \
+    auto operator()(Args... args) -> DECLARE_TYPE(__name, args...) {          \
+      std::call_once(tensorrt_dso_flag, []() {                                \
+        tensorrt_dso_handle = paddle::platform::dynload::GetTensorRtHandle(); \
+      });                                                                     \
+      static void* p_##__name = dlsym(tensorrt_dso_handle, #__name);          \
+      PADDLE_ENFORCE_NOT_NULL(p_##__name,                                     \
+                              platform::errors::Unavailable(                  \
+                                  "Load tensorrt api %s failed", #__name));   \
+      using tensorrt_func = decltype(&::__name);                              \
+      return reinterpret_cast<tensorrt_func>(p_##__name)(args...);            \
+    }                                                                         \
+  };                                                                          \
+  extern DynLoad__##__name __name
+
 #define DECLARE_DYNAMIC_LOAD_TENSORRT_PLUGIN_WRAP(__name)                      \
   struct DynLoad__##__name {                                                   \
     template <typename... Args>                                                \
@@ -76,20 +93,25 @@ extern void* tensorrt_plugin_dso_handle;
 #ifdef NV_TENSORRT_MAJOR
 
 #if (NV_TENSORRT_MAJOR >= 6)
-#define TENSORRT_RAND_ROUTINE_EACH(__macro) \
+#define TENSORRT_RAND_ROUTINE_EACH_POINTER(__macro) \
   __macro(createInferBuilder_INTERNAL);             \
   __macro(createInferRuntime_INTERNAL);             \
   __macro(getPluginRegistry);
 #else
-#define TENSORRT_RAND_ROUTINE_EACH(__macro) \
+#define TENSORRT_RAND_ROUTINE_EACH_POINTER(__macro) \
   __macro(createInferBuilder_INTERNAL);             \
   __macro(createInferRuntime_INTERNAL);
 #endif
 
+#define TENSORRT_RAND_ROUTINE_EACH_NON_POINTER(__macro) \
+  __macro(getInferLibVersion);
+
 #define TENSORRT_PLUGIN_RAND_ROUTINE_EACH(__macro) \
   __macro(initLibNvInferPlugins);
 
-TENSORRT_RAND_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_TENSORRT_WRAP)
+TENSORRT_RAND_ROUTINE_EACH_POINTER(DECLARE_DYNAMIC_LOAD_TENSORRT_POINTER_WRAP)
+TENSORRT_RAND_ROUTINE_EACH_NON_POINTER(
+    DECLARE_DYNAMIC_LOAD_TENSORRT_NON_POINTER_WRAP)
 TENSORRT_PLUGIN_RAND_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_TENSORRT_PLUGIN_WRAP)
 
 #endif  // end of NV_TENSORRT_MAJOR