🐛 [Bug] Windows build broken after 1.2.0 - experimental:filesystem in TRTEngine

[bharrisau]

Bug Description

At some point after v1.2.0 the TensorRT\core\runtime\TRTEngine.h file was added and it brought the #include <experimental/filesystem> with it. I had some issues trying to build due to this. A #define is needed to build, the path cannot implicitly be converted to a std::string, an extra #include is needed to use std::stringbuilder and there is no stdc++fs to link against.

Environment

• Torch-TensorRT Version: master f43be5b
• PyTorch Version: libtorch 1.13.1+cu117
• CPU Architecture: amd64
• OS: Windows 11
• Python version: Python 3.10.9
• CUDA version: 11.8
• cmake version: cmake version 3.21.21080301-MSVC_2
• cl version: Microsoft (R) C/C++ Optimizing Compiler Version 19.30.30706 for x64
• MSVC toolchain version: 14.30.30705

Rough fix

The below changes were needed to get a compile. They probably would break the build on other systems - so maybe need to be based on some platform detection.

Additional CMAKE flag

{
  "name": "CMAKE_CXX_FLAGS",
  "value": "-D_SILENCE_EXPERIMENTAL_FILESYSTEM_DEPRECATION_WARNING",
  "type": "STRING"
}

Code changes

diff --git a/core/runtime/CMakeLists.txt b/core/runtime/CMakeLists.txt
index d21c661a..b3632e8f 100644
--- a/core/runtime/CMakeLists.txt
+++ b/core/runtime/CMakeLists.txt
@@ -33,7 +33,6 @@ target_link_libraries(${lib_name}
         TensorRT::nvinfer
         torch
         core_util
-        stdc++fs
 )

 # Install
diff --git a/core/runtime/TRTEngine.h b/core/runtime/TRTEngine.h
index 5615a824..1a2df1f2 100644
--- a/core/runtime/TRTEngine.h
+++ b/core/runtime/TRTEngine.h
@@ -25,7 +25,7 @@ struct TRTEngine : torch::CustomClassHolder {
   std::string name;
   RTDevice device_info;

-  std::string profile_path_prefix = std::experimental::filesystem::temp_directory_path();
+  std::string profile_path_prefix = std::experimental::filesystem::temp_directory_path().string();

   std::unordered_map<uint64_t, uint64_t> in_binding_map = {}; // TRT IDX -> PYT IDX
   std::unordered_map<uint64_t, uint64_t> out_binding_map = {}; // TRT IDX -> PYT IDX
diff --git a/core/runtime/TRTEngineProfiler.cpp b/core/runtime/TRTEngineProfiler.cpp
index a1159516..7857a467 100644
--- a/core/runtime/TRTEngineProfiler.cpp
+++ b/core/runtime/TRTEngineProfiler.cpp
@@ -1,5 +1,6 @@
 #include <algorithm>
 #include <fstream>
+#include <sstream>
 #include <iomanip>

 #include "core/runtime/TRTEngineProfiler.h"

[narendasan]

@bharrisau can you open a PR for this?

[bharrisau]

I'm 99% sure my changes will break all the other builds though. Did you want me to just take a best guess at hiding the WIN changes behind define/etc?

[narendasan]

Yeah that will at least give us a starting point and we can work to make sure it doesn't effect linux

[Christina-Young-NVIDIA]

Awaiting PR.

[bharrisau]

Done.