[Reland] Upgrade NVTX to NVTX3 (#97582)

PR #90689 replaces NVTX with NVTX3. However, the torch::nvtoolsext is created only when the third party NVTX is used.
 This is clear a logical error. We now move the creation code out of the branch to cover all cases. This should fix the issues reported in the comments of  #90689.

It would be better to move configurations of the failed FRL jobs to CI tests so that we can find such issues early before merging.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97582
Approved by: https://github.com/peterbell10

.gitmodules

@@ -157,3 +157,6 @@
 [submodule "third_party/mimalloc"]
 	path = third_party/mimalloc
 	url = https://github.com/microsoft/mimalloc.git
+[submodule "third_party/NVTX"]
+	path = third_party/NVTX
+	url = https://github.com/NVIDIA/NVTX.git

caffe2/CMakeLists.txt

@@ -1516,7 +1516,8 @@ if(USE_CUDA)
     target_link_libraries(torch_cpu PRIVATE torch::cudart)
   endif()
   target_link_libraries(torch_cuda INTERFACE torch::cudart)
-  target_link_libraries(torch_cuda PUBLIC c10_cuda torch::nvtoolsext)
+  target_link_libraries(torch_cuda PUBLIC c10_cuda)
+  target_link_libraries(torch_cuda PRIVATE torch::nvtoolsext)
 
   target_include_directories(
       torch_cuda INTERFACE $<INSTALL_INTERFACE:include>)
@@ -1571,7 +1572,7 @@ if(BUILD_SHARED_LIBS)
   # not find them, because they're usually in non-standard locations)
   if(USE_CUDA)
     target_link_libraries(torch_global_deps ${Caffe2_PUBLIC_CUDA_DEPENDENCY_LIBS})
-    target_link_libraries(torch_global_deps torch::cudart torch::nvtoolsext)
+    target_link_libraries(torch_global_deps torch::cudart)
   endif()
   if(USE_TBB)
     target_link_libraries(torch_global_deps TBB::tbb)

cmake/TorchConfig.cmake.in

@@ -129,30 +129,18 @@ endif()
 
 if(@USE_CUDA@)
   if(MSVC)
-    if(NOT NVTOOLEXT_HOME)
-      set(NVTOOLEXT_HOME "C:/Program Files/NVIDIA Corporation/NvToolsExt")
-    endif()
-    if(DEFINED ENV{NVTOOLSEXT_PATH})
-      set(NVTOOLEXT_HOME $ENV{NVTOOLSEXT_PATH})
-    endif()
-    set(TORCH_CUDA_LIBRARIES
-      ${NVTOOLEXT_HOME}/lib/x64/nvToolsExt64_1.lib
-      ${CUDA_LIBRARIES})
-    list(APPEND TORCH_INCLUDE_DIRS ${NVTOOLEXT_HOME}/include)
+    set(TORCH_CUDA_LIBRARIES ${CUDA_LIBRARIES})
     find_library(CAFFE2_NVRTC_LIBRARY caffe2_nvrtc PATHS "${TORCH_INSTALL_PREFIX}/lib")
     list(APPEND TORCH_CUDA_LIBRARIES ${CAFFE2_NVRTC_LIBRARY})
   elseif(APPLE)
     set(TORCH_CUDA_LIBRARIES
       ${CUDA_TOOLKIT_ROOT_DIR}/lib/libcudart.dylib
       ${CUDA_TOOLKIT_ROOT_DIR}/lib/libnvrtc.dylib
-      ${CUDA_TOOLKIT_ROOT_DIR}/lib/libnvToolsExt.dylib
       ${CUDA_LIBRARIES})
   else()
-    find_library(LIBNVTOOLSEXT libnvToolsExt.so PATHS ${CUDA_TOOLKIT_ROOT_DIR}/lib64/)
     set(TORCH_CUDA_LIBRARIES
       ${CUDA_CUDA_LIB}
       ${CUDA_NVRTC_LIB}
-      ${LIBNVTOOLSEXT}
       ${CUDA_LIBRARIES})
   endif()
   if(@BUILD_SHARED_LIBS@)

cmake/public/cuda.cmake

@@ -67,10 +67,6 @@ if(NOT CMAKE_CUDA_COMPILER_VERSION STREQUAL CUDAToolkit_VERSION OR
                       "V${CUDAToolkit_VERSION} in '${CUDAToolkit_INCLUDE_DIR}'")
 endif()
 
-if(NOT TARGET CUDA::nvToolsExt)
-  message(FATAL_ERROR "Failed to find nvToolsExt")
-endif()
-
 message(STATUS "Caffe2: CUDA detected: " ${CUDA_VERSION})
 message(STATUS "Caffe2: CUDA nvcc is: " ${CUDA_NVCC_EXECUTABLE})
 message(STATUS "Caffe2: CUDA toolkit directory: " ${CUDA_TOOLKIT_ROOT_DIR})
@@ -216,9 +212,10 @@ endif()
 
 # nvToolsExt
 add_library(torch::nvtoolsext INTERFACE IMPORTED)
-set_property(
-    TARGET torch::nvtoolsext PROPERTY INTERFACE_LINK_LIBRARIES
-    CUDA::nvToolsExt)
+find_path(nvtx3_dir NAMES nvtx3 PATHS "${CUDA_INCLUDE_DIRS}" "${CMAKE_CURRENT_LIST_DIR}/../../third_party/NVTX/c/include" NO_DEFAULT_PATH)
+find_package_handle_standard_args(nvtx3 DEFAULT_MSG nvtx3_dir)
+target_include_directories(torch::nvtoolsext INTERFACE "${nvtx3_dir}")
+
 
 # cublas
 add_library(caffe2::cublas INTERFACE IMPORTED)

setup.py

@@ -186,9 +186,6 @@
 #   NVFUSER_SOURCE_DIR
 #     specify nvfuser root directory
 #
-#   NVTOOLSEXT_PATH (Windows only)
-#     specify where nvtoolsext is installed
-#
 #   ACL_ROOT_DIR
 #     specify where Compute Library is installed
 #

third_party/NVTX

@@ -0,0 +1 @@
+Subproject commit e170594ac7cf1dac584da473d4ca9301087090c1

third_party/nvfuser/csrc/instrumentation.h

@@ -2,7 +2,11 @@
 
 #include <utils.h>
 
+#ifndef FBCODE_CAFFE2
+#include <nvtx3/nvToolsExt.h>
+#else
 #include <nvToolsExt.h>
+#endif
 
 // NOLINTNEXTLINE(modernize-deprecated-headers)
 #include <stdio.h>

torch/CMakeLists.txt

@@ -136,7 +136,6 @@ if(USE_CUDA)
         list(APPEND TORCH_PYTHON_LINK_LIBRARIES torch::cudnn)
         list(APPEND TORCH_PYTHON_COMPILE_DEFINITIONS USE_CUDNN)
     endif()
-
     list(APPEND TORCH_PYTHON_LINK_LIBRARIES torch::nvtoolsext)
 endif()
 

torch/csrc/cuda/shared/nvtx.cpp

@@ -1,7 +1,11 @@
 #ifdef _WIN32
 #include <wchar.h> // _wgetenv for nvtx
 #endif
+#ifndef FBCODE_CAFFE2
+#include <nvtx3/nvToolsExt.h>
+#else
 #include <nvToolsExt.h>
+#endif
 #include <torch/csrc/utils/pybind.h>
 
 namespace torch::cuda::shared {

torch/csrc/profiler/stubs/cuda.cpp

@@ -1,6 +1,10 @@
 #include <sstream>
 
+#ifndef FBCODE_CAFFE2
+#include <nvtx3/nvToolsExt.h>
+#else
 #include <nvToolsExt.h>
+#endif
 
 #include <c10/cuda/CUDAGuard.h>
 #include <c10/util/irange.h>

torch/utils/hipify/cuda_to_hip_mappings.py

@@ -645,6 +645,7 @@ CUDA_INCLUDE_MAP = collections.OrderedDict(
         ("cub/device/device_scan.cuh", ("hipcub/hipcub.hpp", CONV_INCLUDE, API_BLAS)),
         ("cub/device/device_select.cuh", ("hipcub/hipcub.hpp", CONV_INCLUDE, API_BLAS)),
         ("nvToolsExt.h", ("roctracer/roctx.h", CONV_INCLUDE, API_ROCTX)),
+        ("nvtx3/nvToolsExt.h", ("roctracer/roctx.h", CONV_INCLUDE, API_ROCTX)),
         ("nvml.h", ("rocm_smi/rocm_smi.h", CONV_INCLUDE, API_ROCMSMI)),
     ]
 )