Merge topic 'cuda_support_nvcc_from_hpctoolkit'

3cef91a321 CUDA: Always extract CUDA Toolkit root from nvcc verbose output

Acked-by: Kitware Robot <kwrobot@kitware.com>
Reviewed-by: Raul Tambre <raul@tambre.ee>
Tested-by: Raul Tambre <raul@tambre.ee>
Merge-request: !5777

3 files changed, 32 insertions, 38 deletions

diff --git a/Modules/CMakeDetermineCUDACompiler.cmake b/Modules/CMakeDetermineCUDACompiler.cmake
index c77fc3a..468b8a1 100644
--- a/Modules/CMakeDetermineCUDACompiler.cmake
+++ b/Modules/CMakeDetermineCUDACompiler.cmake
@@ -172,23 +172,21 @@ if(NOT CMAKE_CUDA_COMPILER_ID_RUN)
       endif()
     endif()
 
-    # If NVCC is a symlink due to a wrapper script (e.g. ccache or colornvcc), then invoke it to find the
-    # real non-scattered toolkit.
-    if(IS_SYMLINK ${_CUDA_NVCC_EXECUTABLE})
-      execute_process(COMMAND ${_CUDA_NVCC_EXECUTABLE} "-v" "__cmake_determine_cuda" ERROR_VARIABLE NVCC_ERR)
-      if(NVCC_ERR MATCHES " _HERE_=([^\r\n]*)")
-        set(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${CMAKE_MATCH_1}")
-      else()
-        message(FATAL_ERROR "Could not execute nvcc with -v.")
-      endif()
-      unset(NVCC_ERR)
+    # Given that NVCC can be provided by multiple different sources (NVIDIA HPC SDK, CUDA Toolkit, distro)
+    # each of which has a different layout, we need to extract the CUDA toolkit root from the compiler
+    # itself, allowing us to support numerous different scattered toolkit layouts
+    execute_process(COMMAND ${_CUDA_NVCC_EXECUTABLE} "-v" "__cmake_determine_cuda" ERROR_VARIABLE NVCC_ERR)
+    if(NVCC_ERR MATCHES "TOP=([^\r\n]*)")
+      get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${CMAKE_MATCH_1}" ABSOLUTE)
     else()
       get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${_CUDA_NVCC_EXECUTABLE}" DIRECTORY)
+      get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}" DIRECTORY)
     endif()
+    unset(NVCC_ERR)
+
+    set(CMAKE_CUDA_DEVICE_LINKER "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/bin/nvlink${CMAKE_EXECUTABLE_SUFFIX}")
+    set(CMAKE_CUDA_FATBINARY "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/bin/fatbinary${CMAKE_EXECUTABLE_SUFFIX}")
 
-    set(CMAKE_CUDA_DEVICE_LINKER "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/nvlink${CMAKE_EXECUTABLE_SUFFIX}")
-    set(CMAKE_CUDA_FATBINARY "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}/fatbinary${CMAKE_EXECUTABLE_SUFFIX}")
-    get_filename_component(CMAKE_CUDA_COMPILER_TOOLKIT_ROOT "${CMAKE_CUDA_COMPILER_TOOLKIT_ROOT}" DIRECTORY)
 
     # In a non-scattered installation the following are equivalent to CMAKE_CUDA_COMPILER_TOOLKIT_ROOT.
     # We first check for a non-scattered installation to prefer it over a scattered installation.
diff --git a/Modules/FindCUDA.cmake b/Modules/FindCUDA.cmake
index 240f0a5..3d10aef 100644
--- a/Modules/FindCUDA.cmake
+++ b/Modules/FindCUDA.cmake
@@ -834,21 +834,18 @@ if(NOT CUDA_TOOLKIT_ROOT_DIR AND NOT CMAKE_CROSSCOMPILING)
     )
 
   if (CUDA_TOOLKIT_ROOT_DIR_NVCC)
-    # If NVCC is a symlink due to a wrapper script (e.g. ccache or colornvcc), then invoke it to find the
-    # real non-scattered toolkit.
-    if(IS_SYMLINK ${CUDA_TOOLKIT_ROOT_DIR_NVCC})
-      execute_process(COMMAND ${CUDA_TOOLKIT_ROOT_DIR_NVCC} "-v" "__cmake_determine_cuda" ERROR_VARIABLE NVCC_ERR)
-      if(NVCC_ERR MATCHES " _HERE_=([^\r\n]*)")
-        set(CUDA_TOOLKIT_ROOT_DIR_NVCC_PAR "${CMAKE_MATCH_1}")
-      else()
-        message(FATAL_ERROR "Could not execute nvcc with -v.")
-      endif()
-      unset(NVCC_ERR)
+    # Given that NVCC can be provided by multiple different sources (NVIDIA HPC SDK, CUDA Toolkit, distro)
+    # each of which has a different layout, we need to extract the CUDA toolkit root from the compiler
+    # itself, allowing us to support numerous different scattered toolkit layouts
+    execute_process(COMMAND ${CUDA_TOOLKIT_ROOT_DIR_NVCC} "-v" "__cmake_determine_cuda" ERROR_VARIABLE NVCC_ERR)
+    if(NVCC_ERR MATCHES "TOP=([^\r\n]*)")
+      get_filename_component(CUDA_TOOLKIT_ROOT_DIR "${CMAKE_MATCH_1}" ABSOLUTE CACHE)
     else()
-      get_filename_component(CUDA_TOOLKIT_ROOT_DIR_NVCC_PAR "${CUDA_TOOLKIT_ROOT_DIR_NVCC}" DIRECTORY)
+      get_filename_component(CUDA_TOOLKIT_ROOT_DIR "${CUDA_TOOLKIT_ROOT_DIR_NVCC}" DIRECTORY)
+      get_filename_component(CUDA_TOOLKIT_ROOT_DIR "${CUDA_TOOLKIT_ROOT_DIR}" DIRECTORY CACHE)
     endif()
+    unset(NVCC_ERR)
 
-    get_filename_component(CUDA_TOOLKIT_ROOT_DIR "${CUDA_TOOLKIT_ROOT_DIR_NVCC_PAR}" DIRECTORY CACHE)
     string(REGEX REPLACE "[/\\\\]?bin[64]*[/\\\\]?$" "" CUDA_TOOLKIT_ROOT_DIR ${CUDA_TOOLKIT_ROOT_DIR})
     # We need to force this back into the cache.
     set(CUDA_TOOLKIT_ROOT_DIR ${CUDA_TOOLKIT_ROOT_DIR} CACHE PATH "Toolkit location." FORCE)
diff --git a/Modules/FindCUDAToolkit.cmake b/Modules/FindCUDAToolkit.cmake
index ac2516c..83de97b 100644
--- a/Modules/FindCUDAToolkit.cmake
+++ b/Modules/FindCUDAToolkit.cmake
@@ -518,24 +518,23 @@ else()
         )
       endif()
 
-      if(CUDAToolkit_NVCC_EXECUTABLE)
-        # If NVCC is a symlink due to a wrapper script (e.g. ccache or colornvcc), then invoke it to find the
-        # real non-scattered toolkit.
-        if(IS_SYMLINK ${CUDAToolkit_NVCC_EXECUTABLE})
-          execute_process(COMMAND ${CUDAToolkit_NVCC_EXECUTABLE} "-v" "__cmake_determine_cuda" ERROR_VARIABLE NVCC_ERR)
-          if(NVCC_ERR MATCHES " _HERE_=([^\r\n]*)")
-            set(CUDAToolkit_BIN_DIR "${CMAKE_MATCH_1}")
-          else()
-            message(FATAL_ERROR "Could not execute nvcc with -v.")
-          endif()
-          unset(NVCC_ERR)
+      if(EXISTS "${CUDAToolkit_NVCC_EXECUTABLE}")
+        # If NVCC exists  then invoke it to find the toolkit location.
+        # This allows us to support wrapper scripts (e.g. ccache or colornvcc), CUDA Toolkit,
+        # NVIDIA HPC SDK, and distro's splayed layouts
+        execute_process(COMMAND ${CUDAToolkit_NVCC_EXECUTABLE} "-v" "__cmake_determine_cuda" ERROR_VARIABLE NVCC_ERR)
+        if(NVCC_ERR MATCHES "TOP=([^\r\n]*)")
+          get_filename_component(CUDAToolkit_BIN_DIR "${CMAKE_MATCH_1}/bin" ABSOLUTE)
         else()
           get_filename_component(CUDAToolkit_BIN_DIR "${CUDAToolkit_NVCC_EXECUTABLE}" DIRECTORY)
         endif()
+        unset(NVCC_ERR)
 
-        set(CUDAToolkit_BIN_DIR "${CUDAToolkit_BIN_DIR}" CACHE PATH "" FORCE)
         mark_as_advanced(CUDAToolkit_BIN_DIR)
-      elseif(CUDAToolkit_SENTINEL_FILE)
+        set(CUDAToolkit_BIN_DIR "${CUDAToolkit_BIN_DIR}" CACHE PATH "" FORCE)
+      endif()
+
+      if(CUDAToolkit_SENTINEL_FILE)
         get_filename_component(CUDAToolkit_BIN_DIR ${CUDAToolkit_SENTINEL_FILE} DIRECTORY ABSOLUTE)
         set(CUDAToolkit_BIN_DIR "${CUDAToolkit_BIN_DIR}/bin")