Set CMAKE_CUDA_ARCHITECTURES early.

CMake/ClangCompilers.cmake

@@ -19,6 +19,10 @@ if(QMC_OMP)
         CACHE STRING "Offload target architecture")
     set(OPENMP_OFFLOAD_COMPILE_OPTIONS "-fopenmp-targets=${OFFLOAD_TARGET}")
 
+    if(NOT DEFINED OFFLOAD_ARCH AND OFFLOAD_TARGET MATCHES "nvptx64" AND DEFINED CMAKE_CUDA_ARCHITECTURES)
+      set(OFFLOAD_ARCH sm_${CMAKE_CUDA_ARCHITECTURES})
+    endif()
+
     if(DEFINED OFFLOAD_ARCH)
       set(OPENMP_OFFLOAD_COMPILE_OPTIONS
           "${OPENMP_OFFLOAD_COMPILE_OPTIONS} -Xopenmp-target=${OFFLOAD_TARGET} -march=${OFFLOAD_ARCH}")

CMake/NVHPCCompilers.cmake

@@ -8,8 +8,12 @@ if(QMC_OMP)
   if(ENABLE_OFFLOAD AND NOT CMAKE_SYSTEM_NAME STREQUAL "CrayLinuxEnvironment")
     message(WARNING "QMCPACK OpenMP offload is not ready for NVIDIA HPC compiler.")
     if(NOT DEFINED OFFLOAD_ARCH)
-      message(FATAL_ERROR "NVIDIA HPC compiler requires -gpu=ccXX option set based on the target GPU architecture! "
-                          "Please add -DOFFLOAD_ARCH=ccXX to cmake. For example, cc70 is for Volta.")
+      if(DEFINED CMAKE_CUDA_ARCHITECTURES)
+        set(OFFLOAD_ARCH cc${CMAKE_CUDA_ARCHITECTURES})
+      else()
+        message(FATAL_ERROR "NVIDIA HPC compiler requires -gpu=ccXX option set based on the target GPU architecture! "
+                            "Please add -DOFFLOAD_ARCH=ccXX to cmake. For example, cc70 is for Volta.")
+      endif()
     endif()
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mp=gpu")
     set(OPENMP_OFFLOAD_COMPILE_OPTIONS "-gpu=${OFFLOAD_ARCH}")

CMakeLists.txt

@@ -72,6 +72,17 @@ if(ENABLE_CUDA AND QMC_CUDA)
   message(FATAL_ERROR "ENABLE_CUDA=ON and QMC_CUDA=ON can not be set together!")
 endif(ENABLE_CUDA AND QMC_CUDA)
 
+# set CMAKE_CUDA_ARCHITECTURES early such that offload compilers may take advantage of it
+if(ENABLE_CUDA OR QMC_CUDA AND NOT QMC_CUDA2HIP)
+  if(DEFINED CUDA_ARCH)
+    unset(CUDA_ARCH CACHE)
+    message(FATAL_ERROR "CUDA_ARCH option has been removed. Use -DCMAKE_CUDA_ARCHITECTURES=80 if -DCUDA_ARCH=sm_80 was used.")
+  endif()
+  if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+    set(CMAKE_CUDA_ARCHITECTURES 70)
+  endif()
+endif()
+
 #--------------------------------------------------------------------
 # Set compiler-time parameters
 # WALKER_MAX_PROPERTIES max number of observables + 12 or so standard
@@ -665,19 +676,12 @@ if(QMC_CUDA OR ENABLE_CUDA)
     if (CMAKE_VERSION VERSION_LESS 3.18.0)
       message(FATAL_ERROR "QMC_CUDA or ENABLE_CUDA require CMake 3.18.0 or later")
     endif()
-    if(DEFINED CUDA_ARCH)
-      unset(CUDA_ARCH CACHE)
-      message(FATAL_ERROR "CUDA_ARCH option has been removed. Use -DCMAKE_CUDA_ARCHITECTURES=80 if -DCUDA_ARCH=sm_80 was used.")
-    endif()
     # a few production machines use CUDA 10 which only supports C++14.
     if(NOT DEFINED CMAKE_CUDA_STANDARD)
       set(CMAKE_CUDA_STANDARD 14)
     endif()
     set(CMAKE_CUDA_STANDARD_REQUIRED TRUE)
     set(CMAKE_CUDA_EXTENSIONS OFF)
-    if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-      set(CMAKE_CUDA_ARCHITECTURES 70)
-    endif()
     enable_language(CUDA)
     find_package(CUDAToolkit REQUIRED)
     # Automatically set the default NVCC flags