Some changes to the device library to use runtime API

+ Add a simple CUDA compilation step during configure to determine the apporiate arhcitecture flags
+ Allow add_componenet_library to accept a library type different than the DRACO_LIBRARY_TYPE variable
+ Change the GPU_Device class to use the newer CUDA Runtime API
+ Remove the GPU_Module class (explicit module and contexts are not necessary in the Runtime API)
+ Remove "EXTERN C" decorations (these are also no longer necessary in the Runtime API since functions do not need to be loaded explicity)
+ Add a test that shows an example of calling code from both the host and device
+ Consolidate calls to simple GPU test kernels into a "basic_kernels" header and implementation file
+ Compile auxillary GPU test code into a static library callable by the individual gpu tests

config/component_macros.cmake

@@ -253,6 +253,7 @@ endmacro( add_component_executable )
 #   SOURCES      "file1.cc;file2.cc;..."
 #   HEADERS      "file1.hh;file2.hh;..."
 #   LIBRARY_NAME_PREFIX "rtt_"
+#   LIBRARY_TYPE "SHARED"
 #   VENDOR_LIST  "MPI;GSL"
 #   VENDOR_LIBS  "${MPI_CXX_LIBRARIES};${GSL_LIBRARIES}"
 #   VENDOR_INCLUDE_DIRS "${MPI_CXX_INCLUDE_DIR};${GSL_INCLUDE_DIR}"
@@ -279,7 +280,7 @@ macro( add_component_library )
   cmake_parse_arguments(
     acl
     "NOEXPORT"
-    "PREFIX;TARGET;LIBRARY_NAME;LIBRARY_NAME_PREFIX;LINK_LANGUAGE"
+    "PREFIX;TARGET;LIBRARY_NAME;LIBRARY_NAME_PREFIX;LIBRARY_TYPE;LINK_LANGUAGE"
     "HEADERS;SOURCES;TARGET_DEPS;VENDOR_LIST;VENDOR_LIBS;VENDOR_INCLUDE_DIRS"
     ${ARGV}
     )
@@ -305,6 +306,11 @@ macro( add_component_library )
     endif()
   endif()
 
+  # if a library type was not specified use the default Draco setting
+  if(NOT acl_LIBRARY_TYPE)
+    set( acl_LIBRARY_TYPE ${DRACO_LIBRARY_TYPE})
+  endif()
+
   #
   # Create the library and set the properties
   #
@@ -314,7 +320,7 @@ macro( add_component_library )
   # extract project name, minus leading "Lib_"
   string( REPLACE "Lib_" "" folder_name ${acl_TARGET} )
 
-  add_library( ${acl_TARGET} ${DRACO_LIBRARY_TYPE} ${acl_SOURCES} )
+  add_library( ${acl_TARGET} ${acl_LIBRARY_TYPE} ${acl_SOURCES} )
   # Some properties are set at a global scope in compilerEnv.cmake:
   # - C_STANDARD, C_EXTENSIONS, CXX_STANDARD, CXX_EXTENSIONS,
   #   CXX_STANDARD_REQUIRED, and POSITION_INDEPENDENT_CODE

config/query_gpu.cu

@@ -0,0 +1,38 @@
+//----------------------------------*-C++-*----------------------------------//
+/*!
+ * \file   config/query_gpu.cu
+ * \author Alex Long
+ * \brief  Small CUDA code that prints the architecture version, used by CMake
+ * \date   Thu Mat 21 15:53:51 2019
+ * \note   Copyright (C) 2019 Triad National Security, LLC.
+ *         All rights reserved. */
+//---------------------------------------------------------------------------//
+
+// NOTE: This code is from
+// wagonhelm.github.io/articles/2018-03/detecting-cuda-capability-with-cmake
+
+#include <stdio.h>
+
+int main(int argc, char **argv) {
+  cudaDeviceProp dP;
+  float min_cc = 3.0;
+
+  int rc = cudaGetDeviceProperties(&dP, 0);
+  if (rc != cudaSuccess) {
+    cudaError_t error = cudaGetLastError();
+    printf("CUDA error: %s", cudaGetErrorString(error));
+    return rc; /* Failure */
+  }
+  if ((dP.major + (dP.minor / 10)) < min_cc) {
+    printf("Min Compute Capability of %2.1f required:  %d.%d found\n Not "
+           "Building CUDA Code",
+           min_cc, dP.major, dP.minor);
+    return 1; /* Failure */
+  } else {
+    printf("-arch=sm_%d%d", dP.major, dP.minor);
+    return 0; /* Success */
+  }
+}
+//---------------------------------------------------------------------------//
+//  end of query_gpu.cu
+//---------------------------------------------------------------------------//

config/vendor_libraries.cmake

@@ -299,6 +299,32 @@ macro( setupCudaEnv )
     set( CUDA_DBS_STRING "CUDA" CACHE BOOL
       "If CUDA is available, this variable is 'CUDA'")
 
+      set(OUTPUTFILE ${CMAKE_CURRENT_SOURCE_DIR}/config/cuda_script) # No suffix required
+      set(CUDAFILE ${CMAKE_CURRENT_SOURCE_DIR}/config/query_gpu.cu)
+      execute_process(COMMAND nvcc -lcuda ${CUDAFILE} -o ${OUTPUTFILE})
+      execute_process(COMMAND ${OUTPUTFILE}
+                      RESULT_VARIABLE CUDA_RETURN_CODE
+                  OUTPUT_VARIABLE ARCH)
+
+    if(${CUDA_RETURN_CODE} EQUAL 0)
+      set(CUDA_SUCCESS "TRUE")
+    else()
+      set(CUDA_SUCCESS "FALSE")
+    endif()
+
+  if (${CUDA_SUCCESS})
+    message(STATUS "CUDA Architecture: ${ARCH}")
+    set(CMAKE_CUDA_FLAGS "${ARCH} -G -O0")
+    # ARL: These variables are set by the find_package(CUDA) call, which we
+    # currently do not use
+    #message(STATUS "CUDA Version: ${CUDA_VERSION_STRING}")
+    #message(STATUS "CUDA Path: ${CUDA_TOOLKIT_ROOT_DIR}")
+    #message(STATUS "CUDA Libararies: ${CUDA_LIBRARIES}")
+    #message(STATUS "CUDA Performance Primitives: ${CUDA_npp_LIBRARY}")
+  else()
+    message(WARNING ${ARCH})
+  endif()
+
     # message("
     # CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES = ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
     # CMAKE_CUDA_HOST_COMPILER       = ${CMAKE_CUDA_HOST_COMPILER}