Cleaning up cmake for CUDA. Fixing the executable extension of cuda p…

…rograms (now '.x')

CMakeLists.txt

@@ -48,18 +48,14 @@ if(CMAKE_UARCH STREQUAL "ORIN")
   message("Configuring nvcc compiler")
   option(ENABLE_CUDA "Enable building CUDA components" ON)
   enable_language(CUDA)
-  set(CMAKE_VERBOSE_MAKEFILE on)
+  #set(CMAKE_VERBOSE_MAKEFILE on)
   SET(CMAKE_CUDA_COMPILER /usr/local/cuda/bin/nvcc)
-  #SET(CMAKE_CXX_COMPILER /usr/local/cuda/bin/nvcc)
-  #SET(CMAKE_C_COMPILER /usr/local/cuda/bin/nvcc)
   SET(CMAKE_CUDA_ARCHITECTURES 87)
-  #set(CMAKE_CXX_STANDARD 17)
   set(CMAKE_CUDA_STANDARD 17)
-  #SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++17")
   SET(CMAKE_CUDA_FLAGS "${CMAKE_CXX_FLAGS} -arch=sm_87 -std=c++17")
-  #set_target_properties(<target_name>  PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
+
   # https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/
-  # -arch=sm_30  for Kepler
+
 else()
   message("Configuring gnu compilers")
   set(CMAKE_VERBOSE_MAKEFILE off)

cuda_demo/CMakeLists.txt

@@ -21,11 +21,17 @@ message("In cuda_demo/CMakeLists.txt: ${SOURCES}")
 
 foreach( source_file ${SOURCES} )
     get_filename_component(justname ${source_file} NAME)
-    string( REPLACE ".cpp" ".exe" exename ${justname} )
+    string( REPLACE ".cu" ".x" exename ${justname} )
     set(exename "${exename}")
     message("Adding: ${exename}")
     add_executable(${exename} ${source_file})
     target_link_libraries(${exename} /usr/local/cuda/lib64/libcudart.so)
+
+    # The following is only necessary if cuda code is making calls to other
+    # kernel functions defined and compiled in different .cu files
+    # It should be unnecessary with the current organization
+
+    #set_target_properties(<target_name>  PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
 endforeach( source_file ${SOURCES} )
 
 message("Leaving cuda_demo/CMakeLists.txt: ${SOURCES}")

cuda_demo/cuda_relu.cu

@@ -44,10 +44,16 @@ int main()
     small::ReLUActivation(num_channels, image_height, image_width,
                           in, out);
 
+    std::cout << "Checking correctness:\n";
     bool correct = true;
     for (size_t i = 0; i != num_elts; ++i)
     {
         correct &= (fabs(out[i] - check[i]) < 1e-15);
+        if (i < 20)
+        {
+            std::cout << i << ": ReLU(" << in[i] << ") = "
+                      << out[i] << std::endl;
+        }
     }
 
     std::cout << (correct ? "PASSED" : "FAILED") << std::endl;