CUDA: safeCall() Runtime API error invalid device symbol

Question

I'm trying to run the cudaSift in CudaSift project with a Nvidia Tesla M2090 on a remote machine with CUDA 7.5. The machine has 4 of these GPUs, but I'm quite sure from the debuggining that the initialization is done correctly, anyway this is the code:

void InitCuda(int devNum)
{
  int nDevices;
  cudaGetDeviceCount(&nDevices);
  if (!nDevices) {
    std::cerr << "No CUDA devices available" << std::endl;
    return;
  }
  devNum = std::min(nDevices-1, devNum);
  deviceInit(devNum);  
  cudaDeviceProp prop;
  cudaGetDeviceProperties(&prop, devNum);
  printf("Device Number: %d\n", devNum);
  printf("  Device name: %s\n", prop.name);
  printf("  Memory Clock Rate (MHz): %d\n", prop.memoryClockRate/1000);
  printf("  Memory Bus Width (bits): %d\n", prop.memoryBusWidth);
  printf("  Peak Memory Bandwidth (GB/s): %.1f\n\n",
     2.0*prop.memoryClockRate*(prop.memoryBusWidth/8)/1.0e6);
}

I use cmake to generate the makefile, and the make file generate cudaSift without any error.

Anyway, when I run it, the following error is returned:

safeCall() Runtime API error in file </ghome/rzhengac/Downloads/CudaSift-Maxwell/cudaSiftH.cu>, line 42 : invalid device symbol.

At line 42 of cudaSiftH.cu it is:

safeCall(cudaMemcpyToSymbol(d_PointCounter, &totPts, sizeof(int)));

Where:

__device__ unsigned int d_PointCounter[1];
int totPts = 0;

This is the code printed by InitCuda (which make quite sure that everything is fine during initialization):

Device Number: 0
  Device name: Tesla M2090
  Memory Clock Rate (MHz): 1848
  Memory Bus Width (bits): 384
  Peak Memory Bandwidth (GB/s): 177.4

SOLUTION:

As suggested in the comments, I was compiling with the wrong architecture: I had to change each sm_35 in CMakeList.txt into sm_20, and this is the result:

cmake_minimum_required(VERSION 2.6)

project(cudaSift)
set(cudaSift_VERSION_MAJOR 2)
set(cudaSift_VERSION_MINOR 0)
set(cudaSift_VERSION_PATCH 0)

set(CPACK_PACKAGE_VERSION_MAJOR "${cudaSift_VERSION_MAJOR}")
set(CPACK_PACKAGE_VERSION_MINOR "${cudaSift_VERSION_MINOR}")
set(CPACK_PACKAGE_VERSION_PATCH "${cudaSift_VERSION_PATCH}")
set(CPACK_GENERATOR "ZIP")
include(CPack)

find_package(OpenCV REQUIRED)
find_package(CUDA)
if (NOT CUDA_FOUND)
  message(STATUS "CUDA not found. Project will not be built.")
endif(NOT CUDA_FOUND)

if (WIN32)
  set(EXTRA_CXX_FLAGS "/DVERBOSE /D_CRT_SECURE_NO_WARNINGS ")
  list(APPEND CUDA_NVCC_FLAGS "-arch=sm_20;--compiler-options;-O2;-DVERBOSE") 
endif()
if (UNIX)
  if (APPLE)
    set(EXTRA_CXX_FLAGS "-DVERBOSE -msse2")
    list(APPEND CUDA_NVCC_FLAGS "-arch=sm_20;--compiler-options;-O2;-DVERBOSE") 
  else()
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O2 -msse2 ")
    list(APPEND CUDA_NVCC_FLAGS "-lineinfo;--compiler-options;-O2;-DVERBOSE") 
  endif()
endif()

set(cuda_sources
  # dynamic.cu
  cudaImage.cu  
  cudaImage.h  
  cudaSiftH.cu 
  cudaSiftH.h  
  matching.cu  
  cudaSiftD.h  
  cudaSift.h  
  cudautils.h
)  

set(sources
  geomFuncs.cpp  
  mainSift.cpp
)

include_directories(
  ${CMAKE_CURRENT_SOURCE_DIR} 
)

SET(CUDA_SEPARABLE_COMPILATION ON)

cuda_add_executable(cudasift ${cuda_sources} ${sources} OPTIONS -arch=sm_20)

set_target_properties(cudasift PROPERTIES
  COMPILE_FLAGS "${EXTRA_CXX_FLAGS}"               
)

target_link_libraries(cudasift
  /usr/local/cuda/lib64/libcudadevrt.a ${OpenCV_LIBS} 
)

install(FILES 
  ${cuda_sources} 
  ${sources}
  cudaSiftD.cu
  CMakeLists.txt
  Copyright.txt
  DESTINATION .
)
install(FILES data/left.pgm data/righ.pgm
  DESTINATION data
)