Easiest way to test for existence of cuda-capable GPU from cmake?

Question

We have some nightly build machines that have the cuda libraries installed, but which do not have a cuda-capable GPU installed. These machines are capable of building cuda-

Answer 1

I just wrote a pure Python script that does some of the things you seem to need (I took much of this from the pystream project). It's basically just a wrapper for some functions in the CUDA run time library (it uses ctypes). Look at the main() function to see example usage. Also, be aware that I just wrote it, so it's likely to contain bugs. Use with caution.

#!/bin/bash

import sys
import platform
import ctypes

"""
cudart.py: used to access pars of the CUDA runtime library.
Most of this code was lifted from the pystream project (it's BSD licensed):
http://code.google.com/p/pystream

Note that this is likely to only work with CUDA 2.3
To extend to other versions, you may need to edit the DeviceProp Class
"""

cudaSuccess = 0
errorDict = {
    1: 'MissingConfigurationError',
    2: 'MemoryAllocationError',
    3: 'InitializationError',
    4: 'LaunchFailureError',
    5: 'PriorLaunchFailureError',
    6: 'LaunchTimeoutError',
    7: 'LaunchOutOfResourcesError',
    8: 'InvalidDeviceFunctionError',
    9: 'InvalidConfigurationError',
    10: 'InvalidDeviceError',
    11: 'InvalidValueError',
    12: 'InvalidPitchValueError',
    13: 'InvalidSymbolError',
    14: 'MapBufferObjectFailedError',
    15: 'UnmapBufferObjectFailedError',
    16: 'InvalidHostPointerError',
    17: 'InvalidDevicePointerError',
    18: 'InvalidTextureError',
    19: 'InvalidTextureBindingError',
    20: 'InvalidChannelDescriptorError',
    21: 'InvalidMemcpyDirectionError',
    22: 'AddressOfConstantError',
    23: 'TextureFetchFailedError',
    24: 'TextureNotBoundError',
    25: 'SynchronizationError',
    26: 'InvalidFilterSettingError',
    27: 'InvalidNormSettingError',
    28: 'MixedDeviceExecutionError',
    29: 'CudartUnloadingError',
    30: 'UnknownError',
    31: 'NotYetImplementedError',
    32: 'MemoryValueTooLargeError',
    33: 'InvalidResourceHandleError',
    34: 'NotReadyError',
    0x7f: 'StartupFailureError',
    10000: 'ApiFailureBaseError'}


try:
    if platform.system() == "Microsoft":
        _libcudart = ctypes.windll.LoadLibrary('cudart.dll')
    elif platform.system()=="Darwin":
        _libcudart = ctypes.cdll.LoadLibrary('libcudart.dylib')
    else:
        _libcudart = ctypes.cdll.LoadLibrary('libcudart.so')
    _libcudart_error = None
except OSError, e:
    _libcudart_error = e
    _libcudart = None

def _checkCudaStatus(status):
    if status != cudaSuccess:
        eClassString = errorDict[status]
        # Get the class by name from the top level of this module
        eClass = globals()[eClassString]
        raise eClass()

def _checkDeviceNumber(device):
    assert isinstance(device, int), "device number must be an int"
    assert device >= 0, "device number must be greater than 0"
    assert device < 2**8-1, "device number must be < 255"


# cudaDeviceProp
class DeviceProp(ctypes.Structure):
    _fields_ = [
         ("name", 256*ctypes.c_char), #  < ASCII string identifying device
         ("totalGlobalMem", ctypes.c_size_t), #  < Global memory available on device in bytes
         ("sharedMemPerBlock", ctypes.c_size_t), #  < Shared memory available per block in bytes
         ("regsPerBlock", ctypes.c_int), #  < 32-bit registers available per block
         ("warpSize", ctypes.c_int), #  < Warp size in threads
         ("memPitch", ctypes.c_size_t), #  < Maximum pitch in bytes allowed by memory copies
         ("maxThreadsPerBlock", ctypes.c_int), #  < Maximum number of threads per block
         ("maxThreadsDim", 3*ctypes.c_int), #  < Maximum size of each dimension of a block
         ("maxGridSize", 3*ctypes.c_int), #  < Maximum size of each dimension of a grid
         ("clockRate", ctypes.c_int), #  < Clock frequency in kilohertz
         ("totalConstMem", ctypes.c_size_t), #  < Constant memory available on device in bytes
         ("major", ctypes.c_int), #  < Major compute capability
         ("minor", ctypes.c_int), #  < Minor compute capability
         ("textureAlignment", ctypes.c_size_t), #  < Alignment requirement for textures
         ("deviceOverlap", ctypes.c_int), #  < Device can concurrently copy memory and execute a kernel
         ("multiProcessorCount", ctypes.c_int), #  < Number of multiprocessors on device
         ("kernelExecTimeoutEnabled", ctypes.c_int), #  < Specified whether there is a run time limit on kernels
         ("integrated", ctypes.c_int), #  < Device is integrated as opposed to discrete
         ("canMapHostMemory", ctypes.c_int), #  < Device can map host memory with cudaHostAlloc/cudaHostGetDevicePointer
         ("computeMode", ctypes.c_int), #  < Compute mode (See ::cudaComputeMode)
         ("__cudaReserved", 36*ctypes.c_int),
]

    def __str__(self):
        return """NVidia GPU Specifications:
    Name: %s
    Total global mem: %i
    Shared mem per block: %i
    Registers per block: %i
    Warp size: %i
    Mem pitch: %i
    Max threads per block: %i
    Max treads dim: (%i, %i, %i)
    Max grid size: (%i, %i, %i)
    Total const mem: %i
    Compute capability: %i.%i
    Clock Rate (GHz): %f
    Texture alignment: %i
""" % (self.name, self.totalGlobalMem, self.sharedMemPerBlock,
       self.regsPerBlock, self.warpSize, self.memPitch,
       self.maxThreadsPerBlock,
       self.maxThreadsDim[0], self.maxThreadsDim[1], self.maxThreadsDim[2],
       self.maxGridSize[0], self.maxGridSize[1], self.maxGridSize[2],
       self.totalConstMem, self.major, self.minor,
       float(self.clockRate)/1.0e6, self.textureAlignment)

def cudaGetDeviceCount():
    if _libcudart is None: return  0
    deviceCount = ctypes.c_int()
    status = _libcudart.cudaGetDeviceCount(ctypes.byref(deviceCount))
    _checkCudaStatus(status)
    return deviceCount.value

def getDeviceProperties(device):
    if _libcudart is None: return  None
    _checkDeviceNumber(device)
    props = DeviceProp()
    status = _libcudart.cudaGetDeviceProperties(ctypes.byref(props), device)
    _checkCudaStatus(status)
    return props

def getDriverVersion():
    if _libcudart is None: return  None
    version = ctypes.c_int()
    _libcudart.cudaDriverGetVersion(ctypes.byref(version))
    v = "%d.%d" % (version.value//1000,
                   version.value%100)
    return v

def getRuntimeVersion():
    if _libcudart is None: return  None
    version = ctypes.c_int()
    _libcudart.cudaRuntimeGetVersion(ctypes.byref(version))
    v = "%d.%d" % (version.value//1000,
                   version.value%100)
    return v

def getGpuCount():
    count=0
    for ii in range(cudaGetDeviceCount()):
        props = getDeviceProperties(ii)
        if props.major!=9999: count+=1
    return count

def getLoadError():
    return _libcudart_error


version = getDriverVersion()
if version is not None and not version.startswith('2.3'):
    sys.stdout.write("WARNING: Driver version %s may not work with %s\n" %
                     (version, sys.argv[0]))

version = getRuntimeVersion()
if version is not None and not version.startswith('2.3'):
    sys.stdout.write("WARNING: Runtime version %s may not work with %s\n" %
                     (version, sys.argv[0]))


def main():

    sys.stdout.write("Driver version: %s\n" % getDriverVersion())
    sys.stdout.write("Runtime version: %s\n" % getRuntimeVersion())

    nn = cudaGetDeviceCount()
    sys.stdout.write("Device count: %s\n" % nn)

    for ii in range(nn):
        props = getDeviceProperties(ii)
        sys.stdout.write("\nDevice %d:\n" % ii)
        #sys.stdout.write("%s" % props)
        for f_name, f_type in props._fields_:
            attr = props.__getattribute__(f_name)
            sys.stdout.write( "  %s: %s\n" % (f_name, attr))

    gpuCount = getGpuCount()
    if gpuCount > 0:
        sys.stdout.write("\n")
    sys.stdout.write("GPU count: %d\n" % getGpuCount())
    e = getLoadError()
    if e is not None:
        sys.stdout.write("There was an error loading a library:\n%s\n\n" % e)

if __name__=="__main__":
    main()

Answer 2

The answer to this question consists of two parts:

1. A program to detect the presence of a cuda-capable GPU.
2. CMake code to compile, run, and interpret the result of that program at configuration time.

For part 1, the gpu sniffing program, I started with the answer provided by fabrizioM because it is so compact. I quickly discovered that I needed many of the details found in unknown's answer to get it to work well. What I ended up with is the following C source file, which I named has_cuda_gpu.c:

#include <stdio.h>
#include <cuda_runtime.h>

int main() {
    int deviceCount, device;
    int gpuDeviceCount = 0;
    struct cudaDeviceProp properties;
    cudaError_t cudaResultCode = cudaGetDeviceCount(&deviceCount);
    if (cudaResultCode != cudaSuccess) 
        deviceCount = 0;
    /* machines with no GPUs can still report one emulation device */
    for (device = 0; device < deviceCount; ++device) {
        cudaGetDeviceProperties(&properties, device);
        if (properties.major != 9999) /* 9999 means emulation only */
            ++gpuDeviceCount;
    }
    printf("%d GPU CUDA device(s) found\n", gpuDeviceCount);

    /* don't just return the number of gpus, because other runtime cuda
       errors can also yield non-zero return values */
    if (gpuDeviceCount > 0)
        return 0; /* success */
    else
        return 1; /* failure */
}

Notice that the return code is zero in the case where a cuda-enabled GPU is found. This is because on one of my has-cuda-but-no-GPU machines, this program generates a runtime error with non-zero exit code. So any non-zero exit code is interpreted as "cuda does not work on this machine".

You might ask why I don't use cuda emulation mode on non-GPU machines. It is because emulation mode is buggy. I only want to debug my code, and work around bugs in cuda GPU code. I don't have time to debug the emulator.

The second part of the problem is the cmake code to use this test program. After some struggle, I have figured it out. The following block is part of a larger CMakeLists.txt file:

find_package(CUDA)
if(CUDA_FOUND)
    try_run(RUN_RESULT_VAR COMPILE_RESULT_VAR
        ${CMAKE_BINARY_DIR} 
        ${CMAKE_CURRENT_SOURCE_DIR}/has_cuda_gpu.c
        CMAKE_FLAGS 
            -DINCLUDE_DIRECTORIES:STRING=${CUDA_TOOLKIT_INCLUDE}
            -DLINK_LIBRARIES:STRING=${CUDA_CUDART_LIBRARY}
        COMPILE_OUTPUT_VARIABLE COMPILE_OUTPUT_VAR
        RUN_OUTPUT_VARIABLE RUN_OUTPUT_VAR)
    message("${RUN_OUTPUT_VAR}") # Display number of GPUs found
    # COMPILE_RESULT_VAR is TRUE when compile succeeds
    # RUN_RESULT_VAR is zero when a GPU is found
    if(COMPILE_RESULT_VAR AND NOT RUN_RESULT_VAR)
        set(CUDA_HAVE_GPU TRUE CACHE BOOL "Whether CUDA-capable GPU is present")
    else()
        set(CUDA_HAVE_GPU FALSE CACHE BOOL "Whether CUDA-capable GPU is present")
    endif()
endif(CUDA_FOUND)

This sets a CUDA_HAVE_GPU boolean variable in cmake that can subsequently be used to trigger conditional operations.

It took me a long time to figure out that the include and link parameters need to go in the CMAKE_FLAGS stanza, and what the syntax should be. The try_run documentation is very light, but there is more information in the try_compile documentation, which is a closely related command. I still needed to scour the web for examples of try_compile and try_run before getting this to work.

Another tricky but important detail is the third argument to try_run, the "bindir". You should probably always set this to ${CMAKE_BINARY_DIR}. In particular, do not set it to ${CMAKE_CURRENT_BINARY_DIR} if you are in a subdirectory of your project. CMake expects to find the subdirectory CMakeFiles/CMakeTmp within bindir, and spews errors if that directory does not exist. Just use ${CMAKE_BINARY_DIR}, which is one location where those subdirectories seem to naturally reside.

Answer 3

Write a simple program like

#include<cuda.h>

int main (){
    int deviceCount;
    cudaError_t e = cudaGetDeviceCount(&deviceCount);
    return e == cudaSuccess ? deviceCount : -1;
}

and check the return value.

Answer 4

You can compile small GPU query program if cuda was found. here is a simple one you can adopt the needs:

#include <stdlib.h>
#include <stdio.h>
#include <cuda.h>
#include <cuda_runtime.h>

int main(int argc, char** argv) {
  int ct,dev;
  cudaError_t code;
  struct cudaDeviceProp prop;

 cudaGetDeviceCount(&ct);
 code = cudaGetLastError();
 if(code)  printf("%s\n", cudaGetErrorString(code));


if(ct == 0) {
   printf("Cuda device not found.\n");
   exit(0);
}
 printf("Found %i Cuda device(s).\n",ct);

for (dev = 0; dev < ct; ++dev) {
printf("Cuda device %i\n", dev);

cudaGetDeviceProperties(&prop,dev);
printf("\tname : %s\n", prop.name);
 printf("\ttotalGlobablMem: %lu\n", (unsigned long)prop.totalGlobalMem);
printf("\tsharedMemPerBlock: %i\n", prop.sharedMemPerBlock);
printf("\tregsPerBlock: %i\n", prop.regsPerBlock);
printf("\twarpSize: %i\n", prop.warpSize);
printf("\tmemPitch: %i\n", prop.memPitch);
printf("\tmaxThreadsPerBlock: %i\n", prop.maxThreadsPerBlock);
printf("\tmaxThreadsDim: %i, %i, %i\n", prop.maxThreadsDim[0], prop.maxThreadsDim[1], prop.maxThreadsDim[2]);
printf("\tmaxGridSize: %i, %i, %i\n", prop.maxGridSize[0], prop.maxGridSize[1], prop.maxGridSize[2]);
printf("\tclockRate: %i\n", prop.clockRate);
printf("\ttotalConstMem: %i\n", prop.totalConstMem);
printf("\tmajor: %i\n", prop.major);
printf("\tminor: %i\n", prop.minor);
printf("\ttextureAlignment: %i\n", prop.textureAlignment);
printf("\tdeviceOverlap: %i\n", prop.deviceOverlap);
printf("\tmultiProcessorCount: %i\n", prop.multiProcessorCount);
}
}

Answer 5

One useful approach is to run programs that CUDA has installed, such as nvidia-smi, to see what they return.

        find_program(_nvidia_smi "nvidia-smi")
        if (_nvidia_smi)
            set(DETECT_GPU_COUNT_NVIDIA_SMI 0)
            # execute nvidia-smi -L to get a short list of GPUs available
            exec_program(${_nvidia_smi_path} ARGS -L
                OUTPUT_VARIABLE _nvidia_smi_out
                RETURN_VALUE    _nvidia_smi_ret)
            # process the stdout of nvidia-smi
            if (_nvidia_smi_ret EQUAL 0)
                # convert string with newlines to list of strings
                string(REGEX REPLACE "\n" ";" _nvidia_smi_out "${_nvidia_smi_out}")
                foreach(_line ${_nvidia_smi_out})
                    if (_line MATCHES "^GPU [0-9]+:")
                        math(EXPR DETECT_GPU_COUNT_NVIDIA_SMI "${DETECT_GPU_COUNT_NVIDIA_SMI}+1")
                        # the UUID is not very useful for the user, remove it
                        string(REGEX REPLACE " \\(UUID:.*\\)" "" _gpu_info "${_line}")
                        if (NOT _gpu_info STREQUAL "")
                            list(APPEND DETECT_GPU_INFO "${_gpu_info}")
                        endif()
                    endif()
                endforeach()

                check_num_gpu_info(${DETECT_GPU_COUNT_NVIDIA_SMI} DETECT_GPU_INFO)
                set(DETECT_GPU_COUNT ${DETECT_GPU_COUNT_NVIDIA_SMI})
            endif()
        endif()

One might also query linux /proc or lspci. See fully-worked CMake example at https://github.com/gromacs/gromacs/blob/master/cmake/gmxDetectGpu.cmake