Question

我试图在不添加pycuda依赖项的情况下查询CUDA设备。这是我到目前为止所得到的：

import ctypes

cudart = ctypes.cdll.LoadLibrary('libcudart.so')

numDevices = ctypes.c_int()
cudart.cudaGetDeviceCount(ctypes.byref(numDevices))
print 'There are', numDevices.value, 'devices.'

for x in xrange(numDevices.value):
    properties = None # XXX What goes here?
    cudart.cudaGetDeviceProperties(ctypes.byref(properties), x)
    print properties

问题在于我无法创建一个空结构来传递给cudaGetDeviceProperties（）。我想做这样的事情：

properties = cudart.cudaDeviceProp

但是这引发了这个错误：

AttributeError: /usr/local/cuda/lib64/libcudart.so: undefined symbol: cudaDeviceProp

以下是相关的CUDA documentation。

（编辑）

感谢@mhawke，我得到了这个工作。对于任何想要这样做的人，我将为您节省自己打字课程的工作：

class CudaDeviceProp(ctypes.Structure):
    _fields_ = [ 
            ('name', ctypes.c_char * 256),
            ('totalGlobalMem', ctypes.c_size_t),
            ('sharedMemPerBlock', ctypes.c_size_t),
            ('regsPerBlock', ctypes.c_int),
            ('warpSize', ctypes.c_int),
            ('memPitch', ctypes.c_size_t),
            ('maxThreadsPerBlock', ctypes.c_int),
            ('maxThreadsDim', ctypes.c_int * 3), 
            ('maxGridSize', ctypes.c_int * 3), 
            ('clockRate', ctypes.c_int),
            ('totalConstMem', ctypes.c_size_t),
            ('major', ctypes.c_int),
            ('minor', ctypes.c_int),
            ('textureAlignment', ctypes.c_size_t),
            ('texturePitchAlignment', ctypes.c_size_t),
            ('deviceOverlap', ctypes.c_int),
            ('multiProcessorCount', ctypes.c_int),
            ('kernelExecTimeoutEnabled', ctypes.c_int),
            ('integrated', ctypes.c_int),
            ('canMapHostMemory', ctypes.c_int),
            ('computeMode', ctypes.c_int),
            ('maxTexture1D', ctypes.c_int),
            ('maxTexture1DMipmap', ctypes.c_int),
            ('maxTexture1DLinear', ctypes.c_int),
            ('maxTexture2D', ctypes.c_int * 2), 
            ('maxTexture2DMipmap', ctypes.c_int * 2), 
            ('maxTexture2DLinear', ctypes.c_int * 3), 
            ('maxTexture2DGather', ctypes.c_int * 2), 
            ('maxTexture3D', ctypes.c_int * 3), 
            ('maxTexture3DAlt', ctypes.c_int * 3), 
            ('maxTextureCubemap', ctypes.c_int),
            ('maxTexture1DLayered', ctypes.c_int * 2), 
            ('maxTexture2DLayered', ctypes.c_int * 3), 
            ('maxTextureCubemapLayered', ctypes.c_int * 2), 
            ('maxSurface1D', ctypes.c_int),
            ('maxSurface2D', ctypes.c_int * 2), 
            ('maxSurface3D', ctypes.c_int * 3), 
            ('maxSurface1DLayered', ctypes.c_int * 2), 
            ('maxSurface2DLayered', ctypes.c_int * 3), 
            ('maxSurfaceCubemap', ctypes.c_int),
            ('maxSurfaceCubemapLayered', ctypes.c_int * 2), 
            ('surfaceAlignment', ctypes.c_size_t),
            ('concurrentKernels', ctypes.c_int),
            ('ECCEnabled', ctypes.c_int),
            ('pciBusID', ctypes.c_int),
            ('pciDeviceID', ctypes.c_int),
            ('pciDomainID', ctypes.c_int),
            ('tccDriver', ctypes.c_int),
            ('asyncEngineCount', ctypes.c_int),
            ('unifiedAddressing', ctypes.c_int),
            ('memoryClockRate', ctypes.c_int),
            ('memoryBusWidth', ctypes.c_int),
            ('l2CacheSize', ctypes.c_int),
            ('maxThreadsPerMultiProcessor', ctypes.c_int),
            ('streamPrioritiesSupported', ctypes.c_int),
            ('globalL1CacheSupported', ctypes.c_int),
            ('localL1CacheSupported', ctypes.c_int),
            ('sharedMemPerMultiprocessor', ctypes.c_size_t),
            ('regsPerMultiprocessor', ctypes.c_int),
            ('managedMemSupported', ctypes.c_int),
            ('isMultiGpuBoard', ctypes.c_int),
            ('multiGpuBoardGroupID', ctypes.c_int),
            ('singleToDoublePrecisionPerfRatio', ctypes.c_int),
            ('pageableMemoryAccess', ctypes.c_int),
            ('concurrentManagedAccess', ctypes.c_int),
            ]

Answer 1

您需要定义ctypes.Structure的子类，该子类指定cudaDeviceProp结构中的所有字段。然后，您可以将结构的实例传递给函数。请注意，您需要按正确的顺序填写所有字段。其中一些是数组，所以你需要正确地声明它们。

import ctypes

class CudaDeviceProp(ctypes.Structure):
    _fields_ = [('ECCEnabled', ctypes.c_int),
                ('asyncEngineCount', ctypes.c_int),
                ('canMapHostMemory', ctypes.c_int),
                ('clockRate', ctypes.c_int),
                ('computeMode', ctypes.c_int),
                ('concurrentKernels', ctypes.c_int),
                ...
                ('totalGlobalMem', ctypes.c_size_t),
                ('unifiedAddressing', ctypes.c_int),
                ('warpSize', ctypes.c_int)]

properties = CudaDeviceProp()
cudart.cudaGetDeviceProperties(ctypes.byref(properties), 0)

使用ctypes将c结构传递给函数

1 个答案: