如何运行 cuda cooperative template kernel

How to run cuda cooperative template kernel

我试图在 CUDA C++ 中将模板内核作为协作内核启动但未成功,我做错了什么

错误


Error       cannot determine which instance of function template "boolPrepareKernel" is intended    
 

我尝试像下面这样调用内核

 ForBoolKernelArgs<int> fbArgs = ...;

    int device = 0;
    cudaDeviceProp deviceProp;
    cudaGetDeviceProperties(&deviceProp, device);
   cudaLaunchCooperativeKernel((void*)boolPrepareKernel, deviceProp.multiProcessorCount, fFArgs.threads, fbArgs) ;

内核定义如

template <typename TYO>
__global__ void boolPrepareKernel(ForBoolKernelArgs<TYO> fbArgs) {
...
}

我试过参数化启动(在本例中使用 int),例如

    cudaLaunchCooperativeKernel((void*)(<int>boolPrepareKernel), deviceProp.multiProcessorCount, fFArgs.threads, fbArgs) ;

但我收到错误

no instance of overloaded function matches the argument list            argument types are: (<error-type>, int, dim3, ForBoolKernelArgs<int>)

对于建议的案例

cudaLaunchCooperativeKernel((void*)(boolPrepareKernel<int>), deviceProp.multiProcessorCount, fFArgs.threads, fbArgs)

我的错误是

 no instance of overloaded function matches the argument list            argument types are: (void *, int, dim3, ForBoolKernelArgs<int>)

这可能很简单,但我卡住了 - 感谢您的帮助!!

内核启动参考

boolPrepareKernel << <fFArgs.blocks, fFArgs.threads >> > (fbArgs);

有效,但网格同步当然不可用。

这是一个可以编译的最小示例:

$ cat t1954.cu
template <typename TYO>
struct ForBoolKernelArgs
{
    TYO val;
};

template <typename TYO>
__global__ void boolPrepareKernel(ForBoolKernelArgs<TYO> fbArgs) {
}


int main(){
  ForBoolKernelArgs<int> fbArgs;
  void *kernel_args[] = {&fbArgs};
  cudaLaunchCooperativeKernel((void*)(boolPrepareKernel<int>), 1, 1, kernel_args) ;
}
$ nvcc -o t1954 t1954.cu
$

可能您剩下的主要问题是您没有遵循正确的 instructions 来传递内核参数。