获取CUDA错误“声明与先前的”variable_name“不兼容。 [英] Getting CUDA error "declaration is incompatible with previous "variable_name"

问题描述

我试图编译一个程序，包括一个内核与MSVS 2012和CUDA。我使用共享内存，但不同于关于同一问题的此问题 ，我只使用我的变量名为这个内核的共享内存一次，所以没有重定义的问题。用这样的代码：

I'm trying to compile a program including a kernel with MSVS 2012 and CUDA. I use shared memory, but unlike in this question regarding the same problem, I only use my variable name for this kernel's shared memory once, so there's no issue of redefinition. With code like this:

template<typename T>
    __global__ void mykernel(
    const T* __restrict__ data,
    T* __restrict__ results) {
    extern __shared__ T warp_partial_results[];
    /* ... */
    warp_partial_results[lane_id] = something;
    /* ... */
    results[something_else] = warp_partial_results[something_else];
    /* ... */
}

类型（例如float，int，unsigned int），我得到可怕的

which is instantiated for several types (e.g. float, int, unsigned int), I get the dreaded

声明与以前的warp_partial_results不兼容

declaration is incompatible with previous "warp_partial_results"

消息。这会导致什么？

推荐答案

CUDA不立即支持模板化函数中的动态分配共享内存数组， （显然）生成那些extern的实际定义。如果你为多个类型实例化一个模板化的函数，这些定义会发生冲突。

CUDA doesn't immediately 'support' dynamically-allocated shared memory arrays in templated functions, as it (apparently) generates actual definitions of those extern's. If you instantiate a templated function for multiple types, the definitions would conflict.

一个解决方法是通过类的模板专用化形式。请参阅：

A workaround is available in the form of template specialization via classes. See:

http://www.ecse.rpi.edu/~wrf/wiki/ParallelComputingSpring2015/cuda/nvidia/samples/0_Simple/simpleTemplates/sharedmem.cuh

您使用以下解决方法：

template<class T> __global__ void foo( T* g_idata, T* g_odata)
{
    // shared memory
    // the size is determined by the host application

    SharedMem<T> shared;
    T* sdata = shared.getPointer();

    // .. the rest of the code remains unchanged!
}

getPointer（）具有 ^* 每个类型的专用实现，它返回一个不同的指针，例如 extern __shared__ float * shared_mem_float 或 extern __shared__ int * shared_mem_int 等。

the getPointer() has^* a specialized implementation for every type, which returns a different pointer, e.g. extern __shared__ float* shared_mem_float or extern __shared__ int* shared_mem_int etc.

_{（*） - 不是真的。在NVidia提供的头文件中，他们专门针对一些基本类型，就是这样。}

这篇关于获取CUDA错误“声明与先前的”variable_name“不兼容。的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！