CUDA 5.5 cudaMemcpyToSymbol、__constant__ 和超出范围错误

2022-01-10 00:00:00 cuda c++

我正在尝试编译一个具有的 CUDA 示例；

I'm trying to compile a CUDA example which has;

cuda.cu:

__constant__ unsigned VERTICES;
__constant__ unsigned TRIANGLES;

以及main.cpp中对应的代码；

and the corresponding code in main.cpp;

cudaMemcpyToSymbol(VERTICES, &verticesNo, sizeof(int));
cudaMemcpyToSymbol(TRIANGLES, &trianglesNo, sizeof(int));

编译main.cpp时如何避免VERTICES not declared in this scope"错误?

How to avoid "VERTICES not declared in this scope" error when compiling the main.cpp?

TIA.

干杯，

推荐答案

CUDA __constant__ 变量具有文件范围链接.这意味着 cudaMemcpyToSymbol 必须位于定义 __constant__ 变量的同一个 .cu 文件中.

CUDA __constant__ variables have a file scope linkage. That means that the cudaMemcpyToSymbol have to be in the same .cu file where the __constant__ variable is defined.

您可以将包装函数添加到 .cu 文件并从 .cpp 文件中调用它.

You can add a wrapper function to the .cu file and call this one from your .cpp file.

cuda.cu 的示例:

sample for cuda.cu:

__constant__ unsigned VERTICES;
__constant__ unsigned TRIANGLES;

void wrapper_fn(unsigned *verticesNo, unsigned *trianglesNo)
{
  cudaMemcpyToSymbol(VERTICES, verticesNo, sizeof(unsigned));
  cudaMemcpyToSymbol(TRIANGLES, trianglesNo, sizeof(unsigned));
}

然后只在你的 main.cpp 中调用 wrapper_fn.

Then only call wrapper_fn in your main.cpp.

相关文章