调试时共享内存问题
我正在尝试使用 Nsight 调试以下代码：
__device__ void change(int shared[])
{
if(threadIdx.x<10)
shared[threadIdx.x]=threadIdx.x;
}
__global__ void MyK()
{
int shared[10];
change(shared);
__syncthreads();
}
我在 main 方法中调用我的内核，如下所示：
cudaSetDevice(1);
MyK<<<1,20>>>();
当我在 change(shared) 之前放置断点时，我可以看到共享数组是创建并将其值设置为 0。但是，如果我将断点放在 __syncthreads() 之后，调试器会显示以下错误：
cannot resolve name shared
Can't I pass my share array to another device function?
I am trying to use Nsight to debug the following code:
__device__ void change(int shared[])
{
if(threadIdx.x<10)
shared[threadIdx.x]=threadIdx.x;
}
__global__ void MyK()
{
int shared[10];
change(shared);
__syncthreads();
}
I am calling my kernel in the main method like this :
cudaSetDevice(1);
MyK<<<1,20>>>();
When I put a breakpoint before change(shared), I can see that the shared array is created and its values are set to 0. However, if I put the breakpoint after __syncthreads(), the debugger shows the following error:
cannot resolve name shared
Can't I pass my shared array to another device function?
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（3）
您在内存监视窗口中看到“无法解析共享名称”的原因是因为共享数组正在被编译器优化，因为在
change(shared) 之后内核的任何部分都不会使用它。就像前面提到的 @user586831 一样，尝试输出该值作为设备函数的返回值。另外，不确定您是否真正指的是
__shared__数组，还是通过名称shared引用该数组。无论如何，您在上面的代码中没有使用共享内存。int shared只是一个普通的整数数组类型。您需要指定 __shared__ 限定符才能声明共享内存。例如The reason why you see the "Cannot resolve name shared" in the memory watch window is because shared array is being optimized out by the compiler since it is not being used at all by any part of your kernel after
change(shared). Like @user586831 mentioned earlier, try outputing the value as your return value for your device function.Also on another note, not sure if you really meant
__shared__array or referring to the array by its nameshared. Anyway you're not using shared memory in your code above.int sharedis just a normal integer array type. You need to specify the__shared__qualifier in order to declare shared memory. E.g.对某些线程（并非所有线程）调用 __syncthreads() 可能会导致死锁。 <代码>threadIdx.x < 10 调用
_syncthreads()如前所述，您在这里没有使用共享内存。
编译器很聪明，如果您之后不使用该值，则内存位置可能会变得无效。
尝试输出该值作为设备函数的返回值。应该可以正常工作，尤其是在移动/删除
__syncthreads()时。Calling
__syncthreads()for some and not all threads can cause a deadlock.threadIdx.x < 10calls_syncthreads()As previously mentioned you are not using shared memory here.
The compiler is clever if you are not using the value afterwards the memory location can become invalid.
Try outputing the value as your return value for your device function. Should work fine especially if you move/remove
__syncthreads().这是实际的代码还是您从缓冲区声明中省略了
__shared__？另请记住，
__device__函数由编译器内联，并且调试器只能在整个过程中的某个时刻停止。尝试使用至少 16 或 32 个线程的倍数的内核，否则您将无法运行
完整的 SP，这可能会欺骗调试器。
Is that the actual code or you omitted
__shared__from the buffer declaration?Keep also in mind that the
__device__functions get inlined by the compiler and that the debugger can stop only at some point in the whole process.Try to use a kernel of a multiple of at least 16 or 32 threads or otherwise you are not running
a full SP and that might trick the debugger.