CUDA写入其他扭曲看不到的全局内存

Question

I was trying to explain global memory to someone who is new to CUDA. I came up with the following dummy kernel that blocks other threads in other warps until a selected warp sets the global variable to another value:

__global__ void with_sync()
{
    while (threadIdx.x / 32 != 0)
    {
        if (is_done != 0)
        {
            break;
        }
    }

    if (threadIdx.x / 32 == 0)
    {
        is_done = 1;
        printf("I'm done!\n");
    }
}

The variable is_done is declared outside of the function as a __device__ __managed__ int (which, correct me if I'm wrong, means that the the variable will reside in global memory space.

However, when I execute this kernel (1024 1D threads in a single block) like so:

with_sync<<<1, 1024>>>();
cudaDeviceSynchronize();

I'm done is printed out as expected. However, the CUDA program does not terminate (I placed cudaDeviceSynchronize() in the host code for it to wait for all threads). This leads me to wonder if the other warps did not receive the change in the is_done variable. However, I understand that global memory implies that the value can be seen in device level (ie at the very least, all blocks in a grid).

My question is the following: Is there any caching/optimisation done by CUDA that makes it such that this inconsistent global memory view can occur? Is there a way to access the "latest" value from the variable that resides in global memory?

Answer 1

Is there any caching/optimisation done by CUDA that makes it such that this inconsistent global memory view can occur? Is there a way to access the "latest" value from the variable that resides in global memory?

Yes, there is caching behavior. You can modify it with the volatile qualifier .

Here is a worked example:

$ cat t310.cu
#include <stdio.h>

#ifndef USE_VOLATILE
__device__ __managed__ int is_done = 0;
#else
__device__ volatile __managed__ int is_done = 0;
#endif

__global__ void with_sync()
{
    while (threadIdx.x / 32 != 0)
    {
        if (is_done != 0)
        {
            break;
        }
    }

    if (threadIdx.x / 32 == 0)
    {
        is_done = 1;
        printf("I'm done!\n");
    }
}



int main(){

  with_sync<<<1,1024>>>();
  cudaDeviceSynchronize();
}
$ nvcc -o t310 t310.cu
$ ./t310
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
^C
$ nvcc -o t310 t310.cu -DUSE_VOLATILE
$ ./t310
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
I'm done!
$

(In case its unclear the first run above was terminated by Ctrl-C, due to hang)

Tesla P100 PCIE CUDA 10.0, CentOS 7