CUDA 9 shfl 与 shfl_sync

Question

从 CUDA 9 开始，不推荐使用 shfl 指令，应将其替换为 shfl_sync。

但是，当它们的行为不同时，我应该如何更换它们？

代码示例:

__global__
static void shflTest(){
    int tid = threadIdx.x;
    float value = tid + 0.1f;
    int* ivalue = reinterpret_cast<int*>(&value);

    //use the integer shfl
    int ix = __shfl(ivalue[0],5,32);
    int iy = __shfl_sync(ivalue[0],5,32);

    float x = reinterpret_cast<float*>(&ix)[0];
    float y = reinterpret_cast<float*>(&iy)[0];

    if(tid == 0){
        printf("shfl tmp %d %d\n",ix,iy);
        printf("shfl final %f %f\n",x,y);
    }
}

int main()
{
    shflTest<<<1,32>>>();
    cudaDeviceSynchronize();
    return 0;
}

输出:

shfl tmp 1084437299 5
shfl final 5.100000 0.000000

Answer 1

如果您阅读随 CUDA 9RC 副本安装的 CUDA 9RC 编程指南(第 B.15 节)，您将看到新的 __shfl_sync()函数有一个额外的 mask您没有考虑的参数:

CUDA 8:

int __shfl(int var, int srcLane, int width=warpSize);

CUDA 9:

T __shfl_sync(unsigned mask, T var, int srcLane, int width=warpSize);
              ^^^^^^^^^^^^^

此掩码参数的期望值也表示为:

The new *_sync shfl intrinsics take in a mask indicating the threads participating in the call. A bit, representing the thread's lane id, must be set for each participating thread to ensure they are properly converged before the intrinsic is executed by the hardware. All non-exited threads named in mask must execute the same intrinsic with the same mask, or the result is undefined.

因此，如果我们修改您的代码以符合这一点，我们会得到预期的结果:

$ cat t419.cu
#include <stdio.h>

__global__
static void shflTest(int lid){
    int tid = threadIdx.x;
    float value = tid + 0.1f;
    int* ivalue = reinterpret_cast<int*>(&value);

    //use the integer shfl
    int ix = __shfl(ivalue[0],5,32);
    int iy = __shfl_sync(0xFFFFFFFF, ivalue[0],5,32);

    float x = reinterpret_cast<float*>(&ix)[0];
    float y = reinterpret_cast<float*>(&iy)[0];

    if(tid == lid){
        printf("shfl tmp %d %d\n",ix,iy);
        printf("shfl final %f %f\n",x,y);
    }
}

int main()
{
    shflTest<<<1,32>>>(0);
    cudaDeviceSynchronize();
    return 0;
}
$ nvcc -arch=sm_61 -o t419 t419.cu
t419.cu(10): warning: function "__shfl(int, int, int)"
/usr/local/cuda/bin/..//include/sm_30_intrinsics.hpp(152): here was declared deprecated ("__shfl() is deprecated in favor of __shfl_sync() and may be removed in a future release (Use -Wno-deprecated-declarations to suppress this warning).")

$ cuda-memcheck ./t419
========= CUDA-MEMCHECK
shfl tmp 1084437299 1084437299
shfl final 5.100000 5.100000
========= ERROR SUMMARY: 0 errors
$