CUDA扭曲同步问题

Question

在概括将二维数组的值向右移动一个空间（环绕行边界）的内核时，我遇到了扭曲同步问题。完整的代码附在下面。

该代码适用于任意数组宽度、数组高度、线程块数和每个块的线程数。选择33的线程大小（即比完整扭曲更长的线程），第33个线程不会__syncthreads()调用。这会导致输出数据出现问题。该问题仅在存在多个经纱且阵列的宽度大于线程数（例如，宽度 = 35 和 34 线程）时出现。

下面是一个缩小的例子，说明发生了什么（实际上，数组需要有更多的元素才能让内核产生错误）。

初始数组：

0 1 2 3 4 
5 6 7 8 9

预期结果：

4 0 1 2 3
9 5 6 7 8

内核产生：

4 0 1 2 3
8 5 6 7 8

第一行正确完成（如果有多个块，则为每个块），所有后续行都具有重复的倒数第二个值。我已经测试了这两种不同的卡（8600GT 和 GTX280）并得到相同的结果。我想知道这是否只是我的内核的错误，还是无法通过调整我的代码来修复的问题？

完整的源文件包含在下面。

谢谢你。

#include <cstdio>
#include <cstdlib>

// A method to ensure all reads use the same logical layout.
inline __device__ __host__ int loc(int x, int y, int width)
{
  return y*width + x;
}

//kernel to shift all items in a 2D array one position to the right (wrapping around rows)
__global__ void shiftRight ( int* globalArray, int width, int height)
{
  int temp1=0;          //temporary swap variables
  int temp2=0;

  int blockRange=0;     //the number of rows that a single block will shift

  if (height%gridDim.x==0)  //logic to account for awkward array sizes
    blockRange = height/gridDim.x;
  else
    blockRange = (1+height/gridDim.x);

  int yStart = blockIdx.x*blockRange;
  int yEnd = yStart+blockRange; //the end condition for the y-loop
  yEnd = min(height,yEnd);              //make sure that the array doesn't go out of bounds

  for (int y = yStart; y < yEnd ; ++y)
  {
    //do the first read so the swap variables are loaded for the x-loop
    temp1 = globalArray[loc(threadIdx.x,y,width)];
    //Each block shifts an entire row by itself, even if there are more columns than threads
    for (int threadXOffset = threadIdx.x  ; threadXOffset < width ; threadXOffset+=blockDim.x)
    {
      //blockDim.x is added so that we store the next round of values
      //this has to be done now, because the next operation will
      //overwrite one of these values
      temp2 = globalArray[loc((threadXOffset + blockDim.x)%width,y,width)];
      __syncthreads();  //sync before the write to ensure all the values have been read
      globalArray[loc((threadXOffset +1)%width,y,width)] = temp1;
      __syncthreads();  //sync after the write so ensure all the values have been written
      temp1 = temp2;        //swap the storage variables.
    }
    if (threadIdx.x == 0 && y == 0)
      globalArray[loc(12,2,width)]=globalArray[67];
  }
}


int main (int argc, char* argv[])
{
  //set the parameters to be used
  int width = 34;
  int height = 3;
  int threadsPerBlock=33;
  int numBlocks = 1;

  int memSizeInBytes = width*height*sizeof(int);

  //create the host data and assign each element of the array to equal its index
  int* hostData = (int*) malloc (memSizeInBytes);
  for (int y = 0 ; y < height ; ++y)
    for (int x = 0 ; x < width ; ++x)
      hostData [loc(x,y,width)] = loc(x,y,width);

  //create an allocate the device pointers
  int* deviceData;
  cudaMalloc ( &deviceData  ,memSizeInBytes);
  cudaMemset (  deviceData,0,memSizeInBytes);
  cudaMemcpy (  deviceData, hostData, memSizeInBytes, cudaMemcpyHostToDevice);
  cudaThreadSynchronize();

  //launch the kernel
  shiftRight<<<numBlocks,threadsPerBlock>>> (deviceData, width, height);
  cudaThreadSynchronize();

  //copy the device data to a host array
  int* hostDeviceOutput = (int*) malloc (memSizeInBytes);
  cudaMemcpy (hostDeviceOutput, deviceData, memSizeInBytes, cudaMemcpyDeviceToHost); 
  cudaFree (deviceData);

  //Print out the expected/desired device output
  printf("---- Expected Device Output ----\n");
  printf("   | ");
  for (int x = 0 ; x < width ; ++x)
    printf("%4d ",x);
  printf("\n---|-");
  for (int x = 0 ; x < width ; ++x)
    printf("-----");
  for (int y = 0 ; y < height ; ++y)
  {
    printf("\n%2d | ",y);
    for (int x = 0 ; x < width ; ++x)
      printf("%4d ",hostData[loc((x-1+width)%width,y,width)]);
  }
  printf("\n\n");

  printf("---- Actual Device Output ----\n");
  printf("   | ");
  for (int x = 0 ; x < width ; ++x)
    printf("%4d ",x);
  printf("\n---|-");
  for (int x = 0 ; x < width ; ++x)
    printf("-----");
  for (int y = 0 ; y < height ; ++y)
  {
    printf("\n%2d | ",y);
    for (int x = 0 ; x < width ; ++x)
      printf("%4d ",hostDeviceOutput[loc(x,y,width)]);
  }
  printf("\n\n");
}

Answer 1

因为并非所有线程都执行相同数量的循环迭代，所以同步是一个问题！所有线程应该始终命中相同的 __syncthreads()-s。

我建议将最里面的 for 循环转换为如下所示：

for(int blockXOffset=0; blockXOffset < width; blockXOffset+=blockDim.x) {
  int threadXOffset=blockXOffset+threadIdx.x;
  bool isActive=(threadXOffset < width);
  if (isActive) temp2 = globalArray[loc((threadXOffset + blockDim.x)%width,y,width)];
  __syncthreads();
  if (isActive) globalArray[loc((threadXOffset +1)%width,y,width)] = temp1;
  __syncthreads();
  temp1 = temp2;
}