如何实现涉及多个变量的自定义原子函数？

Question

我想在 CUDA 中实现这个原子函数：

__device__ float lowest;   // global var
__device__ int  lowIdx;    // global var
float realNum;   // thread reg var
int index;       // thread reg var

if(realNum < lowest) {
 lowest= realNum;  // the new lowest
 lowIdx= index;    // update the 'low' index
}

我不相信我可以用任何原子函数来做到这一点。我需要为几个指令锁定几个全局内存位置。 我可以用 PTXAS（汇编）代码来实现吗？

Answer 1

正如我在上面的第二条评论中所说，可以将您的两个 32 位数量组合成一个 64 位原子管理数量，并以这种方式处理问题。然后，我们使用arbitrary atomic example 作为粗略指南以原子方式管理 64 位数量。显然，您不能将此想法扩展到两个 32 位数量之外。这是一个例子：

#include <stdio.h>
#define DSIZE 5000
#define nTPB 256

#define cudaCheckErrors(msg) \
    do { \
        cudaError_t __err = cudaGetLastError(); \
        if (__err != cudaSuccess) { \
            fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
                msg, cudaGetErrorString(__err), \
                __FILE__, __LINE__); \
            fprintf(stderr, "*** FAILED - ABORTING\n"); \
            exit(1); \
        } \
    } while (0)

typedef union  {
  float floats[2];                 // floats[0] = lowest
  int ints[2];                     // ints[1] = lowIdx
  unsigned long long int ulong;    // for atomic update
} my_atomics;

__device__ my_atomics test;

__device__ unsigned long long int my_atomicMin(unsigned long long int* address, float val1, int val2)
{
    my_atomics loc, loctest;
    loc.floats[0] = val1;
    loc.ints[1] = val2;
    loctest.ulong = *address;
    while (loctest.floats[0] >  val1) 
      loctest.ulong = atomicCAS(address, loctest.ulong,  loc.ulong);
    return loctest.ulong;
}


__global__ void min_test(const float* data)
{

    int idx = (blockDim.x * blockIdx.x) + threadIdx.x;
    if (idx < DSIZE)
      my_atomicMin(&(test.ulong), data[idx],idx);
}

int main() {

  float *d_data, *h_data;
  my_atomics my_init;
  my_init.floats[0] = 10.0f;
  my_init.ints[1] = DSIZE;

  h_data = (float *)malloc(DSIZE * sizeof(float));
  if (h_data == 0) {printf("malloc fail\n"); return 1;}
  cudaMalloc((void **)&d_data, DSIZE * sizeof(float));
  cudaCheckErrors("cm1 fail");
  // create random floats between 0 and 1
  for (int i = 0; i < DSIZE; i++) h_data[i] = rand()/(float)RAND_MAX;
  cudaMemcpy(d_data, h_data, DSIZE*sizeof(float), cudaMemcpyHostToDevice);
  cudaCheckErrors("cmcp1 fail");
  cudaMemcpyToSymbol(test, &(my_init.ulong), sizeof(unsigned long long int));
  cudaCheckErrors("cmcp2 fail");
  min_test<<<(DSIZE+nTPB-1)/nTPB, nTPB>>>(d_data);
  cudaDeviceSynchronize();
  cudaCheckErrors("kernel fail");

  cudaMemcpyFromSymbol(&(my_init.ulong), test, sizeof(unsigned long long int));
  cudaCheckErrors("cmcp3 fail");

  printf("device min result = %f\n", my_init.floats[0]);
  printf("device idx result = %d\n", my_init.ints[1]);

  float host_val = 10.0f;
  int host_idx = DSIZE;
  for (int i=0; i<DSIZE; i++)
    if (h_data[i] < host_val){
      host_val = h_data[i];
      host_idx = i;
      }

  printf("host min result = %f\n", host_val);
  printf("host idx result = %d\n", host_idx);
  return 0;
}

Here 是一个类似的例子，它对 2 个float 数量进行原子更新。

Answer 2

@Robert Crovella：好主意，但我认为功能应该稍微修改如下：

__device__ unsigned long long int my_atomicMin(unsigned long long int* address, float val1, int val2)
{
    my_atomics loc, loctest, old;
    loc.floats[0] = val1;
    loc.ints[1] = val2;
    loctest.ulong = *address;
    old.ulong = loctest.ulong;
    while (loctest.floats[0] > val1){
        old.ulong = loctest.ulong;
        loctest.ulong = atomicCAS(address, loctest.ulong,  loc.ulong);
    }
    return old.ulong;
}