CUDA Thrust实现的跨步规约

Question

我有一个顶点数组，结构如下：

[x0，y0，z0，空的浮点数，x1，y1，z1，空的浮点数，x2，y2，z2，空的浮点数，...]

我需要使用CUDA找到minX，minY，minZ，maxX，maxY和maxZ。我编写了适当的约简算法，但速度有点慢。我决定使用THRUST库。有一个高度优化的reduce（）方法，甚至更好的是minmax_element（），这是一种同时查找数组的最大值和最小值的方法，但我找不到一种快速的方法来仅使用每个第4个索引。将数据复制到3个分离的数组中不是我要寻找的解决方案。

是否有一种方法（使用Thrust迭代器等技巧）向reduce（）传递步幅？

- aerion

---

也许这篇帖子cuda thrust library repeat vector multiple times可以提供一些提示。 - Vitality

可以使用步进范围、thrust minmax_element和3个thrust调用来生成6个结果，类似于这里。或者使用thrust::reduce_by_key和步进范围迭代器将其缩减为2个调用（一个产生最大结果集，另一个产生最小结果集），相关信息请参见这里。目前我无法想到一种方法来使用单个thrust调用获取所有6个结果。 - Robert Crovella

对于经过优化的代码来说，这个操作很可能受到内存带宽的限制。我认为推力实现可能不如编写良好的内核性能好（因为您不希望重新组织数据），后者可以每个线程加载一个float4结构元素（因此完全协同），并通过并行归约同时计算6个所需结果，通过同时运行6个归约。 - Robert Crovella

如果数组中的元素真的非常密集，那么您应该能够将指向第一个float的指针重新解释为指向float4的指针。将float4 *传递给thrust::reduce，并编写一个特殊的约简函数来计算所有六个值。 - Jared Hoberock

@JaredHoberock 如果缩减二元函数接受两个 float4 值，那么它不是必须生成一个 float4 结果吗？如何将6个 float 数量打包成一个 float4 结果？ - Robert Crovella

@RobertCrovella 这实际上是transform_reduce的一个应用。首先，将每个输入的float4转换为包含单个点的边界框。binary_op将以两个边界框作为输入，并将它们合并成一个包含它们的盒子作为输出。这个一般的想法在这个2D示例中被使用。 - Jared Hoberock

Answer 1

你可以使用步长范围方法和3次调用thrust::minmax_element，在不修改数据存储的情况下获得所需结果。

以下是一个实例：

$ cat t491.cu 
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#include <iostream>
#include <thrust/copy.h>
#include <thrust/iterator/permutation_iterator.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/iterator/transform_iterator.h>
#include <thrust/functional.h>
#include <thrust/extrema.h>
#include <thrust/transform_reduce.h>

#define DSIZE (1048576*2)
#define SSIZE 4
#define FSIZE (DSIZE*SSIZE)
#define nTPB 256
#define BSIZE nTPB
#define nBLKS 64
#define FLOAT_MIN (-99999)
#define FLOAT_MAX 99999


typedef thrust::tuple<float, float, float, float, float, float> tpl6;

struct expand_functor
{
  __host__ __device__
  tpl6 operator()(const float4 a){
    tpl6 result;
    result.get<0>() = a.x;
    result.get<1>() = a.x;
    result.get<2>() = a.y;
    result.get<3>() = a.y;
    result.get<4>() = a.z;
    result.get<5>() = a.z;
    return result;
  }
};


struct minmax3_functor
{
  __host__ __device__
  tpl6 operator()(const tpl6 a, const tpl6 b) {
    tpl6 result;
    result.get<0>() = (a.get<0>() < b.get<0>()) ? a.get<0>():b.get<0>();
    result.get<1>() = (a.get<1>() > b.get<1>()) ? a.get<1>():b.get<1>();
    result.get<2>() = (a.get<2>() < b.get<2>()) ? a.get<2>():b.get<2>();
    result.get<3>() = (a.get<3>() > b.get<3>()) ? a.get<3>():b.get<3>();
    result.get<4>() = (a.get<4>() < b.get<4>()) ? a.get<4>():b.get<4>();
    result.get<5>() = (a.get<5>() > b.get<5>()) ? a.get<5>():b.get<5>();
    return result;
  }
};

__device__ int bcount = 0;
__device__ float xmins[nBLKS];
__device__ float xmaxs[nBLKS];
__device__ float ymins[nBLKS];
__device__ float ymaxs[nBLKS];
__device__ float zmins[nBLKS];
__device__ float zmaxs[nBLKS];

__global__ void my_minmax3(float4 *data, float *results, size_t dsize){
  // assumes power-of-2 threadblock size
  // assumes nBLKS <= nTPB, nBLKS also power-of-2
  __shared__ float xmin[BSIZE], xmax[BSIZE], ymin[BSIZE], ymax[BSIZE], zmin[BSIZE], zmax[BSIZE];
  __shared__ int lblock;

  float my_xmin = FLOAT_MAX;
  float my_ymin = FLOAT_MAX;
  float my_zmin = FLOAT_MAX;
  float my_xmax = FLOAT_MIN;
  float my_ymax = FLOAT_MIN;
  float my_zmax = FLOAT_MIN;
  int idx = threadIdx.x+blockDim.x*blockIdx.x;
  while (idx < dsize){
    float4 my_temp = data[idx];
    if (my_xmin > my_temp.x) my_xmin = my_temp.x;
    if (my_ymin > my_temp.y) my_ymin = my_temp.y;
    if (my_zmin > my_temp.z) my_zmin = my_temp.z;
    if (my_xmax < my_temp.x) my_xmax = my_temp.x;
    if (my_ymax < my_temp.y) my_ymax = my_temp.y;
    if (my_zmax < my_temp.z) my_zmax = my_temp.z;
    idx += blockDim.x*gridDim.x;}
  xmin[threadIdx.x] = my_xmin;
  ymin[threadIdx.x] = my_ymin;
  zmin[threadIdx.x] = my_zmin;
  xmax[threadIdx.x] = my_xmax;
  ymax[threadIdx.x] = my_ymax;
  zmax[threadIdx.x] = my_zmax;
  __syncthreads();
  for (int i = blockDim.x/2; i > 0; i>>=1){
    if (threadIdx.x < i){
      if (xmin[threadIdx.x] > xmin[threadIdx.x+i]) xmin[threadIdx.x] = xmin[threadIdx.x + i];
      if (ymin[threadIdx.x] > ymin[threadIdx.x+i]) ymin[threadIdx.x] = ymin[threadIdx.x + i];
      if (zmin[threadIdx.x] > zmin[threadIdx.x+i]) zmin[threadIdx.x] = zmin[threadIdx.x + i];
      if (xmax[threadIdx.x] < xmax[threadIdx.x+i]) xmax[threadIdx.x] = xmax[threadIdx.x + i];
      if (ymax[threadIdx.x] < ymax[threadIdx.x+i]) ymax[threadIdx.x] = ymax[threadIdx.x + i];
      if (zmax[threadIdx.x] < zmax[threadIdx.x+i]) zmax[threadIdx.x] = zmax[threadIdx.x + i];
      }
    __syncthreads();
    }
  if (!threadIdx.x){
    xmins[blockIdx.x] = xmin[0];
    xmaxs[blockIdx.x] = xmax[0];
    ymins[blockIdx.x] = ymin[0];
    ymaxs[blockIdx.x] = ymax[0];
    zmins[blockIdx.x] = zmin[0];
    zmaxs[blockIdx.x] = zmax[0];
    __threadfence();
    if (atomicAdd(&bcount, 1) == (nBLKS-1)) lblock = 1;
    else lblock = 0;
    }
  __syncthreads();
  if (lblock){ // last block does final reduction
    if (threadIdx.x < nBLKS){
      xmin[threadIdx.x] = xmins[threadIdx.x];
      xmax[threadIdx.x] = xmaxs[threadIdx.x];
      ymin[threadIdx.x] = ymins[threadIdx.x];
      ymax[threadIdx.x] = ymaxs[threadIdx.x];
      zmin[threadIdx.x] = zmins[threadIdx.x];
      zmax[threadIdx.x] = zmaxs[threadIdx.x];}
    __syncthreads();
    for (int i = nBLKS/2; i > 0; i>>=1){
      if (threadIdx.x < i){
        if (xmin[threadIdx.x] > xmin[threadIdx.x+i]) xmin[threadIdx.x] = xmin[threadIdx.x + i];
        if (ymin[threadIdx.x] > ymin[threadIdx.x+i]) ymin[threadIdx.x] = ymin[threadIdx.x + i];
        if (zmin[threadIdx.x] > zmin[threadIdx.x+i]) zmin[threadIdx.x] = zmin[threadIdx.x + i];
        if (xmax[threadIdx.x] < xmax[threadIdx.x+i]) xmax[threadIdx.x] = xmax[threadIdx.x + i];
        if (ymax[threadIdx.x] < ymax[threadIdx.x+i]) ymax[threadIdx.x] = ymax[threadIdx.x + i];
        if (zmax[threadIdx.x] < zmax[threadIdx.x+i]) zmax[threadIdx.x] = zmax[threadIdx.x + i];
        }
      __syncthreads();
      }
    if (!threadIdx.x){
      results[0] = xmin[0];
      results[1] = xmax[0];
      results[2] = ymin[0];
      results[3] = ymax[0];
      results[4] = zmin[0];
      results[5] = zmax[0];
      }

   }
}

template <typename Iterator>
class strided_range
{
    public:

    typedef typename thrust::iterator_difference<Iterator>::type difference_type;

    struct stride_functor : public thrust::unary_function<difference_type,difference_type>
    {
        difference_type stride;

        stride_functor(difference_type stride)
            : stride(stride) {}

        __host__ __device__
        difference_type operator()(const difference_type& i) const
        {
            return stride * i;
        }
    };

    typedef typename thrust::counting_iterator<difference_type>                   CountingIterator;
    typedef typename thrust::transform_iterator<stride_functor, CountingIterator> TransformIterator;
    typedef typename thrust::permutation_iterator<Iterator,TransformIterator>     PermutationIterator;

    // type of the strided_range iterator
    typedef PermutationIterator iterator;

    // construct strided_range for the range [first,last)
    strided_range(Iterator first, Iterator last, difference_type stride)
        : first(first), last(last), stride(stride) {}

    iterator begin(void) const
    {
        return PermutationIterator(first, TransformIterator(CountingIterator(0), stride_functor(stride)));
    }

    iterator end(void) const
    {
        return begin() + ((last - first) + (stride - 1)) / stride;
    }

    protected:
    Iterator first;
    Iterator last;
    difference_type stride;
};

typedef thrust::device_vector<float>::iterator Iter;
typedef strided_range<Iter>::iterator sIter;

int main(){
// set up test data
  cudaEvent_t start, stop;
  float et;
  cudaEventCreate(&start); cudaEventCreate(&stop);
  thrust::host_vector<float> h_vals(FSIZE);
  for (int i = 0; i < DSIZE; i ++) {
    h_vals[i*SSIZE + 0] = rand()/(float)RAND_MAX;
    h_vals[i*SSIZE + 1] = rand()/(float)RAND_MAX;
    h_vals[i*SSIZE + 2] = rand()/(float)RAND_MAX;
    h_vals[i*SSIZE + 3] = 0.0f;}
  thrust::device_vector<float> d_vals = h_vals;
// set up strided ranges
  strided_range<Iter> item_x(d_vals.begin()  , d_vals.end(), SSIZE);
  strided_range<Iter> item_y(d_vals.begin()+1, d_vals.end(), SSIZE);
  strided_range<Iter> item_z(d_vals.begin()+2, d_vals.end(), SSIZE);
// find min and max
  cudaEventRecord(start);
  thrust::pair<sIter, sIter> result_x = thrust::minmax_element(item_x.begin(), item_x.end());
  thrust::pair<sIter, sIter> result_y = thrust::minmax_element(item_y.begin(), item_y.end());
  thrust::pair<sIter, sIter> result_z = thrust::minmax_element(item_z.begin(), item_z.end());
  cudaEventRecord(stop);
  cudaEventSynchronize(stop);
  cudaEventElapsedTime(&et, start, stop);
  std::cout << "thrust elapsed time: " << et << "ms" << std::endl;
  std::cout << "thrust results: " << std::endl;
  std::cout << "x min element = " << *(result_x.first)  << std::endl;
  std::cout << "x max element = " << *(result_x.second) << std::endl;
  std::cout << "y min element = " << *(result_y.first)  << std::endl;
  std::cout << "y max element = " << *(result_y.second) << std::endl;
  std::cout << "z min element = " << *(result_z.first)  << std::endl;
  std::cout << "z max element = " << *(result_z.second) << std::endl;

  float *h_results, *d_results;
  h_results = new float[6];
  cudaMalloc(&d_results, 6*sizeof(float));
  cudaEventRecord(start);
  my_minmax3<<<nBLKS,nTPB>>>((float4 *)thrust::raw_pointer_cast(d_vals.data()), d_results, DSIZE);
  cudaEventRecord(stop);
  cudaEventSynchronize(stop);
  cudaEventElapsedTime(&et, start, stop);
  cudaMemcpy(h_results, d_results, 6*sizeof(float), cudaMemcpyDeviceToHost);
  std::cout << "kernel elapsed time: " << et << "ms" << std::endl;
  std::cout << "kernel results: " << std::endl;
  std::cout << "x min element = " << h_results[0]  << std::endl;
  std::cout << "x max element = " << h_results[1]  << std::endl;
  std::cout << "y min element = " << h_results[2]  << std::endl;
  std::cout << "y max element = " << h_results[3]  << std::endl;
  std::cout << "z min element = " << h_results[4]  << std::endl;
  std::cout << "z max element = " << h_results[5]  << std::endl;

  thrust::device_ptr<float4> dptr_vals = thrust::device_pointer_cast(reinterpret_cast<float4 *>( thrust::raw_pointer_cast(d_vals.data())));
  tpl6 my_init;
  my_init.get<0>() = FLOAT_MAX;
  my_init.get<1>() = FLOAT_MIN;
  my_init.get<2>() = FLOAT_MAX;
  my_init.get<3>() = FLOAT_MIN;
  my_init.get<4>() = FLOAT_MAX;
  my_init.get<5>() = FLOAT_MIN;
  cudaEventRecord(start);
  tpl6 my_result = thrust::transform_reduce(dptr_vals, dptr_vals + DSIZE, expand_functor(),  my_init, minmax3_functor());
  cudaEventRecord(stop);
  cudaEventSynchronize(stop);
  cudaEventElapsedTime(&et, start, stop);
  cudaMemcpy(h_results, d_results, 6*sizeof(float), cudaMemcpyDeviceToHost);
  std::cout << "thrust2 elapsed time: " << et << "ms" << std::endl;
  std::cout << "thrust2 results: " << std::endl;
  std::cout << "x min element = " << my_result.get<0>()  << std::endl;
  std::cout << "x max element = " << my_result.get<1>()  << std::endl;
  std::cout << "y min element = " << my_result.get<2>()  << std::endl;
  std::cout << "y max element = " << my_result.get<3>()  << std::endl;
  std::cout << "z min element = " << my_result.get<4>()  << std::endl;
  std::cout << "z max element = " << my_result.get<5>()  << std::endl;
  return 0;
}
$ nvcc -O3 -arch=sm_20 -o t491 t491.cu
$ ./t491
thrust elapsed time: 3.88784ms
thrust results:
x min element = 1.16788e-06
x max element = 0.999998
y min element = 2.85916e-07
y max element = 1
z min element = 1.72295e-08
z max element = 0.999999
kernel elapsed time: 0.462848ms
kernel results:
x min element = 1.16788e-06
x max element = 0.999998
y min element = 2.85916e-07
y max element = 1
z min element = 1.72295e-08
z max element = 0.999999
thrust2 elapsed time: 1.29728ms
thrust2 results:
x min element = 1.16788e-06
x max element = 0.999998
y min element = 2.85916e-07
y max element = 1
z min element = 1.72295e-08
z max element = 0.999999
$

我已经更新了上面的示例，以进行比较，包括一个“优化”的缩减内核，它在单个内核调用中执行所有6个缩减（最小和最大操作）。

如预期的那样，在我的情况下（RHEL5.5，Quadro5000，CUDA 6.5RC），这种方法运行速度比生成相同结果的3个独立的thrust调用快约5-8倍，具体取决于数据大小。请注意，虽然我在这里选择了一个2的幂（DSIZE）的数据大小，但整个示例对于任意数据大小都可以正确工作。为了简洁起见，我已经放弃了适当的CUDA错误检查。

编辑:根据@JaredHoberock的建议，我添加了第三种方法，允许一次调用thrust::transform_reduce生成所有6个结果。这些是上面的“thrust2”结果。这种方法比第一个（三个thrust调用）方法快约3倍。仍然不如cuda kernel方法快，但也许这种thrust方法可以进一步优化。

- Robert Crovella

---

由于这是一个并行约简问题，请注意Kepler架构引入的Warp Shuffle内在函数，可用于加速先前共享内存模式下的约简。 - Tyson Hilmer

Answer 2

这是一个关于步长范围示例的应用程序。

#include <thrust/iterator/counting_iterator.h>
#include <thrust/iterator/transform_iterator.h>
#include <thrust/iterator/permutation_iterator.h>
#include <thrust/functional.h>

#include <thrust/fill.h>
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>

// for printing
#include <thrust/copy.h>
#include <ostream>

#define STRIDE 2

template <typename Iterator>
class strided_range
{
    public:

    typedef typename thrust::iterator_difference<Iterator>::type difference_type;

    struct stride_functor : public thrust::unary_function<difference_type,difference_type>
    {
        difference_type stride;

        stride_functor(difference_type stride)
            : stride(stride) {}

        __host__ __device__
        difference_type operator()(const difference_type& i) const
        {
            return stride * i;
        }
    };

    typedef typename thrust::counting_iterator<difference_type>                   CountingIterator;
    typedef typename thrust::transform_iterator<stride_functor, CountingIterator> TransformIterator;
    typedef typename thrust::permutation_iterator<Iterator,TransformIterator>     PermutationIterator;

    // type of the strided_range iterator
    typedef PermutationIterator iterator;

    // construct strided_range for the range [first,last)
    strided_range(Iterator first, Iterator last, difference_type stride)
        : first(first), last(last), stride(stride) {}

    iterator begin(void) const
    {
        return PermutationIterator(first, TransformIterator(CountingIterator(0), stride_functor(stride)));
    }

    iterator end(void) const
    {
        return begin() + ((last - first) + (stride - 1)) / stride;
    }

    protected:
    Iterator first;
    Iterator last;
    difference_type stride;
};

int main(void)
{
    const int N = 9;

    thrust::host_vector<int> h_data(N);
    for (int i=0; i<N; i++) h_data[i] = i;

    thrust::device_vector<int> data(h_data);

    typedef thrust::device_vector<int>::iterator Iterator;
    strided_range<Iterator> pos(data.begin(), data.end(), STRIDE);

    int sum = thrust::reduce(pos.begin(), pos.end(), (int) 0, thrust::plus<int>());

    printf("sum = %i\n",sum);

    return 0;
} 

- Vitality

Answer 3

我按照这份文档进行了内核优化：http://www.cuvilib.com/Reduction.pdf 同样的文档（文件大小更小）也可以在这里找到：https://developer.download.nvidia.com/assets/cuda/files/reduction.pdf 现在看起来是这样的：

template<unsigned int blockSize>
__global__ void minMaxReduction_k(float *g_minData, float *g_maxData, float *g_minOutput, float *g_maxOutput, unsigned int n)
{
    extern __shared__ float shared[];
    float* minSdata = (float*)shared; 
    float* maxSdata = (float*)&minSdata[4*blockDim.x]; 

    // reading from global memory, writing to shared memory
    unsigned int tid = threadIdx.x;
    unsigned int i = blockIdx.x*(blockSize*2) + threadIdx.x;
    unsigned int gridSize = blockSize*2*gridDim.x;
    
    minSdata[4*tid] = FLT_MAX;
    minSdata[4*tid+1] = FLT_MAX;
    minSdata[4*tid+2] = FLT_MAX;

    maxSdata[4*tid] = -FLT_MAX;
    maxSdata[4*tid+1] = -FLT_MAX;
    maxSdata[4*tid+2] = -FLT_MAX;
    
    while(i<n){
        minSdata[4*tid] = fminf(fminf(minSdata[4*tid], g_minData[4*i]), g_minData[4*(i+blockDim.x)]);
        minSdata[4*tid+1] = fminf(fminf(minSdata[4*tid+1], g_minData[4*i+1]), g_minData[4*(i+blockDim.x)+1]);
        minSdata[4*tid+2] = fminf(fminf(minSdata[4*tid+2], g_minData[4*i+2]), g_minData[4*(i+blockDim.x)+2]);

        maxSdata[4*tid] = fmaxf(fmaxf(maxSdata[4*tid], g_maxData[4*i]), g_maxData[4*(i+blockDim.x)]);
        maxSdata[4*tid+1] = fmaxf(fmaxf(maxSdata[4*tid+1], g_maxData[4*i+1]), g_maxData[4*(i+blockDim.x)+1]);
        maxSdata[4*tid+2] = fmaxf(fmaxf(maxSdata[4*tid+2], g_maxData[4*i+2]), g_maxData[4*(i+blockDim.x)+2]);

        i+=gridSize;
    }

    __syncthreads();

    if(blockSize >= 1024){
        if(tid < 512){
            minSdata[4*tid] = fminf(minSdata[4*tid], minSdata[4*(tid+512)]);
            minSdata[4*tid+1] = fminf(minSdata[4*tid+1], minSdata[4*(tid+512)+1]);
            minSdata[4*tid+2] = fminf(minSdata[4*tid+2], minSdata[4*(tid+512)+2]);

            maxSdata[4*tid] = fmaxf(maxSdata[4*tid], maxSdata[4*(tid+512)]);
            maxSdata[4*tid+1] = fmaxf(maxSdata[4*tid+1], maxSdata[4*(tid+512)+1]);
            maxSdata[4*tid+2] = fmaxf(maxSdata[4*tid+2], maxSdata[4*(tid+512)+2]);
        }
        __syncthreads();
    }

    if(blockSize >= 512){
        if(tid < 256){
            minSdata[4*tid] = fminf(minSdata[4*tid], minSdata[4*(tid+256)]);
            minSdata[4*tid+1] = fminf(minSdata[4*tid+1], minSdata[4*(tid+256)+1]);
            minSdata[4*tid+2] = fminf(minSdata[4*tid+2], minSdata[4*(tid+256)+2]);

            maxSdata[4*tid] = fmaxf(maxSdata[4*tid], maxSdata[4*(tid+256)]);
            maxSdata[4*tid+1] = fmaxf(maxSdata[4*tid+1], maxSdata[4*(tid+256)+1]);
            maxSdata[4*tid+2] = fmaxf(maxSdata[4*tid+2], maxSdata[4*(tid+256)+2]);
        }
        __syncthreads();
    }

    if(blockSize >= 256){
        if(tid < 128){
            minSdata[4*tid] = fminf(minSdata[4*tid], minSdata[4*(tid+128)]);
            minSdata[4*tid+1] = fminf(minSdata[4*tid+1], minSdata[4*(tid+128)+1]);
            minSdata[4*tid+2] = fminf(minSdata[4*tid+2], minSdata[4*(tid+128)+2]);

            maxSdata[4*tid] = fmaxf(maxSdata[4*tid], maxSdata[4*(tid+128)]);
            maxSdata[4*tid+1] = fmaxf(maxSdata[4*tid+1], maxSdata[4*(tid+128)+1]);
            maxSdata[4*tid+2] = fmaxf(maxSdata[4*tid+2], maxSdata[4*(tid+128)+2]);
        }
        __syncthreads();
    }

    if(blockSize >= 128){
        if(tid < 64){
            minSdata[4*tid] = fminf(minSdata[4*tid], minSdata[4*(tid+64)]);
            minSdata[4*tid+1] = fminf(minSdata[4*tid+1], minSdata[4*(tid+64)+1]);
            minSdata[4*tid+2] = fminf(minSdata[4*tid+2], minSdata[4*(tid+64)+2]);

            maxSdata[4*tid] = fmaxf(maxSdata[4*tid], maxSdata[4*(tid+64)]);
            maxSdata[4*tid+1] = fmaxf(maxSdata[4*tid+1], maxSdata[4*(tid+64)+1]);
            maxSdata[4*tid+2] = fmaxf(maxSdata[4*tid+2], maxSdata[4*(tid+64)+2]);
        }
        __syncthreads();
    }

    if(tid<32){
        if (blockSize >= 64){
            minSdata[4*tid] = fminf(minSdata[4*tid], minSdata[4*(tid+32)]);
            minSdata[4*tid+1] = fminf(minSdata[4*tid+1], minSdata[4*(tid+32)+1]);
            minSdata[4*tid+2] = fminf(minSdata[4*tid+2], minSdata[4*(tid+32)+2]);

            maxSdata[4*tid] = fmaxf(maxSdata[4*tid], maxSdata[4*(tid+32)]);
            maxSdata[4*tid+1] = fmaxf(maxSdata[4*tid+1], maxSdata[4*(tid+32)+1]);
            maxSdata[4*tid+2] = fmaxf(maxSdata[4*tid+2], maxSdata[4*(tid+32)+2]);
        }
        //
        if (blockSize >= 32){
            minSdata[4*tid] = fminf(minSdata[4*tid], minSdata[4*(tid+16)]);
            minSdata[4*tid+1] = fminf(minSdata[4*tid+1], minSdata[4*(tid+16)+1]);
            minSdata[4*tid+2] = fminf(minSdata[4*tid+2], minSdata[4*(tid+16)+2]);

            maxSdata[4*tid] = fmaxf(maxSdata[4*tid], maxSdata[4*(tid+16)]);
            maxSdata[4*tid+1] = fmaxf(maxSdata[4*tid+1], maxSdata[4*(tid+16)+1]);
            maxSdata[4*tid+2] = fmaxf(maxSdata[4*tid+2], maxSdata[4*(tid+16)+2]);
        }
        //
        if (blockSize >= 16){
            minSdata[4*tid] = fminf(minSdata[4*tid], minSdata[4*(tid+8)]);
            minSdata[4*tid+1] = fminf(minSdata[4*tid+1], minSdata[4*(tid+8)+1]);
            minSdata[4*tid+2] = fminf(minSdata[4*tid+2], minSdata[4*(tid+8)+2]);

            maxSdata[4*tid] = fmaxf(maxSdata[4*tid], maxSdata[4*(tid+8)]);
            maxSdata[4*tid+1] = fmaxf(maxSdata[4*tid+1], maxSdata[4*(tid+8)+1]);
            maxSdata[4*tid+2] = fmaxf(maxSdata[4*tid+2], maxSdata[4*(tid+8)+2]);
        }
        //
        if (blockSize >= 8){
            minSdata[4*tid] = fminf(minSdata[4*tid], minSdata[4*(tid+4)]);
            minSdata[4*tid+1] = fminf(minSdata[4*tid+1], minSdata[4*(tid+4)+1]);
            minSdata[4*tid+2] = fminf(minSdata[4*tid+2], minSdata[4*(tid+4)+2]);

            maxSdata[4*tid] = fmaxf(maxSdata[4*tid], maxSdata[4*(tid+4)]);
            maxSdata[4*tid+1] = fmaxf(maxSdata[4*tid+1], maxSdata[4*(tid+4)+1]);
            maxSdata[4*tid+2] = fmaxf(maxSdata[4*tid+2], maxSdata[4*(tid+4)+2]);
        }
        //
        if (blockSize >= 4){
            minSdata[4*tid] = fminf(minSdata[4*tid], minSdata[4*(tid+2)]);
            minSdata[4*tid+1] = fminf(minSdata[4*tid+1], minSdata[4*(tid+2)+1]);
            minSdata[4*tid+2] = fminf(minSdata[4*tid+2], minSdata[4*(tid+2)+2]);

            maxSdata[4*tid] = fmaxf(maxSdata[4*tid], maxSdata[4*(tid+2)]);
            maxSdata[4*tid+1] = fmaxf(maxSdata[4*tid+1], maxSdata[4*(tid+2)+1]);
            maxSdata[4*tid+2] = fmaxf(maxSdata[4*tid+2], maxSdata[4*(tid+2)+2]);
        }
        //
        if (blockSize >= 2){
            minSdata[4*tid] = fminf(minSdata[4*tid], minSdata[4*(tid+1)]);
            minSdata[4*tid+1] = fminf(minSdata[4*tid+1], minSdata[4*(tid+1)+1]);
            minSdata[4*tid+2] = fminf(minSdata[4*tid+2], minSdata[4*(tid+1)+2]);

            maxSdata[4*tid] = fmaxf(maxSdata[4*tid], maxSdata[4*(tid+1)]);
            maxSdata[4*tid+1] = fmaxf(maxSdata[4*tid+1], maxSdata[4*(tid+1)+1]);
            maxSdata[4*tid+2] = fmaxf(maxSdata[4*tid+2], maxSdata[4*(tid+1)+2]);
        }
    }
    // write result for this block to global mem

    if (tid == 0){
        g_minOutput[blockIdx.x] = minSdata[0];
        g_minOutput[blockIdx.x+1] = minSdata[1];
        g_minOutput[blockIdx.x+2] = minSdata[2];

        g_maxOutput[blockIdx.x] = maxSdata[0];
        g_maxOutput[blockIdx.x+1] = maxSdata[1];
        g_maxOutput[blockIdx.x+2] = maxSdata[2];
    }

}

像这样调用：

float *d_minOutput;
    float *d_maxOutput;

    int tempN = n;
    while(tempN>1){
        getNumBlocksAndThreads(tempN, 65535, 1024, blocks, threads);

        cudaMalloc((void **)&d_minOutput, 4*(sizeof(float)*blocks));
        cudaMalloc((void **)&d_maxOutput, 4*(sizeof(float)*blocks));

        int smem = (threads <= 32) ? 2*2*4*threads*sizeof(float) : 2*4*threads*sizeof(float);

        switch(threads){
        case 1024:
            minMaxReduction_k<1024><<< blocks, threads, smem >>>(d_minData, d_maxData, d_minOutput, d_maxOutput, tempN);
            break;
        case 512:
            minMaxReduction_k<512><<< blocks, threads, smem >>>(d_minData, d_maxData, d_minOutput, d_maxOutput, tempN);
            break;
        case 256:
            minMaxReduction_k<256><<< blocks, threads, smem >>>(d_minData, d_maxData, d_minOutput, d_maxOutput, tempN);
            break;
        case 128:
            minMaxReduction_k<128><<< blocks, threads, smem >>>(d_minData, d_maxData, d_minOutput, d_maxOutput, tempN);
            break;
        case 64:
            minMaxReduction_k<64><<< blocks, threads, smem >>>(d_minData, d_maxData, d_minOutput, d_maxOutput, tempN);
            break;
        case 32:
            minMaxReduction_k<32><<< blocks, threads, smem >>>(d_minData, d_maxData, d_minOutput, d_maxOutput, tempN);
            break;
        case 16:
            minMaxReduction_k<16><<< blocks, threads, smem >>>(d_minData, d_maxData, d_minOutput, d_maxOutput, tempN);
            break;
        case 8:
            minMaxReduction_k<8><<< blocks, threads, smem >>>(d_minData, d_maxData, d_minOutput, d_maxOutput, tempN);
            break;
        case 4:
            minMaxReduction_k<4><<< blocks, threads, smem >>>(d_minData, d_maxData, d_minOutput, d_maxOutput, tempN);
            break;
        case 2:
            minMaxReduction_k<2><<< blocks, threads, smem >>>(d_minData, d_maxData, d_minOutput, d_maxOutput, tempN);
            break;
        case 1:
            minMaxReduction_k<1><<< blocks, threads, smem >>>(d_minData, d_maxData, d_minOutput, d_maxOutput, tempN);
            break;
        }

        tempN = blocks;

        cudaMemcpy(d_minData, d_minOutput, 4*(sizeof(float)*blocks), cudaMemcpyDeviceToDevice);
        cudaMemcpy(d_maxData, d_maxOutput, 4*(sizeof(float)*blocks), cudaMemcpyDeviceToDevice);

        cudaFree(d_minOutput);
        cudaFree(d_maxOutput);
    }

辅助程序：

void UniformGrid::getNumBlocksAndThreads(unsigned int n, unsigned int maxBlocks, unsigned int maxThreads, unsigned int &blocks, unsigned int &threads)
{
    //get device capability, to avoid block/grid size excceed the upbound
    cudaDeviceProp prop;
    int device;
    cudaGetDevice(&device);
    cudaGetDeviceProperties(&prop, device);

    threads = (n < maxThreads*2) ? nextPow2((n + 1)/ 2) : maxThreads;
    blocks = (n + (threads * 2 - 1)) / (threads * 2);


    if ((float)threads*blocks > (float)prop.maxGridSize[0] * prop.maxThreadsPerBlock)
    {
        printf("n is too large, please choose a smaller number!\n");
    }

    if (blocks > (unsigned int) prop.maxGridSize[0])
    {
        printf("Grid size <%d> excceeds the device capability <%d>, set block size as %d (original %d)\n",
            blocks, prop.maxGridSize[0], threads*2, threads);

        blocks /= 2;
        threads *= 2;
    }
}

unsigned int UniformGrid::nextPow2(unsigned int x)
{
    --x;
    x |= x >> 1;
    x |= x >> 2;
    x |= x >> 4;
    x |= x >> 8;
    x |= x >> 16;
    return ++x;
}

在几天内，我可能会测试哪个解决方案更快。

- aerion

---

我想你可以通过将重复的if(blockSize >= 1024){}替换为for循环来减少至少80%的代码行数，在调用中完全省略switch(你可以直接调用minMaxReduction_k<threads>...等)。 - avtomaton