非极大值抑制算法（Non-Maximum Suppression，NMS）

https://tcnull.github.io/nms/
https://blog.csdn.net/weicao1990/article/details/103857298

目标检测中检测出了许多的候选框，候选框之间是有重叠的，NMS作用重叠的候选框只保留一个

算法：

1. 将所有候选框放入到集和B
2. 从B中选出分数S最大的bm
3. 将bm 从集和B中移除到集和D
4. 计算bm与B中剩余的候选框之间的IOU。
5. 如果iou>Nt则将其从B中删除。（去除重叠比较多的候选框）
6. 循环直至B为空。
7. D会越来越多。

def box_iou_union_2d(boxes1: Tensor, boxes2: Tensor, eps: float = 0) -> Tuple[Tensor, Tensor]:
    """
    Return intersection-over-union (Jaccard index) and  of boxes.
    Both sets of boxes are expected to be in (x1, y1, x2, y2) format.

    Arguments:
        boxes1: set of boxes (x1, y1, x2, y2)[N, 4]
        boxes2: set of boxes (x1, y1, x2, y2)[M, 4]
        eps: optional small constant for numerical stability

    Returns:
        iou (Tensor[N, M]): the NxM matrix containing the pairwise
            IoU values for every element in boxes1 and boxes2
        union (Tensor[N, M]): the nxM matrix containing the pairwise union
            values
    """
    area1 = box_area(boxes1)
    area2 = box_area(boxes2)

    x1 = torch.max(boxes1[:, None, 0], boxes2[:, 0])  # [N, M]
    y1 = torch.max(boxes1[:, None, 1], boxes2[:, 1])  # [N, M]
    x2 = torch.min(boxes1[:, None, 2], boxes2[:, 2])  # [N, M]
    y2 = torch.min(boxes1[:, None, 3], boxes2[:, 3])  # [N, M]

    inter = ((x2 - x1).clamp(min=0) * (y2 - y1).clamp(min=0)) + eps  # [N, M]
    union = (area1[:, None] + area2 - inter)
    return inter / union, union

def nms_cpu(boxes, scores, thresh):
    """
    Performs non-maximum suppression for 3d boxes on cpu
   
    Args:
        boxes (Tensor): tensor with boxes (x1, y1, x2, y2, (z1, z2))[N, dim * 2]
        scores (Tensor): score for each box [N]
        iou_threshold (float): threshould when boxes are discarded
   
    Returns:
        keep (Tensor): int64 tensor with the indices of the elements that have been kept by NMS,
            sorted in decreasing order of scores
    """
    ious = box_iou(boxes, boxes)
    _, _idx = torch.sort(scores, descending=True)
   
    keep = []
    while _idx.nelement() > 0:
        keep.append(_idx[0])
        # get all elements that were not matched and discard all others.
        non_matches = torch.where((ious[_idx[0]][_idx] <= thresh))[0]
        _idx = _idx[non_matches]
    return torch.tensor(keep).to(boxes).long()

---

template <typename T>
__device__ inline float devIoU(T const* const a, T const* const b) {
  // a, b hold box coords as (y1, x1, y2, x2) with y1 < y2 etc.
  T bottom = max(a[0], b[0]), top = min(a[2], b[2]);
  T left = max(a[1], b[1]), right = min(a[3], b[3]);
  T width = max(right - left, (T)0), height = max(top - bottom, (T)0);
  T interS = width * height;

  T Sa = (a[2] - a[0]) * (a[3] - a[1]);
  T Sb = (b[2] - b[0]) * (b[3] - b[1]);

  return interS / (Sa + Sb - interS);
}

at::Tensor nms_cuda(const at::Tensor& dets, const at::Tensor& scores, float iou_threshold) {
  /* dets expected as (n_dets, dim) where dim=4 in 2D, dim=6 in 3D */
  AT_ASSERTM(dets.type().is_cuda(), "dets must be a CUDA tensor");
  AT_ASSERTM(scores.type().is_cuda(), "scores must be a CUDA tensor");
  at::cuda::CUDAGuard device_guard(dets.device());//管理CUDA设备上下文，并指制定使用的CUDA设备

  bool is_3d = dets.size(1) == 6;
  //按照第一维（索引为0）对scores降序排序，并返回一个包含排序后索引的 Tensor,std::get<1>(...) 提取排序后索引的 Tensor。
  auto order_t = std::get<1>(scores.sort(0, /* descending=*/true));
  //使用排序后的索引 order_t 从 dets 中选择对应的元素，返回一个新的 Tensor; dets_sorted，其中的元素按照排序后的顺序排列。
  auto dets_sorted = dets.index_select(0, order_t);
  //bbox个数
  int dets_num = dets.size(0);
  //该函数用于计算在CUDA中进行并行化计算时所需的最大列块数。其中，dets_num代表待处理的数据数量，threadsPerBlock表示每个CUDA块中的线程数量。
  //函数通过将dets_num除以threadsPerBlock并向上取整得到最大列块数col_blocks。这个函数常用于确定CUDA并行计算中需要启动多少个CUDA块来处理所有数据。
  const int col_blocks = at::cuda::ATenCeilDiv(dets_num, threadsPerBlock);//获取block个数

  //该函数创建了一个名为mask的空Tensor，其大小为dets_num * col_blocks，数据类型为长整型（at::kLong）。该Tensor的属性与dets相同。
  at::Tensor mask = at::empty({dets_num * col_blocks}, dets.options().dtype(at::kLong));//创建一个空容器  D

  dim3 blocks(col_blocks, col_blocks);//定义了一个二维网格,其维度为col_blocks行col_blocks列
  dim3 threads(threadsPerBlock);//定义了一个线程块，其维度为threadsPerBlock；
  cudaStream_t stream = at::cuda::getCurrentCUDAStream();//获取当前的CUDA流，用于异步计算。


  if (is_3d) {
  //std::cout << "performing NMS on 3D boxes in CUDA" << std::endl;
  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
      dets_sorted.type(), "nms_kernel_cuda", [&] {
        nms_kernel_3d<scalar_t><<<blocks, threads, 0, stream>>>(
            dets_num,
            iou_threshold,
            dets_sorted.data_ptr<scalar_t>(),
            (unsigned long long*)mask.data_ptr<int64_t>());
      });
   }
   else {
    //该函数是PyTorch CUDA扩展中的一个宏定义，用于在CUDA代码中处理浮点类型数据（包括float、double、half）的泛型编程。
    //它会根据传入的输入数据类型,自动选择对应的CUDA内核函数进行计算。
    //这样可以避免为每种数据类型编写重复的代码，提高代码的可维护性和可扩展性。
    //相当与模板类
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(
      dets_sorted.type(), "nms_kernel_cuda", [&] {
        nms_kernel<scalar_t><<<blocks, threads, 0, stream>>>(
            dets_num,
            iou_threshold,
            dets_sorted.data_ptr<scalar_t>(),
            (unsigned long long*)mask.data_ptr<int64_t>());
      });

   }

  //将mask_cpu的数据拷贝到主机内存中
  at::Tensor mask_cpu = mask.to(at::kCPU);
  unsigned long long* mask_host = (unsigned long long*)mask_cpu.data_ptr<int64_t>();

  //创建一个空的remv Tensor，其大小为col_blocks，数据类型为unsigned long long。用于记录每个block中哪些线程被标记为无效。
  std::vector<unsigned long long> remv(col_blocks);
  memset(&remv[0], 0, sizeof(unsigned long long) * col_blocks);

  //创建一个空的keep Tensor，其大小为dets_num，数据类型为long。用于存放整个筛选后的索引。
  at::Tensor keep = at::empty({dets_num}, dets.options().dtype(at::kLong).device(at::kCPU));
  int64_t* keep_out = keep.data_ptr<int64_t>();

  int num_to_keep = 0;
  for (int i = 0; i < dets_num; i++) {
    int nblock = i / threadsPerBlock;//第几个block
    int inblock = i % threadsPerBlock;//block中第几个threads

    //判断当前线程是否被标记为无效
    if (!(remv[nblock] & (1ULL << inblock))) {
      keep_out[num_to_keep++] = i;
      unsigned long long* p = mask_host + i * col_blocks;//实际上将所有的形式设置成one hot形式。
      for (int j = nblock; j < col_blocks; j++) {
        remv[j] |= p[j];//按位或运算，并将结果保留在remv数组中
      }
    }
  }

  AT_CUDA_CHECK(cudaGetLastError());//检测CUDA 是否发生错误
  return order_t.index(
      {keep.narrow(/*dim=*/0, /*start=*/0, /*length=*/num_to_keep)//从张量keep中选择制定维度、起始位置和长度的子张量
           .to(order_t.device(), keep.scalar_type())});
}

template <typename T>
__global__ void nms_kernel(const int n_boxes, const float iou_threshold, const T* dev_boxes, unsigned long long* dev_mask) {
  const int row_start = blockIdx.y;
  const int col_start = blockIdx.x;

  // if (row_start > col_start) return;
  const int row_size = min(n_boxes - row_start * threadsPerBlock, threadsPerBlock);//当前block行可以放多少个box
  const int col_size = min(n_boxes - col_start * threadsPerBlock, threadsPerBlock);//获取block列可以放多少个box

  __shared__ T block_boxes[threadsPerBlock * 4];
  if (threadIdx.x < col_size) {//将当前block列的box拷贝到shared memory
    block_boxes[threadIdx.x * 4 + 0] =  dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 4 + 0];
    block_boxes[threadIdx.x * 4 + 1] =  dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 4 + 1];
    block_boxes[threadIdx.x * 4 + 2] =  dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 4 + 2];
    block_boxes[threadIdx.x * 4 + 3] =  dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 4 + 3];
  }
  __syncthreads();//同步

  if (threadIdx.x < row_size) {
    const int cur_box_idx = threadsPerBlock * row_start + threadIdx.x;//当前block行box的索引
    const T* cur_box = dev_boxes + cur_box_idx * 4;//当前block行box的指针
    int i = 0;
    unsigned long long t = 0;
    int start = 0;
    if (row_start == col_start) {
      start = threadIdx.x + 1;//自己的IOU就不算了
    }
    for (i = start; i < col_size; i++) {
      if (devIoU<T>(cur_box, block_boxes + i * 4) > iou_threshold) {//计算各自的IOU
        t |= 1ULL << i;//以二进制的形式表示重叠关系成立
      }
    }
    const int col_blocks = at::cuda::ATenCeilDiv(n_boxes, threadsPerBlock);//列block数
    dev_mask[cur_box_idx * col_blocks + col_start] = t;//将重叠关系写入shared memory
  }
}