scatter_points_cuda.cu

#include <ATen/ATen.h>
#include <ATen/cuda/CUDAContext.h>
#include <torch/types.h>
#include <ATen/cuda/CUDAApplyUtils.cuh>

typedef enum { SUM = 0, MEAN = 1, MAX = 2 } reduce_t;

#define CHECK_CUDA(x)                                                          \
  TORCH_CHECK(x.device().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x)                                                    \
  TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x)                                                         \
  CHECK_CUDA(x);                                                               \
  CHECK_CONTIGUOUS(x)

namespace {
int const threadsPerBlock = 512;
int const maxGridDim = 50000;
} // namespace

__device__ __forceinline__ static void reduceMax(float *address, float val) {
  int *address_as_i = reinterpret_cast<int *>(address);
  int old = *address_as_i, assumed;
  do {
    assumed = old;
    old = atomicCAS(address_as_i, assumed,
                    __float_as_int(fmaxf(val, __int_as_float(assumed))));
  } while (assumed != old || __int_as_float(old) < val);
}

__device__ __forceinline__ static void reduceMax(double *address, double val) {
  unsigned long long *address_as_ull =
      reinterpret_cast<unsigned long long *>(address);
  unsigned long long old = *address_as_ull, assumed;
  do {
    assumed = old;
    old = atomicCAS(
        address_as_ull, assumed,
        __double_as_longlong(fmax(val, __longlong_as_double(assumed))));
  } while (assumed != old || __longlong_as_double(old) < val);
}

// get rid of meaningless warnings when compiling host code
#ifdef __CUDA_ARCH__
__device__ __forceinline__ static void reduceAdd(float *address, float val) {
#if (__CUDA_ARCH__ < 200)
#warning                                                                       \
    "compute capability lower than 2.x. fall back to use CAS version of atomicAdd for float32"
  int *address_as_i = reinterpret_cast<int *>(address);
  int old = *address_as_i, assumed;
  do {
    assumed = old;
    old = atomicCAS(address_as_i, assumed,
                    __float_as_int(val + __int_as_float(assumed)));
  } while (assumed != old);
#else
  atomicAdd(address, val);
#endif
}

__device__ __forceinline__ static void reduceAdd(double *address, double val) {
#if (__CUDA_ARCH__ < 600)
#warning                                                                       \
    "compute capability lower than 6.x. fall back to use CAS version of atomicAdd for float64"
  unsigned long long *address_as_ull =
      reinterpret_cast<unsigned long long *>(address);
  unsigned long long old = *address_as_ull, assumed;
  do {
    assumed = old;
    old = atomicCAS(address_as_ull, assumed,
                    __double_as_longlong(val + __longlong_as_double(assumed)));
  } while (assumed != old);
#else
  atomicAdd(address, val);
#endif
}
#endif

template <typename T_int>
__global__ void coors_id_kernel(const T_int *coors, const T_int *dim,
                                int64_t *coors_id, const int num_input,
                                const int NDim) {
  for (int x = blockIdx.x * blockDim.x + threadIdx.x; x < num_input;
       x += gridDim.x * blockDim.x) {
    const T_int *coor_x = coors + x * NDim;
    auto coor_id = 0;
    for (int i = 0; i < NDim && coor_id != -1; i++) {
      coor_id *= dim[i];
      auto t = static_cast<int64_t>(coor_x[i]);
      coor_id = (t < 0) ? -1 : coor_id + t;
    }
    coors_id[x] = coor_id;
  }
}

template <typename T_int>
__global__ void coors_map_init_kernel(const int64_t *coors_id,
                                      const T_int *coors_id_argsort,
                                      int32_t *coors_map, const int num_input) {
  for (int x = blockIdx.x * blockDim.x + threadIdx.x; x < num_input;
       x += gridDim.x * blockDim.x) {
    auto here = coors_id[coors_id_argsort[x]];
    if (x == 0) {
      if (here == -1) { // there is invalid points
        coors_map[0] = -1;
      } else {
        coors_map[0] = 0;
      }
      continue;
    }
    auto left = coors_id[coors_id_argsort[x - 1]];
    coors_map[x] = (left < here) ? 1 : 0;
  }
}

template <typename T, typename T_int>
__global__ void
feats_reduce_kernel(const T *feats, const T_int *coors, int32_t *coors_map,
                    int32_t *reduce_count, // shall be 0 at initialization
                    T *reduced_feats,      // shall be 0 at initialization
                    T_int *out_coors, const int num_input, const int num_feats,
                    const int NDim, const reduce_t reduce_type) {
  for (int x = blockIdx.x * blockDim.x + threadIdx.x; x < num_input;
       x += gridDim.x * blockDim.x) {
    int32_t reduce_to = coors_map[x];
    if (reduce_to == -1)
      continue;

    const T_int *coors_offset = coors + x * NDim;
    T_int *out_coors_offset = out_coors + reduce_to * NDim;
    for (int i = 0; i < NDim; i++) {
      out_coors_offset[i] = coors_offset[i];
    }

    const T *feats_offset = feats + x * num_feats;
    T *reduced_feats_offset = reduced_feats + reduce_to * num_feats;
    if (reduce_type == reduce_t::MAX) {
      for (int i = 0; i < num_feats; i++) {
        reduceMax(&reduced_feats_offset[i], feats_offset[i]);
      }
    } else {
      if (reduce_type == reduce_t::MEAN) {
        atomicAdd(&reduce_count[reduce_to], static_cast<int32_t>(1));
      }
      for (int i = 0; i < num_feats; i++) {
        reduceAdd(&reduced_feats_offset[i], feats_offset[i]);
      }
    }
  }
}

template <typename T>
__global__ void add_reduce_traceback_grad_kernel(
    T *grad_feats, const T *grad_reduced_feats, const int32_t *coors_map,
    const int32_t *reduce_count, const int num_input, const int num_feats,
    const reduce_t reduce_type) {
  for (int x = blockIdx.x * blockDim.x + threadIdx.x; x < num_input;
       x += gridDim.x * blockDim.x) {
    int32_t reduce_to = coors_map[x];
    if (reduce_to == -1) {
      continue;
    }

    const int input_offset = x * num_feats;
    T *grad_feats_offset = grad_feats + input_offset;
    const int reduced_offset = reduce_to * num_feats;
    const T *grad_reduced_feats_offset = grad_reduced_feats + reduced_offset;

    if (reduce_type == reduce_t::SUM) {
      for (int i = 0; i < num_feats; i++) {
        grad_feats_offset[i] = grad_reduced_feats_offset[i];
      }
    } else if (reduce_type == reduce_t::MEAN) {
      for (int i = 0; i < num_feats; i++) {
        grad_feats_offset[i] = grad_reduced_feats_offset[i] /
                               static_cast<T>(reduce_count[reduce_to]);
      }
    }
  }
}

template <typename T>
__global__ void max_reduce_traceback_scatter_idx_kernel(
    const T *feats, const T *reduced_feats, int32_t *reduce_from,
    const int32_t *coors_map, const int num_input, const int num_feats) {
  for (int x = blockIdx.x * blockDim.x + threadIdx.x; x < num_input;
       x += gridDim.x * blockDim.x) {
    int32_t reduce_to = coors_map[x];

    const int input_offset = x * num_feats;
    const T *feats_offset = feats + input_offset;

    if (reduce_to == -1) {
      continue;
    }

    const int reduced_offset = reduce_to * num_feats;
    const T *reduced_feats_offset = reduced_feats + reduced_offset;
    int32_t *reduce_from_offset = reduce_from + reduced_offset;

    for (int i = 0; i < num_feats; i++) {
      if (feats_offset[i] == reduced_feats_offset[i]) {
        atomicMin(&reduce_from_offset[i], static_cast<int32_t>(x));
      }
    }
  }
}

template <typename T>
__global__ void
max_reduce_scatter_grad_kernel(T *grad_feats, const T *grad_reduced_feats,
                               const int32_t *reduce_from,
                               const int num_reduced, const int num_feats) {
  for (int x = blockIdx.x * blockDim.x + threadIdx.x; x < num_reduced;
       x += gridDim.x * blockDim.x) {

    const int reduced_offset = x * num_feats;
    const int32_t *scatter_to_offset = reduce_from + reduced_offset;
    const T *grad_reduced_feats_offset = grad_reduced_feats + reduced_offset;

    for (int i = 0; i < num_feats; i++) {
      grad_feats[scatter_to_offset[i] * num_feats + i] =
          grad_reduced_feats_offset[i];
    }
  }
}

namespace voxelization {

std::vector<at::Tensor>
dynamic_point_to_voxel_forward_gpu(const at::Tensor &feats,
                                   const at::Tensor &coors,
                                   const reduce_t reduce_type) {
  CHECK_INPUT(feats);
  CHECK_INPUT(coors);

  const int NDim = coors.size(1);
  const int num_input = feats.size(0);
  const int num_feats = feats.size(1);

  auto coors_id = at::empty({num_input}, coors.options().dtype(torch::kInt64));
  auto coor_space_dim = coors.max_values(0) + 1;
  auto coors_map_sorted =
      at::empty({num_input}, coors.options().dtype(torch::kInt32));
  auto coors_map =
      at::empty({num_input}, coors.options().dtype(torch::kInt32));
  auto num_coors = at::zeros({1}, coors.options().dtype(torch::kInt32));

  AT_DISPATCH_INTEGRAL_TYPES(
      coors.scalar_type(), "coors_id_kernel", ([&] {
        dim3 blocks(std::min(at::cuda::ATenCeilDiv(num_input, threadsPerBlock),
                             maxGridDim));
        dim3 threads(threadsPerBlock);
        coors_id_kernel<<<blocks, threads>>>(
            coors.data_ptr<scalar_t>(), coor_space_dim.data_ptr<scalar_t>(),
            coors_id.data_ptr<int64_t>(), num_input, NDim);
      }));
  AT_CUDA_CHECK(cudaGetLastError());

  auto coors_id_argsort = coors_id.argsort();

  AT_DISPATCH_INTEGRAL_TYPES(
      coors_id_argsort.scalar_type(), "coors_map_init_kernel", ([&] {
        dim3 blocks(std::min(at::cuda::ATenCeilDiv(num_input, threadsPerBlock),
                             maxGridDim));
        dim3 threads(threadsPerBlock);
        coors_map_init_kernel<<<blocks, threads>>>(
            coors_id.data_ptr<int64_t>(), coors_id_argsort.data_ptr<scalar_t>(),
            coors_map_sorted.data_ptr<int32_t>(), num_input);
      }));
  AT_CUDA_CHECK(cudaGetLastError());

  coors_map_sorted = coors_map_sorted.cumsum(0, torch::kInt32);
  coors_map.index_put_(coors_id_argsort, coors_map_sorted);

  const int num_coors_cpu =
      coors_map_sorted[-1].cpu().data_ptr<int32_t>()[0] + 1;
  auto out_coors = at::empty({num_coors_cpu, NDim}, coors.options());
  auto reduced_feats =
      at::empty({num_coors_cpu, num_feats}, feats.options());
  auto reduce_count =
      at::zeros({num_coors_cpu}, coors.options().dtype(torch::kInt32));

  AT_DISPATCH_FLOATING_TYPES(
      feats.scalar_type(), "feats_reduce_kernel", ([&] {
        using F_t = scalar_t;
        AT_DISPATCH_INTEGRAL_TYPES(
            coors.scalar_type(), "feats_reduce_kernel", ([&] {
              using I_t = scalar_t;

              if (reduce_type == reduce_t::MAX)
                reduced_feats.fill_(-std::numeric_limits<F_t>::infinity());
              else
                reduced_feats.fill_(static_cast<F_t>(0));

              dim3 blocks(
                  std::min(at::cuda::ATenCeilDiv(num_input, threadsPerBlock),
                           maxGridDim));
              dim3 threads(threadsPerBlock);
              feats_reduce_kernel<<<blocks, threads>>>(
                  feats.data_ptr<F_t>(), coors.data_ptr<I_t>(),
                  coors_map.data_ptr<int32_t>(),
                  reduce_count.data_ptr<int32_t>(),
                  reduced_feats.data_ptr<F_t>(), out_coors.data_ptr<I_t>(),
                  num_input, num_feats, NDim, reduce_type);
              if (reduce_type == reduce_t::MEAN)
                reduced_feats /=
                    reduce_count.unsqueeze(-1).to(reduced_feats.dtype());
            }));
      }));
  AT_CUDA_CHECK(cudaGetLastError());

  return {reduced_feats, out_coors, coors_map, reduce_count};
}

void dynamic_point_to_voxel_backward_gpu(
    at::Tensor &grad_feats, const at::Tensor &grad_reduced_feats,
    const at::Tensor &feats, const at::Tensor &reduced_feats,
    const at::Tensor &coors_map, const at::Tensor &reduce_count,
    const reduce_t reduce_type) {
  CHECK_INPUT(grad_feats);
  CHECK_INPUT(grad_reduced_feats);
  CHECK_INPUT(feats);
  CHECK_INPUT(reduced_feats);
  CHECK_INPUT(coors_map);
  CHECK_INPUT(reduce_count);

  const int num_input = feats.size(0);
  const int num_reduced = reduced_feats.size(0);
  const int num_feats = feats.size(1);

  grad_feats.fill_(0);
  // copy voxel grad to points

  if (reduce_type == reduce_t::MEAN || reduce_type == reduce_t::SUM) {
    AT_DISPATCH_FLOATING_TYPES(
        grad_reduced_feats.scalar_type(), "add_reduce_traceback_grad_kernel",
        ([&] {
          dim3 blocks
              (std::min(at::cuda::ATenCeilDiv(num_input, threadsPerBlock),
                        maxGridDim));
          dim3 threads(threadsPerBlock);
          add_reduce_traceback_grad_kernel<<<blocks, threads>>>(
              grad_feats.data_ptr<scalar_t>(),
              grad_reduced_feats.data_ptr<scalar_t>(),
              coors_map.data_ptr<int32_t>(), reduce_count.data_ptr<int32_t>(),
              num_input, num_feats, reduce_type);
        }));
    AT_CUDA_CHECK(cudaGetLastError());
  } else {
    auto reduce_from = at::full({num_reduced, num_feats}, num_input,
                                coors_map.options().dtype(torch::kInt32));
    AT_DISPATCH_FLOATING_TYPES(
        grad_reduced_feats.scalar_type(),
        "max_reduce_traceback_scatter_idx_kernel", ([&] {
          dim3 blocks
              (std::min(at::cuda::ATenCeilDiv(num_input, threadsPerBlock),
                        maxGridDim));
          dim3 threads(threadsPerBlock);
          max_reduce_traceback_scatter_idx_kernel<<<blocks, threads>>>(
              feats.data_ptr<scalar_t>(), reduced_feats.data_ptr<scalar_t>(),
              reduce_from.data_ptr<int32_t>(), coors_map.data_ptr<int32_t>(),
              num_input, num_feats);
        }));
    AT_CUDA_CHECK(cudaGetLastError());

    AT_DISPATCH_FLOATING_TYPES(
        grad_reduced_feats.scalar_type(),
        "max_reduce_traceback_scatter_idx_kernel", ([&] {
          dim3 blocks(
              std::min(at::cuda::ATenCeilDiv(num_reduced, threadsPerBlock),
                       maxGridDim));
          dim3 threads(threadsPerBlock);
          max_reduce_scatter_grad_kernel<<<blocks, threads>>>(
              grad_feats.data_ptr<scalar_t>(),
              grad_reduced_feats.data_ptr<scalar_t>(),
              reduce_from.data_ptr<int32_t>(), num_reduced, num_feats);
        }));
    AT_CUDA_CHECK(cudaGetLastError());
  }
  return;
}

} // namespace voxelization