Merge pull request #192 from sshaoshuai/master

Release OpenPCDet v0.3.0

pcdet/ops/pointnet2/pointnet2_batch/src/group_points.cpp

+/*
+batch version of point grouping, modified from the original implementation of official PointNet++ codes.
+Written by Shaoshuai Shi
+All Rights Reserved 2018.
+*/
+
+
+#include <torch/serialize/tensor.h>
+#include <cuda.h>
+#include <cuda_runtime_api.h>
+#include <vector>
+#include <THC/THC.h>
+#include "group_points_gpu.h"
+
+extern THCState *state;
+
+
+int group_points_grad_wrapper_fast(int b, int c, int n, int npoints, int nsample, 
+    at::Tensor grad_out_tensor, at::Tensor idx_tensor, at::Tensor grad_points_tensor) {
+
+    float *grad_points = grad_points_tensor.data<float>();
+    const int *idx = idx_tensor.data<int>();
+    const float *grad_out = grad_out_tensor.data<float>();
+
+    group_points_grad_kernel_launcher_fast(b, c, n, npoints, nsample, grad_out, idx, grad_points);
+    return 1;
+}
+
+
+int group_points_wrapper_fast(int b, int c, int n, int npoints, int nsample, 
+    at::Tensor points_tensor, at::Tensor idx_tensor, at::Tensor out_tensor) {
+
+    const float *points = points_tensor.data<float>();
+    const int *idx = idx_tensor.data<int>();
+    float *out = out_tensor.data<float>();
+
+    group_points_kernel_launcher_fast(b, c, n, npoints, nsample, points, idx, out);
+    return 1;
+}

pcdet/ops/pointnet2/pointnet2_batch/src/group_points_gpu.cu

+/*
+batch version of point grouping, modified from the original implementation of official PointNet++ codes.
+Written by Shaoshuai Shi
+All Rights Reserved 2018.
+*/
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "cuda_utils.h"
+#include "group_points_gpu.h"
+
+
+__global__ void group_points_grad_kernel_fast(int b, int c, int n, int npoints, int nsample, 
+    const float *__restrict__ grad_out, const int *__restrict__ idx, float *__restrict__ grad_points) {
+    // grad_out: (B, C, npoints, nsample)
+    // idx: (B, npoints, nsample)
+    // output:
+    //      grad_points: (B, C, N)
+    int bs_idx = blockIdx.z;
+    int c_idx = blockIdx.y;
+    int index = blockIdx.x * blockDim.x + threadIdx.x;
+    int pt_idx = index / nsample;
+    if (bs_idx >= b || c_idx >= c || pt_idx >= npoints) return;
+
+    int sample_idx = index % nsample;
+    grad_out += bs_idx * c * npoints * nsample + c_idx * npoints * nsample + pt_idx * nsample + sample_idx;
+    idx += bs_idx * npoints * nsample + pt_idx * nsample + sample_idx; 
+    
+    atomicAdd(grad_points + bs_idx * c * n + c_idx * n + idx[0] , grad_out[0]);
+}
+
+void group_points_grad_kernel_launcher_fast(int b, int c, int n, int npoints, int nsample, 
+    const float *grad_out, const int *idx, float *grad_points) {
+    // grad_out: (B, C, npoints, nsample)
+    // idx: (B, npoints, nsample)
+    // output:
+    //      grad_points: (B, C, N)
+    cudaError_t err;
+    dim3 blocks(DIVUP(npoints * nsample, THREADS_PER_BLOCK), c, b);  // blockIdx.x(col), blockIdx.y(row)
+    dim3 threads(THREADS_PER_BLOCK);
+
+    group_points_grad_kernel_fast<<<blocks, threads>>>(b, c, n, npoints, nsample, grad_out, idx, grad_points);
+
+    err = cudaGetLastError();
+    if (cudaSuccess != err) {
+        fprintf(stderr, "CUDA kernel failed : %s\n", cudaGetErrorString(err));
+        exit(-1);
+    }
+}
+
+
+__global__ void group_points_kernel_fast(int b, int c, int n, int npoints, int nsample, 
+    const float *__restrict__ points, const int *__restrict__ idx, float *__restrict__ out) {
+    // points: (B, C, N)
+    // idx: (B, npoints, nsample)
+    // output:
+    //      out: (B, C, npoints, nsample)
+    int bs_idx = blockIdx.z;
+    int c_idx = blockIdx.y;
+    int index = blockIdx.x * blockDim.x + threadIdx.x;
+    int pt_idx = index / nsample;
+    if (bs_idx >= b || c_idx >= c || pt_idx >= npoints) return;
+
+    int sample_idx = index % nsample;
+
+    idx += bs_idx * npoints * nsample + pt_idx * nsample + sample_idx; 
+    int in_idx = bs_idx * c * n + c_idx * n + idx[0];
+    int out_idx = bs_idx * c * npoints * nsample + c_idx * npoints * nsample + pt_idx * nsample + sample_idx;
+
+    out[out_idx] = points[in_idx];
+}
+
+
+void group_points_kernel_launcher_fast(int b, int c, int n, int npoints, int nsample, 
+    const float *points, const int *idx, float *out) {
+    // points: (B, C, N)
+    // idx: (B, npoints, nsample)
+    // output:
+    //      out: (B, C, npoints, nsample)
+    cudaError_t err;
+    dim3 blocks(DIVUP(npoints * nsample, THREADS_PER_BLOCK), c, b);  // blockIdx.x(col), blockIdx.y(row)
+    dim3 threads(THREADS_PER_BLOCK);
+
+    group_points_kernel_fast<<<blocks, threads>>>(b, c, n, npoints, nsample, points, idx, out);
+    // cudaDeviceSynchronize();  // for using printf in kernel function
+    err = cudaGetLastError();
+    if (cudaSuccess != err) {
+        fprintf(stderr, "CUDA kernel failed : %s\n", cudaGetErrorString(err));
+        exit(-1);
+    }
+}

pcdet/ops/pointnet2/pointnet2_batch/src/group_points_gpu.h

+#ifndef _GROUP_POINTS_GPU_H
+#define _GROUP_POINTS_GPU_H
+
+#include <torch/serialize/tensor.h>
+#include <cuda.h>
+#include <cuda_runtime_api.h>
+#include <vector>
+
+
+int group_points_wrapper_fast(int b, int c, int n, int npoints, int nsample, 
+    at::Tensor points_tensor, at::Tensor idx_tensor, at::Tensor out_tensor);
+
+void group_points_kernel_launcher_fast(int b, int c, int n, int npoints, int nsample, 
+    const float *points, const int *idx, float *out);
+
+int group_points_grad_wrapper_fast(int b, int c, int n, int npoints, int nsample, 
+    at::Tensor grad_out_tensor, at::Tensor idx_tensor, at::Tensor grad_points_tensor);
+
+void group_points_grad_kernel_launcher_fast(int b, int c, int n, int npoints, int nsample, 
+    const float *grad_out, const int *idx, float *grad_points);
+
+#endif

pcdet/ops/pointnet2/pointnet2_batch/src/interpolate.cpp

+/*
+batch version of point interpolation, modified from the original implementation of official PointNet++ codes.
+Written by Shaoshuai Shi
+All Rights Reserved 2018.
+*/
+
+
+#include <torch/serialize/tensor.h>
+#include <vector>
+#include <THC/THC.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <cuda.h>
+#include <cuda_runtime_api.h>
+#include "interpolate_gpu.h"
+
+extern THCState *state;
+
+
+void three_nn_wrapper_fast(int b, int n, int m, at::Tensor unknown_tensor, 
+    at::Tensor known_tensor, at::Tensor dist2_tensor, at::Tensor idx_tensor) {
+    const float *unknown = unknown_tensor.data<float>();
+    const float *known = known_tensor.data<float>();
+    float *dist2 = dist2_tensor.data<float>();
+    int *idx = idx_tensor.data<int>();
+
+    three_nn_kernel_launcher_fast(b, n, m, unknown, known, dist2, idx);
+}
+
+
+void three_interpolate_wrapper_fast(int b, int c, int m, int n,
+                         at::Tensor points_tensor,
+                         at::Tensor idx_tensor,
+                         at::Tensor weight_tensor,
+                         at::Tensor out_tensor) {
+
+    const float *points = points_tensor.data<float>();
+    const float *weight = weight_tensor.data<float>();
+    float *out = out_tensor.data<float>();
+    const int *idx = idx_tensor.data<int>();
+
+    three_interpolate_kernel_launcher_fast(b, c, m, n, points, idx, weight, out);
+}
+
+void three_interpolate_grad_wrapper_fast(int b, int c, int n, int m,
+                            at::Tensor grad_out_tensor,
+                            at::Tensor idx_tensor,
+                            at::Tensor weight_tensor,
+                            at::Tensor grad_points_tensor) {
+
+    const float *grad_out = grad_out_tensor.data<float>();
+    const float *weight = weight_tensor.data<float>();
+    float *grad_points = grad_points_tensor.data<float>();
+    const int *idx = idx_tensor.data<int>();
+
+    three_interpolate_grad_kernel_launcher_fast(b, c, n, m, grad_out, idx, weight, grad_points);
+}

pcdet/ops/pointnet2/pointnet2_batch/src/interpolate_gpu.cu

+/*
+batch version of point interpolation, modified from the original implementation of official PointNet++ codes.
+Written by Shaoshuai Shi
+All Rights Reserved 2018.
+*/
+
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "cuda_utils.h"
+#include "interpolate_gpu.h"
+
+
+__global__ void three_nn_kernel_fast(int b, int n, int m, const float *__restrict__ unknown, 
+    const float *__restrict__ known, float *__restrict__ dist2, int *__restrict__ idx) {
+    // unknown: (B, N, 3)
+    // known: (B, M, 3)
+    // output: 
+    //      dist2: (B, N, 3)
+    //      idx: (B, N, 3)
+    
+    int bs_idx = blockIdx.y;
+    int pt_idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (bs_idx >= b || pt_idx >= n) return;
+
+    unknown += bs_idx * n * 3 + pt_idx * 3;
+    known += bs_idx * m * 3;
+    dist2 += bs_idx * n * 3 + pt_idx * 3;
+    idx += bs_idx * n * 3 + pt_idx * 3;
+
+    float ux = unknown[0];
+    float uy = unknown[1];
+    float uz = unknown[2];
+
+    double best1 = 1e40, best2 = 1e40, best3 = 1e40;
+    int besti1 = 0, besti2 = 0, besti3 = 0;
+    for (int k = 0; k < m; ++k) {
+        float x = known[k * 3 + 0];
+        float y = known[k * 3 + 1];
+        float z = known[k * 3 + 2];
+        float d = (ux - x) * (ux - x) + (uy - y) * (uy - y) + (uz - z) * (uz - z);
+        if (d < best1) {
+            best3 = best2; besti3 = besti2;
+            best2 = best1; besti2 = besti1;
+            best1 = d; besti1 = k;
+        } 
+        else if (d < best2) {
+            best3 = best2; besti3 = besti2;
+            best2 = d; besti2 = k;
+        } 
+        else if (d < best3) {
+            best3 = d; besti3 = k;
+        }
+    }
+    dist2[0] = best1; dist2[1] = best2; dist2[2] = best3;
+    idx[0] = besti1; idx[1] = besti2; idx[2] = besti3;
+}
+
+
+void three_nn_kernel_launcher_fast(int b, int n, int m, const float *unknown, 
+    const float *known, float *dist2, int *idx) {
+    // unknown: (B, N, 3)
+    // known: (B, M, 3)
+    // output: 
+    //      dist2: (B, N, 3)
+    //      idx: (B, N, 3)
+
+    cudaError_t err;
+    dim3 blocks(DIVUP(n, THREADS_PER_BLOCK), b);  // blockIdx.x(col), blockIdx.y(row)
+    dim3 threads(THREADS_PER_BLOCK);
+
+    three_nn_kernel_fast<<<blocks, threads>>>(b, n, m, unknown, known, dist2, idx);
+
+    err = cudaGetLastError();
+    if (cudaSuccess != err) {
+        fprintf(stderr, "CUDA kernel failed : %s\n", cudaGetErrorString(err));
+        exit(-1);
+    }
+}
+
+
+__global__ void three_interpolate_kernel_fast(int b, int c, int m, int n, const float *__restrict__ points, 
+    const int *__restrict__ idx, const float *__restrict__ weight, float *__restrict__ out) {
+    // points: (B, C, M)
+    // idx: (B, N, 3)
+    // weight: (B, N, 3)
+    // output:
+    //      out: (B, C, N)
+
+    int bs_idx = blockIdx.z;
+    int c_idx = blockIdx.y;
+    int pt_idx = blockIdx.x * blockDim.x + threadIdx.x;
+
+    if (bs_idx >= b || c_idx >= c || pt_idx >= n) return;
+
+    weight += bs_idx * n * 3 + pt_idx * 3;
+    points += bs_idx * c * m + c_idx * m;
+    idx += bs_idx * n * 3 + pt_idx * 3;
+    out += bs_idx * c * n + c_idx * n;
+
+    out[pt_idx] = weight[0] * points[idx[0]] + weight[1] * points[idx[1]] + weight[2] * points[idx[2]];
+}
+
+void three_interpolate_kernel_launcher_fast(int b, int c, int m, int n, 
+    const float *points, const int *idx, const float *weight, float *out) {
+    // points: (B, C, M)
+    // idx: (B, N, 3)
+    // weight: (B, N, 3)
+    // output:
+    //      out: (B, C, N)
+
+    cudaError_t err;
+    dim3 blocks(DIVUP(n, THREADS_PER_BLOCK), c, b);  // blockIdx.x(col), blockIdx.y(row)
+    dim3 threads(THREADS_PER_BLOCK);
+    three_interpolate_kernel_fast<<<blocks, threads>>>(b, c, m, n, points, idx, weight, out);
+
+    err = cudaGetLastError();
+    if (cudaSuccess != err) {
+        fprintf(stderr, "CUDA kernel failed : %s\n", cudaGetErrorString(err));
+        exit(-1);
+    }
+}
+
+
+__global__ void three_interpolate_grad_kernel_fast(int b, int c, int n, int m, const float *__restrict__ grad_out, 
+    const int *__restrict__ idx, const float *__restrict__ weight, float *__restrict__ grad_points) {
+    // grad_out: (B, C, N)
+    // weight: (B, N, 3)
+    // output:
+    //      grad_points: (B, C, M)
+
+    int bs_idx = blockIdx.z;
+    int c_idx = blockIdx.y;
+    int pt_idx = blockIdx.x * blockDim.x + threadIdx.x;
+
+    if (bs_idx >= b || c_idx >= c || pt_idx >= n) return;
+    
+    grad_out += bs_idx * c * n + c_idx * n + pt_idx;
+    weight += bs_idx * n * 3 + pt_idx * 3;
+    grad_points += bs_idx * c * m + c_idx * m;
+    idx += bs_idx * n * 3 + pt_idx * 3;
+
+
+    atomicAdd(grad_points + idx[0], grad_out[0] * weight[0]);
+    atomicAdd(grad_points + idx[1], grad_out[0] * weight[1]);
+    atomicAdd(grad_points + idx[2], grad_out[0] * weight[2]);
+}
+
+void three_interpolate_grad_kernel_launcher_fast(int b, int c, int n, int m, const float *grad_out, 
+    const int *idx, const float *weight, float *grad_points) {
+    // grad_out: (B, C, N)
+    // weight: (B, N, 3)
+    // output:
+    //      grad_points: (B, C, M)
+
+    cudaError_t err;
+    dim3 blocks(DIVUP(n, THREADS_PER_BLOCK), c, b);  // blockIdx.x(col), blockIdx.y(row)
+    dim3 threads(THREADS_PER_BLOCK);
+    three_interpolate_grad_kernel_fast<<<blocks, threads>>>(b, c, n, m, grad_out, idx, weight, grad_points);
+
+    err = cudaGetLastError();
+    if (cudaSuccess != err) {
+        fprintf(stderr, "CUDA kernel failed : %s\n", cudaGetErrorString(err));
+        exit(-1);
+    }
+}

pcdet/ops/pointnet2/pointnet2_batch/src/interpolate_gpu.h

+#ifndef _INTERPOLATE_GPU_H
+#define _INTERPOLATE_GPU_H
+
+#include <torch/serialize/tensor.h>
+#include<vector>
+#include <cuda.h>
+#include <cuda_runtime_api.h>
+
+
+void three_nn_wrapper_fast(int b, int n, int m, at::Tensor unknown_tensor, 
+  at::Tensor known_tensor, at::Tensor dist2_tensor, at::Tensor idx_tensor);
+
+void three_nn_kernel_launcher_fast(int b, int n, int m, const float *unknown,
+	const float *known, float *dist2, int *idx);
+
+
+void three_interpolate_wrapper_fast(int b, int c, int m, int n, at::Tensor points_tensor, 
+    at::Tensor idx_tensor, at::Tensor weight_tensor, at::Tensor out_tensor);
+
+void three_interpolate_kernel_launcher_fast(int b, int c, int m, int n, 
+    const float *points, const int *idx, const float *weight, float *out);
+
+
+void three_interpolate_grad_wrapper_fast(int b, int c, int n, int m, at::Tensor grad_out_tensor, 
+    at::Tensor idx_tensor, at::Tensor weight_tensor, at::Tensor grad_points_tensor);
+
+void three_interpolate_grad_kernel_launcher_fast(int b, int c, int n, int m, const float *grad_out, 
+    const int *idx, const float *weight, float *grad_points);
+
+#endif

pcdet/ops/pointnet2/pointnet2_batch/src/pointnet2_api.cpp

+#include <torch/serialize/tensor.h>
+#include <torch/extension.h>
+
+#include "ball_query_gpu.h"
+#include "group_points_gpu.h"
+#include "sampling_gpu.h"
+#include "interpolate_gpu.h"
+
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("ball_query_wrapper", &ball_query_wrapper_fast, "ball_query_wrapper_fast");
+
+    m.def("group_points_wrapper", &group_points_wrapper_fast, "group_points_wrapper_fast");
+    m.def("group_points_grad_wrapper", &group_points_grad_wrapper_fast, "group_points_grad_wrapper_fast");
+
+    m.def("gather_points_wrapper", &gather_points_wrapper_fast, "gather_points_wrapper_fast");
+    m.def("gather_points_grad_wrapper", &gather_points_grad_wrapper_fast, "gather_points_grad_wrapper_fast");
+
+    m.def("furthest_point_sampling_wrapper", &furthest_point_sampling_wrapper, "furthest_point_sampling_wrapper");
+    
+    m.def("three_nn_wrapper", &three_nn_wrapper_fast, "three_nn_wrapper_fast");
+    m.def("three_interpolate_wrapper", &three_interpolate_wrapper_fast, "three_interpolate_wrapper_fast");
+    m.def("three_interpolate_grad_wrapper", &three_interpolate_grad_wrapper_fast, "three_interpolate_grad_wrapper_fast");
+}

pcdet/ops/pointnet2/pointnet2_batch/src/sampling.cpp

+/*
+batch version of point sampling and gathering, modified from the original implementation of official PointNet++ codes.
+Written by Shaoshuai Shi
+All Rights Reserved 2018.
+*/
+
+
+#include <torch/serialize/tensor.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <vector>
+#include <THC/THC.h>
+
+#include "sampling_gpu.h"
+
+extern THCState *state;
+
+
+int gather_points_wrapper_fast(int b, int c, int n, int npoints, 
+    at::Tensor points_tensor, at::Tensor idx_tensor, at::Tensor out_tensor){
+    const float *points = points_tensor.data<float>();
+    const int *idx = idx_tensor.data<int>();
+    float *out = out_tensor.data<float>();
+
+    gather_points_kernel_launcher_fast(b, c, n, npoints, points, idx, out);
+    return 1;
+}
+
+
+int gather_points_grad_wrapper_fast(int b, int c, int n, int npoints, 
+    at::Tensor grad_out_tensor, at::Tensor idx_tensor, at::Tensor grad_points_tensor) {
+
+    const float *grad_out = grad_out_tensor.data<float>();
+    const int *idx = idx_tensor.data<int>();
+    float *grad_points = grad_points_tensor.data<float>();
+
+    gather_points_grad_kernel_launcher_fast(b, c, n, npoints, grad_out, idx, grad_points);
+    return 1;
+}
+
+
+int furthest_point_sampling_wrapper(int b, int n, int m, 
+    at::Tensor points_tensor, at::Tensor temp_tensor, at::Tensor idx_tensor) {
+
+    const float *points = points_tensor.data<float>();
+    float *temp = temp_tensor.data<float>();
+    int *idx = idx_tensor.data<int>();
+
+    furthest_point_sampling_kernel_launcher(b, n, m, points, temp, idx);
+    return 1;
+}

pcdet/ops/pointnet2/pointnet2_batch/src/sampling_gpu.cu

+/*
+batch version of point sampling and gathering, modified from the original implementation of official PointNet++ codes.
+Written by Shaoshuai Shi
+All Rights Reserved 2018.
+*/
+
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "cuda_utils.h"
+#include "sampling_gpu.h"
+
+
+__global__ void gather_points_kernel_fast(int b, int c, int n, int m, 
+    const float *__restrict__ points, const int *__restrict__ idx, float *__restrict__ out) {
+    // points: (B, C, N)
+    // idx: (B, M)
+    // output:
+    //      out: (B, C, M)
+
+    int bs_idx = blockIdx.z;
+    int c_idx = blockIdx.y;
+    int pt_idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (bs_idx >= b || c_idx >= c || pt_idx >= m) return;
+
+    out += bs_idx * c * m + c_idx * m + pt_idx;
+    idx += bs_idx * m + pt_idx;
+    points += bs_idx * c * n + c_idx * n;
+    out[0] = points[idx[0]];
+}
+
+void gather_points_kernel_launcher_fast(int b, int c, int n, int npoints, 
+    const float *points, const int *idx, float *out) {
+    // points: (B, C, N)
+    // idx: (B, npoints)
+    // output:
+    //      out: (B, C, npoints)
+
+    cudaError_t err;
+    dim3 blocks(DIVUP(npoints, THREADS_PER_BLOCK), c, b);  // blockIdx.x(col), blockIdx.y(row)
+    dim3 threads(THREADS_PER_BLOCK);
+
+    gather_points_kernel_fast<<<blocks, threads>>>(b, c, n, npoints, points, idx, out);
+
+    err = cudaGetLastError();
+    if (cudaSuccess != err) {
+        fprintf(stderr, "CUDA kernel failed : %s\n", cudaGetErrorString(err));
+        exit(-1);
+    }
+}
+
+__global__ void gather_points_grad_kernel_fast(int b, int c, int n, int m, const float *__restrict__ grad_out, 
+    const int *__restrict__ idx, float *__restrict__ grad_points) {
+    // grad_out: (B, C, M)
+    // idx: (B, M)
+    // output:
+    //      grad_points: (B, C, N)
+
+    int bs_idx = blockIdx.z;
+    int c_idx = blockIdx.y;
+    int pt_idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (bs_idx >= b || c_idx >= c || pt_idx >= m) return;
+
+    grad_out += bs_idx * c * m + c_idx * m + pt_idx;
+    idx += bs_idx * m + pt_idx;
+    grad_points += bs_idx * c * n + c_idx * n;
+
+    atomicAdd(grad_points + idx[0], grad_out[0]);
+}
+
+void gather_points_grad_kernel_launcher_fast(int b, int c, int n, int npoints, 
+    const float *grad_out, const int *idx, float *grad_points) {
+    // grad_out: (B, C, npoints)
+    // idx: (B, npoints)
+    // output:
+    //      grad_points: (B, C, N)
+
+    cudaError_t err;
+    dim3 blocks(DIVUP(npoints, THREADS_PER_BLOCK), c, b);  // blockIdx.x(col), blockIdx.y(row)
+    dim3 threads(THREADS_PER_BLOCK);
+
+    gather_points_grad_kernel_fast<<<blocks, threads>>>(b, c, n, npoints, grad_out, idx, grad_points);
+
+    err = cudaGetLastError();
+    if (cudaSuccess != err) {
+        fprintf(stderr, "CUDA kernel failed : %s\n", cudaGetErrorString(err));
+        exit(-1);
+    }
+}
+
+
+__device__ void __update(float *__restrict__ dists, int *__restrict__ dists_i, int idx1, int idx2){
+    const float v1 = dists[idx1], v2 = dists[idx2];
+    const int i1 = dists_i[idx1], i2 = dists_i[idx2];
+    dists[idx1] = max(v1, v2);
+    dists_i[idx1] = v2 > v1 ? i2 : i1;
+}
+
+template <unsigned int block_size>
+__global__ void furthest_point_sampling_kernel(int b, int n, int m, 
+    const float *__restrict__ dataset, float *__restrict__ temp, int *__restrict__ idxs) {
+    // dataset: (B, N, 3)
+    // tmp: (B, N)
+    // output:
+    //      idx: (B, M)
+
+    if (m <= 0) return;
+    __shared__ float dists[block_size];
+    __shared__ int dists_i[block_size];
+
+    int batch_index = blockIdx.x;
+    dataset += batch_index * n * 3;
+    temp += batch_index * n;
+    idxs += batch_index * m;
+
+    int tid = threadIdx.x;
+    const int stride = block_size;
+
+    int old = 0;
+    if (threadIdx.x == 0)
+    idxs[0] = old;
+
+    __syncthreads();
+    for (int j = 1; j < m; j++) {
+    int besti = 0;
+    float best = -1;
+    float x1 = dataset[old * 3 + 0];
+    float y1 = dataset[old * 3 + 1];
+    float z1 = dataset[old * 3 + 2];
+    for (int k = tid; k < n; k += stride) {
+        float x2, y2, z2;
+        x2 = dataset[k * 3 + 0];
+        y2 = dataset[k * 3 + 1];
+        z2 = dataset[k * 3 + 2];
+        // float mag = (x2 * x2) + (y2 * y2) + (z2 * z2);
+        // if (mag <= 1e-3)
+        // continue;
+
+        float d = (x2 - x1) * (x2 - x1) + (y2 - y1) * (y2 - y1) + (z2 - z1) * (z2 - z1);
+        float d2 = min(d, temp[k]);
+        temp[k] = d2;
+        besti = d2 > best ? k : besti;
+        best = d2 > best ? d2 : best;
+    }
+    dists[tid] = best;
+    dists_i[tid] = besti;
+    __syncthreads();
+
+    if (block_size >= 1024) {
+        if (tid < 512) {
+            __update(dists, dists_i, tid, tid + 512);
+        }
+        __syncthreads();
+    }
+
+    if (block_size >= 512) {
+        if (tid < 256) {
+            __update(dists, dists_i, tid, tid + 256);
+        }
+        __syncthreads();
+    }
+    if (block_size >= 256) {
+        if (tid < 128) {
+            __update(dists, dists_i, tid, tid + 128);
+        }
+        __syncthreads();
+    }
+    if (block_size >= 128) {
+        if (tid < 64) {
+            __update(dists, dists_i, tid, tid + 64);
+        }
+        __syncthreads();
+    }
+    if (block_size >= 64) {
+        if (tid < 32) {
+            __update(dists, dists_i, tid, tid + 32);
+        }
+        __syncthreads();
+    }
+    if (block_size >= 32) {
+        if (tid < 16) {
+            __update(dists, dists_i, tid, tid + 16);
+        }
+        __syncthreads();
+    }
+    if (block_size >= 16) {
+        if (tid < 8) {
+            __update(dists, dists_i, tid, tid + 8);
+        }
+        __syncthreads();
+    }
+    if (block_size >= 8) {
+        if (tid < 4) {
+            __update(dists, dists_i, tid, tid + 4);
+        }
+        __syncthreads();
+    }
+    if (block_size >= 4) {
+        if (tid < 2) {
+            __update(dists, dists_i, tid, tid + 2);
+        }
+        __syncthreads();
+    }
+    if (block_size >= 2) {
+        if (tid < 1) {
+            __update(dists, dists_i, tid, tid + 1);
+        }
+        __syncthreads();
+    }
+
+    old = dists_i[0];
+    if (tid == 0)
+        idxs[j] = old;
+    }
+}
+
+void furthest_point_sampling_kernel_launcher(int b, int n, int m, 
+    const float *dataset, float *temp, int *idxs) {
+    // dataset: (B, N, 3)
+    // tmp: (B, N)
+    // output:
+    //      idx: (B, M)
+
+    cudaError_t err;
+    unsigned int n_threads = opt_n_threads(n);
+
+    switch (n_threads) {
+        case 1024:
+        furthest_point_sampling_kernel<1024><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
+        case 512:
+        furthest_point_sampling_kernel<512><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
+        case 256:
+        furthest_point_sampling_kernel<256><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
+        case 128:
+        furthest_point_sampling_kernel<128><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
+        case 64:
+        furthest_point_sampling_kernel<64><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
+        case 32:
+        furthest_point_sampling_kernel<32><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
+        case 16:
+        furthest_point_sampling_kernel<16><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
+        case 8:
+        furthest_point_sampling_kernel<8><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
+        case 4:
+        furthest_point_sampling_kernel<4><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
+        case 2:
+        furthest_point_sampling_kernel<2><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
+        case 1:
+        furthest_point_sampling_kernel<1><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
+        default:
+        furthest_point_sampling_kernel<512><<<b, n_threads>>>(b, n, m, dataset, temp, idxs);
+    }
+
+    err = cudaGetLastError();
+    if (cudaSuccess != err) {
+        fprintf(stderr, "CUDA kernel failed : %s\n", cudaGetErrorString(err));
+        exit(-1);
+    }
+}

pcdet/ops/pointnet2/pointnet2_batch/src/sampling_gpu.h

+#ifndef _SAMPLING_GPU_H
+#define _SAMPLING_GPU_H
+
+#include <torch/serialize/tensor.h>
+#include <ATen/cuda/CUDAContext.h>
+#include<vector>
+
+
+int gather_points_wrapper_fast(int b, int c, int n, int npoints, 
+    at::Tensor points_tensor, at::Tensor idx_tensor, at::Tensor out_tensor);
+
+void gather_points_kernel_launcher_fast(int b, int c, int n, int npoints, 
+    const float *points, const int *idx, float *out);
+
+
+int gather_points_grad_wrapper_fast(int b, int c, int n, int npoints, 
+    at::Tensor grad_out_tensor, at::Tensor idx_tensor, at::Tensor grad_points_tensor);
+
+void gather_points_grad_kernel_launcher_fast(int b, int c, int n, int npoints, 
+    const float *grad_out, const int *idx, float *grad_points);
+
+
+int furthest_point_sampling_wrapper(int b, int n, int m, 
+    at::Tensor points_tensor, at::Tensor temp_tensor, at::Tensor idx_tensor);
+
+void furthest_point_sampling_kernel_launcher(int b, int n, int m, 
+    const float *dataset, float *temp, int *idxs);
+
+#endif

pcdet/ops/pointnet2/pointnet2_stack/pointnet2_modules.py

+from typing import List
+
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+
 from . import pointnet2_utils
-from typing import List
 
 
 class StackSAModuleMSG(nn.Module):
@@ -89,3 +91,47 @@ class StackSAModuleMSG(nn.Module):
 
         return new_xyz, new_features
 
+
+class StackPointnetFPModule(nn.Module):
+    def __init__(self, *, mlp: List[int]):
+        """
+        Args:
+            mlp: list of int
+        """
+        super().__init__()
+        shared_mlps = []
+        for k in range(len(mlp) - 1):
+            shared_mlps.extend([
+                nn.Conv2d(mlp[k], mlp[k + 1], kernel_size=1, bias=False),
+                nn.BatchNorm2d(mlp[k + 1]),
+                nn.ReLU()
+            ])
+        self.mlp = nn.Sequential(*shared_mlps)
+
+    def forward(self, unknown, unknown_batch_cnt, known, known_batch_cnt, unknown_feats=None, known_feats=None):
+        """
+        Args:
+            unknown: (N1 + N2 ..., 3)
+            known: (M1 + M2 ..., 3)
+            unknow_feats: (N1 + N2 ..., C1)
+            known_feats: (M1 + M2 ..., C2)
+
+        Returns:
+            new_features: (N1 + N2 ..., C_out)
+        """
+        dist, idx = pointnet2_utils.three_nn(unknown, unknown_batch_cnt, known, known_batch_cnt)
+        dist_recip = 1.0 / (dist + 1e-8)
+        norm = torch.sum(dist_recip, dim=-1, keepdim=True)
+        weight = dist_recip / norm
+
+        interpolated_feats = pointnet2_utils.three_interpolate(known_feats, idx, weight)
+
+        if unknown_feats is not None:
+            new_features = torch.cat([interpolated_feats, unknown_feats], dim=1)  # (N1 + N2 ..., C2 + C1)
+        else:
+            new_features = interpolated_feats
+        new_features = new_features.permute(1, 0)[None, :, :, None]  # (1, C, N1 + N2 ..., 1)
+        new_features = self.mlp(new_features)
+
+        new_features = new_features.squeeze(dim=0).squeeze(dim=-1).permute(1, 0)  # (N1 + N2 ..., C)
+        return new_features

pcdet/ops/pointnet2/pointnet2_stack/pointnet2_utils.py

 import torch
-from torch.autograd import Variable
-from torch.autograd import Function
 import torch.nn as nn
+from torch.autograd import Function, Variable
 
 from . import pointnet2_stack_cuda as pointnet2
 
@@ -185,5 +184,83 @@ class FurthestPointSampling(Function):
 furthest_point_sample = FurthestPointSampling.apply
 
 
+class ThreeNN(Function):
+    @staticmethod
+    def forward(ctx, unknown, unknown_batch_cnt, known, known_batch_cnt):
+        """
+        Args:
+            ctx:
+            unknown: (N1 + N2..., 3)
+            unknown_batch_cnt: (batch_size), [N1, N2, ...]
+            known: (M1 + M2..., 3)
+            known_batch_cnt: (batch_size), [M1, M2, ...]
+
+        Returns:
+            dist: (N1 + N2 ..., 3)  l2 distance to the three nearest neighbors
+            idx: (N1 + N2 ..., 3)  index of the three nearest neighbors, range [0, M1+M2+...]
+        """
+        assert unknown.shape.__len__() == 2 and unknown.shape[1] == 3
+        assert known.shape.__len__() == 2 and known.shape[1] == 3
+        assert unknown_batch_cnt.__len__() == known_batch_cnt.__len__()
+
+        dist2 = unknown.new_zeros(unknown.shape)
+        idx = unknown_batch_cnt.new_zeros(unknown.shape).int()
+
+        pointnet2.three_nn_wrapper(
+            unknown.contiguous(), unknown_batch_cnt.contiguous(),
+            known.contiguous(), known_batch_cnt.contiguous(), dist2, idx
+        )
+        return torch.sqrt(dist2), idx
+
+    @staticmethod
+    def backward(ctx, a=None, b=None):
+        return None, None
+
+
+three_nn = ThreeNN.apply
+
+
+class ThreeInterpolate(Function):
+
+    @staticmethod
+    def forward(ctx, features: torch.Tensor, idx: torch.Tensor, weight: torch.Tensor):
+        """
+        Args:
+            ctx:
+            features: (M1 + M2 ..., C)
+            idx: [N1 + N2 ..., 3]
+            weight: [N1 + N2 ..., 3]
+
+        Returns:
+            out_tensor: (N1 + N2 ..., C)
+        """
+        assert idx.shape[0] == weight.shape[0] and idx.shape[1] == weight.shape[1] == 3
+
+        ctx.three_interpolate_for_backward = (idx, weight, features.shape[0])
+        output = features.new_zeros((idx.shape[0], features.shape[1]))
+        pointnet2.three_interpolate_wrapper(features.contiguous(), idx.contiguous(), weight.contiguous(), output)
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_out: torch.Tensor):
+        """
+        Args:
+            ctx:
+            grad_out: (N1 + N2 ..., C)
+
+        Returns:
+            grad_features: (M1 + M2 ..., C)
+        """
+        idx, weight, M = ctx.three_interpolate_for_backward
+        grad_features = grad_out.new_zeros((M, grad_out.shape[1]))
+        pointnet2.three_interpolate_grad_wrapper(
+            grad_out.contiguous(), idx.contiguous(), weight.contiguous(), grad_features
+        )
+        return grad_features, None, None
+
+
+three_interpolate = ThreeInterpolate.apply
+
+
 if __name__ == '__main__':
     pass

pcdet/ops/pointnet2/pointnet2_stack/src/ball_query.cpp

@@ -14,8 +14,18 @@ All Rights Reserved 2019-2020.
 
 extern THCState *state;
 
-#define CHECK_CUDA(x) AT_CHECK(x.type().is_cuda(), #x, " must be a CUDAtensor ")
-#define CHECK_CONTIGUOUS(x) AT_CHECK(x.is_contiguous(), #x, " must be contiguous ")
+#define CHECK_CUDA(x) do { \
+  if (!x.type().is_cuda()) { \
+    fprintf(stderr, "%s must be CUDA tensor at %s:%d\n", #x, __FILE__, __LINE__); \
+    exit(-1); \
+  } \
+} while (0)
+#define CHECK_CONTIGUOUS(x) do { \
+  if (!x.is_contiguous()) { \
+    fprintf(stderr, "%s must be contiguous tensor at %s:%d\n", #x, __FILE__, __LINE__); \
+    exit(-1); \
+  } \
+} while (0)
 #define CHECK_INPUT(x) CHECK_CUDA(x);CHECK_CONTIGUOUS(x)
 
 int ball_query_wrapper_stack(int B, int M, float radius, int nsample,

pcdet/ops/pointnet2/pointnet2_stack/src/group_points.cpp

@@ -13,8 +13,18 @@ All Rights Reserved 2019-2020.
 #include "group_points_gpu.h"
 
 extern THCState *state;
-#define CHECK_CUDA(x) AT_CHECK(x.type().is_cuda(), #x, " must be a CUDAtensor ")
-#define CHECK_CONTIGUOUS(x) AT_CHECK(x.is_contiguous(), #x, " must be contiguous ")
+#define CHECK_CUDA(x) do { \
+  if (!x.type().is_cuda()) { \
+    fprintf(stderr, "%s must be CUDA tensor at %s:%d\n", #x, __FILE__, __LINE__); \
+    exit(-1); \
+  } \
+} while (0)
+#define CHECK_CONTIGUOUS(x) do { \
+  if (!x.is_contiguous()) { \
+    fprintf(stderr, "%s must be contiguous tensor at %s:%d\n", #x, __FILE__, __LINE__); \
+    exit(-1); \
+  } \
+} while (0)
 #define CHECK_INPUT(x) CHECK_CUDA(x);CHECK_CONTIGUOUS(x)
 
 

pcdet/ops/pointnet2/pointnet2_stack/src/interpolate.cpp

+/*
+Stacked-batch-data version of point interpolation, modified from the original implementation of official PointNet++ codes.
+Written by Shaoshuai Shi
+All Rights Reserved 2019-2020.
+*/
+
+
+#include <torch/serialize/tensor.h>
+#include <vector>
+#include <THC/THC.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <cuda.h>
+#include <cuda_runtime_api.h>
+#include "interpolate_gpu.h"
+
+extern THCState *state;
+
+#define CHECK_CUDA(x) do { \
+  if (!x.type().is_cuda()) { \
+    fprintf(stderr, "%s must be CUDA tensor at %s:%d\n", #x, __FILE__, __LINE__); \
+    exit(-1); \
+  } \
+} while (0)
+#define CHECK_CONTIGUOUS(x) do { \
+  if (!x.is_contiguous()) { \
+    fprintf(stderr, "%s must be contiguous tensor at %s:%d\n", #x, __FILE__, __LINE__); \
+    exit(-1); \
+  } \
+} while (0)
+#define CHECK_INPUT(x) CHECK_CUDA(x);CHECK_CONTIGUOUS(x)
+
+
+void three_nn_wrapper_stack(at::Tensor unknown_tensor, 
+    at::Tensor unknown_batch_cnt_tensor, at::Tensor known_tensor, 
+    at::Tensor known_batch_cnt_tensor, at::Tensor dist2_tensor, at::Tensor idx_tensor){
+    // unknown: (N1 + N2 ..., 3)
+    // unknown_batch_cnt: (batch_size), [N1, N2, ...]
+    // known: (M1 + M2 ..., 3)
+    // known_batch_cnt: (batch_size), [M1, M2, ...]
+    // Return:
+    // dist: (N1 + N2 ..., 3)  l2 distance to the three nearest neighbors
+    // idx: (N1 + N2 ..., 3)  index of the three nearest neighbors
+    CHECK_INPUT(unknown_tensor);
+    CHECK_INPUT(unknown_batch_cnt_tensor);
+    CHECK_INPUT(known_tensor);
+    CHECK_INPUT(known_batch_cnt_tensor);
+    CHECK_INPUT(dist2_tensor);
+    CHECK_INPUT(idx_tensor);
+
+    int batch_size = unknown_batch_cnt_tensor.size(0);
+    int N = unknown_tensor.size(0);
+    int M = known_tensor.size(0);
+    const float *unknown = unknown_tensor.data<float>();
+    const int *unknown_batch_cnt = unknown_batch_cnt_tensor.data<int>();
+    const float *known = known_tensor.data<float>();
+    const int *known_batch_cnt = known_batch_cnt_tensor.data<int>();
+    float *dist2 = dist2_tensor.data<float>();
+    int *idx = idx_tensor.data<int>();
+
+    three_nn_kernel_launcher_stack(batch_size, N, M, unknown, unknown_batch_cnt, known, known_batch_cnt, dist2, idx);
+}
+
+
+void three_interpolate_wrapper_stack(at::Tensor features_tensor, 
+    at::Tensor idx_tensor, at::Tensor weight_tensor, at::Tensor out_tensor) {
+    // features_tensor: (M1 + M2 ..., C)
+    // idx_tensor: [N1 + N2 ..., 3]
+    // weight_tensor: [N1 + N2 ..., 3]
+    // Return:
+    // out_tensor: (N1 + N2 ..., C)
+    CHECK_INPUT(features_tensor);
+    CHECK_INPUT(idx_tensor);
+    CHECK_INPUT(weight_tensor);
+    CHECK_INPUT(out_tensor);
+
+    int N = out_tensor.size(0);
+    int channels = features_tensor.size(1);
+    const float *features = features_tensor.data<float>();
+    const float *weight = weight_tensor.data<float>();
+    const int *idx = idx_tensor.data<int>();
+    float *out = out_tensor.data<float>();
+
+    three_interpolate_kernel_launcher_stack(N, channels, features, idx, weight, out);
+}
+
+
+void three_interpolate_grad_wrapper_stack(at::Tensor grad_out_tensor, at::Tensor idx_tensor,
+    at::Tensor weight_tensor, at::Tensor grad_features_tensor) {
+    // grad_out_tensor: (N1 + N2 ..., C)
+    // idx_tensor: [N1 + N2 ..., 3]
+    // weight_tensor: [N1 + N2 ..., 3]
+    // Return:
+    // grad_features_tensor: (M1 + M2 ..., C)
+    CHECK_INPUT(grad_out_tensor);
+    CHECK_INPUT(idx_tensor);
+    CHECK_INPUT(weight_tensor);
+    CHECK_INPUT(grad_features_tensor);
+
+    int N = grad_out_tensor.size(0);
+    int channels = grad_out_tensor.size(1);
+    const float *grad_out = grad_out_tensor.data<float>();
+    const float *weight = weight_tensor.data<float>();
+    const int *idx = idx_tensor.data<int>();
+    float *grad_features = grad_features_tensor.data<float>();
+    
+    // printf("N=%d, channels=%d\n", N, channels);
+    three_interpolate_grad_kernel_launcher_stack(N, channels, grad_out, idx, weight, grad_features);
+}
\ No newline at end of file

pcdet/ops/pointnet2/pointnet2_stack/src/interpolate_gpu.cu

+/*
+Stacked-batch-data version of point interpolation, modified from the original implementation of official PointNet++ codes.
+Written by Shaoshuai Shi
+All Rights Reserved 2019-2020.
+*/
+
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "cuda_utils.h"
+#include "interpolate_gpu.h"
+
+
+__global__ void three_nn_kernel_stack(int batch_size, int N, int M, const float *unknown, 
+    const int *unknown_batch_cnt, const float *known, const int *known_batch_cnt,
+    float *dist2, int *idx) {
+    // unknown: (N1 + N2 ..., 3)
+    // unknown_batch_cnt: (batch_size), [N1, N2, ...]
+    // known: (M1 + M2 ..., 3)
+    // known_batch_cnt: (batch_size), [M1, M2, ...]
+    // Return:
+    // dist: (N1 + N2 ..., 3)  l2 distance to the three nearest neighbors
+    // idx: (N1 + N2 ..., 3)  index of the three nearest neighbors
+
+    int pt_idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (pt_idx >= N) return;
+
+    int bs_idx = 0, pt_cnt = unknown_batch_cnt[0];
+    for (int k = 1; k < batch_size; k++){
+        if (pt_idx < pt_cnt) break;
+        pt_cnt += unknown_batch_cnt[k];
+        bs_idx = k;
+    }
+
+    int cur_num_known_points = known_batch_cnt[bs_idx];
+
+    int known_batch_start_idx = 0;
+    for (int k = 0; k < bs_idx; k++) known_batch_start_idx += known_batch_cnt[k];
+
+    known += known_batch_start_idx * 3;
+    unknown += pt_idx * 3;
+    dist2 += pt_idx * 3;
+    idx += pt_idx * 3;
+
+    float ux = unknown[0];
+    float uy = unknown[1];
+    float uz = unknown[2];
+
+    double best1 = 1e40, best2 = 1e40, best3 = 1e40;
+    int besti1 = 0, besti2 = 0, besti3 = 0;
+    for (int k = 0; k < cur_num_known_points; ++k) {
+        float x = known[k * 3 + 0];
+        float y = known[k * 3 + 1];
+        float z = known[k * 3 + 2];
+        float d = (ux - x) * (ux - x) + (uy - y) * (uy - y) + (uz - z) * (uz - z);
+        if (d < best1) {
+            best3 = best2; besti3 = besti2;
+            best2 = best1; besti2 = besti1;
+            best1 = d; besti1 = k;
+        } 
+        else if (d < best2) {
+            best3 = best2; besti3 = besti2;
+            best2 = d; besti2 = k;
+        } 
+        else if (d < best3) {
+            best3 = d; besti3 = k;
+        }
+    }
+    dist2[0] = best1; dist2[1] = best2; dist2[2] = best3;
+    idx[0] = besti1 + known_batch_start_idx; 
+    idx[1] = besti2 + known_batch_start_idx; 
+    idx[2] = besti3 + known_batch_start_idx;
+}
+
+
+void three_nn_kernel_launcher_stack(int batch_size, int N, int M, const float *unknown, 
+    const int *unknown_batch_cnt, const float *known, const int *known_batch_cnt,
+    float *dist2, int *idx) {
+    // unknown: (N1 + N2 ..., 3)
+    // unknown_batch_cnt: (batch_size), [N1, N2, ...]
+    // known: (M1 + M2 ..., 3)
+    // known_batch_cnt: (batch_size), [M1, M2, ...]
+    // Return:
+    // dist: (N1 + N2 ..., 3)  l2 distance to the three nearest neighbors
+    // idx: (N1 + N2 ..., 3)  index of the three nearest neighbors
+
+    cudaError_t err;
+    dim3 blocks(DIVUP(N, THREADS_PER_BLOCK));  // blockIdx.x(col), blockIdx.y(row)
+    dim3 threads(THREADS_PER_BLOCK);
+
+    three_nn_kernel_stack<<<blocks, threads>>>(
+        batch_size, N, M, unknown, unknown_batch_cnt, 
+        known, known_batch_cnt, dist2, idx
+    );
+
+    err = cudaGetLastError();
+    if (cudaSuccess != err) {
+        fprintf(stderr, "CUDA kernel failed : %s\n", cudaGetErrorString(err));
+        exit(-1);
+    }
+}
+
+
+
+__global__ void three_interpolate_kernel_stack(int N, int channels, const float *features, 
+    const int *idx, const float *weight, float *out) {
+    // features: (M1 + M2 ..., C)
+    // idx: [N1 + N2 ..., 3]
+    // weight: [N1 + N2 ..., 3]
+    // Return:
+    // out: (N1 + N2 ..., C)
+
+    int c_idx = blockIdx.y;
+    int pt_idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (pt_idx >= N || c_idx >= channels) return;
+
+    weight += pt_idx * 3;
+    idx += pt_idx * 3;
+    out += pt_idx * channels + c_idx;
+
+    out[0] = weight[0] * features[idx[0] * channels + c_idx] + 
+        weight[1] * features[idx[1] * channels + c_idx] + 
+        weight[2] * features[idx[2] * channels + c_idx];
+}
+
+
+
+void three_interpolate_kernel_launcher_stack(int N, int channels,
+    const float *features, const int *idx, const float *weight, float *out) {
+    // features: (M1 + M2 ..., C)
+    // idx: [N1 + N2 ..., 3]
+    // weight: [N1 + N2 ..., 3]
+    // Return:
+    // out: (N1 + N2 ..., C)
+
+    cudaError_t err;
+    dim3 blocks(DIVUP(N, THREADS_PER_BLOCK), channels);
+    dim3 threads(THREADS_PER_BLOCK);
+    three_interpolate_kernel_stack<<<blocks, threads>>>(N, channels, features, idx, weight, out);
+
+    err = cudaGetLastError();
+    if (cudaSuccess != err) {
+        fprintf(stderr, "CUDA kernel failed : %s\n", cudaGetErrorString(err));
+        exit(-1);
+    }
+}
+
+
+__global__ void three_interpolate_grad_kernel_stack(int N, int channels, const float *grad_out, 
+    const int *idx, const float *weight, float *grad_features) {
+    // grad_out_tensor: (N1 + N2 ..., C)
+    // idx_tensor: [N1 + N2 ..., 3]
+    // weight_tensor: [N1 + N2 ..., 3]
+    // Return:
+    // grad_features_tensor: (M1 + M2 ..., C)
+
+    int c_idx = blockIdx.y;
+    int pt_idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (pt_idx >= N || c_idx >= channels) return;
+
+    grad_out += pt_idx * channels + c_idx;
+    weight += pt_idx * 3;
+    idx += pt_idx * 3;
+    
+    // printf("pt_idx=%d, c_idx=%d, idx=(%d, %d, %d), grad_out=%f\n", pt_idx, c_idx, idx[0], idx[1], idx[2], grad_out[0]);
+
+    atomicAdd(grad_features + idx[0] * channels + c_idx, grad_out[0] * weight[0]);
+    atomicAdd(grad_features + idx[1] * channels + c_idx, grad_out[0] * weight[1]);
+    atomicAdd(grad_features + idx[2] * channels + c_idx, grad_out[0] * weight[2]);
+}
+
+
+void three_interpolate_grad_kernel_launcher_stack(int N, int channels, const float *grad_out, 
+    const int *idx, const float *weight, float *grad_features) {
+    // grad_out_tensor: (N1 + N2 ..., C)
+    // idx_tensor: [N1 + N2 ..., 3]
+    // weight_tensor: [N1 + N2 ..., 3]
+    // Return:
+    // grad_features_tensor: (M1 + M2 ..., C)
+
+    cudaError_t err;
+    dim3 blocks(DIVUP(N, THREADS_PER_BLOCK), channels);  // blockIdx.x(col), blockIdx.y(row)
+    dim3 threads(THREADS_PER_BLOCK);
+    three_interpolate_grad_kernel_stack<<<blocks, threads>>>(
+        N, channels, grad_out, idx, weight, grad_features
+    );
+
+    err = cudaGetLastError();
+    if (cudaSuccess != err) {
+        fprintf(stderr, "CUDA kernel failed : %s\n", cudaGetErrorString(err));
+        exit(-1);
+    }
+}
\ No newline at end of file

pcdet/ops/pointnet2/pointnet2_stack/src/interpolate_gpu.h

+#ifndef _INTERPOLATE_GPU_H
+#define _INTERPOLATE_GPU_H
+
+#include <torch/serialize/tensor.h>
+#include<vector>
+#include <cuda.h>
+#include <cuda_runtime_api.h>
+
+
+void three_nn_wrapper_stack(at::Tensor unknown_tensor, 
+    at::Tensor unknown_batch_cnt_tensor, at::Tensor known_tensor, 
+    at::Tensor known_batch_cnt_tensor, at::Tensor dist2_tensor, at::Tensor idx_tensor);
+
+
+void three_interpolate_wrapper_stack(at::Tensor features_tensor, 
+    at::Tensor idx_tensor, at::Tensor weight_tensor, at::Tensor out_tensor);
+
+
+
+void three_interpolate_grad_wrapper_stack(at::Tensor grad_out_tensor, at::Tensor idx_tensor,
+    at::Tensor weight_tensor, at::Tensor grad_features_tensor);
+
+
+void three_nn_kernel_launcher_stack(int batch_size, int N, int M, const float *unknown, 
+    const int *unknown_batch_cnt, const float *known, const int *known_batch_cnt,
+    float *dist2, int *idx);
+
+
+void three_interpolate_kernel_launcher_stack(int N, int channels,
+    const float *features, const int *idx, const float *weight, float *out);
+
+
+
+void three_interpolate_grad_kernel_launcher_stack(int N, int channels, const float *grad_out, 
+    const int *idx, const float *weight, float *grad_features);
+
+
+
+#endif
\ No newline at end of file

pcdet/ops/pointnet2/pointnet2_stack/src/pointnet2_api.cpp

@@ -4,6 +4,7 @@
 #include "ball_query_gpu.h"
 #include "group_points_gpu.h"
 #include "sampling_gpu.h"
+#include "interpolate_gpu.h"
 
 
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
@@ -13,4 +14,8 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
 
     m.def("group_points_wrapper", &group_points_wrapper_stack, "group_points_wrapper_stack");
     m.def("group_points_grad_wrapper", &group_points_grad_wrapper_stack, "group_points_grad_wrapper_stack");
+
+    m.def("three_nn_wrapper", &three_nn_wrapper_stack, "three_nn_wrapper_stack");
+    m.def("three_interpolate_wrapper", &three_interpolate_wrapper_stack, "three_interpolate_wrapper_stack");
+    m.def("three_interpolate_grad_wrapper", &three_interpolate_grad_wrapper_stack, "three_interpolate_grad_wrapper_stack");
 }

pcdet/ops/pointnet2/pointnet2_stack/src/sampling.cpp

@@ -6,16 +6,32 @@
 #include "sampling_gpu.h"
 
 extern THCState *state;
+#define CHECK_CUDA(x) do { \
+  if (!x.type().is_cuda()) { \
+    fprintf(stderr, "%s must be CUDA tensor at %s:%d\n", #x, __FILE__, __LINE__); \
+    exit(-1); \
+  } \
+} while (0)
+#define CHECK_CONTIGUOUS(x) do { \
+  if (!x.is_contiguous()) { \
+    fprintf(stderr, "%s must be contiguous tensor at %s:%d\n", #x, __FILE__, __LINE__); \
+    exit(-1); \
+  } \
+} while (0)
+#define CHECK_INPUT(x) CHECK_CUDA(x);CHECK_CONTIGUOUS(x)
 
 
 int furthest_point_sampling_wrapper(int b, int n, int m,
     at::Tensor points_tensor, at::Tensor temp_tensor, at::Tensor idx_tensor) {
 
+    CHECK_INPUT(points_tensor);
+    CHECK_INPUT(temp_tensor);
+    CHECK_INPUT(idx_tensor);
+
     const float *points = points_tensor.data<float>();
     float *temp = temp_tensor.data<float>();
     int *idx = idx_tensor.data<int>();
 
-    cudaStream_t stream = THCState_getCurrentStream(state);
-    furthest_point_sampling_kernel_launcher(b, n, m, points, temp, idx, stream);
+    furthest_point_sampling_kernel_launcher(b, n, m, points, temp, idx);
     return 1;
 }

pcdet/ops/pointnet2/pointnet2_stack/src/sampling_gpu.cu

@@ -140,7 +140,7 @@ __global__ void furthest_point_sampling_kernel(int b, int n, int m,
 }
 
 void furthest_point_sampling_kernel_launcher(int b, int n, int m,
-    const float *dataset, float *temp, int *idxs, cudaStream_t stream) {
+    const float *dataset, float *temp, int *idxs) {
     // dataset: (B, N, 3)
     // tmp: (B, N)
     // output:
@@ -151,29 +151,29 @@ void furthest_point_sampling_kernel_launcher(int b, int n, int m,
 
     switch (n_threads) {
         case 1024:
-        furthest_point_sampling_kernel<1024><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs); break;
+        furthest_point_sampling_kernel<1024><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
         case 512:
-        furthest_point_sampling_kernel<512><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs); break;
+        furthest_point_sampling_kernel<512><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
         case 256:
-        furthest_point_sampling_kernel<256><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs); break;
+        furthest_point_sampling_kernel<256><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
         case 128:
-        furthest_point_sampling_kernel<128><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs); break;
+        furthest_point_sampling_kernel<128><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
         case 64:
-        furthest_point_sampling_kernel<64><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs); break;
+        furthest_point_sampling_kernel<64><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
         case 32:
-        furthest_point_sampling_kernel<32><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs); break;
+        furthest_point_sampling_kernel<32><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
         case 16:
-        furthest_point_sampling_kernel<16><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs); break;
+        furthest_point_sampling_kernel<16><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
         case 8:
-        furthest_point_sampling_kernel<8><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs); break;
+        furthest_point_sampling_kernel<8><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
         case 4:
-        furthest_point_sampling_kernel<4><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs); break;
+        furthest_point_sampling_kernel<4><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
         case 2:
-        furthest_point_sampling_kernel<2><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs); break;
+        furthest_point_sampling_kernel<2><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
         case 1:
-        furthest_point_sampling_kernel<1><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs); break;
+        furthest_point_sampling_kernel<1><<<b, n_threads>>>(b, n, m, dataset, temp, idxs); break;
         default:
-        furthest_point_sampling_kernel<512><<<b, n_threads, 0, stream>>>(b, n, m, dataset, temp, idxs);
+        furthest_point_sampling_kernel<512><<<b, n_threads>>>(b, n, m, dataset, temp, idxs);
     }
 
     err = cudaGetLastError();