From 76e870f17c284b8eba0c24dbedfbf3779c8647dc Mon Sep 17 00:00:00 2001
From: Wenhao Wu <79644370+wHao-Wu@users.noreply.github.com>
Date: Mon, 7 Mar 2022 21:08:36 +0800
Subject: [PATCH] [Fix] Fix bugs in Voxelization op (#1746)

* Fix bugs in  Voxelization op

* fix comments

* fix lint

* add comments
---
 mmcv/ops/csrc/pytorch/cpu/voxelization.cpp | 30 +++++++++++++++++-----
 1 file changed, 23 insertions(+), 7 deletions(-)

diff --git a/mmcv/ops/csrc/pytorch/cpu/voxelization.cpp b/mmcv/ops/csrc/pytorch/cpu/voxelization.cpp
index 25cc2b52c..a21f849a0 100644
--- a/mmcv/ops/csrc/pytorch/cpu/voxelization.cpp
+++ b/mmcv/ops/csrc/pytorch/cpu/voxelization.cpp
@@ -26,13 +26,22 @@ void dynamic_voxelize_forward_cpu_kernel(
       coor[ndim_minus_1 - j] = c;
     }
 
-    if (failed)
-      memset(&coors[i][0], -1, NDim * sizeof(T_int));
-    else
-      memcpy(&coors[i][0], &coor[0], NDim * sizeof(T_int));
+    // memcpy and memset will cause problem because of the memory distribution
+    // discontinuity of TensorAccessor, so here using loops to replace memcpy
+    // or memset
+    if (failed) {
+      for (int k = 0; k < NDim; ++k) {
+        coors[i][k] = -1;
+      }
+    } else {
+      for (int k = 0; k < NDim; ++k) {
+        coors[i][k] = coor[k];
+      }
+    }
   }
 
   delete[] coor;
+  return;
 }
 
 template <typename T, typename T_int>
@@ -72,14 +81,21 @@ void hard_voxelize_forward_cpu_kernel(
       voxel_num += 1;
 
       coor_to_voxelidx[coor[i][0]][coor[i][1]][coor[i][2]] = voxelidx;
-      memcpy(&coors[voxelidx][0], &coor[i][0], NDim * sizeof(T_int));
+      // memcpy will cause problem because of the memory distribution
+      // discontinuity of TensorAccessor, so here using loops to replace memcpy
+      for (int k = 0; k < NDim; ++k) {
+        coors[voxelidx][k] = coor[i][k];
+      }
     }
 
     // put points into voxel
     num = num_points_per_voxel[voxelidx];
     if (max_points == -1 || num < max_points) {
-      memcpy(&voxels[voxelidx][num][0], &points[i][0],
-             num_features * sizeof(T));
+      // memcpy will cause problem because of the memory distribution
+      // discontinuity of TensorAccessor, so here using loops to replace memcpy
+      for (int k = 0; k < num_features; ++k) {
+        voxels[voxelidx][num][k] = points[i][k];
+      }
       num_points_per_voxel[voxelidx] += 1;
     }
   }