Merge pull request #73 from JYYCaN/rc4main

add new npu op roiaware_pool3d && fix npu op scatter_points bug

mmcv/ops/csrc/pytorch/npu/roiaware_pool3d.cpp

@@ -0,0 +1,86 @@
+#include "pytorch_npu_helper.hpp"
+using namespace NPU_NAME_SPACE;
+using namespace std;
+
+void roiaware_pool3d_forward_npu(int boxes_num, int pts_num, int channels,
+                                 int max_pts_each_voxel, int out_x, int out_y,
+                                 int out_z, const Tensor rois, const Tensor pts,
+                                 const Tensor pts_feature, Tensor argmax,
+                                 Tensor pts_idx_of_voxels,
+                                 Tensor pooled_features, int pool_method) {
+  at::Tensor rois_cast = rois;
+  at::Tensor pts_cast = pts;
+  at::Tensor pts_feature_cast = pts_feature;
+  at::Tensor pooled_features_cast = pooled_features;
+
+  auto dtype = rois.dtype();
+  if (dtype == at::kHalf) {
+    rois_cast = rois_cast.to(at::kFloat);
+    pts_cast = pts_cast.to(at::kFloat);
+    pts_feature_cast = pts_feature_cast.to(at::kFloat);
+    pooled_features_cast = pooled_features_cast.to(at::kFloat);
+  }
+
+  EXEC_NPU_CMD(aclnnRoiawarePool3d, rois_cast, pts_cast, pts_feature_cast,
+               pool_method, max_pts_each_voxel, out_x, out_y, out_z, argmax,
+               pts_idx_of_voxels, pooled_features_cast);
+
+  if (dtype == at::kHalf) {
+    pooled_features_cast = pooled_features_cast.to(at::kHalf);
+  }
+
+  pooled_features.copy_(pooled_features_cast);
+}
+
+void roiaware_pool3d_backward_npu(int boxes_num, int out_x, int out_y,
+                                  int out_z, int channels,
+                                  int max_pts_each_voxel,
+                                  const Tensor pts_idx_of_voxels,
+                                  const Tensor argmax, const Tensor grad_out,
+                                  Tensor grad_in, int pool_method)
+{
+  int32_t npoints = grad_in.size(0);
+
+  auto dtype = grad_out.dtype();
+  at::Tensor grad_out_cast = grad_out;
+  at::Tensor grad_in_cast = grad_in;
+
+  if (dtype == at::kHalf) {
+    grad_out_cast = grad_out.to(at::kFloat);
+    grad_in_cast = grad_in_cast.to(at::kFloat);
+  }
+
+  if (pool_method == 0) {
+    // maxpool3d
+    EXEC_NPU_CMD(aclnnRoiawareMaxpool3dGrad, argmax, grad_out_cast, boxes_num,
+        out_x, out_y, out_z, channels, npoints, grad_in_cast);
+  } else if (pool_method == 1) {
+    // avgpool3d
+    EXEC_NPU_CMD(aclnnRoiawareAvgpool3dGrad, pts_idx_of_voxels, grad_out_cast,
+                boxes_num, out_x, out_y, out_z, channels, npoints,
+                max_pts_each_voxel, grad_in_cast);
+  }
+
+  if (dtype == at::kHalf) {
+    grad_in_cast = grad_in_cast.to(at::kHalf);
+  }
+
+  grad_in.copy_(grad_in_cast);
+}
+
+void roiaware_pool3d_forward_impl(int boxes_num, int pts_num, int channels,
+                                  int max_pts_each_voxel, int out_x, int out_y,
+                                  int out_z, const Tensor rois,
+                                  const Tensor pts, const Tensor pts_feature,
+                                  Tensor argmax, Tensor pts_idx_of_voxels,
+                                  Tensor pooled_features, int pool_method);
+
+void roiaware_pool3d_backward_impl(int boxes_num, int out_x, int out_y,
+                                   int out_z, int channels,
+                                   int max_pts_each_voxel,
+                                   const Tensor pts_idx_of_voxels,
+                                   const Tensor argmax, const Tensor grad_out,
+                                   Tensor grad_in, int pool_method);
+
+REGISTER_NPU_IMPL(roiaware_pool3d_forward_impl, roiaware_pool3d_forward_npu);
+REGISTER_NPU_IMPL(roiaware_pool3d_backward_impl, roiaware_pool3d_backward_npu);

mmcv/ops/scatter_points.py

@@ -38,15 +38,15 @@ def forward(ctx: Any,
         """
         ctx.device = feats.device.type
         if ctx.device == 'npu':
-            import ads_c
-            voxel_idx = ads_c.point_to_voxel(coors, [], [], 'XYZ')
-            unique_res = ads_c.unique_voxel(voxel_idx)
+            import mx_driving._C
+            voxel_idx = mx_driving._C.point_to_voxel(coors, [], [], 'XYZ')
+            unique_res = mx_driving._C.unique_voxel(voxel_idx)
             num_voxels, uniqued_voxel_idx, prefix_sum, \
                 argsort_coor, _ = unique_res
             voxel_coors = \
-                ads_c.voxel_to_point(uniqued_voxel_idx, [], [], 'XYZ')
+                mx_driving._C.voxel_to_point(uniqued_voxel_idx, [], [], 'XYZ')
             voxel_feats, \
-                compare_mask = ads_c.npu_dynamic_scatter(feats, coors,
+                compare_mask = mx_driving._C.npu_dynamic_scatter(feats, coors,
                                                          prefix_sum,
                                                          argsort_coor,
                                                          num_voxels,