1.x branch fix hccl compile

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

@@ -24,7 +24,7 @@ void bbox_overlaps_npu(const Tensor bboxes1, const Tensor bboxes2, Tensor ious,
     bboxesFP32 = bboxesFP32.to(at::kFloat);
     gtboxesFP32 = gtboxesFP32.to(at::kFloat);
   }
-  c10::SmallVector<int64_t, SIZE> iousSize = {gtboxesFP32.size(0),
+  c10::SmallVector<int64_t, 8> iousSize = {gtboxesFP32.size(0),
                                               bboxesFP32.size(0)};
   if (aligned) {
     iousSize = {gtboxesFP32.size(0), 1};

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

@@ -15,7 +15,7 @@ Tensor fused_bias_leakyrelu_npu(const Tensor &input, const Tensor &bias,
   if (grad == 0) {
     auto input_size = input.sizes();
     int input_length = input_size.size();
-    c10::SmallVector<int64_t, SIZE> input_size_tmp;
+    c10::SmallVector<int64_t, 8> input_size_tmp;
     for (uint64_t i = 0; i < input_size.size(); i++) {
       input_size_tmp.emplace_back(input_size[i]);
     }

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

@@ -13,7 +13,7 @@ Tensor nms_npu(Tensor boxes, Tensor scores, float iou_threshold, int offset) {
       at::empty({}, boxes.options().dtype(at::kFloat)).fill_(0);
   at::Tensor max_outputsize_y =
       at::empty({}, boxes.options().dtype(at::kInt)).fill_(boxes.size(0));
-  c10::SmallVector<int64_t, SIZE> outputsize = {boxes.size(0)};
+  c10::SmallVector<int64_t, 8> outputsize = {boxes.size(0)};
   at::Tensor output =
       at::empty(outputsize, boxes.options().dtype(at::kInt)).fill_(-1);
   OpCommand cmd;

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

@@ -11,7 +11,7 @@ Tensor nms_rotated_npu(const Tensor dets, const Tensor scores,
     detsCast = detsCast.to(at::kFloat);
     scoresCast = scoresCast.to(at::kFloat);
   }
-  c10::SmallVector<int64_t, SIZE> selectedIndexSize = {dets.size(0)};
+  c10::SmallVector<int64_t, 8> selectedIndexSize = {dets.size(0)};
   at::Tensor selectedBox = at::empty_like(dets);
   at::Tensor selectedIndex =
       at::empty(selectedIndexSize, dets.options().dtype(at::kInt));

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

@@ -41,7 +41,7 @@ void roi_align_backward_npu(Tensor grad_output, Tensor rois, Tensor argmax_y,
     roi_end_mode = 0;
   }
   auto shape = grad_input.sizes();
-  c10::SmallVector<int64_t, SIZE> xdiff_shape;
+  c10::SmallVector<int64_t, 8> xdiff_shape;
   for (uint64_t i = 0; i < shape.size(); i++) {
     xdiff_shape.emplace_back(shape[i]);
   }

mmcv/ops/csrc/pytorch/npu/roipoint_pool3d_forward.cpp

@@ -12,12 +12,12 @@ void roipoint_pool3d_forward_impl_npu(int batch_size, int pts_num,
                                       Tensor pooled_empty_flag) {
   auto points_trans = xyz.transpose(1, 2).contiguous();
   auto point_features_trans = pts_feature.transpose(1, 2).contiguous();
-  c10::SmallVector<int64_t, SIZE> features_trans_size = {
+  c10::SmallVector<int64_t, 8> features_trans_size = {
       xyz.size(0), boxes3d.size(1), xyz.size(2) + pts_feature.size(2),
       sampled_pts_num};
   at::Tensor pooled_features_trans =
       at::empty(features_trans_size, xyz.options());
-  c10::SmallVector<int64_t, SIZE> empty_flag_size = {boxes3d.size(0),
+  c10::SmallVector<int64_t, 8> empty_flag_size = {boxes3d.size(0),
                                                      boxes3d.size(1)};
   EXEC_NPU_CMD(aclnnRoipointPool3dForward, points_trans, point_features_trans,
                boxes3d, sampled_pts_num, pooled_features_trans,

setup.py

@@ -446,6 +446,11 @@ def get_mluops_version(file_path):
                         'torch_npu').submodule_search_locations[0] +
                     '/include/third_party/acl/inc'
                 ]
+                extra_compile_args['cxx'] += [
+                    '-I' + importlib.util.find_spec(
+                        'torch_npu').submodule_search_locations[0] +
+                    '/include/third_party/hccl/inc'
+                ]
                 define_macros += [('MMCV_WITH_NPU', None)]
                 extension = NpuExtension
                 if parse_version(torch.__version__) < parse_version('2.1.0'):