fix three_interplote bug.

Author: momo609

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

@@ -12,17 +12,21 @@ void three_interpolate_forward_npu(int b, int c, int m, int n,
   TORCH_CHECK((originDtype == at::kFloat || originDtype == at::kHalf),
               "three_interpolate_forward ascend only support fp32 and fp16.");
 
-  auto point_c_trans = points.transpose(1, 2);
-
+  auto point_c_trans = points.transpose(1, 2).to(at::kFloat);
+  auto weight_cast = weight.to(at::kFloat);
+  auto out_cast = out.to(at::kFloat);
   OpCommand cmd;
   cmd.Name("ThreeInterpolate")
       .Input(point_c_trans)
       .Input(idx)
-      .Input(weight)
-      .Output(out)
+      .Input(weight_cast)
+      .Output(out_cast)
       .Run();
 
-  auto output = out.view({b, n, c}).transpose(1, 2);
+  if (originDtype == at::kHalf) {
+    out_cast = out_cast.to(at::kHalf);
+  }
+  auto output = out_cast.view({b, n, c}).transpose(1, 2);
   auto res = output.contiguous();
   out.copy_(res);
 }