Debug distCUDA2

demo/gs_train.py

@@ -141,7 +141,12 @@ def train(
     checkpoint_iterations = args.checkpoint_iterations 
     debug_from = args.debug_from
 
-    dist2 = torch.clamp_min(distCUDA2(torch.randn((90804, 3)).float().cuda()), 0.0000001)
+    pcd = torch.randn((90804, 3)).float().cuda()
+    print("pcd: ", pcd.shape, pcd.dtype, pcd.min(), pcd.max(), pcd.device)
+    print("distCUDA2: ", distCUDA2(pcd.cpu()))
+    print("distCUDA2: ", distCUDA2(pcd.cuda()))
+
+    dist2 = torch.clamp_min(distCUDA2(pcd.cuda()), 0.0000001)
     print("dist2.shape: ", dist2.shape)