Merge pull request #20 from takagi/remove-onnx-gather

Reduce Gather op when ONNX-exported

torch_dftd/functions/dftd3.py

@@ -109,14 +109,15 @@ def _getc6_impl(
     Zi: Tensor, Zj: Tensor, nci: Tensor, ncj: Tensor, c6ab: Tensor, k3: float = d3_k3
 ) -> Tensor:
     # gather the relevant entries from the table
-    # c6ab (95, 95, 5, 5, 3) --> c6ab_ (n_edges, 5, 5, 3)
-    c6ab_ = c6ab[Zi, Zj].type(nci.dtype)
-    # calculate c6 coefficients
-
-    # cn0, cn1, cn2 (n_edges, 5, 5)
-    cn0 = c6ab_[:, :, :, 0]
-    cn1 = c6ab_[:, :, :, 1]
-    cn2 = c6ab_[:, :, :, 2]
+    # c6ab (95, 95, 5, 5, 3) --> cni (9025, 5, 5, 1)
+    cn0, cn1, cn2 = c6ab.reshape(-1, 5, 5, 3).split(1, dim=3)
+    index = Zi * c6ab.size(1) + Zj
+
+    # cni (9025, 5, 5, 1) --> cni (n_edges, 5, 5)
+    cn0 = cn0.squeeze(dim=3)[index].type(nci.dtype)
+    cn1 = cn1.squeeze(dim=3)[index].type(nci.dtype)
+    cn2 = cn2.squeeze(dim=3)[index].type(nci.dtype)
+
     r = (cn1 - nci[:, None, None]) ** 2 + (cn2 - ncj[:, None, None]) ** 2
 
     n_edges = r.shape[0]