Accelerate angular AEV computation and reduce memory cost (#290)

* Accerate angular AEV computation and reduce memory cost

* reduce number of elementwise product

torchani/aev.py

@@ -9,11 +9,8 @@ from typing import Tuple
 # @torch.jit.script
 def cutoff_cosine(distances, cutoff):
     # type: (Tensor, float) -> Tensor
-    return torch.where(
-        distances <= cutoff,
-        0.5 * torch.cos(math.pi * distances / cutoff) + 0.5,
-        torch.zeros_like(distances)
-    )
+    # assuming all elements in distances are smaller than cutoff
+    return 0.5 * torch.cos(distances * (math.pi / cutoff)) + 0.5
 
 
 # @torch.jit.script
@@ -270,9 +267,8 @@ def compute_aev(species, coordinates, cell, shifts, triu_index, constants, sizes
     num_molecules = species.shape[0]
     num_atoms = species.shape[1]
     num_species_pairs = angular_length // angular_sublength
-    cutoff = max(Rcr, Rca)
 
-    atom_index1, atom_index2, shifts = neighbor_pairs(species == -1, coordinates, cell, shifts, cutoff)
+    atom_index1, atom_index2, shifts = neighbor_pairs(species == -1, coordinates, cell, shifts, Rcr)
     species = species.flatten()
     coordinates = coordinates.flatten(0, 1)
     species1 = species[atom_index1]
@@ -291,6 +287,15 @@ def compute_aev(species, coordinates, cell, shifts, triu_index, constants, sizes
     radial_aev.index_add_(0, index2, radial_terms_)
     radial_aev = radial_aev.reshape(num_molecules, num_atoms, radial_length)
 
+    # Rca is usually much smaller than Rcr, using neighbor list with cutoff=Rcr is a waste of resources
+    # Now we will get a smaller neighbor list that only cares about atoms with distances <= Rca
+    even_closer_indices = (distances <= Rca).nonzero().flatten()
+    atom_index1 = atom_index1.index_select(0, even_closer_indices)
+    atom_index2 = atom_index2.index_select(0, even_closer_indices)
+    species1 = species1.index_select(0, even_closer_indices)
+    species2 = species2.index_select(0, even_closer_indices)
+    vec = vec.index_select(0, even_closer_indices)
+
     # compute angular aev
     central_atom_index, pair_index1, pair_index2, sign1, sign2 = triple_by_molecule(atom_index1, atom_index2)
     vec1 = vec.index_select(0, pair_index1) * sign1.unsqueeze(1).to(vec.dtype)
@@ -338,6 +343,7 @@ class AEVComputer(torch.nn.Module):
         super(AEVComputer, self).__init__()
         self.Rcr = Rcr
         self.Rca = Rca
+        assert Rca <= Rcr, "Current implementation of AEVComputer assumes Rca <= Rcr"
         # convert constant tensors to a ready-to-broadcast shape
         # shape convension (..., EtaR, ShfR)
         self.register_buffer('EtaR', EtaR.view(-1, 1))