Skip to content

pt: fix se_a type_one_side performance degradation #3361

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Feb 29, 2024
Merged

pt: fix se_a type_one_side performance degradation #3361

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 15 additions & 4 deletions deepmd/pt/model/descriptor/se_a.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -472,23 +472,34 @@ def forward(
if self.type_one_side:
ii = embedding_idx
# torch.jit is not happy with slice(None)
ti_mask = torch.ones(nfnl, dtype=torch.bool, device=dmatrix.device)
# ti_mask = torch.ones(nfnl, dtype=torch.bool, device=dmatrix.device)
# applying a mask seems to cause performance degradation
ti_mask = None
else:
# ti: center atom type, ii: neighbor type...
ii = embedding_idx // self.ntypes
ti = embedding_idx % self.ntypes
ti_mask = atype.ravel().eq(ti)
# nfnl x nt
mm = exclude_mask[ti_mask, self.sec[ii] : self.sec[ii + 1]]
if ti_mask is not None:
mm = exclude_mask[ti_mask, self.sec[ii] : self.sec[ii + 1]]
else:
mm = exclude_mask[:, self.sec[ii] : self.sec[ii + 1]]
# nfnl x nt x 4
rr = dmatrix[ti_mask, self.sec[ii] : self.sec[ii + 1], :]
if ti_mask is not None:
rr = dmatrix[ti_mask, self.sec[ii] : self.sec[ii + 1], :]
else:
rr = dmatrix[:, self.sec[ii] : self.sec[ii + 1], :]
rr = rr * mm[:, :, None]
ss = rr[:, :, :1]
# nfnl x nt x ng
gg = ll.forward(ss)
# nfnl x 4 x ng
gr = torch.matmul(rr.permute(0, 2, 1), gg)
xyz_scatter[ti_mask] += gr
if ti_mask is not None:
xyz_scatter[ti_mask] += gr
else:
xyz_scatter += gr

xyz_scatter /= self.nnei
xyz_scatter_1 = xyz_scatter.permute(0, 2, 1)
Expand Down