[Question] Screening

[fishjojo]

This line https://github.com/pyscf/gpu4pyscf/blob/761f85ff22cf04d62167451dbf82df48caefb2e8/gpu4pyscf/lib/gvhf/nr_jk_driver.cu#L184 It seems to me should be inside the following loop https://github.com/pyscf/gpu4pyscf/blob/761f85ff22cf04d62167451dbf82df48caefb2e8/gpu4pyscf/lib/gvhf/nr_jk_driver.cu#L187-L188 and be

offsets.log_q_kl = log_q_kl + bas_kl0;

Maybe I missed something?

[wxj6000]

@fishjojo bas_ij and bas_kl in CUDA kernels are the global indices in each contraction pattern, not the local indices in each bin. So offsets.log_q_ij and offsets.log_q_kl are also should be zero-based correspondingly.

[fishjojo]

I see, I didn't notice bas_kl already has the offset included. Thanks for the explanation.

[fishjojo]

Actually log_q_kl uses the local index. https://github.com/pyscf/gpu4pyscf/blob/761f85ff22cf04d62167451dbf82df48caefb2e8/gpu4pyscf/lib/gvhf/g2e.cu#L64

This is what my original question about.

[wxj6000]

Actually log_q_kl uses the local index.

https://github.com/pyscf/gpu4pyscf/blob/761f85ff22cf04d62167451dbf82df48caefb2e8/gpu4pyscf/lib/gvhf/g2e.cu#L64

This is what my original question about.

This could be a bug.

[wxj6000]

@fishjojo The bug has been fixed.

[fishjojo]

Thanks. I think screening at the shell level is worth more investigation. The density matrix screening may also need more tests.

[wxj6000]

Thanks. I think screening at the shell level is worth more investigation. The density matrix screening may also need more tests.

I agree with the first point. The density matrix screening is quite classical, right?

[fishjojo]

I feel this https://github.com/pyscf/gpu4pyscf/blob/c58c7d57e61aec029d2f07a433a8be823656c14e/gpu4pyscf/scf/hf.py#L135 is a bit unconventional.

[wxj6000]

I feel this

https://github.com/pyscf/gpu4pyscf/blob/c58c7d57e61aec029d2f07a433a8be823656c14e/gpu4pyscf/scf/hf.py#L135

is a bit unconventional.

Well, it is certainly not the most elegant implementation. But the strategy is not something new. https://manual.q-chem.com/5.4/Ch5.S5.SS6.html

[fishjojo]

Maybe it's useful to add a check whether the input density is the density difference and 1e3 could be customizable.