What's the best practice to build 3c2e tensors?

[John-zzh]

Hi developers,

I'm trying to implement TDDFT-ris with density fitting. In general, I need T_pqP,
which is contraced through

" Puv, up,vq - >   Ppq" , eri3c, coeff_p, coeff_q

coeff_p, coeff_q referes to either occupied or virtual blocks of coefficeint matrix, C[:,:nocc], C[:,nocc:]. (could be both)
P being the auxbf dimension, smallest dimension here (compared to u v p q)

How to implement it on GPU?

CPU code is relatively easy, I can build full eri3c given enough memory. Or slice it along the P dimension with limied memory.
I belive GPU memory is more valuable, so I guess there must be some memory-efficeint way to code up this procesure?

[wxj6000]

Hi Zehao,
Yes, GPU memory is quite limited. In GPU4PySCF, CDERI is stored in a sparse format, and is automatically transferred to CPU memory. When it is needed, the sparse CDERI is unpacked into CDERI slices for usage.

Here is an example of performing Lji,jk->Lki.
https://github.com/pyscf/gpu4pyscf/blob/master/gpu4pyscf/df/df_jk.py#L293