GPU-accelerate vppnl_nuc_grad for GTH pseudopotentials#740
Conversation
CUDA kernels for int1e_r2_origi, int1e_r4_origi and ip2 derivatives. GPU-accelerate vppnl_nuc_grad via merged-Cell cross-basis integrals and cupy contraction. Includes test suite covering C, Si, Fe
|
Why not also replace the non-local piece of pp in Fock calculation, at gpu4pyscf/gpu4pyscf/pbc/dft/multigrid.py Line 948 in ad09bb0 gpu4pyscf/gpu4pyscf/pbc/dft/multigrid_v2.py Line 1153 in ad09bb0 gpu4pyscf/gpu4pyscf/pbc/df/rsdf_builder.py Line 932 in ad09bb0 gpu4pyscf/gpu4pyscf/pbc/df/aft.py Line 102 in ad09bb0 |
… remove stale comments
@henryw7 - Sure, though - I'd prefer to do so in a follow up PR to keep current scope, if that's alright. |
* Add _contract_ppnl_gpu and get_pp_nl_gpu in gpu4pyscf/pbc/gto/pseudo/pp_int.py. * Use get_pp_nl_gpu in multigrid.get_pp, multigrid_v2.get_pp, rsdf_builder.get_pp, aft.get_pp. * Add gpu4pyscf/pbc/gto/pseudo/tests/test_pp_int.py covering both branches on C, Si, Fe.
* Add _contract_ppnl_gpu and get_pp_nl_gpu in gpu4pyscf/pbc/gto/pseudo/pp_int.py. * Use get_pp_nl_gpu in multigrid.get_pp, multigrid_v2.get_pp, rsdf_builder.get_pp, aft.get_pp. * Add gpu4pyscf/pbc/gto/pseudo/tests/test_pp_int.py covering both branches on C, Si, Fe.
3 participants
Add CUDA kernels for int1e_r2_origi, int1e_r4_origi and ip2 derivatives in overlap.cu. GPU-accelerate vppnl_nuc_grad via merged-Cell cross-basis integrals and cupy contraction.
Tests added for C, Si, Fe (cross-basis integrals, full gradient vs CPU,
finite difference).