See compute_snap_dgrad.py for a test that compares the dBi/dRj from compute snap (dbirjflag=1) to the sum of dBi/dRj from usual compute snap (dbirjflag=0).
The format of the global array from dbirjflag=1 is as follows.
The first N rows belong to bispectrum components for each atom, if we use bikflag=1. The first K columns correspond to each bispectrum coefficient. The final 3 columns contain reference force components for each atom.
The rows after the first N rows contain dBi/dRj values for all pairs. These values are arranged in row chunks for each atom j, where all the rows in a chunk are associated with the neighbors i of j, as well as the self-terms where i=j. So for atom j, the number of rows is equal to the number of atoms within the SNAP cutoff, plus 1 for the i=j terms, times 3 for each Cartesian component. The total number of dBi/dRj rows is therefore equal to N*(Nneigh+1)*3, and Nneigh may be different for each atom. To facilitate with determining which row belong to which atom pair ij, the last 3 columns contain indices; the 3rd to last column contains global indices of atoms i (the neighbors), the 2nd to last column contains global indices of atoms j, and the last column contains an index 0,1,2 for the Cartesian component. Like the bik rows, the first K columns correspond to each bispectrum coefficient.
Finally, the first column of the last row contains the reference energy.