[dataflow]

Have you tried using precomputed byte offsets rather than indices for all the integers? Those would definitely speed up the non-power-of-2-sized case, but I'd be curious if they might speed this up a little bit too, since they should remove some bit shifting instructions.

[patagurbon]

You often need the indices for other operations (indexing into other arrays). CSC it's transpose CSR, and the doubly compressed forms DCSC/DCSR, are optimized for linear algebra for the most part.