Those are all great questions @liangfu. Question 3 is an interesting one w.r.t to what kinds of scheduling primitives we'll need for sparse operators. One easy workaround is to apply vectorization along a dense tensor dimension if there is one. For many practical examples, tensors won't be sparse along all dimensions. But the question gets more tricky when that assumption does not hold.
And due to the dynamism of varying idx/val arrays, this can create an interesting discussion around how autoTVM will be used on more dynamic operators. Right now TOPI is built around the assumption that shapes are known statically. This will be an interesting implementation challenge, which probably deserves its own RFC. Perhaps @ZihengJiang the `Extension for Scheduling` can point to a follow-up RFC? -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/apache/incubator-tvm/issues/4332#issuecomment-557996282
