A graphblas.apply operation that optimizes this line y = 2 / N * np.exp(-tau / lmax * (a * num + a)) where num is a sparse tensor and everything else is a scalar. If we port the line to the GraphBLAS dialect naively following the PEMDAS order of operations, we'd get a ton of graphblas.apply calls. Ideally, we'd change the ordering to be y = 2 / N * np.exp((-tau / lmax * a) * (num + 1)) so that we'd only have 3 graphblas.apply calls. I'm not sure how much this matters or if it'll all get optimized away anyway via graphblas.apply_generic during lowering. This is not a contrived example as this line is used in GraphWave.
We should review all of our algorithms to see if there are any ops we need.
The minimum work that should be done for this ticket is:
we should review the explicit examples given above.
we should look at all the algorithms in algorithms.py, look at all the scf.for, scf.while, and scf.if ops (and all ops from the scf dialect) to see if we can't abstract out those ops into graphblas ops.
A lot of our algorithms in
algorithms.pyhave workarounds for ops that we don't have yet.Some ops we might need include:
graphblas.concatwith anaxisattribute that will allow us to concatenate/stack tensors (matrices and vectors). https://github.com/metagraph-dev/mlir-graphblas/issues/251 is relevant to this idea.graphblas.applyop that doesn't create a whole new sparse matrix. https://github.com/metagraph-dev/mlir-graphblas/discussions/232#discussioncomment-1762517 is relevant to this idea. This might also apply tographblas.select.graphblas.applyoperation that optimizes this liney = 2 / N * np.exp(-tau / lmax * (a * num + a))wherenumis a sparse tensor and everything else is a scalar. If we port the line to the GraphBLAS dialect naively following the PEMDAS order of operations, we'd get a ton ofgraphblas.applycalls. Ideally, we'd change the ordering to bey = 2 / N * np.exp((-tau / lmax * a) * (num + 1))so that we'd only have 3graphblas.applycalls. I'm not sure how much this matters or if it'll all get optimized away anyway viagraphblas.apply_genericduring lowering. This is not a contrived example as this line is used in GraphWave.We should review all of our algorithms to see if there are any ops we need.
The minimum work that should be done for this ticket is:
algorithms.py, look at all thescf.for,scf.while, andscf.ifops (and all ops from thescfdialect) to see if we can't abstract out those ops into graphblas ops.