pygraphblas author here, here is a good notebook introduction to the very basics of GraphBLAS with Python. The concepts carry naturally to the C API as well:
I'm a graph theorist, and I spend a lot of time on various optimization problems (matchings, packings, coverings, Steiner trees, etc) doing VLSI. AFAICT, it seems like GraphBLAS is geared for information processing about data stored in a graph representation, and doesn't approach the problems I need. Am I wrong?
Right now, the graphs I'm interested in are reasonably small (tens of thousands of edges), and my oldschool approaches are good enough, but if my employer's products continue their grow trajectory, we'll need to leverage GPUs and TPUs if at all possible. Should I be learning about GraphBLAS?
GraphBLAS is meant to solve graph problems (i.e., generally the sort of ones you are interested in, but more emphasis on parallel ones). Perhaps graph analytics may be closer to what it does right now rather than graph theory. At the very basic level, it exploits the matrix graph duality and casts graph problems as matrix problems. For eg., BFS and SSSP are sparse matrix vector multiplications over a suitably defined semiring. The benefit of this is that you can use high performance matrix libraries. The downside is that not everything can be done this way. At least not easily. You may find the book by Kepner and Gilbert useful. Also look at cugraph and gunrock for GPU based graph frameworks.
It's so early in this work that you might be the first person asking those specific questions. I'm no expert but I take it by VLSI you mean doing circuit routing and simulation and optimizing that topology. My gut tells me those would be good fits for the GraphBLAS.
If the base types don't suit your needs, you can always make your own. Tim Davis has talked about how he's made new types that are small nxn matrices as elements of a larger matrix. You can have complex, quaternion, or complex bags of stuff as matrix elements as long as you define the operators you need to work on those types.
Here's an example of using User Defined Types for a shortest path algorithm that also stores backlinks along the shortest path so the shortest path tree is materialized. in Python:
Shortest paths is definitely a big timesink... I'll definitely give that a shot. Thanks for the link :)
As for what I'm doing, I work at D-Wave and a lot of my work is on what can be thought of as "compilers" for our quantum computers. There's tons of similarity between my problem area and compilers for FPGAs, for example, so I call it VLSI because that's the right ballpark.
https://github.com/michelp/pygraphblas/blob/master/pygraphbl...
I also made a quick introduction video for a paper submission you can see here:
https://www.youtube.com/watch?v=JUbXW_f03W0