Jay Zhuang

For reference, the SciPy sparse dot code is extremely simple: ```C++ template void csr_matvec(const I n_row, const I n_col, const I Ap[], const I Aj[], const T Ax[], const T...

> If you run the code through [Binder](https://mybinder.org/v2/gh/spcl/dace/8bb84af561a243ec280bf78c709586c9b7ec568d) with the latest master branch the time goes down to 16.4ms Thanks, it now runs as fast as SciPy on my machine....

> it's possible to write a simple transformation that gets the memory access patterns from the memlets and inserts a prefetch tasklet. Thanks, will come back if I have any...

Oh I forgot to ask -- where can I find the exact code for loop tiling and SIMD transformations, which were used by the SpMV benchmark in the DaCe paper?...

@sancierra @alexnick83 Wonderful, thanks for the detailed replies! Let me take a closer look.

> Is there a particular problem with implementing the benchmark? Thanks, @dionhaefner would be the expert to talk with😃 I am currently busy with other stuff and will try implementing...

> SuiteSparseGraphBLAS.gbset(:nthreads, ) This works well, thanks!

I found it in [Trottenberg's Multigrid book](https://www.elsevier.com/books/multigrid/trottenberg/978-0-08-047956-9) -> Appendix A. An Introduction to Algebraic Multigrid (by Klaus Stuben) -> Section A.4.2 Direct Interpolation -> A.4.2.3 General case (compared to the...

> it should just be setup to use LinearSolve.jl Interesting, then I would argue that `presmoother` and `postsmoother` should also be flexible, generic approximate/iterative solvers. For complicated (say indefinite, nonsymmetric)...

The AMG solve phase (just a few SpMVs) is much easier to port to GPU than the AMG setup phase (contains [coarse node selection algorithms](https://onlinelibrary.wiley.com/doi/abs/10.1002/nla.541)). This strategy is [adopted by...