Tianjian Lu

@machc

Hi, I am still having the issue of downloading the dataset of 1.40625 deg resolution. Please see the following error. `$ wget --no-check-certificate "https://dataserv.ub.tum.de/s/m1524895/download?path=%2F1.40625deg&files=all_1.40625deg.zip" -O all_1.40625deg.zip --2022-07-29 17:41:47-- https://dataserv.ub.tum.de/s/m1524895/download?path=%2F1.40625deg&files=all_1.40625deg.zip` Resolving...

It says the page @ https://climetlab.readthedocs.io/en/latest/examples/09-weatherbench.html does not exist. Can you help check?

cond = np.linalg.cond(A.todense()) print(cond) 1.1747166e+17 It seems the condition number is very large. Is this what you expected?

Looks like the input is symmetric positive definite. I have checked `np.all(np.linalg.eigvals(A.todense()) > 0)` and `np.allclose(A_dense, A_dense.T.conj(), rtol=rtol, atol=atol)` Have you tried using float64?

The nondeterministic behavior may come from multithreading.

Thanks for looking into this! We recently enable the lowering of BCOO dot_general to cuSparse (https://github.com/google/jax/pull/12138). Yes, `indices_sorted=True` is one of the requirements for using cuSparse.

Thanks for the suggestions! Can you also share insights on the cusparse matmat algorithms? Which one shall we use as the default for jax? @marsaev https://docs.nvidia.com/cuda/cusparse/index.html#cusparse-generic-function-spmm

@marsaev @fbusato Thank you. A follow-up question about the cuda versions for these new algorithms. I found this in the release notes ``` [2.5.12. cuSPARSE: Release 11.2 Update 1](https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#cusparse-11.2.1) ......

@marsaev @fbusato Comparing to the default algorithms for SpMV and SpMM, do the four algorithms that provide determinism have trade-offs on accuracy? I broke a few accuracy tests due to...