Tobias Ribizel

Add TAU logger

1

The logger adds timers for all important Ginkgo events based on the [perfstubs](https://github.com/UO-OACISS/perfstubs) library. To make it work, we also need to forward all logger events to the executor. This...

Simplify GMRES kernels

6

This PR separates the step1 and initialize2 kernels into individual reductions (norm and dot) and axpy/scale operations, which allows us to use the simple kernel setup for all of GMRES...

Fix GMRES lucky breakdown

1

Currently we don't correctly deal with the case that the initial guess already solves Ax=b, which leads to NaNs in (CB-)GMRES in #973. This should be fixed before the next...

In #973 we use really large tolerances, because already after a few iterations, Reference and device kernels diverge noticeably. We need to figure out whether this is just an inherent...

In the tests for #973, I noticed that the IDR implementation uses different approaches to initialize the random matrix across executors. To improve testability, we should initialize the data in...

Thread-safety of Ginkgo

6

In the discussion on #993, but also some previous issues like #740 and #565, as well as future plans like #805, there is an underlying thing we have not yet...

Currently, our error reporting from within Ginkgo leaves a bit to be desired. It builds on two things: explicit errors like mismatching dimensions in LinOp::apply and implicit errors like the...

Add OpenMP parallel sync-free triangular solvers

2

This adds a simple sync-free parallel triangular solver to OpenMP. cc @lksriemer

Improve reordering interface

1

As observed in https://github.com/ginkgo-project/ginkgo/discussions/939, our reordering interface is easy to accidentally misuse. Part of the issue is that Permutation violates the LinOp assumptions in some situations. A permutation matrix is...

Public common kernels

4

This PR makes the simple `run_kernel` interface public and adds an example for how to use to allow portable kernel execution.