Tobias Ribizel

## Description This adds a new `multibyte_split` implementation for "nice" non-overlapping delimiters that needs to scan the input only once, and takes full advantage of `byte_range`. To accomplish this, I...

## Description This adds `gdb` pretty printers for `rmm::device_uvector`, `thrust::*_vector`, `thrust::device_reference` and `cudf::*_span`. The implementation is based on https://github.com/NVIDIA/thrust/pull/1631. I will probably backport the thrust-specific changes to there as well,...

When looking at the implementation of device_reference and tuple_of_iterator_references, I discovered that the `swap` support may be unnecessarily constraining (compared to _BitReference from libstdc++, which I originally took as a...

Enable swap(...) for temporaries and between different reference types

1

An attempt at resolving the issues surrounding swap(...) between temporary device references and between different reference(-like) types Closes NVIDIA/cccl#802

Benchmark work estimate

8

This PR adds work estimates to the executor Operations, implements them for a few Dense kernels ~~and outputs them in benchmark loggers.~~ Related to #1784

Add support for profiler hooks in logger

2

Tools like [TAU](https://www.cs.uoregon.edu/research/tau/docs/newguide/bk05rn01.html) or [NVIDIA Toolkit Extensions](https://docs.nvidia.com/cuda/profiler-users-guide/index.html#nvtx) allow applications to log custom events/phases in a nested fashion. We could profit from their tooling by providing loggers wrapping this functionality.

Mixed-Precision SpMV and BLAS Kernels

1

Currently, only ELL, SparsityCsr and RowGatherer support true mixed-precision execution. Adding mixed-precision support for CSR, COO, SELL-P and Hybrid should be straight-forward based on accessors, falling back to our own...

Inconsistent use of integer types

5

It seems to me Ginkgo is sometimes integer types inconsistently. This includes both 32/64 bit as well as `signed`/`unsigned`. Examples: * `dim` returns a `size_type` for all dimensions, even though...

Simplify setup of factory parameters of factory type

1

Currently, setting up stopping criteria requires a lot of boilerplate code: ```cpp gko::ResidualNormReduction::build().with_reduction_factor(...).on(exec) ``` In this example, there are two points that I believe could or should be eliminated completely,...