Adam Lugowski

> On x86 `long double` is an 80-bit "extended precision" type, stored in a 128-bit object. Only on some (most?) x86 Linux. Both macOS and Windows have 64-bit `long double`....

Something that might make sense is the C++23 fixed-width floating-point types: `std::float16_t` through `std::float128_t` and `std::bfloat16_t`. https://en.cppreference.com/w/cpp/types/floating-point

> Given a decimal representation, you always seek the best match. My view is that 1e99999 is indeed representable with a binary64 number: it is infinity. Agreed. However, this assumes...

> @alugowski That's reasonable. Pull request invited. > > I don't consider the current behaviour incorrect but if someone wants to contribute the code for the functionality that @alugowski described,...

I can confirm that the cmake instructions work on an M1 mac: ```bash mkdir build ; cd build cmake -Denable_internal_blaslib=YES .. make ``` Then run one of the tests to...

Ah, I see @maczikasz was specifically referring to using brew's OpenBLAS on M1. That still works, though the Readme instructions are incomplete. One must manually append the link path so...

# Benchmarks I'm attaching the ASV benchmark compare of this PR against the main branch using the bigger benchmarks I mentioned ( https://github.com/scipy/scipy/commit/6b78eacb75408faa2bacb0b4c29e949af6362822 ) Running on M1 Macbook pro (6...

CI errors seem to be `scipy/optimize/tests/test_linprog.py::TestLinprogIPSparse::test_bug_6139`. That test passes for me locally, and I can't figure out how it relates to this change. Any insight appreciated!

I'll write some code for a speedup benchmark. > The CircleCI job shows a build error in `csr.cxx`, that seems more important to sort out. That'll take some thought, there...

Speedup of `tocsr`(mostly sort_elements), two binops and a matmul on a random matrix with 5M nonzeros:  Of course the sparsity pattern matters a lot with any sparse matrix op....