OpenBLAS VS linux

The Fortran implementation is just a reference implementation. The goal of reference BLAS [0] is to provide relatively simple and easy to understand implementations which demonstrate the interface and are intended to give correct results to test against. Perhaps an exceptional Fortran compiler which doesn't yet exist could generate code which rivals hand (or automatically) tuned optimized BLAS libraries like OpenBLAS [1], MKL [2], ATLAS [3], and those based on BLIS [4], but in practice this is not observed.

Justine observed that the threading model for LLaMA makes it impractical to integrate one of these optimized BLAS libraries, so she wrote her own hand-tuned implementations following the same principles they use.

[0] https://en.wikipedia.org/wiki/Basic_Linear_Algebra_Subprogra...

[1] https://github.com/OpenMathLib/OpenBLAS

[2] https://www.intel.com/content/www/us/en/developer/tools/onea...

[3] https://en.wikipedia.org/wiki/Automatically_Tuned_Linear_Alg...

[4]https://en.wikipedia.org/wiki/BLIS_(software)

Yeah. I'm going to be helping to work on expanding CI for OpenBlas and have been diving into this stuff lately. See the discussion in this closed OpenBlas issue gh-1968 [0] for instance. OpenBlas's Skylake kernels do rely on intrinsics [1] for compilers that support them, but there's a wide range of architectures to support, and when hand-tuned assembly kernels work better, that's what are used. For example, [2].

[0] https://github.com/xianyi/OpenBLAS/issues/1968

[1] https://github.com/xianyi/OpenBLAS/blob/develop/kernel/x86_6...

[2] https://github.com/xianyi/OpenBLAS/blob/23693f09a26ffd8b60eb...

We'll have to wait until part 2 to see what they are actually proposing, but they are trying to solve a real problem. To get a sense of things check out the handwritten assembly kernels in OpenBlas [0]. Note the level of granularity. There are micro-optimized implementations for specific chipsets.

If progress in ML will be aided by a proliferation of hyper-specialized hardware, then there really is a scalability issue around developing optimized matmul routines for each specialized chip. To be able to develop a custom ASIC for a particular application and then easily generate the necessary matrix libraries without having to write hand-crafted assembly for each specific case seems like it could be very powerful.

[0] https://github.com/xianyi/OpenBLAS/tree/develop/kernel

libraries mkl_rt not found in ['C:\python\lib', 'C:\', 'C:\python\libs'] ``` Install this and try again. Might need to reboot, never know with Windows https://www.openblas.net/

There isn't any fortran code in the repo there itself but numpy itself can be linked with several numeric libraries. If you look through the wheels for numpy available on pypi, all the latest ones are packaged with OpenBLAS which uses Fortran quite a bit: https://github.com/xianyi/OpenBLAS

Sure - write functions in another language, export C bindings, and then call those functions from Python. An example is NumPy - a lot of its linear algebra functions are implemented in C and Fortran.

Read the official docs yet?

These are a bit easier to see what's going on:

https://github.com/torvalds/linux/commit/d5cf50dafc9dd5faa1e...

https://github.com/torvalds/linux/blob/d5cf50dafc9dd5faa1e61...

Unfortunately Github doesn't have a way to render symbols for whitespace, but you can tell by selecting the spaces that the previous version had leading tabs. Linus changed it so that the tokens `default` and the number e.g. `12` are also separated by a tab. This is tricky, because the token "default" is seven characters, it will always give this added tab a width of 1 char which makes it always layout the same as if it were a space no matter if you use tab widths of 1, 2, 4, or 8.

There was also a Gentoo effort to run atop FreeBSD[0]. The challenge of course is that afaik none of the BSD kernel ABIs are considered stable. The stable interface is the BSD libc. That said, with binfmt_misc, I don't see a reason you couldn't just run (at least some) FreeBSD binaries on Linux with a thin syscall translation layer (rather something like qemu-system) and then your layer hooked via binfmt_misc. I'm not aware of anyone who has done this for FreeBSD, but prior efforts existed as alternate binfmts for SysVr4/5 ELF binaries[2]. Either way would take some elbow grease, but you *might* even be able just reuse binfmt_elf and just have a new interpreter for FreeBSD elf.

[0] https://wiki.gentoo.org/wiki/Gentoo_FreeBSD

[1] https://docs.kernel.org/admin-guide/binfmt-misc.html

[2] https://github.com/torvalds/linux/blob/master/fs/binfmt_elf....

> The original less-than check was deemed incorrect

It was only deemed incorrect because of an information leak. Not because it's a valid use-case for user space to copy smaller portions of *hwrpb into user space. https://github.com/torvalds/linux/commit/21c5977a836e399fc71...

Correct. And the combined work needs to carry the MIT license text and copyright attributions for the MIT software authors. With binary distribution it must also be overt, not hidden in some source code drop, but directly accompanying the binary.

Many people who talk about relicensing never credit the MIT developers or distribute the MIT license text. "Because it's GPL now."

I don't think that you believe that, but many developers do.

Some don't see the need for source code scans for Open Source compliance, because the license.txt says GPL, so it's GPL. Prime example is the Linux kernel. There is code under different licenses in there, but people don't even read https://github.com/torvalds/linux/blob/master/COPYING till the end ("In addition, other licenses may also apply.") and conclude it's simply GPL 2 and nothing else.

Also be aware that sublicensing is not the same as relicensing.

> Does he have something against it?

He notoriously hates GNU Emacs, yes.

https://marc.info/?m=122955159617722

https://github.com/torvalds/linux/blob/master/Documentation/...

So If we would only count code and not comments, it is only 9489 LoC Rust. Which would be about 0.03% and if we take all lines and not only LoC it would be around 0.05%

[0] https://github.com/XAMPPRocky/tokei

[1] https://github.com/torvalds/linux/commit/b401b621758e46812da...

AIUI fsync is built on futex_waitv which has been upstreamed. So this has to be more than that.

https://github.com/torvalds/linux/commit/a0eb2da92b715d0c97b...

git clone --no-checkout --depth 1 https://github.com/torvalds/linux.git $dir