OpenBLAS VS cblas

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?

I found https://github.com/Maluuba/cblas but honestly I cannot speak to its performance.