On Fri, Dec 17, 2021 at 02:43:00PM +0100, Rafael Sadowski wrote:
>
> On Fri Dec 17, 2021 at 01:03:30PM +0800, Kevin Lo wrote:
> > Hi,
> >
> > Please find attached a port of the openblas.
> >
> > Description:
> > OpenBLAS is an optimized BLAS (Basic Linear Algebra Subprograms) library
> > based on GotoBLAS2 1.13 BSD version.
> >
> > It's loosely based on the one at openbsd-wip [1].
> > I tested on i386, amd64, arm, arm64, and mips64.
> >
> > I didn't test on others platforms due to lack of hardware.
> >
> > If somebody is interested, can you give it a try, check if it works for
> > you and give some feedback? Thanks.
> >
> > [1] https://github.com/jasperla/openbsd-wip/tree/master/math/openblas
> >
> > Regards,
> > Kevin
>
> I like to see openblas, I really do. But I think it will be found by
> many other ports when it is imported. Checkout "This port is required
> by" from https://www.freshports.org/math/openblas/
>
> That means a bit more work needs to happen here, doesn't it?
Thanks for the feedback. I don't know if we want to replace
math/{blas,cblas,lapack} dependencies of ports with math/openblas.
I only tested math/{py-numpy,py-scipy}, seems to be working fine.
There's a numpy benchmark is available on the net [1].
I changed the matrix sizes to 2048 [2].
Let's find out how the different implementation compare in terms of
performance on my PC.
$ sysctl hw.model
hw.model=Intel(R) Core(TM) i5-6400 CPU @ 2.70GHz
blas & lapack:
Dotted two 2048x2048 matrices in 6.69 s.
Dotted two vectors of length 262144 in 0.33 ms.
SVD of a 1024x512 matrix in 1.13 s.
Cholesky decomposition of a 1024x1024 matrix in 0.15 s.
Eigendecomposition of a 1024x1024 matrix in 6.79 s.
openblas:
Dotted two 2048x2048 matrices in 0.18 s.
Dotted two vectors of length 262144 in 0.04 ms.
SVD of a 1024x512 matrix in 0.14 s.
Cholesky decomposition of a 1024x1024 matrix in 0.04 s.
Eigendecomposition of a 1024x1024 matrix in 1.41 s.
$ sysctl hw.{model,product}
hw.model=ARM Cortex-A72 r0p3
hw.product=Raspberry Pi 4 Model B Rev 1.1
blas & lapack:
Dotted two 2048x2048 matrices in 49.86 s.
Dotted two vectors of length 262144 in 1.94 ms.
SVD of a 1024x512 matrix in 8.78 s.
Cholesky decomposition of a 1024x1024 matrix in 1.63 s.
Eigendecomposition of a 1024x1024 matrix in 50.11 s.
openblas:
Dotted two 2048x2048 matrices in 3.21 s.
Dotted two vectors of length 262144 in 1.96 ms.
SVD of a 1024x512 matrix in 2.08 s.
Cholesky decomposition of a 1024x1024 matrix in 0.39 s.
Eigendecomposition of a 1024x1024 matrix in 12.34 s.
[1] https://gist.github.com/markus-beuckelmann/8bc25531b11158431a5b09a45abd6276
[2]
https://gist.github.com/markus-beuckelmann/8bc25531b11158431a5b09a45abd6276#file-numpy-benchmark-py-L14
