I did some benchmark somehow the matrix mul performs consistently slower than java ojalgo. about jblas HOT 7 CLOSED

okok,it is fine.I could also become a tester for your binaries in the future too.^^

from jblas.

Hi,

which operating system was this on?

from jblas.

windows_intel64bit
OS Name: Microsoft Windows 10 Home Single Language
OS Version: 10.0.19041 N/A Build 19041

from jblas.

Ah, I see. Yeah, I Windows binaries are unfortunately linked against non-optimized LAPACK and BLAS libraries.

This is mentioned (a bit cryptically) here: https://github.com/jblas-project/jblas/wiki/Supported-Platforms

The problem is that ATLAS (the native BLAS library) has a very complex build process that I could never get to run on Windows. However, I recently started using OpenBLAS, once I find time to run that, performance on Windows should improve a lot. Sorry about that.

from jblas.

BTW,don't worry about it too much since generally it is just around 10 percent slower than pure java ojalgo.

Like>
552035845299
testfinished,with Primitive32Matrix .ojalgo(no try catch)
start test
-- org.jblas INFO Starting temp DLL cleanup task.
-- org.jblas INFO Deleted 4 unused temp DLL libraries from C:\Users\sumur\AppData\Local\Temp
600934105500

or>
start test
-- org.jblas INFO Starting temp DLL cleanup task.
600251319500
testfinished,with matrix .jblas(no try catch)
start test
537240456800
testfinished,with Primitive32Matrix .ojalgo(no try catch)

both of these does not seem too bad,compare to jblas vs apache common.. or naive java2d array..

from jblas.

One of the possibilities is that it take quite some time to also clean temp dll
Maybe we can manually trigger it before the calc.As like a setup process.
Which is honestly quite acceptable when given some documentation and guide.
Compare to boilerplate code for some frameworks...

from jblas.

Yeah, the non-ATLAS code for the low level matrix operations is just very slow, not really optimized in terms of assembly instructions or memory access pattern.

Thanks for the offer for testing, I'll reply to this thread when there is some progress made.

from jblas.