On Monday, 14 November 2016 at 07:42:13 UTC, Picaud Vincent wrote:
Just a word to say that I find the glas benchmark results and
design awesome!
https://github.com/libmir/mir
I am new to D but I think this kind of developments has the
potential to attract a lot of people doing numerical
computations.
I also would like to mention the "libFlame" approach used to
describe and implement linear algebra algorithms like LU,
Cholesky... decompositions.
https://www.cs.utexas.edu/~flame/web/methodology.html
IMHO the ndslice is a perfect match to benefit from this
methodology.
I hope the best future for glas, it is very attractive.
Vincent
Thank you, Vincent!
Yes, I will use libFlame approach, it is a really good. In the
same time Dlang is more user-friendly for SIMD optimization. GLAS
routines probably will follow hypothetical (it is closed source)
Intel MKL approach but without unrolled loops. LibFLAME has 2
kinds of algorithms: blocking and unblocking. GLAS where it is
possible will have 3 kinds: tiny unblocking, register (SIMD)
blocking, and normal blocking.
Mir GLAS was moved to separate repo [1]. It is a "better C"
library now. This means it is written in D but does not requre
DRuntime. It also has standart BLAS API, so GLAS will be able to
replace BLAS in already existing C / Fortran infrastructure.
[1] https://github.com/libmir/mir-glas
Best regards,
Ilya
