Benchmarking mir.ndslice + lubeck against numpy and Julia
no at spam.net
Sun Jan 12 14:41:11 UTC 2020
On Sunday, 12 January 2020 at 12:46:44 UTC, Dennis wrote:
> The first important default setting is the algorithm. Lapack
> offers 'gesdd', a 'more efficient divide-and-conquer approach',
> and 'gesvd', a 'general rectangular approach' . Python and
> Julia default to gesdd, while D's Lubeck defaults to gesvd.
You left out an important detail in your description. gesdd is
more efficient, but at the expense of being less accurate, and
can easily fail on you. I have a strong preference for correct as
the default as I've run into problems with optimized for speed by
default before. Nobody ever digs in and understands the tradeoffs
involved for the default.
More information about the Digitalmars-d