Re: 01/03: gnu: Add QD.

Re: 01/03: gnu: Add QD.

[Mark H Weaver]

Hi Eric,

ericbavier@centurylink.net (Eric Bavier) writes:
> bavier pushed a commit to branch master
> in repository guix.
>
> commit cbc084e1a7c25d1c8944bcb20a989f795acdb096
> Author: Eric Bavier <bavier@cray.com>
> Date:   Wed Jan 31 13:55:37 2018 -0600
>
>     gnu: Add QD.
[...]
> +(define-public qd
> +  (package
> +    (name "qd")
> +    (version "2.3.18")
> +    (source (origin
> +              (method url-fetch)
> +              (uri (string-append "http://crd.lbl.gov/~dhbailey/mpdist/qd-"
> +                                  version ".tar.gz"))
> +              (sha256
> +               (base32
> +                "0vkihcj9fyv2cycq8515713gbs3yskhmivy8bznvx72i6ddnn2c1"))))
> +    (build-system gnu-build-system)
> +    (native-inputs
> +     `(("gfortran" ,gfortran)))
> +    (arguments
> +     `(#:configure-flags `("--disable-enable_fma" ;weird :/

"weird :/" is not a very useful comment.  Can you change it to explain
why you added this flag?  It seems unfortunate to disable fused
multiply-add for this kind of library.

      Mark

Re: 01/03: gnu: Add QD.

[Eric Bavier]

On Sat, 03 Feb 2018 22:55:56 -0500
Mark H Weaver <mhw@netris.org> wrote:

> Hi Eric,
> 
> ericbavier@centurylink.net (Eric Bavier) writes:
> > bavier pushed a commit to branch master
> > in repository guix.
> >
> > commit cbc084e1a7c25d1c8944bcb20a989f795acdb096
> > Author: Eric Bavier <bavier@cray.com>
> > Date:   Wed Jan 31 13:55:37 2018 -0600
> >
> >     gnu: Add QD.  
> [...]
> > +(define-public qd
> > +  (package
> > +    (name "qd")
> > +    (version "2.3.18")
> > +    (source (origin
> > +              (method url-fetch)
> > +              (uri (string-append "http://crd.lbl.gov/~dhbailey/mpdist/qd-"
> > +                                  version ".tar.gz"))
> > +              (sha256
> > +               (base32
> > +                "0vkihcj9fyv2cycq8515713gbs3yskhmivy8bznvx72i6ddnn2c1"))))
> > +    (build-system gnu-build-system)
> > +    (native-inputs
> > +     `(("gfortran" ,gfortran)))
> > +    (arguments
> > +     `(#:configure-flags `("--disable-enable_fma" ;weird :/  
> 
> "weird :/" is not a very useful comment.

Sorry.  I thought it was obvious it was in reference to the
"disable-enable" bit.

>  Can you change it to explain
> why you added this flag?  It seems unfortunate to disable fused
> multiply-add for this kind of library.

The library does not support runtime ISA detection.  I thought FMAs are
something we don't support in our baseline x86-64 builds?

`~Eric

Re: 01/03: gnu: Add QD.

[Ludovic Courtès]

Eric Bavier <ericbavier@centurylink.net> skribis:

> On Sat, 03 Feb 2018 22:55:56 -0500
> Mark H Weaver <mhw@netris.org> wrote:
>
>> Hi Eric,
>> 
>> ericbavier@centurylink.net (Eric Bavier) writes:
>> > bavier pushed a commit to branch master
>> > in repository guix.
>> >
>> > commit cbc084e1a7c25d1c8944bcb20a989f795acdb096
>> > Author: Eric Bavier <bavier@cray.com>
>> > Date:   Wed Jan 31 13:55:37 2018 -0600
>> >
>> >     gnu: Add QD.  
>> [...]
>> > +(define-public qd
>> > +  (package
>> > +    (name "qd")
>> > +    (version "2.3.18")
>> > +    (source (origin
>> > +              (method url-fetch)
>> > +              (uri (string-append "http://crd.lbl.gov/~dhbailey/mpdist/qd-"
>> > +                                  version ".tar.gz"))
>> > +              (sha256
>> > +               (base32
>> > +                "0vkihcj9fyv2cycq8515713gbs3yskhmivy8bznvx72i6ddnn2c1"))))
>> > +    (build-system gnu-build-system)
>> > +    (native-inputs
>> > +     `(("gfortran" ,gfortran)))
>> > +    (arguments
>> > +     `(#:configure-flags `("--disable-enable_fma" ;weird :/  
>> 
>> "weird :/" is not a very useful comment.
>
> Sorry.  I thought it was obvious it was in reference to the
> "disable-enable" bit.
>
>>  Can you change it to explain
>> why you added this flag?  It seems unfortunate to disable fused
>> multiply-add for this kind of library.
>
> The library does not support runtime ISA detection.  I thought FMAs are
> something we don't support in our baseline x86-64 builds?

Indeed.

Ludo'.

Re: 01/03: gnu: Add QD.

[Mark H Weaver]

Hi Eric,

Eric Bavier <ericbavier@centurylink.net> writes:

> On Sat, 03 Feb 2018 22:55:56 -0500
> Mark H Weaver <mhw@netris.org> wrote:
>
>> ericbavier@centurylink.net (Eric Bavier) writes:
>> > +     `(#:configure-flags `("--disable-enable_fma" ;weird :/  
>> 
>> "weird :/" is not a very useful comment.
>
> Sorry.  I thought it was obvious it was in reference to the
> "disable-enable" bit.

I thought that might be the case, but I was still curious why you were
disabling FMA support, and the comment didn't help.

>>  Can you change it to explain
>> why you added this flag?  It seems unfortunate to disable fused
>> multiply-add for this kind of library.
>
> The library does not support runtime ISA detection.  I thought FMAs are
> something we don't support in our baseline x86-64 builds?

Indeed.  I was surprised to learn that FMAs are only supported in very
recent Intel processors.  My Yeeloong with MIPS-based Loongson 2F
processor from 2008 includes support for FMA instructions, and I guessed
that Intel had them long before, but actually Intel didn't add them
until several years later.

       Mark

Re: 01/03: gnu: Add QD.

[Eric Bavier]

On Mon, 05 Feb 2018 20:24:45 -0500
Mark H Weaver <mhw@netris.org> wrote:

> Hi Eric,
> 
> Eric Bavier <ericbavier@centurylink.net> writes:
> 
> > On Sat, 03 Feb 2018 22:55:56 -0500
> > Mark H Weaver <mhw@netris.org> wrote:
> >  
> >> ericbavier@centurylink.net (Eric Bavier) writes:  
> >> > +     `(#:configure-flags `("--disable-enable_fma" ;weird :/    
> >> 
> >> "weird :/" is not a very useful comment.  
> >
> > Sorry.  I thought it was obvious it was in reference to the
> > "disable-enable" bit.  
> 
> I thought that might be the case, but I was still curious why you were
> disabling FMA support, and the comment didn't help.

Sorry about the confusion.  I'll revise it to something more
informative.  I usually try to refrain from "witty" comments.

> 
> >>  Can you change it to explain
> >> why you added this flag?  It seems unfortunate to disable fused
> >> multiply-add for this kind of library.  
> >
> > The library does not support runtime ISA detection.  I thought FMAs are
> > something we don't support in our baseline x86-64 builds?  
> 
> Indeed.  I was surprised to learn that FMAs are only supported in very
> recent Intel processors.  My Yeeloong with MIPS-based Loongson 2F
> processor from 2008 includes support for FMA instructions, and I guessed
> that Intel had them long before, but actually Intel didn't add them
> until several years later.

AMD also had a 4-argument FMA instruction much sooner.

Thank you for the attention to this matter,
`~Eric