Re: libmvec simd math functions in fortran

[Richard Biener]

On Thu, Nov 2, 2017 at 7:42 PM, Toon Moene  wrote:
> On 11/01/2017 05:26 PM, Jakub Jelinek wrote:
>
>> On Wed, Nov 01, 2017 at 04:23:11PM +, Szabolcs Nagy wrote:
>>>
>>> is there a way to get vectorized math functions in fortran?
>>>
>>> in c code there is attribute simd declarations or openmp
>>> declare simd pragma to tell the compiler which functions
>>> have simd variant, but i see no such thing in fortran.
>>
>>
>> !$omp declare simd should work fine in fortran (with -fopenmp
>> or -fopenmp-simd).
>
>
> Note that - if you don't want to change the Fortran code, this - almost two
> years old - proposal would work:
>
> https://gcc.gnu.org/ml/gcc/2016-01/msg00025.html
>
> Obviously, I'll only be able to implement this once retirement comes around
> (i.e., after 2023).

I think a better user-interface would be

gfortran ... -include mathvec

which pulls in a pre-def for the vectorized math functions as available.
Somewhere that mathvec "module" needs to reside, ideally it came
from the local glibc install (but module files are GCC version dependent?).

Not sure if gfortran supports -include for module files, OTOH a
-include of fortran source should also work in all contexts(?).

The implementation challenge is

1) actually being able to declare the relevant math functions appropriately
(say, for -ffixed-form f77 style source)
2) make gfortran pick the FUNCTION_DECLs from those declarations
instead of those generated from the mathbuiltins.def.

Solving 2) would be nice to make this work "manually" anyway.
I'm not quite sure we properly override / merge (for other attributes)
with the builtins.

Richard.

> Kind regards,
>
> --
> Toon Moene - e-mail: t...@moene.org - phone: +31 346 214290
> Saturnushof 14, 3738 XG  Maartensdijk, The Netherlands
> At home: http://moene.org/~toon/; weather: http://moene.org/~hirlam/
> Progress of GNU Fortran: http://gcc.gnu.org/wiki/GFortran#news

Re: libmvec simd math functions in fortran

[Toon Moene]

On 11/01/2017 05:26 PM, Jakub Jelinek wrote:


On Wed, Nov 01, 2017 at 04:23:11PM +, Szabolcs Nagy wrote:

is there a way to get vectorized math functions in fortran?

in c code there is attribute simd declarations or openmp
declare simd pragma to tell the compiler which functions
have simd variant, but i see no such thing in fortran.


!$omp declare simd should work fine in fortran (with -fopenmp
or -fopenmp-simd).


Note that - if you don't want to change the Fortran code, this - almost 
two years old - proposal would work:


https://gcc.gnu.org/ml/gcc/2016-01/msg00025.html

Obviously, I'll only be able to implement this once retirement comes 
around (i.e., after 2023).


Kind regards,

--
Toon Moene - e-mail: t...@moene.org - phone: +31 346 214290
Saturnushof 14, 3738 XG  Maartensdijk, The Netherlands
At home: http://moene.org/~toon/; weather: http://moene.org/~hirlam/
Progress of GNU Fortran: http://gcc.gnu.org/wiki/GFortran#news

Re: libmvec simd math functions in fortran

[Janne Blomqvist]

On Thu, Nov 2, 2017 at 2:49 PM, Szabolcs Nagy  wrote:
> On 01/11/17 16:47, Szabolcs Nagy wrote:
>> On 01/11/17 16:26, Jakub Jelinek wrote:
>>> On Wed, Nov 01, 2017 at 04:23:11PM +, Szabolcs Nagy wrote:
 is there a way to get vectorized math functions in fortran?

 in c code there is attribute simd declarations or openmp
 declare simd pragma to tell the compiler which functions
 have simd variant, but i see no such thing in fortran.
>>>
>>> !$omp declare simd should work fine in fortran (with -fopenmp
>>> or -fopenmp-simd).
>>>
>>
>> 1) i don't want to change the fortran.
>>
>> 2) it does not work for me.
>>
>> i want this to call vector powf in libmvec:
>>
>> subroutine foo(a,b,c)
>>   real(4) a(8000),b(8000),c(8000)
>>   do j=1,8000
>> a(j)=b(j)**c(j)
>>   end do
>> end
>>
>> where do i put
>>
>> !$omp declare simd (powf)
>>
>> ?
>
> to answer my question..
>
> it seems fortran cannot express the type signature
> of mathematical functions because arguments are
> passed by reference.
>
> so there is no way to declare math interfaces
> and then add omp declare simd to them to get
> simd versions.
>
> (it's not clear to me how omp declare simd is
> supposed to work in fortran, but it is not useful
> for vectorizing loops with math functions.)
>
> so gfortran will need a different mechanism to
> do the vectorization, e.g. an option like
> -mveclibabi=glibc, but the list of available
> vector functions need to be specified somewhere.

The fortran frontend turns the

b(j)**c(j)

expression in your example into a "call" to __builtin_powf() (check
with -fdump-tree-original), and similarly for other math intrinsics
that have corresponding builtins. Ideally the middle-end optimizers
should to be able to turn that into calls to the appropriate vector
math library when -mveclibabi= is in effect.



-- 
Janne Blomqvist

Re: libmvec simd math functions in fortran

[Richard Biener]

On Thu, Nov 2, 2017 at 2:07 PM, Richard Biener
 wrote:
> On Thu, Nov 2, 2017 at 1:49 PM, Szabolcs Nagy  wrote:
>> On 01/11/17 16:47, Szabolcs Nagy wrote:
>>> On 01/11/17 16:26, Jakub Jelinek wrote:
 On Wed, Nov 01, 2017 at 04:23:11PM +, Szabolcs Nagy wrote:
> is there a way to get vectorized math functions in fortran?
>
> in c code there is attribute simd declarations or openmp
> declare simd pragma to tell the compiler which functions
> have simd variant, but i see no such thing in fortran.

 !$omp declare simd should work fine in fortran (with -fopenmp
 or -fopenmp-simd).

>>>
>>> 1) i don't want to change the fortran.
>>>
>>> 2) it does not work for me.
>>>
>>> i want this to call vector powf in libmvec:
>>>
>>> subroutine foo(a,b,c)
>>>   real(4) a(8000),b(8000),c(8000)
>>>   do j=1,8000
>>> a(j)=b(j)**c(j)
>>>   end do
>>> end
>>>
>>> where do i put
>>>
>>> !$omp declare simd (powf)
>>>
>>> ?
>>
>> to answer my question..
>>
>> it seems fortran cannot express the type signature
>> of mathematical functions because arguments are
>> passed by reference.
>>
>> so there is no way to declare math interfaces
>> and then add omp declare simd to them to get
>> simd versions.
>
> Not even with BIND(C) or so?
>
>> (it's not clear to me how omp declare simd is
>> supposed to work in fortran, but it is not useful
>> for vectorizing loops with math functions.)
>>
>> so gfortran will need a different mechanism to
>> do the vectorization, e.g. an option like
>> -mveclibabi=glibc, but the list of available
>> vector functions need to be specified somewhere.
>
> I think most useful would be if glibc ships with a
> fortran module that provides the appropriate
> declarations and users can simply include such
> module or such module would be automagically
> included like we have that C pre-def header.
>
> Of course first we have to solve the mystery
> of how to actually declare the libm sin() function
> in the appropriate way.  And if there isn't one,
> provide one by means of an extension
> ($! pragma gcc whatever?)

Always good to CC fortran@ for such stuff anyways.

Richard.

> Richard.

Re: libmvec simd math functions in fortran

[Jakub Jelinek]

On Thu, Nov 02, 2017 at 12:49:18PM +, Szabolcs Nagy wrote:
> to answer my question..
> 
> it seems fortran cannot express the type signature
> of mathematical functions because arguments are
> passed by reference.

Fortran indeed normally does pass by reference, and you can vectorize it
that way (you can use various declare simd clauses to express in detail
how the arguments should be handled), or pass by reference:

interface
  double precision function foo (x, y)
double precision, value :: x, y
!$omp declare simd (foo)
  end function
end interface

and/or using bind(c):

interface
  function bar (x, y) bind(C)
use, intrinsic :: iso_c_binding
real(c_double), value :: x, y
real(c_double) :: bar
!$omp declare simd (bar)
  end function
end interface

etc.

Jakub

Re: libmvec simd math functions in fortran

[Richard Biener]

On Thu, Nov 2, 2017 at 1:49 PM, Szabolcs Nagy  wrote:
> On 01/11/17 16:47, Szabolcs Nagy wrote:
>> On 01/11/17 16:26, Jakub Jelinek wrote:
>>> On Wed, Nov 01, 2017 at 04:23:11PM +, Szabolcs Nagy wrote:
 is there a way to get vectorized math functions in fortran?

 in c code there is attribute simd declarations or openmp
 declare simd pragma to tell the compiler which functions
 have simd variant, but i see no such thing in fortran.
>>>
>>> !$omp declare simd should work fine in fortran (with -fopenmp
>>> or -fopenmp-simd).
>>>
>>
>> 1) i don't want to change the fortran.
>>
>> 2) it does not work for me.
>>
>> i want this to call vector powf in libmvec:
>>
>> subroutine foo(a,b,c)
>>   real(4) a(8000),b(8000),c(8000)
>>   do j=1,8000
>> a(j)=b(j)**c(j)
>>   end do
>> end
>>
>> where do i put
>>
>> !$omp declare simd (powf)
>>
>> ?
>
> to answer my question..
>
> it seems fortran cannot express the type signature
> of mathematical functions because arguments are
> passed by reference.
>
> so there is no way to declare math interfaces
> and then add omp declare simd to them to get
> simd versions.

Not even with BIND(C) or so?

> (it's not clear to me how omp declare simd is
> supposed to work in fortran, but it is not useful
> for vectorizing loops with math functions.)
>
> so gfortran will need a different mechanism to
> do the vectorization, e.g. an option like
> -mveclibabi=glibc, but the list of available
> vector functions need to be specified somewhere.

I think most useful would be if glibc ships with a
fortran module that provides the appropriate
declarations and users can simply include such
module or such module would be automagically
included like we have that C pre-def header.

Of course first we have to solve the mystery
of how to actually declare the libm sin() function
in the appropriate way.  And if there isn't one,
provide one by means of an extension
($! pragma gcc whatever?)

Richard.

Re: libmvec simd math functions in fortran

[Szabolcs Nagy]

On 01/11/17 16:47, Szabolcs Nagy wrote:
> On 01/11/17 16:26, Jakub Jelinek wrote:
>> On Wed, Nov 01, 2017 at 04:23:11PM +, Szabolcs Nagy wrote:
>>> is there a way to get vectorized math functions in fortran?
>>>
>>> in c code there is attribute simd declarations or openmp
>>> declare simd pragma to tell the compiler which functions
>>> have simd variant, but i see no such thing in fortran.
>>
>> !$omp declare simd should work fine in fortran (with -fopenmp
>> or -fopenmp-simd).
>>
> 
> 1) i don't want to change the fortran.
> 
> 2) it does not work for me.
> 
> i want this to call vector powf in libmvec:
> 
> subroutine foo(a,b,c)
>   real(4) a(8000),b(8000),c(8000)
>   do j=1,8000
> a(j)=b(j)**c(j)
>   end do
> end
> 
> where do i put
> 
> !$omp declare simd (powf)
> 
> ?

to answer my question..

it seems fortran cannot express the type signature
of mathematical functions because arguments are
passed by reference.

so there is no way to declare math interfaces
and then add omp declare simd to them to get
simd versions.

(it's not clear to me how omp declare simd is
supposed to work in fortran, but it is not useful
for vectorizing loops with math functions.)

so gfortran will need a different mechanism to
do the vectorization, e.g. an option like
-mveclibabi=glibc, but the list of available
vector functions need to be specified somewhere.

Re: libmvec simd math functions in fortran

[Szabolcs Nagy]

On 01/11/17 16:26, Jakub Jelinek wrote:
> On Wed, Nov 01, 2017 at 04:23:11PM +, Szabolcs Nagy wrote:
>> is there a way to get vectorized math functions in fortran?
>>
>> in c code there is attribute simd declarations or openmp
>> declare simd pragma to tell the compiler which functions
>> have simd variant, but i see no such thing in fortran.
> 
> !$omp declare simd should work fine in fortran (with -fopenmp
> or -fopenmp-simd).
> 

1) i don't want to change the fortran.

2) it does not work for me.

i want this to call vector powf in libmvec:

subroutine foo(a,b,c)
  real(4) a(8000),b(8000),c(8000)
  do j=1,8000
a(j)=b(j)**c(j)
  end do
end

where do i put

!$omp declare simd (powf)

?

Re: libmvec simd math functions in fortran

[Jakub Jelinek]

On Wed, Nov 01, 2017 at 04:23:11PM +, Szabolcs Nagy wrote:
> is there a way to get vectorized math functions in fortran?
> 
> in c code there is attribute simd declarations or openmp
> declare simd pragma to tell the compiler which functions
> have simd variant, but i see no such thing in fortran.

!$omp declare simd should work fine in fortran (with -fopenmp
or -fopenmp-simd).

Jakub