[Qemu-devel] PowerPC Decimal Floating Point

[Qemu-devel] PowerPC Decimal Floating Point

[Tom Musta]

QEMU is currently lacking support for the Decimal Floating Point category of instructions
(see Power ISA V2.07 Chapter 6).  I intend to add this support.

Has this topic been investigated before and if so can anyone point me to that discussion?
A Google search came up empty for me.

My proposal is to incorporate the libdecnumber component of libdfp
(http://www.eglibc.org/cgi-bin/viewvc.cgi/libdfp/trunk/) in a manner analogous to how
softfloat is used for binary floating point.  So, for example, the helper for the dadd
instruction would look something like the following:

   - map FPSCR state to a decContext.
   - convert the contents of the source FPRs to decNumbers (decimal64ToNumber).
   - call decNumberAdd
   - convert the resultant decNumber to DPD (decimal64FromNumber)
   - update FPSCR per the decContext.status and result.

Comments?

Re: [Qemu-devel] [Qemu-ppc] PowerPC Decimal Floating Point

[Alexander Graf]

On 11.02.2014, at 18:14, Tom Musta <tommusta@gmail.com> wrote:

> QEMU is currently lacking support for the Decimal Floating Point category of instructions
> (see Power ISA V2.07 Chapter 6).  I intend to add this support.
> 
> Has this topic been investigated before and if so can anyone point me to that discussion?
> A Google search came up empty for me.
> 
> My proposal is to incorporate the libdecnumber component of libdfp
> (http://www.eglibc.org/cgi-bin/viewvc.cgi/libdfp/trunk/) in a manner analogous to how
> softfloat is used for binary floating point.  So, for example, the helper for the dadd
> instruction would look something like the following:
> 
>   - map FPSCR state to a decContext.
>   - convert the contents of the source FPRs to decNumbers (decimal64ToNumber).
>   - call decNumberAdd
>   - convert the resultant decNumber to DPD (decimal64FromNumber)
>   - update FPSCR per the decContext.status and result.
> 
> Comments?

I think that approach makes a lot of sense, but let's ask Richard and Peter as well.


Alex

Re: [Qemu-devel] [Qemu-ppc] PowerPC Decimal Floating Point

[Richard Henderson]

On 02/11/2014 11:38 PM, Alexander Graf wrote:
>> > My proposal is to incorporate the libdecnumber component of libdfp
>> > (http://www.eglibc.org/cgi-bin/viewvc.cgi/libdfp/trunk/) in a manner analogous to how
>> > softfloat is used for binary floating point.  So, for example, the helper for the dadd
>> > instruction would look something like the following:
>> > 
>> >   - map FPSCR state to a decContext.
>> >   - convert the contents of the source FPRs to decNumbers (decimal64ToNumber).
>> >   - call decNumberAdd
>> >   - convert the resultant decNumber to DPD (decimal64FromNumber)
>> >   - update FPSCR per the decContext.status and result.
>> > 
>> > Comments?
> I think that approach makes a lot of sense, but let's ask Richard and Peter as well.

Reasonable, as far as the implementation details go.

I am a teeny bit concerned which version of libdecnumber is considered most
"upstream".  Ordinarily I'd point to the copy in gcc, but that is of course
GPLv3 + GCC runtime exception.  I see the version from eglibc you quote above
is LGPLv2.1.  I also see that the last change is 18 months old.  On the other
hand, there is only one non-autoconf change in the gcc sources in the same period.

The original import is from IBM sources.  Is there somewhere "upstream" at IBM
that we ought to be importing from instead?


r~

Re: [Qemu-devel] [Qemu-ppc] PowerPC Decimal Floating Point

[Tom Musta]

On 2/14/2014 3:22 PM, Richard Henderson wrote:
> On 02/11/2014 11:38 PM, Alexander Graf wrote:
>>>> My proposal is to incorporate the libdecnumber component of libdfp
>>>> (http://www.eglibc.org/cgi-bin/viewvc.cgi/libdfp/trunk/) in a manner analogous to how
>>>> softfloat is used for binary floating point.  So, for example, the helper for the dadd
>>>> instruction would look something like the following:
>>>>
>>>>   - map FPSCR state to a decContext.
>>>>   - convert the contents of the source FPRs to decNumbers (decimal64ToNumber).
>>>>   - call decNumberAdd
>>>>   - convert the resultant decNumber to DPD (decimal64FromNumber)
>>>>   - update FPSCR per the decContext.status and result.
>>>>
>>>> Comments?
>> I think that approach makes a lot of sense, but let's ask Richard and Peter as well.
> 
> Reasonable, as far as the implementation details go.
> 
> I am a teeny bit concerned which version of libdecnumber is considered most
> "upstream".  Ordinarily I'd point to the copy in gcc, but that is of course
> GPLv3 + GCC runtime exception.  I see the version from eglibc you quote above
> is LGPLv2.1.  I also see that the last change is 18 months old.  On the other
> hand, there is only one non-autoconf change in the gcc sources in the same period.
> 
> The original import is from IBM sources.  Is there somewhere "upstream" at IBM
> that we ought to be importing from instead?
> 

Richard:

I will check with folks inside of IBM here that I have the latest and greatest.

FWIW ... I have about 20 instructions prototyped and running in a test harness
(not QEMU) and comparing results against actual P7/P8 hardware.  So far, this
looks pretty good.  I haven't yet worked through any details of importing the library
into QEMU.