On Mar 3, 2013 1:58 PM, "Manu" <turkeyman@gmail.com> wrote:
>
> Hi people.
>
> So I've been AWOL a while, and I'll still be intermittent for some weeks
yet, but I'm trying to work on std.simd and also my dconf talks in any gaps I can.
> However, in order to do a thorough job, I'm missing some tools. I wonder
if anyone can help me out...?
>
> I need to do some rather strenuous testing of SIMD code gen in D, to make
sure I don't make mistakes and/or lie.
>
> This requires testing against as many compilers+architectures as I can. I have barely any internet access to speak of, I don't really know how,
and I also don't run linux, so I can't practically build myself a suite of toolchains to work with >_<
>
> Ideally, I would really like all these toolchains:
>  GDC: x86, x64, ARM, PPC, MIPS, SH4 <- common toolchains

Do you actually have boxes for these architectures?

Regards
-- 
Iain Buclaw

*(p < e ? p++ : p) = (c & 0x0f) + '0';

Yes, all of them (and more, bonus points for PS3's Cell PPE/SPU
toolchains!) :)
But I don't really need to run it... I just tend to compile with -S. I know
what the resulting code should look like from years of experience.
I can easily verify the proper opcodes are emitted, register assignment is
sensible and function arguments are passed correctly just by looking at the
assembly.

I could probably get you a bare bones compiler (no library) - it *will* be Linux though...  I can give you a login to a server if you like to play with it.  Otherwise I could give cygwin a try...

Regards
-- 
Iain Buclaw

*(p < e ? p++ : p) = (c & 0x0f) + '0';
On Mar 3, 2013 2:58 PM, "Manu" <turkeyman@gmail.com> wrote:

> Yes, all of them (and more, bonus points for PS3's Cell PPE/SPU
> toolchains!) :)
> But I don't really need to run it... I just tend to compile with -S. I
> know what the resulting code should look like from years of experience.
> I can easily verify the proper opcodes are emitted, register assignment is
> sensible and function arguments are passed correctly just by looking at the
> assembly.
>

How does Dan build the MinGW toolchains?
No library is fine for my purposes.
I don't have access to a linux machine atm, but I should do in a week or 2.
If that's the only option, then I can wait till then... just leaves me with
a bit less time.


On 4 March 2013 02:57, Iain Buclaw <ibuclaw@ubuntu.com> wrote:

> I could probably get you a bare bones compiler (no library) - it *will* be Linux though...  I can give you a login to a server if you like to play with it.  Otherwise I could give cygwin a try...
>
> Regards
> --
> Iain Buclaw
>
> *(p < e ? p++ : p) = (c & 0x0f) + '0';
> On Mar 3, 2013 2:58 PM, "Manu" <turkeyman@gmail.com> wrote:
>
>> Yes, all of them (and more, bonus points for PS3's Cell PPE/SPU
>> toolchains!) :)
>> But I don't really need to run it... I just tend to compile with -S. I
>> know what the resulting code should look like from years of experience.
>> I can easily verify the proper opcodes are emitted, register assignment
>> is sensible and function arguments are passed correctly just by looking at
>> the assembly.
>>
>

On Mar 3, 2013 10:48 PM, "Manu" <turkeyman@gmail.com> wrote:
>
> How does Dan build the MinGW toolchains?
> No library is fine for my purposes.
> I don't have access to a linux machine atm, but I should do in a week or
2. If that's the only option, then I can wait till then... just leaves me with a bit less time.
>
>
> On 4 March 2013 02:57, Iain Buclaw <ibuclaw@ubuntu.com> wrote:
>>
>> I could probably get you a bare bones compiler (no library) - it *will*
be Linux though...  I can give you a login to a server if you like to play with it.  Otherwise I could give cygwin a try...
>>
>> Regards
>> --
>> Iain Buclaw
>>
>> *(p < e ? p++ : p) = (c & 0x0f) + '0';
>>
>> On Mar 3, 2013 2:58 PM, "Manu" <turkeyman@gmail.com> wrote:
>>>
>>> Yes, all of them (and more, bonus points for PS3's Cell PPE/SPU
toolchains!) :)
>>> But I don't really need to run it... I just tend to compile with -S. I
know what the resulting code should look like from years of experience.
>>> I can easily verify the proper opcodes are emitted, register assignment
is sensible and function arguments are passed correctly just by looking at the assembly.
>
>

Won't have access to a windows machine until Wednesday.  ;)

As I said, I can give you a login account to a Linux server with these on.   Should be ready Tuesday ish.

Regards
-- 
Iain Buclaw

*(p < e ? p++ : p) = (c & 0x0f) + '0';

Cool, well that'll definitely help! :)


On 4 March 2013 09:02, Iain Buclaw <ibuclaw@ubuntu.com> wrote:

>
> On Mar 3, 2013 10:48 PM, "Manu" <turkeyman@gmail.com> wrote:
> >
> > How does Dan build the MinGW toolchains?
> > No library is fine for my purposes.
> > I don't have access to a linux machine atm, but I should do in a week or
> 2. If that's the only option, then I can wait till then... just leaves me with a bit less time.
> >
> >
> > On 4 March 2013 02:57, Iain Buclaw <ibuclaw@ubuntu.com> wrote:
> >>
> >> I could probably get you a bare bones compiler (no library) - it *will*
> be Linux though...  I can give you a login to a server if you like to play with it.  Otherwise I could give cygwin a try...
> >>
> >> Regards
> >> --
> >> Iain Buclaw
> >>
> >> *(p < e ? p++ : p) = (c & 0x0f) + '0';
> >>
> >> On Mar 3, 2013 2:58 PM, "Manu" <turkeyman@gmail.com> wrote:
> >>>
> >>> Yes, all of them (and more, bonus points for PS3's Cell PPE/SPU
> toolchains!) :)
> >>> But I don't really need to run it... I just tend to compile with -S. I
> know what the resulting code should look like from years of experience.
> >>> I can easily verify the proper opcodes are emitted, register
> assignment is sensible and function arguments are passed correctly just by looking at the assembly.
> >
> >
>
> Won't have access to a windows machine until Wednesday.  ;)
>
> As I said, I can give you a login account to a Linux server with these on.   Should be ready Tuesday ish.
>
> Regards
> --
> Iain Buclaw
>
> *(p < e ? p++ : p) = (c & 0x0f) + '0';
>

...it occurs to me that MIPS doesn't really have a single standardised SIMD instruction set.

There are a few extensions. I've worked on 3 popular MIPS devices; they all
have SIMD, but they are all different.
I'm not sure they are available as options in the standard GCC MIPS arch. I
think they are added via vendor patches.

So don't sorry about MIPS I guess, it sounds like too much effort! I'll hassle the vendors some time in the future...


On 4 March 2013 10:37, Manu <turkeyman@gmail.com> wrote:

> Cool, well that'll definitely help! :)
>
>
> On 4 March 2013 09:02, Iain Buclaw <ibuclaw@ubuntu.com> wrote:
>
>>
>> On Mar 3, 2013 10:48 PM, "Manu" <turkeyman@gmail.com> wrote:
>> >
>> > How does Dan build the MinGW toolchains?
>> > No library is fine for my purposes.
>> > I don't have access to a linux machine atm, but I should do in a week
>> or 2. If that's the only option, then I can wait till then... just leaves me with a bit less time.
>> >
>> >
>> > On 4 March 2013 02:57, Iain Buclaw <ibuclaw@ubuntu.com> wrote:
>> >>
>> >> I could probably get you a bare bones compiler (no library) - it
>> *will* be Linux though...  I can give you a login to a server if you like to play with it.  Otherwise I could give cygwin a try...
>> >>
>> >> Regards
>> >> --
>> >> Iain Buclaw
>> >>
>> >> *(p < e ? p++ : p) = (c & 0x0f) + '0';
>> >>
>> >> On Mar 3, 2013 2:58 PM, "Manu" <turkeyman@gmail.com> wrote:
>> >>>
>> >>> Yes, all of them (and more, bonus points for PS3's Cell PPE/SPU
>> toolchains!) :)
>> >>> But I don't really need to run it... I just tend to compile with -S.
>> I know what the resulting code should look like from years of experience.
>> >>> I can easily verify the proper opcodes are emitted, register
>> assignment is sensible and function arguments are passed correctly just by looking at the assembly.
>> >
>> >
>>
>> Won't have access to a windows machine until Wednesday.  ;)
>>
>> As I said, I can give you a login account to a Linux server with these on.   Should be ready Tuesday ish.
>>
>> Regards
>> --
>> Iain Buclaw
>>
>> *(p < e ? p++ : p) = (c & 0x0f) + '0';
>>
>
>

On Mar 4, 2013 12:42 AM, "Manu" <turkeyman@gmail.com> wrote:
>
> ...it occurs to me that MIPS doesn't really have a single standardised
SIMD instruction set.
>
> There are a few extensions. I've worked on 3 popular MIPS devices; they
all have SIMD, but they are all different.
> I'm not sure they are available as options in the standard GCC MIPS arch.
I think they are added via vendor patches.
>
> So don't sorry about MIPS I guess, it sounds like too much effort! I'll
hassle the vendors some time in the future...
>
>

I can check what the difference is in gcc mips backend later this week... but that is your call.

Regards
-- 
Iain Buclaw

*(p < e ? p++ : p) = (c & 0x0f) + '0';

Am Mon, 4 Mar 2013 10:41:08 +1000
schrieb Manu <turkeyman@gmail.com>:

> ...it occurs to me that MIPS doesn't really have a single standardised SIMD instruction set.
> 
> There are a few extensions. I've worked on 3 popular MIPS devices;
> they all have SIMD, but they are all different.
> I'm not sure they are available as options in the standard GCC MIPS
> arch. I think they are added via vendor patches.
> 
> So don't sorry about MIPS I guess, it sounds like too much effort! I'll hassle the vendors some time in the future...
> 

Is MIPS ASE what you need? Standard gcc supports:

MIPS DSP ASE rev1
MIPS DSP ASE rev2
MIPS SmartMIPS ASE
MIPS Digital Media Extension
MIPS-3D ASE
(optional paired-single floating-point instructions)
http://gcc.gnu.org/onlinedocs/gcc-4.7.2/gcc/MIPS-Options.html

If you have vendor patches for gcc-4.7 I could integrate those. But we need either gcc 4.7 or 4.8 patches for gdc, crosstool-NG+gdc actually only works with 4.7

MIPS-3D and DMX are allegedly common, but not in any device I own (although
I'd like to consider them in my work all the same).
Popular MIPS devices are:
  Android
    Ingenic XBurst is in many existing Android devices, and has its own
SIMD unit.
    MIPS64 chips should be appearing in Android devices soon (which should
support MIPS-3D)
    Google has a toolchain build script
  Sony PSP
    Has a custom (and AWESOME!) SIMD unit
    Toolchain with vendor extensions is(/was) easy to build:
https://github.com/pspdev/psptoolchain
  Sony PS2 is a very interesting research device, but it's old and
irrelevant now
    The toolchain is quite hard to build :( (
https://github.com/ps2dev/ps2toolchain)

While I'm digging up the console build scripts, there are some other well
maintained scripts:
  Sony PS3 - https://github.com/ps3dev/ps3toolchain
    Builds PPE (PPC arch) and SPU (dedicated super-fast SIMD coprocessor,
best simd unit on earth!) toolchains
  Nintendo WiiU, Wii, Gamecube - https://github.com/Vithon/devkitppcsh
    PPC system, but with paired single unit instead of standard VMX

On 4 March 2013 20:54, Johannes Pfau <nospam@example.com> wrote:

> Am Mon, 4 Mar 2013 10:41:08 +1000
> schrieb Manu <turkeyman@gmail.com>:
>
> > ...it occurs to me that MIPS doesn't really have a single standardised SIMD instruction set.
> >
> > There are a few extensions. I've worked on 3 popular MIPS devices;
> > they all have SIMD, but they are all different.
> > I'm not sure they are available as options in the standard GCC MIPS
> > arch. I think they are added via vendor patches.
> >
> > So don't sorry about MIPS I guess, it sounds like too much effort! I'll hassle the vendors some time in the future...
> >
>
> Is MIPS ASE what you need? Standard gcc supports:
>
> MIPS DSP ASE rev1
> MIPS DSP ASE rev2
> MIPS SmartMIPS ASE
> MIPS Digital Media Extension
> MIPS-3D ASE
> (optional paired-single floating-point instructions)
> http://gcc.gnu.org/onlinedocs/gcc-4.7.2/gcc/MIPS-Options.html
>
> If you have vendor patches for gcc-4.7 I could integrate those. But we need either gcc 4.7 or 4.8 patches for gdc, crosstool-NG+gdc actually only works with 4.7
>