headers
George Harvey | 21 Mar 2011 21:40

DECstation 2100 floating point

Hi,

I've been giving 5.1 a try on my DECstation 2100 (PMIN, R2000 CPU). I
had a bit of trouble finding a disk that the ROM would boot from but
apart from that, the network install ran smoothly. However, running a
few test programs, I'm seeing a lot of floating point errors. For
example, the 'flops' benchmark from pkgsrc returns mostly 'nan' and
'inf' values (see sample output below). If I run the same tests on my
5000/133 (3MIN, R3000 CPU), then they work normally so I'm wondering
what the difference is between R2000 and R3000 floating point behaviour?

For info, dmesg shows the CPU types as follows:

On the 2100 it incorrectly shows the CPU as a R3000:
cpu0 at mainbus0: MIPS R3000 CPU (0x220) Rev. 2.0 with MIPS R3010 FPC
Rev. 2.0
On the 5000/133 it shows:
cpu0 at mainbus0: MIPS R3000A CPU (0x230) Rev. 3.0 with MIPS R3010 FPC
Rev. 4.0

George

=== 'flops' output on DECstation 2100 ===
   FLOPS C Program (Double Precision), V2.0 18 Dec 1992

   Module     Error        RunTime      MFLOPS
                            (usec)
     1             nan     76.7969      0.1823
     2     -1.4667e+01      0.0000         inf
     3             inf -719077252129430822595458597950846748799629
(Continue reading)

Permalink | Reply | 8 comments |
headers
George Harvey | 24 Mar 2011 21:27

Re: DECstation 2100 floating point

On Mon, 21 Mar 2011 20:40:14 +0000
George Harvey <fr30 <at> dial.pipex.com> wrote:

> I've been giving 5.1 a try on my DECstation 2100 (PMIN, R2000 CPU). I
> had a bit of trouble finding a disk that the ROM would boot from but
> apart from that, the network install ran smoothly. However, running a
> few test programs, I'm seeing a lot of floating point errors. For
> example, the 'flops' benchmark from pkgsrc returns mostly 'nan' and
> 'inf' values (see sample output below). If I run the same tests on my
> 5000/133 (3MIN, R3000 CPU), then they work normally so I'm wondering
> what the difference is between R2000 and R3000 floating point behaviour?

A few more details...

If I compile 'flops' with no optimisation then it works on my 2100, if
I use -O or -O2 then it produces 'nan' and 'inf' results. However, if I
copy the binaries over to my 5000/133 then both optimised versions work
normally. So it looks as though GCC is generating code that works on a
R3000/R3010 combo but doesn't work on a R2000/R2010. I've had a hunt
round the 'net and the R2000 and R3000 are supposed to be
instruction-set compatible so I don't understand why I'm seeing
different results.

George
Permalink | Reply | 7 comments |
headers
Simon Burge | 24 Mar 2011 23:00
Picon

Re: DECstation 2100 floating point

George Harvey wrote:

> On Mon, 21 Mar 2011 20:40:14 +0000
> George Harvey <fr30 <at> dial.pipex.com> wrote:
> 
> > I've been giving 5.1 a try on my DECstation 2100 (PMIN, R2000 CPU). I
> > had a bit of trouble finding a disk that the ROM would boot from but
> > apart from that, the network install ran smoothly. However, running a
> > few test programs, I'm seeing a lot of floating point errors. For
> > example, the 'flops' benchmark from pkgsrc returns mostly 'nan' and
> > 'inf' values (see sample output below). If I run the same tests on my
> > 5000/133 (3MIN, R3000 CPU), then they work normally so I'm wondering
> > what the difference is between R2000 and R3000 floating point behaviour?
> 
> A few more details...
> 
> If I compile 'flops' with no optimisation then it works on my 2100, if
> I use -O or -O2 then it produces 'nan' and 'inf' results. However, if I
> copy the binaries over to my 5000/133 then both optimised versions work
> normally. So it looks as though GCC is generating code that works on a
> R3000/R3010 combo but doesn't work on a R2000/R2010. I've had a hunt
> round the 'net and the R2000 and R3000 are supposed to be
> instruction-set compatible so I don't understand why I'm seeing
> different results.

Hopefully you've just proved that the problem isn't inside libc or
the kernel since just changing the compile options of the program
itself gets the right results.

There are some R2000-specific options to gcc.  Can you try the
(Continue reading)

Permalink | Reply | 6 comments |
headers
George Harvey | 25 Mar 2011 22:52

Re: DECstation 2100 floating point

On Fri, 25 Mar 2011 09:00:30 +1100
Simon Burge <simonb <at> NetBSD.org> wrote:

> George Harvey wrote:
> 
> > On Mon, 21 Mar 2011 20:40:14 +0000
> > George Harvey <fr30 <at> dial.pipex.com> wrote:
> > 
> > If I compile 'flops' with no optimisation then it works on my 2100, if
> > I use -O or -O2 then it produces 'nan' and 'inf' results. However, if I
> > copy the binaries over to my 5000/133 then both optimised versions work
> > normally. So it looks as though GCC is generating code that works on a
> > R3000/R3010 combo but doesn't work on a R2000/R2010. 
> 
> Hopefully you've just proved that the problem isn't inside libc or
> the kernel since just changing the compile options of the program
> itself gets the right results.
> 
> There are some R2000-specific options to gcc.  Can you try the
> -march=r2000 and/or -mtune=r2000 options and see if that makes any
> difference?

No difference when applied to 'flops.c'. I tried:

-O
-O -march=r2000
-O -mtune=r2000
-O -march=r2000 -mtune=r2000

And all four produced exactly the same binary.
(Continue reading)

Permalink | Reply | 5 comments |
headers
Izumi Tsutsui | 26 Mar 2011 03:53
Picon
Gravatar

Re: DECstation 2100 floating point

> -O
> -O -march=r2000
> -O -mtune=r2000
> -O -march=r2000 -mtune=r2000
> 
> And all four produced exactly the same binary.

Per src/gnu/dist/gcc4/gcc/config/mips/mips.c,
-march=r2000 is the same as r3000.

As mentioned in gcc(1) man page, -O enables following options:
>>         -O turns on the following optimization flags: -fdefer-pop
>>         -fdelayed-branch -fguess-branch-probability -fcprop-registers
>>         -floop-optimize -fif-conversion -fif-conversion2 -ftree-ccp
>>         -ftree-dce -ftree-dominator-opts -ftree-dse -ftree-ter -ftree-lrs
>>         -ftree-sra -ftree-copyrename -ftree-fre -ftree-ch -funit-at-a-time
>>         -fmerge-constants

It might help to disable each options (by -fno-delayed-branch etc)
and which one causes the errors.

Probably we should also check R2000/R2010 errata.
---
Izumi Tsutsui
Permalink | Reply | 4 comments |
headers
George Harvey | 29 Mar 2011 22:31

Re: DECstation 2100 floating point

On Sat, 26 Mar 2011 11:53:00 +0900
Izumi Tsutsui <tsutsui <at> ceres.dti.ne.jp> wrote:

> As mentioned in gcc(1) man page, -O enables following options:
> >>         -O turns on the following optimization flags: -fdefer-pop
> >>         -fdelayed-branch -fguess-branch-probability -fcprop-registers
> >>         -floop-optimize -fif-conversion -fif-conversion2 -ftree-ccp
> >>         -ftree-dce -ftree-dominator-opts -ftree-dse -ftree-ter -ftree-lrs
> >>         -ftree-sra -ftree-copyrename -ftree-fre -ftree-ch -funit-at-a-time
> >>         -fmerge-constants
> 
> It might help to disable each options (by -fno-delayed-branch etc)
> and which one causes the errors.

I tried turning off each option individually and still got 'inf'
results. I then tried turning them all off, and still got an 'inf'
result. The GCC man page says "only optimizations that have a flag are
listed" so I guess it's one of the unlisted ones that's causing the
problem.

George
Permalink | Reply | 3 comments |
headers
Izumi Tsutsui | 31 Mar 2011 15:51
Picon
Gravatar

Re: DECstation 2100 floating point

> I tried turning off each option individually and still got 'inf'
> results. I then tried turning them all off, and still got an 'inf'
> result. The GCC man page says "only optimizations that have a flag are
> listed" so I guess it's one of the unlisted ones that's causing the
> problem.

I tried pkgsrc/benchmark/flops on NetBSD/cobalt 5.99.44 and
it shows a bit strange result:
---
# ./flops 

   FLOPS C Program (Double Precision), V2.0 18 Dec 1992

   Module     Error        RunTime      MFLOPS
                            (usec)
     1     -2.5193e+01      0.4895     28.5980
     2             nan     -0.0000  -4226415.0942
     3      1.4544e-14     33.3001      0.5105
     4      1.6415e+09     31.9524      0.4694
     5      1.8627e+17     69.2567      0.4187
     6     -1.1475e+87     40.3928      0.7179
     7      5.1090e+02     54.9010      0.2186
     8     -2.9167e-01      0.0000  33103448.4122

   Iterations      =    -640000
   NullTime (usec) =     0.0000
   MFLOPS(1)       =     1.5616
   MFLOPS(2)       =     0.3848
   MFLOPS(3)       =     0.6340
   MFLOPS(4)       =     0.8614
(Continue reading)

Permalink | Reply | 2 comments |
headers
George Harvey | 31 Mar 2011 22:47

Re: DECstation 2100 floating point

On Thu, 31 Mar 2011 22:51:35 +0900
Izumi Tsutsui <tsutsui <at> ceres.dti.ne.jp> wrote:

> I tried pkgsrc/benchmark/flops on NetBSD/cobalt 5.99.44 and
> it shows a bit strange result:

That's odd, on my RaQ 2, running NetBSD/Cobalt 5.0.2, with flops
installed from pkgsrc, I get normal-looking results:

-bash-2.05b$ which flops
/usr/pkg/bin/flops
-bash-2.05b$ flops

   FLOPS C Program (Double Precision), V2.0 18 Dec 1992

   Module     Error        RunTime      MFLOPS
                            (usec)
     1      1.3358e-12      0.2935     47.7073
     2      2.0517e-13      0.1487     47.0793
     3      1.7542e-14      0.2976     57.1277
     4     -5.4512e-14      0.2855     52.5427
     5      3.3307e-16      0.4783     60.6303
     6     -1.9040e-14      0.3537     81.9794
     7      2.6034e-11      0.4904     24.4680
     8     -5.4068e-14      0.3737     80.2754

   Iterations      =   64000000
   NullTime (usec) =     0.0000
   MFLOPS(1)       =    49.9517
   MFLOPS(2)       =    41.4133
(Continue reading)

Permalink | Reply | 1 comment |

Gmane