Author Topic: Dealing with REAL10 values  (Read 16088 times)

dedndave

  • Member
  • *****
  • Posts: 8827
  • Still using Abacus 2.0
    • DednDave
Re: Dealing with REAL10 values
« Reply #15 on: April 07, 2013, 08:10:50 PM »
i don't know why not - it's the Real McCoy   :P
80-bit real is the "native" format for the FPU

Gunther

  • Member
  • *****
  • Posts: 3585
  • Forgive your enemies, but never forget their names
Re: Dealing with REAL10 values
« Reply #16 on: April 07, 2013, 08:18:47 PM »
Dave,

i don't know why not - it's the Real McCoy   :P
80-bit real is the "native" format for the FPU

absolutely right and there are good reasons to do so. But times are changing, you know. The extended multi media register are now the new kids on the block. Brave new world.

Gunther
Get your facts first, and then you can distort them.

dedndave

  • Member
  • *****
  • Posts: 8827
  • Still using Abacus 2.0
    • DednDave
Re: Dealing with REAL10 values
« Reply #17 on: April 07, 2013, 09:08:33 PM »
i think SSE supports extended reals, no ?

Gunther

  • Member
  • *****
  • Posts: 3585
  • Forgive your enemies, but never forget their names
Re: Dealing with REAL10 values
« Reply #18 on: April 07, 2013, 09:47:58 PM »
i think SSE supports extended reals, no ?

I think no. There's not enough room inside the registers.

Gunther
Get your facts first, and then you can distort them.

MichaelW

  • Global Moderator
  • Member
  • *****
  • Posts: 1209
Re: Dealing with REAL10 values
« Reply #19 on: April 07, 2013, 10:14:12 PM »
Looking at a very recent Intel manual, it looks like the choices are double-precision , single-precision, and half-precision (this only for conversion with single-precision, and with F16C extensions, whatever that is).

Well Microsoft, here’s another nice mess you’ve gotten us into.

Gunther

  • Member
  • *****
  • Posts: 3585
  • Forgive your enemies, but never forget their names
Re: Dealing with REAL10 values
« Reply #20 on: April 07, 2013, 10:58:08 PM »
Michael,

Looking at a very recent Intel manual, it looks like the choices are double-precision , single-precision, and half-precision (this only for conversion with single-precision, and with F16C extensions, whatever that is).

The so called half-precision is that. On the other hand, the Intel C++ compiler on Microsoft Windows supports extended precision, but requires the /Qlong‑double switch for long double to correspond to the hardware's extended precision format. That's a bit funny. One could find some information about that here.

Gunther
Get your facts first, and then you can distort them.

dedndave

  • Member
  • *****
  • Posts: 8827
  • Still using Abacus 2.0
    • DednDave
Re: Dealing with REAL10 values
« Reply #21 on: April 08, 2013, 12:44:44 AM »
according to this paper,

http://www.cs.fsu.edu/~engelen/courses/HPC-adv/FP.pdf

"double extended" is supported, but it does not appear to be quite the same as an FPU extended real

Gunther

  • Member
  • *****
  • Posts: 3585
  • Forgive your enemies, but never forget their names
Re: Dealing with REAL10 values
« Reply #22 on: April 08, 2013, 12:56:56 AM »
Dave,

"double extended" is supported, but it does not appear to be quite the same as an FPU extended real

of course, it's a system dependency. Thank you for the interesting paper. It's a good summary.

Gunther
Get your facts first, and then you can distort them.

anta40

  • Member
  • ***
  • Posts: 312
Re: Dealing with REAL10 values
« Reply #23 on: April 08, 2013, 01:37:26 PM »
There are no long double (REAL10) values in the pure Windows world.

Do you have the Linux version of the code?
I'd like to give it a try.

Gunther

  • Member
  • *****
  • Posts: 3585
  • Forgive your enemies, but never forget their names
Re: Dealing with REAL10 values
« Reply #24 on: April 09, 2013, 02:59:44 AM »
Hi anta40,

Do you have the Linux version of the code?
I'd like to give it a try.

I'll re-compile the source and upload it - let me say - tomorrow. Do you need 32 or 64 bit Linux?

Gunther
Get your facts first, and then you can distort them.

anta40

  • Member
  • ***
  • Posts: 312
Re: Dealing with REAL10 values
« Reply #25 on: April 09, 2013, 12:59:34 PM »
32-bit is fine. But if you don't mind, both 32-bit and 64-bit would be better.
 :biggrin:

Gunther

  • Member
  • *****
  • Posts: 3585
  • Forgive your enemies, but never forget their names
Re: Dealing with REAL10 values
« Reply #26 on: April 09, 2013, 07:28:44 PM »
Hi anta40,

32-bit is fine. But if you don't mind, both 32-bit and 64-bit would be better.
 :biggrin:

okay, I'll re-compile it for both targets. But under Linux isn't such a problem.

Gunther
Get your facts first, and then you can distort them.

Gunther

  • Member
  • *****
  • Posts: 3585
  • Forgive your enemies, but never forget their names
Re: Dealing with REAL10 values
« Reply #27 on: April 11, 2013, 12:20:40 AM »
Hi anta40,

I've uploaded the Linux sources inside the 64 Bit Assembler sub-forum. You can find it here. The archive contains the 64 bit sources as well the 32 bit sources. 32 bit Linux uses the same ABI like Windows 32 bit; so there are only minimal source changes necessary. But under 64 bit systems I had to change the assembly language source, because the underlying ABI is very different from Windows. I hope that helps.

Gunther
Get your facts first, and then you can distort them.