GNU bug report logs - #68871
Can't use od to print half-precision floats

Previous Next

To add a comment to this bug, you must first unarchive it, by sending
a message to control AT debbugs.gnu.org, with unarchive 68871 in the body.
You can then email your comments to 68871 AT debbugs.gnu.org in the normal way.

Toggle the display of automated, internal messages from the tracker.

Message #5 received at submit <at> debbugs.gnu.org (full text, mbox):

From: "Rozenberg, Eyal (Consultant)" <Eyal.Rozenberg <at> gehealthcare.com>
To: "bug-coreutils <at> gnu.org" <bug-coreutils <at> gnu.org>
Subject: Can't use od to print half-precision floats
Date: Thu, 1 Feb 2024 12:43:47 +0000

[Message part 1 (text/plain, inline)]

If I try to use the od utility to print half-precision (FP16) floating-point values, I get:

$ od -t f2 myfloats.bin
od: invalid type string 'f2';
this system doesn't provide a 2-byte floating point type

I'm not exactly sure what "this system" means, but that should work and print out my floats.

Eyal

PS - This is my first bug-coreutils post, please be gentle.

[Message part 2 (text/html, inline)]

Message #8 received at 68871 <at> debbugs.gnu.org (full text, mbox):

From: Pádraig Brady <P <at> draigBrady.com>
To: "Rozenberg, Eyal (Consultant)" <Eyal.Rozenberg <at> gehealthcare.com>,
 68871 <at> debbugs.gnu.org
Subject: Re: bug#68871: Can't use od to print half-precision floats
Date: Thu, 1 Feb 2024 18:05:10 +0000

[Message part 1 (text/plain, inline)]

On 01/02/2024 12:43, Rozenberg, Eyal (Consultant) wrote:
> If I try to use the od utility to print half-precision (FP16) floating-point values, I get:
> 
> $ od -t f2 myfloats.bin
> od: invalid type string 'f2';
> this system doesn't provide a 2-byte floating point type
> 
> I'm not exactly sure what "this system" means, but that should work and print out my floats.
> 
> Eyal
> 
> PS - This is my first bug-coreutils post, please be gentle.

I just had a read of these:
https://gcc.gnu.org/onlinedocs/gcc/Half-Precision.html
https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0192r4.html

On the face of it it seems that various 16 bit floating point formats have consolidated,
and we could add support for _Float16 in od.
I see gnulib already enables __STDC_WANT_IEC_60559_TYPES_EXT__
so we should be able to do something like the attached.

This is just off the top of my head, and I haven't tested
or thought about this much at all.
Any testing your could do would be appreciated.

thanks,
Pádraig

[od-f16.patch (text/x-patch, attachment)]

Message #11 received at 68871 <at> debbugs.gnu.org (full text, mbox):

From: Paul Eggert <eggert <at> cs.ucla.edu>
To: Pádraig Brady <P <at> draigBrady.com>,
 "Rozenberg, Eyal (Consultant)" <Eyal.Rozenberg <at> gehealthcare.com>,
 68871 <at> debbugs.gnu.org
Subject: Re: bug#68871: Can't use od to print half-precision floats
Date: Thu, 1 Feb 2024 11:08:18 -0800

Thanks for writing that. One suggestion I'd make is to change this:

 #ifndef FLT16_MAX
 /* This is just a place-holder to avoid a few '#if' directives.
    In this case, the type isn't actually used.  */
 typedef float _Float16;
 #endif

to something like this:

 #ifdef FLT16_MAX
 typedef _Float16 float16;
 #else
 /* This is just a place-holder to avoid a few '#if' directives.
    In this case, the type isn't actually used.  */
 typedef float float16;
 #endif

and use 'float16' thereafter. That way, the code doesn't usurp the 
system namespace on older platforms, some of which might define _Float16 
but not FLT16_MAX.

Message #14 received at 68871 <at> debbugs.gnu.org (full text, mbox):

From: Paul Eggert <eggert <at> cs.ucla.edu>
To: Pádraig Brady <P <at> draigBrady.com>,
 "Rozenberg, Eyal (Consultant)" <Eyal.Rozenberg <at> gehealthcare.com>,
 68871 <at> debbugs.gnu.org
Subject: Re: bug#68871: Can't use od to print half-precision floats
Date: Thu, 1 Feb 2024 11:16:42 -0800

Oh, and another thought: suppose someone wants to use od on bfloat16_t 
values? They're popular in machine learning applications, and likely 
will be more popular than float16_t overall. See:

https://sourceware.org/pipermail/libc-alpha/2024-February/154382.html

Message #17 received at 68871 <at> debbugs.gnu.org (full text, mbox):

From: Pádraig Brady <P <at> draigBrady.com>
To: Paul Eggert <eggert <at> cs.ucla.edu>,
 "Rozenberg, Eyal (Consultant)" <Eyal.Rozenberg <at> gehealthcare.com>,
 68871 <at> debbugs.gnu.org
Subject: Re: bug#68871: Can't use od to print half-precision floats
Date: Thu, 1 Feb 2024 21:59:29 +0000

On 01/02/2024 19:16, Paul Eggert wrote:
> Oh, and another thought: suppose someone wants to use od on bfloat16_t
> values? They're popular in machine learning applications, and likely
> will be more popular than float16_t overall. See:
> 
> https://sourceware.org/pipermail/libc-alpha/2024-February/154382.html

True. I suppose we would select between these like:

 -t f2, -t fh = IEEE half precision
        -t fb = brain floating point

bfloat16 looks like a truncated single precision IEEE,
so we should be able to just pad the extra 16 bits with zeros
when converting to single precision internally for processing.

cheers,
Pádraig

Message #20 received at 68871 <at> debbugs.gnu.org (full text, mbox):

From: Paul Eggert <eggert <at> cs.ucla.edu>
To: Pádraig Brady <P <at> draigBrady.com>,
 "Rozenberg, Eyal (Consultant)" <Eyal.Rozenberg <at> gehealthcare.com>,
 68871 <at> debbugs.gnu.org
Subject: Re: bug#68871: Can't use od to print half-precision floats
Date: Thu, 1 Feb 2024 17:47:29 -0800

On 2/1/24 13:59, Pádraig Brady wrote:
> 
> bfloat16 looks like a truncated single precision IEEE,
> so we should be able to just pad the extra 16 bits with zeros
> when converting to single precision internally for processing.

Sounds good. This would mean od could work even the platform doesn't 
support bfloat16_t, since od.c could fall back on the above code (though 
I suppose it could be endianness-dependent).

Message #23 received at 68871 <at> debbugs.gnu.org (full text, mbox):

From: "Rozenberg, Eyal (Consultant)" <Eyal.Rozenberg <at> gehealthcare.com>
To: Paul Eggert <eggert <at> cs.ucla.edu>, Pádraig Brady
 <P <at> draigBrady.com>, "68871 <at> debbugs.gnu.org" <68871 <at> debbugs.gnu.org>
Subject: RE: bug#68871: Can't use od to print half-precision floats
Date: Fri, 2 Feb 2024 14:35:31 +0000

Seems ok to me as a user. Note also that there is the TF32 format (tf for tensor-float), which takes up 32 bits, but only uses 19. So, the switches could be:

-t f2    IEEE half
-t fh    IEEE half
-t fb    brain
-t ft     tensor

or

-t ft4  tensor

(which indicates the 32-bitness)

Eyal

-----Original Message-----
From: Paul Eggert <eggert <at> cs.ucla.edu> 
Sent: Friday, 2 February 2024 3:47
To: Pádraig Brady <P <at> draigBrady.com>; Rozenberg, Eyal (Consultant) <Eyal.Rozenberg <at> gehealthcare.com>; 68871 <at> debbugs.gnu.org
Subject: Re: bug#68871: Can't use od to print half-precision floats

WARNING: This email originated from outside of GE HealthCare. Please validate the sender's email address before clicking on links or attachments as they may not be safe.

On 2/1/24 13:59, Pádraig Brady wrote:
>
> bfloat16 looks like a truncated single precision IEEE, so we should be 
> able to just pad the extra 16 bits with zeros when converting to 
> single precision internally for processing.

Sounds good. This would mean od could work even the platform doesn't support bfloat16_t, since od.c could fall back on the above code (though I suppose it could be endianness-dependent).

Message #28 received at 68871-done <at> debbugs.gnu.org (full text, mbox):

From: Pádraig Brady <P <at> draigBrady.com>
To: Paul Eggert <eggert <at> cs.ucla.edu>,
 "Rozenberg, Eyal (Consultant)" <Eyal.Rozenberg <at> gehealthcare.com>,
 68871-done <at> debbugs.gnu.org
Subject: Re: bug#68871: Can't use od to print half-precision floats
Date: Sun, 4 Feb 2024 15:00:19 +0000

[Message part 1 (text/plain, inline)]

On 02/02/2024 01:47, Paul Eggert wrote:
> On 2/1/24 13:59, Pádraig Brady wrote:
>>
>> bfloat16 looks like a truncated single precision IEEE,
>> so we should be able to just pad the extra 16 bits with zeros
>> when converting to single precision internally for processing.
> 
> Sounds good. This would mean od could work even the platform doesn't
> support bfloat16_t, since od.c could fall back on the above code (though
> I suppose it could be endianness-dependent).

I see that __bf16 is supported by released versions of gcc and clang,
so rather than add conversion complexity to the core od print loop,
the attached relies on compiler support for _Float16 and __bf16 types,
which compilers will expand to more targets going forward.

I tested the attached on:

gcc 13, clang 17 x86 (Both types supported)
gcc 7 aarch64 (Only -fH supported)
gcc 13 ppc(be) (Neither supported (both will be with GCC 14))

I'll commit this later.
Marking this bug as done.

thanks,
Pádraig

[od-half-float.patch (text/x-patch, attachment)]

Message #31 received at 68871 <at> debbugs.gnu.org (full text, mbox):

From: Paul Eggert <eggert <at> cs.ucla.edu>
To: Pádraig Brady <P <at> draigBrady.com>,
 "Rozenberg, Eyal (Consultant)" <Eyal.Rozenberg <at> gehealthcare.com>,
 68871 <at> debbugs.gnu.org
Subject: Re: bug#68871: Can't use od to print half-precision floats
Date: Sun, 4 Feb 2024 13:59:17 -0800

Thanks. One minor comment about this:

> +@item B
> +brain 16 bit float
> +@item H
> +half precision float

It might be helpful explain these two formats, e.g., to cite:

https://en.wikipedia.org/wiki/Bfloat16_floating-point_format

and

https://en.wikipedia.org/wiki/Half-precision_floating-point_format

respectively, since the formats are new compared to the other formats 
being discussed.

Message #34 received at 68871 <at> debbugs.gnu.org (full text, mbox):

From: Pádraig Brady <P <at> draigBrady.com>
To: Paul Eggert <eggert <at> cs.ucla.edu>,
 "Rozenberg, Eyal (Consultant)" <Eyal.Rozenberg <at> gehealthcare.com>,
 68871 <at> debbugs.gnu.org
Subject: Re: bug#68871: Can't use od to print half-precision floats
Date: Mon, 5 Feb 2024 13:37:09 +0000

On 04/02/2024 21:59, Paul Eggert wrote:
> Thanks. One minor comment about this:
> 
>> +@item B
>> +brain 16 bit float
>> +@item H
>> +half precision float
> 
> It might be helpful explain these two formats, e.g., to cite:
> 
> https://en.wikipedia.org/wiki/Bfloat16_floating-point_format
> 
> and
> 
> https://en.wikipedia.org/wiki/Half-precision_floating-point_format
> 
> respectively, since the formats are new compared to the other formats
> being discussed.

Good call.
I added those references and pushed.

cheers,
Pádraig.

Previous Next

GNU bug tracking system
Copyright (C) 1999 Darren O. Benham, 1997,2003 nCipher Corporation Ltd, 1994-97 Ian Jackson.