[R-SIG-Mac] .Machine differences between Intel and M1

[Prof Brian Ripley]

On 21/12/2021 20:46, Simon Urbanek wrote:
> Matt,
> 
> yes, arm64 does not support long doubles. In C the long double type is 64-bit there so has the same precision as doubles (this is allowed by the standard).

And documented in ?.Machine.

However, I see on my M1 Pro

> capabilities("long.double")
long.double
       FALSE

on all the arm64 builds I have installed, including the current CRAN 
distribution (of 4.1.2) and that at mac.r-project.org (see below).  So I 
have no idea why TRUE is reported below (if this really was an arm64 
build run on the M1 Pro).

R version 4.1.2 Patched (2021-12-16 r81394) -- "Bird Hippie"
...
 > capabilities('long.double')
long.double
       FALSE


> Cheers,
> Simon
> 
> 
>> On Dec 22, 2021, at 9:21 AM, Matthew Heun via R-SIG-Mac <r-sig-mac using r-project.org> wrote:
>>
>> All:
>>
>> I'm seeing some test failures on a new M1 Pro machine that I do not see on my Intel machine.  I'm investigating whether the test failures are caused by machine precision differences.  On my M1 Pro machine, differences of large numbers are greater than a specified tolerance.  (On my Intel machine, differences between the supposed same numbers are within tolerance.)

A small number of packages (and R itself) have needed tolerances 
increased for checks on arm64.

>>
>> The output of .Machine shows some differences.  I have 2 questions below, each identified by "****".
>>
>> (1) For sizeof.longdouble, I see the following:
>>
>> Intel machine:
>>
>>> $sizeof.longdouble
>>> [1] 16
>>
>> M1 Pro machine:
>>
>>> $sizeof.longdouble
>>> [1] 8
>>
>>
>>
>> ?.Machine says:
>>
>> sizeof.longdouble	
>> the number of bytes in a C long double type. Will be zero if there is no such type (or its use was disabled when R was built), otherwise possibly12 (most 32-bit builds) or 16 (most 64-bit builds).
>>
>> The M1 Pro uses a 64-bit architecture.  So this result is surprising to me.
>>
>> Furthermore,
>>
>>>> capabilities("long.double")
>>> long.double
>>>        TRUE
>>
>> So somebody thinks that long doubles are supported.
>>
>> **** Is the difference in sizeof.longdouble between the Intel and M1 architectures expected?
>>
>>
>> (2) Also, my M1 Pro machine is missing additional fields (that the Intel machine reports):
>>
>>> $longdouble.eps
>>> [1] 1.084202e-19
>>>
>>> $longdouble.neg.eps
>>> [1] 5.421011e-20
>>>
>>> $longdouble.digits
>>> [1] 64
>>>
>>> $longdouble.rounding
>>> [1] 5
>>>
>>> $longdouble.guard
>>> [1] 0
>>>
>>> $longdouble.ulp.digits
>>> [1] -63
>>>
>>> $longdouble.neg.ulp.digits
>>> [1] -64
>>>
>>> $longdouble.exponent
>>> [1] 15
>>>
>>> $longdouble.min.exp
>>> [1] -16382
>>>
>>> $longdouble.max.exp
>>> [1] 16384
>>
>> **** Is there a reason why the above entries are missing from the output of .Machine on the M1 Pro machine?
>>
>>
>>
>> My R installation is:  4.1.2 Patched (2021/12/16, r81394)
>>
>>
>>
>> Any help will be appreciated.
>>
>>
>>
>> Cheers,
>>
>> Matt


-- 
Brian D. Ripley,                  ripley using stats.ox.ac.uk
Emeritus Professor of Applied Statistics, University of Oxford