[Openmcl-devel] Exposing math functions
martin
brooks.martin at sympatico.ca
Fri Jun 19 18:25:34 PDT 2020
Hi Steve
I didn’t implement exact arithmetic with foreign functions, but I tried.
I spoke with a numerical specialist; they had fast GCD code, but only up to some fixed extended number range, based on a fixed longer representation.
They couldn’t believe that I would ever need more rational precision than that, but suggested that I could put together a C library that overflows to Lisp.
I did some random observations of numbers in my app; I saw many samples in the Lisp-only bignum range.
QED I stayed on the Lisp track.
Martin
> On Jun 19, 2020, at 8:55 PM, Steven Nunez <steve_nunez at yahoo.com> wrote:
>
> Martin, did you do any benchmarks on the difference between, say, CFFI and using the internal code as a template?
>
> On Friday, June 19, 2020, 8:32:22 AM GMT+8, martin <brooks.martin at sympatico.ca> wrote:
>
>
> Hello CCL fans, I will relate a somewhat parallel story.
> My application requires exact arithmetic, i.e. bignum rationals.
> CCL 11 conses like crazy for bignum rationals, whereas CCL 12 has a specialized implementation.
> This happened because I engaged Clozure Associates to solve my bignum rational problem, which they totally did.
> Prior to engaging CA I raided the CCL code to use the low-level allocation and arithmetic %functions, as in the examples below.
>
> Martin
>
>
>> On Jun 18, 2020, at 7:56 PM, Steven Nunez <steve_nunez at yahoo.com <mailto:steve_nunez at yahoo.com>> wrote:
>>
>> Yes, that's been my experience as well. Part of the reason for the question was to try to gain insights from anyone that might have done something similar and to understand why the code for CCL was written as it is. There's some reason why Gary B et. al. did things the way they did, and knowing that will help in my coding efforts.
>>
>>
>>
>>
>> On Friday, June 19, 2020, 5:01:20 AM GMT+8, Ron Garret <ron at flownet.com <mailto:ron at flownet.com>> wrote:
>>
>>
>> The best way to answer performance questions is just to run some benchmarks.
>>
>> On Jun 18, 2020, at 1:58 AM, Steven Nunez <steve_nunez at yahoo.com <mailto:steve_nunez at yahoo.com>> wrote:
>>
>>> Greetings all,
>>>
>>> I've got need of a few functions that are in the standard C library, but not in common lisp. Since the functions are there, but not available, I set about wrapping them. Looking at the code for atanh, I was surprised that CCL's wrapping is so complicated. Now I get that the code, pasted below, has to cater for complex numbers, multiple architectures and the like, so I can likely simplify it for my use case in one architecture and the real domain.
>>>
>>> So, my question is: is this best practice for getting good performance for a wrapper? In Allegro, where I don't have access to the code, things are much simpler, using their FFI, e.g.
>>>
>>> (ff:def-foreign-call (%erf "erf") ((x :double))
>>> :returning :double)
>>> (ff:def-foreign-call (%erff "erff") ((x :float))
>>> :returning :float)
>>> (defun erf (x)
>>> "Returns the error function value for x"
>>> (cond ((typep x 'double-float) (%erf x))
>>> ((typep x 'single-float) (%erff x))))
>>>
>>>
>>> Whereas in the CCL code there's %setf-double-float, and target::with-stack-double-floats, all of which seem undocumented, but look like they're doing some magic to keep things fast and/or efficient. These functions are already in the database, so it's tempting to use the #_ macro to access them, but then I wonder why the CCL guys didn't do this.
>>>
>>> Anyone got any ideas here? I guess I'm wondering as well whether or not I'll get any performance benefit from following the CCL pattern over a simpler approach.
>>>
>>>
>>> #+windows-target
>>> (progn
>>> (defun %double-float-atanh! (n result)
>>> (declare (double-float n result))
>>> (with-stack-double-floats ((temp))
>>> (%setf-double-float TEMP (external-call "atanh" :double n :double))
>>> (%df-check-exception-1 'atanh n (%ffi-exception-status))
>>> (%setf-double-float result TEMP)))
>>>
>>> #+32-bit-target
>>> (defun %single-float-atanh! (n result)
>>> (declare (single-float n result))
>>> (target::with-stack-short-floats ((temp))
>>> #+arm-target (%set-fpscr-status 0)
>>> (%setf-short-float TEMP (external-call "atanhf" :float n :float))
>>> (%sf-check-exception-1 'atanh n (%ffi-exception-status))
>>> (%setf-short-float result TEMP)))
>>>
>>> #+64-bit-target
>>> (defun %single-float-atanh (n)
>>> (declare (single-float n))
>>> (let* ((result (external-call "atanhf" :float n :float)))
>>> (%sf-check-exception-1 'atanh n (%ffi-exception-status))
>>> result)))
>>>
>>> (defun atan (y &optional (x nil x-p))
>>> "Return the arc tangent of Y if X is omitted or Y/X if X is supplied."
>>> (cond (x-p
>>> (cond ((or (typep x 'double-float)
>>> (typep y 'double-float))
>>> (with-stack-double-floats ((dy y)
>>> (dx x))
>>> (%df-atan2 dy dx)))
>>> (t
>>> (when (and (rationalp x) (rationalp y))
>>> ;; rescale arguments so that the maximum absolute value is 1
>>> (let ((x1 (abs x)) (y1 (abs y)))
>>> (cond ((> y1 x1)
>>> (setf x (/ x y1))
>>> (setf y (signum y)))
>>> ((not (zerop x))
>>> (setf y (/ y x1))
>>> (setf x (signum x))))))
>>> #+32-bit-target
>>> (target::with-stack-short-floats ((sy y)
>>> (sx x))
>>> (%sf-atan2! sy sx))
>>> #+64-bit-target
>>> (%sf-atan2 (%short-float y) (%short-float x)))))
>>> ((typep y 'double-float)
>>> (%double-float-atan! y (%make-dfloat)))
>>> ((typep y 'single-float)
>>> #+32-bit-target
>>> (%single-float-atan! y (%make-sfloat))
>>> #+64-bit-target
>>> (%single-float-atan y))
>>> ((typep y 'rational)
>>> (cond ((<= (abs y) most-positive-short-float)
>>> #+32-bit-target
>>> (target::with-stack-short-floats ((sy y))
>>> (%single-float-atan! sy (%make-sfloat)))
>>> #+64-bit-target
>>> (%single-float-atan (%short-float y)))
>>> ((minusp y)
>>> #.(- single-float-half-pi))
>>> (t
>>> single-float-half-pi)))
>>> (t
>>> (let ((r (realpart y))
>>> (i (imagpart y)))
>>> (if (zerop i)
>>> (complex (atan r) i)
>>> (i* (%complex-atanh (complex (- i) r)) -1))))))
>>>
>>>
>>> _______________________________________________
>>> Openmcl-devel mailing list
>>> Openmcl-devel at clozure.com <mailto:Openmcl-devel at clozure.com>
>>> https://lists.clozure.com/mailman/listinfo/openmcl-devel <https://lists.clozure.com/mailman/listinfo/openmcl-devel>
>>
>> _______________________________________________
>> Openmcl-devel mailing list
>> Openmcl-devel at clozure.com <mailto:Openmcl-devel at clozure.com>
>> https://lists.clozure.com/mailman/listinfo/openmcl-devel
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.clozure.com/pipermail/openmcl-devel/attachments/20200619/dd14c61b/attachment.htm>
More information about the Openmcl-devel
mailing list