<html><head></head><body><div class="ydp675aaf1eyahoo-style-wrap" style="font-family:Helvetica Neue, Helvetica, Arial, sans-serif;font-size:13px;"><div></div>
<div dir="ltr" data-setdir="false">Martin, did you do any benchmarks on the difference between, say, CFFI and using the internal code as a template?<br></div><div><br></div>
</div><div id="yahoo_quoted_3120179272" class="yahoo_quoted">
<div style="font-family:'Helvetica Neue', Helvetica, Arial, sans-serif;font-size:13px;color:#26282a;">
<div>
On Friday, June 19, 2020, 8:32:22 AM GMT+8, martin <brooks.martin@sympatico.ca> wrote:
</div>
<div><br></div>
<div><br></div>
<div><div id="yiv4860178186"><div>Hello CCL fans, I will relate a somewhat parallel story.<div class="yiv4860178186">My application requires exact arithmetic, i.e. bignum rationals.</div><div class="yiv4860178186">CCL 11 conses like crazy for bignum rationals, whereas CCL 12 has a specialized implementation.</div><div class="yiv4860178186">This happened because I engaged Clozure Associates to solve my bignum rational problem, which they totally did.</div><div class="yiv4860178186">Prior to engaging CA I raided the CCL code to use the low-level allocation and arithmetic %functions, as in the examples below.</div><div class="yiv4860178186"><br class="yiv4860178186" clear="none"></div><div class="yiv4860178186">Martin</div><div class="yiv4860178186"><br class="yiv4860178186" clear="none"></div><div class="yiv4860178186"><div class="yiv4860178186"><div><br class="yiv4860178186" clear="none"><blockquote class="yiv4860178186" type="cite"><div class="yiv4860178186yqt9195486361" id="yiv4860178186yqt26868"><div class="yiv4860178186">On Jun 18, 2020, at 7:56 PM, Steven Nunez <<a rel="nofollow" shape="rect" class="yiv4860178186" ymailto="mailto:steve_nunez@yahoo.com" target="_blank" href="mailto:steve_nunez@yahoo.com">steve_nunez@yahoo.com</a>> wrote:</div><br class="yiv4860178186Apple-interchange-newline" clear="none"><div class="yiv4860178186"><div class="yiv4860178186"><div class="yiv4860178186ydpde0f8651yahoo-style-wrap" style="font-family:Helvetica Neue, Helvetica, Arial, sans-serif;font-size:13px;"><div class="yiv4860178186"></div>
<div class="yiv4860178186" dir="ltr"><div class="yiv4860178186"><div class="yiv4860178186" dir="ltr">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.<br class="yiv4860178186" clear="none"></div><div class="yiv4860178186"><br class="yiv4860178186" clear="none"><br class="yiv4860178186" clear="none"></div></div><div class="yiv4860178186"><br class="yiv4860178186" clear="none"></div></div><div class="yiv4860178186"><br class="yiv4860178186" clear="none"></div>
</div><div class="yiv4860178186yahoo_quoted" id="yiv4860178186yahoo_quoted_3041486737">
<div class="yiv4860178186" style="font-family:'Helvetica Neue', Helvetica, Arial, sans-serif;font-size:13px;color:#26282a;">
<div class="yiv4860178186">
On Friday, June 19, 2020, 5:01:20 AM GMT+8, Ron Garret <<a rel="nofollow" shape="rect" class="yiv4860178186" ymailto="mailto:ron@flownet.com" target="_blank" href="mailto:ron@flownet.com">ron@flownet.com</a>> wrote:
</div>
<div class="yiv4860178186"><br class="yiv4860178186" clear="none"></div>
<div class="yiv4860178186"><br class="yiv4860178186" clear="none"></div>
<div class="yiv4860178186"><div class="yiv4860178186" id="yiv4860178186"><div class="yiv4860178186">The best way to answer performance questions is just to run some benchmarks.<div class="yiv4860178186"><br class="yiv4860178186" clear="none"><div class="yiv4860178186"><div class="yiv4860178186yqt1136114722" id="yiv4860178186yqt44587"><div class="yiv4860178186">On Jun 18, 2020, at 1:58 AM, Steven Nunez <<a rel="nofollow" shape="rect" class="yiv4860178186" ymailto="mailto:steve_nunez@yahoo.com" target="_blank" href="mailto:steve_nunez@yahoo.com">steve_nunez@yahoo.com</a>> wrote:</div><br class="yiv4860178186Apple-interchange-newline" clear="none"><blockquote class="yiv4860178186" type="cite"><div class="yiv4860178186"><div class="yiv4860178186yahoo-style-wrap" style="font-family:Helvetica Neue, Helvetica, Arial, sans-serif;font-size:13px;"><div class="yiv4860178186" dir="ltr">Greetings all,</div><div class="yiv4860178186" dir="ltr"><br class="yiv4860178186" clear="none"></div><div class="yiv4860178186" dir="ltr">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 <font class="yiv4860178186" face="lucida console, sans-serif">atanh</font>, 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.</div><div class="yiv4860178186" dir="ltr"><br class="yiv4860178186" clear="none"></div><div class="yiv4860178186" dir="ltr">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.</div><div class="yiv4860178186" dir="ltr"><br class="yiv4860178186" clear="none"></div><div class="yiv4860178186" dir="ltr"><div class="yiv4860178186"><font class="yiv4860178186" face="lucida console, sans-serif"> (ff:def-foreign-call (%erf "erf") ((x :double))<br class="yiv4860178186" clear="none"> :returning :double)<br class="yiv4860178186" clear="none"> (ff:def-foreign-call (%erff "erff") ((x :float))<br class="yiv4860178186" clear="none"> :returning :float)<br class="yiv4860178186" clear="none"> (defun erf (x)<br class="yiv4860178186" clear="none"> "Returns the error function value for x"<br class="yiv4860178186" clear="none"> (cond ((typep x 'double-float) (%erf x))<br class="yiv4860178186" clear="none"> ((typep x 'single-float) (%erff x))))</font></div><div class="yiv4860178186"><br class="yiv4860178186" clear="none"></div></div><div class="yiv4860178186" dir="ltr"><br class="yiv4860178186" clear="none"></div><div class="yiv4860178186" dir="ltr">Whereas in the CCL code there's <font class="yiv4860178186" face="lucida console, sans-serif">%setf-double-float</font>, and <font class="yiv4860178186" face="lucida console, sans-serif">target::with-stack-double-floats</font>, 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.<br class="yiv4860178186" clear="none"></div><div class="yiv4860178186" dir="ltr"><br class="yiv4860178186" clear="none"></div><div class="yiv4860178186" dir="ltr">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.<br class="yiv4860178186" clear="none"></div><div class="yiv4860178186" dir="ltr"><br class="yiv4860178186" clear="none"></div><div class="yiv4860178186" dir="ltr"><font class="yiv4860178186" face="lucida console, sans-serif"><br class="yiv4860178186" clear="none"></font></div><div class="yiv4860178186" dir="ltr"><div class="yiv4860178186"><font class="yiv4860178186" face="lucida console, sans-serif">#+windows-target<br class="yiv4860178186" clear="none">(progn<br class="yiv4860178186" clear="none">(defun %double-float-atanh! (n result)<br class="yiv4860178186" clear="none"> (declare (double-float n result))<br class="yiv4860178186" clear="none"> (with-stack-double-floats ((temp))<br class="yiv4860178186" clear="none"> (%setf-double-float TEMP (external-call "atanh" :double n :double))<br class="yiv4860178186" clear="none"> (%df-check-exception-1 'atanh n (%ffi-exception-status))<br class="yiv4860178186" clear="none"> (%setf-double-float result TEMP)))<br class="yiv4860178186" clear="none"><br class="yiv4860178186" clear="none">#+32-bit-target<br class="yiv4860178186" clear="none">(defun %single-float-atanh! (n result)<br class="yiv4860178186" clear="none"> (declare (single-float n result)) <br class="yiv4860178186" clear="none"> (target::with-stack-short-floats ((temp))<br class="yiv4860178186" clear="none"> #+arm-target (%set-fpscr-status 0)<br class="yiv4860178186" clear="none"> (%setf-short-float TEMP (external-call "atanhf" :float n :float))<br class="yiv4860178186" clear="none"> (%sf-check-exception-1 'atanh n (%ffi-exception-status))<br class="yiv4860178186" clear="none"> (%setf-short-float result TEMP)))<br class="yiv4860178186" clear="none"><br class="yiv4860178186" clear="none">#+64-bit-target<br class="yiv4860178186" clear="none">(defun %single-float-atanh (n)<br class="yiv4860178186" clear="none"> (declare (single-float n)) <br class="yiv4860178186" clear="none"> (let* ((result (external-call "atanhf" :float n :float)))<br class="yiv4860178186" clear="none"> (%sf-check-exception-1 'atanh n (%ffi-exception-status))<br class="yiv4860178186" clear="none"></font><div class="yiv4860178186"><font class="yiv4860178186" face="lucida console, sans-serif"> result))</font><font class="yiv4860178186" face="lucida console, sans-serif">)</font></div><div class="yiv4860178186"><font class="yiv4860178186" face="lucida console, sans-serif"><br class="yiv4860178186" clear="none"></font></div><div class="yiv4860178186" dir="ltr"><div class="yiv4860178186"><font class="yiv4860178186" face="lucida console, sans-serif">(defun atan (y &optional (x nil x-p))<br class="yiv4860178186" clear="none"> "Return the arc tangent of Y if X is omitted or Y/X if X is supplied."<br class="yiv4860178186" clear="none"> (cond (x-p<br class="yiv4860178186" clear="none"> (cond ((or (typep x 'double-float)<br class="yiv4860178186" clear="none"> (typep y 'double-float))<br class="yiv4860178186" clear="none"> (with-stack-double-floats ((dy y)<br class="yiv4860178186" clear="none"> (dx x))<br class="yiv4860178186" clear="none"> (%df-atan2 dy dx)))<br class="yiv4860178186" clear="none"> (t<br class="yiv4860178186" clear="none"> (when (and (rationalp x) (rationalp y))<br class="yiv4860178186" clear="none"> ;; rescale arguments so that the maximum absolute value is 1<br class="yiv4860178186" clear="none"> (let ((x1 (abs x)) (y1 (abs y)))<br class="yiv4860178186" clear="none"> (cond ((> y1 x1)<br class="yiv4860178186" clear="none"> (setf x (/ x y1))<br class="yiv4860178186" clear="none"> (setf y (signum y)))<br class="yiv4860178186" clear="none"> ((not (zerop x))<br class="yiv4860178186" clear="none"> (setf y (/ y x1))<br class="yiv4860178186" clear="none"> (setf x (signum x))))))<br class="yiv4860178186" clear="none"> #+32-bit-target<br class="yiv4860178186" clear="none"> (target::with-stack-short-floats ((sy y)<br class="yiv4860178186" clear="none"> (sx x))<br class="yiv4860178186" clear="none"> (%sf-atan2! sy sx))<br class="yiv4860178186" clear="none"> #+64-bit-target<br class="yiv4860178186" clear="none"> (%sf-atan2 (%short-float y) (%short-float x)))))<br class="yiv4860178186" clear="none"> ((typep y 'double-float)<br class="yiv4860178186" clear="none"> (%double-float-atan! y (%make-dfloat)))<br class="yiv4860178186" clear="none"> ((typep y 'single-float)<br class="yiv4860178186" clear="none"> #+32-bit-target<br class="yiv4860178186" clear="none"> (%single-float-atan! y (%make-sfloat))<br class="yiv4860178186" clear="none"> #+64-bit-target<br class="yiv4860178186" clear="none"> (%single-float-atan y))<br class="yiv4860178186" clear="none"> ((typep y 'rational)<br class="yiv4860178186" clear="none"> (cond ((<= (abs y) most-positive-short-float)<br class="yiv4860178186" clear="none"> #+32-bit-target<br class="yiv4860178186" clear="none"> (target::with-stack-short-floats ((sy y))<br class="yiv4860178186" clear="none"> (%single-float-atan! sy (%make-sfloat)))<br class="yiv4860178186" clear="none"> #+64-bit-target<br class="yiv4860178186" clear="none"> (%single-float-atan (%short-float y)))<br class="yiv4860178186" clear="none"> ((minusp y)<br class="yiv4860178186" clear="none"> #.(- single-float-half-pi))<br class="yiv4860178186" clear="none"> (t<br class="yiv4860178186" clear="none"> single-float-half-pi)))<br class="yiv4860178186" clear="none"> (t<br class="yiv4860178186" clear="none"> (let ((r (realpart y))<br class="yiv4860178186" clear="none"> (i (imagpart y)))<br class="yiv4860178186" clear="none"> (if (zerop i)<br class="yiv4860178186" clear="none"> (complex (atan r) i)<br class="yiv4860178186" clear="none"> (i* (%complex-atanh (complex (- i) r)) -1))))))</font></div><div class="yiv4860178186"><br class="yiv4860178186" clear="none"></div></div></div><div class="yiv4860178186"><br class="yiv4860178186" clear="none"></div></div></div></div>_______________________________________________<br class="yiv4860178186" clear="none">Openmcl-devel mailing list<br class="yiv4860178186" clear="none"><a rel="nofollow" shape="rect" class="yiv4860178186" ymailto="mailto:Openmcl-devel@clozure.com" target="_blank" href="mailto:Openmcl-devel@clozure.com">Openmcl-devel@clozure.com</a><br class="yiv4860178186" clear="none"><a rel="nofollow" shape="rect" class="yiv4860178186" target="_blank" href="https://lists.clozure.com/mailman/listinfo/openmcl-devel">https://lists.clozure.com/mailman/listinfo/openmcl-devel</a><br class="yiv4860178186" clear="none"></blockquote></div></div><br class="yiv4860178186" clear="none"></div></div></div></div>
</div>
</div></div>_______________________________________________<br class="yiv4860178186" clear="none">Openmcl-devel mailing list<br class="yiv4860178186" clear="none"><a rel="nofollow" shape="rect" class="yiv4860178186" ymailto="mailto:Openmcl-devel@clozure.com" target="_blank" href="mailto:Openmcl-devel@clozure.com">Openmcl-devel@clozure.com</a><br class="yiv4860178186" clear="none">https://lists.clozure.com/mailman/listinfo/openmcl-devel<br class="yiv4860178186" clear="none"></div></div></blockquote></div><br class="yiv4860178186" clear="none"></div></div></div></div></div>
</div>
</div></body></html>