<html><head><meta http-equiv="Content-Type" content="text/html charset=utf-8"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class=""><br class=""><div><blockquote type="cite" class=""><div class="">On Jan 12, 2017, at 11:04 AM, Steve (Numerics) Canon <<a href="mailto:scanon@apple.com" class="">scanon@apple.com</a>> wrote:</div><br class="Apple-interchange-newline"><div class=""><blockquote type="cite" style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-size-adjust: auto; -webkit-text-stroke-width: 0px;" class=""><br class="Apple-interchange-newline">On Jan 12, 2017, at 12:58 PM, Friedman, Eli via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" class="">llvm-dev@lists.llvm.org</a>> wrote:<br class=""><br class="">On 1/12/2017 9:33 AM, Mehdi Amini via llvm-dev wrote:<br class=""><blockquote type="cite" class=""><blockquote type="cite" class="">On Jan 12, 2017, at 5:03 AM, Antoine Pitrou via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" class="">llvm-dev@lists.llvm.org</a>> wrote:<br class=""><br class="">On Mon, 9 Jan 2017 11:43:17 -0600<br class="">Wei Ding via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" class="">llvm-dev@lists.llvm.org</a>> wrote:<br class=""><blockquote type="cite" class="">Hi,<br class=""><br class="">I want an efficient way to implement function pow in LLVM instead of<br class="">invoking pow() math built-in. For algorithm part, I am clear for the logic.<br class="">But I am not quite sure for which parts of LLVM should I replace built-in<br class="">pow with another efficient pow implementation. Any comments and feedback<br class="">are appreciated. Thanks!<br class=""></blockquote>In Numba, we have decided to optimize some usages of the power function<br class="">in our own front-end, so that LLVM IR gets an already optimized form,<br class="">as we have found that otherwise LLVM may miss some optimization<br class="">opportunities. YMMV.<br class=""></blockquote>It seems to me that it would be more interesting to gather these misoptimization and fix LLVM to catch them.<br class=""><br class=""><blockquote type="cite" class="">(e.g. we detect that the exponent is a compile-time constant and<br class="">transform `x**3` into `x*x*x`)<br class=""></blockquote>This seems definitely in the scope of what LLVM could do, potentially with TTI.<br class=""></blockquote><br class="">LLVM already does this... but only if the pow() call is marked "fast". IEEE 754 pow() is supposed to be correctly rounded, but (x*x)*x has an extra rounding step.<br class=""></blockquote><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""><span style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; float: none; display: inline !important;" class="">pow( ) is not supposed to be correctly rounded.</span><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""><span style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; float: none; display: inline !important;" class="">IEEE 754 recommends that a correctly rounded power function exist [9.2], but does not recommend or require that this be the default pow( ) function. [Note: it was not actually shown until quite late in the IEEE 754 revision process that it was even possible to have a reasonably efficient correctly-rounded pow( ) function, and such a function is at minimum an order of magnitude slower than a good sub-ulp-accurate-but-not-correctly-rounded implementation. Most 754 committee members would recommend that the default pow( ) function *not* be correctly rounded.]</span><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""></div></blockquote><div><br class=""></div><div>So we should use x*x*x even without fast-math...</div><br class=""><blockquote type="cite" class=""><div class=""><span style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; float: none; display: inline !important;" class="">In a good math library, however, pow( ) should absolutely be sub-ulp accurate, which means that it should *not* be implemented via log( ) and exp( ), and indeed, most math libraries don’t do that.</span><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""><span style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; float: none; display: inline !important;" class="">Just to provide some example data, for single-precision x in [1,2) on current OS X:</span><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""><span class="Apple-tab-span" style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: pre; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;"> </span><span style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; float: none; display: inline !important;" class="">The worst-case error of x*x*x is 1.28736 ulp</span><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""><span class="Apple-tab-span" style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: pre; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;"> </span><span style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; float: none; display: inline !important;" class="">The worst-case error of powf(x, 3) is 0.500013 ulp</span><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""><span class="Apple-tab-span" style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: pre; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;"> </span><span style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; float: none; display: inline !important;" class="">The RMS error of x*x*x is 0.585066 ulp</span><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""><span class="Apple-tab-span" style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: pre; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;"> </span><span style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; float: none; display: inline !important;" class="">The RMS error of powf(x,3) is 0.499984 ulp</span><br style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class=""></div></blockquote><div><br class=""></div><div>… or maybe not! :)</div><div><br class=""></div><div>I’m not sure what to think of these results. For instance what’s the perf impact of calling into powf(x, 3) on OSX vs using x*x*x?</div><div>At the source level, what could the user do to decide that 1.28ULP is acceptable for his use-case?</div><div><br class=""></div><div>— </div><div>Mehdi</div></div></body></html>