Hello. Le mer. 18 août 2021 à 09:44, Erik Svensson <erik.svens...@nasdaq.com> a écrit : > > Hi all! > > > > To recap, we use Apache-commons math extensively and we would like to improve > performance by improving performance in the FastMath (now named AccurateMath) > class by using > > Java.lang.Math methods where those are annotated with the > HotSpotIntrinsicCandidate/IntrinsicCandidate annotation. > > One problem is that FM/AM is not a drop-in replacement for jlM as of java 9. > I could add the methods needed easily (if we ignore the performance penalty > for the FMA operation when using BigDecimal) so I could get on with the > actual performance work (NB that FMA is marked as IntrinsicCandidate and that > modern intel cpu:s have integral support for FMA). > > If FM/AM is NOT to be a drop-in replacement for jlM, then I can get on with > finding a solution that enables users to select a math provider for the math > lib, thus being able to select jlM if the additional performance is desired.
I never quite understood the qualification of "drop-in" replacement as switching from "java.lang.Math" to "Fastmath" requires source change and recompilation. A runtime switch would be an interesting feature, but it might have a large adverse effect on performance... > > On another note, FM/AM is advertised as being more accurate than jlM. A unit test[1] compares results from Fastmath with computations that use 40 decimal digits (see "Dfp" class[2]). It ensures that FastMath is within 0.5 ULPs of the high-precision value but does not check whether "Math" is also within that boundary. > Anybody knows if this applies to methods using intrinsics as well? You are welcome to provide benchmark code (to be located in the "commons-math-examples" module where each "sub-module" can require a different version of the JDK), that would give up-to-date answers to those questions with different combinations of JDK version (with/without the annotation): * Accuracy (perhaps using "BigDecimal" instead of "Dfp") * Speed (using JMH) This would be crucial information for selecting the way forward (including whether "AccurateMath" support should be dropped entirely!). Regards, Gilles [1] https://gitbox.apache.org/repos/asf?p=commons-math.git;a=blob;f=commons-math-legacy-core/src/test/java/org/apache/commons/math4/legacy/core/jdkmath/AccurateMathTest.java;h=6cddff2306a9f429c6f8f58d2827bc4a1f35911d;hb=HEAD [2] https://gitbox.apache.org/repos/asf?p=commons-math.git;a=blob;f=commons-math-legacy-core/src/main/java/org/apache/commons/math4/legacy/core/dfp/Dfp.java;h=41f15c91151b16b8005da376e9c3aff54229b967;hb=HEAD --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
