code review request for exp and pow additions
Berg, Michael C
michael.c.berg at intel.com
Wed Jul 20 23:41:16 UTC 2016
There is a header section above each intrinsic that goes into detail about expected behavior. For the brevity of this email thread, I would advise reading there.
The short answer is that each is a mathematical expansion that is equated to x87 answers for the intrinsics that map there (log, log10, sin, cos, tan) as well as the strict math result which includes exp and pow.
For all we compare against strict math results, and with our tests we are not seeing divergence at this time. We have extensive testing especially for sin and cos for large host of corner cases for which we pass. Graal also has testing for each of these math functions in the unit tests. Basically we are providing the same results that JDK9 does when these intrinsics are utilized in either location.
From: graal-dev [mailto:graal-dev-bounces at openjdk.java.net] On Behalf Of Raffaello Giulietti
Sent: Wednesday, July 20, 2016 2:57 PM
To: graal-dev at openjdk.java.net
Subject: Re: code review request for exp and pow additions
just curious about rounding guarantees of the intrinsics on the x86.
On 2016-07-20 20:05, Berg, Michael C wrote:
> Subject: code review request for exp and pow intrinsic additions for x86
> Folks, the code is available as:
> For review.
> This change stages the first installment of intel libm intrinsics (exp, pow).
> For exp the uplift is 4x, the pow the uplift is 2.6x.
More information about the graal-dev